libpcsxcore/new_dynarec/new_dynarec.c

   1 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
   2  *   Mupen64plus - new_dynarec.c                                           *
   3  *   Copyright (C) 2009-2010 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   u_int using_tlb;
 125   u_int stop_after_jal;
 126   extern u_char restore_candidate[512];
 127   extern int cycle_count;
 128
 129   /* registers that may be allocated */
 130   /* 1-31 gpr */
 131 #define HIREG 32 // hi
 132 #define LOREG 33 // lo
 133 #define FSREG 34 // FPU status (FCSR)
 134 #define CSREG 35 // Coprocessor status
 135 #define CCREG 36 // Cycle count
 136 #define INVCP 37 // Pointer to invalid_code
 137 #define MMREG 38 // Pointer to memory_map
 138 #define ROREG 39 // ram offset (if rdram!=0x80000000)
 139 #define TEMPREG 40
 140 #define FTEMP 40 // FPU temporary register
 141 #define PTEMP 41 // Prefetch temporary register
 142 #define TLREG 42 // TLB mapping offset
 143 #define RHASH 43 // Return address hash
 144 #define RHTBL 44 // Return address hash table address
 145 #define RTEMP 45 // JR/JALR address register
 146 #define MAXREG 45
 147 #define AGEN1 46 // Address generation temporary register
 148 #define AGEN2 47 // Address generation temporary register
 149 #define MGEN1 48 // Maptable address generation temporary register
 150 #define MGEN2 49 // Maptable address generation temporary register
 151 #define BTREG 50 // Branch target temporary register
 152
 153   /* instruction types */
 154 #define NOP 0     // No operation
 155 #define LOAD 1    // Load
 156 #define STORE 2   // Store
 157 #define LOADLR 3  // Unaligned load
 158 #define STORELR 4 // Unaligned store
 159 #define MOV 5     // Move
 160 #define ALU 6     // Arithmetic/logic
 161 #define MULTDIV 7 // Multiply/divide
 162 #define SHIFT 8   // Shift by register
 163 #define SHIFTIMM 9// Shift by immediate
 164 #define IMM16 10  // 16-bit immediate
 165 #define RJUMP 11  // Unconditional jump to register
 166 #define UJUMP 12  // Unconditional jump
 167 #define CJUMP 13  // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
 168 #define SJUMP 14  // Conditional branch (regimm format)
 169 #define COP0 15   // Coprocessor 0
 170 #define COP1 16   // Coprocessor 1
 171 #define C1LS 17   // Coprocessor 1 load/store
 172 #define FJUMP 18  // Conditional branch (floating point)
 173 #define FLOAT 19  // Floating point unit
 174 #define FCONV 20  // Convert integer to float
 175 #define FCOMP 21  // Floating point compare (sets FSREG)
 176 #define SYSCALL 22// SYSCALL
 177 #define OTHER 23  // Other
 178 #define SPAN 24   // Branch/delay slot spans 2 pages
 179 #define NI 25     // Not implemented
 180 #define HLECALL 26// PCSX fake opcodes for HLE
 181 #define COP2 27   // Coprocessor 2 move
 182 #define C2LS 28   // Coprocessor 2 load/store
 183 #define C2OP 29   // Coprocessor 2 operation
 184 #define INTCALL 30// Call interpreter to handle rare corner cases
 185
 186   /* stubs */
 187 #define CC_STUB 1
 188 #define FP_STUB 2
 189 #define LOADB_STUB 3
 190 #define LOADH_STUB 4
 191 #define LOADW_STUB 5
 192 #define LOADD_STUB 6
 193 #define LOADBU_STUB 7
 194 #define LOADHU_STUB 8
 195 #define STOREB_STUB 9
 196 #define STOREH_STUB 10
 197 #define STOREW_STUB 11
 198 #define STORED_STUB 12
 199 #define STORELR_STUB 13
 200 #define INVCODE_STUB 14
 201
 202   /* branch codes */
 203 #define TAKEN 1
 204 #define NOTTAKEN 2
 205 #define NULLDS 3
 206
 207 // asm linkage
 208 int new_recompile_block(int addr);
 209 void *get_addr_ht(u_int vaddr);
 210 void invalidate_block(u_int block);
 211 void invalidate_addr(u_int addr);
 212 void remove_hash(int vaddr);
 213 void jump_vaddr();
 214 void dyna_linker();
 215 void dyna_linker_ds();
 216 void verify_code();
 217 void verify_code_vm();
 218 void verify_code_ds();
 219 void cc_interrupt();
 220 void fp_exception();
 221 void fp_exception_ds();
 222 void jump_syscall();
 223 void jump_syscall_hle();
 224 void jump_eret();
 225 void jump_hlecall();
 226 void jump_intcall();
 227 void new_dyna_leave();
 228
 229 // TLB
 230 void TLBWI_new();
 231 void TLBWR_new();
 232 void read_nomem_new();
 233 void read_nomemb_new();
 234 void read_nomemh_new();
 235 void read_nomemd_new();
 236 void write_nomem_new();
 237 void write_nomemb_new();
 238 void write_nomemh_new();
 239 void write_nomemd_new();
 240 void write_rdram_new();
 241 void write_rdramb_new();
 242 void write_rdramh_new();
 243 void write_rdramd_new();
 244 extern u_int memory_map[1048576];
 245
 246 // Needed by assembler
 247 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32);
 248 void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty);
 249 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr);
 250 void load_all_regs(signed char i_regmap[]);
 251 void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
 252 void load_regs_entry(int t);
 253 void load_all_consts(signed char regmap[],int is32,u_int dirty,int i);
 254
 255 int tracedebug=0;
 256
 257 //#define DEBUG_CYCLE_COUNT 1
 258
 259 void nullf() {}
 260 //#define assem_debug printf
 261 //#define inv_debug printf
 262 #define assem_debug nullf
 263 #define inv_debug nullf
 264
 265 static void tlb_hacks()
 266 {
 267 #ifndef DISABLE_TLB
 268   // Goldeneye hack
 269   if (strncmp((char *) ROM_HEADER->nom, "GOLDENEYE",9) == 0)
 270   {
 271     u_int addr;
 272     int n;
 273     switch (ROM_HEADER->Country_code&0xFF)
 274     {
 275       case 0x45: // U
 276         addr=0x34b30;
 277         break;
 278       case 0x4A: // J
 279         addr=0x34b70;
 280         break;
 281       case 0x50: // E
 282         addr=0x329f0;
 283         break;
 284       default:
 285         // Unknown country code
 286         addr=0;
 287         break;
 288     }
 289     u_int rom_addr=(u_int)rom;
 290     #ifdef ROM_COPY
 291     // Since memory_map is 32-bit, on 64-bit systems the rom needs to be
 292     // in the lower 4G of memory to use this hack.  Copy it if necessary.
 293     if((void *)rom>(void *)0xffffffff) {
 294       munmap(ROM_COPY, 67108864);
 295       if(mmap(ROM_COPY, 12582912,
 296               PROT_READ | PROT_WRITE,
 297               MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
 298               -1, 0) <= 0) {printf("mmap() failed\n");}
 299       memcpy(ROM_COPY,rom,12582912);
 300       rom_addr=(u_int)ROM_COPY;
 301     }
 302     #endif
 303     if(addr) {
 304       for(n=0x7F000;n<0x80000;n++) {
 305         memory_map[n]=(((u_int)(rom_addr+addr-0x7F000000))>>2)|0x40000000;
 306       }
 307     }
 308   }
 309 #endif
 310 }
 311
 312 static u_int get_page(u_int vaddr)
 313 {
 314 #ifndef PCSX
 315   u_int page=(vaddr^0x80000000)>>12;
 316 #else
 317   u_int page=vaddr&~0xe0000000;
 318   if (page < 0x1000000)
 319     page &= ~0x0e00000; // RAM mirrors
 320   page>>=12;
 321 #endif
 322 #ifndef DISABLE_TLB
 323   if(page>262143&&tlb_LUT_r[vaddr>>12]) page=(tlb_LUT_r[vaddr>>12]^0x80000000)>>12;
 324 #endif
 325   if(page>2048) page=2048+(page&2047);
 326   return page;
 327 }
 328
 329 static u_int get_vpage(u_int vaddr)
 330 {
 331   u_int vpage=(vaddr^0x80000000)>>12;
 332 #ifndef DISABLE_TLB
 333   if(vpage>262143&&tlb_LUT_r[vaddr>>12]) vpage&=2047; // jump_dirty uses a hash of the virtual address instead
 334 #endif
 335   if(vpage>2048) vpage=2048+(vpage&2047);
 336   return vpage;
 337 }
 338
 339 // Get address from virtual address
 340 // This is called from the recompiled JR/JALR instructions
 341 void *get_addr(u_int vaddr)
 342 {
 343   u_int page=get_page(vaddr);
 344   u_int vpage=get_vpage(vaddr);
 345   struct ll_entry *head;
 346   //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
 347   head=jump_in[page];
 348   while(head!=NULL) {
 349     if(head->vaddr==vaddr&&head->reg32==0) {
 350   //printf("TRACE: count=%d next=%d (get_addr match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 351       int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 352       ht_bin[3]=ht_bin[1];
 353       ht_bin[2]=ht_bin[0];
 354       ht_bin[1]=(int)head->addr;
 355       ht_bin[0]=vaddr;
 356       return head->addr;
 357     }
 358     head=head->next;
 359   }
 360   head=jump_dirty[vpage];
 361   while(head!=NULL) {
 362     if(head->vaddr==vaddr&&head->reg32==0) {
 363       //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 364       // Don't restore blocks which are about to expire from the cache
 365       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
 366       if(verify_dirty(head->addr)) {
 367         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
 368         invalid_code[vaddr>>12]=0;
 369         memory_map[vaddr>>12]|=0x40000000;
 370         if(vpage<2048) {
 371 #ifndef DISABLE_TLB
 372           if(tlb_LUT_r[vaddr>>12]) {
 373             invalid_code[tlb_LUT_r[vaddr>>12]>>12]=0;
 374             memory_map[tlb_LUT_r[vaddr>>12]>>12]|=0x40000000;
 375           }
 376 #endif
 377           restore_candidate[vpage>>3]|=1<<(vpage&7);
 378         }
 379         else restore_candidate[page>>3]|=1<<(page&7);
 380         int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 381         if(ht_bin[0]==vaddr) {
 382           ht_bin[1]=(int)head->addr; // Replace existing entry
 383         }
 384         else
 385         {
 386           ht_bin[3]=ht_bin[1];
 387           ht_bin[2]=ht_bin[0];
 388           ht_bin[1]=(int)head->addr;
 389           ht_bin[0]=vaddr;
 390         }
 391         return head->addr;
 392       }
 393     }
 394     head=head->next;
 395   }
 396   //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
 397   int r=new_recompile_block(vaddr);
 398   if(r==0) return get_addr(vaddr);
 399   // Execute in unmapped page, generate pagefault execption
 400   Status|=2;
 401   Cause=(vaddr<<31)|0x8;
 402   EPC=(vaddr&1)?vaddr-5:vaddr;
 403   BadVAddr=(vaddr&~1);
 404   Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
 405   EntryHi=BadVAddr&0xFFFFE000;
 406   return get_addr_ht(0x80000000);
 407 }
 408 // Look up address in hash table first
 409 void *get_addr_ht(u_int vaddr)
 410 {
 411   //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
 412   int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 413   if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
 414   if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
 415   return get_addr(vaddr);
 416 }
 417
 418 void *get_addr_32(u_int vaddr,u_int flags)
 419 {
 420 #ifdef FORCE32
 421   return get_addr(vaddr);
 422 #else
 423   //printf("TRACE: count=%d next=%d (get_addr_32 %x,flags %x)\n",Count,next_interupt,vaddr,flags);
 424   int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 425   if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
 426   if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
 427   u_int page=get_page(vaddr);
 428   u_int vpage=get_vpage(vaddr);
 429   struct ll_entry *head;
 430   head=jump_in[page];
 431   while(head!=NULL) {
 432     if(head->vaddr==vaddr&&(head->reg32&flags)==0) {
 433       //printf("TRACE: count=%d next=%d (get_addr_32 match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 434       if(head->reg32==0) {
 435         int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 436         if(ht_bin[0]==-1) {
 437           ht_bin[1]=(int)head->addr;
 438           ht_bin[0]=vaddr;
 439         }else if(ht_bin[2]==-1) {
 440           ht_bin[3]=(int)head->addr;
 441           ht_bin[2]=vaddr;
 442         }
 443         //ht_bin[3]=ht_bin[1];
 444         //ht_bin[2]=ht_bin[0];
 445         //ht_bin[1]=(int)head->addr;
 446         //ht_bin[0]=vaddr;
 447       }
 448       return head->addr;
 449     }
 450     head=head->next;
 451   }
 452   head=jump_dirty[vpage];
 453   while(head!=NULL) {
 454     if(head->vaddr==vaddr&&(head->reg32&flags)==0) {
 455       //printf("TRACE: count=%d next=%d (get_addr_32 match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
 456       // Don't restore blocks which are about to expire from the cache
 457       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
 458       if(verify_dirty(head->addr)) {
 459         //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
 460         invalid_code[vaddr>>12]=0;
 461         memory_map[vaddr>>12]|=0x40000000;
 462         if(vpage<2048) {
 463 #ifndef DISABLE_TLB
 464           if(tlb_LUT_r[vaddr>>12]) {
 465             invalid_code[tlb_LUT_r[vaddr>>12]>>12]=0;
 466             memory_map[tlb_LUT_r[vaddr>>12]>>12]|=0x40000000;
 467           }
 468 #endif
 469           restore_candidate[vpage>>3]|=1<<(vpage&7);
 470         }
 471         else restore_candidate[page>>3]|=1<<(page&7);
 472         if(head->reg32==0) {
 473           int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 474           if(ht_bin[0]==-1) {
 475             ht_bin[1]=(int)head->addr;
 476             ht_bin[0]=vaddr;
 477           }else if(ht_bin[2]==-1) {
 478             ht_bin[3]=(int)head->addr;
 479             ht_bin[2]=vaddr;
 480           }
 481           //ht_bin[3]=ht_bin[1];
 482           //ht_bin[2]=ht_bin[0];
 483           //ht_bin[1]=(int)head->addr;
 484           //ht_bin[0]=vaddr;
 485         }
 486         return head->addr;
 487       }
 488     }
 489     head=head->next;
 490   }
 491   //printf("TRACE: count=%d next=%d (get_addr_32 no-match %x,flags %x)\n",Count,next_interupt,vaddr,flags);
 492   int r=new_recompile_block(vaddr);
 493   if(r==0) return get_addr(vaddr);
 494   // Execute in unmapped page, generate pagefault execption
 495   Status|=2;
 496   Cause=(vaddr<<31)|0x8;
 497   EPC=(vaddr&1)?vaddr-5:vaddr;
 498   BadVAddr=(vaddr&~1);
 499   Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
 500   EntryHi=BadVAddr&0xFFFFE000;
 501   return get_addr_ht(0x80000000);
 502 #endif
 503 }
 504
 505 void clear_all_regs(signed char regmap[])
 506 {
 507   int hr;
 508   for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
 509 }
 510
 511 signed char get_reg(signed char regmap[],int r)
 512 {
 513   int hr;
 514   for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap[hr]==r) return hr;
 515   return -1;
 516 }
 517
 518 // Find a register that is available for two consecutive cycles
 519 signed char get_reg2(signed char regmap1[],signed char regmap2[],int r)
 520 {
 521   int hr;
 522   for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
 523   return -1;
 524 }
 525
 526 int count_free_regs(signed char regmap[])
 527 {
 528   int count=0;
 529   int hr;
 530   for(hr=0;hr<HOST_REGS;hr++)
 531   {
 532     if(hr!=EXCLUDE_REG) {
 533       if(regmap[hr]<0) count++;
 534     }
 535   }
 536   return count;
 537 }
 538
 539 void dirty_reg(struct regstat *cur,signed char reg)
 540 {
 541   int hr;
 542   if(!reg) return;
 543   for (hr=0;hr<HOST_REGS;hr++) {
 544     if((cur->regmap[hr]&63)==reg) {
 545       cur->dirty|=1<<hr;
 546     }
 547   }
 548 }
 549
 550 // If we dirty the lower half of a 64 bit register which is now being
 551 // sign-extended, we need to dump the upper half.
