libpcsxcore/new_dynarec/new_dynarec.c

   1 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
   2  *   Mupen64plus - new_dynarec.c                                           *
   3  *   Copyright (C) 2009-2011 Ari64                                         *
   4  *                                                                         *
   5  *   This program is free software; you can redistribute it and/or modify  *
   6  *   it under the terms of the GNU General Public License as published by  *
   7  *   the Free Software Foundation; either version 2 of the License, or     *
   8  *   (at your option) any later version.                                   *
   9  *                                                                         *
  10  *   This program is distributed in the hope that it will be useful,       *
  11  *   but WITHOUT ANY WARRANTY; without even the implied warranty of        *
  12  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the         *
  13  *   GNU General Public License for more details.                          *
  14  *                                                                         *
  15  *   You should have received a copy of the GNU General Public License     *
  16  *   along with this program; if not, write to the                         *
  17  *   Free Software Foundation, Inc.,                                       *
  18  *   51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.          *
  19  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
  20
  21 #include <stdlib.h>
  22 #include <stdint.h> //include for uint64_t
  23 #include <assert.h>
  24
  25 #include "emu_if.h" //emulator interface
  26
  27 #include <sys/mman.h>
  28
  29 #ifdef __i386__
  30 #include "assem_x86.h"
  31 #endif
  32 #ifdef __x86_64__
  33 #include "assem_x64.h"
  34 #endif
  35 #ifdef __arm__
  36 #include "assem_arm.h"
  37 #endif
  38
  39 #define MAXBLOCK 4096
  40 #define MAX_OUTPUT_BLOCK_SIZE 262144
  41 #define CLOCK_DIVIDER 2
  42
  43 struct regstat
  44 {
  45   signed char regmap_entry[HOST_REGS];
  46   signed char regmap[HOST_REGS];
  47   uint64_t was32;
  48   uint64_t is32;
  49   uint64_t wasdirty;
  50   uint64_t dirty;
  51   uint64_t u;
  52   uint64_t uu;
  53   u_int wasconst;
  54   u_int isconst;
  55   uint64_t constmap[HOST_REGS];
  56 };
  57
  58 struct ll_entry
  59 {
  60   u_int vaddr;
  61   u_int reg32;
  62   void *addr;
  63   struct ll_entry *next;
  64 };
  65
  66   u_int start;
  67   u_int *source;
  68   u_int pagelimit;
  69   char insn[MAXBLOCK][10];
  70   u_char itype[MAXBLOCK];
  71   u_char opcode[MAXBLOCK];
  72   u_char opcode2[MAXBLOCK];
  73   u_char bt[MAXBLOCK];
  74   u_char rs1[MAXBLOCK];
  75   u_char rs2[MAXBLOCK];
  76   u_char rt1[MAXBLOCK];
  77   u_char rt2[MAXBLOCK];
  78   u_char us1[MAXBLOCK];
  79   u_char us2[MAXBLOCK];
  80   u_char dep1[MAXBLOCK];
  81   u_char dep2[MAXBLOCK];
  82   u_char lt1[MAXBLOCK];
  83   int imm[MAXBLOCK];
  84   u_int ba[MAXBLOCK];
  85   char likely[MAXBLOCK];
  86   char is_ds[MAXBLOCK];
  87   char ooo[MAXBLOCK];
  88   uint64_t unneeded_reg[MAXBLOCK];
  89   uint64_t unneeded_reg_upper[MAXBLOCK];
  90   uint64_t branch_unneeded_reg[MAXBLOCK];
  91   uint64_t branch_unneeded_reg_upper[MAXBLOCK];
  92   uint64_t p32[MAXBLOCK];
  93   uint64_t pr32[MAXBLOCK];
  94   signed char regmap_pre[MAXBLOCK][HOST_REGS];
  95   signed char regmap[MAXBLOCK][HOST_REGS];
  96   signed char regmap_entry[MAXBLOCK][HOST_REGS];
  97   uint64_t constmap[MAXBLOCK][HOST_REGS];
  98   struct regstat regs[MAXBLOCK];
  99   struct regstat branch_regs[MAXBLOCK];
 100   signed char minimum_free_regs[MAXBLOCK];
 101   u_int needed_reg[MAXBLOCK];
 102   uint64_t requires_32bit[MAXBLOCK];
 103   u_int wont_dirty[MAXBLOCK];
 104   u_int will_dirty[MAXBLOCK];
 105   int ccadj[MAXBLOCK];
 106   int slen;
 107   u_int instr_addr[MAXBLOCK];
 108   u_int link_addr[MAXBLOCK][3];
 109   int linkcount;
 110   u_int stubs[MAXBLOCK*3][8];
 111   int stubcount;
 112   u_int literals[1024][2];
 113   int literalcount;
 114   int is_delayslot;
 115   int cop1_usable;
 116   u_char *out;
 117   struct ll_entry *jump_in[4096];
 118   struct ll_entry *jump_out[4096];
 119   struct ll_entry *jump_dirty[4096];
 120   u_int hash_table[65536][4]  __attribute__((aligned(16)));
 121   char shadow[1048576]  __attribute__((aligned(16)));
 122   void *copy;
 123   int expirep;
 124 #ifndef PCSX
 125   u_int using_tlb;
 126 #else
 127   static const u_int using_tlb=0;
 128 #endif
 129   static u_int sp_in_mirror;
 130   u_int stop_after_jal;
 131   extern u_char restore_candidate[512];
 132   extern int cycle_count;
 133
 134   /* registers that may be allocated */
 135   /* 1-31 gpr */
 136 #define HIREG 32 // hi
 137 #define LOREG 33 // lo
 138 #define FSREG 34 // FPU status (FCSR)
 139 #define CSREG 35 // Coprocessor status
 140 #define CCREG 36 // Cycle count
 141 #define INVCP 37 // Pointer to invalid_code
 142 #define MMREG 38 // Pointer to memory_map
 143 #define ROREG 39 // ram offset (if rdram!=0x80000000)
 144 #define TEMPREG 40
 145 #define FTEMP 40 // FPU temporary register
 146 #define PTEMP 41 // Prefetch temporary register
 147 #define TLREG 42 // TLB mapping offset
 148 #define RHASH 43 // Return address hash
 149 #define RHTBL 44 // Return address hash table address
 150 #define RTEMP 45 // JR/JALR address register
 151 #define MAXREG 45
 152 #define AGEN1 46 // Address generation temporary register
 153 #define AGEN2 47 // Address generation temporary register
 154 #define MGEN1 48 // Maptable address generation temporary register
 155 #define MGEN2 49 // Maptable address generation temporary register
 156 #define BTREG 50 // Branch target temporary register
 157
 158   /* instruction types */
 159 #define NOP 0     // No operation
 160 #define LOAD 1    // Load
 161 #define STORE 2   // Store
 162 #define LOADLR 3  // Unaligned load
 163 #define STORELR 4 // Unaligned store
 164 #define MOV 5     // Move
 165 #define ALU 6     // Arithmetic/logic
 166 #define MULTDIV 7 // Multiply/divide
 167 #define SHIFT 8   // Shift by register
 168 #define SHIFTIMM 9// Shift by immediate
 169 #define IMM16 10  // 16-bit immediate
 170 #define RJUMP 11  // Unconditional jump to register
 171 #define UJUMP 12  // Unconditional jump
 172 #define CJUMP 13  // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
 173 #define SJUMP 14  // Conditional branch (regimm format)
 174 #define COP0 15   // Coprocessor 0
 175 #define COP1 16   // Coprocessor 1
 176 #define C1LS 17   // Coprocessor 1 load/store
 177 #define FJUMP 18  // Conditional branch (floating point)
 178 #define FLOAT 19  // Floating point unit
 179 #define FCONV 20  // Convert integer to float
 180 #define FCOMP 21  // Floating point compare (sets FSREG)
 181 #define SYSCALL 22// SYSCALL
 182 #define OTHER 23  // Other
 183 #define SPAN 24   // Branch/delay slot spans 2 pages
 184 #define NI 25     // Not implemented
 185 #define HLECALL 26// PCSX fake opcodes for HLE
 186 #define COP2 27   // Coprocessor 2 move
 187 #define C2LS 28   // Coprocessor 2 load/store
 188 #define C2OP 29   // Coprocessor 2 operation
 189 #define INTCALL 30// Call interpreter to handle rare corner cases
 190
 191   /* stubs */
 192 #define CC_STUB 1
 193 #define FP_STUB 2
 194 #define LOADB_STUB 3
 195 #define LOADH_STUB 4
 196 #define LOADW_STUB 5
 197 #define LOADD_STUB 6
 198 #define LOADBU_STUB 7
 199 #define LOADHU_STUB 8
 200 #define STOREB_STUB 9
 201 #define STOREH_STUB 10
 202 #define STOREW_STUB 11
 203 #define STORED_STUB 12
 204 #define STORELR_STUB 13
 205 #define INVCODE_STUB 14
 206
 207   /* branch codes */
 208 #define TAKEN 1
 209 #define NOTTAKEN 2
 210 #define NULLDS 3
 211
 212 // asm linkage
 213 int new_recompile_block(int addr);
 214 void *get_addr_ht(u_int vaddr);
 215 void invalidate_block(u_int block);
 216 void invalidate_addr(u_int addr);
 217 void remove_hash(int vaddr);
 218 void jump_vaddr();
 219 void dyna_linker();
 220 void dyna_linker_ds();
 221 void verify_code();
 222 void verify_code_vm();
 223 void verify_code_ds();
 224 void cc_interrupt();
 225 void fp_exception();
 226 void fp_exception_ds();
 227 void jump_syscall();
 228 void jump_syscall_hle();
 229 void jump_eret();
 230 void jump_hlecall();
 231 void jump_intcall();
 232 void new_dyna_leave();
 233
 234 // TLB
 235 void TLBWI_new();
 236 void TLBWR_new();
 237 void read_nomem_new();
 238 void read_nomemb_new();
 239 void read_nomemh_new();
 240 void read_nomemd_new();
 241 void write_nomem_new();
 242 void write_nomemb_new();
 243 void write_nomemh_new();
 244 void write_nomemd_new();
 245 void write_rdram_new();
 246 void write_rdramb_new();
 247 void write_rdramh_new();
 248 void write_rdramd_new();
 249 extern u_int memory_map[1048576];
 250
 251 // Needed by assembler
 252 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32);
 253 void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty);
 254 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr);
 255 void load_all_regs(signed char i_regmap[]);
 256 void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
 257 void load_regs_entry(int t);
 258 void load_all_consts(signed char regmap[],int is32,u_int dirty,int i);
 259
 260 int tracedebug=0;
 261
 262 //#define DEBUG_CYCLE_COUNT 1
 263
 264 void nullf() {}
 265 //#define assem_debug printf
 266 //#define inv_debug printf
 267 #define assem_debug nullf
 268 #define inv_debug nullf
 269
 270 static void tlb_hacks()
 271 {
 272 #ifndef DISABLE_TLB
 273   // Goldeneye hack
 274   if (strncmp((char *) ROM_HEADER->nom, "GOLDENEYE",9) == 0)
 275   {
 276     u_int addr;
 277     int n;
 278     switch (ROM_HEADER->Country_code&0xFF)
 279     {
 280       case 0x45: // U
 281         addr=0x34b30;
 282         break;
 283       case 0x4A: // J
 284         addr=0x34b70;
 285         break;
 286       case 0x50: // E
 287         addr=0x329f0;
 288         break;
 289       default:
 290         // Unknown country code
 291         addr=0;
 292         break;
 293     }
 294     u_int rom_addr=(u_int)rom;
 295     #ifdef ROM_COPY
 296     // Since memory_map is 32-bit, on 64-bit systems the rom needs to be
 297     // in the lower 4G of memory to use this hack.  Copy it if necessary.
 298     if((void *)rom>(void *)0xffffffff) {
 299       munmap(ROM_COPY, 67108864);
 300       if(mmap(ROM_COPY, 12582912,
 301               PROT_READ | PROT_WRITE,
 302               MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
 303               -1, 0) <= 0) {printf("mmap() failed\n");}
 304       memcpy(ROM_COPY,rom,12582912);
 305       rom_addr=(u_int)ROM_COPY;
 306     }
 307     #endif
 308     if(addr) {
 309       for(n=0x7F000;n<0x80000;n++) {
 310         memory_map[n]=(((u_int)(rom_addr+addr-0x7F000000))>>2)|0x40000000;
 311       }
 312     }
 313   }
 314 #endif
 315 }
 316
 317 static u_int get_page(u_int vaddr)
 318 {
 319 #ifndef PCSX
 320   u_int page=(vaddr^0x80000000)>>12;
 321 #else
 322   u_int page=vaddr&~0xe0000000;
 323   if (page < 0x1000000)
 324     page &= ~0x0e00000; // RAM mirrors
 325   page>>=12;
 326 #endif
 327 #ifndef DISABLE_TLB
 328   if(page>262143&&tlb_LUT_r[vaddr>>12]) page=(tlb_LUT_r[vaddr>>12]^0x80000000)>>12;
 329 #endif
 330   if(page>2048) page=2048+(page&2047);
 331   return page;
 332 }
 333
 334 static u_int get_vpage(u_int vaddr)
 335 {
 336   u_int vpage=(vaddr^0x80000000)>>12;
 337 #ifndef DISABLE_TLB
 338   if(vpage>262143&&tlb_LUT_r[vaddr>>12]) vpage&=2047; // jump_dirty uses a hash of the virtual address instead
 339 #endif
 340   if(vpage>2048) vpage=2048+(vpage&2047);
 341   return vpage;
 342 }
 343
 344 // Get address from virtual address
 345 // This is called from the recompiled JR/JALR instructions
 346 void *get_addr(u_int vaddr)
 347 {
 348   u_int page=get_page(vaddr);
 349   u_int vpage=get_vpage(vaddr);
 350   struct ll_entry *head;
 351   //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
 352   head=jump_in[page];
 353   while(head!=NULL) {
 354     if(head->vaddr==vaddr&&head->reg32==0) {
 355   //printf("TRACE: count=%d next=%d (get_addr match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 356       int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 357       ht_bin[3]=ht_bin[1];
 358       ht_bin[2]=ht_bin[0];
 359       ht_bin[1]=(int)head->addr;
 360       ht_bin[0]=vaddr;
 361       return head->addr;
 362     }
 363     head=head->next;
 364   }
 365   head=jump_dirty[vpage];
 366   while(head!=NULL) {
 367     if(head->vaddr==vaddr&&head->reg32==0) {
 368       //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 369       // Don't restore blocks which are about to expire from the cache
 370       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
 371       if(verify_dirty(head->addr)) {
 372         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
 373         invalid_code[vaddr>>12]=0;
 374         memory_map[vaddr>>12]|=0x40000000;
 375         if(vpage<2048) {
 376 #ifndef DISABLE_TLB
 377           if(tlb_LUT_r[vaddr>>12]) {
 378             invalid_code[tlb_LUT_r[vaddr>>12]>>12]=0;
 379             memory_map[tlb_LUT_r[vaddr>>12]>>12]|=0x40000000;
 380           }
 381 #endif
 382           restore_candidate[vpage>>3]|=1<<(vpage&7);
 383         }
 384         else restore_candidate[page>>3]|=1<<(page&7);
 385         int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 386         if(ht_bin[0]==vaddr) {
 387           ht_bin[1]=(int)head->addr; // Replace existing entry
 388         }
 389         else
 390         {
 391           ht_bin[3]=ht_bin[1];
 392           ht_bin[2]=ht_bin[0];
 393           ht_bin[1]=(int)head->addr;
 394           ht_bin[0]=vaddr;
 395         }
 396         return head->addr;
 397       }
 398     }
 399     head=head->next;
 400   }
 401   //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
 402   int r=new_recompile_block(vaddr);
 403   if(r==0) return get_addr(vaddr);
 404   // Execute in unmapped page, generate pagefault execption
 405   Status|=2;
 406   Cause=(vaddr<<31)|0x8;
 407   EPC=(vaddr&1)?vaddr-5:vaddr;
 408   BadVAddr=(vaddr&~1);
 409   Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
 410   EntryHi=BadVAddr&0xFFFFE000;
 411   return get_addr_ht(0x80000000);
 412 }
 413 // Look up address in hash table first
 414 void *get_addr_ht(u_int vaddr)
 415 {
 416   //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
 417   int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 418   if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
 419   if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
 420   return get_addr(vaddr);
 421 }
 422
 423 void *get_addr_32(u_int vaddr,u_int flags)
 424 {
 425 #ifdef FORCE32
 426   return get_addr(vaddr);
 427 #else
 428   //printf("TRACE: count=%d next=%d (get_addr_32 %x,flags %x)\n",Count,next_interupt,vaddr,flags);
 429   int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 430   if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
 431   if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
 432   u_int page=get_page(vaddr);
 433   u_int vpage=get_vpage(vaddr);
 434   struct ll_entry *head;
 435   head=jump_in[page];
 436   while(head!=NULL) {
 437     if(head->vaddr==vaddr&&(head->reg32&flags)==0) {
 438       //printf("TRACE: count=%d next=%d (get_addr_32 match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 439       if(head->reg32==0) {
 440         int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 441         if(ht_bin[0]==-1) {
 442           ht_bin[1]=(int)head->addr;
 443           ht_bin[0]=vaddr;
 444         }else if(ht_bin[2]==-1) {
 445           ht_bin[3]=(int)head->addr;
 446           ht_bin[2]=vaddr;
 447         }
 448         //ht_bin[3]=ht_bin[1];
 449         //ht_bin[2]=ht_bin[0];
 450         //ht_bin[1]=(int)head->addr;
 451         //ht_bin[0]=vaddr;
 452       }
 453       return head->addr;
 454     }
 455     head=head->next;
 456   }
 457   head=jump_dirty[vpage];
 458   while(head!=NULL) {
 459     if(head->vaddr==vaddr&&(head->reg32&flags)==0) {
 460       //printf("TRACE: count=%d next=%d (get_addr_32 match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 461       // Don't restore blocks which are about to expire from the cache
 462       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
 463       if(verify_dirty(head->addr)) {
 464         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
 465         invalid_code[vaddr>>12]=0;
 466         memory_map[vaddr>>12]|=0x40000000;
 467         if(vpage<2048) {
 468 #ifndef DISABLE_TLB
 469           if(tlb_LUT_r[vaddr>>12]) {
 470             invalid_code[tlb_LUT_r[vaddr>>12]>>12]=0;
 471             memory_map[tlb_LUT_r[vaddr>>12]>>12]|=0x40000000;
 472           }
 473 #endif
 474           restore_candidate[vpage>>3]|=1<<(vpage&7);
 475         }
 476         else restore_candidate[page>>3]|=1<<(page&7);
 477         if(head->reg32==0) {
 478           int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 479           if(ht_bin[0]==-1) {
 480             ht_bin[1]=(int)head->addr;
 481             ht_bin[0]=vaddr;
 482           }else if(ht_bin[2]==-1) {
 483             ht_bin[3]=(int)head->addr;
 484             ht_bin[2]=vaddr;
 485           }
 486           //ht_bin[3]=ht_bin[1];
 487           //ht_bin[2]=ht_bin[0];
 488           //ht_bin[1]=(int)head->addr;
 489           //ht_bin[0]=vaddr;
 490         }
 491         return head->addr;
 492       }
 493     }
 494     head=head->next;
 495   }
 496   //printf("TRACE: count=%d next=%d (get_addr_32 no-match %x,flags %x)\n",Count,next_interupt,vaddr,flags);
 497   int r=new_recompile_block(vaddr);
 498   if(r==0) return get_addr(vaddr);
 499   // Execute in unmapped page, generate pagefault execption
 500   Status|=2;
 501   Cause=(vaddr<<31)|0x8;
 502   EPC=(vaddr&1)?vaddr-5:vaddr;
 503   BadVAddr=(vaddr&~1);
 504   Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
 505   EntryHi=BadVAddr&0xFFFFE000;
 506   return get_addr_ht(0x80000000);
 507 #endif
 508 }
 509
 510 void clear_all_regs(signed char regmap[])
 511 {
 512   int hr;
 513   for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
 514 }
 515
 516 signed char get_reg(signed char regmap[],int r)
 517 {
 518   int hr;
 519   for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap[hr]==r) return hr;
 520   return -1;
 521 }
 522
 523 // Find a register that is available for two consecutive cycles
 524 signed char get_reg2(signed char regmap1[],signed char regmap2[],int r)
 525 {
 526   int hr;
 527   for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
 528   return -1;
 529 }
 530
 531 int count_free_regs(signed char regmap[])
 532 {
 533   int count=0;
 534   int hr;
 535   for(hr=0;hr<HOST_REGS;hr++)
 536   {
 537     if(hr!=EXCLUDE_REG) {
 538       if(regmap[hr]<0) count++;
 539     }
 540   }
 541   return count;
 542 }
 543
 544 void dirty_reg(struct regstat *cur,signed char reg)
 545 {
 546   int hr;
 547   if(!reg) return;
 548   for (hr=0;hr<HOST_REGS;hr++) {
 549     if((cur->regmap[hr]&63)==reg) {
 550       cur->dirty|=1<<hr;
 551     }
 552   }
 553 }
 554
 555 // If we dirty the lower half of a 64 bit register which is now being
 556 // sign-extended, we need to dump the upper half.
 557 // Note: Do this only after completion of the instruction, because
 558 // some instructions may need to read the full 64-bit value even if
 559 // overwriting it (eg SLTI, DSRA32).
 560 static void flush_dirty_uppers(struct regstat *cur)
 561 {
 562   int hr,reg;
 563   for (hr=0;hr<HOST_REGS;hr++) {
 564     if((cur->dirty>>hr)&1) {
 565       reg=cur->regmap[hr];
 566       if(reg>=64)
 567         if((cur->is32>>(reg&63))&1) cur->regmap[hr]=-1;
 568     }
 569   }
 570 }
 571
 572 void set_const(struct regstat *cur,signed char reg,uint64_t value)
 573 {
 574   int hr;
 575   if(!reg) return;
 576   for (hr=0;hr<HOST_REGS;hr++) {
 577     if(cur->regmap[hr]==reg) {
 578       cur->isconst|=1<<hr;
 579       cur->constmap[hr]=value;
 580     }
 581     else if((cur->regmap[hr]^64)==reg) {
 582       cur->isconst|=1<<hr;
 583       cur->constmap[hr]=value>>32;
 584     }
 585   }
 586 }
 587
 588 void clear_const(struct regstat *cur,signed char reg)
 589 {
 590   int hr;
 591   if(!reg) return;
 592   for (hr=0;hr<HOST_REGS;hr++) {
 593     if((cur->regmap[hr]&63)==reg) {
 594       cur->isconst&=~(1<<hr);
 595     }
 596   }
 597 }
 598
 599 int is_const(struct regstat *cur,signed char reg)
 600 {
 601   int hr;
 602   if(reg<0) return 0;
 603   if(!reg) return 1;
 604   for (hr=0;hr<HOST_REGS;hr++) {
 605     if((cur->regmap[hr]&63)==reg) {
 606       return (cur->isconst>>hr)&1;
 607     }
 608   }
 609   return 0;
 610 }
 611 uint64_t get_const(struct regstat *cur,signed char reg)
 612 {
 613   int hr;
 614   if(!reg) return 0;
 615   for (hr=0;hr<HOST_REGS;hr++) {
 616     if(cur->regmap[hr]==reg) {
 617       return cur->constmap[hr];
 618     }
 619   }
 620   printf("Unknown constant in r%d\n",reg);
 621   exit(1);
 622 }
 623
 624 // Least soon needed registers
 625 // Look at the next ten instructions and see which registers
 626 // will be used.  Try not to reallocate these.
 627 void lsn(u_char hsn[], int i, int *preferred_reg)
 628 {
 629   int j;
 630   int b=-1;
 631   for(j=0;j<9;j++)
 632   {
 633     if(i+j>=slen) {
 634       j=slen-i-1;
 635       break;
 636     }
 637     if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
 638     {
 639       // Don't go past an unconditonal jump
 640       j++;
 641       break;
 642     }
 643   }
 644   for(;j>=0;j--)
 645   {
 646     if(rs1[i+j]) hsn[rs1[i+j]]=j;
 647     if(rs2[i+j]) hsn[rs2[i+j]]=j;
 648     if(rt1[i+j]) hsn[rt1[i+j]]=j;
 649     if(rt2[i+j]) hsn[rt2[i+j]]=j;
 650     if(itype[i+j]==STORE || itype[i+j]==STORELR) {
 651       // Stores can allocate zero
 652       hsn[rs1[i+j]]=j;
 653       hsn[rs2[i+j]]=j;
 654     }
 655     // On some architectures stores need invc_ptr
 656     #if defined(HOST_IMM8)
 657     if(itype[i+j]==STORE || itype[i+j]==STORELR || (opcode[i+j]&0x3b)==0x39 || (opcode[i+j]&0x3b)==0x3a) {
 658       hsn[INVCP]=j;
 659     }
 660     #endif
 661     if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
 662     {
 663       hsn[CCREG]=j;
 664       b=j;
 665     }
 666   }
 667   if(b>=0)
 668   {
 669     if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
 670     {
 671       // Follow first branch
 672       int t=(ba[i+b]-start)>>2;
 673       j=7-b;if(t+j>=slen) j=slen-t-1;
 674       for(;j>=0;j--)
 675       {
 676         if(rs1[t+j]) if(hsn[rs1[t+j]]>j+b+2) hsn[rs1[t+j]]=j+b+2;
 677         if(rs2[t+j]) if(hsn[rs2[t+j]]>j+b+2) hsn[rs2[t+j]]=j+b+2;
 678         //if(rt1[t+j]) if(hsn[rt1[t+j]]>j+b+2) hsn[rt1[t+j]]=j+b+2;
 679         //if(rt2[t+j]) if(hsn[rt2[t+j]]>j+b+2) hsn[rt2[t+j]]=j+b+2;
 680       }
 681     }
 682     // TODO: preferred register based on backward branch
 683   }
 684   // Delay slot should preferably not overwrite branch conditions or cycle count
 685   if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
 686     if(rs1[i-1]) if(hsn[rs1[i-1]]>1) hsn[rs1[i-1]]=1;
 687     if(rs2[i-1]) if(hsn[rs2[i-1]]>1) hsn[rs2[i-1]]=1;
 688     hsn[CCREG]=1;
 689     // ...or hash tables
 690     hsn[RHASH]=1;
 691     hsn[RHTBL]=1;
 692   }
 693   // Coprocessor load/store needs FTEMP, even if not declared
 694   if(itype[i]==C1LS||itype[i]==C2LS) {
 695     hsn[FTEMP]=0;
 696   }
 697   // Load L/R also uses FTEMP as a temporary register
 698   if(itype[i]==LOADLR) {
 699     hsn[FTEMP]=0;
 700   }
 701   // Also SWL/SWR/SDL/SDR
 702   if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) {
 703     hsn[FTEMP]=0;
 704   }
 705   // Don't remove the TLB registers either
 706   if(itype[i]==LOAD || itype[i]==LOADLR || itype[i]==STORE || itype[i]==STORELR || itype[i]==C1LS || itype[i]==C2LS) {
 707     hsn[TLREG]=0;
 708   }
 709   // Don't remove the miniht registers
 710   if(itype[i]==UJUMP||itype[i]==RJUMP)
 711   {
 712     hsn[RHASH]=0;
 713     hsn[RHTBL]=0;
 714   }
 715 }
 716
 717 // We only want to allocate registers if we're going to use them again soon
 718 int needed_again(int r, int i)
 719 {
 720   int j;
 721   int b=-1;
 722   int rn=10;
 723
 724   if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
 725   {
 726     if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
 727       return 0; // Don't need any registers if exiting the block
 728   }
 729   for(j=0;j<9;j++)
 730   {
 731     if(i+j>=slen) {
 732       j=slen-i-1;
 733       break;
 734     }
 735     if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
 736     {
 737       // Don't go past an unconditonal jump
 738       j++;
 739       break;
 740     }
 741     if(itype[i+j]==SYSCALL||itype[i+j]==HLECALL||itype[i+j]==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
 742     {
 743       break;
 744     }
 745   }
 746   for(;j>=1;j--)
 747   {
 748     if(rs1[i+j]==r) rn=j;
 749     if(rs2[i+j]==r) rn=j;
 750     if((unneeded_reg[i+j]>>r)&1) rn=10;
 751     if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
 752     {
 753       b=j;
 754     }
 755   }
 756   /*
 757   if(b>=0)
 758   {
 759     if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
 760     {
 761       // Follow first branch
 762       int o=rn;
 763       int t=(ba[i+b]-start)>>2;
 764       j=7-b;if(t+j>=slen) j=slen-t-1;
 765       for(;j>=0;j--)
 766       {
 767         if(!((unneeded_reg[t+j]>>r)&1)) {
 768           if(rs1[t+j]==r) if(rn>j+b+2) rn=j+b+2;
 769           if(rs2[t+j]==r) if(rn>j+b+2) rn=j+b+2;
 770         }
 771         else rn=o;
 772       }
 773     }
 774   }*/
 775   if(rn<10) return 1;
 776   return 0;
 777 }
 778
 779 // Try to match register allocations at the end of a loop with those
 780 // at the beginning
 781 int loop_reg(int i, int r, int hr)
 782 {
 783   int j,k;
 784   for(j=0;j<9;j++)
 785   {
 786     if(i+j>=slen) {
 787       j=slen-i-1;
 788       break;
 789     }
 790     if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
 791     {
 792       // Don't go past an unconditonal jump
 793       j++;
 794       break;
 795     }
 796   }
 797   k=0;
 798   if(i>0){
 799     if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)
 800       k--;
 801   }
 802   for(;k<j;k++)
 803   {
 804     if(r<64&&((unneeded_reg[i+k]>>r)&1)) return hr;
 805     if(r>64&&((unneeded_reg_upper[i+k]>>r)&1)) return hr;
 806     if(i+k>=0&&(itype[i+k]==UJUMP||itype[i+k]==CJUMP||itype[i+k]==SJUMP||itype[i+k]==FJUMP))
 807     {
 808       if(ba[i+k]>=start && ba[i+k]<(start+i*4))
 809       {
 810         int t=(ba[i+k]-start)>>2;
 811         int reg=get_reg(regs[t].regmap_entry,r);
 812         if(reg>=0) return reg;
 813         //reg=get_reg(regs[t+1].regmap_entry,r);
 814         //if(reg>=0) return reg;
 815       }
 816     }
 817   }
 818   return hr;
 819 }
 820
 821
 822 // Allocate every register, preserving source/target regs
 823 void alloc_all(struct regstat *cur,int i)
 824 {
 825   int hr;
 826
 827   for(hr=0;hr<HOST_REGS;hr++) {
 828     if(hr!=EXCLUDE_REG) {
 829       if(((cur->regmap[hr]&63)!=rs1[i])&&((cur->regmap[hr]&63)!=rs2[i])&&
 830          ((cur->regmap[hr]&63)!=rt1[i])&&((cur->regmap[hr]&63)!=rt2[i]))
 831       {
 832         cur->regmap[hr]=-1;
 833         cur->dirty&=~(1<<hr);
 834       }
 835       // Don't need zeros
 836       if((cur->regmap[hr]&63)==0)
 837       {
 838         cur->regmap[hr]=-1;
 839         cur->dirty&=~(1<<hr);
 840       }
 841     }
 842   }
 843 }
 844
 845
 846 void div64(int64_t dividend,int64_t divisor)
 847 {
 848   lo=dividend/divisor;
 849   hi=dividend%divisor;
 850   //printf("TRACE: ddiv %8x%8x %8x%8x\n" ,(int)reg[HIREG],(int)(reg[HIREG]>>32)
 851   //                                     ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
 852 }
 853 void divu64(uint64_t dividend,uint64_t divisor)
 854 {
 855   lo=dividend/divisor;
 856   hi=dividend%divisor;
 857   //printf("TRACE: ddivu %8x%8x %8x%8x\n",(int)reg[HIREG],(int)(reg[HIREG]>>32)
 858   //                                     ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
 859 }
 860
 861 void mult64(uint64_t m1,uint64_t m2)
 862 {
 863    unsigned long long int op1, op2, op3, op4;
 864    unsigned long long int result1, result2, result3, result4;
 865    unsigned long long int temp1, temp2, temp3, temp4;
 866    int sign = 0;
 867
 868    if (m1 < 0)
 869      {
 870     op2 = -m1;
 871     sign = 1 - sign;
 872      }
 873    else op2 = m1;
 874    if (m2 < 0)
 875      {
 876     op4 = -m2;
 877     sign = 1 - sign;
 878      }
 879    else op4 = m2;
 880
 881    op1 = op2 & 0xFFFFFFFF;
 882    op2 = (op2 >> 32) & 0xFFFFFFFF;
 883    op3 = op4 & 0xFFFFFFFF;
 884    op4 = (op4 >> 32) & 0xFFFFFFFF;
 885
 886    temp1 = op1 * op3;
 887    temp2 = (temp1 >> 32) + op1 * op4;
 888    temp3 = op2 * op3;
 889    temp4 = (temp3 >> 32) + op2 * op4;
 890
 891    result1 = temp1 & 0xFFFFFFFF;
 892    result2 = temp2 + (temp3 & 0xFFFFFFFF);
 893    result3 = (result2 >> 32) + temp4;
 894    result4 = (result3 >> 32);
 895
 896    lo = result1 | (result2 << 32);
 897    hi = (result3 & 0xFFFFFFFF) | (result4 << 32);
 898    if (sign)
 899      {
 900     hi = ~hi;
 901     if (!lo) hi++;
 902     else lo = ~lo + 1;
 903      }
 904 }
 905
 906 void multu64(uint64_t m1,uint64_t m2)
 907 {
 908    unsigned long long int op1, op2, op3, op4;
 909    unsigned long long int result1, result2, result3, result4;
 910    unsigned long long int temp1, temp2, temp3, temp4;
 911
 912    op1 = m1 & 0xFFFFFFFF;
 913    op2 = (m1 >> 32) & 0xFFFFFFFF;
 914    op3 = m2 & 0xFFFFFFFF;
 915    op4 = (m2 >> 32) & 0xFFFFFFFF;
 916
 917    temp1 = op1 * op3;
 918    temp2 = (temp1 >> 32) + op1 * op4;
 919    temp3 = op2 * op3;
 920    temp4 = (temp3 >> 32) + op2 * op4;
 921
 922    result1 = temp1 & 0xFFFFFFFF;
 923    result2 = temp2 + (temp3 & 0xFFFFFFFF);
 924    result3 = (result2 >> 32) + temp4;
 925    result4 = (result3 >> 32);
 926
 927    lo = result1 | (result2 << 32);
 928    hi = (result3 & 0xFFFFFFFF) | (result4 << 32);
 929
 930   //printf("TRACE: dmultu %8x%8x %8x%8x\n",(int)reg[HIREG],(int)(reg[HIREG]>>32)
 931   //                                      ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
 932 }
 933
 934 uint64_t ldl_merge(uint64_t original,uint64_t loaded,u_int bits)
 935 {
 936   if(bits) {
 937     original<<=64-bits;
 938     original>>=64-bits;
 939     loaded<<=bits;
 940     original|=loaded;
 941   }
 942   else original=loaded;
 943   return original;
 944 }
 945 uint64_t ldr_merge(uint64_t original,uint64_t loaded,u_int bits)
 946 {
 947   if(bits^56) {
 948     original>>=64-(bits^56);
 949     original<<=64-(bits^56);
 950     loaded>>=bits^56;
 951     original|=loaded;
 952   }
 953   else original=loaded;
 954   return original;
 955 }
 956
 957 #ifdef __i386__
 958 #include "assem_x86.c"
 959 #endif
 960 #ifdef __x86_64__
 961 #include "assem_x64.c"
 962 #endif
 963 #ifdef __arm__
 964 #include "assem_arm.c"
 965 #endif
 966
 967 // Add virtual address mapping to linked list
 968 void ll_add(struct ll_entry **head,int vaddr,void *addr)
 969 {
 970   struct ll_entry *new_entry;
 971   new_entry=malloc(sizeof(struct ll_entry));
 972   assert(new_entry!=NULL);
 973   new_entry->vaddr=vaddr;
 974   new_entry->reg32=0;
 975   new_entry->addr=addr;
 976   new_entry->next=*head;
 977   *head=new_entry;
 978 }
 979
 980 // Add virtual address mapping for 32-bit compiled block
 981 void ll_add_32(struct ll_entry **head,int vaddr,u_int reg32,void *addr)
 982 {
 983   ll_add(head,vaddr,addr);
 984 #ifndef FORCE32
 985   (*head)->reg32=reg32;
 986 #endif
 987 }
 988
 989 // Check if an address is already compiled
 990 // but don't return addresses which are about to expire from the cache
 991 void *check_addr(u_int vaddr)
 992 {
 993   u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 994   if(ht_bin[0]==vaddr) {
 995     if(((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
 996       if(isclean(ht_bin[1])) return (void *)ht_bin[1];
 997   }
 998   if(ht_bin[2]==vaddr) {
 999     if(((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
1000       if(isclean(ht_bin[3])) return (void *)ht_bin[3];
1001   }
1002   u_int page=get_page(vaddr);
1003   struct ll_entry *head;
1004   head=jump_in[page];
1005   while(head!=NULL) {
1006     if(head->vaddr==vaddr&&head->reg32==0) {
1007       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1008         // Update existing entry with current address
1009         if(ht_bin[0]==vaddr) {
1010           ht_bin[1]=(int)head->addr;
1011           return head->addr;
1012         }
1013         if(ht_bin[2]==vaddr) {
1014           ht_bin[3]=(int)head->addr;
1015           return head->addr;
1016         }
1017         // Insert into hash table with low priority.
1018         // Don't evict existing entries, as they are probably
1019         // addresses that are being accessed frequently.
1020         if(ht_bin[0]==-1) {
1021           ht_bin[1]=(int)head->addr;
1022           ht_bin[0]=vaddr;
1023         }else if(ht_bin[2]==-1) {
1024           ht_bin[3]=(int)head->addr;
1025           ht_bin[2]=vaddr;
1026         }
1027         return head->addr;
1028       }
1029     }
1030     head=head->next;
1031   }
1032   return 0;
1033 }
1034
1035 void remove_hash(int vaddr)
1036 {
1037   //printf("remove hash: %x\n",vaddr);
1038   int *ht_bin=hash_table[(((vaddr)>>16)^vaddr)&0xFFFF];
1039   if(ht_bin[2]==vaddr) {
1040     ht_bin[2]=ht_bin[3]=-1;
1041   }
1042   if(ht_bin[0]==vaddr) {
1043     ht_bin[0]=ht_bin[2];
1044     ht_bin[1]=ht_bin[3];
1045     ht_bin[2]=ht_bin[3]=-1;
1046   }
1047 }
1048
1049 void ll_remove_matching_addrs(struct ll_entry **head,int addr,int shift)
1050 {
1051   struct ll_entry *next;
1052   while(*head) {
1053     if(((u_int)((*head)->addr)>>shift)==(addr>>shift) ||
1054        ((u_int)((*head)->addr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift))
1055     {
1056       inv_debug("EXP: Remove pointer to %x (%x)\n",(int)(*head)->addr,(*head)->vaddr);
1057       remove_hash((*head)->vaddr);
1058       next=(*head)->next;
1059       free(*head);
1060       *head=next;
1061     }
1062     else
1063     {
1064       head=&((*head)->next);
1065     }
1066   }
1067 }
1068
1069 // Remove all entries from linked list
1070 void ll_clear(struct ll_entry **head)
1071 {
1072   struct ll_entry *cur;
1073   struct ll_entry *next;
1074   if(cur=*head) {
1075     *head=0;
1076     while(cur) {
1077       next=cur->next;
1078       free(cur);
1079       cur=next;
1080     }
1081   }
1082 }
1083
1084 // Dereference the pointers and remove if it matches
1085 void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
1086 {
1087   while(head) {
1088     int ptr=get_pointer(head->addr);
1089     inv_debug("EXP: Lookup pointer to %x at %x (%x)\n",(int)ptr,(int)head->addr,head->vaddr);
1090     if(((ptr>>shift)==(addr>>shift)) ||
1091        (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
1092     {
1093       inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
1094       u_int host_addr=(u_int)kill_pointer(head->addr);
1095       #ifdef __arm__
1096         needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
1097       #endif
1098     }
1099     head=head->next;
1100   }
1101 }
1102
1103 // This is called when we write to a compiled block (see do_invstub)
1104 void invalidate_page(u_int page)
1105 {
1106   struct ll_entry *head;
1107   struct ll_entry *next;
1108   head=jump_in[page];
1109   jump_in[page]=0;
1110   while(head!=NULL) {
1111     inv_debug("INVALIDATE: %x\n",head->vaddr);
1112     remove_hash(head->vaddr);
1113     next=head->next;
1114     free(head);
1115     head=next;
1116   }
1117   head=jump_out[page];
1118   jump_out[page]=0;
1119   while(head!=NULL) {
1120     inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
1121     u_int host_addr=(u_int)kill_pointer(head->addr);
1122     #ifdef __arm__
1123       needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
1124     #endif
1125     next=head->next;
1126     free(head);
1127     head=next;
1128   }
1129 }
1130 void invalidate_block(u_int block)
1131 {
1132   u_int page=get_page(block<<12);
1133   u_int vpage=get_vpage(block<<12);
1134   inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1135   //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1136   u_int first,last;
1137   first=last=page;
1138   struct ll_entry *head;
1139   head=jump_dirty[vpage];
1140   //printf("page=%d vpage=%d\n",page,vpage);
1141   while(head!=NULL) {
1142     u_int start,end;
1143     if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1144       get_bounds((int)head->addr,&start,&end);
1145       //printf("start: %x end: %x\n",start,end);
1146       if(page<2048&&start>=0x80000000&&end<0x80000000+RAM_SIZE) {
1147         if(((start-(u_int)rdram)>>12)<=page&&((end-1-(u_int)rdram)>>12)>=page) {
1148           if((((start-(u_int)rdram)>>12)&2047)<first) first=((start-(u_int)rdram)>>12)&2047;
1149           if((((end-1-(u_int)rdram)>>12)&2047)>last) last=((end-1-(u_int)rdram)>>12)&2047;
1150         }
1151       }
1152 #ifndef DISABLE_TLB
1153       if(page<2048&&(signed int)start>=(signed int)0xC0000000&&(signed int)end>=(signed int)0xC0000000) {
1154         if(((start+memory_map[start>>12]-(u_int)rdram)>>12)<=page&&((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)>=page) {
1155           if((((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047)<first) first=((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047;
1156           if((((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)&2047)>last) last=((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)&2047;
1157         }
1158       }
1159 #endif
1160     }
1161     head=head->next;
1162   }
1163   //printf("first=%d last=%d\n",first,last);
1164   invalidate_page(page);
1165   assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1166   assert(last<page+5);
1167   // Invalidate the adjacent pages if a block crosses a 4K boundary
1168   while(first<page) {
1169     invalidate_page(first);
1170     first++;
1171   }
1172   for(first=page+1;first<last;first++) {
1173     invalidate_page(first);
1174   }
1175   #ifdef __arm__
1176     do_clear_cache();
1177   #endif
1178
1179   // Don't trap writes
1180   invalid_code[block]=1;
1181 #ifdef PCSX
1182   invalid_code[((u_int)0x80000000>>12)|page]=1;
1183 #endif
1184 #ifndef DISABLE_TLB
1185   // If there is a valid TLB entry for this page, remove write protect
1186   if(tlb_LUT_w[block]) {
1187     assert(tlb_LUT_r[block]==tlb_LUT_w[block]);
1188     // CHECK: Is this right?
1189     memory_map[block]=((tlb_LUT_w[block]&0xFFFFF000)-(block<<12)+(unsigned int)rdram-0x80000000)>>2;
1190     u_int real_block=tlb_LUT_w[block]>>12;
1191     invalid_code[real_block]=1;
1192     if(real_block>=0x80000&&real_block<0x80800) memory_map[real_block]=((u_int)rdram-0x80000000)>>2;
1193   }
1194   else if(block>=0x80000&&block<0x80800) memory_map[block]=((u_int)rdram-0x80000000)>>2;
1195 #endif
1196
1197   #ifdef USE_MINI_HT
1198   memset(mini_ht,-1,sizeof(mini_ht));
1199   #endif
1200 }
1201 void invalidate_addr(u_int addr)
1202 {
1203   invalidate_block(addr>>12);
1204 }
1205 // This is called when loading a save state.
1206 // Anything could have changed, so invalidate everything.
1207 void invalidate_all_pages()
1208 {
1209   u_int page,n;
1210   for(page=0;page<4096;page++)
1211     invalidate_page(page);
1212   for(page=0;page<1048576;page++)
1213     if(!invalid_code[page]) {
1214       restore_candidate[(page&2047)>>3]|=1<<(page&7);
1215       restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1216     }
1217   #ifdef __arm__
1218   __clear_cache((void *)BASE_ADDR,(void *)BASE_ADDR+(1<<TARGET_SIZE_2));
1219   #endif
1220   #ifdef USE_MINI_HT
1221   memset(mini_ht,-1,sizeof(mini_ht));
1222   #endif
1223   #ifndef DISABLE_TLB
1224   // TLB
1225   for(page=0;page<0x100000;page++) {
1226     if(tlb_LUT_r[page]) {
1227       memory_map[page]=((tlb_LUT_r[page]&0xFFFFF000)-(page<<12)+(unsigned int)rdram-0x80000000)>>2;
1228       if(!tlb_LUT_w[page]||!invalid_code[page])
1229         memory_map[page]|=0x40000000; // Write protect
1230     }
1231     else memory_map[page]=-1;
1232     if(page==0x80000) page=0xC0000;
1233   }
1234   tlb_hacks();
1235   #endif
1236 }
1237
1238 // Add an entry to jump_out after making a link
1239 void add_link(u_int vaddr,void *src)
1240 {
1241   u_int page=get_page(vaddr);
1242   inv_debug("add_link: %x -> %x (%d)\n",(int)src,vaddr,page);
1243   ll_add(jump_out+page,vaddr,src);
1244   //int ptr=get_pointer(src);
1245   //inv_debug("add_link: Pointer is to %x\n",(int)ptr);
1246 }
1247
1248 // If a code block was found to be unmodified (bit was set in
1249 // restore_candidate) and it remains unmodified (bit is clear
1250 // in invalid_code) then move the entries for that 4K page from
1251 // the dirty list to the clean list.
1252 void clean_blocks(u_int page)
1253 {
1254   struct ll_entry *head;
1255   inv_debug("INV: clean_blocks page=%d\n",page);
1256   head=jump_dirty[page];
1257   while(head!=NULL) {
1258     if(!invalid_code[head->vaddr>>12]) {
1259       // Don't restore blocks which are about to expire from the cache
1260       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1261         u_int start,end;
1262         if(verify_dirty((int)head->addr)) {
1263           //printf("Possibly Restore %x (%x)\n",head->vaddr, (int)head->addr);
1264           u_int i;
1265           u_int inv=0;
1266           get_bounds((int)head->addr,&start,&end);
1267           if(start-(u_int)rdram<RAM_SIZE) {
1268             for(i=(start-(u_int)rdram+0x80000000)>>12;i<=(end-1-(u_int)rdram+0x80000000)>>12;i++) {
1269               inv|=invalid_code[i];
1270             }
1271           }
1272           if((signed int)head->vaddr>=(signed int)0xC0000000) {
1273             u_int addr = (head->vaddr+(memory_map[head->vaddr>>12]<<2));
1274             //printf("addr=%x start=%x end=%x\n",addr,start,end);
1275             if(addr<start||addr>=end) inv=1;
1276           }
1277           else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1278             inv=1;
1279           }
1280           if(!inv) {
1281             void * clean_addr=(void *)get_clean_addr((int)head->addr);
1282             if((((u_int)clean_addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1283               u_int ppage=page;
1284 #ifndef DISABLE_TLB
1285               if(page<2048&&tlb_LUT_r[head->vaddr>>12]) ppage=(tlb_LUT_r[head->vaddr>>12]^0x80000000)>>12;
1286 #endif
1287               inv_debug("INV: Restored %x (%x/%x)\n",head->vaddr, (int)head->addr, (int)clean_addr);
1288               //printf("page=%x, addr=%x\n",page,head->vaddr);
1289               //assert(head->vaddr>>12==(page|0x80000));
1290               ll_add_32(jump_in+ppage,head->vaddr,head->reg32,clean_addr);
1291               int *ht_bin=hash_table[((head->vaddr>>16)^head->vaddr)&0xFFFF];
1292               if(!head->reg32) {
1293                 if(ht_bin[0]==head->vaddr) {
1294                   ht_bin[1]=(int)clean_addr; // Replace existing entry
1295                 }
1296                 if(ht_bin[2]==head->vaddr) {
1297                   ht_bin[3]=(int)clean_addr; // Replace existing entry
1298                 }
1299               }
1300             }
1301           }
1302         }
1303       }
1304     }
1305     head=head->next;
1306   }
1307 }
1308
1309
1310 void mov_alloc(struct regstat *current,int i)
1311 {
1312   // Note: Don't need to actually alloc the source registers
1313   if((~current->is32>>rs1[i])&1) {
1314     //alloc_reg64(current,i,rs1[i]);
1315     alloc_reg64(current,i,rt1[i]);
1316     current->is32&=~(1LL<<rt1[i]);
1317   } else {
1318     //alloc_reg(current,i,rs1[i]);
1319     alloc_reg(current,i,rt1[i]);
1320     current->is32|=(1LL<<rt1[i]);
1321   }
1322   clear_const(current,rs1[i]);
1323   clear_const(current,rt1[i]);
1324   dirty_reg(current,rt1[i]);
1325 }
1326
1327 void shiftimm_alloc(struct regstat *current,int i)
1328 {
1329   clear_const(current,rs1[i]);
1330   clear_const(current,rt1[i]);
1331   if(opcode2[i]<=0x3) // SLL/SRL/SRA
1332   {
1333     if(rt1[i]) {
1334       if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1335       else lt1[i]=rs1[i];
1336       alloc_reg(current,i,rt1[i]);
1337       current->is32|=1LL<<rt1[i];
1338       dirty_reg(current,rt1[i]);
1339     }
1340   }
1341   if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
1342   {
1343     if(rt1[i]) {
1344       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1345       alloc_reg64(current,i,rt1[i]);
1346       current->is32&=~(1LL<<rt1[i]);
1347       dirty_reg(current,rt1[i]);
1348     }
1349   }
1350   if(opcode2[i]==0x3c) // DSLL32
1351   {
1352     if(rt1[i]) {
1353       if(rs1[i]) alloc_reg(current,i,rs1[i]);
1354       alloc_reg64(current,i,rt1[i]);
1355       current->is32&=~(1LL<<rt1[i]);
1356       dirty_reg(current,rt1[i]);
1357     }
1358   }
1359   if(opcode2[i]==0x3e) // DSRL32
1360   {
1361     if(rt1[i]) {
1362       alloc_reg64(current,i,rs1[i]);
1363       if(imm[i]==32) {
1364         alloc_reg64(current,i,rt1[i]);
1365         current->is32&=~(1LL<<rt1[i]);
1366       } else {
1367         alloc_reg(current,i,rt1[i]);
1368         current->is32|=1LL<<rt1[i];
1369       }
1370       dirty_reg(current,rt1[i]);
1371     }
1372   }
1373   if(opcode2[i]==0x3f) // DSRA32
1374   {
1375     if(rt1[i]) {
1376       alloc_reg64(current,i,rs1[i]);
1377       alloc_reg(current,i,rt1[i]);
1378       current->is32|=1LL<<rt1[i];
1379       dirty_reg(current,rt1[i]);
1380     }
1381   }
1382 }
1383
1384 void shift_alloc(struct regstat *current,int i)
1385 {
1386   if(rt1[i]) {
1387     if(opcode2[i]<=0x07) // SLLV/SRLV/SRAV
1388     {
1389       if(rs1[i]) alloc_reg(current,i,rs1[i]);
1390       if(rs2[i]) alloc_reg(current,i,rs2[i]);
1391       alloc_reg(current,i,rt1[i]);
1392       if(rt1[i]==rs2[i]) {
1393         alloc_reg_temp(current,i,-1);
1394         minimum_free_regs[i]=1;
1395       }
1396       current->is32|=1LL<<rt1[i];
1397     } else { // DSLLV/DSRLV/DSRAV
1398       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1399       if(rs2[i]) alloc_reg(current,i,rs2[i]);
1400       alloc_reg64(current,i,rt1[i]);
1401       current->is32&=~(1LL<<rt1[i]);
1402       if(opcode2[i]==0x16||opcode2[i]==0x17) // DSRLV and DSRAV need a temporary register
1403       {
1404         alloc_reg_temp(current,i,-1);
1405         minimum_free_regs[i]=1;
1406       }
1407     }
1408     clear_const(current,rs1[i]);
1409     clear_const(current,rs2[i]);
1410     clear_const(current,rt1[i]);
1411     dirty_reg(current,rt1[i]);
1412   }
1413 }
1414
1415 void alu_alloc(struct regstat *current,int i)
1416 {
1417   if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
1418     if(rt1[i]) {
1419       if(rs1[i]&&rs2[i]) {
1420         alloc_reg(current,i,rs1[i]);
1421         alloc_reg(current,i,rs2[i]);
1422       }
1423       else {
1424         if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1425         if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1426       }
1427       alloc_reg(current,i,rt1[i]);
1428     }
1429     current->is32|=1LL<<rt1[i];
1430   }
1431   if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
1432     if(rt1[i]) {
1433       if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1434       {
1435         alloc_reg64(current,i,rs1[i]);
1436         alloc_reg64(current,i,rs2[i]);
1437         alloc_reg(current,i,rt1[i]);
1438       } else {
1439         alloc_reg(current,i,rs1[i]);
1440         alloc_reg(current,i,rs2[i]);
1441         alloc_reg(current,i,rt1[i]);
1442       }
1443     }
1444     current->is32|=1LL<<rt1[i];
1445   }
1446   if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
1447     if(rt1[i]) {
1448       if(rs1[i]&&rs2[i]) {
1449         alloc_reg(current,i,rs1[i]);
1450         alloc_reg(current,i,rs2[i]);
1451       }
1452       else
1453       {
1454         if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1455         if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1456       }
1457       alloc_reg(current,i,rt1[i]);
1458       if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1459       {
1460         if(!((current->uu>>rt1[i])&1)) {
1461           alloc_reg64(current,i,rt1[i]);
1462         }
1463         if(get_reg(current->regmap,rt1[i]|64)>=0) {
1464           if(rs1[i]&&rs2[i]) {
1465             alloc_reg64(current,i,rs1[i]);
1466             alloc_reg64(current,i,rs2[i]);
1467           }
1468           else
1469           {
1470             // Is is really worth it to keep 64-bit values in registers?
