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"
39 #include "../psxinterpreter.h"
41 #include "emu_if.h" // emulator interface
43 #define noinline __attribute__((noinline,noclone))
45 #define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
48 #define min(a, b) ((b) < (a) ? (b) : (a))
51 #define max(a, b) ((b) > (a) ? (b) : (a))
58 #define assem_debug printf
60 #define assem_debug(...)
62 //#define inv_debug printf
63 #define inv_debug(...)
66 #include "assem_x86.h"
69 #include "assem_x64.h"
72 #include "assem_arm.h"
75 #include "assem_arm64.h"
78 #define RAM_SIZE 0x200000
80 #define MAX_OUTPUT_BLOCK_SIZE 262144
84 u_char translation_cache[1 << TARGET_SIZE_2];
87 struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
88 const void *f[2048 / sizeof(void *)];
92 #ifdef BASE_ADDR_DYNAMIC
93 static struct ndrc_mem *ndrc;
95 static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
96 static struct ndrc_mem *ndrc = &ndrc_;
119 signed char regmap_entry[HOST_REGS];
120 signed char regmap[HOST_REGS];
126 u_int loadedconst; // host regs that have constants loaded
127 u_int waswritten; // MIPS regs that were used as store base before
130 // note: asm depends on this layout
136 struct ll_entry *next;
164 static struct decoded_insn
183 struct ht_entry hash_table[65536] __attribute__((aligned(16)));
184 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
185 struct ll_entry *jump_dirty[4096];
187 static struct ll_entry *jump_out[4096];
189 static u_int *source;
190 static char insn[MAXBLOCK][10];
191 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
192 static uint64_t gte_rt[MAXBLOCK];
193 static uint64_t gte_unneeded[MAXBLOCK];
194 static u_int smrv[32]; // speculated MIPS register values
195 static u_int smrv_strong; // mask or regs that are likely to have correct values
196 static u_int smrv_weak; // same, but somewhat less likely
197 static u_int smrv_strong_next; // same, but after current insn executes
198 static u_int smrv_weak_next;
199 static int imm[MAXBLOCK];
200 static u_int ba[MAXBLOCK];
201 static uint64_t unneeded_reg[MAXBLOCK];
202 static uint64_t branch_unneeded_reg[MAXBLOCK];
203 static signed char regmap_pre[MAXBLOCK][HOST_REGS]; // pre-instruction i?
204 // contains 'real' consts at [i] insn, but may differ from what's actually
205 // loaded in host reg as 'final' value is always loaded, see get_final_value()
206 static uint32_t current_constmap[HOST_REGS];
207 static uint32_t constmap[MAXBLOCK][HOST_REGS];
208 static struct regstat regs[MAXBLOCK];
209 static struct regstat branch_regs[MAXBLOCK];
210 static signed char minimum_free_regs[MAXBLOCK];
211 static u_int needed_reg[MAXBLOCK];
212 static u_int wont_dirty[MAXBLOCK];
213 static u_int will_dirty[MAXBLOCK];
214 static int ccadj[MAXBLOCK];
216 static void *instr_addr[MAXBLOCK];
217 static struct link_entry link_addr[MAXBLOCK];
218 static int linkcount;
219 static struct code_stub stubs[MAXBLOCK*3];
220 static int stubcount;
221 static u_int literals[1024][2];
222 static int literalcount;
223 static int is_delayslot;
224 static char shadow[1048576] __attribute__((aligned(16)));
227 static u_int stop_after_jal;
228 static u_int f1_hack; // 0 - off, ~0 - capture address, else addr
230 static uintptr_t ram_offset;
232 static const uintptr_t ram_offset=0;
235 int new_dynarec_hacks;
236 int new_dynarec_hacks_pergame;
237 int new_dynarec_hacks_old;
238 int new_dynarec_did_compile;
240 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
242 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
243 extern int last_count; // last absolute target, often = next_interupt
245 extern int pending_exception;
246 extern int branch_target;
247 extern uintptr_t mini_ht[32][2];
248 extern u_char restore_candidate[512];
250 /* registers that may be allocated */
252 #define LOREG 32 // lo
253 #define HIREG 33 // hi
254 //#define FSREG 34 // FPU status (FCSR)
255 #define CSREG 35 // Coprocessor status
256 #define CCREG 36 // Cycle count
257 #define INVCP 37 // Pointer to invalid_code
258 //#define MMREG 38 // Pointer to memory_map
259 //#define ROREG 39 // ram offset (if rdram!=0x80000000)
261 #define FTEMP 40 // FPU temporary register
262 #define PTEMP 41 // Prefetch temporary register
263 //#define TLREG 42 // TLB mapping offset
264 #define RHASH 43 // Return address hash
265 #define RHTBL 44 // Return address hash table address
266 #define RTEMP 45 // JR/JALR address register
268 #define AGEN1 46 // Address generation temporary register
269 //#define AGEN2 47 // Address generation temporary register
270 //#define MGEN1 48 // Maptable address generation temporary register
271 //#define MGEN2 49 // Maptable address generation temporary register
272 #define BTREG 50 // Branch target temporary register
274 /* instruction types */
275 #define NOP 0 // No operation
276 #define LOAD 1 // Load
277 #define STORE 2 // Store
278 #define LOADLR 3 // Unaligned load
279 #define STORELR 4 // Unaligned store
280 #define MOV 5 // Move
281 #define ALU 6 // Arithmetic/logic
282 #define MULTDIV 7 // Multiply/divide
283 #define SHIFT 8 // Shift by register
284 #define SHIFTIMM 9// Shift by immediate
285 #define IMM16 10 // 16-bit immediate
286 #define RJUMP 11 // Unconditional jump to register
287 #define UJUMP 12 // Unconditional jump
288 #define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
289 #define SJUMP 14 // Conditional branch (regimm format)
290 #define COP0 15 // Coprocessor 0
291 #define COP1 16 // Coprocessor 1
292 #define C1LS 17 // Coprocessor 1 load/store
293 //#define FJUMP 18 // Conditional branch (floating point)
294 //#define FLOAT 19 // Floating point unit
295 //#define FCONV 20 // Convert integer to float
296 //#define FCOMP 21 // Floating point compare (sets FSREG)
297 #define SYSCALL 22// SYSCALL
298 #define OTHER 23 // Other
299 #define SPAN 24 // Branch/delay slot spans 2 pages
300 #define NI 25 // Not implemented
301 #define HLECALL 26// PCSX fake opcodes for HLE
302 #define COP2 27 // Coprocessor 2 move
303 #define C2LS 28 // Coprocessor 2 load/store
304 #define C2OP 29 // Coprocessor 2 operation
305 #define INTCALL 30// Call interpreter to handle rare corner cases
312 #define DJT_1 (void *)1l // no function, just a label in assem_debug log
313 #define DJT_2 (void *)2l
316 int new_recompile_block(u_int addr);
317 void *get_addr_ht(u_int vaddr);
318 void invalidate_block(u_int block);
319 void invalidate_addr(u_int addr);
320 void remove_hash(int vaddr);
322 void dyna_linker_ds();
324 void verify_code_ds();
327 void fp_exception_ds();
328 void jump_to_new_pc();
329 void call_gteStall();
330 void new_dyna_leave();
332 // Needed by assembler
333 static void wb_register(signed char r,signed char regmap[],uint64_t dirty);
334 static void wb_dirtys(signed char i_regmap[],uint64_t i_dirty);
335 static void wb_needed_dirtys(signed char i_regmap[],uint64_t i_dirty,int addr);
336 static void load_all_regs(signed char i_regmap[]);
337 static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
338 static void load_regs_entry(int t);
339 static void load_all_consts(signed char regmap[],u_int dirty,int i);
340 static u_int get_host_reglist(const signed char *regmap);
342 static int verify_dirty(const u_int *ptr);
343 static int get_final_value(int hr, int i, int *value);
344 static void add_stub(enum stub_type type, void *addr, void *retaddr,
345 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
346 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
347 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
348 static void add_to_linker(void *addr, u_int target, int ext);
349 static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override);
350 static void *get_direct_memhandler(void *table, u_int addr,
351 enum stub_type type, uintptr_t *addr_host);
352 static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
353 static void pass_args(int a0, int a1);
354 static void emit_far_jump(const void *f);
355 static void emit_far_call(const void *f);
357 static void mprotect_w_x(void *start, void *end, int is_x)
361 // *Open* enables write on all memory that was
362 // allocated by sceKernelAllocMemBlockForVM()?
364 sceKernelCloseVMDomain();
366 sceKernelOpenVMDomain();
368 u_long mstart = (u_long)start & ~4095ul;
369 u_long mend = (u_long)end;
370 if (mprotect((void *)mstart, mend - mstart,
371 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
372 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
377 static void start_tcache_write(void *start, void *end)
379 mprotect_w_x(start, end, 0);
382 static void end_tcache_write(void *start, void *end)
384 #if defined(__arm__) || defined(__aarch64__)
385 size_t len = (char *)end - (char *)start;
386 #if defined(__BLACKBERRY_QNX__)
387 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
388 #elif defined(__MACH__)
389 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
391 sceKernelSyncVMDomain(sceBlock, start, len);
393 ctr_flush_invalidate_cache();
394 #elif defined(__aarch64__)
395 // as of 2021, __clear_cache() is still broken on arm64
396 // so here is a custom one :(
397 clear_cache_arm64(start, end);
399 __clear_cache(start, end);
404 mprotect_w_x(start, end, 1);
407 static void *start_block(void)
409 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
410 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
411 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
412 start_tcache_write(out, end);
416 static void end_block(void *start)
418 end_tcache_write(start, out);
421 // also takes care of w^x mappings when patching code
422 static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
424 static void mark_clear_cache(void *target)
426 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
427 u_int mask = 1u << ((offset >> 12) & 31);
428 if (!(needs_clear_cache[offset >> 17] & mask)) {
429 char *start = (char *)((uintptr_t)target & ~4095l);
430 start_tcache_write(start, start + 4095);
431 needs_clear_cache[offset >> 17] |= mask;
435 // Clearing the cache is rather slow on ARM Linux, so mark the areas
436 // that need to be cleared, and then only clear these areas once.
437 static void do_clear_cache(void)
440 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
442 u_int bitmap = needs_clear_cache[i];
445 for (j = 0; j < 32; j++)
448 if (!(bitmap & (1<<j)))
451 start = ndrc->translation_cache + i*131072 + j*4096;
453 for (j++; j < 32; j++) {
454 if (!(bitmap & (1<<j)))
458 end_tcache_write(start, end);
460 needs_clear_cache[i] = 0;
464 //#define DEBUG_CYCLE_COUNT 1
466 #define NO_CYCLE_PENALTY_THR 12
468 int cycle_multiplier; // 100 for 1.0
469 int cycle_multiplier_override;
470 int cycle_multiplier_old;
472 static int CLOCK_ADJUST(int x)
474 int m = cycle_multiplier_override
475 ? cycle_multiplier_override : cycle_multiplier;
477 return (x * m + s * 50) / 100;
480 static int ds_writes_rjump_rs(int i)
482 return dops[i].rs1 != 0 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2);
485 static u_int get_page(u_int vaddr)
487 u_int page=vaddr&~0xe0000000;
488 if (page < 0x1000000)
489 page &= ~0x0e00000; // RAM mirrors
491 if(page>2048) page=2048+(page&2047);
495 // no virtual mem in PCSX
496 static u_int get_vpage(u_int vaddr)
498 return get_page(vaddr);
501 static struct ht_entry *hash_table_get(u_int vaddr)
503 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
506 static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
508 ht_bin->vaddr[1] = ht_bin->vaddr[0];
509 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
510 ht_bin->vaddr[0] = vaddr;
511 ht_bin->tcaddr[0] = tcaddr;
514 // some messy ari64's code, seems to rely on unsigned 32bit overflow
515 static int doesnt_expire_soon(void *tcaddr)
517 u_int diff = (u_int)((u_char *)tcaddr - out) << (32-TARGET_SIZE_2);
518 return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
521 // Get address from virtual address
522 // This is called from the recompiled JR/JALR instructions
523 void noinline *get_addr(u_int vaddr)
525 u_int page=get_page(vaddr);
526 u_int vpage=get_vpage(vaddr);
527 struct ll_entry *head;
528 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
531 if(head->vaddr==vaddr) {
532 //printf("TRACE: count=%d next=%d (get_addr match %x: %p)\n",Count,next_interupt,vaddr,head->addr);
533 hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
538 head=jump_dirty[vpage];
540 if(head->vaddr==vaddr) {
541 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %p)\n",Count,next_interupt,vaddr,head->addr);
542 // Don't restore blocks which are about to expire from the cache
543 if (doesnt_expire_soon(head->addr))
544 if (verify_dirty(head->addr)) {
545 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
546 invalid_code[vaddr>>12]=0;
547 inv_code_start=inv_code_end=~0;
549 restore_candidate[vpage>>3]|=1<<(vpage&7);
551 else restore_candidate[page>>3]|=1<<(page&7);
552 struct ht_entry *ht_bin = hash_table_get(vaddr);
553 if (ht_bin->vaddr[0] == vaddr)
554 ht_bin->tcaddr[0] = head->addr; // Replace existing entry
556 hash_table_add(ht_bin, vaddr, head->addr);
563 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
564 int r=new_recompile_block(vaddr);
565 if(r==0) return get_addr(vaddr);
566 // Execute in unmapped page, generate pagefault execption
568 Cause=(vaddr<<31)|0x8;
569 EPC=(vaddr&1)?vaddr-5:vaddr;
571 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
572 EntryHi=BadVAddr&0xFFFFE000;
573 return get_addr_ht(0x80000000);
575 // Look up address in hash table first
576 void *get_addr_ht(u_int vaddr)
578 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
579 const struct ht_entry *ht_bin = hash_table_get(vaddr);
580 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
581 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
582 return get_addr(vaddr);
585 void clear_all_regs(signed char regmap[])
588 for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
591 static signed char get_reg(const signed char regmap[],int r)
594 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map[hr]==r) return hr;
598 // Find a register that is available for two consecutive cycles
599 static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
602 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map1[hr]==r&®map2[hr]==r) return hr;
606 int count_free_regs(signed char regmap[])
610 for(hr=0;hr<HOST_REGS;hr++)
612 if(hr!=EXCLUDE_REG) {
613 if(regmap[hr]<0) count++;
619 void dirty_reg(struct regstat *cur,signed char reg)
623 for (hr=0;hr<HOST_REGS;hr++) {
624 if((cur->regmap[hr]&63)==reg) {
630 static void set_const(struct regstat *cur, signed char reg, uint32_t value)
634 for (hr=0;hr<HOST_REGS;hr++) {
635 if(cur->regmap[hr]==reg) {
637 current_constmap[hr]=value;
642 static void clear_const(struct regstat *cur, signed char reg)
646 for (hr=0;hr<HOST_REGS;hr++) {
647 if((cur->regmap[hr]&63)==reg) {
648 cur->isconst&=~(1<<hr);
653 static int is_const(struct regstat *cur, signed char reg)
658 for (hr=0;hr<HOST_REGS;hr++) {
659 if((cur->regmap[hr]&63)==reg) {
660 return (cur->isconst>>hr)&1;
666 static uint32_t get_const(struct regstat *cur, signed char reg)
670 for (hr=0;hr<HOST_REGS;hr++) {
671 if(cur->regmap[hr]==reg) {
672 return current_constmap[hr];
675 SysPrintf("Unknown constant in r%d\n",reg);
679 // Least soon needed registers
680 // Look at the next ten instructions and see which registers
681 // will be used. Try not to reallocate these.
682 void lsn(u_char hsn[], int i, int *preferred_reg)
692 if (dops[i+j].is_ujump)
694 // Don't go past an unconditonal jump
701 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
702 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
703 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
704 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
705 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
706 // Stores can allocate zero
707 hsn[dops[i+j].rs1]=j;
708 hsn[dops[i+j].rs2]=j;
710 // On some architectures stores need invc_ptr
711 #if defined(HOST_IMM8)
712 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR || (dops[i+j].opcode&0x3b)==0x39 || (dops[i+j].opcode&0x3b)==0x3a) {
716 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
724 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
726 // Follow first branch
727 int t=(ba[i+b]-start)>>2;
728 j=7-b;if(t+j>=slen) j=slen-t-1;
731 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
732 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
733 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
734 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
737 // TODO: preferred register based on backward branch
739 // Delay slot should preferably not overwrite branch conditions or cycle count
740 if (i > 0 && dops[i-1].is_jump) {
741 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
742 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
748 // Coprocessor load/store needs FTEMP, even if not declared
749 if(dops[i].itype==C1LS||dops[i].itype==C2LS) {
752 // Load L/R also uses FTEMP as a temporary register
753 if(dops[i].itype==LOADLR) {
756 // Also SWL/SWR/SDL/SDR
757 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
760 // Don't remove the miniht registers
761 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
768 // We only want to allocate registers if we're going to use them again soon
769 int needed_again(int r, int i)
775 if (i > 0 && dops[i-1].is_ujump)
777 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
778 return 0; // Don't need any registers if exiting the block
786 if (dops[i+j].is_ujump)
788 // Don't go past an unconditonal jump
792 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
799 if(dops[i+j].rs1==r) rn=j;
800 if(dops[i+j].rs2==r) rn=j;
801 if((unneeded_reg[i+j]>>r)&1) rn=10;
802 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
812 // Try to match register allocations at the end of a loop with those
814 int loop_reg(int i, int r, int hr)
823 if (dops[i+j].is_ujump)
825 // Don't go past an unconditonal jump
832 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
838 if((unneeded_reg[i+k]>>r)&1) return hr;
839 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
841 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
843 int t=(ba[i+k]-start)>>2;
844 int reg=get_reg(regs[t].regmap_entry,r);
845 if(reg>=0) return reg;
846 //reg=get_reg(regs[t+1].regmap_entry,r);
847 //if(reg>=0) return reg;
855 // Allocate every register, preserving source/target regs
856 void alloc_all(struct regstat *cur,int i)
860 for(hr=0;hr<HOST_REGS;hr++) {
861 if(hr!=EXCLUDE_REG) {
862 if(((cur->regmap[hr]&63)!=dops[i].rs1)&&((cur->regmap[hr]&63)!=dops[i].rs2)&&
863 ((cur->regmap[hr]&63)!=dops[i].rt1)&&((cur->regmap[hr]&63)!=dops[i].rt2))
866 cur->dirty&=~(1<<hr);
869 if((cur->regmap[hr]&63)==0)
872 cur->dirty&=~(1<<hr);
879 static int host_tempreg_in_use;
881 static void host_tempreg_acquire(void)
883 assert(!host_tempreg_in_use);
884 host_tempreg_in_use = 1;
887 static void host_tempreg_release(void)
889 host_tempreg_in_use = 0;
892 static void host_tempreg_acquire(void) {}
893 static void host_tempreg_release(void) {}
897 extern void gen_interupt();
898 extern void do_insn_cmp();
899 #define FUNCNAME(f) { f, " " #f }
900 static const struct {
903 } function_names[] = {
904 FUNCNAME(cc_interrupt),
905 FUNCNAME(gen_interupt),
906 FUNCNAME(get_addr_ht),
908 FUNCNAME(jump_handler_read8),
909 FUNCNAME(jump_handler_read16),
910 FUNCNAME(jump_handler_read32),
911 FUNCNAME(jump_handler_write8),
912 FUNCNAME(jump_handler_write16),
913 FUNCNAME(jump_handler_write32),
914 FUNCNAME(invalidate_addr),
915 FUNCNAME(jump_to_new_pc),
916 FUNCNAME(call_gteStall),
917 FUNCNAME(new_dyna_leave),
919 FUNCNAME(pcsx_mtc0_ds),
921 FUNCNAME(do_insn_cmp),
924 FUNCNAME(verify_code),
928 static const char *func_name(const void *a)
931 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
932 if (function_names[i].addr == a)
933 return function_names[i].name;
937 #define func_name(x) ""
941 #include "assem_x86.c"
944 #include "assem_x64.c"
947 #include "assem_arm.c"
950 #include "assem_arm64.c"
953 static void *get_trampoline(const void *f)
957 for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
958 if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
961 if (i == ARRAY_SIZE(ndrc->tramp.f)) {
962 SysPrintf("trampoline table is full, last func %p\n", f);
965 if (ndrc->tramp.f[i] == NULL) {
966 start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
967 ndrc->tramp.f[i] = f;
968 end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
970 return &ndrc->tramp.ops[i];
973 static void emit_far_jump(const void *f)
975 if (can_jump_or_call(f)) {
980 f = get_trampoline(f);
984 static void emit_far_call(const void *f)
986 if (can_jump_or_call(f)) {
991 f = get_trampoline(f);
995 // Add virtual address mapping to linked list
996 void ll_add(struct ll_entry **head,int vaddr,void *addr)
998 struct ll_entry *new_entry;
999 new_entry=malloc(sizeof(struct ll_entry));
1000 assert(new_entry!=NULL);
1001 new_entry->vaddr=vaddr;
1002 new_entry->reg_sv_flags=0;
1003 new_entry->addr=addr;
1004 new_entry->next=*head;
1008 void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
1010 ll_add(head,vaddr,addr);
1011 (*head)->reg_sv_flags=reg_sv_flags;
1014 // Check if an address is already compiled
1015 // but don't return addresses which are about to expire from the cache
1016 void *check_addr(u_int vaddr)
1018 struct ht_entry *ht_bin = hash_table_get(vaddr);
1020 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1021 if (ht_bin->vaddr[i] == vaddr)
1022 if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
1023 if (isclean(ht_bin->tcaddr[i]))
1024 return ht_bin->tcaddr[i];
1026 u_int page=get_page(vaddr);
1027 struct ll_entry *head;
1029 while (head != NULL) {
1030 if (head->vaddr == vaddr) {
1031 if (doesnt_expire_soon(head->addr)) {
1032 // Update existing entry with current address
1033 if (ht_bin->vaddr[0] == vaddr) {
1034 ht_bin->tcaddr[0] = head->addr;
1037 if (ht_bin->vaddr[1] == vaddr) {
1038 ht_bin->tcaddr[1] = head->addr;
1041 // Insert into hash table with low priority.
1042 // Don't evict existing entries, as they are probably
1043 // addresses that are being accessed frequently.
