1 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
2 * Mupen64plus - new_dynarec.c *
3 * Copyright (C) 2009-2010 Ari64 *
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. *
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. *
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 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
22 #include <stdint.h> //include for uint64_t
25 #include "emu_if.h" //emulator interface
30 #include "assem_x86.h"
33 #include "assem_x64.h"
36 #include "assem_arm.h"
40 #define MAX_OUTPUT_BLOCK_SIZE 262144
41 #define CLOCK_DIVIDER 2
45 signed char regmap_entry[HOST_REGS];
46 signed char regmap[HOST_REGS];
55 uint64_t constmap[HOST_REGS];
63 struct ll_entry *next;
69 char insn[MAXBLOCK][10];
70 u_char itype[MAXBLOCK];
71 u_char opcode[MAXBLOCK];
72 u_char opcode2[MAXBLOCK];
80 u_char dep1[MAXBLOCK];
81 u_char dep2[MAXBLOCK];
85 char likely[MAXBLOCK];
87 uint64_t unneeded_reg[MAXBLOCK];
88 uint64_t unneeded_reg_upper[MAXBLOCK];
89 uint64_t branch_unneeded_reg[MAXBLOCK];
90 uint64_t branch_unneeded_reg_upper[MAXBLOCK];
91 uint64_t p32[MAXBLOCK];
92 uint64_t pr32[MAXBLOCK];
93 signed char regmap_pre[MAXBLOCK][HOST_REGS];
94 signed char regmap[MAXBLOCK][HOST_REGS];
95 signed char regmap_entry[MAXBLOCK][HOST_REGS];
96 uint64_t constmap[MAXBLOCK][HOST_REGS];
97 uint64_t known_value[HOST_REGS];
99 struct regstat regs[MAXBLOCK];
100 struct regstat branch_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];
107 u_int instr_addr[MAXBLOCK];
108 u_int link_addr[MAXBLOCK][3];
110 u_int stubs[MAXBLOCK*3][8];
112 u_int literals[1024][2];
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)));
125 u_int stop_after_jal;
126 extern u_char restore_candidate[512];
127 extern int cycle_count;
129 /* registers that may be allocated */
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
138 #define FTEMP 38 // FPU/LDL/LDR temporary register
139 #define PTEMP 39 // Prefetch temporary register
140 #define TLREG 40 // TLB mapping offset
141 #define RHASH 41 // Return address hash
142 #define RHTBL 42 // Return address hash table address
143 #define RTEMP 43 // JR/JALR address register
145 #define AGEN1 44 // Address generation temporary register
146 #define AGEN2 45 // Address generation temporary register
147 #define MGEN1 46 // Maptable address generation temporary register
148 #define MGEN2 47 // Maptable address generation temporary register
149 #define BTREG 48 // Branch target temporary register
151 /* instruction types */
152 #define NOP 0 // No operation
153 #define LOAD 1 // Load
154 #define STORE 2 // Store
155 #define LOADLR 3 // Unaligned load
156 #define STORELR 4 // Unaligned store
157 #define MOV 5 // Move
158 #define ALU 6 // Arithmetic/logic
159 #define MULTDIV 7 // Multiply/divide
160 #define SHIFT 8 // Shift by register
161 #define SHIFTIMM 9// Shift by immediate
162 #define IMM16 10 // 16-bit immediate
163 #define RJUMP 11 // Unconditional jump to register
164 #define UJUMP 12 // Unconditional jump
165 #define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
166 #define SJUMP 14 // Conditional branch (regimm format)
167 #define COP0 15 // Coprocessor 0
168 #define COP1 16 // Coprocessor 1
169 #define C1LS 17 // Coprocessor 1 load/store
170 #define FJUMP 18 // Conditional branch (floating point)
171 #define FLOAT 19 // Floating point unit
172 #define FCONV 20 // Convert integer to float
173 #define FCOMP 21 // Floating point compare (sets FSREG)
174 #define SYSCALL 22// SYSCALL
175 #define OTHER 23 // Other
176 #define SPAN 24 // Branch/delay slot spans 2 pages
177 #define NI 25 // Not implemented
178 #define HLECALL 26// PCSX fake opcodes for HLE
179 #define COP2 27 // Coprocessor 2 move
180 #define C2LS 28 // Coprocessor 2 load/store
181 #define C2OP 29 // Coprocessor 2 operation
190 #define LOADBU_STUB 7
191 #define LOADHU_STUB 8
192 #define STOREB_STUB 9
193 #define STOREH_STUB 10
194 #define STOREW_STUB 11
195 #define STORED_STUB 12
196 #define STORELR_STUB 13
197 #define INVCODE_STUB 14
205 int new_recompile_block(int addr);
206 void *get_addr_ht(u_int vaddr);
207 void invalidate_block(u_int block);
208 void invalidate_addr(u_int addr);
209 void remove_hash(int vaddr);
212 void dyna_linker_ds();
214 void verify_code_vm();
215 void verify_code_ds();
218 void fp_exception_ds();
220 void jump_syscall_hle();
223 void new_dyna_leave();
228 void read_nomem_new();
229 void read_nomemb_new();
230 void read_nomemh_new();
231 void read_nomemd_new();
232 void write_nomem_new();
233 void write_nomemb_new();
234 void write_nomemh_new();
235 void write_nomemd_new();
236 void write_rdram_new();
237 void write_rdramb_new();
238 void write_rdramh_new();
239 void write_rdramd_new();
240 extern u_int memory_map[1048576];
242 // Needed by assembler
243 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32);
244 void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty);
245 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr);
246 void load_all_regs(signed char i_regmap[]);
247 void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
248 void load_regs_entry(int t);
249 void load_all_consts(signed char regmap[],int is32,u_int dirty,int i);
253 //#define DEBUG_CYCLE_COUNT 1
256 //#define assem_debug printf
257 //#define inv_debug printf
258 #define assem_debug nullf
259 #define inv_debug nullf
261 static void tlb_hacks()
265 if (strncmp((char *) ROM_HEADER->nom, "GOLDENEYE",9) == 0)
269 switch (ROM_HEADER->Country_code&0xFF)
281 // Unknown country code
285 u_int rom_addr=(u_int)rom;
287 // Since memory_map is 32-bit, on 64-bit systems the rom needs to be
288 // in the lower 4G of memory to use this hack. Copy it if necessary.
289 if((void *)rom>(void *)0xffffffff) {
290 munmap(ROM_COPY, 67108864);
291 if(mmap(ROM_COPY, 12582912,
292 PROT_READ | PROT_WRITE,
293 MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
294 -1, 0) <= 0) {printf("mmap() failed\n");}
295 memcpy(ROM_COPY,rom,12582912);
296 rom_addr=(u_int)ROM_COPY;
300 for(n=0x7F000;n<0x80000;n++) {
301 memory_map[n]=(((u_int)(rom_addr+addr-0x7F000000))>>2)|0x40000000;
308 static u_int get_page(u_int vaddr)
310 u_int page=(vaddr^0x80000000)>>12;
312 if(page>262143&&tlb_LUT_r[vaddr>>12]) page=(tlb_LUT_r[vaddr>>12]^0x80000000)>>12;
314 if(page>2048) page=2048+(page&2047);
318 static u_int get_vpage(u_int vaddr)
320 u_int vpage=(vaddr^0x80000000)>>12;
322 if(vpage>262143&&tlb_LUT_r[vaddr>>12]) vpage&=2047; // jump_dirty uses a hash of the virtual address instead
324 if(vpage>2048) vpage=2048+(vpage&2047);
328 // Get address from virtual address
329 // This is called from the recompiled JR/JALR instructions
330 void *get_addr(u_int vaddr)
332 u_int page=get_page(vaddr);
333 u_int vpage=get_vpage(vaddr);
334 struct ll_entry *head;
335 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
338 if(head->vaddr==vaddr&&head->reg32==0) {
339 //printf("TRACE: count=%d next=%d (get_addr match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
340 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
343 ht_bin[1]=(int)head->addr;
349 head=jump_dirty[vpage];
351 if(head->vaddr==vaddr&&head->reg32==0) {
352 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
353 // Don't restore blocks which are about to expire from the cache
354 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
355 if(verify_dirty(head->addr)) {
356 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
357 invalid_code[vaddr>>12]=0;
358 memory_map[vaddr>>12]|=0x40000000;
361 if(tlb_LUT_r[vaddr>>12]) {
362 invalid_code[tlb_LUT_r[vaddr>>12]>>12]=0;
363 memory_map[tlb_LUT_r[vaddr>>12]>>12]|=0x40000000;
366 restore_candidate[vpage>>3]|=1<<(vpage&7);
368 else restore_candidate[page>>3]|=1<<(page&7);
369 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
370 if(ht_bin[0]==vaddr) {
371 ht_bin[1]=(int)head->addr; // Replace existing entry
377 ht_bin[1]=(int)head->addr;
385 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
386 int r=new_recompile_block(vaddr);
387 if(r==0) return get_addr(vaddr);
388 // Execute in unmapped page, generate pagefault execption
390 Cause=(vaddr<<31)|0x8;
391 EPC=(vaddr&1)?vaddr-5:vaddr;
393 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
394 EntryHi=BadVAddr&0xFFFFE000;
395 return get_addr_ht(0x80000000);
397 // Look up address in hash table first
398 void *get_addr_ht(u_int vaddr)
400 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
401 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
402 if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
403 if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
404 return get_addr(vaddr);
407 void *get_addr_32(u_int vaddr,u_int flags)
410 return get_addr(vaddr);
412 //printf("TRACE: count=%d next=%d (get_addr_32 %x,flags %x)\n",Count,next_interupt,vaddr,flags);
413 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
414 if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
415 if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
416 u_int page=get_page(vaddr);
417 u_int vpage=get_vpage(vaddr);
418 struct ll_entry *head;
421 if(head->vaddr==vaddr&&(head->reg32&flags)==0) {
422 //printf("TRACE: count=%d next=%d (get_addr_32 match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
424 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
426 ht_bin[1]=(int)head->addr;
428 }else if(ht_bin[2]==-1) {
429 ht_bin[3]=(int)head->addr;
432 //ht_bin[3]=ht_bin[1];
433 //ht_bin[2]=ht_bin[0];
434 //ht_bin[1]=(int)head->addr;
441 head=jump_dirty[vpage];
443 if(head->vaddr==vaddr&&(head->reg32&flags)==0) {
444 //printf("TRACE: count=%d next=%d (get_addr_32 match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
445 // Don't restore blocks which are about to expire from the cache
446 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
447 if(verify_dirty(head->addr)) {
448 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
449 invalid_code[vaddr>>12]=0;
450 memory_map[vaddr>>12]|=0x40000000;
453 if(tlb_LUT_r[vaddr>>12]) {
454 invalid_code[tlb_LUT_r[vaddr>>12]>>12]=0;
455 memory_map[tlb_LUT_r[vaddr>>12]>>12]|=0x40000000;
458 restore_candidate[vpage>>3]|=1<<(vpage&7);
460 else restore_candidate[page>>3]|=1<<(page&7);
462 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
464 ht_bin[1]=(int)head->addr;
466 }else if(ht_bin[2]==-1) {
467 ht_bin[3]=(int)head->addr;
470 //ht_bin[3]=ht_bin[1];
471 //ht_bin[2]=ht_bin[0];
472 //ht_bin[1]=(int)head->addr;
480 //printf("TRACE: count=%d next=%d (get_addr_32 no-match %x,flags %x)\n",Count,next_interupt,vaddr,flags);
481 int r=new_recompile_block(vaddr);
482 if(r==0) return get_addr(vaddr);
483 // Execute in unmapped page, generate pagefault execption
485 Cause=(vaddr<<31)|0x8;
486 EPC=(vaddr&1)?vaddr-5:vaddr;
488 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
489 EntryHi=BadVAddr&0xFFFFE000;
490 return get_addr_ht(0x80000000);
493 void clear_all_regs(signed char regmap[])
496 for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
499 signed char get_reg(signed char regmap[],int r)
502 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map[hr]==r) return hr;
506 // Find a register that is available for two consecutive cycles
507 signed char get_reg2(signed char regmap1[],signed char regmap2[],int r)
510 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map1[hr]==r&®map2[hr]==r) return hr;
514 int count_free_regs(signed char regmap[])
518 for(hr=0;hr<HOST_REGS;hr++)
520 if(hr!=EXCLUDE_REG) {
521 if(regmap[hr]<0) count++;
527 void dirty_reg(struct regstat *cur,signed char reg)
531 for (hr=0;hr<HOST_REGS;hr++) {
532 if((cur->regmap[hr]&63)==reg) {
538 // If we dirty the lower half of a 64 bit register which is now being
539 // sign-extended, we need to dump the upper half.
540 // Note: Do this only after completion of the instruction, because
541 // some instructions may need to read the full 64-bit value even if
542 // overwriting it (eg SLTI, DSRA32).
543 static void flush_dirty_uppers(struct regstat *cur)
546 for (hr=0;hr<HOST_REGS;hr++) {
547 if((cur->dirty>>hr)&1) {
550 if((cur->is32>>(reg&63))&1) cur->regmap[hr]=-1;
555 void set_const(struct regstat *cur,signed char reg,uint64_t value)
559 for (hr=0;hr<HOST_REGS;hr++) {
560 if(cur->regmap[hr]==reg) {
562 cur->constmap[hr]=value;
564 else if((cur->regmap[hr]^64)==reg) {
566 cur->constmap[hr]=value>>32;
571 void clear_const(struct regstat *cur,signed char reg)
575 for (hr=0;hr<HOST_REGS;hr++) {
576 if((cur->regmap[hr]&63)==reg) {
577 cur->isconst&=~(1<<hr);
582 int is_const(struct regstat *cur,signed char reg)
586 for (hr=0;hr<HOST_REGS;hr++) {
587 if((cur->regmap[hr]&63)==reg) {
588 return (cur->isconst>>hr)&1;
593 uint64_t get_const(struct regstat *cur,signed char reg)
597 for (hr=0;hr<HOST_REGS;hr++) {
598 if(cur->regmap[hr]==reg) {
599 return cur->constmap[hr];
602 printf("Unknown constant in r%d\n",reg);
606 // Least soon needed registers
607 // Look at the next ten instructions and see which registers
608 // will be used. Try not to reallocate these.
609 void lsn(u_char hsn[], int i, int *preferred_reg)
619 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
621 // Don't go past an unconditonal jump
628 if(rs1[i+j]) hsn[rs1[i+j]]=j;
629 if(rs2[i+j]) hsn[rs2[i+j]]=j;
630 if(rt1[i+j]) hsn[rt1[i+j]]=j;
631 if(rt2[i+j]) hsn[rt2[i+j]]=j;
632 if(itype[i+j]==STORE || itype[i+j]==STORELR) {
633 // Stores can allocate zero
637 // On some architectures stores need invc_ptr
638 #if defined(HOST_IMM8)
639 if(itype[i+j]==STORE || itype[i+j]==STORELR || (opcode[i+j]&0x3b)==0x39 || (opcode[i+j]&0x3b)==0x3a) {
643 if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
651 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
653 // Follow first branch
654 int t=(ba[i+b]-start)>>2;
655 j=7-b;if(t+j>=slen) j=slen-t-1;
658 if(rs1[t+j]) if(hsn[rs1[t+j]]>j+b+2) hsn[rs1[t+j]]=j+b+2;
659 if(rs2[t+j]) if(hsn[rs2[t+j]]>j+b+2) hsn[rs2[t+j]]=j+b+2;
660 //if(rt1[t+j]) if(hsn[rt1[t+j]]>j+b+2) hsn[rt1[t+j]]=j+b+2;
661 //if(rt2[t+j]) if(hsn[rt2[t+j]]>j+b+2) hsn[rt2[t+j]]=j+b+2;
664 // TODO: preferred register based on backward branch
666 // Delay slot should preferably not overwrite branch conditions or cycle count
667 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
668 if(rs1[i-1]) if(hsn[rs1[i-1]]>1) hsn[rs1[i-1]]=1;
669 if(rs2[i-1]) if(hsn[rs2[i-1]]>1) hsn[rs2[i-1]]=1;
675 // Coprocessor load/store needs FTEMP, even if not declared
676 if(itype[i]==C1LS||itype[i]==C2LS) {
679 // Load L/R also uses FTEMP as a temporary register
680 if(itype[i]==LOADLR) {
683 // Also 64-bit SDL/SDR
684 if(opcode[i]==0x2c||opcode[i]==0x2d) {
687 // Don't remove the TLB registers either
688 if(itype[i]==LOAD || itype[i]==LOADLR || itype[i]==STORE || itype[i]==STORELR || itype[i]==C1LS || itype[i]==C2LS) {
691 // Don't remove the miniht registers
692 if(itype[i]==UJUMP||itype[i]==RJUMP)
699 // We only want to allocate registers if we're going to use them again soon
700 int needed_again(int r, int i)
706 u_char hsn[MAXREG+1];
709 memset(hsn,10,sizeof(hsn));
710 lsn(hsn,i,&preferred_reg);
712 if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
714 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
715 return 0; // Don't need any registers if exiting the block
723 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
725 // Don't go past an unconditonal jump
729 if(itype[i+j]==SYSCALL||itype[i+j]==HLECALL||((source[i+j]&0xfc00003f)==0x0d))
736 if(rs1[i+j]==r) rn=j;
737 if(rs2[i+j]==r) rn=j;
738 if((unneeded_reg[i+j]>>r)&1) rn=10;
739 if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
747 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
749 // Follow first branch
751 int t=(ba[i+b]-start)>>2;
752 j=7-b;if(t+j>=slen) j=slen-t-1;
755 if(!((unneeded_reg[t+j]>>r)&1)) {
756 if(rs1[t+j]==r) if(rn>j+b+2) rn=j+b+2;
757 if(rs2[t+j]==r) if(rn>j+b+2) rn=j+b+2;
763 for(hr=0;hr<HOST_REGS;hr++) {
764 if(hr!=EXCLUDE_REG) {
765 if(rn<hsn[hr]) return 1;
771 // Try to match register allocations at the end of a loop with those
773 int loop_reg(int i, int r, int hr)
782 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
784 // Don't go past an unconditonal jump
791 if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)
796 if(r<64&&((unneeded_reg[i+k]>>r)&1)) return hr;
797 if(r>64&&((unneeded_reg_upper[i+k]>>r)&1)) return hr;
798 if(i+k>=0&&(itype[i+k]==UJUMP||itype[i+k]==CJUMP||itype[i+k]==SJUMP||itype[i+k]==FJUMP))
800 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
802 int t=(ba[i+k]-start)>>2;
803 int reg=get_reg(regs[t].regmap_entry,r);
804 if(reg>=0) return reg;
805 //reg=get_reg(regs[t+1].regmap_entry,r);
806 //if(reg>=0) return reg;
814 // Allocate every register, preserving source/target regs
815 void alloc_all(struct regstat *cur,int i)
819 for(hr=0;hr<HOST_REGS;hr++) {
820 if(hr!=EXCLUDE_REG) {
821 if(((cur->regmap[hr]&63)!=rs1[i])&&((cur->regmap[hr]&63)!=rs2[i])&&
822 ((cur->regmap[hr]&63)!=rt1[i])&&((cur->regmap[hr]&63)!=rt2[i]))
825 cur->dirty&=~(1<<hr);
828 if((cur->regmap[hr]&63)==0)
831 cur->dirty&=~(1<<hr);
838 void div64(int64_t dividend,int64_t divisor)
842 //printf("TRACE: ddiv %8x%8x %8x%8x\n" ,(int)reg[HIREG],(int)(reg[HIREG]>>32)
843 // ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
845 void divu64(uint64_t dividend,uint64_t divisor)
849 //printf("TRACE: ddivu %8x%8x %8x%8x\n",(int)reg[HIREG],(int)(reg[HIREG]>>32)
850 // ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
853 void mult64(uint64_t m1,uint64_t m2)
855 unsigned long long int op1, op2, op3, op4;
856 unsigned long long int result1, result2, result3, result4;
857 unsigned long long int temp1, temp2, temp3, temp4;
873 op1 = op2 & 0xFFFFFFFF;
874 op2 = (op2 >> 32) & 0xFFFFFFFF;
875 op3 = op4 & 0xFFFFFFFF;
876 op4 = (op4 >> 32) & 0xFFFFFFFF;
879 temp2 = (temp1 >> 32) + op1 * op4;
881 temp4 = (temp3 >> 32) + op2 * op4;
883 result1 = temp1 & 0xFFFFFFFF;
884 result2 = temp2 + (temp3 & 0xFFFFFFFF);
885 result3 = (result2 >> 32) + temp4;
886 result4 = (result3 >> 32);
888 lo = result1 | (result2 << 32);
889 hi = (result3 & 0xFFFFFFFF) | (result4 << 32);
898 void multu64(uint64_t m1,uint64_t m2)
900 unsigned long long int op1, op2, op3, op4;
901 unsigned long long int result1, result2, result3, result4;
902 unsigned long long int temp1, temp2, temp3, temp4;
904 op1 = m1 & 0xFFFFFFFF;
905 op2 = (m1 >> 32) & 0xFFFFFFFF;
906 op3 = m2 & 0xFFFFFFFF;
907 op4 = (m2 >> 32) & 0xFFFFFFFF;
910 temp2 = (temp1 >> 32) + op1 * op4;
912 temp4 = (temp3 >> 32) + op2 * op4;
914 result1 = temp1 & 0xFFFFFFFF;
915 result2 = temp2 + (temp3 & 0xFFFFFFFF);
916 result3 = (result2 >> 32) + temp4;
917 result4 = (result3 >> 32);
919 lo = result1 | (result2 << 32);
920 hi = (result3 & 0xFFFFFFFF) | (result4 << 32);
922 //printf("TRACE: dmultu %8x%8x %8x%8x\n",(int)reg[HIREG],(int)(reg[HIREG]>>32)
923 // ,(int)reg[LOREG],(int)(reg[LOREG]>>32));
926 uint64_t ldl_merge(uint64_t original,uint64_t loaded,u_int bits)
934 else original=loaded;
937 uint64_t ldr_merge(uint64_t original,uint64_t loaded,u_int bits)
940 original>>=64-(bits^56);
941 original<<=64-(bits^56);
945 else original=loaded;
950 #include "assem_x86.c"
953 #include "assem_x64.c"
956 #include "assem_arm.c"
959 // Add virtual address mapping to linked list
960 void ll_add(struct ll_entry **head,int vaddr,void *addr)
962 struct ll_entry *new_entry;
963 new_entry=malloc(sizeof(struct ll_entry));
964 assert(new_entry!=NULL);
965 new_entry->vaddr=vaddr;
967 new_entry->addr=addr;
968 new_entry->next=*head;
972 // Add virtual address mapping for 32-bit compiled block
973 void ll_add_32(struct ll_entry **head,int vaddr,u_int reg32,void *addr)
975 ll_add(head,vaddr,addr);
977 (*head)->reg32=reg32;
981 // Check if an address is already compiled
982 // but don't return addresses which are about to expire from the cache
983 void *check_addr(u_int vaddr)
985 u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
986 if(ht_bin[0]==vaddr) {
987 if(((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
988 if(isclean(ht_bin[1])) return (void *)ht_bin[1];
990 if(ht_bin[2]==vaddr) {
991 if(((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
992 if(isclean(ht_bin[3])) return (void *)ht_bin[3];
994 u_int page=get_page(vaddr);
995 struct ll_entry *head;
998 if(head->vaddr==vaddr&&head->reg32==0) {
999 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1000 // Update existing entry with current address
1001 if(ht_bin[0]==vaddr) {
1002 ht_bin[1]=(int)head->addr;
1005 if(ht_bin[2]==vaddr) {
1006 ht_bin[3]=(int)head->addr;
1009 // Insert into hash table with low priority.
1010 // Don't evict existing entries, as they are probably
1011 // addresses that are being accessed frequently.
1013 ht_bin[1]=(int)head->addr;
1015 }else if(ht_bin[2]==-1) {
1016 ht_bin[3]=(int)head->addr;
1027 void remove_hash(int vaddr)
1029 //printf("remove hash: %x\n",vaddr);
1030 int *ht_bin=hash_table[(((vaddr)>>16)^vaddr)&0xFFFF];
1031 if(ht_bin[2]==vaddr) {
1032 ht_bin[2]=ht_bin[3]=-1;
1034 if(ht_bin[0]==vaddr) {
1035 ht_bin[0]=ht_bin[2];
1036 ht_bin[1]=ht_bin[3];
1037 ht_bin[2]=ht_bin[3]=-1;
1041 void ll_remove_matching_addrs(struct ll_entry **head,int addr,int shift)
1043 struct ll_entry *next;
1045 if(((u_int)((*head)->addr)>>shift)==(addr>>shift) ||
1046 ((u_int)((*head)->addr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift))
1048 inv_debug("EXP: Remove pointer to %x (%x)\n",(int)(*head)->addr,(*head)->vaddr);
1049 remove_hash((*head)->vaddr);
1056 head=&((*head)->next);
1061 // Remove all entries from linked list
1062 void ll_clear(struct ll_entry **head)
1064 struct ll_entry *cur;
1065 struct ll_entry *next;
1076 // Dereference the pointers and remove if it matches
1077 void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
1080 int ptr=get_pointer(head->addr);
1081 inv_debug("EXP: Lookup pointer to %x at %x (%x)\n",(int)ptr,(int)head->addr,head->vaddr);
1082 if(((ptr>>shift)==(addr>>shift)) ||
1083 (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
1085 inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
1086 kill_pointer(head->addr);
1092 // This is called when we write to a compiled block (see do_invstub)
1093 int invalidate_page(u_int page)
1096 struct ll_entry *head;
1097 struct ll_entry *next;
1101 inv_debug("INVALIDATE: %x\n",head->vaddr);
1102 remove_hash(head->vaddr);
1107 head=jump_out[page];
1110 inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
1111 kill_pointer(head->addr);
1119 void invalidate_block(u_int block)
1122 u_int page=get_page(block<<12);
1123 u_int vpage=get_vpage(block<<12);
1124 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1125 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1128 struct ll_entry *head;
1129 head=jump_dirty[vpage];
1130 //printf("page=%d vpage=%d\n",page,vpage);
1133 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1134 get_bounds((int)head->addr,&start,&end);
1135 //printf("start: %x end: %x\n",start,end);
1136 if(page<2048&&start>=0x80000000&&end<0x80800000) {
1137 if(((start-(u_int)rdram)>>12)<=page&&((end-1-(u_int)rdram)>>12)>=page) {
1138 if((((start-(u_int)rdram)>>12)&2047)<first) first=((start-(u_int)rdram)>>12)&2047;
1139 if((((end-1-(u_int)rdram)>>12)&2047)>last) last=((end-1-(u_int)rdram)>>12)&2047;
1143 if(page<2048&&(signed int)start>=(signed int)0xC0000000&&(signed int)end>=(signed int)0xC0000000) {
1144 if(((start+memory_map[start>>12]-(u_int)rdram)>>12)<=page&&((end-1+memory_map[(end-1)>>12]-(u_int)rdram)>>12)>=page) {
1145 if((((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047)<first) first=((start+memory_map[start>>12]-(u_int)rdram)>>12)&2047;
1146 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;
1153 //printf("first=%d last=%d\n",first,last);
1154 modified=invalidate_page(page);
1155 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1156 assert(last<page+5);
1157 // Invalidate the adjacent pages if a block crosses a 4K boundary
1159 invalidate_page(first);
1162 for(first=page+1;first<last;first++) {
1163 invalidate_page(first);
1166 // Don't trap writes
1167 invalid_code[block]=1;
1169 // If there is a valid TLB entry for this page, remove write protect
1170 if(tlb_LUT_w[block]) {
1171 assert(tlb_LUT_r[block]==tlb_LUT_w[block]);
1172 // CHECK: Is this right?
1173 memory_map[block]=((tlb_LUT_w[block]&0xFFFFF000)-(block<<12)+(unsigned int)rdram-0x80000000)>>2;
1174 u_int real_block=tlb_LUT_w[block]>>12;
1175 invalid_code[real_block]=1;
1176 if(real_block>=0x80000&&real_block<0x80800) memory_map[real_block]=((u_int)rdram-0x80000000)>>2;
1178 else if(block>=0x80000&&block<0x80800) memory_map[block]=((u_int)rdram-0x80000000)>>2;
1182 __clear_cache((void *)BASE_ADDR,(void *)BASE_ADDR+(1<<TARGET_SIZE_2));
1185 memset(mini_ht,-1,sizeof(mini_ht));
1188 void invalidate_addr(u_int addr)
1190 invalidate_block(addr>>12);
1192 void invalidate_all_pages()
1195 for(page=0;page<4096;page++)
1196 invalidate_page(page);
1197 for(page=0;page<1048576;page++)
1198 if(!invalid_code[page]) {
1199 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1200 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1203 __clear_cache((void *)BASE_ADDR,(void *)BASE_ADDR+(1<<TARGET_SIZE_2));
1206 memset(mini_ht,-1,sizeof(mini_ht));
1210 for(page=0;page<0x100000;page++) {
1211 if(tlb_LUT_r[page]) {
1212 memory_map[page]=((tlb_LUT_r[page]&0xFFFFF000)-(page<<12)+(unsigned int)rdram-0x80000000)>>2;
1213 if(!tlb_LUT_w[page]||!invalid_code[page])
1214 memory_map[page]|=0x40000000; // Write protect
1216 else memory_map[page]=-1;
1217 if(page==0x80000) page=0xC0000;
1223 // Add an entry to jump_out after making a link
1224 void add_link(u_int vaddr,void *src)
1226 u_int page=get_page(vaddr);
1227 inv_debug("add_link: %x -> %x (%d)\n",(int)src,vaddr,page);
1228 ll_add(jump_out+page,vaddr,src);
1229 //int ptr=get_pointer(src);
1230 //inv_debug("add_link: Pointer is to %x\n",(int)ptr);
1233 // If a code block was found to be unmodified (bit was set in
1234 // restore_candidate) and it remains unmodified (bit is clear
1235 // in invalid_code) then move the entries for that 4K page from
1236 // the dirty list to the clean list.
1237 void clean_blocks(u_int page)
1239 struct ll_entry *head;
1240 inv_debug("INV: clean_blocks page=%d\n",page);
1241 head=jump_dirty[page];
1243 if(!invalid_code[head->vaddr>>12]) {
1244 // Don't restore blocks which are about to expire from the cache
1245 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1247 if(verify_dirty((int)head->addr)) {
1248 //printf("Possibly Restore %x (%x)\n",head->vaddr, (int)head->addr);
1251 get_bounds((int)head->addr,&start,&end);
1252 if(start-(u_int)rdram<0x800000) {
1253 for(i=(start-(u_int)rdram+0x80000000)>>12;i<=(end-1-(u_int)rdram+0x80000000)>>12;i++) {
1254 inv|=invalid_code[i];
1257 if((signed int)head->vaddr>=(signed int)0xC0000000) {
1258 u_int addr = (head->vaddr+(memory_map[head->vaddr>>12]<<2));
1259 //printf("addr=%x start=%x end=%x\n",addr,start,end);
1260 if(addr<start||addr>=end) inv=1;
1262 else if((signed int)head->vaddr>=(signed int)0x80800000) {
1266 void * clean_addr=(void *)get_clean_addr((int)head->addr);
1267 if((((u_int)clean_addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1270 if(page<2048&&tlb_LUT_r[head->vaddr>>12]) ppage=(tlb_LUT_r[head->vaddr>>12]^0x80000000)>>12;
1272 inv_debug("INV: Restored %x (%x/%x)\n",head->vaddr, (int)head->addr, (int)clean_addr);
1273 //printf("page=%x, addr=%x\n",page,head->vaddr);
1274 //assert(head->vaddr>>12==(page|0x80000));
1275 ll_add_32(jump_in+ppage,head->vaddr,head->reg32,clean_addr);
1276 int *ht_bin=hash_table[((head->vaddr>>16)^head->vaddr)&0xFFFF];
1278 if(ht_bin[0]==head->vaddr) {
1279 ht_bin[1]=(int)clean_addr; // Replace existing entry
1281 if(ht_bin[2]==head->vaddr) {
1282 ht_bin[3]=(int)clean_addr; // Replace existing entry
1295 void mov_alloc(struct regstat *current,int i)
1297 // Note: Don't need to actually alloc the source registers
1298 if((~current->is32>>rs1[i])&1) {
1299 //alloc_reg64(current,i,rs1[i]);
1300 alloc_reg64(current,i,rt1[i]);
1301 current->is32&=~(1LL<<rt1[i]);
1303 //alloc_reg(current,i,rs1[i]);
1304 alloc_reg(current,i,rt1[i]);
1305 current->is32|=(1LL<<rt1[i]);
1307 clear_const(current,rs1[i]);
1308 clear_const(current,rt1[i]);
1309 dirty_reg(current,rt1[i]);
1312 void shiftimm_alloc(struct regstat *current,int i)
1314 clear_const(current,rs1[i]);
1315 clear_const(current,rt1[i]);
1316 if(opcode2[i]<=0x3) // SLL/SRL/SRA
1319 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1321 alloc_reg(current,i,rt1[i]);
1322 current->is32|=1LL<<rt1[i];
1323 dirty_reg(current,rt1[i]);
1326 if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
1329 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1330 alloc_reg64(current,i,rt1[i]);
1331 current->is32&=~(1LL<<rt1[i]);
1332 dirty_reg(current,rt1[i]);
1335 if(opcode2[i]==0x3c) // DSLL32
1338 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1339 alloc_reg64(current,i,rt1[i]);
1340 current->is32&=~(1LL<<rt1[i]);
1341 dirty_reg(current,rt1[i]);
1344 if(opcode2[i]==0x3e) // DSRL32
1347 alloc_reg64(current,i,rs1[i]);
1349 alloc_reg64(current,i,rt1[i]);
1350 current->is32&=~(1LL<<rt1[i]);
1352 alloc_reg(current,i,rt1[i]);
1353 current->is32|=1LL<<rt1[i];
1355 dirty_reg(current,rt1[i]);
1358 if(opcode2[i]==0x3f) // DSRA32
1361 alloc_reg64(current,i,rs1[i]);
1362 alloc_reg(current,i,rt1[i]);
1363 current->is32|=1LL<<rt1[i];
1364 dirty_reg(current,rt1[i]);
1369 void shift_alloc(struct regstat *current,int i)
1372 if(opcode2[i]<=0x07) // SLLV/SRLV/SRAV
1374 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1375 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1376 alloc_reg(current,i,rt1[i]);
1377 if(rt1[i]==rs2[i]) alloc_reg_temp(current,i,-1);
1378 current->is32|=1LL<<rt1[i];
1379 } else { // DSLLV/DSRLV/DSRAV
1380 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1381 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1382 alloc_reg64(current,i,rt1[i]);
1383 current->is32&=~(1LL<<rt1[i]);
1384 if(opcode2[i]==0x16||opcode2[i]==0x17) // DSRLV and DSRAV need a temporary register
1385 alloc_reg_temp(current,i,-1);
1387 clear_const(current,rs1[i]);
1388 clear_const(current,rs2[i]);
1389 clear_const(current,rt1[i]);
1390 dirty_reg(current,rt1[i]);
1394 void alu_alloc(struct regstat *current,int i)
1396 if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
1398 if(rs1[i]&&rs2[i]) {
1399 alloc_reg(current,i,rs1[i]);
1400 alloc_reg(current,i,rs2[i]);
1403 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1404 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1406 alloc_reg(current,i,rt1[i]);
1408 current->is32|=1LL<<rt1[i];
1410 if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
1412 if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1414 alloc_reg64(current,i,rs1[i]);
1415 alloc_reg64(current,i,rs2[i]);
1416 alloc_reg(current,i,rt1[i]);
1418 alloc_reg(current,i,rs1[i]);
1419 alloc_reg(current,i,rs2[i]);
1420 alloc_reg(current,i,rt1[i]);
1423 current->is32|=1LL<<rt1[i];
1425 if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
1427 if(rs1[i]&&rs2[i]) {
1428 alloc_reg(current,i,rs1[i]);
1429 alloc_reg(current,i,rs2[i]);
1433 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1434 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1436 alloc_reg(current,i,rt1[i]);
1437 if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1439 if(!((current->uu>>rt1[i])&1)) {
1440 alloc_reg64(current,i,rt1[i]);
1442 if(get_reg(current->regmap,rt1[i]|64)>=0) {
1443 if(rs1[i]&&rs2[i]) {
1444 alloc_reg64(current,i,rs1[i]);
1445 alloc_reg64(current,i,rs2[i]);
1449 // Is is really worth it to keep 64-bit values in registers?