1471             #ifdef NATIVE_64BIT
1472             if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1473             if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg64(current,i,rs2[i]);
1474             #endif
1475           }
1476         }
1477         current->is32&=~(1LL<<rt1[i]);
1478       } else {
1479         current->is32|=1LL<<rt1[i];
1480       }
1481     }
1482   }
1483   if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1484     if(rt1[i]) {
1485       if(rs1[i]&&rs2[i]) {
1486         if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1487           alloc_reg64(current,i,rs1[i]);
1488           alloc_reg64(current,i,rs2[i]);
1489           alloc_reg64(current,i,rt1[i]);
1490         } else {
1491           alloc_reg(current,i,rs1[i]);
1492           alloc_reg(current,i,rs2[i]);
1493           alloc_reg(current,i,rt1[i]);
1494         }
1495       }
1496       else {
1497         alloc_reg(current,i,rt1[i]);
1498         if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1499           // DADD used as move, or zeroing
1500           // If we have a 64-bit source, then make the target 64 bits too
1501           if(rs1[i]&&!((current->is32>>rs1[i])&1)) {
1502             if(get_reg(current->regmap,rs1[i])>=0) alloc_reg64(current,i,rs1[i]);
1503             alloc_reg64(current,i,rt1[i]);
1504           } else if(rs2[i]&&!((current->is32>>rs2[i])&1)) {
1505             if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1506             alloc_reg64(current,i,rt1[i]);
1507           }
1508           if(opcode2[i]>=0x2e&&rs2[i]) {
1509             // DSUB used as negation - 64-bit result
1510             // If we have a 32-bit register, extend it to 64 bits
1511             if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1512             alloc_reg64(current,i,rt1[i]);
1513           }
1514         }
1515       }
1516       if(rs1[i]&&rs2[i]) {
1517         current->is32&=~(1LL<<rt1[i]);
1518       } else if(rs1[i]) {
1519         current->is32&=~(1LL<<rt1[i]);
1520         if((current->is32>>rs1[i])&1)
1521           current->is32|=1LL<<rt1[i];
1522       } else if(rs2[i]) {
1523         current->is32&=~(1LL<<rt1[i]);
1524         if((current->is32>>rs2[i])&1)
1525           current->is32|=1LL<<rt1[i];
1526       } else {
1527         current->is32|=1LL<<rt1[i];
1528       }
1529     }
1530   }
1531   clear_const(current,rs1[i]);
1532   clear_const(current,rs2[i]);
1533   clear_const(current,rt1[i]);
1534   dirty_reg(current,rt1[i]);
1535 }
1536
1537 void imm16_alloc(struct regstat *current,int i)
1538 {
1539   if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1540   else lt1[i]=rs1[i];
1541   if(rt1[i]) alloc_reg(current,i,rt1[i]);
1542   if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
1543     current->is32&=~(1LL<<rt1[i]);
1544     if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1545       // TODO: Could preserve the 32-bit flag if the immediate is zero
1546       alloc_reg64(current,i,rt1[i]);
1547       alloc_reg64(current,i,rs1[i]);
1548     }
1549     clear_const(current,rs1[i]);
1550     clear_const(current,rt1[i]);
1551   }
1552   else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
1553     if((~current->is32>>rs1[i])&1) alloc_reg64(current,i,rs1[i]);
1554     current->is32|=1LL<<rt1[i];
1555     clear_const(current,rs1[i]);
1556     clear_const(current,rt1[i]);
1557   }
1558   else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
1559     if(((~current->is32>>rs1[i])&1)&&opcode[i]>0x0c) {
1560       if(rs1[i]!=rt1[i]) {
1561         if(needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1562         alloc_reg64(current,i,rt1[i]);
1563         current->is32&=~(1LL<<rt1[i]);
1564       }
1565     }
1566     else current->is32|=1LL<<rt1[i]; // ANDI clears upper bits
1567     if(is_const(current,rs1[i])) {
1568       int v=get_const(current,rs1[i]);
1569       if(opcode[i]==0x0c) set_const(current,rt1[i],v&imm[i]);
1570       if(opcode[i]==0x0d) set_const(current,rt1[i],v|imm[i]);
1571       if(opcode[i]==0x0e) set_const(current,rt1[i],v^imm[i]);
1572     }
1573     else clear_const(current,rt1[i]);
1574   }
1575   else if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
1576     if(is_const(current,rs1[i])) {
1577       int v=get_const(current,rs1[i]);
1578       set_const(current,rt1[i],v+imm[i]);
1579     }
1580     else clear_const(current,rt1[i]);
1581     current->is32|=1LL<<rt1[i];
1582   }
1583   else {
1584     set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
1585     current->is32|=1LL<<rt1[i];
1586   }
1587   dirty_reg(current,rt1[i]);
1588 }
1589
1590 void load_alloc(struct regstat *current,int i)
1591 {
1592   clear_const(current,rt1[i]);
1593   //if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
1594   if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
1595   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1596   if(rt1[i]&&!((current->u>>rt1[i])&1)) {
1597     alloc_reg(current,i,rt1[i]);
1598     assert(get_reg(current->regmap,rt1[i])>=0);
1599     if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
1600     {
1601       current->is32&=~(1LL<<rt1[i]);
1602       alloc_reg64(current,i,rt1[i]);
1603     }
1604     else if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1605     {
1606       current->is32&=~(1LL<<rt1[i]);
1607       alloc_reg64(current,i,rt1[i]);
1608       alloc_all(current,i);
1609       alloc_reg64(current,i,FTEMP);
1610       minimum_free_regs[i]=HOST_REGS;
1611     }
1612     else current->is32|=1LL<<rt1[i];
1613     dirty_reg(current,rt1[i]);
1614     // If using TLB, need a register for pointer to the mapping table
1615     if(using_tlb) alloc_reg(current,i,TLREG);
1616     // LWL/LWR need a temporary register for the old value
1617     if(opcode[i]==0x22||opcode[i]==0x26)
1618     {
1619       alloc_reg(current,i,FTEMP);
1620       alloc_reg_temp(current,i,-1);
1621       minimum_free_regs[i]=1;
1622     }
1623   }
1624   else
1625   {
1626     // Load to r0 or unneeded register (dummy load)
1627     // but we still need a register to calculate the address
1628     if(opcode[i]==0x22||opcode[i]==0x26)
1629     {
1630       alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1631     }
1632     // If using TLB, need a register for pointer to the mapping table
1633     if(using_tlb) alloc_reg(current,i,TLREG);
1634     alloc_reg_temp(current,i,-1);
1635     minimum_free_regs[i]=1;
1636     if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1637     {
1638       alloc_all(current,i);
1639       alloc_reg64(current,i,FTEMP);
1640       minimum_free_regs[i]=HOST_REGS;
1641     }
1642   }
1643 }
1644
1645 void store_alloc(struct regstat *current,int i)
1646 {
1647   clear_const(current,rs2[i]);
1648   if(!(rs2[i])) current->u&=~1LL; // Allow allocating r0 if necessary
1649   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1650   alloc_reg(current,i,rs2[i]);
1651   if(opcode[i]==0x2c||opcode[i]==0x2d||opcode[i]==0x3f) { // 64-bit SDL/SDR/SD
1652     alloc_reg64(current,i,rs2[i]);
1653     if(rs2[i]) alloc_reg(current,i,FTEMP);
1654   }
1655   // If using TLB, need a register for pointer to the mapping table
1656   if(using_tlb) alloc_reg(current,i,TLREG);
1657   #if defined(HOST_IMM8)
1658   // On CPUs without 32-bit immediates we need a pointer to invalid_code
1659   else alloc_reg(current,i,INVCP);
1660   #endif
1661   if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) { // SWL/SWL/SDL/SDR
1662     alloc_reg(current,i,FTEMP);
1663   }
1664   // We need a temporary register for address generation
1665   alloc_reg_temp(current,i,-1);
1666   minimum_free_regs[i]=1;
1667 }
1668
1669 void c1ls_alloc(struct regstat *current,int i)
1670 {
1671   //clear_const(current,rs1[i]); // FIXME
1672   clear_const(current,rt1[i]);
1673   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1674   alloc_reg(current,i,CSREG); // Status
1675   alloc_reg(current,i,FTEMP);
1676   if(opcode[i]==0x35||opcode[i]==0x3d) { // 64-bit LDC1/SDC1
1677     alloc_reg64(current,i,FTEMP);
1678   }
1679   // If using TLB, need a register for pointer to the mapping table
1680   if(using_tlb) alloc_reg(current,i,TLREG);
1681   #if defined(HOST_IMM8)
1682   // On CPUs without 32-bit immediates we need a pointer to invalid_code
1683   else if((opcode[i]&0x3b)==0x39) // SWC1/SDC1
1684     alloc_reg(current,i,INVCP);
1685   #endif
1686   // We need a temporary register for address generation
1687   alloc_reg_temp(current,i,-1);
1688 }
1689
1690 void c2ls_alloc(struct regstat *current,int i)
1691 {
1692   clear_const(current,rt1[i]);
1693   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1694   alloc_reg(current,i,FTEMP);
1695   // If using TLB, need a register for pointer to the mapping table
1696   if(using_tlb) alloc_reg(current,i,TLREG);
1697   #if defined(HOST_IMM8)
1698   // On CPUs without 32-bit immediates we need a pointer to invalid_code
1699   else if((opcode[i]&0x3b)==0x3a) // SWC2/SDC2
1700     alloc_reg(current,i,INVCP);
1701   #endif
1702   // We need a temporary register for address generation
1703   alloc_reg_temp(current,i,-1);
1704   minimum_free_regs[i]=1;
1705 }
1706
1707 #ifndef multdiv_alloc
1708 void multdiv_alloc(struct regstat *current,int i)
1709 {
1710   //  case 0x18: MULT
1711   //  case 0x19: MULTU
1712   //  case 0x1A: DIV
1713   //  case 0x1B: DIVU
1714   //  case 0x1C: DMULT
1715   //  case 0x1D: DMULTU
1716   //  case 0x1E: DDIV
1717   //  case 0x1F: DDIVU
1718   clear_const(current,rs1[i]);
1719   clear_const(current,rs2[i]);
1720   if(rs1[i]&&rs2[i])
1721   {
1722     if((opcode2[i]&4)==0) // 32-bit
1723     {
1724       current->u&=~(1LL<<HIREG);
1725       current->u&=~(1LL<<LOREG);
1726       alloc_reg(current,i,HIREG);
1727       alloc_reg(current,i,LOREG);
1728       alloc_reg(current,i,rs1[i]);
1729       alloc_reg(current,i,rs2[i]);
1730       current->is32|=1LL<<HIREG;
1731       current->is32|=1LL<<LOREG;
1732       dirty_reg(current,HIREG);
1733       dirty_reg(current,LOREG);
1734     }
1735     else // 64-bit
1736     {
1737       current->u&=~(1LL<<HIREG);
1738       current->u&=~(1LL<<LOREG);
1739       current->uu&=~(1LL<<HIREG);
1740       current->uu&=~(1LL<<LOREG);
1741       alloc_reg64(current,i,HIREG);
1742       //if(HOST_REGS>10) alloc_reg64(current,i,LOREG);
1743       alloc_reg64(current,i,rs1[i]);
1744       alloc_reg64(current,i,rs2[i]);
1745       alloc_all(current,i);
1746       current->is32&=~(1LL<<HIREG);
1747       current->is32&=~(1LL<<LOREG);
1748       dirty_reg(current,HIREG);
1749       dirty_reg(current,LOREG);
1750       minimum_free_regs[i]=HOST_REGS;
1751     }
1752   }
1753   else
1754   {
1755     // Multiply by zero is zero.
1756     // MIPS does not have a divide by zero exception.
1757     // The result is undefined, we return zero.
1758     alloc_reg(current,i,HIREG);
1759     alloc_reg(current,i,LOREG);
1760     current->is32|=1LL<<HIREG;
1761     current->is32|=1LL<<LOREG;
1762     dirty_reg(current,HIREG);
1763     dirty_reg(current,LOREG);
1764   }
1765 }
1766 #endif
1767
1768 void cop0_alloc(struct regstat *current,int i)
1769 {
1770   if(opcode2[i]==0) // MFC0
1771   {
1772     if(rt1[i]) {
1773       clear_const(current,rt1[i]);
1774       alloc_all(current,i);
1775       alloc_reg(current,i,rt1[i]);
1776       current->is32|=1LL<<rt1[i];
1777       dirty_reg(current,rt1[i]);
1778     }
1779   }
1780   else if(opcode2[i]==4) // MTC0
1781   {
1782     if(rs1[i]){
1783       clear_const(current,rs1[i]);
1784       alloc_reg(current,i,rs1[i]);
1785       alloc_all(current,i);
1786     }
1787     else {
1788       alloc_all(current,i); // FIXME: Keep r0
1789       current->u&=~1LL;
1790       alloc_reg(current,i,0);
1791     }
1792   }
1793   else
1794   {
1795     // TLBR/TLBWI/TLBWR/TLBP/ERET
1796     assert(opcode2[i]==0x10);
1797     alloc_all(current,i);
1798   }
1799   minimum_free_regs[i]=HOST_REGS;
1800 }
1801
1802 void cop1_alloc(struct regstat *current,int i)
1803 {
1804   alloc_reg(current,i,CSREG); // Load status
1805   if(opcode2[i]<3) // MFC1/DMFC1/CFC1
1806   {
1807     if(rt1[i]){
1808       clear_const(current,rt1[i]);
1809       if(opcode2[i]==1) {
1810         alloc_reg64(current,i,rt1[i]); // DMFC1
1811         current->is32&=~(1LL<<rt1[i]);
1812       }else{
1813         alloc_reg(current,i,rt1[i]); // MFC1/CFC1
1814         current->is32|=1LL<<rt1[i];
1815       }
1816       dirty_reg(current,rt1[i]);
1817     }
1818     alloc_reg_temp(current,i,-1);
1819   }
1820   else if(opcode2[i]>3) // MTC1/DMTC1/CTC1
1821   {
1822     if(rs1[i]){
1823       clear_const(current,rs1[i]);
1824       if(opcode2[i]==5)
1825         alloc_reg64(current,i,rs1[i]); // DMTC1
1826       else
1827         alloc_reg(current,i,rs1[i]); // MTC1/CTC1
1828       alloc_reg_temp(current,i,-1);
1829     }
1830     else {
1831       current->u&=~1LL;
1832       alloc_reg(current,i,0);
1833       alloc_reg_temp(current,i,-1);
1834     }
1835   }
1836   minimum_free_regs[i]=1;
1837 }
1838 void fconv_alloc(struct regstat *current,int i)
1839 {
1840   alloc_reg(current,i,CSREG); // Load status
1841   alloc_reg_temp(current,i,-1);
1842   minimum_free_regs[i]=1;
1843 }
1844 void float_alloc(struct regstat *current,int i)
1845 {
1846   alloc_reg(current,i,CSREG); // Load status
1847   alloc_reg_temp(current,i,-1);
1848   minimum_free_regs[i]=1;
1849 }
1850 void c2op_alloc(struct regstat *current,int i)
1851 {
1852   alloc_reg_temp(current,i,-1);
1853 }
1854 void fcomp_alloc(struct regstat *current,int i)
1855 {
1856   alloc_reg(current,i,CSREG); // Load status
1857   alloc_reg(current,i,FSREG); // Load flags
1858   dirty_reg(current,FSREG); // Flag will be modified
1859   alloc_reg_temp(current,i,-1);
1860   minimum_free_regs[i]=1;
1861 }
1862
1863 void syscall_alloc(struct regstat *current,int i)
1864 {
1865   alloc_cc(current,i);
1866   dirty_reg(current,CCREG);
1867   alloc_all(current,i);
1868   minimum_free_regs[i]=HOST_REGS;
1869   current->isconst=0;
1870 }
1871
1872 void delayslot_alloc(struct regstat *current,int i)
1873 {
1874   switch(itype[i]) {
1875     case UJUMP:
1876     case CJUMP:
1877     case SJUMP:
1878     case RJUMP:
1879     case FJUMP:
1880     case SYSCALL:
1881     case HLECALL:
1882     case SPAN:
1883       assem_debug("jump in the delay slot.  this shouldn't happen.\n");//exit(1);
1884       printf("Disabled speculative precompilation\n");
1885       stop_after_jal=1;
1886       break;
1887     case IMM16:
1888       imm16_alloc(current,i);
1889       break;
1890     case LOAD:
1891     case LOADLR:
1892       load_alloc(current,i);
1893       break;
1894     case STORE:
1895     case STORELR:
1896       store_alloc(current,i);
1897       break;
1898     case ALU:
1899       alu_alloc(current,i);
1900       break;
1901     case SHIFT:
1902       shift_alloc(current,i);
1903       break;
1904     case MULTDIV:
1905       multdiv_alloc(current,i);
1906       break;
1907     case SHIFTIMM:
1908       shiftimm_alloc(current,i);
1909       break;
1910     case MOV:
1911       mov_alloc(current,i);
1912       break;
1913     case COP0:
1914       cop0_alloc(current,i);
1915       break;
1916     case COP1:
1917     case COP2:
1918       cop1_alloc(current,i);
1919       break;
1920     case C1LS:
1921       c1ls_alloc(current,i);
1922       break;
1923     case C2LS:
1924       c2ls_alloc(current,i);
1925       break;
1926     case FCONV:
1927       fconv_alloc(current,i);
1928       break;
1929     case FLOAT:
1930       float_alloc(current,i);
1931       break;
1932     case FCOMP:
1933       fcomp_alloc(current,i);
1934       break;
1935     case C2OP:
1936       c2op_alloc(current,i);
1937       break;
1938   }
1939 }
1940
1941 // Special case where a branch and delay slot span two pages in virtual memory
1942 static void pagespan_alloc(struct regstat *current,int i)
1943 {
1944   current->isconst=0;
1945   current->wasconst=0;
1946   regs[i].wasconst=0;
1947   minimum_free_regs[i]=HOST_REGS;
1948   alloc_all(current,i);
1949   alloc_cc(current,i);
1950   dirty_reg(current,CCREG);
1951   if(opcode[i]==3) // JAL
1952   {
1953     alloc_reg(current,i,31);
1954     dirty_reg(current,31);
1955   }
1956   if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
1957   {
1958     alloc_reg(current,i,rs1[i]);
1959     if (rt1[i]!=0) {
1960       alloc_reg(current,i,rt1[i]);
1961       dirty_reg(current,rt1[i]);
1962     }
1963   }
1964   if((opcode[i]&0x2E)==4) // BEQ/BNE/BEQL/BNEL
1965   {
1966     if(rs1[i]) alloc_reg(current,i,rs1[i]);
1967     if(rs2[i]) alloc_reg(current,i,rs2[i]);
1968     if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1969     {
1970       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1971       if(rs2[i]) alloc_reg64(current,i,rs2[i]);
1972     }
1973   }
1974   else
1975   if((opcode[i]&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
1976   {
1977     if(rs1[i]) alloc_reg(current,i,rs1[i]);
1978     if(!((current->is32>>rs1[i])&1))
1979     {
1980       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1981     }
1982   }
1983   else
1984   if(opcode[i]==0x11) // BC1
1985   {
1986     alloc_reg(current,i,FSREG);
1987     alloc_reg(current,i,CSREG);
1988   }
1989   //else ...
1990 }
1991
1992 add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e)
1993 {
1994   stubs[stubcount][0]=type;
1995   stubs[stubcount][1]=addr;
1996   stubs[stubcount][2]=retaddr;
1997   stubs[stubcount][3]=a;
1998   stubs[stubcount][4]=b;
1999   stubs[stubcount][5]=c;
2000   stubs[stubcount][6]=d;
2001   stubs[stubcount][7]=e;
2002   stubcount++;
2003 }
2004
2005 // Write out a single register
2006 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32)
2007 {
2008   int hr;
2009   for(hr=0;hr<HOST_REGS;hr++) {
2010     if(hr!=EXCLUDE_REG) {
2011       if((regmap[hr]&63)==r) {
2012         if((dirty>>hr)&1) {
2013           if(regmap[hr]<64) {
2014             emit_storereg(r,hr);
2015 #ifndef FORCE32
2016             if((is32>>regmap[hr])&1) {
2017               emit_sarimm(hr,31,hr);
2018               emit_storereg(r|64,hr);
2019             }
2020 #endif
2021           }else{
2022             emit_storereg(r|64,hr);
2023           }
2024         }
2025       }
2026     }
2027   }
2028 }
2029
2030 int mchecksum()
2031 {
2032   //if(!tracedebug) return 0;
2033   int i;
2034   int sum=0;
2035   for(i=0;i<2097152;i++) {
2036     unsigned int temp=sum;
2037     sum<<=1;
2038     sum|=(~temp)>>31;
2039     sum^=((u_int *)rdram)[i];
2040   }
2041   return sum;
2042 }
2043 int rchecksum()
2044 {
2045   int i;
2046   int sum=0;
2047   for(i=0;i<64;i++)
2048     sum^=((u_int *)reg)[i];
2049   return sum;
2050 }
2051 void rlist()
2052 {
2053   int i;
2054   printf("TRACE: ");
2055   for(i=0;i<32;i++)
2056     printf("r%d:%8x%8x ",i,((int *)(reg+i))[1],((int *)(reg+i))[0]);
2057   printf("\n");
2058 #ifndef DISABLE_COP1
2059   printf("TRACE: ");
2060   for(i=0;i<32;i++)
2061     printf("f%d:%8x%8x ",i,((int*)reg_cop1_simple[i])[1],*((int*)reg_cop1_simple[i]));
2062   printf("\n");
2063 #endif
2064 }
2065
2066 void enabletrace()
2067 {
2068   tracedebug=1;
2069 }
2070
2071 void memdebug(int i)
2072 {
2073   //printf("TRACE: count=%d next=%d (checksum %x) lo=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[LOREG]>>32),(int)reg[LOREG]);
2074   //printf("TRACE: count=%d next=%d (rchecksum %x)\n",Count,next_interupt,rchecksum());
2075   //rlist();
2076   //if(tracedebug) {
2077   //if(Count>=-2084597794) {
2078   if((signed int)Count>=-2084597794&&(signed int)Count<0) {
2079   //if(0) {
2080     printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
2081     //printf("TRACE: count=%d next=%d (checksum %x) Status=%x\n",Count,next_interupt,mchecksum(),Status);
2082     //printf("TRACE: count=%d next=%d (checksum %x) hi=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[HIREG]>>32),(int)reg[HIREG]);
2083     rlist();
2084     #ifdef __i386__
2085     printf("TRACE: %x\n",(&i)[-1]);
2086     #endif
2087     #ifdef __arm__
2088     int j;
2089     printf("TRACE: %x \n",(&j)[10]);
2090     printf("TRACE: %x %x %x %x %x %x %x %x %x %x %x %x %x %x %x %x %x %x %x %x\n",(&j)[1],(&j)[2],(&j)[3],(&j)[4],(&j)[5],(&j)[6],(&j)[7],(&j)[8],(&j)[9],(&j)[10],(&j)[11],(&j)[12],(&j)[13],(&j)[14],(&j)[15],(&j)[16],(&j)[17],(&j)[18],(&j)[19],(&j)[20]);
2091     #endif
2092     //fflush(stdout);
2093   }
2094   //printf("TRACE: %x\n",(&i)[-1]);
2095 }
2096
2097 void tlb_debug(u_int cause, u_int addr, u_int iaddr)
2098 {
2099   printf("TLB Exception: instruction=%x addr=%x cause=%x\n",iaddr, addr, cause);
2100 }
2101
2102 void alu_assemble(int i,struct regstat *i_regs)
2103 {
2104   if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
2105     if(rt1[i]) {
2106       signed char s1,s2,t;
2107       t=get_reg(i_regs->regmap,rt1[i]);
2108       if(t>=0) {
2109         s1=get_reg(i_regs->regmap,rs1[i]);
2110         s2=get_reg(i_regs->regmap,rs2[i]);
2111         if(rs1[i]&&rs2[i]) {
2112           assert(s1>=0);
2113           assert(s2>=0);
2114           if(opcode2[i]&2) emit_sub(s1,s2,t);
2115           else emit_add(s1,s2,t);
2116         }
2117         else if(rs1[i]) {
2118           if(s1>=0) emit_mov(s1,t);
2119           else emit_loadreg(rs1[i],t);
2120         }
2121         else if(rs2[i]) {
2122           if(s2>=0) {
2123             if(opcode2[i]&2) emit_neg(s2,t);
2124             else emit_mov(s2,t);
2125           }
2126           else {
2127             emit_loadreg(rs2[i],t);
2128             if(opcode2[i]&2) emit_neg(t,t);
2129           }
2130         }
2131         else emit_zeroreg(t);
2132       }
2133     }
2134   }
2135   if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2136     if(rt1[i]) {
2137       signed char s1l,s2l,s1h,s2h,tl,th;
2138       tl=get_reg(i_regs->regmap,rt1[i]);
2139       th=get_reg(i_regs->regmap,rt1[i]|64);
2140       if(tl>=0) {
2141         s1l=get_reg(i_regs->regmap,rs1[i]);
2142         s2l=get_reg(i_regs->regmap,rs2[i]);
2143         s1h=get_reg(i_regs->regmap,rs1[i]|64);
2144         s2h=get_reg(i_regs->regmap,rs2[i]|64);
2145         if(rs1[i]&&rs2[i]) {
2146           assert(s1l>=0);
2147           assert(s2l>=0);
2148           if(opcode2[i]&2) emit_subs(s1l,s2l,tl);
2149           else emit_adds(s1l,s2l,tl);
2150           if(th>=0) {
2151             #ifdef INVERTED_CARRY
2152             if(opcode2[i]&2) {if(s1h!=th) emit_mov(s1h,th);emit_sbb(th,s2h);}
2153             #else
2154             if(opcode2[i]&2) emit_sbc(s1h,s2h,th);
2155             #endif
2156             else emit_add(s1h,s2h,th);
2157           }
2158         }
2159         else if(rs1[i]) {
2160           if(s1l>=0) emit_mov(s1l,tl);
2161           else emit_loadreg(rs1[i],tl);
2162           if(th>=0) {
2163             if(s1h>=0) emit_mov(s1h,th);
2164             else emit_loadreg(rs1[i]|64,th);
2165           }
2166         }
2167         else if(rs2[i]) {
2168           if(s2l>=0) {
2169             if(opcode2[i]&2) emit_negs(s2l,tl);
2170             else emit_mov(s2l,tl);
2171           }
2172           else {
2173             emit_loadreg(rs2[i],tl);
2174             if(opcode2[i]&2) emit_negs(tl,tl);
2175           }
2176           if(th>=0) {
2177             #ifdef INVERTED_CARRY
2178             if(s2h>=0) emit_mov(s2h,th);
2179             else emit_loadreg(rs2[i]|64,th);
2180             if(opcode2[i]&2) {
2181               emit_adcimm(-1,th); // x86 has inverted carry flag
2182               emit_not(th,th);
2183             }
2184             #else
2185             if(opcode2[i]&2) {
2186               if(s2h>=0) emit_rscimm(s2h,0,th);
2187               else {
2188                 emit_loadreg(rs2[i]|64,th);
2189                 emit_rscimm(th,0,th);
2190               }
2191             }else{
2192               if(s2h>=0) emit_mov(s2h,th);
2193               else emit_loadreg(rs2[i]|64,th);
2194             }
2195             #endif
2196           }
2197         }
2198         else {
2199           emit_zeroreg(tl);
2200           if(th>=0) emit_zeroreg(th);
2201         }
2202       }
2203     }
2204   }
2205   if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
2206     if(rt1[i]) {
2207       signed char s1l,s1h,s2l,s2h,t;
2208       if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1))
2209       {
2210         t=get_reg(i_regs->regmap,rt1[i]);
2211         //assert(t>=0);
2212         if(t>=0) {
2213           s1l=get_reg(i_regs->regmap,rs1[i]);
2214           s1h=get_reg(i_regs->regmap,rs1[i]|64);
2215           s2l=get_reg(i_regs->regmap,rs2[i]);
2216           s2h=get_reg(i_regs->regmap,rs2[i]|64);
2217           if(rs2[i]==0) // rx<r0
2218           {
2219             assert(s1h>=0);
2220             if(opcode2[i]==0x2a) // SLT
2221               emit_shrimm(s1h,31,t);
2222             else // SLTU (unsigned can not be less than zero)
2223               emit_zeroreg(t);
2224           }
2225           else if(rs1[i]==0) // r0<rx
2226           {
2227             assert(s2h>=0);
2228             if(opcode2[i]==0x2a) // SLT
2229               emit_set_gz64_32(s2h,s2l,t);
2230             else // SLTU (set if not zero)
2231               emit_set_nz64_32(s2h,s2l,t);
2232           }
2233           else {
2234             assert(s1l>=0);assert(s1h>=0);
2235             assert(s2l>=0);assert(s2h>=0);
2236             if(opcode2[i]==0x2a) // SLT
2237               emit_set_if_less64_32(s1h,s1l,s2h,s2l,t);
2238             else // SLTU
2239               emit_set_if_carry64_32(s1h,s1l,s2h,s2l,t);
2240           }
2241         }
2242       } else {
2243         t=get_reg(i_regs->regmap,rt1[i]);
2244         //assert(t>=0);
2245         if(t>=0) {
2246           s1l=get_reg(i_regs->regmap,rs1[i]);
2247           s2l=get_reg(i_regs->regmap,rs2[i]);
2248           if(rs2[i]==0) // rx<r0
2249           {
2250             assert(s1l>=0);
2251             if(opcode2[i]==0x2a) // SLT
2252               emit_shrimm(s1l,31,t);
2253             else // SLTU (unsigned can not be less than zero)
2254               emit_zeroreg(t);
2255           }
2256           else if(rs1[i]==0) // r0<rx
2257           {
2258             assert(s2l>=0);
2259             if(opcode2[i]==0x2a) // SLT
2260               emit_set_gz32(s2l,t);
2261             else // SLTU (set if not zero)
2262               emit_set_nz32(s2l,t);
2263           }
2264           else{
2265             assert(s1l>=0);assert(s2l>=0);
2266             if(opcode2[i]==0x2a) // SLT
2267               emit_set_if_less32(s1l,s2l,t);
2268             else // SLTU
2269               emit_set_if_carry32(s1l,s2l,t);
2270           }
2271         }
2272       }
2273     }
2274   }
2275   if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
2276     if(rt1[i]) {
2277       signed char s1l,s1h,s2l,s2h,th,tl;
2278       tl=get_reg(i_regs->regmap,rt1[i]);
2279       th=get_reg(i_regs->regmap,rt1[i]|64);
2280       if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1)&&th>=0)
2281       {
2282         assert(tl>=0);
2283         if(tl>=0) {
2284           s1l=get_reg(i_regs->regmap,rs1[i]);
2285           s1h=get_reg(i_regs->regmap,rs1[i]|64);
2286           s2l=get_reg(i_regs->regmap,rs2[i]);
2287           s2h=get_reg(i_regs->regmap,rs2[i]|64);
2288           if(rs1[i]&&rs2[i]) {
2289             assert(s1l>=0);assert(s1h>=0);
2290             assert(s2l>=0);assert(s2h>=0);
2291             if(opcode2[i]==0x24) { // AND
2292               emit_and(s1l,s2l,tl);
2293               emit_and(s1h,s2h,th);
2294             } else
2295             if(opcode2[i]==0x25) { // OR
2296               emit_or(s1l,s2l,tl);
2297               emit_or(s1h,s2h,th);
2298             } else
2299             if(opcode2[i]==0x26) { // XOR
2300               emit_xor(s1l,s2l,tl);
2301               emit_xor(s1h,s2h,th);
2302             } else
2303             if(opcode2[i]==0x27) { // NOR
2304               emit_or(s1l,s2l,tl);
2305               emit_or(s1h,s2h,th);
2306               emit_not(tl,tl);
2307               emit_not(th,th);
2308             }
2309           }
2310           else
2311           {
2312             if(opcode2[i]==0x24) { // AND
2313               emit_zeroreg(tl);
2314               emit_zeroreg(th);
2315             } else
2316             if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2317               if(rs1[i]){
2318                 if(s1l>=0) emit_mov(s1l,tl);
2319                 else emit_loadreg(rs1[i],tl);
2320                 if(s1h>=0) emit_mov(s1h,th);
2321                 else emit_loadreg(rs1[i]|64,th);
2322               }
2323               else
2324               if(rs2[i]){
2325                 if(s2l>=0) emit_mov(s2l,tl);
2326                 else emit_loadreg(rs2[i],tl);
2327                 if(s2h>=0) emit_mov(s2h,th);
2328                 else emit_loadreg(rs2[i]|64,th);
2329               }
2330               else{
2331                 emit_zeroreg(tl);
2332                 emit_zeroreg(th);
2333               }
2334             } else
2335             if(opcode2[i]==0x27) { // NOR
2336               if(rs1[i]){
2337                 if(s1l>=0) emit_not(s1l,tl);
2338                 else{
2339                   emit_loadreg(rs1[i],tl);
2340                   emit_not(tl,tl);
2341                 }
2342                 if(s1h>=0) emit_not(s1h,th);
2343                 else{
2344                   emit_loadreg(rs1[i]|64,th);
2345                   emit_not(th,th);
2346                 }
2347               }
2348               else
2349               if(rs2[i]){
2350                 if(s2l>=0) emit_not(s2l,tl);
2351                 else{
2352                   emit_loadreg(rs2[i],tl);
2353                   emit_not(tl,tl);
2354                 }
2355                 if(s2h>=0) emit_not(s2h,th);
2356                 else{
2357                   emit_loadreg(rs2[i]|64,th);
2358                   emit_not(th,th);
2359                 }
2360               }
2361               else {
2362                 emit_movimm(-1,tl);
2363                 emit_movimm(-1,th);
2364               }
2365             }
2366           }
2367         }
2368       }
2369       else
2370       {
2371         // 32 bit
2372         if(tl>=0) {
2373           s1l=get_reg(i_regs->regmap,rs1[i]);
2374           s2l=get_reg(i_regs->regmap,rs2[i]);
2375           if(rs1[i]&&rs2[i]) {
2376             assert(s1l>=0);
2377             assert(s2l>=0);
2378             if(opcode2[i]==0x24) { // AND
2379               emit_and(s1l,s2l,tl);
2380             } else
2381             if(opcode2[i]==0x25) { // OR
2382               emit_or(s1l,s2l,tl);
2383             } else
2384             if(opcode2[i]==0x26) { // XOR
2385               emit_xor(s1l,s2l,tl);
2386             } else
2387             if(opcode2[i]==0x27) { // NOR
2388               emit_or(s1l,s2l,tl);
2389               emit_not(tl,tl);
2390             }
2391           }
2392           else
2393           {
2394             if(opcode2[i]==0x24) { // AND
2395               emit_zeroreg(tl);
2396             } else
2397             if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2398               if(rs1[i]){
2399                 if(s1l>=0) emit_mov(s1l,tl);
2400                 else emit_loadreg(rs1[i],tl); // CHECK: regmap_entry?
2401               }
2402               else
2403               if(rs2[i]){
2404                 if(s2l>=0) emit_mov(s2l,tl);
2405                 else emit_loadreg(rs2[i],tl); // CHECK: regmap_entry?
2406               }
2407               else emit_zeroreg(tl);
2408             } else
2409             if(opcode2[i]==0x27) { // NOR
2410               if(rs1[i]){
2411                 if(s1l>=0) emit_not(s1l,tl);
2412                 else {
2413                   emit_loadreg(rs1[i],tl);
2414                   emit_not(tl,tl);
2415                 }
2416               }
2417               else
2418               if(rs2[i]){
2419                 if(s2l>=0) emit_not(s2l,tl);
2420                 else {
2421                   emit_loadreg(rs2[i],tl);
2422                   emit_not(tl,tl);
2423                 }
2424               }
2425               else emit_movimm(-1,tl);
2426             }
2427           }
2428         }
2429       }
2430     }
2431   }
2432 }
2433
2434 void imm16_assemble(int i,struct regstat *i_regs)
2435 {
2436   if (opcode[i]==0x0f) { // LUI
2437     if(rt1[i]) {
2438       signed char t;
2439       t=get_reg(i_regs->regmap,rt1[i]);
2440       //assert(t>=0);
2441       if(t>=0) {
2442         if(!((i_regs->isconst>>t)&1))
2443           emit_movimm(imm[i]<<16,t);
2444       }
2445     }
2446   }
2447   if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
2448     if(rt1[i]) {
2449       signed char s,t;
2450       t=get_reg(i_regs->regmap,rt1[i]);
2451       s=get_reg(i_regs->regmap,rs1[i]);
2452       if(rs1[i]) {
2453         //assert(t>=0);
2454         //assert(s>=0);
2455         if(t>=0) {
2456           if(!((i_regs->isconst>>t)&1)) {
2457             if(s<0) {
2458               if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2459               emit_addimm(t,imm[i],t);
2460             }else{
2461               if(!((i_regs->wasconst>>s)&1))
2462                 emit_addimm(s,imm[i],t);
2463               else
2464                 emit_movimm(constmap[i][s]+imm[i],t);
2465             }
2466           }
2467         }
2468       } else {
2469         if(t>=0) {
2470           if(!((i_regs->isconst>>t)&1))
2471             emit_movimm(imm[i],t);
2472         }
2473       }
2474     }
2475   }
2476   if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
2477     if(rt1[i]) {
2478       signed char sh,sl,th,tl;
2479       th=get_reg(i_regs->regmap,rt1[i]|64);
2480       tl=get_reg(i_regs->regmap,rt1[i]);
2481       sh=get_reg(i_regs->regmap,rs1[i]|64);
2482       sl=get_reg(i_regs->regmap,rs1[i]);
2483       if(tl>=0) {
2484         if(rs1[i]) {
2485           assert(sh>=0);
2486           assert(sl>=0);
2487           if(th>=0) {
2488             emit_addimm64_32(sh,sl,imm[i],th,tl);
2489           }
2490           else {
2491             emit_addimm(sl,imm[i],tl);
2492           }
2493         } else {
2494           emit_movimm(imm[i],tl);
2495           if(th>=0) emit_movimm(((signed int)imm[i])>>31,th);
2496         }
2497       }
2498     }
2499   }
2500   else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
2501     if(rt1[i]) {
2502       //assert(rs1[i]!=0); // r0 might be valid, but it's probably a bug
2503       signed char sh,sl,t;
2504       t=get_reg(i_regs->regmap,rt1[i]);
2505       sh=get_reg(i_regs->regmap,rs1[i]|64);
2506       sl=get_reg(i_regs->regmap,rs1[i]);
2507       //assert(t>=0);
2508       if(t>=0) {
2509         if(rs1[i]>0) {
2510           if(sh<0) assert((i_regs->was32>>rs1[i])&1);
2511           if(sh<0||((i_regs->was32>>rs1[i])&1)) {
2512             if(opcode[i]==0x0a) { // SLTI
2513               if(sl<0) {
2514                 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2515                 emit_slti32(t,imm[i],t);
2516               }else{
2517                 emit_slti32(sl,imm[i],t);
2518               }
2519             }
2520             else { // SLTIU
2521               if(sl<0) {
2522                 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2523                 emit_sltiu32(t,imm[i],t);
2524               }else{
2525                 emit_sltiu32(sl,imm[i],t);
2526               }
2527             }
2528           }else{ // 64-bit
2529             assert(sl>=0);
2530             if(opcode[i]==0x0a) // SLTI
2531               emit_slti64_32(sh,sl,imm[i],t);
2532             else // SLTIU
2533               emit_sltiu64_32(sh,sl,imm[i],t);
2534           }
2535         }else{
2536           // SLTI(U) with r0 is just stupid,
2537           // nonetheless examples can be found
2538           if(opcode[i]==0x0a) // SLTI
2539             if(0<imm[i]) emit_movimm(1,t);
2540             else emit_zeroreg(t);
2541           else // SLTIU
2542           {
2543             if(imm[i]) emit_movimm(1,t);
2544             else emit_zeroreg(t);
2545           }
2546         }
2547       }
2548     }
2549   }
2550   else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
2551     if(rt1[i]) {
2552       signed char sh,sl,th,tl;
2553       th=get_reg(i_regs->regmap,rt1[i]|64);
2554       tl=get_reg(i_regs->regmap,rt1[i]);
2555       sh=get_reg(i_regs->regmap,rs1[i]|64);
2556       sl=get_reg(i_regs->regmap,rs1[i]);
2557       if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2558         if(opcode[i]==0x0c) //ANDI
2559         {
2560           if(rs1[i]) {
2561             if(sl<0) {
2562               if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2563               emit_andimm(tl,imm[i],tl);
2564             }else{
2565               if(!((i_regs->wasconst>>sl)&1))
2566                 emit_andimm(sl,imm[i],tl);
2567               else
2568                 emit_movimm(constmap[i][sl]&imm[i],tl);
2569             }
2570           }
2571           else
2572             emit_zeroreg(tl);
2573           if(th>=0) emit_zeroreg(th);
2574         }
2575         else
2576         {
2577           if(rs1[i]) {
2578             if(sl<0) {
2579               if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2580             }
2581             if(th>=0) {
2582               if(sh<0) {
2583                 emit_loadreg(rs1[i]|64,th);
2584               }else{
2585                 emit_mov(sh,th);
2586               }
2587             }
2588             if(opcode[i]==0x0d) //ORI
2589             if(sl<0) {
2590               emit_orimm(tl,imm[i],tl);
2591             }else{
2592               if(!((i_regs->wasconst>>sl)&1))
2593                 emit_orimm(sl,imm[i],tl);
2594               else
2595                 emit_movimm(constmap[i][sl]|imm[i],tl);
2596             }
2597             if(opcode[i]==0x0e) //XORI
2598             if(sl<0) {
2599               emit_xorimm(tl,imm[i],tl);
2600             }else{
2601               if(!((i_regs->wasconst>>sl)&1))
2602                 emit_xorimm(sl,imm[i],tl);
2603               else
2604                 emit_movimm(constmap[i][sl]^imm[i],tl);
2605             }
2606           }
2607           else {
2608             emit_movimm(imm[i],tl);
2609             if(th>=0) emit_zeroreg(th);
2610           }
2611         }
2612       }
2613     }
2614   }
2615 }
2616
2617 void shiftimm_assemble(int i,struct regstat *i_regs)
2618 {
2619   if(opcode2[i]<=0x3) // SLL/SRL/SRA
2620   {
2621     if(rt1[i]) {
2622       signed char s,t;
2623       t=get_reg(i_regs->regmap,rt1[i]);
2624       s=get_reg(i_regs->regmap,rs1[i]);
2625       //assert(t>=0);
2626       if(t>=0){
2627         if(rs1[i]==0)
2628         {
2629           emit_zeroreg(t);
2630         }
2631         else
2632         {
2633           if(s<0&&i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2634           if(imm[i]) {
2635             if(opcode2[i]==0) // SLL
2636             {
2637               emit_shlimm(s<0?t:s,imm[i],t);
2638             }
2639             if(opcode2[i]==2) // SRL
2640             {
2641               emit_shrimm(s<0?t:s,imm[i],t);
2642             }
2643             if(opcode2[i]==3) // SRA
2644             {
2645               emit_sarimm(s<0?t:s,imm[i],t);
2646             }
2647           }else{
2648             // Shift by zero
2649             if(s>=0 && s!=t) emit_mov(s,t);
2650           }
2651         }
2652       }
2653       //emit_storereg(rt1[i],t); //DEBUG
2654     }
2655   }
2656   if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
2657   {
2658     if(rt1[i]) {
2659       signed char sh,sl,th,tl;
2660       th=get_reg(i_regs->regmap,rt1[i]|64);
2661       tl=get_reg(i_regs->regmap,rt1[i]);
2662       sh=get_reg(i_regs->regmap,rs1[i]|64);
2663       sl=get_reg(i_regs->regmap,rs1[i]);
2664       if(tl>=0) {
2665         if(rs1[i]==0)
2666         {
2667           emit_zeroreg(tl);
2668           if(th>=0) emit_zeroreg(th);
2669         }
2670         else
2671         {
2672           assert(sl>=0);
2673           assert(sh>=0);
2674           if(imm[i]) {
2675             if(opcode2[i]==0x38) // DSLL
2676             {
2677               if(th>=0) emit_shldimm(sh,sl,imm[i],th);
2678               emit_shlimm(sl,imm[i],tl);
2679             }
2680             if(opcode2[i]==0x3a) // DSRL
2681             {
2682               emit_shrdimm(sl,sh,imm[i],tl);
2683               if(th>=0) emit_shrimm(sh,imm[i],th);
2684             }
2685             if(opcode2[i]==0x3b) // DSRA
2686             {
2687               emit_shrdimm(sl,sh,imm[i],tl);
2688               if(th>=0) emit_sarimm(sh,imm[i],th);
2689             }
2690           }else{
2691             // Shift by zero
2692             if(sl!=tl) emit_mov(sl,tl);
2693             if(th>=0&&sh!=th) emit_mov(sh,th);
2694           }
2695         }
2696       }
2697     }
2698   }
2699   if(opcode2[i]==0x3c) // DSLL32
2700   {
2701     if(rt1[i]) {
2702       signed char sl,tl,th;
2703       tl=get_reg(i_regs->regmap,rt1[i]);
2704       th=get_reg(i_regs->regmap,rt1[i]|64);
2705       sl=get_reg(i_regs->regmap,rs1[i]);
2706       if(th>=0||tl>=0){
2707         assert(tl>=0);
2708         assert(th>=0);
2709         assert(sl>=0);
2710         emit_mov(sl,th);
2711         emit_zeroreg(tl);
2712         if(imm[i]>32)
2713         {
2714           emit_shlimm(th,imm[i]&31,th);
2715         }
2716       }
2717     }
2718   }
2719   if(opcode2[i]==0x3e) // DSRL32
2720   {
2721     if(rt1[i]) {
2722       signed char sh,tl,th;
2723       tl=get_reg(i_regs->regmap,rt1[i]);
2724       th=get_reg(i_regs->regmap,rt1[i]|64);
2725       sh=get_reg(i_regs->regmap,rs1[i]|64);
2726       if(tl>=0){
2727         assert(sh>=0);
2728         emit_mov(sh,tl);
2729         if(th>=0) emit_zeroreg(th);
2730         if(imm[i]>32)
2731         {
2732           emit_shrimm(tl,imm[i]&31,tl);
2733         }
2734       }
2735     }
2736   }
2737   if(opcode2[i]==0x3f) // DSRA32
2738   {
2739     if(rt1[i]) {
2740       signed char sh,tl;
2741       tl=get_reg(i_regs->regmap,rt1[i]);
2742       sh=get_reg(i_regs->regmap,rs1[i]|64);
2743       if(tl>=0){
2744         assert(sh>=0);
2745         emit_mov(sh,tl);
2746         if(imm[i]>32)
2747         {
2748           emit_sarimm(tl,imm[i]&31,tl);
2749         }
2750       }
2751     }
2752   }
2753 }
2754
2755 #ifndef shift_assemble
2756 void shift_assemble(int i,struct regstat *i_regs)
2757 {
2758   printf("Need shift_assemble for this architecture.\n");
2759   exit(1);
2760 }
2761 #endif
2762
2763 void load_assemble(int i,struct regstat *i_regs)
2764 {
2765   int s,th,tl,addr,map=-1;
2766   int offset;
2767   int jaddr=0;
2768   int memtarget=0,c=0;
2769   u_int hr,reglist=0;
2770   th=get_reg(i_regs->regmap,rt1[i]|64);
2771   tl=get_reg(i_regs->regmap,rt1[i]);
2772   s=get_reg(i_regs->regmap,rs1[i]);
2773   offset=imm[i];
2774   for(hr=0;hr<HOST_REGS;hr++) {
2775     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2776   }
2777   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2778   if(s>=0) {
2779     c=(i_regs->wasconst>>s)&1;
2780     if (c) {
2781       memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2782       if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
2783     }
2784   }
2785   //printf("load_assemble: c=%d\n",c);
2786   //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2787   // FIXME: Even if the load is a NOP, we should check for pagefaults...