1044 if (ht_bin->vaddr[0] == -1) {
1045 ht_bin->vaddr[0] = vaddr;
1046 ht_bin->tcaddr[0] = head->addr;
1048 else if (ht_bin->vaddr[1] == -1) {
1049 ht_bin->vaddr[1] = vaddr;
1050 ht_bin->tcaddr[1] = head->addr;
1060 void remove_hash(int vaddr)
1062 //printf("remove hash: %x\n",vaddr);
1063 struct ht_entry *ht_bin = hash_table_get(vaddr);
1064 if (ht_bin->vaddr[1] == vaddr) {
1065 ht_bin->vaddr[1] = -1;
1066 ht_bin->tcaddr[1] = NULL;
1068 if (ht_bin->vaddr[0] == vaddr) {
1069 ht_bin->vaddr[0] = ht_bin->vaddr[1];
1070 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
1071 ht_bin->vaddr[1] = -1;
1072 ht_bin->tcaddr[1] = NULL;
1076 static void ll_remove_matching_addrs(struct ll_entry **head,
1077 uintptr_t base_offs_s, int shift)
1079 struct ll_entry *next;
1081 uintptr_t o1 = (u_char *)(*head)->addr - ndrc->translation_cache;
1082 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1083 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1085 inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
1086 remove_hash((*head)->vaddr);
1093 head=&((*head)->next);
1098 // Remove all entries from linked list
1099 void ll_clear(struct ll_entry **head)
1101 struct ll_entry *cur;
1102 struct ll_entry *next;
1113 // Dereference the pointers and remove if it matches
1114 static void ll_kill_pointers(struct ll_entry *head,
1115 uintptr_t base_offs_s, int shift)
1118 u_char *ptr = get_pointer(head->addr);
1119 uintptr_t o1 = ptr - ndrc->translation_cache;
1120 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1121 inv_debug("EXP: Lookup pointer to %p at %p (%x)\n",ptr,head->addr,head->vaddr);
1122 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1124 inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
1125 void *host_addr=find_extjump_insn(head->addr);
1126 mark_clear_cache(host_addr);
1127 set_jump_target(host_addr, head->addr);
1133 // This is called when we write to a compiled block (see do_invstub)
1134 static void invalidate_page(u_int page)
1136 struct ll_entry *head;
1137 struct ll_entry *next;
1141 inv_debug("INVALIDATE: %x\n",head->vaddr);
1142 remove_hash(head->vaddr);
1147 head=jump_out[page];
1150 inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
1151 void *host_addr=find_extjump_insn(head->addr);
1152 mark_clear_cache(host_addr);
1153 set_jump_target(host_addr, head->addr); // point back to dyna_linker
1160 static void invalidate_block_range(u_int block, u_int first, u_int last)
1162 u_int page=get_page(block<<12);
1163 //printf("first=%d last=%d\n",first,last);
1164 invalidate_page(page);
1165 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1166 assert(last<page+5);
1167 // Invalidate the adjacent pages if a block crosses a 4K boundary
1169 invalidate_page(first);
1172 for(first=page+1;first<last;first++) {
1173 invalidate_page(first);
1177 // Don't trap writes
1178 invalid_code[block]=1;
1181 memset(mini_ht,-1,sizeof(mini_ht));
1185 void invalidate_block(u_int block)
1187 u_int page=get_page(block<<12);
1188 u_int vpage=get_vpage(block<<12);
1189 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1190 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1193 struct ll_entry *head;
1194 head=jump_dirty[vpage];
1195 //printf("page=%d vpage=%d\n",page,vpage);
1197 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1198 u_char *start, *end;
1199 get_bounds(head->addr, &start, &end);
1200 //printf("start: %p end: %p\n", start, end);
1201 if (page < 2048 && start >= rdram && end < rdram+RAM_SIZE) {
1202 if (((start-rdram)>>12) <= page && ((end-1-rdram)>>12) >= page) {
1203 if ((((start-rdram)>>12)&2047) < first) first = ((start-rdram)>>12)&2047;
1204 if ((((end-1-rdram)>>12)&2047) > last) last = ((end-1-rdram)>>12)&2047;
1210 invalidate_block_range(block,first,last);
1213 void invalidate_addr(u_int addr)
1216 // this check is done by the caller
1217 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1218 u_int page=get_vpage(addr);
1219 if(page<2048) { // RAM
1220 struct ll_entry *head;
1221 u_int addr_min=~0, addr_max=0;
1222 u_int mask=RAM_SIZE-1;
1223 u_int addr_main=0x80000000|(addr&mask);
1225 inv_code_start=addr_main&~0xfff;
1226 inv_code_end=addr_main|0xfff;
1229 // must check previous page too because of spans..
1231 inv_code_start-=0x1000;
1233 for(;pg1<=page;pg1++) {
1234 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1235 u_char *start_h, *end_h;
1237 get_bounds(head->addr, &start_h, &end_h);
1238 start = (uintptr_t)start_h - ram_offset;
1239 end = (uintptr_t)end_h - ram_offset;
1240 if(start<=addr_main&&addr_main<end) {
1241 if(start<addr_min) addr_min=start;
1242 if(end>addr_max) addr_max=end;
1244 else if(addr_main<start) {
1245 if(start<inv_code_end)
1246 inv_code_end=start-1;
1249 if(end>inv_code_start)
1255 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1256 inv_code_start=inv_code_end=~0;
1257 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1261 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1262 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1263 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1267 invalidate_block(addr>>12);
1270 // This is called when loading a save state.
1271 // Anything could have changed, so invalidate everything.
1272 void invalidate_all_pages(void)
1275 for(page=0;page<4096;page++)
1276 invalidate_page(page);
1277 for(page=0;page<1048576;page++)
1278 if(!invalid_code[page]) {
1279 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1280 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1283 memset(mini_ht,-1,sizeof(mini_ht));
1288 static void do_invstub(int n)
1291 u_int reglist=stubs[n].a;
1292 set_jump_target(stubs[n].addr, out);
1294 if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
1295 emit_far_call(invalidate_addr);
1296 restore_regs(reglist);
1297 emit_jmp(stubs[n].retaddr); // return address
1300 // Add an entry to jump_out after making a link
1301 // src should point to code by emit_extjump2()
1302 void add_jump_out(u_int vaddr,void *src)
1304 u_int page=get_page(vaddr);
1305 inv_debug("add_jump_out: %p -> %x (%d)\n",src,vaddr,page);
1306 check_extjump2(src);
1307 ll_add(jump_out+page,vaddr,src);
1308 //inv_debug("add_jump_out: to %p\n",get_pointer(src));
1311 // If a code block was found to be unmodified (bit was set in
1312 // restore_candidate) and it remains unmodified (bit is clear
1313 // in invalid_code) then move the entries for that 4K page from
1314 // the dirty list to the clean list.
1315 void clean_blocks(u_int page)
1317 struct ll_entry *head;
1318 inv_debug("INV: clean_blocks page=%d\n",page);
1319 head=jump_dirty[page];
1321 if(!invalid_code[head->vaddr>>12]) {
1322 // Don't restore blocks which are about to expire from the cache
1323 if (doesnt_expire_soon(head->addr)) {
1324 if(verify_dirty(head->addr)) {
1325 u_char *start, *end;
1326 //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
1329 get_bounds(head->addr, &start, &end);
1330 if (start - rdram < RAM_SIZE) {
1331 for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
1332 inv|=invalid_code[i];
1335 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1339 void *clean_addr = get_clean_addr(head->addr);
1340 if (doesnt_expire_soon(clean_addr)) {
1342 inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, clean_addr);
1343 //printf("page=%x, addr=%x\n",page,head->vaddr);
1344 //assert(head->vaddr>>12==(page|0x80000));
1345 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1346 struct ht_entry *ht_bin = hash_table_get(head->vaddr);
1347 if (ht_bin->vaddr[0] == head->vaddr)
1348 ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
1349 if (ht_bin->vaddr[1] == head->vaddr)
1350 ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
1360 /* Register allocation */
1362 // Note: registers are allocated clean (unmodified state)
1363 // if you intend to modify the register, you must call dirty_reg().
1364 static void alloc_reg(struct regstat *cur,int i,signed char reg)
1367 int preferred_reg = (reg&7);
1368 if(reg==CCREG) preferred_reg=HOST_CCREG;
1369 if(reg==PTEMP||reg==FTEMP) preferred_reg=12;
1371 // Don't allocate unused registers
1372 if((cur->u>>reg)&1) return;
1374 // see if it's already allocated
1375 for(hr=0;hr<HOST_REGS;hr++)
1377 if(cur->regmap[hr]==reg) return;
1380 // Keep the same mapping if the register was already allocated in a loop
1381 preferred_reg = loop_reg(i,reg,preferred_reg);
1383 // Try to allocate the preferred register
1384 if(cur->regmap[preferred_reg]==-1) {
1385 cur->regmap[preferred_reg]=reg;
1386 cur->dirty&=~(1<<preferred_reg);
1387 cur->isconst&=~(1<<preferred_reg);
1390 r=cur->regmap[preferred_reg];
1393 cur->regmap[preferred_reg]=reg;
1394 cur->dirty&=~(1<<preferred_reg);
1395 cur->isconst&=~(1<<preferred_reg);
1399 // Clear any unneeded registers
1400 // We try to keep the mapping consistent, if possible, because it
1401 // makes branches easier (especially loops). So we try to allocate
1402 // first (see above) before removing old mappings. If this is not
1403 // possible then go ahead and clear out the registers that are no
1405 for(hr=0;hr<HOST_REGS;hr++)
1410 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1413 // Try to allocate any available register, but prefer
1414 // registers that have not been used recently.
1416 for(hr=0;hr<HOST_REGS;hr++) {
1417 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1418 if(regs[i-1].regmap[hr]!=dops[i-1].rs1&®s[i-1].regmap[hr]!=dops[i-1].rs2&®s[i-1].regmap[hr]!=dops[i-1].rt1&®s[i-1].regmap[hr]!=dops[i-1].rt2) {
1419 cur->regmap[hr]=reg;
1420 cur->dirty&=~(1<<hr);
1421 cur->isconst&=~(1<<hr);
1427 // Try to allocate any available register
1428 for(hr=0;hr<HOST_REGS;hr++) {
1429 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1430 cur->regmap[hr]=reg;
1431 cur->dirty&=~(1<<hr);
1432 cur->isconst&=~(1<<hr);
1437 // Ok, now we have to evict someone
1438 // Pick a register we hopefully won't need soon
1439 u_char hsn[MAXREG+1];
1440 memset(hsn,10,sizeof(hsn));
1442 lsn(hsn,i,&preferred_reg);
1443 //printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",cur->regmap[0],cur->regmap[1],cur->regmap[2],cur->regmap[3],cur->regmap[5],cur->regmap[6],cur->regmap[7]);
1444 //printf("hsn(%x): %d %d %d %d %d %d %d\n",start+i*4,hsn[cur->regmap[0]&63],hsn[cur->regmap[1]&63],hsn[cur->regmap[2]&63],hsn[cur->regmap[3]&63],hsn[cur->regmap[5]&63],hsn[cur->regmap[6]&63],hsn[cur->regmap[7]&63]);
1446 // Don't evict the cycle count at entry points, otherwise the entry
1447 // stub will have to write it.
1448 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1449 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1452 // Alloc preferred register if available
1453 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1454 for(hr=0;hr<HOST_REGS;hr++) {
1455 // Evict both parts of a 64-bit register
1456 if((cur->regmap[hr]&63)==r) {
1458 cur->dirty&=~(1<<hr);
1459 cur->isconst&=~(1<<hr);
1462 cur->regmap[preferred_reg]=reg;
1465 for(r=1;r<=MAXREG;r++)
1467 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1468 for(hr=0;hr<HOST_REGS;hr++) {
1469 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1470 if(cur->regmap[hr]==r) {
1471 cur->regmap[hr]=reg;
1472 cur->dirty&=~(1<<hr);
1473 cur->isconst&=~(1<<hr);
1484 for(r=1;r<=MAXREG;r++)
1487 for(hr=0;hr<HOST_REGS;hr++) {
1488 if(cur->regmap[hr]==r) {
1489 cur->regmap[hr]=reg;
1490 cur->dirty&=~(1<<hr);
1491 cur->isconst&=~(1<<hr);
1498 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1501 // Allocate a temporary register. This is done without regard to
1502 // dirty status or whether the register we request is on the unneeded list
1503 // Note: This will only allocate one register, even if called multiple times
1504 static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1507 int preferred_reg = -1;
1509 // see if it's already allocated
1510 for(hr=0;hr<HOST_REGS;hr++)
1512 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1515 // Try to allocate any available register
1516 for(hr=HOST_REGS-1;hr>=0;hr--) {
1517 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1518 cur->regmap[hr]=reg;
1519 cur->dirty&=~(1<<hr);
1520 cur->isconst&=~(1<<hr);
1525 // Find an unneeded register
1526 for(hr=HOST_REGS-1;hr>=0;hr--)
1532 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1533 cur->regmap[hr]=reg;
1534 cur->dirty&=~(1<<hr);
1535 cur->isconst&=~(1<<hr);
1542 // Ok, now we have to evict someone
1543 // Pick a register we hopefully won't need soon
1544 // TODO: we might want to follow unconditional jumps here
1545 // TODO: get rid of dupe code and make this into a function
1546 u_char hsn[MAXREG+1];
1547 memset(hsn,10,sizeof(hsn));
1549 lsn(hsn,i,&preferred_reg);
1550 //printf("hsn: %d %d %d %d %d %d %d\n",hsn[cur->regmap[0]&63],hsn[cur->regmap[1]&63],hsn[cur->regmap[2]&63],hsn[cur->regmap[3]&63],hsn[cur->regmap[5]&63],hsn[cur->regmap[6]&63],hsn[cur->regmap[7]&63]);
1552 // Don't evict the cycle count at entry points, otherwise the entry
1553 // stub will have to write it.
1554 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1555 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1558 for(r=1;r<=MAXREG;r++)
1560 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1561 for(hr=0;hr<HOST_REGS;hr++) {
1562 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1563 if(cur->regmap[hr]==r) {
1564 cur->regmap[hr]=reg;
1565 cur->dirty&=~(1<<hr);
1566 cur->isconst&=~(1<<hr);
1577 for(r=1;r<=MAXREG;r++)
1580 for(hr=0;hr<HOST_REGS;hr++) {
1581 if(cur->regmap[hr]==r) {
1582 cur->regmap[hr]=reg;
1583 cur->dirty&=~(1<<hr);
1584 cur->isconst&=~(1<<hr);
1591 SysPrintf("This shouldn't happen");abort();
1594 static void mov_alloc(struct regstat *current,int i)
1596 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1597 // logically this is needed but just won't work, no idea why
1598 //alloc_cc(current,i); // for stalls
1599 //dirty_reg(current,CCREG);
1602 // Note: Don't need to actually alloc the source registers
1603 //alloc_reg(current,i,dops[i].rs1);
1604 alloc_reg(current,i,dops[i].rt1);
1606 clear_const(current,dops[i].rs1);
1607 clear_const(current,dops[i].rt1);
1608 dirty_reg(current,dops[i].rt1);
1611 static void shiftimm_alloc(struct regstat *current,int i)
1613 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1616 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1617 else dops[i].lt1=dops[i].rs1;
1618 alloc_reg(current,i,dops[i].rt1);
1619 dirty_reg(current,dops[i].rt1);
1620 if(is_const(current,dops[i].rs1)) {
1621 int v=get_const(current,dops[i].rs1);
1622 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1623 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1624 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1626 else clear_const(current,dops[i].rt1);
1631 clear_const(current,dops[i].rs1);
1632 clear_const(current,dops[i].rt1);
1635 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1639 if(dops[i].opcode2==0x3c) // DSLL32
1643 if(dops[i].opcode2==0x3e) // DSRL32
1647 if(dops[i].opcode2==0x3f) // DSRA32
1653 static void shift_alloc(struct regstat *current,int i)
1656 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1658 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1659 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1660 alloc_reg(current,i,dops[i].rt1);
1661 if(dops[i].rt1==dops[i].rs2) {
1662 alloc_reg_temp(current,i,-1);
1663 minimum_free_regs[i]=1;
1665 } else { // DSLLV/DSRLV/DSRAV
1668 clear_const(current,dops[i].rs1);
1669 clear_const(current,dops[i].rs2);
1670 clear_const(current,dops[i].rt1);
1671 dirty_reg(current,dops[i].rt1);
1675 static void alu_alloc(struct regstat *current,int i)
1677 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1679 if(dops[i].rs1&&dops[i].rs2) {
1680 alloc_reg(current,i,dops[i].rs1);
1681 alloc_reg(current,i,dops[i].rs2);
1684 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1685 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1687 alloc_reg(current,i,dops[i].rt1);
1690 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1692 alloc_reg(current,i,dops[i].rs1);
1693 alloc_reg(current,i,dops[i].rs2);
1694 alloc_reg(current,i,dops[i].rt1);
1697 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1699 if(dops[i].rs1&&dops[i].rs2) {
1700 alloc_reg(current,i,dops[i].rs1);
1701 alloc_reg(current,i,dops[i].rs2);
1705 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1706 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1708 alloc_reg(current,i,dops[i].rt1);
1711 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1714 clear_const(current,dops[i].rs1);
1715 clear_const(current,dops[i].rs2);
1716 clear_const(current,dops[i].rt1);
1717 dirty_reg(current,dops[i].rt1);
1720 static void imm16_alloc(struct regstat *current,int i)
1722 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1723 else dops[i].lt1=dops[i].rs1;
1724 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1725 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1728 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1729 clear_const(current,dops[i].rs1);
1730 clear_const(current,dops[i].rt1);
1732 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1733 if(is_const(current,dops[i].rs1)) {
1734 int v=get_const(current,dops[i].rs1);
1735 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1736 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1737 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1739 else clear_const(current,dops[i].rt1);
1741 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1742 if(is_const(current,dops[i].rs1)) {
1743 int v=get_const(current,dops[i].rs1);
1744 set_const(current,dops[i].rt1,v+imm[i]);
1746 else clear_const(current,dops[i].rt1);
1749 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1751 dirty_reg(current,dops[i].rt1);
1754 static void load_alloc(struct regstat *current,int i)
1756 clear_const(current,dops[i].rt1);
1757 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
1758 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
1759 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1760 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
1761 alloc_reg(current,i,dops[i].rt1);
1762 assert(get_reg(current->regmap,dops[i].rt1)>=0);
1763 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
1767 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1771 dirty_reg(current,dops[i].rt1);
1772 // LWL/LWR need a temporary register for the old value
1773 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1775 alloc_reg(current,i,FTEMP);
1776 alloc_reg_temp(current,i,-1);
1777 minimum_free_regs[i]=1;
1782 // Load to r0 or unneeded register (dummy load)
1783 // but we still need a register to calculate the address
1784 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1786 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1788 alloc_reg_temp(current,i,-1);
1789 minimum_free_regs[i]=1;
1790 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1797 void store_alloc(struct regstat *current,int i)
1799 clear_const(current,dops[i].rs2);
1800 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
1801 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1802 alloc_reg(current,i,dops[i].rs2);
1803 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
1806 #if defined(HOST_IMM8)
1807 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1808 else alloc_reg(current,i,INVCP);
1810 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
1811 alloc_reg(current,i,FTEMP);
1813 // We need a temporary register for address generation
1814 alloc_reg_temp(current,i,-1);
1815 minimum_free_regs[i]=1;
1818 void c1ls_alloc(struct regstat *current,int i)
1820 //clear_const(current,dops[i].rs1); // FIXME
1821 clear_const(current,dops[i].rt1);
1822 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1823 alloc_reg(current,i,CSREG); // Status
1824 alloc_reg(current,i,FTEMP);
1825 if(dops[i].opcode==0x35||dops[i].opcode==0x3d) { // 64-bit LDC1/SDC1
1828 #if defined(HOST_IMM8)
1829 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1830 else if((dops[i].opcode&0x3b)==0x39) // SWC1/SDC1
1831 alloc_reg(current,i,INVCP);
1833 // We need a temporary register for address generation
1834 alloc_reg_temp(current,i,-1);
1837 void c2ls_alloc(struct regstat *current,int i)
1839 clear_const(current,dops[i].rt1);
1840 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1841 alloc_reg(current,i,FTEMP);
1842 #if defined(HOST_IMM8)
1843 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1844 if((dops[i].opcode&0x3b)==0x3a) // SWC2/SDC2
1845 alloc_reg(current,i,INVCP);
1847 // We need a temporary register for address generation
1848 alloc_reg_temp(current,i,-1);
1849 minimum_free_regs[i]=1;
1852 #ifndef multdiv_alloc
1853 void multdiv_alloc(struct regstat *current,int i)
1860 // case 0x1D: DMULTU
1863 clear_const(current,dops[i].rs1);
1864 clear_const(current,dops[i].rs2);
1865 alloc_cc(current,i); // for stalls
1866 if(dops[i].rs1&&dops[i].rs2)
1868 if((dops[i].opcode2&4)==0) // 32-bit
1870 current->u&=~(1LL<<HIREG);
1871 current->u&=~(1LL<<LOREG);
1872 alloc_reg(current,i,HIREG);
1873 alloc_reg(current,i,LOREG);
1874 alloc_reg(current,i,dops[i].rs1);
1875 alloc_reg(current,i,dops[i].rs2);
1876 dirty_reg(current,HIREG);
1877 dirty_reg(current,LOREG);
1886 // Multiply by zero is zero.
1887 // MIPS does not have a divide by zero exception.
1888 // The result is undefined, we return zero.
