1 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
2 * Mupen64plus - new_dynarec.c *
3 * Copyright (C) 2009-2011 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
27 #include <libkern/OSCacheControl.h>
30 #include <3ds_utils.h>
33 #include <psp2/kernel/sysmem.h>
37 #include "new_dynarec_config.h"
38 #include "../psxhle.h" //emulator interface
39 #include "emu_if.h" //emulator interface
42 //#define assem_debug printf
43 //#define inv_debug printf
44 #define assem_debug(...)
45 #define inv_debug(...)
48 #include "assem_x86.h"
51 #include "assem_x64.h"
54 #include "assem_arm.h"
58 #define MAX_OUTPUT_BLOCK_SIZE 262144
62 signed char regmap_entry[HOST_REGS];
63 signed char regmap[HOST_REGS];
72 u_int loadedconst; // host regs that have constants loaded
73 u_int waswritten; // MIPS regs that were used as store base before
76 // note: asm depends on this layout
82 struct ll_entry *next;
87 u_int hash_table[65536][4] __attribute__((aligned(16)));
88 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
89 struct ll_entry *jump_dirty[4096];
91 static struct ll_entry *jump_out[4096];
94 static char insn[MAXBLOCK][10];
95 static u_char itype[MAXBLOCK];
96 static u_char opcode[MAXBLOCK];
97 static u_char opcode2[MAXBLOCK];
98 static u_char bt[MAXBLOCK];
99 static u_char rs1[MAXBLOCK];
100 static u_char rs2[MAXBLOCK];
101 static u_char rt1[MAXBLOCK];
102 static u_char rt2[MAXBLOCK];
103 static u_char us1[MAXBLOCK];
104 static u_char us2[MAXBLOCK];
105 static u_char dep1[MAXBLOCK];
106 static u_char dep2[MAXBLOCK];
107 static u_char lt1[MAXBLOCK];
108 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
109 static uint64_t gte_rt[MAXBLOCK];
110 static uint64_t gte_unneeded[MAXBLOCK];
111 static u_int smrv[32]; // speculated MIPS register values
112 static u_int smrv_strong; // mask or regs that are likely to have correct values
113 static u_int smrv_weak; // same, but somewhat less likely
114 static u_int smrv_strong_next; // same, but after current insn executes
115 static u_int smrv_weak_next;
116 static int imm[MAXBLOCK];
117 static u_int ba[MAXBLOCK];
118 static char likely[MAXBLOCK];
119 static char is_ds[MAXBLOCK];
120 static char ooo[MAXBLOCK];
121 static uint64_t unneeded_reg[MAXBLOCK];
122 static uint64_t unneeded_reg_upper[MAXBLOCK];
123 static uint64_t branch_unneeded_reg[MAXBLOCK];
124 static uint64_t branch_unneeded_reg_upper[MAXBLOCK];
125 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
126 static uint64_t current_constmap[HOST_REGS];
127 static uint64_t constmap[MAXBLOCK][HOST_REGS];
128 static struct regstat regs[MAXBLOCK];
129 static struct regstat branch_regs[MAXBLOCK];
130 static signed char minimum_free_regs[MAXBLOCK];
131 static u_int needed_reg[MAXBLOCK];
132 static u_int wont_dirty[MAXBLOCK];
133 static u_int will_dirty[MAXBLOCK];
134 static int ccadj[MAXBLOCK];
136 static u_int instr_addr[MAXBLOCK];
137 static u_int link_addr[MAXBLOCK][3];
138 static int linkcount;
139 static u_int stubs[MAXBLOCK*3][8];
140 static int stubcount;
141 static u_int literals[1024][2];
142 static int literalcount;
143 static int is_delayslot;
144 static int cop1_usable;
145 static char shadow[1048576] __attribute__((aligned(16)));
148 static u_int stop_after_jal;
150 static u_int ram_offset;
152 static const u_int ram_offset=0;
155 int new_dynarec_hacks;
156 int new_dynarec_did_compile;
157 extern u_char restore_candidate[512];
158 extern int cycle_count;
160 /* registers that may be allocated */
162 #define HIREG 32 // hi
163 #define LOREG 33 // lo
164 #define FSREG 34 // FPU status (FCSR)
165 #define CSREG 35 // Coprocessor status
166 #define CCREG 36 // Cycle count
167 #define INVCP 37 // Pointer to invalid_code
168 //#define MMREG 38 // Pointer to memory_map
169 #define ROREG 39 // ram offset (if rdram!=0x80000000)
171 #define FTEMP 40 // FPU temporary register
172 #define PTEMP 41 // Prefetch temporary register
173 //#define TLREG 42 // TLB mapping offset
174 #define RHASH 43 // Return address hash
175 #define RHTBL 44 // Return address hash table address
176 #define RTEMP 45 // JR/JALR address register
178 #define AGEN1 46 // Address generation temporary register
179 //#define AGEN2 47 // Address generation temporary register
180 //#define MGEN1 48 // Maptable address generation temporary register
181 //#define MGEN2 49 // Maptable address generation temporary register
182 #define BTREG 50 // Branch target temporary register
184 /* instruction types */
185 #define NOP 0 // No operation
186 #define LOAD 1 // Load
187 #define STORE 2 // Store
188 #define LOADLR 3 // Unaligned load
189 #define STORELR 4 // Unaligned store
190 #define MOV 5 // Move
191 #define ALU 6 // Arithmetic/logic
192 #define MULTDIV 7 // Multiply/divide
193 #define SHIFT 8 // Shift by register
194 #define SHIFTIMM 9// Shift by immediate
195 #define IMM16 10 // 16-bit immediate
196 #define RJUMP 11 // Unconditional jump to register
197 #define UJUMP 12 // Unconditional jump
198 #define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
199 #define SJUMP 14 // Conditional branch (regimm format)
200 #define COP0 15 // Coprocessor 0
201 #define COP1 16 // Coprocessor 1
202 #define C1LS 17 // Coprocessor 1 load/store
203 #define FJUMP 18 // Conditional branch (floating point)
204 #define FLOAT 19 // Floating point unit
205 #define FCONV 20 // Convert integer to float
206 #define FCOMP 21 // Floating point compare (sets FSREG)
207 #define SYSCALL 22// SYSCALL
208 #define OTHER 23 // Other
209 #define SPAN 24 // Branch/delay slot spans 2 pages
210 #define NI 25 // Not implemented
211 #define HLECALL 26// PCSX fake opcodes for HLE
212 #define COP2 27 // Coprocessor 2 move
213 #define C2LS 28 // Coprocessor 2 load/store
214 #define C2OP 29 // Coprocessor 2 operation
215 #define INTCALL 30// Call interpreter to handle rare corner cases
224 #define LOADBU_STUB 7
225 #define LOADHU_STUB 8
226 #define STOREB_STUB 9
227 #define STOREH_STUB 10
228 #define STOREW_STUB 11
229 #define STORED_STUB 12
230 #define STORELR_STUB 13
231 #define INVCODE_STUB 14
239 int new_recompile_block(int addr);
240 void *get_addr_ht(u_int vaddr);
241 void invalidate_block(u_int block);
242 void invalidate_addr(u_int addr);
243 void remove_hash(int vaddr);
245 void dyna_linker_ds();
247 void verify_code_vm();
248 void verify_code_ds();
251 void fp_exception_ds();
252 void jump_syscall_hle();
255 void new_dyna_leave();
257 // Needed by assembler
258 static void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32);
259 static void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty);
260 static void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr);
261 static void load_all_regs(signed char i_regmap[]);
262 static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
263 static void load_regs_entry(int t);
264 static void load_all_consts(signed char regmap[],int is32,u_int dirty,int i);
266 static int verify_dirty(u_int *ptr);
267 static int get_final_value(int hr, int i, int *value);
268 static void add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e);
269 static void add_to_linker(int addr,int target,int ext);
271 static int tracedebug=0;
273 static void mprotect_w_x(void *start, void *end, int is_x)
277 // *Open* enables write on all memory that was
278 // allocated by sceKernelAllocMemBlockForVM()?
280 sceKernelCloseVMDomain();
282 sceKernelOpenVMDomain();
284 u_long mstart = (u_long)start & ~4095ul;
285 u_long mend = (u_long)end;
286 if (mprotect((void *)mstart, mend - mstart,
287 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
288 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
293 static void start_tcache_write(void *start, void *end)
295 mprotect_w_x(start, end, 0);
298 static void end_tcache_write(void *start, void *end)
301 size_t len = (char *)end - (char *)start;
302 #if defined(__BLACKBERRY_QNX__)
303 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
304 #elif defined(__MACH__)
305 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
307 sceKernelSyncVMDomain(sceBlock, start, len);
309 ctr_flush_invalidate_cache();
311 __clear_cache(start, end);
316 mprotect_w_x(start, end, 1);
319 static void *start_block(void)
321 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
322 if (end > (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2))
323 end = (u_char *)BASE_ADDR + (1<<TARGET_SIZE_2);
324 start_tcache_write(out, end);
328 static void end_block(void *start)
330 end_tcache_write(start, out);
333 //#define DEBUG_CYCLE_COUNT 1
335 #define NO_CYCLE_PENALTY_THR 12
337 int cycle_multiplier; // 100 for 1.0
339 static int CLOCK_ADJUST(int x)
342 return (x * cycle_multiplier + s * 50) / 100;
345 static u_int get_page(u_int vaddr)
347 u_int page=vaddr&~0xe0000000;
348 if (page < 0x1000000)
349 page &= ~0x0e00000; // RAM mirrors
351 if(page>2048) page=2048+(page&2047);
355 // no virtual mem in PCSX
356 static u_int get_vpage(u_int vaddr)
358 return get_page(vaddr);
361 // Get address from virtual address
362 // This is called from the recompiled JR/JALR instructions
363 void *get_addr(u_int vaddr)
365 u_int page=get_page(vaddr);
366 u_int vpage=get_vpage(vaddr);
367 struct ll_entry *head;
368 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
371 if(head->vaddr==vaddr) {
372 //printf("TRACE: count=%d next=%d (get_addr match %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
373 u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
376 ht_bin[1]=(u_int)head->addr;
382 head=jump_dirty[vpage];
384 if(head->vaddr==vaddr) {
385 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %x)\n",Count,next_interupt,vaddr,(int)head->addr);
386 // Don't restore blocks which are about to expire from the cache
387 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
388 if(verify_dirty(head->addr)) {
389 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
390 invalid_code[vaddr>>12]=0;
391 inv_code_start=inv_code_end=~0;
393 restore_candidate[vpage>>3]|=1<<(vpage&7);
395 else restore_candidate[page>>3]|=1<<(page&7);
396 u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
397 if(ht_bin[0]==vaddr) {
398 ht_bin[1]=(u_int)head->addr; // Replace existing entry
404 ht_bin[1]=(int)head->addr;
412 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
413 int r=new_recompile_block(vaddr);
414 if(r==0) return get_addr(vaddr);
415 // Execute in unmapped page, generate pagefault execption
417 Cause=(vaddr<<31)|0x8;
418 EPC=(vaddr&1)?vaddr-5:vaddr;
420 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
421 EntryHi=BadVAddr&0xFFFFE000;
422 return get_addr_ht(0x80000000);
424 // Look up address in hash table first
425 void *get_addr_ht(u_int vaddr)
427 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
428 u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
429 if(ht_bin[0]==vaddr) return (void *)ht_bin[1];
430 if(ht_bin[2]==vaddr) return (void *)ht_bin[3];
431 return get_addr(vaddr);
434 void clear_all_regs(signed char regmap[])
437 for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
440 signed char get_reg(signed char regmap[],int r)
443 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map[hr]==r) return hr;
447 // Find a register that is available for two consecutive cycles
448 signed char get_reg2(signed char regmap1[],signed char regmap2[],int r)
451 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map1[hr]==r&®map2[hr]==r) return hr;
455 int count_free_regs(signed char regmap[])
459 for(hr=0;hr<HOST_REGS;hr++)
461 if(hr!=EXCLUDE_REG) {
462 if(regmap[hr]<0) count++;
468 void dirty_reg(struct regstat *cur,signed char reg)
472 for (hr=0;hr<HOST_REGS;hr++) {
473 if((cur->regmap[hr]&63)==reg) {
479 // If we dirty the lower half of a 64 bit register which is now being
480 // sign-extended, we need to dump the upper half.
481 // Note: Do this only after completion of the instruction, because
482 // some instructions may need to read the full 64-bit value even if
483 // overwriting it (eg SLTI, DSRA32).
484 static void flush_dirty_uppers(struct regstat *cur)
487 for (hr=0;hr<HOST_REGS;hr++) {
488 if((cur->dirty>>hr)&1) {
491 if((cur->is32>>(reg&63))&1) cur->regmap[hr]=-1;
496 void set_const(struct regstat *cur,signed char reg,uint64_t value)
500 for (hr=0;hr<HOST_REGS;hr++) {
501 if(cur->regmap[hr]==reg) {
503 current_constmap[hr]=value;
505 else if((cur->regmap[hr]^64)==reg) {
507 current_constmap[hr]=value>>32;
512 void clear_const(struct regstat *cur,signed char reg)
516 for (hr=0;hr<HOST_REGS;hr++) {
517 if((cur->regmap[hr]&63)==reg) {
518 cur->isconst&=~(1<<hr);
523 int is_const(struct regstat *cur,signed char reg)
528 for (hr=0;hr<HOST_REGS;hr++) {
529 if((cur->regmap[hr]&63)==reg) {
530 return (cur->isconst>>hr)&1;
535 uint64_t get_const(struct regstat *cur,signed char reg)
539 for (hr=0;hr<HOST_REGS;hr++) {
540 if(cur->regmap[hr]==reg) {
541 return current_constmap[hr];
544 SysPrintf("Unknown constant in r%d\n",reg);
548 // Least soon needed registers
549 // Look at the next ten instructions and see which registers
550 // will be used. Try not to reallocate these.
551 void lsn(u_char hsn[], int i, int *preferred_reg)
561 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
563 // Don't go past an unconditonal jump
570 if(rs1[i+j]) hsn[rs1[i+j]]=j;
571 if(rs2[i+j]) hsn[rs2[i+j]]=j;
572 if(rt1[i+j]) hsn[rt1[i+j]]=j;
573 if(rt2[i+j]) hsn[rt2[i+j]]=j;
574 if(itype[i+j]==STORE || itype[i+j]==STORELR) {
575 // Stores can allocate zero
579 // On some architectures stores need invc_ptr
580 #if defined(HOST_IMM8)
581 if(itype[i+j]==STORE || itype[i+j]==STORELR || (opcode[i+j]&0x3b)==0x39 || (opcode[i+j]&0x3b)==0x3a) {
585 if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
593 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
595 // Follow first branch
596 int t=(ba[i+b]-start)>>2;
597 j=7-b;if(t+j>=slen) j=slen-t-1;
600 if(rs1[t+j]) if(hsn[rs1[t+j]]>j+b+2) hsn[rs1[t+j]]=j+b+2;
601 if(rs2[t+j]) if(hsn[rs2[t+j]]>j+b+2) hsn[rs2[t+j]]=j+b+2;
602 //if(rt1[t+j]) if(hsn[rt1[t+j]]>j+b+2) hsn[rt1[t+j]]=j+b+2;
603 //if(rt2[t+j]) if(hsn[rt2[t+j]]>j+b+2) hsn[rt2[t+j]]=j+b+2;
606 // TODO: preferred register based on backward branch
608 // Delay slot should preferably not overwrite branch conditions or cycle count
609 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
610 if(rs1[i-1]) if(hsn[rs1[i-1]]>1) hsn[rs1[i-1]]=1;
611 if(rs2[i-1]) if(hsn[rs2[i-1]]>1) hsn[rs2[i-1]]=1;
617 // Coprocessor load/store needs FTEMP, even if not declared
618 if(itype[i]==C1LS||itype[i]==C2LS) {
621 // Load L/R also uses FTEMP as a temporary register
622 if(itype[i]==LOADLR) {
625 // Also SWL/SWR/SDL/SDR
626 if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) {
629 // Don't remove the miniht registers
630 if(itype[i]==UJUMP||itype[i]==RJUMP)
637 // We only want to allocate registers if we're going to use them again soon
638 int needed_again(int r, int i)
644 if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
646 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
647 return 0; // Don't need any registers if exiting the block
655 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
657 // Don't go past an unconditonal jump
661 if(itype[i+j]==SYSCALL||itype[i+j]==HLECALL||itype[i+j]==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
668 if(rs1[i+j]==r) rn=j;
669 if(rs2[i+j]==r) rn=j;
670 if((unneeded_reg[i+j]>>r)&1) rn=10;
671 if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP||itype[i+j]==FJUMP))
679 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
681 // Follow first branch
683 int t=(ba[i+b]-start)>>2;
684 j=7-b;if(t+j>=slen) j=slen-t-1;
687 if(!((unneeded_reg[t+j]>>r)&1)) {
688 if(rs1[t+j]==r) if(rn>j+b+2) rn=j+b+2;
689 if(rs2[t+j]==r) if(rn>j+b+2) rn=j+b+2;
700 // Try to match register allocations at the end of a loop with those
702 int loop_reg(int i, int r, int hr)
711 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
713 // Don't go past an unconditonal jump
720 if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)
725 if(r<64&&((unneeded_reg[i+k]>>r)&1)) return hr;
726 if(r>64&&((unneeded_reg_upper[i+k]>>r)&1)) return hr;
727 if(i+k>=0&&(itype[i+k]==UJUMP||itype[i+k]==CJUMP||itype[i+k]==SJUMP||itype[i+k]==FJUMP))
729 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
731 int t=(ba[i+k]-start)>>2;
732 int reg=get_reg(regs[t].regmap_entry,r);
733 if(reg>=0) return reg;
734 //reg=get_reg(regs[t+1].regmap_entry,r);
735 //if(reg>=0) return reg;
743 // Allocate every register, preserving source/target regs
744 void alloc_all(struct regstat *cur,int i)
748 for(hr=0;hr<HOST_REGS;hr++) {
749 if(hr!=EXCLUDE_REG) {
750 if(((cur->regmap[hr]&63)!=rs1[i])&&((cur->regmap[hr]&63)!=rs2[i])&&
751 ((cur->regmap[hr]&63)!=rt1[i])&&((cur->regmap[hr]&63)!=rt2[i]))
754 cur->dirty&=~(1<<hr);
757 if((cur->regmap[hr]&63)==0)
760 cur->dirty&=~(1<<hr);
767 #include "assem_x86.c"
770 #include "assem_x64.c"
773 #include "assem_arm.c"
776 // Add virtual address mapping to linked list
777 void ll_add(struct ll_entry **head,int vaddr,void *addr)
779 struct ll_entry *new_entry;
780 new_entry=malloc(sizeof(struct ll_entry));
781 assert(new_entry!=NULL);
782 new_entry->vaddr=vaddr;
783 new_entry->reg_sv_flags=0;
784 new_entry->addr=addr;
785 new_entry->next=*head;
789 void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
791 ll_add(head,vaddr,addr);
792 (*head)->reg_sv_flags=reg_sv_flags;
795 // Check if an address is already compiled
796 // but don't return addresses which are about to expire from the cache
797 void *check_addr(u_int vaddr)
799 u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
800 if(ht_bin[0]==vaddr) {
801 if(((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
802 if(isclean(ht_bin[1])) return (void *)ht_bin[1];
804 if(ht_bin[2]==vaddr) {
805 if(((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2)))
806 if(isclean(ht_bin[3])) return (void *)ht_bin[3];
808 u_int page=get_page(vaddr);
809 struct ll_entry *head;
812 if(head->vaddr==vaddr) {
813 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
814 // Update existing entry with current address
815 if(ht_bin[0]==vaddr) {
816 ht_bin[1]=(int)head->addr;
819 if(ht_bin[2]==vaddr) {
820 ht_bin[3]=(int)head->addr;
823 // Insert into hash table with low priority.
824 // Don't evict existing entries, as they are probably
825 // addresses that are being accessed frequently.
827 ht_bin[1]=(int)head->addr;
829 }else if(ht_bin[2]==-1) {
830 ht_bin[3]=(int)head->addr;
841 void remove_hash(int vaddr)
843 //printf("remove hash: %x\n",vaddr);
844 u_int *ht_bin=hash_table[(((vaddr)>>16)^vaddr)&0xFFFF];
845 if(ht_bin[2]==vaddr) {
846 ht_bin[2]=ht_bin[3]=-1;
848 if(ht_bin[0]==vaddr) {
851 ht_bin[2]=ht_bin[3]=-1;
855 void ll_remove_matching_addrs(struct ll_entry **head,int addr,int shift)
857 struct ll_entry *next;
859 if(((u_int)((*head)->addr)>>shift)==(addr>>shift) ||
860 ((u_int)((*head)->addr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift))
862 inv_debug("EXP: Remove pointer to %x (%x)\n",(int)(*head)->addr,(*head)->vaddr);
863 remove_hash((*head)->vaddr);
870 head=&((*head)->next);
875 // Remove all entries from linked list
876 void ll_clear(struct ll_entry **head)
878 struct ll_entry *cur;
879 struct ll_entry *next;
890 // Dereference the pointers and remove if it matches
891 static void ll_kill_pointers(struct ll_entry *head,int addr,int shift)
894 int ptr=get_pointer(head->addr);
895 inv_debug("EXP: Lookup pointer to %x at %x (%x)\n",(int)ptr,(int)head->addr,head->vaddr);
896 if(((ptr>>shift)==(addr>>shift)) ||
897 (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
899 inv_debug("EXP: Kill pointer at %x (%x)\n",(int)head->addr,head->vaddr);
900 void *host_addr=find_extjump_insn(head->addr);
902 mark_clear_cache(host_addr);
904 set_jump_target((int)host_addr,(int)head->addr);
910 // This is called when we write to a compiled block (see do_invstub)
911 void invalidate_page(u_int page)
913 struct ll_entry *head;
914 struct ll_entry *next;
918 inv_debug("INVALIDATE: %x\n",head->vaddr);
919 remove_hash(head->vaddr);
927 inv_debug("INVALIDATE: kill pointer to %x (%x)\n",head->vaddr,(int)head->addr);
928 void *host_addr=find_extjump_insn(head->addr);
930 mark_clear_cache(host_addr);
932 set_jump_target((int)host_addr,(int)head->addr);
939 static void invalidate_block_range(u_int block, u_int first, u_int last)
941 u_int page=get_page(block<<12);
942 //printf("first=%d last=%d\n",first,last);
943 invalidate_page(page);
944 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
946 // Invalidate the adjacent pages if a block crosses a 4K boundary
948 invalidate_page(first);
951 for(first=page+1;first<last;first++) {
952 invalidate_page(first);
959 invalid_code[block]=1;
962 memset(mini_ht,-1,sizeof(mini_ht));
966 void invalidate_block(u_int block)
968 u_int page=get_page(block<<12);
969 u_int vpage=get_vpage(block<<12);
970 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
971 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
974 struct ll_entry *head;
975 head=jump_dirty[vpage];
976 //printf("page=%d vpage=%d\n",page,vpage);
979 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
980 get_bounds((int)head->addr,&start,&end);
981 //printf("start: %x end: %x\n",start,end);
982 if(page<2048&&start>=(u_int)rdram&&end<(u_int)rdram+RAM_SIZE) {
983 if(((start-(u_int)rdram)>>12)<=page&&((end-1-(u_int)rdram)>>12)>=page) {
984 if((((start-(u_int)rdram)>>12)&2047)<first) first=((start-(u_int)rdram)>>12)&2047;
985 if((((end-1-(u_int)rdram)>>12)&2047)>last) last=((end-1-(u_int)rdram)>>12)&2047;
991 invalidate_block_range(block,first,last);
994 void invalidate_addr(u_int addr)
997 // this check is done by the caller
998 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
999 u_int page=get_vpage(addr);
1000 if(page<2048) { // RAM
1001 struct ll_entry *head;
1002 u_int addr_min=~0, addr_max=0;
1003 u_int mask=RAM_SIZE-1;
1004 u_int addr_main=0x80000000|(addr&mask);
1006 inv_code_start=addr_main&~0xfff;
1007 inv_code_end=addr_main|0xfff;
1010 // must check previous page too because of spans..
1012 inv_code_start-=0x1000;
1014 for(;pg1<=page;pg1++) {
1015 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1017 get_bounds((int)head->addr,&start,&end);
1022 if(start<=addr_main&&addr_main<end) {
1023 if(start<addr_min) addr_min=start;
1024 if(end>addr_max) addr_max=end;
1026 else if(addr_main<start) {
1027 if(start<inv_code_end)
1028 inv_code_end=start-1;
1031 if(end>inv_code_start)
1037 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1038 inv_code_start=inv_code_end=~0;
1039 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1043 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1044 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1045 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1049 invalidate_block(addr>>12);
1052 // This is called when loading a save state.
1053 // Anything could have changed, so invalidate everything.
1054 void invalidate_all_pages()
1057 for(page=0;page<4096;page++)
1058 invalidate_page(page);
1059 for(page=0;page<1048576;page++)
1060 if(!invalid_code[page]) {
1061 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1062 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1065 memset(mini_ht,-1,sizeof(mini_ht));
1069 // Add an entry to jump_out after making a link
1070 void add_link(u_int vaddr,void *src)
1072 u_int page=get_page(vaddr);
1073 inv_debug("add_link: %x -> %x (%d)\n",(int)src,vaddr,page);
1074 int *ptr=(int *)(src+4);
1075 assert((*ptr&0x0fff0000)==0x059f0000);
1077 ll_add(jump_out+page,vaddr,src);
1078 //int ptr=get_pointer(src);
1079 //inv_debug("add_link: Pointer is to %x\n",(int)ptr);
1082 // If a code block was found to be unmodified (bit was set in
1083 // restore_candidate) and it remains unmodified (bit is clear
1084 // in invalid_code) then move the entries for that 4K page from
1085 // the dirty list to the clean list.
1086 void clean_blocks(u_int page)
1088 struct ll_entry *head;
1089 inv_debug("INV: clean_blocks page=%d\n",page);
1090 head=jump_dirty[page];
1092 if(!invalid_code[head->vaddr>>12]) {
1093 // Don't restore blocks which are about to expire from the cache
1094 if((((u_int)head->addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1096 if(verify_dirty(head->addr)) {
1097 //printf("Possibly Restore %x (%x)\n",head->vaddr, (int)head->addr);
1100 get_bounds((int)head->addr,&start,&end);
1101 if(start-(u_int)rdram<RAM_SIZE) {
1102 for(i=(start-(u_int)rdram+0x80000000)>>12;i<=(end-1-(u_int)rdram+0x80000000)>>12;i++) {
1103 inv|=invalid_code[i];
1106 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1110 void * clean_addr=(void *)get_clean_addr((int)head->addr);
1111 if((((u_int)clean_addr-(u_int)out)<<(32-TARGET_SIZE_2))>0x60000000+(MAX_OUTPUT_BLOCK_SIZE<<(32-TARGET_SIZE_2))) {
1113 inv_debug("INV: Restored %x (%x/%x)\n",head->vaddr, (int)head->addr, (int)clean_addr);
1114 //printf("page=%x, addr=%x\n",page,head->vaddr);
1115 //assert(head->vaddr>>12==(page|0x80000));
1116 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1117 u_int *ht_bin=hash_table[((head->vaddr>>16)^head->vaddr)&0xFFFF];
1118 if(ht_bin[0]==head->vaddr) {
1119 ht_bin[1]=(u_int)clean_addr; // Replace existing entry
1121 if(ht_bin[2]==head->vaddr) {
1122 ht_bin[3]=(u_int)clean_addr; // Replace existing entry
1134 void mov_alloc(struct regstat *current,int i)
1136 // Note: Don't need to actually alloc the source registers
1137 if((~current->is32>>rs1[i])&1) {
1138 //alloc_reg64(current,i,rs1[i]);
1139 alloc_reg64(current,i,rt1[i]);
1140 current->is32&=~(1LL<<rt1[i]);
1142 //alloc_reg(current,i,rs1[i]);
1143 alloc_reg(current,i,rt1[i]);
1144 current->is32|=(1LL<<rt1[i]);
1146 clear_const(current,rs1[i]);
1147 clear_const(current,rt1[i]);
1148 dirty_reg(current,rt1[i]);
1151 void shiftimm_alloc(struct regstat *current,int i)
1153 if(opcode2[i]<=0x3) // SLL/SRL/SRA
1156 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1158 alloc_reg(current,i,rt1[i]);
1159 current->is32|=1LL<<rt1[i];
1160 dirty_reg(current,rt1[i]);
1161 if(is_const(current,rs1[i])) {
1162 int v=get_const(current,rs1[i]);
1163 if(opcode2[i]==0x00) set_const(current,rt1[i],v<<imm[i]);
1164 if(opcode2[i]==0x02) set_const(current,rt1[i],(u_int)v>>imm[i]);
1165 if(opcode2[i]==0x03) set_const(current,rt1[i],v>>imm[i]);
1167 else clear_const(current,rt1[i]);
1172 clear_const(current,rs1[i]);
1173 clear_const(current,rt1[i]);
1176 if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
1179 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1180 alloc_reg64(current,i,rt1[i]);
1181 current->is32&=~(1LL<<rt1[i]);
1182 dirty_reg(current,rt1[i]);
1185 if(opcode2[i]==0x3c) // DSLL32
1188 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1189 alloc_reg64(current,i,rt1[i]);
1190 current->is32&=~(1LL<<rt1[i]);
1191 dirty_reg(current,rt1[i]);
1194 if(opcode2[i]==0x3e) // DSRL32
1197 alloc_reg64(current,i,rs1[i]);
1199 alloc_reg64(current,i,rt1[i]);
1200 current->is32&=~(1LL<<rt1[i]);
1202 alloc_reg(current,i,rt1[i]);
1203 current->is32|=1LL<<rt1[i];
1205 dirty_reg(current,rt1[i]);
1208 if(opcode2[i]==0x3f) // DSRA32
1211 alloc_reg64(current,i,rs1[i]);
1212 alloc_reg(current,i,rt1[i]);
1213 current->is32|=1LL<<rt1[i];
1214 dirty_reg(current,rt1[i]);
1219 void shift_alloc(struct regstat *current,int i)
1222 if(opcode2[i]<=0x07) // SLLV/SRLV/SRAV
1224 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1225 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1226 alloc_reg(current,i,rt1[i]);
1227 if(rt1[i]==rs2[i]) {
1228 alloc_reg_temp(current,i,-1);
1229 minimum_free_regs[i]=1;
1231 current->is32|=1LL<<rt1[i];
1232 } else { // DSLLV/DSRLV/DSRAV
1233 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1234 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1235 alloc_reg64(current,i,rt1[i]);
1236 current->is32&=~(1LL<<rt1[i]);
1237 if(opcode2[i]==0x16||opcode2[i]==0x17) // DSRLV and DSRAV need a temporary register
1239 alloc_reg_temp(current,i,-1);
1240 minimum_free_regs[i]=1;
1243 clear_const(current,rs1[i]);
1244 clear_const(current,rs2[i]);
1245 clear_const(current,rt1[i]);
1246 dirty_reg(current,rt1[i]);
1250 void alu_alloc(struct regstat *current,int i)
1252 if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
1254 if(rs1[i]&&rs2[i]) {
1255 alloc_reg(current,i,rs1[i]);
1256 alloc_reg(current,i,rs2[i]);
1259 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1260 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1262 alloc_reg(current,i,rt1[i]);
1264 current->is32|=1LL<<rt1[i];
1266 if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
1268 if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1270 alloc_reg64(current,i,rs1[i]);
1271 alloc_reg64(current,i,rs2[i]);
1272 alloc_reg(current,i,rt1[i]);
1274 alloc_reg(current,i,rs1[i]);
1275 alloc_reg(current,i,rs2[i]);
1276 alloc_reg(current,i,rt1[i]);
1279 current->is32|=1LL<<rt1[i];
1281 if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
1283 if(rs1[i]&&rs2[i]) {
1284 alloc_reg(current,i,rs1[i]);
1285 alloc_reg(current,i,rs2[i]);
1289 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1290 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1292 alloc_reg(current,i,rt1[i]);
1293 if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1295 if(!((current->uu>>rt1[i])&1)) {
1296 alloc_reg64(current,i,rt1[i]);
1298 if(get_reg(current->regmap,rt1[i]|64)>=0) {
1299 if(rs1[i]&&rs2[i]) {
1300 alloc_reg64(current,i,rs1[i]);
1301 alloc_reg64(current,i,rs2[i]);
1305 // Is is really worth it to keep 64-bit values in registers?