2788 #ifdef PCSX
2789   if(tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80)
2790     ||rt1[i]==0) {
2791       // could be FIFO, must perform the read
2792       // ||dummy read
2793       assem_debug("(forced read)\n");
2794       tl=get_reg(i_regs->regmap,-1);
2795       assert(tl>=0);
2796   }
2797 #endif
2798   if(offset||s<0||c) addr=tl;
2799   else addr=s;
2800   //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2801  if(tl>=0) {
2802   //printf("load_assemble: c=%d\n",c);
2803   //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2804   assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2805   reglist&=~(1<<tl);
2806   if(th>=0) reglist&=~(1<<th);
2807   if(!using_tlb) {
2808     if(!c) {
2809       #ifdef RAM_OFFSET
2810       map=get_reg(i_regs->regmap,ROREG);
2811       if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG);
2812       #endif
2813 //#define R29_HACK 1
2814       #ifdef R29_HACK
2815       // Strmnnrmn's speed hack
2816       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2817       #endif
2818       {
2819         #ifdef PCSX
2820         if(sp_in_mirror&&rs1[i]==29) {
2821           emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2822           emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
2823         }
2824         else
2825         #endif
2826         emit_cmpimm(addr,RAM_SIZE);
2827         jaddr=(int)out;
2828         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2829         // Hint to branch predictor that the branch is unlikely to be taken
2830         if(rs1[i]>=28)
2831           emit_jno_unlikely(0);
2832         else
2833         #endif
2834         emit_jno(0);
2835       }
2836     }
2837   }else{ // using tlb
2838     int x=0;
2839     if (opcode[i]==0x20||opcode[i]==0x24) x=3; // LB/LBU
2840     if (opcode[i]==0x21||opcode[i]==0x25) x=2; // LH/LHU
2841     map=get_reg(i_regs->regmap,TLREG);
2842     assert(map>=0);
2843     reglist&=~(1<<map);
2844     map=do_tlb_r(addr,tl,map,x,-1,-1,c,constmap[i][s]+offset);
2845     do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr);
2846   }
2847   int dummy=(rt1[i]==0)||(tl!=get_reg(i_regs->regmap,rt1[i])); // ignore loads to r0 and unneeded reg
2848   if (opcode[i]==0x20) { // LB
2849     if(!c||memtarget) {
2850       if(!dummy) {
2851         #ifdef HOST_IMM_ADDR32
2852         if(c)
2853           emit_movsbl_tlb((constmap[i][s]+offset)^3,map,tl);
2854         else
2855         #endif
2856         {
2857           //emit_xorimm(addr,3,tl);
2858           //gen_tlb_addr_r(tl,map);
2859           //emit_movsbl_indexed((int)rdram-0x80000000,tl,tl);
2860           int x=0,a=tl;
2861 #ifdef BIG_ENDIAN_MIPS
2862           if(!c) emit_xorimm(addr,3,tl);
2863           else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2864 #else
2865           if(!c) a=addr;
2866 #endif
2867 #ifdef PCSX
2868           if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
2869 #endif
2870           emit_movsbl_indexed_tlb(x,a,map,tl);
2871         }
2872       }
2873       if(jaddr)
2874         add_stub(LOADB_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2875     }
2876     else
2877       inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2878   }
2879   if (opcode[i]==0x21) { // LH
2880     if(!c||memtarget) {
2881       if(!dummy) {
2882         #ifdef HOST_IMM_ADDR32
2883         if(c)
2884           emit_movswl_tlb((constmap[i][s]+offset)^2,map,tl);
2885         else
2886         #endif
2887         {
2888           int x=0,a=tl;
2889 #ifdef BIG_ENDIAN_MIPS
2890           if(!c) emit_xorimm(addr,2,tl);
2891           else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2892 #else
2893           if(!c) a=addr;
2894 #endif
2895 #ifdef PCSX
2896           if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
2897 #endif
2898           //#ifdef
2899           //emit_movswl_indexed_tlb(x,tl,map,tl);
2900           //else
2901           if(map>=0) {
2902             gen_tlb_addr_r(a,map);
2903             emit_movswl_indexed(x,a,tl);
2904           }else{
2905             #ifdef RAM_OFFSET
2906             emit_movswl_indexed(x,a,tl);
2907             #else
2908             emit_movswl_indexed((int)rdram-0x80000000+x,a,tl);
2909             #endif
2910           }
2911         }
2912       }
2913       if(jaddr)
2914         add_stub(LOADH_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2915     }
2916     else
2917       inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2918   }
2919   if (opcode[i]==0x23) { // LW
2920     if(!c||memtarget) {
2921       if(!dummy) {
2922         int a=addr;
2923 #ifdef PCSX
2924         if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
2925 #endif
2926         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2927         #ifdef HOST_IMM_ADDR32
2928         if(c)
2929           emit_readword_tlb(constmap[i][s]+offset,map,tl);
2930         else
2931         #endif
2932         emit_readword_indexed_tlb(0,a,map,tl);
2933       }
2934       if(jaddr)
2935         add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2936     }
2937     else
2938       inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2939   }
2940   if (opcode[i]==0x24) { // LBU
2941     if(!c||memtarget) {
2942       if(!dummy) {
2943         #ifdef HOST_IMM_ADDR32
2944         if(c)
2945           emit_movzbl_tlb((constmap[i][s]+offset)^3,map,tl);
2946         else
2947         #endif
2948         {
2949           //emit_xorimm(addr,3,tl);
2950           //gen_tlb_addr_r(tl,map);
2951           //emit_movzbl_indexed((int)rdram-0x80000000,tl,tl);
2952           int x=0,a=tl;
2953 #ifdef BIG_ENDIAN_MIPS
2954           if(!c) emit_xorimm(addr,3,tl);
2955           else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2956 #else
2957           if(!c) a=addr;
2958 #endif
2959 #ifdef PCSX
2960           if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
2961 #endif
2962           emit_movzbl_indexed_tlb(x,a,map,tl);
2963         }
2964       }
2965       if(jaddr)
2966         add_stub(LOADBU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2967     }
2968     else
2969       inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2970   }
2971   if (opcode[i]==0x25) { // LHU
2972     if(!c||memtarget) {
2973       if(!dummy) {
2974         #ifdef HOST_IMM_ADDR32
2975         if(c)
2976           emit_movzwl_tlb((constmap[i][s]+offset)^2,map,tl);
2977         else
2978         #endif
2979         {
2980           int x=0,a=tl;
2981 #ifdef BIG_ENDIAN_MIPS
2982           if(!c) emit_xorimm(addr,2,tl);
2983           else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2984 #else
2985           if(!c) a=addr;
2986 #endif
2987 #ifdef PCSX
2988           if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
2989 #endif
2990           //#ifdef
2991           //emit_movzwl_indexed_tlb(x,tl,map,tl);
2992           //#else
2993           if(map>=0) {
2994             gen_tlb_addr_r(a,map);
2995             emit_movzwl_indexed(x,a,tl);
2996           }else{
2997             #ifdef RAM_OFFSET
2998             emit_movzwl_indexed(x,a,tl);
2999             #else
3000             emit_movzwl_indexed((int)rdram-0x80000000+x,a,tl);
3001             #endif
3002           }
3003         }
3004       }
3005       if(jaddr)
3006         add_stub(LOADHU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3007     }
3008     else
3009       inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
3010   }
3011   if (opcode[i]==0x27) { // LWU
3012     assert(th>=0);
3013     if(!c||memtarget) {
3014       if(!dummy) {
3015         int a=addr;
3016 #ifdef PCSX
3017         if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
3018 #endif
3019         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
3020         #ifdef HOST_IMM_ADDR32
3021         if(c)
3022           emit_readword_tlb(constmap[i][s]+offset,map,tl);
3023         else
3024         #endif
3025         emit_readword_indexed_tlb(0,a,map,tl);
3026       }
3027       if(jaddr)
3028         add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3029     }
3030     else {
3031       inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
3032     }
3033     emit_zeroreg(th);
3034   }
3035   if (opcode[i]==0x37) { // LD
3036     if(!c||memtarget) {
3037       if(!dummy) {
3038         int a=addr;
3039 #ifdef PCSX
3040         if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
3041 #endif
3042         //gen_tlb_addr_r(tl,map);
3043         //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,addr,th);
3044         //emit_readword_indexed((int)rdram-0x7FFFFFFC,addr,tl);
3045         #ifdef HOST_IMM_ADDR32
3046         if(c)
3047           emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
3048         else
3049         #endif
3050         emit_readdword_indexed_tlb(0,a,map,th,tl);
3051       }
3052       if(jaddr)
3053         add_stub(LOADD_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3054     }
3055     else
3056       inline_readstub(LOADD_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
3057   }
3058  }
3059   //emit_storereg(rt1[i],tl); // DEBUG
3060   //if(opcode[i]==0x23)
3061   //if(opcode[i]==0x24)
3062   //if(opcode[i]==0x23||opcode[i]==0x24)
3063   /*if(opcode[i]==0x21||opcode[i]==0x23||opcode[i]==0x24)
3064   {
3065     //emit_pusha();
3066     save_regs(0x100f);
3067         emit_readword((int)&last_count,ECX);
3068         #ifdef __i386__
3069         if(get_reg(i_regs->regmap,CCREG)<0)
3070           emit_loadreg(CCREG,HOST_CCREG);
3071         emit_add(HOST_CCREG,ECX,HOST_CCREG);
3072         emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3073         emit_writeword(HOST_CCREG,(int)&Count);
3074         #endif
3075         #ifdef __arm__
3076         if(get_reg(i_regs->regmap,CCREG)<0)
3077           emit_loadreg(CCREG,0);
3078         else
3079           emit_mov(HOST_CCREG,0);
3080         emit_add(0,ECX,0);
3081         emit_addimm(0,2*ccadj[i],0);
3082         emit_writeword(0,(int)&Count);
3083         #endif
3084     emit_call((int)memdebug);
3085     //emit_popa();
3086     restore_regs(0x100f);
3087   }/**/
3088 }
3089
3090 #ifndef loadlr_assemble
3091 void loadlr_assemble(int i,struct regstat *i_regs)
3092 {
3093   printf("Need loadlr_assemble for this architecture.\n");
3094   exit(1);
3095 }
3096 #endif
3097
3098 void store_assemble(int i,struct regstat *i_regs)
3099 {
3100   int s,th,tl,map=-1;
3101   int addr,temp;
3102   int offset;
3103   int jaddr=0,jaddr2,type;
3104   int memtarget=0,c=0;
3105   int agr=AGEN1+(i&1);
3106   u_int hr,reglist=0;
3107   th=get_reg(i_regs->regmap,rs2[i]|64);
3108   tl=get_reg(i_regs->regmap,rs2[i]);
3109   s=get_reg(i_regs->regmap,rs1[i]);
3110   temp=get_reg(i_regs->regmap,agr);
3111   if(temp<0) temp=get_reg(i_regs->regmap,-1);
3112   offset=imm[i];
3113   if(s>=0) {
3114     c=(i_regs->wasconst>>s)&1;
3115     if(c) {
3116       memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3117       if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
3118     }
3119   }
3120   assert(tl>=0);
3121   assert(temp>=0);
3122   for(hr=0;hr<HOST_REGS;hr++) {
3123     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3124   }
3125   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3126   if(offset||s<0||c) addr=temp;
3127   else addr=s;
3128   if(!using_tlb) {
3129     if(!c) {
3130       #ifdef PCSX
3131       if(sp_in_mirror&&rs1[i]==29) {
3132         emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
3133         emit_cmpimm(HOST_TEMPREG,RAM_SIZE);
3134       }
3135       else
3136       #endif
3137       #ifdef R29_HACK
3138       // Strmnnrmn's speed hack
3139       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3140       #endif
3141       emit_cmpimm(addr,RAM_SIZE);
3142       #ifdef DESTRUCTIVE_SHIFT
3143       if(s==addr) emit_mov(s,temp);
3144       #endif
3145       #ifdef R29_HACK
3146       memtarget=1;
3147       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3148       #endif
3149       {
3150         jaddr=(int)out;
3151         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
3152         // Hint to branch predictor that the branch is unlikely to be taken
3153         if(rs1[i]>=28)
3154           emit_jno_unlikely(0);
3155         else
3156         #endif
3157         emit_jno(0);
3158       }
3159     }
3160   }else{ // using tlb
3161     int x=0;
3162     if (opcode[i]==0x28) x=3; // SB
3163     if (opcode[i]==0x29) x=2; // SH
3164     map=get_reg(i_regs->regmap,TLREG);
3165     assert(map>=0);
3166     reglist&=~(1<<map);
3167     map=do_tlb_w(addr,temp,map,x,c,constmap[i][s]+offset);
3168     do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
3169   }
3170
3171   if (opcode[i]==0x28) { // SB
3172     if(!c||memtarget) {
3173       int x=0,a=temp;
3174 #ifdef BIG_ENDIAN_MIPS
3175       if(!c) emit_xorimm(addr,3,temp);
3176       else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
3177 #else
3178       if(!c) a=addr;
3179 #endif
3180 #ifdef PCSX
3181       if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
3182 #endif
3183       //gen_tlb_addr_w(temp,map);
3184       //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
3185       emit_writebyte_indexed_tlb(tl,x,a,map,a);
3186     }
3187     type=STOREB_STUB;
3188   }
3189   if (opcode[i]==0x29) { // SH
3190     if(!c||memtarget) {
3191       int x=0,a=temp;
3192 #ifdef BIG_ENDIAN_MIPS
3193       if(!c) emit_xorimm(addr,2,temp);
3194       else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
3195 #else
3196       if(!c) a=addr;
3197 #endif
3198 #ifdef PCSX
3199       if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
3200 #endif
3201       //#ifdef
3202       //emit_writehword_indexed_tlb(tl,x,temp,map,temp);
3203       //#else
3204       if(map>=0) {
3205         gen_tlb_addr_w(a,map);
3206         emit_writehword_indexed(tl,x,a);
3207       }else
3208         emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
3209     }
3210     type=STOREH_STUB;
3211   }
3212   if (opcode[i]==0x2B) { // SW
3213     if(!c||memtarget) {
3214       int a=addr;
3215 #ifdef PCSX
3216       if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
3217 #endif
3218       //emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
3219       emit_writeword_indexed_tlb(tl,0,a,map,temp);
3220     }
3221     type=STOREW_STUB;
3222   }
3223   if (opcode[i]==0x3F) { // SD
3224     if(!c||memtarget) {
3225       int a=addr;
3226 #ifdef PCSX
3227       if(sp_in_mirror&&rs1[i]==29) a=HOST_TEMPREG;
3228 #endif
3229       if(rs2[i]) {
3230         assert(th>=0);
3231         //emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
3232         //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,addr);
3233         emit_writedword_indexed_tlb(th,tl,0,a,map,temp);
3234       }else{
3235         // Store zero
3236         //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
3237         //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
3238         emit_writedword_indexed_tlb(tl,tl,0,a,map,temp);
3239       }
3240     }
3241     type=STORED_STUB;
3242   }
3243   if(!using_tlb) {
3244     if(!c||memtarget) {
3245       #ifdef DESTRUCTIVE_SHIFT
3246       // The x86 shift operation is 'destructive'; it overwrites the
3247       // source register, so we need to make a copy first and use that.
3248       addr=temp;
3249       #endif
3250       #if defined(HOST_IMM8)
3251       int ir=get_reg(i_regs->regmap,INVCP);
3252       assert(ir>=0);
3253       emit_cmpmem_indexedsr12_reg(ir,addr,1);
3254       #else
3255       emit_cmpmem_indexedsr12_imm((int)invalid_code,addr,1);
3256       #endif
3257       #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3258       emit_callne(invalidate_addr_reg[addr]);
3259       #else
3260       jaddr2=(int)out;
3261       emit_jne(0);
3262       add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3263       #endif
3264     }
3265   }
3266   if(jaddr) {
3267     add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3268   } else if(c&&!memtarget) {
3269     inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
3270   }
3271   //if(opcode[i]==0x2B || opcode[i]==0x3F)
3272   //if(opcode[i]==0x2B || opcode[i]==0x28)
3273   //if(opcode[i]==0x2B || opcode[i]==0x29)
3274   //if(opcode[i]==0x2B)
3275   /*if(opcode[i]==0x2B || opcode[i]==0x28 || opcode[i]==0x29 || opcode[i]==0x3F)
3276   {
3277     //emit_pusha();
3278     save_regs(0x100f);
3279         emit_readword((int)&last_count,ECX);
3280         #ifdef __i386__
3281         if(get_reg(i_regs->regmap,CCREG)<0)
3282           emit_loadreg(CCREG,HOST_CCREG);
3283         emit_add(HOST_CCREG,ECX,HOST_CCREG);
3284         emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3285         emit_writeword(HOST_CCREG,(int)&Count);
3286         #endif
3287         #ifdef __arm__
3288         if(get_reg(i_regs->regmap,CCREG)<0)
3289           emit_loadreg(CCREG,0);
3290         else
3291           emit_mov(HOST_CCREG,0);
3292         emit_add(0,ECX,0);
3293         emit_addimm(0,2*ccadj[i],0);
3294         emit_writeword(0,(int)&Count);
3295         #endif
3296     emit_call((int)memdebug);
3297     //emit_popa();
3298     restore_regs(0x100f);
3299   }/**/
3300 }
3301
3302 void storelr_assemble(int i,struct regstat *i_regs)
3303 {
3304   int s,th,tl;
3305   int temp;
3306   int temp2;
3307   int offset;
3308   int jaddr=0,jaddr2;
3309   int case1,case2,case3;
3310   int done0,done1,done2;
3311   int memtarget=0,c=0;
3312   int agr=AGEN1+(i&1);
3313   u_int hr,reglist=0;
3314   th=get_reg(i_regs->regmap,rs2[i]|64);
3315   tl=get_reg(i_regs->regmap,rs2[i]);
3316   s=get_reg(i_regs->regmap,rs1[i]);
3317   temp=get_reg(i_regs->regmap,agr);
3318   if(temp<0) temp=get_reg(i_regs->regmap,-1);
3319   offset=imm[i];
3320   if(s>=0) {
3321     c=(i_regs->isconst>>s)&1;
3322     if(c) {
3323       memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3324       if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
3325     }
3326   }
3327   assert(tl>=0);
3328   for(hr=0;hr<HOST_REGS;hr++) {
3329     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3330   }
3331   assert(temp>=0);
3332   if(!using_tlb) {
3333     if(!c) {
3334       emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3335       if(!offset&&s!=temp) emit_mov(s,temp);
3336       jaddr=(int)out;
3337       emit_jno(0);
3338     }
3339     else
3340     {
3341       if(!memtarget||!rs1[i]) {
3342         jaddr=(int)out;
3343         emit_jmp(0);
3344       }
3345     }
3346     #ifdef RAM_OFFSET
3347     int map=get_reg(i_regs->regmap,ROREG);
3348     if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG);
3349     gen_tlb_addr_w(temp,map);
3350     #else
3351     if((u_int)rdram!=0x80000000)
3352       emit_addimm_no_flags((u_int)rdram-(u_int)0x80000000,temp);
3353     #endif
3354   }else{ // using tlb
3355     int map=get_reg(i_regs->regmap,TLREG);
3356     assert(map>=0);
3357     reglist&=~(1<<map);
3358     map=do_tlb_w(c||s<0||offset?temp:s,temp,map,0,c,constmap[i][s]+offset);
3359     if(!c&&!offset&&s>=0) emit_mov(s,temp);
3360     do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
3361     if(!jaddr&&!memtarget) {
3362       jaddr=(int)out;
3363       emit_jmp(0);
3364     }
3365     gen_tlb_addr_w(temp,map);
3366   }
3367
3368   if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
3369     temp2=get_reg(i_regs->regmap,FTEMP);
3370     if(!rs2[i]) temp2=th=tl;
3371   }
3372
3373 #ifndef BIG_ENDIAN_MIPS
3374     emit_xorimm(temp,3,temp);
3375 #endif
3376   emit_testimm(temp,2);
3377   case2=(int)out;
3378   emit_jne(0);
3379   emit_testimm(temp,1);
3380   case1=(int)out;
3381   emit_jne(0);
3382   // 0
3383   if (opcode[i]==0x2A) { // SWL
3384     emit_writeword_indexed(tl,0,temp);
3385   }
3386   if (opcode[i]==0x2E) { // SWR
3387     emit_writebyte_indexed(tl,3,temp);
3388   }
3389   if (opcode[i]==0x2C) { // SDL
3390     emit_writeword_indexed(th,0,temp);
3391     if(rs2[i]) emit_mov(tl,temp2);
3392   }
3393   if (opcode[i]==0x2D) { // SDR
3394     emit_writebyte_indexed(tl,3,temp);
3395     if(rs2[i]) emit_shldimm(th,tl,24,temp2);
3396   }
3397   done0=(int)out;
3398   emit_jmp(0);
3399   // 1
3400   set_jump_target(case1,(int)out);
3401   if (opcode[i]==0x2A) { // SWL
3402     // Write 3 msb into three least significant bytes
3403     if(rs2[i]) emit_rorimm(tl,8,tl);
3404     emit_writehword_indexed(tl,-1,temp);
3405     if(rs2[i]) emit_rorimm(tl,16,tl);
3406     emit_writebyte_indexed(tl,1,temp);
3407     if(rs2[i]) emit_rorimm(tl,8,tl);
3408   }
3409   if (opcode[i]==0x2E) { // SWR
3410     // Write two lsb into two most significant bytes
3411     emit_writehword_indexed(tl,1,temp);
3412   }
3413   if (opcode[i]==0x2C) { // SDL
3414     if(rs2[i]) emit_shrdimm(tl,th,8,temp2);
3415     // Write 3 msb into three least significant bytes
3416     if(rs2[i]) emit_rorimm(th,8,th);
3417     emit_writehword_indexed(th,-1,temp);
3418     if(rs2[i]) emit_rorimm(th,16,th);
3419     emit_writebyte_indexed(th,1,temp);
3420     if(rs2[i]) emit_rorimm(th,8,th);
3421   }
3422   if (opcode[i]==0x2D) { // SDR
3423     if(rs2[i]) emit_shldimm(th,tl,16,temp2);
3424     // Write two lsb into two most significant bytes
3425     emit_writehword_indexed(tl,1,temp);
3426   }
3427   done1=(int)out;
3428   emit_jmp(0);
3429   // 2
3430   set_jump_target(case2,(int)out);
3431   emit_testimm(temp,1);
3432   case3=(int)out;
3433   emit_jne(0);
3434   if (opcode[i]==0x2A) { // SWL
3435     // Write two msb into two least significant bytes
3436     if(rs2[i]) emit_rorimm(tl,16,tl);
3437     emit_writehword_indexed(tl,-2,temp);
3438     if(rs2[i]) emit_rorimm(tl,16,tl);
3439   }
3440   if (opcode[i]==0x2E) { // SWR
3441     // Write 3 lsb into three most significant bytes
3442     emit_writebyte_indexed(tl,-1,temp);
3443     if(rs2[i]) emit_rorimm(tl,8,tl);
3444     emit_writehword_indexed(tl,0,temp);
3445     if(rs2[i]) emit_rorimm(tl,24,tl);
3446   }
3447   if (opcode[i]==0x2C) { // SDL
3448     if(rs2[i]) emit_shrdimm(tl,th,16,temp2);
3449     // Write two msb into two least significant bytes
3450     if(rs2[i]) emit_rorimm(th,16,th);
3451     emit_writehword_indexed(th,-2,temp);
3452     if(rs2[i]) emit_rorimm(th,16,th);
3453   }
3454   if (opcode[i]==0x2D) { // SDR
3455     if(rs2[i]) emit_shldimm(th,tl,8,temp2);
3456     // Write 3 lsb into three most significant bytes
3457     emit_writebyte_indexed(tl,-1,temp);
3458     if(rs2[i]) emit_rorimm(tl,8,tl);
3459     emit_writehword_indexed(tl,0,temp);
3460     if(rs2[i]) emit_rorimm(tl,24,tl);
3461   }
3462   done2=(int)out;
3463   emit_jmp(0);
3464   // 3
3465   set_jump_target(case3,(int)out);
3466   if (opcode[i]==0x2A) { // SWL
3467     // Write msb into least significant byte
3468     if(rs2[i]) emit_rorimm(tl,24,tl);
3469     emit_writebyte_indexed(tl,-3,temp);
3470     if(rs2[i]) emit_rorimm(tl,8,tl);
3471   }
3472   if (opcode[i]==0x2E) { // SWR
3473     // Write entire word
3474     emit_writeword_indexed(tl,-3,temp);
3475   }
3476   if (opcode[i]==0x2C) { // SDL
3477     if(rs2[i]) emit_shrdimm(tl,th,24,temp2);
3478     // Write msb into least significant byte
3479     if(rs2[i]) emit_rorimm(th,24,th);
3480     emit_writebyte_indexed(th,-3,temp);
3481     if(rs2[i]) emit_rorimm(th,8,th);
3482   }
3483   if (opcode[i]==0x2D) { // SDR
3484     if(rs2[i]) emit_mov(th,temp2);
3485     // Write entire word
3486     emit_writeword_indexed(tl,-3,temp);
3487   }
3488   set_jump_target(done0,(int)out);
3489   set_jump_target(done1,(int)out);
3490   set_jump_target(done2,(int)out);
3491   if (opcode[i]==0x2C) { // SDL
3492     emit_testimm(temp,4);
3493     done0=(int)out;
3494     emit_jne(0);
3495     emit_andimm(temp,~3,temp);
3496     emit_writeword_indexed(temp2,4,temp);
3497     set_jump_target(done0,(int)out);
3498   }
3499   if (opcode[i]==0x2D) { // SDR
3500     emit_testimm(temp,4);
3501     done0=(int)out;
3502     emit_jeq(0);
3503     emit_andimm(temp,~3,temp);
3504     emit_writeword_indexed(temp2,-4,temp);
3505     set_jump_target(done0,(int)out);
3506   }
3507   if(!c||!memtarget)
3508     add_stub(STORELR_STUB,jaddr,(int)out,i,(int)i_regs,temp,ccadj[i],reglist);
3509   if(!using_tlb) {
3510     #ifdef RAM_OFFSET
3511     int map=get_reg(i_regs->regmap,ROREG);
3512     if(map<0) map=HOST_TEMPREG;
3513     gen_orig_addr_w(temp,map);
3514     #else
3515     emit_addimm_no_flags((u_int)0x80000000-(u_int)rdram,temp);
3516     #endif
3517     #if defined(HOST_IMM8)
3518     int ir=get_reg(i_regs->regmap,INVCP);
3519     assert(ir>=0);
3520     emit_cmpmem_indexedsr12_reg(ir,temp,1);
3521     #else
3522     emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
3523     #endif
3524     #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3525     emit_callne(invalidate_addr_reg[temp]);
3526     #else
3527     jaddr2=(int)out;
3528     emit_jne(0);
3529     add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3530     #endif
3531   }
3532   /*
3533     emit_pusha();
3534     //save_regs(0x100f);
3535         emit_readword((int)&last_count,ECX);
3536         if(get_reg(i_regs->regmap,CCREG)<0)
3537           emit_loadreg(CCREG,HOST_CCREG);
3538         emit_add(HOST_CCREG,ECX,HOST_CCREG);
3539         emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3540         emit_writeword(HOST_CCREG,(int)&Count);
3541     emit_call((int)memdebug);
3542     emit_popa();
3543     //restore_regs(0x100f);
3544   /**/
3545 }
3546
3547 void c1ls_assemble(int i,struct regstat *i_regs)
3548 {
3549 #ifndef DISABLE_COP1
3550   int s,th,tl;
3551   int temp,ar;
3552   int map=-1;
3553   int offset;
3554   int c=0;
3555   int jaddr,jaddr2=0,jaddr3,type;
3556   int agr=AGEN1+(i&1);
3557   u_int hr,reglist=0;
3558   th=get_reg(i_regs->regmap,FTEMP|64);
3559   tl=get_reg(i_regs->regmap,FTEMP);
3560   s=get_reg(i_regs->regmap,rs1[i]);
3561   temp=get_reg(i_regs->regmap,agr);
3562   if(temp<0) temp=get_reg(i_regs->regmap,-1);
3563   offset=imm[i];
3564   assert(tl>=0);
3565   assert(rs1[i]>0);
3566   assert(temp>=0);
3567   for(hr=0;hr<HOST_REGS;hr++) {
3568     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3569   }
3570   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3571   if (opcode[i]==0x31||opcode[i]==0x35) // LWC1/LDC1
3572   {
3573     // Loads use a temporary register which we need to save
3574     reglist|=1<<temp;
3575   }
3576   if (opcode[i]==0x39||opcode[i]==0x3D) // SWC1/SDC1
3577     ar=temp;
3578   else // LWC1/LDC1
3579     ar=tl;
3580   //if(s<0) emit_loadreg(rs1[i],ar); //address_generation does this now
3581   //else c=(i_regs->wasconst>>s)&1;
3582   if(s>=0) c=(i_regs->wasconst>>s)&1;
3583   // Check cop1 unusable
3584   if(!cop1_usable) {
3585     signed char rs=get_reg(i_regs->regmap,CSREG);
3586     assert(rs>=0);
3587     emit_testimm(rs,0x20000000);
3588     jaddr=(int)out;
3589     emit_jeq(0);
3590     add_stub(FP_STUB,jaddr,(int)out,i,rs,(int)i_regs,is_delayslot,0);
3591     cop1_usable=1;
3592   }
3593   if (opcode[i]==0x39) { // SWC1 (get float address)
3594     emit_readword((int)&reg_cop1_simple[(source[i]>>16)&0x1f],tl);
3595   }
3596   if (opcode[i]==0x3D) { // SDC1 (get double address)
3597     emit_readword((int)&reg_cop1_double[(source[i]>>16)&0x1f],tl);
3598   }
3599   // Generate address + offset
3600   if(!using_tlb) {
3601     if(!c)
3602       emit_cmpimm(offset||c||s<0?ar:s,RAM_SIZE);
3603   }
3604   else
3605   {
3606     map=get_reg(i_regs->regmap,TLREG);
3607     assert(map>=0);
3608     reglist&=~(1<<map);
3609     if (opcode[i]==0x31||opcode[i]==0x35) { // LWC1/LDC1
3610       map=do_tlb_r(offset||c||s<0?ar:s,ar,map,0,-1,-1,c,constmap[i][s]+offset);
3611     }
3612     if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3613       map=do_tlb_w(offset||c||s<0?ar:s,ar,map,0,c,constmap[i][s]+offset);
3614     }
3615   }
3616   if (opcode[i]==0x39) { // SWC1 (read float)
3617     emit_readword_indexed(0,tl,tl);
3618   }
3619   if (opcode[i]==0x3D) { // SDC1 (read double)
3620     emit_readword_indexed(4,tl,th);
3621     emit_readword_indexed(0,tl,tl);
3622   }
3623   if (opcode[i]==0x31) { // LWC1 (get target address)
3624     emit_readword((int)&reg_cop1_simple[(source[i]>>16)&0x1f],temp);
3625   }
3626   if (opcode[i]==0x35) { // LDC1 (get target address)
3627     emit_readword((int)&reg_cop1_double[(source[i]>>16)&0x1f],temp);
3628   }
3629   if(!using_tlb) {
3630     if(!c) {
3631       jaddr2=(int)out;
3632       emit_jno(0);
3633     }
3634     else if(((signed int)(constmap[i][s]+offset))>=(signed int)0x80000000+RAM_SIZE) {
3635       jaddr2=(int)out;
3636       emit_jmp(0); // inline_readstub/inline_writestub?  Very rare case
3637     }
3638     #ifdef DESTRUCTIVE_SHIFT
3639     if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3640       if(!offset&&!c&&s>=0) emit_mov(s,ar);
3641     }
3642     #endif
3643   }else{
3644     if (opcode[i]==0x31||opcode[i]==0x35) { // LWC1/LDC1
3645       do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr2);
3646     }
3647     if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3648       do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr2);
3649     }
3650   }
3651   if (opcode[i]==0x31) { // LWC1
3652     //if(s>=0&&!c&&!offset) emit_mov(s,tl);
3653     //gen_tlb_addr_r(ar,map);
3654     //emit_readword_indexed((int)rdram-0x80000000,tl,tl);
3655     #ifdef HOST_IMM_ADDR32
3656     if(c) emit_readword_tlb(constmap[i][s]+offset,map,tl);
3657     else
3658     #endif
3659     emit_readword_indexed_tlb(0,offset||c||s<0?tl:s,map,tl);
3660     type=LOADW_STUB;
3661   }
3662   if (opcode[i]==0x35) { // LDC1
3663     assert(th>=0);
3664     //if(s>=0&&!c&&!offset) emit_mov(s,tl);
3665     //gen_tlb_addr_r(ar,map);
3666     //emit_readword_indexed((int)rdram-0x80000000,tl,th);
3667     //emit_readword_indexed((int)rdram-0x7FFFFFFC,tl,tl);
3668     #ifdef HOST_IMM_ADDR32
3669     if(c) emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
3670     else
3671     #endif
3672     emit_readdword_indexed_tlb(0,offset||c||s<0?tl:s,map,th,tl);
3673     type=LOADD_STUB;
3674   }
3675   if (opcode[i]==0x39) { // SWC1
3676     //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
3677     emit_writeword_indexed_tlb(tl,0,offset||c||s<0?temp:s,map,temp);
3678     type=STOREW_STUB;
3679   }
3680   if (opcode[i]==0x3D) { // SDC1
3681     assert(th>=0);
3682     //emit_writeword_indexed(th,(int)rdram-0x80000000,temp);
3683     //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
3684     emit_writedword_indexed_tlb(th,tl,0,offset||c||s<0?temp:s,map,temp);
3685     type=STORED_STUB;
3686   }
3687   if(!using_tlb) {
3688     if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3689       #ifndef DESTRUCTIVE_SHIFT
3690       temp=offset||c||s<0?ar:s;
3691       #endif
3692       #if defined(HOST_IMM8)
3693       int ir=get_reg(i_regs->regmap,INVCP);
3694       assert(ir>=0);
3695       emit_cmpmem_indexedsr12_reg(ir,temp,1);
3696       #else
3697       emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
3698       #endif
3699       #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3700       emit_callne(invalidate_addr_reg[temp]);
3701       #else
3702       jaddr3=(int)out;
3703       emit_jne(0);
3704       add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3705       #endif
3706     }
3707   }
3708   if(jaddr2) add_stub(type,jaddr2,(int)out,i,offset||c||s<0?ar:s,(int)i_regs,ccadj[i],reglist);
3709   if (opcode[i]==0x31) { // LWC1 (write float)
3710     emit_writeword_indexed(tl,0,temp);
3711   }
3712   if (opcode[i]==0x35) { // LDC1 (write double)
3713     emit_writeword_indexed(th,4,temp);
3714     emit_writeword_indexed(tl,0,temp);
3715   }
3716   //if(opcode[i]==0x39)
3717   /*if(opcode[i]==0x39||opcode[i]==0x31)
3718   {
3719     emit_pusha();
3720         emit_readword((int)&last_count,ECX);
3721         if(get_reg(i_regs->regmap,CCREG)<0)
3722           emit_loadreg(CCREG,HOST_CCREG);
3723         emit_add(HOST_CCREG,ECX,HOST_CCREG);
3724         emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3725         emit_writeword(HOST_CCREG,(int)&Count);
3726     emit_call((int)memdebug);
3727     emit_popa();
3728   }/**/
3729 #else
3730   cop1_unusable(i, i_regs);
3731 #endif
3732 }
3733
3734 void c2ls_assemble(int i,struct regstat *i_regs)
3735 {
3736   int s,tl;
3737   int ar;
3738   int offset;
3739   int memtarget=0,c=0;
3740   int jaddr2=0,jaddr3,type;
3741   int agr=AGEN1+(i&1);
3742   u_int hr,reglist=0;
3743   u_int copr=(source[i]>>16)&0x1f;
3744   s=get_reg(i_regs->regmap,rs1[i]);
3745   tl=get_reg(i_regs->regmap,FTEMP);
3746   offset=imm[i];
3747   assert(rs1[i]>0);
3748   assert(tl>=0);
3749   assert(!using_tlb);
3750
3751   for(hr=0;hr<HOST_REGS;hr++) {
3752     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3753   }
3754   if(i_regs->regmap[HOST_CCREG]==CCREG)
3755     reglist&=~(1<<HOST_CCREG);
3756
3757   // get the address
3758   if (opcode[i]==0x3a) { // SWC2
3759     ar=get_reg(i_regs->regmap,agr);
3760     if(ar<0) ar=get_reg(i_regs->regmap,-1);
3761     reglist|=1<<ar;
3762   } else { // LWC2
3763     ar=tl;
3764   }
3765   if(s>=0) c=(i_regs->wasconst>>s)&1;
3766   memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3767   if (!offset&&!c&&s>=0) ar=s;
3768   assert(ar>=0);
3769
3770   if (opcode[i]==0x3a) { // SWC2
3771     cop2_get_dreg(copr,tl,HOST_TEMPREG);
3772     type=STOREW_STUB;
3773   }
3774   else
3775     type=LOADW_STUB;
3776
3777   if(c&&!memtarget) {
3778     jaddr2=(int)out;
3779     emit_jmp(0); // inline_readstub/inline_writestub?
3780   }
3781   else {
3782     if(!c) {
3783       emit_cmpimm(offset||c||s<0?ar:s,RAM_SIZE);
3784       jaddr2=(int)out;
3785       emit_jno(0);
3786     }
3787     if (opcode[i]==0x32) { // LWC2
3788       #ifdef HOST_IMM_ADDR32
3789       if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl);
3790       else
3791       #endif
3792       emit_readword_indexed(0,ar,tl);
3793     }
3794     if (opcode[i]==0x3a) { // SWC2
3795       #ifdef DESTRUCTIVE_SHIFT
3796       if(!offset&&!c&&s>=0) emit_mov(s,ar);
3797       #endif
3798       emit_writeword_indexed(tl,0,ar);
3799     }
3800   }
3801   if(jaddr2)
3802     add_stub(type,jaddr2,(int)out,i,ar,(int)i_regs,ccadj[i],reglist);
3803   if (opcode[i]==0x3a) { // SWC2
3804 #if defined(HOST_IMM8)
3805     int ir=get_reg(i_regs->regmap,INVCP);
3806     assert(ir>=0);
3807     emit_cmpmem_indexedsr12_reg(ir,ar,1);
3808 #else
3809     emit_cmpmem_indexedsr12_imm((int)invalid_code,ar,1);
3810 #endif
3811     #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3812     emit_callne(invalidate_addr_reg[ar]);
3813     #else
3814     jaddr3=(int)out;
3815     emit_jne(0);
3816     add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3817     #endif
3818   }
3819   if (opcode[i]==0x32) { // LWC2
3820     cop2_put_dreg(copr,tl,HOST_TEMPREG);
3821   }
3822 }
3823
3824 #ifndef multdiv_assemble
3825 void multdiv_assemble(int i,struct regstat *i_regs)
3826 {
3827   printf("Need multdiv_assemble for this architecture.\n");
3828   exit(1);
3829 }
3830 #endif
3831
3832 void mov_assemble(int i,struct regstat *i_regs)
3833 {
3834   //if(opcode2[i]==0x10||opcode2[i]==0x12) { // MFHI/MFLO
3835   //if(opcode2[i]==0x11||opcode2[i]==0x13) { // MTHI/MTLO
3836   if(rt1[i]) {
3837     signed char sh,sl,th,tl;
3838     th=get_reg(i_regs->regmap,rt1[i]|64);
3839     tl=get_reg(i_regs->regmap,rt1[i]);
3840     //assert(tl>=0);
3841     if(tl>=0) {
3842       sh=get_reg(i_regs->regmap,rs1[i]|64);
3843       sl=get_reg(i_regs->regmap,rs1[i]);
3844       if(sl>=0) emit_mov(sl,tl);
3845       else emit_loadreg(rs1[i],tl);
3846       if(th>=0) {
3847         if(sh>=0) emit_mov(sh,th);
3848         else emit_loadreg(rs1[i]|64,th);
3849       }
3850     }
3851   }
3852 }
3853
3854 #ifndef fconv_assemble
3855 void fconv_assemble(int i,struct regstat *i_regs)
3856 {
3857   printf("Need fconv_assemble for this architecture.\n");
3858   exit(1);
3859 }
3860 #endif
3861
3862 #if 0
3863 void float_assemble(int i,struct regstat *i_regs)
3864 {
3865   printf("Need float_assemble for this architecture.\n");
3866   exit(1);
3867 }
3868 #endif
3869
3870 void syscall_assemble(int i,struct regstat *i_regs)
3871 {
3872   signed char ccreg=get_reg(i_regs->regmap,CCREG);
3873   assert(ccreg==HOST_CCREG);
3874   assert(!is_delayslot);
3875   emit_movimm(start+i*4,EAX); // Get PC
3876   emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // CHECK: is this right?  There should probably be an extra cycle...
3877   emit_jmp((int)jump_syscall_hle); // XXX
3878 }
3879
3880 void hlecall_assemble(int i,struct regstat *i_regs)
3881 {
3882   signed char ccreg=get_reg(i_regs->regmap,CCREG);
3883   assert(ccreg==HOST_CCREG);
3884   assert(!is_delayslot);
3885   emit_movimm(start+i*4+4,0); // Get PC
3886   emit_movimm((int)psxHLEt[source[i]&7],1);
3887   emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // XXX
3888   emit_jmp((int)jump_hlecall);
3889 }
3890
3891 void intcall_assemble(int i,struct regstat *i_regs)
3892 {
3893   signed char ccreg=get_reg(i_regs->regmap,CCREG);
3894   assert(ccreg==HOST_CCREG);
3895   assert(!is_delayslot);
3896   emit_movimm(start+i*4,0); // Get PC
3897   emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG);
3898   emit_jmp((int)jump_intcall);
3899 }
3900
3901 void ds_assemble(int i,struct regstat *i_regs)
3902 {
3903   is_delayslot=1;
3904   switch(itype[i]) {
3905     case ALU:
3906       alu_assemble(i,i_regs);break;
3907     case IMM16:
3908       imm16_assemble(i,i_regs);break;
3909     case SHIFT:
3910       shift_assemble(i,i_regs);break;
3911     case SHIFTIMM:
3912       shiftimm_assemble(i,i_regs);break;
3913     case LOAD:
3914       load_assemble(i,i_regs);break;
3915     case LOADLR:
3916       loadlr_assemble(i,i_regs);break;
3917     case STORE:
3918       store_assemble(i,i_regs);break;
3919     case STORELR:
3920       storelr_assemble(i,i_regs);break;
3921     case COP0:
3922       cop0_assemble(i,i_regs);break;
3923     case COP1:
3924       cop1_assemble(i,i_regs);break;
3925     case C1LS:
3926       c1ls_assemble(i,i_regs);break;
3927     case COP2:
3928       cop2_assemble(i,i_regs);break;
3929     case C2LS:
3930       c2ls_assemble(i,i_regs);break;
3931     case C2OP:
3932       c2op_assemble(i,i_regs);break;
3933     case FCONV:
3934       fconv_assemble(i,i_regs);break;
3935     case FLOAT:
3936       float_assemble(i,i_regs);break;
3937     case FCOMP:
3938       fcomp_assemble(i,i_regs);break;
3939     case MULTDIV:
3940       multdiv_assemble(i,i_regs);break;
3941     case MOV:
3942       mov_assemble(i,i_regs);break;
3943     case SYSCALL:
3944     case HLECALL:
3945     case INTCALL:
3946     case SPAN:
3947     case UJUMP:
3948     case RJUMP:
3949     case CJUMP:
3950     case SJUMP:
3951     case FJUMP:
3952       printf("Jump in the delay slot.  This is probably a bug.\n");
3953   }
3954   is_delayslot=0;
3955 }
3956
3957 // Is the branch target a valid internal jump?
3958 int internal_branch(uint64_t i_is32,int addr)
3959 {
3960   if(addr&1) return 0; // Indirect (register) jump
3961   if(addr>=start && addr<start+slen*4-4)
3962   {
3963     int t=(addr-start)>>2;
3964     // Delay slots are not valid branch targets
3965     //if(t>0&&(itype[t-1]==RJUMP||itype[t-1]==UJUMP||itype[t-1]==CJUMP||itype[t-1]==SJUMP||itype[t-1]==FJUMP)) return 0;
3966     // 64 -> 32 bit transition requires a recompile
3967     /*if(is32[t]&~unneeded_reg_upper[t]&~i_is32)
3968     {
3969       if(requires_32bit[t]&~i_is32) printf("optimizable: no\n");
3970       else printf("optimizable: yes\n");
3971     }*/
3972     //if(is32[t]&~unneeded_reg_upper[t]&~i_is32) return 0;
3973 #ifndef FORCE32
3974     if(requires_32bit[t]&~i_is32) return 0;
3975     else
3976 #endif
3977       return 1;
3978   }
3979   return 0;
3980 }
3981
3982 #ifndef wb_invalidate
3983 void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t is32,
3984   uint64_t u,uint64_t uu)
3985 {
3986   int hr;
3987   for(hr=0;hr<HOST_REGS;hr++) {
3988     if(hr!=EXCLUDE_REG) {
3989       if(pre[hr]!=entry[hr]) {
3990         if(pre[hr]>=0) {
3991           if((dirty>>hr)&1) {
3992             if(get_reg(entry,pre[hr])<0) {
3993               if(pre[hr]<64) {
3994                 if(!((u>>pre[hr])&1)) {
3995                   emit_storereg(pre[hr],hr);
3996                   if( ((is32>>pre[hr])&1) && !((uu>>pre[hr])&1) ) {
3997                     emit_sarimm(hr,31,hr);
3998                     emit_storereg(pre[hr]|64,hr);
3999                   }
4000                 }
4001               }else{
4002                 if(!((uu>>(pre[hr]&63))&1) && !((is32>>(pre[hr]&63))&1)) {
4003                   emit_storereg(pre[hr],hr);
4004                 }
4005               }
4006             }
4007           }
4008         }
4009       }
4010     }
4011   }
4012   // Move from one register to another (no writeback)
4013   for(hr=0;hr<HOST_REGS;hr++) {
4014     if(hr!=EXCLUDE_REG) {
4015       if(pre[hr]!=entry[hr]) {
4016         if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
4017           int nr;
4018           if((nr=get_reg(entry,pre[hr]))>=0) {
4019             emit_mov(hr,nr);
4020           }
4021         }
4022       }
4023     }
4024   }
4025 }
4026 #endif
4027
4028 // Load the specified registers
4029 // This only loads the registers given as arguments because
4030 // we don't want to load things that will be overwritten
4031 void load_regs(signed char entry[],signed char regmap[],int is32,int rs1,int rs2)
4032 {
4033   int hr;
4034   // Load 32-bit regs
4035   for(hr=0;hr<HOST_REGS;hr++) {
4036     if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4037       if(entry[hr]!=regmap[hr]) {
4038         if(regmap[hr]==rs1||regmap[hr]==rs2)
4039         {
4040           if(regmap[hr]==0) {
4041             emit_zeroreg(hr);
4042           }
4043           else
4044           {
4045             emit_loadreg(regmap[hr],hr);
4046           }
4047         }
4048       }
4049     }
4050   }
4051   //Load 64-bit regs
4052   for(hr=0;hr<HOST_REGS;hr++) {
4053     if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4054       if(entry[hr]!=regmap[hr]) {
4055         if(regmap[hr]-64==rs1||regmap[hr]-64==rs2)
4056         {
4057           assert(regmap[hr]!=64);
4058           if((is32>>(regmap[hr]&63))&1) {
4059             int lr=get_reg(regmap,regmap[hr]-64);
4060             if(lr>=0)
4061               emit_sarimm(lr,31,hr);
4062             else
4063               emit_loadreg(regmap[hr],hr);
4064           }
4065           else
4066           {
4067             emit_loadreg(regmap[hr],hr);
4068           }
4069         }
4070       }
4071     }
4072   }
4073 }
4074
4075 // Load registers prior to the start of a loop
4076 // so that they are not loaded within the loop
4077 static void loop_preload(signed char pre[],signed char entry[])
4078 {
4079   int hr;
4080   for(hr=0;hr<HOST_REGS;hr++) {
4081     if(hr!=EXCLUDE_REG) {
4082       if(pre[hr]!=entry[hr]) {
4083         if(entry[hr]>=0) {
4084           if(get_reg(pre,entry[hr])<0) {
4085             assem_debug("loop preload:\n");
4086             //printf("loop preload: %d\n",hr);
4087             if(entry[hr]==0) {
4088               emit_zeroreg(hr);
4089             }
4090             else if(entry[hr]<TEMPREG)
4091             {
4092               emit_loadreg(entry[hr],hr);
4093             }
4094             else if(entry[hr]-64<TEMPREG)
4095             {
4096               emit_loadreg(entry[hr],hr);
4097             }
4098           }
4099         }
4100       }
4101     }
4102   }
4103 }
4104
4105 // Generate address for load/store instruction
4106 // goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4107 void address_generation(int i,struct regstat *i_regs,signed char entry[])
4108 {
4109   if(itype[i]==LOAD||itype[i]==LOADLR||itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS||itype[i]==C2LS) {
4110     int ra=-1;
4111     int agr=AGEN1+(i&1);
4112     int mgr=MGEN1+(i&1);
4113     if(itype[i]==LOAD) {
4114       ra=get_reg(i_regs->regmap,rt1[i]);
4115       if(ra<0) ra=get_reg(i_regs->regmap,-1);
4116       assert(ra>=0);
4117     }
4118     if(itype[i]==LOADLR) {
4119       ra=get_reg(i_regs->regmap,FTEMP);
4120     }
4121     if(itype[i]==STORE||itype[i]==STORELR) {
4122       ra=get_reg(i_regs->regmap,agr);
4123       if(ra<0) ra=get_reg(i_regs->regmap,-1);
4124     }
4125     if(itype[i]==C1LS||itype[i]==C2LS) {
4126       if ((opcode[i]&0x3b)==0x31||(opcode[i]&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4127         ra=get_reg(i_regs->regmap,FTEMP);
4128       else { // SWC1/SDC1/SWC2/SDC2
4129         ra=get_reg(i_regs->regmap,agr);
4130         if(ra<0) ra=get_reg(i_regs->regmap,-1);
4131       }
4132     }
4133     int rs=get_reg(i_regs->regmap,rs1[i]);
4134     int rm=get_reg(i_regs->regmap,TLREG);
4135     if(ra>=0) {
4136       int offset=imm[i];
4137       int c=(i_regs->wasconst>>rs)&1;
4138       if(rs1[i]==0) {
4139         // Using r0 as a base address
4140         /*if(rm>=0) {
4141           if(!entry||entry[rm]!=mgr) {
4142             generate_map_const(offset,rm);
4143           } // else did it in the previous cycle
4144         }*/
4145         if(!entry||entry[ra]!=agr) {
4146           if (opcode[i]==0x22||opcode[i]==0x26) {
4147             emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4148           }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
4149             emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4150           }else{
4151             emit_movimm(offset,ra);
4152           }
4153         } // else did it in the previous cycle
4154       }
4155       else if(rs<0) {
4156         if(!entry||entry[ra]!=rs1[i])
4157           emit_loadreg(rs1[i],ra);
4158         //if(!entry||entry[ra]!=rs1[i])
4159         //  printf("poor load scheduling!\n");
4160       }
4161       else if(c) {
4162         if(rm>=0) {
4163           if(!entry||entry[rm]!=mgr) {
4164             if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a) {
4165               // Stores to memory go thru the mapper to detect self-modifying
4166               // code, loads don't.
4167               if((unsigned int)(constmap[i][rs]+offset)>=0xC0000000 ||
4168                  (unsigned int)(constmap[i][rs]+offset)<0x80000000+RAM_SIZE )
4169                 generate_map_const(constmap[i][rs]+offset,rm);
4170             }else{
4171               if((signed int)(constmap[i][rs]+offset)>=(signed int)0xC0000000)
4172                 generate_map_const(constmap[i][rs]+offset,rm);
4173             }
4174           }
4175         }
4176         if(rs1[i]!=rt1[i]||itype[i]!=LOAD) {
4177           if(!entry||entry[ra]!=agr) {
4178             if (opcode[i]==0x22||opcode[i]==0x26) {
4179               emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4180             }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
4181               emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4182             }else{
4183               #ifdef HOST_IMM_ADDR32
4184               if((itype[i]!=LOAD&&(opcode[i]&0x3b)!=0x31&&(opcode[i]&0x3b)!=0x32) || // LWC1/LDC1/LWC2/LDC2
4185                  (using_tlb&&((signed int)constmap[i][rs]+offset)>=(signed int)0xC0000000))
4186               #endif
4187               emit_movimm(constmap[i][rs]+offset,ra);
4188             }
4189           } // else did it in the previous cycle
4190         } // else load_consts already did it
4191       }
4192       if(offset&&!c&&rs1[i]) {
4193         if(rs>=0) {
4194           emit_addimm(rs,offset,ra);
4195         }else{
4196           emit_addimm(ra,offset,ra);
4197         }
4198       }
4199     }
4200   }
4201   // Preload constants for next instruction
4202   if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) {
4203     int agr,ra;
4204     #ifndef HOST_IMM_ADDR32
4205     // Mapper entry
4206     agr=MGEN1+((i+1)&1);
4207     ra=get_reg(i_regs->regmap,agr);
4208     if(ra>=0) {
4209       int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
4210       int offset=imm[i+1];
4211       int c=(regs[i+1].wasconst>>rs)&1;
4212       if(c) {
4213         if(itype[i+1]==STORE||itype[i+1]==STORELR
4214            ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1, SWC2/SDC2
4215           // Stores to memory go thru the mapper to detect self-modifying
4216           // code, loads don't.