 552 // Note: Do this only after completion of the instruction, because
 553 // some instructions may need to read the full 64-bit value even if
 554 // overwriting it (eg SLTI, DSRA32).
 555 static void flush_dirty_uppers(struct regstat *cur)
 556 {
 557   int hr,reg;
 558   for (hr=0;hr<HOST_REGS;hr++) {
 559     if((cur->dirty>>hr)&1) {
 560       reg=cur->regmap[hr];
 561       if(reg>=64)
 562         if((cur->is32>>(reg&63))&1) cur->regmap[hr]=-1;
 563     }
 564   }
 565 }
 566
 567 void set_const(struct regstat *cur,signed char reg,uint64_t value)
 568 {
 569   int hr;
 570   if(!reg) return;
 571   for (hr=0;hr<HOST_REGS;hr++) {
 572     if(cur->regmap[hr]==reg) {
 573       cur->isconst|=1<<hr;
 574       cur->constmap[hr]=value;
 575     }
 576     else if((cur->regmap[hr]^64)==reg) {
 577       cur->isconst|=1<<hr;
 578       cur->constmap[hr]=value>>32;
 579     }
 580   }
 581 }
 582
 583 void clear_const(struct regstat *cur,signed char reg)
 584 {
 585   int hr;
 586   if(!reg) return;
 587   for (hr=0;hr<HOST_REGS;hr++) {
 588     if((cur->regmap[hr]&63)==reg) {
 589       cur->isconst&=~(1<<hr);
 590     }
 591   }
 592 }
 593
 594 int is_const(struct regstat *cur,signed char reg)
 595 {
 596   int hr;
 597   if(!reg) return 1;
 598   for (hr=0;hr<HOST_REGS;hr++) {
 599     if((cur->regmap[hr]&63)==reg) {
 600       return (cur->isconst>>hr)&1;
 601     }
 602   }
 603   return 0;
 604 }
 605 uint64_t get_const(struct regstat *cur,signed char reg)
 606 {
 607   int hr;
 608   if(!reg) return 0;
 609   for (hr=0;hr<HOST_REGS;hr++) {
 610     if(cur->regmap[hr]==reg) {
 611       return cur->constmap[hr];
 612     }
 613   }
 614   printf("Unknown constant in r%d\n",reg);
 615   exit(1);
 616 }
 617
 618 // Least soon needed registers
 619 // Look at the next ten instructions and see which registers
 620 // will be used.  Try not to reallocate these.
 621 void lsn(u_char hsn[], int i, int *preferred_reg)
 622 {
 623   int j;
 624   int b=-1;
 625   for(j=0;j<9;j++)
 626   {
 627     if(i+j>=slen) {
 628       j=slen-i-1;
 629       break;
 630     }
 631     if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
 632     {
 633       // Don't go past an unconditonal jump
 634       j++;
 635       break;
 636     }
 637   }
 638   for(;j>=0;j--)
 639   {
 640     if(rs1[i+j]) hsn[rs1[i+j]]=j;
 641     if(rs2[i+j]) hsn[rs2[i+j]]=j;
 642     if(rt1[i+j]) hsn[rt1[i+j]]=j;
 643     if(rt2[i+j]) hsn[rt2[i+j]]=j;
 644     if(itype[i+j]==STORE || itype[i+j]==STORELR) {
 645       // Stores can allocate zero
 646       hsn[rs1[i+j]]=j;
 647       hsn[rs2[i+j]]=j;
 648     }
 649     // On some architectures stores need invc_ptr
 650     #if defined(HOST_IMM8)
 651     if(itype[i+j]==STORE || itype[i+j]==STORELR || (opcode[i+j]&0x3b)==0x39 || (opcode[i+j]&0x3b)==0x3a) {
 652       hsn[INVCP]=j;
 653     }
 654     #endif
 655     if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
 656     {
 657       hsn[CCREG]=j;
 658       b=j;
 659     }
 660   }
 661   if(b>=0)
 662   {
 663     if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
 664     {
 665       // Follow first branch
 666       int t=(ba[i+b]-start)>>2;
 667       j=7-b;if(t+j>=slen) j=slen-t-1;
 668       for(;j>=0;j--)
 669       {
 670         if(rs1[t+j]) if(hsn[rs1[t+j]]>j+b+2) hsn[rs1[t+j]]=j+b+2;
 671         if(rs2[t+j]) if(hsn[rs2[t+j]]>j+b+2) hsn[rs2[t+j]]=j+b+2;
 672         //if(rt1[t+j]) if(hsn[rt1[t+j]]>j+b+2) hsn[rt1[t+j]]=j+b+2;
 673         //if(rt2[t+j]) if(hsn[rt2[t+j]]>j+b+2) hsn[rt2[t+j]]=j+b+2;
 674       }
 675     }
 676     // TODO: preferred register based on backward branch
 677   }
 678   // Delay slot should preferably not overwrite branch conditions or cycle count
 679   if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
 680     if(rs1[i-1]) if(hsn[rs1[i-1]]>1) hsn[rs1[i-1]]=1;
 681     if(rs2[i-1]) if(hsn[rs2[i-1]]>1) hsn[rs2[i-1]]=1;
 682     hsn[CCREG]=1;
 683     // ...or hash tables
 684     hsn[RHASH]=1;
 685     hsn[RHTBL]=1;
 686   }
 687   // Coprocessor load/store needs FTEMP, even if not declared
 688   if(itype[i]==C1LS||itype[i]==C2LS) {
 689     hsn[FTEMP]=0;
 690   }
 691   // Load L/R also uses FTEMP as a temporary register
 692   if(itype[i]==LOADLR) {
 693     hsn[FTEMP]=0;
 694   }
 695   // Also SWL/SWR/SDL/SDR
 696   if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) {
 697     hsn[FTEMP]=0;
 698   }
 699   // Don't remove the TLB registers either
 700   if(itype[i]==LOAD || itype[i]==LOADLR || itype[i]==STORE || itype[i]==STORELR || itype[i]==C1LS || itype[i]==C2LS) {
 701     hsn[TLREG]=0;
 702   }
 703   // Don't remove the miniht registers
 704   if(itype[i]==UJUMP||itype[i]==RJUMP)
 705   {
 706     hsn[RHASH]=0;
 707     hsn[RHTBL]=0;
 708   }
 709 }
 710
 711 // We only want to allocate registers if we're going to use them again soon
 712 int needed_again(int r, int i)
 713 {
 714   int j;
 715   int b=-1;
 716   int rn=10;
 717   int hr;
 718   u_char hsn[MAXREG+1];
 719   int preferred_reg;
 720
 721   memset(hsn,10,sizeof(hsn));
 722   lsn(hsn,i,&preferred_reg);
 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   for(hr=0;hr<HOST_REGS;hr++) {
 776     if(hr!=EXCLUDE_REG) {
 777       if(rn<hsn[hr]) return 1;
 778     }
 779   }
 780   return 0;
 781 }
 782
 783 // Try to match register allocations at the end of a loop with those
 784 // at the beginning
 785 int loop_reg(int i, int r, int hr)
 786 {
 787   int j,k;
 788   for(j=0;j<9;j++)
 789   {
 790     if(i+j>=slen) {
 791       j=slen-i-1;
 792       break;
 793     }
 794     if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
 795     {
 796       // Don't go past an unconditonal jump
 797       j++;
 798       break;
 799     }
 800   }
 801   k=0;
 802   if(i>0){
 803     if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)
 804       k--;
 805   }
 806   for(;k<j;k++)
 807   {
 808     if(r<64&&((unneeded_reg[i+k]>>r)&1)) return hr;
 809     if(r>64&&((unneeded_reg_upper[i+k]>>r)&1)) return hr;
 810     if(i+k>=0&&(itype[i+k]==UJUMP||itype[i+k]==CJUMP||itype[i+k]==SJUMP||itype[i+k]==FJUMP))
 811     {
 812       if(ba[i+k]>=start && ba[i+k]<(start+i*4))
 813       {
 814         int t=(ba[i+k]-start)>>2;
 815         int reg=get_reg(regs[t].regmap_entry,r);
 816         if(reg>=0) return reg;
 817         //reg=get_reg(regs[t+1].regmap_entry,r);
 818         //if(reg>=0) return reg;
 819       }
 820     }
 821   }
 822   return hr;
 823 }
 824
 825
 826 // Allocate every register, preserving source/target regs
 827 void alloc_all(struct regstat *cur,int i)
 828 {
 829   int hr;
 830
 831   for(hr=0;hr<HOST_REGS;hr++) {
 832     if(hr!=EXCLUDE_REG) {
 833       if(((cur->regmap[hr]&63)!=rs1[i])&&((cur->regmap[hr]&63)!=rs2[i])&&
 834          ((cur->regmap[hr]&63)!=rt1[i])&&((cur->regmap[hr]&63)!=rt2[i]))
 835       {
 836         cur->regmap[hr]=-1;
 837         cur->dirty&=~(1<<hr);
 838       }
 839       // Don't need zeros
 840       if((cur->regmap[hr]&63)==0)
 841       {
 842         cur->regmap[hr]=-1;
 843         cur->dirty&=~(1<<hr);
 844       }
 845     }
 846   }
 847 }
 848
 849
 850 void div64(int64_t dividend,int64_t divisor)
 851 {
 852   lo=dividend/divisor;
 853   hi=dividend%divisor;
 854   //printf("TRACE: ddiv %8x%8x %8x%8x\n" ,(int)reg[HIREG],(int)(reg[HIREG]>>32)
 855   //                                     ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
 856 }
 857 void divu64(uint64_t dividend,uint64_t divisor)
 858 {
 859   lo=dividend/divisor;
 860   hi=dividend%divisor;
 861   //printf("TRACE: ddivu %8x%8x %8x%8x\n",(int)reg[HIREG],(int)(reg[HIREG]>>32)
 862   //                                     ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
 863 }
 864
 865 void mult64(uint64_t m1,uint64_t m2)
 866 {
 867    unsigned long long int op1, op2, op3, op4;
 868    unsigned long long int result1, result2, result3, result4;
 869    unsigned long long int temp1, temp2, temp3, temp4;
 870    int sign = 0;
 871
 872    if (m1 < 0)
 873      {
 874     op2 = -m1;
 875     sign = 1 - sign;
 876      }
 877    else op2 = m1;
 878    if (m2 < 0)
 879      {
 880     op4 = -m2;
 881     sign = 1 - sign;
 882      }
 883    else op4 = m2;
 884
 885    op1 = op2 & 0xFFFFFFFF;
 886    op2 = (op2 >> 32) & 0xFFFFFFFF;
 887    op3 = op4 & 0xFFFFFFFF;
 888    op4 = (op4 >> 32) & 0xFFFFFFFF;
 889
 890    temp1 = op1 * op3;
 891    temp2 = (temp1 >> 32) + op1 * op4;
 892    temp3 = op2 * op3;
 893    temp4 = (temp3 >> 32) + op2 * op4;
 894
 895    result1 = temp1 & 0xFFFFFFFF;
 896    result2 = temp2 + (temp3 & 0xFFFFFFFF);
 897    result3 = (result2 >> 32) + temp4;
 898    result4 = (result3 >> 32);
 899
 900    lo = result1 | (result2 << 32);
 901    hi = (result3 & 0xFFFFFFFF) | (result4 << 32);
 902    if (sign)
 903      {
 904     hi = ~hi;
 905     if (!lo) hi++;
 906     else lo = ~lo + 1;
 907      }
 908 }
 909
 910 void multu64(uint64_t m1,uint64_t m2)
 911 {
 912    unsigned long long int op1, op2, op3, op4;
 913    unsigned long long int result1, result2, result3, result4;
 914    unsigned long long int temp1, temp2, temp3, temp4;
 915
 916    op1 = m1 & 0xFFFFFFFF;
 917    op2 = (m1 >> 32) & 0xFFFFFFFF;
 918    op3 = m2 & 0xFFFFFFFF;
 919    op4 = (m2 >> 32) & 0xFFFFFFFF;
 920
 921    temp1 = op1 * op3;
 922    temp2 = (temp1 >> 32) + op1 * op4;
 923    temp3 = op2 * op3;
 924    temp4 = (temp3 >> 32) + op2 * op4;
 925
 926    result1 = temp1 & 0xFFFFFFFF;
 927    result2 = temp2 + (temp3 & 0xFFFFFFFF);
 928    result3 = (result2 >> 32) + temp4;
 929    result4 = (result3 >> 32);
 930
 931    lo = result1 | (result2 << 32);
 932    hi = (result3 & 0xFFFFFFFF) | (result4 << 32);
 933
 934   //printf("TRACE: dmultu %8x%8x %8x%8x\n",(int)reg[HIREG],(int)(reg[HIREG]>>32)
 935   //                                      ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
 936 }
 937
 938 uint64_t ldl_merge(uint64_t original,uint64_t loaded,u_int bits)
 939 {
 940   if(bits) {
 941     original<<=64-bits;
 942     original>>=64-bits;
 943     loaded<<=bits;
 944     original|=loaded;
 945   }
 946   else original=loaded;
 947   return original;
 948 }
 949 uint64_t ldr_merge(uint64_t original,uint64_t loaded,u_int bits)
 950 {
 951   if(bits^56) {
 952     original>>=64-(bits^56);
 953     original<<=64-(bits^56);
 954     loaded>>=bits^56;
 955     original|=loaded;
 956   }
 957   else original=loaded;
 958   return original;
 959 }
 960
 961 #ifdef __i386__
 962 #include "assem_x86.c"
 963 #endif
 964 #ifdef __x86_64__
 965 #include "assem_x64.c"
 966 #endif
 967 #ifdef __arm__
 968 #include "assem_arm.c"
 969 #endif
 970
 971 // Add virtual address mapping to linked list
 972 void ll_add(struct ll_entry **head,int vaddr,void *addr)
 973 {
 974   struct ll_entry *new_entry;
 975   new_entry=malloc(sizeof(struct ll_entry));
 976   assert(new_entry!=NULL);
 977   new_entry->vaddr=vaddr;
 978   new_entry->reg32=0;
 979   new_entry->addr=addr;
 980   new_entry->next=*head;
 981   *head=new_entry;
 982 }
 983
 984 // Add virtual address mapping for 32-bit compiled block
 985 void ll_add_32(struct ll_entry **head,int vaddr,u_int reg32,void *addr)
 986 {
 987   ll_add(head,vaddr,addr);
 988 #ifndef FORCE32
 989   (*head)->reg32=reg32;
 990 #endif
 991 }
 992
 993 // Check if an address is already compiled
 994 // but don't return addresses which are about to expire from the cache
 995 void *check_addr(u_int vaddr)
 996 {
 997   u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
 998   if(ht_bin[0]==vaddr) {
 999     if(((ht_bin[1]-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[1])) return (void *)ht_bin[1];
1001   }
1002   if(ht_bin[2]==vaddr) {
1003     if(((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
1004       if(isclean(ht_bin[3])) return (void *)ht_bin[3];
1005   }
1006   u_int page=get_page(vaddr);
1007   struct ll_entry *head;
1008   head=jump_in[page];
1009   while(head!=NULL) {
1010     if(head->vaddr==vaddr&&head->reg32==0) {
1011       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1012         // Update existing entry with current address
1013         if(ht_bin[0]==vaddr) {
1014           ht_bin[1]=(int)head->addr;
1015           return head->addr;
1016         }
1017         if(ht_bin[2]==vaddr) {
1018           ht_bin[3]=(int)head->addr;
1019           return head->addr;
1020         }
1021         // Insert into hash table with low priority.