1889 alloc_reg(current,i,HIREG);
1890 alloc_reg(current,i,LOREG);
1891 dirty_reg(current,HIREG);
1892 dirty_reg(current,LOREG);
1897 void cop0_alloc(struct regstat *current,int i)
1899 if(dops[i].opcode2==0) // MFC0
1902 clear_const(current,dops[i].rt1);
1903 alloc_all(current,i);
1904 alloc_reg(current,i,dops[i].rt1);
1905 dirty_reg(current,dops[i].rt1);
1908 else if(dops[i].opcode2==4) // MTC0
1911 clear_const(current,dops[i].rs1);
1912 alloc_reg(current,i,dops[i].rs1);
1913 alloc_all(current,i);
1916 alloc_all(current,i); // FIXME: Keep r0
1918 alloc_reg(current,i,0);
1923 // TLBR/TLBWI/TLBWR/TLBP/ERET
1924 assert(dops[i].opcode2==0x10);
1925 alloc_all(current,i);
1927 minimum_free_regs[i]=HOST_REGS;
1930 static void cop2_alloc(struct regstat *current,int i)
1932 if (dops[i].opcode2 < 3) // MFC2/CFC2
1934 alloc_cc(current,i); // for stalls
1935 dirty_reg(current,CCREG);
1937 clear_const(current,dops[i].rt1);
1938 alloc_reg(current,i,dops[i].rt1);
1939 dirty_reg(current,dops[i].rt1);
1942 else if (dops[i].opcode2 > 3) // MTC2/CTC2
1945 clear_const(current,dops[i].rs1);
1946 alloc_reg(current,i,dops[i].rs1);
1950 alloc_reg(current,i,0);
1953 alloc_reg_temp(current,i,-1);
1954 minimum_free_regs[i]=1;
1957 void c2op_alloc(struct regstat *current,int i)
1959 alloc_cc(current,i); // for stalls
1960 dirty_reg(current,CCREG);
1961 alloc_reg_temp(current,i,-1);
1964 void syscall_alloc(struct regstat *current,int i)
1966 alloc_cc(current,i);
1967 dirty_reg(current,CCREG);
1968 alloc_all(current,i);
1969 minimum_free_regs[i]=HOST_REGS;
1973 void delayslot_alloc(struct regstat *current,int i)
1975 switch(dops[i].itype) {
1983 assem_debug("jump in the delay slot. this shouldn't happen.\n");//abort();
1984 SysPrintf("Disabled speculative precompilation\n");
1988 imm16_alloc(current,i);
1992 load_alloc(current,i);
1996 store_alloc(current,i);
1999 alu_alloc(current,i);
2002 shift_alloc(current,i);
2005 multdiv_alloc(current,i);
2008 shiftimm_alloc(current,i);
2011 mov_alloc(current,i);
2014 cop0_alloc(current,i);
2019 cop2_alloc(current,i);
2022 c1ls_alloc(current,i);
2025 c2ls_alloc(current,i);
2028 c2op_alloc(current,i);
2033 // Special case where a branch and delay slot span two pages in virtual memory
2034 static void pagespan_alloc(struct regstat *current,int i)
2037 current->wasconst=0;
2039 minimum_free_regs[i]=HOST_REGS;
2040 alloc_all(current,i);
2041 alloc_cc(current,i);
2042 dirty_reg(current,CCREG);
2043 if(dops[i].opcode==3) // JAL
2045 alloc_reg(current,i,31);
2046 dirty_reg(current,31);
2048 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
2050 alloc_reg(current,i,dops[i].rs1);
2051 if (dops[i].rt1!=0) {
2052 alloc_reg(current,i,dops[i].rt1);
2053 dirty_reg(current,dops[i].rt1);
2056 if((dops[i].opcode&0x2E)==4) // BEQ/BNE/BEQL/BNEL
2058 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2059 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
2062 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
2064 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2069 static void add_stub(enum stub_type type, void *addr, void *retaddr,
2070 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2072 assert(stubcount < ARRAY_SIZE(stubs));
2073 stubs[stubcount].type = type;
2074 stubs[stubcount].addr = addr;
2075 stubs[stubcount].retaddr = retaddr;
2076 stubs[stubcount].a = a;
2077 stubs[stubcount].b = b;
2078 stubs[stubcount].c = c;
2079 stubs[stubcount].d = d;
2080 stubs[stubcount].e = e;
2084 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2085 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2087 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2090 // Write out a single register
2091 static void wb_register(signed char r,signed char regmap[],uint64_t dirty)
2094 for(hr=0;hr<HOST_REGS;hr++) {
2095 if(hr!=EXCLUDE_REG) {
2096 if((regmap[hr]&63)==r) {
2098 assert(regmap[hr]<64);
2099 emit_storereg(r,hr);
2106 static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2108 //if(dirty_pre==dirty) return;
2110 for(hr=0;hr<HOST_REGS;hr++) {
2111 if(hr!=EXCLUDE_REG) {
2113 if(((~u)>>(reg&63))&1) {
2115 if(((dirty_pre&~dirty)>>hr)&1) {
2117 emit_storereg(reg,hr);
2130 static void pass_args(int a0, int a1)
2134 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2136 else if(a0!=0&&a1==0) {
2138 if (a0>=0) emit_mov(a0,0);
2141 if(a0>=0&&a0!=0) emit_mov(a0,0);
2142 if(a1>=0&&a1!=1) emit_mov(a1,1);
2146 static void alu_assemble(int i,struct regstat *i_regs)
2148 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2150 signed char s1,s2,t;
2151 t=get_reg(i_regs->regmap,dops[i].rt1);
2153 s1=get_reg(i_regs->regmap,dops[i].rs1);
2154 s2=get_reg(i_regs->regmap,dops[i].rs2);
2155 if(dops[i].rs1&&dops[i].rs2) {
2158 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2159 else emit_add(s1,s2,t);
2161 else if(dops[i].rs1) {
2162 if(s1>=0) emit_mov(s1,t);
2163 else emit_loadreg(dops[i].rs1,t);
2165 else if(dops[i].rs2) {
2167 if(dops[i].opcode2&2) emit_neg(s2,t);
2168 else emit_mov(s2,t);
2171 emit_loadreg(dops[i].rs2,t);
2172 if(dops[i].opcode2&2) emit_neg(t,t);
2175 else emit_zeroreg(t);
2179 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2182 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2184 signed char s1l,s2l,t;
2186 t=get_reg(i_regs->regmap,dops[i].rt1);
2189 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2190 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2191 if(dops[i].rs2==0) // rx<r0
2193 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2195 emit_shrimm(s1l,31,t);
2197 else // SLTU (unsigned can not be less than zero, 0<0)
2200 else if(dops[i].rs1==0) // r0<rx
2203 if(dops[i].opcode2==0x2a) // SLT
2204 emit_set_gz32(s2l,t);
2205 else // SLTU (set if not zero)
2206 emit_set_nz32(s2l,t);
2209 assert(s1l>=0);assert(s2l>=0);
2210 if(dops[i].opcode2==0x2a) // SLT
2211 emit_set_if_less32(s1l,s2l,t);
2213 emit_set_if_carry32(s1l,s2l,t);
2219 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2221 signed char s1l,s2l,tl;
2222 tl=get_reg(i_regs->regmap,dops[i].rt1);
2225 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2226 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2227 if(dops[i].rs1&&dops[i].rs2) {
2230 if(dops[i].opcode2==0x24) { // AND
2231 emit_and(s1l,s2l,tl);
2233 if(dops[i].opcode2==0x25) { // OR
2234 emit_or(s1l,s2l,tl);
2236 if(dops[i].opcode2==0x26) { // XOR
2237 emit_xor(s1l,s2l,tl);
2239 if(dops[i].opcode2==0x27) { // NOR
2240 emit_or(s1l,s2l,tl);
2246 if(dops[i].opcode2==0x24) { // AND
2249 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2251 if(s1l>=0) emit_mov(s1l,tl);
2252 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2256 if(s2l>=0) emit_mov(s2l,tl);
2257 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2259 else emit_zeroreg(tl);
2261 if(dops[i].opcode2==0x27) { // NOR
2263 if(s1l>=0) emit_not(s1l,tl);
2265 emit_loadreg(dops[i].rs1,tl);
2271 if(s2l>=0) emit_not(s2l,tl);
2273 emit_loadreg(dops[i].rs2,tl);
2277 else emit_movimm(-1,tl);
2286 void imm16_assemble(int i,struct regstat *i_regs)
2288 if (dops[i].opcode==0x0f) { // LUI
2291 t=get_reg(i_regs->regmap,dops[i].rt1);
2294 if(!((i_regs->isconst>>t)&1))
2295 emit_movimm(imm[i]<<16,t);
2299 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2302 t=get_reg(i_regs->regmap,dops[i].rt1);
2303 s=get_reg(i_regs->regmap,dops[i].rs1);
2308 if(!((i_regs->isconst>>t)&1)) {
2310 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2311 emit_addimm(t,imm[i],t);
2313 if(!((i_regs->wasconst>>s)&1))
2314 emit_addimm(s,imm[i],t);
2316 emit_movimm(constmap[i][s]+imm[i],t);
2322 if(!((i_regs->isconst>>t)&1))
2323 emit_movimm(imm[i],t);
2328 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2331 tl=get_reg(i_regs->regmap,dops[i].rt1);
2332 sl=get_reg(i_regs->regmap,dops[i].rs1);
2336 emit_addimm(sl,imm[i],tl);
2338 emit_movimm(imm[i],tl);
2343 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2345 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2347 t=get_reg(i_regs->regmap,dops[i].rt1);
2348 sl=get_reg(i_regs->regmap,dops[i].rs1);
2352 if(dops[i].opcode==0x0a) { // SLTI
2354 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2355 emit_slti32(t,imm[i],t);
2357 emit_slti32(sl,imm[i],t);
2362 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2363 emit_sltiu32(t,imm[i],t);
2365 emit_sltiu32(sl,imm[i],t);
2369 // SLTI(U) with r0 is just stupid,
2370 // nonetheless examples can be found
2371 if(dops[i].opcode==0x0a) // SLTI
2372 if(0<imm[i]) emit_movimm(1,t);
2373 else emit_zeroreg(t);
2376 if(imm[i]) emit_movimm(1,t);
2377 else emit_zeroreg(t);
2383 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2386 tl=get_reg(i_regs->regmap,dops[i].rt1);
2387 sl=get_reg(i_regs->regmap,dops[i].rs1);
2388 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2389 if(dops[i].opcode==0x0c) //ANDI
2393 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2394 emit_andimm(tl,imm[i],tl);
2396 if(!((i_regs->wasconst>>sl)&1))
2397 emit_andimm(sl,imm[i],tl);
2399 emit_movimm(constmap[i][sl]&imm[i],tl);
2409 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2411 if(dops[i].opcode==0x0d) { // ORI
2413 emit_orimm(tl,imm[i],tl);
2415 if(!((i_regs->wasconst>>sl)&1))
2416 emit_orimm(sl,imm[i],tl);
2418 emit_movimm(constmap[i][sl]|imm[i],tl);
2421 if(dops[i].opcode==0x0e) { // XORI
2423 emit_xorimm(tl,imm[i],tl);
2425 if(!((i_regs->wasconst>>sl)&1))
2426 emit_xorimm(sl,imm[i],tl);
2428 emit_movimm(constmap[i][sl]^imm[i],tl);
2433 emit_movimm(imm[i],tl);
2441 void shiftimm_assemble(int i,struct regstat *i_regs)
2443 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2447 t=get_reg(i_regs->regmap,dops[i].rt1);
2448 s=get_reg(i_regs->regmap,dops[i].rs1);
2450 if(t>=0&&!((i_regs->isconst>>t)&1)){
2457 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2459 if(dops[i].opcode2==0) // SLL
2461 emit_shlimm(s<0?t:s,imm[i],t);
2463 if(dops[i].opcode2==2) // SRL
2465 emit_shrimm(s<0?t:s,imm[i],t);
2467 if(dops[i].opcode2==3) // SRA
2469 emit_sarimm(s<0?t:s,imm[i],t);
2473 if(s>=0 && s!=t) emit_mov(s,t);
2477 //emit_storereg(dops[i].rt1,t); //DEBUG
2480 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2484 if(dops[i].opcode2==0x3c) // DSLL32
2488 if(dops[i].opcode2==0x3e) // DSRL32
2492 if(dops[i].opcode2==0x3f) // DSRA32
2498 #ifndef shift_assemble
2499 static void shift_assemble(int i,struct regstat *i_regs)
2501 signed char s,t,shift;
2502 if (dops[i].rt1 == 0)
2504 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2505 t = get_reg(i_regs->regmap, dops[i].rt1);
2506 s = get_reg(i_regs->regmap, dops[i].rs1);
2507 shift = get_reg(i_regs->regmap, dops[i].rs2);
2513 else if(dops[i].rs2==0) {
2515 if(s!=t) emit_mov(s,t);
2518 host_tempreg_acquire();
2519 emit_andimm(shift,31,HOST_TEMPREG);
2520 switch(dops[i].opcode2) {
2522 emit_shl(s,HOST_TEMPREG,t);
2525 emit_shr(s,HOST_TEMPREG,t);
2528 emit_sar(s,HOST_TEMPREG,t);
2533 host_tempreg_release();
2547 static int get_ptr_mem_type(u_int a)
2549 if(a < 0x00200000) {
2550 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2551 // return wrong, must use memhandler for BIOS self-test to pass
2552 // 007 does similar stuff from a00 mirror, weird stuff
2556 if(0x1f800000 <= a && a < 0x1f801000)
2558 if(0x80200000 <= a && a < 0x80800000)
2560 if(0xa0000000 <= a && a < 0xa0200000)
2565 static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
2570 if(((smrv_strong|smrv_weak)>>mr)&1) {
2571 type=get_ptr_mem_type(smrv[mr]);
2572 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2575 // use the mirror we are running on
2576 type=get_ptr_mem_type(start);
2577 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2580 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2581 host_tempreg_acquire();
2582 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2583 addr=*addr_reg_override=HOST_TEMPREG;
2586 else if(type==MTYPE_0000) { // RAM 0 mirror
2587 host_tempreg_acquire();
2588 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2589 addr=*addr_reg_override=HOST_TEMPREG;
2592 else if(type==MTYPE_A000) { // RAM A mirror
2593 host_tempreg_acquire();
2594 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2595 addr=*addr_reg_override=HOST_TEMPREG;
2598 else if(type==MTYPE_1F80) { // scratchpad
2599 if (psxH == (void *)0x1f800000) {
2600 host_tempreg_acquire();
2601 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2602 emit_cmpimm(HOST_TEMPREG,0x1000);
2603 host_tempreg_release();
2608 // do the usual RAM check, jump will go to the right handler
2615 emit_cmpimm(addr,RAM_SIZE);
2617 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2618 // Hint to branch predictor that the branch is unlikely to be taken
2620 emit_jno_unlikely(0);
2625 host_tempreg_acquire();
2626 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2627 addr=*addr_reg_override=HOST_TEMPREG;
2634 // return memhandler, or get directly accessable address and return 0
2635 static void *get_direct_memhandler(void *table, u_int addr,
2636 enum stub_type type, uintptr_t *addr_host)
2638 uintptr_t l1, l2 = 0;
2639 l1 = ((uintptr_t *)table)[addr>>12];
2640 if ((l1 & (1ul << (sizeof(l1)*8-1))) == 0) {
2641 uintptr_t v = l1 << 1;
2642 *addr_host = v + addr;
2647 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2648 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2649 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2650 l2=((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2652 l2=((uintptr_t *)l1)[(addr&0xfff)/4];
2653 if ((l2 & (1<<31)) == 0) {
2654 uintptr_t v = l2 << 1;
2655 *addr_host = v + (addr&0xfff);
2658 return (void *)(l2 << 1);
2662 static u_int get_host_reglist(const signed char *regmap)
2664 u_int reglist = 0, hr;
2665 for (hr = 0; hr < HOST_REGS; hr++) {
2666 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2672 static u_int reglist_exclude(u_int reglist, int r1, int r2)
2675 reglist &= ~(1u << r1);
2677 reglist &= ~(1u << r2);
2681 // find a temp caller-saved register not in reglist (so assumed to be free)
2682 static int reglist_find_free(u_int reglist)
2684 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2687 return __builtin_ctz(free_regs);
2690 static void load_assemble(int i, const struct regstat *i_regs)
2695 int memtarget=0,c=0;
2696 int fastio_reg_override=-1;
2697 u_int reglist=get_host_reglist(i_regs->regmap);
2698 tl=get_reg(i_regs->regmap,dops[i].rt1);
2699 s=get_reg(i_regs->regmap,dops[i].rs1);
2701 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2703 c=(i_regs->wasconst>>s)&1;
2705 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2708 //printf("load_assemble: c=%d\n",c);
2709 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2710 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2711 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2713 // could be FIFO, must perform the read
2715 assem_debug("(forced read)\n");
2716 tl=get_reg(i_regs->regmap,-1);
2719 if(offset||s<0||c) addr=tl;
2721 //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2723 //printf("load_assemble: c=%d\n",c);
2724 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2725 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2729 // Strmnnrmn's speed hack
2730 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2733 jaddr=emit_fastpath_cmp_jump(i,addr,&fastio_reg_override);
2736 else if(ram_offset&&memtarget) {
2737 host_tempreg_acquire();
2738 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2739 fastio_reg_override=HOST_TEMPREG;
2741 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2742 if (dops[i].opcode==0x20) { // LB
2748 if(fastio_reg_override>=0) a=fastio_reg_override;
2750 emit_movsbl_indexed(x,a,tl);
2754 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2757 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2759 if (dops[i].opcode==0x21) { // LH
2764 if(fastio_reg_override>=0) a=fastio_reg_override;
2765 emit_movswl_indexed(x,a,tl);
2768 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2771 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2773 if (dops[i].opcode==0x23) { // LW
2777 if(fastio_reg_override>=0) a=fastio_reg_override;
2778 emit_readword_indexed(0,a,tl);
2781 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2784 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2786 if (dops[i].opcode==0x24) { // LBU
2791 if(fastio_reg_override>=0) a=fastio_reg_override;
2793 emit_movzbl_indexed(x,a,tl);
2796 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2799 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2801 if (dops[i].opcode==0x25) { // LHU
2806 if(fastio_reg_override>=0) a=fastio_reg_override;
2807 emit_movzwl_indexed(x,a,tl);
2810 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2813 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2815 if (dops[i].opcode==0x27) { // LWU
2818 if (dops[i].opcode==0x37) { // LD
2822 if (fastio_reg_override == HOST_TEMPREG)
2823 host_tempreg_release();
2826 #ifndef loadlr_assemble
2827 static void loadlr_assemble(int i, const struct regstat *i_regs)
2829 int s,tl,temp,temp2,addr;
2832 int memtarget=0,c=0;
2833 int fastio_reg_override=-1;
2834 u_int reglist=get_host_reglist(i_regs->regmap);
2835 tl=get_reg(i_regs->regmap,dops[i].rt1);
2836 s=get_reg(i_regs->regmap,dops[i].rs1);
2837 temp=get_reg(i_regs->regmap,-1);
2838 temp2=get_reg(i_regs->regmap,FTEMP);
2839 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
2843 if(offset||s<0||c) addr=temp2;
2846 c=(i_regs->wasconst>>s)&1;
2848 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2852 emit_shlimm(addr,3,temp);
2853 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2854 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
2856 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
2858 jaddr=emit_fastpath_cmp_jump(i,temp2,&fastio_reg_override);
2861 if(ram_offset&&memtarget) {
2862 host_tempreg_acquire();
2863 emit_addimm(temp2,ram_offset,HOST_TEMPREG);
2864 fastio_reg_override=HOST_TEMPREG;
2866 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2867 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
2869 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
2872 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
2875 if(fastio_reg_override>=0) a=fastio_reg_override;
2876 emit_readword_indexed(0,a,temp2);
2877 if(fastio_reg_override==HOST_TEMPREG) host_tempreg_release();
2878 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj[i],reglist);
2881 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj[i],reglist);
2884 emit_andimm(temp,24,temp);
2885 if (dops[i].opcode==0x22) // LWL
2886 emit_xorimm(temp,24,temp);
2887 host_tempreg_acquire();
2888 emit_movimm(-1,HOST_TEMPREG);
2889 if (dops[i].opcode==0x26) {
2890 emit_shr(temp2,temp,temp2);
2891 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
2893 emit_shl(temp2,temp,temp2);
2894 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
2896 host_tempreg_release();
2897 emit_or(temp2,tl,tl);
2899 //emit_storereg(dops[i].rt1,tl); // DEBUG
2901 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
2907 void store_assemble(int i, const struct regstat *i_regs)
2913 enum stub_type type;
2914 int memtarget=0,c=0;
2915 int agr=AGEN1+(i&1);
2916 int fastio_reg_override=-1;
2917 u_int reglist=get_host_reglist(i_regs->regmap);
2918 tl=get_reg(i_regs->regmap,dops[i].rs2);
2919 s=get_reg(i_regs->regmap,dops[i].rs1);
2920 temp=get_reg(i_regs->regmap,agr);
2921 if(temp<0) temp=get_reg(i_regs->regmap,-1);
2924 c=(i_regs->wasconst>>s)&1;
2926 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2931 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2932 if(offset||s<0||c) addr=temp;
2935 jaddr=emit_fastpath_cmp_jump(i,addr,&fastio_reg_override);
2937 else if(ram_offset&&memtarget) {
2938 host_tempreg_acquire();
2939 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2940 fastio_reg_override=HOST_TEMPREG;
2943 if (dops[i].opcode==0x28) { // SB
2947 if(fastio_reg_override>=0) a=fastio_reg_override;
2948 emit_writebyte_indexed(tl,x,a);
2952 if (dops[i].opcode==0x29) { // SH
2956 if(fastio_reg_override>=0) a=fastio_reg_override;
2957 emit_writehword_indexed(tl,x,a);
2961 if (dops[i].opcode==0x2B) { // SW
2964 if(fastio_reg_override>=0) a=fastio_reg_override;
2965 emit_writeword_indexed(tl,0,a);
2969 if (dops[i].opcode==0x3F) { // SD
2973 if(fastio_reg_override==HOST_TEMPREG)
2974 host_tempreg_release();
2976 // PCSX store handlers don't check invcode again
2978 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2981 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
2983 #ifdef DESTRUCTIVE_SHIFT
2984 // The x86 shift operation is 'destructive'; it overwrites the
2985 // source register, so we need to make a copy first and use that.
2988 #if defined(HOST_IMM8)
2989 int ir=get_reg(i_regs->regmap,INVCP);
2991 emit_cmpmem_indexedsr12_reg(ir,addr,1);
2993 emit_cmpmem_indexedsr12_imm(invalid_code,addr,1);
2995 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
2996 emit_callne(invalidate_addr_reg[addr]);
3000 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3004 u_int addr_val=constmap[i][s]+offset;
3006 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
3007 } else if(c&&!memtarget) {
3008 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj[i],reglist);
3010 // basic current block modification detection..