4217           if((unsigned int)(constmap[i+1][rs]+offset)>=0xC0000000 ||
4218              (unsigned int)(constmap[i+1][rs]+offset)<0x80000000+RAM_SIZE )
4219             generate_map_const(constmap[i+1][rs]+offset,ra);
4220         }else{
4221           if((signed int)(constmap[i+1][rs]+offset)>=(signed int)0xC0000000)
4222             generate_map_const(constmap[i+1][rs]+offset,ra);
4223         }
4224       }
4225       /*else if(rs1[i]==0) {
4226         generate_map_const(offset,ra);
4227       }*/
4228     }
4229     #endif
4230     // Actual address
4231     agr=AGEN1+((i+1)&1);
4232     ra=get_reg(i_regs->regmap,agr);
4233     if(ra>=0) {
4234       int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
4235       int offset=imm[i+1];
4236       int c=(regs[i+1].wasconst>>rs)&1;
4237       if(c&&(rs1[i+1]!=rt1[i+1]||itype[i+1]!=LOAD)) {
4238         if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
4239           emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4240         }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
4241           emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4242         }else{
4243           #ifdef HOST_IMM_ADDR32
4244           if((itype[i+1]!=LOAD&&(opcode[i+1]&0x3b)!=0x31&&(opcode[i+1]&0x3b)!=0x32) || // LWC1/LDC1/LWC2/LDC2
4245              (using_tlb&&((signed int)constmap[i+1][rs]+offset)>=(signed int)0xC0000000))
4246           #endif
4247           emit_movimm(constmap[i+1][rs]+offset,ra);
4248         }
4249       }
4250       else if(rs1[i+1]==0) {
4251         // Using r0 as a base address
4252         if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
4253           emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4254         }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
4255           emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4256         }else{
4257           emit_movimm(offset,ra);
4258         }
4259       }
4260     }
4261   }
4262 }
4263
4264 int get_final_value(int hr, int i, int *value)
4265 {
4266   int reg=regs[i].regmap[hr];
4267   while(i<slen-1) {
4268     if(regs[i+1].regmap[hr]!=reg) break;
4269     if(!((regs[i+1].isconst>>hr)&1)) break;
4270     if(bt[i+1]) break;
4271     i++;
4272   }
4273   if(i<slen-1) {
4274     if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP) {
4275       *value=constmap[i][hr];
4276       return 1;
4277     }
4278     if(!bt[i+1]) {
4279       if(itype[i+1]==UJUMP||itype[i+1]==RJUMP||itype[i+1]==CJUMP||itype[i+1]==SJUMP) {
4280         // Load in delay slot, out-of-order execution
4281         if(itype[i+2]==LOAD&&rs1[i+2]==reg&&rt1[i+2]==reg&&((regs[i+1].wasconst>>hr)&1))
4282         {
4283           #ifdef HOST_IMM_ADDR32
4284           if(!using_tlb||((signed int)constmap[i][hr]+imm[i+2])<(signed int)0xC0000000) return 0;
4285           #endif
4286           // Precompute load address
4287           *value=constmap[i][hr]+imm[i+2];
4288           return 1;
4289         }
4290       }
4291       if(itype[i+1]==LOAD&&rs1[i+1]==reg&&rt1[i+1]==reg)
4292       {
4293         #ifdef HOST_IMM_ADDR32
4294         if(!using_tlb||((signed int)constmap[i][hr]+imm[i+1])<(signed int)0xC0000000) return 0;
4295         #endif
4296         // Precompute load address
4297         *value=constmap[i][hr]+imm[i+1];
4298         //printf("c=%x imm=%x\n",(int)constmap[i][hr],imm[i+1]);
4299         return 1;
4300       }
4301     }
4302   }
4303   *value=constmap[i][hr];
4304   //printf("c=%x\n",(int)constmap[i][hr]);
4305   if(i==slen-1) return 1;
4306   if(reg<64) {
4307     return !((unneeded_reg[i+1]>>reg)&1);
4308   }else{
4309     return !((unneeded_reg_upper[i+1]>>reg)&1);
4310   }
4311 }
4312
4313 // Load registers with known constants
4314 void load_consts(signed char pre[],signed char regmap[],int is32,int i)
4315 {
4316   int hr;
4317   // Load 32-bit regs
4318   for(hr=0;hr<HOST_REGS;hr++) {
4319     if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4320       //if(entry[hr]!=regmap[hr]) {
4321       if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
4322         if(((regs[i].isconst>>hr)&1)&&regmap[hr]<64&&regmap[hr]>0) {
4323           int value;
4324           if(get_final_value(hr,i,&value)) {
4325             if(value==0) {
4326               emit_zeroreg(hr);
4327             }
4328             else {
4329               emit_movimm(value,hr);
4330             }
4331           }
4332         }
4333       }
4334     }
4335   }
4336   // Load 64-bit regs
4337   for(hr=0;hr<HOST_REGS;hr++) {
4338     if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4339       //if(entry[hr]!=regmap[hr]) {
4340       if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
4341         if(((regs[i].isconst>>hr)&1)&&regmap[hr]>64) {
4342           if((is32>>(regmap[hr]&63))&1) {
4343             int lr=get_reg(regmap,regmap[hr]-64);
4344             assert(lr>=0);
4345             emit_sarimm(lr,31,hr);
4346           }
4347           else
4348           {
4349             int value;
4350             if(get_final_value(hr,i,&value)) {
4351               if(value==0) {
4352                 emit_zeroreg(hr);
4353               }
4354               else {
4355                 emit_movimm(value,hr);
4356               }
4357             }
4358           }
4359         }
4360       }
4361     }
4362   }
4363 }
4364 void load_all_consts(signed char regmap[],int is32,u_int dirty,int i)
4365 {
4366   int hr;
4367   // Load 32-bit regs
4368   for(hr=0;hr<HOST_REGS;hr++) {
4369     if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4370       if(((regs[i].isconst>>hr)&1)&&regmap[hr]<64&&regmap[hr]>0) {
4371         int value=constmap[i][hr];
4372         if(value==0) {
4373           emit_zeroreg(hr);
4374         }
4375         else {
4376           emit_movimm(value,hr);
4377         }
4378       }
4379     }
4380   }
4381   // Load 64-bit regs
4382   for(hr=0;hr<HOST_REGS;hr++) {
4383     if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4384       if(((regs[i].isconst>>hr)&1)&&regmap[hr]>64) {
4385         if((is32>>(regmap[hr]&63))&1) {
4386           int lr=get_reg(regmap,regmap[hr]-64);
4387           assert(lr>=0);
4388           emit_sarimm(lr,31,hr);
4389         }
4390         else
4391         {
4392           int value=constmap[i][hr];
4393           if(value==0) {
4394             emit_zeroreg(hr);
4395           }
4396           else {
4397             emit_movimm(value,hr);
4398           }
4399         }
4400       }
4401     }
4402   }
4403 }
4404
4405 // Write out all dirty registers (except cycle count)
4406 void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty)
4407 {
4408   int hr;
4409   for(hr=0;hr<HOST_REGS;hr++) {
4410     if(hr!=EXCLUDE_REG) {
4411       if(i_regmap[hr]>0) {
4412         if(i_regmap[hr]!=CCREG) {
4413           if((i_dirty>>hr)&1) {
4414             if(i_regmap[hr]<64) {
4415               emit_storereg(i_regmap[hr],hr);
4416 #ifndef FORCE32
4417               if( ((i_is32>>i_regmap[hr])&1) ) {
4418                 #ifdef DESTRUCTIVE_WRITEBACK
4419                 emit_sarimm(hr,31,hr);
4420                 emit_storereg(i_regmap[hr]|64,hr);
4421                 #else
4422                 emit_sarimm(hr,31,HOST_TEMPREG);
4423                 emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4424                 #endif
4425               }
4426 #endif
4427             }else{
4428               if( !((i_is32>>(i_regmap[hr]&63))&1) ) {
4429                 emit_storereg(i_regmap[hr],hr);
4430               }
4431             }
4432           }
4433         }
4434       }
4435     }
4436   }
4437 }
4438 // Write out dirty registers that we need to reload (pair with load_needed_regs)
4439 // This writes the registers not written by store_regs_bt
4440 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4441 {
4442   int hr;
4443   int t=(addr-start)>>2;
4444   for(hr=0;hr<HOST_REGS;hr++) {
4445     if(hr!=EXCLUDE_REG) {
4446       if(i_regmap[hr]>0) {
4447         if(i_regmap[hr]!=CCREG) {
4448           if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1) && !(((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
4449             if((i_dirty>>hr)&1) {
4450               if(i_regmap[hr]<64) {
4451                 emit_storereg(i_regmap[hr],hr);
4452 #ifndef FORCE32
4453                 if( ((i_is32>>i_regmap[hr])&1) ) {
4454                   #ifdef DESTRUCTIVE_WRITEBACK
4455                   emit_sarimm(hr,31,hr);
4456                   emit_storereg(i_regmap[hr]|64,hr);
4457                   #else
4458                   emit_sarimm(hr,31,HOST_TEMPREG);
4459                   emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4460                   #endif
4461                 }
4462 #endif
4463               }else{
4464                 if( !((i_is32>>(i_regmap[hr]&63))&1) ) {
4465                   emit_storereg(i_regmap[hr],hr);
4466                 }
4467               }
4468             }
4469           }
4470         }
4471       }
4472     }
4473   }
4474 }
4475
4476 // Load all registers (except cycle count)
4477 void load_all_regs(signed char i_regmap[])
4478 {
4479   int hr;
4480   for(hr=0;hr<HOST_REGS;hr++) {
4481     if(hr!=EXCLUDE_REG) {
4482       if(i_regmap[hr]==0) {
4483         emit_zeroreg(hr);
4484       }
4485       else
4486       if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4487       {
4488         emit_loadreg(i_regmap[hr],hr);
4489       }
4490     }
4491   }
4492 }
4493
4494 // Load all current registers also needed by next instruction
4495 void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
4496 {
4497   int hr;
4498   for(hr=0;hr<HOST_REGS;hr++) {
4499     if(hr!=EXCLUDE_REG) {
4500       if(get_reg(next_regmap,i_regmap[hr])>=0) {
4501         if(i_regmap[hr]==0) {
4502           emit_zeroreg(hr);
4503         }
4504         else
4505         if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4506         {
4507           emit_loadreg(i_regmap[hr],hr);
4508         }
4509       }
4510     }
4511   }
4512 }
4513
4514 // Load all regs, storing cycle count if necessary
4515 void load_regs_entry(int t)
4516 {
4517   int hr;
4518   if(is_ds[t]) emit_addimm(HOST_CCREG,CLOCK_DIVIDER,HOST_CCREG);
4519   else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t]*CLOCK_DIVIDER,HOST_CCREG);
4520   if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4521     emit_storereg(CCREG,HOST_CCREG);
4522   }
4523   // Load 32-bit regs
4524   for(hr=0;hr<HOST_REGS;hr++) {
4525     if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4526       if(regs[t].regmap_entry[hr]==0) {
4527         emit_zeroreg(hr);
4528       }
4529       else if(regs[t].regmap_entry[hr]!=CCREG)
4530       {
4531         emit_loadreg(regs[t].regmap_entry[hr],hr);
4532       }
4533     }
4534   }
4535   // Load 64-bit regs
4536   for(hr=0;hr<HOST_REGS;hr++) {
4537     if(regs[t].regmap_entry[hr]>=64&&regs[t].regmap_entry[hr]<TEMPREG+64) {
4538       assert(regs[t].regmap_entry[hr]!=64);
4539       if((regs[t].was32>>(regs[t].regmap_entry[hr]&63))&1) {
4540         int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4541         if(lr<0) {
4542           emit_loadreg(regs[t].regmap_entry[hr],hr);
4543         }
4544         else
4545         {
4546           emit_sarimm(lr,31,hr);
4547         }
4548       }
4549       else
4550       {
4551         emit_loadreg(regs[t].regmap_entry[hr],hr);
4552       }
4553     }
4554   }
4555 }
4556
4557 // Store dirty registers prior to branch
4558 void store_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4559 {
4560   if(internal_branch(i_is32,addr))
4561   {
4562     int t=(addr-start)>>2;
4563     int hr;
4564     for(hr=0;hr<HOST_REGS;hr++) {
4565       if(hr!=EXCLUDE_REG) {
4566         if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4567           if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1) || (((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
4568             if((i_dirty>>hr)&1) {
4569               if(i_regmap[hr]<64) {
4570                 if(!((unneeded_reg[t]>>i_regmap[hr])&1)) {
4571                   emit_storereg(i_regmap[hr],hr);
4572                   if( ((i_is32>>i_regmap[hr])&1) && !((unneeded_reg_upper[t]>>i_regmap[hr])&1) ) {
4573                     #ifdef DESTRUCTIVE_WRITEBACK
4574                     emit_sarimm(hr,31,hr);
4575                     emit_storereg(i_regmap[hr]|64,hr);
4576                     #else
4577                     emit_sarimm(hr,31,HOST_TEMPREG);
4578                     emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4579                     #endif
4580                   }
4581                 }
4582               }else{
4583                 if( !((i_is32>>(i_regmap[hr]&63))&1) && !((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1) ) {
4584                   emit_storereg(i_regmap[hr],hr);
4585                 }
4586               }
4587             }
4588           }
4589         }
4590       }
4591     }
4592   }
4593   else
4594   {
4595     // Branch out of this block, write out all dirty regs
4596     wb_dirtys(i_regmap,i_is32,i_dirty);
4597   }
4598 }
4599
4600 // Load all needed registers for branch target
4601 void load_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4602 {
4603   //if(addr>=start && addr<(start+slen*4))
4604   if(internal_branch(i_is32,addr))
4605   {
4606     int t=(addr-start)>>2;
4607     int hr;
4608     // Store the cycle count before loading something else
4609     if(i_regmap[HOST_CCREG]!=CCREG) {
4610       assert(i_regmap[HOST_CCREG]==-1);
4611     }
4612     if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4613       emit_storereg(CCREG,HOST_CCREG);
4614     }
4615     // Load 32-bit regs
4616     for(hr=0;hr<HOST_REGS;hr++) {
4617       if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4618         #ifdef DESTRUCTIVE_WRITEBACK
4619         if(i_regmap[hr]!=regs[t].regmap_entry[hr] || ( !((regs[t].dirty>>hr)&1) && ((i_dirty>>hr)&1) && (((i_is32&~unneeded_reg_upper[t])>>i_regmap[hr])&1) ) || (((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
4620         #else
4621         if(i_regmap[hr]!=regs[t].regmap_entry[hr] ) {
4622         #endif
4623           if(regs[t].regmap_entry[hr]==0) {
4624             emit_zeroreg(hr);
4625           }
4626           else if(regs[t].regmap_entry[hr]!=CCREG)
4627           {
4628             emit_loadreg(regs[t].regmap_entry[hr],hr);
4629           }
4630         }
4631       }
4632     }
4633     //Load 64-bit regs
4634     for(hr=0;hr<HOST_REGS;hr++) {
4635       if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=64&&regs[t].regmap_entry[hr]<TEMPREG+64) {
4636         if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4637           assert(regs[t].regmap_entry[hr]!=64);
4638           if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
4639             int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4640             if(lr<0) {
4641               emit_loadreg(regs[t].regmap_entry[hr],hr);
4642             }
4643             else
4644             {
4645               emit_sarimm(lr,31,hr);
4646             }
4647           }
4648           else
4649           {
4650             emit_loadreg(regs[t].regmap_entry[hr],hr);
4651           }
4652         }
4653         else if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
4654           int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4655           assert(lr>=0);
4656           emit_sarimm(lr,31,hr);
4657         }
4658       }
4659     }
4660   }
4661 }
4662
4663 int match_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4664 {
4665   if(addr>=start && addr<start+slen*4-4)
4666   {
4667     int t=(addr-start)>>2;
4668     int hr;
4669     if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4670     for(hr=0;hr<HOST_REGS;hr++)
4671     {
4672       if(hr!=EXCLUDE_REG)
4673       {
4674         if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4675         {
4676           if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4677           {
4678             return 0;
4679           }
4680           else
4681           if((i_dirty>>hr)&1)
4682           {
4683             if(i_regmap[hr]<TEMPREG)
4684             {
4685               if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4686                 return 0;
4687             }
4688             else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4689             {
4690               if(!((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1))
4691                 return 0;
4692             }
4693           }
4694         }
4695         else // Same register but is it 32-bit or dirty?
4696         if(i_regmap[hr]>=0)
4697         {
4698           if(!((regs[t].dirty>>hr)&1))
4699           {
4700             if((i_dirty>>hr)&1)
4701             {
4702               if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4703               {
4704                 //printf("%x: dirty no match\n",addr);
4705                 return 0;
4706               }
4707             }
4708           }
4709           if((((regs[t].was32^i_is32)&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)
4710           {
4711             //printf("%x: is32 no match\n",addr);
4712             return 0;
4713           }
4714         }
4715       }
4716     }
4717     //if(is32[t]&~unneeded_reg_upper[t]&~i_is32) return 0;
4718 #ifndef FORCE32
4719     if(requires_32bit[t]&~i_is32) return 0;
4720 #endif
4721     // Delay slots are not valid branch targets
4722     //if(t>0&&(itype[t-1]==RJUMP||itype[t-1]==UJUMP||itype[t-1]==CJUMP||itype[t-1]==SJUMP||itype[t-1]==FJUMP)) return 0;
4723     // Delay slots require additional processing, so do not match
4724     if(is_ds[t]) return 0;
4725   }
4726   else
4727   {
4728     int hr;
4729     for(hr=0;hr<HOST_REGS;hr++)
4730     {
4731       if(hr!=EXCLUDE_REG)
4732       {
4733         if(i_regmap[hr]>=0)
4734         {
4735           if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4736           {
4737             if((i_dirty>>hr)&1)
4738             {
4739               return 0;
4740             }
4741           }
4742         }
4743       }
4744     }
4745   }
4746   return 1;
4747 }
4748
4749 // Used when a branch jumps into the delay slot of another branch
4750 void ds_assemble_entry(int i)
4751 {
4752   int t=(ba[i]-start)>>2;
4753   if(!instr_addr[t]) instr_addr[t]=(u_int)out;
4754   assem_debug("Assemble delay slot at %x\n",ba[i]);
4755   assem_debug("<->\n");
4756   if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4757     wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty,regs[t].was32);
4758   load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,rs1[t],rs2[t]);
4759   address_generation(t,&regs[t],regs[t].regmap_entry);
4760   if(itype[t]==STORE||itype[t]==STORELR||(opcode[t]&0x3b)==0x39||(opcode[t]&0x3b)==0x3a)
4761     load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,INVCP,INVCP);
4762   cop1_usable=0;
4763   is_delayslot=0;
4764   switch(itype[t]) {
4765     case ALU:
4766       alu_assemble(t,&regs[t]);break;
4767     case IMM16:
4768       imm16_assemble(t,&regs[t]);break;
4769     case SHIFT:
4770       shift_assemble(t,&regs[t]);break;
4771     case SHIFTIMM:
4772       shiftimm_assemble(t,&regs[t]);break;
4773     case LOAD:
4774       load_assemble(t,&regs[t]);break;
4775     case LOADLR:
4776       loadlr_assemble(t,&regs[t]);break;
4777     case STORE:
4778       store_assemble(t,&regs[t]);break;
4779     case STORELR:
4780       storelr_assemble(t,&regs[t]);break;
4781     case COP0:
4782       cop0_assemble(t,&regs[t]);break;
4783     case COP1:
4784       cop1_assemble(t,&regs[t]);break;
4785     case C1LS:
4786       c1ls_assemble(t,&regs[t]);break;
4787     case COP2:
4788       cop2_assemble(t,&regs[t]);break;
4789     case C2LS:
4790       c2ls_assemble(t,&regs[t]);break;
4791     case C2OP:
4792       c2op_assemble(t,&regs[t]);break;
4793     case FCONV:
4794       fconv_assemble(t,&regs[t]);break;
4795     case FLOAT:
4796       float_assemble(t,&regs[t]);break;
4797     case FCOMP:
4798       fcomp_assemble(t,&regs[t]);break;
4799     case MULTDIV:
4800       multdiv_assemble(t,&regs[t]);break;
4801     case MOV:
4802       mov_assemble(t,&regs[t]);break;
4803     case SYSCALL:
4804     case HLECALL:
4805     case INTCALL:
4806     case SPAN:
4807     case UJUMP:
4808     case RJUMP:
4809     case CJUMP:
4810     case SJUMP:
4811     case FJUMP:
4812       printf("Jump in the delay slot.  This is probably a bug.\n");
4813   }
4814   store_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
4815   load_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
4816   if(internal_branch(regs[t].is32,ba[i]+4))
4817     assem_debug("branch: internal\n");
4818   else
4819     assem_debug("branch: external\n");
4820   assert(internal_branch(regs[t].is32,ba[i]+4));
4821   add_to_linker((int)out,ba[i]+4,internal_branch(regs[t].is32,ba[i]+4));
4822   emit_jmp(0);
4823 }
4824
4825 void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
4826 {
4827   int count;
4828   int jaddr;
4829   int idle=0;
4830   if(itype[i]==RJUMP)
4831   {
4832     *adj=0;
4833   }
4834   //if(ba[i]>=start && ba[i]<(start+slen*4))
4835   if(internal_branch(branch_regs[i].is32,ba[i]))
4836   {
4837     int t=(ba[i]-start)>>2;
4838     if(is_ds[t]) *adj=-1; // Branch into delay slot adds an extra cycle
4839     else *adj=ccadj[t];
4840   }
4841   else
4842   {
4843     *adj=0;
4844   }
4845   count=ccadj[i];
4846   if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4847     // Idle loop
4848     if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4849     idle=(int)out;
4850     //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4851     emit_andimm(HOST_CCREG,3,HOST_CCREG);
4852     jaddr=(int)out;
4853     emit_jmp(0);
4854   }
4855   else if(*adj==0||invert) {
4856     emit_addimm_and_set_flags(CLOCK_DIVIDER*(count+2),HOST_CCREG);
4857     jaddr=(int)out;
4858     emit_jns(0);
4859   }
4860   else
4861   {
4862     emit_cmpimm(HOST_CCREG,-CLOCK_DIVIDER*(count+2));
4863     jaddr=(int)out;
4864     emit_jns(0);
4865   }
4866   add_stub(CC_STUB,jaddr,idle?idle:(int)out,(*adj==0||invert||idle)?0:(count+2),i,addr,taken,0);
4867 }
4868
4869 void do_ccstub(int n)
4870 {
4871   literal_pool(256);
4872   assem_debug("do_ccstub %x\n",start+stubs[n][4]*4);
4873   set_jump_target(stubs[n][1],(int)out);
4874   int i=stubs[n][4];
4875   if(stubs[n][6]==NULLDS) {
4876     // Delay slot instruction is nullified ("likely" branch)
4877     wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
4878   }
4879   else if(stubs[n][6]!=TAKEN) {
4880     wb_dirtys(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty);
4881   }
4882   else {
4883     if(internal_branch(branch_regs[i].is32,ba[i]))
4884       wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
4885   }
4886   if(stubs[n][5]!=-1)
4887   {
4888     // Save PC as return address
4889     emit_movimm(stubs[n][5],EAX);
4890     emit_writeword(EAX,(int)&pcaddr);
4891   }
4892   else
4893   {
4894     // Return address depends on which way the branch goes
4895     if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
4896     {
4897       int s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4898       int s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
4899       int s2l=get_reg(branch_regs[i].regmap,rs2[i]);
4900       int s2h=get_reg(branch_regs[i].regmap,rs2[i]|64);
4901       if(rs1[i]==0)
4902       {
4903         s1l=s2l;s1h=s2h;
4904         s2l=s2h=-1;
4905       }
4906       else if(rs2[i]==0)
4907       {
4908         s2l=s2h=-1;
4909       }
4910       if((branch_regs[i].is32>>rs1[i])&(branch_regs[i].is32>>rs2[i])&1) {
4911         s1h=s2h=-1;
4912       }
4913       assert(s1l>=0);
4914       #ifdef DESTRUCTIVE_WRITEBACK
4915       if(rs1[i]) {
4916         if((branch_regs[i].dirty>>s1l)&(branch_regs[i].is32>>rs1[i])&1)
4917           emit_loadreg(rs1[i],s1l);
4918       }
4919       else {
4920         if((branch_regs[i].dirty>>s1l)&(branch_regs[i].is32>>rs2[i])&1)
4921           emit_loadreg(rs2[i],s1l);
4922       }
4923       if(s2l>=0)
4924         if((branch_regs[i].dirty>>s2l)&(branch_regs[i].is32>>rs2[i])&1)
4925           emit_loadreg(rs2[i],s2l);
4926       #endif
4927       int hr=0;
4928       int addr=-1,alt=-1,ntaddr=-1;
4929       while(hr<HOST_REGS)
4930       {
4931         if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4932            (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4933            (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4934         {
4935           addr=hr++;break;
4936         }
4937         hr++;
4938       }
4939       while(hr<HOST_REGS)
4940       {
4941         if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4942            (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4943            (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4944         {
4945           alt=hr++;break;
4946         }
4947         hr++;
4948       }
4949       if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
4950       {
4951         while(hr<HOST_REGS)
4952         {
4953           if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4954              (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4955              (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4956           {
4957             ntaddr=hr;break;
4958           }
4959           hr++;
4960         }
4961         assert(hr<HOST_REGS);
4962       }
4963       if((opcode[i]&0x2f)==4) // BEQ
4964       {
4965         #ifdef HAVE_CMOV_IMM
4966         if(s1h<0) {
4967           if(s2l>=0) emit_cmp(s1l,s2l);
4968           else emit_test(s1l,s1l);
4969           emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4970         }
4971         else
4972         #endif
4973         {
4974           emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4975           if(s1h>=0) {
4976             if(s2h>=0) emit_cmp(s1h,s2h);
4977             else emit_test(s1h,s1h);
4978             emit_cmovne_reg(alt,addr);
4979           }
4980           if(s2l>=0) emit_cmp(s1l,s2l);
4981           else emit_test(s1l,s1l);
4982           emit_cmovne_reg(alt,addr);
4983         }
4984       }
4985       if((opcode[i]&0x2f)==5) // BNE
4986       {
4987         #ifdef HAVE_CMOV_IMM
4988         if(s1h<0) {
4989           if(s2l>=0) emit_cmp(s1l,s2l);
4990           else emit_test(s1l,s1l);
4991           emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4992         }
4993         else
4994         #endif
4995         {
4996           emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4997           if(s1h>=0) {
4998             if(s2h>=0) emit_cmp(s1h,s2h);
4999             else emit_test(s1h,s1h);
5000             emit_cmovne_reg(alt,addr);
5001           }
5002           if(s2l>=0) emit_cmp(s1l,s2l);
5003           else emit_test(s1l,s1l);
5004           emit_cmovne_reg(alt,addr);
5005         }
5006       }
5007       if((opcode[i]&0x2f)==6) // BLEZ
5008       {
5009         //emit_movimm(ba[i],alt);
5010         //emit_movimm(start+i*4+8,addr);
5011         emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5012         emit_cmpimm(s1l,1);
5013         if(s1h>=0) emit_mov(addr,ntaddr);
5014         emit_cmovl_reg(alt,addr);
5015         if(s1h>=0) {
5016           emit_test(s1h,s1h);
5017           emit_cmovne_reg(ntaddr,addr);
5018           emit_cmovs_reg(alt,addr);
5019         }
5020       }
5021       if((opcode[i]&0x2f)==7) // BGTZ
5022       {
5023         //emit_movimm(ba[i],addr);
5024         //emit_movimm(start+i*4+8,ntaddr);
5025         emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5026         emit_cmpimm(s1l,1);
5027         if(s1h>=0) emit_mov(addr,alt);
5028         emit_cmovl_reg(ntaddr,addr);
5029         if(s1h>=0) {
5030           emit_test(s1h,s1h);
5031           emit_cmovne_reg(alt,addr);
5032           emit_cmovs_reg(ntaddr,addr);
5033         }
5034       }
5035       if((opcode[i]==1)&&(opcode2[i]&0x2D)==0) // BLTZ
5036       {
5037         //emit_movimm(ba[i],alt);
5038         //emit_movimm(start+i*4+8,addr);
5039         emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5040         if(s1h>=0) emit_test(s1h,s1h);
5041         else emit_test(s1l,s1l);
5042         emit_cmovs_reg(alt,addr);
5043       }
5044       if((opcode[i]==1)&&(opcode2[i]&0x2D)==1) // BGEZ
5045       {
5046         //emit_movimm(ba[i],addr);
5047         //emit_movimm(start+i*4+8,alt);
5048         emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5049         if(s1h>=0) emit_test(s1h,s1h);
5050         else emit_test(s1l,s1l);
5051         emit_cmovs_reg(alt,addr);
5052       }
5053       if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
5054         if(source[i]&0x10000) // BC1T
5055         {
5056           //emit_movimm(ba[i],alt);
5057           //emit_movimm(start+i*4+8,addr);
5058           emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5059           emit_testimm(s1l,0x800000);
5060           emit_cmovne_reg(alt,addr);
5061         }
5062         else // BC1F
5063         {
5064           //emit_movimm(ba[i],addr);
5065           //emit_movimm(start+i*4+8,alt);
5066           emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5067           emit_testimm(s1l,0x800000);
5068           emit_cmovne_reg(alt,addr);
5069         }
5070       }
5071       emit_writeword(addr,(int)&pcaddr);
5072     }
5073     else
5074     if(itype[i]==RJUMP)
5075     {
5076       int r=get_reg(branch_regs[i].regmap,rs1[i]);
5077       if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
5078         r=get_reg(branch_regs[i].regmap,RTEMP);
5079       }
5080       emit_writeword(r,(int)&pcaddr);
5081     }
5082     else {printf("Unknown branch type in do_ccstub\n");exit(1);}
5083   }
5084   // Update cycle count
5085   assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5086   if(stubs[n][3]) emit_addimm(HOST_CCREG,CLOCK_DIVIDER*stubs[n][3],HOST_CCREG);
5087   emit_call((int)cc_interrupt);
5088   if(stubs[n][3]) emit_addimm(HOST_CCREG,-CLOCK_DIVIDER*stubs[n][3],HOST_CCREG);
5089   if(stubs[n][6]==TAKEN) {
5090     if(internal_branch(branch_regs[i].is32,ba[i]))
5091       load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
5092     else if(itype[i]==RJUMP) {
5093       if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5094         emit_readword((int)&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5095       else
5096         emit_loadreg(rs1[i],get_reg(branch_regs[i].regmap,rs1[i]));
5097     }
5098   }else if(stubs[n][6]==NOTTAKEN) {
5099     if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5100     else load_all_regs(branch_regs[i].regmap);
5101   }else if(stubs[n][6]==NULLDS) {
5102     // Delay slot instruction is nullified ("likely" branch)
5103     if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5104     else load_all_regs(regs[i].regmap);
5105   }else{
5106     load_all_regs(branch_regs[i].regmap);
5107   }
5108   emit_jmp(stubs[n][2]); // return address
5109
5110   /* This works but uses a lot of memory...
5111   emit_readword((int)&last_count,ECX);
5112   emit_add(HOST_CCREG,ECX,EAX);
5113   emit_writeword(EAX,(int)&Count);
5114   emit_call((int)gen_interupt);
5115   emit_readword((int)&Count,HOST_CCREG);
5116   emit_readword((int)&next_interupt,EAX);
5117   emit_readword((int)&pending_exception,EBX);
5118   emit_writeword(EAX,(int)&last_count);
5119   emit_sub(HOST_CCREG,EAX,HOST_CCREG);
5120   emit_test(EBX,EBX);
5121   int jne_instr=(int)out;
5122   emit_jne(0);
5123   if(stubs[n][3]) emit_addimm(HOST_CCREG,-2*stubs[n][3],HOST_CCREG);
5124   load_all_regs(branch_regs[i].regmap);
5125   emit_jmp(stubs[n][2]); // return address
5126   set_jump_target(jne_instr,(int)out);
5127   emit_readword((int)&pcaddr,EAX);
5128   // Call get_addr_ht instead of doing the hash table here.
5129   // This code is executed infrequently and takes up a lot of space
5130   // so smaller is better.
5131   emit_storereg(CCREG,HOST_CCREG);
5132   emit_pushreg(EAX);
5133   emit_call((int)get_addr_ht);
5134   emit_loadreg(CCREG,HOST_CCREG);
5135   emit_addimm(ESP,4,ESP);
5136   emit_jmpreg(EAX);*/
5137 }
5138
5139 add_to_linker(int addr,int target,int ext)
5140 {
5141   link_addr[linkcount][0]=addr;
5142   link_addr[linkcount][1]=target;
5143   link_addr[linkcount][2]=ext;
5144   linkcount++;
5145 }
5146
5147 void ujump_assemble(int i,struct regstat *i_regs)
5148 {
5149   signed char *i_regmap=i_regs->regmap;
5150   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5151   address_generation(i+1,i_regs,regs[i].regmap_entry);
5152   #ifdef REG_PREFETCH
5153   int temp=get_reg(branch_regs[i].regmap,PTEMP);
5154   if(rt1[i]==31&&temp>=0)
5155   {
5156     int return_address=start+i*4+8;
5157     if(get_reg(branch_regs[i].regmap,31)>0)
5158     if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5159   }
5160   #endif
5161   if(rt1[i]==31) {
5162     int rt;
5163     unsigned int return_address;
5164     rt=get_reg(branch_regs[i].regmap,31);
5165     assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5166     //assert(rt>=0);
5167     return_address=start+i*4+8;
5168     if(rt>=0) {
5169       #ifdef USE_MINI_HT
5170       if(internal_branch(branch_regs[i].is32,return_address)&&rt1[i+1]!=31) {
5171         int temp=-1; // note: must be ds-safe
5172         #ifdef HOST_TEMPREG
5173         temp=HOST_TEMPREG;
5174         #endif
5175         if(temp>=0) do_miniht_insert(return_address,rt,temp);
5176         else emit_movimm(return_address,rt);
5177       }
5178       else
5179       #endif
5180       {
5181         #ifdef REG_PREFETCH
5182         if(temp>=0)
5183         {
5184           if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5185         }
5186         #endif
5187         emit_movimm(return_address,rt); // PC into link register
5188         #ifdef IMM_PREFETCH
5189         emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5190         #endif
5191       }
5192     }
5193   }
5194   ds_assemble(i+1,i_regs);
5195   uint64_t bc_unneeded=branch_regs[i].u;
5196   uint64_t bc_unneeded_upper=branch_regs[i].uu;
5197   bc_unneeded|=1|(1LL<<rt1[i]);
5198   bc_unneeded_upper|=1|(1LL<<rt1[i]);
5199   wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5200                 bc_unneeded,bc_unneeded_upper);
5201   load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5202   int cc,adj;
5203   cc=get_reg(branch_regs[i].regmap,CCREG);
5204   assert(cc==HOST_CCREG);
5205   store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5206   #ifdef REG_PREFETCH
5207   if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
5208   #endif
5209   do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5210   if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5211   load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5212   if(internal_branch(branch_regs[i].is32,ba[i]))
5213     assem_debug("branch: internal\n");
5214   else
5215     assem_debug("branch: external\n");
5216   if(internal_branch(branch_regs[i].is32,ba[i])&&is_ds[(ba[i]-start)>>2]) {
5217     ds_assemble_entry(i);
5218   }
5219   else {
5220     add_to_linker((int)out,ba[i],internal_branch(branch_regs[i].is32,ba[i]));
5221     emit_jmp(0);
5222   }
5223 }
5224
5225 void rjump_assemble(int i,struct regstat *i_regs)
5226 {
5227   signed char *i_regmap=i_regs->regmap;
5228   int temp;
5229   int rs,cc,adj;
5230   rs=get_reg(branch_regs[i].regmap,rs1[i]);
5231   assert(rs>=0);
5232   if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
5233     // Delay slot abuse, make a copy of the branch address register
5234     temp=get_reg(branch_regs[i].regmap,RTEMP);
5235     assert(temp>=0);
5236     assert(regs[i].regmap[temp]==RTEMP);
5237     emit_mov(rs,temp);
5238     rs=temp;
5239   }
5240   address_generation(i+1,i_regs,regs[i].regmap_entry);
5241   #ifdef REG_PREFETCH
5242   if(rt1[i]==31)
5243   {
5244     if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5245       int return_address=start+i*4+8;
5246       if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5247     }
5248   }
5249   #endif
5250   #ifdef USE_MINI_HT
5251   if(rs1[i]==31) {
5252     int rh=get_reg(regs[i].regmap,RHASH);
5253     if(rh>=0) do_preload_rhash(rh);
5254   }
5255   #endif
5256   ds_assemble(i+1,i_regs);
5257   uint64_t bc_unneeded=branch_regs[i].u;
5258   uint64_t bc_unneeded_upper=branch_regs[i].uu;
5259   bc_unneeded|=1|(1LL<<rt1[i]);
5260   bc_unneeded_upper|=1|(1LL<<rt1[i]);
5261   bc_unneeded&=~(1LL<<rs1[i]);
5262   wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5263                 bc_unneeded,bc_unneeded_upper);
5264   load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],CCREG);
5265   if(rt1[i]!=0) {
5266     int rt,return_address;
5267     assert(rt1[i+1]!=rt1[i]);
5268     assert(rt2[i+1]!=rt1[i]);
5269     rt=get_reg(branch_regs[i].regmap,rt1[i]);
5270     assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5271     assert(rt>=0);
5272     return_address=start+i*4+8;
5273     #ifdef REG_PREFETCH
5274     if(temp>=0)
5275     {
5276       if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5277     }
5278     #endif
5279     emit_movimm(return_address,rt); // PC into link register
5280     #ifdef IMM_PREFETCH
5281     emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5282     #endif
5283   }
5284   cc=get_reg(branch_regs[i].regmap,CCREG);
5285   assert(cc==HOST_CCREG);
5286   #ifdef USE_MINI_HT
5287   int rh=get_reg(branch_regs[i].regmap,RHASH);
5288   int ht=get_reg(branch_regs[i].regmap,RHTBL);
5289   if(rs1[i]==31) {
5290     if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5291     do_preload_rhtbl(ht);
5292     do_rhash(rs,rh);
5293   }
5294   #endif
5295   store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
5296   #ifdef DESTRUCTIVE_WRITEBACK
5297   if((branch_regs[i].dirty>>rs)&(branch_regs[i].is32>>rs1[i])&1) {
5298     if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
5299       emit_loadreg(rs1[i],rs);
5300     }
5301   }
5302   #endif
5303   #ifdef REG_PREFETCH
5304   if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
5305   #endif
5306   #ifdef USE_MINI_HT
5307   if(rs1[i]==31) {
5308     do_miniht_load(ht,rh);
5309   }
5310   #endif
5311   //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5312   //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5313   //assert(adj==0);
5314   emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
5315   add_stub(CC_STUB,(int)out,jump_vaddr_reg[rs],0,i,-1,TAKEN,0);
5316   emit_jns(0);
5317   //load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
5318   #ifdef USE_MINI_HT
5319   if(rs1[i]==31) {
5320     do_miniht_jump(rs,rh,ht);
5321   }
5322   else
5323   #endif
5324   {
5325     //if(rs!=EAX) emit_mov(rs,EAX);
5326     //emit_jmp((int)jump_vaddr_eax);
5327     emit_jmp(jump_vaddr_reg[rs]);
5328   }
5329   /* Check hash table
5330   temp=!rs;
5331   emit_mov(rs,temp);
5332   emit_shrimm(rs,16,rs);
5333   emit_xor(temp,rs,rs);
5334   emit_movzwl_reg(rs,rs);
5335   emit_shlimm(rs,4,rs);
5336   emit_cmpmem_indexed((int)hash_table,rs,temp);
5337   emit_jne((int)out+14);
5338   emit_readword_indexed((int)hash_table+4,rs,rs);
5339   emit_jmpreg(rs);
5340   emit_cmpmem_indexed((int)hash_table+8,rs,temp);
5341   emit_addimm_no_flags(8,rs);
5342   emit_jeq((int)out-17);
5343   // No hit on hash table, call compiler
5344   emit_pushreg(temp);
5345 //DEBUG >
5346 #ifdef DEBUG_CYCLE_COUNT
5347   emit_readword((int)&last_count,ECX);
5348   emit_add(HOST_CCREG,ECX,HOST_CCREG);
5349   emit_readword((int)&next_interupt,ECX);
5350   emit_writeword(HOST_CCREG,(int)&Count);
5351   emit_sub(HOST_CCREG,ECX,HOST_CCREG);
5352   emit_writeword(ECX,(int)&last_count);
5353 #endif
5354 //DEBUG <
5355   emit_storereg(CCREG,HOST_CCREG);
5356   emit_call((int)get_addr);
5357   emit_loadreg(CCREG,HOST_CCREG);
5358   emit_addimm(ESP,4,ESP);
5359   emit_jmpreg(EAX);*/
5360   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5361   if(rt1[i]!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5362   #endif
5363 }
5364
5365 void cjump_assemble(int i,struct regstat *i_regs)
5366 {
5367   signed char *i_regmap=i_regs->regmap;
5368   int cc;
5369   int match;
5370   match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5371   assem_debug("match=%d\n",match);
5372   int s1h,s1l,s2h,s2l;
5373   int prev_cop1_usable=cop1_usable;
5374   int unconditional=0,nop=0;
5375   int only32=0;
5376   int invert=0;
5377   int internal=internal_branch(branch_regs[i].is32,ba[i]);
5378   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5379   if(!match) invert=1;
5380   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5381   if(i>(ba[i]-start)>>2) invert=1;
5382   #endif
5383
5384   if(ooo[i]) {
5385     s1l=get_reg(branch_regs[i].regmap,rs1[i]);
5386     s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
5387     s2l=get_reg(branch_regs[i].regmap,rs2[i]);
5388     s2h=get_reg(branch_regs[i].regmap,rs2[i]|64);
5389   }
5390   else {
5391     s1l=get_reg(i_regmap,rs1[i]);
5392     s1h=get_reg(i_regmap,rs1[i]|64);
5393     s2l=get_reg(i_regmap,rs2[i]);
5394     s2h=get_reg(i_regmap,rs2[i]|64);
5395   }
5396   if(rs1[i]==0&&rs2[i]==0)
5397   {
5398     if(opcode[i]&1) nop=1;
5399     else unconditional=1;
5400     //assert(opcode[i]!=5);
5401     //assert(opcode[i]!=7);
5402     //assert(opcode[i]!=0x15);
5403     //assert(opcode[i]!=0x17);
5404   }
5405   else if(rs1[i]==0)
5406   {
5407     s1l=s2l;s1h=s2h;
5408     s2l=s2h=-1;
5409     only32=(regs[i].was32>>rs2[i])&1;
5410   }
5411   else if(rs2[i]==0)
5412   {
5413     s2l=s2h=-1;
5414     only32=(regs[i].was32>>rs1[i])&1;
5415   }
5416   else {
5417     only32=(regs[i].was32>>rs1[i])&(regs[i].was32>>rs2[i])&1;
5418   }
5419
5420   if(ooo[i]) {
5421     // Out of order execution (delay slot first)
5422     //printf("OOOE\n");
5423     address_generation(i+1,i_regs,regs[i].regmap_entry);
5424     ds_assemble(i+1,i_regs);
5425     int adj;
5426     uint64_t bc_unneeded=branch_regs[i].u;
5427     uint64_t bc_unneeded_upper=branch_regs[i].uu;
5428     bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5429     bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5430     bc_unneeded|=1;
5431     bc_unneeded_upper|=1;
5432     wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5433                   bc_unneeded,bc_unneeded_upper);
5434     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs2[i]);
5435     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5436     cc=get_reg(branch_regs[i].regmap,CCREG);
5437     assert(cc==HOST_CCREG);
5438     if(unconditional)
5439       store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5440     //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5441     //assem_debug("cycle count (adj)\n");
5442     if(unconditional) {
5443       do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5444       if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5445         if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5446         load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5447         if(internal)
5448           assem_debug("branch: internal\n");
5449         else
5450           assem_debug("branch: external\n");
5451         if(internal&&is_ds[(ba[i]-start)>>2]) {
5452           ds_assemble_entry(i);
5453         }
5454         else {
5455           add_to_linker((int)out,ba[i],internal);
5456           emit_jmp(0);
5457         }
5458         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5459         if(((u_int)out)&7) emit_addnop(0);
5460         #endif
5461       }
5462     }
5463     else if(nop) {
5464       emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5465       int jaddr=(int)out;
5466       emit_jns(0);
5467       add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5468     }
5469     else {
5470       int taken=0,nottaken=0,nottaken1=0;
5471       do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5472       if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5473       if(!only32)
5474       {
5475         assert(s1h>=0);
5476         if(opcode[i]==4) // BEQ
5477         {
5478           if(s2h>=0) emit_cmp(s1h,s2h);
5479           else emit_test(s1h,s1h);
5480           nottaken1=(int)out;
5481           emit_jne(1);
5482         }
5483         if(opcode[i]==5) // BNE
5484         {
5485           if(s2h>=0) emit_cmp(s1h,s2h);
5486           else emit_test(s1h,s1h);
5487           if(invert) taken=(int)out;
5488           else add_to_linker((int)out,ba[i],internal);
5489           emit_jne(0);
5490         }
5491         if(opcode[i]==6) // BLEZ
5492         {
5493           emit_test(s1h,s1h);
5494           if(invert) taken=(int)out;
5495           else add_to_linker((int)out,ba[i],internal);
5496           emit_js(0);
5497           nottaken1=(int)out;
5498           emit_jne(1);
5499         }
5500         if(opcode[i]==7) // BGTZ
5501         {
5502           emit_test(s1h,s1h);
5503           nottaken1=(int)out;
5504           emit_js(1);
5505           if(invert) taken=(int)out;
5506           else add_to_linker((int)out,ba[i],internal);
5507           emit_jne(0);
5508         }
5509       } // if(!only32)
5510
5511       //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5512       assert(s1l>=0);
5513       if(opcode[i]==4) // BEQ
5514       {
5515         if(s2l>=0) emit_cmp(s1l,s2l);
5516         else emit_test(s1l,s1l);
5517         if(invert){
5518           nottaken=(int)out;
5519           emit_jne(1);
5520         }else{
5521           add_to_linker((int)out,ba[i],internal);
5522           emit_jeq(0);
5523         }
5524       }
5525       if(opcode[i]==5) // BNE
5526       {
5527         if(s2l>=0) emit_cmp(s1l,s2l);
5528         else emit_test(s1l,s1l);
5529         if(invert){
5530           nottaken=(int)out;
5531           emit_jeq(1);
5532         }else{
5533           add_to_linker((int)out,ba[i],internal);
5534           emit_jne(0);
5535         }
5536       }
5537       if(opcode[i]==6) // BLEZ
5538       {
5539         emit_cmpimm(s1l,1);
5540         if(invert){
5541           nottaken=(int)out;
5542           emit_jge(1);
5543         }else{
5544           add_to_linker((int)out,ba[i],internal);
5545           emit_jl(0);
5546         }
5547       }
5548       if(opcode[i]==7) // BGTZ
5549       {
5550         emit_cmpimm(s1l,1);
5551         if(invert){
5552           nottaken=(int)out;
5553           emit_jl(1);
5554         }else{
5555           add_to_linker((int)out,ba[i],internal);
5556           emit_jge(0);
5557         }
5558       }
5559       if(invert) {
5560         if(taken) set_jump_target(taken,(int)out);
5561         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5562         if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5563           if(adj) {
5564             emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5565             add_to_linker((int)out,ba[i],internal);
5566           }else{
5567             emit_addnop(13);
5568             add_to_linker((int)out,ba[i],internal*2);
5569           }
5570           emit_jmp(0);
5571         }else
5572         #endif
5573         {
5574           if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5575           store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5576           load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5577           if(internal)
5578             assem_debug("branch: internal\n");
5579           else
5580             assem_debug("branch: external\n");
5581           if(internal&&is_ds[(ba[i]-start)>>2]) {
5582             ds_assemble_entry(i);
5583           }
5584           else {
5585             add_to_linker((int)out,ba[i],internal);
5586             emit_jmp(0);
5587           }
5588         }
5589         set_jump_target(nottaken,(int)out);
5590       }
5591
5592       if(nottaken1) set_jump_target(nottaken1,(int)out);
5593       if(adj) {
5594         if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
5595       }
5596     } // (!unconditional)
5597   } // if(ooo)
5598   else
5599   {
5600     // In-order execution (branch first)
5601     //if(likely[i]) printf("IOL\n");
5602     //else
5603     //printf("IOE\n");
5604     int taken=0,nottaken=0,nottaken1=0;
5605     if(!unconditional&&!nop) {
5606       if(!only32)
5607       {
5608         assert(s1h>=0);
5609         if((opcode[i]&0x2f)==4) // BEQ
5610         {
5611           if(s2h>=0) emit_cmp(s1h,s2h);
5612           else emit_test(s1h,s1h);
5613           nottaken1=(int)out;
5614           emit_jne(2);
5615         }
5616         if((opcode[i]&0x2f)==5) // BNE
5617         {
5618           if(s2h>=0) emit_cmp(s1h,s2h);
5619           else emit_test(s1h,s1h);
5620           taken=(int)out;
5621           emit_jne(1);
5622         }
5623         if((opcode[i]&0x2f)==6) // BLEZ
5624         {
5625           emit_test(s1h,s1h);
5626           taken=(int)out;
5627           emit_js(1);
5628           nottaken1=(int)out;
5629           emit_jne(2);
5630         }
5631         if((opcode[i]&0x2f)==7) // BGTZ
5632         {
5633           emit_test(s1h,s1h);
5634           nottaken1=(int)out;
5635           emit_js(2);
5636           taken=(int)out;
5637           emit_jne(1);
5638         }
5639       } // if(!only32)
5640
5641       //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5642       assert(s1l>=0);
5643       if((opcode[i]&0x2f)==4) // BEQ
5644       {
5645         if(s2l>=0) emit_cmp(s1l,s2l);
5646         else emit_test(s1l,s1l);
5647         nottaken=(int)out;
5648         emit_jne(2);
5649       }
5650       if((opcode[i]&0x2f)==5) // BNE
5651       {
5652         if(s2l>=0) emit_cmp(s1l,s2l);
5653         else emit_test(s1l,s1l);
5654         nottaken=(int)out;
5655         emit_jeq(2);
5656       }
5657       if((opcode[i]&0x2f)==6) // BLEZ
5658       {
5659         emit_cmpimm(s1l,1);
5660         nottaken=(int)out;
5661         emit_jge(2);
5662       }
5663       if((opcode[i]&0x2f)==7) // BGTZ
5664       {
5665         emit_cmpimm(s1l,1);
5666         nottaken=(int)out;
5667         emit_jl(2);
5668       }
5669     } // if(!unconditional)
5670     int adj;
5671     uint64_t ds_unneeded=branch_regs[i].u;
5672     uint64_t ds_unneeded_upper=branch_regs[i].uu;
5673     ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5674     ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
5675     if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
5676     ds_unneeded|=1;
5677     ds_unneeded_upper|=1;
5678     // branch taken
5679     if(!nop) {
5680       if(taken) set_jump_target(taken,(int)out);
5681       assem_debug("1:\n");
5682       wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5683                     ds_unneeded,ds_unneeded_upper);
5684       // load regs
5685       load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5686       address_generation(i+1,&branch_regs[i],0);
5687       load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
5688       ds_assemble(i+1,&branch_regs[i]);
5689       cc=get_reg(branch_regs[i].regmap,CCREG);
5690       if(cc==-1) {
5691         emit_loadreg(CCREG,cc=HOST_CCREG);
5692         // CHECK: Is the following instruction (fall thru) allocated ok?