1022         // Don't evict existing entries, as they are probably
1023         // addresses that are being accessed frequently.
1024         if(ht_bin[0]==-1) {
1025           ht_bin[1]=(int)head->addr;
1026           ht_bin[0]=vaddr;
1027         }else if(ht_bin[2]==-1) {
1028           ht_bin[3]=(int)head->addr;
1029           ht_bin[2]=vaddr;
1030         }
1031         return head->addr;
1032       }
1033     }
1034     head=head->next;
1035   }
1036   return 0;
1037 }
1038
1039 void remove_hash(int vaddr)
1040 {
1041   //printf("remove hash: %x\n",vaddr);
1042   int *ht_bin=hash_table[(((vaddr)>>16)^vaddr)&0xFFFF];
1043   if(ht_bin[2]==vaddr) {
1044     ht_bin[2]=ht_bin[3]=-1;
1045   }
1046   if(ht_bin[0]==vaddr) {
1047     ht_bin[0]=ht_bin[2];
1048     ht_bin[1]=ht_bin[3];
1049     ht_bin[2]=ht_bin[3]=-1;
1050   }
1051 }
1052
1053 void ll_remove_matching_addrs(struct ll_entry **head,int addr,int shift)
1054 {
1055   struct ll_entry *next;
1056   while(*head) {
1057     if(((u_int)((*head)->addr)>>shift)==(addr>>shift) ||
1058        ((u_int)((*head)->addr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift))
1059     {
1060       inv_debug("EXP: Remove pointer to %x (%x)\n",(int)(*head)->addr,(*head)->vaddr);
1061       remove_hash((*head)->vaddr);
1062       next=(*head)->next;
1063       free(*head);
1064       *head=next;
1065     }
1066     else
1067     {
1068       head=&((*head)->next);
1069     }
1070   }
1071 }
1072
1073 // Remove all entries from linked list
1074 void ll_clear(struct ll_entry **head)
1075 {
1076   struct ll_entry *cur;
1077   struct ll_entry *next;
1078   if(cur=*head) {
1079     *head=0;
1080     while(cur) {
1081       next=cur->next;
1082       free(cur);
1083       cur=next;
1084     }
1085   }
1086 }
1087
1088 // Dereference the pointers and remove if it matches
1089 void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
1090 {
1091   while(head) {
1092     int ptr=get_pointer(head->addr);
1093     inv_debug("EXP: Lookup pointer to %x at %x (%x)\n",(int)ptr,(int)head->addr,head->vaddr);
1094     if(((ptr>>shift)==(addr>>shift)) ||
1095        (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
1096     {
1097       inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
1098       u_int host_addr=(u_int)kill_pointer(head->addr);
1099       #ifdef __arm__
1100         needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
1101       #endif
1102     }
1103     head=head->next;
1104   }
1105 }
1106
1107 // This is called when we write to a compiled block (see do_invstub)
1108 void invalidate_page(u_int page)
1109 {
1110   struct ll_entry *head;
1111   struct ll_entry *next;
1112   head=jump_in[page];
1113   jump_in[page]=0;
1114   while(head!=NULL) {
1115     inv_debug("INVALIDATE: %x\n",head->vaddr);
1116     remove_hash(head->vaddr);
1117     next=head->next;
1118     free(head);
1119     head=next;
1120   }
1121   head=jump_out[page];
1122   jump_out[page]=0;
1123   while(head!=NULL) {
1124     inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
1125     u_int host_addr=(u_int)kill_pointer(head->addr);
1126     #ifdef __arm__
1127       needs_clear_cache[(host_addr-(u_int)BASE_ADDR)>>17]|=1<<(((host_addr-(u_int)BASE_ADDR)>>12)&31);
1128     #endif
1129     next=head->next;
1130     free(head);
1131     head=next;
1132   }
1133 }
1134 void invalidate_block(u_int block)
1135 {
1136   u_int page=get_page(block<<12);
1137   u_int vpage=get_vpage(block<<12);
1138   inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1139   //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1140   u_int first,last;
1141   first=last=page;
1142   struct ll_entry *head;
1143   head=jump_dirty[vpage];
1144   //printf("page=%d vpage=%d\n",page,vpage);
1145   while(head!=NULL) {
1146     u_int start,end;
1147     if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1148       get_bounds((int)head->addr,&start,&end);
1149       //printf("start: %x end: %x\n",start,end);
1150       if(page<2048&&start>=0x80000000&&end<0x80000000+RAM_SIZE) {
1151         if(((start-(u_int)rdram)>>12)<=page&&((end-1-(u_int)rdram)>>12)>=page) {
1152           if((((start-(u_int)rdram)>>12)&2047)<first) first=((start-(u_int)rdram)>>12)&2047;
1153           if((((end-1-(u_int)rdram)>>12)&2047)>last) last=((end-1-(u_int)rdram)>>12)&2047;
1154         }
1155       }
1156 #ifndef DISABLE_TLB
1157       if(page<2048&&(signed int)start>=(signed int)0xC0000000&&(signed int)end>=(signed int)0xC0000000) {
1158         if(((start+memory_map[start>>12]-(u_int)rdram)>>12)<=page&&((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)>=page) {
1159           if((((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047)<first) first=((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047;
1160           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;
1161         }
1162       }
1163 #endif
1164     }
1165     head=head->next;
1166   }
1167   //printf("first=%d last=%d\n",first,last);
1168   invalidate_page(page);
1169   assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1170   assert(last<page+5);
1171   // Invalidate the adjacent pages if a block crosses a 4K boundary
1172   while(first<page) {
1173     invalidate_page(first);
1174     first++;
1175   }
1176   for(first=page+1;first<last;first++) {
1177     invalidate_page(first);
1178   }
1179   #ifdef __arm__
1180     do_clear_cache();
1181   #endif
1182
1183   // Don't trap writes
1184   invalid_code[block]=1;
1185 #ifndef DISABLE_TLB
1186   // If there is a valid TLB entry for this page, remove write protect
1187   if(tlb_LUT_w[block]) {
1188     assert(tlb_LUT_r[block]==tlb_LUT_w[block]);
1189     // CHECK: Is this right?
1190     memory_map[block]=((tlb_LUT_w[block]&0xFFFFF000)-(block<<12)+(unsigned int)rdram-0x80000000)>>2;
1191     u_int real_block=tlb_LUT_w[block]>>12;
1192     invalid_code[real_block]=1;
1193     if(real_block>=0x80000&&real_block<0x80800) memory_map[real_block]=((u_int)rdram-0x80000000)>>2;
1194   }
1195   else if(block>=0x80000&&block<0x80800) memory_map[block]=((u_int)rdram-0x80000000)>>2;
1196 #endif
1197
1198   #ifdef USE_MINI_HT
1199   memset(mini_ht,-1,sizeof(mini_ht));
1200   #endif
1201 }
1202 void invalidate_addr(u_int addr)
1203 {
1204   invalidate_block(addr>>12);
1205 }
1206 // This is called when loading a save state.
1207 // Anything could have changed, so invalidate everything.
1208 void invalidate_all_pages()
1209 {
1210   u_int page,n;
1211   for(page=0;page<4096;page++)
1212     invalidate_page(page);
1213   for(page=0;page<1048576;page++)
1214     if(!invalid_code[page]) {
1215       restore_candidate[(page&2047)>>3]|=1<<(page&7);
1216       restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1217     }
1218   #ifdef __arm__
1219   __clear_cache((void *)BASE_ADDR,(void *)BASE_ADDR+(1<<TARGET_SIZE_2));
1220   #endif
1221   #ifdef USE_MINI_HT
1222   memset(mini_ht,-1,sizeof(mini_ht));
1223   #endif
1224   #ifndef DISABLE_TLB
1225   // TLB
1226   for(page=0;page<0x100000;page++) {
1227     if(tlb_LUT_r[page]) {
1228       memory_map[page]=((tlb_LUT_r[page]&0xFFFFF000)-(page<<12)+(unsigned int)rdram-0x80000000)>>2;
1229       if(!tlb_LUT_w[page]||!invalid_code[page])
1230         memory_map[page]|=0x40000000; // Write protect
1231     }
1232     else memory_map[page]=-1;
1233     if(page==0x80000) page=0xC0000;
1234   }
1235   tlb_hacks();
1236   #endif
1237 }
1238
1239 // Add an entry to jump_out after making a link
1240 void add_link(u_int vaddr,void *src)
1241 {
1242   u_int page=get_page(vaddr);
1243   inv_debug("add_link: %x -> %x (%d)\n",(int)src,vaddr,page);
1244   ll_add(jump_out+page,vaddr,src);
1245   //int ptr=get_pointer(src);
1246   //inv_debug("add_link: Pointer is to %x\n",(int)ptr);
1247 }
1248
1249 // If a code block was found to be unmodified (bit was set in
1250 // restore_candidate) and it remains unmodified (bit is clear
1251 // in invalid_code) then move the entries for that 4K page from
1252 // the dirty list to the clean list.