1451 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1452 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg64(current,i,rs2[i]);
1456 current->is32&=~(1LL<<rt1[i]);
1458 current->is32|=1LL<<rt1[i];
1462 if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1464 if(rs1[i]&&rs2[i]) {
1465 if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1466 alloc_reg64(current,i,rs1[i]);
1467 alloc_reg64(current,i,rs2[i]);
1468 alloc_reg64(current,i,rt1[i]);
1470 alloc_reg(current,i,rs1[i]);
1471 alloc_reg(current,i,rs2[i]);
1472 alloc_reg(current,i,rt1[i]);
1476 alloc_reg(current,i,rt1[i]);
1477 if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1478 // DADD used as move, or zeroing
1479 // If we have a 64-bit source, then make the target 64 bits too
1480 if(rs1[i]&&!((current->is32>>rs1[i])&1)) {
1481 if(get_reg(current->regmap,rs1[i])>=0) alloc_reg64(current,i,rs1[i]);
1482 alloc_reg64(current,i,rt1[i]);
1483 } else if(rs2[i]&&!((current->is32>>rs2[i])&1)) {
1484 if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1485 alloc_reg64(current,i,rt1[i]);
1487 if(opcode2[i]>=0x2e&&rs2[i]) {
1488 // DSUB used as negation - 64-bit result
1489 // If we have a 32-bit register, extend it to 64 bits
1490 if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1491 alloc_reg64(current,i,rt1[i]);
1495 if(rs1[i]&&rs2[i]) {
1496 current->is32&=~(1LL<<rt1[i]);
1498 current->is32&=~(1LL<<rt1[i]);
1499 if((current->is32>>rs1[i])&1)
1500 current->is32|=1LL<<rt1[i];
1502 current->is32&=~(1LL<<rt1[i]);
1503 if((current->is32>>rs2[i])&1)
1504 current->is32|=1LL<<rt1[i];
1506 current->is32|=1LL<<rt1[i];
1510 clear_const(current,rs1[i]);
1511 clear_const(current,rs2[i]);
1512 clear_const(current,rt1[i]);
1513 dirty_reg(current,rt1[i]);
1516 void imm16_alloc(struct regstat *current,int i)
1518 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1520 if(rt1[i]) alloc_reg(current,i,rt1[i]);
1521 if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
1522 current->is32&=~(1LL<<rt1[i]);
1523 if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1524 // TODO: Could preserve the 32-bit flag if the immediate is zero
1525 alloc_reg64(current,i,rt1[i]);
1526 alloc_reg64(current,i,rs1[i]);
1528 clear_const(current,rs1[i]);
1529 clear_const(current,rt1[i]);
1531 else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
1532 if((~current->is32>>rs1[i])&1) alloc_reg64(current,i,rs1[i]);
1533 current->is32|=1LL<<rt1[i];
1534 clear_const(current,rs1[i]);
1535 clear_const(current,rt1[i]);
1537 else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
1538 if(((~current->is32>>rs1[i])&1)&&opcode[i]>0x0c) {
1539 if(rs1[i]!=rt1[i]) {
1540 if(needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1541 alloc_reg64(current,i,rt1[i]);
1542 current->is32&=~(1LL<<rt1[i]);
1545 else current->is32|=1LL<<rt1[i]; // ANDI clears upper bits
1546 if(is_const(current,rs1[i])) {
1547 int v=get_const(current,rs1[i]);
1548 if(opcode[i]==0x0c) set_const(current,rt1[i],v&imm[i]);
1549 if(opcode[i]==0x0d) set_const(current,rt1[i],v|imm[i]);
1550 if(opcode[i]==0x0e) set_const(current,rt1[i],v^imm[i]);
1552 else clear_const(current,rt1[i]);
1554 else if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
1555 if(is_const(current,rs1[i])) {
1556 int v=get_const(current,rs1[i]);
1557 set_const(current,rt1[i],v+imm[i]);
1559 else clear_const(current,rt1[i]);
1560 current->is32|=1LL<<rt1[i];
1563 set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
1564 current->is32|=1LL<<rt1[i];
1566 dirty_reg(current,rt1[i]);
1569 void load_alloc(struct regstat *current,int i)
1571 clear_const(current,rt1[i]);
1572 //if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
1573 if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
1574 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1576 alloc_reg(current,i,rt1[i]);
1577 if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
1579 current->is32&=~(1LL<<rt1[i]);
1580 alloc_reg64(current,i,rt1[i]);
1582 else if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1584 current->is32&=~(1LL<<rt1[i]);
1585 alloc_reg64(current,i,rt1[i]);
1586 alloc_all(current,i);
1587 alloc_reg64(current,i,FTEMP);
1589 else current->is32|=1LL<<rt1[i];
1590 dirty_reg(current,rt1[i]);
1591 // If using TLB, need a register for pointer to the mapping table
1592 if(using_tlb) alloc_reg(current,i,TLREG);
1593 // LWL/LWR need a temporary register for the old value
1594 if(opcode[i]==0x22||opcode[i]==0x26)
1596 alloc_reg(current,i,FTEMP);
1597 alloc_reg_temp(current,i,-1);
1602 // Load to r0 (dummy load)
1603 // but we still need a register to calculate the address
1604 alloc_reg_temp(current,i,-1);
1608 void store_alloc(struct regstat *current,int i)
1610 clear_const(current,rs2[i]);
1611 if(!(rs2[i])) current->u&=~1LL; // Allow allocating r0 if necessary
1612 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1613 alloc_reg(current,i,rs2[i]);
1614 if(opcode[i]==0x2c||opcode[i]==0x2d||opcode[i]==0x3f) { // 64-bit SDL/SDR/SD
1615 alloc_reg64(current,i,rs2[i]);
1616 if(rs2[i]) alloc_reg(current,i,FTEMP);
1618 // If using TLB, need a register for pointer to the mapping table
1619 if(using_tlb) alloc_reg(current,i,TLREG);
1620 #if defined(HOST_IMM8)
1621 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1622 else alloc_reg(current,i,INVCP);
1624 if(opcode[i]==0x2c||opcode[i]==0x2d) { // 64-bit SDL/SDR
1625 alloc_reg(current,i,FTEMP);
1627 // We need a temporary register for address generation
1628 alloc_reg_temp(current,i,-1);
1631 void c1ls_alloc(struct regstat *current,int i)
1633 //clear_const(current,rs1[i]); // FIXME
1634 clear_const(current,rt1[i]);
1635 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1636 alloc_reg(current,i,CSREG); // Status
1637 alloc_reg(current,i,FTEMP);
1638 if(opcode[i]==0x35||opcode[i]==0x3d) { // 64-bit LDC1/SDC1
1639 alloc_reg64(current,i,FTEMP);
1641 // If using TLB, need a register for pointer to the mapping table
1642 if(using_tlb) alloc_reg(current,i,TLREG);
1643 #if defined(HOST_IMM8)
1644 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1645 else if((opcode[i]&0x3b)==0x39) // SWC1/SDC1
1646 alloc_reg(current,i,INVCP);
1648 // We need a temporary register for address generation
1649 alloc_reg_temp(current,i,-1);
1652 void c2ls_alloc(struct regstat *current,int i)
1654 clear_const(current,rt1[i]);
1655 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1656 alloc_reg(current,i,FTEMP);
1657 // If using TLB, need a register for pointer to the mapping table
1658 if(using_tlb) alloc_reg(current,i,TLREG);
1659 #if defined(HOST_IMM8)
1660 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1661 else if((opcode[i]&0x3b)==0x3a) // SWC2/SDC2
1662 alloc_reg(current,i,INVCP);
1664 // We need a temporary register for address generation
1665 alloc_reg_temp(current,i,-1);
1668 #ifndef multdiv_alloc
1669 void multdiv_alloc(struct regstat *current,int i)
1676 // case 0x1D: DMULTU
1679 clear_const(current,rs1[i]);
1680 clear_const(current,rs2[i]);
1683 if((opcode2[i]&4)==0) // 32-bit
1685 current->u&=~(1LL<<HIREG);
1686 current->u&=~(1LL<<LOREG);
1687 alloc_reg(current,i,HIREG);
1688 alloc_reg(current,i,LOREG);
1689 alloc_reg(current,i,rs1[i]);
1690 alloc_reg(current,i,rs2[i]);
1691 current->is32|=1LL<<HIREG;
1692 current->is32|=1LL<<LOREG;
1693 dirty_reg(current,HIREG);
1694 dirty_reg(current,LOREG);
1698 current->u&=~(1LL<<HIREG);
1699 current->u&=~(1LL<<LOREG);
1700 current->uu&=~(1LL<<HIREG);
1701 current->uu&=~(1LL<<LOREG);
1702 alloc_reg64(current,i,HIREG);
1703 //if(HOST_REGS>10) alloc_reg64(current,i,LOREG);
1704 alloc_reg64(current,i,rs1[i]);
1705 alloc_reg64(current,i,rs2[i]);
1706 alloc_all(current,i);
1707 current->is32&=~(1LL<<HIREG);
1708 current->is32&=~(1LL<<LOREG);
1709 dirty_reg(current,HIREG);
1710 dirty_reg(current,LOREG);
1715 // Multiply by zero is zero.
1716 // MIPS does not have a divide by zero exception.
1717 // The result is undefined, we return zero.
1718 alloc_reg(current,i,HIREG);
1719 alloc_reg(current,i,LOREG);
1720 current->is32|=1LL<<HIREG;
1721 current->is32|=1LL<<LOREG;
1722 dirty_reg(current,HIREG);
1723 dirty_reg(current,LOREG);
1728 void cop0_alloc(struct regstat *current,int i)
1730 if(opcode2[i]==0) // MFC0
1733 clear_const(current,rt1[i]);
1734 alloc_all(current,i);
1735 alloc_reg(current,i,rt1[i]);
1736 current->is32|=1LL<<rt1[i];
1737 dirty_reg(current,rt1[i]);
1740 else if(opcode2[i]==4) // MTC0
1743 clear_const(current,rs1[i]);
1744 alloc_reg(current,i,rs1[i]);
1745 alloc_all(current,i);
1748 alloc_all(current,i); // FIXME: Keep r0
1750 alloc_reg(current,i,0);
1755 // TLBR/TLBWI/TLBWR/TLBP/ERET
1756 assert(opcode2[i]==0x10);
1757 alloc_all(current,i);
1761 void cop1_alloc(struct regstat *current,int i)
1763 alloc_reg(current,i,CSREG); // Load status
1764 if(opcode2[i]<3) // MFC1/DMFC1/CFC1
1767 clear_const(current,rt1[i]);
1769 alloc_reg64(current,i,rt1[i]); // DMFC1
1770 current->is32&=~(1LL<<rt1[i]);
1772 alloc_reg(current,i,rt1[i]); // MFC1/CFC1
1773 current->is32|=1LL<<rt1[i];
1775 dirty_reg(current,rt1[i]);
1776 alloc_reg_temp(current,i,-1);
1778 else if(opcode2[i]>3) // MTC1/DMTC1/CTC1
1781 clear_const(current,rs1[i]);
1783 alloc_reg64(current,i,rs1[i]); // DMTC1
1785 alloc_reg(current,i,rs1[i]); // MTC1/CTC1
1786 alloc_reg_temp(current,i,-1);
1790 alloc_reg(current,i,0);
1791 alloc_reg_temp(current,i,-1);
1795 void fconv_alloc(struct regstat *current,int i)
1797 alloc_reg(current,i,CSREG); // Load status
1798 alloc_reg_temp(current,i,-1);
1800 void float_alloc(struct regstat *current,int i)
1802 alloc_reg(current,i,CSREG); // Load status
1803 alloc_reg_temp(current,i,-1);
1805 void c2op_alloc(struct regstat *current,int i)
1807 alloc_reg_temp(current,i,-1);
1809 void fcomp_alloc(struct regstat *current,int i)
1811 alloc_reg(current,i,CSREG); // Load status
1812 alloc_reg(current,i,FSREG); // Load flags
1813 dirty_reg(current,FSREG); // Flag will be modified
1814 alloc_reg_temp(current,i,-1);
1817 void syscall_alloc(struct regstat *current,int i)
1819 alloc_cc(current,i);
1820 dirty_reg(current,CCREG);
1821 alloc_all(current,i);
1825 void delayslot_alloc(struct regstat *current,int i)
1836 assem_debug("jump in the delay slot. this shouldn't happen.\n");//exit(1);
1837 printf("Disabled speculative precompilation\n");
1841 imm16_alloc(current,i);
1845 load_alloc(current,i);
1849 store_alloc(current,i);
1852 alu_alloc(current,i);
1855 shift_alloc(current,i);
1858 multdiv_alloc(current,i);
1861 shiftimm_alloc(current,i);
1864 mov_alloc(current,i);
1867 cop0_alloc(current,i);
1871 cop1_alloc(current,i);
1874 c1ls_alloc(current,i);
1877 c2ls_alloc(current,i);
1880 fconv_alloc(current,i);
1883 float_alloc(current,i);
1886 fcomp_alloc(current,i);
1889 c2op_alloc(current,i);
1894 // Special case where a branch and delay slot span two pages in virtual memory
1895 static void pagespan_alloc(struct regstat *current,int i)
1898 current->wasconst=0;
1900 alloc_all(current,i);
1901 alloc_cc(current,i);
1902 dirty_reg(current,CCREG);
1903 if(opcode[i]==3) // JAL
1905 alloc_reg(current,i,31);
1906 dirty_reg(current,31);
1908 if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
1910 alloc_reg(current,i,rs1[i]);
1912 alloc_reg(current,i,rt1[i]);
1913 dirty_reg(current,rt1[i]);
1916 if((opcode[i]&0x2E)==4) // BEQ/BNE/BEQL/BNEL
1918 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1919 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1920 if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1922 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1923 if(rs2[i]) alloc_reg64(current,i,rs2[i]);
1927 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
1929 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1930 if(!((current->is32>>rs1[i])&1))
1932 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1936 if(opcode[i]==0x11) // BC1
1938 alloc_reg(current,i,FSREG);
1939 alloc_reg(current,i,CSREG);
1944 add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e)
1946 stubs[stubcount][0]=type;
1947 stubs[stubcount][1]=addr;
1948 stubs[stubcount][2]=retaddr;
1949 stubs[stubcount][3]=a;
1950 stubs[stubcount][4]=b;
1951 stubs[stubcount][5]=c;
1952 stubs[stubcount][6]=d;
1953 stubs[stubcount][7]=e;
1957 // Write out a single register
1958 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32)
1961 for(hr=0;hr<HOST_REGS;hr++) {
1962 if(hr!=EXCLUDE_REG) {
1963 if((regmap[hr]&63)==r) {
1966 emit_storereg(r,hr);
1968 if((is32>>regmap[hr])&1) {
1969 emit_sarimm(hr,31,hr);
1970 emit_storereg(r|64,hr);
1974 emit_storereg(r|64,hr);
1984 //if(!tracedebug) return 0;
1987 for(i=0;i<2097152;i++) {
1988 unsigned int temp=sum;
1991 sum^=((u_int *)rdram)[i];
2000 sum^=((u_int *)reg)[i];
2008 printf("r%d:%8x%8x ",i,((int *)(reg+i))[1],((int *)(reg+i))[0]);
2010 #ifndef DISABLE_COP1
2013 printf("f%d:%8x%8x ",i,((int*)reg_cop1_simple[i])[1],*((int*)reg_cop1_simple[i]));
2023 void memdebug(int i)
2025 //printf("TRACE: count=%d next=%d (checksum %x) lo=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[LOREG]>>32),(int)reg[LOREG]);
2026 //printf("TRACE: count=%d next=%d (rchecksum %x)\n",Count,next_interupt,rchecksum());
2029 //if(Count>=-2084597794) {
2030 if((signed int)Count>=-2084597794&&(signed int)Count<0) {
2032 printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
2033 //printf("TRACE: count=%d next=%d (checksum %x) Status=%x\n",Count,next_interupt,mchecksum(),Status);
2034 //printf("TRACE: count=%d next=%d (checksum %x) hi=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[HIREG]>>32),(int)reg[HIREG]);
2037 printf("TRACE: %x\n",(&i)[-1]);
2041 printf("TRACE: %x \n",(&j)[10]);
2042 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]);
2046 //printf("TRACE: %x\n",(&i)[-1]);
2049 void tlb_debug(u_int cause, u_int addr, u_int iaddr)
2051 printf("TLB Exception: instruction=%x addr=%x cause=%x\n",iaddr, addr, cause);
2054 void alu_assemble(int i,struct regstat *i_regs)
2056 if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
2058 signed char s1,s2,t;
2059 t=get_reg(i_regs->regmap,rt1[i]);
2061 s1=get_reg(i_regs->regmap,rs1[i]);
2062 s2=get_reg(i_regs->regmap,rs2[i]);
2063 if(rs1[i]&&rs2[i]) {
2066 if(opcode2[i]&2) emit_sub(s1,s2,t);
2067 else emit_add(s1,s2,t);
2070 if(s1>=0) emit_mov(s1,t);
2071 else emit_loadreg(rs1[i],t);
2075 if(opcode2[i]&2) emit_neg(s2,t);
2076 else emit_mov(s2,t);
2079 emit_loadreg(rs2[i],t);
2080 if(opcode2[i]&2) emit_neg(t,t);
2083 else emit_zeroreg(t);
2087 if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2089 signed char s1l,s2l,s1h,s2h,tl,th;
2090 tl=get_reg(i_regs->regmap,rt1[i]);
2091 th=get_reg(i_regs->regmap,rt1[i]|64);
2093 s1l=get_reg(i_regs->regmap,rs1[i]);
2094 s2l=get_reg(i_regs->regmap,rs2[i]);
2095 s1h=get_reg(i_regs->regmap,rs1[i]|64);
2096 s2h=get_reg(i_regs->regmap,rs2[i]|64);
2097 if(rs1[i]&&rs2[i]) {
2100 if(opcode2[i]&2) emit_subs(s1l,s2l,tl);
2101 else emit_adds(s1l,s2l,tl);
2103 #ifdef INVERTED_CARRY
2104 if(opcode2[i]&2) {if(s1h!=th) emit_mov(s1h,th);emit_sbb(th,s2h);}
2106 if(opcode2[i]&2) emit_sbc(s1h,s2h,th);
2108 else emit_add(s1h,s2h,th);
2112 if(s1l>=0) emit_mov(s1l,tl);
2113 else emit_loadreg(rs1[i],tl);
2115 if(s1h>=0) emit_mov(s1h,th);
2116 else emit_loadreg(rs1[i]|64,th);
2121 if(opcode2[i]&2) emit_negs(s2l,tl);
2122 else emit_mov(s2l,tl);
2125 emit_loadreg(rs2[i],tl);
2126 if(opcode2[i]&2) emit_negs(tl,tl);
2129 #ifdef INVERTED_CARRY
2130 if(s2h>=0) emit_mov(s2h,th);
2131 else emit_loadreg(rs2[i]|64,th);
2133 emit_adcimm(-1,th); // x86 has inverted carry flag
2138 if(s2h>=0) emit_rscimm(s2h,0,th);
2140 emit_loadreg(rs2[i]|64,th);
2141 emit_rscimm(th,0,th);
2144 if(s2h>=0) emit_mov(s2h,th);
2145 else emit_loadreg(rs2[i]|64,th);
2152 if(th>=0) emit_zeroreg(th);
2157 if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
2159 signed char s1l,s1h,s2l,s2h,t;
2160 if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1))
2162 t=get_reg(i_regs->regmap,rt1[i]);
2165 s1l=get_reg(i_regs->regmap,rs1[i]);
2166 s1h=get_reg(i_regs->regmap,rs1[i]|64);
2167 s2l=get_reg(i_regs->regmap,rs2[i]);
2168 s2h=get_reg(i_regs->regmap,rs2[i]|64);
2169 if(rs2[i]==0) // rx<r0
2172 if(opcode2[i]==0x2a) // SLT
2173 emit_shrimm(s1h,31,t);
2174 else // SLTU (unsigned can not be less than zero)
2177 else if(rs1[i]==0) // r0<rx
2180 if(opcode2[i]==0x2a) // SLT
2181 emit_set_gz64_32(s2h,s2l,t);
2182 else // SLTU (set if not zero)
2183 emit_set_nz64_32(s2h,s2l,t);
2186 assert(s1l>=0);assert(s1h>=0);
2187 assert(s2l>=0);assert(s2h>=0);
2188 if(opcode2[i]==0x2a) // SLT
2189 emit_set_if_less64_32(s1h,s1l,s2h,s2l,t);
2191 emit_set_if_carry64_32(s1h,s1l,s2h,s2l,t);
2195 t=get_reg(i_regs->regmap,rt1[i]);
2198 s1l=get_reg(i_regs->regmap,rs1[i]);
2199 s2l=get_reg(i_regs->regmap,rs2[i]);
2200 if(rs2[i]==0) // rx<r0
2203 if(opcode2[i]==0x2a) // SLT
2204 emit_shrimm(s1l,31,t);
2205 else // SLTU (unsigned can not be less than zero)
2208 else if(rs1[i]==0) // r0<rx
2211 if(opcode2[i]==0x2a) // SLT
2212 emit_set_gz32(s2l,t);
2213 else // SLTU (set if not zero)
2214 emit_set_nz32(s2l,t);
2217 assert(s1l>=0);assert(s2l>=0);
2218 if(opcode2[i]==0x2a) // SLT
2219 emit_set_if_less32(s1l,s2l,t);
2221 emit_set_if_carry32(s1l,s2l,t);
2227 if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
2229 signed char s1l,s1h,s2l,s2h,th,tl;
2230 tl=get_reg(i_regs->regmap,rt1[i]);
2231 th=get_reg(i_regs->regmap,rt1[i]|64);
2232 if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1)&&th>=0)
2236 s1l=get_reg(i_regs->regmap,rs1[i]);
2237 s1h=get_reg(i_regs->regmap,rs1[i]|64);
2238 s2l=get_reg(i_regs->regmap,rs2[i]);
2239 s2h=get_reg(i_regs->regmap,rs2[i]|64);
2240 if(rs1[i]&&rs2[i]) {
2241 assert(s1l>=0);assert(s1h>=0);
2242 assert(s2l>=0);assert(s2h>=0);
2243 if(opcode2[i]==0x24) { // AND
2244 emit_and(s1l,s2l,tl);
2245 emit_and(s1h,s2h,th);
2247 if(opcode2[i]==0x25) { // OR
2248 emit_or(s1l,s2l,tl);
2249 emit_or(s1h,s2h,th);
2251 if(opcode2[i]==0x26) { // XOR
2252 emit_xor(s1l,s2l,tl);
2253 emit_xor(s1h,s2h,th);
2255 if(opcode2[i]==0x27) { // NOR
2256 emit_or(s1l,s2l,tl);
2257 emit_or(s1h,s2h,th);
2264 if(opcode2[i]==0x24) { // AND
2268 if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2270 if(s1l>=0) emit_mov(s1l,tl);
2271 else emit_loadreg(rs1[i],tl);
2272 if(s1h>=0) emit_mov(s1h,th);
2273 else emit_loadreg(rs1[i]|64,th);
2277 if(s2l>=0) emit_mov(s2l,tl);
2278 else emit_loadreg(rs2[i],tl);
2279 if(s2h>=0) emit_mov(s2h,th);
2280 else emit_loadreg(rs2[i]|64,th);
2287 if(opcode2[i]==0x27) { // NOR
2289 if(s1l>=0) emit_not(s1l,tl);
2291 emit_loadreg(rs1[i],tl);
2294 if(s1h>=0) emit_not(s1h,th);
2296 emit_loadreg(rs1[i]|64,th);
2302 if(s2l>=0) emit_not(s2l,tl);
2304 emit_loadreg(rs2[i],tl);
2307 if(s2h>=0) emit_not(s2h,th);
2309 emit_loadreg(rs2[i]|64,th);
2325 s1l=get_reg(i_regs->regmap,rs1[i]);
2326 s2l=get_reg(i_regs->regmap,rs2[i]);
2327 if(rs1[i]&&rs2[i]) {
2330 if(opcode2[i]==0x24) { // AND
2331 emit_and(s1l,s2l,tl);
2333 if(opcode2[i]==0x25) { // OR
2334 emit_or(s1l,s2l,tl);
2336 if(opcode2[i]==0x26) { // XOR
2337 emit_xor(s1l,s2l,tl);
2339 if(opcode2[i]==0x27) { // NOR
2340 emit_or(s1l,s2l,tl);
2346 if(opcode2[i]==0x24) { // AND
2349 if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2351 if(s1l>=0) emit_mov(s1l,tl);
2352 else emit_loadreg(rs1[i],tl); // CHECK: regmap_entry?
2356 if(s2l>=0) emit_mov(s2l,tl);
2357 else emit_loadreg(rs2[i],tl); // CHECK: regmap_entry?
2359 else emit_zeroreg(tl);
2361 if(opcode2[i]==0x27) { // NOR
2363 if(s1l>=0) emit_not(s1l,tl);
2365 emit_loadreg(rs1[i],tl);
2371 if(s2l>=0) emit_not(s2l,tl);
2373 emit_loadreg(rs2[i],tl);
2377 else emit_movimm(-1,tl);
2386 void imm16_assemble(int i,struct regstat *i_regs)
2388 if (opcode[i]==0x0f) { // LUI
2391 t=get_reg(i_regs->regmap,rt1[i]);
2394 if(!((i_regs->isconst>>t)&1))
2395 emit_movimm(imm[i]<<16,t);
2399 if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
2402 t=get_reg(i_regs->regmap,rt1[i]);
2403 s=get_reg(i_regs->regmap,rs1[i]);
2408 if(!((i_regs->isconst>>t)&1)) {
2410 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2411 emit_addimm(t,imm[i],t);
2413 if(!((i_regs->wasconst>>s)&1))
2414 emit_addimm(s,imm[i],t);
2416 emit_movimm(constmap[i][s]+imm[i],t);
2422 if(!((i_regs->isconst>>t)&1))
2423 emit_movimm(imm[i],t);
2428 if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
2430 signed char sh,sl,th,tl;
2431 th=get_reg(i_regs->regmap,rt1[i]|64);
2432 tl=get_reg(i_regs->regmap,rt1[i]);
2433 sh=get_reg(i_regs->regmap,rs1[i]|64);
2434 sl=get_reg(i_regs->regmap,rs1[i]);
2440 emit_addimm64_32(sh,sl,imm[i],th,tl);
2443 emit_addimm(sl,imm[i],tl);
2446 emit_movimm(imm[i],tl);
2447 if(th>=0) emit_movimm(((signed int)imm[i])>>31,th);
2452 else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
2454 //assert(rs1[i]!=0); // r0 might be valid, but it's probably a bug
2455 signed char sh,sl,t;
2456 t=get_reg(i_regs->regmap,rt1[i]);
2457 sh=get_reg(i_regs->regmap,rs1[i]|64);
2458 sl=get_reg(i_regs->regmap,rs1[i]);
2462 if(sh<0) assert((i_regs->was32>>rs1[i])&1);
2463 if(sh<0||((i_regs->was32>>rs1[i])&1)) {
2464 if(opcode[i]==0x0a) { // SLTI
2466 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2467 emit_slti32(t,imm[i],t);
2469 emit_slti32(sl,imm[i],t);
2474 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2475 emit_sltiu32(t,imm[i],t);
2477 emit_sltiu32(sl,imm[i],t);
2482 if(opcode[i]==0x0a) // SLTI
2483 emit_slti64_32(sh,sl,imm[i],t);
2485 emit_sltiu64_32(sh,sl,imm[i],t);
2488 // SLTI(U) with r0 is just stupid,
2489 // nonetheless examples can be found
2490 if(opcode[i]==0x0a) // SLTI
2491 if(0<imm[i]) emit_movimm(1,t);
2492 else emit_zeroreg(t);
2495 if(imm[i]) emit_movimm(1,t);
2496 else emit_zeroreg(t);
2502 else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
2504 signed char sh,sl,th,tl;
2505 th=get_reg(i_regs->regmap,rt1[i]|64);
2506 tl=get_reg(i_regs->regmap,rt1[i]);
2507 sh=get_reg(i_regs->regmap,rs1[i]|64);
2508 sl=get_reg(i_regs->regmap,rs1[i]);
2509 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2510 if(opcode[i]==0x0c) //ANDI
2514 if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2515 emit_andimm(tl,imm[i],tl);
2517 if(!((i_regs->wasconst>>sl)&1))
2518 emit_andimm(sl,imm[i],tl);
2520 emit_movimm(constmap[i][sl]&imm[i],tl);
2525 if(th>=0) emit_zeroreg(th);
2531 if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2535 emit_loadreg(rs1[i]|64,th);
2540 if(opcode[i]==0x0d) //ORI
2542 emit_orimm(tl,imm[i],tl);
2544 if(!((i_regs->wasconst>>sl)&1))
2545 emit_orimm(sl,imm[i],tl);
2547 emit_movimm(constmap[i][sl]|imm[i],tl);
2549 if(opcode[i]==0x0e) //XORI
2551 emit_xorimm(tl,imm[i],tl);
2553 if(!((i_regs->wasconst>>sl)&1))
2554 emit_xorimm(sl,imm[i],tl);
2556 emit_movimm(constmap[i][sl]^imm[i],tl);
2560 emit_movimm(imm[i],tl);
2561 if(th>=0) emit_zeroreg(th);
2569 void shiftimm_assemble(int i,struct regstat *i_regs)
2571 if(opcode2[i]<=0x3) // SLL/SRL/SRA
2575 t=get_reg(i_regs->regmap,rt1[i]);
2576 s=get_reg(i_regs->regmap,rs1[i]);
2585 if(s<0&&i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2587 if(opcode2[i]==0) // SLL
2589 emit_shlimm(s<0?t:s,imm[i],t);
2591 if(opcode2[i]==2) // SRL
2593 emit_shrimm(s<0?t:s,imm[i],t);
2595 if(opcode2[i]==3) // SRA
2597 emit_sarimm(s<0?t:s,imm[i],t);
2601 if(s>=0 && s!=t) emit_mov(s,t);
2605 //emit_storereg(rt1[i],t); //DEBUG
2608 if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
2611 signed char sh,sl,th,tl;
2612 th=get_reg(i_regs->regmap,rt1[i]|64);
2613 tl=get_reg(i_regs->regmap,rt1[i]);
2614 sh=get_reg(i_regs->regmap,rs1[i]|64);
2615 sl=get_reg(i_regs->regmap,rs1[i]);
2620 if(th>=0) emit_zeroreg(th);
2627 if(opcode2[i]==0x38) // DSLL
2629 if(th>=0) emit_shldimm(sh,sl,imm[i],th);
2630 emit_shlimm(sl,imm[i],tl);
2632 if(opcode2[i]==0x3a) // DSRL
2634 emit_shrdimm(sl,sh,imm[i],tl);
2635 if(th>=0) emit_shrimm(sh,imm[i],th);
2637 if(opcode2[i]==0x3b) // DSRA
2639 emit_shrdimm(sl,sh,imm[i],tl);
2640 if(th>=0) emit_sarimm(sh,imm[i],th);
2644 if(sl!=tl) emit_mov(sl,tl);
2645 if(th>=0&&sh!=th) emit_mov(sh,th);
2651 if(opcode2[i]==0x3c) // DSLL32
2654 signed char sl,tl,th;
2655 tl=get_reg(i_regs->regmap,rt1[i]);
2656 th=get_reg(i_regs->regmap,rt1[i]|64);
2657 sl=get_reg(i_regs->regmap,rs1[i]);
2666 emit_shlimm(th,imm[i]&31,th);
2671 if(opcode2[i]==0x3e) // DSRL32
2674 signed char sh,tl,th;
2675 tl=get_reg(i_regs->regmap,rt1[i]);
2676 th=get_reg(i_regs->regmap,rt1[i]|64);
2677 sh=get_reg(i_regs->regmap,rs1[i]|64);
2681 if(th>=0) emit_zeroreg(th);
2684 emit_shrimm(tl,imm[i]&31,tl);
2689 if(opcode2[i]==0x3f) // DSRA32
2693 tl=get_reg(i_regs->regmap,rt1[i]);
2694 sh=get_reg(i_regs->regmap,rs1[i]|64);
2700 emit_sarimm(tl,imm[i]&31,tl);
2707 #ifndef shift_assemble
2708 void shift_assemble(int i,struct regstat *i_regs)
2710 printf("Need shift_assemble for this architecture.\n");
2715 void load_assemble(int i,struct regstat *i_regs)
2717 int s,th,tl,addr,map=-1;
2720 int memtarget=0,c=0;
2722 th=get_reg(i_regs->regmap,rt1[i]|64);
2723 tl=get_reg(i_regs->regmap,rt1[i]);
2724 s=get_reg(i_regs->regmap,rs1[i]);
2726 for(hr=0;hr<HOST_REGS;hr++) {
2727 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2729 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2731 c=(i_regs->wasconst>>s)&1;
2732 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80800000;
2733 if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
2735 //printf("load_assemble: c=%d\n",c);
2736 //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2737 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2740 if(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80) {
2741 // could be FIFO, must perform the read
2742 assem_debug("(forced read)\n");
2743 tl=get_reg(i_regs->regmap,-1);
2747 if(offset||s<0||c) addr=tl;
2754 if(th>=0) reglist&=~(1<<th);
2757 //#define R29_HACK 1
2759 // Strmnnrmn's speed hack
2760 if(rs1[i]!=29||start<0x80001000||start>=0x80800000)
2763 emit_cmpimm(addr,0x800000);
2765 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2766 // Hint to branch predictor that the branch is unlikely to be taken
2768 emit_jno_unlikely(0);
2776 if (opcode[i]==0x20||opcode[i]==0x24) x=3; // LB/LBU
2777 if (opcode[i]==0x21||opcode[i]==0x25) x=2; // LH/LHU
2778 map=get_reg(i_regs->regmap,TLREG);
2780 map=do_tlb_r(addr,tl,map,x,-1,-1,c,constmap[i][s]+offset);
2781 do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr);
2783 if (opcode[i]==0x20) { // LB
2785 #ifdef HOST_IMM_ADDR32
2787 emit_movsbl_tlb((constmap[i][s]+offset)^3,map,tl);
2791 //emit_xorimm(addr,3,tl);
2792 //gen_tlb_addr_r(tl,map);
2793 //emit_movsbl_indexed((int)rdram-0x80000000,tl,tl);
2795 #ifdef BIG_ENDIAN_MIPS
2796 if(!c) emit_xorimm(addr,3,tl);
2797 else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2799 if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
2800 else if (tl!=addr) emit_mov(addr,tl);
2802 emit_movsbl_indexed_tlb(x,tl,map,tl);
2805 add_stub(LOADB_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2808 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2810 if (opcode[i]==0x21) { // LH
2812 #ifdef HOST_IMM_ADDR32
2814 emit_movswl_tlb((constmap[i][s]+offset)^2,map,tl);
2819 #ifdef BIG_ENDIAN_MIPS
2820 if(!c) emit_xorimm(addr,2,tl);
2821 else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2823 if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
2824 else if (tl!=addr) emit_mov(addr,tl);
2827 //emit_movswl_indexed_tlb(x,tl,map,tl);
2830 gen_tlb_addr_r(tl,map);
2831 emit_movswl_indexed(x,tl,tl);
2833 emit_movswl_indexed((int)rdram-0x80000000+x,tl,tl);
2836 add_stub(LOADH_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2839 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2841 if (opcode[i]==0x23) { // LW
2843 //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2844 #ifdef HOST_IMM_ADDR32
2846 emit_readword_tlb(constmap[i][s]+offset,map,tl);
2849 emit_readword_indexed_tlb(0,addr,map,tl);
2851 add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2854 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2856 if (opcode[i]==0x24) { // LBU
2858 #ifdef HOST_IMM_ADDR32
2860 emit_movzbl_tlb((constmap[i][s]+offset)^3,map,tl);
2864 //emit_xorimm(addr,3,tl);
2865 //gen_tlb_addr_r(tl,map);
2866 //emit_movzbl_indexed((int)rdram-0x80000000,tl,tl);
2868 #ifdef BIG_ENDIAN_MIPS
2869 if(!c) emit_xorimm(addr,3,tl);
2870 else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2872 if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
2873 else if (tl!=addr) emit_mov(addr,tl);
2875 emit_movzbl_indexed_tlb(x,tl,map,tl);
2878 add_stub(LOADBU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2881 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2883 if (opcode[i]==0x25) { // LHU
2885 #ifdef HOST_IMM_ADDR32
2887 emit_movzwl_tlb((constmap[i][s]+offset)^2,map,tl);
2892 #ifdef BIG_ENDIAN_MIPS
2893 if(!c) emit_xorimm(addr,2,tl);
2894 else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2896 if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
2897 else if (tl!=addr) emit_mov(addr,tl);
2900 //emit_movzwl_indexed_tlb(x,tl,map,tl);
2903 gen_tlb_addr_r(tl,map);
2904 emit_movzwl_indexed(x,tl,tl);
2906 emit_movzwl_indexed((int)rdram-0x80000000+x,tl,tl);
2908 add_stub(LOADHU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2912 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2914 if (opcode[i]==0x27) { // LWU
2917 //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2918 #ifdef HOST_IMM_ADDR32
2920 emit_readword_tlb(constmap[i][s]+offset,map,tl);
2923 emit_readword_indexed_tlb(0,addr,map,tl);
2925 add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2928 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2932 if (opcode[i]==0x37) { // LD
2934 //gen_tlb_addr_r(tl,map);
2935 //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,addr,th);
2936 //emit_readword_indexed((int)rdram-0x7FFFFFFC,addr,tl);
2937 #ifdef HOST_IMM_ADDR32
2939 emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
2942 emit_readdword_indexed_tlb(0,addr,map,th,tl);
2944 add_stub(LOADD_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2947 inline_readstub(LOADD_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2949 //emit_storereg(rt1[i],tl); // DEBUG
2951 //if(opcode[i]==0x23)
2952 //if(opcode[i]==0x24)
2953 //if(opcode[i]==0x23||opcode[i]==0x24)
2954 /*if(opcode[i]==0x21||opcode[i]==0x23||opcode[i]==0x24)
2958 emit_readword((int)&last_count,ECX);
2960 if(get_reg(i_regs->regmap,CCREG)<0)
2961 emit_loadreg(CCREG,HOST_CCREG);
2962 emit_add(HOST_CCREG,ECX,HOST_CCREG);
2963 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
2964 emit_writeword(HOST_CCREG,(int)&Count);
2967 if(get_reg(i_regs->regmap,CCREG)<0)
2968 emit_loadreg(CCREG,0);
2970 emit_mov(HOST_CCREG,0);
2972 emit_addimm(0,2*ccadj[i],0);
2973 emit_writeword(0,(int)&Count);
2975 emit_call((int)memdebug);
2977 restore_regs(0x100f);
2981 #ifndef loadlr_assemble
2982 void loadlr_assemble(int i,struct regstat *i_regs)
2984 printf("Need loadlr_assemble for this architecture.\n");
2989 void store_assemble(int i,struct regstat *i_regs)
2994 int jaddr=0,jaddr2,type;
2995 int memtarget=0,c=0;
2996 int agr=AGEN1+(i&1);
2998 th=get_reg(i_regs->regmap,rs2[i]|64);
2999 tl=get_reg(i_regs->regmap,rs2[i]);
3000 s=get_reg(i_regs->regmap,rs1[i]);
3001 temp=get_reg(i_regs->regmap,agr);
3002 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3005 c=(i_regs->wasconst>>s)&1;
3006 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80800000;
3007 if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
3011 for(hr=0;hr<HOST_REGS;hr++) {
3012 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3014 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3015 if(offset||s<0||c) addr=temp;
3020 // Strmnnrmn's speed hack
3022 if(rs1[i]!=29||start<0x80001000||start>=0x80800000)
3024 emit_cmpimm(addr,0x800000);
3025 #ifdef DESTRUCTIVE_SHIFT
3026 if(s==addr) emit_mov(s,temp);
3029 if(rs1[i]!=29||start<0x80001000||start>=0x80800000)
3033 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
3034 // Hint to branch predictor that the branch is unlikely to be taken
3036 emit_jno_unlikely(0);
3044 if (opcode[i]==0x28) x=3; // SB
3045 if (opcode[i]==0x29) x=2; // SH
3046 map=get_reg(i_regs->regmap,TLREG);
3048 map=do_tlb_w(addr,temp,map,x,c,constmap[i][s]+offset);
3049 do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
3052 if (opcode[i]==0x28) { // SB
3055 #ifdef BIG_ENDIAN_MIPS
3056 if(!c) emit_xorimm(addr,3,temp);
3057 else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
3059 if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
3060 else if (addr!=temp) emit_mov(addr,temp);
3062 //gen_tlb_addr_w(temp,map);
3063 //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
3064 emit_writebyte_indexed_tlb(tl,x,temp,map,temp);
3068 if (opcode[i]==0x29) { // SH
3071 #ifdef BIG_ENDIAN_MIPS
3072 if(!c) emit_xorimm(addr,2,temp);
3073 else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
3075 if(c) x=(constmap[i][s]+offset)-(constmap[i][s]+offset);
3076 else if (addr!=temp) emit_mov(addr,temp);
3079 //emit_writehword_indexed_tlb(tl,x,temp,map,temp);
3082 gen_tlb_addr_w(temp,map);
3083 emit_writehword_indexed(tl,x,temp);
3085 emit_writehword_indexed(tl,(int)rdram-0x80000000+x,temp);
3089 if (opcode[i]==0x2B) { // SW
3091 //emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
3092 emit_writeword_indexed_tlb(tl,0,addr,map,temp);
3095 if (opcode[i]==0x3F) { // SD
3099 //emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
3100 //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,addr);
3101 emit_writedword_indexed_tlb(th,tl,0,addr,map,temp);
3104 //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
3105 //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
3106 emit_writedword_indexed_tlb(tl,tl,0,addr,map,temp);
3111 if(!using_tlb&&(!c||memtarget))
3112 // addr could be a temp, make sure it survives STORE*_STUB
3115 add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
3116 } else if(!memtarget) {
3117 inline_writestub(type,i,constmap[i][s]+offset,i_regs->regmap,rs2[i],ccadj[i],reglist);
3121 #ifdef DESTRUCTIVE_SHIFT
3122 // The x86 shift operation is 'destructive'; it overwrites the
3123 // source register, so we need to make a copy first and use that.