5693       }
5694       assert(cc==HOST_CCREG);
5695       store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5696       do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5697       assem_debug("cycle count (adj)\n");
5698       if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5699       load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5700       if(internal)
5701         assem_debug("branch: internal\n");
5702       else
5703         assem_debug("branch: external\n");
5704       if(internal&&is_ds[(ba[i]-start)>>2]) {
5705         ds_assemble_entry(i);
5706       }
5707       else {
5708         add_to_linker((int)out,ba[i],internal);
5709         emit_jmp(0);
5710       }
5711     }
5712     // branch not taken
5713     cop1_usable=prev_cop1_usable;
5714     if(!unconditional) {
5715       if(nottaken1) set_jump_target(nottaken1,(int)out);
5716       set_jump_target(nottaken,(int)out);
5717       assem_debug("2:\n");
5718       if(!likely[i]) {
5719         wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5720                       ds_unneeded,ds_unneeded_upper);
5721         load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5722         address_generation(i+1,&branch_regs[i],0);
5723         load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5724         ds_assemble(i+1,&branch_regs[i]);
5725       }
5726       cc=get_reg(branch_regs[i].regmap,CCREG);
5727       if(cc==-1&&!likely[i]) {
5728         // Cycle count isn't in a register, temporarily load it then write it out
5729         emit_loadreg(CCREG,HOST_CCREG);
5730         emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
5731         int jaddr=(int)out;
5732         emit_jns(0);
5733         add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5734         emit_storereg(CCREG,HOST_CCREG);
5735       }
5736       else{
5737         cc=get_reg(i_regmap,CCREG);
5738         assert(cc==HOST_CCREG);
5739         emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5740         int jaddr=(int)out;
5741         emit_jns(0);
5742         add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5743       }
5744     }
5745   }
5746 }
5747
5748 void sjump_assemble(int i,struct regstat *i_regs)
5749 {
5750   signed char *i_regmap=i_regs->regmap;
5751   int cc;
5752   int match;
5753   match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5754   assem_debug("smatch=%d\n",match);
5755   int s1h,s1l;
5756   int prev_cop1_usable=cop1_usable;
5757   int unconditional=0,nevertaken=0;
5758   int only32=0;
5759   int invert=0;
5760   int internal=internal_branch(branch_regs[i].is32,ba[i]);
5761   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5762   if(!match) invert=1;
5763   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5764   if(i>(ba[i]-start)>>2) invert=1;
5765   #endif
5766
5767   //if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
5768   //assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
5769
5770   if(ooo[i]) {
5771     s1l=get_reg(branch_regs[i].regmap,rs1[i]);
5772     s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
5773   }
5774   else {
5775     s1l=get_reg(i_regmap,rs1[i]);
5776     s1h=get_reg(i_regmap,rs1[i]|64);
5777   }
5778   if(rs1[i]==0)
5779   {
5780     if(opcode2[i]&1) unconditional=1;
5781     else nevertaken=1;
5782     // These are never taken (r0 is never less than zero)
5783     //assert(opcode2[i]!=0);
5784     //assert(opcode2[i]!=2);
5785     //assert(opcode2[i]!=0x10);
5786     //assert(opcode2[i]!=0x12);
5787   }
5788   else {
5789     only32=(regs[i].was32>>rs1[i])&1;
5790   }
5791
5792   if(ooo[i]) {
5793     // Out of order execution (delay slot first)
5794     //printf("OOOE\n");
5795     address_generation(i+1,i_regs,regs[i].regmap_entry);
5796     ds_assemble(i+1,i_regs);
5797     int adj;
5798     uint64_t bc_unneeded=branch_regs[i].u;
5799     uint64_t bc_unneeded_upper=branch_regs[i].uu;
5800     bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5801     bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5802     bc_unneeded|=1;
5803     bc_unneeded_upper|=1;
5804     wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5805                   bc_unneeded,bc_unneeded_upper);
5806     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs1[i]);
5807     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5808     if(rt1[i]==31) {
5809       int rt,return_address;
5810       rt=get_reg(branch_regs[i].regmap,31);
5811       assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5812       if(rt>=0) {
5813         // Save the PC even if the branch is not taken
5814         return_address=start+i*4+8;
5815         emit_movimm(return_address,rt); // PC into link register
5816         #ifdef IMM_PREFETCH
5817         if(!nevertaken) emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5818         #endif
5819       }
5820     }
5821     cc=get_reg(branch_regs[i].regmap,CCREG);
5822     assert(cc==HOST_CCREG);
5823     if(unconditional)
5824       store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5825     //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5826     assem_debug("cycle count (adj)\n");
5827     if(unconditional) {
5828       do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5829       if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5830         if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5831         load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5832         if(internal)
5833           assem_debug("branch: internal\n");
5834         else
5835           assem_debug("branch: external\n");
5836         if(internal&&is_ds[(ba[i]-start)>>2]) {
5837           ds_assemble_entry(i);
5838         }
5839         else {
5840           add_to_linker((int)out,ba[i],internal);
5841           emit_jmp(0);
5842         }
5843         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5844         if(((u_int)out)&7) emit_addnop(0);
5845         #endif
5846       }
5847     }
5848     else if(nevertaken) {
5849       emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5850       int jaddr=(int)out;
5851       emit_jns(0);
5852       add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5853     }
5854     else {
5855       int nottaken=0;
5856       do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5857       if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5858       if(!only32)
5859       {
5860         assert(s1h>=0);
5861         if((opcode2[i]&0xf)==0) // BLTZ/BLTZAL
5862         {
5863           emit_test(s1h,s1h);
5864           if(invert){
5865             nottaken=(int)out;
5866             emit_jns(1);
5867           }else{
5868             add_to_linker((int)out,ba[i],internal);
5869             emit_js(0);
5870           }
5871         }
5872         if((opcode2[i]&0xf)==1) // BGEZ/BLTZAL
5873         {
5874           emit_test(s1h,s1h);
5875           if(invert){
5876             nottaken=(int)out;
5877             emit_js(1);
5878           }else{
5879             add_to_linker((int)out,ba[i],internal);
5880             emit_jns(0);
5881           }
5882         }
5883       } // if(!only32)
5884       else
5885       {
5886         assert(s1l>=0);
5887         if((opcode2[i]&0xf)==0) // BLTZ/BLTZAL
5888         {
5889           emit_test(s1l,s1l);
5890           if(invert){
5891             nottaken=(int)out;
5892             emit_jns(1);
5893           }else{
5894             add_to_linker((int)out,ba[i],internal);
5895             emit_js(0);
5896           }
5897         }
5898         if((opcode2[i]&0xf)==1) // BGEZ/BLTZAL
5899         {
5900           emit_test(s1l,s1l);
5901           if(invert){
5902             nottaken=(int)out;
5903             emit_js(1);
5904           }else{
5905             add_to_linker((int)out,ba[i],internal);
5906             emit_jns(0);
5907           }
5908         }
5909       } // if(!only32)
5910
5911       if(invert) {
5912         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5913         if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5914           if(adj) {
5915             emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5916             add_to_linker((int)out,ba[i],internal);
5917           }else{
5918             emit_addnop(13);
5919             add_to_linker((int)out,ba[i],internal*2);
5920           }
5921           emit_jmp(0);
5922         }else
5923         #endif
5924         {
5925           if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5926           store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5927           load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5928           if(internal)
5929             assem_debug("branch: internal\n");
5930           else
5931             assem_debug("branch: external\n");
5932           if(internal&&is_ds[(ba[i]-start)>>2]) {
5933             ds_assemble_entry(i);
5934           }
5935           else {
5936             add_to_linker((int)out,ba[i],internal);
5937             emit_jmp(0);
5938           }
5939         }
5940         set_jump_target(nottaken,(int)out);
5941       }
5942
5943       if(adj) {
5944         if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
5945       }
5946     } // (!unconditional)
5947   } // if(ooo)
5948   else
5949   {
5950     // In-order execution (branch first)
5951     //printf("IOE\n");
5952     int nottaken=0;
5953     if(rt1[i]==31) {
5954       int rt,return_address;
5955       rt=get_reg(branch_regs[i].regmap,31);
5956       if(rt>=0) {
5957         // Save the PC even if the branch is not taken
5958         return_address=start+i*4+8;
5959         emit_movimm(return_address,rt); // PC into link register
5960         #ifdef IMM_PREFETCH
5961         emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5962         #endif
5963       }
5964     }
5965     if(!unconditional) {
5966       //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5967       if(!only32)
5968       {
5969         assert(s1h>=0);
5970         if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5971         {
5972           emit_test(s1h,s1h);
5973           nottaken=(int)out;
5974           emit_jns(1);
5975         }
5976         if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5977         {
5978           emit_test(s1h,s1h);
5979           nottaken=(int)out;
5980           emit_js(1);
5981         }
5982       } // if(!only32)
5983       else
5984       {
5985         assert(s1l>=0);
5986         if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5987         {
5988           emit_test(s1l,s1l);
5989           nottaken=(int)out;
5990           emit_jns(1);
5991         }
5992         if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5993         {
5994           emit_test(s1l,s1l);
5995           nottaken=(int)out;
5996           emit_js(1);
5997         }
5998       }
5999     } // if(!unconditional)
6000     int adj;
6001     uint64_t ds_unneeded=branch_regs[i].u;
6002     uint64_t ds_unneeded_upper=branch_regs[i].uu;
6003     ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6004     ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6005     if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
6006     ds_unneeded|=1;
6007     ds_unneeded_upper|=1;
6008     // branch taken
6009     if(!nevertaken) {
6010       //assem_debug("1:\n");
6011       wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6012                     ds_unneeded,ds_unneeded_upper);
6013       // load regs
6014       load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
6015       address_generation(i+1,&branch_regs[i],0);
6016       load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
6017       ds_assemble(i+1,&branch_regs[i]);
6018       cc=get_reg(branch_regs[i].regmap,CCREG);
6019       if(cc==-1) {
6020         emit_loadreg(CCREG,cc=HOST_CCREG);
6021         // CHECK: Is the following instruction (fall thru) allocated ok?
6022       }
6023       assert(cc==HOST_CCREG);
6024       store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6025       do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
6026       assem_debug("cycle count (adj)\n");
6027       if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
6028       load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6029       if(internal)
6030         assem_debug("branch: internal\n");
6031       else
6032         assem_debug("branch: external\n");
6033       if(internal&&is_ds[(ba[i]-start)>>2]) {
6034         ds_assemble_entry(i);
6035       }
6036       else {
6037         add_to_linker((int)out,ba[i],internal);
6038         emit_jmp(0);
6039       }
6040     }
6041     // branch not taken
6042     cop1_usable=prev_cop1_usable;
6043     if(!unconditional) {
6044       set_jump_target(nottaken,(int)out);
6045       assem_debug("1:\n");
6046       if(!likely[i]) {
6047         wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6048                       ds_unneeded,ds_unneeded_upper);
6049         load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
6050         address_generation(i+1,&branch_regs[i],0);
6051         load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
6052         ds_assemble(i+1,&branch_regs[i]);
6053       }
6054       cc=get_reg(branch_regs[i].regmap,CCREG);
6055       if(cc==-1&&!likely[i]) {
6056         // Cycle count isn't in a register, temporarily load it then write it out
6057         emit_loadreg(CCREG,HOST_CCREG);
6058         emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
6059         int jaddr=(int)out;
6060         emit_jns(0);
6061         add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
6062         emit_storereg(CCREG,HOST_CCREG);
6063       }
6064       else{
6065         cc=get_reg(i_regmap,CCREG);
6066         assert(cc==HOST_CCREG);
6067         emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
6068         int jaddr=(int)out;
6069         emit_jns(0);
6070         add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
6071       }
6072     }
6073   }
6074 }
6075
6076 void fjump_assemble(int i,struct regstat *i_regs)
6077 {
6078   signed char *i_regmap=i_regs->regmap;
6079   int cc;
6080   int match;
6081   match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6082   assem_debug("fmatch=%d\n",match);
6083   int fs,cs;
6084   int eaddr;
6085   int invert=0;
6086   int internal=internal_branch(branch_regs[i].is32,ba[i]);
6087   if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
6088   if(!match) invert=1;
6089   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
6090   if(i>(ba[i]-start)>>2) invert=1;
6091   #endif
6092
6093   if(ooo[i]) {
6094     fs=get_reg(branch_regs[i].regmap,FSREG);
6095     address_generation(i+1,i_regs,regs[i].regmap_entry); // Is this okay?
6096   }
6097   else {
6098     fs=get_reg(i_regmap,FSREG);
6099   }
6100
6101   // Check cop1 unusable
6102   if(!cop1_usable) {
6103     cs=get_reg(i_regmap,CSREG);
6104     assert(cs>=0);
6105     emit_testimm(cs,0x20000000);
6106     eaddr=(int)out;
6107     emit_jeq(0);
6108     add_stub(FP_STUB,eaddr,(int)out,i,cs,(int)i_regs,0,0);
6109     cop1_usable=1;
6110   }
6111
6112   if(ooo[i]) {
6113     // Out of order execution (delay slot first)
6114     //printf("OOOE\n");
6115     ds_assemble(i+1,i_regs);
6116     int adj;
6117     uint64_t bc_unneeded=branch_regs[i].u;
6118     uint64_t bc_unneeded_upper=branch_regs[i].uu;
6119     bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
6120     bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
6121     bc_unneeded|=1;
6122     bc_unneeded_upper|=1;
6123     wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6124                   bc_unneeded,bc_unneeded_upper);
6125     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs1[i]);
6126     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
6127     cc=get_reg(branch_regs[i].regmap,CCREG);
6128     assert(cc==HOST_CCREG);
6129     do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
6130     assem_debug("cycle count (adj)\n");
6131     if(1) {
6132       int nottaken=0;
6133       if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
6134       if(1) {
6135         assert(fs>=0);
6136         emit_testimm(fs,0x800000);
6137         if(source[i]&0x10000) // BC1T
6138         {
6139           if(invert){
6140             nottaken=(int)out;
6141             emit_jeq(1);
6142           }else{
6143             add_to_linker((int)out,ba[i],internal);
6144             emit_jne(0);
6145           }
6146         }
6147         else // BC1F
6148           if(invert){
6149             nottaken=(int)out;
6150             emit_jne(1);
6151           }else{
6152             add_to_linker((int)out,ba[i],internal);
6153             emit_jeq(0);
6154           }
6155         {
6156         }
6157       } // if(!only32)
6158
6159       if(invert) {
6160         if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
6161         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
6162         else if(match) emit_addnop(13);
6163         #endif
6164         store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6165         load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6166         if(internal)
6167           assem_debug("branch: internal\n");
6168         else
6169           assem_debug("branch: external\n");
6170         if(internal&&is_ds[(ba[i]-start)>>2]) {
6171           ds_assemble_entry(i);
6172         }
6173         else {
6174           add_to_linker((int)out,ba[i],internal);
6175           emit_jmp(0);
6176         }
6177         set_jump_target(nottaken,(int)out);
6178       }
6179
6180       if(adj) {
6181         if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
6182       }
6183     } // (!unconditional)
6184   } // if(ooo)
6185   else
6186   {
6187     // In-order execution (branch first)
6188     //printf("IOE\n");
6189     int nottaken=0;
6190     if(1) {
6191       //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
6192       if(1) {
6193         assert(fs>=0);
6194         emit_testimm(fs,0x800000);
6195         if(source[i]&0x10000) // BC1T
6196         {
6197           nottaken=(int)out;
6198           emit_jeq(1);
6199         }
6200         else // BC1F
6201         {
6202           nottaken=(int)out;
6203           emit_jne(1);
6204         }
6205       }
6206     } // if(!unconditional)
6207     int adj;
6208     uint64_t ds_unneeded=branch_regs[i].u;
6209     uint64_t ds_unneeded_upper=branch_regs[i].uu;
6210     ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6211     ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6212     if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
6213     ds_unneeded|=1;
6214     ds_unneeded_upper|=1;
6215     // branch taken
6216     //assem_debug("1:\n");
6217     wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6218                   ds_unneeded,ds_unneeded_upper);
6219     // load regs
6220     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
6221     address_generation(i+1,&branch_regs[i],0);
6222     load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
6223     ds_assemble(i+1,&branch_regs[i]);
6224     cc=get_reg(branch_regs[i].regmap,CCREG);
6225     if(cc==-1) {
6226       emit_loadreg(CCREG,cc=HOST_CCREG);
6227       // CHECK: Is the following instruction (fall thru) allocated ok?
6228     }
6229     assert(cc==HOST_CCREG);
6230     store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6231     do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
6232     assem_debug("cycle count (adj)\n");
6233     if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
6234     load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6235     if(internal)
6236       assem_debug("branch: internal\n");
6237     else
6238       assem_debug("branch: external\n");
6239     if(internal&&is_ds[(ba[i]-start)>>2]) {
6240       ds_assemble_entry(i);
6241     }
6242     else {
6243       add_to_linker((int)out,ba[i],internal);
6244       emit_jmp(0);
6245     }
6246
6247     // branch not taken
6248     if(1) { // <- FIXME (don't need this)
6249       set_jump_target(nottaken,(int)out);
6250       assem_debug("1:\n");
6251       if(!likely[i]) {
6252         wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6253                       ds_unneeded,ds_unneeded_upper);
6254         load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
6255         address_generation(i+1,&branch_regs[i],0);
6256         load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
6257         ds_assemble(i+1,&branch_regs[i]);
6258       }
6259       cc=get_reg(branch_regs[i].regmap,CCREG);
6260       if(cc==-1&&!likely[i]) {
6261         // Cycle count isn't in a register, temporarily load it then write it out
6262         emit_loadreg(CCREG,HOST_CCREG);
6263         emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
6264         int jaddr=(int)out;
6265         emit_jns(0);
6266         add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
6267         emit_storereg(CCREG,HOST_CCREG);
6268       }
6269       else{
6270         cc=get_reg(i_regmap,CCREG);
6271         assert(cc==HOST_CCREG);
6272         emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
6273         int jaddr=(int)out;
6274         emit_jns(0);
6275         add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
6276       }
6277     }
6278   }
6279 }
6280
6281 static void pagespan_assemble(int i,struct regstat *i_regs)
6282 {
6283   int s1l=get_reg(i_regs->regmap,rs1[i]);
6284   int s1h=get_reg(i_regs->regmap,rs1[i]|64);
6285   int s2l=get_reg(i_regs->regmap,rs2[i]);
6286   int s2h=get_reg(i_regs->regmap,rs2[i]|64);
6287   void *nt_branch=NULL;
6288   int taken=0;
6289   int nottaken=0;
6290   int unconditional=0;
6291   if(rs1[i]==0)
6292   {
6293     s1l=s2l;s1h=s2h;
6294     s2l=s2h=-1;
6295   }
6296   else if(rs2[i]==0)
6297   {
6298     s2l=s2h=-1;
6299   }
6300   if((i_regs->is32>>rs1[i])&(i_regs->is32>>rs2[i])&1) {
6301     s1h=s2h=-1;
6302   }
6303   int hr=0;
6304   int addr,alt,ntaddr;
6305   if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
6306   else {
6307     while(hr<HOST_REGS)
6308     {
6309       if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
6310          (i_regs->regmap[hr]&63)!=rs1[i] &&
6311          (i_regs->regmap[hr]&63)!=rs2[i] )
6312       {
6313         addr=hr++;break;
6314       }
6315       hr++;
6316     }
6317   }
6318   while(hr<HOST_REGS)
6319   {
6320     if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
6321        (i_regs->regmap[hr]&63)!=rs1[i] &&
6322        (i_regs->regmap[hr]&63)!=rs2[i] )
6323     {
6324       alt=hr++;break;
6325     }
6326     hr++;
6327   }
6328   if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
6329   {
6330     while(hr<HOST_REGS)
6331     {
6332       if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
6333          (i_regs->regmap[hr]&63)!=rs1[i] &&
6334          (i_regs->regmap[hr]&63)!=rs2[i] )
6335       {
6336         ntaddr=hr;break;
6337       }
6338       hr++;
6339     }
6340   }
6341   assert(hr<HOST_REGS);
6342   if((opcode[i]&0x2e)==4||opcode[i]==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
6343     load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
6344   }
6345   emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
6346   if(opcode[i]==2) // J
6347   {
6348     unconditional=1;
6349   }
6350   if(opcode[i]==3) // JAL
6351   {
6352     // TODO: mini_ht
6353     int rt=get_reg(i_regs->regmap,31);
6354     emit_movimm(start+i*4+8,rt);
6355     unconditional=1;
6356   }
6357   if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
6358   {
6359     emit_mov(s1l,addr);
6360     if(opcode2[i]==9) // JALR
6361     {
6362       int rt=get_reg(i_regs->regmap,rt1[i]);
6363       emit_movimm(start+i*4+8,rt);
6364     }
6365   }
6366   if((opcode[i]&0x3f)==4) // BEQ
6367   {
6368     if(rs1[i]==rs2[i])
6369     {
6370       unconditional=1;
6371     }
6372     else
6373     #ifdef HAVE_CMOV_IMM
6374     if(s1h<0) {
6375       if(s2l>=0) emit_cmp(s1l,s2l);
6376       else emit_test(s1l,s1l);
6377       emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
6378     }
6379     else
6380     #endif
6381     {
6382       assert(s1l>=0);
6383       emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6384       if(s1h>=0) {
6385         if(s2h>=0) emit_cmp(s1h,s2h);
6386         else emit_test(s1h,s1h);
6387         emit_cmovne_reg(alt,addr);
6388       }
6389       if(s2l>=0) emit_cmp(s1l,s2l);
6390       else emit_test(s1l,s1l);
6391       emit_cmovne_reg(alt,addr);
6392     }
6393   }
6394   if((opcode[i]&0x3f)==5) // BNE
6395   {
6396     #ifdef HAVE_CMOV_IMM
6397     if(s1h<0) {
6398       if(s2l>=0) emit_cmp(s1l,s2l);
6399       else emit_test(s1l,s1l);
6400       emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
6401     }
6402     else
6403     #endif
6404     {
6405       assert(s1l>=0);
6406       emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
6407       if(s1h>=0) {
6408         if(s2h>=0) emit_cmp(s1h,s2h);
6409         else emit_test(s1h,s1h);
6410         emit_cmovne_reg(alt,addr);
6411       }
6412       if(s2l>=0) emit_cmp(s1l,s2l);
6413       else emit_test(s1l,s1l);
6414       emit_cmovne_reg(alt,addr);
6415     }
6416   }
6417   if((opcode[i]&0x3f)==0x14) // BEQL
6418   {
6419     if(s1h>=0) {
6420       if(s2h>=0) emit_cmp(s1h,s2h);
6421       else emit_test(s1h,s1h);
6422       nottaken=(int)out;
6423       emit_jne(0);
6424     }
6425     if(s2l>=0) emit_cmp(s1l,s2l);
6426     else emit_test(s1l,s1l);
6427     if(nottaken) set_jump_target(nottaken,(int)out);
6428     nottaken=(int)out;
6429     emit_jne(0);
6430   }
6431   if((opcode[i]&0x3f)==0x15) // BNEL
6432   {
6433     if(s1h>=0) {
6434       if(s2h>=0) emit_cmp(s1h,s2h);
6435       else emit_test(s1h,s1h);
6436       taken=(int)out;
6437       emit_jne(0);
6438     }
6439     if(s2l>=0) emit_cmp(s1l,s2l);
6440     else emit_test(s1l,s1l);
6441     nottaken=(int)out;
6442     emit_jeq(0);
6443     if(taken) set_jump_target(taken,(int)out);
6444   }
6445   if((opcode[i]&0x3f)==6) // BLEZ
6446   {
6447     emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6448     emit_cmpimm(s1l,1);
6449     if(s1h>=0) emit_mov(addr,ntaddr);
6450     emit_cmovl_reg(alt,addr);
6451     if(s1h>=0) {
6452       emit_test(s1h,s1h);
6453       emit_cmovne_reg(ntaddr,addr);
6454       emit_cmovs_reg(alt,addr);
6455     }
6456   }
6457   if((opcode[i]&0x3f)==7) // BGTZ
6458   {
6459     emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
6460     emit_cmpimm(s1l,1);
6461     if(s1h>=0) emit_mov(addr,alt);
6462     emit_cmovl_reg(ntaddr,addr);
6463     if(s1h>=0) {
6464       emit_test(s1h,s1h);
6465       emit_cmovne_reg(alt,addr);
6466       emit_cmovs_reg(ntaddr,addr);
6467     }
6468   }
6469   if((opcode[i]&0x3f)==0x16) // BLEZL
6470   {
6471     assert((opcode[i]&0x3f)!=0x16);
6472   }
6473   if((opcode[i]&0x3f)==0x17) // BGTZL
6474   {
6475     assert((opcode[i]&0x3f)!=0x17);
6476   }
6477   assert(opcode[i]!=1); // BLTZ/BGEZ
6478
6479   //FIXME: Check CSREG
6480   if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
6481     if((source[i]&0x30000)==0) // BC1F
6482     {
6483       emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6484       emit_testimm(s1l,0x800000);
6485       emit_cmovne_reg(alt,addr);
6486     }
6487     if((source[i]&0x30000)==0x10000) // BC1T
6488     {
6489       emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6490       emit_testimm(s1l,0x800000);
6491       emit_cmovne_reg(alt,addr);
6492     }
6493     if((source[i]&0x30000)==0x20000) // BC1FL
6494     {
6495       emit_testimm(s1l,0x800000);
6496       nottaken=(int)out;
6497       emit_jne(0);
6498     }
6499     if((source[i]&0x30000)==0x30000) // BC1TL
6500     {
6501       emit_testimm(s1l,0x800000);
6502       nottaken=(int)out;
6503       emit_jeq(0);
6504     }
6505   }
6506
6507   assert(i_regs->regmap[HOST_CCREG]==CCREG);
6508   wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
6509   if(likely[i]||unconditional)
6510   {
6511     emit_movimm(ba[i],HOST_BTREG);
6512   }
6513   else if(addr!=HOST_BTREG)
6514   {
6515     emit_mov(addr,HOST_BTREG);
6516   }
6517   void *branch_addr=out;
6518   emit_jmp(0);
6519   int target_addr=start+i*4+5;
6520   void *stub=out;
6521   void *compiled_target_addr=check_addr(target_addr);
6522   emit_extjump_ds((int)branch_addr,target_addr);
6523   if(compiled_target_addr) {
6524     set_jump_target((int)branch_addr,(int)compiled_target_addr);
6525     add_link(target_addr,stub);
6526   }
6527   else set_jump_target((int)branch_addr,(int)stub);
6528   if(likely[i]) {
6529     // Not-taken path
6530     set_jump_target((int)nottaken,(int)out);
6531     wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
6532     void *branch_addr=out;
6533     emit_jmp(0);
6534     int target_addr=start+i*4+8;
6535     void *stub=out;
6536     void *compiled_target_addr=check_addr(target_addr);
6537     emit_extjump_ds((int)branch_addr,target_addr);
6538     if(compiled_target_addr) {
6539       set_jump_target((int)branch_addr,(int)compiled_target_addr);
6540       add_link(target_addr,stub);
6541     }
6542     else set_jump_target((int)branch_addr,(int)stub);
6543   }
6544 }
6545
6546 // Assemble the delay slot for the above
6547 static void pagespan_ds()
6548 {
6549   assem_debug("initial delay slot:\n");
6550   u_int vaddr=start+1;
6551   u_int page=get_page(vaddr);
6552   u_int vpage=get_vpage(vaddr);
6553   ll_add(jump_dirty+vpage,vaddr,(void *)out);
6554   do_dirty_stub_ds();
6555   ll_add(jump_in+page,vaddr,(void *)out);
6556   assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
6557   if(regs[0].regmap[HOST_CCREG]!=CCREG)
6558     wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty,regs[0].was32);
6559   if(regs[0].regmap[HOST_BTREG]!=BTREG)
6560     emit_writeword(HOST_BTREG,(int)&branch_target);
6561   load_regs(regs[0].regmap_entry,regs[0].regmap,regs[0].was32,rs1[0],rs2[0]);
6562   address_generation(0,&regs[0],regs[0].regmap_entry);
6563   if(itype[0]==STORE||itype[0]==STORELR||(opcode[0]&0x3b)==0x39||(opcode[0]&0x3b)==0x3a)
6564     load_regs(regs[0].regmap_entry,regs[0].regmap,regs[0].was32,INVCP,INVCP);
6565   cop1_usable=0;
6566   is_delayslot=0;
6567   switch(itype[0]) {
6568     case ALU:
6569       alu_assemble(0,&regs[0]);break;
6570     case IMM16:
6571       imm16_assemble(0,&regs[0]);break;
6572     case SHIFT:
6573       shift_assemble(0,&regs[0]);break;
6574     case SHIFTIMM:
6575       shiftimm_assemble(0,&regs[0]);break;
6576     case LOAD:
6577       load_assemble(0,&regs[0]);break;
6578     case LOADLR:
6579       loadlr_assemble(0,&regs[0]);break;
6580     case STORE:
6581       store_assemble(0,&regs[0]);break;
6582     case STORELR:
6583       storelr_assemble(0,&regs[0]);break;
6584     case COP0:
6585       cop0_assemble(0,&regs[0]);break;
6586     case COP1:
6587       cop1_assemble(0,&regs[0]);break;
6588     case C1LS:
6589       c1ls_assemble(0,&regs[0]);break;
6590     case COP2:
6591       cop2_assemble(0,&regs[0]);break;
6592     case C2LS:
6593       c2ls_assemble(0,&regs[0]);break;
6594     case C2OP:
6595       c2op_assemble(0,&regs[0]);break;
6596     case FCONV:
6597       fconv_assemble(0,&regs[0]);break;
6598     case FLOAT:
6599       float_assemble(0,&regs[0]);break;
6600     case FCOMP:
6601       fcomp_assemble(0,&regs[0]);break;
6602     case MULTDIV:
6603       multdiv_assemble(0,&regs[0]);break;
6604     case MOV:
6605       mov_assemble(0,&regs[0]);break;
6606     case SYSCALL:
6607     case HLECALL:
6608     case INTCALL:
6609     case SPAN:
6610     case UJUMP:
6611     case RJUMP:
6612     case CJUMP:
6613     case SJUMP:
6614     case FJUMP:
6615       printf("Jump in the delay slot.  This is probably a bug.\n");
6616   }
6617   int btaddr=get_reg(regs[0].regmap,BTREG);
6618   if(btaddr<0) {
6619     btaddr=get_reg(regs[0].regmap,-1);
6620     emit_readword((int)&branch_target,btaddr);
6621   }
6622   assert(btaddr!=HOST_CCREG);
6623   if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
6624 #ifdef HOST_IMM8
6625   emit_movimm(start+4,HOST_TEMPREG);
6626   emit_cmp(btaddr,HOST_TEMPREG);
6627 #else
6628   emit_cmpimm(btaddr,start+4);
6629 #endif
6630   int branch=(int)out;
6631   emit_jeq(0);
6632   store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,-1);
6633   emit_jmp(jump_vaddr_reg[btaddr]);
6634   set_jump_target(branch,(int)out);
6635   store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
6636   load_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
6637 }
6638
6639 // Basic liveness analysis for MIPS registers
6640 void unneeded_registers(int istart,int iend,int r)
6641 {
6642   int i;
6643   uint64_t u,uu,b,bu;
6644   uint64_t temp_u,temp_uu;
6645   uint64_t tdep;
6646   if(iend==slen-1) {
6647     u=1;uu=1;
6648   }else{
6649     u=unneeded_reg[iend+1];
6650     uu=unneeded_reg_upper[iend+1];
6651     u=1;uu=1;
6652   }
6653   for (i=iend;i>=istart;i--)
6654   {
6655     //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6656     if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
6657     {
6658       // If subroutine call, flag return address as a possible branch target
6659       if(rt1[i]==31 && i<slen-2) bt[i+2]=1;
6660
6661       if(ba[i]<start || ba[i]>=(start+slen*4))
6662       {
6663         // Branch out of this block, flush all regs
6664         u=1;
6665         uu=1;
6666         /* Hexagon hack
6667         if(itype[i]==UJUMP&&rt1[i]==31)
6668         {
6669           uu=u=0x300C00F; // Discard at, v0-v1, t6-t9
6670         }
6671         if(itype[i]==RJUMP&&rs1[i]==31)
6672         {
6673           uu=u=0x300C0F3; // Discard at, a0-a3, t6-t9
6674         }
6675         if(start>0x80000400&&start<0x80000000+RAM_SIZE) {
6676           if(itype[i]==UJUMP&&rt1[i]==31)
6677           {
6678             //uu=u=0x30300FF0FLL; // Discard at, v0-v1, t0-t9, lo, hi
6679             uu=u=0x300FF0F; // Discard at, v0-v1, t0-t9
6680           }
6681           if(itype[i]==RJUMP&&rs1[i]==31)
6682           {
6683             //uu=u=0x30300FFF3LL; // Discard at, a0-a3, t0-t9, lo, hi
6684             uu=u=0x300FFF3; // Discard at, a0-a3, t0-t9
6685           }
6686         }*/
6687         branch_unneeded_reg[i]=u;
6688         branch_unneeded_reg_upper[i]=uu;
6689         // Merge in delay slot
6690         tdep=(~uu>>rt1[i+1])&1;
6691         u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6692         uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6693         u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6694         uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6695         uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6696         u|=1;uu|=1;
6697         // If branch is "likely" (and conditional)
6698         // then we skip the delay slot on the fall-thru path
6699         if(likely[i]) {
6700           if(i<slen-1) {
6701             u&=unneeded_reg[i+2];
6702             uu&=unneeded_reg_upper[i+2];
6703           }
6704           else
6705           {
6706             u=1;
6707             uu=1;
6708           }
6709         }
6710       }
6711       else
6712       {
6713         // Internal branch, flag target
6714         bt[(ba[i]-start)>>2]=1;
6715         if(ba[i]<=start+i*4) {
6716           // Backward branch
6717           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6718           {
6719             // Unconditional branch
6720             temp_u=1;temp_uu=1;
6721           } else {
6722             // Conditional branch (not taken case)
6723             temp_u=unneeded_reg[i+2];
6724             temp_uu=unneeded_reg_upper[i+2];
6725           }
6726           // Merge in delay slot
6727           tdep=(~temp_uu>>rt1[i+1])&1;
6728           temp_u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6729           temp_uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6730           temp_u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6731           temp_uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6732           temp_uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6733           temp_u|=1;temp_uu|=1;
6734           // If branch is "likely" (and conditional)
6735           // then we skip the delay slot on the fall-thru path
6736           if(likely[i]) {
6737             if(i<slen-1) {
6738               temp_u&=unneeded_reg[i+2];
6739               temp_uu&=unneeded_reg_upper[i+2];
6740             }
6741             else
6742             {
6743               temp_u=1;
6744               temp_uu=1;
6745             }
6746           }
6747           tdep=(~temp_uu>>rt1[i])&1;
6748           temp_u|=(1LL<<rt1[i])|(1LL<<rt2[i]);
6749           temp_uu|=(1LL<<rt1[i])|(1LL<<rt2[i]);
6750           temp_u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
6751           temp_uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
6752           temp_uu&=~((tdep<<dep1[i])|(tdep<<dep2[i]));
6753           temp_u|=1;temp_uu|=1;
6754           unneeded_reg[i]=temp_u;
6755           unneeded_reg_upper[i]=temp_uu;
6756           // Only go three levels deep.  This recursion can take an
6757           // excessive amount of time if there are a lot of nested loops.
6758           if(r<2) {
6759             unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6760           }else{
6761             unneeded_reg[(ba[i]-start)>>2]=1;
6762             unneeded_reg_upper[(ba[i]-start)>>2]=1;
6763           }
6764         } /*else*/ if(1) {
6765           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6766           {
6767             // Unconditional branch
6768             u=unneeded_reg[(ba[i]-start)>>2];
6769             uu=unneeded_reg_upper[(ba[i]-start)>>2];
6770             branch_unneeded_reg[i]=u;
6771             branch_unneeded_reg_upper[i]=uu;
6772         //u=1;
6773         //uu=1;
6774         //branch_unneeded_reg[i]=u;
6775         //branch_unneeded_reg_upper[i]=uu;
6776             // Merge in delay slot
6777             tdep=(~uu>>rt1[i+1])&1;
6778             u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6779             uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6780             u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6781             uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6782             uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6783             u|=1;uu|=1;
6784           } else {
6785             // Conditional branch
6786             b=unneeded_reg[(ba[i]-start)>>2];
6787             bu=unneeded_reg_upper[(ba[i]-start)>>2];
6788             branch_unneeded_reg[i]=b;
6789             branch_unneeded_reg_upper[i]=bu;
6790         //b=1;
6791         //bu=1;
6792         //branch_unneeded_reg[i]=b;
6793         //branch_unneeded_reg_upper[i]=bu;
6794             // Branch delay slot
6795             tdep=(~uu>>rt1[i+1])&1;
6796             b|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6797             bu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6798             b&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6799             bu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6800             bu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6801             b|=1;bu|=1;
6802             // If branch is "likely" then we skip the
6803             // delay slot on the fall-thru path
6804             if(likely[i]) {
6805               u=b;
6806               uu=bu;
6807               if(i<slen-1) {
6808                 u&=unneeded_reg[i+2];
6809                 uu&=unneeded_reg_upper[i+2];
6810         //u=1;
6811         //uu=1;
6812               }
6813             } else {
6814               u&=b;
6815               uu&=bu;
6816         //u=1;
6817         //uu=1;
6818             }
6819             if(i<slen-1) {
6820               branch_unneeded_reg[i]&=unneeded_reg[i+2];
6821               branch_unneeded_reg_upper[i]&=unneeded_reg_upper[i+2];
6822         //branch_unneeded_reg[i]=1;
6823         //branch_unneeded_reg_upper[i]=1;
6824             } else {
6825               branch_unneeded_reg[i]=1;
6826               branch_unneeded_reg_upper[i]=1;
6827             }
6828           }
6829         }
6830       }
6831     }
6832     else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
6833     {
6834       // SYSCALL instruction (software interrupt)
6835       u=1;
6836       uu=1;
6837     }
6838     else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
6839     {
6840       // ERET instruction (return from interrupt)
6841       u=1;
6842       uu=1;
6843     }
6844     //u=uu=1; // DEBUG
6845     tdep=(~uu>>rt1[i])&1;
6846     // Written registers are unneeded
6847     u|=1LL<<rt1[i];
6848     u|=1LL<<rt2[i];
6849     uu|=1LL<<rt1[i];
6850     uu|=1LL<<rt2[i];
6851     // Accessed registers are needed
6852     u&=~(1LL<<rs1[i]);
6853     u&=~(1LL<<rs2[i]);
6854     uu&=~(1LL<<us1[i]);
6855     uu&=~(1LL<<us2[i]);
6856     // Source-target dependencies
6857     uu&=~(tdep<<dep1[i]);
6858     uu&=~(tdep<<dep2[i]);
6859     // R0 is always unneeded
6860     u|=1;uu|=1;
6861     // Save it
6862     unneeded_reg[i]=u;
6863     unneeded_reg_upper[i]=uu;
6864     /*
6865     printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6866     printf("U:");
6867     int r;
6868     for(r=1;r<=CCREG;r++) {
6869       if((unneeded_reg[i]>>r)&1) {
6870         if(r==HIREG) printf(" HI");
6871         else if(r==LOREG) printf(" LO");
6872         else printf(" r%d",r);
6873       }
6874     }
6875     printf(" UU:");
6876     for(r=1;r<=CCREG;r++) {
6877       if(((unneeded_reg_upper[i]&~unneeded_reg[i])>>r)&1) {
6878         if(r==HIREG) printf(" HI");
6879         else if(r==LOREG) printf(" LO");
6880         else printf(" r%d",r);
6881       }
6882     }
6883     printf("\n");*/
6884   }
6885 #ifdef FORCE32
6886   for (i=iend;i>=istart;i--)
6887   {
6888     unneeded_reg_upper[i]=branch_unneeded_reg_upper[i]=-1LL;
6889   }
6890 #endif
6891 }
6892
6893 // Identify registers which are likely to contain 32-bit values
6894 // This is used to predict whether any branches will jump to a
6895 // location with 64-bit values in registers.
6896 static void provisional_32bit()
6897 {
6898   int i,j;
6899   uint64_t is32=1;
6900   uint64_t lastbranch=1;
6901
6902   for(i=0;i<slen;i++)
6903   {
6904     if(i>0) {
6905       if(itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP) {
6906         if(i>1) is32=lastbranch;
6907         else is32=1;
6908       }
6909     }
6910     if(i>1)
6911     {
6912       if(itype[i-2]==CJUMP||itype[i-2]==SJUMP||itype[i-2]==FJUMP) {
6913         if(likely[i-2]) {
6914           if(i>2) is32=lastbranch;
6915           else is32=1;
6916         }
6917       }
6918       if((opcode[i-2]&0x2f)==0x05) // BNE/BNEL
6919       {
6920         if(rs1[i-2]==0||rs2[i-2]==0)
6921         {
6922           if(rs1[i-2]) {
6923             is32|=1LL<<rs1[i-2];
6924           }
6925           if(rs2[i-2]) {
6926             is32|=1LL<<rs2[i-2];
6927           }
6928         }
6929       }
6930     }
6931     // If something jumps here with 64-bit values
6932     // then promote those registers to 64 bits
6933     if(bt[i])
6934     {
6935       uint64_t temp_is32=is32;
6936       for(j=i-1;j>=0;j--)
6937       {
6938         if(ba[j]==start+i*4)
6939           //temp_is32&=branch_regs[j].is32;
6940           temp_is32&=p32[j];
6941       }
6942       for(j=i;j<slen;j++)
6943       {
6944         if(ba[j]==start+i*4)
6945           temp_is32=1;
6946       }
6947       is32=temp_is32;
6948     }
6949     int type=itype[i];
6950     int op=opcode[i];
6951     int op2=opcode2[i];
6952     int rt=rt1[i];
6953     int s1=rs1[i];
6954     int s2=rs2[i];
6955     if(type==UJUMP||type==RJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
6956       // Branches don't write registers, consider the delay slot instead.
6957       type=itype[i+1];
6958       op=opcode[i+1];
6959       op2=opcode2[i+1];
6960       rt=rt1[i+1];
6961       s1=rs1[i+1];
6962       s2=rs2[i+1];
6963       lastbranch=is32;
6964     }
6965     switch(type) {
6966       case LOAD:
6967         if(opcode[i]==0x27||opcode[i]==0x37|| // LWU/LD
6968            opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
6969           is32&=~(1LL<<rt);
6970         else
6971           is32|=1LL<<rt;
6972         break;
6973       case STORE:
6974       case STORELR:
6975         break;
6976       case LOADLR:
6977         if(op==0x1a||op==0x1b) is32&=~(1LL<<rt); // LDR/LDL
6978         if(op==0x22) is32|=1LL<<rt; // LWL
6979         break;
6980       case IMM16:
6981         if (op==0x08||op==0x09|| // ADDI/ADDIU
6982             op==0x0a||op==0x0b|| // SLTI/SLTIU
6983             op==0x0c|| // ANDI
6984             op==0x0f)  // LUI
6985         {
6986           is32|=1LL<<rt;
6987         }
6988         if(op==0x18||op==0x19) { // DADDI/DADDIU
6989           is32&=~(1LL<<rt);
6990           //if(imm[i]==0)
6991           //  is32|=((is32>>s1)&1LL)<<rt;
6992         }
6993         if(op==0x0d||op==0x0e) { // ORI/XORI
6994           uint64_t sr=((is32>>s1)&1LL);
6995           is32&=~(1LL<<rt);
6996           is32|=sr<<rt;
6997         }
6998         break;
6999       case UJUMP:
7000         break;
7001       case RJUMP:
7002         break;
7003       case CJUMP:
7004         break;
7005       case SJUMP:
7006         break;
7007       case FJUMP:
7008         break;
7009       case ALU:
7010         if(op2>=0x20&&op2<=0x23) { // ADD/ADDU/SUB/SUBU
7011           is32|=1LL<<rt;
7012         }
7013         if(op2==0x2a||op2==0x2b) { // SLT/SLTU
7014           is32|=1LL<<rt;
7015         }
7016         else if(op2>=0x24&&op2<=0x27) { // AND/OR/XOR/NOR
7017           uint64_t sr=((is32>>s1)&(is32>>s2)&1LL);
7018           is32&=~(1LL<<rt);
7019           is32|=sr<<rt;
7020         }
7021         else if(op2>=0x2c&&op2<=0x2d) { // DADD/DADDU
7022           if(s1==0&&s2==0) {
7023             is32|=1LL<<rt;
7024           }
7025           else if(s2==0) {
7026             uint64_t sr=((is32>>s1)&1LL);
7027             is32&=~(1LL<<rt);
7028             is32|=sr<<rt;
7029           }
7030           else if(s1==0) {
7031             uint64_t sr=((is32>>s2)&1LL);
7032             is32&=~(1LL<<rt);
7033             is32|=sr<<rt;
7034           }
7035           else {
7036             is32&=~(1LL<<rt);
7037           }
7038         }
7039         else if(op2>=0x2e&&op2<=0x2f) { // DSUB/DSUBU
7040           if(s1==0&&s2==0) {
7041             is32|=1LL<<rt;
7042           }
7043           else if(s2==0) {
7044             uint64_t sr=((is32>>s1)&1LL);
7045             is32&=~(1LL<<rt);
7046             is32|=sr<<rt;
7047           }
7048           else {
7049             is32&=~(1LL<<rt);
7050           }
7051         }
7052         break;
7053       case MULTDIV:
7054         if (op2>=0x1c&&op2<=0x1f) { // DMULT/DMULTU/DDIV/DDIVU
7055           is32&=~((1LL<<HIREG)|(1LL<<LOREG));
7056         }
7057         else {
7058           is32|=(1LL<<HIREG)|(1LL<<LOREG);
7059         }
7060         break;
7061       case MOV:
7062         {
7063           uint64_t sr=((is32>>s1)&1LL);
7064           is32&=~(1LL<<rt);
7065           is32|=sr<<rt;
7066         }
7067         break;
7068       case SHIFT:
7069         if(op2>=0x14&&op2<=0x17) is32&=~(1LL<<rt); // DSLLV/DSRLV/DSRAV
7070         else is32|=1LL<<rt; // SLLV/SRLV/SRAV
7071         break;
7072       case SHIFTIMM:
7073         is32|=1LL<<rt;
7074         // DSLL/DSRL/DSRA/DSLL32/DSRL32 but not DSRA32 have 64-bit result
7075         if(op2>=0x38&&op2<0x3f) is32&=~(1LL<<rt);
7076         break;
7077       case COP0:
7078         if(op2==0) is32|=1LL<<rt; // MFC0
7079         break;
7080       case COP1:
7081       case COP2:
7082         if(op2==0) is32|=1LL<<rt; // MFC1
7083         if(op2==1) is32&=~(1LL<<rt); // DMFC1
7084         if(op2==2) is32|=1LL<<rt; // CFC1
7085         break;
7086       case C1LS:
7087       case C2LS:
7088         break;
7089       case FLOAT:
7090       case FCONV:
7091         break;
7092       case FCOMP:
7093         break;
7094       case C2OP:
7095       case SYSCALL:
7096       case HLECALL:
7097         break;
7098       default:
7099         break;
7100     }
7101     is32|=1;
7102     p32[i]=is32;
7103
7104     if(i>0)
7105     {
7106       if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
7107       {
7108         if(rt1[i-1]==31) // JAL/JALR
7109         {
7110           // Subroutine call will return here, don't alloc any registers
7111           is32=1;
7112         }
7113         else if(i+1<slen)
7114         {
7115           // Internal branch will jump here, match registers to caller
7116           is32=0x3FFFFFFFFLL;
7117         }
7118       }
7119     }
7120   }
7121 }
7122
7123 // Identify registers which may be assumed to contain 32-bit values
7124 // and where optimizations will rely on this.
7125 // This is used to determine whether backward branches can safely
7126 // jump to a location with 64-bit values in registers.
7127 static void provisional_r32()
7128 {
7129   u_int r32=0;
7130   int i;
7131
7132   for (i=slen-1;i>=0;i--)
7133   {
7134     int hr;
7135     if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
7136     {
7137       if(ba[i]<start || ba[i]>=(start+slen*4))
7138       {
7139         // Branch out of this block, don't need anything
7140         r32=0;
7141       }
7142       else
7143       {
7144         // Internal branch
7145         // Need whatever matches the target
7146         // (and doesn't get overwritten by the delay slot instruction)
7147         r32=0;
7148         int t=(ba[i]-start)>>2;
7149         if(ba[i]>start+i*4) {
7150           // Forward branch
7151           //if(!(requires_32bit[t]&~regs[i].was32))
7152           //  r32|=requires_32bit[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
7153           if(!(pr32[t]&~regs[i].was32))
7154             r32|=pr32[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
7155         }else{
7156           // Backward branch
7157           if(!(regs[t].was32&~unneeded_reg_upper[t]&~regs[i].was32))
7158             r32|=regs[t].was32&~unneeded_reg_upper[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
7159         }
7160       }
7161       // Conditional branch may need registers for following instructions
7162       if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
7163       {
7164         if(i<slen-2) {
7165           //r32|=requires_32bit[i+2];
7166           r32|=pr32[i+2];
7167           r32&=regs[i].was32;
7168           // Mark this address as a branch target since it may be called
7169           // upon return from interrupt
7170           //bt[i+2]=1;
7171         }
7172       }
7173       // Merge in delay slot
7174       if(!likely[i]) {
7175         // These are overwritten unless the branch is "likely"
7176         // and the delay slot is nullified if not taken
7177         r32&=~(1LL<<rt1[i+1]);
7178         r32&=~(1LL<<rt2[i+1]);
7179       }
7180       // Assume these are needed (delay slot)
7181       if(us1[i+1]>0)
7182       {
7183         if((regs[i].was32>>us1[i+1])&1) r32|=1LL<<us1[i+1];
7184       }
7185       if(us2[i+1]>0)
7186       {
7187         if((regs[i].was32>>us2[i+1])&1) r32|=1LL<<us2[i+1];
7188       }
7189       if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1))
7190       {
7191         if((regs[i].was32>>dep1[i+1])&1) r32|=1LL<<dep1[i+1];
7192       }
7193       if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1))
7194       {
7195         if((regs[i].was32>>dep2[i+1])&1) r32|=1LL<<dep2[i+1];
7196       }
7197     }
7198     else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
7199     {
7200       // SYSCALL instruction (software interrupt)
7201       r32=0;
7202     }
7203     else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
7204     {
7205       // ERET instruction (return from interrupt)
7206       r32=0;
7207     }
7208     // Check 32 bits
7209     r32&=~(1LL<<rt1[i]);
7210     r32&=~(1LL<<rt2[i]);
7211     if(us1[i]>0)
7212     {
7213       if((regs[i].was32>>us1[i])&1) r32|=1LL<<us1[i];
7214     }
7215     if(us2[i]>0)
7216     {
7217       if((regs[i].was32>>us2[i])&1) r32|=1LL<<us2[i];
7218     }
7219     if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1))
7220     {
7221       if((regs[i].was32>>dep1[i])&1) r32|=1LL<<dep1[i];
7222     }
7223     if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1))
7224     {
7225       if((regs[i].was32>>dep2[i])&1) r32|=1LL<<dep2[i];
7226     }
7227     //requires_32bit[i]=r32;
7228     pr32[i]=r32;
7229
7230     // Dirty registers which are 32-bit, require 32-bit input
7231     // as they will be written as 32-bit values
7232     for(hr=0;hr<HOST_REGS;hr++)
7233     {
7234       if(regs[i].regmap_entry[hr]>0&&regs[i].regmap_entry[hr]<64) {
7235         if((regs[i].was32>>regs[i].regmap_entry[hr])&(regs[i].wasdirty>>hr)&1) {
7236           if(!((unneeded_reg_upper[i]>>regs[i].regmap_entry[hr])&1))
7237           pr32[i]|=1LL<<regs[i].regmap_entry[hr];
7238           //requires_32bit[i]|=1LL<<regs[i].regmap_entry[hr];
7239         }
7240       }
7241     }
7242   }
7243 }
7244
7245 // Write back dirty registers as soon as we will no longer modify them,
7246 // so that we don't end up with lots of writes at the branches.