1253 void clean_blocks(u_int page)
1254 {
1255   struct ll_entry *head;
1256   inv_debug("INV: clean_blocks page=%d\n",page);
1257   head=jump_dirty[page];
1258   while(head!=NULL) {
1259     if(!invalid_code[head->vaddr>>12]) {
1260       // Don't restore blocks which are about to expire from the cache
1261       if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1262         u_int start,end;
1263         if(verify_dirty((int)head->addr)) {
1264           //printf("Possibly Restore %x (%x)\n",head->vaddr, (int)head->addr);
1265           u_int i;
1266           u_int inv=0;
1267           get_bounds((int)head->addr,&start,&end);
1268           if(start-(u_int)rdram<RAM_SIZE) {
1269             for(i=(start-(u_int)rdram+0x80000000)>>12;i<=(end-1-(u_int)rdram+0x80000000)>>12;i++) {
1270               inv|=invalid_code[i];
1271             }
1272           }
1273           if((signed int)head->vaddr>=(signed int)0xC0000000) {
1274             u_int addr = (head->vaddr+(memory_map[head->vaddr>>12]<<2));
1275             //printf("addr=%x start=%x end=%x\n",addr,start,end);
1276             if(addr<start||addr>=end) inv=1;
1277           }
1278           else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1279             inv=1;
1280           }
1281           if(!inv) {
1282             void * clean_addr=(void *)get_clean_addr((int)head->addr);
1283             if((((u_int)clean_addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1284               u_int ppage=page;
1285 #ifndef DISABLE_TLB
1286               if(page<2048&&tlb_LUT_r[head->vaddr>>12]) ppage=(tlb_LUT_r[head->vaddr>>12]^0x80000000)>>12;
1287 #endif
1288               inv_debug("INV: Restored %x (%x/%x)\n",head->vaddr, (int)head->addr, (int)clean_addr);
1289               //printf("page=%x, addr=%x\n",page,head->vaddr);
1290               //assert(head->vaddr>>12==(page|0x80000));
1291               ll_add_32(jump_in+ppage,head->vaddr,head->reg32,clean_addr);
1292               int *ht_bin=hash_table[((head->vaddr>>16)^head->vaddr)&0xFFFF];
1293               if(!head->reg32) {
1294                 if(ht_bin[0]==head->vaddr) {
1295                   ht_bin[1]=(int)clean_addr; // Replace existing entry
1296                 }
1297                 if(ht_bin[2]==head->vaddr) {
1298                   ht_bin[3]=(int)clean_addr; // Replace existing entry
1299                 }
1300               }
1301             }
1302           }
1303         }
1304       }
1305     }
1306     head=head->next;
1307   }
1308 }
1309
1310
1311 void mov_alloc(struct regstat *current,int i)
1312 {
1313   // Note: Don't need to actually alloc the source registers
1314   if((~current->is32>>rs1[i])&1) {
1315     //alloc_reg64(current,i,rs1[i]);
1316     alloc_reg64(current,i,rt1[i]);
1317     current->is32&=~(1LL<<rt1[i]);
1318   } else {
1319     //alloc_reg(current,i,rs1[i]);
1320     alloc_reg(current,i,rt1[i]);
1321     current->is32|=(1LL<<rt1[i]);
1322   }
1323   clear_const(current,rs1[i]);
1324   clear_const(current,rt1[i]);
1325   dirty_reg(current,rt1[i]);
1326 }
1327
1328 void shiftimm_alloc(struct regstat *current,int i)
1329 {
1330   clear_const(current,rs1[i]);
1331   clear_const(current,rt1[i]);
1332   if(opcode2[i]<=0x3) // SLL/SRL/SRA
1333   {
1334     if(rt1[i]) {
1335       if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1336       else lt1[i]=rs1[i];
1337       alloc_reg(current,i,rt1[i]);
1338       current->is32|=1LL<<rt1[i];
1339       dirty_reg(current,rt1[i]);
1340     }
1341   }
1342   if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
1343   {
1344     if(rt1[i]) {
1345       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1346       alloc_reg64(current,i,rt1[i]);
1347       current->is32&=~(1LL<<rt1[i]);
1348       dirty_reg(current,rt1[i]);
1349     }
1350   }
1351   if(opcode2[i]==0x3c) // DSLL32
1352   {
1353     if(rt1[i]) {
1354       if(rs1[i]) alloc_reg(current,i,rs1[i]);
1355       alloc_reg64(current,i,rt1[i]);
1356       current->is32&=~(1LL<<rt1[i]);
1357       dirty_reg(current,rt1[i]);
1358     }
1359   }
1360   if(opcode2[i]==0x3e) // DSRL32
1361   {
1362     if(rt1[i]) {
1363       alloc_reg64(current,i,rs1[i]);
1364       if(imm[i]==32) {
1365         alloc_reg64(current,i,rt1[i]);
1366         current->is32&=~(1LL<<rt1[i]);
1367       } else {
1368         alloc_reg(current,i,rt1[i]);
1369         current->is32|=1LL<<rt1[i];
1370       }
1371       dirty_reg(current,rt1[i]);
1372     }
1373   }
1374   if(opcode2[i]==0x3f) // DSRA32
1375   {
1376     if(rt1[i]) {
1377       alloc_reg64(current,i,rs1[i]);
1378       alloc_reg(current,i,rt1[i]);
1379       current->is32|=1LL<<rt1[i];
1380       dirty_reg(current,rt1[i]);
1381     }
1382   }
1383 }
1384
1385 void shift_alloc(struct regstat *current,int i)
1386 {
1387   if(rt1[i]) {
1388     if(opcode2[i]<=0x07) // SLLV/SRLV/SRAV
1389     {
1390       if(rs1[i]) alloc_reg(current,i,rs1[i]);
1391       if(rs2[i]) alloc_reg(current,i,rs2[i]);
1392       alloc_reg(current,i,rt1[i]);
1393       if(rt1[i]==rs2[i]) {
1394         alloc_reg_temp(current,i,-1);
1395         minimum_free_regs[i]=1;
1396       }
1397       current->is32|=1LL<<rt1[i];
1398     } else { // DSLLV/DSRLV/DSRAV
1399       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1400       if(rs2[i]) alloc_reg(current,i,rs2[i]);
1401       alloc_reg64(current,i,rt1[i]);
1402       current->is32&=~(1LL<<rt1[i]);
1403       if(opcode2[i]==0x16||opcode2[i]==0x17) // DSRLV and DSRAV need a temporary register
1404       {
1405         alloc_reg_temp(current,i,-1);
1406         minimum_free_regs[i]=1;
1407       }
1408     }
1409     clear_const(current,rs1[i]);
1410     clear_const(current,rs2[i]);
1411     clear_const(current,rt1[i]);
1412     dirty_reg(current,rt1[i]);
1413   }
1414 }
1415
1416 void alu_alloc(struct regstat *current,int i)
1417 {
1418   if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
1419     if(rt1[i]) {
1420       if(rs1[i]&&rs2[i]) {
1421         alloc_reg(current,i,rs1[i]);
1422         alloc_reg(current,i,rs2[i]);
1423       }
1424       else {
1425         if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1426         if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1427       }
1428       alloc_reg(current,i,rt1[i]);
1429     }
1430     current->is32|=1LL<<rt1[i];
1431   }
1432   if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
1433     if(rt1[i]) {
1434       if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1435       {
1436         alloc_reg64(current,i,rs1[i]);
1437         alloc_reg64(current,i,rs2[i]);
1438         alloc_reg(current,i,rt1[i]);
1439       } else {
1440         alloc_reg(current,i,rs1[i]);
1441         alloc_reg(current,i,rs2[i]);
1442         alloc_reg(current,i,rt1[i]);
1443       }
1444     }
1445     current->is32|=1LL<<rt1[i];
1446   }
1447   if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
1448     if(rt1[i]) {
1449       if(rs1[i]&&rs2[i]) {
1450         alloc_reg(current,i,rs1[i]);
1451         alloc_reg(current,i,rs2[i]);
1452       }
1453       else
1454       {
1455         if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1456         if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1457       }
1458       alloc_reg(current,i,rt1[i]);
1459       if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1460       {
1461         if(!((current->uu>>rt1[i])&1)) {
1462           alloc_reg64(current,i,rt1[i]);
1463         }
1464         if(get_reg(current->regmap,rt1[i]|64)>=0) {
1465           if(rs1[i]&&rs2[i]) {
1466             alloc_reg64(current,i,rs1[i]);
1467             alloc_reg64(current,i,rs2[i]);
1468           }
1469           else
1470           {
1471             // Is is really worth it to keep 64-bit values in registers?
1472             #ifdef NATIVE_64BIT
1473             if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1474             if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg64(current,i,rs2[i]);
1475             #endif
1476           }
1477         }
1478         current->is32&=~(1LL<<rt1[i]);
1479       } else {
1480         current->is32|=1LL<<rt1[i];
1481       }
1482     }
1483   }
1484   if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1485     if(rt1[i]) {
1486       if(rs1[i]&&rs2[i]) {
1487         if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1488           alloc_reg64(current,i,rs1[i]);
1489           alloc_reg64(current,i,rs2[i]);
1490           alloc_reg64(current,i,rt1[i]);
1491         } else {
1492           alloc_reg(current,i,rs1[i]);
1493           alloc_reg(current,i,rs2[i]);
1494           alloc_reg(current,i,rt1[i]);
1495         }
1496       }
1497       else {
1498         alloc_reg(current,i,rt1[i]);
1499         if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1500           // DADD used as move, or zeroing
1501           // If we have a 64-bit source, then make the target 64 bits too
1502           if(rs1[i]&&!((current->is32>>rs1[i])&1)) {
1503             if(get_reg(current->regmap,rs1[i])>=0) alloc_reg64(current,i,rs1[i]);
1504             alloc_reg64(current,i,rt1[i]);
1505           } else if(rs2[i]&&!((current->is32>>rs2[i])&1)) {
1506             if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1507             alloc_reg64(current,i,rt1[i]);
1508           }
1509           if(opcode2[i]>=0x2e&&rs2[i]) {
1510             // DSUB used as negation - 64-bit result
1511             // If we have a 32-bit register, extend it to 64 bits
1512             if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1513             alloc_reg64(current,i,rt1[i]);
1514           }
1515         }
1516       }
1517       if(rs1[i]&&rs2[i]) {
1518         current->is32&=~(1LL<<rt1[i]);
1519       } else if(rs1[i]) {
1520         current->is32&=~(1LL<<rt1[i]);
1521         if((current->is32>>rs1[i])&1)
1522           current->is32|=1LL<<rt1[i];
1523       } else if(rs2[i]) {
1524         current->is32&=~(1LL<<rt1[i]);
1525         if((current->is32>>rs2[i])&1)
1526           current->is32|=1LL<<rt1[i];
1527       } else {
1528         current->is32|=1LL<<rt1[i];
1529       }
1530     }
1531   }
1532   clear_const(current,rs1[i]);
1533   clear_const(current,rs2[i]);
1534   clear_const(current,rt1[i]);
1535   dirty_reg(current,rt1[i]);
1536 }
1537
1538 void imm16_alloc(struct regstat *current,int i)
1539 {
1540   if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1541   else lt1[i]=rs1[i];
1542   if(rt1[i]) alloc_reg(current,i,rt1[i]);
1543   if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
1544     current->is32&=~(1LL<<rt1[i]);
1545     if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1546       // TODO: Could preserve the 32-bit flag if the immediate is zero
1547       alloc_reg64(current,i,rt1[i]);
1548       alloc_reg64(current,i,rs1[i]);
1549     }
1550     clear_const(current,rs1[i]);
1551     clear_const(current,rt1[i]);
1552   }
1553   else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
1554     if((~current->is32>>rs1[i])&1) alloc_reg64(current,i,rs1[i]);
1555     current->is32|=1LL<<rt1[i];
1556     clear_const(current,rs1[i]);
1557     clear_const(current,rt1[i]);
1558   }
1559   else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
1560     if(((~current->is32>>rs1[i])&1)&&opcode[i]>0x0c) {
1561       if(rs1[i]!=rt1[i]) {
1562         if(needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1563         alloc_reg64(current,i,rt1[i]);
1564         current->is32&=~(1LL<<rt1[i]);
1565       }
1566     }
1567     else current->is32|=1LL<<rt1[i]; // ANDI clears upper bits
1568     if(is_const(current,rs1[i])) {
1569       int v=get_const(current,rs1[i]);
1570       if(opcode[i]==0x0c) set_const(current,rt1[i],v&imm[i]);
1571       if(opcode[i]==0x0d) set_const(current,rt1[i],v|imm[i]);
1572       if(opcode[i]==0x0e) set_const(current,rt1[i],v^imm[i]);
1573     }
1574     else clear_const(current,rt1[i]);
1575   }
1576   else if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
1577     if(is_const(current,rs1[i])) {
1578       int v=get_const(current,rs1[i]);
1579       set_const(current,rt1[i],v+imm[i]);
1580     }
1581     else clear_const(current,rt1[i]);
1582     current->is32|=1LL<<rt1[i];
1583   }
1584   else {
1585     set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
1586     current->is32|=1LL<<rt1[i];
1587   }
1588   dirty_reg(current,rt1[i]);
1589 }
1590
1591 void load_alloc(struct regstat *current,int i)
1592 {
1593   clear_const(current,rt1[i]);
1594   //if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
1595   if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
1596   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1597   if(rt1[i]) {
1598     alloc_reg(current,i,rt1[i]);
1599     if(get_reg(current->regmap,rt1[i])<0) {
1600       // dummy load, but we still need a register to calculate the address
1601       alloc_reg_temp(current,i,-1);
1602       minimum_free_regs[i]=1;
1603     }
1604     if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
1605     {
1606       current->is32&=~(1LL<<rt1[i]);
1607       alloc_reg64(current,i,rt1[i]);
1608     }
1609     else if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1610     {
1611       current->is32&=~(1LL<<rt1[i]);
1612       alloc_reg64(current,i,rt1[i]);
1613       alloc_all(current,i);
1614       alloc_reg64(current,i,FTEMP);
1615       minimum_free_regs[i]=HOST_REGS;
1616     }
1617     else current->is32|=1LL<<rt1[i];
1618     dirty_reg(current,rt1[i]);
1619     // If using TLB, need a register for pointer to the mapping table
1620     if(using_tlb) alloc_reg(current,i,TLREG);
1621     // LWL/LWR need a temporary register for the old value
1622     if(opcode[i]==0x22||opcode[i]==0x26)
1623     {
1624       alloc_reg(current,i,FTEMP);
1625       alloc_reg_temp(current,i,-1);
1626       minimum_free_regs[i]=1;
1627     }
1628   }
1629   else
1630   {
1631     // Load to r0 (dummy load)
1632     // but we still need a register to calculate the address
1633     if(opcode[i]==0x22||opcode[i]==0x26)
1634     {
1635       alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1636     }
1637     alloc_reg_temp(current,i,-1);
1638     minimum_free_regs[i]=1;
1639     if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1640     {
1641       alloc_all(current,i);
1642       alloc_reg64(current,i,FTEMP);
1643       minimum_free_regs[i]=HOST_REGS;
1644     }
1645   }
1646 }
1647
1648 void store_alloc(struct regstat *current,int i)
1649 {
1650   clear_const(current,rs2[i]);
1651   if(!(rs2[i])) current->u&=~1LL; // Allow allocating r0 if necessary
1652   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1653   alloc_reg(current,i,rs2[i]);
1654   if(opcode[i]==0x2c||opcode[i]==0x2d||opcode[i]==0x3f) { // 64-bit SDL/SDR/SD
1655     alloc_reg64(current,i,rs2[i]);
1656     if(rs2[i]) alloc_reg(current,i,FTEMP);
1657   }
1658   // If using TLB, need a register for pointer to the mapping table
1659   if(using_tlb) alloc_reg(current,i,TLREG);
1660   #if defined(HOST_IMM8)
1661   // On CPUs without 32-bit immediates we need a pointer to invalid_code
1662   else alloc_reg(current,i,INVCP);
1663   #endif
1664   if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) { // SWL/SWL/SDL/SDR
1665     alloc_reg(current,i,FTEMP);
1666   }
1667   // We need a temporary register for address generation
1668   alloc_reg_temp(current,i,-1);
1669   minimum_free_regs[i]=1;
1670 }
1671
1672 void c1ls_alloc(struct regstat *current,int i)
1673 {
1674   //clear_const(current,rs1[i]); // FIXME
1675   clear_const(current,rt1[i]);
1676   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1677   alloc_reg(current,i,CSREG); // Status
1678   alloc_reg(current,i,FTEMP);
1679   if(opcode[i]==0x35||opcode[i]==0x3d) { // 64-bit LDC1/SDC1
1680     alloc_reg64(current,i,FTEMP);
1681   }
1682   // If using TLB, need a register for pointer to the mapping table
1683   if(using_tlb) alloc_reg(current,i,TLREG);
1684   #if defined(HOST_IMM8)
1685   // On CPUs without 32-bit immediates we need a pointer to invalid_code
1686   else if((opcode[i]&0x3b)==0x39) // SWC1/SDC1
1687     alloc_reg(current,i,INVCP);
1688   #endif
1689   // We need a temporary register for address generation
1690   alloc_reg_temp(current,i,-1);
1691 }
1692
1693 void c2ls_alloc(struct regstat *current,int i)
1694 {
1695   clear_const(current,rt1[i]);
1696   if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1697   alloc_reg(current,i,FTEMP);
1698   // If using TLB, need a register for pointer to the mapping table
1699   if(using_tlb) alloc_reg(current,i,TLREG);
1700   #if defined(HOST_IMM8)
1701   // On CPUs without 32-bit immediates we need a pointer to invalid_code
1702   else if((opcode[i]&0x3b)==0x3a) // SWC2/SDC2
1703     alloc_reg(current,i,INVCP);
1704   #endif
1705   // We need a temporary register for address generation
1706   alloc_reg_temp(current,i,-1);
1707   minimum_free_regs[i]=1;
1708 }
1709
1710 #ifndef multdiv_alloc
1711 void multdiv_alloc(struct regstat *current,int i)
1712 {
1713   //  case 0x18: MULT
1714   //  case 0x19: MULTU
1715   //  case 0x1A: DIV
1716   //  case 0x1B: DIVU
1717   //  case 0x1C: DMULT
1718   //  case 0x1D: DMULTU
1719   //  case 0x1E: DDIV
1720   //  case 0x1F: DDIVU
1721   clear_const(current,rs1[i]);
1722   clear_const(current,rs2[i]);
1723   if(rs1[i]&&rs2[i])
1724   {
1725     if((opcode2[i]&4)==0) // 32-bit
1726     {
1727       current->u&=~(1LL<<HIREG);
1728       current->u&=~(1LL<<LOREG);
1729       alloc_reg(current,i,HIREG);
1730       alloc_reg(current,i,LOREG);
1731       alloc_reg(current,i,rs1[i]);
1732       alloc_reg(current,i,rs2[i]);
1733       current->is32|=1LL<<HIREG;
1734       current->is32|=1LL<<LOREG;
1735       dirty_reg(current,HIREG);
1736       dirty_reg(current,LOREG);
1737     }
1738     else // 64-bit
1739     {
1740       current->u&=~(1LL<<HIREG);
1741       current->u&=~(1LL<<LOREG);
1742       current->uu&=~(1LL<<HIREG);
1743       current->uu&=~(1LL<<LOREG);
1744       alloc_reg64(current,i,HIREG);
1745       //if(HOST_REGS>10) alloc_reg64(current,i,LOREG);
1746       alloc_reg64(current,i,rs1[i]);
1747       alloc_reg64(current,i,rs2[i]);
1748       alloc_all(current,i);
1749       current->is32&=~(1LL<<HIREG);
1750       current->is32&=~(1LL<<LOREG);
1751       dirty_reg(current,HIREG);
1752       dirty_reg(current,LOREG);
1753       minimum_free_regs[i]=HOST_REGS;
1754     }
1755   }
1756   else
1757   {
1758     // Multiply by zero is zero.