3126 #if defined(HOST_IMM8)
3127 int ir=get_reg(i_regs->regmap,INVCP);
3129 emit_cmpmem_indexedsr12_reg(ir,addr,1);
3131 emit_cmpmem_indexedsr12_imm((int)invalid_code,addr,1);
3135 add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3138 //if(opcode[i]==0x2B || opcode[i]==0x3F)
3139 //if(opcode[i]==0x2B || opcode[i]==0x28)
3140 //if(opcode[i]==0x2B || opcode[i]==0x29)
3141 //if(opcode[i]==0x2B)
3142 /*if(opcode[i]==0x2B || opcode[i]==0x28 || opcode[i]==0x29 || opcode[i]==0x3F)
3146 emit_readword((int)&last_count,ECX);
3148 if(get_reg(i_regs->regmap,CCREG)<0)
3149 emit_loadreg(CCREG,HOST_CCREG);
3150 emit_add(HOST_CCREG,ECX,HOST_CCREG);
3151 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3152 emit_writeword(HOST_CCREG,(int)&Count);
3155 if(get_reg(i_regs->regmap,CCREG)<0)
3156 emit_loadreg(CCREG,0);
3158 emit_mov(HOST_CCREG,0);
3160 emit_addimm(0,2*ccadj[i],0);
3161 emit_writeword(0,(int)&Count);
3163 emit_call((int)memdebug);
3165 restore_regs(0x100f);
3169 void storelr_assemble(int i,struct regstat *i_regs)
3176 int case1,case2,case3;
3177 int done0,done1,done2;
3180 th=get_reg(i_regs->regmap,rs2[i]|64);
3181 tl=get_reg(i_regs->regmap,rs2[i]);
3182 s=get_reg(i_regs->regmap,rs1[i]);
3183 temp=get_reg(i_regs->regmap,-1);
3186 c=(i_regs->isconst>>s)&1;
3187 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80800000;
3188 if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1;
3191 for(hr=0;hr<HOST_REGS;hr++) {
3192 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3198 emit_cmpimm(s<0||offset?temp:s,0x800000);
3199 if(!offset&&s!=temp) emit_mov(s,temp);
3205 if(!memtarget||!rs1[i]) {
3210 if((u_int)rdram!=0x80000000)
3211 emit_addimm_no_flags((u_int)rdram-(u_int)0x80000000,temp);
3213 int map=get_reg(i_regs->regmap,TLREG);
3215 map=do_tlb_w(c||s<0||offset?temp:s,temp,map,0,c,constmap[i][s]+offset);
3216 if(!c&&!offset&&s>=0) emit_mov(s,temp);
3217 do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr);
3218 if(!jaddr&&!memtarget) {
3222 gen_tlb_addr_w(temp,map);
3225 if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
3226 temp2=get_reg(i_regs->regmap,FTEMP);
3227 if(!rs2[i]) temp2=th=tl;
3230 #ifndef BIG_ENDIAN_MIPS
3231 emit_xorimm(temp,3,temp);
3233 emit_testimm(temp,2);
3236 emit_testimm(temp,1);
3240 if (opcode[i]==0x2A) { // SWL
3241 emit_writeword_indexed(tl,0,temp);
3243 if (opcode[i]==0x2E) { // SWR
3244 emit_writebyte_indexed(tl,3,temp);
3246 if (opcode[i]==0x2C) { // SDL
3247 emit_writeword_indexed(th,0,temp);
3248 if(rs2[i]) emit_mov(tl,temp2);
3250 if (opcode[i]==0x2D) { // SDR
3251 emit_writebyte_indexed(tl,3,temp);
3252 if(rs2[i]) emit_shldimm(th,tl,24,temp2);
3257 set_jump_target(case1,(int)out);
3258 if (opcode[i]==0x2A) { // SWL
3259 // Write 3 msb into three least significant bytes
3260 if(rs2[i]) emit_rorimm(tl,8,tl);
3261 emit_writehword_indexed(tl,-1,temp);
3262 if(rs2[i]) emit_rorimm(tl,16,tl);
3263 emit_writebyte_indexed(tl,1,temp);
3264 if(rs2[i]) emit_rorimm(tl,8,tl);
3266 if (opcode[i]==0x2E) { // SWR
3267 // Write two lsb into two most significant bytes
3268 emit_writehword_indexed(tl,1,temp);
3270 if (opcode[i]==0x2C) { // SDL
3271 if(rs2[i]) emit_shrdimm(tl,th,8,temp2);
3272 // Write 3 msb into three least significant bytes
3273 if(rs2[i]) emit_rorimm(th,8,th);
3274 emit_writehword_indexed(th,-1,temp);
3275 if(rs2[i]) emit_rorimm(th,16,th);
3276 emit_writebyte_indexed(th,1,temp);
3277 if(rs2[i]) emit_rorimm(th,8,th);
3279 if (opcode[i]==0x2D) { // SDR
3280 if(rs2[i]) emit_shldimm(th,tl,16,temp2);
3281 // Write two lsb into two most significant bytes
3282 emit_writehword_indexed(tl,1,temp);
3287 set_jump_target(case2,(int)out);
3288 emit_testimm(temp,1);
3291 if (opcode[i]==0x2A) { // SWL
3292 // Write two msb into two least significant bytes
3293 if(rs2[i]) emit_rorimm(tl,16,tl);
3294 emit_writehword_indexed(tl,-2,temp);
3295 if(rs2[i]) emit_rorimm(tl,16,tl);
3297 if (opcode[i]==0x2E) { // SWR
3298 // Write 3 lsb into three most significant bytes
3299 emit_writebyte_indexed(tl,-1,temp);
3300 if(rs2[i]) emit_rorimm(tl,8,tl);
3301 emit_writehword_indexed(tl,0,temp);
3302 if(rs2[i]) emit_rorimm(tl,24,tl);
3304 if (opcode[i]==0x2C) { // SDL
3305 if(rs2[i]) emit_shrdimm(tl,th,16,temp2);
3306 // Write two msb into two least significant bytes
3307 if(rs2[i]) emit_rorimm(th,16,th);
3308 emit_writehword_indexed(th,-2,temp);
3309 if(rs2[i]) emit_rorimm(th,16,th);
3311 if (opcode[i]==0x2D) { // SDR
3312 if(rs2[i]) emit_shldimm(th,tl,8,temp2);
3313 // Write 3 lsb into three most significant bytes
3314 emit_writebyte_indexed(tl,-1,temp);
3315 if(rs2[i]) emit_rorimm(tl,8,tl);
3316 emit_writehword_indexed(tl,0,temp);
3317 if(rs2[i]) emit_rorimm(tl,24,tl);
3322 set_jump_target(case3,(int)out);
3323 if (opcode[i]==0x2A) { // SWL
3324 // Write msb into least significant byte
3325 if(rs2[i]) emit_rorimm(tl,24,tl);
3326 emit_writebyte_indexed(tl,-3,temp);
3327 if(rs2[i]) emit_rorimm(tl,8,tl);
3329 if (opcode[i]==0x2E) { // SWR
3330 // Write entire word
3331 emit_writeword_indexed(tl,-3,temp);
3333 if (opcode[i]==0x2C) { // SDL
3334 if(rs2[i]) emit_shrdimm(tl,th,24,temp2);
3335 // Write msb into least significant byte
3336 if(rs2[i]) emit_rorimm(th,24,th);
3337 emit_writebyte_indexed(th,-3,temp);
3338 if(rs2[i]) emit_rorimm(th,8,th);
3340 if (opcode[i]==0x2D) { // SDR
3341 if(rs2[i]) emit_mov(th,temp2);
3342 // Write entire word
3343 emit_writeword_indexed(tl,-3,temp);
3345 set_jump_target(done0,(int)out);
3346 set_jump_target(done1,(int)out);
3347 set_jump_target(done2,(int)out);
3348 if (opcode[i]==0x2C) { // SDL
3349 emit_testimm(temp,4);
3352 emit_andimm(temp,~3,temp);
3353 emit_writeword_indexed(temp2,4,temp);
3354 set_jump_target(done0,(int)out);
3356 if (opcode[i]==0x2D) { // SDR
3357 emit_testimm(temp,4);
3360 emit_andimm(temp,~3,temp);
3361 emit_writeword_indexed(temp2,-4,temp);
3362 set_jump_target(done0,(int)out);
3365 add_stub(STORELR_STUB,jaddr,(int)out,0,(int)i_regs,rs2[i],ccadj[i],reglist);
3368 emit_addimm_no_flags((u_int)0x80000000-(u_int)rdram,temp);
3369 #if defined(HOST_IMM8)
3370 int ir=get_reg(i_regs->regmap,INVCP);
3372 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3374 emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
3378 add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3382 //save_regs(0x100f);
3383 emit_readword((int)&last_count,ECX);
3384 if(get_reg(i_regs->regmap,CCREG)<0)
3385 emit_loadreg(CCREG,HOST_CCREG);
3386 emit_add(HOST_CCREG,ECX,HOST_CCREG);
3387 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3388 emit_writeword(HOST_CCREG,(int)&Count);
3389 emit_call((int)memdebug);
3391 //restore_regs(0x100f);
3395 void c1ls_assemble(int i,struct regstat *i_regs)
3397 #ifndef DISABLE_COP1
3403 int jaddr,jaddr2=0,jaddr3,type;
3404 int agr=AGEN1+(i&1);
3406 th=get_reg(i_regs->regmap,FTEMP|64);
3407 tl=get_reg(i_regs->regmap,FTEMP);
3408 s=get_reg(i_regs->regmap,rs1[i]);
3409 temp=get_reg(i_regs->regmap,agr);
3410 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3415 for(hr=0;hr<HOST_REGS;hr++) {
3416 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3418 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3419 if (opcode[i]==0x31||opcode[i]==0x35) // LWC1/LDC1
3421 // Loads use a temporary register which we need to save
3424 if (opcode[i]==0x39||opcode[i]==0x3D) // SWC1/SDC1
3428 //if(s<0) emit_loadreg(rs1[i],ar); //address_generation does this now
3429 //else c=(i_regs->wasconst>>s)&1;
3430 if(s>=0) c=(i_regs->wasconst>>s)&1;
3431 // Check cop1 unusable
3433 signed char rs=get_reg(i_regs->regmap,CSREG);
3435 emit_testimm(rs,0x20000000);
3438 add_stub(FP_STUB,jaddr,(int)out,i,rs,(int)i_regs,is_delayslot,0);
3441 if (opcode[i]==0x39) { // SWC1 (get float address)
3442 emit_readword((int)®_cop1_simple[(source[i]>>16)&0x1f],tl);
3444 if (opcode[i]==0x3D) { // SDC1 (get double address)
3445 emit_readword((int)®_cop1_double[(source[i]>>16)&0x1f],tl);
3447 // Generate address + offset
3450 emit_cmpimm(offset||c||s<0?ar:s,0x800000);
3454 map=get_reg(i_regs->regmap,TLREG);
3456 if (opcode[i]==0x31||opcode[i]==0x35) { // LWC1/LDC1
3457 map=do_tlb_r(offset||c||s<0?ar:s,ar,map,0,-1,-1,c,constmap[i][s]+offset);
3459 if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3460 map=do_tlb_w(offset||c||s<0?ar:s,ar,map,0,c,constmap[i][s]+offset);
3463 if (opcode[i]==0x39) { // SWC1 (read float)
3464 emit_readword_indexed(0,tl,tl);
3466 if (opcode[i]==0x3D) { // SDC1 (read double)
3467 emit_readword_indexed(4,tl,th);
3468 emit_readword_indexed(0,tl,tl);
3470 if (opcode[i]==0x31) { // LWC1 (get target address)
3471 emit_readword((int)®_cop1_simple[(source[i]>>16)&0x1f],temp);
3473 if (opcode[i]==0x35) { // LDC1 (get target address)
3474 emit_readword((int)®_cop1_double[(source[i]>>16)&0x1f],temp);
3481 else if(((signed int)(constmap[i][s]+offset))>=(signed int)0x80800000) {
3483 emit_jmp(0); // inline_readstub/inline_writestub? Very rare case
3485 #ifdef DESTRUCTIVE_SHIFT
3486 if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3487 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3491 if (opcode[i]==0x31||opcode[i]==0x35) { // LWC1/LDC1
3492 do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr2);
3494 if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3495 do_tlb_w_branch(map,c,constmap[i][s]+offset,&jaddr2);
3498 if (opcode[i]==0x31) { // LWC1
3499 //if(s>=0&&!c&&!offset) emit_mov(s,tl);
3500 //gen_tlb_addr_r(ar,map);
3501 //emit_readword_indexed((int)rdram-0x80000000,tl,tl);
3502 #ifdef HOST_IMM_ADDR32
3503 if(c) emit_readword_tlb(constmap[i][s]+offset,map,tl);
3506 emit_readword_indexed_tlb(0,offset||c||s<0?tl:s,map,tl);
3509 if (opcode[i]==0x35) { // LDC1
3511 //if(s>=0&&!c&&!offset) emit_mov(s,tl);
3512 //gen_tlb_addr_r(ar,map);
3513 //emit_readword_indexed((int)rdram-0x80000000,tl,th);
3514 //emit_readword_indexed((int)rdram-0x7FFFFFFC,tl,tl);
3515 #ifdef HOST_IMM_ADDR32
3516 if(c) emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
3519 emit_readdword_indexed_tlb(0,offset||c||s<0?tl:s,map,th,tl);
3522 if (opcode[i]==0x39) { // SWC1
3523 //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
3524 emit_writeword_indexed_tlb(tl,0,offset||c||s<0?temp:s,map,temp);
3527 if (opcode[i]==0x3D) { // SDC1
3529 //emit_writeword_indexed(th,(int)rdram-0x80000000,temp);
3530 //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
3531 emit_writedword_indexed_tlb(th,tl,0,offset||c||s<0?temp:s,map,temp);
3535 if (opcode[i]==0x39||opcode[i]==0x3D) { // SWC1/SDC1
3536 #ifndef DESTRUCTIVE_SHIFT
3537 temp=offset||c||s<0?ar:s;
3539 #if defined(HOST_IMM8)
3540 int ir=get_reg(i_regs->regmap,INVCP);
3542 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3544 emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
3548 add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3551 if(jaddr2) add_stub(type,jaddr2,(int)out,i,offset||c||s<0?ar:s,(int)i_regs,ccadj[i],reglist);
3552 if (opcode[i]==0x31) { // LWC1 (write float)
3553 emit_writeword_indexed(tl,0,temp);
3555 if (opcode[i]==0x35) { // LDC1 (write double)
3556 emit_writeword_indexed(th,4,temp);
3557 emit_writeword_indexed(tl,0,temp);
3559 //if(opcode[i]==0x39)
3560 /*if(opcode[i]==0x39||opcode[i]==0x31)
3563 emit_readword((int)&last_count,ECX);
3564 if(get_reg(i_regs->regmap,CCREG)<0)
3565 emit_loadreg(CCREG,HOST_CCREG);
3566 emit_add(HOST_CCREG,ECX,HOST_CCREG);
3567 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3568 emit_writeword(HOST_CCREG,(int)&Count);
3569 emit_call((int)memdebug);
3573 cop1_unusable(i, i_regs);
3577 void c2ls_assemble(int i,struct regstat *i_regs)
3583 int jaddr,jaddr2=0,jaddr3,type;
3584 int agr=AGEN1+(i&1);
3586 u_int copr=(source[i]>>16)&0x1f;
3587 s=get_reg(i_regs->regmap,rs1[i]);
3588 tl=get_reg(i_regs->regmap,FTEMP);
3594 for(hr=0;hr<HOST_REGS;hr++) {
3595 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3597 if(i_regs->regmap[HOST_CCREG]==CCREG)
3598 reglist&=~(1<<HOST_CCREG);
3601 if (opcode[i]==0x3a) { // SWC2
3602 ar=get_reg(i_regs->regmap,agr);
3603 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3608 if (!offset&&!c&&s>=0) ar=s;
3611 if (opcode[i]==0x3a) { // SWC2
3612 cop2_get_dreg(copr,tl,HOST_TEMPREG);
3614 if(s>=0) c=(i_regs->wasconst>>s)&1;
3616 emit_cmpimm(offset||c||s<0?ar:s,0x800000);
3620 else if(((signed int)(constmap[i][s]+offset))>=(signed int)0x80800000) {
3622 emit_jmp(0); // inline_readstub/inline_writestub? Very rare case
3624 if (opcode[i]==0x32) { // LWC2
3625 #ifdef HOST_IMM_ADDR32
3626 if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl);
3629 emit_readword_indexed(0,ar,tl);
3632 if (opcode[i]==0x3a) { // SWC2
3633 #ifdef DESTRUCTIVE_SHIFT
3634 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3636 emit_writeword_indexed(tl,0,ar);
3640 add_stub(type,jaddr2,(int)out,i,ar,(int)i_regs,ccadj[i],reglist);
3641 if (opcode[i]==0x3a) { // SWC2
3642 #if defined(HOST_IMM8)
3643 int ir=get_reg(i_regs->regmap,INVCP);
3645 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3647 emit_cmpmem_indexedsr12_imm((int)invalid_code,ar,1);
3651 add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3653 if (opcode[i]==0x32) { // LWC2
3654 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3658 #ifndef multdiv_assemble
3659 void multdiv_assemble(int i,struct regstat *i_regs)
3661 printf("Need multdiv_assemble for this architecture.\n");
3666 void mov_assemble(int i,struct regstat *i_regs)
3668 //if(opcode2[i]==0x10||opcode2[i]==0x12) { // MFHI/MFLO
3669 //if(opcode2[i]==0x11||opcode2[i]==0x13) { // MTHI/MTLO
3672 signed char sh,sl,th,tl;
3673 th=get_reg(i_regs->regmap,rt1[i]|64);
3674 tl=get_reg(i_regs->regmap,rt1[i]);
3677 sh=get_reg(i_regs->regmap,rs1[i]|64);
3678 sl=get_reg(i_regs->regmap,rs1[i]);
3679 if(sl>=0) emit_mov(sl,tl);
3680 else emit_loadreg(rs1[i],tl);
3682 if(sh>=0) emit_mov(sh,th);
3683 else emit_loadreg(rs1[i]|64,th);
3689 #ifndef fconv_assemble
3690 void fconv_assemble(int i,struct regstat *i_regs)
3692 printf("Need fconv_assemble for this architecture.\n");
3698 void float_assemble(int i,struct regstat *i_regs)
3700 printf("Need float_assemble for this architecture.\n");
3705 void syscall_assemble(int i,struct regstat *i_regs)
3707 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3708 assert(ccreg==HOST_CCREG);
3709 assert(!is_delayslot);
3710 emit_movimm(start+i*4,EAX); // Get PC
3711 emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3712 emit_jmp((int)jump_syscall_hle); // XXX
3715 void hlecall_assemble(int i,struct regstat *i_regs)
3717 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3718 assert(ccreg==HOST_CCREG);
3719 assert(!is_delayslot);
3720 emit_movimm(start+i*4+4,0); // Get PC
3721 emit_movimm((int)psxHLEt[source[i]&7],1);
3722 emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // XXX
3723 emit_jmp((int)jump_hlecall);
3726 void ds_assemble(int i,struct regstat *i_regs)
3731 alu_assemble(i,i_regs);break;
3733 imm16_assemble(i,i_regs);break;
3735 shift_assemble(i,i_regs);break;
3737 shiftimm_assemble(i,i_regs);break;
3739 load_assemble(i,i_regs);break;
3741 loadlr_assemble(i,i_regs);break;
3743 store_assemble(i,i_regs);break;
3745 storelr_assemble(i,i_regs);break;
3747 cop0_assemble(i,i_regs);break;
3749 cop1_assemble(i,i_regs);break;
3751 c1ls_assemble(i,i_regs);break;
3753 cop2_assemble(i,i_regs);break;
3755 c2ls_assemble(i,i_regs);break;
3757 c2op_assemble(i,i_regs);break;
3759 fconv_assemble(i,i_regs);break;
3761 float_assemble(i,i_regs);break;
3763 fcomp_assemble(i,i_regs);break;
3765 multdiv_assemble(i,i_regs);break;
3767 mov_assemble(i,i_regs);break;
3776 printf("Jump in the delay slot. This is probably a bug.\n");
3781 // Is the branch target a valid internal jump?
3782 int internal_branch(uint64_t i_is32,int addr)
3784 if(addr&1) return 0; // Indirect (register) jump
3785 if(addr>=start && addr<start+slen*4-4)
3787 int t=(addr-start)>>2;
3788 // Delay slots are not valid branch targets
3789 //if(t>0&&(itype[t-1]==RJUMP||itype[t-1]==UJUMP||itype[t-1]==CJUMP||itype[t-1]==SJUMP||itype[t-1]==FJUMP)) return 0;
3790 // 64 -> 32 bit transition requires a recompile
3791 /*if(is32[t]&~unneeded_reg_upper[t]&~i_is32)
3793 if(requires_32bit[t]&~i_is32) printf("optimizable: no\n");
3794 else printf("optimizable: yes\n");
3796 //if(is32[t]&~unneeded_reg_upper[t]&~i_is32) return 0;
3797 if(requires_32bit[t]&~i_is32) return 0;
3803 #ifndef wb_invalidate
3804 void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t is32,
3805 uint64_t u,uint64_t uu)
3808 for(hr=0;hr<HOST_REGS;hr++) {
3809 if(hr!=EXCLUDE_REG) {
3810 if(pre[hr]!=entry[hr]) {
3813 if(get_reg(entry,pre[hr])<0) {
3815 if(!((u>>pre[hr])&1)) {
3816 emit_storereg(pre[hr],hr);
3817 if( ((is32>>pre[hr])&1) && !((uu>>pre[hr])&1) ) {
3818 emit_sarimm(hr,31,hr);
3819 emit_storereg(pre[hr]|64,hr);
3823 if(!((uu>>(pre[hr]&63))&1) && !((is32>>(pre[hr]&63))&1)) {
3824 emit_storereg(pre[hr],hr);
3833 // Move from one register to another (no writeback)
3834 for(hr=0;hr<HOST_REGS;hr++) {
3835 if(hr!=EXCLUDE_REG) {
3836 if(pre[hr]!=entry[hr]) {
3837 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
3839 if((nr=get_reg(entry,pre[hr]))>=0) {
3849 // Load the specified registers
3850 // This only loads the registers given as arguments because
3851 // we don't want to load things that will be overwritten
3852 void load_regs(signed char entry[],signed char regmap[],int is32,int rs1,int rs2)
3856 for(hr=0;hr<HOST_REGS;hr++) {
3857 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3858 if(entry[hr]!=regmap[hr]) {
3859 if(regmap[hr]==rs1||regmap[hr]==rs2)
3866 emit_loadreg(regmap[hr],hr);
3873 for(hr=0;hr<HOST_REGS;hr++) {
3874 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3875 if(entry[hr]!=regmap[hr]) {
3876 if(regmap[hr]-64==rs1||regmap[hr]-64==rs2)
3878 assert(regmap[hr]!=64);
3879 if((is32>>(regmap[hr]&63))&1) {
3880 int lr=get_reg(regmap,regmap[hr]-64);
3882 emit_sarimm(lr,31,hr);
3884 emit_loadreg(regmap[hr],hr);
3888 emit_loadreg(regmap[hr],hr);
3896 // Load registers prior to the start of a loop
3897 // so that they are not loaded within the loop
3898 static void loop_preload(signed char pre[],signed char entry[])
3901 for(hr=0;hr<HOST_REGS;hr++) {
3902 if(hr!=EXCLUDE_REG) {
3903 if(pre[hr]!=entry[hr]) {
3905 if(get_reg(pre,entry[hr])<0) {
3906 assem_debug("loop preload:\n");
3907 //printf("loop preload: %d\n",hr);
3911 else if(entry[hr]<TEMPREG)
3913 emit_loadreg(entry[hr],hr);
3915 else if(entry[hr]-64<TEMPREG)
3917 emit_loadreg(entry[hr],hr);
3926 // Generate address for load/store instruction
3927 // goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
3928 void address_generation(int i,struct regstat *i_regs,signed char entry[])
3930 if(itype[i]==LOAD||itype[i]==LOADLR||itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS||itype[i]==C2LS) {
3932 int agr=AGEN1+(i&1);
3933 int mgr=MGEN1+(i&1);
3934 if(itype[i]==LOAD) {
3935 ra=get_reg(i_regs->regmap,rt1[i]);
3936 //if(rt1[i]) assert(ra>=0);
3938 if(itype[i]==LOADLR) {
3939 ra=get_reg(i_regs->regmap,FTEMP);
3941 if(itype[i]==STORE||itype[i]==STORELR) {
3942 ra=get_reg(i_regs->regmap,agr);
3943 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3945 if(itype[i]==C1LS||itype[i]==C2LS) {
3946 if ((opcode[i]&0x3b)==0x31||(opcode[i]&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
3947 ra=get_reg(i_regs->regmap,FTEMP);
3949 ra=get_reg(i_regs->regmap,agr);
3950 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3953 int rs=get_reg(i_regs->regmap,rs1[i]);
3954 int rm=get_reg(i_regs->regmap,TLREG);
3957 int c=(i_regs->wasconst>>rs)&1;
3959 // Using r0 as a base address
3961 if(!entry||entry[rm]!=mgr) {
3962 generate_map_const(offset,rm);
3963 } // else did it in the previous cycle
3965 if(!entry||entry[ra]!=agr) {
3966 if (opcode[i]==0x22||opcode[i]==0x26) {
3967 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
3968 }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
3969 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
3971 emit_movimm(offset,ra);
3973 } // else did it in the previous cycle
3976 if(!entry||entry[ra]!=rs1[i])
3977 emit_loadreg(rs1[i],ra);
3978 //if(!entry||entry[ra]!=rs1[i])
3979 // printf("poor load scheduling!\n");
3983 if(!entry||entry[rm]!=mgr) {
3984 if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a) {
3985 // Stores to memory go thru the mapper to detect self-modifying
3986 // code, loads don't.
3987 if((unsigned int)(constmap[i][rs]+offset)>=0xC0000000 ||
3988 (unsigned int)(constmap[i][rs]+offset)<0x80800000 )
3989 generate_map_const(constmap[i][rs]+offset,rm);
3991 if((signed int)(constmap[i][rs]+offset)>=(signed int)0xC0000000)
3992 generate_map_const(constmap[i][rs]+offset,rm);
3996 if(rs1[i]!=rt1[i]||itype[i]!=LOAD) {
3997 if(!entry||entry[ra]!=agr) {
3998 if (opcode[i]==0x22||opcode[i]==0x26) {
3999 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4000 }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
4001 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4003 #ifdef HOST_IMM_ADDR32
4004 if((itype[i]!=LOAD&&(opcode[i]&0x3b)!=0x31&&(opcode[i]&0x3b)!=0x32) || // LWC1/LDC1/LWC2/LDC2
4005 (using_tlb&&((signed int)constmap[i][rs]+offset)>=(signed int)0xC0000000))
4007 emit_movimm(constmap[i][rs]+offset,ra);
4009 } // else did it in the previous cycle
4010 } // else load_consts already did it
4012 if(offset&&!c&&rs1[i]) {
4014 emit_addimm(rs,offset,ra);
4016 emit_addimm(ra,offset,ra);
4021 // Preload constants for next instruction
4022 if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) {
4024 #ifndef HOST_IMM_ADDR32
4026 agr=MGEN1+((i+1)&1);
4027 ra=get_reg(i_regs->regmap,agr);
4029 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
4030 int offset=imm[i+1];
4031 int c=(regs[i+1].wasconst>>rs)&1;
4033 if(itype[i+1]==STORE||itype[i+1]==STORELR
4034 ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1, SWC2/SDC2
4035 // Stores to memory go thru the mapper to detect self-modifying
4036 // code, loads don't.