1307 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1308 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg64(current,i,rs2[i]);
1312 current->is32&=~(1LL<<rt1[i]);
1314 current->is32|=1LL<<rt1[i];
1318 if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1320 if(rs1[i]&&rs2[i]) {
1321 if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1322 alloc_reg64(current,i,rs1[i]);
1323 alloc_reg64(current,i,rs2[i]);
1324 alloc_reg64(current,i,rt1[i]);
1326 alloc_reg(current,i,rs1[i]);
1327 alloc_reg(current,i,rs2[i]);
1328 alloc_reg(current,i,rt1[i]);
1332 alloc_reg(current,i,rt1[i]);
1333 if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1334 // DADD used as move, or zeroing
1335 // If we have a 64-bit source, then make the target 64 bits too
1336 if(rs1[i]&&!((current->is32>>rs1[i])&1)) {
1337 if(get_reg(current->regmap,rs1[i])>=0) alloc_reg64(current,i,rs1[i]);
1338 alloc_reg64(current,i,rt1[i]);
1339 } else if(rs2[i]&&!((current->is32>>rs2[i])&1)) {
1340 if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1341 alloc_reg64(current,i,rt1[i]);
1343 if(opcode2[i]>=0x2e&&rs2[i]) {
1344 // DSUB used as negation - 64-bit result
1345 // If we have a 32-bit register, extend it to 64 bits
1346 if(get_reg(current->regmap,rs2[i])>=0) alloc_reg64(current,i,rs2[i]);
1347 alloc_reg64(current,i,rt1[i]);
1351 if(rs1[i]&&rs2[i]) {
1352 current->is32&=~(1LL<<rt1[i]);
1354 current->is32&=~(1LL<<rt1[i]);
1355 if((current->is32>>rs1[i])&1)
1356 current->is32|=1LL<<rt1[i];
1358 current->is32&=~(1LL<<rt1[i]);
1359 if((current->is32>>rs2[i])&1)
1360 current->is32|=1LL<<rt1[i];
1362 current->is32|=1LL<<rt1[i];
1366 clear_const(current,rs1[i]);
1367 clear_const(current,rs2[i]);
1368 clear_const(current,rt1[i]);
1369 dirty_reg(current,rt1[i]);
1372 void imm16_alloc(struct regstat *current,int i)
1374 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1376 if(rt1[i]) alloc_reg(current,i,rt1[i]);
1377 if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
1378 current->is32&=~(1LL<<rt1[i]);
1379 if(!((current->uu>>rt1[i])&1)||get_reg(current->regmap,rt1[i]|64)>=0) {
1380 // TODO: Could preserve the 32-bit flag if the immediate is zero
1381 alloc_reg64(current,i,rt1[i]);
1382 alloc_reg64(current,i,rs1[i]);
1384 clear_const(current,rs1[i]);
1385 clear_const(current,rt1[i]);
1387 else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
1388 if((~current->is32>>rs1[i])&1) alloc_reg64(current,i,rs1[i]);
1389 current->is32|=1LL<<rt1[i];
1390 clear_const(current,rs1[i]);
1391 clear_const(current,rt1[i]);
1393 else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
1394 if(((~current->is32>>rs1[i])&1)&&opcode[i]>0x0c) {
1395 if(rs1[i]!=rt1[i]) {
1396 if(needed_again(rs1[i],i)) alloc_reg64(current,i,rs1[i]);
1397 alloc_reg64(current,i,rt1[i]);
1398 current->is32&=~(1LL<<rt1[i]);
1401 else current->is32|=1LL<<rt1[i]; // ANDI clears upper bits
1402 if(is_const(current,rs1[i])) {
1403 int v=get_const(current,rs1[i]);
1404 if(opcode[i]==0x0c) set_const(current,rt1[i],v&imm[i]);
1405 if(opcode[i]==0x0d) set_const(current,rt1[i],v|imm[i]);
1406 if(opcode[i]==0x0e) set_const(current,rt1[i],v^imm[i]);
1408 else clear_const(current,rt1[i]);
1410 else if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
1411 if(is_const(current,rs1[i])) {
1412 int v=get_const(current,rs1[i]);
1413 set_const(current,rt1[i],v+imm[i]);
1415 else clear_const(current,rt1[i]);
1416 current->is32|=1LL<<rt1[i];
1419 set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
1420 current->is32|=1LL<<rt1[i];
1422 dirty_reg(current,rt1[i]);
1425 void load_alloc(struct regstat *current,int i)
1427 clear_const(current,rt1[i]);
1428 //if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
1429 if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
1430 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1431 if(rt1[i]&&!((current->u>>rt1[i])&1)) {
1432 alloc_reg(current,i,rt1[i]);
1433 assert(get_reg(current->regmap,rt1[i])>=0);
1434 if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
1436 current->is32&=~(1LL<<rt1[i]);
1437 alloc_reg64(current,i,rt1[i]);
1439 else if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1441 current->is32&=~(1LL<<rt1[i]);
1442 alloc_reg64(current,i,rt1[i]);
1443 alloc_all(current,i);
1444 alloc_reg64(current,i,FTEMP);
1445 minimum_free_regs[i]=HOST_REGS;
1447 else current->is32|=1LL<<rt1[i];
1448 dirty_reg(current,rt1[i]);
1449 // LWL/LWR need a temporary register for the old value
1450 if(opcode[i]==0x22||opcode[i]==0x26)
1452 alloc_reg(current,i,FTEMP);
1453 alloc_reg_temp(current,i,-1);
1454 minimum_free_regs[i]=1;
1459 // Load to r0 or unneeded register (dummy load)
1460 // but we still need a register to calculate the address
1461 if(opcode[i]==0x22||opcode[i]==0x26)
1463 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1465 alloc_reg_temp(current,i,-1);
1466 minimum_free_regs[i]=1;
1467 if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1469 alloc_all(current,i);
1470 alloc_reg64(current,i,FTEMP);
1471 minimum_free_regs[i]=HOST_REGS;
1476 void store_alloc(struct regstat *current,int i)
1478 clear_const(current,rs2[i]);
1479 if(!(rs2[i])) current->u&=~1LL; // Allow allocating r0 if necessary
1480 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1481 alloc_reg(current,i,rs2[i]);
1482 if(opcode[i]==0x2c||opcode[i]==0x2d||opcode[i]==0x3f) { // 64-bit SDL/SDR/SD
1483 alloc_reg64(current,i,rs2[i]);
1484 if(rs2[i]) alloc_reg(current,i,FTEMP);
1486 #if defined(HOST_IMM8)
1487 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1488 else alloc_reg(current,i,INVCP);
1490 if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) { // SWL/SWL/SDL/SDR
1491 alloc_reg(current,i,FTEMP);
1493 // We need a temporary register for address generation
1494 alloc_reg_temp(current,i,-1);
1495 minimum_free_regs[i]=1;
1498 void c1ls_alloc(struct regstat *current,int i)
1500 //clear_const(current,rs1[i]); // FIXME
1501 clear_const(current,rt1[i]);
1502 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1503 alloc_reg(current,i,CSREG); // Status
1504 alloc_reg(current,i,FTEMP);
1505 if(opcode[i]==0x35||opcode[i]==0x3d) { // 64-bit LDC1/SDC1
1506 alloc_reg64(current,i,FTEMP);
1508 #if defined(HOST_IMM8)
1509 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1510 else if((opcode[i]&0x3b)==0x39) // SWC1/SDC1
1511 alloc_reg(current,i,INVCP);
1513 // We need a temporary register for address generation
1514 alloc_reg_temp(current,i,-1);
1517 void c2ls_alloc(struct regstat *current,int i)
1519 clear_const(current,rt1[i]);
1520 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1521 alloc_reg(current,i,FTEMP);
1522 #if defined(HOST_IMM8)
1523 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1524 if((opcode[i]&0x3b)==0x3a) // SWC2/SDC2
1525 alloc_reg(current,i,INVCP);
1527 // We need a temporary register for address generation
1528 alloc_reg_temp(current,i,-1);
1529 minimum_free_regs[i]=1;
1532 #ifndef multdiv_alloc
1533 void multdiv_alloc(struct regstat *current,int i)
1540 // case 0x1D: DMULTU
1543 clear_const(current,rs1[i]);
1544 clear_const(current,rs2[i]);
1547 if((opcode2[i]&4)==0) // 32-bit
1549 current->u&=~(1LL<<HIREG);
1550 current->u&=~(1LL<<LOREG);
1551 alloc_reg(current,i,HIREG);
1552 alloc_reg(current,i,LOREG);
1553 alloc_reg(current,i,rs1[i]);
1554 alloc_reg(current,i,rs2[i]);
1555 current->is32|=1LL<<HIREG;
1556 current->is32|=1LL<<LOREG;
1557 dirty_reg(current,HIREG);
1558 dirty_reg(current,LOREG);
1562 current->u&=~(1LL<<HIREG);
1563 current->u&=~(1LL<<LOREG);
1564 current->uu&=~(1LL<<HIREG);
1565 current->uu&=~(1LL<<LOREG);
1566 alloc_reg64(current,i,HIREG);
1567 //if(HOST_REGS>10) alloc_reg64(current,i,LOREG);
1568 alloc_reg64(current,i,rs1[i]);
1569 alloc_reg64(current,i,rs2[i]);
1570 alloc_all(current,i);
1571 current->is32&=~(1LL<<HIREG);
1572 current->is32&=~(1LL<<LOREG);
1573 dirty_reg(current,HIREG);
1574 dirty_reg(current,LOREG);
1575 minimum_free_regs[i]=HOST_REGS;
1580 // Multiply by zero is zero.
1581 // MIPS does not have a divide by zero exception.
1582 // The result is undefined, we return zero.
1583 alloc_reg(current,i,HIREG);
1584 alloc_reg(current,i,LOREG);
1585 current->is32|=1LL<<HIREG;
1586 current->is32|=1LL<<LOREG;
1587 dirty_reg(current,HIREG);
1588 dirty_reg(current,LOREG);
1593 void cop0_alloc(struct regstat *current,int i)
1595 if(opcode2[i]==0) // MFC0
1598 clear_const(current,rt1[i]);
1599 alloc_all(current,i);
1600 alloc_reg(current,i,rt1[i]);
1601 current->is32|=1LL<<rt1[i];
1602 dirty_reg(current,rt1[i]);
1605 else if(opcode2[i]==4) // MTC0
1608 clear_const(current,rs1[i]);
1609 alloc_reg(current,i,rs1[i]);
1610 alloc_all(current,i);
1613 alloc_all(current,i); // FIXME: Keep r0
1615 alloc_reg(current,i,0);
1620 // TLBR/TLBWI/TLBWR/TLBP/ERET
1621 assert(opcode2[i]==0x10);
1622 alloc_all(current,i);
1624 minimum_free_regs[i]=HOST_REGS;
1627 void cop1_alloc(struct regstat *current,int i)
1629 alloc_reg(current,i,CSREG); // Load status
1630 if(opcode2[i]<3) // MFC1/DMFC1/CFC1
1633 clear_const(current,rt1[i]);
1635 alloc_reg64(current,i,rt1[i]); // DMFC1
1636 current->is32&=~(1LL<<rt1[i]);
1638 alloc_reg(current,i,rt1[i]); // MFC1/CFC1
1639 current->is32|=1LL<<rt1[i];
1641 dirty_reg(current,rt1[i]);
1643 alloc_reg_temp(current,i,-1);
1645 else if(opcode2[i]>3) // MTC1/DMTC1/CTC1
1648 clear_const(current,rs1[i]);
1650 alloc_reg64(current,i,rs1[i]); // DMTC1
1652 alloc_reg(current,i,rs1[i]); // MTC1/CTC1
1653 alloc_reg_temp(current,i,-1);
1657 alloc_reg(current,i,0);
1658 alloc_reg_temp(current,i,-1);
1661 minimum_free_regs[i]=1;
1663 void fconv_alloc(struct regstat *current,int i)
1665 alloc_reg(current,i,CSREG); // Load status
1666 alloc_reg_temp(current,i,-1);
1667 minimum_free_regs[i]=1;
1669 void float_alloc(struct regstat *current,int i)
1671 alloc_reg(current,i,CSREG); // Load status
1672 alloc_reg_temp(current,i,-1);
1673 minimum_free_regs[i]=1;
1675 void c2op_alloc(struct regstat *current,int i)
1677 alloc_reg_temp(current,i,-1);
1679 void fcomp_alloc(struct regstat *current,int i)
1681 alloc_reg(current,i,CSREG); // Load status
1682 alloc_reg(current,i,FSREG); // Load flags
1683 dirty_reg(current,FSREG); // Flag will be modified
1684 alloc_reg_temp(current,i,-1);
1685 minimum_free_regs[i]=1;
1688 void syscall_alloc(struct regstat *current,int i)
1690 alloc_cc(current,i);
1691 dirty_reg(current,CCREG);
1692 alloc_all(current,i);
1693 minimum_free_regs[i]=HOST_REGS;
1697 void delayslot_alloc(struct regstat *current,int i)
1708 assem_debug("jump in the delay slot. this shouldn't happen.\n");//exit(1);
1709 SysPrintf("Disabled speculative precompilation\n");
1713 imm16_alloc(current,i);
1717 load_alloc(current,i);
1721 store_alloc(current,i);
1724 alu_alloc(current,i);
1727 shift_alloc(current,i);
1730 multdiv_alloc(current,i);
1733 shiftimm_alloc(current,i);
1736 mov_alloc(current,i);
1739 cop0_alloc(current,i);
1743 cop1_alloc(current,i);
1746 c1ls_alloc(current,i);
1749 c2ls_alloc(current,i);
1752 fconv_alloc(current,i);
1755 float_alloc(current,i);
1758 fcomp_alloc(current,i);
1761 c2op_alloc(current,i);
1766 // Special case where a branch and delay slot span two pages in virtual memory
1767 static void pagespan_alloc(struct regstat *current,int i)
1770 current->wasconst=0;
1772 minimum_free_regs[i]=HOST_REGS;
1773 alloc_all(current,i);
1774 alloc_cc(current,i);
1775 dirty_reg(current,CCREG);
1776 if(opcode[i]==3) // JAL
1778 alloc_reg(current,i,31);
1779 dirty_reg(current,31);
1781 if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
1783 alloc_reg(current,i,rs1[i]);
1785 alloc_reg(current,i,rt1[i]);
1786 dirty_reg(current,rt1[i]);
1789 if((opcode[i]&0x2E)==4) // BEQ/BNE/BEQL/BNEL
1791 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1792 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1793 if(!((current->is32>>rs1[i])&(current->is32>>rs2[i])&1))
1795 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1796 if(rs2[i]) alloc_reg64(current,i,rs2[i]);
1800 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
1802 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1803 if(!((current->is32>>rs1[i])&1))
1805 if(rs1[i]) alloc_reg64(current,i,rs1[i]);
1809 if(opcode[i]==0x11) // BC1
1811 alloc_reg(current,i,FSREG);
1812 alloc_reg(current,i,CSREG);
1817 static void add_stub(int type,int addr,int retaddr,int a,int b,int c,int d,int e)
1819 stubs[stubcount][0]=type;
1820 stubs[stubcount][1]=addr;
1821 stubs[stubcount][2]=retaddr;
1822 stubs[stubcount][3]=a;
1823 stubs[stubcount][4]=b;
1824 stubs[stubcount][5]=c;
1825 stubs[stubcount][6]=d;
1826 stubs[stubcount][7]=e;
1830 // Write out a single register
1831 void wb_register(signed char r,signed char regmap[],uint64_t dirty,uint64_t is32)
1834 for(hr=0;hr<HOST_REGS;hr++) {
1835 if(hr!=EXCLUDE_REG) {
1836 if((regmap[hr]&63)==r) {
1839 emit_storereg(r,hr);
1841 emit_storereg(r|64,hr);
1851 //if(!tracedebug) return 0;
1854 for(i=0;i<2097152;i++) {
1855 unsigned int temp=sum;
1858 sum^=((u_int *)rdram)[i];
1867 sum^=((u_int *)reg)[i];
1875 printf("r%d:%8x%8x ",i,((int *)(reg+i))[1],((int *)(reg+i))[0]);
1884 void memdebug(int i)
1886 //printf("TRACE: count=%d next=%d (checksum %x) lo=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[LOREG]>>32),(int)reg[LOREG]);
1887 //printf("TRACE: count=%d next=%d (rchecksum %x)\n",Count,next_interupt,rchecksum());
1890 //if(Count>=-2084597794) {
1891 if((signed int)Count>=-2084597794&&(signed int)Count<0) {
1893 printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
1894 //printf("TRACE: count=%d next=%d (checksum %x) Status=%x\n",Count,next_interupt,mchecksum(),Status);
1895 //printf("TRACE: count=%d next=%d (checksum %x) hi=%8x%8x\n",Count,next_interupt,mchecksum(),(int)(reg[HIREG]>>32),(int)reg[HIREG]);
1898 printf("TRACE: %x\n",(&i)[-1]);
1902 printf("TRACE: %x \n",(&j)[10]);
1903 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]);
1907 //printf("TRACE: %x\n",(&i)[-1]);
1910 void alu_assemble(int i,struct regstat *i_regs)
1912 if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
1914 signed char s1,s2,t;
1915 t=get_reg(i_regs->regmap,rt1[i]);
1917 s1=get_reg(i_regs->regmap,rs1[i]);
1918 s2=get_reg(i_regs->regmap,rs2[i]);
1919 if(rs1[i]&&rs2[i]) {
1922 if(opcode2[i]&2) emit_sub(s1,s2,t);
1923 else emit_add(s1,s2,t);
1926 if(s1>=0) emit_mov(s1,t);
1927 else emit_loadreg(rs1[i],t);
1931 if(opcode2[i]&2) emit_neg(s2,t);
1932 else emit_mov(s2,t);
1935 emit_loadreg(rs2[i],t);
1936 if(opcode2[i]&2) emit_neg(t,t);
1939 else emit_zeroreg(t);
1943 if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1945 signed char s1l,s2l,s1h,s2h,tl,th;
1946 tl=get_reg(i_regs->regmap,rt1[i]);
1947 th=get_reg(i_regs->regmap,rt1[i]|64);
1949 s1l=get_reg(i_regs->regmap,rs1[i]);
1950 s2l=get_reg(i_regs->regmap,rs2[i]);
1951 s1h=get_reg(i_regs->regmap,rs1[i]|64);
1952 s2h=get_reg(i_regs->regmap,rs2[i]|64);
1953 if(rs1[i]&&rs2[i]) {
1956 if(opcode2[i]&2) emit_subs(s1l,s2l,tl);
1957 else emit_adds(s1l,s2l,tl);
1959 #ifdef INVERTED_CARRY
1960 if(opcode2[i]&2) {if(s1h!=th) emit_mov(s1h,th);emit_sbb(th,s2h);}
1962 if(opcode2[i]&2) emit_sbc(s1h,s2h,th);
1964 else emit_add(s1h,s2h,th);
1968 if(s1l>=0) emit_mov(s1l,tl);
1969 else emit_loadreg(rs1[i],tl);
1971 if(s1h>=0) emit_mov(s1h,th);
1972 else emit_loadreg(rs1[i]|64,th);
1977 if(opcode2[i]&2) emit_negs(s2l,tl);
1978 else emit_mov(s2l,tl);
1981 emit_loadreg(rs2[i],tl);
1982 if(opcode2[i]&2) emit_negs(tl,tl);
1985 #ifdef INVERTED_CARRY
1986 if(s2h>=0) emit_mov(s2h,th);
1987 else emit_loadreg(rs2[i]|64,th);
1989 emit_adcimm(-1,th); // x86 has inverted carry flag
1994 if(s2h>=0) emit_rscimm(s2h,0,th);
1996 emit_loadreg(rs2[i]|64,th);
1997 emit_rscimm(th,0,th);
2000 if(s2h>=0) emit_mov(s2h,th);
2001 else emit_loadreg(rs2[i]|64,th);
2008 if(th>=0) emit_zeroreg(th);
2013 if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
2015 signed char s1l,s1h,s2l,s2h,t;
2016 if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1))
2018 t=get_reg(i_regs->regmap,rt1[i]);
2021 s1l=get_reg(i_regs->regmap,rs1[i]);
2022 s1h=get_reg(i_regs->regmap,rs1[i]|64);
2023 s2l=get_reg(i_regs->regmap,rs2[i]);
2024 s2h=get_reg(i_regs->regmap,rs2[i]|64);
2025 if(rs2[i]==0) // rx<r0
2028 if(opcode2[i]==0x2a) // SLT
2029 emit_shrimm(s1h,31,t);
2030 else // SLTU (unsigned can not be less than zero)
2033 else if(rs1[i]==0) // r0<rx
2036 if(opcode2[i]==0x2a) // SLT
2037 emit_set_gz64_32(s2h,s2l,t);
2038 else // SLTU (set if not zero)
2039 emit_set_nz64_32(s2h,s2l,t);
2042 assert(s1l>=0);assert(s1h>=0);
2043 assert(s2l>=0);assert(s2h>=0);
2044 if(opcode2[i]==0x2a) // SLT
2045 emit_set_if_less64_32(s1h,s1l,s2h,s2l,t);
2047 emit_set_if_carry64_32(s1h,s1l,s2h,s2l,t);
2051 t=get_reg(i_regs->regmap,rt1[i]);
2054 s1l=get_reg(i_regs->regmap,rs1[i]);
2055 s2l=get_reg(i_regs->regmap,rs2[i]);
2056 if(rs2[i]==0) // rx<r0
2059 if(opcode2[i]==0x2a) // SLT
2060 emit_shrimm(s1l,31,t);
2061 else // SLTU (unsigned can not be less than zero)
2064 else if(rs1[i]==0) // r0<rx
2067 if(opcode2[i]==0x2a) // SLT
2068 emit_set_gz32(s2l,t);
2069 else // SLTU (set if not zero)
2070 emit_set_nz32(s2l,t);
2073 assert(s1l>=0);assert(s2l>=0);
2074 if(opcode2[i]==0x2a) // SLT
2075 emit_set_if_less32(s1l,s2l,t);
2077 emit_set_if_carry32(s1l,s2l,t);
2083 if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
2085 signed char s1l,s1h,s2l,s2h,th,tl;
2086 tl=get_reg(i_regs->regmap,rt1[i]);
2087 th=get_reg(i_regs->regmap,rt1[i]|64);
2088 if(!((i_regs->was32>>rs1[i])&(i_regs->was32>>rs2[i])&1)&&th>=0)
2092 s1l=get_reg(i_regs->regmap,rs1[i]);
2093 s1h=get_reg(i_regs->regmap,rs1[i]|64);
2094 s2l=get_reg(i_regs->regmap,rs2[i]);
2095 s2h=get_reg(i_regs->regmap,rs2[i]|64);
2096 if(rs1[i]&&rs2[i]) {
2097 assert(s1l>=0);assert(s1h>=0);
2098 assert(s2l>=0);assert(s2h>=0);
2099 if(opcode2[i]==0x24) { // AND
2100 emit_and(s1l,s2l,tl);
2101 emit_and(s1h,s2h,th);
2103 if(opcode2[i]==0x25) { // OR
2104 emit_or(s1l,s2l,tl);
2105 emit_or(s1h,s2h,th);
2107 if(opcode2[i]==0x26) { // XOR
2108 emit_xor(s1l,s2l,tl);
2109 emit_xor(s1h,s2h,th);
2111 if(opcode2[i]==0x27) { // NOR
2112 emit_or(s1l,s2l,tl);
2113 emit_or(s1h,s2h,th);
2120 if(opcode2[i]==0x24) { // AND
2124 if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2126 if(s1l>=0) emit_mov(s1l,tl);
2127 else emit_loadreg(rs1[i],tl);
2128 if(s1h>=0) emit_mov(s1h,th);
2129 else emit_loadreg(rs1[i]|64,th);
2133 if(s2l>=0) emit_mov(s2l,tl);
2134 else emit_loadreg(rs2[i],tl);
2135 if(s2h>=0) emit_mov(s2h,th);
2136 else emit_loadreg(rs2[i]|64,th);
2143 if(opcode2[i]==0x27) { // NOR
2145 if(s1l>=0) emit_not(s1l,tl);
2147 emit_loadreg(rs1[i],tl);
2150 if(s1h>=0) emit_not(s1h,th);
2152 emit_loadreg(rs1[i]|64,th);
2158 if(s2l>=0) emit_not(s2l,tl);
2160 emit_loadreg(rs2[i],tl);
2163 if(s2h>=0) emit_not(s2h,th);
2165 emit_loadreg(rs2[i]|64,th);
2181 s1l=get_reg(i_regs->regmap,rs1[i]);
2182 s2l=get_reg(i_regs->regmap,rs2[i]);
2183 if(rs1[i]&&rs2[i]) {
2186 if(opcode2[i]==0x24) { // AND
2187 emit_and(s1l,s2l,tl);
2189 if(opcode2[i]==0x25) { // OR
2190 emit_or(s1l,s2l,tl);
2192 if(opcode2[i]==0x26) { // XOR
2193 emit_xor(s1l,s2l,tl);
2195 if(opcode2[i]==0x27) { // NOR
2196 emit_or(s1l,s2l,tl);
2202 if(opcode2[i]==0x24) { // AND
2205 if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2207 if(s1l>=0) emit_mov(s1l,tl);
2208 else emit_loadreg(rs1[i],tl); // CHECK: regmap_entry?
2212 if(s2l>=0) emit_mov(s2l,tl);
2213 else emit_loadreg(rs2[i],tl); // CHECK: regmap_entry?
2215 else emit_zeroreg(tl);
2217 if(opcode2[i]==0x27) { // NOR
2219 if(s1l>=0) emit_not(s1l,tl);
2221 emit_loadreg(rs1[i],tl);
2227 if(s2l>=0) emit_not(s2l,tl);
2229 emit_loadreg(rs2[i],tl);
2233 else emit_movimm(-1,tl);
2242 void imm16_assemble(int i,struct regstat *i_regs)
2244 if (opcode[i]==0x0f) { // LUI
2247 t=get_reg(i_regs->regmap,rt1[i]);
2250 if(!((i_regs->isconst>>t)&1))
2251 emit_movimm(imm[i]<<16,t);
2255 if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
2258 t=get_reg(i_regs->regmap,rt1[i]);
2259 s=get_reg(i_regs->regmap,rs1[i]);
2264 if(!((i_regs->isconst>>t)&1)) {
2266 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2267 emit_addimm(t,imm[i],t);
2269 if(!((i_regs->wasconst>>s)&1))
2270 emit_addimm(s,imm[i],t);
2272 emit_movimm(constmap[i][s]+imm[i],t);
2278 if(!((i_regs->isconst>>t)&1))
2279 emit_movimm(imm[i],t);
2284 if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
2286 signed char sh,sl,th,tl;
2287 th=get_reg(i_regs->regmap,rt1[i]|64);
2288 tl=get_reg(i_regs->regmap,rt1[i]);
2289 sh=get_reg(i_regs->regmap,rs1[i]|64);
2290 sl=get_reg(i_regs->regmap,rs1[i]);
2296 emit_addimm64_32(sh,sl,imm[i],th,tl);
2299 emit_addimm(sl,imm[i],tl);
2302 emit_movimm(imm[i],tl);
2303 if(th>=0) emit_movimm(((signed int)imm[i])>>31,th);
2308 else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
2310 //assert(rs1[i]!=0); // r0 might be valid, but it's probably a bug
2311 signed char sh,sl,t;
2312 t=get_reg(i_regs->regmap,rt1[i]);
2313 sh=get_reg(i_regs->regmap,rs1[i]|64);
2314 sl=get_reg(i_regs->regmap,rs1[i]);
2318 if(sh<0) assert((i_regs->was32>>rs1[i])&1);
2319 if(sh<0||((i_regs->was32>>rs1[i])&1)) {
2320 if(opcode[i]==0x0a) { // SLTI
2322 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2323 emit_slti32(t,imm[i],t);
2325 emit_slti32(sl,imm[i],t);
2330 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2331 emit_sltiu32(t,imm[i],t);
2333 emit_sltiu32(sl,imm[i],t);
2338 if(opcode[i]==0x0a) // SLTI
2339 emit_slti64_32(sh,sl,imm[i],t);
2341 emit_sltiu64_32(sh,sl,imm[i],t);
2344 // SLTI(U) with r0 is just stupid,
2345 // nonetheless examples can be found
2346 if(opcode[i]==0x0a) // SLTI
2347 if(0<imm[i]) emit_movimm(1,t);
2348 else emit_zeroreg(t);
2351 if(imm[i]) emit_movimm(1,t);
2352 else emit_zeroreg(t);
2358 else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
2360 signed char sh,sl,th,tl;
2361 th=get_reg(i_regs->regmap,rt1[i]|64);
2362 tl=get_reg(i_regs->regmap,rt1[i]);
2363 sh=get_reg(i_regs->regmap,rs1[i]|64);
2364 sl=get_reg(i_regs->regmap,rs1[i]);
2365 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2366 if(opcode[i]==0x0c) //ANDI
2370 if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2371 emit_andimm(tl,imm[i],tl);
2373 if(!((i_regs->wasconst>>sl)&1))
2374 emit_andimm(sl,imm[i],tl);
2376 emit_movimm(constmap[i][sl]&imm[i],tl);
2381 if(th>=0) emit_zeroreg(th);
2387 if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2391 emit_loadreg(rs1[i]|64,th);
2396 if(opcode[i]==0x0d) { // ORI
2398 emit_orimm(tl,imm[i],tl);
2400 if(!((i_regs->wasconst>>sl)&1))
2401 emit_orimm(sl,imm[i],tl);
2403 emit_movimm(constmap[i][sl]|imm[i],tl);
2406 if(opcode[i]==0x0e) { // XORI
2408 emit_xorimm(tl,imm[i],tl);
2410 if(!((i_regs->wasconst>>sl)&1))
2411 emit_xorimm(sl,imm[i],tl);
2413 emit_movimm(constmap[i][sl]^imm[i],tl);
2418 emit_movimm(imm[i],tl);
2419 if(th>=0) emit_zeroreg(th);
2427 void shiftimm_assemble(int i,struct regstat *i_regs)
2429 if(opcode2[i]<=0x3) // SLL/SRL/SRA
2433 t=get_reg(i_regs->regmap,rt1[i]);
2434 s=get_reg(i_regs->regmap,rs1[i]);
2436 if(t>=0&&!((i_regs->isconst>>t)&1)){
2443 if(s<0&&i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2445 if(opcode2[i]==0) // SLL
2447 emit_shlimm(s<0?t:s,imm[i],t);
2449 if(opcode2[i]==2) // SRL
2451 emit_shrimm(s<0?t:s,imm[i],t);
2453 if(opcode2[i]==3) // SRA
2455 emit_sarimm(s<0?t:s,imm[i],t);
2459 if(s>=0 && s!=t) emit_mov(s,t);
2463 //emit_storereg(rt1[i],t); //DEBUG
2466 if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
2469 signed char sh,sl,th,tl;
2470 th=get_reg(i_regs->regmap,rt1[i]|64);
2471 tl=get_reg(i_regs->regmap,rt1[i]);
2472 sh=get_reg(i_regs->regmap,rs1[i]|64);
2473 sl=get_reg(i_regs->regmap,rs1[i]);
2478 if(th>=0) emit_zeroreg(th);
2485 if(opcode2[i]==0x38) // DSLL
2487 if(th>=0) emit_shldimm(sh,sl,imm[i],th);
2488 emit_shlimm(sl,imm[i],tl);
2490 if(opcode2[i]==0x3a) // DSRL
2492 emit_shrdimm(sl,sh,imm[i],tl);
2493 if(th>=0) emit_shrimm(sh,imm[i],th);
2495 if(opcode2[i]==0x3b) // DSRA
2497 emit_shrdimm(sl,sh,imm[i],tl);
2498 if(th>=0) emit_sarimm(sh,imm[i],th);
2502 if(sl!=tl) emit_mov(sl,tl);
2503 if(th>=0&&sh!=th) emit_mov(sh,th);
2509 if(opcode2[i]==0x3c) // DSLL32
2512 signed char sl,tl,th;
2513 tl=get_reg(i_regs->regmap,rt1[i]);
2514 th=get_reg(i_regs->regmap,rt1[i]|64);
2515 sl=get_reg(i_regs->regmap,rs1[i]);
2524 emit_shlimm(th,imm[i]&31,th);
2529 if(opcode2[i]==0x3e) // DSRL32
2532 signed char sh,tl,th;
2533 tl=get_reg(i_regs->regmap,rt1[i]);
2534 th=get_reg(i_regs->regmap,rt1[i]|64);
2535 sh=get_reg(i_regs->regmap,rs1[i]|64);
2539 if(th>=0) emit_zeroreg(th);
2542 emit_shrimm(tl,imm[i]&31,tl);
2547 if(opcode2[i]==0x3f) // DSRA32
2551 tl=get_reg(i_regs->regmap,rt1[i]);
2552 sh=get_reg(i_regs->regmap,rs1[i]|64);
2558 emit_sarimm(tl,imm[i]&31,tl);
2565 #ifndef shift_assemble
2566 void shift_assemble(int i,struct regstat *i_regs)
2568 printf("Need shift_assemble for this architecture.\n");
2573 void load_assemble(int i,struct regstat *i_regs)
2575 int s,th,tl,addr,map=-1;
2578 int memtarget=0,c=0;
2579 int fastload_reg_override=0;
2581 th=get_reg(i_regs->regmap,rt1[i]|64);
2582 tl=get_reg(i_regs->regmap,rt1[i]);
2583 s=get_reg(i_regs->regmap,rs1[i]);
2585 for(hr=0;hr<HOST_REGS;hr++) {
2586 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2588 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2590 c=(i_regs->wasconst>>s)&1;
2592 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2595 //printf("load_assemble: c=%d\n",c);
2596 //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2597 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2598 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2600 // could be FIFO, must perform the read
2602 assem_debug("(forced read)\n");
2603 tl=get_reg(i_regs->regmap,-1);
2606 if(offset||s<0||c) addr=tl;
2608 //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2610 //printf("load_assemble: c=%d\n",c);
2611 //if(c) printf("load_assemble: const=%x\n",(int)constmap[i][s]+offset);
2612 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2614 if(th>=0) reglist&=~(1<<th);
2617 map=get_reg(i_regs->regmap,ROREG);
2618 if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG);
2621 // Strmnnrmn's speed hack
2622 if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2625 jaddr=emit_fastpath_cmp_jump(i,addr,&fastload_reg_override);
2628 else if(ram_offset&&memtarget) {
2629 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2630 fastload_reg_override=HOST_TEMPREG;
2632 int dummy=(rt1[i]==0)||(tl!=get_reg(i_regs->regmap,rt1[i])); // ignore loads to r0 and unneeded reg
2633 if (opcode[i]==0x20) { // LB
2636 #ifdef HOST_IMM_ADDR32
2638 emit_movsbl_tlb((constmap[i][s]+offset)^3,map,tl);
2642 //emit_xorimm(addr,3,tl);
2643 //emit_movsbl_indexed((int)rdram-0x80000000,tl,tl);
2645 #ifdef BIG_ENDIAN_MIPS
2646 if(!c) emit_xorimm(addr,3,tl);
2647 else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2651 if(fastload_reg_override) a=fastload_reg_override;
2653 emit_movsbl_indexed_tlb(x,a,map,tl);
2657 add_stub(LOADB_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2660 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2662 if (opcode[i]==0x21) { // LH
2665 #ifdef HOST_IMM_ADDR32
2667 emit_movswl_tlb((constmap[i][s]+offset)^2,map,tl);
2672 #ifdef BIG_ENDIAN_MIPS
2673 if(!c) emit_xorimm(addr,2,tl);
2674 else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2678 if(fastload_reg_override) a=fastload_reg_override;
2680 //emit_movswl_indexed_tlb(x,tl,map,tl);
2683 emit_movswl_indexed(x,a,tl);
2685 #if 1 //def RAM_OFFSET
2686 emit_movswl_indexed(x,a,tl);
2688 emit_movswl_indexed((int)rdram-0x80000000+x,a,tl);
2694 add_stub(LOADH_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2697 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2699 if (opcode[i]==0x23) { // LW
2703 if(fastload_reg_override) a=fastload_reg_override;
2704 //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2705 #ifdef HOST_IMM_ADDR32
2707 emit_readword_tlb(constmap[i][s]+offset,map,tl);
2710 emit_readword_indexed_tlb(0,a,map,tl);
2713 add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2716 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2718 if (opcode[i]==0x24) { // LBU
2721 #ifdef HOST_IMM_ADDR32
2723 emit_movzbl_tlb((constmap[i][s]+offset)^3,map,tl);
2727 //emit_xorimm(addr,3,tl);
2728 //emit_movzbl_indexed((int)rdram-0x80000000,tl,tl);
2730 #ifdef BIG_ENDIAN_MIPS
2731 if(!c) emit_xorimm(addr,3,tl);
2732 else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2736 if(fastload_reg_override) a=fastload_reg_override;
2738 emit_movzbl_indexed_tlb(x,a,map,tl);
2742 add_stub(LOADBU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2745 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2747 if (opcode[i]==0x25) { // LHU
2750 #ifdef HOST_IMM_ADDR32
2752 emit_movzwl_tlb((constmap[i][s]+offset)^2,map,tl);
2757 #ifdef BIG_ENDIAN_MIPS
2758 if(!c) emit_xorimm(addr,2,tl);
2759 else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2763 if(fastload_reg_override) a=fastload_reg_override;
2765 //emit_movzwl_indexed_tlb(x,tl,map,tl);
2768 emit_movzwl_indexed(x,a,tl);
2770 #if 1 //def RAM_OFFSET
2771 emit_movzwl_indexed(x,a,tl);
2773 emit_movzwl_indexed((int)rdram-0x80000000+x,a,tl);
2779 add_stub(LOADHU_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2782 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2784 if (opcode[i]==0x27) { // LWU
2789 if(fastload_reg_override) a=fastload_reg_override;
2790 //emit_readword_indexed((int)rdram-0x80000000,addr,tl);
2791 #ifdef HOST_IMM_ADDR32
2793 emit_readword_tlb(constmap[i][s]+offset,map,tl);
2796 emit_readword_indexed_tlb(0,a,map,tl);
2799 add_stub(LOADW_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2802 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2806 if (opcode[i]==0x37) { // LD
2810 if(fastload_reg_override) a=fastload_reg_override;
2811 //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,addr,th);
2812 //emit_readword_indexed((int)rdram-0x7FFFFFFC,addr,tl);
2813 #ifdef HOST_IMM_ADDR32
2815 emit_readdword_tlb(constmap[i][s]+offset,map,th,tl);
2818 emit_readdword_indexed_tlb(0,a,map,th,tl);
2821 add_stub(LOADD_STUB,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2824 inline_readstub(LOADD_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2827 //emit_storereg(rt1[i],tl); // DEBUG
2828 //if(opcode[i]==0x23)
2829 //if(opcode[i]==0x24)
2830 //if(opcode[i]==0x23||opcode[i]==0x24)
2831 /*if(opcode[i]==0x21||opcode[i]==0x23||opcode[i]==0x24)
2835 emit_readword((int)&last_count,ECX);
2837 if(get_reg(i_regs->regmap,CCREG)<0)
2838 emit_loadreg(CCREG,HOST_CCREG);
2839 emit_add(HOST_CCREG,ECX,HOST_CCREG);
2840 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
2841 emit_writeword(HOST_CCREG,(int)&Count);
2844 if(get_reg(i_regs->regmap,CCREG)<0)
2845 emit_loadreg(CCREG,0);
2847 emit_mov(HOST_CCREG,0);
2849 emit_addimm(0,2*ccadj[i],0);
2850 emit_writeword(0,(int)&Count);
2852 emit_call((int)memdebug);
2854 restore_regs(0x100f);
2858 #ifndef loadlr_assemble
2859 void loadlr_assemble(int i,struct regstat *i_regs)
2861 printf("Need loadlr_assemble for this architecture.\n");
2866 void store_assemble(int i,struct regstat *i_regs)
2872 int memtarget=0,c=0;
2873 int agr=AGEN1+(i&1);
2874 int faststore_reg_override=0;
2876 th=get_reg(i_regs->regmap,rs2[i]|64);
2877 tl=get_reg(i_regs->regmap,rs2[i]);
2878 s=get_reg(i_regs->regmap,rs1[i]);
2879 temp=get_reg(i_regs->regmap,agr);
2880 if(temp<0) temp=get_reg(i_regs->regmap,-1);
2883 c=(i_regs->wasconst>>s)&1;
2885 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2890 for(hr=0;hr<HOST_REGS;hr++) {
2891 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2893 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2894 if(offset||s<0||c) addr=temp;
2897 jaddr=emit_fastpath_cmp_jump(i,addr,&faststore_reg_override);
2899 else if(ram_offset&&memtarget) {
2900 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2901 faststore_reg_override=HOST_TEMPREG;
2904 if (opcode[i]==0x28) { // SB
2907 #ifdef BIG_ENDIAN_MIPS
2908 if(!c) emit_xorimm(addr,3,temp);
2909 else x=((constmap[i][s]+offset)^3)-(constmap[i][s]+offset);
2913 if(faststore_reg_override) a=faststore_reg_override;
2914 //emit_writebyte_indexed(tl,(int)rdram-0x80000000,temp);
2915 emit_writebyte_indexed_tlb(tl,x,a,map,a);
2919 if (opcode[i]==0x29) { // SH
2922 #ifdef BIG_ENDIAN_MIPS
2923 if(!c) emit_xorimm(addr,2,temp);
2924 else x=((constmap[i][s]+offset)^2)-(constmap[i][s]+offset);
2928 if(faststore_reg_override) a=faststore_reg_override;
2930 //emit_writehword_indexed_tlb(tl,x,temp,map,temp);
2933 emit_writehword_indexed(tl,x,a);
2935 //emit_writehword_indexed(tl,(int)rdram-0x80000000+x,a);
2936 emit_writehword_indexed(tl,x,a);
2940 if (opcode[i]==0x2B) { // SW
2943 if(faststore_reg_override) a=faststore_reg_override;
2944 //emit_writeword_indexed(tl,(int)rdram-0x80000000,addr);
2945 emit_writeword_indexed_tlb(tl,0,a,map,temp);
2949 if (opcode[i]==0x3F) { // SD
2952 if(faststore_reg_override) a=faststore_reg_override;
2955 //emit_writeword_indexed(th,(int)rdram-0x80000000,addr);
2956 //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,addr);
2957 emit_writedword_indexed_tlb(th,tl,0,a,map,temp);
2960 //emit_writeword_indexed(tl,(int)rdram-0x80000000,temp);
2961 //emit_writeword_indexed(tl,(int)rdram-0x7FFFFFFC,temp);
2962 emit_writedword_indexed_tlb(tl,tl,0,a,map,temp);
2968 // PCSX store handlers don't check invcode again
2970 add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2973 if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
2975 #ifdef DESTRUCTIVE_SHIFT
2976 // The x86 shift operation is 'destructive'; it overwrites the
2977 // source register, so we need to make a copy first and use that.