3011 // not looking back as that should be in mips cache already
3012 // (see Spyro2 title->attract mode)
3013 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3014 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3015 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3016 if(i_regs->regmap==regs[i].regmap) {
3017 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3018 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3019 emit_movimm(start+i*4+4,0);
3020 emit_writeword(0,&pcaddr);
3021 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3022 emit_far_call(get_addr_ht);
3028 static void storelr_assemble(int i, const struct regstat *i_regs)
3034 void *case1, *case2, *case3;
3035 void *done0, *done1, *done2;
3036 int memtarget=0,c=0;
3037 int agr=AGEN1+(i&1);
3038 u_int reglist=get_host_reglist(i_regs->regmap);
3039 tl=get_reg(i_regs->regmap,dops[i].rs2);
3040 s=get_reg(i_regs->regmap,dops[i].rs1);
3041 temp=get_reg(i_regs->regmap,agr);
3042 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3045 c=(i_regs->isconst>>s)&1;
3047 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3053 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3054 if(!offset&&s!=temp) emit_mov(s,temp);
3060 if(!memtarget||!dops[i].rs1) {
3066 emit_addimm_no_flags(ram_offset,temp);
3068 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3072 emit_xorimm(temp,3,temp);
3073 emit_testimm(temp,2);
3076 emit_testimm(temp,1);
3080 if (dops[i].opcode==0x2A) { // SWL
3081 emit_writeword_indexed(tl,0,temp);
3083 else if (dops[i].opcode==0x2E) { // SWR
3084 emit_writebyte_indexed(tl,3,temp);
3091 set_jump_target(case1, out);
3092 if (dops[i].opcode==0x2A) { // SWL
3093 // Write 3 msb into three least significant bytes
3094 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3095 emit_writehword_indexed(tl,-1,temp);
3096 if(dops[i].rs2) emit_rorimm(tl,16,tl);
3097 emit_writebyte_indexed(tl,1,temp);
3098 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3100 else if (dops[i].opcode==0x2E) { // SWR
3101 // Write two lsb into two most significant bytes
3102 emit_writehword_indexed(tl,1,temp);
3107 set_jump_target(case2, out);
3108 emit_testimm(temp,1);
3111 if (dops[i].opcode==0x2A) { // SWL
3112 // Write two msb into two least significant bytes
3113 if(dops[i].rs2) emit_rorimm(tl,16,tl);
3114 emit_writehword_indexed(tl,-2,temp);
3115 if(dops[i].rs2) emit_rorimm(tl,16,tl);
3117 else if (dops[i].opcode==0x2E) { // SWR
3118 // Write 3 lsb into three most significant bytes
3119 emit_writebyte_indexed(tl,-1,temp);
3120 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3121 emit_writehword_indexed(tl,0,temp);
3122 if(dops[i].rs2) emit_rorimm(tl,24,tl);
3127 set_jump_target(case3, out);
3128 if (dops[i].opcode==0x2A) { // SWL
3129 // Write msb into least significant byte
3130 if(dops[i].rs2) emit_rorimm(tl,24,tl);
3131 emit_writebyte_indexed(tl,-3,temp);
3132 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3134 else if (dops[i].opcode==0x2E) { // SWR
3135 // Write entire word
3136 emit_writeword_indexed(tl,-3,temp);
3138 set_jump_target(done0, out);
3139 set_jump_target(done1, out);
3140 set_jump_target(done2, out);
3142 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj[i],reglist);
3143 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3144 emit_addimm_no_flags(-ram_offset,temp);
3145 #if defined(HOST_IMM8)
3146 int ir=get_reg(i_regs->regmap,INVCP);
3148 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3150 emit_cmpmem_indexedsr12_imm(invalid_code,temp,1);
3152 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3153 emit_callne(invalidate_addr_reg[temp]);
3157 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3162 static void cop0_assemble(int i,struct regstat *i_regs)
3164 if(dops[i].opcode2==0) // MFC0
3166 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3167 u_int copr=(source[i]>>11)&0x1f;
3168 //assert(t>=0); // Why does this happen? OOT is weird
3169 if(t>=0&&dops[i].rt1!=0) {
3170 emit_readword(®_cop0[copr],t);
3173 else if(dops[i].opcode2==4) // MTC0
3175 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3176 char copr=(source[i]>>11)&0x1f;
3178 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3179 if(copr==9||copr==11||copr==12||copr==13) {
3180 emit_readword(&last_count,HOST_TEMPREG);
3181 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3182 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3183 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3184 emit_writeword(HOST_CCREG,&Count);
3186 // What a mess. The status register (12) can enable interrupts,
3187 // so needs a special case to handle a pending interrupt.
3188 // The interrupt must be taken immediately, because a subsequent
3189 // instruction might disable interrupts again.
3190 if(copr==12||copr==13) {
3192 // burn cycles to cause cc_interrupt, which will
3193 // reschedule next_interupt. Relies on CCREG from above.
3194 assem_debug("MTC0 DS %d\n", copr);
3195 emit_writeword(HOST_CCREG,&last_count);
3196 emit_movimm(0,HOST_CCREG);
3197 emit_storereg(CCREG,HOST_CCREG);
3198 emit_loadreg(dops[i].rs1,1);
3199 emit_movimm(copr,0);
3200 emit_far_call(pcsx_mtc0_ds);
3201 emit_loadreg(dops[i].rs1,s);
3204 emit_movimm(start+i*4+4,HOST_TEMPREG);
3205 emit_writeword(HOST_TEMPREG,&pcaddr);
3206 emit_movimm(0,HOST_TEMPREG);
3207 emit_writeword(HOST_TEMPREG,&pending_exception);
3210 emit_loadreg(dops[i].rs1,1);
3213 emit_movimm(copr,0);
3214 emit_far_call(pcsx_mtc0);
3215 if(copr==9||copr==11||copr==12||copr==13) {
3216 emit_readword(&Count,HOST_CCREG);
3217 emit_readword(&next_interupt,HOST_TEMPREG);
3218 emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3219 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3220 emit_writeword(HOST_TEMPREG,&last_count);
3221 emit_storereg(CCREG,HOST_CCREG);
3223 if(copr==12||copr==13) {
3224 assert(!is_delayslot);
3225 emit_readword(&pending_exception,14);
3229 emit_readword(&pcaddr, 0);
3230 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3231 emit_far_call(get_addr_ht);
3233 set_jump_target(jaddr, out);
3235 emit_loadreg(dops[i].rs1,s);
3239 assert(dops[i].opcode2==0x10);
3240 //if((source[i]&0x3f)==0x10) // RFE
3242 emit_readword(&Status,0);
3243 emit_andimm(0,0x3c,1);
3244 emit_andimm(0,~0xf,0);
3245 emit_orrshr_imm(1,2,0);
3246 emit_writeword(0,&Status);
3251 static void cop1_unusable(int i,struct regstat *i_regs)
3253 // XXX: should just just do the exception instead
3258 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3262 static void cop1_assemble(int i,struct regstat *i_regs)
3264 cop1_unusable(i, i_regs);
3267 static void c1ls_assemble(int i,struct regstat *i_regs)
3269 cop1_unusable(i, i_regs);
3273 static void do_cop1stub(int n)
3276 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3277 set_jump_target(stubs[n].addr, out);
3279 // int rs=stubs[n].b;
3280 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3283 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3284 //if(i_regs!=®s[i]) printf("oops: regs[i]=%x i_regs=%x",(int)®s[i],(int)i_regs);
3286 //else {printf("fp exception in delay slot\n");}
3287 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3288 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3289 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3290 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3291 emit_far_jump(ds?fp_exception_ds:fp_exception);
3294 static int cop2_is_stalling_op(int i, int *cycles)
3296 if (dops[i].opcode == 0x3a) { // SWC2
3300 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3304 if (dops[i].itype == C2OP) {
3305 *cycles = gte_cycletab[source[i] & 0x3f];
3308 // ... what about MTC2/CTC2/LWC2?
3313 static void log_gte_stall(int stall, u_int cycle)
3315 if ((u_int)stall <= 44)
3316 printf("x stall %2d %u\n", stall, cycle + last_count);
3319 static void emit_log_gte_stall(int i, int stall, u_int reglist)
3323 emit_movimm(stall, 0);
3325 emit_mov(HOST_TEMPREG, 0);
3326 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
3327 emit_far_call(log_gte_stall);
3328 restore_regs(reglist);
3332 static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3334 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3335 int rtmp = reglist_find_free(reglist);
3337 if (HACK_ENABLED(NDHACK_NO_STALLS))
3339 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3340 // happens occasionally... cc evicted? Don't bother then
3341 //printf("no cc %08x\n", start + i*4);
3345 for (j = i - 1; j >= 0; j--) {
3346 //if (dops[j].is_ds) break;
3347 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3352 cycles_passed = CLOCK_ADJUST(ccadj[i] - ccadj[j]);
3353 if (other_gte_op_cycles >= 0)
3354 stall = other_gte_op_cycles - cycles_passed;
3355 else if (cycles_passed >= 44)
3356 stall = 0; // can't stall
3357 if (stall == -MAXBLOCK && rtmp >= 0) {
3358 // unknown stall, do the expensive runtime check
3359 assem_debug("; cop2_do_stall_check\n");
3362 emit_movimm(gte_cycletab[op], 0);
3363 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
3364 emit_far_call(call_gteStall);
3365 restore_regs(reglist);
3367 host_tempreg_acquire();
3368 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3369 emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
3370 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3371 emit_cmpimm(HOST_TEMPREG, 44);
3372 emit_cmovb_reg(rtmp, HOST_CCREG);
3373 //emit_log_gte_stall(i, 0, reglist);
3374 host_tempreg_release();
3377 else if (stall > 0) {
3378 //emit_log_gte_stall(i, stall, reglist);
3379 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3382 // save gteBusyCycle, if needed
3383 if (gte_cycletab[op] == 0)
3385 other_gte_op_cycles = -1;
3386 for (j = i + 1; j < slen; j++) {
3387 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3389 if (dops[j].is_jump) {
3391 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3396 if (other_gte_op_cycles >= 0)
3397 // will handle stall when assembling that op
3399 cycles_passed = CLOCK_ADJUST(ccadj[min(j, slen -1)] - ccadj[i]);
3400 if (cycles_passed >= 44)
3402 assem_debug("; save gteBusyCycle\n");
3403 host_tempreg_acquire();
3405 emit_readword(&last_count, HOST_TEMPREG);
3406 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3407 emit_addimm(HOST_TEMPREG, CLOCK_ADJUST(ccadj[i]), HOST_TEMPREG);
3408 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3409 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3411 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + gte_cycletab[op], HOST_TEMPREG);
3412 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3414 host_tempreg_release();
3417 static int is_mflohi(int i)
3419 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3422 static int check_multdiv(int i, int *cycles)
3424 if (dops[i].itype != MULTDIV)
3426 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3427 *cycles = 11; // approx from 7 11 14
3433 static void multdiv_prepare_stall(int i, const struct regstat *i_regs)
3435 int j, found = 0, c = 0;
3436 if (HACK_ENABLED(NDHACK_NO_STALLS))
3438 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3439 // happens occasionally... cc evicted? Don't bother then
3442 for (j = i + 1; j < slen; j++) {
3445 if ((found = is_mflohi(j)))
3447 if (dops[j].is_jump) {
3449 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3455 // handle all in multdiv_do_stall()
3457 check_multdiv(i, &c);
3459 assem_debug("; muldiv prepare stall %d\n", c);
3460 host_tempreg_acquire();
3461 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + c, HOST_TEMPREG);
3462 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3463 host_tempreg_release();
3466 static void multdiv_do_stall(int i, const struct regstat *i_regs)
3468 int j, known_cycles = 0;
3469 u_int reglist = get_host_reglist(i_regs->regmap);
3470 int rtmp = get_reg(i_regs->regmap, -1);
3472 rtmp = reglist_find_free(reglist);
3473 if (HACK_ENABLED(NDHACK_NO_STALLS))
3475 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3476 // happens occasionally... cc evicted? Don't bother then
3477 //printf("no cc/rtmp %08x\n", start + i*4);
3481 for (j = i - 1; j >= 0; j--) {
3482 if (dops[j].is_ds) break;
3483 if (check_multdiv(j, &known_cycles) || dops[j].bt)
3486 // already handled by this op
3491 if (known_cycles > 0) {
3492 known_cycles -= CLOCK_ADJUST(ccadj[i] - ccadj[j]);
3493 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3494 if (known_cycles > 0)
3495 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3498 assem_debug("; muldiv stall unresolved\n");
3499 host_tempreg_acquire();
3500 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3501 emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
3502 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3503 emit_cmpimm(HOST_TEMPREG, 37);
3504 emit_cmovb_reg(rtmp, HOST_CCREG);
3505 //emit_log_gte_stall(i, 0, reglist);
3506 host_tempreg_release();
3509 static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3519 emit_readword(®_cop2d[copr],tl);
3520 emit_signextend16(tl,tl);
3521 emit_writeword(tl,®_cop2d[copr]); // hmh
3528 emit_readword(®_cop2d[copr],tl);
3529 emit_andimm(tl,0xffff,tl);
3530 emit_writeword(tl,®_cop2d[copr]);
3533 emit_readword(®_cop2d[14],tl); // SXY2
3534 emit_writeword(tl,®_cop2d[copr]);
3538 c2op_mfc2_29_assemble(tl,temp);
3541 emit_readword(®_cop2d[copr],tl);
3546 static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3550 emit_readword(®_cop2d[13],temp); // SXY1
3551 emit_writeword(sl,®_cop2d[copr]);
3552 emit_writeword(temp,®_cop2d[12]); // SXY0
3553 emit_readword(®_cop2d[14],temp); // SXY2
3554 emit_writeword(sl,®_cop2d[14]);
3555 emit_writeword(temp,®_cop2d[13]); // SXY1
3558 emit_andimm(sl,0x001f,temp);
3559 emit_shlimm(temp,7,temp);
3560 emit_writeword(temp,®_cop2d[9]);
3561 emit_andimm(sl,0x03e0,temp);
3562 emit_shlimm(temp,2,temp);
3563 emit_writeword(temp,®_cop2d[10]);
3564 emit_andimm(sl,0x7c00,temp);
3565 emit_shrimm(temp,3,temp);
3566 emit_writeword(temp,®_cop2d[11]);
3567 emit_writeword(sl,®_cop2d[28]);
3570 emit_xorsar_imm(sl,sl,31,temp);
3571 #if defined(HAVE_ARMV5) || defined(__aarch64__)
3572 emit_clz(temp,temp);
3574 emit_movs(temp,HOST_TEMPREG);
3575 emit_movimm(0,temp);
3576 emit_jeq((int)out+4*4);
3577 emit_addpl_imm(temp,1,temp);
3578 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3579 emit_jns((int)out-2*4);
3581 emit_writeword(sl,®_cop2d[30]);
3582 emit_writeword(temp,®_cop2d[31]);
3587 emit_writeword(sl,®_cop2d[copr]);
3592 static void c2ls_assemble(int i, const struct regstat *i_regs)
3597 int memtarget=0,c=0;
3599 enum stub_type type;
3600 int agr=AGEN1+(i&1);
3601 int fastio_reg_override=-1;
3602 u_int reglist=get_host_reglist(i_regs->regmap);
3603 u_int copr=(source[i]>>16)&0x1f;
3604 s=get_reg(i_regs->regmap,dops[i].rs1);
3605 tl=get_reg(i_regs->regmap,FTEMP);
3607 assert(dops[i].rs1>0);
3610 if(i_regs->regmap[HOST_CCREG]==CCREG)
3611 reglist&=~(1<<HOST_CCREG);
3614 if (dops[i].opcode==0x3a) { // SWC2
3615 ar=get_reg(i_regs->regmap,agr);
3616 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3621 if(s>=0) c=(i_regs->wasconst>>s)&1;
3622 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3623 if (!offset&&!c&&s>=0) ar=s;
3626 cop2_do_stall_check(0, i, i_regs, reglist);
3628 if (dops[i].opcode==0x3a) { // SWC2
3629 cop2_get_dreg(copr,tl,-1);
3637 emit_jmp(0); // inline_readstub/inline_writestub?
3641 jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
3643 else if(ram_offset&&memtarget) {
3644 host_tempreg_acquire();
3645 emit_addimm(ar,ram_offset,HOST_TEMPREG);
3646 fastio_reg_override=HOST_TEMPREG;
3648 if (dops[i].opcode==0x32) { // LWC2
3650 if(fastio_reg_override>=0) a=fastio_reg_override;
3651 emit_readword_indexed(0,a,tl);
3653 if (dops[i].opcode==0x3a) { // SWC2
3654 #ifdef DESTRUCTIVE_SHIFT
3655 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3658 if(fastio_reg_override>=0) a=fastio_reg_override;
3659 emit_writeword_indexed(tl,0,a);
3662 if(fastio_reg_override==HOST_TEMPREG)
3663 host_tempreg_release();
3665 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj[i],reglist);
3666 if(dops[i].opcode==0x3a) // SWC2
3667 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3668 #if defined(HOST_IMM8)
3669 int ir=get_reg(i_regs->regmap,INVCP);
3671 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3673 emit_cmpmem_indexedsr12_imm(invalid_code,ar,1);
3675 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3676 emit_callne(invalidate_addr_reg[ar]);
3680 add_stub(INVCODE_STUB,jaddr3,out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3683 if (dops[i].opcode==0x32) { // LWC2
3684 host_tempreg_acquire();
3685 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3686 host_tempreg_release();
3690 static void cop2_assemble(int i, const struct regstat *i_regs)
3692 u_int copr = (source[i]>>11) & 0x1f;
3693 signed char temp = get_reg(i_regs->regmap, -1);
3695 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3696 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3697 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3698 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3699 reglist = reglist_exclude(reglist, tl, -1);
3701 cop2_do_stall_check(0, i, i_regs, reglist);
3703 if (dops[i].opcode2==0) { // MFC2
3704 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3705 if(tl>=0&&dops[i].rt1!=0)
3706 cop2_get_dreg(copr,tl,temp);
3708 else if (dops[i].opcode2==4) { // MTC2
3709 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3710 cop2_put_dreg(copr,sl,temp);
3712 else if (dops[i].opcode2==2) // CFC2
3714 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3715 if(tl>=0&&dops[i].rt1!=0)
3716 emit_readword(®_cop2c[copr],tl);
3718 else if (dops[i].opcode2==6) // CTC2
3720 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3729 emit_signextend16(sl,temp);
3732 c2op_ctc2_31_assemble(sl,temp);
3738 emit_writeword(temp,®_cop2c[copr]);
3743 static void do_unalignedwritestub(int n)
3745 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
3747 set_jump_target(stubs[n].addr, out);
3750 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3751 int addr=stubs[n].b;
3752 u_int reglist=stubs[n].e;
3753 signed char *i_regmap=i_regs->regmap;
3754 int temp2=get_reg(i_regmap,FTEMP);
3756 rt=get_reg(i_regmap,dops[i].rs2);
3759 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
3761 reglist&=~(1<<temp2);
3763 // don't bother with it and call write handler
3766 int cc=get_reg(i_regmap,CCREG);
3768 emit_loadreg(CCREG,2);
3769 emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n].d+1),2);
3770 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
3771 emit_addimm(0,-CLOCK_ADJUST((int)stubs[n].d+1),cc<0?2:cc);
3773 emit_storereg(CCREG,2);
3774 restore_regs(reglist);
3775 emit_jmp(stubs[n].retaddr); // return address
3778 #ifndef multdiv_assemble
3779 void multdiv_assemble(int i,struct regstat *i_regs)
3781 printf("Need multdiv_assemble for this architecture.\n");
3786 static void mov_assemble(int i,struct regstat *i_regs)
3788 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
3789 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
3792 tl=get_reg(i_regs->regmap,dops[i].rt1);
3795 sl=get_reg(i_regs->regmap,dops[i].rs1);
3796 if(sl>=0) emit_mov(sl,tl);
3797 else emit_loadreg(dops[i].rs1,tl);
3800 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
3801 multdiv_do_stall(i, i_regs);
3804 // call interpreter, exception handler, things that change pc/regs/cycles ...
3805 static void call_c_cpu_handler(int i, const struct regstat *i_regs, u_int pc, void *func)
3807 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3808 assert(ccreg==HOST_CCREG);
3809 assert(!is_delayslot);
3812 emit_movimm(pc,3); // Get PC
3813 emit_readword(&last_count,2);
3814 emit_writeword(3,&psxRegs.pc);
3815 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
3816 emit_add(2,HOST_CCREG,2);
3817 emit_writeword(2,&psxRegs.cycle);
3818 emit_far_call(func);
3819 emit_far_jump(jump_to_new_pc);
3822 static void syscall_assemble(int i,struct regstat *i_regs)
3824 emit_movimm(0x20,0); // cause code
3825 emit_movimm(0,1); // not in delay slot
3826 call_c_cpu_handler(i,i_regs,start+i*4,psxException);
3829 static void hlecall_assemble(int i,struct regstat *i_regs)
3831 void *hlefunc = psxNULL;
3832 uint32_t hleCode = source[i] & 0x03ffffff;
3833 if (hleCode < ARRAY_SIZE(psxHLEt))
3834 hlefunc = psxHLEt[hleCode];
3836 call_c_cpu_handler(i,i_regs,start+i*4+4,hlefunc);
3839 static void intcall_assemble(int i,struct regstat *i_regs)
3841 call_c_cpu_handler(i,i_regs,start+i*4,execI);
3844 static void speculate_mov(int rs,int rt)
3847 smrv_strong_next|=1<<rt;
3852 static void speculate_mov_weak(int rs,int rt)
3855 smrv_weak_next|=1<<rt;
3860 static void speculate_register_values(int i)
3863 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
3864 // gp,sp are likely to stay the same throughout the block
3865 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
3866 smrv_weak_next=~smrv_strong_next;
3867 //printf(" llr %08x\n", smrv[4]);
3869 smrv_strong=smrv_strong_next;
3870 smrv_weak=smrv_weak_next;
3871 switch(dops[i].itype) {
3873 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
3874 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
3875 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
3876 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
3878 smrv_strong_next&=~(1<<dops[i].rt1);
3879 smrv_weak_next&=~(1<<dops[i].rt1);
3883 smrv_strong_next&=~(1<<dops[i].rt1);
3884 smrv_weak_next&=~(1<<dops[i].rt1);
3887 if(dops[i].rt1&&is_const(®s[i],dops[i].rt1)) {
3888 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
3890 if(get_final_value(hr,i,&value))
3891 smrv[dops[i].rt1]=value;
3892 else smrv[dops[i].rt1]=constmap[i][hr];
3893 smrv_strong_next|=1<<dops[i].rt1;
3897 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
3898 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
3902 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
3903 // special case for BIOS
3904 smrv[dops[i].rt1]=0xa0000000;
3905 smrv_strong_next|=1<<dops[i].rt1;
3912 smrv_strong_next&=~(1<<dops[i].rt1);
3913 smrv_weak_next&=~(1<<dops[i].rt1);
3917 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
3918 smrv_strong_next&=~(1<<dops[i].rt1);
3919 smrv_weak_next&=~(1<<dops[i].rt1);
3923 if (dops[i].opcode==0x32) { // LWC2
3924 smrv_strong_next&=~(1<<dops[i].rt1);
3925 smrv_weak_next&=~(1<<dops[i].rt1);
3931 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
3932 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
3936 static void ds_assemble(int i,struct regstat *i_regs)
3938 speculate_register_values(i);
3940 switch(dops[i].itype) {
3942 alu_assemble(i,i_regs);break;
3944 imm16_assemble(i,i_regs);break;
3946 shift_assemble(i,i_regs);break;
3948 shiftimm_assemble(i,i_regs);break;
3950 load_assemble(i,i_regs);break;
3952 loadlr_assemble(i,i_regs);break;
3954 store_assemble(i,i_regs);break;
3956 storelr_assemble(i,i_regs);break;
3958 cop0_assemble(i,i_regs);break;
3960 cop1_assemble(i,i_regs);break;
3962 c1ls_assemble(i,i_regs);break;
3964 cop2_assemble(i,i_regs);break;
3966 c2ls_assemble(i,i_regs);break;
3968 c2op_assemble(i,i_regs);break;
3970 multdiv_assemble(i,i_regs);
3971 multdiv_prepare_stall(i,i_regs);
3974 mov_assemble(i,i_regs);break;
3983 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
3988 // Is the branch target a valid internal jump?