4037 if((unsigned int)(constmap[i+1][rs]+offset)>=0xC0000000 ||
4038 (unsigned int)(constmap[i+1][rs]+offset)<0x80800000 )
4039 generate_map_const(constmap[i+1][rs]+offset,ra);
4041 if((signed int)(constmap[i+1][rs]+offset)>=(signed int)0xC0000000)
4042 generate_map_const(constmap[i+1][rs]+offset,ra);
4045 /*else if(rs1[i]==0) {
4046 generate_map_const(offset,ra);
4051 agr=AGEN1+((i+1)&1);
4052 ra=get_reg(i_regs->regmap,agr);
4054 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
4055 int offset=imm[i+1];
4056 int c=(regs[i+1].wasconst>>rs)&1;
4057 if(c&&(rs1[i+1]!=rt1[i+1]||itype[i+1]!=LOAD)) {
4058 if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
4059 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4060 }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
4061 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4063 #ifdef HOST_IMM_ADDR32
4064 if((itype[i+1]!=LOAD&&(opcode[i+1]&0x3b)!=0x31&&(opcode[i+1]&0x3b)!=0x32) || // LWC1/LDC1/LWC2/LDC2
4065 (using_tlb&&((signed int)constmap[i+1][rs]+offset)>=(signed int)0xC0000000))
4067 emit_movimm(constmap[i+1][rs]+offset,ra);
4070 else if(rs1[i+1]==0) {
4071 // Using r0 as a base address
4072 if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
4073 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4074 }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
4075 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4077 emit_movimm(offset,ra);
4084 int get_final_value(int hr, int i, int *value)
4086 int reg=regs[i].regmap[hr];
4088 if(regs[i+1].regmap[hr]!=reg) break;
4089 if(!((regs[i+1].isconst>>hr)&1)) break;
4094 if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP) {
4095 *value=constmap[i][hr];
4099 if(itype[i+1]==UJUMP||itype[i+1]==RJUMP||itype[i+1]==CJUMP||itype[i+1]==SJUMP) {
4100 // Load in delay slot, out-of-order execution
4101 if(itype[i+2]==LOAD&&rs1[i+2]==reg&&rt1[i+2]==reg&&((regs[i+1].wasconst>>hr)&1))
4103 #ifdef HOST_IMM_ADDR32
4104 if(!using_tlb||((signed int)constmap[i][hr]+imm[i+2])<(signed int)0xC0000000) return 0;
4106 // Precompute load address
4107 *value=constmap[i][hr]+imm[i+2];
4111 if(itype[i+1]==LOAD&&rs1[i+1]==reg&&rt1[i+1]==reg)
4113 #ifdef HOST_IMM_ADDR32
4114 if(!using_tlb||((signed int)constmap[i][hr]+imm[i+1])<(signed int)0xC0000000) return 0;
4116 // Precompute load address
4117 *value=constmap[i][hr]+imm[i+1];
4118 //printf("c=%x imm=%x\n",(int)constmap[i][hr],imm[i+1]);
4123 *value=constmap[i][hr];
4124 //printf("c=%x\n",(int)constmap[i][hr]);
4125 if(i==slen-1) return 1;
4127 return !((unneeded_reg[i+1]>>reg)&1);
4129 return !((unneeded_reg_upper[i+1]>>reg)&1);
4133 // Load registers with known constants
4134 void load_consts(signed char pre[],signed char regmap[],int is32,int i)
4138 for(hr=0;hr<HOST_REGS;hr++) {
4139 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4140 //if(entry[hr]!=regmap[hr]) {
4141 if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
4142 if(((regs[i].isconst>>hr)&1)&®map[hr]<64&®map[hr]>0) {
4144 if(get_final_value(hr,i,&value)) {
4149 emit_movimm(value,hr);
4157 for(hr=0;hr<HOST_REGS;hr++) {
4158 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4159 //if(entry[hr]!=regmap[hr]) {
4160 if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
4161 if(((regs[i].isconst>>hr)&1)&®map[hr]>64) {
4162 if((is32>>(regmap[hr]&63))&1) {
4163 int lr=get_reg(regmap,regmap[hr]-64);
4165 emit_sarimm(lr,31,hr);
4170 if(get_final_value(hr,i,&value)) {
4175 emit_movimm(value,hr);
4184 void load_all_consts(signed char regmap[],int is32,u_int dirty,int i)
4188 for(hr=0;hr<HOST_REGS;hr++) {
4189 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
4190 if(((regs[i].isconst>>hr)&1)&®map[hr]<64&®map[hr]>0) {
4191 int value=constmap[i][hr];
4196 emit_movimm(value,hr);
4202 for(hr=0;hr<HOST_REGS;hr++) {
4203 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
4204 if(((regs[i].isconst>>hr)&1)&®map[hr]>64) {
4205 if((is32>>(regmap[hr]&63))&1) {
4206 int lr=get_reg(regmap,regmap[hr]-64);
4208 emit_sarimm(lr,31,hr);
4212 int value=constmap[i][hr];
4217 emit_movimm(value,hr);
4225 // Write out all dirty registers (except cycle count)
4226 void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty)
4229 for(hr=0;hr<HOST_REGS;hr++) {
4230 if(hr!=EXCLUDE_REG) {
4231 if(i_regmap[hr]>0) {
4232 if(i_regmap[hr]!=CCREG) {
4233 if((i_dirty>>hr)&1) {
4234 if(i_regmap[hr]<64) {
4235 emit_storereg(i_regmap[hr],hr);
4237 if( ((i_is32>>i_regmap[hr])&1) ) {
4238 #ifdef DESTRUCTIVE_WRITEBACK
4239 emit_sarimm(hr,31,hr);
4240 emit_storereg(i_regmap[hr]|64,hr);
4242 emit_sarimm(hr,31,HOST_TEMPREG);
4243 emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4248 if( !((i_is32>>(i_regmap[hr]&63))&1) ) {
4249 emit_storereg(i_regmap[hr],hr);
4258 // Write out dirty registers that we need to reload (pair with load_needed_regs)
4259 // This writes the registers not written by store_regs_bt
4260 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4263 int t=(addr-start)>>2;
4264 for(hr=0;hr<HOST_REGS;hr++) {
4265 if(hr!=EXCLUDE_REG) {
4266 if(i_regmap[hr]>0) {
4267 if(i_regmap[hr]!=CCREG) {
4268 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1) && !(((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
4269 if((i_dirty>>hr)&1) {
4270 if(i_regmap[hr]<64) {
4271 emit_storereg(i_regmap[hr],hr);
4273 if( ((i_is32>>i_regmap[hr])&1) ) {
4274 #ifdef DESTRUCTIVE_WRITEBACK
4275 emit_sarimm(hr,31,hr);
4276 emit_storereg(i_regmap[hr]|64,hr);
4278 emit_sarimm(hr,31,HOST_TEMPREG);
4279 emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4284 if( !((i_is32>>(i_regmap[hr]&63))&1) ) {
4285 emit_storereg(i_regmap[hr],hr);
4296 // Load all registers (except cycle count)
4297 void load_all_regs(signed char i_regmap[])
4300 for(hr=0;hr<HOST_REGS;hr++) {
4301 if(hr!=EXCLUDE_REG) {
4302 if(i_regmap[hr]==0) {
4306 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG)
4308 emit_loadreg(i_regmap[hr],hr);
4314 // Load all current registers also needed by next instruction
4315 void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
4318 for(hr=0;hr<HOST_REGS;hr++) {
4319 if(hr!=EXCLUDE_REG) {
4320 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4321 if(i_regmap[hr]==0) {
4325 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG)
4327 emit_loadreg(i_regmap[hr],hr);
4334 // Load all regs, storing cycle count if necessary
4335 void load_regs_entry(int t)
4338 if(is_ds[t]) emit_addimm(HOST_CCREG,CLOCK_DIVIDER,HOST_CCREG);
4339 else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t]*CLOCK_DIVIDER,HOST_CCREG);
4340 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4341 emit_storereg(CCREG,HOST_CCREG);
4344 for(hr=0;hr<HOST_REGS;hr++) {
4345 if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<64) {
4346 if(regs[t].regmap_entry[hr]==0) {
4349 else if(regs[t].regmap_entry[hr]!=CCREG)
4351 emit_loadreg(regs[t].regmap_entry[hr],hr);
4356 for(hr=0;hr<HOST_REGS;hr++) {
4357 if(regs[t].regmap_entry[hr]>=64) {
4358 assert(regs[t].regmap_entry[hr]!=64);
4359 if((regs[t].was32>>(regs[t].regmap_entry[hr]&63))&1) {
4360 int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4362 emit_loadreg(regs[t].regmap_entry[hr],hr);
4366 emit_sarimm(lr,31,hr);
4371 emit_loadreg(regs[t].regmap_entry[hr],hr);
4377 // Store dirty registers prior to branch
4378 void store_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4380 if(internal_branch(i_is32,addr))
4382 int t=(addr-start)>>2;
4384 for(hr=0;hr<HOST_REGS;hr++) {
4385 if(hr!=EXCLUDE_REG) {
4386 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4387 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1) || (((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
4388 if((i_dirty>>hr)&1) {
4389 if(i_regmap[hr]<64) {
4390 if(!((unneeded_reg[t]>>i_regmap[hr])&1)) {
4391 emit_storereg(i_regmap[hr],hr);
4392 if( ((i_is32>>i_regmap[hr])&1) && !((unneeded_reg_upper[t]>>i_regmap[hr])&1) ) {
4393 #ifdef DESTRUCTIVE_WRITEBACK
4394 emit_sarimm(hr,31,hr);
4395 emit_storereg(i_regmap[hr]|64,hr);
4397 emit_sarimm(hr,31,HOST_TEMPREG);
4398 emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4403 if( !((i_is32>>(i_regmap[hr]&63))&1) && !((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1) ) {
4404 emit_storereg(i_regmap[hr],hr);
4415 // Branch out of this block, write out all dirty regs
4416 wb_dirtys(i_regmap,i_is32,i_dirty);
4420 // Load all needed registers for branch target
4421 void load_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4423 //if(addr>=start && addr<(start+slen*4))
4424 if(internal_branch(i_is32,addr))
4426 int t=(addr-start)>>2;
4428 // Store the cycle count before loading something else
4429 if(i_regmap[HOST_CCREG]!=CCREG) {
4430 assert(i_regmap[HOST_CCREG]==-1);
4432 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4433 emit_storereg(CCREG,HOST_CCREG);
4436 for(hr=0;hr<HOST_REGS;hr++) {
4437 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<64) {
4438 #ifdef DESTRUCTIVE_WRITEBACK
4439 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || ( !((regs[t].dirty>>hr)&1) && ((i_dirty>>hr)&1) && (((i_is32&~unneeded_reg_upper[t])>>i_regmap[hr])&1) ) || (((i_is32&~regs[t].was32&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)) {
4441 if(i_regmap[hr]!=regs[t].regmap_entry[hr] ) {
4443 if(regs[t].regmap_entry[hr]==0) {
4446 else if(regs[t].regmap_entry[hr]!=CCREG)
4448 emit_loadreg(regs[t].regmap_entry[hr],hr);
4454 for(hr=0;hr<HOST_REGS;hr++) {
4455 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=64) {
4456 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4457 assert(regs[t].regmap_entry[hr]!=64);
4458 if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
4459 int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4461 emit_loadreg(regs[t].regmap_entry[hr],hr);
4465 emit_sarimm(lr,31,hr);
4470 emit_loadreg(regs[t].regmap_entry[hr],hr);
4473 else if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
4474 int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4476 emit_sarimm(lr,31,hr);
4483 int match_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4485 if(addr>=start && addr<start+slen*4-4)
4487 int t=(addr-start)>>2;
4489 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4490 for(hr=0;hr<HOST_REGS;hr++)
4494 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4496 if(regs[t].regmap_entry[hr]!=-1)
4505 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4510 if(!((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1))
4515 else // Same register but is it 32-bit or dirty?
4518 if(!((regs[t].dirty>>hr)&1))
4522 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4524 //printf("%x: dirty no match\n",addr);
4529 if((((regs[t].was32^i_is32)&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)
4531 //printf("%x: is32 no match\n",addr);
4537 //if(is32[t]&~unneeded_reg_upper[t]&~i_is32) return 0;
4538 if(requires_32bit[t]&~i_is32) return 0;
4539 // Delay slots are not valid branch targets
4540 //if(t>0&&(itype[t-1]==RJUMP||itype[t-1]==UJUMP||itype[t-1]==CJUMP||itype[t-1]==SJUMP||itype[t-1]==FJUMP)) return 0;
4541 // Delay slots require additional processing, so do not match
4542 if(is_ds[t]) return 0;
4547 for(hr=0;hr<HOST_REGS;hr++)
4553 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4567 // Used when a branch jumps into the delay slot of another branch
4568 void ds_assemble_entry(int i)
4570 int t=(ba[i]-start)>>2;
4571 if(!instr_addr[t]) instr_addr[t]=(u_int)out;
4572 assem_debug("Assemble delay slot at %x\n",ba[i]);
4573 assem_debug("<->\n");
4574 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&®s[t].regmap[HOST_CCREG]!=CCREG)
4575 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty,regs[t].was32);
4576 load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,rs1[t],rs2[t]);
4577 address_generation(t,®s[t],regs[t].regmap_entry);
4578 if(itype[t]==STORE||itype[t]==STORELR||(opcode[t]&0x3b)==0x39||(opcode[t]&0x3b)==0x3a)
4579 load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,INVCP,INVCP);
4584 alu_assemble(t,®s[t]);break;
4586 imm16_assemble(t,®s[t]);break;
4588 shift_assemble(t,®s[t]);break;
4590 shiftimm_assemble(t,®s[t]);break;
4592 load_assemble(t,®s[t]);break;
4594 loadlr_assemble(t,®s[t]);break;
4596 store_assemble(t,®s[t]);break;
4598 storelr_assemble(t,®s[t]);break;
4600 cop0_assemble(t,®s[t]);break;
4602 cop1_assemble(t,®s[t]);break;
4604 c1ls_assemble(t,®s[t]);break;
4606 cop2_assemble(t,®s[t]);break;
4608 c2ls_assemble(t,®s[t]);break;
4610 c2op_assemble(t,®s[t]);break;
4612 fconv_assemble(t,®s[t]);break;
4614 float_assemble(t,®s[t]);break;
4616 fcomp_assemble(t,®s[t]);break;
4618 multdiv_assemble(t,®s[t]);break;
4620 mov_assemble(t,®s[t]);break;
4629 printf("Jump in the delay slot. This is probably a bug.\n");
4631 store_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
4632 load_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
4633 if(internal_branch(regs[t].is32,ba[i]+4))
4634 assem_debug("branch: internal\n");
4636 assem_debug("branch: external\n");
4637 assert(internal_branch(regs[t].is32,ba[i]+4));
4638 add_to_linker((int)out,ba[i]+4,internal_branch(regs[t].is32,ba[i]+4));
4642 void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
4651 //if(ba[i]>=start && ba[i]<(start+slen*4))
4652 if(internal_branch(branch_regs[i].is32,ba[i]))
4654 int t=(ba[i]-start)>>2;
4655 if(is_ds[t]) *adj=-1; // Branch into delay slot adds an extra cycle
4663 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4665 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4667 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4668 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4672 else if(*adj==0||invert) {
4673 emit_addimm_and_set_flags(CLOCK_DIVIDER*(count+2),HOST_CCREG);
4679 emit_cmpimm(HOST_CCREG,-2*(count+2));
4683 add_stub(CC_STUB,jaddr,idle?idle:(int)out,(*adj==0||invert||idle)?0:(count+2),i,addr,taken,0);
4686 void do_ccstub(int n)
4689 assem_debug("do_ccstub %x\n",start+stubs[n][4]*4);
4690 set_jump_target(stubs[n][1],(int)out);
4692 if(stubs[n][6]==NULLDS) {
4693 // Delay slot instruction is nullified ("likely" branch)
4694 wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
4696 else if(stubs[n][6]!=TAKEN) {
4697 wb_dirtys(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty);
4700 if(internal_branch(branch_regs[i].is32,ba[i]))
4701 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
4705 // Save PC as return address
4706 emit_movimm(stubs[n][5],EAX);
4707 emit_writeword(EAX,(int)&pcaddr);
4711 // Return address depends on which way the branch goes
4712 if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
4714 int s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4715 int s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
4716 int s2l=get_reg(branch_regs[i].regmap,rs2[i]);
4717 int s2h=get_reg(branch_regs[i].regmap,rs2[i]|64);
4727 if((branch_regs[i].is32>>rs1[i])&(branch_regs[i].is32>>rs2[i])&1) {
4731 #ifdef DESTRUCTIVE_WRITEBACK
4733 if((branch_regs[i].dirty>>s1l)&(branch_regs[i].is32>>rs1[i])&1)
4734 emit_loadreg(rs1[i],s1l);
4737 if((branch_regs[i].dirty>>s1l)&(branch_regs[i].is32>>rs2[i])&1)
4738 emit_loadreg(rs2[i],s1l);
4741 if((branch_regs[i].dirty>>s2l)&(branch_regs[i].is32>>rs2[i])&1)
4742 emit_loadreg(rs2[i],s2l);
4745 int addr,alt,ntaddr;
4748 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4749 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4750 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4758 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4759 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4760 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4766 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
4770 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4771 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4772 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4778 assert(hr<HOST_REGS);
4780 if((opcode[i]&0x2f)==4) // BEQ
4782 #ifdef HAVE_CMOV_IMM
4784 if(s2l>=0) emit_cmp(s1l,s2l);
4785 else emit_test(s1l,s1l);
4786 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4791 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4793 if(s2h>=0) emit_cmp(s1h,s2h);
4794 else emit_test(s1h,s1h);
4795 emit_cmovne_reg(alt,addr);
4797 if(s2l>=0) emit_cmp(s1l,s2l);
4798 else emit_test(s1l,s1l);
4799 emit_cmovne_reg(alt,addr);
4802 if((opcode[i]&0x2f)==5) // BNE
4804 #ifdef HAVE_CMOV_IMM
4806 if(s2l>=0) emit_cmp(s1l,s2l);
4807 else emit_test(s1l,s1l);
4808 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4813 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4815 if(s2h>=0) emit_cmp(s1h,s2h);
4816 else emit_test(s1h,s1h);
4817 emit_cmovne_reg(alt,addr);
4819 if(s2l>=0) emit_cmp(s1l,s2l);
4820 else emit_test(s1l,s1l);
4821 emit_cmovne_reg(alt,addr);
4824 if((opcode[i]&0x2f)==6) // BLEZ
4826 //emit_movimm(ba[i],alt);
4827 //emit_movimm(start+i*4+8,addr);
4828 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4830 if(s1h>=0) emit_mov(addr,ntaddr);
4831 emit_cmovl_reg(alt,addr);
4834 emit_cmovne_reg(ntaddr,addr);
4835 emit_cmovs_reg(alt,addr);
4838 if((opcode[i]&0x2f)==7) // BGTZ
4840 //emit_movimm(ba[i],addr);
4841 //emit_movimm(start+i*4+8,ntaddr);
4842 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
4844 if(s1h>=0) emit_mov(addr,alt);
4845 emit_cmovl_reg(ntaddr,addr);
4848 emit_cmovne_reg(alt,addr);
4849 emit_cmovs_reg(ntaddr,addr);
4852 if((opcode[i]==1)&&(opcode2[i]&0x2D)==0) // BLTZ
4854 //emit_movimm(ba[i],alt);
4855 //emit_movimm(start+i*4+8,addr);
4856 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4857 if(s1h>=0) emit_test(s1h,s1h);
4858 else emit_test(s1l,s1l);
4859 emit_cmovs_reg(alt,addr);
4861 if((opcode[i]==1)&&(opcode2[i]&0x2D)==1) // BGEZ
4863 //emit_movimm(ba[i],addr);
4864 //emit_movimm(start+i*4+8,alt);
4865 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4866 if(s1h>=0) emit_test(s1h,s1h);
4867 else emit_test(s1l,s1l);
4868 emit_cmovs_reg(alt,addr);
4870 if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
4871 if(source[i]&0x10000) // BC1T
4873 //emit_movimm(ba[i],alt);
4874 //emit_movimm(start+i*4+8,addr);
4875 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4876 emit_testimm(s1l,0x800000);
4877 emit_cmovne_reg(alt,addr);
4881 //emit_movimm(ba[i],addr);
4882 //emit_movimm(start+i*4+8,alt);
4883 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4884 emit_testimm(s1l,0x800000);
4885 emit_cmovne_reg(alt,addr);
4888 emit_writeword(addr,(int)&pcaddr);
4893 int r=get_reg(branch_regs[i].regmap,rs1[i]);
4894 if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
4895 r=get_reg(branch_regs[i].regmap,RTEMP);
4897 emit_writeword(r,(int)&pcaddr);
4899 else {printf("Unknown branch type in do_ccstub\n");exit(1);}
4901 // Update cycle count
4902 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
4903 if(stubs[n][3]) emit_addimm(HOST_CCREG,CLOCK_DIVIDER*stubs[n][3],HOST_CCREG);
4904 emit_call((int)cc_interrupt);
4905 if(stubs[n][3]) emit_addimm(HOST_CCREG,-CLOCK_DIVIDER*stubs[n][3],HOST_CCREG);
4906 if(stubs[n][6]==TAKEN) {
4907 if(internal_branch(branch_regs[i].is32,ba[i]))
4908 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
4909 else if(itype[i]==RJUMP) {
4910 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
4911 emit_readword((int)&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
4913 emit_loadreg(rs1[i],get_reg(branch_regs[i].regmap,rs1[i]));
4915 }else if(stubs[n][6]==NOTTAKEN) {
4916 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
4917 else load_all_regs(branch_regs[i].regmap);
4918 }else if(stubs[n][6]==NULLDS) {
4919 // Delay slot instruction is nullified ("likely" branch)
4920 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
4921 else load_all_regs(regs[i].regmap);
4923 load_all_regs(branch_regs[i].regmap);
4925 emit_jmp(stubs[n][2]); // return address
4927 /* This works but uses a lot of memory...
4928 emit_readword((int)&last_count,ECX);
4929 emit_add(HOST_CCREG,ECX,EAX);
4930 emit_writeword(EAX,(int)&Count);
4931 emit_call((int)gen_interupt);
4932 emit_readword((int)&Count,HOST_CCREG);
4933 emit_readword((int)&next_interupt,EAX);
4934 emit_readword((int)&pending_exception,EBX);
4935 emit_writeword(EAX,(int)&last_count);
4936 emit_sub(HOST_CCREG,EAX,HOST_CCREG);
4938 int jne_instr=(int)out;
4940 if(stubs[n][3]) emit_addimm(HOST_CCREG,-2*stubs[n][3],HOST_CCREG);
4941 load_all_regs(branch_regs[i].regmap);
4942 emit_jmp(stubs[n][2]); // return address
4943 set_jump_target(jne_instr,(int)out);
4944 emit_readword((int)&pcaddr,EAX);
4945 // Call get_addr_ht instead of doing the hash table here.
4946 // This code is executed infrequently and takes up a lot of space
4947 // so smaller is better.
4948 emit_storereg(CCREG,HOST_CCREG);
4950 emit_call((int)get_addr_ht);
4951 emit_loadreg(CCREG,HOST_CCREG);
4952 emit_addimm(ESP,4,ESP);
4956 add_to_linker(int addr,int target,int ext)
4958 link_addr[linkcount][0]=addr;
4959 link_addr[linkcount][1]=target;
4960 link_addr[linkcount][2]=ext;
4964 void ujump_assemble(int i,struct regstat *i_regs)
4966 signed char *i_regmap=i_regs->regmap;
4967 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4968 address_generation(i+1,i_regs,regs[i].regmap_entry);
4970 int temp=get_reg(branch_regs[i].regmap,PTEMP);
4971 if(rt1[i]==31&&temp>=0)
4973 int return_address=start+i*4+8;
4974 if(get_reg(branch_regs[i].regmap,31)>0)
4975 if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
4978 ds_assemble(i+1,i_regs);
4979 uint64_t bc_unneeded=branch_regs[i].u;
4980 uint64_t bc_unneeded_upper=branch_regs[i].uu;
4981 bc_unneeded|=1|(1LL<<rt1[i]);
4982 bc_unneeded_upper|=1|(1LL<<rt1[i]);
4983 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
4984 bc_unneeded,bc_unneeded_upper);
4985 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
4988 unsigned int return_address;
4989 assert(rt1[i+1]!=31);
4990 assert(rt2[i+1]!=31);
4991 rt=get_reg(branch_regs[i].regmap,31);
4992 assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
4994 return_address=start+i*4+8;
4997 if(internal_branch(branch_regs[i].is32,return_address)) {
4999 if(temp==EXCLUDE_REG||temp>=HOST_REGS||
5000 branch_regs[i].regmap[temp]>=0)
5002 temp=get_reg(branch_regs[i].regmap,-1);
5005 if(temp<0) temp=HOST_TEMPREG;
5007 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5008 else emit_movimm(return_address,rt);
5016 if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5019 emit_movimm(return_address,rt); // PC into link register
5021 emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5027 cc=get_reg(branch_regs[i].regmap,CCREG);
5028 assert(cc==HOST_CCREG);
5029 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5031 if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
5033 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5034 if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5035 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5036 if(internal_branch(branch_regs[i].is32,ba[i]))
5037 assem_debug("branch: internal\n");
5039 assem_debug("branch: external\n");
5040 if(internal_branch(branch_regs[i].is32,ba[i])&&is_ds[(ba[i]-start)>>2]) {
5041 ds_assemble_entry(i);
5044 add_to_linker((int)out,ba[i],internal_branch(branch_regs[i].is32,ba[i]));
5049 void rjump_assemble(int i,struct regstat *i_regs)
5051 signed char *i_regmap=i_regs->regmap;
5054 rs=get_reg(branch_regs[i].regmap,rs1[i]);
5056 if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
5057 // Delay slot abuse, make a copy of the branch address register
5058 temp=get_reg(branch_regs[i].regmap,RTEMP);
5060 assert(regs[i].regmap[temp]==RTEMP);
5064 address_generation(i+1,i_regs,regs[i].regmap_entry);
5068 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5069 int return_address=start+i*4+8;
5070 if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5076 int rh=get_reg(regs[i].regmap,RHASH);
5077 if(rh>=0) do_preload_rhash(rh);
5080 ds_assemble(i+1,i_regs);
5081 uint64_t bc_unneeded=branch_regs[i].u;
5082 uint64_t bc_unneeded_upper=branch_regs[i].uu;
5083 bc_unneeded|=1|(1LL<<rt1[i]);
5084 bc_unneeded_upper|=1|(1LL<<rt1[i]);
5085 bc_unneeded&=~(1LL<<rs1[i]);
5086 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5087 bc_unneeded,bc_unneeded_upper);
5088 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],CCREG);
5090 int rt,return_address;
5091 assert(rt1[i+1]!=rt1[i]);
5092 assert(rt2[i+1]!=rt1[i]);
5093 rt=get_reg(branch_regs[i].regmap,rt1[i]);
5094 assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5096 return_address=start+i*4+8;
5100 if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
5103 emit_movimm(return_address,rt); // PC into link register
5105 emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5108 cc=get_reg(branch_regs[i].regmap,CCREG);
5109 assert(cc==HOST_CCREG);
5111 int rh=get_reg(branch_regs[i].regmap,RHASH);
5112 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5114 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5115 do_preload_rhtbl(ht);
5119 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
5120 #ifdef DESTRUCTIVE_WRITEBACK
5121 if((branch_regs[i].dirty>>rs)&(branch_regs[i].is32>>rs1[i])&1) {
5122 if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
5123 emit_loadreg(rs1[i],rs);
5128 if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
5132 do_miniht_load(ht,rh);
5135 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5136 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5138 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
5139 add_stub(CC_STUB,(int)out,jump_vaddr_reg[rs],0,i,-1,TAKEN,0);
5141 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
5144 do_miniht_jump(rs,rh,ht);
5149 //if(rs!=EAX) emit_mov(rs,EAX);
5150 //emit_jmp((int)jump_vaddr_eax);
5151 emit_jmp(jump_vaddr_reg[rs]);
5156 emit_shrimm(rs,16,rs);
5157 emit_xor(temp,rs,rs);
5158 emit_movzwl_reg(rs,rs);
5159 emit_shlimm(rs,4,rs);
5160 emit_cmpmem_indexed((int)hash_table,rs,temp);
5161 emit_jne((int)out+14);
5162 emit_readword_indexed((int)hash_table+4,rs,rs);
5164 emit_cmpmem_indexed((int)hash_table+8,rs,temp);
5165 emit_addimm_no_flags(8,rs);
5166 emit_jeq((int)out-17);
5167 // No hit on hash table, call compiler
5170 #ifdef DEBUG_CYCLE_COUNT
5171 emit_readword((int)&last_count,ECX);
5172 emit_add(HOST_CCREG,ECX,HOST_CCREG);
5173 emit_readword((int)&next_interupt,ECX);
5174 emit_writeword(HOST_CCREG,(int)&Count);
5175 emit_sub(HOST_CCREG,ECX,HOST_CCREG);
5176 emit_writeword(ECX,(int)&last_count);
5179 emit_storereg(CCREG,HOST_CCREG);
5180 emit_call((int)get_addr);
5181 emit_loadreg(CCREG,HOST_CCREG);
5182 emit_addimm(ESP,4,ESP);
5184 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5185 if(rt1[i]!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5189 void cjump_assemble(int i,struct regstat *i_regs)
5191 signed char *i_regmap=i_regs->regmap;
5194 match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5195 assem_debug("match=%d\n",match);
5196 int s1h,s1l,s2h,s2l;
5197 int prev_cop1_usable=cop1_usable;
5198 int unconditional=0,nop=0;
5202 int internal=internal_branch(branch_regs[i].is32,ba[i]);
5203 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5204 if(likely[i]) ooo=0;
5205 if(!match) invert=1;
5206 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5207 if(i>(ba[i]-start)>>2) invert=1;
5211 if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
5212 (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1])))
5214 // Write-after-read dependency prevents out of order execution
5215 // First test branch condition, then execute delay slot, then branch
5220 s1l=get_reg(branch_regs[i].regmap,rs1[i]);
5221 s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
5222 s2l=get_reg(branch_regs[i].regmap,rs2[i]);
5223 s2h=get_reg(branch_regs[i].regmap,rs2[i]|64);
5226 s1l=get_reg(i_regmap,rs1[i]);
5227 s1h=get_reg(i_regmap,rs1[i]|64);
5228 s2l=get_reg(i_regmap,rs2[i]);
5229 s2h=get_reg(i_regmap,rs2[i]|64);
5231 if(rs1[i]==0&&rs2[i]==0)
5233 if(opcode[i]&1) nop=1;
5234 else unconditional=1;
5235 //assert(opcode[i]!=5);
5236 //assert(opcode[i]!=7);
5237 //assert(opcode[i]!=0x15);
5238 //assert(opcode[i]!=0x17);
5244 only32=(regs[i].was32>>rs2[i])&1;
5249 only32=(regs[i].was32>>rs1[i])&1;
5252 only32=(regs[i].was32>>rs1[i])&(regs[i].was32>>rs2[i])&1;
5256 // Out of order execution (delay slot first)
5258 address_generation(i+1,i_regs,regs[i].regmap_entry);
5259 ds_assemble(i+1,i_regs);
5261 uint64_t bc_unneeded=branch_regs[i].u;
5262 uint64_t bc_unneeded_upper=branch_regs[i].uu;
5263 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5264 bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5266 bc_unneeded_upper|=1;
5267 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5268 bc_unneeded,bc_unneeded_upper);
5269 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs2[i]);
5270 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5271 cc=get_reg(branch_regs[i].regmap,CCREG);
5272 assert(cc==HOST_CCREG);
5274 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5275 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5276 //assem_debug("cycle count (adj)\n");
5278 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5279 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5280 if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5281 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5283 assem_debug("branch: internal\n");
5285 assem_debug("branch: external\n");
5286 if(internal&&is_ds[(ba[i]-start)>>2]) {
5287 ds_assemble_entry(i);
5290 add_to_linker((int)out,ba[i],internal);
5293 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5294 if(((u_int)out)&7) emit_addnop(0);
5299 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5302 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5305 int taken=0,nottaken=0,nottaken1=0;
5306 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5307 if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5311 if(opcode[i]==4) // BEQ
5313 if(s2h>=0) emit_cmp(s1h,s2h);
5314 else emit_test(s1h,s1h);
5318 if(opcode[i]==5) // BNE
5320 if(s2h>=0) emit_cmp(s1h,s2h);
5321 else emit_test(s1h,s1h);
5322 if(invert) taken=(int)out;
5323 else add_to_linker((int)out,ba[i],internal);
5326 if(opcode[i]==6) // BLEZ
5329 if(invert) taken=(int)out;
5330 else add_to_linker((int)out,ba[i],internal);
5335 if(opcode[i]==7) // BGTZ
5340 if(invert) taken=(int)out;
5341 else add_to_linker((int)out,ba[i],internal);
5346 //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5348 if(opcode[i]==4) // BEQ
5350 if(s2l>=0) emit_cmp(s1l,s2l);
5351 else emit_test(s1l,s1l);
5356 add_to_linker((int)out,ba[i],internal);
5360 if(opcode[i]==5) // BNE
5362 if(s2l>=0) emit_cmp(s1l,s2l);
5363 else emit_test(s1l,s1l);
5368 add_to_linker((int)out,ba[i],internal);
5372 if(opcode[i]==6) // BLEZ
5379 add_to_linker((int)out,ba[i],internal);
5383 if(opcode[i]==7) // BGTZ
5390 add_to_linker((int)out,ba[i],internal);
5395 if(taken) set_jump_target(taken,(int)out);
5396 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5397 if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5399 emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5400 add_to_linker((int)out,ba[i],internal);
5403 add_to_linker((int)out,ba[i],internal*2);
5409 if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5410 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5411 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5413 assem_debug("branch: internal\n");
5415 assem_debug("branch: external\n");
5416 if(internal&&is_ds[(ba[i]-start)>>2]) {
5417 ds_assemble_entry(i);
5420 add_to_linker((int)out,ba[i],internal);
5424 set_jump_target(nottaken,(int)out);
5427 if(nottaken1) set_jump_target(nottaken1,(int)out);
5429 if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
5431 } // (!unconditional)
5435 // In-order execution (branch first)
5436 //if(likely[i]) printf("IOL\n");
5439 int taken=0,nottaken=0,nottaken1=0;
5440 if(!unconditional&&!nop) {
5444 if((opcode[i]&0x2f)==4) // BEQ
5446 if(s2h>=0) emit_cmp(s1h,s2h);
5447 else emit_test(s1h,s1h);
5451 if((opcode[i]&0x2f)==5) // BNE
5453 if(s2h>=0) emit_cmp(s1h,s2h);
5454 else emit_test(s1h,s1h);
5458 if((opcode[i]&0x2f)==6) // BLEZ
5466 if((opcode[i]&0x2f)==7) // BGTZ
5476 //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5478 if((opcode[i]&0x2f)==4) // BEQ
5480 if(s2l>=0) emit_cmp(s1l,s2l);
5481 else emit_test(s1l,s1l);
5485 if((opcode[i]&0x2f)==5) // BNE
5487 if(s2l>=0) emit_cmp(s1l,s2l);
5488 else emit_test(s1l,s1l);
5492 if((opcode[i]&0x2f)==6) // BLEZ
5498 if((opcode[i]&0x2f)==7) // BGTZ
5504 } // if(!unconditional)
5506 uint64_t ds_unneeded=branch_regs[i].u;
5507 uint64_t ds_unneeded_upper=branch_regs[i].uu;
5508 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5509 ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
5510 if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
5512 ds_unneeded_upper|=1;
5515 if(taken) set_jump_target(taken,(int)out);
5516 assem_debug("1:\n");
5517 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5518 ds_unneeded,ds_unneeded_upper);
5520 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5521 address_generation(i+1,&branch_regs[i],0);
5522 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
5523 ds_assemble(i+1,&branch_regs[i]);
5524 cc=get_reg(branch_regs[i].regmap,CCREG);
5526 emit_loadreg(CCREG,cc=HOST_CCREG);
5527 // CHECK: Is the following instruction (fall thru) allocated ok?