2980 #if defined(HOST_IMM8)
2981 int ir=get_reg(i_regs->regmap,INVCP);
2983 emit_cmpmem_indexedsr12_reg(ir,addr,1);
2985 emit_cmpmem_indexedsr12_imm((int)invalid_code,addr,1);
2987 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
2988 emit_callne(invalidate_addr_reg[addr]);
2990 int jaddr2=(int)out;
2992 add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),addr,0,0,0);
2996 u_int addr_val=constmap[i][s]+offset;
2998 add_stub(type,jaddr,(int)out,i,addr,(int)i_regs,ccadj[i],reglist);
2999 } else if(c&&!memtarget) {
3000 inline_writestub(type,i,addr_val,i_regs->regmap,rs2[i],ccadj[i],reglist);
3002 // basic current block modification detection..
3003 // not looking back as that should be in mips cache already
3004 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3005 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3006 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3007 if(i_regs->regmap==regs[i].regmap) {
3008 load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i);
3009 wb_dirtys(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty);
3010 emit_movimm(start+i*4+4,0);
3011 emit_writeword(0,(int)&pcaddr);
3012 emit_jmp((int)do_interrupt);
3015 //if(opcode[i]==0x2B || opcode[i]==0x3F)
3016 //if(opcode[i]==0x2B || opcode[i]==0x28)
3017 //if(opcode[i]==0x2B || opcode[i]==0x29)
3018 //if(opcode[i]==0x2B)
3019 /*if(opcode[i]==0x2B || opcode[i]==0x28 || opcode[i]==0x29 || opcode[i]==0x3F)
3027 emit_readword((int)&last_count,ECX);
3029 if(get_reg(i_regs->regmap,CCREG)<0)
3030 emit_loadreg(CCREG,HOST_CCREG);
3031 emit_add(HOST_CCREG,ECX,HOST_CCREG);
3032 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3033 emit_writeword(HOST_CCREG,(int)&Count);
3036 if(get_reg(i_regs->regmap,CCREG)<0)
3037 emit_loadreg(CCREG,0);
3039 emit_mov(HOST_CCREG,0);
3041 emit_addimm(0,2*ccadj[i],0);
3042 emit_writeword(0,(int)&Count);
3044 emit_call((int)memdebug);
3049 restore_regs(0x100f);
3054 void storelr_assemble(int i,struct regstat *i_regs)
3061 int case1,case2,case3;
3062 int done0,done1,done2;
3063 int memtarget=0,c=0;
3064 int agr=AGEN1+(i&1);
3066 th=get_reg(i_regs->regmap,rs2[i]|64);
3067 tl=get_reg(i_regs->regmap,rs2[i]);
3068 s=get_reg(i_regs->regmap,rs1[i]);
3069 temp=get_reg(i_regs->regmap,agr);
3070 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3073 c=(i_regs->isconst>>s)&1;
3075 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3079 for(hr=0;hr<HOST_REGS;hr++) {
3080 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3084 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3085 if(!offset&&s!=temp) emit_mov(s,temp);
3091 if(!memtarget||!rs1[i]) {
3097 int map=get_reg(i_regs->regmap,ROREG);
3098 if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG);
3100 if((u_int)rdram!=0x80000000)
3101 emit_addimm_no_flags((u_int)rdram-(u_int)0x80000000,temp);
3104 if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
3105 temp2=get_reg(i_regs->regmap,FTEMP);
3106 if(!rs2[i]) temp2=th=tl;
3109 #ifndef BIG_ENDIAN_MIPS
3110 emit_xorimm(temp,3,temp);
3112 emit_testimm(temp,2);
3115 emit_testimm(temp,1);
3119 if (opcode[i]==0x2A) { // SWL
3120 emit_writeword_indexed(tl,0,temp);
3122 if (opcode[i]==0x2E) { // SWR
3123 emit_writebyte_indexed(tl,3,temp);
3125 if (opcode[i]==0x2C) { // SDL
3126 emit_writeword_indexed(th,0,temp);
3127 if(rs2[i]) emit_mov(tl,temp2);
3129 if (opcode[i]==0x2D) { // SDR
3130 emit_writebyte_indexed(tl,3,temp);
3131 if(rs2[i]) emit_shldimm(th,tl,24,temp2);
3136 set_jump_target(case1,(int)out);
3137 if (opcode[i]==0x2A) { // SWL
3138 // Write 3 msb into three least significant bytes
3139 if(rs2[i]) emit_rorimm(tl,8,tl);
3140 emit_writehword_indexed(tl,-1,temp);
3141 if(rs2[i]) emit_rorimm(tl,16,tl);
3142 emit_writebyte_indexed(tl,1,temp);
3143 if(rs2[i]) emit_rorimm(tl,8,tl);
3145 if (opcode[i]==0x2E) { // SWR
3146 // Write two lsb into two most significant bytes
3147 emit_writehword_indexed(tl,1,temp);
3149 if (opcode[i]==0x2C) { // SDL
3150 if(rs2[i]) emit_shrdimm(tl,th,8,temp2);
3151 // Write 3 msb into three least significant bytes
3152 if(rs2[i]) emit_rorimm(th,8,th);
3153 emit_writehword_indexed(th,-1,temp);
3154 if(rs2[i]) emit_rorimm(th,16,th);
3155 emit_writebyte_indexed(th,1,temp);
3156 if(rs2[i]) emit_rorimm(th,8,th);
3158 if (opcode[i]==0x2D) { // SDR
3159 if(rs2[i]) emit_shldimm(th,tl,16,temp2);
3160 // Write two lsb into two most significant bytes
3161 emit_writehword_indexed(tl,1,temp);
3166 set_jump_target(case2,(int)out);
3167 emit_testimm(temp,1);
3170 if (opcode[i]==0x2A) { // SWL
3171 // Write two msb into two least significant bytes
3172 if(rs2[i]) emit_rorimm(tl,16,tl);
3173 emit_writehword_indexed(tl,-2,temp);
3174 if(rs2[i]) emit_rorimm(tl,16,tl);
3176 if (opcode[i]==0x2E) { // SWR
3177 // Write 3 lsb into three most significant bytes
3178 emit_writebyte_indexed(tl,-1,temp);
3179 if(rs2[i]) emit_rorimm(tl,8,tl);
3180 emit_writehword_indexed(tl,0,temp);
3181 if(rs2[i]) emit_rorimm(tl,24,tl);
3183 if (opcode[i]==0x2C) { // SDL
3184 if(rs2[i]) emit_shrdimm(tl,th,16,temp2);
3185 // Write two msb into two least significant bytes
3186 if(rs2[i]) emit_rorimm(th,16,th);
3187 emit_writehword_indexed(th,-2,temp);
3188 if(rs2[i]) emit_rorimm(th,16,th);
3190 if (opcode[i]==0x2D) { // SDR
3191 if(rs2[i]) emit_shldimm(th,tl,8,temp2);
3192 // Write 3 lsb into three most significant bytes
3193 emit_writebyte_indexed(tl,-1,temp);
3194 if(rs2[i]) emit_rorimm(tl,8,tl);
3195 emit_writehword_indexed(tl,0,temp);
3196 if(rs2[i]) emit_rorimm(tl,24,tl);
3201 set_jump_target(case3,(int)out);
3202 if (opcode[i]==0x2A) { // SWL
3203 // Write msb into least significant byte
3204 if(rs2[i]) emit_rorimm(tl,24,tl);
3205 emit_writebyte_indexed(tl,-3,temp);
3206 if(rs2[i]) emit_rorimm(tl,8,tl);
3208 if (opcode[i]==0x2E) { // SWR
3209 // Write entire word
3210 emit_writeword_indexed(tl,-3,temp);
3212 if (opcode[i]==0x2C) { // SDL
3213 if(rs2[i]) emit_shrdimm(tl,th,24,temp2);
3214 // Write msb into least significant byte
3215 if(rs2[i]) emit_rorimm(th,24,th);
3216 emit_writebyte_indexed(th,-3,temp);
3217 if(rs2[i]) emit_rorimm(th,8,th);
3219 if (opcode[i]==0x2D) { // SDR
3220 if(rs2[i]) emit_mov(th,temp2);
3221 // Write entire word
3222 emit_writeword_indexed(tl,-3,temp);
3224 set_jump_target(done0,(int)out);
3225 set_jump_target(done1,(int)out);
3226 set_jump_target(done2,(int)out);
3227 if (opcode[i]==0x2C) { // SDL
3228 emit_testimm(temp,4);
3231 emit_andimm(temp,~3,temp);
3232 emit_writeword_indexed(temp2,4,temp);
3233 set_jump_target(done0,(int)out);
3235 if (opcode[i]==0x2D) { // SDR
3236 emit_testimm(temp,4);
3239 emit_andimm(temp,~3,temp);
3240 emit_writeword_indexed(temp2,-4,temp);
3241 set_jump_target(done0,(int)out);
3244 add_stub(STORELR_STUB,jaddr,(int)out,i,(int)i_regs,temp,ccadj[i],reglist);
3245 if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
3247 int map=get_reg(i_regs->regmap,ROREG);
3248 if(map<0) map=HOST_TEMPREG;
3249 gen_orig_addr_w(temp,map);
3251 emit_addimm_no_flags((u_int)0x80000000-(u_int)rdram,temp);
3253 #if defined(HOST_IMM8)
3254 int ir=get_reg(i_regs->regmap,INVCP);
3256 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3258 emit_cmpmem_indexedsr12_imm((int)invalid_code,temp,1);
3260 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3261 emit_callne(invalidate_addr_reg[temp]);
3263 int jaddr2=(int)out;
3265 add_stub(INVCODE_STUB,jaddr2,(int)out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3270 //save_regs(0x100f);
3271 emit_readword((int)&last_count,ECX);
3272 if(get_reg(i_regs->regmap,CCREG)<0)
3273 emit_loadreg(CCREG,HOST_CCREG);
3274 emit_add(HOST_CCREG,ECX,HOST_CCREG);
3275 emit_addimm(HOST_CCREG,2*ccadj[i],HOST_CCREG);
3276 emit_writeword(HOST_CCREG,(int)&Count);
3277 emit_call((int)memdebug);
3279 //restore_regs(0x100f);
3283 void c1ls_assemble(int i,struct regstat *i_regs)
3285 cop1_unusable(i, i_regs);
3288 void c2ls_assemble(int i,struct regstat *i_regs)
3293 int memtarget=0,c=0;
3295 int agr=AGEN1+(i&1);
3296 int fastio_reg_override=0;
3298 u_int copr=(source[i]>>16)&0x1f;
3299 s=get_reg(i_regs->regmap,rs1[i]);
3300 tl=get_reg(i_regs->regmap,FTEMP);
3305 for(hr=0;hr<HOST_REGS;hr++) {
3306 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3308 if(i_regs->regmap[HOST_CCREG]==CCREG)
3309 reglist&=~(1<<HOST_CCREG);
3312 if (opcode[i]==0x3a) { // SWC2
3313 ar=get_reg(i_regs->regmap,agr);
3314 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3319 if(s>=0) c=(i_regs->wasconst>>s)&1;
3320 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3321 if (!offset&&!c&&s>=0) ar=s;
3324 if (opcode[i]==0x3a) { // SWC2
3325 cop2_get_dreg(copr,tl,HOST_TEMPREG);
3333 emit_jmp(0); // inline_readstub/inline_writestub?
3337 jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
3339 else if(ram_offset&&memtarget) {
3340 emit_addimm(ar,ram_offset,HOST_TEMPREG);
3341 fastio_reg_override=HOST_TEMPREG;
3343 if (opcode[i]==0x32) { // LWC2
3344 #ifdef HOST_IMM_ADDR32
3345 if(c) emit_readword_tlb(constmap[i][s]+offset,-1,tl);
3349 if(fastio_reg_override) a=fastio_reg_override;
3350 emit_readword_indexed(0,a,tl);
3352 if (opcode[i]==0x3a) { // SWC2
3353 #ifdef DESTRUCTIVE_SHIFT
3354 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3357 if(fastio_reg_override) a=fastio_reg_override;
3358 emit_writeword_indexed(tl,0,a);
3362 add_stub(type,jaddr2,(int)out,i,ar,(int)i_regs,ccadj[i],reglist);
3363 if(opcode[i]==0x3a) // SWC2
3364 if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
3365 #if defined(HOST_IMM8)
3366 int ir=get_reg(i_regs->regmap,INVCP);
3368 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3370 emit_cmpmem_indexedsr12_imm((int)invalid_code,ar,1);
3372 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3373 emit_callne(invalidate_addr_reg[ar]);
3375 int jaddr3=(int)out;
3377 add_stub(INVCODE_STUB,jaddr3,(int)out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3380 if (opcode[i]==0x32) { // LWC2
3381 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3385 #ifndef multdiv_assemble
3386 void multdiv_assemble(int i,struct regstat *i_regs)
3388 printf("Need multdiv_assemble for this architecture.\n");
3393 void mov_assemble(int i,struct regstat *i_regs)
3395 //if(opcode2[i]==0x10||opcode2[i]==0x12) { // MFHI/MFLO
3396 //if(opcode2[i]==0x11||opcode2[i]==0x13) { // MTHI/MTLO
3398 signed char sh,sl,th,tl;
3399 th=get_reg(i_regs->regmap,rt1[i]|64);
3400 tl=get_reg(i_regs->regmap,rt1[i]);
3403 sh=get_reg(i_regs->regmap,rs1[i]|64);
3404 sl=get_reg(i_regs->regmap,rs1[i]);
3405 if(sl>=0) emit_mov(sl,tl);
3406 else emit_loadreg(rs1[i],tl);
3408 if(sh>=0) emit_mov(sh,th);
3409 else emit_loadreg(rs1[i]|64,th);
3415 #ifndef fconv_assemble
3416 void fconv_assemble(int i,struct regstat *i_regs)
3418 printf("Need fconv_assemble for this architecture.\n");
3424 void float_assemble(int i,struct regstat *i_regs)
3426 printf("Need float_assemble for this architecture.\n");
3431 void syscall_assemble(int i,struct regstat *i_regs)
3433 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3434 assert(ccreg==HOST_CCREG);
3435 assert(!is_delayslot);
3437 emit_movimm(start+i*4,EAX); // Get PC
3438 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3439 emit_jmp((int)jump_syscall_hle); // XXX
3442 void hlecall_assemble(int i,struct regstat *i_regs)
3444 extern void psxNULL();
3445 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3446 assert(ccreg==HOST_CCREG);
3447 assert(!is_delayslot);
3449 emit_movimm(start+i*4+4,0); // Get PC
3450 uint32_t hleCode = source[i] & 0x03ffffff;
3451 if (hleCode >= (sizeof(psxHLEt) / sizeof(psxHLEt[0])))
3452 emit_movimm((int)psxNULL,1);
3454 emit_movimm((int)psxHLEt[hleCode],1);
3455 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
3456 emit_jmp((int)jump_hlecall);
3459 void intcall_assemble(int i,struct regstat *i_regs)
3461 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3462 assert(ccreg==HOST_CCREG);
3463 assert(!is_delayslot);
3465 emit_movimm(start+i*4,0); // Get PC
3466 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3467 emit_jmp((int)jump_intcall);
3470 void ds_assemble(int i,struct regstat *i_regs)
3472 speculate_register_values(i);
3476 alu_assemble(i,i_regs);break;
3478 imm16_assemble(i,i_regs);break;
3480 shift_assemble(i,i_regs);break;
3482 shiftimm_assemble(i,i_regs);break;
3484 load_assemble(i,i_regs);break;
3486 loadlr_assemble(i,i_regs);break;
3488 store_assemble(i,i_regs);break;
3490 storelr_assemble(i,i_regs);break;
3492 cop0_assemble(i,i_regs);break;
3494 cop1_assemble(i,i_regs);break;
3496 c1ls_assemble(i,i_regs);break;
3498 cop2_assemble(i,i_regs);break;
3500 c2ls_assemble(i,i_regs);break;
3502 c2op_assemble(i,i_regs);break;
3504 fconv_assemble(i,i_regs);break;
3506 float_assemble(i,i_regs);break;
3508 fcomp_assemble(i,i_regs);break;
3510 multdiv_assemble(i,i_regs);break;
3512 mov_assemble(i,i_regs);break;
3522 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
3527 // Is the branch target a valid internal jump?
3528 int internal_branch(uint64_t i_is32,int addr)
3530 if(addr&1) return 0; // Indirect (register) jump
3531 if(addr>=start && addr<start+slen*4-4)
3533 //int t=(addr-start)>>2;
3534 // Delay slots are not valid branch targets
3535 //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;
3536 // 64 -> 32 bit transition requires a recompile
3537 /*if(is32[t]&~unneeded_reg_upper[t]&~i_is32)
3539 if(requires_32bit[t]&~i_is32) printf("optimizable: no\n");
3540 else printf("optimizable: yes\n");
3542 //if(is32[t]&~unneeded_reg_upper[t]&~i_is32) return 0;
3548 #ifndef wb_invalidate
3549 void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t is32,
3550 uint64_t u,uint64_t uu)
3553 for(hr=0;hr<HOST_REGS;hr++) {
3554 if(hr!=EXCLUDE_REG) {
3555 if(pre[hr]!=entry[hr]) {
3558 if(get_reg(entry,pre[hr])<0) {
3560 if(!((u>>pre[hr])&1)) {
3561 emit_storereg(pre[hr],hr);
3562 if( ((is32>>pre[hr])&1) && !((uu>>pre[hr])&1) ) {
3563 emit_sarimm(hr,31,hr);
3564 emit_storereg(pre[hr]|64,hr);
3568 if(!((uu>>(pre[hr]&63))&1) && !((is32>>(pre[hr]&63))&1)) {
3569 emit_storereg(pre[hr],hr);
3578 // Move from one register to another (no writeback)
3579 for(hr=0;hr<HOST_REGS;hr++) {
3580 if(hr!=EXCLUDE_REG) {
3581 if(pre[hr]!=entry[hr]) {
3582 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
3584 if((nr=get_reg(entry,pre[hr]))>=0) {
3594 // Load the specified registers
3595 // This only loads the registers given as arguments because
3596 // we don't want to load things that will be overwritten
3597 void load_regs(signed char entry[],signed char regmap[],int is32,int rs1,int rs2)
3601 for(hr=0;hr<HOST_REGS;hr++) {
3602 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3603 if(entry[hr]!=regmap[hr]) {
3604 if(regmap[hr]==rs1||regmap[hr]==rs2)
3611 emit_loadreg(regmap[hr],hr);
3618 for(hr=0;hr<HOST_REGS;hr++) {
3619 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3620 if(entry[hr]!=regmap[hr]) {
3621 if(regmap[hr]-64==rs1||regmap[hr]-64==rs2)
3623 assert(regmap[hr]!=64);
3624 if((is32>>(regmap[hr]&63))&1) {
3625 int lr=get_reg(regmap,regmap[hr]-64);
3627 emit_sarimm(lr,31,hr);
3629 emit_loadreg(regmap[hr],hr);
3633 emit_loadreg(regmap[hr],hr);
3641 // Load registers prior to the start of a loop
3642 // so that they are not loaded within the loop
3643 static void loop_preload(signed char pre[],signed char entry[])
3646 for(hr=0;hr<HOST_REGS;hr++) {
3647 if(hr!=EXCLUDE_REG) {
3648 if(pre[hr]!=entry[hr]) {
3650 if(get_reg(pre,entry[hr])<0) {
3651 assem_debug("loop preload:\n");
3652 //printf("loop preload: %d\n",hr);
3656 else if(entry[hr]<TEMPREG)
3658 emit_loadreg(entry[hr],hr);
3660 else if(entry[hr]-64<TEMPREG)
3662 emit_loadreg(entry[hr],hr);
3671 // Generate address for load/store instruction
3672 // goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
3673 void address_generation(int i,struct regstat *i_regs,signed char entry[])
3675 if(itype[i]==LOAD||itype[i]==LOADLR||itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS||itype[i]==C2LS) {
3677 int agr=AGEN1+(i&1);
3678 if(itype[i]==LOAD) {
3679 ra=get_reg(i_regs->regmap,rt1[i]);
3680 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3683 if(itype[i]==LOADLR) {
3684 ra=get_reg(i_regs->regmap,FTEMP);
3686 if(itype[i]==STORE||itype[i]==STORELR) {
3687 ra=get_reg(i_regs->regmap,agr);
3688 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3690 if(itype[i]==C1LS||itype[i]==C2LS) {
3691 if ((opcode[i]&0x3b)==0x31||(opcode[i]&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
3692 ra=get_reg(i_regs->regmap,FTEMP);
3693 else { // SWC1/SDC1/SWC2/SDC2
3694 ra=get_reg(i_regs->regmap,agr);
3695 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3698 int rs=get_reg(i_regs->regmap,rs1[i]);
3701 int c=(i_regs->wasconst>>rs)&1;
3703 // Using r0 as a base address
3704 if(!entry||entry[ra]!=agr) {
3705 if (opcode[i]==0x22||opcode[i]==0x26) {
3706 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
3707 }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
3708 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
3710 emit_movimm(offset,ra);
3712 } // else did it in the previous cycle
3715 if(!entry||entry[ra]!=rs1[i])
3716 emit_loadreg(rs1[i],ra);
3717 //if(!entry||entry[ra]!=rs1[i])
3718 // printf("poor load scheduling!\n");
3721 if(rs1[i]!=rt1[i]||itype[i]!=LOAD) {
3722 if(!entry||entry[ra]!=agr) {
3723 if (opcode[i]==0x22||opcode[i]==0x26) {
3724 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
3725 }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
3726 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
3728 #ifdef HOST_IMM_ADDR32
3729 if((itype[i]!=LOAD&&(opcode[i]&0x3b)!=0x31&&(opcode[i]&0x3b)!=0x32)) // LWC1/LDC1/LWC2/LDC2
3731 emit_movimm(constmap[i][rs]+offset,ra);
3732 regs[i].loadedconst|=1<<ra;
3734 } // else did it in the previous cycle
3735 } // else load_consts already did it
3737 if(offset&&!c&&rs1[i]) {
3739 emit_addimm(rs,offset,ra);
3741 emit_addimm(ra,offset,ra);
3746 // Preload constants for next instruction
3747 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) {
3750 agr=AGEN1+((i+1)&1);
3751 ra=get_reg(i_regs->regmap,agr);
3753 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
3754 int offset=imm[i+1];
3755 int c=(regs[i+1].wasconst>>rs)&1;
3756 if(c&&(rs1[i+1]!=rt1[i+1]||itype[i+1]!=LOAD)) {
3757 if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
3758 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
3759 }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
3760 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
3762 #ifdef HOST_IMM_ADDR32
3763 if((itype[i+1]!=LOAD&&(opcode[i+1]&0x3b)!=0x31&&(opcode[i+1]&0x3b)!=0x32)) // LWC1/LDC1/LWC2/LDC2
3765 emit_movimm(constmap[i+1][rs]+offset,ra);
3766 regs[i+1].loadedconst|=1<<ra;
3769 else if(rs1[i+1]==0) {
3770 // Using r0 as a base address
3771 if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
3772 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
3773 }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
3774 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
3776 emit_movimm(offset,ra);
3783 static int get_final_value(int hr, int i, int *value)
3785 int reg=regs[i].regmap[hr];
3787 if(regs[i+1].regmap[hr]!=reg) break;
3788 if(!((regs[i+1].isconst>>hr)&1)) break;
3793 if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP) {
3794 *value=constmap[i][hr];
3798 if(itype[i+1]==UJUMP||itype[i+1]==RJUMP||itype[i+1]==CJUMP||itype[i+1]==SJUMP) {
3799 // Load in delay slot, out-of-order execution
3800 if(itype[i+2]==LOAD&&rs1[i+2]==reg&&rt1[i+2]==reg&&((regs[i+1].wasconst>>hr)&1))
3802 // Precompute load address
3803 *value=constmap[i][hr]+imm[i+2];
3807 if(itype[i+1]==LOAD&&rs1[i+1]==reg&&rt1[i+1]==reg)
3809 // Precompute load address
3810 *value=constmap[i][hr]+imm[i+1];
3811 //printf("c=%x imm=%x\n",(int)constmap[i][hr],imm[i+1]);
3816 *value=constmap[i][hr];
3817 //printf("c=%x\n",(int)constmap[i][hr]);
3818 if(i==slen-1) return 1;
3820 return !((unneeded_reg[i+1]>>reg)&1);
3822 return !((unneeded_reg_upper[i+1]>>reg)&1);
3826 // Load registers with known constants
3827 void load_consts(signed char pre[],signed char regmap[],int is32,int i)
3830 // propagate loaded constant flags
3832 regs[i].loadedconst=0;
3834 for(hr=0;hr<HOST_REGS;hr++) {
3835 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
3836 &®map[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
3838 regs[i].loadedconst|=1<<hr;
3843 for(hr=0;hr<HOST_REGS;hr++) {
3844 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3845 //if(entry[hr]!=regmap[hr]) {
3846 if(!((regs[i].loadedconst>>hr)&1)) {
3847 if(((regs[i].isconst>>hr)&1)&®map[hr]<64&®map[hr]>0) {
3848 int value,similar=0;
3849 if(get_final_value(hr,i,&value)) {
3850 // see if some other register has similar value
3851 for(hr2=0;hr2<HOST_REGS;hr2++) {
3852 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
3853 if(is_similar_value(value,constmap[i][hr2])) {
3861 if(get_final_value(hr2,i,&value2)) // is this needed?