1759     // MIPS does not have a divide by zero exception.
1760     // The result is undefined, we return zero.
1761     alloc_reg(current,i,HIREG);
1762     alloc_reg(current,i,LOREG);
1763     current->is32|=1LL<<HIREG;
1764     current->is32|=1LL<<LOREG;
1765     dirty_reg(current,HIREG);
1766     dirty_reg(current,LOREG);
1767   }
1768 }
1769 #endif
1770
1771 void cop0_alloc(struct regstat *current,int i)
1772 {
1773   if(opcode2[i]==0) // MFC0
1774   {
1775     if(rt1[i]) {
1776       clear_const(current,rt1[i]);
1777       alloc_all(current,i);
1778       alloc_reg(current,i,rt1[i]);
1779       current->is32|=1LL<<rt1[i];
1780       dirty_reg(current,rt1[i]);
1781     }
1782   }
1783   else if(opcode2[i]==4) // MTC0
1784   {
1785     if(rs1[i]){
1786       clear_const(current,rs1[i]);
1787       alloc_reg(current,i,rs1[i]);
1788       alloc_all(current,i);
1789     }
1790     else {
1791       alloc_all(current,i); // FIXME: Keep r0
1792       current->u&=~1LL;
1793       alloc_reg(current,i,0);
1794     }
1795   }
1796   else
1797   {
1798     // TLBR/TLBWI/TLBWR/TLBP/ERET
1799     assert(opcode2[i]==0x10);
1800     alloc_all(current,i);
1801   }
1802   minimum_free_regs[i]=HOST_REGS;
1803 }
1804
1805 void cop1_alloc(struct regstat *current,int i)
1806 {
1807   alloc_reg(current,i,CSREG); // Load status
1808   if(opcode2[i]<3) // MFC1/DMFC1/CFC1
1809   {
1810     if(rt1[i]){
1811       clear_const(current,rt1[i]);
1812       if(opcode2[i]==1) {
1813         alloc_reg64(current,i,rt1[i]); // DMFC1
1814         current->is32&=~(1LL<<rt1[i]);
1815       }else{
1816         alloc_reg(current,i,rt1[i]); // MFC1/CFC1
1817         current->is32|=1LL<<rt1[i];
1818       }
1819       dirty_reg(current,rt1[i]);
1820     }
1821     alloc_reg_temp(current,i,-1);
1822   }
1823   else if(opcode2[i]>3) // MTC1/DMTC1/CTC1
1824   {
1825     if(rs1[i]){
1826       clear_const(current,rs1[i]);
1827       if(opcode2[i]==5)
1828         alloc_reg64(current,i,rs1[i]); // DMTC1
1829       else
1830         alloc_reg(current,i,rs1[i]); // MTC1/CTC1
1831       alloc_reg_temp(current,i,-1);
1832     }
1833     else {
1834       current->u&=~1LL;
1835       alloc_reg(current,i,0);
1836       alloc_reg_temp(current,i,-1);
1837     }
1838   }
1839   minimum_free_regs[i]=1;
1840 }
1841 void fconv_alloc(struct regstat *current,int i)
1842 {
1843   alloc_reg(current,i,CSREG); // Load status
1844   alloc_reg_temp(current,i,-1);
1845   minimum_free_regs[i]=1;
1846 }
1847 void float_alloc(struct regstat *current,int i)
1848 {
1849   alloc_reg(current,i,CSREG); // Load status
1850   alloc_reg_temp(current,i,-1);
1851   minimum_free_regs[i]=1;
1852 }
1853 void c2op_alloc(struct regstat *current,int i)
1854 {
1855   alloc_reg_temp(current,i,-1);
1856 }
1857 void fcomp_alloc(struct regstat *current,int i)
1858 {
1859   alloc_reg(current,i,CSREG); // Load status
1860   alloc_reg(current,i,FSREG); // Load flags
1861   dirty_reg(current,FSREG); // Flag will be modified
1862   alloc_reg_temp(current,i,-1);
1863   minimum_free_regs[i]=1;
1864 }
1865
1866 void syscall_alloc(struct regstat *current,int i)
1867 {
1868   alloc_cc(current,i);
1869   dirty_reg(current,CCREG);
1870   alloc_all(current,i);
1871   minimum_free_regs[i]=HOST_REGS;
1872   current->isconst=0;
1873 }
1874
1875 void delayslot_alloc(struct regstat *current,int i)
1876 {
1877   switch(itype[i]) {
1878     case UJUMP:
1879     case CJUMP:
1880     case SJUMP:
1881     case RJUMP:
1882     case FJUMP:
1883     case SYSCALL:
1884     case HLECALL:
1885     case SPAN:
1886       assem_debug("jump in the delay slot.  this shouldn't happen.\n");//exit(1);
1887       printf("Disabled speculative precompilation\n");
1888       stop_after_jal=1;
1889       break;
1890     case IMM16:
1891       imm16_alloc(current,i);
1892       break;
1893     case LOAD:
1894     case LOADLR:
1895       load_alloc(current,i);
1896       break;
1897     case STORE:
1898     case STORELR:
1899       store_alloc(current,i);
1900       break;
1901     case ALU:
1902       alu_alloc(current,i);
1903       break;
1904     case SHIFT:
1905       shift_alloc(current,i);
1906       break;
1907     case MULTDIV:
1908       multdiv_alloc(current,i);
1909       break;
1910     case SHIFTIMM:
1911       shiftimm_alloc(current,i);
1912       break;
1913     case MOV:
1914       mov_alloc(current,i);
1915       break;
1916     case COP0:
1917       cop0_alloc(current,i);
1918       break;
1919     case COP1:
1920     case COP2:
1921       cop1_alloc(current,i);
1922       break;
1923     case C1LS:
1924       c1ls_alloc(current,i);
1925       break;
1926     case C2LS:
1927       c2ls_alloc(current,i);
1928       break;
1929     case FCONV:
1930       fconv_alloc(current,i);
1931       break;
1932     case FLOAT:
1933       float_alloc(current,i);
1934       break;
1935     case FCOMP:
1936       fcomp_alloc(current,i);
1937       break;
1938     case C2OP:
1939       c2op_alloc(current,i);
1940       break;
1941   }
1942 }
1943
1944 // Special case where a branch and delay slot span two pages in virtual memory
1945 static void pagespan_alloc(struct regstat *current,int i)
1946 {
1947   current->isconst=0;
1948   current->wasconst=0;
1949   regs[i].wasconst=0;
1950   minimum_free_regs[i]=HOST_REGS;
1951   alloc_all(current,i);
1952   alloc_cc(current,i);
1953   dirty_reg(current,CCREG);
1954   if(opcode[i]==3) // JAL
1955   {
1956     alloc_reg(current,i,31);
1957     dirty_reg(current,31);
1958   }
1959   if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
1960   {
1961     alloc_reg(current,i,rs1[i]);
1962     if (rt1[i]!=0) {
1963       alloc_reg(current,i,rt1[i]);
1964       dirty_reg(current,rt1[i]);
1965     }
1966   }
1967   if((opcode[i]&0x2E)==4) // BEQ/BNE/BEQL/BNEL
1968   {
1969     if(rs1[i]) alloc_reg(current,i,rs1[i]);
1970     if(rs2[i]) alloc_reg(current,i,rs2[i]);
1971     if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1972     {
1973       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1974       if(rs2[i]) alloc_reg64(current,i,rs2[i]);
1975     }
1976   }
1977   else
1978   if((opcode[i]&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
1979   {
1980     if(rs1[i]) alloc_reg(current,i,rs1[i]);
1981     if(!((current->is32>>rs1[i])&1))
1982     {
1983       if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1984     }
1985   }
1986   else
1987   if(opcode[i]==0x11) // BC1
1988   {
1989     alloc_reg(current,i,FSREG);
1990     alloc_reg(current,i,CSREG);
1991   }
1992   //else ...
1993 }
1994
1995 add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e)
1996 {
1997   stubs[stubcount][0]=type;
1998   stubs[stubcount][1]=addr;
1999   stubs[stubcount][2]=retaddr;
2000   stubs[stubcount][3]=a;
2001   stubs[stubcount][4]=b;
2002   stubs[stubcount][5]=c;
2003   stubs[stubcount][6]=d;
2004   stubs[stubcount][7]=e;
2005   stubcount++;
2006 }
2007
2008 // Write out a single register
2009 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32)
2010 {
2011   int hr;
2012   for(hr=0;hr<HOST_REGS;hr++) {
2013     if(hr!=EXCLUDE_REG) {
2014       if((regmap[hr]&63)==r) {
2015         if((dirty>>hr)&1) {
2016           if(regmap[hr]<64) {
2017             emit_storereg(r,hr);
2018 #ifndef FORCE32
2019             if((is32>>regmap[hr])&1) {
2020               emit_sarimm(hr,31,hr);
2021               emit_storereg(r|64,hr);
2022             }
2023 #endif
2024           }else{
2025             emit_storereg(r|64,hr);
2026           }
2027         }
2028       }
2029     }
2030   }
2031 }
2032
2033 int mchecksum()
2034 {
2035   //if(!tracedebug) return 0;
2036   int i;
2037   int sum=0;
2038   for(i=0;i<2097152;i++) {
2039     unsigned int temp=sum;
2040     sum<<=1;
2041     sum|=(~temp)>>31;
2042     sum^=((u_int *)rdram)[i];
2043   }
2044   return sum;
2045 }
2046 int rchecksum()
2047 {
2048   int i;
2049   int sum=0;
2050   for(i=0;i<64;i++)
2051     sum^=((u_int *)reg)[i];
2052   return sum;
2053 }
2054 void rlist()
2055 {
2056   int i;
2057   printf("TRACE: ");
2058   for(i=0;i<32;i++)
2059     printf("r%d:%8x%8x ",i,((int *)(reg+i))[1],((int *)(reg+i))[0]);
2060   printf("\n");
2061 #ifndef DISABLE_COP1
2062   printf("TRACE: ");
2063   for(i=0;i<32;i++)
2064     printf("f%d:%8x%8x ",i,((int*)reg_cop1_simple[i])[1],*((int*)reg_cop1_simple[i]));
2065   printf("\n");
2066 #endif
2067 }
2068
2069 void enabletrace()
2070 {
2071   tracedebug=1;
2072 }
2073
2074 void memdebug(int i)
2075 {
2076   //printf("TRACE: count=%d next=%d (checksum %x) lo=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[LOREG]>>32),(int)reg[LOREG]);
2077   //printf("TRACE: count=%d next=%d (rchecksum %x)\n",Count,next_interupt,rchecksum());
2078   //rlist();
2079   //if(tracedebug) {
2080   //if(Count>=-2084597794) {
2081   if((signed int)Count>=-2084597794&&(signed int)Count<0) {
2082   //if(0) {
2083     printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
2084     //printf("TRACE: count=%d next=%d (checksum %x) Status=%x\n",Count,next_interupt,mchecksum(),Status);
2085     //printf("TRACE: count=%d next=%d (checksum %x) hi=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[HIREG]>>32),(int)reg[HIREG]);
2086     rlist();
2087     #ifdef __i386__
2088     printf("TRACE: %x\n",(&i)[-1]);
2089     #endif
2090     #ifdef __arm__
2091     int j;
2092     printf("TRACE: %x \n",(&j)[10]);
2093     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]);
2094     #endif
2095     //fflush(stdout);
2096   }
2097   //printf("TRACE: %x\n",(&i)[-1]);
2098 }
2099
2100 void tlb_debug(u_int cause, u_int addr, u_int iaddr)
2101 {
2102   printf("TLB Exception: instruction=%x addr=%x cause=%x\n",iaddr, addr, cause);
2103 }
2104
2105 void alu_assemble(int i,struct regstat *i_regs)
2106 {
2107   if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
2108     if(rt1[i]) {
2109       signed char s1,s2,t;
2110       t=get_reg(i_regs->regmap,rt1[i]);
2111       if(t>=0) {
2112         s1=get_reg(i_regs->regmap,rs1[i]);
2113         s2=get_reg(i_regs->regmap,rs2[i]);
2114         if(rs1[i]&&rs2[i]) {
2115           assert(s1>=0);
2116           assert(s2>=0);
2117           if(opcode2[i]&2) emit_sub(s1,s2,t);
2118           else emit_add(s1,s2,t);
2119         }
2120         else if(rs1[i]) {
2121           if(s1>=0) emit_mov(s1,t);
2122           else emit_loadreg(rs1[i],t);
2123         }
2124         else if(rs2[i]) {
2125           if(s2>=0) {
2126             if(opcode2[i]&2) emit_neg(s2,t);
2127             else emit_mov(s2,t);
2128           }
2129           else {
2130             emit_loadreg(rs2[i],t);
2131             if(opcode2[i]&2) emit_neg(t,t);
2132           }
2133         }
2134         else emit_zeroreg(t);
2135       }
2136     }
2137   }
2138   if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2139     if(rt1[i]) {
2140       signed char s1l,s2l,s1h,s2h,tl,th;
2141       tl=get_reg(i_regs->regmap,rt1[i]);
2142       th=get_reg(i_regs->regmap,rt1[i]|64);
2143       if(tl>=0) {
2144         s1l=get_reg(i_regs->regmap,rs1[i]);
2145         s2l=get_reg(i_regs->regmap,rs2[i]);
2146         s1h=get_reg(i_regs->regmap,rs1[i]|64);
2147         s2h=get_reg(i_regs->regmap,rs2[i]|64);
2148         if(rs1[i]&&rs2[i]) {
2149           assert(s1l>=0);
2150           assert(s2l>=0);
2151           if(opcode2[i]&2) emit_subs(s1l,s2l,tl);
2152           else emit_adds(s1l,s2l,tl);
2153           if(th>=0) {
2154             #ifdef INVERTED_CARRY
2155             if(opcode2[i]&2) {if(s1h!=th) emit_mov(s1h,th);emit_sbb(th,s2h);}
2156             #else
2157             if(opcode2[i]&2) emit_sbc(s1h,s2h,th);
2158             #endif
2159             else emit_add(s1h,s2h,th);
2160           }
2161         }
2162         else if(rs1[i]) {
2163           if(s1l>=0) emit_mov(s1l,tl);
2164           else emit_loadreg(rs1[i],tl);
2165           if(th>=0) {
2166             if(s1h>=0) emit_mov(s1h,th);
2167             else emit_loadreg(rs1[i]|64,th);
2168           }
2169         }
2170         else if(rs2[i]) {
2171           if(s2l>=0) {
2172             if(opcode2[i]&2) emit_negs(s2l,tl);
2173             else emit_mov(s2l,tl);
2174           }
2175           else {
2176             emit_loadreg(rs2[i],tl);
2177             if(opcode2[i]&2) emit_negs(tl,tl);
2178           }
2179           if(th>=0) {
2180             #ifdef INVERTED_CARRY
2181             if(s2h>=0) emit_mov(s2h,th);
2182             else emit_loadreg(rs2[i]|64,th);
2183             if(opcode2[i]&2) {
2184               emit_adcimm(-1,th); // x86 has inverted carry flag
2185               emit_not(th,th);
2186             }
2187             #else
2188             if(opcode2[i]&2) {
2189               if(s2h>=0) emit_rscimm(s2h,0,th);
2190               else {
2191                 emit_loadreg(rs2[i]|64,th);
2192                 emit_rscimm(th,0,th);
2193               }
2194             }else{
2195               if(s2h>=0) emit_mov(s2h,th);
2196               else emit_loadreg(rs2[i]|64,th);
2197             }
2198             #endif
2199           }
2200         }
2201         else {
2202           emit_zeroreg(tl);