5529 assert(cc==HOST_CCREG);
5530 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5531 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5532 assem_debug("cycle count (adj)\n");
5533 if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5534 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5536 assem_debug("branch: internal\n");
5538 assem_debug("branch: external\n");
5539 if(internal&&is_ds[(ba[i]-start)>>2]) {
5540 ds_assemble_entry(i);
5543 add_to_linker((int)out,ba[i],internal);
5548 cop1_usable=prev_cop1_usable;
5549 if(!unconditional) {
5550 if(nottaken1) set_jump_target(nottaken1,(int)out);
5551 set_jump_target(nottaken,(int)out);
5552 assem_debug("2:\n");
5554 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5555 ds_unneeded,ds_unneeded_upper);
5556 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5557 address_generation(i+1,&branch_regs[i],0);
5558 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5559 ds_assemble(i+1,&branch_regs[i]);
5561 cc=get_reg(branch_regs[i].regmap,CCREG);
5562 if(cc==-1&&!likely[i]) {
5563 // Cycle count isn't in a register, temporarily load it then write it out
5564 emit_loadreg(CCREG,HOST_CCREG);
5565 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
5568 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5569 emit_storereg(CCREG,HOST_CCREG);
5572 cc=get_reg(i_regmap,CCREG);
5573 assert(cc==HOST_CCREG);
5574 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5577 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5583 void sjump_assemble(int i,struct regstat *i_regs)
5585 signed char *i_regmap=i_regs->regmap;
5588 match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5589 assem_debug("smatch=%d\n",match);
5591 int prev_cop1_usable=cop1_usable;
5592 int unconditional=0,nevertaken=0;
5596 int internal=internal_branch(branch_regs[i].is32,ba[i]);
5597 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5598 if(likely[i]) ooo=0;
5599 if(!match) invert=1;
5600 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5601 if(i>(ba[i]-start)>>2) invert=1;
5604 //if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
5605 assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
5608 if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))
5610 // Write-after-read dependency prevents out of order execution
5611 // First test branch condition, then execute delay slot, then branch
5614 // TODO: Conditional branches w/link must execute in-order so that
5615 // condition test and write to r31 occur before cycle count test
5618 s1l=get_reg(branch_regs[i].regmap,rs1[i]);
5619 s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
5622 s1l=get_reg(i_regmap,rs1[i]);
5623 s1h=get_reg(i_regmap,rs1[i]|64);
5627 if(opcode2[i]&1) unconditional=1;
5629 // These are never taken (r0 is never less than zero)
5630 //assert(opcode2[i]!=0);
5631 //assert(opcode2[i]!=2);
5632 //assert(opcode2[i]!=0x10);
5633 //assert(opcode2[i]!=0x12);
5636 only32=(regs[i].was32>>rs1[i])&1;
5640 // Out of order execution (delay slot first)
5642 address_generation(i+1,i_regs,regs[i].regmap_entry);
5643 ds_assemble(i+1,i_regs);
5645 uint64_t bc_unneeded=branch_regs[i].u;
5646 uint64_t bc_unneeded_upper=branch_regs[i].uu;
5647 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5648 bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5650 bc_unneeded_upper|=1;
5651 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5652 bc_unneeded,bc_unneeded_upper);
5653 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs1[i]);
5654 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5656 int rt,return_address;
5657 assert(rt1[i+1]!=31);
5658 assert(rt2[i+1]!=31);
5659 rt=get_reg(branch_regs[i].regmap,31);
5660 assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5662 // Save the PC even if the branch is not taken
5663 return_address=start+i*4+8;
5664 emit_movimm(return_address,rt); // PC into link register
5666 if(!nevertaken) emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5670 cc=get_reg(branch_regs[i].regmap,CCREG);
5671 assert(cc==HOST_CCREG);
5673 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5674 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5675 assem_debug("cycle count (adj)\n");
5677 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5678 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5679 if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5680 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5682 assem_debug("branch: internal\n");
5684 assem_debug("branch: external\n");
5685 if(internal&&is_ds[(ba[i]-start)>>2]) {
5686 ds_assemble_entry(i);
5689 add_to_linker((int)out,ba[i],internal);
5692 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5693 if(((u_int)out)&7) emit_addnop(0);
5697 else if(nevertaken) {
5698 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5701 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5705 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5706 if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5710 if(opcode2[i]==0) // BLTZ
5717 add_to_linker((int)out,ba[i],internal);
5721 if(opcode2[i]==1) // BGEZ
5728 add_to_linker((int)out,ba[i],internal);
5736 if(opcode2[i]==0) // BLTZ
5743 add_to_linker((int)out,ba[i],internal);
5747 if(opcode2[i]==1) // BGEZ
5754 add_to_linker((int)out,ba[i],internal);
5761 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5762 if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5764 emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5765 add_to_linker((int)out,ba[i],internal);
5768 add_to_linker((int)out,ba[i],internal*2);
5774 if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
5775 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5776 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5778 assem_debug("branch: internal\n");
5780 assem_debug("branch: external\n");
5781 if(internal&&is_ds[(ba[i]-start)>>2]) {
5782 ds_assemble_entry(i);
5785 add_to_linker((int)out,ba[i],internal);
5789 set_jump_target(nottaken,(int)out);
5793 if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
5795 } // (!unconditional)
5799 // In-order execution (branch first)
5802 if(!unconditional) {
5803 //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5807 if((opcode2[i]&0x1d)==0) // BLTZ/BLTZL
5813 if((opcode2[i]&0x1d)==1) // BGEZ/BGEZL
5823 if((opcode2[i]&0x1d)==0) // BLTZ/BLTZL
5829 if((opcode2[i]&0x1d)==1) // BGEZ/BGEZL
5836 } // if(!unconditional)
5838 uint64_t ds_unneeded=branch_regs[i].u;
5839 uint64_t ds_unneeded_upper=branch_regs[i].uu;
5840 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5841 ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
5842 if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
5844 ds_unneeded_upper|=1;
5847 //assem_debug("1:\n");
5848 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5849 ds_unneeded,ds_unneeded_upper);
5851 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5852 address_generation(i+1,&branch_regs[i],0);
5853 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
5854 ds_assemble(i+1,&branch_regs[i]);
5855 cc=get_reg(branch_regs[i].regmap,CCREG);
5857 emit_loadreg(CCREG,cc=HOST_CCREG);
5858 // CHECK: Is the following instruction (fall thru) allocated ok?
5860 assert(cc==HOST_CCREG);
5861 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5862 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5863 assem_debug("cycle count (adj)\n");
5864 if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5865 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5867 assem_debug("branch: internal\n");
5869 assem_debug("branch: external\n");
5870 if(internal&&is_ds[(ba[i]-start)>>2]) {
5871 ds_assemble_entry(i);
5874 add_to_linker((int)out,ba[i],internal);
5879 cop1_usable=prev_cop1_usable;
5880 if(!unconditional) {
5881 set_jump_target(nottaken,(int)out);
5882 assem_debug("1:\n");
5884 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5885 ds_unneeded,ds_unneeded_upper);
5886 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5887 address_generation(i+1,&branch_regs[i],0);
5888 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5889 ds_assemble(i+1,&branch_regs[i]);
5891 cc=get_reg(branch_regs[i].regmap,CCREG);
5892 if(cc==-1&&!likely[i]) {
5893 // Cycle count isn't in a register, temporarily load it then write it out
5894 emit_loadreg(CCREG,HOST_CCREG);
5895 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
5898 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5899 emit_storereg(CCREG,HOST_CCREG);
5902 cc=get_reg(i_regmap,CCREG);
5903 assert(cc==HOST_CCREG);
5904 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
5907 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5913 void fjump_assemble(int i,struct regstat *i_regs)
5915 signed char *i_regmap=i_regs->regmap;
5918 match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5919 assem_debug("fmatch=%d\n",match);
5924 int internal=internal_branch(branch_regs[i].is32,ba[i]);
5925 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5926 if(likely[i]) ooo=0;
5927 if(!match) invert=1;
5928 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5929 if(i>(ba[i]-start)>>2) invert=1;
5933 if(itype[i+1]==FCOMP)
5935 // Write-after-read dependency prevents out of order execution
5936 // First test branch condition, then execute delay slot, then branch
5941 fs=get_reg(branch_regs[i].regmap,FSREG);
5942 address_generation(i+1,i_regs,regs[i].regmap_entry); // Is this okay?
5945 fs=get_reg(i_regmap,FSREG);
5948 // Check cop1 unusable
5950 cs=get_reg(i_regmap,CSREG);
5952 emit_testimm(cs,0x20000000);
5955 add_stub(FP_STUB,eaddr,(int)out,i,cs,(int)i_regs,0,0);
5960 // Out of order execution (delay slot first)
5962 ds_assemble(i+1,i_regs);
5964 uint64_t bc_unneeded=branch_regs[i].u;
5965 uint64_t bc_unneeded_upper=branch_regs[i].uu;
5966 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5967 bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5969 bc_unneeded_upper|=1;
5970 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5971 bc_unneeded,bc_unneeded_upper);
5972 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs1[i]);
5973 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5974 cc=get_reg(branch_regs[i].regmap,CCREG);
5975 assert(cc==HOST_CCREG);
5976 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5977 assem_debug("cycle count (adj)\n");
5980 if(adj&&!invert) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
5983 emit_testimm(fs,0x800000);
5984 if(source[i]&0x10000) // BC1T
5990 add_to_linker((int)out,ba[i],internal);
5999 add_to_linker((int)out,ba[i],internal);
6007 if(adj) emit_addimm(cc,-CLOCK_DIVIDER*adj,cc);
6008 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
6009 else if(match) emit_addnop(13);
6011 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6012 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6014 assem_debug("branch: internal\n");
6016 assem_debug("branch: external\n");
6017 if(internal&&is_ds[(ba[i]-start)>>2]) {
6018 ds_assemble_entry(i);
6021 add_to_linker((int)out,ba[i],internal);
6024 set_jump_target(nottaken,(int)out);
6028 if(!invert) emit_addimm(cc,CLOCK_DIVIDER*adj,cc);
6030 } // (!unconditional)
6034 // In-order execution (branch first)
6038 //printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
6041 emit_testimm(fs,0x800000);
6042 if(source[i]&0x10000) // BC1T
6053 } // if(!unconditional)
6055 uint64_t ds_unneeded=branch_regs[i].u;
6056 uint64_t ds_unneeded_upper=branch_regs[i].uu;
6057 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6058 ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6059 if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
6061 ds_unneeded_upper|=1;
6063 //assem_debug("1:\n");
6064 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6065 ds_unneeded,ds_unneeded_upper);
6067 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
6068 address_generation(i+1,&branch_regs[i],0);
6069 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
6070 ds_assemble(i+1,&branch_regs[i]);
6071 cc=get_reg(branch_regs[i].regmap,CCREG);
6073 emit_loadreg(CCREG,cc=HOST_CCREG);
6074 // CHECK: Is the following instruction (fall thru) allocated ok?
6076 assert(cc==HOST_CCREG);
6077 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6078 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
6079 assem_debug("cycle count (adj)\n");
6080 if(adj) emit_addimm(cc,CLOCK_DIVIDER*(ccadj[i]+2-adj),cc);
6081 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
6083 assem_debug("branch: internal\n");
6085 assem_debug("branch: external\n");
6086 if(internal&&is_ds[(ba[i]-start)>>2]) {
6087 ds_assemble_entry(i);
6090 add_to_linker((int)out,ba[i],internal);
6095 if(1) { // <- FIXME (don't need this)
6096 set_jump_target(nottaken,(int)out);
6097 assem_debug("1:\n");
6099 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
6100 ds_unneeded,ds_unneeded_upper);
6101 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
6102 address_generation(i+1,&branch_regs[i],0);
6103 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
6104 ds_assemble(i+1,&branch_regs[i]);
6106 cc=get_reg(branch_regs[i].regmap,CCREG);
6107 if(cc==-1&&!likely[i]) {
6108 // Cycle count isn't in a register, temporarily load it then write it out
6109 emit_loadreg(CCREG,HOST_CCREG);
6110 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
6113 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
6114 emit_storereg(CCREG,HOST_CCREG);
6117 cc=get_reg(i_regmap,CCREG);
6118 assert(cc==HOST_CCREG);
6119 emit_addimm_and_set_flags(CLOCK_DIVIDER*(ccadj[i]+2),cc);
6122 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
6128 static void pagespan_assemble(int i,struct regstat *i_regs)
6130 int s1l=get_reg(i_regs->regmap,rs1[i]);
6131 int s1h=get_reg(i_regs->regmap,rs1[i]|64);
6132 int s2l=get_reg(i_regs->regmap,rs2[i]);
6133 int s2h=get_reg(i_regs->regmap,rs2[i]|64);
6134 void *nt_branch=NULL;
6137 int unconditional=0;
6147 if((i_regs->is32>>rs1[i])&(i_regs->is32>>rs2[i])&1) {
6151 int addr,alt,ntaddr;
6152 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
6156 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
6157 (i_regs->regmap[hr]&63)!=rs1[i] &&
6158 (i_regs->regmap[hr]&63)!=rs2[i] )
6167 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
6168 (i_regs->regmap[hr]&63)!=rs1[i] &&
6169 (i_regs->regmap[hr]&63)!=rs2[i] )
6175 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
6179 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
6180 (i_regs->regmap[hr]&63)!=rs1[i] &&
6181 (i_regs->regmap[hr]&63)!=rs2[i] )
6188 assert(hr<HOST_REGS);
6189 if((opcode[i]&0x2e)==4||opcode[i]==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
6190 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
6192 emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i]+2),HOST_CCREG);
6193 if(opcode[i]==2) // J
6197 if(opcode[i]==3) // JAL
6200 int rt=get_reg(i_regs->regmap,31);
6201 emit_movimm(start+i*4+8,rt);
6204 if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
6207 if(opcode2[i]==9) // JALR
6209 int rt=get_reg(i_regs->regmap,rt1[i]);
6210 emit_movimm(start+i*4+8,rt);
6213 if((opcode[i]&0x3f)==4) // BEQ
6220 #ifdef HAVE_CMOV_IMM
6222 if(s2l>=0) emit_cmp(s1l,s2l);
6223 else emit_test(s1l,s1l);
6224 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
6230 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6232 if(s2h>=0) emit_cmp(s1h,s2h);
6233 else emit_test(s1h,s1h);
6234 emit_cmovne_reg(alt,addr);
6236 if(s2l>=0) emit_cmp(s1l,s2l);
6237 else emit_test(s1l,s1l);
6238 emit_cmovne_reg(alt,addr);
6241 if((opcode[i]&0x3f)==5) // BNE
6243 #ifdef HAVE_CMOV_IMM
6245 if(s2l>=0) emit_cmp(s1l,s2l);
6246 else emit_test(s1l,s1l);
6247 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
6253 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
6255 if(s2h>=0) emit_cmp(s1h,s2h);
6256 else emit_test(s1h,s1h);
6257 emit_cmovne_reg(alt,addr);
6259 if(s2l>=0) emit_cmp(s1l,s2l);
6260 else emit_test(s1l,s1l);
6261 emit_cmovne_reg(alt,addr);
6264 if((opcode[i]&0x3f)==0x14) // BEQL
6267 if(s2h>=0) emit_cmp(s1h,s2h);
6268 else emit_test(s1h,s1h);
6272 if(s2l>=0) emit_cmp(s1l,s2l);
6273 else emit_test(s1l,s1l);
6274 if(nottaken) set_jump_target(nottaken,(int)out);
6278 if((opcode[i]&0x3f)==0x15) // BNEL
6281 if(s2h>=0) emit_cmp(s1h,s2h);
6282 else emit_test(s1h,s1h);
6286 if(s2l>=0) emit_cmp(s1l,s2l);
6287 else emit_test(s1l,s1l);
6290 if(taken) set_jump_target(taken,(int)out);
6292 if((opcode[i]&0x3f)==6) // BLEZ
6294 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6296 if(s1h>=0) emit_mov(addr,ntaddr);
6297 emit_cmovl_reg(alt,addr);
6300 emit_cmovne_reg(ntaddr,addr);
6301 emit_cmovs_reg(alt,addr);
6304 if((opcode[i]&0x3f)==7) // BGTZ
6306 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
6308 if(s1h>=0) emit_mov(addr,alt);
6309 emit_cmovl_reg(ntaddr,addr);
6312 emit_cmovne_reg(alt,addr);
6313 emit_cmovs_reg(ntaddr,addr);
6316 if((opcode[i]&0x3f)==0x16) // BLEZL
6318 assert((opcode[i]&0x3f)!=0x16);
6320 if((opcode[i]&0x3f)==0x17) // BGTZL
6322 assert((opcode[i]&0x3f)!=0x17);
6324 assert(opcode[i]!=1); // BLTZ/BGEZ
6326 //FIXME: Check CSREG
6327 if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
6328 if((source[i]&0x30000)==0) // BC1F
6330 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6331 emit_testimm(s1l,0x800000);
6332 emit_cmovne_reg(alt,addr);
6334 if((source[i]&0x30000)==0x10000) // BC1T
6336 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6337 emit_testimm(s1l,0x800000);
6338 emit_cmovne_reg(alt,addr);
6340 if((source[i]&0x30000)==0x20000) // BC1FL
6342 emit_testimm(s1l,0x800000);
6346 if((source[i]&0x30000)==0x30000) // BC1TL
6348 emit_testimm(s1l,0x800000);
6354 assert(i_regs->regmap[HOST_CCREG]==CCREG);
6355 wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
6356 if(likely[i]||unconditional)
6358 emit_movimm(ba[i],HOST_BTREG);
6360 else if(addr!=HOST_BTREG)
6362 emit_mov(addr,HOST_BTREG);
6364 void *branch_addr=out;
6366 int target_addr=start+i*4+5;
6368 void *compiled_target_addr=check_addr(target_addr);
6369 emit_extjump_ds((int)branch_addr,target_addr);
6370 if(compiled_target_addr) {
6371 set_jump_target((int)branch_addr,(int)compiled_target_addr);
6372 add_link(target_addr,stub);
6374 else set_jump_target((int)branch_addr,(int)stub);
6377 set_jump_target((int)nottaken,(int)out);
6378 wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
6379 void *branch_addr=out;
6381 int target_addr=start+i*4+8;
6383 void *compiled_target_addr=check_addr(target_addr);
6384 emit_extjump_ds((int)branch_addr,target_addr);
6385 if(compiled_target_addr) {
6386 set_jump_target((int)branch_addr,(int)compiled_target_addr);
6387 add_link(target_addr,stub);
6389 else set_jump_target((int)branch_addr,(int)stub);
6393 // Assemble the delay slot for the above
6394 static void pagespan_ds()
6396 assem_debug("initial delay slot:\n");
6397 u_int vaddr=start+1;
6398 u_int page=get_page(vaddr);
6399 u_int vpage=get_vpage(vaddr);
6400 ll_add(jump_dirty+vpage,vaddr,(void *)out);
6402 ll_add(jump_in+page,vaddr,(void *)out);
6403 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
6404 if(regs[0].regmap[HOST_CCREG]!=CCREG)
6405 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty,regs[0].was32);
6406 if(regs[0].regmap[HOST_BTREG]!=BTREG)
6407 emit_writeword(HOST_BTREG,(int)&branch_target);
6408 load_regs(regs[0].regmap_entry,regs[0].regmap,regs[0].was32,rs1[0],rs2[0]);
6409 address_generation(0,®s[0],regs[0].regmap_entry);
6410 if(itype[0]==STORE||itype[0]==STORELR||(opcode[0]&0x3b)==0x39||(opcode[0]&0x3b)==0x3a)
6411 load_regs(regs[0].regmap_entry,regs[0].regmap,regs[0].was32,INVCP,INVCP);
6416 alu_assemble(0,®s[0]);break;
6418 imm16_assemble(0,®s[0]);break;
6420 shift_assemble(0,®s[0]);break;
6422 shiftimm_assemble(0,®s[0]);break;
6424 load_assemble(0,®s[0]);break;
6426 loadlr_assemble(0,®s[0]);break;
6428 store_assemble(0,®s[0]);break;
6430 storelr_assemble(0,®s[0]);break;
6432 cop0_assemble(0,®s[0]);break;
6434 cop1_assemble(0,®s[0]);break;
6436 c1ls_assemble(0,®s[0]);break;
6438 cop2_assemble(0,®s[0]);break;
6440 c2ls_assemble(0,®s[0]);break;
6442 c2op_assemble(0,®s[0]);break;
6444 fconv_assemble(0,®s[0]);break;
6446 float_assemble(0,®s[0]);break;
6448 fcomp_assemble(0,®s[0]);break;
6450 multdiv_assemble(0,®s[0]);break;
6452 mov_assemble(0,®s[0]);break;
6461 printf("Jump in the delay slot. This is probably a bug.\n");
6463 int btaddr=get_reg(regs[0].regmap,BTREG);
6465 btaddr=get_reg(regs[0].regmap,-1);
6466 emit_readword((int)&branch_target,btaddr);
6468 assert(btaddr!=HOST_CCREG);
6469 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
6471 emit_movimm(start+4,HOST_TEMPREG);
6472 emit_cmp(btaddr,HOST_TEMPREG);
6474 emit_cmpimm(btaddr,start+4);
6476 int branch=(int)out;
6478 store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,-1);
6479 emit_jmp(jump_vaddr_reg[btaddr]);
6480 set_jump_target(branch,(int)out);
6481 store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
6482 load_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
6485 // Basic liveness analysis for MIPS registers
6486 void unneeded_registers(int istart,int iend,int r)
6490 uint64_t temp_u,temp_uu;
6495 u=unneeded_reg[iend+1];
6496 uu=unneeded_reg_upper[iend+1];
6499 for (i=iend;i>=istart;i--)
6501 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6502 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
6504 // If subroutine call, flag return address as a possible branch target
6505 if(rt1[i]==31 && i<slen-2) bt[i+2]=1;
6507 if(ba[i]<start || ba[i]>=(start+slen*4))
6509 // Branch out of this block, flush all regs
6513 if(itype[i]==UJUMP&&rt1[i]==31)
6515 uu=u=0x300C00F; // Discard at, v0-v1, t6-t9
6517 if(itype[i]==RJUMP&&rs1[i]==31)
6519 uu=u=0x300C0F3; // Discard at, a0-a3, t6-t9
6521 if(start>0x80000400&&start<0x80800000) {
6522 if(itype[i]==UJUMP&&rt1[i]==31)
6524 //uu=u=0x30300FF0FLL; // Discard at, v0-v1, t0-t9, lo, hi
6525 uu=u=0x300FF0F; // Discard at, v0-v1, t0-t9
6527 if(itype[i]==RJUMP&&rs1[i]==31)
6529 //uu=u=0x30300FFF3LL; // Discard at, a0-a3, t0-t9, lo, hi
6530 uu=u=0x300FFF3; // Discard at, a0-a3, t0-t9
6533 branch_unneeded_reg[i]=u;
6534 branch_unneeded_reg_upper[i]=uu;
6535 // Merge in delay slot
6536 tdep=(~uu>>rt1[i+1])&1;
6537 u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6538 uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6539 u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6540 uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6541 uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6543 // If branch is "likely" (and conditional)
6544 // then we skip the delay slot on the fall-thru path
6547 u&=unneeded_reg[i+2];
6548 uu&=unneeded_reg_upper[i+2];
6559 // Internal branch, flag target
6560 bt[(ba[i]-start)>>2]=1;
6561 if(ba[i]<=start+i*4) {
6563 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6565 // Unconditional branch
6568 // Conditional branch (not taken case)
6569 temp_u=unneeded_reg[i+2];
6570 temp_uu=unneeded_reg_upper[i+2];
6572 // Merge in delay slot
6573 tdep=(~temp_uu>>rt1[i+1])&1;
6574 temp_u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6575 temp_uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6576 temp_u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6577 temp_uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6578 temp_uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6579 temp_u|=1;temp_uu|=1;
6580 // If branch is "likely" (and conditional)
6581 // then we skip the delay slot on the fall-thru path
6584 temp_u&=unneeded_reg[i+2];
6585 temp_uu&=unneeded_reg_upper[i+2];
6593 tdep=(~temp_uu>>rt1[i])&1;
6594 temp_u|=(1LL<<rt1[i])|(1LL<<rt2[i]);
6595 temp_uu|=(1LL<<rt1[i])|(1LL<<rt2[i]);
6596 temp_u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
6597 temp_uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
6598 temp_uu&=~((tdep<<dep1[i])|(tdep<<dep2[i]));
6599 temp_u|=1;temp_uu|=1;
6600 unneeded_reg[i]=temp_u;
6601 unneeded_reg_upper[i]=temp_uu;
6602 // Only go three levels deep. This recursion can take an
6603 // excessive amount of time if there are a lot of nested loops.
6605 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6607 unneeded_reg[(ba[i]-start)>>2]=1;
6608 unneeded_reg_upper[(ba[i]-start)>>2]=1;
6611 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6613 // Unconditional branch
6614 u=unneeded_reg[(ba[i]-start)>>2];
6615 uu=unneeded_reg_upper[(ba[i]-start)>>2];
6616 branch_unneeded_reg[i]=u;
6617 branch_unneeded_reg_upper[i]=uu;
6620 //branch_unneeded_reg[i]=u;
6621 //branch_unneeded_reg_upper[i]=uu;
6622 // Merge in delay slot
6623 tdep=(~uu>>rt1[i+1])&1;
6624 u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6625 uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6626 u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6627 uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6628 uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6631 // Conditional branch
6632 b=unneeded_reg[(ba[i]-start)>>2];
6633 bu=unneeded_reg_upper[(ba[i]-start)>>2];
6634 branch_unneeded_reg[i]=b;
6635 branch_unneeded_reg_upper[i]=bu;
6638 //branch_unneeded_reg[i]=b;
6639 //branch_unneeded_reg_upper[i]=bu;
6640 // Branch delay slot
6641 tdep=(~uu>>rt1[i+1])&1;
6642 b|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6643 bu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6644 b&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6645 bu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6646 bu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6648 // If branch is "likely" then we skip the
6649 // delay slot on the fall-thru path
6654 u&=unneeded_reg[i+2];
6655 uu&=unneeded_reg_upper[i+2];
6666 branch_unneeded_reg[i]&=unneeded_reg[i+2];
6667 branch_unneeded_reg_upper[i]&=unneeded_reg_upper[i+2];
6668 //branch_unneeded_reg[i]=1;
6669 //branch_unneeded_reg_upper[i]=1;
6671 branch_unneeded_reg[i]=1;
6672 branch_unneeded_reg_upper[i]=1;
6678 else if(itype[i]==SYSCALL||itype[i]==HLECALL)
6680 // SYSCALL instruction (software interrupt)
6684 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
6686 // ERET instruction (return from interrupt)
6691 tdep=(~uu>>rt1[i])&1;
6692 // Written registers are unneeded
6697 // Accessed registers are needed
6702 // Source-target dependencies
6703 uu&=~(tdep<<dep1[i]);
6704 uu&=~(tdep<<dep2[i]);
6705 // R0 is always unneeded
6709 unneeded_reg_upper[i]=uu;
6711 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6714 for(r=1;r<=CCREG;r++) {
6715 if((unneeded_reg[i]>>r)&1) {
6716 if(r==HIREG) printf(" HI");
6717 else if(r==LOREG) printf(" LO");
6718 else printf(" r%d",r);
6722 for(r=1;r<=CCREG;r++) {
6723 if(((unneeded_reg_upper[i]&~unneeded_reg[i])>>r)&1) {
6724 if(r==HIREG) printf(" HI");
6725 else if(r==LOREG) printf(" LO");
6726 else printf(" r%d",r);
6732 for (i=iend;i>=istart;i--)
6734 unneeded_reg_upper[i]=branch_unneeded_reg_upper[i]=-1LL;
6739 // Identify registers which are likely to contain 32-bit values
6740 // This is used to predict whether any branches will jump to a
6741 // location with 64-bit values in registers.
6742 static void provisional_32bit()
6746 uint64_t lastbranch=1;
6751 if(itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP) {
6752 if(i>1) is32=lastbranch;
6758 if(itype[i-2]==CJUMP||itype[i-2]==SJUMP||itype[i-2]==FJUMP) {
6760 if(i>2) is32=lastbranch;
6764 if((opcode[i-2]&0x2f)==0x05) // BNE/BNEL
6766 if(rs1[i-2]==0||rs2[i-2]==0)
6769 is32|=1LL<<rs1[i-2];
6772 is32|=1LL<<rs2[i-2];
6777 // If something jumps here with 64-bit values
6778 // then promote those registers to 64 bits
6781 uint64_t temp_is32=is32;
6784 if(ba[j]==start+i*4)
6785 //temp_is32&=branch_regs[j].is32;
6790 if(ba[j]==start+i*4)
6801 if(type==UJUMP||type==RJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
6802 // Branches don't write registers, consider the delay slot instead.
6813 if(opcode[i]==0x27||opcode[i]==0x37|| // LWU/LD
6814 opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
6823 if(op==0x1a||op==0x1b) is32&=~(1LL<<rt); // LDR/LDL
6824 if(op==0x22) is32|=1LL<<rt; // LWL
6827 if (op==0x08||op==0x09|| // ADDI/ADDIU
6828 op==0x0a||op==0x0b|| // SLTI/SLTIU
6834 if(op==0x18||op==0x19) { // DADDI/DADDIU
6837 // is32|=((is32>>s1)&1LL)<<rt;
6839 if(op==0x0d||op==0x0e) { // ORI/XORI
6840 uint64_t sr=((is32>>s1)&1LL);
6856 if(op2>=0x20&&op2<=0x23) { // ADD/ADDU/SUB/SUBU
6859 if(op2==0x2a||op2==0x2b) { // SLT/SLTU
6862 else if(op2>=0x24&&op2<=0x27) { // AND/OR/XOR/NOR
6863 uint64_t sr=((is32>>s1)&(is32>>s2)&1LL);
6867 else if(op2>=0x2c&&op2<=0x2d) { // DADD/DADDU
6872 uint64_t sr=((is32>>s1)&1LL);
6877 uint64_t sr=((is32>>s2)&1LL);
6885 else if(op2>=0x2e&&op2<=0x2f) { // DSUB/DSUBU
6890 uint64_t sr=((is32>>s1)&1LL);
6900 if (op2>=0x1c&&op2<=0x1f) { // DMULT/DMULTU/DDIV/DDIVU
6901 is32&=~((1LL<<HIREG)|(1LL<<LOREG));
6904 is32|=(1LL<<HIREG)|(1LL<<LOREG);
6909 uint64_t sr=((is32>>s1)&1LL);
6915 if(op2>=0x14&&op2<=0x17) is32&=~(1LL<<rt); // DSLLV/DSRLV/DSRAV
6916 else is32|=1LL<<rt; // SLLV/SRLV/SRAV
6920 // DSLL/DSRL/DSRA/DSLL32/DSRL32 but not DSRA32 have 64-bit result
6921 if(op2>=0x38&&op2<0x3f) is32&=~(1LL<<rt);
6924 if(op2==0) is32|=1LL<<rt; // MFC0
6928 if(op2==0) is32|=1LL<<rt; // MFC1
6929 if(op2==1) is32&=~(1LL<<rt); // DMFC1
6930 if(op2==2) is32|=1LL<<rt; // CFC1
6952 if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
6954 if(rt1[i-1]==31) // JAL/JALR
6956 // Subroutine call will return here, don't alloc any registers
6961 // Internal branch will jump here, match registers to caller
6969 // Identify registers which may be assumed to contain 32-bit values
6970 // and where optimizations will rely on this.
6971 // This is used to determine whether backward branches can safely
6972 // jump to a location with 64-bit values in registers.
6973 static void provisional_r32()
6978 for (i=slen-1;i>=0;i--)
6981 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
6983 if(ba[i]<start || ba[i]>=(start+slen*4))
6985 // Branch out of this block, don't need anything
6991 // Need whatever matches the target
6992 // (and doesn't get overwritten by the delay slot instruction)
6994 int t=(ba[i]-start)>>2;
6995 if(ba[i]>start+i*4) {
6997 //if(!(requires_32bit[t]&~regs[i].was32))
6998 // r32|=requires_32bit[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
6999 if(!(pr32[t]&~regs[i].was32))
7000 r32|=pr32[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
7003 if(!(regs[t].was32&~unneeded_reg_upper[t]&~regs[i].was32))
7004 r32|=regs[t].was32&~unneeded_reg_upper[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
7007 // Conditional branch may need registers for following instructions
7008 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
7011 //r32|=requires_32bit[i+2];
7014 // Mark this address as a branch target since it may be called
7015 // upon return from interrupt
7019 // Merge in delay slot
7021 // These are overwritten unless the branch is "likely"
7022 // and the delay slot is nullified if not taken
7023 r32&=~(1LL<<rt1[i+1]);
7024 r32&=~(1LL<<rt2[i+1]);
7026 // Assume these are needed (delay slot)
7029 if((regs[i].was32>>us1[i+1])&1) r32|=1LL<<us1[i+1];
7033 if((regs[i].was32>>us2[i+1])&1) r32|=1LL<<us2[i+1];
7035 if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1))
7037 if((regs[i].was32>>dep1[i+1])&1) r32|=1LL<<dep1[i+1];
7039 if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1))
7041 if((regs[i].was32>>dep2[i+1])&1) r32|=1LL<<dep2[i+1];
7044 else if(itype[i]==SYSCALL||itype[i]==HLECALL)
7046 // SYSCALL instruction (software interrupt)
7049 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
7051 // ERET instruction (return from interrupt)
7055 r32&=~(1LL<<rt1[i]);
7056 r32&=~(1LL<<rt2[i]);
7059 if((regs[i].was32>>us1[i])&1) r32|=1LL<<us1[i];
7063 if((regs[i].was32>>us2[i])&1) r32|=1LL<<us2[i];
7065 if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1))
7067 if((regs[i].was32>>dep1[i])&1) r32|=1LL<<dep1[i];
7069 if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1))
7071 if((regs[i].was32>>dep2[i])&1) r32|=1LL<<dep2[i];
7073 //requires_32bit[i]=r32;
7076 // Dirty registers which are 32-bit, require 32-bit input
7077 // as they will be written as 32-bit values
7078 for(hr=0;hr<HOST_REGS;hr++)
7080 if(regs[i].regmap_entry[hr]>0&®s[i].regmap_entry[hr]<64) {
7081 if((regs[i].was32>>regs[i].regmap_entry[hr])&(regs[i].wasdirty>>hr)&1) {
7082 if(!((unneeded_reg_upper[i]>>regs[i].regmap_entry[hr])&1))
7083 pr32[i]|=1LL<<regs[i].regmap_entry[hr];
7084 //requires_32bit[i]|=1LL<<regs[i].regmap_entry[hr];
7091 // Write back dirty registers as soon as we will no longer modify them,
7092 // so that we don't end up with lots of writes at the branches.
7093 void clean_registers(int istart,int iend,int wr)
7097 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
7098 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
7100 will_dirty_i=will_dirty_next=0;
7101 wont_dirty_i=wont_dirty_next=0;
7103 will_dirty_i=will_dirty_next=will_dirty[iend+1];
7104 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
7106 for (i=iend;i>=istart;i--)
7108 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
7110 if(ba[i]<start || ba[i]>=(start+slen*4))
7112 // Branch out of this block, flush all regs
7113 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
7115 // Unconditional branch
7118 // Merge in delay slot (will dirty)
7119 for(r=0;r<HOST_REGS;r++) {
7120 if(r!=EXCLUDE_REG) {
7121 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7122 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7123 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7124 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7125 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7126 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7127 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7128 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7129 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7130 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7131 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7132 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7133 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7134 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7140 // Conditional branch
7142 wont_dirty_i=wont_dirty_next;
7143 // Merge in delay slot (will dirty)
7144 for(r=0;r<HOST_REGS;r++) {
7145 if(r!=EXCLUDE_REG) {
7147 // Might not dirty if likely branch is not taken
7148 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7149 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7150 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7151 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7152 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7153 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
7154 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7155 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7156 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7157 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7158 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7159 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7160 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7161 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7166 // Merge in delay slot (wont dirty)
7167 for(r=0;r<HOST_REGS;r++) {
7168 if(r!=EXCLUDE_REG) {
7169 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7170 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7171 if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7172 if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7173 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7174 if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7175 if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7176 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7177 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7178 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7182 #ifndef DESTRUCTIVE_WRITEBACK
7183 branch_regs[i].dirty&=wont_dirty_i;
7185 branch_regs[i].dirty|=will_dirty_i;
7191 if(ba[i]<=start+i*4) {
7193 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
7195 // Unconditional branch
7198 // Merge in delay slot (will dirty)
7199 for(r=0;r<HOST_REGS;r++) {
7200 if(r!=EXCLUDE_REG) {
7201 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7202 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7203 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7204 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7205 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7206 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
7207 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7208 if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7209 if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7210 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7211 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7212 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7213 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
7214 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7218 // Conditional branch (not taken case)
7219 temp_will_dirty=will_dirty_next;
7220 temp_wont_dirty=wont_dirty_next;
7221 // Merge in delay slot (will dirty)
7222 for(r=0;r<HOST_REGS;r++) {
7223 if(r!=EXCLUDE_REG) {
7225 // Will not dirty if likely branch is not taken
7226 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7227 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7228 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7229 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7230 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7231 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
7232 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7233 //if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
7234 //if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
7235 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
7236 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
7237 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
7238 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
7239 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
7244 // Merge in delay slot (wont dirty)
7245 for(r=0;r<HOST_REGS;r++) {
7246 if(r!=EXCLUDE_REG) {
7247 if((regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
7248 if((regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
7249 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
7250 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
7251 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
7252 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
7253 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
7254 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
7255 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
7256 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
7259 // Deal with changed mappings
7261 for(r=0;r<HOST_REGS;r++) {
7262 if(r!=EXCLUDE_REG) {
7263 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
7264 temp_will_dirty&=~(1<<r);
7265 temp_wont_dirty&=~(1<<r);
7266 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
7267 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7268 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7270 temp_will_dirty|=1<<r;
7271 temp_wont_dirty|=1<<r;
7278 will_dirty[i]=temp_will_dirty;
7279 wont_dirty[i]=temp_wont_dirty;
7280 clean_registers((ba[i]-start)>>2,i-1,0);
7282 // Limit recursion. It can take an excessive amount
7283 // of time if there are a lot of nested loops.