3862 emit_movimm_from(value2,hr2,value,hr);
3864 emit_movimm(value,hr);
3870 emit_movimm(value,hr);
3873 regs[i].loadedconst|=1<<hr;
3879 for(hr=0;hr<HOST_REGS;hr++) {
3880 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3881 //if(entry[hr]!=regmap[hr]) {
3882 if(i==0||!((regs[i-1].isconst>>hr)&1)||pre[hr]!=regmap[hr]||bt[i]) {
3883 if(((regs[i].isconst>>hr)&1)&®map[hr]>64) {
3884 if((is32>>(regmap[hr]&63))&1) {
3885 int lr=get_reg(regmap,regmap[hr]-64);
3887 emit_sarimm(lr,31,hr);
3892 if(get_final_value(hr,i,&value)) {
3897 emit_movimm(value,hr);
3906 void load_all_consts(signed char regmap[],int is32,u_int dirty,int i)
3910 for(hr=0;hr<HOST_REGS;hr++) {
3911 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
3912 if(((regs[i].isconst>>hr)&1)&®map[hr]<64&®map[hr]>0) {
3913 int value=constmap[i][hr];
3918 emit_movimm(value,hr);
3924 for(hr=0;hr<HOST_REGS;hr++) {
3925 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
3926 if(((regs[i].isconst>>hr)&1)&®map[hr]>64) {
3927 if((is32>>(regmap[hr]&63))&1) {
3928 int lr=get_reg(regmap,regmap[hr]-64);
3930 emit_sarimm(lr,31,hr);
3934 int value=constmap[i][hr];
3939 emit_movimm(value,hr);
3947 // Write out all dirty registers (except cycle count)
3948 void wb_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty)
3951 for(hr=0;hr<HOST_REGS;hr++) {
3952 if(hr!=EXCLUDE_REG) {
3953 if(i_regmap[hr]>0) {
3954 if(i_regmap[hr]!=CCREG) {
3955 if((i_dirty>>hr)&1) {
3956 if(i_regmap[hr]<64) {
3957 emit_storereg(i_regmap[hr],hr);
3959 if( !((i_is32>>(i_regmap[hr]&63))&1) ) {
3960 emit_storereg(i_regmap[hr],hr);
3969 // Write out dirty registers that we need to reload (pair with load_needed_regs)
3970 // This writes the registers not written by store_regs_bt
3971 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
3974 int t=(addr-start)>>2;
3975 for(hr=0;hr<HOST_REGS;hr++) {
3976 if(hr!=EXCLUDE_REG) {
3977 if(i_regmap[hr]>0) {
3978 if(i_regmap[hr]!=CCREG) {
3979 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)) {
3980 if((i_dirty>>hr)&1) {
3981 if(i_regmap[hr]<64) {
3982 emit_storereg(i_regmap[hr],hr);
3984 if( !((i_is32>>(i_regmap[hr]&63))&1) ) {
3985 emit_storereg(i_regmap[hr],hr);
3996 // Load all registers (except cycle count)
3997 void load_all_regs(signed char i_regmap[])
4000 for(hr=0;hr<HOST_REGS;hr++) {
4001 if(hr!=EXCLUDE_REG) {
4002 if(i_regmap[hr]==0) {
4006 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4008 emit_loadreg(i_regmap[hr],hr);
4014 // Load all current registers also needed by next instruction
4015 void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
4018 for(hr=0;hr<HOST_REGS;hr++) {
4019 if(hr!=EXCLUDE_REG) {
4020 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4021 if(i_regmap[hr]==0) {
4025 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4027 emit_loadreg(i_regmap[hr],hr);
4034 // Load all regs, storing cycle count if necessary
4035 void load_regs_entry(int t)
4038 if(is_ds[t]) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4039 else if(ccadj[t]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[t]),HOST_CCREG);
4040 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4041 emit_storereg(CCREG,HOST_CCREG);
4044 for(hr=0;hr<HOST_REGS;hr++) {
4045 if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4046 if(regs[t].regmap_entry[hr]==0) {
4049 else if(regs[t].regmap_entry[hr]!=CCREG)
4051 emit_loadreg(regs[t].regmap_entry[hr],hr);
4056 for(hr=0;hr<HOST_REGS;hr++) {
4057 if(regs[t].regmap_entry[hr]>=64&®s[t].regmap_entry[hr]<TEMPREG+64) {
4058 assert(regs[t].regmap_entry[hr]!=64);
4059 if((regs[t].was32>>(regs[t].regmap_entry[hr]&63))&1) {
4060 int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4062 emit_loadreg(regs[t].regmap_entry[hr],hr);
4066 emit_sarimm(lr,31,hr);
4071 emit_loadreg(regs[t].regmap_entry[hr],hr);
4077 // Store dirty registers prior to branch
4078 void store_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4080 if(internal_branch(i_is32,addr))
4082 int t=(addr-start)>>2;
4084 for(hr=0;hr<HOST_REGS;hr++) {
4085 if(hr!=EXCLUDE_REG) {
4086 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4087 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)) {
4088 if((i_dirty>>hr)&1) {
4089 if(i_regmap[hr]<64) {
4090 if(!((unneeded_reg[t]>>i_regmap[hr])&1)) {
4091 emit_storereg(i_regmap[hr],hr);
4092 if( ((i_is32>>i_regmap[hr])&1) && !((unneeded_reg_upper[t]>>i_regmap[hr])&1) ) {
4093 #ifdef DESTRUCTIVE_WRITEBACK
4094 emit_sarimm(hr,31,hr);
4095 emit_storereg(i_regmap[hr]|64,hr);
4097 emit_sarimm(hr,31,HOST_TEMPREG);
4098 emit_storereg(i_regmap[hr]|64,HOST_TEMPREG);
4103 if( !((i_is32>>(i_regmap[hr]&63))&1) && !((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1) ) {
4104 emit_storereg(i_regmap[hr],hr);
4115 // Branch out of this block, write out all dirty regs
4116 wb_dirtys(i_regmap,i_is32,i_dirty);
4120 // Load all needed registers for branch target
4121 void load_regs_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4123 //if(addr>=start && addr<(start+slen*4))
4124 if(internal_branch(i_is32,addr))
4126 int t=(addr-start)>>2;
4128 // Store the cycle count before loading something else
4129 if(i_regmap[HOST_CCREG]!=CCREG) {
4130 assert(i_regmap[HOST_CCREG]==-1);
4132 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4133 emit_storereg(CCREG,HOST_CCREG);
4136 for(hr=0;hr<HOST_REGS;hr++) {
4137 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4138 #ifdef DESTRUCTIVE_WRITEBACK
4139 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)) {
4141 if(i_regmap[hr]!=regs[t].regmap_entry[hr] ) {
4143 if(regs[t].regmap_entry[hr]==0) {
4146 else if(regs[t].regmap_entry[hr]!=CCREG)
4148 emit_loadreg(regs[t].regmap_entry[hr],hr);
4154 for(hr=0;hr<HOST_REGS;hr++) {
4155 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=64&®s[t].regmap_entry[hr]<TEMPREG+64) {
4156 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4157 assert(regs[t].regmap_entry[hr]!=64);
4158 if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
4159 int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4161 emit_loadreg(regs[t].regmap_entry[hr],hr);
4165 emit_sarimm(lr,31,hr);
4170 emit_loadreg(regs[t].regmap_entry[hr],hr);
4173 else if((i_is32>>(regs[t].regmap_entry[hr]&63))&1) {
4174 int lr=get_reg(regs[t].regmap_entry,regs[t].regmap_entry[hr]-64);
4176 emit_sarimm(lr,31,hr);
4183 int match_bt(signed char i_regmap[],uint64_t i_is32,uint64_t i_dirty,int addr)
4185 if(addr>=start && addr<start+slen*4-4)
4187 int t=(addr-start)>>2;
4189 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4190 for(hr=0;hr<HOST_REGS;hr++)
4194 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4196 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4203 if(i_regmap[hr]<TEMPREG)
4205 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4208 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4210 if(!((unneeded_reg_upper[t]>>(i_regmap[hr]&63))&1))
4215 else // Same register but is it 32-bit or dirty?
4218 if(!((regs[t].dirty>>hr)&1))
4222 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4224 //printf("%x: dirty no match\n",addr);
4229 if((((regs[t].was32^i_is32)&~unneeded_reg_upper[t])>>(i_regmap[hr]&63))&1)
4231 //printf("%x: is32 no match\n",addr);
4237 //if(is32[t]&~unneeded_reg_upper[t]&~i_is32) return 0;
4238 // Delay slots are not valid branch targets
4239 //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;
4240 // Delay slots require additional processing, so do not match
4241 if(is_ds[t]) return 0;
4246 for(hr=0;hr<HOST_REGS;hr++)
4252 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4267 static void drc_dbg_emit_do_cmp(int i)
4269 extern void do_insn_cmp();
4273 for(hr=0;hr<HOST_REGS;hr++)
4274 if(regs[i].regmap[hr]>=0) reglist|=1<<hr;
4276 emit_movimm(start+i*4,0);
4277 emit_writeword(0,(int)&pcaddr);
4278 emit_call((int)do_insn_cmp);
4279 //emit_readword((int)&cycle,0);
4280 //emit_addimm(0,2,0);
4281 //emit_writeword(0,(int)&cycle);
4282 restore_regs(reglist);
4285 #define drc_dbg_emit_do_cmp(x)
4288 // Used when a branch jumps into the delay slot of another branch
4289 void ds_assemble_entry(int i)
4291 int t=(ba[i]-start)>>2;
4292 if(!instr_addr[t]) instr_addr[t]=(u_int)out;
4293 assem_debug("Assemble delay slot at %x\n",ba[i]);
4294 assem_debug("<->\n");
4295 drc_dbg_emit_do_cmp(t);
4296 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&®s[t].regmap[HOST_CCREG]!=CCREG)
4297 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty,regs[t].was32);
4298 load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,rs1[t],rs2[t]);
4299 address_generation(t,®s[t],regs[t].regmap_entry);
4300 if(itype[t]==STORE||itype[t]==STORELR||(opcode[t]&0x3b)==0x39||(opcode[t]&0x3b)==0x3a)
4301 load_regs(regs[t].regmap_entry,regs[t].regmap,regs[t].was32,INVCP,INVCP);
4306 alu_assemble(t,®s[t]);break;
4308 imm16_assemble(t,®s[t]);break;
4310 shift_assemble(t,®s[t]);break;
4312 shiftimm_assemble(t,®s[t]);break;
4314 load_assemble(t,®s[t]);break;
4316 loadlr_assemble(t,®s[t]);break;
4318 store_assemble(t,®s[t]);break;
4320 storelr_assemble(t,®s[t]);break;
4322 cop0_assemble(t,®s[t]);break;
4324 cop1_assemble(t,®s[t]);break;
4326 c1ls_assemble(t,®s[t]);break;
4328 cop2_assemble(t,®s[t]);break;
4330 c2ls_assemble(t,®s[t]);break;
4332 c2op_assemble(t,®s[t]);break;
4334 fconv_assemble(t,®s[t]);break;
4336 float_assemble(t,®s[t]);break;
4338 fcomp_assemble(t,®s[t]);break;
4340 multdiv_assemble(t,®s[t]);break;
4342 mov_assemble(t,®s[t]);break;
4352 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4354 store_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
4355 load_regs_bt(regs[t].regmap,regs[t].is32,regs[t].dirty,ba[i]+4);
4356 if(internal_branch(regs[t].is32,ba[i]+4))
4357 assem_debug("branch: internal\n");
4359 assem_debug("branch: external\n");
4360 assert(internal_branch(regs[t].is32,ba[i]+4));
4361 add_to_linker((int)out,ba[i]+4,internal_branch(regs[t].is32,ba[i]+4));
4365 void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
4375 //if(ba[i]>=start && ba[i]<(start+slen*4))
4376 if(internal_branch(branch_regs[i].is32,ba[i]))
4379 if(is_ds[t]) *adj=-1; // Branch into delay slot adds an extra cycle
4387 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4389 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4391 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4392 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4396 else if(*adj==0||invert) {
4397 int cycles=CLOCK_ADJUST(count+2);
4401 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4402 cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
4404 emit_addimm_and_set_flags(cycles,HOST_CCREG);
4410 emit_cmpimm(HOST_CCREG,-CLOCK_ADJUST(count+2));
4414 add_stub(CC_STUB,jaddr,idle?idle:(int)out,(*adj==0||invert||idle)?0:(count+2),i,addr,taken,0);
4417 void do_ccstub(int n)
4420 assem_debug("do_ccstub %x\n",start+stubs[n][4]*4);
4421 set_jump_target(stubs[n][1],(int)out);
4423 if(stubs[n][6]==NULLDS) {
4424 // Delay slot instruction is nullified ("likely" branch)
4425 wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
4427 else if(stubs[n][6]!=TAKEN) {
4428 wb_dirtys(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty);
4431 if(internal_branch(branch_regs[i].is32,ba[i]))
4432 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
4436 // Save PC as return address
4437 emit_movimm(stubs[n][5],EAX);
4438 emit_writeword(EAX,(int)&pcaddr);
4442 // Return address depends on which way the branch goes
4443 if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
4445 int s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4446 int s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
4447 int s2l=get_reg(branch_regs[i].regmap,rs2[i]);
4448 int s2h=get_reg(branch_regs[i].regmap,rs2[i]|64);
4458 if((branch_regs[i].is32>>rs1[i])&(branch_regs[i].is32>>rs2[i])&1) {
4462 #ifdef DESTRUCTIVE_WRITEBACK
4464 if((branch_regs[i].dirty>>s1l)&(branch_regs[i].is32>>rs1[i])&1)
4465 emit_loadreg(rs1[i],s1l);
4468 if((branch_regs[i].dirty>>s1l)&(branch_regs[i].is32>>rs2[i])&1)
4469 emit_loadreg(rs2[i],s1l);
4472 if((branch_regs[i].dirty>>s2l)&(branch_regs[i].is32>>rs2[i])&1)
4473 emit_loadreg(rs2[i],s2l);
4476 int addr=-1,alt=-1,ntaddr=-1;
4479 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4480 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4481 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4489 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4490 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4491 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4497 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
4501 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4502 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4503 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4509 assert(hr<HOST_REGS);
4511 if((opcode[i]&0x2f)==4) // BEQ
4513 #ifdef HAVE_CMOV_IMM
4515 if(s2l>=0) emit_cmp(s1l,s2l);
4516 else emit_test(s1l,s1l);
4517 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4522 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4524 if(s2h>=0) emit_cmp(s1h,s2h);
4525 else emit_test(s1h,s1h);
4526 emit_cmovne_reg(alt,addr);
4528 if(s2l>=0) emit_cmp(s1l,s2l);
4529 else emit_test(s1l,s1l);
4530 emit_cmovne_reg(alt,addr);
4533 if((opcode[i]&0x2f)==5) // BNE
4535 #ifdef HAVE_CMOV_IMM
4537 if(s2l>=0) emit_cmp(s1l,s2l);
4538 else emit_test(s1l,s1l);
4539 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4544 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4546 if(s2h>=0) emit_cmp(s1h,s2h);
4547 else emit_test(s1h,s1h);
4548 emit_cmovne_reg(alt,addr);
4550 if(s2l>=0) emit_cmp(s1l,s2l);
4551 else emit_test(s1l,s1l);
4552 emit_cmovne_reg(alt,addr);
4555 if((opcode[i]&0x2f)==6) // BLEZ
4557 //emit_movimm(ba[i],alt);
4558 //emit_movimm(start+i*4+8,addr);
4559 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4561 if(s1h>=0) emit_mov(addr,ntaddr);
4562 emit_cmovl_reg(alt,addr);
4565 emit_cmovne_reg(ntaddr,addr);
4566 emit_cmovs_reg(alt,addr);
4569 if((opcode[i]&0x2f)==7) // BGTZ
4571 //emit_movimm(ba[i],addr);
4572 //emit_movimm(start+i*4+8,ntaddr);
4573 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
4575 if(s1h>=0) emit_mov(addr,alt);
4576 emit_cmovl_reg(ntaddr,addr);
4579 emit_cmovne_reg(alt,addr);
4580 emit_cmovs_reg(ntaddr,addr);
4583 if((opcode[i]==1)&&(opcode2[i]&0x2D)==0) // BLTZ
4585 //emit_movimm(ba[i],alt);
4586 //emit_movimm(start+i*4+8,addr);
4587 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4588 if(s1h>=0) emit_test(s1h,s1h);
4589 else emit_test(s1l,s1l);
4590 emit_cmovs_reg(alt,addr);
4592 if((opcode[i]==1)&&(opcode2[i]&0x2D)==1) // BGEZ
4594 //emit_movimm(ba[i],addr);
4595 //emit_movimm(start+i*4+8,alt);
4596 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4597 if(s1h>=0) emit_test(s1h,s1h);
4598 else emit_test(s1l,s1l);
4599 emit_cmovs_reg(alt,addr);
4601 if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
4602 if(source[i]&0x10000) // BC1T
4604 //emit_movimm(ba[i],alt);
4605 //emit_movimm(start+i*4+8,addr);
4606 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4607 emit_testimm(s1l,0x800000);
4608 emit_cmovne_reg(alt,addr);
4612 //emit_movimm(ba[i],addr);
4613 //emit_movimm(start+i*4+8,alt);
4614 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4615 emit_testimm(s1l,0x800000);
4616 emit_cmovne_reg(alt,addr);
4619 emit_writeword(addr,(int)&pcaddr);
4624 int r=get_reg(branch_regs[i].regmap,rs1[i]);
4625 if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
4626 r=get_reg(branch_regs[i].regmap,RTEMP);
4628 emit_writeword(r,(int)&pcaddr);
4630 else {SysPrintf("Unknown branch type in do_ccstub\n");exit(1);}
4632 // Update cycle count
4633 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
4634 if(stubs[n][3]) emit_addimm(HOST_CCREG,CLOCK_ADJUST((int)stubs[n][3]),HOST_CCREG);
4635 emit_call((int)cc_interrupt);
4636 if(stubs[n][3]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((int)stubs[n][3]),HOST_CCREG);
4637 if(stubs[n][6]==TAKEN) {
4638 if(internal_branch(branch_regs[i].is32,ba[i]))
4639 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
4640 else if(itype[i]==RJUMP) {
4641 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
4642 emit_readword((int)&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
4644 emit_loadreg(rs1[i],get_reg(branch_regs[i].regmap,rs1[i]));
4646 }else if(stubs[n][6]==NOTTAKEN) {
4647 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
4648 else load_all_regs(branch_regs[i].regmap);
4649 }else if(stubs[n][6]==NULLDS) {
4650 // Delay slot instruction is nullified ("likely" branch)
4651 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
4652 else load_all_regs(regs[i].regmap);
4654 load_all_regs(branch_regs[i].regmap);
4656 emit_jmp(stubs[n][2]); // return address
4658 /* This works but uses a lot of memory...
4659 emit_readword((int)&last_count,ECX);
4660 emit_add(HOST_CCREG,ECX,EAX);
4661 emit_writeword(EAX,(int)&Count);
4662 emit_call((int)gen_interupt);
4663 emit_readword((int)&Count,HOST_CCREG);
4664 emit_readword((int)&next_interupt,EAX);
4665 emit_readword((int)&pending_exception,EBX);
4666 emit_writeword(EAX,(int)&last_count);
4667 emit_sub(HOST_CCREG,EAX,HOST_CCREG);
4669 int jne_instr=(int)out;
4671 if(stubs[n][3]) emit_addimm(HOST_CCREG,-2*stubs[n][3],HOST_CCREG);
4672 load_all_regs(branch_regs[i].regmap);
4673 emit_jmp(stubs[n][2]); // return address
4674 set_jump_target(jne_instr,(int)out);
4675 emit_readword((int)&pcaddr,EAX);
4676 // Call get_addr_ht instead of doing the hash table here.
4677 // This code is executed infrequently and takes up a lot of space
4678 // so smaller is better.
4679 emit_storereg(CCREG,HOST_CCREG);
4681 emit_call((int)get_addr_ht);
4682 emit_loadreg(CCREG,HOST_CCREG);
4683 emit_addimm(ESP,4,ESP);
4687 static void add_to_linker(int addr,int target,int ext)
4689 link_addr[linkcount][0]=addr;
4690 link_addr[linkcount][1]=target;
4691 link_addr[linkcount][2]=ext;
4695 static void ujump_assemble_write_ra(int i)
4698 unsigned int return_address;
4699 rt=get_reg(branch_regs[i].regmap,31);
4700 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]);
4702 return_address=start+i*4+8;
4705 if(internal_branch(branch_regs[i].is32,return_address)&&rt1[i+1]!=31) {
4706 int temp=-1; // note: must be ds-safe
4710 if(temp>=0) do_miniht_insert(return_address,rt,temp);
4711 else emit_movimm(return_address,rt);
4719 if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
4722 emit_movimm(return_address,rt); // PC into link register
4724 emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
4730 void ujump_assemble(int i,struct regstat *i_regs)
4733 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4734 address_generation(i+1,i_regs,regs[i].regmap_entry);
4736 int temp=get_reg(branch_regs[i].regmap,PTEMP);
4737 if(rt1[i]==31&&temp>=0)
4739 signed char *i_regmap=i_regs->regmap;
4740 int return_address=start+i*4+8;
4741 if(get_reg(branch_regs[i].regmap,31)>0)
4742 if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
4745 if(rt1[i]==31&&(rt1[i]==rs1[i+1]||rt1[i]==rs2[i+1])) {
4746 ujump_assemble_write_ra(i); // writeback ra for DS
4749 ds_assemble(i+1,i_regs);
4750 uint64_t bc_unneeded=branch_regs[i].u;
4751 uint64_t bc_unneeded_upper=branch_regs[i].uu;
4752 bc_unneeded|=1|(1LL<<rt1[i]);
4753 bc_unneeded_upper|=1|(1LL<<rt1[i]);
4754 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
4755 bc_unneeded,bc_unneeded_upper);
4756 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
4757 if(!ra_done&&rt1[i]==31)
4758 ujump_assemble_write_ra(i);
4760 cc=get_reg(branch_regs[i].regmap,CCREG);
4761 assert(cc==HOST_CCREG);
4762 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
4764 if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
4766 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
4767 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
4768 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
4769 if(internal_branch(branch_regs[i].is32,ba[i]))
4770 assem_debug("branch: internal\n");
4772 assem_debug("branch: external\n");
4773 if(internal_branch(branch_regs[i].is32,ba[i])&&is_ds[(ba[i]-start)>>2]) {
4774 ds_assemble_entry(i);
4777 add_to_linker((int)out,ba[i],internal_branch(branch_regs[i].is32,ba[i]));
4782 static void rjump_assemble_write_ra(int i)
4784 int rt,return_address;
4785 assert(rt1[i+1]!=rt1[i]);
4786 assert(rt2[i+1]!=rt1[i]);
4787 rt=get_reg(branch_regs[i].regmap,rt1[i]);
4788 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]);
4790 return_address=start+i*4+8;
4794 if(i_regmap[temp]!=PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
4797 emit_movimm(return_address,rt); // PC into link register
4799 emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
4803 void rjump_assemble(int i,struct regstat *i_regs)
4808 rs=get_reg(branch_regs[i].regmap,rs1[i]);
4810 if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
4811 // Delay slot abuse, make a copy of the branch address register
4812 temp=get_reg(branch_regs[i].regmap,RTEMP);
4814 assert(regs[i].regmap[temp]==RTEMP);
4818 address_generation(i+1,i_regs,regs[i].regmap_entry);
4822 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
4823 signed char *i_regmap=i_regs->regmap;
4824 int return_address=start+i*4+8;
4825 if(i_regmap[temp]==PTEMP) emit_movimm((int)hash_table[((return_address>>16)^return_address)&0xFFFF],temp);
4831 int rh=get_reg(regs[i].regmap,RHASH);
4832 if(rh>=0) do_preload_rhash(rh);
4835 if(rt1[i]!=0&&(rt1[i]==rs1[i+1]||rt1[i]==rs2[i+1])) {
4836 rjump_assemble_write_ra(i);
4839 ds_assemble(i+1,i_regs);
4840 uint64_t bc_unneeded=branch_regs[i].u;
4841 uint64_t bc_unneeded_upper=branch_regs[i].uu;
4842 bc_unneeded|=1|(1LL<<rt1[i]);
4843 bc_unneeded_upper|=1|(1LL<<rt1[i]);
4844 bc_unneeded&=~(1LL<<rs1[i]);
4845 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
4846 bc_unneeded,bc_unneeded_upper);
4847 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],CCREG);
4848 if(!ra_done&&rt1[i]!=0)
4849 rjump_assemble_write_ra(i);
4850 cc=get_reg(branch_regs[i].regmap,CCREG);
4851 assert(cc==HOST_CCREG);
4854 int rh=get_reg(branch_regs[i].regmap,RHASH);
4855 int ht=get_reg(branch_regs[i].regmap,RHTBL);
4857 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
4858 do_preload_rhtbl(ht);
4862 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
4863 #ifdef DESTRUCTIVE_WRITEBACK
4864 if((branch_regs[i].dirty>>rs)&(branch_regs[i].is32>>rs1[i])&1) {
4865 if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
4866 emit_loadreg(rs1[i],rs);
4871 if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
4875 do_miniht_load(ht,rh);
4878 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
4879 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
4881 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
4882 add_stub(CC_STUB,(int)out,jump_vaddr_reg[rs],0,i,-1,TAKEN,0);
4883 if(itype[i+1]==COP0&&(source[i+1]&0x3f)==0x10)
4884 // special case for RFE
4888 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,-1);
4891 do_miniht_jump(rs,rh,ht);
4896 //if(rs!=EAX) emit_mov(rs,EAX);
4897 //emit_jmp((int)jump_vaddr_eax);
4898 emit_jmp(jump_vaddr_reg[rs]);
4903 emit_shrimm(rs,16,rs);
4904 emit_xor(temp,rs,rs);
4905 emit_movzwl_reg(rs,rs);
4906 emit_shlimm(rs,4,rs);
4907 emit_cmpmem_indexed((int)hash_table,rs,temp);
4908 emit_jne((int)out+14);
4909 emit_readword_indexed((int)hash_table+4,rs,rs);
4911 emit_cmpmem_indexed((int)hash_table+8,rs,temp);
4912 emit_addimm_no_flags(8,rs);
4913 emit_jeq((int)out-17);
4914 // No hit on hash table, call compiler
4917 #ifdef DEBUG_CYCLE_COUNT
4918 emit_readword((int)&last_count,ECX);
4919 emit_add(HOST_CCREG,ECX,HOST_CCREG);
4920 emit_readword((int)&next_interupt,ECX);
4921 emit_writeword(HOST_CCREG,(int)&Count);
4922 emit_sub(HOST_CCREG,ECX,HOST_CCREG);
4923 emit_writeword(ECX,(int)&last_count);
4926 emit_storereg(CCREG,HOST_CCREG);
4927 emit_call((int)get_addr);
4928 emit_loadreg(CCREG,HOST_CCREG);
4929 emit_addimm(ESP,4,ESP);
4931 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4932 if(rt1[i]!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
4936 void cjump_assemble(int i,struct regstat *i_regs)
4938 signed char *i_regmap=i_regs->regmap;
4941 match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
4942 assem_debug("match=%d\n",match);
4943 int s1h,s1l,s2h,s2l;
4944 int prev_cop1_usable=cop1_usable;
4945 int unconditional=0,nop=0;
4948 int internal=internal_branch(branch_regs[i].is32,ba[i]);
4949 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4950 if(!match) invert=1;
4951 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4952 if(i>(ba[i]-start)>>2) invert=1;
4956 s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4957 s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
4958 s2l=get_reg(branch_regs[i].regmap,rs2[i]);
4959 s2h=get_reg(branch_regs[i].regmap,rs2[i]|64);
4962 s1l=get_reg(i_regmap,rs1[i]);
4963 s1h=get_reg(i_regmap,rs1[i]|64);
4964 s2l=get_reg(i_regmap,rs2[i]);
4965 s2h=get_reg(i_regmap,rs2[i]|64);
4967 if(rs1[i]==0&&rs2[i]==0)
4969 if(opcode[i]&1) nop=1;
4970 else unconditional=1;
4971 //assert(opcode[i]!=5);
4972 //assert(opcode[i]!=7);
4973 //assert(opcode[i]!=0x15);
4974 //assert(opcode[i]!=0x17);
4980 only32=(regs[i].was32>>rs2[i])&1;
4985 only32=(regs[i].was32>>rs1[i])&1;
4988 only32=(regs[i].was32>>rs1[i])&(regs[i].was32>>rs2[i])&1;
4992 // Out of order execution (delay slot first)
4994 address_generation(i+1,i_regs,regs[i].regmap_entry);
4995 ds_assemble(i+1,i_regs);
4997 uint64_t bc_unneeded=branch_regs[i].u;
4998 uint64_t bc_unneeded_upper=branch_regs[i].uu;
4999 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5000 bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5002 bc_unneeded_upper|=1;
5003 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5004 bc_unneeded,bc_unneeded_upper);
5005 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs2[i]);
5006 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5007 cc=get_reg(branch_regs[i].regmap,CCREG);
5008 assert(cc==HOST_CCREG);
5010 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5011 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5012 //assem_debug("cycle count (adj)\n");
5014 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5015 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5016 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5017 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5019 assem_debug("branch: internal\n");
5021 assem_debug("branch: external\n");
5022 if(internal&&is_ds[(ba[i]-start)>>2]) {
5023 ds_assemble_entry(i);
5026 add_to_linker((int)out,ba[i],internal);
5029 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5030 if(((u_int)out)&7) emit_addnop(0);
5035 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5038 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5041 int taken=0,nottaken=0,nottaken1=0;
5042 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5043 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5047 if(opcode[i]==4) // BEQ
5049 if(s2h>=0) emit_cmp(s1h,s2h);
5050 else emit_test(s1h,s1h);
5054 if(opcode[i]==5) // BNE
5056 if(s2h>=0) emit_cmp(s1h,s2h);
5057 else emit_test(s1h,s1h);
5058 if(invert) taken=(int)out;
5059 else add_to_linker((int)out,ba[i],internal);
5062 if(opcode[i]==6) // BLEZ
5065 if(invert) taken=(int)out;
5066 else add_to_linker((int)out,ba[i],internal);
5071 if(opcode[i]==7) // BGTZ
5076 if(invert) taken=(int)out;
5077 else add_to_linker((int)out,ba[i],internal);
5082 //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]);
5084 if(opcode[i]==4) // BEQ
5086 if(s2l>=0) emit_cmp(s1l,s2l);
5087 else emit_test(s1l,s1l);
5092 add_to_linker((int)out,ba[i],internal);
5096 if(opcode[i]==5) // BNE
5098 if(s2l>=0) emit_cmp(s1l,s2l);
5099 else emit_test(s1l,s1l);
5104 add_to_linker((int)out,ba[i],internal);
5108 if(opcode[i]==6) // BLEZ
5115 add_to_linker((int)out,ba[i],internal);
5119 if(opcode[i]==7) // BGTZ
5126 add_to_linker((int)out,ba[i],internal);
5131 if(taken) set_jump_target(taken,(int)out);
5132 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5133 if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5135 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5136 add_to_linker((int)out,ba[i],internal);
5139 add_to_linker((int)out,ba[i],internal*2);
5145 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5146 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5147 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5149 assem_debug("branch: internal\n");
5151 assem_debug("branch: external\n");
5152 if(internal&&is_ds[(ba[i]-start)>>2]) {
5153 ds_assemble_entry(i);
5156 add_to_linker((int)out,ba[i],internal);
5160 set_jump_target(nottaken,(int)out);
5163 if(nottaken1) set_jump_target(nottaken1,(int)out);
5165 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5167 } // (!unconditional)
5171 // In-order execution (branch first)
5172 //if(likely[i]) printf("IOL\n");
5175 int taken=0,nottaken=0,nottaken1=0;
5176 if(!unconditional&&!nop) {
5180 if((opcode[i]&0x2f)==4) // BEQ
5182 if(s2h>=0) emit_cmp(s1h,s2h);
5183 else emit_test(s1h,s1h);
5187 if((opcode[i]&0x2f)==5) // BNE
5189 if(s2h>=0) emit_cmp(s1h,s2h);
5190 else emit_test(s1h,s1h);
5194 if((opcode[i]&0x2f)==6) // BLEZ
5202 if((opcode[i]&0x2f)==7) // BGTZ
5212 //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]);
5214 if((opcode[i]&0x2f)==4) // BEQ
5216 if(s2l>=0) emit_cmp(s1l,s2l);
5217 else emit_test(s1l,s1l);
5221 if((opcode[i]&0x2f)==5) // BNE
5223 if(s2l>=0) emit_cmp(s1l,s2l);
5224 else emit_test(s1l,s1l);
5228 if((opcode[i]&0x2f)==6) // BLEZ
5234 if((opcode[i]&0x2f)==7) // BGTZ
5240 } // if(!unconditional)
5242 uint64_t ds_unneeded=branch_regs[i].u;
5243 uint64_t ds_unneeded_upper=branch_regs[i].uu;
5244 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5245 ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
5246 if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
5248 ds_unneeded_upper|=1;
5251 if(taken) set_jump_target(taken,(int)out);
5252 assem_debug("1:\n");
5253 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5254 ds_unneeded,ds_unneeded_upper);
5256 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5257 address_generation(i+1,&branch_regs[i],0);
5258 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
5259 ds_assemble(i+1,&branch_regs[i]);
5260 cc=get_reg(branch_regs[i].regmap,CCREG);
5262 emit_loadreg(CCREG,cc=HOST_CCREG);
5263 // CHECK: Is the following instruction (fall thru) allocated ok?
5265 assert(cc==HOST_CCREG);
5266 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5267 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5268 assem_debug("cycle count (adj)\n");
5269 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5270 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5272 assem_debug("branch: internal\n");
5274 assem_debug("branch: external\n");
5275 if(internal&&is_ds[(ba[i]-start)>>2]) {
5276 ds_assemble_entry(i);
5279 add_to_linker((int)out,ba[i],internal);
5284 cop1_usable=prev_cop1_usable;
5285 if(!unconditional) {
5286 if(nottaken1) set_jump_target(nottaken1,(int)out);
5287 set_jump_target(nottaken,(int)out);
5288 assem_debug("2:\n");
5290 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5291 ds_unneeded,ds_unneeded_upper);
5292 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5293 address_generation(i+1,&branch_regs[i],0);
5294 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5295 ds_assemble(i+1,&branch_regs[i]);
5297 cc=get_reg(branch_regs[i].regmap,CCREG);
5298 if(cc==-1&&!likely[i]) {
5299 // Cycle count isn't in a register, temporarily load it then write it out
5300 emit_loadreg(CCREG,HOST_CCREG);
5301 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5304 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5305 emit_storereg(CCREG,HOST_CCREG);
5308 cc=get_reg(i_regmap,CCREG);
5309 assert(cc==HOST_CCREG);
5310 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5313 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5319 void sjump_assemble(int i,struct regstat *i_regs)
5321 signed char *i_regmap=i_regs->regmap;
5324 match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5325 assem_debug("smatch=%d\n",match);
5327 int prev_cop1_usable=cop1_usable;
5328 int unconditional=0,nevertaken=0;
5331 int internal=internal_branch(branch_regs[i].is32,ba[i]);
5332 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5333 if(!match) invert=1;
5334 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5335 if(i>(ba[i]-start)>>2) invert=1;
5338 //if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
5339 //assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
5342 s1l=get_reg(branch_regs[i].regmap,rs1[i]);
5343 s1h=get_reg(branch_regs[i].regmap,rs1[i]|64);
5346 s1l=get_reg(i_regmap,rs1[i]);
5347 s1h=get_reg(i_regmap,rs1[i]|64);
5351 if(opcode2[i]&1) unconditional=1;
5353 // These are never taken (r0 is never less than zero)
5354 //assert(opcode2[i]!=0);
5355 //assert(opcode2[i]!=2);
5356 //assert(opcode2[i]!=0x10);
5357 //assert(opcode2[i]!=0x12);
5360 only32=(regs[i].was32>>rs1[i])&1;
5364 // Out of order execution (delay slot first)
5366 address_generation(i+1,i_regs,regs[i].regmap_entry);
5367 ds_assemble(i+1,i_regs);
5369 uint64_t bc_unneeded=branch_regs[i].u;
5370 uint64_t bc_unneeded_upper=branch_regs[i].uu;
5371 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5372 bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5374 bc_unneeded_upper|=1;
5375 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5376 bc_unneeded,bc_unneeded_upper);
5377 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs1[i]);
5378 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5380 int rt,return_address;
5381 rt=get_reg(branch_regs[i].regmap,31);
5382 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]);
5384 // Save the PC even if the branch is not taken
5385 return_address=start+i*4+8;
5386 emit_movimm(return_address,rt); // PC into link register
5388 if(!nevertaken) emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5392 cc=get_reg(branch_regs[i].regmap,CCREG);
5393 assert(cc==HOST_CCREG);
5395 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5396 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5397 assem_debug("cycle count (adj)\n");
5399 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5400 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5401 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5402 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5404 assem_debug("branch: internal\n");
5406 assem_debug("branch: external\n");
5407 if(internal&&is_ds[(ba[i]-start)>>2]) {
5408 ds_assemble_entry(i);
5411 add_to_linker((int)out,ba[i],internal);
5414 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5415 if(((u_int)out)&7) emit_addnop(0);
5419 else if(nevertaken) {
5420 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5423 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5427 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5428 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5432 if((opcode2[i]&0xf)==0) // BLTZ/BLTZAL
5439 add_to_linker((int)out,ba[i],internal);
5443 if((opcode2[i]&0xf)==1) // BGEZ/BLTZAL
5450 add_to_linker((int)out,ba[i],internal);
5458 if((opcode2[i]&0xf)==0) // BLTZ/BLTZAL
5465 add_to_linker((int)out,ba[i],internal);
5469 if((opcode2[i]&0xf)==1) // BGEZ/BLTZAL
5476 add_to_linker((int)out,ba[i],internal);
5483 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5484 if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5486 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5487 add_to_linker((int)out,ba[i],internal);
5490 add_to_linker((int)out,ba[i],internal*2);
5496 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5497 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5498 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5500 assem_debug("branch: internal\n");
5502 assem_debug("branch: external\n");
5503 if(internal&&is_ds[(ba[i]-start)>>2]) {
5504 ds_assemble_entry(i);
5507 add_to_linker((int)out,ba[i],internal);
5511 set_jump_target(nottaken,(int)out);
5515 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5517 } // (!unconditional)
5521 // In-order execution (branch first)
5525 int rt,return_address;
5526 rt=get_reg(branch_regs[i].regmap,31);
5528 // Save the PC even if the branch is not taken
5529 return_address=start+i*4+8;
5530 emit_movimm(return_address,rt); // PC into link register
5532 emit_prefetch(hash_table[((return_address>>16)^return_address)&0xFFFF]);
5536 if(!unconditional) {
5537 //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]);
5541 if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5547 if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5557 if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5563 if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5570 } // if(!unconditional)
5572 uint64_t ds_unneeded=branch_regs[i].u;
5573 uint64_t ds_unneeded_upper=branch_regs[i].uu;
5574 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5575 ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
5576 if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
5578 ds_unneeded_upper|=1;
5581 //assem_debug("1:\n");
5582 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5583 ds_unneeded,ds_unneeded_upper);
5585 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5586 address_generation(i+1,&branch_regs[i],0);
5587 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
5588 ds_assemble(i+1,&branch_regs[i]);
5589 cc=get_reg(branch_regs[i].regmap,CCREG);
5591 emit_loadreg(CCREG,cc=HOST_CCREG);
5592 // CHECK: Is the following instruction (fall thru) allocated ok?