2203           if(th>=0) emit_zeroreg(th);
2204         }
2205       }
2206     }
2207   }
2208   if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
2209     if(rt1[i]) {
2210       signed char s1l,s1h,s2l,s2h,t;
2211       if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1))
2212       {
2213         t=get_reg(i_regs->regmap,rt1[i]);
2214         //assert(t>=0);
2215         if(t>=0) {
2216           s1l=get_reg(i_regs->regmap,rs1[i]);
2217           s1h=get_reg(i_regs->regmap,rs1[i]|64);
2218           s2l=get_reg(i_regs->regmap,rs2[i]);
2219           s2h=get_reg(i_regs->regmap,rs2[i]|64);
2220           if(rs2[i]==0) // rx<r0
2221           {
2222             assert(s1h>=0);
2223             if(opcode2[i]==0x2a) // SLT
2224               emit_shrimm(s1h,31,t);
2225             else // SLTU (unsigned can not be less than zero)
2226               emit_zeroreg(t);
2227           }
2228           else if(rs1[i]==0) // r0<rx
2229           {
2230             assert(s2h>=0);
2231             if(opcode2[i]==0x2a) // SLT
2232               emit_set_gz64_32(s2h,s2l,t);
2233             else // SLTU (set if not zero)
2234               emit_set_nz64_32(s2h,s2l,t);
2235           }
2236           else {
2237             assert(s1l>=0);assert(s1h>=0);
2238             assert(s2l>=0);assert(s2h>=0);
2239             if(opcode2[i]==0x2a) // SLT
2240               emit_set_if_less64_32(s1h,s1l,s2h,s2l,t);
2241             else // SLTU
2242               emit_set_if_carry64_32(s1h,s1l,s2h,s2l,t);
2243           }
2244         }
2245       } else {
2246         t=get_reg(i_regs->regmap,rt1[i]);
2247         //assert(t>=0);
2248         if(t>=0) {
2249           s1l=get_reg(i_regs->regmap,rs1[i]);
2250           s2l=get_reg(i_regs->regmap,rs2[i]);
2251           if(rs2[i]==0) // rx<r0
2252           {
2253             assert(s1l>=0);
2254             if(opcode2[i]==0x2a) // SLT
2255               emit_shrimm(s1l,31,t);
2256             else // SLTU (unsigned can not be less than zero)
2257               emit_zeroreg(t);
2258           }
2259           else if(rs1[i]==0) // r0<rx
2260           {
2261             assert(s2l>=0);
2262             if(opcode2[i]==0x2a) // SLT
2263               emit_set_gz32(s2l,t);
2264             else // SLTU (set if not zero)
2265               emit_set_nz32(s2l,t);
2266           }
2267           else{
2268             assert(s1l>=0);assert(s2l>=0);
2269             if(opcode2[i]==0x2a) // SLT
2270               emit_set_if_less32(s1l,s2l,t);
2271             else // SLTU
2272               emit_set_if_carry32(s1l,s2l,t);
2273           }
2274         }
2275       }
2276     }
2277   }
2278   if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
2279     if(rt1[i]) {
2280       signed char s1l,s1h,s2l,s2h,th,tl;
2281       tl=get_reg(i_regs->regmap,rt1[i]);
2282       th=get_reg(i_regs->regmap,rt1[i]|64);
2283       if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1)&&th>=0)
2284       {
2285         assert(tl>=0);
2286         if(tl>=0) {
2287           s1l=get_reg(i_regs->regmap,rs1[i]);
2288           s1h=get_reg(i_regs->regmap,rs1[i]|64);
2289           s2l=get_reg(i_regs->regmap,rs2[i]);
2290           s2h=get_reg(i_regs->regmap,rs2[i]|64);
2291           if(rs1[i]&&rs2[i]) {
2292             assert(s1l>=0);assert(s1h>=0);
2293             assert(s2l>=0);assert(s2h>=0);
2294             if(opcode2[i]==0x24) { // AND
2295               emit_and(s1l,s2l,tl);
2296               emit_and(s1h,s2h,th);
2297             } else
2298             if(opcode2[i]==0x25) { // OR
2299               emit_or(s1l,s2l,tl);
2300               emit_or(s1h,s2h,th);
2301             } else
2302             if(opcode2[i]==0x26) { // XOR
2303               emit_xor(s1l,s2l,tl);
2304               emit_xor(s1h,s2h,th);
2305             } else
2306             if(opcode2[i]==0x27) { // NOR
2307               emit_or(s1l,s2l,tl);
2308               emit_or(s1h,s2h,th);
2309               emit_not(tl,tl);
2310               emit_not(th,th);
2311             }
2312           }
2313           else
2314           {
2315             if(opcode2[i]==0x24) { // AND
2316               emit_zeroreg(tl);
2317               emit_zeroreg(th);
2318             } else
2319             if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2320               if(rs1[i]){
2321                 if(s1l>=0) emit_mov(s1l,tl);
2322                 else emit_loadreg(rs1[i],tl);
2323                 if(s1h>=0) emit_mov(s1h,th);
2324                 else emit_loadreg(rs1[i]|64,th);
2325               }
2326               else
2327               if(rs2[i]){
2328                 if(s2l>=0) emit_mov(s2l,tl);
2329                 else emit_loadreg(rs2[i],tl);
2330                 if(s2h>=0) emit_mov(s2h,th);
2331                 else emit_loadreg(rs2[i]|64,th);
2332               }
2333               else{
2334                 emit_zeroreg(tl);
2335                 emit_zeroreg(th);
2336               }
2337             } else
2338             if(opcode2[i]==0x27) { // NOR
2339               if(rs1[i]){
2340                 if(s1l>=0) emit_not(s1l,tl);
2341                 else{
2342                   emit_loadreg(rs1[i],tl);
2343                   emit_not(tl,tl);
2344                 }
2345                 if(s1h>=0) emit_not(s1h,th);
2346                 else{
2347                   emit_loadreg(rs1[i]|64,th);
2348                   emit_not(th,th);
2349                 }
2350               }
2351               else
2352               if(rs2[i]){
2353                 if(s2l>=0) emit_not(s2l,tl);
2354                 else{
2355                   emit_loadreg(rs2[i],tl);
2356                   emit_not(tl,tl);
2357                 }
2358                 if(s2h>=0) emit_not(s2h,th);
2359                 else{
2360                   emit_loadreg(rs2[i]|64,th);
2361                   emit_not(th,th);
2362                 }
2363               }
2364               else {
2365                 emit_movimm(-1,tl);
2366                 emit_movimm(-1,th);
2367               }
2368             }
2369           }
2370         }
2371       }
2372       else
2373       {
2374         // 32 bit
2375         if(tl>=0) {
2376           s1l=get_reg(i_regs->regmap,rs1[i]);
2377           s2l=get_reg(i_regs->regmap,rs2[i]);
2378           if(rs1[i]&&rs2[i]) {
2379             assert(s1l>=0);
2380             assert(s2l>=0);
2381             if(opcode2[i]==0x24) { // AND
2382               emit_and(s1l,s2l,tl);
2383             } else
2384             if(opcode2[i]==0x25) { // OR
2385               emit_or(s1l,s2l,tl);
2386             } else
2387             if(opcode2[i]==0x26) { // XOR
2388               emit_xor(s1l,s2l,tl);
2389             } else
2390             if(opcode2[i]==0x27) { // NOR
2391               emit_or(s1l,s2l,tl);
2392               emit_not(tl,tl);
2393             }
2394           }
2395           else
2396           {
2397             if(opcode2[i]==0x24) { // AND
2398               emit_zeroreg(tl);
2399             } else
2400             if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2401               if(rs1[i]){
2402                 if(s1l>=0) emit_mov(s1l,tl);
2403                 else emit_loadreg(rs1[i],tl); // CHECK: regmap_entry?
2404               }
2405               else
2406               if(rs2[i]){
2407                 if(s2l>=0) emit_mov(s2l,tl);
2408                 else emit_loadreg(rs2[i],tl); // CHECK: regmap_entry?
2409               }
2410               else emit_zeroreg(tl);
2411             } else
2412             if(opcode2[i]==0x27) { // NOR
2413               if(rs1[i]){
2414                 if(s1l>=0) emit_not(s1l,tl);
2415                 else {
2416                   emit_loadreg(rs1[i],tl);
2417                   emit_not(tl,tl);
2418                 }
2419               }
2420               else
2421               if(rs2[i]){
2422                 if(s2l>=0) emit_not(s2l,tl);
2423                 else {
2424                   emit_loadreg(rs2[i],tl);
2425                   emit_not(tl,tl);
2426                 }
2427               }
2428               else emit_movimm(-1,tl);
2429             }
2430           }
2431         }
2432       }
2433     }
2434   }
2435 }
2436
2437 void imm16_assemble(int i,struct regstat *i_regs)
2438 {
2439   if (opcode[i]==0x0f) { // LUI
2440     if(rt1[i]) {
2441       signed char t;
2442       t=get_reg(i_regs->regmap,rt1[i]);
2443       //assert(t>=0);
2444       if(t>=0) {
2445         if(!((i_regs->isconst>>t)&1))
2446           emit_movimm(imm[i]<<16,t);
2447       }
2448     }
2449   }
2450   if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
2451     if(rt1[i]) {
2452       signed char s,t;
2453       t=get_reg(i_regs->regmap,rt1[i]);
2454       s=get_reg(i_regs->regmap,rs1[i]);
2455       if(rs1[i]) {
2456         //assert(t>=0);
2457         //assert(s>=0);
2458         if(t>=0) {
2459           if(!((i_regs->isconst>>t)&1)) {
2460             if(s<0) {
2461               if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2462               emit_addimm(t,imm[i],t);
2463             }else{
2464               if(!((i_regs->wasconst>>s)&1))
2465                 emit_addimm(s,imm[i],t);
2466               else
2467                 emit_movimm(constmap[i][s]+imm[i],t);
2468             }
2469           }
2470         }
2471       } else {
2472         if(t>=0) {
2473           if(!((i_regs->isconst>>t)&1))
2474             emit_movimm(imm[i],t);
2475         }
2476       }
2477     }
2478   }
2479   if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
2480     if(rt1[i]) {
2481       signed char sh,sl,th,tl;
2482       th=get_reg(i_regs->regmap,rt1[i]|64);
2483       tl=get_reg(i_regs->regmap,rt1[i]);
2484       sh=get_reg(i_regs->regmap,rs1[i]|64);
2485       sl=get_reg(i_regs->regmap,rs1[i]);
2486       if(tl>=0) {
2487         if(rs1[i]) {
2488           assert(sh>=0);
2489           assert(sl>=0);
2490           if(th>=0) {
2491             emit_addimm64_32(sh,sl,imm[i],th,tl);
2492           }
2493           else {
2494             emit_addimm(sl,imm[i],tl);
2495           }
2496         } else {
2497           emit_movimm(imm[i],tl);
2498           if(th>=0) emit_movimm(((signed int)imm[i])>>31,th);
2499         }
2500       }
2501     }
2502   }
2503   else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
2504     if(rt1[i]) {
2505       //assert(rs1[i]!=0); // r0 might be valid, but it's probably a bug
2506       signed char sh,sl,t;
2507       t=get_reg(i_regs->regmap,rt1[i]);
2508       sh=get_reg(i_regs->regmap,rs1[i]|64);
2509       sl=get_reg(i_regs->regmap,rs1[i]);
2510       //assert(t>=0);
2511       if(t>=0) {
2512         if(rs1[i]>0) {
2513           if(sh<0) assert((i_regs->was32>>rs1[i])&1);
2514           if(sh<0||((i_regs->was32>>rs1[i])&1)) {
2515             if(opcode[i]==0x0a) { // SLTI
2516               if(sl<0) {
2517                 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2518                 emit_slti32(t,imm[i],t);
2519               }else{
2520                 emit_slti32(sl,imm[i],t);
2521               }
2522             }
2523             else { // SLTIU
2524               if(sl<0) {
2525                 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2526                 emit_sltiu32(t,imm[i],t);
2527               }else{
2528                 emit_sltiu32(sl,imm[i],t);
2529               }
2530             }
2531           }else{ // 64-bit
2532             assert(sl>=0);
2533             if(opcode[i]==0x0a) // SLTI
2534               emit_slti64_32(sh,sl,imm[i],t);
2535             else // SLTIU
2536               emit_sltiu64_32(sh,sl,imm[i],t);
2537           }
2538         }else{
2539           // SLTI(U) with r0 is just stupid,
2540           // nonetheless examples can be found
2541           if(opcode[i]==0x0a) // SLTI
2542             if(0<imm[i]) emit_movimm(1,t);
2543             else emit_zeroreg(t);
2544           else // SLTIU
2545           {
2546             if(imm[i]) emit_movimm(1,t);
2547             else emit_zeroreg(t);
2548           }
2549         }
2550       }
2551     }
2552   }
2553   else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
2554     if(rt1[i]) {
2555       signed char sh,sl,th,tl;
2556       th=get_reg(i_regs->regmap,rt1[i]|64);
2557       tl=get_reg(i_regs->regmap,rt1[i]);
2558       sh=get_reg(i_regs->regmap,rs1[i]|64);
2559       sl=get_reg(i_regs->regmap,rs1[i]);
2560       if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2561         if(opcode[i]==0x0c) //ANDI
2562         {
2563           if(rs1[i]) {
2564             if(sl<0) {
2565               if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2566               emit_andimm(tl,imm[i],tl);
2567             }else{
2568               if(!((i_regs->wasconst>>sl)&1))
2569                 emit_andimm(sl,imm[i],tl);
2570               else
2571                 emit_movimm(constmap[i][sl]&imm[i],tl);
2572             }
2573           }
2574           else
2575             emit_zeroreg(tl);
2576           if(th>=0) emit_zeroreg(th);
2577         }
2578         else