7284 will_dirty[(ba[i]-start)>>2]=0;
7285 wont_dirty[(ba[i]-start)>>2]=-1;
7290 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
7292 // Unconditional branch
7295 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
7296 for(r=0;r<HOST_REGS;r++) {
7297 if(r!=EXCLUDE_REG) {
7298 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
7299 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
7300 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
7305 // Merge in delay slot
7306 for(r=0;r<HOST_REGS;r++) {
7307 if(r!=EXCLUDE_REG) {
7308 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7309 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7310 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7311 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7312 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7313 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7314 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7315 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7316 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7317 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7318 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7319 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7320 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7321 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7325 // Conditional branch
7326 will_dirty_i=will_dirty_next;
7327 wont_dirty_i=wont_dirty_next;
7328 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
7329 for(r=0;r<HOST_REGS;r++) {
7330 if(r!=EXCLUDE_REG) {
7331 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
7332 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
7333 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
7337 will_dirty_i&=~(1<<r);
7339 // Treat delay slot as part of branch too
7340 /*if(regs[i+1].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
7341 will_dirty[i+1]&=will_dirty[(ba[i]-start)>>2]&(1<<r);
7342 wont_dirty[i+1]|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
7346 will_dirty[i+1]&=~(1<<r);
7351 // Merge in delay slot
7352 for(r=0;r<HOST_REGS;r++) {
7353 if(r!=EXCLUDE_REG) {
7355 // Might not dirty if likely branch is not taken
7356 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7357 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7358 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7359 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7360 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7361 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7362 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7363 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7364 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7365 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
7366 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
7367 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7368 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7369 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7374 // Merge in delay slot
7375 for(r=0;r<HOST_REGS;r++) {
7376 if(r!=EXCLUDE_REG) {
7377 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7378 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7379 if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7380 if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7381 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7382 if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7383 if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7384 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
7385 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
7386 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7390 #ifndef DESTRUCTIVE_WRITEBACK
7391 branch_regs[i].dirty&=wont_dirty_i;
7393 branch_regs[i].dirty|=will_dirty_i;
7398 else if(itype[i]==SYSCALL||itype[i]==HLECALL)
7400 // SYSCALL instruction (software interrupt)
7404 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
7406 // ERET instruction (return from interrupt)
7410 will_dirty_next=will_dirty_i;
7411 wont_dirty_next=wont_dirty_i;
7412 for(r=0;r<HOST_REGS;r++) {
7413 if(r!=EXCLUDE_REG) {
7414 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
7415 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
7416 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
7417 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
7418 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
7419 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
7420 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
7421 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
7423 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=FJUMP)
7425 // Don't store a register immediately after writing it,
7426 // may prevent dual-issue.
7427 if((regs[i].regmap[r]&63)==rt1[i-1]) wont_dirty_i|=1<<r;
7428 if((regs[i].regmap[r]&63)==rt2[i-1]) wont_dirty_i|=1<<r;
7434 will_dirty[i]=will_dirty_i;
7435 wont_dirty[i]=wont_dirty_i;
7436 // Mark registers that won't be dirtied as not dirty
7438 /*printf("wr (%d,%d) %x will:",istart,iend,start+i*4);
7439 for(r=0;r<HOST_REGS;r++) {
7440 if((will_dirty_i>>r)&1) {
7446 //if(i==istart||(itype[i-1]!=RJUMP&&itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=FJUMP)) {
7447 regs[i].dirty|=will_dirty_i;
7448 #ifndef DESTRUCTIVE_WRITEBACK
7449 regs[i].dirty&=wont_dirty_i;
7450 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
7452 if(i<iend-1&&itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
7453 for(r=0;r<HOST_REGS;r++) {
7454 if(r!=EXCLUDE_REG) {
7455 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
7456 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
7457 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);/*assert(!((wont_dirty_i>>r)&1));*/}
7465 for(r=0;r<HOST_REGS;r++) {
7466 if(r!=EXCLUDE_REG) {
7467 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
7468 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
7469 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);/*assert(!((wont_dirty_i>>r)&1));*/}
7477 // Deal with changed mappings
7478 temp_will_dirty=will_dirty_i;
7479 temp_wont_dirty=wont_dirty_i;
7480 for(r=0;r<HOST_REGS;r++) {
7481 if(r!=EXCLUDE_REG) {
7483 if(regs[i].regmap[r]==regmap_pre[i][r]) {
7485 #ifndef DESTRUCTIVE_WRITEBACK
7486 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
7488 regs[i].wasdirty|=will_dirty_i&(1<<r);
7491 else if((nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
7492 // Register moved to a different register
7493 will_dirty_i&=~(1<<r);
7494 wont_dirty_i&=~(1<<r);
7495 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
7496 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
7498 #ifndef DESTRUCTIVE_WRITEBACK
7499 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
7501 regs[i].wasdirty|=will_dirty_i&(1<<r);
7505 will_dirty_i&=~(1<<r);
7506 wont_dirty_i&=~(1<<r);
7507 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
7508 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7509 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
7512 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);/*assert(!((will_dirty>>r)&1));*/
7521 void disassemble_inst(int i)
7523 if (bt[i]) printf("*"); else printf(" ");
7526 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
7528 printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):*ba);break;
7530 printf (" %x: %s r%d,%8x\n",start+i*4,insn[i],rs1[i],start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14));break;
7532 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
7535 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i]);
7537 printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
7540 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],ba[i]);break;
7542 if(opcode[i]==0xf) //LUI
7543 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],rt1[i],imm[i]&0xffff);
7545 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
7549 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
7553 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rs2[i],rs1[i],imm[i]);
7557 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i],rs2[i]);
7560 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rs1[i],rs2[i]);
7563 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
7566 if((opcode2[i]&0x1d)==0x10)
7567 printf (" %x: %s r%d\n",start+i*4,insn[i],rt1[i]);
7568 else if((opcode2[i]&0x1d)==0x11)
7569 printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
7571 printf (" %x: %s\n",start+i*4,insn[i]);
7575 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC0
7576 else if(opcode2[i]==4)
7577 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC0
7578 else printf (" %x: %s\n",start+i*4,insn[i]);
7582 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC1
7583 else if(opcode2[i]>3)
7584 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC1
7585 else printf (" %x: %s\n",start+i*4,insn[i]);
7589 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC2
7590 else if(opcode2[i]>3)
7591 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC2
7592 else printf (" %x: %s\n",start+i*4,insn[i]);
7595 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
7598 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
7601 //printf (" %s %8x\n",insn[i],source[i]);
7602 printf (" %x: %s\n",start+i*4,insn[i]);
7606 void new_dynarec_init()
7608 printf("Init new dynarec\n");
7609 out=(u_char *)BASE_ADDR;
7610 if (mmap (out, 1<<TARGET_SIZE_2,
7611 PROT_READ | PROT_WRITE | PROT_EXEC,
7612 MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS,
7613 -1, 0) <= 0) {printf("mmap() failed\n");}
7615 rdword=&readmem_dword;
7616 fake_pc.f.r.rs=&readmem_dword;
7617 fake_pc.f.r.rt=&readmem_dword;
7618 fake_pc.f.r.rd=&readmem_dword;
7621 for(n=0x80000;n<0x80800;n++)
7623 for(n=0;n<65536;n++)
7624 hash_table[n][0]=hash_table[n][2]=-1;
7625 memset(mini_ht,-1,sizeof(mini_ht));
7626 memset(restore_candidate,0,sizeof(restore_candidate));
7628 expirep=16384; // Expiry pointer, +2 blocks
7629 pending_exception=0;
7632 // Copy this into local area so we don't have to put it in every literal pool
7633 invc_ptr=invalid_code;
7638 for(n=0;n<524288;n++) // 0 .. 0x7FFFFFFF
7640 for(n=524288;n<526336;n++) // 0x80000000 .. 0x807FFFFF
7641 memory_map[n]=((u_int)rdram-0x80000000)>>2;
7642 for(n=526336;n<1048576;n++) // 0x80800000 .. 0xFFFFFFFF
7645 for(n=0;n<0x8000;n++) { // 0 .. 0x7FFFFFFF
7646 writemem[n] = write_nomem_new;
7647 writememb[n] = write_nomemb_new;
7648 writememh[n] = write_nomemh_new;
7650 writememd[n] = write_nomemd_new;
7652 readmem[n] = read_nomem_new;
7653 readmemb[n] = read_nomemb_new;
7654 readmemh[n] = read_nomemh_new;
7656 readmemd[n] = read_nomemd_new;
7659 for(n=0x8000;n<0x8080;n++) { // 0x80000000 .. 0x807FFFFF
7660 writemem[n] = write_rdram_new;
7661 writememb[n] = write_rdramb_new;
7662 writememh[n] = write_rdramh_new;
7664 writememd[n] = write_rdramd_new;
7667 for(n=0xC000;n<0x10000;n++) { // 0xC0000000 .. 0xFFFFFFFF
7668 writemem[n] = write_nomem_new;
7669 writememb[n] = write_nomemb_new;
7670 writememh[n] = write_nomemh_new;
7672 writememd[n] = write_nomemd_new;
7674 readmem[n] = read_nomem_new;
7675 readmemb[n] = read_nomemb_new;
7676 readmemh[n] = read_nomemh_new;
7678 readmemd[n] = read_nomemd_new;
7686 void new_dynarec_cleanup()
7689 if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0) {printf("munmap() failed\n");}
7690 for(n=0;n<4096;n++) ll_clear(jump_in+n);
7691 for(n=0;n<4096;n++) ll_clear(jump_out+n);
7692 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
7694 if (munmap (ROM_COPY, 67108864) < 0) {printf("munmap() failed\n");}
7698 int new_recompile_block(int addr)
7701 if(addr==0x800cd050) {
7703 for(block=0x80000;block<0x80800;block++) invalidate_block(block);
7705 for(n=0;n<=2048;n++) ll_clear(jump_dirty+n);
7708 //if(Count==365117028) tracedebug=1;
7709 assem_debug("NOTCOMPILED: addr = %x -> %x\n", (int)addr, (int)out);
7710 //printf("NOTCOMPILED: addr = %x -> %x\n", (int)addr, (int)out);
7711 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
7713 //printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
7714 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
7715 /*if(Count>=312978186) {
7719 start = (u_int)addr&~3;
7720 //assert(((u_int)addr&1)==0);
7722 if (Config.HLE && start == 0x80001000) {
7723 // XXX: is this enough? Maybe check hleSoftCall?
7724 u_int beginning=(u_int)out;
7725 u_int page=get_page(start);
7726 ll_add(jump_in+page,start,out);
7727 invalid_code[start>>12]=0;
7728 emit_movimm(start,0);
7729 emit_writeword(0,(int)&pcaddr);
7730 emit_jmp((int)new_dyna_leave);
7732 __clear_cache((void *)beginning,out);
7736 else if ((u_int)addr < 0x00200000) {
7737 // used for BIOS calls mostly?
7738 source = (u_int *)((u_int)rdram+start-0);
7739 pagelimit = 0x00200000;
7744 if ((int)addr >= 0xa4000000 && (int)addr < 0xa4001000) {
7745 source = (u_int *)((u_int)SP_DMEM+start-0xa4000000);
7746 pagelimit = 0xa4001000;
7750 if ((int)addr >= 0x80000000 && (int)addr < 0x80800000) {
7751 source = (u_int *)((u_int)rdram+start-0x80000000);
7752 pagelimit = 0x80800000;
7755 else if ((signed int)addr >= (signed int)0xC0000000) {
7756 //printf("addr=%x mm=%x\n",(u_int)addr,(memory_map[start>>12]<<2));
7757 //if(tlb_LUT_r[start>>12])
7758 //source = (u_int *)(((int)rdram)+(tlb_LUT_r[start>>12]&0xFFFFF000)+(((int)addr)&0xFFF)-0x80000000);
7759 if((signed int)memory_map[start>>12]>=0) {
7760 source = (u_int *)((u_int)(start+(memory_map[start>>12]<<2)));
7761 pagelimit=(start+4096)&0xFFFFF000;
7762 int map=memory_map[start>>12];
7765 //printf("start: %x next: %x\n",map,memory_map[pagelimit>>12]);
7766 if((map&0xBFFFFFFF)==(memory_map[pagelimit>>12]&0xBFFFFFFF)) pagelimit+=4096;
7768 assem_debug("pagelimit=%x\n",pagelimit);
7769 assem_debug("mapping=%x (%x)\n",memory_map[start>>12],(memory_map[start>>12]<<2)+start);
7772 assem_debug("Compile at unmapped memory address: %x \n", (int)addr);
7773 //assem_debug("start: %x next: %x\n",memory_map[start>>12],memory_map[(start+4096)>>12]);
7774 return 1; // Caller will invoke exception handler
7776 //printf("source= %x\n",(int)source);
7780 printf("Compile at bogus memory address: %x \n", (int)addr);
7784 /* Pass 1: disassemble */
7785 /* Pass 2: register dependencies, branch targets */
7786 /* Pass 3: register allocation */
7787 /* Pass 4: branch dependencies */
7788 /* Pass 5: pre-alloc */
7789 /* Pass 6: optimize clean/dirty state */
7790 /* Pass 7: flag 32-bit registers */
7791 /* Pass 8: assembly */
7792 /* Pass 9: linker */
7793 /* Pass 10: garbage collection / free memory */
7797 unsigned int type,op,op2;
7799 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
7801 /* Pass 1 disassembly */
7803 for(i=0;!done;i++) {
7804 bt[i]=0;likely[i]=0;op2=0;
7805 opcode[i]=op=source[i]>>26;
7808 case 0x00: strcpy(insn[i],"special"); type=NI;
7812 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
7813 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
7814 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
7815 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
7816 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
7817 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
7818 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
7819 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
7820 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
7821 case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
7822 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
7823 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
7824 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
7825 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
7826 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
7827 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
7828 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
7829 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
7830 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
7831 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
7832 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
7833 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
7834 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
7835 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
7836 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
7837 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
7838 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
7839 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
7840 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
7841 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
7842 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
7843 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
7844 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
7845 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
7846 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
7847 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
7848 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
7849 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
7850 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
7851 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
7852 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
7853 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
7854 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
7855 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
7856 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
7857 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
7858 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
7859 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
7860 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
7861 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
7862 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
7863 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
7866 case 0x01: strcpy(insn[i],"regimm"); type=NI;
7867 op2=(source[i]>>16)&0x1f;
7870 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
7871 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
7872 case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
7873 case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
7874 case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
7875 case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
7876 case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
7877 case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
7878 case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
7879 case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
7880 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
7881 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
7882 case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
7883 case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
7886 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
7887 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
7888 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
7889 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
7890 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
7891 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
7892 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
7893 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
7894 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
7895 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
7896 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
7897 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
7898 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
7899 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
7900 case 0x10: strcpy(insn[i],"cop0"); type=NI;
7901 op2=(source[i]>>21)&0x1f;
7904 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
7905 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
7906 case 0x10: strcpy(insn[i],"tlb"); type=NI;
7907 switch(source[i]&0x3f)
7909 case 0x01: strcpy(insn[i],"TLBR"); type=COP0; break;
7910 case 0x02: strcpy(insn[i],"TLBWI"); type=COP0; break;
7911 case 0x06: strcpy(insn[i],"TLBWR"); type=COP0; break;
7912 case 0x08: strcpy(insn[i],"TLBP"); type=COP0; break;
7913 case 0x18: strcpy(insn[i],"ERET"); type=COP0; break;
7917 case 0x11: strcpy(insn[i],"cop1"); type=NI;
7918 op2=(source[i]>>21)&0x1f;
7921 case 0x00: strcpy(insn[i],"MFC1"); type=COP1; break;
7922 case 0x01: strcpy(insn[i],"DMFC1"); type=COP1; break;
7923 case 0x02: strcpy(insn[i],"CFC1"); type=COP1; break;
7924 case 0x04: strcpy(insn[i],"MTC1"); type=COP1; break;
7925 case 0x05: strcpy(insn[i],"DMTC1"); type=COP1; break;
7926 case 0x06: strcpy(insn[i],"CTC1"); type=COP1; break;
7927 case 0x08: strcpy(insn[i],"BC1"); type=FJUMP;
7928 switch((source[i]>>16)&0x3)
7930 case 0x00: strcpy(insn[i],"BC1F"); break;
7931 case 0x01: strcpy(insn[i],"BC1T"); break;
7932 case 0x02: strcpy(insn[i],"BC1FL"); break;
7933 case 0x03: strcpy(insn[i],"BC1TL"); break;
7936 case 0x10: strcpy(insn[i],"C1.S"); type=NI;
7937 switch(source[i]&0x3f)
7939 case 0x00: strcpy(insn[i],"ADD.S"); type=FLOAT; break;
7940 case 0x01: strcpy(insn[i],"SUB.S"); type=FLOAT; break;
7941 case 0x02: strcpy(insn[i],"MUL.S"); type=FLOAT; break;
7942 case 0x03: strcpy(insn[i],"DIV.S"); type=FLOAT; break;
7943 case 0x04: strcpy(insn[i],"SQRT.S"); type=FLOAT; break;
7944 case 0x05: strcpy(insn[i],"ABS.S"); type=FLOAT; break;
7945 case 0x06: strcpy(insn[i],"MOV.S"); type=FLOAT; break;
7946 case 0x07: strcpy(insn[i],"NEG.S"); type=FLOAT; break;
7947 case 0x08: strcpy(insn[i],"ROUND.L.S"); type=FCONV; break;
7948 case 0x09: strcpy(insn[i],"TRUNC.L.S"); type=FCONV; break;
7949 case 0x0A: strcpy(insn[i],"CEIL.L.S"); type=FCONV; break;
7950 case 0x0B: strcpy(insn[i],"FLOOR.L.S"); type=FCONV; break;
7951 case 0x0C: strcpy(insn[i],"ROUND.W.S"); type=FCONV; break;
7952 case 0x0D: strcpy(insn[i],"TRUNC.W.S"); type=FCONV; break;
7953 case 0x0E: strcpy(insn[i],"CEIL.W.S"); type=FCONV; break;
7954 case 0x0F: strcpy(insn[i],"FLOOR.W.S"); type=FCONV; break;
7955 case 0x21: strcpy(insn[i],"CVT.D.S"); type=FCONV; break;
7956 case 0x24: strcpy(insn[i],"CVT.W.S"); type=FCONV; break;
7957 case 0x25: strcpy(insn[i],"CVT.L.S"); type=FCONV; break;
7958 case 0x30: strcpy(insn[i],"C.F.S"); type=FCOMP; break;
7959 case 0x31: strcpy(insn[i],"C.UN.S"); type=FCOMP; break;
7960 case 0x32: strcpy(insn[i],"C.EQ.S"); type=FCOMP; break;
7961 case 0x33: strcpy(insn[i],"C.UEQ.S"); type=FCOMP; break;
7962 case 0x34: strcpy(insn[i],"C.OLT.S"); type=FCOMP; break;
7963 case 0x35: strcpy(insn[i],"C.ULT.S"); type=FCOMP; break;
7964 case 0x36: strcpy(insn[i],"C.OLE.S"); type=FCOMP; break;
7965 case 0x37: strcpy(insn[i],"C.ULE.S"); type=FCOMP; break;
7966 case 0x38: strcpy(insn[i],"C.SF.S"); type=FCOMP; break;
7967 case 0x39: strcpy(insn[i],"C.NGLE.S"); type=FCOMP; break;
7968 case 0x3A: strcpy(insn[i],"C.SEQ.S"); type=FCOMP; break;
7969 case 0x3B: strcpy(insn[i],"C.NGL.S"); type=FCOMP; break;
7970 case 0x3C: strcpy(insn[i],"C.LT.S"); type=FCOMP; break;
7971 case 0x3D: strcpy(insn[i],"C.NGE.S"); type=FCOMP; break;
7972 case 0x3E: strcpy(insn[i],"C.LE.S"); type=FCOMP; break;
7973 case 0x3F: strcpy(insn[i],"C.NGT.S"); type=FCOMP; break;
7976 case 0x11: strcpy(insn[i],"C1.D"); type=NI;
7977 switch(source[i]&0x3f)
7979 case 0x00: strcpy(insn[i],"ADD.D"); type=FLOAT; break;
7980 case 0x01: strcpy(insn[i],"SUB.D"); type=FLOAT; break;
7981 case 0x02: strcpy(insn[i],"MUL.D"); type=FLOAT; break;
7982 case 0x03: strcpy(insn[i],"DIV.D"); type=FLOAT; break;
7983 case 0x04: strcpy(insn[i],"SQRT.D"); type=FLOAT; break;
7984 case 0x05: strcpy(insn[i],"ABS.D"); type=FLOAT; break;
7985 case 0x06: strcpy(insn[i],"MOV.D"); type=FLOAT; break;
7986 case 0x07: strcpy(insn[i],"NEG.D"); type=FLOAT; break;
7987 case 0x08: strcpy(insn[i],"ROUND.L.D"); type=FCONV; break;
7988 case 0x09: strcpy(insn[i],"TRUNC.L.D"); type=FCONV; break;
7989 case 0x0A: strcpy(insn[i],"CEIL.L.D"); type=FCONV; break;
7990 case 0x0B: strcpy(insn[i],"FLOOR.L.D"); type=FCONV; break;
7991 case 0x0C: strcpy(insn[i],"ROUND.W.D"); type=FCONV; break;
7992 case 0x0D: strcpy(insn[i],"TRUNC.W.D"); type=FCONV; break;
7993 case 0x0E: strcpy(insn[i],"CEIL.W.D"); type=FCONV; break;
7994 case 0x0F: strcpy(insn[i],"FLOOR.W.D"); type=FCONV; break;
7995 case 0x20: strcpy(insn[i],"CVT.S.D"); type=FCONV; break;
7996 case 0x24: strcpy(insn[i],"CVT.W.D"); type=FCONV; break;
7997 case 0x25: strcpy(insn[i],"CVT.L.D"); type=FCONV; break;
7998 case 0x30: strcpy(insn[i],"C.F.D"); type=FCOMP; break;
7999 case 0x31: strcpy(insn[i],"C.UN.D"); type=FCOMP; break;
8000 case 0x32: strcpy(insn[i],"C.EQ.D"); type=FCOMP; break;
8001 case 0x33: strcpy(insn[i],"C.UEQ.D"); type=FCOMP; break;
8002 case 0x34: strcpy(insn[i],"C.OLT.D"); type=FCOMP; break;
8003 case 0x35: strcpy(insn[i],"C.ULT.D"); type=FCOMP; break;
8004 case 0x36: strcpy(insn[i],"C.OLE.D"); type=FCOMP; break;
8005 case 0x37: strcpy(insn[i],"C.ULE.D"); type=FCOMP; break;
8006 case 0x38: strcpy(insn[i],"C.SF.D"); type=FCOMP; break;
8007 case 0x39: strcpy(insn[i],"C.NGLE.D"); type=FCOMP; break;
8008 case 0x3A: strcpy(insn[i],"C.SEQ.D"); type=FCOMP; break;
8009 case 0x3B: strcpy(insn[i],"C.NGL.D"); type=FCOMP; break;
8010 case 0x3C: strcpy(insn[i],"C.LT.D"); type=FCOMP; break;
8011 case 0x3D: strcpy(insn[i],"C.NGE.D"); type=FCOMP; break;
8012 case 0x3E: strcpy(insn[i],"C.LE.D"); type=FCOMP; break;
8013 case 0x3F: strcpy(insn[i],"C.NGT.D"); type=FCOMP; break;
8016 case 0x14: strcpy(insn[i],"C1.W"); type=NI;
8017 switch(source[i]&0x3f)
8019 case 0x20: strcpy(insn[i],"CVT.S.W"); type=FCONV; break;
8020 case 0x21: strcpy(insn[i],"CVT.D.W"); type=FCONV; break;
8023 case 0x15: strcpy(insn[i],"C1.L"); type=NI;
8024 switch(source[i]&0x3f)
8026 case 0x20: strcpy(insn[i],"CVT.S.L"); type=FCONV; break;
8027 case 0x21: strcpy(insn[i],"CVT.D.L"); type=FCONV; break;
8032 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
8033 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
8034 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
8035 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
8036 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
8037 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
8038 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
8039 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
8040 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
8041 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
8042 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
8043 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
8044 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
8045 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
8046 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
8047 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
8048 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
8049 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
8050 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
8051 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
8052 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
8053 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
8054 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
8055 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
8056 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
8057 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
8058 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
8059 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
8060 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
8061 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
8062 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
8063 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
8064 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
8065 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
8067 case 0x12: strcpy(insn[i],"COP2"); type=NI;
8068 op2=(source[i]>>21)&0x1f;
8071 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
8072 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
8073 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
8074 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
8076 if (gte_handlers[source[i]&0x3f]!=NULL) {
8077 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
8083 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
8084 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
8085 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
8087 default: strcpy(insn[i],"???"); type=NI;
8088 printf("NI %08x @%08x\n", source[i], addr + i*4);
8093 /* Get registers/immediates */
8101 rs1[i]=(source[i]>>21)&0x1f;
8103 rt1[i]=(source[i]>>16)&0x1f;
8105 imm[i]=(short)source[i];
8109 rs1[i]=(source[i]>>21)&0x1f;
8110 rs2[i]=(source[i]>>16)&0x1f;
8113 imm[i]=(short)source[i];
8114 if(op==0x2c||op==0x2d||op==0x3f) us1[i]=rs2[i]; // 64-bit SDL/SDR/SD
8117 // LWL/LWR only load part of the register,
8118 // therefore the target register must be treated as a source too
8119 rs1[i]=(source[i]>>21)&0x1f;
8120 rs2[i]=(source[i]>>16)&0x1f;
8121 rt1[i]=(source[i]>>16)&0x1f;
8123 imm[i]=(short)source[i];
8124 if(op==0x1a||op==0x1b) us1[i]=rs2[i]; // LDR/LDL
8125 if(op==0x26) dep1[i]=rt1[i]; // LWR
8128 if (op==0x0f) rs1[i]=0; // LUI instruction has no source register
8129 else rs1[i]=(source[i]>>21)&0x1f;
8131 rt1[i]=(source[i]>>16)&0x1f;
8133 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
8134 imm[i]=(unsigned short)source[i];
8136 imm[i]=(short)source[i];
8138 if(op==0x18||op==0x19) us1[i]=rs1[i]; // DADDI/DADDIU
8139 if(op==0x0a||op==0x0b) us1[i]=rs1[i]; // SLTI/SLTIU
8140 if(op==0x0d||op==0x0e) dep1[i]=rs1[i]; // ORI/XORI
8147 // The JAL instruction writes to r31.
8154 rs1[i]=(source[i]>>21)&0x1f;
8158 // The JALR instruction writes to rd.
8160 rt1[i]=(source[i]>>11)&0x1f;
8165 rs1[i]=(source[i]>>21)&0x1f;
8166 rs2[i]=(source[i]>>16)&0x1f;
8169 if(op&2) { // BGTZ/BLEZ
8177 rs1[i]=(source[i]>>21)&0x1f;
8182 if(op2&0x10) { // BxxAL
8184 // NOTE: If the branch is not taken, r31 is still overwritten
8186 likely[i]=(op2&2)>>1;
8193 likely[i]=((source[i])>>17)&1;
8196 rs1[i]=(source[i]>>21)&0x1f; // source
8197 rs2[i]=(source[i]>>16)&0x1f; // subtract amount
8198 rt1[i]=(source[i]>>11)&0x1f; // destination
8200 if(op2==0x2a||op2==0x2b) { // SLT/SLTU
8201 us1[i]=rs1[i];us2[i]=rs2[i];
8203 else if(op2>=0x24&&op2<=0x27) { // AND/OR/XOR/NOR
8204 dep1[i]=rs1[i];dep2[i]=rs2[i];
8206 else if(op2>=0x2c&&op2<=0x2f) { // DADD/DSUB
8207 dep1[i]=rs1[i];dep2[i]=rs2[i];
8211 rs1[i]=(source[i]>>21)&0x1f; // source
8212 rs2[i]=(source[i]>>16)&0x1f; // divisor
8215 if (op2>=0x1c&&op2<=0x1f) { // DMULT/DMULTU/DDIV/DDIVU
8216 us1[i]=rs1[i];us2[i]=rs2[i];
8224 if(op2==0x10) rs1[i]=HIREG; // MFHI
8225 if(op2==0x11) rt1[i]=HIREG; // MTHI
8226 if(op2==0x12) rs1[i]=LOREG; // MFLO
8227 if(op2==0x13) rt1[i]=LOREG; // MTLO
8228 if((op2&0x1d)==0x10) rt1[i]=(source[i]>>11)&0x1f; // MFxx
8229 if((op2&0x1d)==0x11) rs1[i]=(source[i]>>21)&0x1f; // MTxx
8233 rs1[i]=(source[i]>>16)&0x1f; // target of shift
8234 rs2[i]=(source[i]>>21)&0x1f; // shift amount
8235 rt1[i]=(source[i]>>11)&0x1f; // destination
8237 // DSLLV/DSRLV/DSRAV are 64-bit
8238 if(op2>=0x14&&op2<=0x17) us1[i]=rs1[i];
8241 rs1[i]=(source[i]>>16)&0x1f;
8243 rt1[i]=(source[i]>>11)&0x1f;
8245 imm[i]=(source[i]>>6)&0x1f;
8246 // DSxx32 instructions
8247 if(op2>=0x3c) imm[i]|=0x20;
8248 // DSLL/DSRL/DSRA/DSRA32/DSRL32 but not DSLL32 require 64-bit source
8249 if(op2>=0x38&&op2!=0x3c) us1[i]=rs1[i];
8256 if(op2==0) rt1[i]=(source[i]>>16)&0x1F; // MFC0
8257 if(op2==4) rs1[i]=(source[i]>>16)&0x1F; // MTC0
8258 if(op2==4&&((source[i]>>11)&0x1f)==12) rt2[i]=CSREG; // Status
8259 if(op2==16) if((source[i]&0x3f)==0x18) rs2[i]=CCREG; // ERET
8267 if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
8268 if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
8269 if(op2==5) us1[i]=rs1[i]; // DMTC1
8273 rs1[i]=(source[i]>>21)&0x1F;
8277 imm[i]=(short)source[i];
8280 rs1[i]=(source[i]>>21)&0x1F;
8284 imm[i]=(short)source[i];
8312 /* Calculate branch target addresses */
8314 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
8315 else if(type==CJUMP&&rs1[i]==rs2[i]&&(op&1))
8316 ba[i]=start+i*4+8; // Ignore never taken branch
8317 else if(type==SJUMP&&rs1[i]==0&&!(op2&1))
8318 ba[i]=start+i*4+8; // Ignore never taken branch
8319 else if(type==CJUMP||type==SJUMP||type==FJUMP)
8320 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
8322 /* Is this the end of the block? */
8323 if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
8324 if(rt1[i-1]==0) { // Continue past subroutine call (JAL)
8326 // Does the block continue due to a branch?
8329 if(ba[j]==start+i*4+4) done=j=0;
8330 if(ba[j]==start+i*4+8) done=j=0;
8334 if(stop_after_jal) done=1;
8336 if((source[i+1]&0xfc00003f)==0x0d) done=1;
8338 // Don't recompile stuff that's already compiled
8339 if(check_addr(start+i*4+4)) done=1;
8340 // Don't get too close to the limit
8341 if(i>MAXBLOCK/2) done=1;
8343 if(i>0&&itype[i-1]==SYSCALL&&stop_after_jal) done=1;
8344 if(itype[i-1]==HLECALL) done=1;
8345 assert(i<MAXBLOCK-1);
8346 if(start+i*4==pagelimit-4) done=1;
8347 assert(start+i*4<pagelimit);
8348 if (i==MAXBLOCK-1) done=1;
8349 // Stop if we're compiling junk
8350 if(itype[i]==NI&&opcode[i]==0x11) {
8351 done=stop_after_jal=1;
8352 printf("Disabled speculative precompilation\n");
8356 if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==RJUMP||itype[i-1]==FJUMP) {
8357 if(start+i*4==pagelimit) {
8363 /* Pass 2 - Register dependencies and branch targets */
8365 unneeded_registers(0,slen-1,0);
8367 /* Pass 3 - Register allocation */
8369 struct regstat current; // Current register allocations/status
8372 current.u=unneeded_reg[0];
8373 current.uu=unneeded_reg_upper[0];
8374 clear_all_regs(current.regmap);
8375 alloc_reg(¤t,0,CCREG);
8376 dirty_reg(¤t,CCREG);
8383 provisional_32bit();
8386 // First instruction is delay slot
8391 unneeded_reg_upper[0]=1;
8392 current.regmap[HOST_BTREG]=BTREG;
8400 for(hr=0;hr<HOST_REGS;hr++)
8402 // Is this really necessary?
8403 if(current.regmap[hr]==0) current.regmap[hr]=-1;
8409 if((opcode[i-2]&0x2f)==0x05) // BNE/BNEL
8411 if(rs1[i-2]==0||rs2[i-2]==0)
8414 current.is32|=1LL<<rs1[i-2];
8415 int hr=get_reg(current.regmap,rs1[i-2]|64);
8416 if(hr>=0) current.regmap[hr]=-1;
8419 current.is32|=1LL<<rs2[i-2];
8420 int hr=get_reg(current.regmap,rs2[i-2]|64);
8421 if(hr>=0) current.regmap[hr]=-1;
8426 // If something jumps here with 64-bit values
8427 // then promote those registers to 64 bits
8430 uint64_t temp_is32=current.is32;
8433 if(ba[j]==start+i*4)
8434 temp_is32&=branch_regs[j].is32;
8438 if(ba[j]==start+i*4)
8442 if(temp_is32!=current.is32) {
8443 //printf("dumping 32-bit regs (%x)\n",start+i*4);
8444 #ifdef DESTRUCTIVE_WRITEBACK
8445 for(hr=0;hr<HOST_REGS;hr++)
8447 int r=current.regmap[hr];
8450 if((current.dirty>>hr)&((current.is32&~temp_is32)>>r)&1) {
8452 //printf("restore %d\n",r);
8457 current.is32=temp_is32;
8461 memset(p32, 0xff, sizeof(p32));
8465 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
8466 regs[i].wasconst=current.isconst;
8467 regs[i].was32=current.is32;
8468 regs[i].wasdirty=current.dirty;
8469 #ifdef DESTRUCTIVE_WRITEBACK
8470 // To change a dirty register from 32 to 64 bits, we must write
8471 // it out during the previous cycle (for branches, 2 cycles)
8472 if(i<slen-1&&bt[i+1]&&itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP)
8474 uint64_t temp_is32=current.is32;
8477 if(ba[j]==start+i*4+4)
8478 temp_is32&=branch_regs[j].is32;
8482 if(ba[j]==start+i*4+4)
8486 if(temp_is32!=current.is32) {
8487 //printf("pre-dumping 32-bit regs (%x)\n",start+i*4);
8488 for(hr=0;hr<HOST_REGS;hr++)
8490 int r=current.regmap[hr];
8493 if((current.dirty>>hr)&((current.is32&~temp_is32)>>(r&63))&1) {
8494 if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP)
8496 if(rs1[i]!=(r&63)&&rs2[i]!=(r&63))
8498 //printf("dump %d/r%d\n",hr,r);
8499 current.regmap[hr]=-1;
8500 if(get_reg(current.regmap,r|64)>=0)
8501 current.regmap[get_reg(current.regmap,r|64)]=-1;
8509 else if(i<slen-2&&bt[i+2]&&(source[i-1]>>16)!=0x1000&&(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP))
8511 uint64_t temp_is32=current.is32;
8514 if(ba[j]==start+i*4+8)
8515 temp_is32&=branch_regs[j].is32;
8519 if(ba[j]==start+i*4+8)
8523 if(temp_is32!=current.is32) {
8524 //printf("pre-dumping 32-bit regs (%x)\n",start+i*4);
8525 for(hr=0;hr<HOST_REGS;hr++)
8527 int r=current.regmap[hr];
8530 if((current.dirty>>hr)&((current.is32&~temp_is32)>>(r&63))&1) {
8531 if(rs1[i]!=(r&63)&&rs2[i]!=(r&63)&&rs1[i+1]!=(r&63)&&rs2[i+1]!=(r&63))
8533 //printf("dump %d/r%d\n",hr,r);
8534 current.regmap[hr]=-1;
8535 if(get_reg(current.regmap,r|64)>=0)
8536 current.regmap[get_reg(current.regmap,r|64)]=-1;
8544 if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP) {
8546 current.u=unneeded_reg[i+1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
8547 current.uu=unneeded_reg_upper[i+1]&~((1LL<<us1[i])|(1LL<<us2[i]));
8548 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8557 current.u=branch_unneeded_reg[i]&~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
8558 current.uu=branch_unneeded_reg_upper[i]&~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
8559 if((~current.uu>>rt1[i+1])&1) current.uu&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
8560 current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
8561 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8564 } else { printf("oops, branch at end of block with no delay slot\n");exit(1); }
8568 ds=0; // Skip delay slot, already allocated as part of branch
8569 // ...but we need to alloc it in case something jumps here
8571 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
8572 current.uu=branch_unneeded_reg_upper[i-1]&unneeded_reg_upper[i+1];
8574 current.u=branch_unneeded_reg[i-1];
8575 current.uu=branch_unneeded_reg_upper[i-1];
8577 current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
8578 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8579 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8582 struct regstat temp;
8583 memcpy(&temp,¤t,sizeof(current));
8584 temp.wasdirty=temp.dirty;
8585 temp.was32=temp.is32;
8586 // TODO: Take into account unconditional branches, as below
8587 delayslot_alloc(&temp,i);
8588 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
8589 regs[i].wasdirty=temp.wasdirty;
8590 regs[i].was32=temp.was32;
8591 regs[i].dirty=temp.dirty;
8592 regs[i].is32=temp.is32;
8596 // Create entry (branch target) regmap
8597 for(hr=0;hr<HOST_REGS;hr++)
8599 int r=temp.regmap[hr];
8601 if(r!=regmap_pre[i][hr]) {
8602 regs[i].regmap_entry[hr]=-1;
8607 if((current.u>>r)&1) {
8608 regs[i].regmap_entry[hr]=-1;
8609 regs[i].regmap[hr]=-1;
8610 //Don't clear regs in the delay slot as the branch might need them
8611 //current.regmap[hr]=-1;
8613 regs[i].regmap_entry[hr]=r;
8616 if((current.uu>>(r&63))&1) {
8617 regs[i].regmap_entry[hr]=-1;
8618 regs[i].regmap[hr]=-1;
8619 //Don't clear regs in the delay slot as the branch might need them
8620 //current.regmap[hr]=-1;
8622 regs[i].regmap_entry[hr]=r;
8626 // First instruction expects CCREG to be allocated
8627 if(i==0&&hr==HOST_CCREG)
8628 regs[i].regmap_entry[hr]=CCREG;
8630 regs[i].regmap_entry[hr]=-1;
8634 else { // Not delay slot
8637 //current.isconst=0; // DEBUG
8638 //current.wasconst=0; // DEBUG
8639 //regs[i].wasconst=0; // DEBUG
8640 clear_const(¤t,rt1[i]);
8641 alloc_cc(¤t,i);
8642 dirty_reg(¤t,CCREG);
8644 alloc_reg(¤t,i,31);
8645 dirty_reg(¤t,31);
8646 assert(rs1[i+1]!=31&&rs2[i+1]!=31);
8648 alloc_reg(¤t,i,PTEMP);
8650 //current.is32|=1LL<<rt1[i];
8652 delayslot_alloc(¤t,i+1);
8653 //current.isconst=0; // DEBUG
8655 //printf("i=%d, isconst=%x\n",i,current.isconst);
8658 //current.isconst=0;
8659 //current.wasconst=0;
8660 //regs[i].wasconst=0;
8661 clear_const(¤t,rs1[i]);
8662 clear_const(¤t,rt1[i]);
8663 alloc_cc(¤t,i);
8664 dirty_reg(¤t,CCREG);
8665 if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
8666 alloc_reg(¤t,i,rs1[i]);
8668 alloc_reg(¤t,i,rt1[i]);
8669 dirty_reg(¤t,rt1[i]);
8670 assert(rs1[i+1]!=31&&rs2[i+1]!=31);
8672 alloc_reg(¤t,i,PTEMP);
8676 if(rs1[i]==31) { // JALR
8677 alloc_reg(¤t,i,RHASH);
8678 #ifndef HOST_IMM_ADDR32
8679 alloc_reg(¤t,i,RHTBL);
8683 delayslot_alloc(¤t,i+1);
8685 // The delay slot overwrites our source register,
8686 // allocate a temporary register to hold the old value.