3989 static int internal_branch(int addr)
3991 if(addr&1) return 0; // Indirect (register) jump
3992 if(addr>=start && addr<start+slen*4-4)
3999 static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4002 for(hr=0;hr<HOST_REGS;hr++) {
4003 if(hr!=EXCLUDE_REG) {
4004 if(pre[hr]!=entry[hr]) {
4007 if(get_reg(entry,pre[hr])<0) {
4009 if(!((u>>pre[hr])&1))
4010 emit_storereg(pre[hr],hr);
4017 // Move from one register to another (no writeback)
4018 for(hr=0;hr<HOST_REGS;hr++) {
4019 if(hr!=EXCLUDE_REG) {
4020 if(pre[hr]!=entry[hr]) {
4021 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
4023 if((nr=get_reg(entry,pre[hr]))>=0) {
4032 // Load the specified registers
4033 // This only loads the registers given as arguments because
4034 // we don't want to load things that will be overwritten
4035 static void load_regs(signed char entry[],signed char regmap[],int rs1,int rs2)
4039 for(hr=0;hr<HOST_REGS;hr++) {
4040 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4041 if(entry[hr]!=regmap[hr]) {
4042 if(regmap[hr]==rs1||regmap[hr]==rs2)
4049 emit_loadreg(regmap[hr],hr);
4057 // Load registers prior to the start of a loop
4058 // so that they are not loaded within the loop
4059 static void loop_preload(signed char pre[],signed char entry[])
4062 for(hr=0;hr<HOST_REGS;hr++) {
4063 if(hr!=EXCLUDE_REG) {
4064 if(pre[hr]!=entry[hr]) {
4066 if(get_reg(pre,entry[hr])<0) {
4067 assem_debug("loop preload:\n");
4068 //printf("loop preload: %d\n",hr);
4072 else if(entry[hr]<TEMPREG)
4074 emit_loadreg(entry[hr],hr);
4076 else if(entry[hr]-64<TEMPREG)
4078 emit_loadreg(entry[hr],hr);
4087 // Generate address for load/store instruction
4088 // goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4089 void address_generation(int i,struct regstat *i_regs,signed char entry[])
4091 if(dops[i].itype==LOAD||dops[i].itype==LOADLR||dops[i].itype==STORE||dops[i].itype==STORELR||dops[i].itype==C1LS||dops[i].itype==C2LS) {
4093 int agr=AGEN1+(i&1);
4094 if(dops[i].itype==LOAD) {
4095 ra=get_reg(i_regs->regmap,dops[i].rt1);
4096 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4099 if(dops[i].itype==LOADLR) {
4100 ra=get_reg(i_regs->regmap,FTEMP);
4102 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4103 ra=get_reg(i_regs->regmap,agr);
4104 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4106 if(dops[i].itype==C1LS||dops[i].itype==C2LS) {
4107 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4108 ra=get_reg(i_regs->regmap,FTEMP);
4109 else { // SWC1/SDC1/SWC2/SDC2
4110 ra=get_reg(i_regs->regmap,agr);
4111 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4114 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4117 int c=(i_regs->wasconst>>rs)&1;
4118 if(dops[i].rs1==0) {
4119 // Using r0 as a base address
4120 if(!entry||entry[ra]!=agr) {
4121 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4122 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4123 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4124 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4126 emit_movimm(offset,ra);
4128 } // else did it in the previous cycle
4131 if(!entry||entry[ra]!=dops[i].rs1)
4132 emit_loadreg(dops[i].rs1,ra);
4133 //if(!entry||entry[ra]!=dops[i].rs1)
4134 // printf("poor load scheduling!\n");
4137 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4138 if(!entry||entry[ra]!=agr) {
4139 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4140 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4141 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4142 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4144 emit_movimm(constmap[i][rs]+offset,ra);
4145 regs[i].loadedconst|=1<<ra;
4147 } // else did it in the previous cycle
4148 } // else load_consts already did it
4150 if(offset&&!c&&dops[i].rs1) {
4152 emit_addimm(rs,offset,ra);
4154 emit_addimm(ra,offset,ra);
4159 // Preload constants for next instruction
4160 if(dops[i+1].itype==LOAD||dops[i+1].itype==LOADLR||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS||dops[i+1].itype==C2LS) {
4163 agr=AGEN1+((i+1)&1);
4164 ra=get_reg(i_regs->regmap,agr);
4166 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4167 int offset=imm[i+1];
4168 int c=(regs[i+1].wasconst>>rs)&1;
4169 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4170 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4171 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4172 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4173 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4175 emit_movimm(constmap[i+1][rs]+offset,ra);
4176 regs[i+1].loadedconst|=1<<ra;
4179 else if(dops[i+1].rs1==0) {
4180 // Using r0 as a base address
4181 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4182 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4183 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4184 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4186 emit_movimm(offset,ra);
4193 static int get_final_value(int hr, int i, int *value)
4195 int reg=regs[i].regmap[hr];
4197 if(regs[i+1].regmap[hr]!=reg) break;
4198 if(!((regs[i+1].isconst>>hr)&1)) break;
4199 if(dops[i+1].bt) break;
4203 if (dops[i].is_jump) {
4204 *value=constmap[i][hr];
4208 if (dops[i+1].is_jump) {
4209 // Load in delay slot, out-of-order execution
4210 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4212 // Precompute load address
4213 *value=constmap[i][hr]+imm[i+2];
4217 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4219 // Precompute load address
4220 *value=constmap[i][hr]+imm[i+1];
4221 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4226 *value=constmap[i][hr];
4227 //printf("c=%lx\n",(long)constmap[i][hr]);
4228 if(i==slen-1) return 1;
4230 return !((unneeded_reg[i+1]>>reg)&1);
4233 // Load registers with known constants
4234 static void load_consts(signed char pre[],signed char regmap[],int i)
4237 // propagate loaded constant flags
4238 if(i==0||dops[i].bt)
4239 regs[i].loadedconst=0;
4241 for(hr=0;hr<HOST_REGS;hr++) {
4242 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4243 &®map[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4245 regs[i].loadedconst|=1<<hr;
4250 for(hr=0;hr<HOST_REGS;hr++) {
4251 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4252 //if(entry[hr]!=regmap[hr]) {
4253 if(!((regs[i].loadedconst>>hr)&1)) {
4254 assert(regmap[hr]<64);
4255 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
4256 int value,similar=0;
4257 if(get_final_value(hr,i,&value)) {
4258 // see if some other register has similar value
4259 for(hr2=0;hr2<HOST_REGS;hr2++) {
4260 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4261 if(is_similar_value(value,constmap[i][hr2])) {
4269 if(get_final_value(hr2,i,&value2)) // is this needed?
4270 emit_movimm_from(value2,hr2,value,hr);
4272 emit_movimm(value,hr);
4278 emit_movimm(value,hr);
4281 regs[i].loadedconst|=1<<hr;
4288 void load_all_consts(signed char regmap[], u_int dirty, int i)
4292 for(hr=0;hr<HOST_REGS;hr++) {
4293 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
4294 assert(regmap[hr] < 64);
4295 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
4296 int value=constmap[i][hr];
4301 emit_movimm(value,hr);
4308 // Write out all dirty registers (except cycle count)
4309 static void wb_dirtys(signed char i_regmap[],uint64_t i_dirty)
4312 for(hr=0;hr<HOST_REGS;hr++) {
4313 if(hr!=EXCLUDE_REG) {
4314 if(i_regmap[hr]>0) {
4315 if(i_regmap[hr]!=CCREG) {
4316 if((i_dirty>>hr)&1) {
4317 assert(i_regmap[hr]<64);
4318 emit_storereg(i_regmap[hr],hr);
4326 // Write out dirty registers that we need to reload (pair with load_needed_regs)
4327 // This writes the registers not written by store_regs_bt
4328 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_dirty,int addr)
4331 int t=(addr-start)>>2;
4332 for(hr=0;hr<HOST_REGS;hr++) {
4333 if(hr!=EXCLUDE_REG) {
4334 if(i_regmap[hr]>0) {
4335 if(i_regmap[hr]!=CCREG) {
4336 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4337 if((i_dirty>>hr)&1) {
4338 assert(i_regmap[hr]<64);
4339 emit_storereg(i_regmap[hr],hr);
4348 // Load all registers (except cycle count)
4349 void load_all_regs(signed char i_regmap[])
4352 for(hr=0;hr<HOST_REGS;hr++) {
4353 if(hr!=EXCLUDE_REG) {
4354 if(i_regmap[hr]==0) {
4358 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4360 emit_loadreg(i_regmap[hr],hr);
4366 // Load all current registers also needed by next instruction
4367 void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
4370 for(hr=0;hr<HOST_REGS;hr++) {
4371 if(hr!=EXCLUDE_REG) {
4372 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4373 if(i_regmap[hr]==0) {
4377 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4379 emit_loadreg(i_regmap[hr],hr);
4386 // Load all regs, storing cycle count if necessary
4387 void load_regs_entry(int t)
4390 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4391 else if(ccadj[t]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[t]),HOST_CCREG);
4392 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4393 emit_storereg(CCREG,HOST_CCREG);
4396 for(hr=0;hr<HOST_REGS;hr++) {
4397 if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4398 if(regs[t].regmap_entry[hr]==0) {
4401 else if(regs[t].regmap_entry[hr]!=CCREG)
4403 emit_loadreg(regs[t].regmap_entry[hr],hr);
4409 // Store dirty registers prior to branch
4410 void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4412 if(internal_branch(addr))
4414 int t=(addr-start)>>2;
4416 for(hr=0;hr<HOST_REGS;hr++) {
4417 if(hr!=EXCLUDE_REG) {
4418 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4419 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4420 if((i_dirty>>hr)&1) {
4421 assert(i_regmap[hr]<64);
4422 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4423 emit_storereg(i_regmap[hr],hr);
4432 // Branch out of this block, write out all dirty regs
4433 wb_dirtys(i_regmap,i_dirty);
4437 // Load all needed registers for branch target
4438 static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4440 //if(addr>=start && addr<(start+slen*4))
4441 if(internal_branch(addr))
4443 int t=(addr-start)>>2;
4445 // Store the cycle count before loading something else
4446 if(i_regmap[HOST_CCREG]!=CCREG) {
4447 assert(i_regmap[HOST_CCREG]==-1);
4449 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4450 emit_storereg(CCREG,HOST_CCREG);
4453 for(hr=0;hr<HOST_REGS;hr++) {
4454 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4455 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4456 if(regs[t].regmap_entry[hr]==0) {
4459 else if(regs[t].regmap_entry[hr]!=CCREG)
4461 emit_loadreg(regs[t].regmap_entry[hr],hr);
4469 static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4471 if(addr>=start && addr<start+slen*4-4)
4473 int t=(addr-start)>>2;
4475 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4476 for(hr=0;hr<HOST_REGS;hr++)
4480 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4482 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4489 if(i_regmap[hr]<TEMPREG)
4491 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4494 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4500 else // Same register but is it 32-bit or dirty?
4503 if(!((regs[t].dirty>>hr)&1))
4507 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4509 //printf("%x: dirty no match\n",addr);
4517 // Delay slots are not valid branch targets
4518 //if(t>0&&(dops[t-1].is_jump) return 0;
4519 // Delay slots require additional processing, so do not match
4520 if(dops[t].is_ds) return 0;
4525 for(hr=0;hr<HOST_REGS;hr++)
4531 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4546 static void drc_dbg_emit_do_cmp(int i)
4548 extern void do_insn_cmp();
4550 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4552 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4554 // write out changed consts to match the interpreter
4555 if (i > 0 && !dops[i].bt) {
4556 for (hr = 0; hr < HOST_REGS; hr++) {
4557 int reg = regs[i-1].regmap[hr];
4558 if (hr == EXCLUDE_REG || reg < 0)
4560 if (!((regs[i-1].isconst >> hr) & 1))
4562 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4564 emit_movimm(constmap[i-1][hr],0);
4565 emit_storereg(reg, 0);
4568 emit_movimm(start+i*4,0);
4569 emit_writeword(0,&pcaddr);
4570 emit_far_call(do_insn_cmp);
4571 //emit_readword(&cycle,0);
4572 //emit_addimm(0,2,0);
4573 //emit_writeword(0,&cycle);
4575 restore_regs(reglist);
4576 assem_debug("\\\\do_insn_cmp\n");
4579 #define drc_dbg_emit_do_cmp(x)
4582 // Used when a branch jumps into the delay slot of another branch
4583 static void ds_assemble_entry(int i)
4585 int t=(ba[i]-start)>>2;
4587 instr_addr[t] = out;
4588 assem_debug("Assemble delay slot at %x\n",ba[i]);
4589 assem_debug("<->\n");
4590 drc_dbg_emit_do_cmp(t);
4591 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&®s[t].regmap[HOST_CCREG]!=CCREG)
4592 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4593 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4594 address_generation(t,®s[t],regs[t].regmap_entry);
4595 if(dops[t].itype==STORE||dops[t].itype==STORELR||(dops[t].opcode&0x3b)==0x39||(dops[t].opcode&0x3b)==0x3a)
4596 load_regs(regs[t].regmap_entry,regs[t].regmap,INVCP,INVCP);
4598 switch(dops[t].itype) {
4600 alu_assemble(t,®s[t]);break;
4602 imm16_assemble(t,®s[t]);break;
4604 shift_assemble(t,®s[t]);break;
4606 shiftimm_assemble(t,®s[t]);break;
4608 load_assemble(t,®s[t]);break;
4610 loadlr_assemble(t,®s[t]);break;
4612 store_assemble(t,®s[t]);break;
4614 storelr_assemble(t,®s[t]);break;
4616 cop0_assemble(t,®s[t]);break;
4618 cop1_assemble(t,®s[t]);break;
4620 c1ls_assemble(t,®s[t]);break;
4622 cop2_assemble(t,®s[t]);break;
4624 c2ls_assemble(t,®s[t]);break;
4626 c2op_assemble(t,®s[t]);break;
4628 multdiv_assemble(t,®s[t]);
4629 multdiv_prepare_stall(i,®s[t]);
4632 mov_assemble(t,®s[t]);break;
4641 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4643 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4644 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4645 if(internal_branch(ba[i]+4))
4646 assem_debug("branch: internal\n");
4648 assem_debug("branch: external\n");
4649 assert(internal_branch(ba[i]+4));
4650 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4654 static void emit_extjump(void *addr, u_int target)
4656 emit_extjump2(addr, target, dyna_linker);
4659 static void emit_extjump_ds(void *addr, u_int target)
4661 emit_extjump2(addr, target, dyna_linker_ds);
4664 // Load 2 immediates optimizing for small code size
4665 static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4667 emit_movimm(imm1,rt1);
4668 emit_movimm_from(imm1,rt1,imm2,rt2);
4671 void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
4677 if(dops[i].itype==RJUMP)
4681 //if(ba[i]>=start && ba[i]<(start+slen*4))
4682 if(internal_branch(ba[i]))
4685 if(dops[t].is_ds) *adj=-1; // Branch into delay slot adds an extra cycle
4693 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4695 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4697 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4698 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4702 else if(*adj==0||invert) {
4703 int cycles=CLOCK_ADJUST(count+2);
4708 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4709 cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
4712 emit_addimm_and_set_flags(cycles,HOST_CCREG);
4718 emit_cmpimm(HOST_CCREG,-CLOCK_ADJUST(count+2));
4722 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:(count+2),i,addr,taken,0);
4725 static void do_ccstub(int n)
4728 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
4729 set_jump_target(stubs[n].addr, out);
4731 if(stubs[n].d==NULLDS) {
4732 // Delay slot instruction is nullified ("likely" branch)
4733 wb_dirtys(regs[i].regmap,regs[i].dirty);
4735 else if(stubs[n].d!=TAKEN) {
4736 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
4739 if(internal_branch(ba[i]))
4740 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4744 // Save PC as return address
4745 emit_movimm(stubs[n].c,EAX);
4746 emit_writeword(EAX,&pcaddr);
4750 // Return address depends on which way the branch goes
4751 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
4753 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
4754 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
4760 else if(dops[i].rs2==0)
4765 #ifdef DESTRUCTIVE_WRITEBACK
4767 if((branch_regs[i].dirty>>s1l)&&1)
4768 emit_loadreg(dops[i].rs1,s1l);
4771 if((branch_regs[i].dirty>>s1l)&1)
4772 emit_loadreg(dops[i].rs2,s1l);
4775 if((branch_regs[i].dirty>>s2l)&1)
4776 emit_loadreg(dops[i].rs2,s2l);
4779 int addr=-1,alt=-1,ntaddr=-1;
4782 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4783 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4784 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4792 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4793 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4794 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4800 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
4804 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4805 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4806 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4812 assert(hr<HOST_REGS);
4814 if((dops[i].opcode&0x2f)==4) // BEQ
4816 #ifdef HAVE_CMOV_IMM
4817 if(s2l>=0) emit_cmp(s1l,s2l);
4818 else emit_test(s1l,s1l);
4819 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4821 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4822 if(s2l>=0) emit_cmp(s1l,s2l);
4823 else emit_test(s1l,s1l);
4824 emit_cmovne_reg(alt,addr);
4827 if((dops[i].opcode&0x2f)==5) // BNE
4829 #ifdef HAVE_CMOV_IMM
4830 if(s2l>=0) emit_cmp(s1l,s2l);
4831 else emit_test(s1l,s1l);
4832 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4834 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4835 if(s2l>=0) emit_cmp(s1l,s2l);
4836 else emit_test(s1l,s1l);
4837 emit_cmovne_reg(alt,addr);
4840 if((dops[i].opcode&0x2f)==6) // BLEZ
4842 //emit_movimm(ba[i],alt);
4843 //emit_movimm(start+i*4+8,addr);
4844 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4846 emit_cmovl_reg(alt,addr);
4848 if((dops[i].opcode&0x2f)==7) // BGTZ
4850 //emit_movimm(ba[i],addr);
4851 //emit_movimm(start+i*4+8,ntaddr);
4852 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
4854 emit_cmovl_reg(ntaddr,addr);
4856 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
4858 //emit_movimm(ba[i],alt);
4859 //emit_movimm(start+i*4+8,addr);
4860 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4862 emit_cmovs_reg(alt,addr);
4864 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
4866 //emit_movimm(ba[i],addr);
4867 //emit_movimm(start+i*4+8,alt);
4868 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4870 emit_cmovs_reg(alt,addr);
4872 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
4873 if(source[i]&0x10000) // BC1T
4875 //emit_movimm(ba[i],alt);
4876 //emit_movimm(start+i*4+8,addr);
4877 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4878 emit_testimm(s1l,0x800000);
4879 emit_cmovne_reg(alt,addr);
4883 //emit_movimm(ba[i],addr);
4884 //emit_movimm(start+i*4+8,alt);
4885 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4886 emit_testimm(s1l,0x800000);
4887 emit_cmovne_reg(alt,addr);
4890 emit_writeword(addr,&pcaddr);
4893 if(dops[i].itype==RJUMP)
4895 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
4896 if (ds_writes_rjump_rs(i)) {
4897 r=get_reg(branch_regs[i].regmap,RTEMP);
4899 emit_writeword(r,&pcaddr);
4901 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
4903 // Update cycle count
4904 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
4905 if(stubs[n].a) emit_addimm(HOST_CCREG,CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
4906 emit_far_call(cc_interrupt);
4907 if(stubs[n].a) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
4908 if(stubs[n].d==TAKEN) {
4909 if(internal_branch(ba[i]))
4910 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
4911 else if(dops[i].itype==RJUMP) {
4912 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
4913 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
4915 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
4917 }else if(stubs[n].d==NOTTAKEN) {
4918 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
4919 else load_all_regs(branch_regs[i].regmap);
4920 }else if(stubs[n].d==NULLDS) {
4921 // Delay slot instruction is nullified ("likely" branch)
4922 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
4923 else load_all_regs(regs[i].regmap);
4925 load_all_regs(branch_regs[i].regmap);
4927 if (stubs[n].retaddr)
4928 emit_jmp(stubs[n].retaddr);
4930 do_jump_vaddr(stubs[n].e);
4933 static void add_to_linker(void *addr, u_int target, int ext)
4935 assert(linkcount < ARRAY_SIZE(link_addr));
4936 link_addr[linkcount].addr = addr;
4937 link_addr[linkcount].target = target;
4938 link_addr[linkcount].ext = ext;
4942 static void ujump_assemble_write_ra(int i)
4945 unsigned int return_address;
4946 rt=get_reg(branch_regs[i].regmap,31);
4947 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]);
4949 return_address=start+i*4+8;
4952 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
4953 int temp=-1; // note: must be ds-safe
4957 if(temp>=0) do_miniht_insert(return_address,rt,temp);
4958 else emit_movimm(return_address,rt);
4966 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4969 emit_movimm(return_address,rt); // PC into link register
4971 emit_prefetch(hash_table_get(return_address));
4977 static void ujump_assemble(int i,struct regstat *i_regs)
4980 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4981 address_generation(i+1,i_regs,regs[i].regmap_entry);
4983 int temp=get_reg(branch_regs[i].regmap,PTEMP);
4984 if(dops[i].rt1==31&&temp>=0)
4986 signed char *i_regmap=i_regs->regmap;
4987 int return_address=start+i*4+8;
4988 if(get_reg(branch_regs[i].regmap,31)>0)
4989 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4992 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
4993 ujump_assemble_write_ra(i); // writeback ra for DS
4996 ds_assemble(i+1,i_regs);
4997 uint64_t bc_unneeded=branch_regs[i].u;
4998 bc_unneeded|=1|(1LL<<dops[i].rt1);
4999 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5000 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5001 if(!ra_done&&dops[i].rt1==31)
5002 ujump_assemble_write_ra(i);
5004 cc=get_reg(branch_regs[i].regmap,CCREG);
5005 assert(cc==HOST_CCREG);
5006 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5008 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5010 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5011 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5012 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5013 if(internal_branch(ba[i]))
5014 assem_debug("branch: internal\n");
5016 assem_debug("branch: external\n");
5017 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5018 ds_assemble_entry(i);
5021 add_to_linker(out,ba[i],internal_branch(ba[i]));
5026 static void rjump_assemble_write_ra(int i)
5028 int rt,return_address;
5029 assert(dops[i+1].rt1!=dops[i].rt1);
5030 assert(dops[i+1].rt2!=dops[i].rt1);
5031 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5032 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]);
5034 return_address=start+i*4+8;
5038 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5041 emit_movimm(return_address,rt); // PC into link register
5043 emit_prefetch(hash_table_get(return_address));
5047 static void rjump_assemble(int i,struct regstat *i_regs)
5052 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5054 if (ds_writes_rjump_rs(i)) {
5055 // Delay slot abuse, make a copy of the branch address register
5056 temp=get_reg(branch_regs[i].regmap,RTEMP);
5058 assert(regs[i].regmap[temp]==RTEMP);
5062 address_generation(i+1,i_regs,regs[i].regmap_entry);
5066 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5067 signed char *i_regmap=i_regs->regmap;
5068 int return_address=start+i*4+8;
5069 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5074 if(dops[i].rs1==31) {
5075 int rh=get_reg(regs[i].regmap,RHASH);
5076 if(rh>=0) do_preload_rhash(rh);
5079 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5080 rjump_assemble_write_ra(i);
5083 ds_assemble(i+1,i_regs);
5084 uint64_t bc_unneeded=branch_regs[i].u;
5085 bc_unneeded|=1|(1LL<<dops[i].rt1);
5086 bc_unneeded&=~(1LL<<dops[i].rs1);
5087 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5088 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5089 if(!ra_done&&dops[i].rt1!=0)
5090 rjump_assemble_write_ra(i);
5091 cc=get_reg(branch_regs[i].regmap,CCREG);
5092 assert(cc==HOST_CCREG);
5095 int rh=get_reg(branch_regs[i].regmap,RHASH);
5096 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5097 if(dops[i].rs1==31) {
5098 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5099 do_preload_rhtbl(ht);
5103 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5104 #ifdef DESTRUCTIVE_WRITEBACK
5105 if((branch_regs[i].dirty>>rs)&1) {
5106 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5107 emit_loadreg(dops[i].rs1,rs);
5112 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5115 if(dops[i].rs1==31) {
5116 do_miniht_load(ht,rh);
5119 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5120 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5122 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5123 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5124 if(dops[i+1].itype==COP0&&(source[i+1]&0x3f)==0x10)
5125 // special case for RFE
5129 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5131 if(dops[i].rs1==31) {
5132 do_miniht_jump(rs,rh,ht);
5139 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5140 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5144 static void cjump_assemble(int i,struct regstat *i_regs)
5146 signed char *i_regmap=i_regs->regmap;
5149 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5150 assem_debug("match=%d\n",match);
5152 int unconditional=0,nop=0;
5154 int internal=internal_branch(ba[i]);
5155 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5156 if(!match) invert=1;
5157 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5158 if(i>(ba[i]-start)>>2) invert=1;
5161 invert=1; // because of near cond. branches
5165 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5166 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5169 s1l=get_reg(i_regmap,dops[i].rs1);
5170 s2l=get_reg(i_regmap,dops[i].rs2);
5172 if(dops[i].rs1==0&&dops[i].rs2==0)
5174 if(dops[i].opcode&1) nop=1;
5175 else unconditional=1;
5176 //assert(dops[i].opcode!=5);
5177 //assert(dops[i].opcode!=7);
5178 //assert(dops[i].opcode!=0x15);
5179 //assert(dops[i].opcode!=0x17);
5181 else if(dops[i].rs1==0)
5186 else if(dops[i].rs2==0)
5192 // Out of order execution (delay slot first)
5194 address_generation(i+1,i_regs,regs[i].regmap_entry);
5195 ds_assemble(i+1,i_regs);
5197 uint64_t bc_unneeded=branch_regs[i].u;
5198 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5200 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5201 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5202 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5203 cc=get_reg(branch_regs[i].regmap,CCREG);
5204 assert(cc==HOST_CCREG);
5206 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5207 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5208 //assem_debug("cycle count (adj)\n");
5210 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5211 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5212 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5213 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5215 assem_debug("branch: internal\n");
5217 assem_debug("branch: external\n");
5218 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5219 ds_assemble_entry(i);
5222 add_to_linker(out,ba[i],internal);
5225 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5226 if(((u_int)out)&7) emit_addnop(0);
5231 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5234 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5237 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5238 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5239 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5241 //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]);
5243 if(dops[i].opcode==4) // BEQ
5245 if(s2l>=0) emit_cmp(s1l,s2l);
5246 else emit_test(s1l,s1l);
5251 add_to_linker(out,ba[i],internal);
5255 if(dops[i].opcode==5) // BNE
5257 if(s2l>=0) emit_cmp(s1l,s2l);
5258 else emit_test(s1l,s1l);
5263 add_to_linker(out,ba[i],internal);
5267 if(dops[i].opcode==6) // BLEZ
5274 add_to_linker(out,ba[i],internal);
5278 if(dops[i].opcode==7) // BGTZ
5285 add_to_linker(out,ba[i],internal);
5290 if(taken) set_jump_target(taken, out);
5291 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5292 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5294 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5295 add_to_linker(out,ba[i],internal);
5298 add_to_linker(out,ba[i],internal*2);
5304 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5305 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5306 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5308 assem_debug("branch: internal\n");
5310 assem_debug("branch: external\n");
5311 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5312 ds_assemble_entry(i);
5315 add_to_linker(out,ba[i],internal);
5319 set_jump_target(nottaken, out);
5322 if(nottaken1) set_jump_target(nottaken1, out);
5324 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5326 } // (!unconditional)
5330 // In-order execution (branch first)
5331 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5332 if(!unconditional&&!nop) {
5333 //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]);
5335 if((dops[i].opcode&0x2f)==4) // BEQ
5337 if(s2l>=0) emit_cmp(s1l,s2l);
5338 else emit_test(s1l,s1l);
5342 if((dops[i].opcode&0x2f)==5) // BNE
5344 if(s2l>=0) emit_cmp(s1l,s2l);
5345 else emit_test(s1l,s1l);
5349 if((dops[i].opcode&0x2f)==6) // BLEZ
5355 if((dops[i].opcode&0x2f)==7) // BGTZ
5361 } // if(!unconditional)
5363 uint64_t ds_unneeded=branch_regs[i].u;
5364 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5368 if(taken) set_jump_target(taken, out);
5369 assem_debug("1:\n");
5370 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5372 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5373 address_generation(i+1,&branch_regs[i],0);
5374 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5375 ds_assemble(i+1,&branch_regs[i]);
5376 cc=get_reg(branch_regs[i].regmap,CCREG);
5378 emit_loadreg(CCREG,cc=HOST_CCREG);
5379 // CHECK: Is the following instruction (fall thru) allocated ok?