7247 void clean_registers(int istart,int iend,int wr)
7248 {
7249   int i;
7250   int r;
7251   u_int will_dirty_i,will_dirty_next,temp_will_dirty;
7252   u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
7253   if(iend==slen-1) {
7254     will_dirty_i=will_dirty_next=0;
7255     wont_dirty_i=wont_dirty_next=0;
7256   }else{
7257     will_dirty_i=will_dirty_next=will_dirty[iend+1];
7258     wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
7259   }
7260   for (i=iend;i>=istart;i--)
7261   {
7262     if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
7263     {
7264       if(ba[i]<start || ba[i]>=(start+slen*4))
7265       {
7266         // Branch out of this block, flush all regs
7267         if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
7268         {
7269           // Unconditional branch
7270           will_dirty_i=0;
7271           wont_dirty_i=0;
7272           // Merge in delay slot (will dirty)
7273           for(r=0;r<HOST_REGS;r++) {
7274             if(r!=EXCLUDE_REG) {
7275               if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7276               if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7277               if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7278               if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7279               if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7280               if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7281               if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7282               if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7283               if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7284               if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7285               if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7286               if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7287               if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7288               if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7289             }
7290           }
7291         }
7292         else
7293         {
7294           // Conditional branch
7295           will_dirty_i=0;
7296           wont_dirty_i=wont_dirty_next;
7297           // Merge in delay slot (will dirty)
7298           for(r=0;r<HOST_REGS;r++) {
7299             if(r!=EXCLUDE_REG) {
7300               if(!likely[i]) {
7301                 // Might not dirty if likely branch is not taken
7302                 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7303                 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7304                 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7305                 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7306                 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7307                 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
7308                 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7309                 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7310                 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7311                 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7312                 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7313                 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7314                 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7315                 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7316               }
7317             }
7318           }
7319         }
7320         // Merge in delay slot (wont dirty)
7321         for(r=0;r<HOST_REGS;r++) {
7322           if(r!=EXCLUDE_REG) {
7323             if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7324             if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7325             if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7326             if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7327             if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7328             if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7329             if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7330             if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7331             if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7332             if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7333           }
7334         }
7335         if(wr) {
7336           #ifndef DESTRUCTIVE_WRITEBACK
7337           branch_regs[i].dirty&=wont_dirty_i;
7338           #endif
7339           branch_regs[i].dirty|=will_dirty_i;
7340         }
7341       }
7342       else
7343       {
7344         // Internal branch
7345         if(ba[i]<=start+i*4) {
7346           // Backward branch
7347           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
7348           {
7349             // Unconditional branch
7350             temp_will_dirty=0;
7351             temp_wont_dirty=0;
7352             // Merge in delay slot (will dirty)
7353             for(r=0;r<HOST_REGS;r++) {
7354               if(r!=EXCLUDE_REG) {
7355                 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7356                 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7357                 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7358                 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7359                 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7360                 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
7361                 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7362                 if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7363                 if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7364                 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7365                 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7366                 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7367                 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
7368                 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7369               }
7370             }
7371           } else {
7372             // Conditional branch (not taken case)
7373             temp_will_dirty=will_dirty_next;
7374             temp_wont_dirty=wont_dirty_next;
7375             // Merge in delay slot (will dirty)
7376             for(r=0;r<HOST_REGS;r++) {
7377               if(r!=EXCLUDE_REG) {
7378                 if(!likely[i]) {
7379                   // Will not dirty if likely branch is not taken
7380                   if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7381                   if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7382                   if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7383                   if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7384                   if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7385                   if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
7386                   if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7387                   //if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7388                   //if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7389                   if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7390                   if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7391                   if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7392                   if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
7393                   if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7394                 }
7395               }
7396             }
7397           }
7398           // Merge in delay slot (wont dirty)
7399           for(r=0;r<HOST_REGS;r++) {
7400             if(r!=EXCLUDE_REG) {
7401               if((regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
7402               if((regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
7403               if((regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
7404               if((regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
7405               if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
7406               if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
7407               if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
7408               if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
7409               if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
7410               if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
7411             }
7412           }
7413           // Deal with changed mappings
7414           if(i<iend) {
7415             for(r=0;r<HOST_REGS;r++) {
7416               if(r!=EXCLUDE_REG) {
7417                 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
7418                   temp_will_dirty&=~(1<<r);
7419                   temp_wont_dirty&=~(1<<r);
7420                   if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
7421                     temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7422                     temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7423                   } else {
7424                     temp_will_dirty|=1<<r;
7425                     temp_wont_dirty|=1<<r;
7426                   }
7427                 }
7428               }
7429             }
7430           }
7431           if(wr) {
7432             will_dirty[i]=temp_will_dirty;
7433             wont_dirty[i]=temp_wont_dirty;
7434             clean_registers((ba[i]-start)>>2,i-1,0);
7435           }else{
7436             // Limit recursion.  It can take an excessive amount
7437             // of time if there are a lot of nested loops.
7438             will_dirty[(ba[i]-start)>>2]=0;
7439             wont_dirty[(ba[i]-start)>>2]=-1;
7440           }
7441         }
7442         /*else*/ if(1)
7443         {
7444           if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
7445           {
7446             // Unconditional branch
7447             will_dirty_i=0;
7448             wont_dirty_i=0;
7449           //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
7450             for(r=0;r<HOST_REGS;r++) {
7451               if(r!=EXCLUDE_REG) {
7452                 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
7453                   will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
7454                   wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
7455                 }
7456                 if(branch_regs[i].regmap[r]>=0) {
7457                   will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
7458                   wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
7459                 }
7460               }
7461             }
7462           //}
7463             // Merge in delay slot
7464             for(r=0;r<HOST_REGS;r++) {
7465               if(r!=EXCLUDE_REG) {
7466                 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7467                 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7468                 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7469                 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7470                 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7471                 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7472                 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7473                 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7474                 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7475                 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7476                 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7477                 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7478                 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7479                 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7480               }
7481             }
7482           } else {
7483             // Conditional branch
7484             will_dirty_i=will_dirty_next;
7485             wont_dirty_i=wont_dirty_next;
7486           //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
7487             for(r=0;r<HOST_REGS;r++) {
7488               if(r!=EXCLUDE_REG) {
7489                 signed char target_reg=branch_regs[i].regmap[r];
7490                 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
7491                   will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
7492                   wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
7493                 }
7494                 else if(target_reg>=0) {
7495                   will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
7496                   wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
7497                 }
7498                 // Treat delay slot as part of branch too
7499                 /*if(regs[i+1].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
7500                   will_dirty[i+1]&=will_dirty[(ba[i]-start)>>2]&(1<<r);
7501                   wont_dirty[i+1]|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
7502                 }
7503                 else
7504                 {
7505                   will_dirty[i+1]&=~(1<<r);
7506                 }*/
7507               }
7508             }
7509           //}
7510             // Merge in delay slot
7511             for(r=0;r<HOST_REGS;r++) {
7512               if(r!=EXCLUDE_REG) {
7513                 if(!likely[i]) {
7514                   // Might not dirty if likely branch is not taken
7515                   if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7516                   if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7517                   if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7518                   if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7519                   if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7520                   if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7521                   if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7522                   //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7523                   //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7524                   if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7525                   if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7526                   if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7527                   if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7528                   if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7529                 }
7530               }
7531             }
7532           }
7533           // Merge in delay slot (won't dirty)
7534           for(r=0;r<HOST_REGS;r++) {
7535             if(r!=EXCLUDE_REG) {
7536               if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7537               if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7538               if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7539               if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7540               if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7541               if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7542               if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7543               if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7544               if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7545               if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7546             }
7547           }
7548           if(wr) {
7549             #ifndef DESTRUCTIVE_WRITEBACK
7550             branch_regs[i].dirty&=wont_dirty_i;
7551             #endif
7552             branch_regs[i].dirty|=will_dirty_i;
7553           }
7554         }
7555       }
7556     }
7557     else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
7558     {
7559       // SYSCALL instruction (software interrupt)
7560       will_dirty_i=0;
7561       wont_dirty_i=0;
7562     }
7563     else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
7564     {
7565       // ERET instruction (return from interrupt)
7566       will_dirty_i=0;
7567       wont_dirty_i=0;
7568     }
7569     will_dirty_next=will_dirty_i;
7570     wont_dirty_next=wont_dirty_i;
7571     for(r=0;r<HOST_REGS;r++) {
7572       if(r!=EXCLUDE_REG) {
7573         if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7574         if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7575         if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7576         if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7577         if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7578         if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7579         if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7580         if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7581         if(i>istart) {
7582           if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=FJUMP)
7583           {
7584             // Don't store a register immediately after writing it,
7585             // may prevent dual-issue.
7586             if((regs[i].regmap[r]&63)==rt1[i-1]) wont_dirty_i|=1<<r;
7587             if((regs[i].regmap[r]&63)==rt2[i-1]) wont_dirty_i|=1<<r;
7588           }
7589         }
7590       }
7591     }
7592     // Save it
7593     will_dirty[i]=will_dirty_i;
7594     wont_dirty[i]=wont_dirty_i;
7595     // Mark registers that won't be dirtied as not dirty
7596     if(wr) {
7597       /*printf("wr (%d,%d) %x will:",istart,iend,start+i*4);
7598       for(r=0;r<HOST_REGS;r++) {
7599         if((will_dirty_i>>r)&1) {
7600           printf(" r%d",r);
7601         }
7602       }
7603       printf("\n");*/
7604
7605       //if(i==istart||(itype[i-1]!=RJUMP&&itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=FJUMP)) {
7606         regs[i].dirty|=will_dirty_i;
7607         #ifndef DESTRUCTIVE_WRITEBACK
7608         regs[i].dirty&=wont_dirty_i;
7609         if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
7610         {
7611           if(i<iend-1&&itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
7612             for(r=0;r<HOST_REGS;r++) {
7613               if(r!=EXCLUDE_REG) {
7614                 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
7615                   regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
7616                 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);/*assert(!((wont_dirty_i>>r)&1));*/}
7617               }
7618             }
7619           }
7620         }
7621         else
7622         {
7623           if(i<iend) {
7624             for(r=0;r<HOST_REGS;r++) {
7625               if(r!=EXCLUDE_REG) {
7626                 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
7627                   regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
7628                 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);/*assert(!((wont_dirty_i>>r)&1));*/}
7629               }
7630             }
7631           }
7632         }
7633         #endif
7634       //}
7635     }
7636     // Deal with changed mappings
7637     temp_will_dirty=will_dirty_i;
7638     temp_wont_dirty=wont_dirty_i;
7639     for(r=0;r<HOST_REGS;r++) {
7640       if(r!=EXCLUDE_REG) {
7641         int nr;
7642         if(regs[i].regmap[r]==regmap_pre[i][r]) {
7643           if(wr) {
7644             #ifndef DESTRUCTIVE_WRITEBACK
7645             regs[i].wasdirty&=wont_dirty_i|~(1<<r);
7646             #endif
7647             regs[i].wasdirty|=will_dirty_i&(1<<r);
7648           }
7649         }
7650         else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
7651           // Register moved to a different register
7652           will_dirty_i&=~(1<<r);
7653           wont_dirty_i&=~(1<<r);
7654           will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
7655           wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
7656           if(wr) {
7657             #ifndef DESTRUCTIVE_WRITEBACK
7658             regs[i].wasdirty&=wont_dirty_i|~(1<<r);
7659             #endif
7660             regs[i].wasdirty|=will_dirty_i&(1<<r);
7661           }
7662         }
7663         else {
7664           will_dirty_i&=~(1<<r);
7665           wont_dirty_i&=~(1<<r);
7666           if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
7667             will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7668             wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7669           } else {
7670             wont_dirty_i|=1<<r;
7671             /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);/*assert(!((will_dirty>>r)&1));*/
7672           }
7673         }
7674       }
7675     }
7676   }
7677 }
7678
7679   /* disassembly */
7680 void disassemble_inst(int i)
7681 {
7682     if (bt[i]) printf("*"); else printf(" ");
7683     switch(itype[i]) {
7684       case UJUMP:
7685         printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
7686       case CJUMP:
7687         printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):*ba);break;
7688       case SJUMP:
7689         printf (" %x: %s r%d,%8x\n",start+i*4,insn[i],rs1[i],start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14));break;
7690       case FJUMP:
7691         printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
7692       case RJUMP:
7693         if (opcode[i]==0x9&&rt1[i]!=31)
7694           printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i]);
7695         else
7696           printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
7697         break;
7698       case SPAN:
7699         printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],ba[i]);break;
7700       case IMM16:
7701         if(opcode[i]==0xf) //LUI
7702           printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],rt1[i],imm[i]&0xffff);
7703         else
7704           printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
7705         break;
7706       case LOAD:
7707       case LOADLR:
7708         printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
7709         break;
7710       case STORE:
7711       case STORELR:
7712         printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rs2[i],rs1[i],imm[i]);
7713         break;
7714       case ALU:
7715       case SHIFT:
7716         printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i],rs2[i]);
7717         break;
7718       case MULTDIV:
7719         printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rs1[i],rs2[i]);
7720         break;
7721       case SHIFTIMM:
7722         printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
7723         break;
7724       case MOV:
7725         if((opcode2[i]&0x1d)==0x10)
7726           printf (" %x: %s r%d\n",start+i*4,insn[i],rt1[i]);
7727         else if((opcode2[i]&0x1d)==0x11)
7728           printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
7729         else
7730           printf (" %x: %s\n",start+i*4,insn[i]);
7731         break;
7732       case COP0:
7733         if(opcode2[i]==0)
7734           printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC0
7735         else if(opcode2[i]==4)
7736           printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC0
7737         else printf (" %x: %s\n",start+i*4,insn[i]);
7738         break;
7739       case COP1:
7740         if(opcode2[i]<3)
7741           printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC1
7742         else if(opcode2[i]>3)
7743           printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC1
7744         else printf (" %x: %s\n",start+i*4,insn[i]);
7745         break;
7746       case COP2:
7747         if(opcode2[i]<3)
7748           printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC2
7749         else if(opcode2[i]>3)
7750           printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC2
7751         else printf (" %x: %s\n",start+i*4,insn[i]);
7752         break;
7753       case C1LS:
7754         printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
7755         break;
7756       case C2LS:
7757         printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
7758         break;
7759       case INTCALL:
7760         printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
7761         break;
7762       default:
7763         //printf (" %s %8x\n",insn[i],source[i]);
7764         printf (" %x: %s\n",start+i*4,insn[i]);
7765     }
7766 }
7767
7768 // clear the state completely, instead of just marking
7769 // things invalid like invalidate_all_pages() does
7770 void new_dynarec_clear_full()
7771 {
7772   int n;
7773   out=(u_char *)BASE_ADDR;
7774   memset(invalid_code,1,sizeof(invalid_code));
7775   memset(hash_table,0xff,sizeof(hash_table));
7776   memset(mini_ht,-1,sizeof(mini_ht));
7777   memset(restore_candidate,0,sizeof(restore_candidate));
7778   memset(shadow,0,sizeof(shadow));
7779   copy=shadow;
7780   expirep=16384; // Expiry pointer, +2 blocks
7781   pending_exception=0;
7782   literalcount=0;
7783   stop_after_jal=0;
7784   // TLB
7785 #ifndef DISABLE_TLB
7786   using_tlb=0;
7787 #endif
7788   sp_in_mirror=0;
7789   for(n=0;n<524288;n++) // 0 .. 0x7FFFFFFF
7790     memory_map[n]=-1;
7791   for(n=524288;n<526336;n++) // 0x80000000 .. 0x807FFFFF
7792     memory_map[n]=((u_int)rdram-0x80000000)>>2;
7793   for(n=526336;n<1048576;n++) // 0x80800000 .. 0xFFFFFFFF
7794     memory_map[n]=-1;
7795   for(n=0;n<4096;n++) ll_clear(jump_in+n);
7796   for(n=0;n<4096;n++) ll_clear(jump_out+n);
7797   for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
7798 }
7799
7800 void new_dynarec_init()
7801 {
7802   printf("Init new dynarec\n");
7803   out=(u_char *)BASE_ADDR;
7804   if (mmap (out, 1<<TARGET_SIZE_2,
7805             PROT_READ | PROT_WRITE | PROT_EXEC,
7806             MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
7807             -1, 0) <= 0) {printf("mmap() failed\n");}
7808 #ifdef MUPEN64
7809   rdword=&readmem_dword;
7810   fake_pc.f.r.rs=&readmem_dword;
7811   fake_pc.f.r.rt=&readmem_dword;
7812   fake_pc.f.r.rd=&readmem_dword;
7813 #endif
7814   int n;
7815   new_dynarec_clear_full();
7816 #ifdef HOST_IMM8
7817   // Copy this into local area so we don't have to put it in every literal pool
7818   invc_ptr=invalid_code;
7819 #endif
7820 #ifdef MUPEN64
7821   for(n=0;n<0x8000;n++) { // 0 .. 0x7FFFFFFF
7822     writemem[n] = write_nomem_new;
7823     writememb[n] = write_nomemb_new;
7824     writememh[n] = write_nomemh_new;
7825 #ifndef FORCE32
7826     writememd[n] = write_nomemd_new;
7827 #endif
7828     readmem[n] = read_nomem_new;
7829     readmemb[n] = read_nomemb_new;
7830     readmemh[n] = read_nomemh_new;
7831 #ifndef FORCE32
7832     readmemd[n] = read_nomemd_new;
7833 #endif
7834   }
7835   for(n=0x8000;n<0x8080;n++) { // 0x80000000 .. 0x807FFFFF
7836     writemem[n] = write_rdram_new;
7837     writememb[n] = write_rdramb_new;
7838     writememh[n] = write_rdramh_new;
7839 #ifndef FORCE32
7840     writememd[n] = write_rdramd_new;
7841 #endif
7842   }
7843   for(n=0xC000;n<0x10000;n++) { // 0xC0000000 .. 0xFFFFFFFF
7844     writemem[n] = write_nomem_new;
7845     writememb[n] = write_nomemb_new;
7846     writememh[n] = write_nomemh_new;
7847 #ifndef FORCE32
7848     writememd[n] = write_nomemd_new;
7849 #endif
7850     readmem[n] = read_nomem_new;
7851     readmemb[n] = read_nomemb_new;
7852     readmemh[n] = read_nomemh_new;
7853 #ifndef FORCE32
7854     readmemd[n] = read_nomemd_new;
7855 #endif
7856   }
7857 #endif
7858   tlb_hacks();
7859   arch_init();
7860 }
7861
7862 void new_dynarec_cleanup()
7863 {
7864   int n;
7865   if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0) {printf("munmap() failed\n");}
7866   for(n=0;n<4096;n++) ll_clear(jump_in+n);
7867   for(n=0;n<4096;n++) ll_clear(jump_out+n);
7868   for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
7869   #ifdef ROM_COPY
7870   if (munmap (ROM_COPY, 67108864) < 0) {printf("munmap() failed\n");}
7871   #endif
7872 }
7873
7874 int new_recompile_block(int addr)
7875 {
7876 /*
7877   if(addr==0x800cd050) {
7878     int block;
7879     for(block=0x80000;block<0x80800;block++) invalidate_block(block);
7880     int n;
7881     for(n=0;n<=2048;n++) ll_clear(jump_dirty+n);
7882   }
7883 */
7884   //if(Count==365117028) tracedebug=1;
7885   assem_debug("NOTCOMPILED: addr = %x -> %x\n", (int)addr, (int)out);
7886   //printf("NOTCOMPILED: addr = %x -> %x\n", (int)addr, (int)out);
7887   //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
7888   //if(debug)
7889   //printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
7890   //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
7891   /*if(Count>=312978186) {
7892     rlist();
7893   }*/
7894   //rlist();
7895   start = (u_int)addr&~3;
7896   //assert(((u_int)addr&1)==0);
7897 #ifdef PCSX
7898   if(!sp_in_mirror&&(signed int)(psxRegs.GPR.n.sp&0xffe00000)>0x80200000&&
7899      0x10000<=psxRegs.GPR.n.sp&&(psxRegs.GPR.n.sp&~0xe0e00000)<RAM_SIZE) {
7900     printf("SP hack enabled (%08x), @%08x\n", psxRegs.GPR.n.sp, psxRegs.pc);
7901     sp_in_mirror=1;
7902   }
7903   if (Config.HLE && start == 0x80001000) // hlecall
7904   {
7905     // XXX: is this enough? Maybe check hleSoftCall?
7906     u_int beginning=(u_int)out;
7907     u_int page=get_page(start);
7908     invalid_code[start>>12]=0;
7909     emit_movimm(start,0);
7910     emit_writeword(0,(int)&pcaddr);
7911     emit_jmp((int)new_dyna_leave);
7912 #ifdef __arm__
7913     __clear_cache((void *)beginning,out);
7914 #endif
7915     ll_add(jump_in+page,start,(void *)beginning);
7916     return 0;
7917   }
7918   else if ((u_int)addr < 0x00200000 ||
7919     (0xa0000000 <= addr && addr < 0xa0200000)) {
7920     // used for BIOS calls mostly?
7921     source = (u_int *)((u_int)rdram+(start&0x1fffff));
7922     pagelimit = (addr&0xa0000000)|0x00200000;
7923   }
7924   else if (!Config.HLE && (
7925 /*    (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
7926     (0xbfc00000 <= addr && addr < 0xbfc80000))) {
7927     // BIOS
7928     source = (u_int *)((u_int)psxR+(start&0x7ffff));
7929     pagelimit = (addr&0xfff00000)|0x80000;
7930   }
7931   else
7932 #endif
7933 #ifdef MUPEN64
7934   if ((int)addr >= 0xa4000000 && (int)addr < 0xa4001000) {
7935     source = (u_int *)((u_int)SP_DMEM+start-0xa4000000);
7936     pagelimit = 0xa4001000;
7937   }
7938   else
7939 #endif
7940   if ((int)addr >= 0x80000000 && (int)addr < 0x80000000+RAM_SIZE) {
7941     source = (u_int *)((u_int)rdram+start-0x80000000);
7942     pagelimit = 0x80000000+RAM_SIZE;
7943   }
7944 #ifndef DISABLE_TLB
7945   else if ((signed int)addr >= (signed int)0xC0000000) {
7946     //printf("addr=%x mm=%x\n",(u_int)addr,(memory_map[start>>12]<<2));
7947     //if(tlb_LUT_r[start>>12])
7948       //source = (u_int *)(((int)rdram)+(tlb_LUT_r[start>>12]&0xFFFFF000)+(((int)addr)&0xFFF)-0x80000000);
7949     if((signed int)memory_map[start>>12]>=0) {
7950       source = (u_int *)((u_int)(start+(memory_map[start>>12]<<2)));
7951       pagelimit=(start+4096)&0xFFFFF000;
7952       int map=memory_map[start>>12];
7953       int i;
7954       for(i=0;i<5;i++) {
7955         //printf("start: %x next: %x\n",map,memory_map[pagelimit>>12]);
7956         if((map&0xBFFFFFFF)==(memory_map[pagelimit>>12]&0xBFFFFFFF)) pagelimit+=4096;
7957       }
7958       assem_debug("pagelimit=%x\n",pagelimit);
7959       assem_debug("mapping=%x (%x)\n",memory_map[start>>12],(memory_map[start>>12]<<2)+start);
7960     }
7961     else {
7962       assem_debug("Compile at unmapped memory address: %x \n", (int)addr);
7963       //assem_debug("start: %x next: %x\n",memory_map[start>>12],memory_map[(start+4096)>>12]);
7964       return -1; // Caller will invoke exception handler
7965     }
7966     //printf("source= %x\n",(int)source);
7967   }
7968 #endif
7969   else {
7970     printf("Compile at bogus memory address: %x \n", (int)addr);
7971     exit(1);
7972   }
7973
7974   /* Pass 1: disassemble */
7975   /* Pass 2: register dependencies, branch targets */
7976   /* Pass 3: register allocation */
7977   /* Pass 4: branch dependencies */
7978   /* Pass 5: pre-alloc */
7979   /* Pass 6: optimize clean/dirty state */
7980   /* Pass 7: flag 32-bit registers */
7981   /* Pass 8: assembly */
7982   /* Pass 9: linker */
7983   /* Pass 10: garbage collection / free memory */
7984
7985   int i,j;
7986   int done=0;
7987   unsigned int type,op,op2;
7988
7989   //printf("addr = %x source = %x %x\n", addr,source,source[0]);
7990
7991   /* Pass 1 disassembly */
7992
7993   for(i=0;!done;i++) {
7994     bt[i]=0;likely[i]=0;ooo[i]=0;op2=0;
7995     minimum_free_regs[i]=0;
7996     opcode[i]=op=source[i]>>26;
7997     switch(op)
7998     {
7999       case 0x00: strcpy(insn[i],"special"); type=NI;
8000         op2=source[i]&0x3f;
8001         switch(op2)
8002         {
8003           case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
8004           case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
8005           case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
8006           case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
8007           case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
8008           case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
8009           case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
8010           case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
8011           case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
8012           case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
8013           case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
8014           case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
8015           case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
8016           case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
8017           case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
8018           case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
8019           case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
8020           case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
8021           case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
8022           case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
8023           case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
8024           case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
8025           case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
8026           case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
8027           case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
8028           case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
8029           case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
8030           case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
8031           case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
8032           case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
8033           case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
8034           case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
8035           case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
8036           case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
8037           case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
8038 #ifndef FORCE32
8039           case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
8040           case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
8041           case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
8042           case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
8043           case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
8044           case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
8045           case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
8046           case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
8047           case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
8048           case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
8049           case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
8050           case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
8051           case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
8052           case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
8053           case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
8054           case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
8055           case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
8056 #endif
8057         }
8058         break;
8059       case 0x01: strcpy(insn[i],"regimm"); type=NI;
8060         op2=(source[i]>>16)&0x1f;
8061         switch(op2)
8062         {
8063           case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
8064           case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
8065           case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
8066           case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
8067           case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
8068           case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
8069           case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
8070           case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
8071           case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
8072           case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
8073           case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
8074           case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
8075           case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
8076           case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
8077         }
8078         break;
8079       case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
8080       case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
8081       case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
8082       case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
8083       case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
8084       case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
8085       case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
8086       case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
8087       case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
8088       case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
8089       case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
8090       case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
8091       case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
8092       case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
8093       case 0x10: strcpy(insn[i],"cop0"); type=NI;
8094         op2=(source[i]>>21)&0x1f;
8095         switch(op2)
8096         {
8097           case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
8098           case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
8099           case 0x10: strcpy(insn[i],"tlb"); type=NI;
8100           switch(source[i]&0x3f)
8101           {
8102             case 0x01: strcpy(insn[i],"TLBR"); type=COP0; break;
8103             case 0x02: strcpy(insn[i],"TLBWI"); type=COP0; break;
8104             case 0x06: strcpy(insn[i],"TLBWR"); type=COP0; break;
8105             case 0x08: strcpy(insn[i],"TLBP"); type=COP0; break;
8106 #ifdef PCSX
8107             case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
8108 #else
8109             case 0x18: strcpy(insn[i],"ERET"); type=COP0; break;
8110 #endif
8111           }
8112         }
8113         break;
8114       case 0x11: strcpy(insn[i],"cop1"); type=NI;
8115         op2=(source[i]>>21)&0x1f;
8116         switch(op2)
8117         {
8118           case 0x00: strcpy(insn[i],"MFC1"); type=COP1; break;
8119           case 0x01: strcpy(insn[i],"DMFC1"); type=COP1; break;
8120           case 0x02: strcpy(insn[i],"CFC1"); type=COP1; break;
8121           case 0x04: strcpy(insn[i],"MTC1"); type=COP1; break;
8122           case 0x05: strcpy(insn[i],"DMTC1"); type=COP1; break;
8123           case 0x06: strcpy(insn[i],"CTC1"); type=COP1; break;
8124           case 0x08: strcpy(insn[i],"BC1"); type=FJUMP;
8125           switch((source[i]>>16)&0x3)
8126           {
8127             case 0x00: strcpy(insn[i],"BC1F"); break;
8128             case 0x01: strcpy(insn[i],"BC1T"); break;
8129             case 0x02: strcpy(insn[i],"BC1FL"); break;
8130             case 0x03: strcpy(insn[i],"BC1TL"); break;
8131           }
8132           break;
8133           case 0x10: strcpy(insn[i],"C1.S"); type=NI;
8134           switch(source[i]&0x3f)
8135           {
8136             case 0x00: strcpy(insn[i],"ADD.S"); type=FLOAT; break;
8137             case 0x01: strcpy(insn[i],"SUB.S"); type=FLOAT; break;
8138             case 0x02: strcpy(insn[i],"MUL.S"); type=FLOAT; break;
8139             case 0x03: strcpy(insn[i],"DIV.S"); type=FLOAT; break;
8140             case 0x04: strcpy(insn[i],"SQRT.S"); type=FLOAT; break;
8141             case 0x05: strcpy(insn[i],"ABS.S"); type=FLOAT; break;
8142             case 0x06: strcpy(insn[i],"MOV.S"); type=FLOAT; break;
8143             case 0x07: strcpy(insn[i],"NEG.S"); type=FLOAT; break;
8144             case 0x08: strcpy(insn[i],"ROUND.L.S"); type=FCONV; break;
8145             case 0x09: strcpy(insn[i],"TRUNC.L.S"); type=FCONV; break;
8146             case 0x0A: strcpy(insn[i],"CEIL.L.S"); type=FCONV; break;
8147             case 0x0B: strcpy(insn[i],"FLOOR.L.S"); type=FCONV; break;
8148             case 0x0C: strcpy(insn[i],"ROUND.W.S"); type=FCONV; break;
8149             case 0x0D: strcpy(insn[i],"TRUNC.W.S"); type=FCONV; break;
8150             case 0x0E: strcpy(insn[i],"CEIL.W.S"); type=FCONV; break;
8151             case 0x0F: strcpy(insn[i],"FLOOR.W.S"); type=FCONV; break;
8152             case 0x21: strcpy(insn[i],"CVT.D.S"); type=FCONV; break;
8153             case 0x24: strcpy(insn[i],"CVT.W.S"); type=FCONV; break;
8154             case 0x25: strcpy(insn[i],"CVT.L.S"); type=FCONV; break;
8155             case 0x30: strcpy(insn[i],"C.F.S"); type=FCOMP; break;
8156             case 0x31: strcpy(insn[i],"C.UN.S"); type=FCOMP; break;
8157             case 0x32: strcpy(insn[i],"C.EQ.S"); type=FCOMP; break;
8158             case 0x33: strcpy(insn[i],"C.UEQ.S"); type=FCOMP; break;
8159             case 0x34: strcpy(insn[i],"C.OLT.S"); type=FCOMP; break;
8160             case 0x35: strcpy(insn[i],"C.ULT.S"); type=FCOMP; break;
8161             case 0x36: strcpy(insn[i],"C.OLE.S"); type=FCOMP; break;
8162             case 0x37: strcpy(insn[i],"C.ULE.S"); type=FCOMP; break;
8163             case 0x38: strcpy(insn[i],"C.SF.S"); type=FCOMP; break;
8164             case 0x39: strcpy(insn[i],"C.NGLE.S"); type=FCOMP; break;
8165             case 0x3A: strcpy(insn[i],"C.SEQ.S"); type=FCOMP; break;
8166             case 0x3B: strcpy(insn[i],"C.NGL.S"); type=FCOMP; break;
8167             case 0x3C: strcpy(insn[i],"C.LT.S"); type=FCOMP; break;
8168             case 0x3D: strcpy(insn[i],"C.NGE.S"); type=FCOMP; break;
8169             case 0x3E: strcpy(insn[i],"C.LE.S"); type=FCOMP; break;
8170             case 0x3F: strcpy(insn[i],"C.NGT.S"); type=FCOMP; break;
8171           }
8172           break;
8173           case 0x11: strcpy(insn[i],"C1.D"); type=NI;
8174           switch(source[i]&0x3f)
8175           {
8176             case 0x00: strcpy(insn[i],"ADD.D"); type=FLOAT; break;
8177             case 0x01: strcpy(insn[i],"SUB.D"); type=FLOAT; break;
8178             case 0x02: strcpy(insn[i],"MUL.D"); type=FLOAT; break;
8179             case 0x03: strcpy(insn[i],"DIV.D"); type=FLOAT; break;
8180             case 0x04: strcpy(insn[i],"SQRT.D"); type=FLOAT; break;
8181             case 0x05: strcpy(insn[i],"ABS.D"); type=FLOAT; break;
8182             case 0x06: strcpy(insn[i],"MOV.D"); type=FLOAT; break;
8183             case 0x07: strcpy(insn[i],"NEG.D"); type=FLOAT; break;
8184             case 0x08: strcpy(insn[i],"ROUND.L.D"); type=FCONV; break;
8185             case 0x09: strcpy(insn[i],"TRUNC.L.D"); type=FCONV; break;
8186             case 0x0A: strcpy(insn[i],"CEIL.L.D"); type=FCONV; break;
8187             case 0x0B: strcpy(insn[i],"FLOOR.L.D"); type=FCONV; break;
8188             case 0x0C: strcpy(insn[i],"ROUND.W.D"); type=FCONV; break;
8189             case 0x0D: strcpy(insn[i],"TRUNC.W.D"); type=FCONV; break;
8190             case 0x0E: strcpy(insn[i],"CEIL.W.D"); type=FCONV; break;
8191             case 0x0F: strcpy(insn[i],"FLOOR.W.D"); type=FCONV; break;
8192             case 0x20: strcpy(insn[i],"CVT.S.D"); type=FCONV; break;
8193             case 0x24: strcpy(insn[i],"CVT.W.D"); type=FCONV; break;
8194             case 0x25: strcpy(insn[i],"CVT.L.D"); type=FCONV; break;
8195             case 0x30: strcpy(insn[i],"C.F.D"); type=FCOMP; break;
8196             case 0x31: strcpy(insn[i],"C.UN.D"); type=FCOMP; break;
8197             case 0x32: strcpy(insn[i],"C.EQ.D"); type=FCOMP; break;
8198             case 0x33: strcpy(insn[i],"C.UEQ.D"); type=FCOMP; break;
8199             case 0x34: strcpy(insn[i],"C.OLT.D"); type=FCOMP; break;
8200             case 0x35: strcpy(insn[i],"C.ULT.D"); type=FCOMP; break;
8201             case 0x36: strcpy(insn[i],"C.OLE.D"); type=FCOMP; break;
8202             case 0x37: strcpy(insn[i],"C.ULE.D"); type=FCOMP; break;
8203             case 0x38: strcpy(insn[i],"C.SF.D"); type=FCOMP; break;
8204             case 0x39: strcpy(insn[i],"C.NGLE.D"); type=FCOMP; break;
8205             case 0x3A: strcpy(insn[i],"C.SEQ.D"); type=FCOMP; break;
8206             case 0x3B: strcpy(insn[i],"C.NGL.D"); type=FCOMP; break;
8207             case 0x3C: strcpy(insn[i],"C.LT.D"); type=FCOMP; break;
8208             case 0x3D: strcpy(insn[i],"C.NGE.D"); type=FCOMP; break;
8209             case 0x3E: strcpy(insn[i],"C.LE.D"); type=FCOMP; break;
8210             case 0x3F: strcpy(insn[i],"C.NGT.D"); type=FCOMP; break;
8211           }
8212           break;
8213           case 0x14: strcpy(insn[i],"C1.W"); type=NI;
8214           switch(source[i]&0x3f)
8215           {
8216             case 0x20: strcpy(insn[i],"CVT.S.W"); type=FCONV; break;
8217             case 0x21: strcpy(insn[i],"CVT.D.W"); type=FCONV; break;
8218           }
8219           break;
8220           case 0x15: strcpy(insn[i],"C1.L"); type=NI;
8221           switch(source[i]&0x3f)
8222           {
8223             case 0x20: strcpy(insn[i],"CVT.S.L"); type=FCONV; break;
8224             case 0x21: strcpy(insn[i],"CVT.D.L"); type=FCONV; break;
8225           }
8226           break;
8227         }
8228         break;
8229 #ifndef FORCE32
8230       case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
8231       case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
8232       case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
8233       case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
8234       case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
8235       case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
8236       case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
8237       case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
8238 #endif
8239       case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
8240       case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
8241       case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
8242       case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
8243       case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
8244       case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
8245       case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
8246       case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
8247       case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
8248       case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
8249       case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
8250       case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
8251 #ifndef FORCE32
8252       case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
8253       case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
8254 #endif
8255       case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
8256       case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
8257       case 0x30: strcpy(insn[i],"LL"); type=NI; break;
8258       case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
8259 #ifndef FORCE32
8260       case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
8261       case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
8262       case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
8263 #endif
8264       case 0x38: strcpy(insn[i],"SC"); type=NI; break;
8265       case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
8266 #ifndef FORCE32
8267       case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
8268       case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
8269       case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
8270 #endif
8271 #ifdef PCSX
8272       case 0x12: strcpy(insn[i],"COP2"); type=NI;
8273         // note: COP MIPS-1 encoding differs from MIPS32
8274         op2=(source[i]>>21)&0x1f;
8275         if (source[i]&0x3f) {
8276           if (gte_handlers[source[i]&0x3f]!=NULL) {
8277             snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
8278             type=C2OP;
8279           }
8280         }
8281         else switch(op2)
8282         {
8283           case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
8284           case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
8285           case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
8286           case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
8287         }
8288         break;
8289       case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
8290       case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
8291       case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
8292 #endif
8293       default: strcpy(insn[i],"???"); type=NI;
8294         printf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
8295         break;
8296     }
8297     itype[i]=type;
8298     opcode2[i]=op2;
8299     /* Get registers/immediates */
8300     lt1[i]=0;
8301     us1[i]=0;
8302     us2[i]=0;
8303     dep1[i]=0;
8304     dep2[i]=0;
8305     switch(type) {
8306       case LOAD:
8307         rs1[i]=(source[i]>>21)&0x1f;
8308         rs2[i]=0;
8309         rt1[i]=(source[i]>>16)&0x1f;
8310         rt2[i]=0;
8311         imm[i]=(short)source[i];
8312         break;
8313       case STORE:
8314       case STORELR:
8315         rs1[i]=(source[i]>>21)&0x1f;
8316         rs2[i]=(source[i]>>16)&0x1f;
8317         rt1[i]=0;
8318         rt2[i]=0;
8319         imm[i]=(short)source[i];
8320         if(op==0x2c||op==0x2d||op==0x3f) us1[i]=rs2[i]; // 64-bit SDL/SDR/SD
8321         break;
8322       case LOADLR:
8323         // LWL/LWR only load part of the register,
8324         // therefore the target register must be treated as a source too
8325         rs1[i]=(source[i]>>21)&0x1f;
8326         rs2[i]=(source[i]>>16)&0x1f;
8327         rt1[i]=(source[i]>>16)&0x1f;
8328         rt2[i]=0;
8329         imm[i]=(short)source[i];
8330         if(op==0x1a||op==0x1b) us1[i]=rs2[i]; // LDR/LDL
8331         if(op==0x26) dep1[i]=rt1[i]; // LWR
8332         break;
8333       case IMM16:
8334         if (op==0x0f) rs1[i]=0; // LUI instruction has no source register
8335         else rs1[i]=(source[i]>>21)&0x1f;
8336         rs2[i]=0;
8337         rt1[i]=(source[i]>>16)&0x1f;
8338         rt2[i]=0;
8339         if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
8340           imm[i]=(unsigned short)source[i];
8341         }else{
8342           imm[i]=(short)source[i];
8343         }
8344         if(op==0x18||op==0x19) us1[i]=rs1[i]; // DADDI/DADDIU
8345         if(op==0x0a||op==0x0b) us1[i]=rs1[i]; // SLTI/SLTIU
8346         if(op==0x0d||op==0x0e) dep1[i]=rs1[i]; // ORI/XORI
8347         break;
8348       case UJUMP:
8349         rs1[i]=0;
8350         rs2[i]=0;
8351         rt1[i]=0;
8352         rt2[i]=0;
8353         // The JAL instruction writes to r31.
8354         if (op&1) {
8355           rt1[i]=31;
8356         }
8357         rs2[i]=CCREG;
8358         break;
8359       case RJUMP:
8360         rs1[i]=(source[i]>>21)&0x1f;
8361         rs2[i]=0;
8362         rt1[i]=0;
8363         rt2[i]=0;
8364         // The JALR instruction writes to rd.
8365         if (op2&1) {
8366           rt1[i]=(source[i]>>11)&0x1f;
8367         }
8368         rs2[i]=CCREG;
8369         break;
8370       case CJUMP:
8371         rs1[i]=(source[i]>>21)&0x1f;
8372         rs2[i]=(source[i]>>16)&0x1f;
8373         rt1[i]=0;
8374         rt2[i]=0;
8375         if(op&2) { // BGTZ/BLEZ
8376           rs2[i]=0;
8377         }
8378         us1[i]=rs1[i];
8379         us2[i]=rs2[i];
8380         likely[i]=op>>4;
8381         break;
8382       case SJUMP:
8383         rs1[i]=(source[i]>>21)&0x1f;
8384         rs2[i]=CCREG;
8385         rt1[i]=0;
8386         rt2[i]=0;
8387         us1[i]=rs1[i];
8388         if(op2&0x10) { // BxxAL
8389           rt1[i]=31;
8390           // NOTE: If the branch is not taken, r31 is still overwritten
8391         }
8392         likely[i]=(op2&2)>>1;
8393         break;
8394       case FJUMP:
8395         rs1[i]=FSREG;
8396         rs2[i]=CSREG;
8397         rt1[i]=0;
8398         rt2[i]=0;
8399         likely[i]=((source[i])>>17)&1;
8400         break;
8401       case ALU:
8402         rs1[i]=(source[i]>>21)&0x1f; // source
8403         rs2[i]=(source[i]>>16)&0x1f; // subtract amount
8404         rt1[i]=(source[i]>>11)&0x1f; // destination
8405         rt2[i]=0;
8406         if(op2==0x2a||op2==0x2b) { // SLT/SLTU
8407           us1[i]=rs1[i];us2[i]=rs2[i];
8408         }
8409         else if(op2>=0x24&&op2<=0x27) { // AND/OR/XOR/NOR
8410           dep1[i]=rs1[i];dep2[i]=rs2[i];
8411         }
8412         else if(op2>=0x2c&&op2<=0x2f) { // DADD/DSUB
8413           dep1[i]=rs1[i];dep2[i]=rs2[i];
8414         }
8415         break;
8416       case MULTDIV:
8417         rs1[i]=(source[i]>>21)&0x1f; // source
8418         rs2[i]=(source[i]>>16)&0x1f; // divisor
8419         rt1[i]=HIREG;
8420         rt2[i]=LOREG;
8421         if (op2>=0x1c&&op2<=0x1f) { // DMULT/DMULTU/DDIV/DDIVU
8422           us1[i]=rs1[i];us2[i]=rs2[i];
8423         }
8424         break;
8425       case MOV:
8426         rs1[i]=0;
8427         rs2[i]=0;
8428         rt1[i]=0;
8429         rt2[i]=0;
8430         if(op2==0x10) rs1[i]=HIREG; // MFHI
8431         if(op2==0x11) rt1[i]=HIREG; // MTHI
8432         if(op2==0x12) rs1[i]=LOREG; // MFLO
8433         if(op2==0x13) rt1[i]=LOREG; // MTLO
8434         if((op2&0x1d)==0x10) rt1[i]=(source[i]>>11)&0x1f; // MFxx
8435         if((op2&0x1d)==0x11) rs1[i]=(source[i]>>21)&0x1f; // MTxx
8436         dep1[i]=rs1[i];
8437         break;
8438       case SHIFT:
8439         rs1[i]=(source[i]>>16)&0x1f; // target of shift
8440         rs2[i]=(source[i]>>21)&0x1f; // shift amount
8441         rt1[i]=(source[i]>>11)&0x1f; // destination
8442         rt2[i]=0;
8443         // DSLLV/DSRLV/DSRAV are 64-bit
8444         if(op2>=0x14&&op2<=0x17) us1[i]=rs1[i];
8445         break;
8446       case SHIFTIMM:
8447         rs1[i]=(source[i]>>16)&0x1f;
8448         rs2[i]=0;
8449         rt1[i]=(source[i]>>11)&0x1f;
8450         rt2[i]=0;
8451         imm[i]=(source[i]>>6)&0x1f;
8452         // DSxx32 instructions
8453         if(op2>=0x3c) imm[i]|=0x20;
8454         // DSLL/DSRL/DSRA/DSRA32/DSRL32 but not DSLL32 require 64-bit source
8455         if(op2>=0x38&&op2!=0x3c) us1[i]=rs1[i];
8456         break;
8457       case COP0:
8458         rs1[i]=0;
8459         rs2[i]=0;
8460         rt1[i]=0;
8461         rt2[i]=0;
8462         if(op2==0) rt1[i]=(source[i]>>16)&0x1F; // MFC0
8463         if(op2==4) rs1[i]=(source[i]>>16)&0x1F; // MTC0
8464         if(op2==4&&((source[i]>>11)&0x1f)==12) rt2[i]=CSREG; // Status
8465         if(op2==16) if((source[i]&0x3f)==0x18) rs2[i]=CCREG; // ERET
8466         break;
8467       case COP1:
8468       case COP2:
8469         rs1[i]=0;
8470         rs2[i]=0;
8471         rt1[i]=0;
8472         rt2[i]=0;
8473         if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
8474         if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
8475         if(op2==5) us1[i]=rs1[i]; // DMTC1
8476         rs2[i]=CSREG;
8477         break;
8478       case C1LS:
8479         rs1[i]=(source[i]>>21)&0x1F;
8480         rs2[i]=CSREG;
8481         rt1[i]=0;
8482         rt2[i]=0;
8483         imm[i]=(short)source[i];
8484         break;
8485       case C2LS:
8486         rs1[i]=(source[i]>>21)&0x1F;
8487         rs2[i]=0;
8488         rt1[i]=0;
8489         rt2[i]=0;
8490         imm[i]=(short)source[i];
8491         break;
8492       case FLOAT:
8493       case FCONV:
8494         rs1[i]=0;
8495         rs2[i]=CSREG;
8496         rt1[i]=0;
8497         rt2[i]=0;
8498         break;
8499       case FCOMP:
8500         rs1[i]=FSREG;
8501         rs2[i]=CSREG;
8502         rt1[i]=FSREG;
8503         rt2[i]=0;
8504         break;
8505       case SYSCALL:
8506       case HLECALL:
8507       case INTCALL:
8508         rs1[i]=CCREG;
8509         rs2[i]=0;
8510         rt1[i]=0;
8511         rt2[i]=0;
8512         break;
8513       default:
8514         rs1[i]=0;
8515         rs2[i]=0;
8516         rt1[i]=0;
8517         rt2[i]=0;
8518     }
8519     /* Calculate branch target addresses */
8520     if(type==UJUMP)
8521       ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
8522     else if(type==CJUMP&&rs1[i]==rs2[i]&&(op&1))
8523       ba[i]=start+i*4+8; // Ignore never taken branch
8524     else if(type==SJUMP&&rs1[i]==0&&!(op2&1))
8525       ba[i]=start+i*4+8; // Ignore never taken branch
8526     else if(type==CJUMP||type==SJUMP||type==FJUMP)
8527       ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
8528     else ba[i]=-1;
8529 #ifdef PCSX
8530     if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
8531       int do_in_intrp=0;
8532       // branch in delay slot?
8533       if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
8534         // don't handle first branch and call interpreter if it's hit
8535         printf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
8536         do_in_intrp=1;
8537       }
8538       // basic load delay detection
8539       else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&rt1[i]!=0) {
8540         int t=(ba[i-1]-start)/4;
8541         if(0 <= t && t < i &&(rt1[i]==rs1[t]||rt1[i]==rs2[t])&&itype[t]!=CJUMP&&itype[t]!=SJUMP) {
8542           // jump target wants DS result - potential load delay effect
8543           printf("load delay @%08x (%08x)\n", addr + i*4, addr);
8544           do_in_intrp=1;
8545           bt[t+1]=1; // expected return from interpreter
8546         }
8547         else if(i>=2&&rt1[i-2]==2&&rt1[i]==2&&rs1[i]!=2&&rs2[i]!=2&&rs1[i-1]!=2&&rs2[i-1]!=2&&
8548               !(i>=3&&(itype[i-3]==RJUMP||itype[i-3]==UJUMP||itype[i-3]==CJUMP||itype[i-3]==SJUMP))) {
8549           // v0 overwrite like this is a sign of trouble, bail out
8550           printf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
8551           do_in_intrp=1;
8552         }
8553       }
8554       if(do_in_intrp) {
8555         rs1[i-1]=CCREG;
8556         rs2[i-1]=rt1[i-1]=rt2[i-1]=0;
8557         ba[i-1]=-1;
8558         itype[i-1]=INTCALL;
8559         done=2;
8560         i--; // don't compile the DS
8561       }
8562     }
8563 #endif
8564     /* Is this the end of the block? */
8565     if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
8566       if(rt1[i-1]==0) { // Continue past subroutine call (JAL)
8567         done=2;
8568       }
8569       else {
8570         if(stop_after_jal) done=1;
8571         // Stop on BREAK
8572         if((source[i+1]&0xfc00003f)==0x0d) done=1;
8573       }
8574       // Don't recompile stuff that's already compiled
8575       if(check_addr(start+i*4+4)) done=1;
8576       // Don't get too close to the limit
8577       if(i>MAXBLOCK/2) done=1;
8578     }
8579     if(itype[i]==SYSCALL&&stop_after_jal) done=1;
8580     if(itype[i]==HLECALL||itype[i]==INTCALL) done=2;
8581     if(done==2) {
8582       // Does the block continue due to a branch?
8583       for(j=i-1;j>=0;j--)
8584       {
8585         if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
8586         if(ba[j]==start+i*4+4) done=j=0;
8587         if(ba[j]==start+i*4+8) done=j=0;
8588       }
8589     }
8590     //assert(i<MAXBLOCK-1);
8591     if(start+i*4==pagelimit-4) done=1;
8592     assert(start+i*4<pagelimit);
8593     if (i==MAXBLOCK-1) done=1;
8594     // Stop if we're compiling junk
8595     if(itype[i]==NI&&opcode[i]==0x11) {
8596       done=stop_after_jal=1;
8597       printf("Disabled speculative precompilation\n");
8598     }
8599   }
8600   slen=i;
8601   if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==RJUMP||itype[i-1]==FJUMP) {
8602     if(start+i*4==pagelimit) {
8603       itype[i-1]=SPAN;
8604     }
8605   }
8606   assert(slen>0);
8607
8608   /* Pass 2 - Register dependencies and branch targets */
8609
8610   unneeded_registers(0,slen-1,0);
8611
8612   /* Pass 3 - Register allocation */
8613
8614   struct regstat current; // Current register allocations/status
8615   current.is32=1;
8616   current.dirty=0;
8617   current.u=unneeded_reg[0];
8618   current.uu=unneeded_reg_upper[0];
8619   clear_all_regs(current.regmap);
8620   alloc_reg(&current,0,CCREG);
8621   dirty_reg(&current,CCREG);
8622   current.isconst=0;
8623   current.wasconst=0;
8624   int ds=0;
8625   int cc=0;
8626   int hr=-1;
8627
8628 #ifndef FORCE32
8629   provisional_32bit();
8630 #endif
8631   if((u_int)addr&1) {
8632     // First instruction is delay slot
8633     cc=-1;
8634     bt[1]=1;
8635     ds=1;
8636     unneeded_reg[0]=1;
8637     unneeded_reg_upper[0]=1;
8638     current.regmap[HOST_BTREG]=BTREG;
8639   }
8640
8641   for(i=0;i<slen;i++)
8642   {
8643     if(bt[i])
8644     {
8645       int hr;
8646       for(hr=0;hr<HOST_REGS;hr++)
8647       {
8648         // Is this really necessary?