2579         {
2580           if(rs1[i]) {
2581             if(sl<0) {
2582               if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2583             }
2584             if(th>=0) {
2585               if(sh<0) {
2586                 emit_loadreg(rs1[i]|64,th);
2587               }else{
2588                 emit_mov(sh,th);
2589               }
2590             }
2591             if(opcode[i]==0x0d) //ORI
2592             if(sl<0) {
2593               emit_orimm(tl,imm[i],tl);
2594             }else{
2595               if(!((i_regs->wasconst>>sl)&1))
2596                 emit_orimm(sl,imm[i],tl);
2597               else
2598                 emit_movimm(constmap[i][sl]|imm[i],tl);
2599             }
2600             if(opcode[i]==0x0e) //XORI
2601             if(sl<0) {
2602               emit_xorimm(tl,imm[i],tl);
2603             }else{
2604               if(!((i_regs->wasconst>>sl)&1))
2605                 emit_xorimm(sl,imm[i],tl);
2606               else
2607                 emit_movimm(constmap[i][sl]^imm[i],tl);
2608             }
2609           }
2610           else {
2611             emit_movimm(imm[i],tl);
2612             if(th>=0) emit_zeroreg(th);
2613           }
2614         }
2615       }
2616     }
2617   }
2618 }
2619
2620 void shiftimm_assemble(int i,struct regstat *i_regs)
2621 {
2622   if(opcode2[i]<=0x3) // SLL/SRL/SRA
2623   {
2624     if(rt1[i]) {
2625       signed char s,t;
2626       t=get_reg(i_regs->regmap,rt1[i]);
2627       s=get_reg(i_regs->regmap,rs1[i]);
2628       //assert(t>=0);
2629       if(t>=0){
2630         if(rs1[i]==0)
2631         {
2632           emit_zeroreg(t);
2633         }
2634         else
2635         {
2636           if(s<0&&i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2637           if(imm[i]) {
2638             if(opcode2[i]==0) // SLL
2639             {
2640               emit_shlimm(s<0?t:s,imm[i],t);
2641             }
2642             if(opcode2[i]==2) // SRL
2643             {
2644               emit_shrimm(s<0?t:s,imm[i],t);
2645             }
2646             if(opcode2[i]==3) // SRA
2647             {
2648               emit_sarimm(s<0?t:s,imm[i],t);
2649             }
2650           }else{
2651             // Shift by zero
2652             if(s>=0 && s!=t) emit_mov(s,t);
2653           }
2654         }
2655       }
2656       //emit_storereg(rt1[i],t); //DEBUG
2657     }
2658   }
2659   if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
2660   {
2661     if(rt1[i]) {
2662       signed char sh,sl,th,tl;
2663       th=get_reg(i_regs->regmap,rt1[i]|64);
2664       tl=get_reg(i_regs->regmap,rt1[i]);
2665       sh=get_reg(i_regs->regmap,rs1[i]|64);
2666       sl=get_reg(i_regs->regmap,rs1[i]);
2667       if(tl>=0) {
2668         if(rs1[i]==0)
2669         {
2670           emit_zeroreg(tl);
2671           if(th>=0) emit_zeroreg(th);
2672         }
2673         else
2674         {
2675           assert(sl>=0);
2676           assert(sh>=0);
2677           if(imm[i]) {
2678             if(opcode2[i]==0x38) // DSLL
2679             {
2680               if(th>=0) emit_shldimm(sh,sl,imm[i],th);
2681               emit_shlimm(sl,imm[i],tl);
2682             }
2683             if(opcode2[i]==0x3a) // DSRL
2684             {
2685               emit_shrdimm(sl,sh,imm[i],tl);
2686               if(th>=0) emit_shrimm(sh,imm[i],th);
2687             }
2688             if(opcode2[i]==0x3b) // DSRA
2689             {
2690               emit_shrdimm(sl,sh,imm[i],tl);
2691               if(th>=0) emit_sarimm(sh,imm[i],th);
2692             }
2693           }else{
2694             // Shift by zero
2695             if(sl!=tl) emit_mov(sl,tl);
2696             if(th>=0&&sh!=th) emit_mov(sh,th);
2697           }
2698         }
2699       }
2700     }
2701   }
2702   if(opcode2[i]==0x3c) // DSLL32
2703   {
2704     if(rt1[i]) {
2705       signed char sl,tl,th;
2706       tl=get_reg(i_regs->regmap,rt1[i]);
2707       th=get_reg(i_regs->regmap,rt1[i]|64);
2708       sl=get_reg(i_regs->regmap,rs1[i]);
2709       if(th>=0||tl>=0){
2710         assert(tl>=0);
2711         assert(th>=0);
2712         assert(sl>=0);
2713         emit_mov(sl,th);
2714         emit_zeroreg(tl);
2715         if(imm[i]>32)
2716         {
2717           emit_shlimm(th,imm[i]&31,th);
2718         }
2719       }
2720     }
2721   }
2722   if(opcode2[i]==0x3e) // DSRL32
2723   {
2724     if(rt1[i]) {
2725       signed char sh,tl,th;
2726       tl=get_reg(i_regs->regmap,rt1[i]);
2727       th=get_reg(i_regs->regmap,rt1[i]|64);
2728       sh=get_reg(i_regs->regmap,rs1[i]|64);
2729       if(tl>=0){
2730         assert(sh>=0);
2731         emit_mov(sh,tl);
2732         if(th>=0) emit_zeroreg(th);
2733         if(imm[i]>32)
2734         {
2735           emit_shrimm(tl,imm[i]&31,tl);
2736         }
2737       }
2738     }
2739   }
2740   if(opcode2[i]==0x3f) // DSRA32
2741   {
2742     if(rt1[i]) {
2743       signed char sh,tl;
2744       tl=get_reg(i_regs->regmap,rt1[i]);
2745       sh=get_reg(i_regs->regmap,rs1[i]|64);
2746       if(tl>=0){
2747         assert(sh>=0);
2748         emit_mov(sh,tl);
2749         if(imm[i]>32)
2750         {
2751           emit_sarimm(tl,imm[i]&31,tl);
2752         }
2753       }
2754     }
2755   }
2756 }
2757
2758 #ifndef shift_assemble
2759 void shift_assemble(int i,struct regstat *i_regs)
2760 {
2761   printf("Need shift_assemble for this architecture.\n");
2762   exit(1);
2763 }
2764 #endif
2765
2766 void load_assemble(int i,struct regstat *i_regs)
2767 {
2768   int s,th,tl,addr,map=-1;
2769   int offset;
2770   int jaddr=0;
2771   int memtarget=0,c=0;
2772   u_int hr,reglist=0;
2773   th=get_reg(i_regs->regmap,rt1[i]|64);
2774   tl=get_reg(i_regs->regmap,rt1[i]);
2775   s=get_reg(i_regs->regmap,rs1[i]);
2776   offset=imm[i];
2777   for(hr=0;hr<HOST_REGS;hr++) {
2778     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2779   }
2780   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2781   if(s>=0) {
2782     c=(i_regs->wasconst>>s)&1;
2783     memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2784     if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
2785   }
2786   //printf("load_assemble: c=%d\n",c);
2787   //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2788   // FIXME: Even if the load is a NOP, we should check for pagefaults...
2789 #ifdef PCSX
2790   if(tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80)
2791     ||rt1[i]==0) {
2792       // could be FIFO, must perform the read
2793       // ||dummy read
2794       assem_debug("(forced read)\n");
2795       tl=get_reg(i_regs->regmap,-1);
2796       assert(tl>=0);
2797   }
2798 #endif
2799   if(offset||s<0||c) addr=tl;
2800   else addr=s;
2801   //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2802  if(tl>=0) {
2803   //printf("load_assemble: c=%d\n",c);
2804   //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2805   assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2806   reglist&=~(1<<tl);
2807   if(th>=0) reglist&=~(1<<th);
2808   if(!using_tlb) {
2809     if(!c) {
2810       #ifdef RAM_OFFSET
2811       map=get_reg(i_regs->regmap,ROREG);
2812       if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG);
2813       #endif
2814 //#define R29_HACK 1
2815       #ifdef R29_HACK
2816       // Strmnnrmn's speed hack
2817       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2818       #endif
2819       {
2820         emit_cmpimm(addr,RAM_SIZE);
2821         jaddr=(int)out;
2822         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2823         // Hint to branch predictor that the branch is unlikely to be taken
2824         if(rs1[i]>=28)
2825           emit_jno_unlikely(0);
2826         else
2827         #endif
2828         emit_jno(0);
2829       }
2830     }
2831   }else{ // using tlb
2832     int x=0;
2833     if (opcode[i]==0x20||opcode[i]==0x24) x=3; // LB/LBU
2834     if (opcode[i]==0x21||opcode[i]==0x25) x=2; // LH/LHU
2835     map=get_reg(i_regs->regmap,TLREG);
2836     assert(map>=0);
2837     map=do_tlb_r(addr,tl,map,x,-1,-1,c,constmap[i][s]+offset);
2838     do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr);
2839   }
2840   int dummy=(rt1[i]==0)||(tl!=get_reg(i_regs->regmap,rt1[i])); // ignore loads to r0 and unneeded reg
2841   if (opcode[i]==0x20) { // LB
2842     if(!c||memtarget) {
2843       if(!dummy) {
2844         #ifdef HOST_IMM_ADDR32
2845         if(c)
2846           emit_movsbl_tlb((constmap[i][s]+offset)^3,map,tl);
2847         else
2848         #endif
2849         {
2850           //emit_xorimm(addr,3,tl);
2851           //gen_tlb_addr_r(tl,map);
2852           //emit_movsbl_indexed((int)rdram-0x80000000,tl,tl);
2853           int x=0,a=tl;
2854 #ifdef BIG_ENDIAN_MIPS
2855           if(!c) emit_xorimm(addr,3,tl);
2856           else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2857 #else
2858           if(!c) a=addr;
2859 #endif
2860           emit_movsbl_indexed_tlb(x,a,map,tl);
2861         }
2862       }
2863       if(jaddr)
2864         add_stub(LOADB_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2865     }
2866     else
2867       inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2868   }
2869   if (opcode[i]==0x21) { // LH
2870     if(!c||memtarget) {
2871       if(!dummy) {
2872         #ifdef HOST_IMM_ADDR32
2873         if(c)
2874           emit_movswl_tlb((constmap[i][s]+offset)^2,map,tl);
2875         else
2876         #endif
2877         {
2878           int x=0,a=tl;
2879 #ifdef BIG_ENDIAN_MIPS
2880           if(!c) emit_xorimm(addr,2,tl);
2881           else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2882 #else
2883           if(!c) a=addr;
2884 #endif
2885           //#ifdef
2886           //emit_movswl_indexed_tlb(x,tl,map,tl);
2887           //else
2888           if(map>=0) {
2889             gen_tlb_addr_r(a,map);
2890             emit_movswl_indexed(x,a,tl);
2891           }else{
2892             #ifdef RAM_OFFSET
2893             emit_movswl_indexed(x,a,tl);
2894             #else
2895             emit_movswl_indexed((int)rdram-0x80000000+x,a,tl);
2896             #endif
2897           }
2898         }
2899       }
2900       if(jaddr)
2901         add_stub(LOADH_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2902     }
2903     else
2904       inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2905   }
2906   if (opcode[i]==0x23) { // LW
2907     if(!c||memtarget) {
2908       if(!dummy) {
2909         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2910         #ifdef HOST_IMM_ADDR32
2911         if(c)
2912           emit_readword_tlb(constmap[i][s]+offset,map,tl);
2913         else
2914         #endif
2915         emit_readword_indexed_tlb(0,addr,map,tl);
2916       }
2917       if(jaddr)
2918         add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2919     }
2920     else
2921       inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2922   }
2923   if (opcode[i]==0x24) { // LBU
2924     if(!c||memtarget) {
2925       if(!dummy) {
2926         #ifdef HOST_IMM_ADDR32
2927         if(c)
2928           emit_movzbl_tlb((constmap[i][s]+offset)^3,map,tl);
2929         else
2930         #endif
2931         {
2932           //emit_xorimm(addr,3,tl);
2933           //gen_tlb_addr_r(tl,map);
2934           //emit_movzbl_indexed((int)rdram-0x80000000,tl,tl);
2935           int x=0,a=tl;
2936 #ifdef BIG_ENDIAN_MIPS
2937           if(!c) emit_xorimm(addr,3,tl);
2938           else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2939 #else
2940           if(!c) a=addr;
2941 #endif
2942           emit_movzbl_indexed_tlb(x,a,map,tl);
2943         }
2944       }
2945       if(jaddr)
2946         add_stub(LOADBU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2947     }
2948     else
2949       inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2950   }
2951   if (opcode[i]==0x25) { // LHU
2952     if(!c||memtarget) {
2953       if(!dummy) {
2954         #ifdef HOST_IMM_ADDR32
2955         if(c)
2956           emit_movzwl_tlb((constmap[i][s]+offset)^2,map,tl);
2957         else
2958         #endif
2959         {
2960           int x=0,a=tl;
2961 #ifdef BIG_ENDIAN_MIPS
2962           if(!c) emit_xorimm(addr,2,tl);
2963           else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2964 #else
2965           if(!c) a=addr;
2966 #endif
2967           //#ifdef
2968           //emit_movzwl_indexed_tlb(x,tl,map,tl);
2969           //#else
2970           if(map>=0) {
2971             gen_tlb_addr_r(a,map);
2972             emit_movzwl_indexed(x,a,tl);
2973           }else{
2974             #ifdef RAM_OFFSET
2975             emit_movzwl_indexed(x,a,tl);
2976             #else
2977             emit_movzwl_indexed((int)rdram-0x80000000+x,a,tl);
2978             #endif
2979           }
2980         }
2981       }
2982       if(jaddr)
2983         add_stub(LOADHU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2984     }
2985     else
2986       inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2987   }
2988   if (opcode[i]==0x27) { // LWU
2989     assert(th>=0);
2990     if(!c||memtarget) {
2991       if(!dummy) {