5381 assert(cc==HOST_CCREG);
5382 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5383 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5384 assem_debug("cycle count (adj)\n");
5385 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5386 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5388 assem_debug("branch: internal\n");
5390 assem_debug("branch: external\n");
5391 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5392 ds_assemble_entry(i);
5395 add_to_linker(out,ba[i],internal);
5400 if(!unconditional) {
5401 if(nottaken1) set_jump_target(nottaken1, out);
5402 set_jump_target(nottaken, out);
5403 assem_debug("2:\n");
5404 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5405 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5406 address_generation(i+1,&branch_regs[i],0);
5407 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5408 ds_assemble(i+1,&branch_regs[i]);
5409 cc=get_reg(branch_regs[i].regmap,CCREG);
5411 // Cycle count isn't in a register, temporarily load it then write it out
5412 emit_loadreg(CCREG,HOST_CCREG);
5413 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5416 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5417 emit_storereg(CCREG,HOST_CCREG);
5420 cc=get_reg(i_regmap,CCREG);
5421 assert(cc==HOST_CCREG);
5422 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5425 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5431 static void sjump_assemble(int i,struct regstat *i_regs)
5433 signed char *i_regmap=i_regs->regmap;
5436 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5437 assem_debug("smatch=%d\n",match);
5439 int unconditional=0,nevertaken=0;
5441 int internal=internal_branch(ba[i]);
5442 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5443 if(!match) invert=1;
5444 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5445 if(i>(ba[i]-start)>>2) invert=1;
5448 invert=1; // because of near cond. branches
5451 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5452 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5455 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5458 s1l=get_reg(i_regmap,dops[i].rs1);
5462 if(dops[i].opcode2&1) unconditional=1;
5464 // These are never taken (r0 is never less than zero)
5465 //assert(dops[i].opcode2!=0);
5466 //assert(dops[i].opcode2!=2);
5467 //assert(dops[i].opcode2!=0x10);
5468 //assert(dops[i].opcode2!=0x12);
5472 // Out of order execution (delay slot first)
5474 address_generation(i+1,i_regs,regs[i].regmap_entry);
5475 ds_assemble(i+1,i_regs);
5477 uint64_t bc_unneeded=branch_regs[i].u;
5478 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5480 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5481 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5482 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5483 if(dops[i].rt1==31) {
5484 int rt,return_address;
5485 rt=get_reg(branch_regs[i].regmap,31);
5486 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]);
5488 // Save the PC even if the branch is not taken
5489 return_address=start+i*4+8;
5490 emit_movimm(return_address,rt); // PC into link register
5492 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5496 cc=get_reg(branch_regs[i].regmap,CCREG);
5497 assert(cc==HOST_CCREG);
5499 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5500 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5501 assem_debug("cycle count (adj)\n");
5503 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5504 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5505 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5506 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5508 assem_debug("branch: internal\n");
5510 assem_debug("branch: external\n");
5511 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5512 ds_assemble_entry(i);
5515 add_to_linker(out,ba[i],internal);
5518 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5519 if(((u_int)out)&7) emit_addnop(0);
5523 else if(nevertaken) {
5524 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5527 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5530 void *nottaken = NULL;
5531 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5532 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5535 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5542 add_to_linker(out,ba[i],internal);
5546 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5553 add_to_linker(out,ba[i],internal);
5560 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5561 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5563 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5564 add_to_linker(out,ba[i],internal);
5567 add_to_linker(out,ba[i],internal*2);
5573 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5574 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5575 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5577 assem_debug("branch: internal\n");
5579 assem_debug("branch: external\n");
5580 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5581 ds_assemble_entry(i);
5584 add_to_linker(out,ba[i],internal);
5588 set_jump_target(nottaken, out);
5592 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5594 } // (!unconditional)
5598 // In-order execution (branch first)
5600 void *nottaken = NULL;
5601 if(dops[i].rt1==31) {
5602 int rt,return_address;
5603 rt=get_reg(branch_regs[i].regmap,31);
5605 // Save the PC even if the branch is not taken
5606 return_address=start+i*4+8;
5607 emit_movimm(return_address,rt); // PC into link register
5609 emit_prefetch(hash_table_get(return_address));
5613 if(!unconditional) {
5614 //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]);
5616 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5622 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5628 } // if(!unconditional)
5630 uint64_t ds_unneeded=branch_regs[i].u;
5631 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5635 //assem_debug("1:\n");
5636 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5638 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5639 address_generation(i+1,&branch_regs[i],0);
5640 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5641 ds_assemble(i+1,&branch_regs[i]);
5642 cc=get_reg(branch_regs[i].regmap,CCREG);
5644 emit_loadreg(CCREG,cc=HOST_CCREG);
5645 // CHECK: Is the following instruction (fall thru) allocated ok?
5647 assert(cc==HOST_CCREG);
5648 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5649 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5650 assem_debug("cycle count (adj)\n");
5651 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5652 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5654 assem_debug("branch: internal\n");
5656 assem_debug("branch: external\n");
5657 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5658 ds_assemble_entry(i);
5661 add_to_linker(out,ba[i],internal);
5666 if(!unconditional) {
5667 set_jump_target(nottaken, out);
5668 assem_debug("1:\n");
5669 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5670 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5671 address_generation(i+1,&branch_regs[i],0);
5672 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5673 ds_assemble(i+1,&branch_regs[i]);
5674 cc=get_reg(branch_regs[i].regmap,CCREG);
5676 // Cycle count isn't in a register, temporarily load it then write it out
5677 emit_loadreg(CCREG,HOST_CCREG);
5678 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5681 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5682 emit_storereg(CCREG,HOST_CCREG);
5685 cc=get_reg(i_regmap,CCREG);
5686 assert(cc==HOST_CCREG);
5687 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5690 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5696 static void pagespan_assemble(int i,struct regstat *i_regs)
5698 int s1l=get_reg(i_regs->regmap,dops[i].rs1);
5699 int s2l=get_reg(i_regs->regmap,dops[i].rs2);
5701 void *nottaken = NULL;
5702 int unconditional=0;
5708 else if(dops[i].rs2==0)
5713 int addr=-1,alt=-1,ntaddr=-1;
5714 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5718 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5719 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5720 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5729 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5730 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5731 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5737 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5741 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5742 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5743 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5750 assert(hr<HOST_REGS);
5751 if((dops[i].opcode&0x2e)==4||dops[i].opcode==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5752 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
5754 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5755 if(dops[i].opcode==2) // J
5759 if(dops[i].opcode==3) // JAL
5762 int rt=get_reg(i_regs->regmap,31);
5763 emit_movimm(start+i*4+8,rt);
5766 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
5769 if(dops[i].opcode2==9) // JALR
5771 int rt=get_reg(i_regs->regmap,dops[i].rt1);
5772 emit_movimm(start+i*4+8,rt);
5775 if((dops[i].opcode&0x3f)==4) // BEQ
5777 if(dops[i].rs1==dops[i].rs2)
5782 #ifdef HAVE_CMOV_IMM
5784 if(s2l>=0) emit_cmp(s1l,s2l);
5785 else emit_test(s1l,s1l);
5786 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5792 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5793 if(s2l>=0) emit_cmp(s1l,s2l);
5794 else emit_test(s1l,s1l);
5795 emit_cmovne_reg(alt,addr);
5798 if((dops[i].opcode&0x3f)==5) // BNE
5800 #ifdef HAVE_CMOV_IMM
5801 if(s2l>=0) emit_cmp(s1l,s2l);
5802 else emit_test(s1l,s1l);
5803 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5806 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5807 if(s2l>=0) emit_cmp(s1l,s2l);
5808 else emit_test(s1l,s1l);
5809 emit_cmovne_reg(alt,addr);
5812 if((dops[i].opcode&0x3f)==0x14) // BEQL
5814 if(s2l>=0) emit_cmp(s1l,s2l);
5815 else emit_test(s1l,s1l);
5816 if(nottaken) set_jump_target(nottaken, out);
5820 if((dops[i].opcode&0x3f)==0x15) // BNEL
5822 if(s2l>=0) emit_cmp(s1l,s2l);
5823 else emit_test(s1l,s1l);
5826 if(taken) set_jump_target(taken, out);
5828 if((dops[i].opcode&0x3f)==6) // BLEZ
5830 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5832 emit_cmovl_reg(alt,addr);
5834 if((dops[i].opcode&0x3f)==7) // BGTZ
5836 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5838 emit_cmovl_reg(ntaddr,addr);
5840 if((dops[i].opcode&0x3f)==0x16) // BLEZL
5842 assert((dops[i].opcode&0x3f)!=0x16);
5844 if((dops[i].opcode&0x3f)==0x17) // BGTZL
5846 assert((dops[i].opcode&0x3f)!=0x17);
5848 assert(dops[i].opcode!=1); // BLTZ/BGEZ
5850 //FIXME: Check CSREG
5851 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5852 if((source[i]&0x30000)==0) // BC1F
5854 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5855 emit_testimm(s1l,0x800000);
5856 emit_cmovne_reg(alt,addr);
5858 if((source[i]&0x30000)==0x10000) // BC1T
5860 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5861 emit_testimm(s1l,0x800000);
5862 emit_cmovne_reg(alt,addr);
5864 if((source[i]&0x30000)==0x20000) // BC1FL
5866 emit_testimm(s1l,0x800000);
5870 if((source[i]&0x30000)==0x30000) // BC1TL
5872 emit_testimm(s1l,0x800000);
5878 assert(i_regs->regmap[HOST_CCREG]==CCREG);
5879 wb_dirtys(regs[i].regmap,regs[i].dirty);
5882 emit_movimm(ba[i],HOST_BTREG);
5884 else if(addr!=HOST_BTREG)
5886 emit_mov(addr,HOST_BTREG);
5888 void *branch_addr=out;
5890 int target_addr=start+i*4+5;
5892 void *compiled_target_addr=check_addr(target_addr);
5893 emit_extjump_ds(branch_addr, target_addr);
5894 if(compiled_target_addr) {
5895 set_jump_target(branch_addr, compiled_target_addr);
5896 add_jump_out(target_addr,stub);
5898 else set_jump_target(branch_addr, stub);
5901 // Assemble the delay slot for the above
5902 static void pagespan_ds()
5904 assem_debug("initial delay slot:\n");
5905 u_int vaddr=start+1;
5906 u_int page=get_page(vaddr);
5907 u_int vpage=get_vpage(vaddr);
5908 ll_add(jump_dirty+vpage,vaddr,(void *)out);
5909 do_dirty_stub_ds(slen*4);
5910 ll_add(jump_in+page,vaddr,(void *)out);
5911 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
5912 if(regs[0].regmap[HOST_CCREG]!=CCREG)
5913 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
5914 if(regs[0].regmap[HOST_BTREG]!=BTREG)
5915 emit_writeword(HOST_BTREG,&branch_target);
5916 load_regs(regs[0].regmap_entry,regs[0].regmap,dops[0].rs1,dops[0].rs2);
5917 address_generation(0,®s[0],regs[0].regmap_entry);
5918 if(dops[0].itype==STORE||dops[0].itype==STORELR||(dops[0].opcode&0x3b)==0x39||(dops[0].opcode&0x3b)==0x3a)
5919 load_regs(regs[0].regmap_entry,regs[0].regmap,INVCP,INVCP);
5921 switch(dops[0].itype) {
5923 alu_assemble(0,®s[0]);break;
5925 imm16_assemble(0,®s[0]);break;
5927 shift_assemble(0,®s[0]);break;
5929 shiftimm_assemble(0,®s[0]);break;
5931 load_assemble(0,®s[0]);break;
5933 loadlr_assemble(0,®s[0]);break;
5935 store_assemble(0,®s[0]);break;
5937 storelr_assemble(0,®s[0]);break;
5939 cop0_assemble(0,®s[0]);break;
5941 cop1_assemble(0,®s[0]);break;
5943 c1ls_assemble(0,®s[0]);break;
5945 cop2_assemble(0,®s[0]);break;
5947 c2ls_assemble(0,®s[0]);break;
5949 c2op_assemble(0,®s[0]);break;
5951 multdiv_assemble(0,®s[0]);
5952 multdiv_prepare_stall(0,®s[0]);
5955 mov_assemble(0,®s[0]);break;
5964 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5966 int btaddr=get_reg(regs[0].regmap,BTREG);
5968 btaddr=get_reg(regs[0].regmap,-1);
5969 emit_readword(&branch_target,btaddr);
5971 assert(btaddr!=HOST_CCREG);
5972 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
5974 host_tempreg_acquire();
5975 emit_movimm(start+4,HOST_TEMPREG);
5976 emit_cmp(btaddr,HOST_TEMPREG);
5977 host_tempreg_release();
5979 emit_cmpimm(btaddr,start+4);
5983 store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
5984 do_jump_vaddr(btaddr);
5985 set_jump_target(branch, out);
5986 store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
5987 load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
5990 // Basic liveness analysis for MIPS registers
5991 void unneeded_registers(int istart,int iend,int r)
5994 uint64_t u,gte_u,b,gte_b;
5995 uint64_t temp_u,temp_gte_u=0;
5996 uint64_t gte_u_unknown=0;
5997 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
6001 gte_u=gte_u_unknown;
6003 //u=unneeded_reg[iend+1];
6005 gte_u=gte_unneeded[iend+1];
6008 for (i=iend;i>=istart;i--)
6010 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6013 // If subroutine call, flag return address as a possible branch target
6014 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
6016 if(ba[i]<start || ba[i]>=(start+slen*4))
6018 // Branch out of this block, flush all regs
6020 gte_u=gte_u_unknown;
6021 branch_unneeded_reg[i]=u;
6022 // Merge in delay slot
6023 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6024 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6027 gte_u&=~gte_rs[i+1];
6031 // Internal branch, flag target
6032 dops[(ba[i]-start)>>2].bt=1;
6033 if(ba[i]<=start+i*4) {
6035 if(dops[i].is_ujump)
6037 // Unconditional branch
6041 // Conditional branch (not taken case)
6042 temp_u=unneeded_reg[i+2];
6043 temp_gte_u&=gte_unneeded[i+2];
6045 // Merge in delay slot
6046 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6047 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6049 temp_gte_u|=gte_rt[i+1];
6050 temp_gte_u&=~gte_rs[i+1];
6051 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
6052 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
6054 temp_gte_u|=gte_rt[i];
6055 temp_gte_u&=~gte_rs[i];
6056 unneeded_reg[i]=temp_u;
6057 gte_unneeded[i]=temp_gte_u;
6058 // Only go three levels deep. This recursion can take an
6059 // excessive amount of time if there are a lot of nested loops.
6061 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6063 unneeded_reg[(ba[i]-start)>>2]=1;
6064 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
6067 if (dops[i].is_ujump)
6069 // Unconditional branch
6070 u=unneeded_reg[(ba[i]-start)>>2];
6071 gte_u=gte_unneeded[(ba[i]-start)>>2];
6072 branch_unneeded_reg[i]=u;
6073 // Merge in delay slot
6074 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6075 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6078 gte_u&=~gte_rs[i+1];
6080 // Conditional branch
6081 b=unneeded_reg[(ba[i]-start)>>2];
6082 gte_b=gte_unneeded[(ba[i]-start)>>2];
6083 branch_unneeded_reg[i]=b;
6084 // Branch delay slot
6085 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6086 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6089 gte_b&=~gte_rs[i+1];
6093 branch_unneeded_reg[i]&=unneeded_reg[i+2];
6095 branch_unneeded_reg[i]=1;
6101 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6103 // SYSCALL instruction (software interrupt)
6106 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6108 // ERET instruction (return from interrupt)
6112 // Written registers are unneeded
6113 u|=1LL<<dops[i].rt1;
6114 u|=1LL<<dops[i].rt2;
6116 // Accessed registers are needed
6117 u&=~(1LL<<dops[i].rs1);
6118 u&=~(1LL<<dops[i].rs2);
6120 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
6121 gte_u|=gte_rs[i]>e_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
6122 // Source-target dependencies
6123 // R0 is always unneeded
6127 gte_unneeded[i]=gte_u;
6129 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6132 for(r=1;r<=CCREG;r++) {
6133 if((unneeded_reg[i]>>r)&1) {
6134 if(r==HIREG) printf(" HI");
6135 else if(r==LOREG) printf(" LO");
6136 else printf(" r%d",r);
6144 // Write back dirty registers as soon as we will no longer modify them,
6145 // so that we don't end up with lots of writes at the branches.