5594 assert(cc==HOST_CCREG);
5595 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5596 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5597 assem_debug("cycle count (adj)\n");
5598 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5599 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5601 assem_debug("branch: internal\n");
5603 assem_debug("branch: external\n");
5604 if(internal&&is_ds[(ba[i]-start)>>2]) {
5605 ds_assemble_entry(i);
5608 add_to_linker((int)out,ba[i],internal);
5613 cop1_usable=prev_cop1_usable;
5614 if(!unconditional) {
5615 set_jump_target(nottaken,(int)out);
5616 assem_debug("1:\n");
5618 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5619 ds_unneeded,ds_unneeded_upper);
5620 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5621 address_generation(i+1,&branch_regs[i],0);
5622 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5623 ds_assemble(i+1,&branch_regs[i]);
5625 cc=get_reg(branch_regs[i].regmap,CCREG);
5626 if(cc==-1&&!likely[i]) {
5627 // Cycle count isn't in a register, temporarily load it then write it out
5628 emit_loadreg(CCREG,HOST_CCREG);
5629 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5632 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5633 emit_storereg(CCREG,HOST_CCREG);
5636 cc=get_reg(i_regmap,CCREG);
5637 assert(cc==HOST_CCREG);
5638 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5641 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5647 void fjump_assemble(int i,struct regstat *i_regs)
5649 signed char *i_regmap=i_regs->regmap;
5652 match=match_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5653 assem_debug("fmatch=%d\n",match);
5657 int internal=internal_branch(branch_regs[i].is32,ba[i]);
5658 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5659 if(!match) invert=1;
5660 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5661 if(i>(ba[i]-start)>>2) invert=1;
5665 fs=get_reg(branch_regs[i].regmap,FSREG);
5666 address_generation(i+1,i_regs,regs[i].regmap_entry); // Is this okay?
5669 fs=get_reg(i_regmap,FSREG);
5672 // Check cop1 unusable
5674 cs=get_reg(i_regmap,CSREG);
5676 emit_testimm(cs,0x20000000);
5679 add_stub(FP_STUB,eaddr,(int)out,i,cs,(int)i_regs,0,0);
5684 // Out of order execution (delay slot first)
5686 ds_assemble(i+1,i_regs);
5688 uint64_t bc_unneeded=branch_regs[i].u;
5689 uint64_t bc_unneeded_upper=branch_regs[i].uu;
5690 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5691 bc_unneeded_upper&=~((1LL<<us1[i])|(1LL<<us2[i]));
5693 bc_unneeded_upper|=1;
5694 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5695 bc_unneeded,bc_unneeded_upper);
5696 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i],rs1[i]);
5697 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5698 cc=get_reg(branch_regs[i].regmap,CCREG);
5699 assert(cc==HOST_CCREG);
5700 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5701 assem_debug("cycle count (adj)\n");
5704 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5707 emit_testimm(fs,0x800000);
5708 if(source[i]&0x10000) // BC1T
5714 add_to_linker((int)out,ba[i],internal);
5723 add_to_linker((int)out,ba[i],internal);
5731 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5732 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5733 else if(match) emit_addnop(13);
5735 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5736 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5738 assem_debug("branch: internal\n");
5740 assem_debug("branch: external\n");
5741 if(internal&&is_ds[(ba[i]-start)>>2]) {
5742 ds_assemble_entry(i);
5745 add_to_linker((int)out,ba[i],internal);
5748 set_jump_target(nottaken,(int)out);
5752 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5754 } // (!unconditional)
5758 // In-order execution (branch first)
5762 //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]);
5765 emit_testimm(fs,0x800000);
5766 if(source[i]&0x10000) // BC1T
5777 } // if(!unconditional)
5779 uint64_t ds_unneeded=branch_regs[i].u;
5780 uint64_t ds_unneeded_upper=branch_regs[i].uu;
5781 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5782 ds_unneeded_upper&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
5783 if((~ds_unneeded_upper>>rt1[i+1])&1) ds_unneeded_upper&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
5785 ds_unneeded_upper|=1;
5787 //assem_debug("1:\n");
5788 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5789 ds_unneeded,ds_unneeded_upper);
5791 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5792 address_generation(i+1,&branch_regs[i],0);
5793 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,INVCP);
5794 ds_assemble(i+1,&branch_regs[i]);
5795 cc=get_reg(branch_regs[i].regmap,CCREG);
5797 emit_loadreg(CCREG,cc=HOST_CCREG);
5798 // CHECK: Is the following instruction (fall thru) allocated ok?
5800 assert(cc==HOST_CCREG);
5801 store_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5802 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5803 assem_debug("cycle count (adj)\n");
5804 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5805 load_regs_bt(branch_regs[i].regmap,branch_regs[i].is32,branch_regs[i].dirty,ba[i]);
5807 assem_debug("branch: internal\n");
5809 assem_debug("branch: external\n");
5810 if(internal&&is_ds[(ba[i]-start)>>2]) {
5811 ds_assemble_entry(i);
5814 add_to_linker((int)out,ba[i],internal);
5819 if(1) { // <- FIXME (don't need this)
5820 set_jump_target(nottaken,(int)out);
5821 assem_debug("1:\n");
5823 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,regs[i].is32,
5824 ds_unneeded,ds_unneeded_upper);
5825 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,rs1[i+1],rs2[i+1]);
5826 address_generation(i+1,&branch_regs[i],0);
5827 load_regs(regs[i].regmap,branch_regs[i].regmap,regs[i].was32,CCREG,CCREG);
5828 ds_assemble(i+1,&branch_regs[i]);
5830 cc=get_reg(branch_regs[i].regmap,CCREG);
5831 if(cc==-1&&!likely[i]) {
5832 // Cycle count isn't in a register, temporarily load it then write it out
5833 emit_loadreg(CCREG,HOST_CCREG);
5834 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5837 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,NOTTAKEN,0);
5838 emit_storereg(CCREG,HOST_CCREG);
5841 cc=get_reg(i_regmap,CCREG);
5842 assert(cc==HOST_CCREG);
5843 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5846 add_stub(CC_STUB,jaddr,(int)out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5852 static void pagespan_assemble(int i,struct regstat *i_regs)
5854 int s1l=get_reg(i_regs->regmap,rs1[i]);
5855 int s1h=get_reg(i_regs->regmap,rs1[i]|64);
5856 int s2l=get_reg(i_regs->regmap,rs2[i]);
5857 int s2h=get_reg(i_regs->regmap,rs2[i]|64);
5860 int unconditional=0;
5870 if((i_regs->is32>>rs1[i])&(i_regs->is32>>rs2[i])&1) {
5874 int addr=-1,alt=-1,ntaddr=-1;
5875 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5879 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5880 (i_regs->regmap[hr]&63)!=rs1[i] &&
5881 (i_regs->regmap[hr]&63)!=rs2[i] )
5890 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5891 (i_regs->regmap[hr]&63)!=rs1[i] &&
5892 (i_regs->regmap[hr]&63)!=rs2[i] )
5898 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
5902 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5903 (i_regs->regmap[hr]&63)!=rs1[i] &&
5904 (i_regs->regmap[hr]&63)!=rs2[i] )
5911 assert(hr<HOST_REGS);
5912 if((opcode[i]&0x2e)==4||opcode[i]==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5913 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
5915 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5916 if(opcode[i]==2) // J
5920 if(opcode[i]==3) // JAL
5923 int rt=get_reg(i_regs->regmap,31);
5924 emit_movimm(start+i*4+8,rt);
5927 if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
5930 if(opcode2[i]==9) // JALR
5932 int rt=get_reg(i_regs->regmap,rt1[i]);
5933 emit_movimm(start+i*4+8,rt);
5936 if((opcode[i]&0x3f)==4) // BEQ
5943 #ifdef HAVE_CMOV_IMM
5945 if(s2l>=0) emit_cmp(s1l,s2l);
5946 else emit_test(s1l,s1l);
5947 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5953 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5955 if(s2h>=0) emit_cmp(s1h,s2h);
5956 else emit_test(s1h,s1h);
5957 emit_cmovne_reg(alt,addr);
5959 if(s2l>=0) emit_cmp(s1l,s2l);
5960 else emit_test(s1l,s1l);
5961 emit_cmovne_reg(alt,addr);
5964 if((opcode[i]&0x3f)==5) // BNE
5966 #ifdef HAVE_CMOV_IMM
5968 if(s2l>=0) emit_cmp(s1l,s2l);
5969 else emit_test(s1l,s1l);
5970 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5976 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5978 if(s2h>=0) emit_cmp(s1h,s2h);
5979 else emit_test(s1h,s1h);
5980 emit_cmovne_reg(alt,addr);
5982 if(s2l>=0) emit_cmp(s1l,s2l);
5983 else emit_test(s1l,s1l);
5984 emit_cmovne_reg(alt,addr);
5987 if((opcode[i]&0x3f)==0x14) // BEQL
5990 if(s2h>=0) emit_cmp(s1h,s2h);
5991 else emit_test(s1h,s1h);
5995 if(s2l>=0) emit_cmp(s1l,s2l);
5996 else emit_test(s1l,s1l);
5997 if(nottaken) set_jump_target(nottaken,(int)out);
6001 if((opcode[i]&0x3f)==0x15) // BNEL
6004 if(s2h>=0) emit_cmp(s1h,s2h);
6005 else emit_test(s1h,s1h);
6009 if(s2l>=0) emit_cmp(s1l,s2l);
6010 else emit_test(s1l,s1l);
6013 if(taken) set_jump_target(taken,(int)out);
6015 if((opcode[i]&0x3f)==6) // BLEZ
6017 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6019 if(s1h>=0) emit_mov(addr,ntaddr);
6020 emit_cmovl_reg(alt,addr);
6023 emit_cmovne_reg(ntaddr,addr);
6024 emit_cmovs_reg(alt,addr);
6027 if((opcode[i]&0x3f)==7) // BGTZ
6029 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
6031 if(s1h>=0) emit_mov(addr,alt);
6032 emit_cmovl_reg(ntaddr,addr);
6035 emit_cmovne_reg(alt,addr);
6036 emit_cmovs_reg(ntaddr,addr);
6039 if((opcode[i]&0x3f)==0x16) // BLEZL
6041 assert((opcode[i]&0x3f)!=0x16);
6043 if((opcode[i]&0x3f)==0x17) // BGTZL
6045 assert((opcode[i]&0x3f)!=0x17);
6047 assert(opcode[i]!=1); // BLTZ/BGEZ
6049 //FIXME: Check CSREG
6050 if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
6051 if((source[i]&0x30000)==0) // BC1F
6053 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6054 emit_testimm(s1l,0x800000);
6055 emit_cmovne_reg(alt,addr);
6057 if((source[i]&0x30000)==0x10000) // BC1T
6059 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6060 emit_testimm(s1l,0x800000);
6061 emit_cmovne_reg(alt,addr);
6063 if((source[i]&0x30000)==0x20000) // BC1FL
6065 emit_testimm(s1l,0x800000);
6069 if((source[i]&0x30000)==0x30000) // BC1TL
6071 emit_testimm(s1l,0x800000);
6077 assert(i_regs->regmap[HOST_CCREG]==CCREG);
6078 wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
6079 if(likely[i]||unconditional)
6081 emit_movimm(ba[i],HOST_BTREG);
6083 else if(addr!=HOST_BTREG)
6085 emit_mov(addr,HOST_BTREG);
6087 void *branch_addr=out;
6089 int target_addr=start+i*4+5;
6091 void *compiled_target_addr=check_addr(target_addr);
6092 emit_extjump_ds((int)branch_addr,target_addr);
6093 if(compiled_target_addr) {
6094 set_jump_target((int)branch_addr,(int)compiled_target_addr);
6095 add_link(target_addr,stub);
6097 else set_jump_target((int)branch_addr,(int)stub);
6100 set_jump_target((int)nottaken,(int)out);
6101 wb_dirtys(regs[i].regmap,regs[i].is32,regs[i].dirty);
6102 void *branch_addr=out;
6104 int target_addr=start+i*4+8;
6106 void *compiled_target_addr=check_addr(target_addr);
6107 emit_extjump_ds((int)branch_addr,target_addr);
6108 if(compiled_target_addr) {
6109 set_jump_target((int)branch_addr,(int)compiled_target_addr);
6110 add_link(target_addr,stub);
6112 else set_jump_target((int)branch_addr,(int)stub);
6116 // Assemble the delay slot for the above
6117 static void pagespan_ds()
6119 assem_debug("initial delay slot:\n");
6120 u_int vaddr=start+1;
6121 u_int page=get_page(vaddr);
6122 u_int vpage=get_vpage(vaddr);
6123 ll_add(jump_dirty+vpage,vaddr,(void *)out);
6125 ll_add(jump_in+page,vaddr,(void *)out);
6126 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
6127 if(regs[0].regmap[HOST_CCREG]!=CCREG)
6128 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty,regs[0].was32);
6129 if(regs[0].regmap[HOST_BTREG]!=BTREG)
6130 emit_writeword(HOST_BTREG,(int)&branch_target);
6131 load_regs(regs[0].regmap_entry,regs[0].regmap,regs[0].was32,rs1[0],rs2[0]);
6132 address_generation(0,®s[0],regs[0].regmap_entry);
6133 if(itype[0]==STORE||itype[0]==STORELR||(opcode[0]&0x3b)==0x39||(opcode[0]&0x3b)==0x3a)
6134 load_regs(regs[0].regmap_entry,regs[0].regmap,regs[0].was32,INVCP,INVCP);
6139 alu_assemble(0,®s[0]);break;
6141 imm16_assemble(0,®s[0]);break;
6143 shift_assemble(0,®s[0]);break;
6145 shiftimm_assemble(0,®s[0]);break;
6147 load_assemble(0,®s[0]);break;
6149 loadlr_assemble(0,®s[0]);break;
6151 store_assemble(0,®s[0]);break;
6153 storelr_assemble(0,®s[0]);break;
6155 cop0_assemble(0,®s[0]);break;
6157 cop1_assemble(0,®s[0]);break;
6159 c1ls_assemble(0,®s[0]);break;
6161 cop2_assemble(0,®s[0]);break;
6163 c2ls_assemble(0,®s[0]);break;
6165 c2op_assemble(0,®s[0]);break;
6167 fconv_assemble(0,®s[0]);break;
6169 float_assemble(0,®s[0]);break;
6171 fcomp_assemble(0,®s[0]);break;
6173 multdiv_assemble(0,®s[0]);break;
6175 mov_assemble(0,®s[0]);break;
6185 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
6187 int btaddr=get_reg(regs[0].regmap,BTREG);
6189 btaddr=get_reg(regs[0].regmap,-1);
6190 emit_readword((int)&branch_target,btaddr);
6192 assert(btaddr!=HOST_CCREG);
6193 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
6195 emit_movimm(start+4,HOST_TEMPREG);
6196 emit_cmp(btaddr,HOST_TEMPREG);
6198 emit_cmpimm(btaddr,start+4);
6200 int branch=(int)out;
6202 store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,-1);
6203 emit_jmp(jump_vaddr_reg[btaddr]);
6204 set_jump_target(branch,(int)out);
6205 store_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
6206 load_regs_bt(regs[0].regmap,regs[0].is32,regs[0].dirty,start+4);
6209 // Basic liveness analysis for MIPS registers
6210 void unneeded_registers(int istart,int iend,int r)
6213 uint64_t u,uu,gte_u,b,bu,gte_bu;
6214 uint64_t temp_u,temp_uu,temp_gte_u=0;
6216 uint64_t gte_u_unknown=0;
6217 if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
6221 gte_u=gte_u_unknown;
6223 u=unneeded_reg[iend+1];
6224 uu=unneeded_reg_upper[iend+1];
6226 gte_u=gte_unneeded[iend+1];
6229 for (i=iend;i>=istart;i--)
6231 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6232 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
6234 // If subroutine call, flag return address as a possible branch target
6235 if(rt1[i]==31 && i<slen-2) bt[i+2]=1;
6237 if(ba[i]<start || ba[i]>=(start+slen*4))
6239 // Branch out of this block, flush all regs
6242 gte_u=gte_u_unknown;
6244 if(itype[i]==UJUMP&&rt1[i]==31)
6246 uu=u=0x300C00F; // Discard at, v0-v1, t6-t9
6248 if(itype[i]==RJUMP&&rs1[i]==31)
6250 uu=u=0x300C0F3; // Discard at, a0-a3, t6-t9
6252 if(start>0x80000400&&start<0x80000000+RAM_SIZE) {
6253 if(itype[i]==UJUMP&&rt1[i]==31)
6255 //uu=u=0x30300FF0FLL; // Discard at, v0-v1, t0-t9, lo, hi
6256 uu=u=0x300FF0F; // Discard at, v0-v1, t0-t9
6258 if(itype[i]==RJUMP&&rs1[i]==31)
6260 //uu=u=0x30300FFF3LL; // Discard at, a0-a3, t0-t9, lo, hi
6261 uu=u=0x300FFF3; // Discard at, a0-a3, t0-t9
6264 branch_unneeded_reg[i]=u;
6265 branch_unneeded_reg_upper[i]=uu;
6266 // Merge in delay slot
6267 tdep=(~uu>>rt1[i+1])&1;
6268 u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6269 uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6270 u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6271 uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6272 uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6275 gte_u&=~gte_rs[i+1];
6276 // If branch is "likely" (and conditional)
6277 // then we skip the delay slot on the fall-thru path
6280 u&=unneeded_reg[i+2];
6281 uu&=unneeded_reg_upper[i+2];
6282 gte_u&=gte_unneeded[i+2];
6288 gte_u=gte_u_unknown;
6294 // Internal branch, flag target
6295 bt[(ba[i]-start)>>2]=1;
6296 if(ba[i]<=start+i*4) {
6298 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6300 // Unconditional branch
6304 // Conditional branch (not taken case)
6305 temp_u=unneeded_reg[i+2];
6306 temp_uu=unneeded_reg_upper[i+2];
6307 temp_gte_u&=gte_unneeded[i+2];
6309 // Merge in delay slot
6310 tdep=(~temp_uu>>rt1[i+1])&1;
6311 temp_u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6312 temp_uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6313 temp_u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6314 temp_uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6315 temp_uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6316 temp_u|=1;temp_uu|=1;
6317 temp_gte_u|=gte_rt[i+1];
6318 temp_gte_u&=~gte_rs[i+1];
6319 // If branch is "likely" (and conditional)
6320 // then we skip the delay slot on the fall-thru path
6323 temp_u&=unneeded_reg[i+2];
6324 temp_uu&=unneeded_reg_upper[i+2];
6325 temp_gte_u&=gte_unneeded[i+2];
6331 temp_gte_u=gte_u_unknown;
6334 tdep=(~temp_uu>>rt1[i])&1;
6335 temp_u|=(1LL<<rt1[i])|(1LL<<rt2[i]);
6336 temp_uu|=(1LL<<rt1[i])|(1LL<<rt2[i]);
6337 temp_u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
6338 temp_uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
6339 temp_uu&=~((tdep<<dep1[i])|(tdep<<dep2[i]));
6340 temp_u|=1;temp_uu|=1;
6341 temp_gte_u|=gte_rt[i];
6342 temp_gte_u&=~gte_rs[i];
6343 unneeded_reg[i]=temp_u;
6344 unneeded_reg_upper[i]=temp_uu;
6345 gte_unneeded[i]=temp_gte_u;
6346 // Only go three levels deep. This recursion can take an
6347 // excessive amount of time if there are a lot of nested loops.
6349 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6351 unneeded_reg[(ba[i]-start)>>2]=1;
6352 unneeded_reg_upper[(ba[i]-start)>>2]=1;
6353 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
6356 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6358 // Unconditional branch
6359 u=unneeded_reg[(ba[i]-start)>>2];
6360 uu=unneeded_reg_upper[(ba[i]-start)>>2];
6361 gte_u=gte_unneeded[(ba[i]-start)>>2];
6362 branch_unneeded_reg[i]=u;
6363 branch_unneeded_reg_upper[i]=uu;
6366 //branch_unneeded_reg[i]=u;
6367 //branch_unneeded_reg_upper[i]=uu;
6368 // Merge in delay slot
6369 tdep=(~uu>>rt1[i+1])&1;
6370 u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6371 uu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6372 u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6373 uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6374 uu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6377 gte_u&=~gte_rs[i+1];
6379 // Conditional branch
6380 b=unneeded_reg[(ba[i]-start)>>2];
6381 bu=unneeded_reg_upper[(ba[i]-start)>>2];
6382 gte_bu=gte_unneeded[(ba[i]-start)>>2];
6383 branch_unneeded_reg[i]=b;
6384 branch_unneeded_reg_upper[i]=bu;
6387 //branch_unneeded_reg[i]=b;
6388 //branch_unneeded_reg_upper[i]=bu;
6389 // Branch delay slot
6390 tdep=(~uu>>rt1[i+1])&1;
6391 b|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6392 bu|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
6393 b&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
6394 bu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
6395 bu&=~((tdep<<dep1[i+1])|(tdep<<dep2[i+1]));
6397 gte_bu|=gte_rt[i+1];
6398 gte_bu&=~gte_rs[i+1];
6399 // If branch is "likely" then we skip the
6400 // delay slot on the fall-thru path
6406 u&=unneeded_reg[i+2];
6407 uu&=unneeded_reg_upper[i+2];
6408 gte_u&=gte_unneeded[i+2];
6420 branch_unneeded_reg[i]&=unneeded_reg[i+2];
6421 branch_unneeded_reg_upper[i]&=unneeded_reg_upper[i+2];
6422 //branch_unneeded_reg[i]=1;
6423 //branch_unneeded_reg_upper[i]=1;
6425 branch_unneeded_reg[i]=1;
6426 branch_unneeded_reg_upper[i]=1;
6432 else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
6434 // SYSCALL instruction (software interrupt)
6438 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
6440 // ERET instruction (return from interrupt)
6445 tdep=(~uu>>rt1[i])&1;
6446 // Written registers are unneeded
6452 // Accessed registers are needed
6458 if(gte_rs[i]&&rt1[i]&&(unneeded_reg[i+1]&(1ll<<rt1[i])))
6459 gte_u|=gte_rs[i]>e_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
6460 // Source-target dependencies
6461 uu&=~(tdep<<dep1[i]);
6462 uu&=~(tdep<<dep2[i]);
6463 // R0 is always unneeded
6467 unneeded_reg_upper[i]=uu;
6468 gte_unneeded[i]=gte_u;
6470 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6473 for(r=1;r<=CCREG;r++) {
6474 if((unneeded_reg[i]>>r)&1) {
6475 if(r==HIREG) printf(" HI");
6476 else if(r==LOREG) printf(" LO");
6477 else printf(" r%d",r);
6481 for(r=1;r<=CCREG;r++) {
6482 if(((unneeded_reg_upper[i]&~unneeded_reg[i])>>r)&1) {
6483 if(r==HIREG) printf(" HI");
6484 else if(r==LOREG) printf(" LO");
6485 else printf(" r%d",r);
6490 for (i=iend;i>=istart;i--)
6492 unneeded_reg_upper[i]=branch_unneeded_reg_upper[i]=-1LL;
6496 // Write back dirty registers as soon as we will no longer modify them,
6497 // so that we don't end up with lots of writes at the branches.
6498 void clean_registers(int istart,int iend,int wr)
6502 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
6503 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
6505 will_dirty_i=will_dirty_next=0;
6506 wont_dirty_i=wont_dirty_next=0;
6508 will_dirty_i=will_dirty_next=will_dirty[iend+1];
6509 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
6511 for (i=iend;i>=istart;i--)
6513 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
6515 if(ba[i]<start || ba[i]>=(start+slen*4))
6517 // Branch out of this block, flush all regs
6518 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6520 // Unconditional branch
6523 // Merge in delay slot (will dirty)
6524 for(r=0;r<HOST_REGS;r++) {
6525 if(r!=EXCLUDE_REG) {
6526 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6527 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6528 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6529 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6530 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6531 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6532 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6533 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6534 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6535 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6536 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6537 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6538 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6539 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6545 // Conditional branch
6547 wont_dirty_i=wont_dirty_next;
6548 // Merge in delay slot (will dirty)
6549 for(r=0;r<HOST_REGS;r++) {
6550 if(r!=EXCLUDE_REG) {
6552 // Might not dirty if likely branch is not taken
6553 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6554 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6555 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6556 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6557 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6558 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
6559 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6560 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6561 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6562 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6563 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6564 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6565 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6566 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6571 // Merge in delay slot (wont dirty)
6572 for(r=0;r<HOST_REGS;r++) {
6573 if(r!=EXCLUDE_REG) {
6574 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6575 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6576 if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
6577 if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
6578 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6579 if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6580 if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6581 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
6582 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
6583 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6587 #ifndef DESTRUCTIVE_WRITEBACK
6588 branch_regs[i].dirty&=wont_dirty_i;
6590 branch_regs[i].dirty|=will_dirty_i;
6596 if(ba[i]<=start+i*4) {
6598 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6600 // Unconditional branch
6603 // Merge in delay slot (will dirty)
6604 for(r=0;r<HOST_REGS;r++) {
6605 if(r!=EXCLUDE_REG) {
6606 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
6607 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
6608 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
6609 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
6610 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6611 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6612 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6613 if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
6614 if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
6615 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
6616 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
6617 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6618 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6619 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6623 // Conditional branch (not taken case)
6624 temp_will_dirty=will_dirty_next;
6625 temp_wont_dirty=wont_dirty_next;
6626 // Merge in delay slot (will dirty)
6627 for(r=0;r<HOST_REGS;r++) {
6628 if(r!=EXCLUDE_REG) {
6630 // Will not dirty if likely branch is not taken
6631 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
6632 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
6633 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
6634 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
6635 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6636 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
6637 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6638 //if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
6639 //if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
6640 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
6641 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
6642 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6643 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6644 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6649 // Merge in delay slot (wont dirty)
6650 for(r=0;r<HOST_REGS;r++) {
6651 if(r!=EXCLUDE_REG) {
6652 if((regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
6653 if((regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
6654 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
6655 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
6656 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6657 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
6658 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
6659 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
6660 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
6661 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6664 // Deal with changed mappings
6666 for(r=0;r<HOST_REGS;r++) {
6667 if(r!=EXCLUDE_REG) {
6668 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
6669 temp_will_dirty&=~(1<<r);
6670 temp_wont_dirty&=~(1<<r);
6671 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6672 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6673 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6675 temp_will_dirty|=1<<r;
6676 temp_wont_dirty|=1<<r;
6683 will_dirty[i]=temp_will_dirty;
6684 wont_dirty[i]=temp_wont_dirty;
6685 clean_registers((ba[i]-start)>>2,i-1,0);
6687 // Limit recursion. It can take an excessive amount
6688 // of time if there are a lot of nested loops.
6689 will_dirty[(ba[i]-start)>>2]=0;
6690 wont_dirty[(ba[i]-start)>>2]=-1;
6695 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
6697 // Unconditional branch
6700 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6701 for(r=0;r<HOST_REGS;r++) {
6702 if(r!=EXCLUDE_REG) {
6703 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6704 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
6705 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6707 if(branch_regs[i].regmap[r]>=0) {
6708 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6709 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6714 // Merge in delay slot
6715 for(r=0;r<HOST_REGS;r++) {
6716 if(r!=EXCLUDE_REG) {
6717 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6718 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6719 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6720 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6721 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6722 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6723 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6724 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6725 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6726 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6727 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6728 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6729 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6730 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6734 // Conditional branch
6735 will_dirty_i=will_dirty_next;
6736 wont_dirty_i=wont_dirty_next;
6737 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6738 for(r=0;r<HOST_REGS;r++) {
6739 if(r!=EXCLUDE_REG) {
6740 signed char target_reg=branch_regs[i].regmap[r];
6741 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6742 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
6743 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6745 else if(target_reg>=0) {
6746 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6747 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6749 // Treat delay slot as part of branch too
6750 /*if(regs[i+1].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6751 will_dirty[i+1]&=will_dirty[(ba[i]-start)>>2]&(1<<r);
6752 wont_dirty[i+1]|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6756 will_dirty[i+1]&=~(1<<r);
6761 // Merge in delay slot
6762 for(r=0;r<HOST_REGS;r++) {
6763 if(r!=EXCLUDE_REG) {
6765 // Might not dirty if likely branch is not taken
6766 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6767 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6768 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6769 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6770 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6771 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6772 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6773 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6774 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6775 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6776 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6777 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6778 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6779 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6784 // Merge in delay slot (won't dirty)
6785 for(r=0;r<HOST_REGS;r++) {
6786 if(r!=EXCLUDE_REG) {
6787 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6788 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6789 if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
6790 if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
6791 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6792 if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6793 if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6794 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
6795 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
6796 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6800 #ifndef DESTRUCTIVE_WRITEBACK
6801 branch_regs[i].dirty&=wont_dirty_i;
6803 branch_regs[i].dirty|=will_dirty_i;
6808 else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
6810 // SYSCALL instruction (software interrupt)
6814 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
6816 // ERET instruction (return from interrupt)
6820 will_dirty_next=will_dirty_i;
6821 wont_dirty_next=wont_dirty_i;
6822 for(r=0;r<HOST_REGS;r++) {
6823 if(r!=EXCLUDE_REG) {
6824 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6825 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6826 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6827 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6828 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6829 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6830 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6831 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6833 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=FJUMP)
6835 // Don't store a register immediately after writing it,
6836 // may prevent dual-issue.