8690 delayslot_alloc(¤t,i+1);
8692 alloc_reg(¤t,i,RTEMP);
8694 //current.isconst=0; // DEBUG
8698 //current.isconst=0;
8699 //current.wasconst=0;
8700 //regs[i].wasconst=0;
8701 clear_const(¤t,rs1[i]);
8702 clear_const(¤t,rs2[i]);
8703 if((opcode[i]&0x3E)==4) // BEQ/BNE
8705 alloc_cc(¤t,i);
8706 dirty_reg(¤t,CCREG);
8707 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8708 if(rs2[i]) alloc_reg(¤t,i,rs2[i]);
8709 if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8711 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8712 if(rs2[i]) alloc_reg64(¤t,i,rs2[i]);
8714 if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
8715 (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1]))) {
8716 // The delay slot overwrites one of our conditions.
8717 // Allocate the branch condition registers instead.
8718 // Note that such a sequence of instructions could
8719 // be considered a bug since the branch can not be
8720 // re-executed if an exception occurs.
8724 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8725 if(rs2[i]) alloc_reg(¤t,i,rs2[i]);
8726 if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8728 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8729 if(rs2[i]) alloc_reg64(¤t,i,rs2[i]);
8732 else delayslot_alloc(¤t,i+1);
8735 if((opcode[i]&0x3E)==6) // BLEZ/BGTZ
8737 alloc_cc(¤t,i);
8738 dirty_reg(¤t,CCREG);
8739 alloc_reg(¤t,i,rs1[i]);
8740 if(!(current.is32>>rs1[i]&1))
8742 alloc_reg64(¤t,i,rs1[i]);
8744 if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
8745 // The delay slot overwrites one of our conditions.
8746 // Allocate the branch condition registers instead.
8747 // Note that such a sequence of instructions could
8748 // be considered a bug since the branch can not be
8749 // re-executed if an exception occurs.
8753 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8754 if(!((current.is32>>rs1[i])&1))
8756 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8759 else delayslot_alloc(¤t,i+1);
8762 // Don't alloc the delay slot yet because we might not execute it
8763 if((opcode[i]&0x3E)==0x14) // BEQL/BNEL
8768 alloc_cc(¤t,i);
8769 dirty_reg(¤t,CCREG);
8770 alloc_reg(¤t,i,rs1[i]);
8771 alloc_reg(¤t,i,rs2[i]);
8772 if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8774 alloc_reg64(¤t,i,rs1[i]);
8775 alloc_reg64(¤t,i,rs2[i]);
8779 if((opcode[i]&0x3E)==0x16) // BLEZL/BGTZL
8784 alloc_cc(¤t,i);
8785 dirty_reg(¤t,CCREG);
8786 alloc_reg(¤t,i,rs1[i]);
8787 if(!(current.is32>>rs1[i]&1))
8789 alloc_reg64(¤t,i,rs1[i]);
8793 //current.isconst=0;
8796 //current.isconst=0;
8797 //current.wasconst=0;
8798 //regs[i].wasconst=0;
8799 clear_const(¤t,rs1[i]);
8800 clear_const(¤t,rt1[i]);
8801 //if((opcode2[i]&0x1E)==0x0) // BLTZ/BGEZ
8802 if((opcode2[i]&0x0E)==0x0) // BLTZ/BGEZ
8804 alloc_cc(¤t,i);
8805 dirty_reg(¤t,CCREG);
8806 alloc_reg(¤t,i,rs1[i]);
8807 if(!(current.is32>>rs1[i]&1))
8809 alloc_reg64(¤t,i,rs1[i]);
8811 if (rt1[i]==31) { // BLTZAL/BGEZAL
8812 alloc_reg(¤t,i,31);
8813 dirty_reg(¤t,31);
8814 assert(rs1[i+1]!=31&&rs2[i+1]!=31);
8815 //#ifdef REG_PREFETCH
8816 //alloc_reg(¤t,i,PTEMP);
8818 //current.is32|=1LL<<rt1[i];
8820 if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
8821 // The delay slot overwrites the branch condition.
8822 // Allocate the branch condition registers instead.
8823 // Note that such a sequence of instructions could
8824 // be considered a bug since the branch can not be
8825 // re-executed if an exception occurs.
8829 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8830 if(!((current.is32>>rs1[i])&1))
8832 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8835 else delayslot_alloc(¤t,i+1);
8838 // Don't alloc the delay slot yet because we might not execute it
8839 if((opcode2[i]&0x1E)==0x2) // BLTZL/BGEZL
8844 alloc_cc(¤t,i);
8845 dirty_reg(¤t,CCREG);
8846 alloc_reg(¤t,i,rs1[i]);
8847 if(!(current.is32>>rs1[i]&1))
8849 alloc_reg64(¤t,i,rs1[i]);
8853 //current.isconst=0;
8859 if(likely[i]==0) // BC1F/BC1T
8861 // TODO: Theoretically we can run out of registers here on x86.
8862 // The delay slot can allocate up to six, and we need to check
8863 // CSREG before executing the delay slot. Possibly we can drop
8864 // the cycle count and then reload it after checking that the
8865 // FPU is in a usable state, or don't do out-of-order execution.
8866 alloc_cc(¤t,i);
8867 dirty_reg(¤t,CCREG);
8868 alloc_reg(¤t,i,FSREG);
8869 alloc_reg(¤t,i,CSREG);
8870 if(itype[i+1]==FCOMP) {
8871 // The delay slot overwrites the branch condition.
8872 // Allocate the branch condition registers instead.
8873 // Note that such a sequence of instructions could
8874 // be considered a bug since the branch can not be
8875 // re-executed if an exception occurs.
8876 alloc_cc(¤t,i);
8877 dirty_reg(¤t,CCREG);
8878 alloc_reg(¤t,i,CSREG);
8879 alloc_reg(¤t,i,FSREG);
8882 delayslot_alloc(¤t,i+1);
8883 alloc_reg(¤t,i+1,CSREG);
8887 // Don't alloc the delay slot yet because we might not execute it
8888 if(likely[i]) // BC1FL/BC1TL
8890 alloc_cc(¤t,i);
8891 dirty_reg(¤t,CCREG);
8892 alloc_reg(¤t,i,CSREG);
8893 alloc_reg(¤t,i,FSREG);
8899 imm16_alloc(¤t,i);
8903 load_alloc(¤t,i);
8907 store_alloc(¤t,i);
8910 alu_alloc(¤t,i);
8913 shift_alloc(¤t,i);
8916 multdiv_alloc(¤t,i);
8919 shiftimm_alloc(¤t,i);
8922 mov_alloc(¤t,i);
8925 cop0_alloc(¤t,i);
8929 cop1_alloc(¤t,i);
8932 c1ls_alloc(¤t,i);
8935 c2ls_alloc(¤t,i);
8938 c2op_alloc(¤t,i);
8941 fconv_alloc(¤t,i);
8944 float_alloc(¤t,i);
8947 fcomp_alloc(¤t,i);
8951 syscall_alloc(¤t,i);
8954 pagespan_alloc(¤t,i);
8958 // Drop the upper half of registers that have become 32-bit
8959 current.uu|=current.is32&((1LL<<rt1[i])|(1LL<<rt2[i]));
8960 if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP) {
8961 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8962 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8965 current.uu|=current.is32&((1LL<<rt1[i+1])|(1LL<<rt2[i+1]));
8966 current.uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
8967 if((~current.uu>>rt1[i+1])&1) current.uu&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
8968 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8972 // Create entry (branch target) regmap
8973 for(hr=0;hr<HOST_REGS;hr++)
8976 r=current.regmap[hr];
8978 if(r!=regmap_pre[i][hr]) {
8979 // TODO: delay slot (?)
8980 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
8981 if(or<0||(r&63)>=TEMPREG){
8982 regs[i].regmap_entry[hr]=-1;
8986 // Just move it to a different register
8987 regs[i].regmap_entry[hr]=r;
8988 // If it was dirty before, it's still dirty
8989 if((regs[i].wasdirty>>or)&1) dirty_reg(¤t,r&63);
8996 regs[i].regmap_entry[hr]=0;
9000 if((current.u>>r)&1) {
9001 regs[i].regmap_entry[hr]=-1;
9002 //regs[i].regmap[hr]=-1;
9003 current.regmap[hr]=-1;
9005 regs[i].regmap_entry[hr]=r;
9008 if((current.uu>>(r&63))&1) {
9009 regs[i].regmap_entry[hr]=-1;
9010 //regs[i].regmap[hr]=-1;
9011 current.regmap[hr]=-1;
9013 regs[i].regmap_entry[hr]=r;
9017 // Branches expect CCREG to be allocated at the target
9018 if(regmap_pre[i][hr]==CCREG)
9019 regs[i].regmap_entry[hr]=CCREG;
9021 regs[i].regmap_entry[hr]=-1;
9024 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
9026 /* Branch post-alloc */
9029 current.was32=current.is32;
9030 current.wasdirty=current.dirty;
9031 switch(itype[i-1]) {
9033 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9034 branch_regs[i-1].isconst=0;
9035 branch_regs[i-1].wasconst=0;
9036 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
9037 branch_regs[i-1].uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
9038 alloc_cc(&branch_regs[i-1],i-1);
9039 dirty_reg(&branch_regs[i-1],CCREG);
9040 if(rt1[i-1]==31) { // JAL
9041 alloc_reg(&branch_regs[i-1],i-1,31);
9042 dirty_reg(&branch_regs[i-1],31);
9043 branch_regs[i-1].is32|=1LL<<31;
9045 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9046 memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9049 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9050 branch_regs[i-1].isconst=0;
9051 branch_regs[i-1].wasconst=0;
9052 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
9053 branch_regs[i-1].uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
9054 alloc_cc(&branch_regs[i-1],i-1);
9055 dirty_reg(&branch_regs[i-1],CCREG);
9056 alloc_reg(&branch_regs[i-1],i-1,rs1[i-1]);
9057 if(rt1[i-1]!=0) { // JALR
9058 alloc_reg(&branch_regs[i-1],i-1,rt1[i-1]);
9059 dirty_reg(&branch_regs[i-1],rt1[i-1]);
9060 branch_regs[i-1].is32|=1LL<<rt1[i-1];
9063 if(rs1[i-1]==31) { // JALR
9064 alloc_reg(&branch_regs[i-1],i-1,RHASH);
9065 #ifndef HOST_IMM_ADDR32
9066 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
9070 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9071 memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9074 if((opcode[i-1]&0x3E)==4) // BEQ/BNE
9076 alloc_cc(¤t,i-1);
9077 dirty_reg(¤t,CCREG);
9078 if((rs1[i-1]&&(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]))||
9079 (rs2[i-1]&&(rs2[i-1]==rt1[i]||rs2[i-1]==rt2[i]))) {
9080 // The delay slot overwrote one of our conditions
9081 // Delay slot goes after the test (in order)
9082 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
9083 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
9084 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9087 delayslot_alloc(¤t,i);
9092 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
9093 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
9094 // Alloc the branch condition registers
9095 if(rs1[i-1]) alloc_reg(¤t,i-1,rs1[i-1]);
9096 if(rs2[i-1]) alloc_reg(¤t,i-1,rs2[i-1]);
9097 if(!((current.is32>>rs1[i-1])&(current.is32>>rs2[i-1])&1))
9099 if(rs1[i-1]) alloc_reg64(¤t,i-1,rs1[i-1]);
9100 if(rs2[i-1]) alloc_reg64(¤t,i-1,rs2[i-1]);
9103 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9104 branch_regs[i-1].isconst=0;
9105 branch_regs[i-1].wasconst=0;
9106 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
9107 memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9110 if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
9112 alloc_cc(¤t,i-1);
9113 dirty_reg(¤t,CCREG);
9114 if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
9115 // The delay slot overwrote the branch condition
9116 // Delay slot goes after the test (in order)
9117 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
9118 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
9119 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9122 delayslot_alloc(¤t,i);
9127 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
9128 current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
9129 // Alloc the branch condition register
9130 alloc_reg(¤t,i-1,rs1[i-1]);
9131 if(!(current.is32>>rs1[i-1]&1))
9133 alloc_reg64(¤t,i-1,rs1[i-1]);
9136 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9137 branch_regs[i-1].isconst=0;
9138 branch_regs[i-1].wasconst=0;
9139 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
9140 memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9143 // Alloc the delay slot in case the branch is taken
9144 if((opcode[i-1]&0x3E)==0x14) // BEQL/BNEL
9146 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9147 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9148 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9149 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9150 alloc_cc(&branch_regs[i-1],i);
9151 dirty_reg(&branch_regs[i-1],CCREG);
9152 delayslot_alloc(&branch_regs[i-1],i);
9153 branch_regs[i-1].isconst=0;
9154 alloc_reg(¤t,i,CCREG); // Not taken path
9155 dirty_reg(¤t,CCREG);
9156 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9159 if((opcode[i-1]&0x3E)==0x16) // BLEZL/BGTZL
9161 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9162 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9163 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9164 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9165 alloc_cc(&branch_regs[i-1],i);
9166 dirty_reg(&branch_regs[i-1],CCREG);
9167 delayslot_alloc(&branch_regs[i-1],i);
9168 branch_regs[i-1].isconst=0;
9169 alloc_reg(¤t,i,CCREG); // Not taken path
9170 dirty_reg(¤t,CCREG);
9171 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9175 //if((opcode2[i-1]&0x1E)==0) // BLTZ/BGEZ
9176 if((opcode2[i-1]&0x0E)==0) // BLTZ/BGEZ
9178 alloc_cc(¤t,i-1);
9179 dirty_reg(¤t,CCREG);
9180 if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
9181 // The delay slot overwrote the branch condition
9182 // Delay slot goes after the test (in order)
9183 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
9184 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
9185 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
9188 delayslot_alloc(¤t,i);
9193 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
9194 current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
9195 // Alloc the branch condition register
9196 alloc_reg(¤t,i-1,rs1[i-1]);
9197 if(!(current.is32>>rs1[i-1]&1))
9199 alloc_reg64(¤t,i-1,rs1[i-1]);
9202 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9203 branch_regs[i-1].isconst=0;
9204 branch_regs[i-1].wasconst=0;
9205 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
9206 memcpy(constmap[i],constmap[i-1],sizeof(current.constmap));
9209 // Alloc the delay slot in case the branch is taken
9210 if((opcode2[i-1]&0x1E)==2) // BLTZL/BGEZL
9212 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9213 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9214 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9215 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9216 alloc_cc(&branch_regs[i-1],i);
9217 dirty_reg(&branch_regs[i-1],CCREG);
9218 delayslot_alloc(&branch_regs[i-1],i);
9219 branch_regs[i-1].isconst=0;
9220 alloc_reg(¤t,i,CCREG); // Not taken path
9221 dirty_reg(¤t,CCREG);
9222 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9224 // FIXME: BLTZAL/BGEZAL
9225 if(opcode2[i-1]&0x10) { // BxxZAL
9226 alloc_reg(&branch_regs[i-1],i-1,31);
9227 dirty_reg(&branch_regs[i-1],31);
9228 branch_regs[i-1].is32|=1LL<<31;
9232 if(likely[i-1]==0) // BC1F/BC1T
9234 alloc_cc(¤t,i-1);
9235 dirty_reg(¤t,CCREG);
9236 if(itype[i]==FCOMP) {
9237 // The delay slot overwrote the branch condition
9238 // Delay slot goes after the test (in order)
9239 delayslot_alloc(¤t,i);
9244 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
9245 current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
9246 // Alloc the branch condition register
9247 alloc_reg(¤t,i-1,FSREG);
9249 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9250 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
9254 // Alloc the delay slot in case the branch is taken
9255 memcpy(&branch_regs[i-1],¤t,sizeof(current));
9256 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9257 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
9258 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
9259 alloc_cc(&branch_regs[i-1],i);
9260 dirty_reg(&branch_regs[i-1],CCREG);
9261 delayslot_alloc(&branch_regs[i-1],i);
9262 branch_regs[i-1].isconst=0;
9263 alloc_reg(¤t,i,CCREG); // Not taken path
9264 dirty_reg(¤t,CCREG);
9265 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
9270 if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
9272 if(rt1[i-1]==31) // JAL/JALR
9274 // Subroutine call will return here, don't alloc any registers
9277 clear_all_regs(current.regmap);
9278 alloc_reg(¤t,i,CCREG);
9279 dirty_reg(¤t,CCREG);
9283 // Internal branch will jump here, match registers to caller
9284 current.is32=0x3FFFFFFFFLL;
9286 clear_all_regs(current.regmap);
9287 alloc_reg(¤t,i,CCREG);
9288 dirty_reg(¤t,CCREG);
9291 if(ba[j]==start+i*4+4) {
9292 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
9293 current.is32=branch_regs[j].is32;
9294 current.dirty=branch_regs[j].dirty;
9299 if(ba[j]==start+i*4+4) {
9300 for(hr=0;hr<HOST_REGS;hr++) {
9301 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
9302 current.regmap[hr]=-1;
9304 current.is32&=branch_regs[j].is32;
9305 current.dirty&=branch_regs[j].dirty;
9314 // Count cycles in between branches
9316 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP||itype[i]==SYSCALL||itype[i]==HLECALL))
9325 flush_dirty_uppers(¤t);
9327 regs[i].is32=current.is32;
9328 regs[i].dirty=current.dirty;
9329 regs[i].isconst=current.isconst;
9330 memcpy(constmap[i],current.constmap,sizeof(current.constmap));
9332 for(hr=0;hr<HOST_REGS;hr++) {
9333 if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
9334 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
9335 regs[i].wasconst&=~(1<<hr);
9339 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
9342 /* Pass 4 - Cull unused host registers */
9346 for (i=slen-1;i>=0;i--)
9349 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9351 if(ba[i]<start || ba[i]>=(start+slen*4))
9353 // Branch out of this block, don't need anything
9359 // Need whatever matches the target
9361 int t=(ba[i]-start)>>2;
9362 for(hr=0;hr<HOST_REGS;hr++)
9364 if(regs[i].regmap_entry[hr]>=0) {
9365 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
9369 // Conditional branch may need registers for following instructions
9370 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9373 nr|=needed_reg[i+2];
9374 for(hr=0;hr<HOST_REGS;hr++)
9376 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
9377 //if((regmap_entry[i+2][hr])>=0) if(!((nr>>hr)&1)) printf("%x-bogus(%d=%d)\n",start+i*4,hr,regmap_entry[i+2][hr]);
9381 // Don't need stuff which is overwritten
9382 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
9383 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
9384 // Merge in delay slot
9385 for(hr=0;hr<HOST_REGS;hr++)
9388 // These are overwritten unless the branch is "likely"
9389 // and the delay slot is nullified if not taken
9390 if(rt1[i+1]&&rt1[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9391 if(rt2[i+1]&&rt2[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9393 if(us1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9394 if(us2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9395 if(rs1[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
9396 if(rs2[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
9397 if(us1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9398 if(us2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9399 if(rs1[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9400 if(rs2[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9401 if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1)) {
9402 if(dep1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9403 if(dep2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9405 if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1)) {
9406 if(dep1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9407 if(dep2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9409 if(itype[i+1]==STORE || itype[i+1]==STORELR || (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) {
9410 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
9411 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
9415 else if(itype[i]==SYSCALL||itype[i]==HLECALL)
9417 // SYSCALL instruction (software interrupt)
9420 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
9422 // ERET instruction (return from interrupt)
9428 for(hr=0;hr<HOST_REGS;hr++) {
9429 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
9430 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
9431 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
9432 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
9436 for(hr=0;hr<HOST_REGS;hr++)
9438 // Overwritten registers are not needed
9439 if(rt1[i]&&rt1[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9440 if(rt2[i]&&rt2[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9441 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
9442 // Source registers are needed
9443 if(us1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9444 if(us2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9445 if(rs1[i]==regmap_pre[i][hr]) nr|=1<<hr;
9446 if(rs2[i]==regmap_pre[i][hr]) nr|=1<<hr;
9447 if(us1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9448 if(us2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9449 if(rs1[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9450 if(rs2[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
9451 if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1)) {
9452 if(dep1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9453 if(dep1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9455 if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1)) {
9456 if(dep2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9457 if(dep2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9459 if(itype[i]==STORE || itype[i]==STORELR || (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) {
9460 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
9461 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
9463 // Don't store a register immediately after writing it,
9464 // may prevent dual-issue.
9465 // But do so if this is a branch target, otherwise we
9466 // might have to load the register before the branch.
9467 if(i>0&&!bt[i]&&((regs[i].wasdirty>>hr)&1)) {
9468 if((regmap_pre[i][hr]>0&®map_pre[i][hr]<64&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1)) ||
9469 (regmap_pre[i][hr]>64&&!((unneeded_reg_upper[i]>>(regmap_pre[i][hr]&63))&1)) ) {
9470 if(rt1[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9471 if(rt2[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9473 if((regs[i].regmap_entry[hr]>0&®s[i].regmap_entry[hr]<64&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1)) ||
9474 (regs[i].regmap_entry[hr]>64&&!((unneeded_reg_upper[i]>>(regs[i].regmap_entry[hr]&63))&1)) ) {
9475 if(rt1[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9476 if(rt2[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9480 // Cycle count is needed at branches. Assume it is needed at the target too.
9481 if(i==0||bt[i]||itype[i]==CJUMP||itype[i]==FJUMP||itype[i]==SPAN) {
9482 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
9483 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
9488 // Deallocate unneeded registers
9489 for(hr=0;hr<HOST_REGS;hr++)
9492 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
9493 if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
9494 (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9495 (regs[i].regmap[hr]&63)!=PTEMP && (regs[i].regmap[hr]&63)!=CCREG)
9497 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9500 regs[i].regmap[hr]=-1;
9501 regs[i].isconst&=~(1<<hr);
9502 if(i<slen-2) regmap_pre[i+2][hr]=-1;
9506 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9508 int d1=0,d2=0,map=0,temp=0;
9509 if(get_reg(regs[i].regmap,rt1[i+1]|64)>=0||get_reg(branch_regs[i].regmap,rt1[i+1]|64)>=0)
9515 if(itype[i+1]==LOAD || itype[i+1]==LOADLR ||
9516 itype[i+1]==STORE || itype[i+1]==STORELR ||
9517 itype[i+1]==C1LS || itype[i+1]==C2LS)
9520 if(itype[i+1]==STORE || itype[i+1]==STORELR ||
9521 (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
9524 if(itype[i+1]==LOADLR || itype[i+1]==STORELR ||
9525 itype[i+1]==C1LS || itype[i+1]==C2LS)
9527 if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
9528 (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9529 (regs[i].regmap[hr]&63)!=rt1[i+1] && (regs[i].regmap[hr]&63)!=rt2[i+1] &&
9530 (regs[i].regmap[hr]^64)!=us1[i+1] && (regs[i].regmap[hr]^64)!=us2[i+1] &&
9531 (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
9532 regs[i].regmap[hr]!=rs1[i+1] && regs[i].regmap[hr]!=rs2[i+1] &&
9533 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
9534 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
9535 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
9536 regs[i].regmap[hr]!=map )
9538 regs[i].regmap[hr]=-1;
9539 regs[i].isconst&=~(1<<hr);
9540 if((branch_regs[i].regmap[hr]&63)!=rs1[i] && (branch_regs[i].regmap[hr]&63)!=rs2[i] &&
9541 (branch_regs[i].regmap[hr]&63)!=rt1[i] && (branch_regs[i].regmap[hr]&63)!=rt2[i] &&
9542 (branch_regs[i].regmap[hr]&63)!=rt1[i+1] && (branch_regs[i].regmap[hr]&63)!=rt2[i+1] &&
9543 (branch_regs[i].regmap[hr]^64)!=us1[i+1] && (branch_regs[i].regmap[hr]^64)!=us2[i+1] &&
9544 (branch_regs[i].regmap[hr]^64)!=d1 && (branch_regs[i].regmap[hr]^64)!=d2 &&
9545 branch_regs[i].regmap[hr]!=rs1[i+1] && branch_regs[i].regmap[hr]!=rs2[i+1] &&
9546 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
9547 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
9548 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
9549 branch_regs[i].regmap[hr]!=map)
9551 branch_regs[i].regmap[hr]=-1;
9552 branch_regs[i].regmap_entry[hr]=-1;
9553 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9555 if(!likely[i]&&i<slen-2) {
9556 regmap_pre[i+2][hr]=-1;
9567 int d1=0,d2=0,map=-1,temp=-1;
9568 if(get_reg(regs[i].regmap,rt1[i]|64)>=0)
9574 if(itype[i]==LOAD || itype[i]==LOADLR ||
9575 itype[i]==STORE || itype[i]==STORELR ||
9576 itype[i]==C1LS || itype[i]==C2LS)
9578 } else if(itype[i]==STORE || itype[i]==STORELR ||
9579 (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
9582 if(itype[i]==LOADLR || itype[i]==STORELR ||
9583 itype[i]==C1LS || itype[i]==C2LS)
9585 if((regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9586 (regs[i].regmap[hr]^64)!=us1[i] && (regs[i].regmap[hr]^64)!=us2[i] &&
9587 (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
9588 regs[i].regmap[hr]!=rs1[i] && regs[i].regmap[hr]!=rs2[i] &&
9589 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map &&
9590 (itype[i]!=SPAN||regs[i].regmap[hr]!=CCREG))
9592 if(i<slen-1&&!is_ds[i]) {
9593 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]!=-1)
9594 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
9595 if(regs[i].regmap[hr]<64||!((regs[i].was32>>(regs[i].regmap[hr]&63))&1))
9597 printf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
9598 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
9600 regmap_pre[i+1][hr]=-1;
9601 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
9603 regs[i].regmap[hr]=-1;
9604 regs[i].isconst&=~(1<<hr);
9612 /* Pass 5 - Pre-allocate registers */
9614 // If a register is allocated during a loop, try to allocate it for the
9615 // entire loop, if possible. This avoids loading/storing registers
9616 // inside of the loop.
9618 signed char f_regmap[HOST_REGS];
9619 clear_all_regs(f_regmap);
9620 for(i=0;i<slen-1;i++)
9622 if(itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9624 if(ba[i]>=start && ba[i]<(start+i*4))
9625 if(itype[i+1]==NOP||itype[i+1]==MOV||itype[i+1]==ALU
9626 ||itype[i+1]==SHIFTIMM||itype[i+1]==IMM16||itype[i+1]==LOAD
9627 ||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
9628 ||itype[i+1]==SHIFT||itype[i+1]==COP1||itype[i+1]==FLOAT
9629 ||itype[i+1]==FCOMP||itype[i+1]==FCONV
9630 ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
9632 int t=(ba[i]-start)>>2;
9633 if(t>0&&(itype[t-1]!=UJUMP&&itype[t-1]!=RJUMP&&itype[t-1]!=CJUMP&&itype[t-1]!=SJUMP&&itype[t-1]!=FJUMP)) // loop_preload can't handle jumps into delay slots
9634 if(t<2||(itype[t-2]!=UJUMP)) // call/ret assumes no registers allocated
9635 for(hr=0;hr<HOST_REGS;hr++)
9637 if(regs[i].regmap[hr]>64) {
9638 if(!((regs[i].dirty>>hr)&1))
9639 f_regmap[hr]=regs[i].regmap[hr];
9640 else f_regmap[hr]=-1;
9642 else if(regs[i].regmap[hr]>=0) f_regmap[hr]=regs[i].regmap[hr];
9643 if(branch_regs[i].regmap[hr]>64) {
9644 if(!((branch_regs[i].dirty>>hr)&1))
9645 f_regmap[hr]=branch_regs[i].regmap[hr];
9646 else f_regmap[hr]=-1;
9648 else if(branch_regs[i].regmap[hr]>=0) f_regmap[hr]=branch_regs[i].regmap[hr];
9649 if(itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
9650 ||itype[i+1]==SHIFT||itype[i+1]==COP1||itype[i+1]==FLOAT
9651 ||itype[i+1]==FCOMP||itype[i+1]==FCONV
9652 ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
9654 // Test both in case the delay slot is ooo,
9655 // could be done better...
9656 if(count_free_regs(branch_regs[i].regmap)<2
9657 ||count_free_regs(regs[i].regmap)<2)
9658 f_regmap[hr]=branch_regs[i].regmap[hr];
9660 // Avoid dirty->clean transition
9661 // #ifdef DESTRUCTIVE_WRITEBACK here?
9662 if(t>0) if(get_reg(regmap_pre[t],f_regmap[hr])>=0) if((regs[t].wasdirty>>get_reg(regmap_pre[t],f_regmap[hr]))&1) f_regmap[hr]=-1;
9663 if(f_regmap[hr]>0) {
9664 if(regs[t].regmap_entry[hr]<0) {
9668 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
9669 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
9670 if(r>63&&((unneeded_reg_upper[j]>>(r&63))&1)) break;
9672 // NB This can exclude the case where the upper-half
9673 // register is lower numbered than the lower-half
9674 // register. Not sure if it's worth fixing...
9675 if(get_reg(regs[j].regmap,r&63)<0) break;
9676 if(regs[j].is32&(1LL<<(r&63))) break;
9678 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
9679 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
9681 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
9682 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
9684 if(get_reg(regs[i].regmap,r&63)<0) break;
9685 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
9688 while(k>1&®s[k-1].regmap[hr]==-1) {
9689 if(itype[k-1]==STORE||itype[k-1]==STORELR
9690 ||itype[k-1]==C1LS||itype[k-1]==SHIFT||itype[k-1]==COP1
9691 ||itype[k-1]==FLOAT||itype[k-1]==FCONV||itype[k-1]==FCOMP
9692 ||itype[k-1]==COP2||itype[k-1]==C2LS||itype[k-1]==C2OP) {
9693 if(count_free_regs(regs[k-1].regmap)<2) {
9694 //printf("no free regs for store %x\n",start+(k-1)*4);
9699 if(itype[k-1]!=NOP&&itype[k-1]!=MOV&&itype[k-1]!=ALU&&itype[k-1]!=SHIFTIMM&&itype[k-1]!=IMM16&&itype[k-1]!=LOAD) break;
9700 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
9701 //printf("no-match due to different register\n");
9704 if(itype[k-2]==UJUMP||itype[k-2]==RJUMP||itype[k-2]==CJUMP||itype[k-2]==SJUMP||itype[k-2]==FJUMP) {
9705 //printf("no-match due to branch\n");
9708 // call/ret fast path assumes no registers allocated
9709 if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)) {
9713 // NB This can exclude the case where the upper-half
9714 // register is lower numbered than the lower-half
9715 // register. Not sure if it's worth fixing...