6146 void clean_registers(int istart,int iend,int wr)
6150 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
6151 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
6153 will_dirty_i=will_dirty_next=0;
6154 wont_dirty_i=wont_dirty_next=0;
6156 will_dirty_i=will_dirty_next=will_dirty[iend+1];
6157 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
6159 for (i=iend;i>=istart;i--)
6163 if(ba[i]<start || ba[i]>=(start+slen*4))
6165 // Branch out of this block, flush all regs
6166 if (dops[i].is_ujump)
6168 // Unconditional branch
6171 // Merge in delay slot (will dirty)
6172 for(r=0;r<HOST_REGS;r++) {
6173 if(r!=EXCLUDE_REG) {
6174 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6175 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6176 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6177 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6178 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6179 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6180 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6181 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6182 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6183 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6184 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6185 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6186 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6187 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6193 // Conditional branch
6195 wont_dirty_i=wont_dirty_next;
6196 // Merge in delay slot (will dirty)
6197 for(r=0;r<HOST_REGS;r++) {
6198 if(r!=EXCLUDE_REG) {
6199 if (1) { // !dops[i].likely) {
6200 // Might not dirty if likely branch is not taken
6201 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6202 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6203 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6204 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6205 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6206 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
6207 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6208 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6209 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6210 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6211 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6212 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6213 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6214 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6219 // Merge in delay slot (wont dirty)
6220 for(r=0;r<HOST_REGS;r++) {
6221 if(r!=EXCLUDE_REG) {
6222 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6223 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6224 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6225 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6226 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6227 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6228 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6229 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6230 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6231 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6235 #ifndef DESTRUCTIVE_WRITEBACK
6236 branch_regs[i].dirty&=wont_dirty_i;
6238 branch_regs[i].dirty|=will_dirty_i;
6244 if(ba[i]<=start+i*4) {
6246 if (dops[i].is_ujump)
6248 // Unconditional branch
6251 // Merge in delay slot (will dirty)
6252 for(r=0;r<HOST_REGS;r++) {
6253 if(r!=EXCLUDE_REG) {
6254 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6255 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6256 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6257 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6258 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6259 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6260 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6261 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6262 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6263 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6264 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6265 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6266 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6267 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6271 // Conditional branch (not taken case)
6272 temp_will_dirty=will_dirty_next;
6273 temp_wont_dirty=wont_dirty_next;
6274 // Merge in delay slot (will dirty)
6275 for(r=0;r<HOST_REGS;r++) {
6276 if(r!=EXCLUDE_REG) {
6277 if (1) { // !dops[i].likely) {
6278 // Will not dirty if likely branch is not taken
6279 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6280 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6281 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6282 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6283 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6284 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
6285 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6286 //if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6287 //if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6288 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6289 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6290 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6291 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6292 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6297 // Merge in delay slot (wont dirty)
6298 for(r=0;r<HOST_REGS;r++) {
6299 if(r!=EXCLUDE_REG) {
6300 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6301 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6302 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6303 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6304 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6305 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6306 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6307 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6308 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6309 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6312 // Deal with changed mappings
6314 for(r=0;r<HOST_REGS;r++) {
6315 if(r!=EXCLUDE_REG) {
6316 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
6317 temp_will_dirty&=~(1<<r);
6318 temp_wont_dirty&=~(1<<r);
6319 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6320 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6321 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6323 temp_will_dirty|=1<<r;
6324 temp_wont_dirty|=1<<r;
6331 will_dirty[i]=temp_will_dirty;
6332 wont_dirty[i]=temp_wont_dirty;
6333 clean_registers((ba[i]-start)>>2,i-1,0);
6335 // Limit recursion. It can take an excessive amount
6336 // of time if there are a lot of nested loops.
6337 will_dirty[(ba[i]-start)>>2]=0;
6338 wont_dirty[(ba[i]-start)>>2]=-1;
6343 if (dops[i].is_ujump)
6345 // Unconditional branch
6348 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6349 for(r=0;r<HOST_REGS;r++) {
6350 if(r!=EXCLUDE_REG) {
6351 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6352 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
6353 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6355 if(branch_regs[i].regmap[r]>=0) {
6356 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6357 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6362 // Merge in delay slot
6363 for(r=0;r<HOST_REGS;r++) {
6364 if(r!=EXCLUDE_REG) {
6365 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6366 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6367 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6368 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6369 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6370 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6371 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6372 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6373 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6374 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6375 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6376 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6377 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6378 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6382 // Conditional branch
6383 will_dirty_i=will_dirty_next;
6384 wont_dirty_i=wont_dirty_next;
6385 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6386 for(r=0;r<HOST_REGS;r++) {
6387 if(r!=EXCLUDE_REG) {
6388 signed char target_reg=branch_regs[i].regmap[r];
6389 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6390 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
6391 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6393 else if(target_reg>=0) {
6394 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6395 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6400 // Merge in delay slot
6401 for(r=0;r<HOST_REGS;r++) {
6402 if(r!=EXCLUDE_REG) {
6403 if (1) { // !dops[i].likely) {
6404 // Might not dirty if likely branch is not taken
6405 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6406 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6407 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6408 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6409 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6410 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6411 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6412 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6413 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6414 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6415 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6416 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6417 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6418 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6423 // Merge in delay slot (won't dirty)
6424 for(r=0;r<HOST_REGS;r++) {
6425 if(r!=EXCLUDE_REG) {
6426 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6427 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6428 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6429 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6430 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6431 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6432 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6433 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6434 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6435 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6439 #ifndef DESTRUCTIVE_WRITEBACK
6440 branch_regs[i].dirty&=wont_dirty_i;
6442 branch_regs[i].dirty|=will_dirty_i;
6447 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6449 // SYSCALL instruction (software interrupt)
6453 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6455 // ERET instruction (return from interrupt)
6459 will_dirty_next=will_dirty_i;
6460 wont_dirty_next=wont_dirty_i;
6461 for(r=0;r<HOST_REGS;r++) {
6462 if(r!=EXCLUDE_REG) {
6463 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6464 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6465 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6466 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6467 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6468 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6469 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6470 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6472 if (!dops[i].is_jump)
6474 // Don't store a register immediately after writing it,
6475 // may prevent dual-issue.
6476 if((regs[i].regmap[r]&63)==dops[i-1].rt1) wont_dirty_i|=1<<r;
6477 if((regs[i].regmap[r]&63)==dops[i-1].rt2) wont_dirty_i|=1<<r;
6483 will_dirty[i]=will_dirty_i;
6484 wont_dirty[i]=wont_dirty_i;
6485 // Mark registers that won't be dirtied as not dirty
6487 regs[i].dirty|=will_dirty_i;
6488 #ifndef DESTRUCTIVE_WRITEBACK
6489 regs[i].dirty&=wont_dirty_i;
6492 if (i < iend-1 && !dops[i].is_ujump) {
6493 for(r=0;r<HOST_REGS;r++) {
6494 if(r!=EXCLUDE_REG) {
6495 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
6496 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
6497 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6505 for(r=0;r<HOST_REGS;r++) {
6506 if(r!=EXCLUDE_REG) {
6507 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
6508 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
6509 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6517 // Deal with changed mappings
6518 temp_will_dirty=will_dirty_i;
6519 temp_wont_dirty=wont_dirty_i;
6520 for(r=0;r<HOST_REGS;r++) {
6521 if(r!=EXCLUDE_REG) {
6523 if(regs[i].regmap[r]==regmap_pre[i][r]) {
6525 #ifndef DESTRUCTIVE_WRITEBACK
6526 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6528 regs[i].wasdirty|=will_dirty_i&(1<<r);
6531 else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
6532 // Register moved to a different register
6533 will_dirty_i&=~(1<<r);
6534 wont_dirty_i&=~(1<<r);
6535 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
6536 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
6538 #ifndef DESTRUCTIVE_WRITEBACK
6539 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6541 regs[i].wasdirty|=will_dirty_i&(1<<r);
6545 will_dirty_i&=~(1<<r);
6546 wont_dirty_i&=~(1<<r);
6547 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6548 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6549 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6552 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
6562 void disassemble_inst(int i)
6564 if (dops[i].bt) printf("*"); else printf(" ");
6565 switch(dops[i].itype) {
6567 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6569 printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):*ba);break;
6571 printf (" %x: %s r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14));break;
6573 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6574 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6576 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6579 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,ba[i]);break;
6581 if(dops[i].opcode==0xf) //LUI
6582 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6584 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6588 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6592 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6596 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6599 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6602 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6605 if((dops[i].opcode2&0x1d)==0x10)
6606 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6607 else if((dops[i].opcode2&0x1d)==0x11)
6608 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6610 printf (" %x: %s\n",start+i*4,insn[i]);
6613 if(dops[i].opcode2==0)
6614 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6615 else if(dops[i].opcode2==4)
6616 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6617 else printf (" %x: %s\n",start+i*4,insn[i]);
6620 if(dops[i].opcode2<3)
6621 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6622 else if(dops[i].opcode2>3)
6623 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6624 else printf (" %x: %s\n",start+i*4,insn[i]);
6627 if(dops[i].opcode2<3)
6628 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6629 else if(dops[i].opcode2>3)
6630 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6631 else printf (" %x: %s\n",start+i*4,insn[i]);
6634 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6637 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6640 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6643 //printf (" %s %8x\n",insn[i],source[i]);
6644 printf (" %x: %s\n",start+i*4,insn[i]);
6648 static void disassemble_inst(int i) {}
6651 #define DRC_TEST_VAL 0x74657374
6653 static void new_dynarec_test(void)
6655 int (*testfunc)(void);
6660 // check structure linkage
6661 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6663 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6666 SysPrintf("testing if we can run recompiled code...\n");
6667 ((volatile u_int *)out)[0]++; // make cache dirty
6669 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6670 out = ndrc->translation_cache;
6671 beginning = start_block();
6672 emit_movimm(DRC_TEST_VAL + i, 0); // test
6675 end_block(beginning);
6676 testfunc = beginning;
6677 ret[i] = testfunc();
6680 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6681 SysPrintf("test passed.\n");
6683 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6684 out = ndrc->translation_cache;
6687 // clear the state completely, instead of just marking
6688 // things invalid like invalidate_all_pages() does
6689 void new_dynarec_clear_full(void)
6692 out = ndrc->translation_cache;
6693 memset(invalid_code,1,sizeof(invalid_code));
6694 memset(hash_table,0xff,sizeof(hash_table));
6695 memset(mini_ht,-1,sizeof(mini_ht));
6696 memset(restore_candidate,0,sizeof(restore_candidate));
6697 memset(shadow,0,sizeof(shadow));
6699 expirep=16384; // Expiry pointer, +2 blocks
6700 pending_exception=0;
6703 inv_code_start=inv_code_end=~0;
6706 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6707 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6708 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6710 cycle_multiplier_old = cycle_multiplier;
6711 new_dynarec_hacks_old = new_dynarec_hacks;
6714 void new_dynarec_init(void)
6716 SysPrintf("Init new dynarec\n");
6718 #ifdef BASE_ADDR_DYNAMIC
6720 sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
6722 SysPrintf("sceKernelAllocMemBlockForVM failed\n");
6723 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6725 SysPrintf("sceKernelGetMemBlockBase failed\n");
6727 uintptr_t desired_addr = 0;
6730 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6732 ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
6733 PROT_READ | PROT_WRITE | PROT_EXEC,
6734 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
6735 if (ndrc == MAP_FAILED) {
6736 SysPrintf("mmap() failed: %s\n", strerror(errno));
6741 #ifndef NO_WRITE_EXEC
6742 // not all systems allow execute in data segment by default
6743 if (mprotect(ndrc, sizeof(ndrc->translation_cache) + sizeof(ndrc->tramp.ops),
6744 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6745 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6748 out = ndrc->translation_cache;
6749 cycle_multiplier=200;
6750 new_dynarec_clear_full();
6752 // Copy this into local area so we don't have to put it in every literal pool
6753 invc_ptr=invalid_code;
6758 ram_offset=(uintptr_t)rdram-0x80000000;
6761 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6764 void new_dynarec_cleanup(void)
6767 #ifdef BASE_ADDR_DYNAMIC
6769 sceKernelFreeMemBlock(sceBlock);
6772 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6773 SysPrintf("munmap() failed\n");
6776 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6777 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6778 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6780 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
6784 static u_int *get_source_start(u_int addr, u_int *limit)
6786 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6787 cycle_multiplier_override = 0;
6789 if (addr < 0x00200000 ||
6790 (0xa0000000 <= addr && addr < 0xa0200000))
6792 // used for BIOS calls mostly?
6793 *limit = (addr&0xa0000000)|0x00200000;
6794 return (u_int *)(rdram + (addr&0x1fffff));
6796 else if (!Config.HLE && (
6797 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6798 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6800 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6801 // but timings in PCSX are too tied to the interpreter's BIAS
6802 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6803 cycle_multiplier_override = 200;
6805 *limit = (addr & 0xfff00000) | 0x80000;
6806 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6808 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6809 *limit = (addr & 0x80600000) + 0x00200000;
6810 return (u_int *)(rdram + (addr&0x1fffff));
6815 static u_int scan_for_ret(u_int addr)
6820 mem = get_source_start(addr, &limit);
6824 if (limit > addr + 0x1000)
6825 limit = addr + 0x1000;
6826 for (; addr < limit; addr += 4, mem++) {
6827 if (*mem == 0x03e00008) // jr $ra
6833 struct savestate_block {
6838 static int addr_cmp(const void *p1_, const void *p2_)
6840 const struct savestate_block *p1 = p1_, *p2 = p2_;
6841 return p1->addr - p2->addr;
6844 int new_dynarec_save_blocks(void *save, int size)
6846 struct savestate_block *blocks = save;
6847 int maxcount = size / sizeof(blocks[0]);
6848 struct savestate_block tmp_blocks[1024];
6849 struct ll_entry *head;
6850 int p, s, d, o, bcnt;
6854 for (p = 0; p < ARRAY_SIZE(jump_in); p++) {
6856 for (head = jump_in[p]; head != NULL; head = head->next) {
6857 tmp_blocks[bcnt].addr = head->vaddr;
6858 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
6863 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6865 addr = tmp_blocks[0].addr;
6866 for (s = d = 0; s < bcnt; s++) {
6867 if (tmp_blocks[s].addr < addr)
6869 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6870 tmp_blocks[d++] = tmp_blocks[s];
6871 addr = scan_for_ret(tmp_blocks[s].addr);
6874 if (o + d > maxcount)
6876 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
6880 return o * sizeof(blocks[0]);
6883 void new_dynarec_load_blocks(const void *save, int size)
6885 const struct savestate_block *blocks = save;
6886 int count = size / sizeof(blocks[0]);
6887 u_int regs_save[32];
6891 get_addr(psxRegs.pc);
6893 // change GPRs for speculation to at least partially work..
6894 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6895 for (i = 1; i < 32; i++)
6896 psxRegs.GPR.r[i] = 0x80000000;
6898 for (b = 0; b < count; b++) {
6899 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6901 psxRegs.GPR.r[i] = 0x1f800000;
6904 get_addr(blocks[b].addr);
6906 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6908 psxRegs.GPR.r[i] = 0x80000000;
6912 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6915 int new_recompile_block(u_int addr)
6917 u_int pagelimit = 0;
6918 u_int state_rflags = 0;
6921 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
6922 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
6924 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
6926 // this is just for speculation
6927 for (i = 1; i < 32; i++) {
6928 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
6929 state_rflags |= 1 << i;
6932 start = (u_int)addr&~3;
6933 //assert(((u_int)addr&1)==0); // start-in-delay-slot flag
6934 new_dynarec_did_compile=1;
6935 if (Config.HLE && start == 0x80001000) // hlecall
6937 // XXX: is this enough? Maybe check hleSoftCall?
6938 void *beginning=start_block();
6939 u_int page=get_page(start);
6941 invalid_code[start>>12]=0;
6942 emit_movimm(start,0);
6943 emit_writeword(0,&pcaddr);
6944 emit_far_jump(new_dyna_leave);
6946 end_block(beginning);
6947 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
6950 else if (f1_hack == ~0u || (f1_hack != 0 && start == f1_hack)) {
6951 void *beginning = start_block();
6952 u_int page = get_page(start);
6953 emit_readword(&psxRegs.GPR.n.sp, 0);
6954 emit_readptr(&mem_rtab, 1);
6955 emit_shrimm(0, 12, 2);
6956 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
6957 emit_addimm(0, 0x18, 0);
6958 emit_adds_ptr(1, 1, 1);
6959 emit_ldr_dualindexed(1, 0, 0);
6960 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
6961 emit_far_call(get_addr_ht);
6962 emit_jmpreg(0); // jr k0
6964 end_block(beginning);
6966 ll_add_flags(jump_in + page, start, state_rflags, beginning);
6967 SysPrintf("F1 hack to %08x\n", start);
6972 source = get_source_start(start, &pagelimit);
6973 if (source == NULL) {
6974 SysPrintf("Compile at bogus memory address: %08x\n", addr);
6978 /* Pass 1: disassemble */
6979 /* Pass 2: register dependencies, branch targets */
6980 /* Pass 3: register allocation */
6981 /* Pass 4: branch dependencies */
6982 /* Pass 5: pre-alloc */
6983 /* Pass 6: optimize clean/dirty state */
6984 /* Pass 7: flag 32-bit registers */
6985 /* Pass 8: assembly */
6986 /* Pass 9: linker */
6987 /* Pass 10: garbage collection / free memory */
6991 unsigned int type,op,op2;
6993 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
6995 /* Pass 1 disassembly */
6997 for(i=0;!done;i++) {
7001 minimum_free_regs[i]=0;
7002 dops[i].opcode=op=source[i]>>26;
7005 case 0x00: strcpy(insn[i],"special"); type=NI;
7009 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
7010 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
7011 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
7012 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
7013 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
7014 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
7015 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
7016 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
7017 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
7018 case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
7019 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
7020 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
7021 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
7022 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
7023 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
7024 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
7025 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
7026 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
7027 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
7028 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
7029 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
7030 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
7031 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
7032 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
7033 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
7034 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
7035 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
7036 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
7037 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
7038 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
7039 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
7040 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
7041 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
7042 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
7043 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
7045 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
7046 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
7047 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
7048 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
7049 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
7050 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
7051 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
7052 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
7053 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
7054 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
7055 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
7056 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
7057 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
7058 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
7059 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
7060 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
7061 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
7065 case 0x01: strcpy(insn[i],"regimm"); type=NI;
7066 op2=(source[i]>>16)&0x1f;
7069 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
7070 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
7071 //case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
7072 //case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
7073 //case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
7074 //case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
7075 //case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
7076 //case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
7077 //case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
7078 //case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
7079 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
7080 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
7081 //case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
7082 //case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
7085 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
7086 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
7087 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
7088 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
7089 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
7090 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
7091 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
7092 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
7093 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
7094 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
7095 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
7096 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
7097 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
7098 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
7099 case 0x10: strcpy(insn[i],"cop0"); type=NI;
7100 op2=(source[i]>>21)&0x1f;
7103 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
7104 case 0x02: strcpy(insn[i],"CFC0"); type=COP0; break;
7105 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
7106 case 0x06: strcpy(insn[i],"CTC0"); type=COP0; break;
7107 case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
7110 case 0x11: strcpy(insn[i],"cop1"); type=COP1;
7111 op2=(source[i]>>21)&0x1f;
7114 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
7115 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
7116 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
7117 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
7118 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
7119 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
7120 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
7121 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
7123 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
7124 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
7125 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
7126 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
7127 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
7128 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
7129 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
7131 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
7133 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
7134 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
7135 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
7136 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
7138 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
7139 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
7141 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
7142 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
7143 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
7144 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
7146 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
7147 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
7148 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
7150 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
7151 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
7153 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
7154 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
7155 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
7157 case 0x12: strcpy(insn[i],"COP2"); type=NI;
7158 op2=(source[i]>>21)&0x1f;
7160 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
7161 if (gte_handlers[source[i]&0x3f]!=NULL) {
7162 if (gte_regnames[source[i]&0x3f]!=NULL)
7163 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
7165 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
7171 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
7172 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
7173 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
7174 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
7177 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
7178 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
7179 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
7180 default: strcpy(insn[i],"???"); type=NI;
7181 SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
7185 dops[i].opcode2=op2;
7186 /* Get registers/immediates */
7188 gte_rs[i]=gte_rt[i]=0;
7191 dops[i].rs1=(source[i]>>21)&0x1f;
7193 dops[i].rt1=(source[i]>>16)&0x1f;
7195 imm[i]=(short)source[i];
7199 dops[i].rs1=(source[i]>>21)&0x1f;
7200 dops[i].rs2=(source[i]>>16)&0x1f;
7203 imm[i]=(short)source[i];
7206 // LWL/LWR only load part of the register,
7207 // therefore the target register must be treated as a source too
7208 dops[i].rs1=(source[i]>>21)&0x1f;
7209 dops[i].rs2=(source[i]>>16)&0x1f;
7210 dops[i].rt1=(source[i]>>16)&0x1f;
7212 imm[i]=(short)source[i];
7215 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
7216 else dops[i].rs1=(source[i]>>21)&0x1f;
7218 dops[i].rt1=(source[i]>>16)&0x1f;
7220 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
7221 imm[i]=(unsigned short)source[i];
7223 imm[i]=(short)source[i];
7231 // The JAL instruction writes to r31.
7238 dops[i].rs1=(source[i]>>21)&0x1f;
7242 // The JALR instruction writes to rd.