6837 if((regs[i].regmap[r]&63)==rt1[i-1]) wont_dirty_i|=1<<r;
6838 if((regs[i].regmap[r]&63)==rt2[i-1]) wont_dirty_i|=1<<r;
6844 will_dirty[i]=will_dirty_i;
6845 wont_dirty[i]=wont_dirty_i;
6846 // Mark registers that won't be dirtied as not dirty
6848 /*printf("wr (%d,%d) %x will:",istart,iend,start+i*4);
6849 for(r=0;r<HOST_REGS;r++) {
6850 if((will_dirty_i>>r)&1) {
6856 //if(i==istart||(itype[i-1]!=RJUMP&&itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=FJUMP)) {
6857 regs[i].dirty|=will_dirty_i;
6858 #ifndef DESTRUCTIVE_WRITEBACK
6859 regs[i].dirty&=wont_dirty_i;
6860 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
6862 if(i<iend-1&&itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
6863 for(r=0;r<HOST_REGS;r++) {
6864 if(r!=EXCLUDE_REG) {
6865 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
6866 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
6867 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6875 for(r=0;r<HOST_REGS;r++) {
6876 if(r!=EXCLUDE_REG) {
6877 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
6878 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
6879 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6887 // Deal with changed mappings
6888 temp_will_dirty=will_dirty_i;
6889 temp_wont_dirty=wont_dirty_i;
6890 for(r=0;r<HOST_REGS;r++) {
6891 if(r!=EXCLUDE_REG) {
6893 if(regs[i].regmap[r]==regmap_pre[i][r]) {
6895 #ifndef DESTRUCTIVE_WRITEBACK
6896 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6898 regs[i].wasdirty|=will_dirty_i&(1<<r);
6901 else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
6902 // Register moved to a different register
6903 will_dirty_i&=~(1<<r);
6904 wont_dirty_i&=~(1<<r);
6905 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
6906 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
6908 #ifndef DESTRUCTIVE_WRITEBACK
6909 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6911 regs[i].wasdirty|=will_dirty_i&(1<<r);
6915 will_dirty_i&=~(1<<r);
6916 wont_dirty_i&=~(1<<r);
6917 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6918 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6919 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6922 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
6932 void disassemble_inst(int i)
6934 if (bt[i]) printf("*"); else printf(" ");
6937 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6939 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;
6941 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;
6943 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6945 if (opcode[i]==0x9&&rt1[i]!=31)
6946 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i]);
6948 printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
6951 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],ba[i]);break;
6953 if(opcode[i]==0xf) //LUI
6954 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],rt1[i],imm[i]&0xffff);
6956 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
6960 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
6964 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rs2[i],rs1[i],imm[i]);
6968 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i],rs2[i]);
6971 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rs1[i],rs2[i]);
6974 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
6977 if((opcode2[i]&0x1d)==0x10)
6978 printf (" %x: %s r%d\n",start+i*4,insn[i],rt1[i]);
6979 else if((opcode2[i]&0x1d)==0x11)
6980 printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
6982 printf (" %x: %s\n",start+i*4,insn[i]);
6986 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC0
6987 else if(opcode2[i]==4)
6988 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC0
6989 else printf (" %x: %s\n",start+i*4,insn[i]);
6993 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC1
6994 else if(opcode2[i]>3)
6995 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC1
6996 else printf (" %x: %s\n",start+i*4,insn[i]);
7000 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC2
7001 else if(opcode2[i]>3)
7002 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC2
7003 else printf (" %x: %s\n",start+i*4,insn[i]);
7006 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
7009 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
7012 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
7015 //printf (" %s %8x\n",insn[i],source[i]);
7016 printf (" %x: %s\n",start+i*4,insn[i]);
7020 static void disassemble_inst(int i) {}
7023 #define DRC_TEST_VAL 0x74657374
7025 static int new_dynarec_test(void)
7027 int (*testfunc)(void) = (void *)out;
7031 beginning = start_block();
7032 emit_movimm(DRC_TEST_VAL,0); // test
7035 end_block(beginning);
7036 SysPrintf("testing if we can run recompiled code..\n");
7038 if (ret == DRC_TEST_VAL)
7039 SysPrintf("test passed.\n");
7041 SysPrintf("test failed: %08x\n", ret);
7042 out=(u_char *)BASE_ADDR;
7043 return ret == DRC_TEST_VAL;
7046 // clear the state completely, instead of just marking
7047 // things invalid like invalidate_all_pages() does
7048 void new_dynarec_clear_full()
7051 out=(u_char *)BASE_ADDR;
7052 memset(invalid_code,1,sizeof(invalid_code));
7053 memset(hash_table,0xff,sizeof(hash_table));
7054 memset(mini_ht,-1,sizeof(mini_ht));
7055 memset(restore_candidate,0,sizeof(restore_candidate));
7056 memset(shadow,0,sizeof(shadow));
7058 expirep=16384; // Expiry pointer, +2 blocks
7059 pending_exception=0;
7062 inv_code_start=inv_code_end=~0;
7064 for(n=0;n<4096;n++) ll_clear(jump_in+n);
7065 for(n=0;n<4096;n++) ll_clear(jump_out+n);
7066 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
7069 void new_dynarec_init()
7071 SysPrintf("Init new dynarec\n");
7073 // allocate/prepare a buffer for translation cache
7074 // see assem_arm.h for some explanation
7075 #if defined(BASE_ADDR_FIXED)
7076 if (mmap (translation_cache, 1 << TARGET_SIZE_2,
7077 PROT_READ | PROT_WRITE | PROT_EXEC,
7078 MAP_PRIVATE | MAP_ANONYMOUS,
7079 -1, 0) != translation_cache) {
7080 SysPrintf("mmap() failed: %s\n", strerror(errno));
7081 SysPrintf("disable BASE_ADDR_FIXED and recompile\n");
7084 #elif defined(BASE_ADDR_DYNAMIC)
7086 sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
7088 SysPrintf("sceKernelAllocMemBlockForVM failed\n");
7089 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&translation_cache);
7091 SysPrintf("sceKernelGetMemBlockBase failed\n");
7093 translation_cache = mmap (NULL, 1 << TARGET_SIZE_2,
7094 PROT_READ | PROT_WRITE | PROT_EXEC,
7095 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
7096 if (translation_cache == MAP_FAILED) {
7097 SysPrintf("mmap() failed: %s\n", strerror(errno));
7102 #ifndef NO_WRITE_EXEC
7103 // not all systems allow execute in data segment by default
7104 if (mprotect((void *)BASE_ADDR, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
7105 SysPrintf("mprotect() failed: %s\n", strerror(errno));
7108 out=(u_char *)BASE_ADDR;
7109 cycle_multiplier=200;
7110 new_dynarec_clear_full();
7112 // Copy this into local area so we don't have to put it in every literal pool
7113 invc_ptr=invalid_code;
7118 ram_offset=(u_int)rdram-0x80000000;
7121 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
7124 void new_dynarec_cleanup()
7127 #if defined(BASE_ADDR_FIXED) || defined(BASE_ADDR_DYNAMIC)
7129 sceKernelFreeMemBlock(sceBlock);
7132 if (munmap ((void *)BASE_ADDR, 1<<TARGET_SIZE_2) < 0)
7133 SysPrintf("munmap() failed\n");
7136 for(n=0;n<4096;n++) ll_clear(jump_in+n);
7137 for(n=0;n<4096;n++) ll_clear(jump_out+n);
7138 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
7140 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
7144 static u_int *get_source_start(u_int addr, u_int *limit)
7146 if (addr < 0x00200000 ||
7147 (0xa0000000 <= addr && addr < 0xa0200000)) {
7148 // used for BIOS calls mostly?
7149 *limit = (addr&0xa0000000)|0x00200000;
7150 return (u_int *)((u_int)rdram + (addr&0x1fffff));
7152 else if (!Config.HLE && (
7153 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
7154 (0xbfc00000 <= addr && addr < 0xbfc80000))) {
7156 *limit = (addr & 0xfff00000) | 0x80000;
7157 return (u_int *)((u_int)psxR + (addr&0x7ffff));
7159 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
7160 *limit = (addr & 0x80600000) + 0x00200000;
7161 return (u_int *)((u_int)rdram + (addr&0x1fffff));
7166 static u_int scan_for_ret(u_int addr)
7171 mem = get_source_start(addr, &limit);
7175 if (limit > addr + 0x1000)
7176 limit = addr + 0x1000;
7177 for (; addr < limit; addr += 4, mem++) {
7178 if (*mem == 0x03e00008) // jr $ra
7184 struct savestate_block {
7189 static int addr_cmp(const void *p1_, const void *p2_)
7191 const struct savestate_block *p1 = p1_, *p2 = p2_;
7192 return p1->addr - p2->addr;
7195 int new_dynarec_save_blocks(void *save, int size)
7197 struct savestate_block *blocks = save;
7198 int maxcount = size / sizeof(blocks[0]);
7199 struct savestate_block tmp_blocks[1024];
7200 struct ll_entry *head;
7201 int p, s, d, o, bcnt;
7205 for (p = 0; p < sizeof(jump_in) / sizeof(jump_in[0]); p++) {
7207 for (head = jump_in[p]; head != NULL; head = head->next) {
7208 tmp_blocks[bcnt].addr = head->vaddr;
7209 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
7214 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
7216 addr = tmp_blocks[0].addr;
7217 for (s = d = 0; s < bcnt; s++) {
7218 if (tmp_blocks[s].addr < addr)
7220 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
7221 tmp_blocks[d++] = tmp_blocks[s];
7222 addr = scan_for_ret(tmp_blocks[s].addr);
7225 if (o + d > maxcount)
7227 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
7231 return o * sizeof(blocks[0]);
7234 void new_dynarec_load_blocks(const void *save, int size)
7236 const struct savestate_block *blocks = save;
7237 int count = size / sizeof(blocks[0]);
7238 u_int regs_save[32];
7242 get_addr(psxRegs.pc);
7244 // change GPRs for speculation to at least partially work..
7245 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
7246 for (i = 1; i < 32; i++)
7247 psxRegs.GPR.r[i] = 0x80000000;
7249 for (b = 0; b < count; b++) {
7250 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
7252 psxRegs.GPR.r[i] = 0x1f800000;
7255 get_addr(blocks[b].addr);
7257 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
7259 psxRegs.GPR.r[i] = 0x80000000;
7263 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
7266 int new_recompile_block(int addr)
7268 u_int pagelimit = 0;
7269 u_int state_rflags = 0;
7272 assem_debug("NOTCOMPILED: addr = %x -> %x\n", (int)addr, (int)out);
7273 //printf("NOTCOMPILED: addr = %x -> %x\n", (int)addr, (int)out);
7274 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
7276 //printf("TRACE: count=%d next=%d (checksum %x)\n",Count,next_interupt,mchecksum());
7277 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
7278 /*if(Count>=312978186) {
7283 // this is just for speculation
7284 for (i = 1; i < 32; i++) {
7285 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
7286 state_rflags |= 1 << i;
7289 start = (u_int)addr&~3;
7290 //assert(((u_int)addr&1)==0);
7291 new_dynarec_did_compile=1;
7292 if (Config.HLE && start == 0x80001000) // hlecall
7294 // XXX: is this enough? Maybe check hleSoftCall?
7295 void *beginning=start_block();
7296 u_int page=get_page(start);
7298 invalid_code[start>>12]=0;
7299 emit_movimm(start,0);
7300 emit_writeword(0,(int)&pcaddr);
7301 emit_jmp((int)new_dyna_leave);
7303 end_block(beginning);
7304 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
7308 source = get_source_start(start, &pagelimit);
7309 if (source == NULL) {
7310 SysPrintf("Compile at bogus memory address: %08x\n", addr);
7314 /* Pass 1: disassemble */
7315 /* Pass 2: register dependencies, branch targets */
7316 /* Pass 3: register allocation */
7317 /* Pass 4: branch dependencies */
7318 /* Pass 5: pre-alloc */
7319 /* Pass 6: optimize clean/dirty state */
7320 /* Pass 7: flag 32-bit registers */
7321 /* Pass 8: assembly */
7322 /* Pass 9: linker */
7323 /* Pass 10: garbage collection / free memory */
7327 unsigned int type,op,op2;
7329 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
7331 /* Pass 1 disassembly */
7333 for(i=0;!done;i++) {
7334 bt[i]=0;likely[i]=0;ooo[i]=0;op2=0;
7335 minimum_free_regs[i]=0;
7336 opcode[i]=op=source[i]>>26;
7339 case 0x00: strcpy(insn[i],"special"); type=NI;
7343 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
7344 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
7345 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
7346 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
7347 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
7348 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
7349 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
7350 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
7351 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
7352 case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
7353 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
7354 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
7355 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
7356 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
7357 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
7358 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
7359 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
7360 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
7361 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
7362 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
7363 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
7364 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
7365 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
7366 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
7367 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
7368 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
7369 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
7370 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
7371 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
7372 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
7373 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
7374 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
7375 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
7376 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
7377 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
7379 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
7380 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
7381 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
7382 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
7383 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
7384 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
7385 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
7386 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
7387 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
7388 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
7389 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
7390 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
7391 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
7392 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
7393 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
7394 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
7395 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
7399 case 0x01: strcpy(insn[i],"regimm"); type=NI;
7400 op2=(source[i]>>16)&0x1f;
7403 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
7404 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
7405 case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
7406 case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
7407 case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
7408 case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
7409 case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
7410 case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
7411 case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
7412 case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
7413 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
7414 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
7415 case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
7416 case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
7419 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
7420 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
7421 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
7422 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
7423 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
7424 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
7425 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
7426 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
7427 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
7428 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
7429 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
7430 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
7431 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
7432 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
7433 case 0x10: strcpy(insn[i],"cop0"); type=NI;
7434 op2=(source[i]>>21)&0x1f;
7437 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
7438 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
7439 case 0x10: strcpy(insn[i],"tlb"); type=NI;
7440 switch(source[i]&0x3f)
7442 case 0x01: strcpy(insn[i],"TLBR"); type=COP0; break;
7443 case 0x02: strcpy(insn[i],"TLBWI"); type=COP0; break;
7444 case 0x06: strcpy(insn[i],"TLBWR"); type=COP0; break;
7445 case 0x08: strcpy(insn[i],"TLBP"); type=COP0; break;
7446 case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
7447 //case 0x18: strcpy(insn[i],"ERET"); type=COP0; break;
7451 case 0x11: strcpy(insn[i],"cop1"); type=NI;
7452 op2=(source[i]>>21)&0x1f;
7455 case 0x00: strcpy(insn[i],"MFC1"); type=COP1; break;
7456 case 0x01: strcpy(insn[i],"DMFC1"); type=COP1; break;
7457 case 0x02: strcpy(insn[i],"CFC1"); type=COP1; break;
7458 case 0x04: strcpy(insn[i],"MTC1"); type=COP1; break;
7459 case 0x05: strcpy(insn[i],"DMTC1"); type=COP1; break;
7460 case 0x06: strcpy(insn[i],"CTC1"); type=COP1; break;
7461 case 0x08: strcpy(insn[i],"BC1"); type=FJUMP;
7462 switch((source[i]>>16)&0x3)
7464 case 0x00: strcpy(insn[i],"BC1F"); break;
7465 case 0x01: strcpy(insn[i],"BC1T"); break;
7466 case 0x02: strcpy(insn[i],"BC1FL"); break;
7467 case 0x03: strcpy(insn[i],"BC1TL"); break;
7470 case 0x10: strcpy(insn[i],"C1.S"); type=NI;
7471 switch(source[i]&0x3f)
7473 case 0x00: strcpy(insn[i],"ADD.S"); type=FLOAT; break;
7474 case 0x01: strcpy(insn[i],"SUB.S"); type=FLOAT; break;
7475 case 0x02: strcpy(insn[i],"MUL.S"); type=FLOAT; break;
7476 case 0x03: strcpy(insn[i],"DIV.S"); type=FLOAT; break;
7477 case 0x04: strcpy(insn[i],"SQRT.S"); type=FLOAT; break;
7478 case 0x05: strcpy(insn[i],"ABS.S"); type=FLOAT; break;
7479 case 0x06: strcpy(insn[i],"MOV.S"); type=FLOAT; break;
7480 case 0x07: strcpy(insn[i],"NEG.S"); type=FLOAT; break;
7481 case 0x08: strcpy(insn[i],"ROUND.L.S"); type=FCONV; break;
7482 case 0x09: strcpy(insn[i],"TRUNC.L.S"); type=FCONV; break;
7483 case 0x0A: strcpy(insn[i],"CEIL.L.S"); type=FCONV; break;
7484 case 0x0B: strcpy(insn[i],"FLOOR.L.S"); type=FCONV; break;
7485 case 0x0C: strcpy(insn[i],"ROUND.W.S"); type=FCONV; break;
7486 case 0x0D: strcpy(insn[i],"TRUNC.W.S"); type=FCONV; break;
7487 case 0x0E: strcpy(insn[i],"CEIL.W.S"); type=FCONV; break;
7488 case 0x0F: strcpy(insn[i],"FLOOR.W.S"); type=FCONV; break;
7489 case 0x21: strcpy(insn[i],"CVT.D.S"); type=FCONV; break;
7490 case 0x24: strcpy(insn[i],"CVT.W.S"); type=FCONV; break;
7491 case 0x25: strcpy(insn[i],"CVT.L.S"); type=FCONV; break;
7492 case 0x30: strcpy(insn[i],"C.F.S"); type=FCOMP; break;
7493 case 0x31: strcpy(insn[i],"C.UN.S"); type=FCOMP; break;
7494 case 0x32: strcpy(insn[i],"C.EQ.S"); type=FCOMP; break;
7495 case 0x33: strcpy(insn[i],"C.UEQ.S"); type=FCOMP; break;
7496 case 0x34: strcpy(insn[i],"C.OLT.S"); type=FCOMP; break;
7497 case 0x35: strcpy(insn[i],"C.ULT.S"); type=FCOMP; break;
7498 case 0x36: strcpy(insn[i],"C.OLE.S"); type=FCOMP; break;
7499 case 0x37: strcpy(insn[i],"C.ULE.S"); type=FCOMP; break;
7500 case 0x38: strcpy(insn[i],"C.SF.S"); type=FCOMP; break;
7501 case 0x39: strcpy(insn[i],"C.NGLE.S"); type=FCOMP; break;
7502 case 0x3A: strcpy(insn[i],"C.SEQ.S"); type=FCOMP; break;
7503 case 0x3B: strcpy(insn[i],"C.NGL.S"); type=FCOMP; break;
7504 case 0x3C: strcpy(insn[i],"C.LT.S"); type=FCOMP; break;
7505 case 0x3D: strcpy(insn[i],"C.NGE.S"); type=FCOMP; break;
7506 case 0x3E: strcpy(insn[i],"C.LE.S"); type=FCOMP; break;
7507 case 0x3F: strcpy(insn[i],"C.NGT.S"); type=FCOMP; break;
7510 case 0x11: strcpy(insn[i],"C1.D"); type=NI;
7511 switch(source[i]&0x3f)
7513 case 0x00: strcpy(insn[i],"ADD.D"); type=FLOAT; break;
7514 case 0x01: strcpy(insn[i],"SUB.D"); type=FLOAT; break;
7515 case 0x02: strcpy(insn[i],"MUL.D"); type=FLOAT; break;
7516 case 0x03: strcpy(insn[i],"DIV.D"); type=FLOAT; break;
7517 case 0x04: strcpy(insn[i],"SQRT.D"); type=FLOAT; break;
7518 case 0x05: strcpy(insn[i],"ABS.D"); type=FLOAT; break;
7519 case 0x06: strcpy(insn[i],"MOV.D"); type=FLOAT; break;
7520 case 0x07: strcpy(insn[i],"NEG.D"); type=FLOAT; break;
7521 case 0x08: strcpy(insn[i],"ROUND.L.D"); type=FCONV; break;
7522 case 0x09: strcpy(insn[i],"TRUNC.L.D"); type=FCONV; break;
7523 case 0x0A: strcpy(insn[i],"CEIL.L.D"); type=FCONV; break;
7524 case 0x0B: strcpy(insn[i],"FLOOR.L.D"); type=FCONV; break;
7525 case 0x0C: strcpy(insn[i],"ROUND.W.D"); type=FCONV; break;
7526 case 0x0D: strcpy(insn[i],"TRUNC.W.D"); type=FCONV; break;
7527 case 0x0E: strcpy(insn[i],"CEIL.W.D"); type=FCONV; break;
7528 case 0x0F: strcpy(insn[i],"FLOOR.W.D"); type=FCONV; break;
7529 case 0x20: strcpy(insn[i],"CVT.S.D"); type=FCONV; break;
7530 case 0x24: strcpy(insn[i],"CVT.W.D"); type=FCONV; break;
7531 case 0x25: strcpy(insn[i],"CVT.L.D"); type=FCONV; break;
7532 case 0x30: strcpy(insn[i],"C.F.D"); type=FCOMP; break;
7533 case 0x31: strcpy(insn[i],"C.UN.D"); type=FCOMP; break;
7534 case 0x32: strcpy(insn[i],"C.EQ.D"); type=FCOMP; break;
7535 case 0x33: strcpy(insn[i],"C.UEQ.D"); type=FCOMP; break;
7536 case 0x34: strcpy(insn[i],"C.OLT.D"); type=FCOMP; break;
7537 case 0x35: strcpy(insn[i],"C.ULT.D"); type=FCOMP; break;
7538 case 0x36: strcpy(insn[i],"C.OLE.D"); type=FCOMP; break;
7539 case 0x37: strcpy(insn[i],"C.ULE.D"); type=FCOMP; break;
7540 case 0x38: strcpy(insn[i],"C.SF.D"); type=FCOMP; break;
7541 case 0x39: strcpy(insn[i],"C.NGLE.D"); type=FCOMP; break;
7542 case 0x3A: strcpy(insn[i],"C.SEQ.D"); type=FCOMP; break;
7543 case 0x3B: strcpy(insn[i],"C.NGL.D"); type=FCOMP; break;
7544 case 0x3C: strcpy(insn[i],"C.LT.D"); type=FCOMP; break;
7545 case 0x3D: strcpy(insn[i],"C.NGE.D"); type=FCOMP; break;
7546 case 0x3E: strcpy(insn[i],"C.LE.D"); type=FCOMP; break;
7547 case 0x3F: strcpy(insn[i],"C.NGT.D"); type=FCOMP; break;
7550 case 0x14: strcpy(insn[i],"C1.W"); type=NI;
7551 switch(source[i]&0x3f)
7553 case 0x20: strcpy(insn[i],"CVT.S.W"); type=FCONV; break;
7554 case 0x21: strcpy(insn[i],"CVT.D.W"); type=FCONV; break;
7557 case 0x15: strcpy(insn[i],"C1.L"); type=NI;
7558 switch(source[i]&0x3f)
7560 case 0x20: strcpy(insn[i],"CVT.S.L"); type=FCONV; break;
7561 case 0x21: strcpy(insn[i],"CVT.D.L"); type=FCONV; break;
7567 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
7568 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
7569 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
7570 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
7571 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
7572 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
7573 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
7574 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
7576 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
7577 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
7578 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
7579 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
7580 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
7581 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
7582 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
7584 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
7586 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
7587 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
7588 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
7589 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
7591 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
7592 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
7594 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
7595 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
7596 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
7597 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
7599 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
7600 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
7601 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
7603 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
7604 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
7606 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
7607 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
7608 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
7610 case 0x12: strcpy(insn[i],"COP2"); type=NI;
7611 op2=(source[i]>>21)&0x1f;
7613 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
7614 if (gte_handlers[source[i]&0x3f]!=NULL) {
7615 if (gte_regnames[source[i]&0x3f]!=NULL)
7616 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
7618 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
7624 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
7625 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
7626 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
7627 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
7630 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
7631 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
7632 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
7633 default: strcpy(insn[i],"???"); type=NI;
7634 SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
7639 /* Get registers/immediates */
7645 gte_rs[i]=gte_rt[i]=0;
7648 rs1[i]=(source[i]>>21)&0x1f;
7650 rt1[i]=(source[i]>>16)&0x1f;
7652 imm[i]=(short)source[i];
7656 rs1[i]=(source[i]>>21)&0x1f;
7657 rs2[i]=(source[i]>>16)&0x1f;
7660 imm[i]=(short)source[i];
7661 if(op==0x2c||op==0x2d||op==0x3f) us1[i]=rs2[i]; // 64-bit SDL/SDR/SD
7664 // LWL/LWR only load part of the register,
7665 // therefore the target register must be treated as a source too
7666 rs1[i]=(source[i]>>21)&0x1f;
7667 rs2[i]=(source[i]>>16)&0x1f;
7668 rt1[i]=(source[i]>>16)&0x1f;
7670 imm[i]=(short)source[i];
7671 if(op==0x1a||op==0x1b) us1[i]=rs2[i]; // LDR/LDL
7672 if(op==0x26) dep1[i]=rt1[i]; // LWR
7675 if (op==0x0f) rs1[i]=0; // LUI instruction has no source register
7676 else rs1[i]=(source[i]>>21)&0x1f;
7678 rt1[i]=(source[i]>>16)&0x1f;
7680 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
7681 imm[i]=(unsigned short)source[i];
7683 imm[i]=(short)source[i];
7685 if(op==0x18||op==0x19) us1[i]=rs1[i]; // DADDI/DADDIU
7686 if(op==0x0a||op==0x0b) us1[i]=rs1[i]; // SLTI/SLTIU
7687 if(op==0x0d||op==0x0e) dep1[i]=rs1[i]; // ORI/XORI
7694 // The JAL instruction writes to r31.
7701 rs1[i]=(source[i]>>21)&0x1f;
7705 // The JALR instruction writes to rd.
7707 rt1[i]=(source[i]>>11)&0x1f;
7712 rs1[i]=(source[i]>>21)&0x1f;
7713 rs2[i]=(source[i]>>16)&0x1f;
7716 if(op&2) { // BGTZ/BLEZ
7724 rs1[i]=(source[i]>>21)&0x1f;
7729 if(op2&0x10) { // BxxAL
7731 // NOTE: If the branch is not taken, r31 is still overwritten
7733 likely[i]=(op2&2)>>1;
7740 likely[i]=((source[i])>>17)&1;
7743 rs1[i]=(source[i]>>21)&0x1f; // source
7744 rs2[i]=(source[i]>>16)&0x1f; // subtract amount
7745 rt1[i]=(source[i]>>11)&0x1f; // destination
7747 if(op2==0x2a||op2==0x2b) { // SLT/SLTU
7748 us1[i]=rs1[i];us2[i]=rs2[i];
7750 else if(op2>=0x24&&op2<=0x27) { // AND/OR/XOR/NOR
7751 dep1[i]=rs1[i];dep2[i]=rs2[i];
7753 else if(op2>=0x2c&&op2<=0x2f) { // DADD/DSUB
7754 dep1[i]=rs1[i];dep2[i]=rs2[i];
7758 rs1[i]=(source[i]>>21)&0x1f; // source
7759 rs2[i]=(source[i]>>16)&0x1f; // divisor
7762 if (op2>=0x1c&&op2<=0x1f) { // DMULT/DMULTU/DDIV/DDIVU
7763 us1[i]=rs1[i];us2[i]=rs2[i];
7771 if(op2==0x10) rs1[i]=HIREG; // MFHI
7772 if(op2==0x11) rt1[i]=HIREG; // MTHI
7773 if(op2==0x12) rs1[i]=LOREG; // MFLO
7774 if(op2==0x13) rt1[i]=LOREG; // MTLO
7775 if((op2&0x1d)==0x10) rt1[i]=(source[i]>>11)&0x1f; // MFxx
7776 if((op2&0x1d)==0x11) rs1[i]=(source[i]>>21)&0x1f; // MTxx
7780 rs1[i]=(source[i]>>16)&0x1f; // target of shift
7781 rs2[i]=(source[i]>>21)&0x1f; // shift amount
7782 rt1[i]=(source[i]>>11)&0x1f; // destination
7784 // DSLLV/DSRLV/DSRAV are 64-bit
7785 if(op2>=0x14&&op2<=0x17) us1[i]=rs1[i];
7788 rs1[i]=(source[i]>>16)&0x1f;
7790 rt1[i]=(source[i]>>11)&0x1f;
7792 imm[i]=(source[i]>>6)&0x1f;
7793 // DSxx32 instructions
7794 if(op2>=0x3c) imm[i]|=0x20;
7795 // DSLL/DSRL/DSRA/DSRA32/DSRL32 but not DSLL32 require 64-bit source
7796 if(op2>=0x38&&op2!=0x3c) us1[i]=rs1[i];
7803 if(op2==0) rt1[i]=(source[i]>>16)&0x1F; // MFC0
7804 if(op2==4) rs1[i]=(source[i]>>16)&0x1F; // MTC0
7805 if(op2==4&&((source[i]>>11)&0x1f)==12) rt2[i]=CSREG; // Status
7806 if(op2==16) if((source[i]&0x3f)==0x18) rs2[i]=CCREG; // ERET
7813 if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
7814 if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
7815 if(op2==5) us1[i]=rs1[i]; // DMTC1
7823 if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC2/CFC2
7824 if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC2/CTC2
7826 int gr=(source[i]>>11)&0x1F;
7829 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
7830 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
7831 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
7832 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
7836 rs1[i]=(source[i]>>21)&0x1F;
7840 imm[i]=(short)source[i];
7843 rs1[i]=(source[i]>>21)&0x1F;
7847 imm[i]=(short)source[i];
7848 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
7849 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
7856 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
7857 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
7858 gte_rt[i]|=1ll<<63; // every op changes flags
7859 if((source[i]&0x3f)==GTE_MVMVA) {
7860 int v = (source[i] >> 15) & 3;
7861 gte_rs[i]&=~0xe3fll;
7862 if(v==3) gte_rs[i]|=0xe00ll;
7863 else gte_rs[i]|=3ll<<(v*2);
7893 /* Calculate branch target addresses */
7895 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7896 else if(type==CJUMP&&rs1[i]==rs2[i]&&(op&1))
7897 ba[i]=start+i*4+8; // Ignore never taken branch
7898 else if(type==SJUMP&&rs1[i]==0&&!(op2&1))
7899 ba[i]=start+i*4+8; // Ignore never taken branch
7900 else if(type==CJUMP||type==SJUMP||type==FJUMP)
7901 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7903 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==FJUMP)) {
7905 // branch in delay slot?
7906 if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP||type==FJUMP) {
7907 // don't handle first branch and call interpreter if it's hit
7908 SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
7911 // basic load delay detection
7912 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&rt1[i]!=0) {
7913 int t=(ba[i-1]-start)/4;
7914 if(0 <= t && t < i &&(rt1[i]==rs1[t]||rt1[i]==rs2[t])&&itype[t]!=CJUMP&&itype[t]!=SJUMP) {
7915 // jump target wants DS result - potential load delay effect
7916 SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
7918 bt[t+1]=1; // expected return from interpreter
7920 else if(i>=2&&rt1[i-2]==2&&rt1[i]==2&&rs1[i]!=2&&rs2[i]!=2&&rs1[i-1]!=2&&rs2[i-1]!=2&&
7921 !(i>=3&&(itype[i-3]==RJUMP||itype[i-3]==UJUMP||itype[i-3]==CJUMP||itype[i-3]==SJUMP))) {
7922 // v0 overwrite like this is a sign of trouble, bail out
7923 SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
7929 rs2[i-1]=rt1[i-1]=rt2[i-1]=0;
7933 i--; // don't compile the DS
7936 /* Is this the end of the block? */
7937 if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
7938 if(rt1[i-1]==0) { // Continue past subroutine call (JAL)
7942 if(stop_after_jal) done=1;
7944 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7946 // Don't recompile stuff that's already compiled
7947 if(check_addr(start+i*4+4)) done=1;
7948 // Don't get too close to the limit
7949 if(i>MAXBLOCK/2) done=1;
7951 if(itype[i]==SYSCALL&&stop_after_jal) done=1;
7952 if(itype[i]==HLECALL||itype[i]==INTCALL) done=2;
7954 // Does the block continue due to a branch?
7957 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7958 if(ba[j]==start+i*4+4) done=j=0;
7959 if(ba[j]==start+i*4+8) done=j=0;
7962 //assert(i<MAXBLOCK-1);
7963 if(start+i*4==pagelimit-4) done=1;
7964 assert(start+i*4<pagelimit);
7965 if (i==MAXBLOCK-1) done=1;
7966 // Stop if we're compiling junk
7967 if(itype[i]==NI&&opcode[i]==0x11) {
7968 done=stop_after_jal=1;
7969 SysPrintf("Disabled speculative precompilation\n");
7973 if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==RJUMP||itype[i-1]==FJUMP) {
7974 if(start+i*4==pagelimit) {
7980 /* Pass 2 - Register dependencies and branch targets */
7982 unneeded_registers(0,slen-1,0);
7984 /* Pass 3 - Register allocation */
7986 struct regstat current; // Current register allocations/status
7989 current.u=unneeded_reg[0];
7990 current.uu=unneeded_reg_upper[0];
7991 clear_all_regs(current.regmap);
7992 alloc_reg(¤t,0,CCREG);
7993 dirty_reg(¤t,CCREG);
7996 current.waswritten=0;
8002 // First instruction is delay slot
8007 unneeded_reg_upper[0]=1;
8008 current.regmap[HOST_BTREG]=BTREG;
8016 for(hr=0;hr<HOST_REGS;hr++)
8018 // Is this really necessary?
8019 if(current.regmap[hr]==0) current.regmap[hr]=-1;
8022 current.waswritten=0;
8026 if((opcode[i-2]&0x2f)==0x05) // BNE/BNEL
8028 if(rs1[i-2]==0||rs2[i-2]==0)
8031 current.is32|=1LL<<rs1[i-2];
8032 int hr=get_reg(current.regmap,rs1[i-2]|64);
8033 if(hr>=0) current.regmap[hr]=-1;
8036 current.is32|=1LL<<rs2[i-2];
8037 int hr=get_reg(current.regmap,rs2[i-2]|64);
8038 if(hr>=0) current.regmap[hr]=-1;
8045 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
8046 regs[i].wasconst=current.isconst;
8047 regs[i].was32=current.is32;
8048 regs[i].wasdirty=current.dirty;
8049 regs[i].loadedconst=0;
8050 if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP) {
8052 current.u=unneeded_reg[i+1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
8053 current.uu=unneeded_reg_upper[i+1]&~((1LL<<us1[i])|(1LL<<us2[i]));
8054 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8063 current.u=branch_unneeded_reg[i]&~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
8064 current.uu=branch_unneeded_reg_upper[i]&~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
8065 if((~current.uu>>rt1[i+1])&1) current.uu&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
8066 current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
8067 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8070 } else { SysPrintf("oops, branch at end of block with no delay slot\n");exit(1); }
8074 ds=0; // Skip delay slot, already allocated as part of branch
8075 // ...but we need to alloc it in case something jumps here
8077 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
8078 current.uu=branch_unneeded_reg_upper[i-1]&unneeded_reg_upper[i+1];
8080 current.u=branch_unneeded_reg[i-1];
8081 current.uu=branch_unneeded_reg_upper[i-1];
8083 current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
8084 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8085 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8088 struct regstat temp;
8089 memcpy(&temp,¤t,sizeof(current));
8090 temp.wasdirty=temp.dirty;
8091 temp.was32=temp.is32;
8092 // TODO: Take into account unconditional branches, as below
8093 delayslot_alloc(&temp,i);
8094 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
8095 regs[i].wasdirty=temp.wasdirty;
8096 regs[i].was32=temp.was32;
8097 regs[i].dirty=temp.dirty;
8098 regs[i].is32=temp.is32;
8102 // Create entry (branch target) regmap
8103 for(hr=0;hr<HOST_REGS;hr++)
8105 int r=temp.regmap[hr];
8107 if(r!=regmap_pre[i][hr]) {
8108 regs[i].regmap_entry[hr]=-1;
8113 if((current.u>>r)&1) {
8114 regs[i].regmap_entry[hr]=-1;
8115 regs[i].regmap[hr]=-1;
8116 //Don't clear regs in the delay slot as the branch might need them
8117 //current.regmap[hr]=-1;
8119 regs[i].regmap_entry[hr]=r;
8122 if((current.uu>>(r&63))&1) {
8123 regs[i].regmap_entry[hr]=-1;
8124 regs[i].regmap[hr]=-1;
8125 //Don't clear regs in the delay slot as the branch might need them
8126 //current.regmap[hr]=-1;
8128 regs[i].regmap_entry[hr]=r;
8132 // First instruction expects CCREG to be allocated
8133 if(i==0&&hr==HOST_CCREG)
8134 regs[i].regmap_entry[hr]=CCREG;
8136 regs[i].regmap_entry[hr]=-1;
8140 else { // Not delay slot
8143 //current.isconst=0; // DEBUG
8144 //current.wasconst=0; // DEBUG
8145 //regs[i].wasconst=0; // DEBUG
8146 clear_const(¤t,rt1[i]);
8147 alloc_cc(¤t,i);
8148 dirty_reg(¤t,CCREG);
8150 alloc_reg(¤t,i,31);
8151 dirty_reg(¤t,31);
8152 //assert(rs1[i+1]!=31&&rs2[i+1]!=31);
8153 //assert(rt1[i+1]!=rt1[i]);
8155 alloc_reg(¤t,i,PTEMP);
8157 //current.is32|=1LL<<rt1[i];
8160 delayslot_alloc(¤t,i+1);
8161 //current.isconst=0; // DEBUG
8163 //printf("i=%d, isconst=%x\n",i,current.isconst);
8166 //current.isconst=0;
8167 //current.wasconst=0;
8168 //regs[i].wasconst=0;
8169 clear_const(¤t,rs1[i]);
8170 clear_const(¤t,rt1[i]);
8171 alloc_cc(¤t,i);
8172 dirty_reg(¤t,CCREG);
8173 if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
8174 alloc_reg(¤t,i,rs1[i]);
8176 alloc_reg(¤t,i,rt1[i]);
8177 dirty_reg(¤t,rt1[i]);
8178 assert(rs1[i+1]!=rt1[i]&&rs2[i+1]!=rt1[i]);
8179 assert(rt1[i+1]!=rt1[i]);
8181 alloc_reg(¤t,i,PTEMP);
8185 if(rs1[i]==31) { // JALR
8186 alloc_reg(¤t,i,RHASH);
8187 #ifndef HOST_IMM_ADDR32
8188 alloc_reg(¤t,i,RHTBL);
8192 delayslot_alloc(¤t,i+1);
8194 // The delay slot overwrites our source register,
8195 // allocate a temporary register to hold the old value.