2992         //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2993         #ifdef HOST_IMM_ADDR32
2994         if(c)
2995           emit_readword_tlb(constmap[i][s]+offset,map,tl);
2996         else
2997         #endif
2998         emit_readword_indexed_tlb(0,addr,map,tl);
2999       }
3000       if(jaddr)
3001         add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3002     }
3003     else {
3004       inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
3005     }
3006     emit_zeroreg(th);
3007   }
3008   if (opcode[i]==0x37) { // LD
3009     if(!c||memtarget) {
3010       if(!dummy) {
3011         //gen_tlb_addr_r(tl,map);
3012         //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,addr,th);
3013         //emit_readword_indexed((int)rdram-0x7FFFFFFC,addr,tl);
3014         #ifdef HOST_IMM_ADDR32
3015         if(c)
3016           emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
3017         else
3018         #endif
3019         emit_readdword_indexed_tlb(0,addr,map,th,tl);
3020       }
3021       if(jaddr)
3022         add_stub(LOADD_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3023     }
3024     else
3025       inline_readstub(LOADD_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
3026   }
3027  }
3028   //emit_storereg(rt1[i],tl); // DEBUG
3029   //if(opcode[i]==0x23)
3030   //if(opcode[i]==0x24)
3031   //if(opcode[i]==0x23||opcode[i]==0x24)
3032   /*if(opcode[i]==0x21||opcode[i]==0x23||opcode[i]==0x24)
3033   {
3034     //emit_pusha();
3035     save_regs(0x100f);
3036         emit_readword((int)&last_count,ECX);
3037         #ifdef __i386__
3038         if(get_reg(i_regs->regmap,CCREG)<0)
3039           emit_loadreg(CCREG,HOST_CCREG);
3040         emit_add(HOST_CCREG,ECX,HOST_CCREG);
3041         emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3042         emit_writeword(HOST_CCREG,(int)&Count);
3043         #endif
3044         #ifdef __arm__
3045         if(get_reg(i_regs->regmap,CCREG)<0)
3046           emit_loadreg(CCREG,0);
3047         else
3048           emit_mov(HOST_CCREG,0);
3049         emit_add(0,ECX,0);
3050         emit_addimm(0,2*ccadj[i],0);
3051         emit_writeword(0,(int)&Count);
3052         #endif
3053     emit_call((int)memdebug);
3054     //emit_popa();
3055     restore_regs(0x100f);
3056   }/**/
3057 }
3058
3059 #ifndef loadlr_assemble
3060 void loadlr_assemble(int i,struct regstat *i_regs)
3061 {
3062   printf("Need loadlr_assemble for this architecture.\n");
3063   exit(1);
3064 }
3065 #endif
3066
3067 void store_assemble(int i,struct regstat *i_regs)
3068 {
3069   int s,th,tl,map=-1;
3070   int addr,temp;
3071   int offset;
3072   int jaddr=0,jaddr2,type;
3073   int memtarget=0,c=0;
3074   int agr=AGEN1+(i&1);
3075   u_int hr,reglist=0;
3076   th=get_reg(i_regs->regmap,rs2[i]|64);
3077   tl=get_reg(i_regs->regmap,rs2[i]);
3078   s=get_reg(i_regs->regmap,rs1[i]);
3079   temp=get_reg(i_regs->regmap,agr);
3080   if(temp<0) temp=get_reg(i_regs->regmap,-1);
3081   offset=imm[i];
3082   if(s>=0) {
3083     c=(i_regs->wasconst>>s)&1;
3084     memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3085     if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
3086   }
3087   assert(tl>=0);
3088   assert(temp>=0);
3089   for(hr=0;hr<HOST_REGS;hr++) {
3090     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3091   }
3092   if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3093   if(offset||s<0||c) addr=temp;
3094   else addr=s;
3095   if(!using_tlb) {
3096     if(!c) {
3097       #ifdef R29_HACK
3098       // Strmnnrmn's speed hack
3099       memtarget=1;
3100       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3101       #endif
3102       emit_cmpimm(addr,RAM_SIZE);
3103       #ifdef DESTRUCTIVE_SHIFT
3104       if(s==addr) emit_mov(s,temp);
3105       #endif
3106       #ifdef R29_HACK
3107       if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3108       #endif
3109       {
3110         jaddr=(int)out;
3111         #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
3112         // Hint to branch predictor that the branch is unlikely to be taken
3113         if(rs1[i]>=28)
3114           emit_jno_unlikely(0);
3115         else
3116         #endif
3117         emit_jno(0);
3118       }
3119     }
3120   }else{ // using tlb
3121     int x=0;
3122     if (opcode[i]==0x28) x=3; // SB
3123     if (opcode[i]==0x29) x=2; // SH
3124     map=get_reg(i_regs->regmap,TLREG);
3125     assert(map>=0);
3126     map=do_tlb_w(addr,temp,map,x,c,constmap[i][s]+offset);
3127     do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
3128   }
3129
3130   if (opcode[i]==0x28) { // SB
3131     if(!c||memtarget) {
3132       int x=0,a=temp;
3133 #ifdef BIG_ENDIAN_MIPS
3134       if(!c) emit_xorimm(addr,3,temp);
3135       else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
3136 #else
3137       if(!c) a=addr;
3138 #endif
3139       //gen_tlb_addr_w(temp,map);
3140       //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
3141       emit_writebyte_indexed_tlb(tl,x,a,map,a);
3142     }
3143     type=STOREB_STUB;
3144   }
3145   if (opcode[i]==0x29) { // SH
3146     if(!c||memtarget) {
3147       int x=0,a=temp;
3148 #ifdef BIG_ENDIAN_MIPS
3149       if(!c) emit_xorimm(addr,2,temp);
3150       else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
3151 #else
3152       if(!c) a=addr;
3153 #endif
3154       //#ifdef
3155       //emit_writehword_indexed_tlb(tl,x,temp,map,temp);
3156       //#else
3157       if(map>=0) {
3158         gen_tlb_addr_w(a,map);
3159         emit_writehword_indexed(tl,x,a);
3160       }else
3161         emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
3162     }
3163     type=STOREH_STUB;
3164   }
3165   if (opcode[i]==0x2B) { // SW
3166     if(!c||memtarget)
3167       //emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
3168       emit_writeword_indexed_tlb(tl,0,addr,map,temp);
3169     type=STOREW_STUB;
3170   }
3171   if (opcode[i]==0x3F) { // SD
3172     if(!c||memtarget) {
3173       if(rs2[i]) {
3174         assert(th>=0);
3175         //emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
3176         //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,addr);
3177         emit_writedword_indexed_tlb(th,tl,0,addr,map,temp);
3178       }else{
3179         // Store zero
3180         //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
3181         //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
3182         emit_writedword_indexed_tlb(tl,tl,0,addr,map,temp);
3183       }
3184     }
3185     type=STORED_STUB;
3186   }
3187   if(!using_tlb) {
3188     if(!c||memtarget) {
3189       #ifdef DESTRUCTIVE_SHIFT
3190       // The x86 shift operation is 'destructive'; it overwrites the
3191       // source register, so we need to make a copy first and use that.
3192       addr=temp;
3193       #endif
3194       #if defined(HOST_IMM8)
3195       int ir=get_reg(i_regs->regmap,INVCP);
3196       assert(ir>=0);
3197       emit_cmpmem_indexedsr12_reg(ir,addr,1);
3198       #else
3199       emit_cmpmem_indexedsr12_imm((int)invalid_code,addr,1);
3200       #endif
3201       #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3202       emit_callne(invalidate_addr_reg[addr]);
3203       #else
3204       jaddr2=(int)out;
3205       emit_jne(0);
3206       add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3207       #endif
3208     }
3209   }
3210   if(jaddr) {
3211     add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3212   } else if(c&&!memtarget) {
3213     inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
3214   }
3215   //if(opcode[i]==0x2B || opcode[i]==0x3F)
3216   //if(opcode[i]==0x2B || opcode[i]==0x28)
3217   //if(opcode[i]==0x2B || opcode[i]==0x29)
3218   //if(opcode[i]==0x2B)
3219   /*if(opcode[i]==0x2B || opcode[i]==0x28 || opcode[i]==0x29 || opcode[i]==0x3F)
3220   {
3221     //emit_pusha();
3222     save_regs(0x100f);
3223         emit_readword((int)&last_count,ECX);
3224         #ifdef __i386__
3225         if(get_reg(i_regs->regmap,CCREG)<0)
3226           emit_loadreg(CCREG,HOST_CCREG);
3227         emit_add(HOST_CCREG,ECX,HOST_CCREG);
3228         emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3229         emit_writeword(HOST_CCREG,(int)&Count);
3230         #endif
3231         #ifdef __arm__
3232         if(get_reg(i_regs->regmap,CCREG)<0)
3233           emit_loadreg(CCREG,0);
3234         else
3235           emit_mov(HOST_CCREG,0);
3236         emit_add(0,ECX,0);
3237         emit_addimm(0,2*ccadj[i],0);
3238         emit_writeword(0,(int)&Count);
3239         #endif
3240     emit_call((int)memdebug);
3241     //emit_popa();
3242     restore_regs(0x100f);
3243   }/**/
3244 }
3245
3246 void storelr_assemble(int i,struct regstat *i_regs)
3247 {
3248   int s,th,tl;
3249   int temp;
3250   int temp2;
3251   int offset;
3252   int jaddr=0,jaddr2;
3253   int case1,case2,case3;
3254   int done0,done1,done2;
3255   int memtarget,c=0;
3256   int agr=AGEN1+(i&1);
3257   u_int hr,reglist=0;
3258   th=get_reg(i_regs->regmap,rs2[i]|64);
3259   tl=get_reg(i_regs->regmap,rs2[i]);
3260   s=get_reg(i_regs->regmap,rs1[i]);
3261   temp=get_reg(i_regs->regmap,agr);
3262   if(temp<0) temp=get_reg(i_regs->regmap,-1);
3263   offset=imm[i];
3264   if(s>=0) {
3265     c=(i_regs->isconst>>s)&1;
3266     memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3267     if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
3268   }
3269   assert(tl>=0);
3270   for(hr=0;hr<HOST_REGS;hr++) {
3271     if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3272   }
3273   assert(temp>=0);
3274   if(!using_tlb) {
3275     if(!c) {
3276       emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3277       if(!offset&&s!=temp) emit_mov(s,temp);
3278       jaddr=(int)out;
3279       emit_jno(0);
3280     }
3281     else
3282     {
3283       if(!memtarget||!rs1[i]) {
3284         jaddr=(int)out;
3285         emit_jmp(0);
3286       }
3287     }
3288     #ifdef RAM_OFFSET
3289     int map=get_reg(i_regs->regmap,ROREG);
3290     if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG);
3291     gen_tlb_addr_w(temp,map);
3292     #else
3293     if((u_int)rdram!=0x80000000)
3294       emit_addimm_no_flags((u_int)rdram-(u_int)0x80000000,temp);
3295     #endif
3296   }else{ // using tlb
3297     int map=get_reg(i_regs->regmap,TLREG);
3298     assert(map>=0);
3299     map=do_tlb_w(c||s<0||offset?temp:s,temp,map,0,c,constmap[i][s]+offset);
3300     if(!c&&!offset&&s>=0) emit_mov(s,temp);
3301     do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
3302     if(!jaddr&&!memtarget) {
3303       jaddr=(int)out;
3304       emit_jmp(0);
3305     }
3306     gen_tlb_addr_w(temp,map);
3307   }
3308
3309   if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
3310     temp2=get_reg(i_regs->regmap,FTEMP);
3311     if(!rs2[i]) temp2=th=tl;
3312   }
3313
3314 #ifndef BIG_ENDIAN_MIPS
3315     emit_xorimm(temp,3,temp);
3316 #endif
3317   emit_testimm(temp,2);
3318   case2=(int)out;
3319   emit_jne(0);
3320   emit_testimm(temp,1);
3321   case1=(int)out;
3322   emit_jne(0);
3323   // 0
3324   if (opcode[i]==0x2A) { // SWL
3325     emit_writeword_indexed(tl,0,temp);
3326   }
3327   if (opcode[i]==0x2E) { // SWR
3328     emit_writebyte_indexed(tl,3,temp);
3329   }
3330   if (opcode[i]==0x2C) { // SDL
3331     emit_writeword_indexed(th,0,temp);
3332     if(rs2[i]) emit_mov(tl,temp2);
3333   }
3334   if (opcode[i]==0x2D) { // SDR
3335     emit_writebyte_indexed(tl,3,temp);
3336     if(rs2[i]) emit_shldimm(th,tl,24,temp2);
3337   }
3338   done0=(int)out;
3339   emit_jmp(0);
3340   // 1
3341   set_jump_target(case1,(int)out);
3342   if (opcode[i]==0x2A) { // SWL
3343     // Write 3 msb into three least significant bytes
3344     if(rs2[i]) emit_rorimm(tl,8,tl);
3345     emit_writehword_indexed(tl,-1,temp);
3346     if(rs2[i]) emit_rorimm(tl,16,tl);
3347     emit_writebyte_indexed(tl,1,temp);
3348     if(rs2[i]) emit_rorimm(tl,8,tl);
3349   }
3350   if (opcode[i]==0x2E) { // SWR
3351     // Write two lsb into two most significant bytes
3352     emit_writehword_indexed(tl,1,temp);
3353   }
3354   if (opcode[i]==0x2C) { // SDL
3355     if(rs2[i]) emit_shrdimm(tl,th,8,temp2);
3356     // Write 3 msb into three least significant bytes
3357     if(rs2[i]) emit_rorimm(th,8,th);
3358     emit_writehword_indexed(th,-1,temp);
3359     if(rs2[i]) emit_rorimm(th,16,th);
3360     emit_writebyte_indexed(th,1,temp);
3361     if(rs2[i]) emit_rorimm(th,8,th);
3362   }
3363   if (opcode[i]==0x2D) { // SDR
3364     if(rs2[i]) emit_shldimm(th,tl,16,temp2);
3365     // Write two lsb into two most significant bytes
3366     emit_writehword_indexed(tl,1,temp);
3367   }
3368   done1=(int)out;
3369   emit_jmp(0);
3370   // 2
3371   set_jump_target(case2,(int)out);
3372   emit_testimm(temp,1);
3373   case3=(int)out;
3374   emit_jne(0);
3375   if (opcode[i]==0x2A) { // SWL
3376     // Write two msb into two least significant bytes
3377     if(rs2[i]) emit_rorimm(tl,16,tl);
3378     emit_writehword_indexed(tl,-2,temp);
3379     if(rs2[i]) emit_rorimm(tl,16,tl);
3380   }
3381   if (opcode[i]==