8649         if(current.regmap[hr]==0) current.regmap[hr]=-1;
8650       }
8651       current.isconst=0;
8652     }
8653     if(i>1)
8654     {
8655       if((opcode[i-2]&0x2f)==0x05) // BNE/BNEL
8656       {
8657         if(rs1[i-2]==0||rs2[i-2]==0)
8658         {
8659           if(rs1[i-2]) {
8660             current.is32|=1LL<<rs1[i-2];
8661             int hr=get_reg(current.regmap,rs1[i-2]|64);
8662             if(hr>=0) current.regmap[hr]=-1;
8663           }
8664           if(rs2[i-2]) {
8665             current.is32|=1LL<<rs2[i-2];
8666             int hr=get_reg(current.regmap,rs2[i-2]|64);
8667             if(hr>=0) current.regmap[hr]=-1;
8668           }
8669         }
8670       }
8671     }
8672 #ifndef FORCE32
8673     // If something jumps here with 64-bit values
8674     // then promote those registers to 64 bits
8675     if(bt[i])
8676     {
8677       uint64_t temp_is32=current.is32;
8678       for(j=i-1;j>=0;j--)
8679       {
8680         if(ba[j]==start+i*4)
8681           temp_is32&=branch_regs[j].is32;
8682       }
8683       for(j=i;j<slen;j++)
8684       {
8685         if(ba[j]==start+i*4)
8686           //temp_is32=1;
8687           temp_is32&=p32[j];
8688       }
8689       if(temp_is32!=current.is32) {
8690         //printf("dumping 32-bit regs (%x)\n",start+i*4);
8691         #ifdef DESTRUCTIVE_WRITEBACK
8692         for(hr=0;hr<HOST_REGS;hr++)
8693         {
8694           int r=current.regmap[hr];
8695           if(r>0&&r<64)
8696           {
8697             if((current.dirty>>hr)&((current.is32&~temp_is32)>>r)&1) {
8698               temp_is32|=1LL<<r;
8699               //printf("restore %d\n",r);
8700             }
8701           }
8702         }
8703         #endif
8704         current.is32=temp_is32;
8705       }
8706     }
8707 #else
8708     current.is32=-1LL;
8709 #endif
8710
8711     memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
8712     regs[i].wasconst=current.isconst;
8713     regs[i].was32=current.is32;
8714     regs[i].wasdirty=current.dirty;
8715     #if defined(DESTRUCTIVE_WRITEBACK) && !defined(FORCE32)
8716     // To change a dirty register from 32 to 64 bits, we must write
8717     // it out during the previous cycle (for branches, 2 cycles)
8718     if(i<slen-1&&bt[i+1]&&itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP)
8719     {
8720       uint64_t temp_is32=current.is32;
8721       for(j=i-1;j>=0;j--)
8722       {
8723         if(ba[j]==start+i*4+4)
8724           temp_is32&=branch_regs[j].is32;
8725       }
8726       for(j=i;j<slen;j++)
8727       {
8728         if(ba[j]==start+i*4+4)
8729           //temp_is32=1;
8730           temp_is32&=p32[j];
8731       }
8732       if(temp_is32!=current.is32) {
8733         //printf("pre-dumping 32-bit regs (%x)\n",start+i*4);
8734         for(hr=0;hr<HOST_REGS;hr++)
8735         {
8736           int r=current.regmap[hr];
8737           if(r>0)
8738           {
8739             if((current.dirty>>hr)&((current.is32&~temp_is32)>>(r&63))&1) {
8740               if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP)
8741               {
8742                 if(rs1[i]!=(r&63)&&rs2[i]!=(r&63))
8743                 {
8744                   //printf("dump %d/r%d\n",hr,r);
8745                   current.regmap[hr]=-1;
8746                   if(get_reg(current.regmap,r|64)>=0)
8747                     current.regmap[get_reg(current.regmap,r|64)]=-1;
8748                 }
8749               }
8750             }
8751           }
8752         }
8753       }
8754     }
8755     else if(i<slen-2&&bt[i+2]&&(source[i-1]>>16)!=0x1000&&(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP))
8756     {
8757       uint64_t temp_is32=current.is32;
8758       for(j=i-1;j>=0;j--)
8759       {
8760         if(ba[j]==start+i*4+8)
8761           temp_is32&=branch_regs[j].is32;
8762       }
8763       for(j=i;j<slen;j++)
8764       {
8765         if(ba[j]==start+i*4+8)
8766           //temp_is32=1;
8767           temp_is32&=p32[j];
8768       }
8769       if(temp_is32!=current.is32) {
8770         //printf("pre-dumping 32-bit regs (%x)\n",start+i*4);
8771         for(hr=0;hr<HOST_REGS;hr++)
8772         {
8773           int r=current.regmap[hr];
8774           if(r>0)
8775           {
8776             if((current.dirty>>hr)&((current.is32&~temp_is32)>>(r&63))&1) {
8777               if(rs1[i]!=(r&63)&&rs2[i]!=(r&63)&&rs1[i+1]!=(r&63)&&rs2[i+1]!=(r&63))
8778               {
8779                 //printf("dump %d/r%d\n",hr,r);
8780                 current.regmap[hr]=-1;
8781                 if(get_reg(current.regmap,r|64)>=0)
8782                   current.regmap[get_reg(current.regmap,r|64)]=-1;
8783               }
8784             }
8785           }
8786         }
8787       }
8788     }
8789     #endif
8790     if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP) {
8791       if(i+1<slen) {
8792         current.u=unneeded_reg[i+1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
8793         current.uu=unneeded_reg_upper[i+1]&~((1LL<<us1[i])|(1LL<<us2[i]));
8794         if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8795         current.u|=1;
8796         current.uu|=1;
8797       } else {
8798         current.u=1;
8799         current.uu=1;
8800       }
8801     } else {
8802       if(i+1<slen) {
8803         current.u=branch_unneeded_reg[i]&~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
8804         current.uu=branch_unneeded_reg_upper[i]&~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
8805         if((~current.uu>>rt1[i+1])&1) current.uu&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
8806         current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
8807         current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8808         current.u|=1;
8809         current.uu|=1;
8810       } else { printf("oops, branch at end of block with no delay slot\n");exit(1); }
8811     }
8812     is_ds[i]=ds;
8813     if(ds) {
8814       ds=0; // Skip delay slot, already allocated as part of branch
8815       // ...but we need to alloc it in case something jumps here
8816       if(i+1<slen) {
8817         current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
8818         current.uu=branch_unneeded_reg_upper[i-1]&unneeded_reg_upper[i+1];
8819       }else{
8820         current.u=branch_unneeded_reg[i-1];
8821         current.uu=branch_unneeded_reg_upper[i-1];
8822       }
8823       current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
8824       current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8825       if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8826       current.u|=1;
8827       current.uu|=1;
8828       struct regstat temp;
8829       memcpy(&temp,&current,sizeof(current));
8830       temp.wasdirty=temp.dirty;
8831       temp.was32=temp.is32;
8832       // TODO: Take into account unconditional branches, as below
8833       delayslot_alloc(&temp,i);
8834       memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
8835       regs[i].wasdirty=temp.wasdirty;
8836       regs[i].was32=temp.was32;
8837       regs[i].dirty=temp.dirty;
8838       regs[i].is32=temp.is32;
8839       regs[i].isconst=0;
8840       regs[i].wasconst=0;
8841       current.isconst=0;
8842       // Create entry (branch target) regmap
8843       for(hr=0;hr<HOST_REGS;hr++)
8844       {
8845         int r=temp.regmap[hr];
8846         if(r>=0) {
8847           if(r!=regmap_pre[i][hr]) {
8848             regs[i].regmap_entry[hr]=-1;
8849           }
8850           else
8851           {
8852             if(r<64){
8853               if((current.u>>r)&1) {
8854                 regs[i].regmap_entry[hr]=-1;
8855                 regs[i].regmap[hr]=-1;
8856                 //Don't clear regs in the delay slot as the branch might need them
8857                 //current.regmap[hr]=-1;
8858               }else
8859                 regs[i].regmap_entry[hr]=r;
8860             }
8861             else {
8862               if((current.uu>>(r&63))&1) {
8863                 regs[i].regmap_entry[hr]=-1;
8864                 regs[i].regmap[hr]=-1;
8865                 //Don't clear regs in the delay slot as the branch might need them
8866                 //current.regmap[hr]=-1;
8867               }else
8868                 regs[i].regmap_entry[hr]=r;
8869             }
8870           }
8871         } else {
8872           // First instruction expects CCREG to be allocated
8873           if(i==0&&hr==HOST_CCREG)
8874             regs[i].regmap_entry[hr]=CCREG;
8875           else
8876             regs[i].regmap_entry[hr]=-1;
8877         }
8878       }
8879     }
8880     else { // Not delay slot
8881       switch(itype[i]) {
8882         case UJUMP:
8883           //current.isconst=0; // DEBUG
8884           //current.wasconst=0; // DEBUG
8885           //regs[i].wasconst=0; // DEBUG
8886           clear_const(&current,rt1[i]);
8887           alloc_cc(&current,i);
8888           dirty_reg(&current,CCREG);
8889           ooo[i]=1;
8890           delayslot_alloc(&current,i+1);
8891           if (rt1[i]==31) {
8892             alloc_reg(&current,i,31);
8893             dirty_reg(&current,31);
8894             //assert(rs1[i+1]!=31&&rs2[i+1]!=31);
8895             //assert(rt1[i+1]!=rt1[i]);
8896             #ifdef REG_PREFETCH
8897             alloc_reg(&current,i,PTEMP);
8898             #endif
8899             //current.is32|=1LL<<rt1[i];
8900           }
8901           //current.isconst=0; // DEBUG
8902           ds=1;
8903           //printf("i=%d, isconst=%x\n",i,current.isconst);
8904           break;
8905         case RJUMP:
8906           //current.isconst=0;
8907           //current.wasconst=0;
8908           //regs[i].wasconst=0;
8909           clear_const(&current,rs1[i]);
8910           clear_const(&current,rt1[i]);
8911           alloc_cc(&current,i);
8912           dirty_reg(&current,CCREG);
8913           if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
8914             alloc_reg(&current,i,rs1[i]);
8915             if (rt1[i]!=0) {
8916               alloc_reg(&current,i,rt1[i]);
8917               dirty_reg(&current,rt1[i]);
8918               assert(rs1[i+1]!=rt1[i]&&rs2[i+1]!=rt1[i]);
8919               assert(rt1[i+1]!=rt1[i]);
8920               #ifdef REG_PREFETCH
8921               alloc_reg(&current,i,PTEMP);
8922               #endif
8923             }
8924             #ifdef USE_MINI_HT
8925             if(rs1[i]==31) { // JALR
8926               alloc_reg(&current,i,RHASH);
8927               #ifndef HOST_IMM_ADDR32
8928               alloc_reg(&current,i,RHTBL);
8929               #endif
8930             }
8931             #endif
8932             delayslot_alloc(&current,i+1);
8933           } else {
8934             // The delay slot overwrites our source register,
8935             // allocate a temporary register to hold the old value.
8936             current.isconst=0;
8937             current.wasconst=0;
8938             regs[i].wasconst=0;
8939             delayslot_alloc(&current,i+1);
8940             current.isconst=0;
8941             alloc_reg(&current,i,RTEMP);
8942           }
8943           //current.isconst=0; // DEBUG
8944           ooo[i]=1;
8945           ds=1;
8946           break;
8947         case CJUMP:
8948           //current.isconst=0;
8949           //current.wasconst=0;
8950           //regs[i].wasconst=0;
8951           clear_const(&current,rs1[i]);
8952           clear_const(&current,rs2[i]);
8953           if((opcode[i]&0x3E)==4) // BEQ/BNE
8954           {
8955             alloc_cc(&current,i);
8956             dirty_reg(&current,CCREG);
8957             if(rs1[i]) alloc_reg(&current,i,rs1[i]);
8958             if(rs2[i]) alloc_reg(&current,i,rs2[i]);
8959             if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8960             {
8961               if(rs1[i]) alloc_reg64(&current,i,rs1[i]);
8962               if(rs2[i]) alloc_reg64(&current,i,rs2[i]);
8963             }
8964             if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
8965                (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1]))) {
8966               // The delay slot overwrites one of our conditions.
8967               // Allocate the branch condition registers instead.
8968               current.isconst=0;
8969               current.wasconst=0;
8970               regs[i].wasconst=0;
8971               if(rs1[i]) alloc_reg(&current,i,rs1[i]);
8972               if(rs2[i]) alloc_reg(&current,i,rs2[i]);
8973               if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8974               {
8975                 if(rs1[i]) alloc_reg64(&current,i,rs1[i]);
8976                 if(rs2[i]) alloc_reg64(&current,i,rs2[i]);
8977               }
8978             }
8979             else
8980             {
8981               ooo[i]=1;
8982               delayslot_alloc(&current,i+1);
8983             }
8984           }
8985           else
8986           if((opcode[i]&0x3E)==6) // BLEZ/BGTZ
8987           {
8988             alloc_cc(&current,i);
8989             dirty_reg(&current,CCREG);
8990             alloc_reg(&current,i,rs1[i]);
8991             if(!(current.is32>>rs1[i]&1))
8992             {
8993               alloc_reg64(&current,i,rs1[i]);
8994             }
8995             if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
8996               // The delay slot overwrites one of our conditions.
8997               // Allocate the branch condition registers instead.
8998               current.isconst=0;
8999               current.wasconst=0;
9000               regs[i].wasconst=0;
9001               if(rs1[i]) alloc_reg(&current,i,rs1[i]);
9002               if(!((current.is32>>rs1[i])&1))
9003               {
9004                 if(rs1[i]) alloc_reg64(&current,i,rs1[i]);
9005               }
9006             }
9007             else
9008             {
9009               ooo[i]=1;
9010               delayslot_alloc(&current,i+1);
9011             }
9012           }
9013           else
9014           // Don't alloc the delay slot yet because we might not execute it
9015           if((opcode[i]&0x3E)==0x14) // BEQL/BNEL
9016           {
9017             current.isconst=0;
9018             current.wasconst=0;
9019             regs[i].wasconst=0;
9020             alloc_cc(&current,i);
9021             dirty_reg(&current,CCREG);
9022             alloc_reg(&current,i,rs1[i]);
9023             alloc_reg(&current,i,rs2[i]);
9024             if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
9025             {
9026               alloc_reg64(&current,i,rs1[i]);
9027               alloc_reg64(&current,i,rs2[i]);
9028             }
9029           }
9030           else
9031           if((opcode[i]&0x3E)==0x16) // BLEZL/BGTZL
9032           {
9033             current.isconst=0;
9034             current.wasconst=0;
9035             regs[i].wasconst=0;
9036             alloc_cc(&current,i);
9037             dirty_reg(&current,CCREG);
9038             alloc_reg(&current,i,rs1[i]);
9039             if(!(current.is32>>rs1[i]&1))
9040             {
9041               alloc_reg64(&current,i,rs1[i]);
9042             }
9043           }
9044           ds=1;
9045           //current.isconst=0;
9046           break;
9047         case SJUMP:
9048           //current.isconst=0;
9049           //current.wasconst=0;
9050           //regs[i].wasconst=0;
9051           clear_const(&current,rs1[i]);
9052           clear_const(&current,rt1[i]);
9053           //if((opcode2[i]&0x1E)==0x0) // BLTZ/BGEZ
9054           if((opcode2[i]&0x0E)==0x0) // BLTZ/BGEZ
9055           {
9056             alloc_cc(&current,i);
9057             dirty_reg(&current,CCREG);
9058             alloc_reg(&current,i,rs1[i]);
9059             if(!(current.is32>>rs1[i]&1))
9060             {
9061               alloc_reg64(&current,i,rs1[i]);
9062             }
9063             if (rt1[i]==31) { // BLTZAL/BGEZAL
9064               alloc_reg(&current,i,31);
9065               dirty_reg(&current,31);
9066               //#ifdef REG_PREFETCH
9067               //alloc_reg(&current,i,PTEMP);
9068               //#endif
9069               //current.is32|=1LL<<rt1[i];
9070             }
9071             if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) // The delay slot overwrites the branch condition.
9072                ||(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))) { // DS touches $ra
9073               // Allocate the branch condition registers instead.
9074               current.isconst=0;
9075               current.wasconst=0;
9076               regs[i].wasconst=0;
9077               if(rs1[i]) alloc_reg(&current,i,rs1[i]);
9078               if(!((current.is32>>rs1[i])&1))
9079               {
9080                 if(rs1[i]) alloc_reg64(&current,i,rs1[i]);
9081               }
9082             }
9083             else
9084             {
9085               ooo[i]=1;
9086               delayslot_alloc(&current,i+1);
9087             }
9088           }
9089           else
9090           // Don't alloc the delay slot yet because we might not execute it
9091           if((opcode2[i]&0x1E)==0x2) // BLTZL/BGEZL
9092           {
9093             current.isconst=0;
9094             current.wasconst=0;
9095             regs[i].wasconst=0;
9096             alloc_cc(&current,i);
9097             dirty_reg(&current,CCREG);
9098             alloc_reg(&current,i,rs1[i]);
9099             if(!(current.is32>>rs1[i]&1))
9100             {
9101               alloc_reg64(&current,i,rs1[i]);
9102             }
9103           }
9104           ds=1;
9105           //current.isconst=0;
9106           break;
9107         case FJUMP:
9108           current.isconst=0;
9109           current.wasconst=0;
9110           regs[i].wasconst=0;
9111           if(likely[i]==0) // BC1F/BC1T
9112           {
9113             // TODO: Theoretically we can run out of registers here on x86.
9114             // The delay slot can allocate up to six, and we need to check
9115             // CSREG before executing the delay slot.  Possibly we can drop
9116             // the cycle count and then reload it after checking that the
9117             // FPU is in a usable state, or don't do out-of-order execution.
9118             alloc_cc(&current,i);
9119             dirty_reg(&current,CCREG);
9120             alloc_reg(&current,i,FSREG);
9121             alloc_reg(&current,i,CSREG);
9122             if(itype[i+1]==FCOMP) {
9123               // The delay slot overwrites the branch condition.
9124               // Allocate the branch condition registers instead.
9125               alloc_cc(&current,i);
9126               dirty_reg(&current,CCREG);
9127               alloc_reg(&current,i,CSREG);
9128               alloc_reg(&current,i,FSREG);
9129             }
9130             else {
9131               ooo[i]=1;
9132               delayslot_alloc(&current,i+1);
9133               alloc_reg(&current,i+1,CSREG);
9134             }
9135           }
9136           else
9137           // Don't alloc the delay slot yet because we might not execute it
9138           if(likely[i]) // BC1FL/BC1TL
9139           {
9140             alloc_cc(&current,i);
9141             dirty_reg(&current,CCREG);
9142             alloc_reg(&current,i,CSREG);
9143             alloc_reg(&current,i,FSREG);
9144           }
9145           ds=1;
9146           current.isconst=0;
9147           break;
9148         case IMM16:
9149           imm16_alloc(&current,i);
9150           break;
9151         case LOAD:
9152         case LOADLR:
9153           load_alloc(&current,i);
9154           break;
9155         case STORE:
9156         case STORELR:
9157           store_alloc(&current,i);
9158           break;
9159         case ALU:
9160           alu_alloc(&current,i);
9161           break;
9162         case SHIFT:
9163           shift_alloc(&current,i);
9164           break;
9165         case MULTDIV:
9166           multdiv_alloc(&current,i);
9167           break;
9168         case SHIFTIMM:
9169           shiftimm_alloc(&current,i);
9170           break;
9171         case MOV:
9172           mov_alloc(&current,i);
9173           break;
9174         case COP0:
9175           cop0_alloc(&current,i);
9176           break;
9177         case COP1:
9178         case COP2:
9179           cop1_alloc(&current,i);
9180           break;
9181         case C1LS:
9182           c1ls_alloc(&current,i);
9183           break;
9184         case C2LS:
9185           c2ls_alloc(&current,i);
9186           break;
9187         case C2OP:
9188           c2op_alloc(&current,i);
9189           break;
9190         case FCONV:
9191           fconv_alloc(&current,i);
9192           break;
9193         case FLOAT:
9194           float_alloc(&current,i);
9195           break;
9196         case FCOMP:
9197           fcomp_alloc(&current,i);
9198           break;
9199         case SYSCALL:
9200         case HLECALL:
9201         case INTCALL:
9202           syscall_alloc(&current,i);
9203           break;
9204         case SPAN:
9205           pagespan_alloc(&current,i);
9206           break;
9207       }
9208
9209       // Drop the upper half of registers that have become 32-bit
9210       current.uu|=current.is32&((1LL<<rt1[i])|(1LL<<rt2[i]));
9211       if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP) {
9212         current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
9213         if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9214         current.uu|=1;
9215       } else {
9216         current.uu|=current.is32&((1LL<<rt1[i+1])|(1LL<<rt2[i+1]));
9217         current.uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
9218         if((~current.uu>>rt1[i+1])&1) current.uu&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
9219         current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
9220         current.uu|=1;
9221       }
9222
9223       // Create entry (branch target) regmap
9224       for(hr=0;hr<HOST_REGS;hr++)
9225       {
9226         int r,or,er;
9227         r=current.regmap[hr];
9228         if(r>=0) {
9229           if(r!=regmap_pre[i][hr]) {
9230             // TODO: delay slot (?)
9231             or=get_reg(regmap_pre[i],r); // Get old mapping for this register
9232             if(or<0||(r&63)>=TEMPREG){
9233               regs[i].regmap_entry[hr]=-1;
9234             }
9235             else
9236             {
9237               // Just move it to a different register
9238               regs[i].regmap_entry[hr]=r;
9239               // If it was dirty before, it's still dirty
9240               if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r&63);
9241             }
9242           }
9243           else
9244           {
9245             // Unneeded
9246             if(r==0){
9247               regs[i].regmap_entry[hr]=0;
9248             }
9249             else
9250             if(r<64){
9251               if((current.u>>r)&1) {
9252                 regs[i].regmap_entry[hr]=-1;
9253                 //regs[i].regmap[hr]=-1;
9254                 current.regmap[hr]=-1;
9255               }else
9256                 regs[i].regmap_entry[hr]=r;
9257             }
9258             else {
9259               if((current.uu>>(r&63))&1) {
9260                 regs[i].regmap_entry[hr]=-1;
9261                 //regs[i].regmap[hr]=-1;
9262                 current.regmap[hr]=-1;
9263               }else
9264                 regs[i].regmap_entry[hr]=r;
9265             }
9266           }
9267         } else {
9268           // Branches expect CCREG to be allocated at the target
9269           if(regmap_pre[i][hr]==CCREG)
9270             regs[i].regmap_entry[hr]=CCREG;
9271           else
9272             regs[i].regmap_entry[hr]=-1;
9273         }
9274       }
9275       memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
9276     }
9277     /* Branch post-alloc */
9278     if(i>0)
9279     {
9280       current.was32=current.is32;
9281       current.wasdirty=current.dirty;
9282       switch(itype[i-1]) {
9283         case UJUMP:
9284           memcpy(&branch_regs[i-1],&current,sizeof(current));
9285           branch_regs[i-1].isconst=0;
9286           branch_regs[i-1].wasconst=0;
9287           branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
9288           branch_regs[i-1].uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
9289           alloc_cc(&branch_regs[i-1],i-1);
9290           dirty_reg(&branch_regs[i-1],CCREG);
9291           if(rt1[i-1]==31) { // JAL
9292             alloc_reg(&branch_regs[i-1],i-1,31);
9293             dirty_reg(&branch_regs[i-1],31);
9294             branch_regs[i-1].is32|=1LL<<31;
9295           }
9296           memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9297           memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9298           break;
9299         case RJUMP:
9300           memcpy(&branch_regs[i-1],&current,sizeof(current));
9301           branch_regs[i-1].isconst=0;
9302           branch_regs[i-1].wasconst=0;
9303           branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
9304           branch_regs[i-1].uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
9305           alloc_cc(&branch_regs[i-1],i-1);
9306           dirty_reg(&branch_regs[i-1],CCREG);
9307           alloc_reg(&branch_regs[i-1],i-1,rs1[i-1]);
9308           if(rt1[i-1]!=0) { // JALR
9309             alloc_reg(&branch_regs[i-1],i-1,rt1[i-1]);
9310             dirty_reg(&branch_regs[i-1],rt1[i-1]);
9311             branch_regs[i-1].is32|=1LL<<rt1[i-1];
9312           }
9313           #ifdef USE_MINI_HT
9314           if(rs1[i-1]==31) { // JALR
9315             alloc_reg(&branch_regs[i-1],i-1,RHASH);
9316             #ifndef HOST_IMM_ADDR32
9317             alloc_reg(&branch_regs[i-1],i-1,RHTBL);
9318             #endif
9319           }
9320           #endif
9321           memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9322           memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9323           break;
9324         case CJUMP:
9325           if((opcode[i-1]&0x3E)==4) // BEQ/BNE
9326           {
9327             alloc_cc(&current,i-1);
9328             dirty_reg(&current,CCREG);
9329             if((rs1[i-1]&&(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]))||
9330                (rs2[i-1]&&(rs2[i-1]==rt1[i]||rs2[i-1]==rt2[i]))) {
9331               // The delay slot overwrote one of our conditions
9332               // Delay slot goes after the test (in order)
9333               current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
9334               current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
9335               if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9336               current.u|=1;
9337               current.uu|=1;
9338               delayslot_alloc(&current,i);
9339               current.isconst=0;
9340             }
9341             else
9342             {
9343               current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
9344               current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
9345               // Alloc the branch condition registers
9346               if(rs1[i-1]) alloc_reg(&current,i-1,rs1[i-1]);
9347               if(rs2[i-1]) alloc_reg(&current,i-1,rs2[i-1]);
9348               if(!((current.is32>>rs1[i-1])&(current.is32>>rs2[i-1])&1))
9349               {
9350                 if(rs1[i-1]) alloc_reg64(&current,i-1,rs1[i-1]);
9351                 if(rs2[i-1]) alloc_reg64(&current,i-1,rs2[i-1]);
9352               }
9353             }
9354             memcpy(&branch_regs[i-1],&current,sizeof(current));
9355             branch_regs[i-1].isconst=0;
9356             branch_regs[i-1].wasconst=0;
9357             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
9358             memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9359           }
9360           else
9361           if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
9362           {
9363             alloc_cc(&current,i-1);
9364             dirty_reg(&current,CCREG);
9365             if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
9366               // The delay slot overwrote the branch condition
9367               // Delay slot goes after the test (in order)
9368               current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
9369               current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
9370               if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9371               current.u|=1;
9372               current.uu|=1;
9373               delayslot_alloc(&current,i);
9374               current.isconst=0;
9375             }
9376             else
9377             {
9378               current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
9379               current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
9380               // Alloc the branch condition register
9381               alloc_reg(&current,i-1,rs1[i-1]);
9382               if(!(current.is32>>rs1[i-1]&1))
9383               {
9384                 alloc_reg64(&current,i-1,rs1[i-1]);
9385               }
9386             }
9387             memcpy(&branch_regs[i-1],&current,sizeof(current));
9388             branch_regs[i-1].isconst=0;
9389             branch_regs[i-1].wasconst=0;
9390             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
9391             memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9392           }
9393           else
9394           // Alloc the delay slot in case the branch is taken
9395           if((opcode[i-1]&0x3E)==0x14) // BEQL/BNEL
9396           {
9397             memcpy(&branch_regs[i-1],&current,sizeof(current));
9398             branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9399             branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9400             if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9401             alloc_cc(&branch_regs[i-1],i);
9402             dirty_reg(&branch_regs[i-1],CCREG);
9403             delayslot_alloc(&branch_regs[i-1],i);
9404             branch_regs[i-1].isconst=0;
9405             alloc_reg(&current,i,CCREG); // Not taken path
9406             dirty_reg(&current,CCREG);
9407             memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9408           }
9409           else
9410           if((opcode[i-1]&0x3E)==0x16) // BLEZL/BGTZL
9411           {
9412             memcpy(&branch_regs[i-1],&current,sizeof(current));
9413             branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9414             branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9415             if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9416             alloc_cc(&branch_regs[i-1],i);
9417             dirty_reg(&branch_regs[i-1],CCREG);
9418             delayslot_alloc(&branch_regs[i-1],i);
9419             branch_regs[i-1].isconst=0;
9420             alloc_reg(&current,i,CCREG); // Not taken path
9421             dirty_reg(&current,CCREG);
9422             memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9423           }
9424           break;
9425         case SJUMP:
9426           //if((opcode2[i-1]&0x1E)==0) // BLTZ/BGEZ
9427           if((opcode2[i-1]&0x0E)==0) // BLTZ/BGEZ
9428           {
9429             alloc_cc(&current,i-1);
9430             dirty_reg(&current,CCREG);
9431             if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
9432               // The delay slot overwrote the branch condition
9433               // Delay slot goes after the test (in order)
9434               current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
9435               current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
9436               if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9437               current.u|=1;
9438               current.uu|=1;
9439               delayslot_alloc(&current,i);
9440               current.isconst=0;
9441             }
9442             else
9443             {
9444               current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
9445               current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
9446               // Alloc the branch condition register
9447               alloc_reg(&current,i-1,rs1[i-1]);
9448               if(!(current.is32>>rs1[i-1]&1))
9449               {
9450                 alloc_reg64(&current,i-1,rs1[i-1]);
9451               }
9452             }
9453             memcpy(&branch_regs[i-1],&current,sizeof(current));
9454             branch_regs[i-1].isconst=0;
9455             branch_regs[i-1].wasconst=0;
9456             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
9457             memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9458           }
9459           else
9460           // Alloc the delay slot in case the branch is taken
9461           if((opcode2[i-1]&0x1E)==2) // BLTZL/BGEZL
9462           {
9463             memcpy(&branch_regs[i-1],&current,sizeof(current));
9464             branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9465             branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9466             if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9467             alloc_cc(&branch_regs[i-1],i);
9468             dirty_reg(&branch_regs[i-1],CCREG);
9469             delayslot_alloc(&branch_regs[i-1],i);
9470             branch_regs[i-1].isconst=0;
9471             alloc_reg(&current,i,CCREG); // Not taken path
9472             dirty_reg(&current,CCREG);
9473             memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9474           }
9475           // FIXME: BLTZAL/BGEZAL
9476           if(opcode2[i-1]&0x10) { // BxxZAL
9477             alloc_reg(&branch_regs[i-1],i-1,31);
9478             dirty_reg(&branch_regs[i-1],31);
9479             branch_regs[i-1].is32|=1LL<<31;
9480           }
9481           break;
9482         case FJUMP:
9483           if(likely[i-1]==0) // BC1F/BC1T
9484           {
9485             alloc_cc(&current,i-1);
9486             dirty_reg(&current,CCREG);
9487             if(itype[i]==FCOMP) {
9488               // The delay slot overwrote the branch condition
9489               // Delay slot goes after the test (in order)
9490               delayslot_alloc(&current,i);
9491               current.isconst=0;
9492             }
9493             else
9494             {
9495               current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
9496               current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
9497               // Alloc the branch condition register
9498               alloc_reg(&current,i-1,FSREG);
9499             }
9500             memcpy(&branch_regs[i-1],&current,sizeof(current));
9501             memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
9502           }
9503           else // BC1FL/BC1TL
9504           {
9505             // Alloc the delay slot in case the branch is taken
9506             memcpy(&branch_regs[i-1],&current,sizeof(current));
9507             branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9508             branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9509             if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9510             alloc_cc(&branch_regs[i-1],i);
9511             dirty_reg(&branch_regs[i-1],CCREG);
9512             delayslot_alloc(&branch_regs[i-1],i);
9513             branch_regs[i-1].isconst=0;
9514             alloc_reg(&current,i,CCREG); // Not taken path
9515             dirty_reg(&current,CCREG);
9516             memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9517           }
9518           break;
9519       }
9520
9521       if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
9522       {
9523         if(rt1[i-1]==31) // JAL/JALR
9524         {
9525           // Subroutine call will return here, don't alloc any registers
9526           current.is32=1;
9527           current.dirty=0;
9528           clear_all_regs(current.regmap);
9529           alloc_reg(&current,i,CCREG);
9530           dirty_reg(&current,CCREG);
9531         }
9532         else if(i+1<slen)
9533         {
9534           // Internal branch will jump here, match registers to caller
9535           current.is32=0x3FFFFFFFFLL;
9536           current.dirty=0;
9537           clear_all_regs(current.regmap);
9538           alloc_reg(&current,i,CCREG);
9539           dirty_reg(&current,CCREG);
9540           for(j=i-1;j>=0;j--)
9541           {
9542             if(ba[j]==start+i*4+4) {
9543               memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
9544               current.is32=branch_regs[j].is32;
9545               current.dirty=branch_regs[j].dirty;
9546               break;
9547             }
9548           }
9549           while(j>=0) {
9550             if(ba[j]==start+i*4+4) {
9551               for(hr=0;hr<HOST_REGS;hr++) {
9552                 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
9553                   current.regmap[hr]=-1;
9554                 }
9555                 current.is32&=branch_regs[j].is32;
9556                 current.dirty&=branch_regs[j].dirty;
9557               }
9558             }
9559             j--;
9560           }
9561         }
9562       }
9563     }
9564
9565     // Count cycles in between branches
9566     ccadj[i]=cc;
9567     if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP||itype[i]==SYSCALL||itype[i]==HLECALL))
9568     {
9569       cc=0;
9570     }
9571 #ifdef PCSX
9572     else if(/*itype[i]==LOAD||*/itype[i]==STORE||itype[i]==C1LS) // load causes weird timing issues
9573     {
9574       cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
9575     }
9576     else if(itype[i]==C2LS)
9577     {
9578       cc+=4;
9579     }
9580 #endif
9581     else
9582     {
9583       cc++;
9584     }
9585
9586     flush_dirty_uppers(&current);
9587     if(!is_ds[i]) {
9588       regs[i].is32=current.is32;
9589       regs[i].dirty=current.dirty;
9590       regs[i].isconst=current.isconst;
9591       memcpy(constmap[i],current.constmap,sizeof(current.constmap));
9592     }
9593     for(hr=0;hr<HOST_REGS;hr++) {
9594       if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
9595         if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
9596           regs[i].wasconst&=~(1<<hr);
9597         }
9598       }
9599     }
9600     if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
9601   }
9602
9603   /* Pass 4 - Cull unused host registers */
9604
9605   uint64_t nr=0;
9606
9607   for (i=slen-1;i>=0;i--)
9608   {
9609     int hr;
9610     if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9611     {
9612       if(ba[i]<start || ba[i]>=(start+slen*4))
9613       {
9614         // Branch out of this block, don't need anything
9615         nr=0;
9616       }
9617       else
9618       {
9619         // Internal branch
9620         // Need whatever matches the target
9621         nr=0;
9622         int t=(ba[i]-start)>>2;
9623         for(hr=0;hr<HOST_REGS;hr++)
9624         {
9625           if(regs[i].regmap_entry[hr]>=0) {
9626             if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
9627           }
9628         }
9629       }
9630       // Conditional branch may need registers for following instructions
9631       if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9632       {
9633         if(i<slen-2) {
9634           nr|=needed_reg[i+2];
9635           for(hr=0;hr<HOST_REGS;hr++)
9636           {
9637             if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
9638             //if((regmap_entry[i+2][hr])>=0) if(!((nr>>hr)&1)) printf("%x-bogus(%d=%d)\n",start+i*4,hr,regmap_entry[i+2][hr]);
9639           }
9640         }
9641       }
9642       // Don't need stuff which is overwritten
9643       //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
9644       //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
9645       // Merge in delay slot
9646       for(hr=0;hr<HOST_REGS;hr++)
9647       {
9648         if(!likely[i]) {
9649           // These are overwritten unless the branch is "likely"
9650           // and the delay slot is nullified if not taken
9651           if(rt1[i+1]&&rt1[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9652           if(rt2[i+1]&&rt2[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9653         }
9654         if(us1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9655         if(us2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9656         if(rs1[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
9657         if(rs2[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
9658         if(us1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9659         if(us2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9660         if(rs1[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9661         if(rs2[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9662         if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1)) {
9663           if(dep1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9664           if(dep2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9665         }
9666         if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1)) {
9667           if(dep1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9668           if(dep2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9669         }
9670         if(itype[i+1]==STORE || itype[i+1]==STORELR || (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) {
9671           if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
9672           if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
9673         }
9674       }
9675     }
9676     else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
9677     {
9678       // SYSCALL instruction (software interrupt)
9679       nr=0;
9680     }
9681     else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
9682     {
9683       // ERET instruction (return from interrupt)
9684       nr=0;
9685     }
9686     else // Non-branch
9687     {
9688       if(i<slen-1) {
9689         for(hr=0;hr<HOST_REGS;hr++) {
9690           if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
9691           if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
9692           if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
9693           if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
9694         }
9695       }
9696     }
9697     for(hr=0;hr<HOST_REGS;hr++)
9698     {
9699       // Overwritten registers are not needed
9700       if(rt1[i]&&rt1[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9701       if(rt2[i]&&rt2[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9702       if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9703       // Source registers are needed
9704       if(us1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9705       if(us2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9706       if(rs1[i]==regmap_pre[i][hr]) nr|=1<<hr;
9707       if(rs2[i]==regmap_pre[i][hr]) nr|=1<<hr;
9708       if(us1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9709       if(us2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9710       if(rs1[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9711       if(rs2[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9712       if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1)) {
9713         if(dep1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9714         if(dep1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9715       }
9716       if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1)) {
9717         if(dep2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9718         if(dep2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9719       }
9720       if(itype[i]==STORE || itype[i]==STORELR || (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) {
9721         if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
9722         if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
9723       }
9724       // Don't store a register immediately after writing it,
9725       // may prevent dual-issue.
9726       // But do so if this is a branch target, otherwise we
9727       // might have to load the register before the branch.
9728       if(i>0&&!bt[i]&&((regs[i].wasdirty>>hr)&1)) {
9729         if((regmap_pre[i][hr]>0&&regmap_pre[i][hr]<64&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1)) ||
9730            (regmap_pre[i][hr]>64&&!((unneeded_reg_upper[i]>>(regmap_pre[i][hr]&63))&1)) ) {
9731           if(rt1[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9732           if(rt2[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9733         }
9734         if((regs[i].regmap_entry[hr]>0&&regs[i].regmap_entry[hr]<64&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1)) ||
9735            (regs[i].regmap_entry[hr]>64&&!((unneeded_reg_upper[i]>>(regs[i].regmap_entry[hr]&63))&1)) ) {
9736           if(rt1[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9737           if(rt2[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9738         }
9739       }
9740     }
9741     // Cycle count is needed at branches.  Assume it is needed at the target too.
9742     if(i==0||bt[i]||itype[i]==CJUMP||itype[i]==FJUMP||itype[i]==SPAN) {
9743       if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
9744       if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
9745     }
9746     // Save it
9747     needed_reg[i]=nr;
9748
9749     // Deallocate unneeded registers
9750     for(hr=0;hr<HOST_REGS;hr++)
9751     {
9752       if(!((nr>>hr)&1)) {
9753         if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
9754         if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
9755            (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9756            (regs[i].regmap[hr]&63)!=PTEMP && (regs[i].regmap[hr]&63)!=CCREG)
9757         {
9758           if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9759           {
9760             if(likely[i]) {
9761               regs[i].regmap[hr]=-1;
9762               regs[i].isconst&=~(1<<hr);
9763               if(i<slen-2) {
9764                 regmap_pre[i+2][hr]=-1;
9765                 regs[i+2].wasconst&=~(1<<hr);
9766               }
9767             }
9768           }
9769         }
9770         if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9771         {
9772           int d1=0,d2=0,map=0,temp=0;
9773           if(get_reg(regs[i].regmap,rt1[i+1]|64)>=0||get_reg(branch_regs[i].regmap,rt1[i+1]|64)>=0)
9774           {
9775             d1=dep1[i+1];
9776             d2=dep2[i+1];
9777           }
9778           if(using_tlb) {
9779             if(itype[i+1]==LOAD || itype[i+1]==LOADLR ||
9780                itype[i+1]==STORE || itype[i+1]==STORELR ||
9781                itype[i+1]==C1LS || itype[i+1]==C2LS)
9782             map=TLREG;
9783           } else
9784           if(itype[i+1]==STORE || itype[i+1]==STORELR ||
9785              (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
9786             map=INVCP;
9787           }
9788           if(itype[i+1]==LOADLR || itype[i+1]==STORELR ||
9789              itype[i+1]==C1LS || itype[i+1]==C2LS)
9790             temp=FTEMP;
9791           if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
9792              (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9793              (regs[i].regmap[hr]&63)!=rt1[i+1] && (regs[i].regmap[hr]&63)!=rt2[i+1] &&
9794              (regs[i].regmap[hr]^64)!=us1[i+1] && (regs[i].regmap[hr]^64)!=us2[i+1] &&
9795              (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
9796              regs[i].regmap[hr]!=rs1[i+1] && regs[i].regmap[hr]!=rs2[i+1] &&
9797              (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
9798              regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
9799              regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
9800              regs[i].regmap[hr]!=map )
9801           {
9802             regs[i].regmap[hr]=-1;
9803             regs[i].isconst&=~(1<<hr);
9804             if((branch_regs[i].regmap[hr]&63)!=rs1[i] && (branch_regs[i].regmap[hr]&63)!=rs2[i] &&
9805                (branch_regs[i].regmap[hr]&63)!=rt1[i] && (branch_regs[i].regmap[hr]&63)!=rt2[i] &&
9806                (branch_regs[i].regmap[hr]&63)!=rt1[i+1] && (branch_regs[i].regmap[hr]&63)!=rt2[i+1] &&
9807                (branch_regs[i].regmap[hr]^64)!=us1[i+1] && (branch_regs[i].regmap[hr]^64)!=us2[i+1] &&
9808                (branch_regs[i].regmap[hr]^64)!=d1 && (branch_regs[i].regmap[hr]^64)!=d2 &&
9809                branch_regs[i].regmap[hr]!=rs1[i+1] && branch_regs[i].regmap[hr]!=rs2[i+1] &&
9810                (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
9811                branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
9812                branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
9813                branch_regs[i].regmap[hr]!=map)
9814             {
9815               branch_regs[i].regmap[hr]=-1;
9816               branch_regs[i].regmap_entry[hr]=-1;
9817               if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9818               {
9819                 if(!likely[i]&&i<slen-2) {
9820                   regmap_pre[i+2][hr]=-1;
9821                   regs[i+2].wasconst&=~(1<<hr);
9822                 }
9823               }
9824             }
9825           }
9826         }
9827         else
9828         {
9829           // Non-branch
9830           if(i>0)
9831           {
9832             int d1=0,d2=0,map=-1,temp=-1;
9833             if(get_reg(regs[i].regmap,rt1[i]|64)>=0)
9834             {
9835               d1=dep1[i];
9836               d2=dep2[i];
9837             }
9838             if(using_tlb) {
9839               if(itype[i]==LOAD || itype[i]==LOADLR ||
9840                  itype[i]==STORE || itype[i]==STORELR ||
9841                  itype[i]==C1LS || itype[i]==C2LS)
9842               map=TLREG;
9843             } else if(itype[i]==STORE || itype[i]==STORELR ||
9844                       (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
9845               map=INVCP;
9846             }
9847             if(itype[i]==LOADLR || itype[i]==STORELR ||
9848                itype[i]==C1LS || itype[i]==C2LS)
9849               temp=FTEMP;
9850             if((regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9851                (regs[i].regmap[hr]^64)!=us1[i] && (regs[i].regmap[hr]^64)!=us2[i] &&
9852                (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
9853                regs[i].regmap[hr]!=rs1[i] && regs[i].regmap[hr]!=rs2[i] &&
9854                (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map &&
9855                (itype[i]!=SPAN||regs[i].regmap[hr]!=CCREG))
9856             {
9857               if(i<slen-1&&!is_ds[i]) {
9858                 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]!=-1)
9859                 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
9860                 if(regs[i].regmap[hr]<64||!((regs[i].was32>>(regs[i].regmap[hr]&63))&1))
9861                 {
9862                   printf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
9863                   assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
9864                 }
9865                 regmap_pre[i+1][hr]=-1;
9866                 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
9867                 regs[i+1].wasconst&=~(1<<hr);
9868               }
9869               regs[i].regmap[hr]=-1;
9870               regs[i].isconst&=~(1<<hr);
9871             }
9872           }
9873         }
9874       }
9875     }
9876   }
9877
9878   /* Pass 5 - Pre-allocate registers */
9879
9880   // If a register is allocated during a loop, try to allocate it for the
9881   // entire loop, if possible.  This avoids loading/storing registers
9882   // inside of the loop.
9883
9884   signed char f_regmap[HOST_REGS];
9885   clear_all_regs(f_regmap);
9886   for(i=0;i<slen-1;i++)
9887   {
9888     if(itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9889     {
9890       if(ba[i]>=start && ba[i]<(start+i*4))
9891       if(itype[i+1]==NOP||itype[i+1]==MOV||itype[i+1]==ALU
9892       ||itype[i+1]==SHIFTIMM||itype[i+1]==IMM16||itype[i+1]==LOAD
9893       ||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
9894       ||itype[i+1]==SHIFT||itype[i+1]==COP1||itype[i+1]==FLOAT
9895       ||itype[i+1]==FCOMP||itype[i+1]==FCONV
9896       ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
9897       {
9898         int t=(ba[i]-start)>>2;
9899         if(t>0&&(itype[t-1]!=UJUMP&&itype[t-1]!=RJUMP&&itype[t-1]!=CJUMP&&itype[t-1]!=SJUMP&&itype[t-1]!=FJUMP)) // loop_preload can't handle jumps into delay slots
9900         if(t<2||(itype[t-2]!=UJUMP)) // call/ret assumes no registers allocated
9901         for(hr=0;hr<HOST_REGS;hr++)
9902         {
9903           if(regs[i].regmap[hr]>64) {
9904             if(!((regs[i].dirty>>hr)&1))
9905               f_regmap[hr]=regs[i].regmap[hr];
9906             else f_regmap[hr]=-1;
9907           }
9908           else if(regs[i].regmap[hr]>=0) {
9909             if(f_regmap[hr]!=regs[i].regmap[hr]) {
9910               // dealloc old register
9911               int n;
9912               for(n=0;n<HOST_REGS;n++)
9913               {
9914                 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
9915               }
9916               // and alloc new one
9917               f_regmap[hr]=regs[i].regmap[hr];
9918             }
9919           }
9920           if(branch_regs[i].regmap[hr]>64) {
9921             if(!((branch_regs[i].dirty>>hr)&1))
9922               f_regmap[hr]=branch_regs[i].regmap[hr];
9923             else f_regmap[hr]=-1;
9924           }
9925           else if(branch_regs[i].regmap[hr]>=0) {
9926             if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
9927               // dealloc old register
9928               int n;
9929               for(n=0;n<HOST_REGS;n++)
9930               {
9931                 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
9932               }
9933               // and alloc new one
9934               f_regmap[hr]=branch_regs[i].regmap[hr];
9935             }
9936           }
9937           if(ooo[i]) {
9938             if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
9939               f_regmap[hr]=branch_regs[i].regmap[hr];
9940           }else{
9941             if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
9942               f_regmap[hr]=branch_regs[i].regmap[hr];
9943           }
9944           // Avoid dirty->clean transition
9945           #ifdef DESTRUCTIVE_WRITEBACK
9946           if(t>0) if(get_reg(regmap_pre[t],f_regmap[hr])>=0) if((regs[t].wasdirty>>get_reg(regmap_pre[t],f_regmap[hr]))&1) f_regmap[hr]=-1;
9947           #endif
9948           // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
9949           // case above, however it's always a good idea.  We can't hoist the
9950           // load if the register was already allocated, so there's no point
9951           // wasting time analyzing most of these cases.  It only "succeeds"
9952           // when the mapping was different and the load can be replaced with
9953           // a mov, which is of negligible benefit.  So such cases are
9954           // skipped below.
9955           if(f_regmap[hr]>0) {
9956             if(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0) {
9957               int r=f_regmap[hr];
9958               for(j=t;j<=i;j++)
9959               {
9960                 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
9961                 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
9962                 if(r>63&&((unneeded_reg_upper[j]>>(r&63))&1)) break;
9963                 if(r>63) {
9964                   // NB This can exclude the case where the upper-half
9965                   // register is lower numbered than the lower-half
9966                   // register.  Not sure if it's worth fixing...
9967                   if(get_reg(regs[j].regmap,r&63)<0) break;
9968                   if(get_reg(regs[j].regmap_entry,r&63)<0) break;
9969                   if(regs[j].is32&(1LL<<(r&63))) break;
9970                 }
9971                 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
9972                   //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
9973                   int k;
9974                   if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
9975                     if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
9976                     if(r>63) {
9977                       if(get_reg(regs[i].regmap,r&63)<0) break;
9978                       if(get_reg(branch_regs[i].regmap,r&63)<0) break;
9979                     }
9980                     k=i;
9981                     while(k>1&&regs[k-1].regmap[hr]==-1) {
9982                       if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
9983                         //printf("no free regs for store %x\n",start+(k-1)*4);
9984                         break;
9985                       }
9986                       if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
9987                         //printf("no-match due to different register\n");
9988                         break;
9989                       }
9990                       if(itype[k-2]==UJUMP||itype[k-2]==RJUMP||itype[k-2]==CJUMP||itype[k-2]==SJUMP||itype[k-2]==FJUMP) {
9991                         //printf("no-match due to branch\n");
9992                         break;
9993                       }
9994                       // call/ret fast path assumes no registers allocated
9995                       if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)) {
9996                         break;
9997                       }
9998                       if(r>63) {
9999                         // NB This can exclude the case where the upper-half
10000                         // register is lower numbered than the lower-half
10001                         // register.  Not sure if it's worth fixing...