8199 delayslot_alloc(¤t,i+1);
8201 alloc_reg(¤t,i,RTEMP);
8203 //current.isconst=0; // DEBUG
8208 //current.isconst=0;
8209 //current.wasconst=0;
8210 //regs[i].wasconst=0;
8211 clear_const(¤t,rs1[i]);
8212 clear_const(¤t,rs2[i]);
8213 if((opcode[i]&0x3E)==4) // BEQ/BNE
8215 alloc_cc(¤t,i);
8216 dirty_reg(¤t,CCREG);
8217 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8218 if(rs2[i]) alloc_reg(¤t,i,rs2[i]);
8219 if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8221 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8222 if(rs2[i]) alloc_reg64(¤t,i,rs2[i]);
8224 if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
8225 (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1]))) {
8226 // The delay slot overwrites one of our conditions.
8227 // Allocate the branch condition registers instead.
8231 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8232 if(rs2[i]) alloc_reg(¤t,i,rs2[i]);
8233 if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8235 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8236 if(rs2[i]) alloc_reg64(¤t,i,rs2[i]);
8242 delayslot_alloc(¤t,i+1);
8246 if((opcode[i]&0x3E)==6) // BLEZ/BGTZ
8248 alloc_cc(¤t,i);
8249 dirty_reg(¤t,CCREG);
8250 alloc_reg(¤t,i,rs1[i]);
8251 if(!(current.is32>>rs1[i]&1))
8253 alloc_reg64(¤t,i,rs1[i]);
8255 if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
8256 // The delay slot overwrites one of our conditions.
8257 // Allocate the branch condition registers instead.
8261 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8262 if(!((current.is32>>rs1[i])&1))
8264 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8270 delayslot_alloc(¤t,i+1);
8274 // Don't alloc the delay slot yet because we might not execute it
8275 if((opcode[i]&0x3E)==0x14) // BEQL/BNEL
8280 alloc_cc(¤t,i);
8281 dirty_reg(¤t,CCREG);
8282 alloc_reg(¤t,i,rs1[i]);
8283 alloc_reg(¤t,i,rs2[i]);
8284 if(!((current.is32>>rs1[i])&(current.is32>>rs2[i])&1))
8286 alloc_reg64(¤t,i,rs1[i]);
8287 alloc_reg64(¤t,i,rs2[i]);
8291 if((opcode[i]&0x3E)==0x16) // BLEZL/BGTZL
8296 alloc_cc(¤t,i);
8297 dirty_reg(¤t,CCREG);
8298 alloc_reg(¤t,i,rs1[i]);
8299 if(!(current.is32>>rs1[i]&1))
8301 alloc_reg64(¤t,i,rs1[i]);
8305 //current.isconst=0;
8308 //current.isconst=0;
8309 //current.wasconst=0;
8310 //regs[i].wasconst=0;
8311 clear_const(¤t,rs1[i]);
8312 clear_const(¤t,rt1[i]);
8313 //if((opcode2[i]&0x1E)==0x0) // BLTZ/BGEZ
8314 if((opcode2[i]&0x0E)==0x0) // BLTZ/BGEZ
8316 alloc_cc(¤t,i);
8317 dirty_reg(¤t,CCREG);
8318 alloc_reg(¤t,i,rs1[i]);
8319 if(!(current.is32>>rs1[i]&1))
8321 alloc_reg64(¤t,i,rs1[i]);
8323 if (rt1[i]==31) { // BLTZAL/BGEZAL
8324 alloc_reg(¤t,i,31);
8325 dirty_reg(¤t,31);
8326 //#ifdef REG_PREFETCH
8327 //alloc_reg(¤t,i,PTEMP);
8329 //current.is32|=1LL<<rt1[i];
8331 if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) // The delay slot overwrites the branch condition.
8332 ||(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))) { // DS touches $ra
8333 // Allocate the branch condition registers instead.
8337 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
8338 if(!((current.is32>>rs1[i])&1))
8340 if(rs1[i]) alloc_reg64(¤t,i,rs1[i]);
8346 delayslot_alloc(¤t,i+1);
8350 // Don't alloc the delay slot yet because we might not execute it
8351 if((opcode2[i]&0x1E)==0x2) // BLTZL/BGEZL
8356 alloc_cc(¤t,i);
8357 dirty_reg(¤t,CCREG);
8358 alloc_reg(¤t,i,rs1[i]);
8359 if(!(current.is32>>rs1[i]&1))
8361 alloc_reg64(¤t,i,rs1[i]);
8365 //current.isconst=0;
8371 if(likely[i]==0) // BC1F/BC1T
8373 // TODO: Theoretically we can run out of registers here on x86.
8374 // The delay slot can allocate up to six, and we need to check
8375 // CSREG before executing the delay slot. Possibly we can drop
8376 // the cycle count and then reload it after checking that the
8377 // FPU is in a usable state, or don't do out-of-order execution.
8378 alloc_cc(¤t,i);
8379 dirty_reg(¤t,CCREG);
8380 alloc_reg(¤t,i,FSREG);
8381 alloc_reg(¤t,i,CSREG);
8382 if(itype[i+1]==FCOMP) {
8383 // The delay slot overwrites the branch condition.
8384 // Allocate the branch condition registers instead.
8385 alloc_cc(¤t,i);
8386 dirty_reg(¤t,CCREG);
8387 alloc_reg(¤t,i,CSREG);
8388 alloc_reg(¤t,i,FSREG);
8392 delayslot_alloc(¤t,i+1);
8393 alloc_reg(¤t,i+1,CSREG);
8397 // Don't alloc the delay slot yet because we might not execute it
8398 if(likely[i]) // BC1FL/BC1TL
8400 alloc_cc(¤t,i);
8401 dirty_reg(¤t,CCREG);
8402 alloc_reg(¤t,i,CSREG);
8403 alloc_reg(¤t,i,FSREG);
8409 imm16_alloc(¤t,i);
8413 load_alloc(¤t,i);
8417 store_alloc(¤t,i);
8420 alu_alloc(¤t,i);
8423 shift_alloc(¤t,i);
8426 multdiv_alloc(¤t,i);
8429 shiftimm_alloc(¤t,i);
8432 mov_alloc(¤t,i);
8435 cop0_alloc(¤t,i);
8439 cop1_alloc(¤t,i);
8442 c1ls_alloc(¤t,i);
8445 c2ls_alloc(¤t,i);
8448 c2op_alloc(¤t,i);
8451 fconv_alloc(¤t,i);
8454 float_alloc(¤t,i);
8457 fcomp_alloc(¤t,i);
8462 syscall_alloc(¤t,i);
8465 pagespan_alloc(¤t,i);
8469 // Drop the upper half of registers that have become 32-bit
8470 current.uu|=current.is32&((1LL<<rt1[i])|(1LL<<rt2[i]));
8471 if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP&&itype[i]!=FJUMP) {
8472 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8473 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8476 current.uu|=current.is32&((1LL<<rt1[i+1])|(1LL<<rt2[i+1]));
8477 current.uu&=~((1LL<<us1[i+1])|(1LL<<us2[i+1]));
8478 if((~current.uu>>rt1[i+1])&1) current.uu&=~((1LL<<dep1[i+1])|(1LL<<dep2[i+1]));
8479 current.uu&=~((1LL<<us1[i])|(1LL<<us2[i]));
8483 // Create entry (branch target) regmap
8484 for(hr=0;hr<HOST_REGS;hr++)
8487 r=current.regmap[hr];
8489 if(r!=regmap_pre[i][hr]) {
8490 // TODO: delay slot (?)
8491 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
8492 if(or<0||(r&63)>=TEMPREG){
8493 regs[i].regmap_entry[hr]=-1;
8497 // Just move it to a different register
8498 regs[i].regmap_entry[hr]=r;
8499 // If it was dirty before, it's still dirty
8500 if((regs[i].wasdirty>>or)&1) dirty_reg(¤t,r&63);
8507 regs[i].regmap_entry[hr]=0;
8511 if((current.u>>r)&1) {
8512 regs[i].regmap_entry[hr]=-1;
8513 //regs[i].regmap[hr]=-1;
8514 current.regmap[hr]=-1;
8516 regs[i].regmap_entry[hr]=r;
8519 if((current.uu>>(r&63))&1) {
8520 regs[i].regmap_entry[hr]=-1;
8521 //regs[i].regmap[hr]=-1;
8522 current.regmap[hr]=-1;
8524 regs[i].regmap_entry[hr]=r;
8528 // Branches expect CCREG to be allocated at the target
8529 if(regmap_pre[i][hr]==CCREG)
8530 regs[i].regmap_entry[hr]=CCREG;
8532 regs[i].regmap_entry[hr]=-1;
8535 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
8538 if(i>0&&(itype[i-1]==STORE||itype[i-1]==STORELR||(itype[i-1]==C2LS&&opcode[i-1]==0x3a))&&(u_int)imm[i-1]<0x800)
8539 current.waswritten|=1<<rs1[i-1];
8540 current.waswritten&=~(1<<rt1[i]);
8541 current.waswritten&=~(1<<rt2[i]);
8542 if((itype[i]==STORE||itype[i]==STORELR||(itype[i]==C2LS&&opcode[i]==0x3a))&&(u_int)imm[i]>=0x800)
8543 current.waswritten&=~(1<<rs1[i]);
8545 /* Branch post-alloc */
8548 current.was32=current.is32;
8549 current.wasdirty=current.dirty;
8550 switch(itype[i-1]) {
8552 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8553 branch_regs[i-1].isconst=0;
8554 branch_regs[i-1].wasconst=0;
8555 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
8556 branch_regs[i-1].uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
8557 alloc_cc(&branch_regs[i-1],i-1);
8558 dirty_reg(&branch_regs[i-1],CCREG);
8559 if(rt1[i-1]==31) { // JAL
8560 alloc_reg(&branch_regs[i-1],i-1,31);
8561 dirty_reg(&branch_regs[i-1],31);
8562 branch_regs[i-1].is32|=1LL<<31;
8564 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8565 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
8568 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8569 branch_regs[i-1].isconst=0;
8570 branch_regs[i-1].wasconst=0;
8571 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
8572 branch_regs[i-1].uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
8573 alloc_cc(&branch_regs[i-1],i-1);
8574 dirty_reg(&branch_regs[i-1],CCREG);
8575 alloc_reg(&branch_regs[i-1],i-1,rs1[i-1]);
8576 if(rt1[i-1]!=0) { // JALR
8577 alloc_reg(&branch_regs[i-1],i-1,rt1[i-1]);
8578 dirty_reg(&branch_regs[i-1],rt1[i-1]);
8579 branch_regs[i-1].is32|=1LL<<rt1[i-1];
8582 if(rs1[i-1]==31) { // JALR
8583 alloc_reg(&branch_regs[i-1],i-1,RHASH);
8584 #ifndef HOST_IMM_ADDR32
8585 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
8589 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8590 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
8593 if((opcode[i-1]&0x3E)==4) // BEQ/BNE
8595 alloc_cc(¤t,i-1);
8596 dirty_reg(¤t,CCREG);
8597 if((rs1[i-1]&&(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]))||
8598 (rs2[i-1]&&(rs2[i-1]==rt1[i]||rs2[i-1]==rt2[i]))) {
8599 // The delay slot overwrote one of our conditions
8600 // Delay slot goes after the test (in order)
8601 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
8602 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
8603 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8606 delayslot_alloc(¤t,i);
8611 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
8612 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i-1])|(1LL<<us2[i-1]));
8613 // Alloc the branch condition registers
8614 if(rs1[i-1]) alloc_reg(¤t,i-1,rs1[i-1]);
8615 if(rs2[i-1]) alloc_reg(¤t,i-1,rs2[i-1]);
8616 if(!((current.is32>>rs1[i-1])&(current.is32>>rs2[i-1])&1))
8618 if(rs1[i-1]) alloc_reg64(¤t,i-1,rs1[i-1]);
8619 if(rs2[i-1]) alloc_reg64(¤t,i-1,rs2[i-1]);
8622 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8623 branch_regs[i-1].isconst=0;
8624 branch_regs[i-1].wasconst=0;
8625 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8626 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
8629 if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
8631 alloc_cc(¤t,i-1);
8632 dirty_reg(¤t,CCREG);
8633 if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
8634 // The delay slot overwrote the branch condition
8635 // Delay slot goes after the test (in order)
8636 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
8637 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
8638 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8641 delayslot_alloc(¤t,i);
8646 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
8647 current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
8648 // Alloc the branch condition register
8649 alloc_reg(¤t,i-1,rs1[i-1]);
8650 if(!(current.is32>>rs1[i-1]&1))
8652 alloc_reg64(¤t,i-1,rs1[i-1]);
8655 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8656 branch_regs[i-1].isconst=0;
8657 branch_regs[i-1].wasconst=0;
8658 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8659 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
8662 // Alloc the delay slot in case the branch is taken
8663 if((opcode[i-1]&0x3E)==0x14) // BEQL/BNEL
8665 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8666 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8667 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8668 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
8669 alloc_cc(&branch_regs[i-1],i);
8670 dirty_reg(&branch_regs[i-1],CCREG);
8671 delayslot_alloc(&branch_regs[i-1],i);
8672 branch_regs[i-1].isconst=0;
8673 alloc_reg(¤t,i,CCREG); // Not taken path
8674 dirty_reg(¤t,CCREG);
8675 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8678 if((opcode[i-1]&0x3E)==0x16) // BLEZL/BGTZL
8680 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8681 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8682 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8683 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
8684 alloc_cc(&branch_regs[i-1],i);
8685 dirty_reg(&branch_regs[i-1],CCREG);
8686 delayslot_alloc(&branch_regs[i-1],i);
8687 branch_regs[i-1].isconst=0;
8688 alloc_reg(¤t,i,CCREG); // Not taken path
8689 dirty_reg(¤t,CCREG);
8690 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8694 //if((opcode2[i-1]&0x1E)==0) // BLTZ/BGEZ
8695 if((opcode2[i-1]&0x0E)==0) // BLTZ/BGEZ
8697 alloc_cc(¤t,i-1);
8698 dirty_reg(¤t,CCREG);
8699 if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
8700 // The delay slot overwrote the branch condition
8701 // Delay slot goes after the test (in order)
8702 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
8703 current.uu=branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i]));
8704 if((~current.uu>>rt1[i])&1) current.uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]));
8707 delayslot_alloc(¤t,i);
8712 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
8713 current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
8714 // Alloc the branch condition register
8715 alloc_reg(¤t,i-1,rs1[i-1]);
8716 if(!(current.is32>>rs1[i-1]&1))
8718 alloc_reg64(¤t,i-1,rs1[i-1]);
8721 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8722 branch_regs[i-1].isconst=0;
8723 branch_regs[i-1].wasconst=0;
8724 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8725 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
8728 // Alloc the delay slot in case the branch is taken
8729 if((opcode2[i-1]&0x1E)==2) // BLTZL/BGEZL
8731 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8732 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8733 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8734 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
8735 alloc_cc(&branch_regs[i-1],i);
8736 dirty_reg(&branch_regs[i-1],CCREG);
8737 delayslot_alloc(&branch_regs[i-1],i);
8738 branch_regs[i-1].isconst=0;
8739 alloc_reg(¤t,i,CCREG); // Not taken path
8740 dirty_reg(¤t,CCREG);
8741 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8743 // FIXME: BLTZAL/BGEZAL
8744 if(opcode2[i-1]&0x10) { // BxxZAL
8745 alloc_reg(&branch_regs[i-1],i-1,31);
8746 dirty_reg(&branch_regs[i-1],31);
8747 branch_regs[i-1].is32|=1LL<<31;
8751 if(likely[i-1]==0) // BC1F/BC1T
8753 alloc_cc(¤t,i-1);
8754 dirty_reg(¤t,CCREG);
8755 if(itype[i]==FCOMP) {
8756 // The delay slot overwrote the branch condition
8757 // Delay slot goes after the test (in order)
8758 delayslot_alloc(¤t,i);
8763 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
8764 current.uu=branch_unneeded_reg_upper[i-1]&~(1LL<<us1[i-1]);
8765 // Alloc the branch condition register
8766 alloc_reg(¤t,i-1,FSREG);
8768 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8769 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8773 // Alloc the delay slot in case the branch is taken
8774 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8775 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8776 branch_regs[i-1].uu=(branch_unneeded_reg_upper[i-1]&~((1LL<<us1[i])|(1LL<<us2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
8777 if((~branch_regs[i-1].uu>>rt1[i])&1) branch_regs[i-1].uu&=~((1LL<<dep1[i])|(1LL<<dep2[i]))|1;
8778 alloc_cc(&branch_regs[i-1],i);
8779 dirty_reg(&branch_regs[i-1],CCREG);
8780 delayslot_alloc(&branch_regs[i-1],i);
8781 branch_regs[i-1].isconst=0;
8782 alloc_reg(¤t,i,CCREG); // Not taken path
8783 dirty_reg(¤t,CCREG);
8784 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8789 if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
8791 if(rt1[i-1]==31) // JAL/JALR
8793 // Subroutine call will return here, don't alloc any registers
8796 clear_all_regs(current.regmap);
8797 alloc_reg(¤t,i,CCREG);
8798 dirty_reg(¤t,CCREG);
8802 // Internal branch will jump here, match registers to caller
8803 current.is32=0x3FFFFFFFFLL;
8805 clear_all_regs(current.regmap);
8806 alloc_reg(¤t,i,CCREG);
8807 dirty_reg(¤t,CCREG);
8810 if(ba[j]==start+i*4+4) {
8811 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
8812 current.is32=branch_regs[j].is32;
8813 current.dirty=branch_regs[j].dirty;
8818 if(ba[j]==start+i*4+4) {
8819 for(hr=0;hr<HOST_REGS;hr++) {
8820 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
8821 current.regmap[hr]=-1;
8823 current.is32&=branch_regs[j].is32;
8824 current.dirty&=branch_regs[j].dirty;
8833 // Count cycles in between branches
8835 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))
8839 #if !defined(DRC_DBG)
8840 else if(itype[i]==C2OP&>e_cycletab[source[i]&0x3f]>2)
8842 // GTE runs in parallel until accessed, divide by 2 for a rough guess
8843 cc+=gte_cycletab[source[i]&0x3f]/2;
8845 else if(/*itype[i]==LOAD||itype[i]==STORE||*/itype[i]==C1LS) // load,store causes weird timing issues
8847 cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
8849 else if(i>1&&itype[i]==STORE&&itype[i-1]==STORE&&itype[i-2]==STORE&&!bt[i])
8853 else if(itype[i]==C2LS)
8863 flush_dirty_uppers(¤t);
8865 regs[i].is32=current.is32;
8866 regs[i].dirty=current.dirty;
8867 regs[i].isconst=current.isconst;
8868 memcpy(constmap[i],current_constmap,sizeof(current_constmap));
8870 for(hr=0;hr<HOST_REGS;hr++) {
8871 if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
8872 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
8873 regs[i].wasconst&=~(1<<hr);
8877 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
8878 regs[i].waswritten=current.waswritten;
8881 /* Pass 4 - Cull unused host registers */
8885 for (i=slen-1;i>=0;i--)
8888 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
8890 if(ba[i]<start || ba[i]>=(start+slen*4))
8892 // Branch out of this block, don't need anything
8898 // Need whatever matches the target
8900 int t=(ba[i]-start)>>2;
8901 for(hr=0;hr<HOST_REGS;hr++)
8903 if(regs[i].regmap_entry[hr]>=0) {
8904 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
8908 // Conditional branch may need registers for following instructions
8909 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
8912 nr|=needed_reg[i+2];
8913 for(hr=0;hr<HOST_REGS;hr++)
8915 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
8916 //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]);
8920 // Don't need stuff which is overwritten
8921 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8922 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8923 // Merge in delay slot
8924 for(hr=0;hr<HOST_REGS;hr++)
8927 // These are overwritten unless the branch is "likely"
8928 // and the delay slot is nullified if not taken
8929 if(rt1[i+1]&&rt1[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8930 if(rt2[i+1]&&rt2[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8932 if(us1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8933 if(us2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8934 if(rs1[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
8935 if(rs2[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
8936 if(us1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8937 if(us2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8938 if(rs1[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
8939 if(rs2[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
8940 if(dep1[i+1]&&!((unneeded_reg_upper[i]>>dep1[i+1])&1)) {
8941 if(dep1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8942 if(dep2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8944 if(dep2[i+1]&&!((unneeded_reg_upper[i]>>dep2[i+1])&1)) {
8945 if(dep1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8946 if(dep2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8948 if(itype[i+1]==STORE || itype[i+1]==STORELR || (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) {
8949 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8950 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8954 else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
8956 // SYSCALL instruction (software interrupt)
8959 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
8961 // ERET instruction (return from interrupt)
8967 for(hr=0;hr<HOST_REGS;hr++) {
8968 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8969 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8970 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8971 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8975 for(hr=0;hr<HOST_REGS;hr++)
8977 // Overwritten registers are not needed
8978 if(rt1[i]&&rt1[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8979 if(rt2[i]&&rt2[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8980 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8981 // Source registers are needed
8982 if(us1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8983 if(us2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8984 if(rs1[i]==regmap_pre[i][hr]) nr|=1<<hr;
8985 if(rs2[i]==regmap_pre[i][hr]) nr|=1<<hr;
8986 if(us1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8987 if(us2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8988 if(rs1[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
8989 if(rs2[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
8990 if(dep1[i]&&!((unneeded_reg_upper[i]>>dep1[i])&1)) {
8991 if(dep1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8992 if(dep1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8994 if(dep2[i]&&!((unneeded_reg_upper[i]>>dep2[i])&1)) {
8995 if(dep2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8996 if(dep2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8998 if(itype[i]==STORE || itype[i]==STORELR || (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) {
8999 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
9000 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
9002 // Don't store a register immediately after writing it,
9003 // may prevent dual-issue.
9004 // But do so if this is a branch target, otherwise we
9005 // might have to load the register before the branch.
9006 if(i>0&&!bt[i]&&((regs[i].wasdirty>>hr)&1)) {
9007 if((regmap_pre[i][hr]>0&®map_pre[i][hr]<64&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1)) ||
9008 (regmap_pre[i][hr]>64&&!((unneeded_reg_upper[i]>>(regmap_pre[i][hr]&63))&1)) ) {
9009 if(rt1[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9010 if(rt2[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
9012 if((regs[i].regmap_entry[hr]>0&®s[i].regmap_entry[hr]<64&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1)) ||
9013 (regs[i].regmap_entry[hr]>64&&!((unneeded_reg_upper[i]>>(regs[i].regmap_entry[hr]&63))&1)) ) {
9014 if(rt1[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9015 if(rt2[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
9019 // Cycle count is needed at branches. Assume it is needed at the target too.
9020 if(i==0||bt[i]||itype[i]==CJUMP||itype[i]==FJUMP||itype[i]==SPAN) {
9021 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
9022 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
9027 // Deallocate unneeded registers
9028 for(hr=0;hr<HOST_REGS;hr++)
9031 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
9032 if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
9033 (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9034 (regs[i].regmap[hr]&63)!=PTEMP && (regs[i].regmap[hr]&63)!=CCREG)
9036 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9039 regs[i].regmap[hr]=-1;
9040 regs[i].isconst&=~(1<<hr);
9042 regmap_pre[i+2][hr]=-1;
9043 regs[i+2].wasconst&=~(1<<hr);
9048 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9050 int d1=0,d2=0,map=0,temp=0;
9051 if(get_reg(regs[i].regmap,rt1[i+1]|64)>=0||get_reg(branch_regs[i].regmap,rt1[i+1]|64)>=0)
9056 if(itype[i+1]==STORE || itype[i+1]==STORELR ||
9057 (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
9060 if(itype[i+1]==LOADLR || itype[i+1]==STORELR ||
9061 itype[i+1]==C1LS || itype[i+1]==C2LS)
9063 if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
9064 (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9065 (regs[i].regmap[hr]&63)!=rt1[i+1] && (regs[i].regmap[hr]&63)!=rt2[i+1] &&
9066 (regs[i].regmap[hr]^64)!=us1[i+1] && (regs[i].regmap[hr]^64)!=us2[i+1] &&
9067 (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
9068 regs[i].regmap[hr]!=rs1[i+1] && regs[i].regmap[hr]!=rs2[i+1] &&
9069 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
9070 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
9071 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
9072 regs[i].regmap[hr]!=map )
9074 regs[i].regmap[hr]=-1;
9075 regs[i].isconst&=~(1<<hr);
9076 if((branch_regs[i].regmap[hr]&63)!=rs1[i] && (branch_regs[i].regmap[hr]&63)!=rs2[i] &&
9077 (branch_regs[i].regmap[hr]&63)!=rt1[i] && (branch_regs[i].regmap[hr]&63)!=rt2[i] &&
9078 (branch_regs[i].regmap[hr]&63)!=rt1[i+1] && (branch_regs[i].regmap[hr]&63)!=rt2[i+1] &&
9079 (branch_regs[i].regmap[hr]^64)!=us1[i+1] && (branch_regs[i].regmap[hr]^64)!=us2[i+1] &&
9080 (branch_regs[i].regmap[hr]^64)!=d1 && (branch_regs[i].regmap[hr]^64)!=d2 &&
9081 branch_regs[i].regmap[hr]!=rs1[i+1] && branch_regs[i].regmap[hr]!=rs2[i+1] &&
9082 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
9083 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
9084 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
9085 branch_regs[i].regmap[hr]!=map)
9087 branch_regs[i].regmap[hr]=-1;
9088 branch_regs[i].regmap_entry[hr]=-1;
9089 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
9091 if(!likely[i]&&i<slen-2) {
9092 regmap_pre[i+2][hr]=-1;
9093 regs[i+2].wasconst&=~(1<<hr);
9104 int d1=0,d2=0,map=-1,temp=-1;
9105 if(get_reg(regs[i].regmap,rt1[i]|64)>=0)
9110 if(itype[i]==STORE || itype[i]==STORELR ||
9111 (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
9114 if(itype[i]==LOADLR || itype[i]==STORELR ||
9115 itype[i]==C1LS || itype[i]==C2LS)
9117 if((regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
9118 (regs[i].regmap[hr]^64)!=us1[i] && (regs[i].regmap[hr]^64)!=us2[i] &&
9119 (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
9120 regs[i].regmap[hr]!=rs1[i] && regs[i].regmap[hr]!=rs2[i] &&
9121 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map &&
9122 (itype[i]!=SPAN||regs[i].regmap[hr]!=CCREG))
9124 if(i<slen-1&&!is_ds[i]) {
9125 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]!=-1)
9126 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
9127 if(regs[i].regmap[hr]<64||!((regs[i].was32>>(regs[i].regmap[hr]&63))&1))
9129 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
9130 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
9132 regmap_pre[i+1][hr]=-1;
9133 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
9134 regs[i+1].wasconst&=~(1<<hr);
9136 regs[i].regmap[hr]=-1;
9137 regs[i].isconst&=~(1<<hr);
9145 /* Pass 5 - Pre-allocate registers */
9147 // If a register is allocated during a loop, try to allocate it for the
9148 // entire loop, if possible. This avoids loading/storing registers
9149 // inside of the loop.
9151 signed char f_regmap[HOST_REGS];
9152 clear_all_regs(f_regmap);
9153 for(i=0;i<slen-1;i++)
9155 if(itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9157 if(ba[i]>=start && ba[i]<(start+i*4))
9158 if(itype[i+1]==NOP||itype[i+1]==MOV||itype[i+1]==ALU
9159 ||itype[i+1]==SHIFTIMM||itype[i+1]==IMM16||itype[i+1]==LOAD
9160 ||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
9161 ||itype[i+1]==SHIFT||itype[i+1]==COP1||itype[i+1]==FLOAT
9162 ||itype[i+1]==FCOMP||itype[i+1]==FCONV
9163 ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
9165 int t=(ba[i]-start)>>2;
9166 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
9167 if(t<2||(itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||rt1[t-2]!=31) // call/ret assumes no registers allocated
9168 for(hr=0;hr<HOST_REGS;hr++)
9170 if(regs[i].regmap[hr]>64) {
9171 if(!((regs[i].dirty>>hr)&1))
9172 f_regmap[hr]=regs[i].regmap[hr];
9173 else f_regmap[hr]=-1;
9175 else if(regs[i].regmap[hr]>=0) {
9176 if(f_regmap[hr]!=regs[i].regmap[hr]) {
9177 // dealloc old register
9179 for(n=0;n<HOST_REGS;n++)
9181 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
9183 // and alloc new one
9184 f_regmap[hr]=regs[i].regmap[hr];
9187 if(branch_regs[i].regmap[hr]>64) {
9188 if(!((branch_regs[i].dirty>>hr)&1))
9189 f_regmap[hr]=branch_regs[i].regmap[hr];
9190 else f_regmap[hr]=-1;
9192 else if(branch_regs[i].regmap[hr]>=0) {
9193 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
9194 // dealloc old register
9196 for(n=0;n<HOST_REGS;n++)
9198 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
9200 // and alloc new one
9201 f_regmap[hr]=branch_regs[i].regmap[hr];
9205 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
9206 f_regmap[hr]=branch_regs[i].regmap[hr];
9208 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
9209 f_regmap[hr]=branch_regs[i].regmap[hr];
9211 // Avoid dirty->clean transition
9212 #ifdef DESTRUCTIVE_WRITEBACK
9213 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;
9215 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
9216 // case above, however it's always a good idea. We can't hoist the
9217 // load if the register was already allocated, so there's no point
9218 // wasting time analyzing most of these cases. It only "succeeds"
9219 // when the mapping was different and the load can be replaced with
9220 // a mov, which is of negligible benefit. So such cases are
9222 if(f_regmap[hr]>0) {
9223 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
9227 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
9228 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
9229 if(r>63&&((unneeded_reg_upper[j]>>(r&63))&1)) break;
9231 // NB This can exclude the case where the upper-half
9232 // register is lower numbered than the lower-half
9233 // register. Not sure if it's worth fixing...
9234 if(get_reg(regs[j].regmap,r&63)<0) break;
9235 if(get_reg(regs[j].regmap_entry,r&63)<0) break;
9236 if(regs[j].is32&(1LL<<(r&63))) break;
9238 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
9239 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
9241 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
9242 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
9244 if(get_reg(regs[i].regmap,r&63)<0) break;
9245 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
9248 while(k>1&®s[k-1].regmap[hr]==-1) {
9249 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
9250 //printf("no free regs for store %x\n",start+(k-1)*4);
9253 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
9254 //printf("no-match due to different register\n");
9257 if(itype[k-2]==UJUMP||itype[k-2]==RJUMP||itype[k-2]==CJUMP||itype[k-2]==SJUMP||itype[k-2]==FJUMP) {
9258 //printf("no-match due to branch\n");
9261 // call/ret fast path assumes no registers allocated
9262 if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)&&rt1[k-3]==31) {
9266 // NB This can exclude the case where the upper-half
9267 // register is lower numbered than the lower-half
9268 // register. Not sure if it's worth fixing...