9716 if(get_reg(regs[k-1].regmap,r&63)<0) break;
9717 if(regs[k-1].is32&(1LL<<(r&63))) break;
9722 if((regs[k].is32&(1LL<<f_regmap[hr]))!=
9723 (regs[i+2].was32&(1LL<<f_regmap[hr]))) {
9724 //printf("bad match after branch\n");
9728 if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
9729 //printf("Extend r%d, %x ->\n",hr,start+k*4);
9731 regs[k].regmap_entry[hr]=f_regmap[hr];
9732 regs[k].regmap[hr]=f_regmap[hr];
9733 regmap_pre[k+1][hr]=f_regmap[hr];
9734 regs[k].wasdirty&=~(1<<hr);
9735 regs[k].dirty&=~(1<<hr);
9736 regs[k].wasdirty|=(1<<hr)®s[k-1].dirty;
9737 regs[k].dirty|=(1<<hr)®s[k].wasdirty;
9738 regs[k].wasconst&=~(1<<hr);
9739 regs[k].isconst&=~(1<<hr);
9744 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
9747 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
9748 if(regs[i-1].regmap[hr]==f_regmap[hr]&®map_pre[i][hr]==f_regmap[hr]) {
9749 //printf("OK fill %x (r%d)\n",start+i*4,hr);
9750 regs[i].regmap_entry[hr]=f_regmap[hr];
9751 regs[i].regmap[hr]=f_regmap[hr];
9752 regs[i].wasdirty&=~(1<<hr);
9753 regs[i].dirty&=~(1<<hr);
9754 regs[i].wasdirty|=(1<<hr)®s[i-1].dirty;
9755 regs[i].dirty|=(1<<hr)®s[i-1].dirty;
9756 regs[i].wasconst&=~(1<<hr);
9757 regs[i].isconst&=~(1<<hr);
9758 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
9759 branch_regs[i].wasdirty&=~(1<<hr);
9760 branch_regs[i].wasdirty|=(1<<hr)®s[i].dirty;
9761 branch_regs[i].regmap[hr]=f_regmap[hr];
9762 branch_regs[i].dirty&=~(1<<hr);
9763 branch_regs[i].dirty|=(1<<hr)®s[i].dirty;
9764 branch_regs[i].wasconst&=~(1<<hr);
9765 branch_regs[i].isconst&=~(1<<hr);
9766 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
9767 regmap_pre[i+2][hr]=f_regmap[hr];
9768 regs[i+2].wasdirty&=~(1<<hr);
9769 regs[i+2].wasdirty|=(1<<hr)®s[i].dirty;
9770 assert((branch_regs[i].is32&(1LL<<f_regmap[hr]))==
9771 (regs[i+2].was32&(1LL<<f_regmap[hr])));
9776 regs[k].regmap_entry[hr]=f_regmap[hr];
9777 regs[k].regmap[hr]=f_regmap[hr];
9778 regmap_pre[k+1][hr]=f_regmap[hr];
9779 regs[k+1].wasdirty&=~(1<<hr);
9780 regs[k].dirty&=~(1<<hr);
9781 regs[k].wasconst&=~(1<<hr);
9782 regs[k].isconst&=~(1<<hr);
9784 if(regs[j].regmap[hr]==f_regmap[hr])
9785 regs[j].regmap_entry[hr]=f_regmap[hr];
9789 if(regs[j].regmap[hr]>=0)
9791 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
9792 //printf("no-match due to different register\n");
9795 if((regs[j+1].is32&(1LL<<f_regmap[hr]))!=(regs[j].is32&(1LL<<f_regmap[hr]))) {
9796 //printf("32/64 mismatch %x %d\n",start+j*4,hr);
9799 if(itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS
9800 ||itype[j]==SHIFT||itype[j]==COP1||itype[j]==FLOAT
9801 ||itype[j]==FCOMP||itype[j]==FCONV
9802 ||itype[j]==COP2||itype[j]==C2LS||itype[j]==C2OP) {
9803 if(count_free_regs(regs[j].regmap)<2) {
9804 //printf("No free regs for store %x\n",start+j*4);
9808 else if(itype[j]!=NOP&&itype[j]!=MOV&&itype[j]!=ALU&&itype[j]!=SHIFTIMM&&itype[j]!=IMM16&&itype[j]!=LOAD) break;
9809 if(f_regmap[hr]>=64) {
9810 if(regs[j].is32&(1LL<<(f_regmap[hr]&63))) {
9815 if(get_reg(regs[j].regmap,f_regmap[hr]&63)<0) {
9827 for(hr=0;hr<HOST_REGS;hr++)
9829 if(hr!=EXCLUDE_REG) {
9830 if(regs[i].regmap[hr]>64) {
9831 if(!((regs[i].dirty>>hr)&1))
9832 f_regmap[hr]=regs[i].regmap[hr];
9834 else if(regs[i].regmap[hr]>=0) f_regmap[hr]=regs[i].regmap[hr];
9835 else if(regs[i].regmap[hr]<0) count++;
9838 // Try to restore cycle count at branch targets
9840 for(j=i;j<slen-1;j++) {
9841 if(regs[j].regmap[HOST_CCREG]!=-1) break;
9842 if(itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS
9843 ||itype[j]==SHIFT||itype[j]==COP1||itype[j]==FLOAT
9844 ||itype[j]==FCOMP||itype[j]==FCONV
9845 ||itype[j]==COP2||itype[j]==C2LS||itype[j]==C2OP) {
9846 if(count_free_regs(regs[j].regmap)<2) {
9847 //printf("no free regs for store %x\n",start+j*4);
9852 if(itype[j]!=NOP&&itype[j]!=MOV&&itype[j]!=ALU&&itype[j]!=SHIFTIMM&&itype[j]!=IMM16&&itype[j]!=LOAD) break;
9854 if(regs[j].regmap[HOST_CCREG]==CCREG) {
9856 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
9858 regs[k].regmap_entry[HOST_CCREG]=CCREG;
9859 regs[k].regmap[HOST_CCREG]=CCREG;
9860 regmap_pre[k+1][HOST_CCREG]=CCREG;
9861 regs[k+1].wasdirty|=1<<HOST_CCREG;
9862 regs[k].dirty|=1<<HOST_CCREG;
9863 regs[k].wasconst&=~(1<<HOST_CCREG);
9864 regs[k].isconst&=~(1<<HOST_CCREG);
9867 regs[j].regmap_entry[HOST_CCREG]=CCREG;
9869 // Work backwards from the branch target
9870 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
9872 //printf("Extend backwards\n");
9875 while(regs[k-1].regmap[HOST_CCREG]==-1) {
9876 if(itype[k-1]==STORE||itype[k-1]==STORELR||itype[k-1]==C1LS
9877 ||itype[k-1]==SHIFT||itype[k-1]==COP1||itype[k-1]==FLOAT
9878 ||itype[k-1]==FCONV||itype[k-1]==FCOMP
9879 ||itype[k-1]==COP2||itype[k-1]==C2LS||itype[k-1]==C2OP) {
9880 if(count_free_regs(regs[k-1].regmap)<2) {
9881 //printf("no free regs for store %x\n",start+(k-1)*4);
9886 if(itype[k-1]!=NOP&&itype[k-1]!=MOV&&itype[k-1]!=ALU&&itype[k-1]!=SHIFTIMM&&itype[k-1]!=IMM16&&itype[k-1]!=LOAD) break;
9889 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
9890 //printf("Extend CC, %x ->\n",start+k*4);
9892 regs[k].regmap_entry[HOST_CCREG]=CCREG;
9893 regs[k].regmap[HOST_CCREG]=CCREG;
9894 regmap_pre[k+1][HOST_CCREG]=CCREG;
9895 regs[k+1].wasdirty|=1<<HOST_CCREG;
9896 regs[k].dirty|=1<<HOST_CCREG;
9897 regs[k].wasconst&=~(1<<HOST_CCREG);
9898 regs[k].isconst&=~(1<<HOST_CCREG);
9903 //printf("Fail Extend CC, %x ->\n",start+k*4);
9907 if(itype[i]!=STORE&&itype[i]!=STORELR&&itype[i]!=C1LS&&itype[i]!=SHIFT&&
9908 itype[i]!=NOP&&itype[i]!=MOV&&itype[i]!=ALU&&itype[i]!=SHIFTIMM&&
9909 itype[i]!=IMM16&&itype[i]!=LOAD&&itype[i]!=COP1&&itype[i]!=FLOAT&&
9910 itype[i]!=FCONV&&itype[i]!=FCOMP&&
9911 itype[i]!=COP2&&itype[i]!=C2LS&&itype[i]!=C2OP)
9913 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
9918 // This allocates registers (if possible) one instruction prior
9919 // to use, which can avoid a load-use penalty on certain CPUs.
9920 for(i=0;i<slen-1;i++)
9922 if(!i||(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP))
9926 if(itype[i]==ALU||itype[i]==MOV||itype[i]==LOAD||itype[i]==SHIFTIMM||itype[i]==IMM16
9927 ||((itype[i]==COP1||itype[i]==COP2)&&opcode2[i]<3))
9930 if((hr=get_reg(regs[i+1].regmap,rs1[i+1]))>=0)
9932 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9934 regs[i].regmap[hr]=regs[i+1].regmap[hr];
9935 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
9936 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
9937 regs[i].isconst&=~(1<<hr);
9938 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9939 constmap[i][hr]=constmap[i+1][hr];
9940 regs[i+1].wasdirty&=~(1<<hr);
9941 regs[i].dirty&=~(1<<hr);
9946 if((hr=get_reg(regs[i+1].regmap,rs2[i+1]))>=0)
9948 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9950 regs[i].regmap[hr]=regs[i+1].regmap[hr];
9951 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
9952 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
9953 regs[i].isconst&=~(1<<hr);
9954 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9955 constmap[i][hr]=constmap[i+1][hr];
9956 regs[i+1].wasdirty&=~(1<<hr);
9957 regs[i].dirty&=~(1<<hr);
9961 if(itype[i+1]==LOAD&&rs1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
9962 if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
9964 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9966 regs[i].regmap[hr]=rs1[i+1];
9967 regmap_pre[i+1][hr]=rs1[i+1];
9968 regs[i+1].regmap_entry[hr]=rs1[i+1];
9969 regs[i].isconst&=~(1<<hr);
9970 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9971 constmap[i][hr]=constmap[i+1][hr];
9972 regs[i+1].wasdirty&=~(1<<hr);
9973 regs[i].dirty&=~(1<<hr);
9977 if(lt1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
9978 if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
9980 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9982 regs[i].regmap[hr]=rs1[i+1];
9983 regmap_pre[i+1][hr]=rs1[i+1];
9984 regs[i+1].regmap_entry[hr]=rs1[i+1];
9985 regs[i].isconst&=~(1<<hr);
9986 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9987 constmap[i][hr]=constmap[i+1][hr];
9988 regs[i+1].wasdirty&=~(1<<hr);
9989 regs[i].dirty&=~(1<<hr);
9993 #ifndef HOST_IMM_ADDR32
9994 if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) {
9995 hr=get_reg(regs[i+1].regmap,TLREG);
9997 int sr=get_reg(regs[i+1].regmap,rs1[i+1]);
9998 if(sr>=0&&((regs[i+1].wasconst>>sr)&1)) {
10000 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
10002 regs[i].regmap[hr]=MGEN1+((i+1)&1);
10003 regmap_pre[i+1][hr]=MGEN1+((i+1)&1);
10004 regs[i+1].regmap_entry[hr]=MGEN1+((i+1)&1);
10005 regs[i].isconst&=~(1<<hr);
10006 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10007 constmap[i][hr]=constmap[i+1][hr];
10008 regs[i+1].wasdirty&=~(1<<hr);
10009 regs[i].dirty&=~(1<<hr);
10011 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
10013 // move it to another register
10014 regs[i+1].regmap[hr]=-1;
10015 regmap_pre[i+2][hr]=-1;
10016 regs[i+1].regmap[nr]=TLREG;
10017 regmap_pre[i+2][nr]=TLREG;
10018 regs[i].regmap[nr]=MGEN1+((i+1)&1);
10019 regmap_pre[i+1][nr]=MGEN1+((i+1)&1);
10020 regs[i+1].regmap_entry[nr]=MGEN1+((i+1)&1);
10021 regs[i].isconst&=~(1<<nr);
10022 regs[i+1].isconst&=~(1<<nr);
10023 regs[i].dirty&=~(1<<nr);
10024 regs[i+1].wasdirty&=~(1<<nr);
10025 regs[i+1].dirty&=~(1<<nr);
10026 regs[i+2].wasdirty&=~(1<<nr);
10032 if(itype[i+1]==STORE||itype[i+1]==STORELR
10033 ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
10034 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
10035 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
10036 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
10037 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
10039 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
10041 regs[i].regmap[hr]=rs1[i+1];
10042 regmap_pre[i+1][hr]=rs1[i+1];
10043 regs[i+1].regmap_entry[hr]=rs1[i+1];
10044 regs[i].isconst&=~(1<<hr);
10045 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10046 constmap[i][hr]=constmap[i+1][hr];
10047 regs[i+1].wasdirty&=~(1<<hr);
10048 regs[i].dirty&=~(1<<hr);
10052 if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
10053 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
10055 hr=get_reg(regs[i+1].regmap,FTEMP);
10057 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
10059 regs[i].regmap[hr]=rs1[i+1];
10060 regmap_pre[i+1][hr]=rs1[i+1];
10061 regs[i+1].regmap_entry[hr]=rs1[i+1];
10062 regs[i].isconst&=~(1<<hr);
10063 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
10064 constmap[i][hr]=constmap[i+1][hr];
10065 regs[i+1].wasdirty&=~(1<<hr);
10066 regs[i].dirty&=~(1<<hr);
10068 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
10070 // move it to another register
10071 regs[i+1].regmap[hr]=-1;
10072 regmap_pre[i+2][hr]=-1;
10073 regs[i+1].regmap[nr]=FTEMP;
10074 regmap_pre[i+2][nr]=FTEMP;
10075 regs[i].regmap[nr]=rs1[i+1];
10076 regmap_pre[i+1][nr]=rs1[i+1];
10077 regs[i+1].regmap_entry[nr]=rs1[i+1];
10078 regs[i].isconst&=~(1<<nr);
10079 regs[i+1].isconst&=~(1<<nr);
10080 regs[i].dirty&=~(1<<nr);
10081 regs[i+1].wasdirty&=~(1<<nr);
10082 regs[i+1].dirty&=~(1<<nr);
10083 regs[i+2].wasdirty&=~(1<<nr);
10087 if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR/*||itype[i+1]==C1LS||||itype[i+1]==C2LS*/) {
10088 if(itype[i+1]==LOAD)
10089 hr=get_reg(regs[i+1].regmap,rt1[i+1]);
10090 if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
10091 hr=get_reg(regs[i+1].regmap,FTEMP);
10092 if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
10093 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
10094 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
10096 if(hr>=0&®s[i].regmap[hr]<0) {
10097 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
10098 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
10099 regs[i].regmap[hr]=AGEN1+((i+1)&1);
10100 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
10101 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
10102 regs[i].isconst&=~(1<<hr);
10103 regs[i+1].wasdirty&=~(1<<hr);
10104 regs[i].dirty&=~(1<<hr);
10113 /* Pass 6 - Optimize clean/dirty state */
10114 clean_registers(0,slen-1,1);
10116 /* Pass 7 - Identify 32-bit registers */
10122 for (i=slen-1;i>=0;i--)
10125 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
10127 if(ba[i]<start || ba[i]>=(start+slen*4))
10129 // Branch out of this block, don't need anything
10135 // Need whatever matches the target
10136 // (and doesn't get overwritten by the delay slot instruction)
10138 int t=(ba[i]-start)>>2;
10139 if(ba[i]>start+i*4) {
10141 if(!(requires_32bit[t]&~regs[i].was32))
10142 r32|=requires_32bit[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
10145 //if(!(regs[t].was32&~unneeded_reg_upper[t]&~regs[i].was32))
10146 // r32|=regs[t].was32&~unneeded_reg_upper[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
10147 if(!(pr32[t]&~regs[i].was32))
10148 r32|=pr32[t]&(~(1LL<<rt1[i+1]))&(~(1LL<<rt2[i+1]));
10151 // Conditional branch may need registers for following instructions
10152 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
10155 r32|=requires_32bit[i+2];
10156 r32&=regs[i].was32;
10157 // Mark this address as a branch target since it may be called
10158 // upon return from interrupt
10162 // Merge in delay slot
10164 // These are overwritten unless the branch is "likely"
10165 // and the delay slot is nullified if not taken
10166 r32&=~(1LL<<rt1[i+1]);
10167 r32&=~(1LL<<rt2[i+1]);
10169 // Assume these are needed (delay slot)
10172 if((regs[i].was32>>us1[i+1])&1) r32|=1LL<<us1[i+1];
10176 if((regs[i].was32>>us2[i+1])&1) r32|=1LL<<us2[i+1];
10178 if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1))
10180 if((regs[i].was32>>dep1[i+1])&1) r32|=1LL<<dep1[i+1];
10182 if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1))
10184 if((regs[i].was32>>dep2[i+1])&1) r32|=1LL<<dep2[i+1];
10187 else if(itype[i]==SYSCALL||itype[i]==HLECALL)
10189 // SYSCALL instruction (software interrupt)
10192 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
10194 // ERET instruction (return from interrupt)
10198 r32&=~(1LL<<rt1[i]);
10199 r32&=~(1LL<<rt2[i]);
10202 if((regs[i].was32>>us1[i])&1) r32|=1LL<<us1[i];
10206 if((regs[i].was32>>us2[i])&1) r32|=1LL<<us2[i];
10208 if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1))
10210 if((regs[i].was32>>dep1[i])&1) r32|=1LL<<dep1[i];
10212 if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1))
10214 if((regs[i].was32>>dep2[i])&1) r32|=1LL<<dep2[i];
10216 requires_32bit[i]=r32;
10218 // Dirty registers which are 32-bit, require 32-bit input
10219 // as they will be written as 32-bit values
10220 for(hr=0;hr<HOST_REGS;hr++)
10222 if(regs[i].regmap_entry[hr]>0&®s[i].regmap_entry[hr]<64) {
10223 if((regs[i].was32>>regs[i].regmap_entry[hr])&(regs[i].wasdirty>>hr)&1) {
10224 if(!((unneeded_reg_upper[i]>>regs[i].regmap_entry[hr])&1))
10225 requires_32bit[i]|=1LL<<regs[i].regmap_entry[hr];
10229 //requires_32bit[i]=is32[i]&~unneeded_reg_upper[i]; // DEBUG
10232 if(itype[slen-1]==SPAN) {
10233 bt[slen-1]=1; // Mark as a branch target so instruction can restart after exception
10236 /* Debug/disassembly */
10237 if((void*)assem_debug==(void*)printf)
10238 for(i=0;i<slen;i++)
10242 for(r=1;r<=CCREG;r++) {
10243 if((unneeded_reg[i]>>r)&1) {
10244 if(r==HIREG) printf(" HI");
10245 else if(r==LOREG) printf(" LO");
10246 else printf(" r%d",r);
10251 for(r=1;r<=CCREG;r++) {
10252 if(((unneeded_reg_upper[i]&~unneeded_reg[i])>>r)&1) {
10253 if(r==HIREG) printf(" HI");
10254 else if(r==LOREG) printf(" LO");
10255 else printf(" r%d",r);
10259 for(r=0;r<=CCREG;r++) {
10260 //if(((is32[i]>>r)&(~unneeded_reg[i]>>r))&1) {
10261 if((regs[i].was32>>r)&1) {
10262 if(r==CCREG) printf(" CC");
10263 else if(r==HIREG) printf(" HI");
10264 else if(r==LOREG) printf(" LO");
10265 else printf(" r%d",r);
10270 #if defined(__i386__) || defined(__x86_64__)
10271 printf("pre: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7]);
10274 printf("pre: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][4],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7],regmap_pre[i][8],regmap_pre[i][9],regmap_pre[i][10],regmap_pre[i][12]);
10277 if(needed_reg[i]&1) printf("eax ");
10278 if((needed_reg[i]>>1)&1) printf("ecx ");
10279 if((needed_reg[i]>>2)&1) printf("edx ");
10280 if((needed_reg[i]>>3)&1) printf("ebx ");
10281 if((needed_reg[i]>>5)&1) printf("ebp ");
10282 if((needed_reg[i]>>6)&1) printf("esi ");
10283 if((needed_reg[i]>>7)&1) printf("edi ");
10285 for(r=0;r<=CCREG;r++) {
10286 //if(((requires_32bit[i]>>r)&(~unneeded_reg[i]>>r))&1) {
10287 if((requires_32bit[i]>>r)&1) {
10288 if(r==CCREG) printf(" CC");
10289 else if(r==HIREG) printf(" HI");
10290 else if(r==LOREG) printf(" LO");
10291 else printf(" r%d",r);
10296 for(r=0;r<=CCREG;r++) {
10297 //if(((requires_32bit[i]>>r)&(~unneeded_reg[i]>>r))&1) {
10298 if((pr32[i]>>r)&1) {
10299 if(r==CCREG) printf(" CC");
10300 else if(r==HIREG) printf(" HI");
10301 else if(r==LOREG) printf(" LO");
10302 else printf(" r%d",r);
10305 if(pr32[i]!=requires_32bit[i]) printf(" OOPS");
10307 #if defined(__i386__) || defined(__x86_64__)
10308 printf("entry: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7]);
10310 if(regs[i].wasdirty&1) printf("eax ");
10311 if((regs[i].wasdirty>>1)&1) printf("ecx ");
10312 if((regs[i].wasdirty>>2)&1) printf("edx ");
10313 if((regs[i].wasdirty>>3)&1) printf("ebx ");
10314 if((regs[i].wasdirty>>5)&1) printf("ebp ");
10315 if((regs[i].wasdirty>>6)&1) printf("esi ");
10316 if((regs[i].wasdirty>>7)&1) printf("edi ");
10319 printf("entry: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[4],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7],regs[i].regmap_entry[8],regs[i].regmap_entry[9],regs[i].regmap_entry[10],regs[i].regmap_entry[12]);
10321 if(regs[i].wasdirty&1) printf("r0 ");
10322 if((regs[i].wasdirty>>1)&1) printf("r1 ");
10323 if((regs[i].wasdirty>>2)&1) printf("r2 ");
10324 if((regs[i].wasdirty>>3)&1) printf("r3 ");
10325 if((regs[i].wasdirty>>4)&1) printf("r4 ");
10326 if((regs[i].wasdirty>>5)&1) printf("r5 ");
10327 if((regs[i].wasdirty>>6)&1) printf("r6 ");
10328 if((regs[i].wasdirty>>7)&1) printf("r7 ");
10329 if((regs[i].wasdirty>>8)&1) printf("r8 ");
10330 if((regs[i].wasdirty>>9)&1) printf("r9 ");
10331 if((regs[i].wasdirty>>10)&1) printf("r10 ");
10332 if((regs[i].wasdirty>>12)&1) printf("r12 ");
10335 disassemble_inst(i);
10336 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
10337 #if defined(__i386__) || defined(__x86_64__)
10338 printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7]);
10339 if(regs[i].dirty&1) printf("eax ");
10340 if((regs[i].dirty>>1)&1) printf("ecx ");
10341 if((regs[i].dirty>>2)&1) printf("edx ");
10342 if((regs[i].dirty>>3)&1) printf("ebx ");
10343 if((regs[i].dirty>>5)&1) printf("ebp ");
10344 if((regs[i].dirty>>6)&1) printf("esi ");
10345 if((regs[i].dirty>>7)&1) printf("edi ");
10348 printf("r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[4],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7],regs[i].regmap[8],regs[i].regmap[9],regs[i].regmap[10],regs[i].regmap[12]);
10349 if(regs[i].dirty&1) printf("r0 ");
10350 if((regs[i].dirty>>1)&1) printf("r1 ");
10351 if((regs[i].dirty>>2)&1) printf("r2 ");
10352 if((regs[i].dirty>>3)&1) printf("r3 ");
10353 if((regs[i].dirty>>4)&1) printf("r4 ");
10354 if((regs[i].dirty>>5)&1) printf("r5 ");
10355 if((regs[i].dirty>>6)&1) printf("r6 ");
10356 if((regs[i].dirty>>7)&1) printf("r7 ");
10357 if((regs[i].dirty>>8)&1) printf("r8 ");
10358 if((regs[i].dirty>>9)&1) printf("r9 ");
10359 if((regs[i].dirty>>10)&1) printf("r10 ");
10360 if((regs[i].dirty>>12)&1) printf("r12 ");
10363 if(regs[i].isconst) {
10364 printf("constants: ");
10365 #if defined(__i386__) || defined(__x86_64__)
10366 if(regs[i].isconst&1) printf("eax=%x ",(int)constmap[i][0]);
10367 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(int)constmap[i][1]);
10368 if((regs[i].isconst>>2)&1) printf("edx=%x ",(int)constmap[i][2]);
10369 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(int)constmap[i][3]);
10370 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(int)constmap[i][5]);
10371 if((regs[i].isconst>>6)&1) printf("esi=%x ",(int)constmap[i][6]);
10372 if((regs[i].isconst>>7)&1) printf("edi=%x ",(int)constmap[i][7]);
10375 if(regs[i].isconst&1) printf("r0=%x ",(int)constmap[i][0]);
10376 if((regs[i].isconst>>1)&1) printf("r1=%x ",(int)constmap[i][1]);
10377 if((regs[i].isconst>>2)&1) printf("r2=%x ",(int)constmap[i][2]);
10378 if((regs[i].isconst>>3)&1) printf("r3=%x ",(int)constmap[i][3]);
10379 if((regs[i].isconst>>4)&1) printf("r4=%x ",(int)constmap[i][4]);
10380 if((regs[i].isconst>>5)&1) printf("r5=%x ",(int)constmap[i][5]);
10381 if((regs[i].isconst>>6)&1) printf("r6=%x ",(int)constmap[i][6]);
10382 if((regs[i].isconst>>7)&1) printf("r7=%x ",(int)constmap[i][7]);
10383 if((regs[i].isconst>>8)&1) printf("r8=%x ",(int)constmap[i][8]);
10384 if((regs[i].isconst>>9)&1) printf("r9=%x ",(int)constmap[i][9]);
10385 if((regs[i].isconst>>10)&1) printf("r10=%x ",(int)constmap[i][10]);
10386 if((regs[i].isconst>>12)&1) printf("r12=%x ",(int)constmap[i][12]);
10392 for(r=0;r<=CCREG;r++) {
10393 if((regs[i].is32>>r)&1) {
10394 if(r==CCREG) printf(" CC");
10395 else if(r==HIREG) printf(" HI");
10396 else if(r==LOREG) printf(" LO");
10397 else printf(" r%d",r);
10403 for(r=0;r<=CCREG;r++) {
10404 if((p32[i]>>r)&1) {
10405 if(r==CCREG) printf(" CC");
10406 else if(r==HIREG) printf(" HI");
10407 else if(r==LOREG) printf(" LO");
10408 else printf(" r%d",r);
10411 if(p32[i]!=regs[i].is32) printf(" NO MATCH\n");
10412 else printf("\n");*/
10413 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
10414 #if defined(__i386__) || defined(__x86_64__)
10415 printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
10416 if(branch_regs[i].dirty&1) printf("eax ");
10417 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
10418 if((branch_regs[i].dirty>>2)&1) printf("edx ");
10419 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
10420 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
10421 if((branch_regs[i].dirty>>6)&1) printf("esi ");
10422 if((branch_regs[i].dirty>>7)&1) printf("edi ");
10425 printf("branch(%d): r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[4],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7],branch_regs[i].regmap[8],branch_regs[i].regmap[9],branch_regs[i].regmap[10],branch_regs[i].regmap[12]);
10426 if(branch_regs[i].dirty&1) printf("r0 ");
10427 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
10428 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
10429 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
10430 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
10431 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
10432 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
10433 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
10434 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
10435 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
10436 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
10437 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
10441 for(r=0;r<=CCREG;r++) {
10442 if((branch_regs[i].is32>>r)&1) {
10443 if(r==CCREG) printf(" CC");
10444 else if(r==HIREG) printf(" HI");
10445 else if(r==LOREG) printf(" LO");
10446 else printf(" r%d",r);
10454 /* Pass 8 - Assembly */
10455 linkcount=0;stubcount=0;
10456 ds=0;is_delayslot=0;
10458 uint64_t is32_pre=0;
10460 u_int beginning=(u_int)out;
10461 if((u_int)addr&1) {
10465 for(i=0;i<slen;i++)
10467 //if(ds) printf("ds: ");
10468 if((void*)assem_debug==(void*)printf) disassemble_inst(i);
10470 ds=0; // Skip delay slot
10471 if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
10474 #ifndef DESTRUCTIVE_WRITEBACK
10475 if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
10477 wb_sx(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,is32_pre,regs[i].was32,
10478 unneeded_reg[i],unneeded_reg_upper[i]);
10479 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,is32_pre,
10480 unneeded_reg[i],unneeded_reg_upper[i]);
10482 is32_pre=regs[i].is32;
10483 dirty_pre=regs[i].dirty;
10486 if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
10488 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32,
10489 unneeded_reg[i],unneeded_reg_upper[i]);
10490 loop_preload(regmap_pre[i],regs[i].regmap_entry);
10492 // branch target entry point
10493 instr_addr[i]=(u_int)out;
10494 assem_debug("<->\n");
10496 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
10497 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32);
10498 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i],rs2[i]);
10499 address_generation(i,®s[i],regs[i].regmap_entry);
10500 load_consts(regmap_pre[i],regs[i].regmap,regs[i].was32,i);
10501 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
10503 // Load the delay slot registers if necessary
10504 if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i])
10505 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
10506 if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i])
10507 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
10508 if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a)
10509 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
10513 // Preload registers for following instruction
10514 if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i])
10515 if(rs1[i+1]!=rt1[i]&&rs1[i+1]!=rt2[i])
10516 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
10517 if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i])
10518 if(rs2[i+1]!=rt1[i]&&rs2[i+1]!=rt2[i])
10519 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
10521 // TODO: if(is_ooo(i)) address_generation(i+1);
10522 if(itype[i]==CJUMP||itype[i]==FJUMP)
10523 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
10524 if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a)
10525 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
10526 if(bt[i]) cop1_usable=0;
10530 alu_assemble(i,®s[i]);break;
10532 imm16_assemble(i,®s[i]);break;
10534 shift_assemble(i,®s[i]);break;
10536 shiftimm_assemble(i,®s[i]);break;
10538 load_assemble(i,®s[i]);break;
10540 loadlr_assemble(i,®s[i]);break;
10542 store_assemble(i,®s[i]);break;
10544 storelr_assemble(i,®s[i]);break;
10546 cop0_assemble(i,®s[i]);break;
10548 cop1_assemble(i,®s[i]);break;
10550 c1ls_assemble(i,®s[i]);break;
10552 cop2_assemble(i,®s[i]);break;
10554 c2ls_assemble(i,®s[i]);break;
10556 c2op_assemble(i,®s[i]);break;
10558 fconv_assemble(i,®s[i]);break;
10560 float_assemble(i,®s[i]);break;
10562 fcomp_assemble(i,®s[i]);break;
10564 multdiv_assemble(i,®s[i]);break;
10566 mov_assemble(i,®s[i]);break;
10568 syscall_assemble(i,®s[i]);break;
10570 hlecall_assemble(i,®s[i]);break;
10572 ujump_assemble(i,®s[i]);ds=1;break;
10574 rjump_assemble(i,®s[i]);ds=1;break;
10576 cjump_assemble(i,®s[i]);ds=1;break;
10578 sjump_assemble(i,®s[i]);ds=1;break;
10580 fjump_assemble(i,®s[i]);ds=1;break;
10582 pagespan_assemble(i,®s[i]);break;
10584 if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
10585 literal_pool(1024);
10587 literal_pool_jumpover(256);
10590 //assert(itype[i-2]==UJUMP||itype[i-2]==RJUMP||(source[i-2]>>16)==0x1000);
10591 // If the block did not end with an unconditional branch,
10592 // add a jump to the next instruction.
10594 if(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000&&itype[i-1]!=SPAN) {
10595 assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP);
10597 if(itype[i-2]!=CJUMP&&itype[i-2]!=SJUMP&&itype[i-2]!=FJUMP) {
10598 store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
10599 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
10600 emit_loadreg(CCREG,HOST_CCREG);
10601 emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i-1]+1),HOST_CCREG);
10603 else if(!likely[i-2])
10605 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].is32,branch_regs[i-2].dirty,start+i*4);
10606 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
10610 store_regs_bt(regs[i-2].regmap,regs[i-2].is32,regs[i-2].dirty,start+i*4);
10611 assert(regs[i-2].regmap[HOST_CCREG]==CCREG);
10613 add_to_linker((int)out,start+i*4,0);
10620 assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP);
10621 store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
10622 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
10623 emit_loadreg(CCREG,HOST_CCREG);
10624 emit_addimm(HOST_CCREG,CLOCK_DIVIDER*(ccadj[i-1]+1),HOST_CCREG);
10625 add_to_linker((int)out,start+i*4,0);
10629 // TODO: delay slot stubs?
10631 for(i=0;i<stubcount;i++)
10633 switch(stubs[i][0])
10641 do_readstub(i);break;
10646 do_writestub(i);break;
10648 do_ccstub(i);break;
10650 do_invstub(i);break;
10652 do_cop1stub(i);break;
10654 do_unalignedwritestub(i);break;
10658 /* Pass 9 - Linker */
10659 for(i=0;i<linkcount;i++)
10661 assem_debug("%8x -> %8x\n",link_addr[i][0],link_addr[i][1]);
10663 if(!link_addr[i][2])
10666 void *addr=check_addr(link_addr[i][1]);
10667 emit_extjump(link_addr[i][0],link_addr[i][1]);
10669 set_jump_target(link_addr[i][0],(int)addr);
10670 add_link(link_addr[i][1],stub);
10672 else set_jump_target(link_addr[i][0],(int)stub);
10677 int target=(link_addr[i][1]-start)>>2;
10678 assert(target>=0&&target<slen);
10679 assert(instr_addr[target]);
10680 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
10681 //set_jump_target_fillslot(link_addr[i][0],instr_addr[target],link_addr[i][2]>>1);
10683 set_jump_target(link_addr[i][0],instr_addr[target]);
10687 // External Branch Targets (jump_in)
10688 if(copy+slen*4>(void *)shadow+sizeof(shadow)) copy=shadow;
10689 for(i=0;i<slen;i++)
10693 if(instr_addr[i]) // TODO - delay slots (=null)
10695 u_int vaddr=start+i*4;
10696 u_int page=get_page(vaddr);
10697 u_int vpage=get_vpage(vaddr);
10699 //if(!(is32[i]&(~unneeded_reg_upper[i])&~(1LL<<CCREG)))
10700 if(!requires_32bit[i])
10702 assem_debug("%8x (%d) <- %8x\n",instr_addr[i],i,start+i*4);
10703 assem_debug("jump_in: %x\n",start+i*4);
10704 ll_add(jump_dirty+vpage,vaddr,(void *)out);
10705 int entry_point=do_dirty_stub(i);
10706 ll_add(jump_in+page,vaddr,(void *)entry_point);
10707 // If there was an existing entry in the hash table,
10708 // replace it with the new address.
10709 // Don't add new entries. We'll insert the
10710 // ones that actually get used in check_addr().
10711 int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
10712 if(ht_bin[0]==vaddr) {
10713 ht_bin[1]=entry_point;
10715 if(ht_bin[2]==vaddr) {
10716 ht_bin[3]=entry_point;
10721 u_int r=requires_32bit[i]|!!(requires_32bit[i]>>32);
10722 assem_debug("%8x (%d) <- %8x\n",instr_addr[i],i,start+i*4);
10723 assem_debug("jump_in: %x (restricted - %x)\n",start+i*4,r);
10724 //int entry_point=(int)out;
10725 ////assem_debug("entry_point: %x\n",entry_point);
10726 //load_regs_entry(i);
10727 //if(entry_point==(int)out)
10728 // entry_point=instr_addr[i];
10730 // emit_jmp(instr_addr[i]);
10731 //ll_add_32(jump_in+page,vaddr,r,(void *)entry_point);
10732 ll_add_32(jump_dirty+vpage,vaddr,r,(void *)out);
10733 int entry_point=do_dirty_stub(i);
10734 ll_add_32(jump_in+page,vaddr,r,(void *)entry_point);
10739 // Write out the literal pool if necessary
10741 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
10743 if(((u_int)out)&7) emit_addnop(13);
10745 assert((u_int)out-beginning<MAX_OUTPUT_BLOCK_SIZE);
10746 //printf("shadow buffer: %x-%x\n",(int)copy,(int)copy+slen*4);
10747 memcpy(copy,source,slen*4);
10751 __clear_cache((void *)beginning,out);
10754 // If we're within 256K of the end of the buffer,
10755 // start over from the beginning. (Is 256K enough?)
10756 if((int)out>BASE_ADDR+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE) out=(u_char *)BASE_ADDR;
10758 // Trap writes to any of the pages we compiled
10759 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
10761 #ifndef DISABLE_TLB
10762 memory_map[i]|=0x40000000;
10763 if((signed int)start>=(signed int)0xC0000000) {
10765 j=(((u_int)i<<12)+(memory_map[i]<<2)-(u_int)rdram+(u_int)0x80000000)>>12;
10767 memory_map[j]|=0x40000000;
10768 //printf("write protect physical page: %x (virtual %x)\n",j<<12,start);
10773 /* Pass 10 - Free memory by expiring oldest blocks */
10775 int end=((((int)out-BASE_ADDR)>>(TARGET_SIZE_2-16))+16384)&65535;
10776 while(expirep!=end)
10778 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
10779 int base=BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
10780 inv_debug("EXP: Phase %d\n",expirep);
10781 switch((expirep>>11)&3)
10784 // Clear jump_in and jump_dirty
10785 ll_remove_matching_addrs(jump_in+(expirep&2047),base,shift);
10786 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base,shift);
10787 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base,shift);
10788 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base,shift);
10792 ll_kill_pointers(jump_out[expirep&2047],base,shift);
10793 ll_kill_pointers(jump_out[(expirep&2047)+2048],base,shift);
10796 // Clear hash table
10797 for(i=0;i<32;i++) {
10798 int *ht_bin=hash_table[((expirep&2047)<<5)+i];
10799 if((ht_bin[3]>>shift)==(base>>shift) ||
10800 ((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
10801 inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[2],ht_bin[3]);
10802 ht_bin[2]=ht_bin[3]=-1;
10804 if((ht_bin[1]>>shift)==(base>>shift) ||
10805 ((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
10806 inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[0],ht_bin[1]);
10807 ht_bin[0]=ht_bin[2];
10808 ht_bin[1]=ht_bin[3];
10809 ht_bin[2]=ht_bin[3]=-1;
10816 if((expirep&2047)==0)
10817 __clear_cache((void *)BASE_ADDR,(void *)BASE_ADDR+(1<<TARGET_SIZE_2));
10819 ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
10820 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
10823 expirep=(expirep+1)&65535;
10828 // vim:shiftwidth=2:expandtab