10002                         if(get_reg(regs[k-1].regmap,r&63)<0) break;
10003                         if(regs[k-1].is32&(1LL<<(r&63))) break;
10004                       }
10005                       k--;
10006                     }
10007                     if(i<slen-1) {
10008                       if((regs[k].is32&(1LL<<f_regmap[hr]))!=
10009                         (regs[i+2].was32&(1LL<<f_regmap[hr]))) {
10010                         //printf("bad match after branch\n");
10011                         break;
10012                       }
10013                     }
10014                     if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
10015                       //printf("Extend r%d, %x ->\n",hr,start+k*4);
10016                       while(k<i) {
10017                         regs[k].regmap_entry[hr]=f_regmap[hr];
10018                         regs[k].regmap[hr]=f_regmap[hr];
10019                         regmap_pre[k+1][hr]=f_regmap[hr];
10020                         regs[k].wasdirty&=~(1<<hr);
10021                         regs[k].dirty&=~(1<<hr);
10022                         regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
10023                         regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
10024                         regs[k].wasconst&=~(1<<hr);
10025                         regs[k].isconst&=~(1<<hr);
10026                         k++;
10027                       }
10028                     }
10029                     else {
10030                       //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
10031                       break;
10032                     }
10033                     assert(regs[i-1].regmap[hr]==f_regmap[hr]);
10034                     if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
10035                       //printf("OK fill %x (r%d)\n",start+i*4,hr);
10036                       regs[i].regmap_entry[hr]=f_regmap[hr];
10037                       regs[i].regmap[hr]=f_regmap[hr];
10038                       regs[i].wasdirty&=~(1<<hr);
10039                       regs[i].dirty&=~(1<<hr);
10040                       regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
10041                       regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
10042                       regs[i].wasconst&=~(1<<hr);
10043                       regs[i].isconst&=~(1<<hr);
10044                       branch_regs[i].regmap_entry[hr]=f_regmap[hr];
10045                       branch_regs[i].wasdirty&=~(1<<hr);
10046                       branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
10047                       branch_regs[i].regmap[hr]=f_regmap[hr];
10048                       branch_regs[i].dirty&=~(1<<hr);
10049                       branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
10050                       branch_regs[i].wasconst&=~(1<<hr);
10051                       branch_regs[i].isconst&=~(1<<hr);
10052                       if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
10053                         regmap_pre[i+2][hr]=f_regmap[hr];
10054                         regs[i+2].wasdirty&=~(1<<hr);
10055                         regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
10056                         assert((branch_regs[i].is32&(1LL<<f_regmap[hr]))==
10057                           (regs[i+2].was32&(1LL<<f_regmap[hr])));
10058                       }
10059                     }
10060                   }
10061                   for(k=t;k<j;k++) {
10062                     // Alloc register clean at beginning of loop,
10063                     // but may dirty it in pass 6
10064                     regs[k].regmap_entry[hr]=f_regmap[hr];
10065                     regs[k].regmap[hr]=f_regmap[hr];
10066                     regs[k].dirty&=~(1<<hr);
10067                     regs[k].wasconst&=~(1<<hr);
10068                     regs[k].isconst&=~(1<<hr);
10069                     if(itype[k]==UJUMP||itype[k]==RJUMP||itype[k]==CJUMP||itype[k]==SJUMP||itype[k]==FJUMP) {
10070                       branch_regs[k].regmap_entry[hr]=f_regmap[hr];
10071                       branch_regs[k].regmap[hr]=f_regmap[hr];
10072                       branch_regs[k].dirty&=~(1<<hr);
10073                       branch_regs[k].wasconst&=~(1<<hr);
10074                       branch_regs[k].isconst&=~(1<<hr);
10075                       if(itype[k]!=RJUMP&&itype[k]!=UJUMP&&(source[k]>>16)!=0x1000) {
10076                         regmap_pre[k+2][hr]=f_regmap[hr];
10077                         regs[k+2].wasdirty&=~(1<<hr);
10078                         assert((branch_regs[k].is32&(1LL<<f_regmap[hr]))==
10079                           (regs[k+2].was32&(1LL<<f_regmap[hr])));
10080                       }
10081                     }
10082                     else
10083                     {
10084                       regmap_pre[k+1][hr]=f_regmap[hr];
10085                       regs[k+1].wasdirty&=~(1<<hr);
10086                     }
10087                   }
10088                   if(regs[j].regmap[hr]==f_regmap[hr])
10089                     regs[j].regmap_entry[hr]=f_regmap[hr];
10090                   break;
10091                 }
10092                 if(j==i) break;
10093                 if(regs[j].regmap[hr]>=0)
10094                   break;
10095                 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
10096                   //printf("no-match due to different register\n");
10097                   break;
10098                 }
10099                 if((regs[j+1].is32&(1LL<<f_regmap[hr]))!=(regs[j].is32&(1LL<<f_regmap[hr]))) {
10100                   //printf("32/64 mismatch %x %d\n",start+j*4,hr);
10101                   break;
10102                 }
10103                 if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
10104                 {
10105                   // Stop on unconditional branch
10106                   break;
10107                 }
10108                 if(itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP)
10109                 {
10110                   if(ooo[j]) {
10111                     if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
10112                       break;
10113                   }else{
10114                     if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
10115                       break;
10116                   }
10117                   if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
10118                     //printf("no-match due to different register (branch)\n");
10119                     break;
10120                   }
10121                 }
10122                 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
10123                   //printf("No free regs for store %x\n",start+j*4);
10124                   break;
10125                 }
10126                 if(f_regmap[hr]>=64) {
10127                   if(regs[j].is32&(1LL<<(f_regmap[hr]&63))) {
10128                     break;
10129                   }
10130                   else
10131                   {
10132                     if(get_reg(regs[j].regmap,f_regmap[hr]&63)<0) {
10133                       break;
10134                     }
10135                   }
10136                 }
10137               }
10138             }
10139           }
10140         }
10141       }
10142     }else{
10143       int count=0;
10144       for(hr=0;hr<HOST_REGS;hr++)
10145       {
10146         if(hr!=EXCLUDE_REG) {
10147           if(regs[i].regmap[hr]>64) {
10148             if(!((regs[i].dirty>>hr)&1))
10149               f_regmap[hr]=regs[i].regmap[hr];
10150           }
10151           else if(regs[i].regmap[hr]>=0) {
10152             if(f_regmap[hr]!=regs[i].regmap[hr]) {
10153               // dealloc old register
10154               int n;
10155               for(n=0;n<HOST_REGS;n++)
10156               {
10157                 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
10158               }
10159               // and alloc new one
10160               f_regmap[hr]=regs[i].regmap[hr];
10161             }
10162           }
10163           else if(regs[i].regmap[hr]<0) count++;
10164         }
10165       }
10166       // Try to restore cycle count at branch targets
10167       if(bt[i]) {
10168         for(j=i;j<slen-1;j++) {
10169           if(regs[j].regmap[HOST_CCREG]!=-1) break;
10170           if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
10171             //printf("no free regs for store %x\n",start+j*4);
10172             break;
10173           }
10174         }
10175         if(regs[j].regmap[HOST_CCREG]==CCREG) {
10176           int k=i;
10177           //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
10178           while(k<j) {
10179             regs[k].regmap_entry[HOST_CCREG]=CCREG;
10180             regs[k].regmap[HOST_CCREG]=CCREG;
10181             regmap_pre[k+1][HOST_CCREG]=CCREG;
10182             regs[k+1].wasdirty|=1<<HOST_CCREG;
10183             regs[k].dirty|=1<<HOST_CCREG;
10184             regs[k].wasconst&=~(1<<HOST_CCREG);
10185             regs[k].isconst&=~(1<<HOST_CCREG);
10186             k++;
10187           }
10188           regs[j].regmap_entry[HOST_CCREG]=CCREG;
10189         }
10190         // Work backwards from the branch target
10191         if(j>i&&f_regmap[HOST_CCREG]==CCREG)
10192         {
10193           //printf("Extend backwards\n");
10194           int k;
10195           k=i;
10196           while(regs[k-1].regmap[HOST_CCREG]==-1) {
10197             if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
10198               //printf("no free regs for store %x\n",start+(k-1)*4);
10199               break;
10200             }
10201             k--;
10202           }
10203           if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
10204             //printf("Extend CC, %x ->\n",start+k*4);
10205             while(k<=i) {
10206               regs[k].regmap_entry[HOST_CCREG]=CCREG;
10207               regs[k].regmap[HOST_CCREG]=CCREG;
10208               regmap_pre[k+1][HOST_CCREG]=CCREG;
10209               regs[k+1].wasdirty|=1<<HOST_CCREG;
10210               regs[k].dirty|=1<<HOST_CCREG;
10211               regs[k].wasconst&=~(1<<HOST_CCREG);
10212               regs[k].isconst&=~(1<<HOST_CCREG);
10213               k++;
10214             }
10215           }
10216           else {
10217             //printf("Fail Extend CC, %x ->\n",start+k*4);
10218           }
10219         }
10220       }
10221       if(itype[i]!=STORE&&itype[i]!=STORELR&&itype[i]!=C1LS&&itype[i]!=SHIFT&&
10222          itype[i]!=NOP&&itype[i]!=MOV&&itype[i]!=ALU&&itype[i]!=SHIFTIMM&&
10223          itype[i]!=IMM16&&itype[i]!=LOAD&&itype[i]!=COP1&&itype[i]!=FLOAT&&
10224          itype[i]!=FCONV&&itype[i]!=FCOMP)
10225       {
10226         memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
10227       }
10228     }
10229   }
10230
10231   // Cache memory offset or tlb map pointer if a register is available
10232   #ifndef HOST_IMM_ADDR32
10233   #ifndef RAM_OFFSET
10234   if(using_tlb)
10235   #endif
10236   {
10237     int earliest_available[HOST_REGS];
10238     int loop_start[HOST_REGS];
10239     int score[HOST_REGS];
10240     int end[HOST_REGS];
10241     int reg=using_tlb?MMREG:ROREG;
10242
10243     // Init
10244     for(hr=0;hr<HOST_REGS;hr++) {
10245       score[hr]=0;earliest_available[hr]=0;
10246       loop_start[hr]=MAXBLOCK;
10247     }
10248     for(i=0;i<slen-1;i++)
10249     {
10250       // Can't do anything if no registers are available
10251       if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i]) {
10252         for(hr=0;hr<HOST_REGS;hr++) {
10253           score[hr]=0;earliest_available[hr]=i+1;
10254           loop_start[hr]=MAXBLOCK;
10255         }
10256       }
10257       if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
10258         if(!ooo[i]) {
10259           if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1]) {
10260             for(hr=0;hr<HOST_REGS;hr++) {
10261               score[hr]=0;earliest_available[hr]=i+1;
10262               loop_start[hr]=MAXBLOCK;
10263             }
10264           }
10265         }
10266       }
10267       // Mark unavailable registers
10268       for(hr=0;hr<HOST_REGS;hr++) {
10269         if(regs[i].regmap[hr]>=0) {
10270           score[hr]=0;earliest_available[hr]=i+1;
10271           loop_start[hr]=MAXBLOCK;
10272         }
10273         if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
10274           if(branch_regs[i].regmap[hr]>=0) {
10275             score[hr]=0;earliest_available[hr]=i+2;
10276             loop_start[hr]=MAXBLOCK;
10277           }
10278         }
10279       }
10280       // No register allocations after unconditional jumps
10281       if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
10282       {
10283         for(hr=0;hr<HOST_REGS;hr++) {
10284           score[hr]=0;earliest_available[hr]=i+2;
10285           loop_start[hr]=MAXBLOCK;
10286         }
10287         i++; // Skip delay slot too
10288         //printf("skip delay slot: %x\n",start+i*4);
10289       }
10290       else
10291       // Possible match
10292       if(itype[i]==LOAD||itype[i]==LOADLR||
10293          itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS) {
10294         for(hr=0;hr<HOST_REGS;hr++) {
10295           if(hr!=EXCLUDE_REG) {
10296             end[hr]=i-1;
10297             for(j=i;j<slen-1;j++) {
10298               if(regs[j].regmap[hr]>=0) break;
10299               if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
10300                 if(branch_regs[j].regmap[hr]>=0) break;
10301                 if(ooo[j]) {
10302                   if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1]) break;
10303                 }else{
10304                   if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1]) break;
10305                 }
10306               }
10307               else if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) break;
10308               if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
10309                 int t=(ba[j]-start)>>2;
10310                 if(t<j&&t>=earliest_available[hr]) {
10311                   // Score a point for hoisting loop invariant
10312                   if(t<loop_start[hr]) loop_start[hr]=t;
10313                   //printf("set loop_start: i=%x j=%x (%x)\n",start+i*4,start+j*4,start+t*4);
10314                   score[hr]++;
10315                   end[hr]=j;
10316                 }
10317                 else if(t<j) {
10318                   if(regs[t].regmap[hr]==reg) {
10319                     // Score a point if the branch target matches this register
10320                     score[hr]++;
10321                     end[hr]=j;
10322                   }
10323                 }
10324                 if(itype[j+1]==LOAD||itype[j+1]==LOADLR||
10325                    itype[j+1]==STORE||itype[j+1]==STORELR||itype[j+1]==C1LS) {
10326                   score[hr]++;
10327                   end[hr]=j;
10328                 }
10329               }
10330               if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
10331               {
10332                 // Stop on unconditional branch
10333                 break;
10334               }
10335               else
10336               if(itype[j]==LOAD||itype[j]==LOADLR||
10337                  itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS) {
10338                 score[hr]++;
10339                 end[hr]=j;
10340               }
10341             }
10342           }
10343         }
10344         // Find highest score and allocate that register
10345         int maxscore=0;
10346         for(hr=0;hr<HOST_REGS;hr++) {
10347           if(hr!=EXCLUDE_REG) {
10348             if(score[hr]>score[maxscore]) {
10349               maxscore=hr;
10350               //printf("highest score: %d %d (%x->%x)\n",score[hr],hr,start+i*4,start+end[hr]*4);
10351             }
10352           }
10353         }
10354         if(score[maxscore]>1)
10355         {
10356           if(i<loop_start[maxscore]) loop_start[maxscore]=i;
10357           for(j=loop_start[maxscore];j<slen&&j<=end[maxscore];j++) {
10358             //if(regs[j].regmap[maxscore]>=0) {printf("oops: %x %x was %d=%d\n",loop_start[maxscore]*4+start,j*4+start,maxscore,regs[j].regmap[maxscore]);}
10359             assert(regs[j].regmap[maxscore]<0);
10360             if(j>loop_start[maxscore]) regs[j].regmap_entry[maxscore]=reg;
10361             regs[j].regmap[maxscore]=reg;
10362             regs[j].dirty&=~(1<<maxscore);
10363             regs[j].wasconst&=~(1<<maxscore);
10364             regs[j].isconst&=~(1<<maxscore);
10365             if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
10366               branch_regs[j].regmap[maxscore]=reg;
10367               branch_regs[j].wasdirty&=~(1<<maxscore);
10368               branch_regs[j].dirty&=~(1<<maxscore);
10369               branch_regs[j].wasconst&=~(1<<maxscore);
10370               branch_regs[j].isconst&=~(1<<maxscore);
10371               if(itype[j]!=RJUMP&&itype[j]!=UJUMP&&(source[j]>>16)!=0x1000) {
10372                 regmap_pre[j+2][maxscore]=reg;
10373                 regs[j+2].wasdirty&=~(1<<maxscore);
10374               }
10375               // loop optimization (loop_preload)
10376               int t=(ba[j]-start)>>2;
10377               if(t==loop_start[maxscore]) regs[t].regmap_entry[maxscore]=reg;
10378             }
10379             else
10380             {
10381               if(j<1||(itype[j-1]!=RJUMP&&itype[j-1]!=UJUMP&&itype[j-1]!=CJUMP&&itype[j-1]!=SJUMP&&itype[j-1]!=FJUMP)) {
10382                 regmap_pre[j+1][maxscore]=reg;
10383                 regs[j+1].wasdirty&=~(1<<maxscore);
10384               }
10385             }
10386           }
10387           i=j-1;
10388           if(itype[j-1]==RJUMP||itype[j-1]==UJUMP||itype[j-1]==CJUMP||itype[j-1]==SJUMP||itype[j-1]==FJUMP) i++; // skip delay slot
10389           for(hr=0;hr<HOST_REGS;hr++) {
10390             score[hr]=0;earliest_available[hr]=i+i;
10391             loop_start[hr]=MAXBLOCK;
10392           }
10393         }
10394       }
10395     }
10396   }
10397   #endif
10398
10399   // This allocates registers (if possible) one instruction prior
10400   // to use, which can avoid a load-use penalty on certain CPUs.
10401   for(i=0;i<slen-1;i++)
10402   {
10403     if(!i||(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP))
10404     {
10405       if(!bt[i+1])
10406       {
10407         if(itype[i]==ALU||itype[i]==MOV||itype[i]==LOAD||itype[i]==SHIFTIMM||itype[i]==IMM16
10408            ||((itype[i]==COP1||itype[i]==COP2)&&opcode2[i]<3))
10409         {
10410           if(rs1[i+1]) {
10411             if((hr=get_reg(regs[i+1].regmap,rs1[i+1]))>=0)
10412             {
10413               if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10414               {
10415                 regs[i].regmap[hr]=regs[i+1].regmap[hr];
10416                 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
10417                 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
10418                 regs[i].isconst&=~(1<<hr);
10419                 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10420                 constmap[i][hr]=constmap[i+1][hr];
10421                 regs[i+1].wasdirty&=~(1<<hr);
10422                 regs[i].dirty&=~(1<<hr);
10423               }
10424             }
10425           }
10426           if(rs2[i+1]) {
10427             if((hr=get_reg(regs[i+1].regmap,rs2[i+1]))>=0)
10428             {
10429               if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10430               {
10431                 regs[i].regmap[hr]=regs[i+1].regmap[hr];
10432                 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
10433                 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
10434                 regs[i].isconst&=~(1<<hr);
10435                 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10436                 constmap[i][hr]=constmap[i+1][hr];
10437                 regs[i+1].wasdirty&=~(1<<hr);
10438                 regs[i].dirty&=~(1<<hr);
10439               }
10440             }
10441           }
10442           if(itype[i+1]==LOAD&&rs1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
10443             if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
10444             {
10445               if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10446               {
10447                 regs[i].regmap[hr]=rs1[i+1];
10448                 regmap_pre[i+1][hr]=rs1[i+1];
10449                 regs[i+1].regmap_entry[hr]=rs1[i+1];
10450                 regs[i].isconst&=~(1<<hr);
10451                 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10452                 constmap[i][hr]=constmap[i+1][hr];
10453                 regs[i+1].wasdirty&=~(1<<hr);
10454                 regs[i].dirty&=~(1<<hr);
10455               }
10456             }
10457           }
10458           if(lt1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
10459             if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
10460             {
10461               if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10462               {
10463                 regs[i].regmap[hr]=rs1[i+1];
10464                 regmap_pre[i+1][hr]=rs1[i+1];
10465                 regs[i+1].regmap_entry[hr]=rs1[i+1];
10466                 regs[i].isconst&=~(1<<hr);
10467                 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10468                 constmap[i][hr]=constmap[i+1][hr];
10469                 regs[i+1].wasdirty&=~(1<<hr);
10470                 regs[i].dirty&=~(1<<hr);
10471               }
10472             }
10473           }
10474           #ifndef HOST_IMM_ADDR32
10475           if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) {
10476             hr=get_reg(regs[i+1].regmap,TLREG);
10477             if(hr>=0) {
10478               int sr=get_reg(regs[i+1].regmap,rs1[i+1]);
10479               if(sr>=0&&((regs[i+1].wasconst>>sr)&1)) {
10480                 int nr;
10481                 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10482                 {
10483                   regs[i].regmap[hr]=MGEN1+((i+1)&1);
10484                   regmap_pre[i+1][hr]=MGEN1+((i+1)&1);
10485                   regs[i+1].regmap_entry[hr]=MGEN1+((i+1)&1);
10486                   regs[i].isconst&=~(1<<hr);
10487                   regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10488                   constmap[i][hr]=constmap[i+1][hr];
10489                   regs[i+1].wasdirty&=~(1<<hr);
10490                   regs[i].dirty&=~(1<<hr);
10491                 }
10492                 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
10493                 {
10494                   // move it to another register
10495                   regs[i+1].regmap[hr]=-1;
10496                   regmap_pre[i+2][hr]=-1;
10497                   regs[i+1].regmap[nr]=TLREG;
10498                   regmap_pre[i+2][nr]=TLREG;
10499                   regs[i].regmap[nr]=MGEN1+((i+1)&1);
10500                   regmap_pre[i+1][nr]=MGEN1+((i+1)&1);
10501                   regs[i+1].regmap_entry[nr]=MGEN1+((i+1)&1);
10502                   regs[i].isconst&=~(1<<nr);
10503                   regs[i+1].isconst&=~(1<<nr);
10504                   regs[i].dirty&=~(1<<nr);
10505                   regs[i+1].wasdirty&=~(1<<nr);
10506                   regs[i+1].dirty&=~(1<<nr);
10507                   regs[i+2].wasdirty&=~(1<<nr);
10508                 }
10509               }
10510             }
10511           }
10512           #endif
10513           if(itype[i+1]==STORE||itype[i+1]==STORELR
10514              ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
10515             if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
10516               hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
10517               if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
10518               else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
10519               assert(hr>=0);
10520               if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10521               {
10522                 regs[i].regmap[hr]=rs1[i+1];
10523                 regmap_pre[i+1][hr]=rs1[i+1];
10524                 regs[i+1].regmap_entry[hr]=rs1[i+1];
10525                 regs[i].isconst&=~(1<<hr);
10526                 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10527                 constmap[i][hr]=constmap[i+1][hr];
10528                 regs[i+1].wasdirty&=~(1<<hr);
10529                 regs[i].dirty&=~(1<<hr);
10530               }
10531             }
10532           }
10533           if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
10534             if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
10535               int nr;
10536               hr=get_reg(regs[i+1].regmap,FTEMP);
10537               assert(hr>=0);
10538               if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
10539               {
10540                 regs[i].regmap[hr]=rs1[i+1];
10541                 regmap_pre[i+1][hr]=rs1[i+1];
10542                 regs[i+1].regmap_entry[hr]=rs1[i+1];
10543                 regs[i].isconst&=~(1<<hr);
10544                 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10545                 constmap[i][hr]=constmap[i+1][hr];
10546                 regs[i+1].wasdirty&=~(1<<hr);
10547                 regs[i].dirty&=~(1<<hr);
10548               }
10549               else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
10550               {
10551                 // move it to another register
10552                 regs[i+1].regmap[hr]=-1;
10553                 regmap_pre[i+2][hr]=-1;
10554                 regs[i+1].regmap[nr]=FTEMP;
10555                 regmap_pre[i+2][nr]=FTEMP;
10556                 regs[i].regmap[nr]=rs1[i+1];
10557                 regmap_pre[i+1][nr]=rs1[i+1];
10558                 regs[i+1].regmap_entry[nr]=rs1[i+1];
10559                 regs[i].isconst&=~(1<<nr);
10560                 regs[i+1].isconst&=~(1<<nr);
10561                 regs[i].dirty&=~(1<<nr);
10562                 regs[i+1].wasdirty&=~(1<<nr);
10563                 regs[i+1].dirty&=~(1<<nr);
10564                 regs[i+2].wasdirty&=~(1<<nr);
10565               }
10566             }
10567           }
10568           if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR/*||itype[i+1]==C1LS||||itype[i+1]==C2LS*/) {
10569             if(itype[i+1]==LOAD)
10570               hr=get_reg(regs[i+1].regmap,rt1[i+1]);
10571             if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
10572               hr=get_reg(regs[i+1].regmap,FTEMP);
10573             if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
10574               hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
10575               if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
10576             }
10577             if(hr>=0&&regs[i].regmap[hr]<0) {
10578               int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
10579               if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
10580                 regs[i].regmap[hr]=AGEN1+((i+1)&1);
10581                 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
10582                 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
10583                 regs[i].isconst&=~(1<<hr);
10584                 regs[i+1].wasdirty&=~(1<<hr);
10585                 regs[i].dirty&=~(1<<hr);
10586               }
10587             }
10588           }
10589         }
10590       }
10591     }
10592   }
10593
10594   /* Pass 6 - Optimize clean/dirty state */
10595   clean_registers(0,slen-1,1);
10596
10597   /* Pass 7 - Identify 32-bit registers */
10598 #ifndef FORCE32
10599   provisional_r32();
10600
10601   u_int r32=0;
10602
10603   for (i=slen-1;i>=0;i--)
10604   {
10605     int hr;
10606     if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
10607     {
10608       if(ba[i]<start || ba[i]>=(start+slen*4))
10609       {
10610         // Branch out of this block, don't need anything
10611         r32=0;
10612       }
10613       else
10614       {
10615         // Internal branch
10616         // Need whatever matches the target
10617         // (and doesn't get overwritten by the delay slot instruction)
10618         r32=0;
10619         int t=(ba[i]-start)>>2;
10620         if(ba[i]>start+i*4) {
10621           // Forward branch
10622           if(!(requires_32bit[t]&~regs[i].was32))
10623             r32|=requires_32bit[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
10624         }else{
10625           // Backward branch
10626           //if(!(regs[t].was32&~unneeded_reg_upper[t]&~regs[i].was32))
10627           //  r32|=regs[t].was32&~unneeded_reg_upper[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
10628           if(!(pr32[t]&~regs[i].was32))
10629             r32|=pr32[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
10630         }
10631       }
10632       // Conditional branch may need registers for following instructions
10633       if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
10634       {
10635         if(i<slen-2) {
10636           r32|=requires_32bit[i+2];
10637           r32&=regs[i].was32;
10638           // Mark this address as a branch target since it may be called
10639           // upon return from interrupt
10640           bt[i+2]=1;
10641         }
10642       }
10643       // Merge in delay slot
10644       if(!likely[i]) {
10645         // These are overwritten unless the branch is "likely"
10646         // and the delay slot is nullified if not taken
10647         r32&=~(1LL<<rt1[i+1]);
10648         r32&=~(1LL<<rt2[i+1]);
10649       }
10650       // Assume these are needed (delay slot)
10651       if(us1[i+1]>0)
10652       {
10653         if((regs[i].was32>>us1[i+1])&1) r32|=1LL<<us1[i+1];
10654       }
10655       if(us2[i+1]>0)
10656       {
10657         if((regs[i].was32>>us2[i+1])&1) r32|=1LL<<us2[i+1];
10658       }
10659       if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1))
10660       {
10661         if((regs[i].was32>>dep1[i+1])&1) r32|=1LL<<dep1[i+1];
10662       }
10663       if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1))
10664       {
10665         if((regs[i].was32>>dep2[i+1])&1) r32|=1LL<<dep2[i+1];
10666       }
10667     }
10668     else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
10669     {
10670       // SYSCALL instruction (software interrupt)
10671       r32=0;
10672     }
10673     else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
10674     {
10675       // ERET instruction (return from interrupt)
10676       r32=0;
10677     }
10678     // Check 32 bits
10679     r32&=~(1LL<<rt1[i]);
10680     r32&=~(1LL<<rt2[i]);
10681     if(us1[i]>0)
10682     {
10683       if((regs[i].was32>>us1[i])&1) r32|=1LL<<us1[i];
10684     }
10685     if(us2[i]>0)
10686     {
10687       if((regs[i].was32>>us2[i])&1) r32|=1LL<<us2[i];
10688     }
10689     if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1))
10690     {
10691       if((regs[i].was32>>dep1[i])&1) r32|=1LL<<dep1[i];
10692     }
10693     if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1))
10694     {
10695       if((regs[i].was32>>dep2[i])&1) r32|=1LL<<dep2[i];
10696     }
10697     requires_32bit[i]=r32;
10698
10699     // Dirty registers which are 32-bit, require 32-bit input
10700     // as they will be written as 32-bit values
10701     for(hr=0;hr<HOST_REGS;hr++)
10702     {
10703       if(regs[i].regmap_entry[hr]>0&&regs[i].regmap_entry[hr]<64) {
10704         if((regs[i].was32>>regs[i].regmap_entry[hr])&(regs[i].wasdirty>>hr)&1) {
10705           if(!((unneeded_reg_upper[i]>>regs[i].regmap_entry[hr])&1))
10706           requires_32bit[i]|=1LL<<regs[i].regmap_entry[hr];
10707         }
10708       }
10709     }
10710     //requires_32bit[i]=is32[i]&~unneeded_reg_upper[i]; // DEBUG
10711   }
10712 #else
10713   for (i=slen-1;i>=0;i--)
10714   {
10715     if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
10716     {
10717       // Conditional branch
10718       if((source[i]>>16)!=0x1000&&i<slen-2) {
10719         // Mark this address as a branch target since it may be called
10720         // upon return from interrupt
10721         bt[i+2]=1;
10722       }
10723     }
10724   }
10725 #endif
10726
10727   if(itype[slen-1]==SPAN) {
10728     bt[slen-1]=1; // Mark as a branch target so instruction can restart after exception
10729   }
10730
10731   /* Debug/disassembly */
10732   if((void*)assem_debug==(void*)printf)
10733   for(i=0;i<slen;i++)
10734   {
10735     printf("U:");
10736     int r;
10737     for(r=1;r<=CCREG;r++) {
10738       if((unneeded_reg[i]>>r)&1) {
10739         if(r==HIREG) printf(" HI");
10740         else if(r==LOREG) printf(" LO");
10741         else printf(" r%d",r);
10742       }
10743     }
10744 #ifndef FORCE32
10745     printf(" UU:");
10746     for(r=1;r<=CCREG;r++) {
10747       if(((unneeded_reg_upper[i]&~unneeded_reg[i])>>r)&1) {
10748         if(r==HIREG) printf(" HI");
10749         else if(r==LOREG) printf(" LO");
10750         else printf(" r%d",r);
10751       }
10752     }
10753     printf(" 32:");
10754     for(r=0;r<=CCREG;r++) {
10755       //if(((is32[i]>>r)&(~unneeded_reg[i]>>r))&1) {
10756       if((regs[i].was32>>r)&1) {
10757         if(r==CCREG) printf(" CC");
10758         else if(r==HIREG) printf(" HI");
10759         else if(r==LOREG) printf(" LO");
10760         else printf(" r%d",r);
10761       }
10762     }
10763 #endif
10764     printf("\n");
10765     #if defined(__i386__) || defined(__x86_64__)
10766     printf("pre: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7]);
10767     #endif
10768     #ifdef __arm__
10769     printf("pre: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][4],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7],regmap_pre[i][8],regmap_pre[i][9],regmap_pre[i][10],regmap_pre[i][12]);
10770     #endif
10771     printf("needs: ");
10772     if(needed_reg[i]&1) printf("eax ");
10773     if((needed_reg[i]>>1)&1) printf("ecx ");
10774     if((needed_reg[i]>>2)&1) printf("edx ");
10775     if((needed_reg[i]>>3)&1) printf("ebx ");
10776     if((needed_reg[i]>>5)&1) printf("ebp ");
10777     if((needed_reg[i]>>6)&1) printf("esi ");
10778     if((needed_reg[i]>>7)&1) printf("edi ");
10779     printf("r:");
10780     for(r=0;r<=CCREG;r++) {
10781       //if(((requires_32bit[i]>>r)&(~unneeded_reg[i]>>r))&1) {
10782       if((requires_32bit[i]>>r)&1) {
10783         if(r==CCREG) printf(" CC");
10784         else if(r==HIREG) printf(" HI");
10785         else if(r==LOREG) printf(" LO");
10786         else printf(" r%d",r);
10787       }
10788     }
10789     printf("\n");
10790     /*printf("pr:");
10791     for(r=0;r<=CCREG;r++) {
10792       //if(((requires_32bit[i]>>r)&(~unneeded_reg[i]>>r))&1) {
10793       if((pr32[i]>>r)&1) {
10794         if(r==CCREG) printf(" CC");
10795         else if(r==HIREG) printf(" HI");
10796         else if(r==LOREG) printf(" LO");
10797         else printf(" r%d",r);
10798       }
10799     }
10800     if(pr32[i]!=requires_32bit[i]) printf(" OOPS");
10801     printf("\n");*/
10802     #if defined(__i386__) || defined(__x86_64__)
10803     printf("entry: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7]);
10804     printf("dirty: ");
10805     if(regs[i].wasdirty&1) printf("eax ");
10806     if((regs[i].wasdirty>>1)&1) printf("ecx ");
10807     if((regs[i].wasdirty>>2)&1) printf("edx ");
10808     if((regs[i].wasdirty>>3)&1) printf("ebx ");
10809     if((regs[i].wasdirty>>5)&1) printf("ebp ");
10810     if((regs[i].wasdirty>>6)&1) printf("esi ");
10811     if((regs[i].wasdirty>>7)&1) printf("edi ");
10812     #endif
10813     #ifdef __arm__
10814     printf("entry: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[4],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7],regs[i].regmap_entry[8],regs[i].regmap_entry[9],regs[i].regmap_entry[10],regs[i].regmap_entry[12]);
10815     printf("dirty: ");
10816     if(regs[i].wasdirty&1) printf("r0 ");
10817     if((regs[i].wasdirty>>1)&1) printf("r1 ");
10818     if((regs[i].wasdirty>>2)&1) printf("r2 ");
10819     if((regs[i].wasdirty>>3)&1) printf("r3 ");
10820     if((regs[i].wasdirty>>4)&1) printf("r4 ");
10821     if((regs[i].wasdirty>>5)&1) printf("r5 ");
10822     if((regs[i].wasdirty>>6)&1) printf("r6 ");
10823     if((regs[i].wasdirty>>7)&1) printf("r7 ");
10824     if((regs[i].wasdirty>>8)&1) printf("r8 ");
10825     if((regs[i].wasdirty>>9)&1) printf("r9 ");
10826     if((regs[i].wasdirty>>10)&1) printf("r10 ");
10827     if((regs[i].wasdirty>>12)&1) printf("r12 ");
10828     #endif
10829     printf("\n");
10830     disassemble_inst(i);
10831     //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
10832     #if defined(__i386__) || defined(__x86_64__)
10833     printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7]);
10834     if(regs[i].dirty&1) printf("eax ");
10835     if((regs[i].dirty>>1)&1) printf("ecx ");
10836     if((regs[i].dirty>>2)&1) printf("edx ");
10837     if((regs[i].dirty>>3)&1) printf("ebx ");
10838     if((regs[i].dirty>>5)&1) printf("ebp ");
10839     if((regs[i].dirty>>6)&1) printf("esi ");
10840     if((regs[i].dirty>>7)&1) printf("edi ");
10841     #endif
10842     #ifdef __arm__
10843     printf("r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[4],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7],regs[i].regmap[8],regs[i].regmap[9],regs[i].regmap[10],regs[i].regmap[12]);
10844     if(regs[i].dirty&1) printf("r0 ");
10845     if((regs[i].dirty>>1)&1) printf("r1 ");
10846     if((regs[i].dirty>>2)&1) printf("r2 ");
10847     if((regs[i].dirty>>3)&1) printf("r3 ");
10848     if((regs[i].dirty>>4)&1) printf("r4 ");
10849     if((regs[i].dirty>>5)&1) printf("r5 ");
10850     if((regs[i].dirty>>6)&1) printf("r6 ");
10851     if((regs[i].dirty>>7)&1) printf("r7 ");
10852     if((regs[i].dirty>>8)&1) printf("r8 ");
10853     if((regs[i].dirty>>9)&1) printf("r9 ");
10854     if((regs[i].dirty>>10)&1) printf("r10 ");
10855     if((regs[i].dirty>>12)&1) printf("r12 ");
10856     #endif
10857     printf("\n");
10858     if(regs[i].isconst) {
10859       printf("constants: ");
10860       #if defined(__i386__) || defined(__x86_64__)
10861       if(regs[i].isconst&1) printf("eax=%x ",(int)constmap[i][0]);
10862       if((regs[i].isconst>>1)&1) printf("ecx=%x ",(int)constmap[i][1]);
10863       if((regs[i].isconst>>2)&1) printf("edx=%x ",(int)constmap[i][2]);
10864       if((regs[i].isconst>>3)&1) printf("ebx=%x ",(int)constmap[i][3]);
10865       if((regs[i].isconst>>5)&1) printf("ebp=%x ",(int)constmap[i][5]);
10866       if((regs[i].isconst>>6)&1) printf("esi=%x ",(int)constmap[i][6]);
10867       if((regs[i].isconst>>7)&1) printf("edi=%x ",(int)constmap[i][7]);
10868       #endif
10869       #ifdef __arm__
10870       if(regs[i].isconst&1) printf("r0=%x ",(int)constmap[i][0]);
10871       if((regs[i].isconst>>1)&1) printf("r1=%x ",(int)constmap[i][1]);
10872       if((regs[i].isconst>>2)&1) printf("r2=%x ",(int)constmap[i][2]);
10873       if((regs[i].isconst>>3)&1) printf("r3=%x ",(int)constmap[i][3]);
10874       if((regs[i].isconst>>4)&1) printf("r4=%x ",(int)constmap[i][4]);
10875       if((regs[i].isconst>>5)&1) printf("r5=%x ",(int)constmap[i][5]);
10876       if((regs[i].isconst>>6)&1) printf("r6=%x ",(int)constmap[i][6]);
10877       if((regs[i].isconst>>7)&1) printf("r7=%x ",(int)constmap[i][7]);
10878       if((regs[i].isconst>>8)&1) printf("r8=%x ",(int)constmap[i][8]);
10879       if((regs[i].isconst>>9)&1) printf("r9=%x ",(int)constmap[i][9]);
10880       if((regs[i].isconst>>10)&1) printf("r10=%x ",(int)constmap[i][10]);
10881       if((regs[i].isconst>>12)&1) printf("r12=%x ",(int)constmap[i][12]);
10882       #endif
10883       printf("\n");
10884     }
10885 #ifndef FORCE32
10886     printf(" 32:");
10887     for(r=0;r<=CCREG;r++) {
10888       if((regs[i].is32>>r)&1) {
10889         if(r==CCREG) printf(" CC");
10890         else if(r==HIREG) printf(" HI");
10891         else if(r==LOREG) printf(" LO");
10892         else printf(" r%d",r);
10893       }
10894     }
10895     printf("\n");
10896 #endif
10897     /*printf(" p32:");
10898     for(r=0;r<=CCREG;r++) {
10899       if((p32[i]>>r)&1) {
10900         if(r==CCREG) printf(" CC");
10901         else if(r==HIREG) printf(" HI");
10902         else if(r==LOREG) printf(" LO");
10903         else printf(" r%d",r);
10904       }
10905     }
10906     if(p32[i]!=regs[i].is32) printf(" NO MATCH\n");
10907     else printf("\n");*/
10908     if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
10909       #if defined(__i386__) || defined(__x86_64__)
10910       printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
10911       if(branch_regs[i].dirty&1) printf("eax ");
10912       if((branch_regs[i].dirty>>1)&1) printf("ecx ");
10913       if((branch_regs[i].dirty>>2)&1) printf("edx ");
10914       if((branch_regs[i].dirty>>3)&1) printf("ebx ");
10915       if((branch_regs[i].dirty>>5)&1) printf("ebp ");
10916       if((branch_regs[i].dirty>>6)&1) printf("esi ");
10917       if((branch_regs[i].dirty>>7)&1) printf("edi ");
10918       #endif
10919       #ifdef __arm__
10920       printf("branch(%d): r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[4],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7],branch_regs[i].regmap[8],branch_regs[i].regmap[9],branch_regs[i].regmap[10],branch_regs[i].regmap[12]);
10921       if(branch_regs[i].dirty&1) printf("r0 ");
10922       if((branch_regs[i].dirty>>1)&1) printf("r1 ");
10923       if((branch_regs[i].dirty>>2)&1) printf("r2 ");
10924       if((branch_regs[i].dirty>>3)&1) printf("r3 ");
10925       if((branch_regs[i].dirty>>4)&1) printf("r4 ");
10926       if((branch_regs[i].dirty>>5)&1) printf("r5 ");
10927       if((branch_regs[i].dirty>>6)&1) printf("r6 ");
10928       if((branch_regs[i].dirty>>7)&1) printf("r7 ");
10929       if((branch_regs[i].dirty>>8)&1) printf("r8 ");
10930       if((branch_regs[i].dirty>>9)&1) printf("r9 ");
10931       if((branch_regs[i].dirty>>10)&1) printf("r10 ");
10932       if((branch_regs[i].dirty>>12)&1) printf("r12 ");
10933       #endif
10934 #ifndef FORCE32
10935       printf(" 32:");
10936       for(r=0;r<=CCREG;r++) {
10937         if((branch_regs[i].is32>>r)&1) {
10938           if(r==CCREG) printf(" CC");
10939           else if(r==HIREG) printf(" HI");
10940           else if(r==LOREG) printf(" LO");
10941           else printf(" r%d",r);
10942         }
10943       }
10944       printf("\n");
10945 #endif
10946     }
10947   }
10948
10949   /* Pass 8 - Assembly */
10950   linkcount=0;stubcount=0;
10951   ds=0;is_delayslot=0;
10952   cop1_usable=0;
10953   uint64_t is32_pre=0;
10954   u_int dirty_pre=0;
10955   u_int beginning=(u_int)out;
10956   if((u_int)addr&1) {
10957     ds=1;
10958     pagespan_ds();
10959   }
10960   u_int instr_addr0_override=0;
10961
10962 #ifdef PCSX
10963   if (start == 0x80030000) {
10964     // nasty hack for fastbios thing
10965     instr_addr0_override=(u_int)out;
10966     emit_movimm(start,0);
10967     emit_readword((int)&pcaddr,1);
10968     emit_writeword(0,(int)&pcaddr);
10969     emit_cmp(0,1);
10970     emit_jne((int)new_dyna_leave);
10971   }
10972 #endif
10973   for(i=0;i<slen;i++)
10974   {
10975     //if(ds) printf("ds: ");
10976     if((void*)assem_debug==(void*)printf) disassemble_inst(i);
10977     if(ds) {
10978       ds=0; // Skip delay slot
10979       if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
10980       instr_addr[i]=0;
10981     } else {
10982       #ifndef DESTRUCTIVE_WRITEBACK
10983       if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
10984       {
10985         wb_sx(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,is32_pre,regs[i].was32,
10986               unneeded_reg[i],unneeded_reg_upper[i]);
10987         wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,is32_pre,
10988               unneeded_reg[i],unneeded_reg_upper[i]);
10989       }
10990       if((itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)&&!likely[i]) {
10991         is32_pre=branch_regs[i].is32;
10992         dirty_pre=branch_regs[i].dirty;
10993       }else{
10994         is32_pre=regs[i].is32;
10995         dirty_pre=regs[i].dirty;
10996       }
10997       #endif
10998       // write back
10999       if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
11000       {
11001         wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32,
11002                       unneeded_reg[i],unneeded_reg_upper[i]);
11003         loop_preload(regmap_pre[i],regs[i].regmap_entry);
11004       }
11005       // branch target entry point
11006       instr_addr[i]=(u_int)out;
11007       assem_debug("<->\n");
11008       // load regs
11009       if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
11010         wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32);
11011       load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i],rs2[i]);
11012       address_generation(i,&regs[i],regs[i].regmap_entry);
11013       load_consts(regmap_pre[i],regs[i].regmap,regs[i].was32,i);
11014       if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
11015       {
11016         // Load the delay slot registers if necessary
11017         if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i]&&(rs1[i+1]!=rt1[i]||rt1[i]==0))
11018           load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
11019         if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i]&&(rs2[i+1]!=rt1[i]||rt1[i]==0))
11020           load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
11021         if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a)
11022           load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
11023       }
11024       else if(i+1<slen)
11025       {
11026         // Preload registers for following instruction
11027         if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i])
11028           if(rs1[i+1]!=rt1[i]&&rs1[i+1]!=rt2[i])
11029             load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
11030         if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i])
11031           if(rs2[i+1]!=rt1[i]&&rs2[i+1]!=rt2[i])
11032             load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
11033       }
11034       // TODO: if(is_ooo(i)) address_generation(i+1);
11035       if(itype[i]==CJUMP||itype[i]==FJUMP)
11036         load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
11037       if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a)
11038         load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
11039       if(bt[i]) cop1_usable=0;
11040       // assemble
11041       switch(itype[i]) {
11042         case ALU:
11043           alu_assemble(i,&regs[i]);break;
11044         case IMM16:
11045           imm16_assemble(i,&regs[i]);break;
11046         case SHIFT:
11047           shift_assemble(i,&regs[i]);break;
11048         case SHIFTIMM:
11049           shiftimm_assemble(i,&regs[i]);break;
11050         case LOAD:
11051           load_assemble(i,&regs[i]);break;
11052         case LOADLR:
11053           loadlr_assemble(i,&regs[i]);break;
11054         case STORE:
11055           store_assemble(i,&regs[i]);break;
11056         case STORELR:
11057           storelr_assemble(i,&regs[i]);break;
11058         case COP0:
11059           cop0_assemble(i,&regs[i]);break;
11060         case COP1:
11061           cop1_assemble(i,&regs[i]);break;
11062         case C1LS:
11063           c1ls_assemble(i,&regs[i]);break;
11064         case COP2:
11065           cop2_assemble(i,&regs[i]);break;
11066         case C2LS:
11067           c2ls_assemble(i,&regs[i]);break;
11068         case C2OP:
11069           c2op_assemble(i,&regs[i]);break;
11070         case FCONV:
11071           fconv_assemble(i,&regs[i]);break;
11072         case FLOAT:
11073           float_assemble(i,&regs[i]);break;
11074         case FCOMP:
11075           fcomp_assemble(i,&regs[i]);break;
11076         case MULTDIV:
11077           multdiv_assemble(i,&regs[i]);break;
11078         case MOV:
11079           mov_assemble(i,&regs[i]);break;
11080         case SYSCALL:
11081           syscall_assemble(i,&regs[i]);break;
11082         case HLECALL:
11083           hlecall_assemble(i,&regs[i]);break;
11084         case INTCALL:
11085           intcall_assemble(i,&regs[i]);break;
11086         case UJUMP:
11087           ujump_assemble(i,&regs[i]);ds=1;break;
11088         case RJUMP:
11089           rjump_assemble(i,&regs[i]);ds=1;break;
11090         case CJUMP:
11091           cjump_assemble(i,&regs[i]);ds=1;break;
11092         case SJUMP:
11093           sjump_assemble(i,&regs[i]);ds=1;break;
11094         case FJUMP:
11095           fjump_assemble(i,&regs[i]);ds=1;break;
11096         case SPAN:
11097           pagespan_assemble(i,&regs[i]);break;
11098       }
11099       if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
11100         literal_pool(1024);
11101       else
11102         literal_pool_jumpover(256);
11103     }
11104   }
11105   //assert(itype[i-2]==UJUMP||itype[i-2]==RJUMP||(source[i-2]>>16)==0x1000);
11106   // If the block did not end with an unconditional branch,
11107   // add a jump to the next instruction.
11108   if(i>1) {
11109     if(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000&&itype[i-1]!=SPAN) {
11110       assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP);
11111       assert(i==slen);
11112       if(itype[i-2]!=CJUMP&&itype[i-2]!=SJUMP&&itype[i-2]!=FJUMP) {
11113         store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
11114         if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
11115           emit_loadreg(CCREG,HOST_CCREG);
11116         emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i-1]+1),HOST_CCREG);
11117       }
11118       else if(!likely[i-2])
11119       {
11120         store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].is32,branch_regs[i-2].dirty,start+i*4);
11121         assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
11122       }
11123       else
11124       {
11125         store_regs_bt(regs[i-2].regmap,regs[i-2].is32,regs[i-2].dirty,start+i*4);
11126         assert(regs[i-2].regmap[HOST_CCREG]==CCREG);
11127       }
11128       add_to_linker((int)out,start+i*4,0);
11129       emit_jmp(0);
11130     }
11131   }
11132   else
11133   {
11134     assert(i>0);
11135     assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP);
11136     store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
11137     if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
11138       emit_loadreg(CCREG,HOST_CCREG);
11139     emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i-1]+1),HOST_CCREG);
11140     add_to_linker((int)out,start+i*4,0);
11141     emit_jmp(0);
11142   }
11143
11144   // TODO: delay slot stubs?
11145   // Stubs
11146   for(i=0;i<stubcount;i++)
11147   {
11148     switch(stubs[i][0])
11149     {
11150       case LOADB_STUB:
11151       case LOADH_STUB:
11152       case LOADW_STUB:
11153       case LOADD_STUB:
11154       case LOADBU_STUB:
11155       case LOADHU_STUB:
11156         do_readstub(i);break;
11157       case STOREB_STUB:
11158       case STOREH_STUB:
11159       case STOREW_STUB:
11160       case STORED_STUB:
11161         do_writestub(i);break;
11162       case CC_STUB:
11163         do_ccstub(i);break;
11164       case INVCODE_STUB:
11165         do_invstub(i);break;
11166       case FP_STUB:
11167         do_cop1stub(i);break;
11168       case STORELR_STUB:
11169         do_unalignedwritestub(i);break;
11170     }
11171   }
11172
11173   if (instr_addr0_override)
11174     instr_addr[0] = instr_addr0_override;
11175
11176   /* Pass 9 - Linker */
11177   for(i=0;i<linkcount;i++)
11178   {
11179     assem_debug("%8x -> %8x\n",link_addr[i][0],link_addr[i][1]);
11180     literal_pool(64);
11181     if(!link_addr[i][2])
11182     {
11183       void *stub=out;
11184       void *addr=check_addr(link_addr[i][1]);
11185       emit_extjump(link_addr[i][0],link_addr[i][1]);
11186       if(addr) {
11187         set_jump_target(link_addr[i][0],(int)addr);
11188         add_link(link_addr[i][1],stub);
11189       }
11190       else set_jump_target(link_addr[i][0],(int)stub);
11191     }
11192     else
11193     {
11194       // Internal branch
11195       int target=(link_addr[i][1]-start)>>2;
11196       assert(target>=0&&target<slen);
11197       assert(instr_addr[target]);
11198       //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
11199       //set_jump_target_fillslot(link_addr[i][0],instr_addr[target],link_addr[i][2]>>1);
11200       //#else
11201       set_jump_target(link_addr[i][0],instr_addr[target]);
11202       //#endif
11203     }
11204   }
11205   // External Branch Targets (jump_in)
11206   if(copy+slen*4>(void *)shadow+sizeof(shadow)) copy=shadow;
11207   for(i=0;i<slen;i++)
11208   {
11209     if(bt[i]||i==0)
11210     {
11211       if(instr_addr[i]) // TODO - delay slots (=null)
11212       {
11213         u_int vaddr=start+i*4;
11214         u_int page=get_page(vaddr);
11215         u_int vpage=get_vpage(vaddr);
11216         literal_pool(256);
11217         //if(!(is32[i]&(~unneeded_reg_upper[i])&~(1LL<<CCREG)))
11218 #ifndef FORCE32
11219         if(!requires_32bit[i])
11220 #else
11221         if(1)
11222 #endif
11223         {
11224           assem_debug("%8x (%d) <- %8x\n",instr_addr[i],i,start+i*4);
11225           assem_debug("jump_in: %x\n",start+i*4);
11226           ll_add(jump_dirty+vpage,vaddr,(void *)out);
11227           int entry_point=do_dirty_stub(i);
11228           ll_add(jump_in+page,vaddr,(void *)entry_point);
11229           // If there was an existing entry in the hash table,
11230           // replace it with the new address.
11231           // Don't add new entries.  We'll insert the
11232           // ones that actually get used in check_addr().
11233           int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
11234           if(ht_bin[0]==vaddr) {
11235             ht_bin[1]=entry_point;
11236           }
11237           if(ht_bin[2]==vaddr) {
11238             ht_bin[3]=entry_point;
11239           }
11240         }
11241         else
11242         {
11243           u_int r=requires_32bit[i]|!!(requires_32bit[i]>>32);
11244           assem_debug("%8x (%d) <- %8x\n",instr_addr[i],i,start+i*4);
11245           assem_debug("jump_in: %x (restricted - %x)\n",start+i*4,r);
11246           //int entry_point=(int)out;
11247           ////assem_debug("entry_point: %x\n",entry_point);
11248           //load_regs_entry(i);
11249           //if(entry_point==(int)out)
11250           //  entry_point=instr_addr[i];
11251           //else
11252           //  emit_jmp(instr_addr[i]);
11253           //ll_add_32(jump_in+page,vaddr,r,(void *)entry_point);
11254           ll_add_32(jump_dirty+vpage,vaddr,r,(void *)out);
11255           int entry_point=do_dirty_stub(i);
11256           ll_add_32(jump_in+page,vaddr,r,(void *)entry_point);
11257         }
11258       }
11259     }
11260   }
11261   // Write out the literal pool if necessary
11262   literal_pool(0);
11263   #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
11264   // Align code
11265   if(((u_int)out)&7) emit_addnop(13);
11266   #endif
11267   assert((u_int)out-beginning<MAX_OUTPUT_BLOCK_SIZE);
11268   //printf("shadow buffer: %x-%x\n",(int)copy,(int)copy+slen*4);
11269   memcpy(copy,source,slen*4);
11270   copy+=slen*4;
11271
11272   #ifdef __arm__
11273   __clear_cache((void *)beginning,out);
11274   #endif
11275
11276   // If we're within 256K of the end of the buffer,
11277   // start over from the beginning. (Is 256K enough?)
11278   if((int)out>BASE_ADDR+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE) out=(u_char *)BASE_ADDR;
11279
11280   // Trap writes to any of the pages we compiled
11281   for(i=start>>12;i<=(start+slen*4)>>12;i++) {
11282     invalid_code[i]=0;
11283 #ifndef DISABLE_TLB
11284     memory_map[i]|=0x40000000;
11285     if((signed int)start>=(signed int)0xC0000000) {
11286       assert(using_tlb);
11287       j=(((u_int)i<<12)+(memory_map[i]<<2)-(u_int)rdram+(u_int)0x80000000)>>12;
11288       invalid_code[j]=0;
11289       memory_map[j]|=0x40000000;
11290       //printf("write protect physical page: %x (virtual %x)\n",j<<12,start);
11291     }
11292 #endif
11293   }
11294 #ifdef PCSX
11295   // PCSX maps all RAM mirror invalid_code tests to 0x80000000..0x80000000+RAM_SIZE
11296   if(get_page(start)<(RAM_SIZE>>12))
11297     for(i=start>>12;i<=(start+slen*4)>>12;i++)
11298       invalid_code[((u_int)0x80000000>>12)|i]=0;
11299 #endif
11300
11301   /* Pass 10 - Free memory by expiring oldest blocks */
11302
11303   int end=((((int)out-BASE_ADDR)>>(TARGET_SIZE_2-16))+16384)&65535;
11304   while(expirep!=end)
11305   {
11306     int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
11307     int base=BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
11308     inv_debug("EXP: Phase %d\n",expirep);
11309     switch((expirep>>11)&3)
11310     {
11311       case 0:
11312         // Clear jump_in and jump_dirty
11313         ll_remove_matching_addrs(jump_in+(expirep&2047),base,shift);
11314         ll_remove_matching_addrs(jump_dirty+(expirep&2047),base,shift);
11315         ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base,shift);
11316         ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base,shift);
11317         break;
11318       case 1:
11319         // Clear pointers
11320         ll_kill_pointers(jump_out[expirep&2047],base,shift);
11321         ll_kill_pointers(jump_out[(expirep&2047)+2048],base,shift);
11322         break;
11323       case 2:
11324         // Clear hash table
11325         for(i=0;i<32;i++) {
11326           int *ht_bin=hash_table[((expirep&2047)<<5)+i];
11327           if((ht_bin[3]>>shift)==(base>>shift) ||
11328              ((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
11329             inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[2],ht_bin[3]);
11330             ht_bin[2]=ht_bin[3]=-1;
11331           }
11332           if((ht_bin[1]>>shift)==(base>>shift) ||
11333              ((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
11334             inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[0],ht_bin[1]);
11335             ht_bin[0]=ht_bin[2];
11336             ht_bin[1]=ht_bin[3];
11337             ht_bin[2]=ht_bin[3]=-1;
11338           }
11339         }
11340         break;
11341       case 3:
11342         // Clear jump_out
11343         #ifdef __arm__
11344         if((expirep&2047)==0)
11345           do_clear_cache();
11346         #endif
11347         ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
11348         ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
11349         break;
11350     }
11351     expirep=(expirep+1)&65535;
11352   }
11353   return 0;
11354 }
11355
11356 // vim:shiftwidth=2:expandtab