9269 if(get_reg(regs[k-1].regmap,r&63)<0) break;
9270 if(regs[k-1].is32&(1LL<<(r&63))) break;
9275 if((regs[k].is32&(1LL<<f_regmap[hr]))!=
9276 (regs[i+2].was32&(1LL<<f_regmap[hr]))) {
9277 //printf("bad match after branch\n");
9281 if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
9282 //printf("Extend r%d, %x ->\n",hr,start+k*4);
9284 regs[k].regmap_entry[hr]=f_regmap[hr];
9285 regs[k].regmap[hr]=f_regmap[hr];
9286 regmap_pre[k+1][hr]=f_regmap[hr];
9287 regs[k].wasdirty&=~(1<<hr);
9288 regs[k].dirty&=~(1<<hr);
9289 regs[k].wasdirty|=(1<<hr)®s[k-1].dirty;
9290 regs[k].dirty|=(1<<hr)®s[k].wasdirty;
9291 regs[k].wasconst&=~(1<<hr);
9292 regs[k].isconst&=~(1<<hr);
9297 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
9300 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
9301 if(regs[i-1].regmap[hr]==f_regmap[hr]&®map_pre[i][hr]==f_regmap[hr]) {
9302 //printf("OK fill %x (r%d)\n",start+i*4,hr);
9303 regs[i].regmap_entry[hr]=f_regmap[hr];
9304 regs[i].regmap[hr]=f_regmap[hr];
9305 regs[i].wasdirty&=~(1<<hr);
9306 regs[i].dirty&=~(1<<hr);
9307 regs[i].wasdirty|=(1<<hr)®s[i-1].dirty;
9308 regs[i].dirty|=(1<<hr)®s[i-1].dirty;
9309 regs[i].wasconst&=~(1<<hr);
9310 regs[i].isconst&=~(1<<hr);
9311 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
9312 branch_regs[i].wasdirty&=~(1<<hr);
9313 branch_regs[i].wasdirty|=(1<<hr)®s[i].dirty;
9314 branch_regs[i].regmap[hr]=f_regmap[hr];
9315 branch_regs[i].dirty&=~(1<<hr);
9316 branch_regs[i].dirty|=(1<<hr)®s[i].dirty;
9317 branch_regs[i].wasconst&=~(1<<hr);
9318 branch_regs[i].isconst&=~(1<<hr);
9319 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
9320 regmap_pre[i+2][hr]=f_regmap[hr];
9321 regs[i+2].wasdirty&=~(1<<hr);
9322 regs[i+2].wasdirty|=(1<<hr)®s[i].dirty;
9323 assert((branch_regs[i].is32&(1LL<<f_regmap[hr]))==
9324 (regs[i+2].was32&(1LL<<f_regmap[hr])));
9329 // Alloc register clean at beginning of loop,
9330 // but may dirty it in pass 6
9331 regs[k].regmap_entry[hr]=f_regmap[hr];
9332 regs[k].regmap[hr]=f_regmap[hr];
9333 regs[k].dirty&=~(1<<hr);
9334 regs[k].wasconst&=~(1<<hr);
9335 regs[k].isconst&=~(1<<hr);
9336 if(itype[k]==UJUMP||itype[k]==RJUMP||itype[k]==CJUMP||itype[k]==SJUMP||itype[k]==FJUMP) {
9337 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
9338 branch_regs[k].regmap[hr]=f_regmap[hr];
9339 branch_regs[k].dirty&=~(1<<hr);
9340 branch_regs[k].wasconst&=~(1<<hr);
9341 branch_regs[k].isconst&=~(1<<hr);
9342 if(itype[k]!=RJUMP&&itype[k]!=UJUMP&&(source[k]>>16)!=0x1000) {
9343 regmap_pre[k+2][hr]=f_regmap[hr];
9344 regs[k+2].wasdirty&=~(1<<hr);
9345 assert((branch_regs[k].is32&(1LL<<f_regmap[hr]))==
9346 (regs[k+2].was32&(1LL<<f_regmap[hr])));
9351 regmap_pre[k+1][hr]=f_regmap[hr];
9352 regs[k+1].wasdirty&=~(1<<hr);
9355 if(regs[j].regmap[hr]==f_regmap[hr])
9356 regs[j].regmap_entry[hr]=f_regmap[hr];
9360 if(regs[j].regmap[hr]>=0)
9362 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
9363 //printf("no-match due to different register\n");
9366 if((regs[j+1].is32&(1LL<<f_regmap[hr]))!=(regs[j].is32&(1LL<<f_regmap[hr]))) {
9367 //printf("32/64 mismatch %x %d\n",start+j*4,hr);
9370 if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
9372 // Stop on unconditional branch
9375 if(itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP)
9378 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
9381 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
9384 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
9385 //printf("no-match due to different register (branch)\n");
9389 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
9390 //printf("No free regs for store %x\n",start+j*4);
9393 if(f_regmap[hr]>=64) {
9394 if(regs[j].is32&(1LL<<(f_regmap[hr]&63))) {
9399 if(get_reg(regs[j].regmap,f_regmap[hr]&63)<0) {
9410 // Non branch or undetermined branch target
9411 for(hr=0;hr<HOST_REGS;hr++)
9413 if(hr!=EXCLUDE_REG) {
9414 if(regs[i].regmap[hr]>64) {
9415 if(!((regs[i].dirty>>hr)&1))
9416 f_regmap[hr]=regs[i].regmap[hr];
9418 else if(regs[i].regmap[hr]>=0) {
9419 if(f_regmap[hr]!=regs[i].regmap[hr]) {
9420 // dealloc old register
9422 for(n=0;n<HOST_REGS;n++)
9424 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
9426 // and alloc new one
9427 f_regmap[hr]=regs[i].regmap[hr];
9432 // Try to restore cycle count at branch targets
9434 for(j=i;j<slen-1;j++) {
9435 if(regs[j].regmap[HOST_CCREG]!=-1) break;
9436 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
9437 //printf("no free regs for store %x\n",start+j*4);
9441 if(regs[j].regmap[HOST_CCREG]==CCREG) {
9443 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
9445 regs[k].regmap_entry[HOST_CCREG]=CCREG;
9446 regs[k].regmap[HOST_CCREG]=CCREG;
9447 regmap_pre[k+1][HOST_CCREG]=CCREG;
9448 regs[k+1].wasdirty|=1<<HOST_CCREG;
9449 regs[k].dirty|=1<<HOST_CCREG;
9450 regs[k].wasconst&=~(1<<HOST_CCREG);
9451 regs[k].isconst&=~(1<<HOST_CCREG);
9454 regs[j].regmap_entry[HOST_CCREG]=CCREG;
9456 // Work backwards from the branch target
9457 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
9459 //printf("Extend backwards\n");
9462 while(regs[k-1].regmap[HOST_CCREG]==-1) {
9463 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
9464 //printf("no free regs for store %x\n",start+(k-1)*4);
9469 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
9470 //printf("Extend CC, %x ->\n",start+k*4);
9472 regs[k].regmap_entry[HOST_CCREG]=CCREG;
9473 regs[k].regmap[HOST_CCREG]=CCREG;
9474 regmap_pre[k+1][HOST_CCREG]=CCREG;
9475 regs[k+1].wasdirty|=1<<HOST_CCREG;
9476 regs[k].dirty|=1<<HOST_CCREG;
9477 regs[k].wasconst&=~(1<<HOST_CCREG);
9478 regs[k].isconst&=~(1<<HOST_CCREG);
9483 //printf("Fail Extend CC, %x ->\n",start+k*4);
9487 if(itype[i]!=STORE&&itype[i]!=STORELR&&itype[i]!=C1LS&&itype[i]!=SHIFT&&
9488 itype[i]!=NOP&&itype[i]!=MOV&&itype[i]!=ALU&&itype[i]!=SHIFTIMM&&
9489 itype[i]!=IMM16&&itype[i]!=LOAD&&itype[i]!=COP1&&itype[i]!=FLOAT&&
9490 itype[i]!=FCONV&&itype[i]!=FCOMP)
9492 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
9497 // Cache memory offset or tlb map pointer if a register is available
9498 #ifndef HOST_IMM_ADDR32
9503 int earliest_available[HOST_REGS];
9504 int loop_start[HOST_REGS];
9505 int score[HOST_REGS];
9510 for(hr=0;hr<HOST_REGS;hr++) {
9511 score[hr]=0;earliest_available[hr]=0;
9512 loop_start[hr]=MAXBLOCK;
9514 for(i=0;i<slen-1;i++)
9516 // Can't do anything if no registers are available
9517 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i]) {
9518 for(hr=0;hr<HOST_REGS;hr++) {
9519 score[hr]=0;earliest_available[hr]=i+1;
9520 loop_start[hr]=MAXBLOCK;
9523 if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
9525 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1]) {
9526 for(hr=0;hr<HOST_REGS;hr++) {
9527 score[hr]=0;earliest_available[hr]=i+1;
9528 loop_start[hr]=MAXBLOCK;
9532 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1]) {
9533 for(hr=0;hr<HOST_REGS;hr++) {
9534 score[hr]=0;earliest_available[hr]=i+1;
9535 loop_start[hr]=MAXBLOCK;
9540 // Mark unavailable registers
9541 for(hr=0;hr<HOST_REGS;hr++) {
9542 if(regs[i].regmap[hr]>=0) {
9543 score[hr]=0;earliest_available[hr]=i+1;
9544 loop_start[hr]=MAXBLOCK;
9546 if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
9547 if(branch_regs[i].regmap[hr]>=0) {
9548 score[hr]=0;earliest_available[hr]=i+2;
9549 loop_start[hr]=MAXBLOCK;
9553 // No register allocations after unconditional jumps
9554 if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
9556 for(hr=0;hr<HOST_REGS;hr++) {
9557 score[hr]=0;earliest_available[hr]=i+2;
9558 loop_start[hr]=MAXBLOCK;
9560 i++; // Skip delay slot too
9561 //printf("skip delay slot: %x\n",start+i*4);
9565 if(itype[i]==LOAD||itype[i]==LOADLR||
9566 itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS) {
9567 for(hr=0;hr<HOST_REGS;hr++) {
9568 if(hr!=EXCLUDE_REG) {
9570 for(j=i;j<slen-1;j++) {
9571 if(regs[j].regmap[hr]>=0) break;
9572 if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
9573 if(branch_regs[j].regmap[hr]>=0) break;
9575 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1]) break;
9577 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1]) break;
9580 else if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) break;
9581 if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
9582 int t=(ba[j]-start)>>2;
9583 if(t<j&&t>=earliest_available[hr]) {
9584 if(t==1||(t>1&&itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||(t>1&&rt1[t-2]!=31)) { // call/ret assumes no registers allocated
9585 // Score a point for hoisting loop invariant
9586 if(t<loop_start[hr]) loop_start[hr]=t;
9587 //printf("set loop_start: i=%x j=%x (%x)\n",start+i*4,start+j*4,start+t*4);
9593 if(regs[t].regmap[hr]==reg) {
9594 // Score a point if the branch target matches this register
9599 if(itype[j+1]==LOAD||itype[j+1]==LOADLR||
9600 itype[j+1]==STORE||itype[j+1]==STORELR||itype[j+1]==C1LS) {
9605 if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
9607 // Stop on unconditional branch
9611 if(itype[j]==LOAD||itype[j]==LOADLR||
9612 itype[j]==STORE||itype[j]==STORELR||itype[j]==C1LS) {
9619 // Find highest score and allocate that register
9621 for(hr=0;hr<HOST_REGS;hr++) {
9622 if(hr!=EXCLUDE_REG) {
9623 if(score[hr]>score[maxscore]) {
9625 //printf("highest score: %d %d (%x->%x)\n",score[hr],hr,start+i*4,start+end[hr]*4);
9629 if(score[maxscore]>1)
9631 if(i<loop_start[maxscore]) loop_start[maxscore]=i;
9632 for(j=loop_start[maxscore];j<slen&&j<=end[maxscore];j++) {
9633 //if(regs[j].regmap[maxscore]>=0) {printf("oops: %x %x was %d=%d\n",loop_start[maxscore]*4+start,j*4+start,maxscore,regs[j].regmap[maxscore]);}
9634 assert(regs[j].regmap[maxscore]<0);
9635 if(j>loop_start[maxscore]) regs[j].regmap_entry[maxscore]=reg;
9636 regs[j].regmap[maxscore]=reg;
9637 regs[j].dirty&=~(1<<maxscore);
9638 regs[j].wasconst&=~(1<<maxscore);
9639 regs[j].isconst&=~(1<<maxscore);
9640 if(itype[j]==UJUMP||itype[j]==RJUMP||itype[j]==CJUMP||itype[j]==SJUMP||itype[j]==FJUMP) {
9641 branch_regs[j].regmap[maxscore]=reg;
9642 branch_regs[j].wasdirty&=~(1<<maxscore);
9643 branch_regs[j].dirty&=~(1<<maxscore);
9644 branch_regs[j].wasconst&=~(1<<maxscore);
9645 branch_regs[j].isconst&=~(1<<maxscore);
9646 if(itype[j]!=RJUMP&&itype[j]!=UJUMP&&(source[j]>>16)!=0x1000) {
9647 regmap_pre[j+2][maxscore]=reg;
9648 regs[j+2].wasdirty&=~(1<<maxscore);
9650 // loop optimization (loop_preload)
9651 int t=(ba[j]-start)>>2;
9652 if(t==loop_start[maxscore]) {
9653 if(t==1||(t>1&&itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||(t>1&&rt1[t-2]!=31)) // call/ret assumes no registers allocated
9654 regs[t].regmap_entry[maxscore]=reg;
9659 if(j<1||(itype[j-1]!=RJUMP&&itype[j-1]!=UJUMP&&itype[j-1]!=CJUMP&&itype[j-1]!=SJUMP&&itype[j-1]!=FJUMP)) {
9660 regmap_pre[j+1][maxscore]=reg;
9661 regs[j+1].wasdirty&=~(1<<maxscore);
9666 if(itype[j-1]==RJUMP||itype[j-1]==UJUMP||itype[j-1]==CJUMP||itype[j-1]==SJUMP||itype[j-1]==FJUMP) i++; // skip delay slot
9667 for(hr=0;hr<HOST_REGS;hr++) {
9668 score[hr]=0;earliest_available[hr]=i+i;
9669 loop_start[hr]=MAXBLOCK;
9677 // This allocates registers (if possible) one instruction prior
9678 // to use, which can avoid a load-use penalty on certain CPUs.
9679 for(i=0;i<slen-1;i++)
9681 if(!i||(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP))
9685 if(itype[i]==ALU||itype[i]==MOV||itype[i]==LOAD||itype[i]==SHIFTIMM||itype[i]==IMM16
9686 ||((itype[i]==COP1||itype[i]==COP2)&&opcode2[i]<3))
9689 if((hr=get_reg(regs[i+1].regmap,rs1[i+1]))>=0)
9691 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9693 regs[i].regmap[hr]=regs[i+1].regmap[hr];
9694 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
9695 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
9696 regs[i].isconst&=~(1<<hr);
9697 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9698 constmap[i][hr]=constmap[i+1][hr];
9699 regs[i+1].wasdirty&=~(1<<hr);
9700 regs[i].dirty&=~(1<<hr);
9705 if((hr=get_reg(regs[i+1].regmap,rs2[i+1]))>=0)
9707 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9709 regs[i].regmap[hr]=regs[i+1].regmap[hr];
9710 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
9711 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
9712 regs[i].isconst&=~(1<<hr);
9713 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9714 constmap[i][hr]=constmap[i+1][hr];
9715 regs[i+1].wasdirty&=~(1<<hr);
9716 regs[i].dirty&=~(1<<hr);
9720 // Preload target address for load instruction (non-constant)
9721 if(itype[i+1]==LOAD&&rs1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
9722 if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
9724 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9726 regs[i].regmap[hr]=rs1[i+1];
9727 regmap_pre[i+1][hr]=rs1[i+1];
9728 regs[i+1].regmap_entry[hr]=rs1[i+1];
9729 regs[i].isconst&=~(1<<hr);
9730 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9731 constmap[i][hr]=constmap[i+1][hr];
9732 regs[i+1].wasdirty&=~(1<<hr);
9733 regs[i].dirty&=~(1<<hr);
9737 // Load source into target register
9738 if(lt1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
9739 if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
9741 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9743 regs[i].regmap[hr]=rs1[i+1];
9744 regmap_pre[i+1][hr]=rs1[i+1];
9745 regs[i+1].regmap_entry[hr]=rs1[i+1];
9746 regs[i].isconst&=~(1<<hr);
9747 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9748 constmap[i][hr]=constmap[i+1][hr];
9749 regs[i+1].wasdirty&=~(1<<hr);
9750 regs[i].dirty&=~(1<<hr);
9754 // Address for store instruction (non-constant)
9755 if(itype[i+1]==STORE||itype[i+1]==STORELR
9756 ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
9757 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
9758 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
9759 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
9760 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
9762 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9764 regs[i].regmap[hr]=rs1[i+1];
9765 regmap_pre[i+1][hr]=rs1[i+1];
9766 regs[i+1].regmap_entry[hr]=rs1[i+1];
9767 regs[i].isconst&=~(1<<hr);
9768 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9769 constmap[i][hr]=constmap[i+1][hr];
9770 regs[i+1].wasdirty&=~(1<<hr);
9771 regs[i].dirty&=~(1<<hr);
9775 if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
9776 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
9778 hr=get_reg(regs[i+1].regmap,FTEMP);
9780 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
9782 regs[i].regmap[hr]=rs1[i+1];
9783 regmap_pre[i+1][hr]=rs1[i+1];
9784 regs[i+1].regmap_entry[hr]=rs1[i+1];
9785 regs[i].isconst&=~(1<<hr);
9786 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
9787 constmap[i][hr]=constmap[i+1][hr];
9788 regs[i+1].wasdirty&=~(1<<hr);
9789 regs[i].dirty&=~(1<<hr);
9791 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
9793 // move it to another register
9794 regs[i+1].regmap[hr]=-1;
9795 regmap_pre[i+2][hr]=-1;
9796 regs[i+1].regmap[nr]=FTEMP;
9797 regmap_pre[i+2][nr]=FTEMP;
9798 regs[i].regmap[nr]=rs1[i+1];
9799 regmap_pre[i+1][nr]=rs1[i+1];
9800 regs[i+1].regmap_entry[nr]=rs1[i+1];
9801 regs[i].isconst&=~(1<<nr);
9802 regs[i+1].isconst&=~(1<<nr);
9803 regs[i].dirty&=~(1<<nr);
9804 regs[i+1].wasdirty&=~(1<<nr);
9805 regs[i+1].dirty&=~(1<<nr);
9806 regs[i+2].wasdirty&=~(1<<nr);
9810 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*/) {
9811 if(itype[i+1]==LOAD)
9812 hr=get_reg(regs[i+1].regmap,rt1[i+1]);
9813 if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
9814 hr=get_reg(regs[i+1].regmap,FTEMP);
9815 if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
9816 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
9817 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
9819 if(hr>=0&®s[i].regmap[hr]<0) {
9820 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
9821 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
9822 regs[i].regmap[hr]=AGEN1+((i+1)&1);
9823 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
9824 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
9825 regs[i].isconst&=~(1<<hr);
9826 regs[i+1].wasdirty&=~(1<<hr);
9827 regs[i].dirty&=~(1<<hr);
9836 /* Pass 6 - Optimize clean/dirty state */
9837 clean_registers(0,slen-1,1);
9839 /* Pass 7 - Identify 32-bit registers */
9840 for (i=slen-1;i>=0;i--)
9842 if(itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
9844 // Conditional branch
9845 if((source[i]>>16)!=0x1000&&i<slen-2) {
9846 // Mark this address as a branch target since it may be called
9847 // upon return from interrupt
9853 if(itype[slen-1]==SPAN) {
9854 bt[slen-1]=1; // Mark as a branch target so instruction can restart after exception
9858 /* Debug/disassembly */
9863 for(r=1;r<=CCREG;r++) {
9864 if((unneeded_reg[i]>>r)&1) {
9865 if(r==HIREG) printf(" HI");
9866 else if(r==LOREG) printf(" LO");
9867 else printf(" r%d",r);
9871 #if defined(__i386__) || defined(__x86_64__)
9872 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]);
9875 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]);
9878 if(needed_reg[i]&1) printf("eax ");
9879 if((needed_reg[i]>>1)&1) printf("ecx ");
9880 if((needed_reg[i]>>2)&1) printf("edx ");
9881 if((needed_reg[i]>>3)&1) printf("ebx ");
9882 if((needed_reg[i]>>5)&1) printf("ebp ");
9883 if((needed_reg[i]>>6)&1) printf("esi ");
9884 if((needed_reg[i]>>7)&1) printf("edi ");
9886 #if defined(__i386__) || defined(__x86_64__)
9887 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]);
9889 if(regs[i].wasdirty&1) printf("eax ");
9890 if((regs[i].wasdirty>>1)&1) printf("ecx ");
9891 if((regs[i].wasdirty>>2)&1) printf("edx ");
9892 if((regs[i].wasdirty>>3)&1) printf("ebx ");
9893 if((regs[i].wasdirty>>5)&1) printf("ebp ");
9894 if((regs[i].wasdirty>>6)&1) printf("esi ");
9895 if((regs[i].wasdirty>>7)&1) printf("edi ");
9898 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]);
9900 if(regs[i].wasdirty&1) printf("r0 ");
9901 if((regs[i].wasdirty>>1)&1) printf("r1 ");
9902 if((regs[i].wasdirty>>2)&1) printf("r2 ");
9903 if((regs[i].wasdirty>>3)&1) printf("r3 ");
9904 if((regs[i].wasdirty>>4)&1) printf("r4 ");
9905 if((regs[i].wasdirty>>5)&1) printf("r5 ");
9906 if((regs[i].wasdirty>>6)&1) printf("r6 ");
9907 if((regs[i].wasdirty>>7)&1) printf("r7 ");
9908 if((regs[i].wasdirty>>8)&1) printf("r8 ");
9909 if((regs[i].wasdirty>>9)&1) printf("r9 ");
9910 if((regs[i].wasdirty>>10)&1) printf("r10 ");
9911 if((regs[i].wasdirty>>12)&1) printf("r12 ");
9914 disassemble_inst(i);
9915 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
9916 #if defined(__i386__) || defined(__x86_64__)
9917 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]);
9918 if(regs[i].dirty&1) printf("eax ");
9919 if((regs[i].dirty>>1)&1) printf("ecx ");
9920 if((regs[i].dirty>>2)&1) printf("edx ");
9921 if((regs[i].dirty>>3)&1) printf("ebx ");
9922 if((regs[i].dirty>>5)&1) printf("ebp ");
9923 if((regs[i].dirty>>6)&1) printf("esi ");
9924 if((regs[i].dirty>>7)&1) printf("edi ");
9927 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]);
9928 if(regs[i].dirty&1) printf("r0 ");
9929 if((regs[i].dirty>>1)&1) printf("r1 ");
9930 if((regs[i].dirty>>2)&1) printf("r2 ");
9931 if((regs[i].dirty>>3)&1) printf("r3 ");
9932 if((regs[i].dirty>>4)&1) printf("r4 ");
9933 if((regs[i].dirty>>5)&1) printf("r5 ");
9934 if((regs[i].dirty>>6)&1) printf("r6 ");
9935 if((regs[i].dirty>>7)&1) printf("r7 ");
9936 if((regs[i].dirty>>8)&1) printf("r8 ");
9937 if((regs[i].dirty>>9)&1) printf("r9 ");
9938 if((regs[i].dirty>>10)&1) printf("r10 ");
9939 if((regs[i].dirty>>12)&1) printf("r12 ");
9942 if(regs[i].isconst) {
9943 printf("constants: ");
9944 #if defined(__i386__) || defined(__x86_64__)
9945 if(regs[i].isconst&1) printf("eax=%x ",(int)constmap[i][0]);
9946 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(int)constmap[i][1]);
9947 if((regs[i].isconst>>2)&1) printf("edx=%x ",(int)constmap[i][2]);
9948 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(int)constmap[i][3]);
9949 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(int)constmap[i][5]);
9950 if((regs[i].isconst>>6)&1) printf("esi=%x ",(int)constmap[i][6]);
9951 if((regs[i].isconst>>7)&1) printf("edi=%x ",(int)constmap[i][7]);
9954 if(regs[i].isconst&1) printf("r0=%x ",(int)constmap[i][0]);
9955 if((regs[i].isconst>>1)&1) printf("r1=%x ",(int)constmap[i][1]);
9956 if((regs[i].isconst>>2)&1) printf("r2=%x ",(int)constmap[i][2]);
9957 if((regs[i].isconst>>3)&1) printf("r3=%x ",(int)constmap[i][3]);
9958 if((regs[i].isconst>>4)&1) printf("r4=%x ",(int)constmap[i][4]);
9959 if((regs[i].isconst>>5)&1) printf("r5=%x ",(int)constmap[i][5]);
9960 if((regs[i].isconst>>6)&1) printf("r6=%x ",(int)constmap[i][6]);
9961 if((regs[i].isconst>>7)&1) printf("r7=%x ",(int)constmap[i][7]);
9962 if((regs[i].isconst>>8)&1) printf("r8=%x ",(int)constmap[i][8]);
9963 if((regs[i].isconst>>9)&1) printf("r9=%x ",(int)constmap[i][9]);
9964 if((regs[i].isconst>>10)&1) printf("r10=%x ",(int)constmap[i][10]);
9965 if((regs[i].isconst>>12)&1) printf("r12=%x ",(int)constmap[i][12]);
9969 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP) {
9970 #if defined(__i386__) || defined(__x86_64__)
9971 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]);
9972 if(branch_regs[i].dirty&1) printf("eax ");
9973 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
9974 if((branch_regs[i].dirty>>2)&1) printf("edx ");
9975 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
9976 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
9977 if((branch_regs[i].dirty>>6)&1) printf("esi ");
9978 if((branch_regs[i].dirty>>7)&1) printf("edi ");
9981 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]);
9982 if(branch_regs[i].dirty&1) printf("r0 ");
9983 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
9984 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
9985 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
9986 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
9987 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
9988 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
9989 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
9990 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
9991 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
9992 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
9993 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
9999 /* Pass 8 - Assembly */
10000 linkcount=0;stubcount=0;
10001 ds=0;is_delayslot=0;
10003 uint64_t is32_pre=0;
10005 void *beginning=start_block();
10006 if((u_int)addr&1) {
10010 u_int instr_addr0_override=0;
10012 if (start == 0x80030000) {
10013 // nasty hack for fastbios thing
10014 // override block entry to this code
10015 instr_addr0_override=(u_int)out;
10016 emit_movimm(start,0);
10017 // abuse io address var as a flag that we
10018 // have already returned here once
10019 emit_readword((int)&address,1);
10020 emit_writeword(0,(int)&pcaddr);
10021 emit_writeword(0,(int)&address);
10023 emit_jne((int)new_dyna_leave);
10025 for(i=0;i<slen;i++)
10027 //if(ds) printf("ds: ");
10028 disassemble_inst(i);
10030 ds=0; // Skip delay slot
10031 if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
10034 speculate_register_values(i);
10035 #ifndef DESTRUCTIVE_WRITEBACK
10036 if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
10038 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,is32_pre,
10039 unneeded_reg[i],unneeded_reg_upper[i]);
10041 if((itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)&&!likely[i]) {
10042 is32_pre=branch_regs[i].is32;
10043 dirty_pre=branch_regs[i].dirty;
10045 is32_pre=regs[i].is32;
10046 dirty_pre=regs[i].dirty;
10050 if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
10052 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32,
10053 unneeded_reg[i],unneeded_reg_upper[i]);
10054 loop_preload(regmap_pre[i],regs[i].regmap_entry);
10056 // branch target entry point
10057 instr_addr[i]=(u_int)out;
10058 assem_debug("<->\n");
10059 drc_dbg_emit_do_cmp(i);
10062 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
10063 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty,regs[i].was32);
10064 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i],rs2[i]);
10065 address_generation(i,®s[i],regs[i].regmap_entry);
10066 load_consts(regmap_pre[i],regs[i].regmap,regs[i].was32,i);
10067 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP||itype[i]==FJUMP)
10069 // Load the delay slot registers if necessary
10070 if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i]&&(rs1[i+1]!=rt1[i]||rt1[i]==0))
10071 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
10072 if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i]&&(rs2[i+1]!=rt1[i]||rt1[i]==0))
10073 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
10074 if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a)
10075 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
10079 // Preload registers for following instruction
10080 if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i])
10081 if(rs1[i+1]!=rt1[i]&&rs1[i+1]!=rt2[i])
10082 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs1[i+1],rs1[i+1]);
10083 if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i])
10084 if(rs2[i+1]!=rt1[i]&&rs2[i+1]!=rt2[i])
10085 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,rs2[i+1],rs2[i+1]);
10087 // TODO: if(is_ooo(i)) address_generation(i+1);
10088 if(itype[i]==CJUMP||itype[i]==FJUMP)
10089 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,CCREG,CCREG);
10090 if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a)
10091 load_regs(regs[i].regmap_entry,regs[i].regmap,regs[i].was32,INVCP,INVCP);
10092 if(bt[i]) cop1_usable=0;
10096 alu_assemble(i,®s[i]);break;
10098 imm16_assemble(i,®s[i]);break;
10100 shift_assemble(i,®s[i]);break;
10102 shiftimm_assemble(i,®s[i]);break;
10104 load_assemble(i,®s[i]);break;
10106 loadlr_assemble(i,®s[i]);break;
10108 store_assemble(i,®s[i]);break;
10110 storelr_assemble(i,®s[i]);break;
10112 cop0_assemble(i,®s[i]);break;
10114 cop1_assemble(i,®s[i]);break;
10116 c1ls_assemble(i,®s[i]);break;
10118 cop2_assemble(i,®s[i]);break;
10120 c2ls_assemble(i,®s[i]);break;
10122 c2op_assemble(i,®s[i]);break;
10124 fconv_assemble(i,®s[i]);break;
10126 float_assemble(i,®s[i]);break;
10128 fcomp_assemble(i,®s[i]);break;
10130 multdiv_assemble(i,®s[i]);break;
10132 mov_assemble(i,®s[i]);break;
10134 syscall_assemble(i,®s[i]);break;
10136 hlecall_assemble(i,®s[i]);break;
10138 intcall_assemble(i,®s[i]);break;
10140 ujump_assemble(i,®s[i]);ds=1;break;
10142 rjump_assemble(i,®s[i]);ds=1;break;
10144 cjump_assemble(i,®s[i]);ds=1;break;
10146 sjump_assemble(i,®s[i]);ds=1;break;
10148 fjump_assemble(i,®s[i]);ds=1;break;
10150 pagespan_assemble(i,®s[i]);break;
10152 if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
10153 literal_pool(1024);
10155 literal_pool_jumpover(256);
10158 //assert(itype[i-2]==UJUMP||itype[i-2]==RJUMP||(source[i-2]>>16)==0x1000);
10159 // If the block did not end with an unconditional branch,
10160 // add a jump to the next instruction.
10162 if(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000&&itype[i-1]!=SPAN) {
10163 assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP);
10165 if(itype[i-2]!=CJUMP&&itype[i-2]!=SJUMP&&itype[i-2]!=FJUMP) {
10166 store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
10167 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
10168 emit_loadreg(CCREG,HOST_CCREG);
10169 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
10171 else if(!likely[i-2])
10173 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].is32,branch_regs[i-2].dirty,start+i*4);
10174 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
10178 store_regs_bt(regs[i-2].regmap,regs[i-2].is32,regs[i-2].dirty,start+i*4);
10179 assert(regs[i-2].regmap[HOST_CCREG]==CCREG);
10181 add_to_linker((int)out,start+i*4,0);
10188 assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP&&itype[i-1]!=FJUMP);
10189 store_regs_bt(regs[i-1].regmap,regs[i-1].is32,regs[i-1].dirty,start+i*4);
10190 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
10191 emit_loadreg(CCREG,HOST_CCREG);
10192 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
10193 add_to_linker((int)out,start+i*4,0);
10197 // TODO: delay slot stubs?
10199 for(i=0;i<stubcount;i++)
10201 switch(stubs[i][0])
10209 do_readstub(i);break;
10214 do_writestub(i);break;
10216 do_ccstub(i);break;
10218 do_invstub(i);break;
10220 do_cop1stub(i);break;
10222 do_unalignedwritestub(i);break;
10226 if (instr_addr0_override)
10227 instr_addr[0] = instr_addr0_override;
10229 /* Pass 9 - Linker */
10230 for(i=0;i<linkcount;i++)
10232 assem_debug("%8x -> %8x\n",link_addr[i][0],link_addr[i][1]);
10234 if(!link_addr[i][2])
10237 void *addr=check_addr(link_addr[i][1]);
10238 emit_extjump(link_addr[i][0],link_addr[i][1]);
10240 set_jump_target(link_addr[i][0],(int)addr);
10241 add_link(link_addr[i][1],stub);
10243 else set_jump_target(link_addr[i][0],(int)stub);
10248 int target=(link_addr[i][1]-start)>>2;
10249 assert(target>=0&&target<slen);
10250 assert(instr_addr[target]);
10251 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
10252 //set_jump_target_fillslot(link_addr[i][0],instr_addr[target],link_addr[i][2]>>1);
10254 set_jump_target(link_addr[i][0],instr_addr[target]);
10258 // External Branch Targets (jump_in)
10259 if(copy+slen*4>(void *)shadow+sizeof(shadow)) copy=shadow;
10260 for(i=0;i<slen;i++)
10264 if(instr_addr[i]) // TODO - delay slots (=null)
10266 u_int vaddr=start+i*4;
10267 u_int page=get_page(vaddr);
10268 u_int vpage=get_vpage(vaddr);
10271 assem_debug("%8x (%d) <- %8x\n",instr_addr[i],i,start+i*4);
10272 assem_debug("jump_in: %x\n",start+i*4);
10273 ll_add(jump_dirty+vpage,vaddr,(void *)out);
10274 int entry_point=do_dirty_stub(i);
10275 ll_add_flags(jump_in+page,vaddr,state_rflags,(void *)entry_point);
10276 // If there was an existing entry in the hash table,
10277 // replace it with the new address.
10278 // Don't add new entries. We'll insert the
10279 // ones that actually get used in check_addr().
10280 u_int *ht_bin=hash_table[((vaddr>>16)^vaddr)&0xFFFF];
10281 if(ht_bin[0]==vaddr) {
10282 ht_bin[1]=entry_point;
10284 if(ht_bin[2]==vaddr) {
10285 ht_bin[3]=entry_point;
10291 // Write out the literal pool if necessary
10293 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
10295 if(((u_int)out)&7) emit_addnop(13);
10297 assert((u_int)out-(u_int)beginning<MAX_OUTPUT_BLOCK_SIZE);
10298 //printf("shadow buffer: %x-%x\n",(int)copy,(int)copy+slen*4);
10299 memcpy(copy,source,slen*4);
10302 end_block(beginning);
10304 // If we're within 256K of the end of the buffer,
10305 // start over from the beginning. (Is 256K enough?)
10306 if((u_int)out>(u_int)BASE_ADDR+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE) out=(u_char *)BASE_ADDR;
10308 // Trap writes to any of the pages we compiled
10309 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
10312 inv_code_start=inv_code_end=~0;
10314 // for PCSX we need to mark all mirrors too
10315 if(get_page(start)<(RAM_SIZE>>12))
10316 for(i=start>>12;i<=(start+slen*4)>>12;i++)
10317 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
10318 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
10319 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
10321 /* Pass 10 - Free memory by expiring oldest blocks */
10323 int end=((((int)out-(int)BASE_ADDR)>>(TARGET_SIZE_2-16))+16384)&65535;
10324 while(expirep!=end)
10326 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
10327 int base=(int)BASE_ADDR+((expirep>>13)<<shift); // Base address of this block
10328 inv_debug("EXP: Phase %d\n",expirep);
10329 switch((expirep>>11)&3)
10332 // Clear jump_in and jump_dirty
10333 ll_remove_matching_addrs(jump_in+(expirep&2047),base,shift);
10334 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base,shift);
10335 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base,shift);
10336 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base,shift);
10340 ll_kill_pointers(jump_out[expirep&2047],base,shift);
10341 ll_kill_pointers(jump_out[(expirep&2047)+2048],base,shift);
10344 // Clear hash table
10345 for(i=0;i<32;i++) {
10346 u_int *ht_bin=hash_table[((expirep&2047)<<5)+i];
10347 if((ht_bin[3]>>shift)==(base>>shift) ||
10348 ((ht_bin[3]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
10349 inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[2],ht_bin[3]);
10350 ht_bin[2]=ht_bin[3]=-1;
10352 if((ht_bin[1]>>shift)==(base>>shift) ||
10353 ((ht_bin[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
10354 inv_debug("EXP: Remove hash %x -> %x\n",ht_bin[0],ht_bin[1]);
10355 ht_bin[0]=ht_bin[2];
10356 ht_bin[1]=ht_bin[3];
10357 ht_bin[2]=ht_bin[3]=-1;
10364 if((expirep&2047)==0)
10367 ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
10368 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
10371 expirep=(expirep+1)&65535;
10376 // vim:shiftwidth=2:expandtab