7244 dops[i].rt1=(source[i]>>11)&0x1f;
7249 dops[i].rs1=(source[i]>>21)&0x1f;
7250 dops[i].rs2=(source[i]>>16)&0x1f;
7253 if(op&2) { // BGTZ/BLEZ
7258 dops[i].rs1=(source[i]>>21)&0x1f;
7262 if(op2&0x10) { // BxxAL
7264 // NOTE: If the branch is not taken, r31 is still overwritten
7268 dops[i].rs1=(source[i]>>21)&0x1f; // source
7269 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
7270 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7274 dops[i].rs1=(source[i]>>21)&0x1f; // source
7275 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
7284 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
7285 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
7286 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
7287 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
7288 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
7289 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
7292 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
7293 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
7294 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7298 dops[i].rs1=(source[i]>>16)&0x1f;
7300 dops[i].rt1=(source[i]>>11)&0x1f;
7302 imm[i]=(source[i]>>6)&0x1f;
7303 // DSxx32 instructions
7304 if(op2>=0x3c) imm[i]|=0x20;
7311 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
7312 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
7313 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
7314 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
7321 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
7322 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
7330 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
7331 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
7333 int gr=(source[i]>>11)&0x1F;
7336 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
7337 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
7338 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
7339 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
7343 dops[i].rs1=(source[i]>>21)&0x1F;
7347 imm[i]=(short)source[i];
7350 dops[i].rs1=(source[i]>>21)&0x1F;
7354 imm[i]=(short)source[i];
7355 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
7356 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
7363 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
7364 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
7365 gte_rt[i]|=1ll<<63; // every op changes flags
7366 if((source[i]&0x3f)==GTE_MVMVA) {
7367 int v = (source[i] >> 15) & 3;
7368 gte_rs[i]&=~0xe3fll;
7369 if(v==3) gte_rs[i]|=0xe00ll;
7370 else gte_rs[i]|=3ll<<(v*2);
7387 /* Calculate branch target addresses */
7389 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7390 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
7391 ba[i]=start+i*4+8; // Ignore never taken branch
7392 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
7393 ba[i]=start+i*4+8; // Ignore never taken branch
7394 else if(type==CJUMP||type==SJUMP)
7395 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7398 /* simplify always (not)taken branches */
7399 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
7400 dops[i].rs1 = dops[i].rs2 = 0;
7402 dops[i].itype = type = UJUMP;
7403 dops[i].rs2 = CCREG;
7406 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
7407 dops[i].itype = type = UJUMP;
7409 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
7410 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
7412 /* messy cases to just pass over to the interpreter */
7413 if (i > 0 && dops[i-1].is_jump) {
7415 // branch in delay slot?
7416 if (dops[i].is_jump) {
7417 // don't handle first branch and call interpreter if it's hit
7418 SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
7421 // basic load delay detection
7422 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
7423 int t=(ba[i-1]-start)/4;
7424 if(0 <= t && t < i &&(dops[i].rt1==dops[t].rs1||dops[i].rt1==dops[t].rs2)&&dops[t].itype!=CJUMP&&dops[t].itype!=SJUMP) {
7425 // jump target wants DS result - potential load delay effect
7426 SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
7428 dops[t+1].bt=1; // expected return from interpreter
7430 else if(i>=2&&dops[i-2].rt1==2&&dops[i].rt1==2&&dops[i].rs1!=2&&dops[i].rs2!=2&&dops[i-1].rs1!=2&&dops[i-1].rs2!=2&&
7431 !(i>=3&&dops[i-3].is_jump)) {
7432 // v0 overwrite like this is a sign of trouble, bail out
7433 SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
7438 dops[i-1].rs1=CCREG;
7439 dops[i-1].rs2=dops[i-1].rt1=dops[i-1].rt2=0;
7441 dops[i-1].itype=INTCALL;
7443 i--; // don't compile the DS
7447 /* Is this the end of the block? */
7448 if (i > 0 && dops[i-1].is_ujump) {
7449 if(dops[i-1].rt1==0) { // Continue past subroutine call (JAL)
7453 if(stop_after_jal) done=1;
7455 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7457 // Don't recompile stuff that's already compiled
7458 if(check_addr(start+i*4+4)) done=1;
7459 // Don't get too close to the limit
7460 if(i>MAXBLOCK/2) done=1;
7462 if(dops[i].itype==SYSCALL&&stop_after_jal) done=1;
7463 if(dops[i].itype==HLECALL||dops[i].itype==INTCALL) done=2;
7465 // Does the block continue due to a branch?
7468 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7469 if(ba[j]==start+i*4+4) done=j=0;
7470 if(ba[j]==start+i*4+8) done=j=0;
7473 //assert(i<MAXBLOCK-1);
7474 if(start+i*4==pagelimit-4) done=1;
7475 assert(start+i*4<pagelimit);
7476 if (i==MAXBLOCK-1) done=1;
7477 // Stop if we're compiling junk
7478 if(dops[i].itype==NI&&dops[i].opcode==0x11) {
7479 done=stop_after_jal=1;
7480 SysPrintf("Disabled speculative precompilation\n");
7484 if (dops[i-1].is_jump) {
7485 if(start+i*4==pagelimit) {
7486 dops[i-1].itype=SPAN;
7491 /* spacial hack(s) */
7492 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
7493 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
7494 && dops[i-7].itype == STORE)
7497 if (dops[i].itype == IMM16)
7499 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
7500 if (dops[i].itype == STORELR && dops[i].rs1 == 6
7501 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
7503 SysPrintf("F1 hack from %08x\n", start);
7508 /* Pass 2 - Register dependencies and branch targets */
7510 unneeded_registers(0,slen-1,0);
7512 /* Pass 3 - Register allocation */
7514 struct regstat current; // Current register allocations/status
7516 current.u=unneeded_reg[0];
7517 clear_all_regs(current.regmap);
7518 alloc_reg(¤t,0,CCREG);
7519 dirty_reg(¤t,CCREG);
7522 current.waswritten=0;
7528 // First instruction is delay slot
7533 current.regmap[HOST_BTREG]=BTREG;
7541 for(hr=0;hr<HOST_REGS;hr++)
7543 // Is this really necessary?
7544 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7547 current.waswritten=0;
7550 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7551 regs[i].wasconst=current.isconst;
7552 regs[i].wasdirty=current.dirty;
7553 regs[i].loadedconst=0;
7554 if (!dops[i].is_jump) {
7556 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7563 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7564 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7566 } else { SysPrintf("oops, branch at end of block with no delay slot\n");abort(); }
7570 ds=0; // Skip delay slot, already allocated as part of branch
7571 // ...but we need to alloc it in case something jumps here
7573 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7575 current.u=branch_unneeded_reg[i-1];
7577 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7579 struct regstat temp;
7580 memcpy(&temp,¤t,sizeof(current));
7581 temp.wasdirty=temp.dirty;
7582 // TODO: Take into account unconditional branches, as below
7583 delayslot_alloc(&temp,i);
7584 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7585 regs[i].wasdirty=temp.wasdirty;
7586 regs[i].dirty=temp.dirty;
7590 // Create entry (branch target) regmap
7591 for(hr=0;hr<HOST_REGS;hr++)
7593 int r=temp.regmap[hr];
7595 if(r!=regmap_pre[i][hr]) {
7596 regs[i].regmap_entry[hr]=-1;
7601 if((current.u>>r)&1) {
7602 regs[i].regmap_entry[hr]=-1;
7603 regs[i].regmap[hr]=-1;
7604 //Don't clear regs in the delay slot as the branch might need them
7605 //current.regmap[hr]=-1;
7607 regs[i].regmap_entry[hr]=r;
7610 // First instruction expects CCREG to be allocated
7611 if(i==0&&hr==HOST_CCREG)
7612 regs[i].regmap_entry[hr]=CCREG;
7614 regs[i].regmap_entry[hr]=-1;
7618 else { // Not delay slot
7619 switch(dops[i].itype) {
7621 //current.isconst=0; // DEBUG
7622 //current.wasconst=0; // DEBUG
7623 //regs[i].wasconst=0; // DEBUG
7624 clear_const(¤t,dops[i].rt1);
7625 alloc_cc(¤t,i);
7626 dirty_reg(¤t,CCREG);
7627 if (dops[i].rt1==31) {
7628 alloc_reg(¤t,i,31);
7629 dirty_reg(¤t,31);
7630 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7631 //assert(dops[i+1].rt1!=dops[i].rt1);
7633 alloc_reg(¤t,i,PTEMP);
7637 delayslot_alloc(¤t,i+1);
7638 //current.isconst=0; // DEBUG
7640 //printf("i=%d, isconst=%x\n",i,current.isconst);
7643 //current.isconst=0;
7644 //current.wasconst=0;
7645 //regs[i].wasconst=0;
7646 clear_const(¤t,dops[i].rs1);
7647 clear_const(¤t,dops[i].rt1);
7648 alloc_cc(¤t,i);
7649 dirty_reg(¤t,CCREG);
7650 if (!ds_writes_rjump_rs(i)) {
7651 alloc_reg(¤t,i,dops[i].rs1);
7652 if (dops[i].rt1!=0) {
7653 alloc_reg(¤t,i,dops[i].rt1);
7654 dirty_reg(¤t,dops[i].rt1);
7655 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7656 assert(dops[i+1].rt1!=dops[i].rt1);
7658 alloc_reg(¤t,i,PTEMP);
7662 if(dops[i].rs1==31) { // JALR
7663 alloc_reg(¤t,i,RHASH);
7664 alloc_reg(¤t,i,RHTBL);
7667 delayslot_alloc(¤t,i+1);
7669 // The delay slot overwrites our source register,
7670 // allocate a temporary register to hold the old value.
7674 delayslot_alloc(¤t,i+1);
7676 alloc_reg(¤t,i,RTEMP);
7678 //current.isconst=0; // DEBUG
7683 //current.isconst=0;
7684 //current.wasconst=0;
7685 //regs[i].wasconst=0;
7686 clear_const(¤t,dops[i].rs1);
7687 clear_const(¤t,dops[i].rs2);
7688 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7690 alloc_cc(¤t,i);
7691 dirty_reg(¤t,CCREG);
7692 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7693 if(dops[i].rs2) alloc_reg(¤t,i,dops[i].rs2);
7694 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7695 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7696 // The delay slot overwrites one of our conditions.
7697 // Allocate the branch condition registers instead.
7701 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7702 if(dops[i].rs2) alloc_reg(¤t,i,dops[i].rs2);
7707 delayslot_alloc(¤t,i+1);
7711 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7713 alloc_cc(¤t,i);
7714 dirty_reg(¤t,CCREG);
7715 alloc_reg(¤t,i,dops[i].rs1);
7716 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7717 // The delay slot overwrites one of our conditions.
7718 // Allocate the branch condition registers instead.
7722 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7727 delayslot_alloc(¤t,i+1);
7731 // Don't alloc the delay slot yet because we might not execute it
7732 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7737 alloc_cc(¤t,i);
7738 dirty_reg(¤t,CCREG);
7739 alloc_reg(¤t,i,dops[i].rs1);
7740 alloc_reg(¤t,i,dops[i].rs2);
7743 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7748 alloc_cc(¤t,i);
7749 dirty_reg(¤t,CCREG);
7750 alloc_reg(¤t,i,dops[i].rs1);
7753 //current.isconst=0;
7756 //current.isconst=0;
7757 //current.wasconst=0;
7758 //regs[i].wasconst=0;
7759 clear_const(¤t,dops[i].rs1);
7760 clear_const(¤t,dops[i].rt1);
7761 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7762 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7764 alloc_cc(¤t,i);
7765 dirty_reg(¤t,CCREG);
7766 alloc_reg(¤t,i,dops[i].rs1);
7767 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7768 alloc_reg(¤t,i,31);
7769 dirty_reg(¤t,31);
7770 //#ifdef REG_PREFETCH
7771 //alloc_reg(¤t,i,PTEMP);
7774 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7775 ||(dops[i].rt1==31&&(dops[i+1].rs1==31||dops[i+1].rs2==31||dops[i+1].rt1==31||dops[i+1].rt2==31))) { // DS touches $ra
7776 // Allocate the branch condition registers instead.
7780 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7785 delayslot_alloc(¤t,i+1);
7789 // Don't alloc the delay slot yet because we might not execute it
7790 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7795 alloc_cc(¤t,i);
7796 dirty_reg(¤t,CCREG);
7797 alloc_reg(¤t,i,dops[i].rs1);
7800 //current.isconst=0;
7803 imm16_alloc(¤t,i);
7807 load_alloc(¤t,i);
7811 store_alloc(¤t,i);
7814 alu_alloc(¤t,i);
7817 shift_alloc(¤t,i);
7820 multdiv_alloc(¤t,i);
7823 shiftimm_alloc(¤t,i);
7826 mov_alloc(¤t,i);
7829 cop0_alloc(¤t,i);
7834 cop2_alloc(¤t,i);
7837 c1ls_alloc(¤t,i);
7840 c2ls_alloc(¤t,i);
7843 c2op_alloc(¤t,i);
7848 syscall_alloc(¤t,i);
7851 pagespan_alloc(¤t,i);
7855 // Create entry (branch target) regmap
7856 for(hr=0;hr<HOST_REGS;hr++)
7859 r=current.regmap[hr];
7861 if(r!=regmap_pre[i][hr]) {
7862 // TODO: delay slot (?)
7863 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7864 if(or<0||(r&63)>=TEMPREG){
7865 regs[i].regmap_entry[hr]=-1;
7869 // Just move it to a different register
7870 regs[i].regmap_entry[hr]=r;
7871 // If it was dirty before, it's still dirty
7872 if((regs[i].wasdirty>>or)&1) dirty_reg(¤t,r&63);
7879 regs[i].regmap_entry[hr]=0;
7884 if((current.u>>r)&1) {
7885 regs[i].regmap_entry[hr]=-1;
7886 //regs[i].regmap[hr]=-1;
7887 current.regmap[hr]=-1;
7889 regs[i].regmap_entry[hr]=r;
7893 // Branches expect CCREG to be allocated at the target
7894 if(regmap_pre[i][hr]==CCREG)
7895 regs[i].regmap_entry[hr]=CCREG;
7897 regs[i].regmap_entry[hr]=-1;
7900 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7903 if(i>0&&(dops[i-1].itype==STORE||dops[i-1].itype==STORELR||(dops[i-1].itype==C2LS&&dops[i-1].opcode==0x3a))&&(u_int)imm[i-1]<0x800)
7904 current.waswritten|=1<<dops[i-1].rs1;
7905 current.waswritten&=~(1<<dops[i].rt1);
7906 current.waswritten&=~(1<<dops[i].rt2);
7907 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
7908 current.waswritten&=~(1<<dops[i].rs1);
7910 /* Branch post-alloc */
7913 current.wasdirty=current.dirty;
7914 switch(dops[i-1].itype) {
7916 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7917 branch_regs[i-1].isconst=0;
7918 branch_regs[i-1].wasconst=0;
7919 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7920 alloc_cc(&branch_regs[i-1],i-1);
7921 dirty_reg(&branch_regs[i-1],CCREG);
7922 if(dops[i-1].rt1==31) { // JAL
7923 alloc_reg(&branch_regs[i-1],i-1,31);
7924 dirty_reg(&branch_regs[i-1],31);
7926 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7927 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7930 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7931 branch_regs[i-1].isconst=0;
7932 branch_regs[i-1].wasconst=0;
7933 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7934 alloc_cc(&branch_regs[i-1],i-1);
7935 dirty_reg(&branch_regs[i-1],CCREG);
7936 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7937 if(dops[i-1].rt1!=0) { // JALR
7938 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7939 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7942 if(dops[i-1].rs1==31) { // JALR
7943 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7944 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7947 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7948 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7951 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7953 alloc_cc(¤t,i-1);
7954 dirty_reg(¤t,CCREG);
7955 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7956 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7957 // The delay slot overwrote one of our conditions
7958 // Delay slot goes after the test (in order)
7959 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7961 delayslot_alloc(¤t,i);
7966 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7967 // Alloc the branch condition registers
7968 if(dops[i-1].rs1) alloc_reg(¤t,i-1,dops[i-1].rs1);
7969 if(dops[i-1].rs2) alloc_reg(¤t,i-1,dops[i-1].rs2);
7971 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7972 branch_regs[i-1].isconst=0;
7973 branch_regs[i-1].wasconst=0;
7974 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7975 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7978 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7980 alloc_cc(¤t,i-1);
7981 dirty_reg(¤t,CCREG);
7982 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7983 // The delay slot overwrote the branch condition
7984 // Delay slot goes after the test (in order)
7985 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7987 delayslot_alloc(¤t,i);
7992 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7993 // Alloc the branch condition register
7994 alloc_reg(¤t,i-1,dops[i-1].rs1);
7996 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7997 branch_regs[i-1].isconst=0;
7998 branch_regs[i-1].wasconst=0;
7999 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8000 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8003 // Alloc the delay slot in case the branch is taken
8004 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
8006 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8007 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
8008 alloc_cc(&branch_regs[i-1],i);
8009 dirty_reg(&branch_regs[i-1],CCREG);
8010 delayslot_alloc(&branch_regs[i-1],i);
8011 branch_regs[i-1].isconst=0;
8012 alloc_reg(¤t,i,CCREG); // Not taken path
8013 dirty_reg(¤t,CCREG);
8014 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8017 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
8019 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8020 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
8021 alloc_cc(&branch_regs[i-1],i);
8022 dirty_reg(&branch_regs[i-1],CCREG);
8023 delayslot_alloc(&branch_regs[i-1],i);
8024 branch_regs[i-1].isconst=0;
8025 alloc_reg(¤t,i,CCREG); // Not taken path
8026 dirty_reg(¤t,CCREG);
8027 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8031 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
8032 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
8034 alloc_cc(¤t,i-1);
8035 dirty_reg(¤t,CCREG);
8036 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
8037 // The delay slot overwrote the branch condition
8038 // Delay slot goes after the test (in order)
8039 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8041 delayslot_alloc(¤t,i);
8046 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
8047 // Alloc the branch condition register
8048 alloc_reg(¤t,i-1,dops[i-1].rs1);
8050 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8051 branch_regs[i-1].isconst=0;
8052 branch_regs[i-1].wasconst=0;
8053 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8054 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8057 // Alloc the delay slot in case the branch is taken
8058 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
8060 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8061 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
8062 alloc_cc(&branch_regs[i-1],i);
8063 dirty_reg(&branch_regs[i-1],CCREG);
8064 delayslot_alloc(&branch_regs[i-1],i);
8065 branch_regs[i-1].isconst=0;
8066 alloc_reg(¤t,i,CCREG); // Not taken path
8067 dirty_reg(¤t,CCREG);
8068 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8070 // FIXME: BLTZAL/BGEZAL
8071 if(dops[i-1].opcode2&0x10) { // BxxZAL
8072 alloc_reg(&branch_regs[i-1],i-1,31);
8073 dirty_reg(&branch_regs[i-1],31);
8078 if (dops[i-1].is_ujump)
8080 if(dops[i-1].rt1==31) // JAL/JALR
8082 // Subroutine call will return here, don't alloc any registers
8084 clear_all_regs(current.regmap);
8085 alloc_reg(¤t,i,CCREG);
8086 dirty_reg(¤t,CCREG);
8090 // Internal branch will jump here, match registers to caller
8092 clear_all_regs(current.regmap);
8093 alloc_reg(¤t,i,CCREG);
8094 dirty_reg(¤t,CCREG);
8097 if(ba[j]==start+i*4+4) {
8098 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
8099 current.dirty=branch_regs[j].dirty;
8104 if(ba[j]==start+i*4+4) {
8105 for(hr=0;hr<HOST_REGS;hr++) {
8106 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
8107 current.regmap[hr]=-1;
8109 current.dirty&=branch_regs[j].dirty;
8118 // Count cycles in between branches
8120 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
8124 #if !defined(DRC_DBG)
8125 else if(dops[i].itype==C2OP&>e_cycletab[source[i]&0x3f]>2)
8127 // this should really be removed since the real stalls have been implemented,
8128 // but doing so causes sizeable perf regression against the older version
8129 u_int gtec = gte_cycletab[source[i] & 0x3f];
8130 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
8132 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
8136 else if(dops[i].itype==C2LS)
8138 // same as with C2OP
8139 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
8147 if(!dops[i].is_ds) {
8148 regs[i].dirty=current.dirty;
8149 regs[i].isconst=current.isconst;
8150 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
8152 for(hr=0;hr<HOST_REGS;hr++) {
8153 if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
8154 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
8155 regs[i].wasconst&=~(1<<hr);
8159 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
8160 regs[i].waswritten=current.waswritten;
8163 /* Pass 4 - Cull unused host registers */
8167 for (i=slen-1;i>=0;i--)
8172 if(ba[i]<start || ba[i]>=(start+slen*4))
8174 // Branch out of this block, don't need anything
8180 // Need whatever matches the target
8182 int t=(ba[i]-start)>>2;
8183 for(hr=0;hr<HOST_REGS;hr++)
8185 if(regs[i].regmap_entry[hr]>=0) {
8186 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
8190 // Conditional branch may need registers for following instructions
8191 if (!dops[i].is_ujump)
8194 nr|=needed_reg[i+2];
8195 for(hr=0;hr<HOST_REGS;hr++)
8197 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
8198 //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]);
8202 // Don't need stuff which is overwritten
8203 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8204 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8205 // Merge in delay slot
8206 for(hr=0;hr<HOST_REGS;hr++)
8208 if(dops[i+1].rt1&&dops[i+1].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8209 if(dops[i+1].rt2&&dops[i+1].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8210 if(dops[i+1].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8211 if(dops[i+1].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8212 if(dops[i+1].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8213 if(dops[i+1].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8214 if(dops[i+1].itype==STORE || dops[i+1].itype==STORELR || (dops[i+1].opcode&0x3b)==0x39 || (dops[i+1].opcode&0x3b)==0x3a) {
8215 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8216 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8220 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8222 // SYSCALL instruction (software interrupt)
8225 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8227 // ERET instruction (return from interrupt)
8233 for(hr=0;hr<HOST_REGS;hr++) {
8234 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8235 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8236 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8237 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8241 for(hr=0;hr<HOST_REGS;hr++)
8243 // Overwritten registers are not needed
8244 if(dops[i].rt1&&dops[i].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8245 if(dops[i].rt2&&dops[i].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8246 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8247 // Source registers are needed
8248 if(dops[i].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8249 if(dops[i].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8250 if(dops[i].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8251 if(dops[i].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8252 if(dops[i].itype==STORE || dops[i].itype==STORELR || (dops[i].opcode&0x3b)==0x39 || (dops[i].opcode&0x3b)==0x3a) {
8253 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8254 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8256 // Don't store a register immediately after writing it,
8257 // may prevent dual-issue.
8258 // But do so if this is a branch target, otherwise we
8259 // might have to load the register before the branch.
8260 if(i>0&&!dops[i].bt&&((regs[i].wasdirty>>hr)&1)) {
8261 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
8262 if(dops[i-1].rt1==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8263 if(dops[i-1].rt2==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8265 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
8266 if(dops[i-1].rt1==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8267 if(dops[i-1].rt2==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8271 // Cycle count is needed at branches. Assume it is needed at the target too.
8272 if(i==0||dops[i].bt||dops[i].itype==CJUMP||dops[i].itype==SPAN) {
8273 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8274 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8279 // Deallocate unneeded registers
8280 for(hr=0;hr<HOST_REGS;hr++)
8283 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8287 if(dops[i+1].itype==STORE || dops[i+1].itype==STORELR ||
8288 (dops[i+1].opcode&0x3b)==0x39 || (dops[i+1].opcode&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
8291 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR ||
8292 dops[i+1].itype==C1LS || dops[i+1].itype==C2LS)
8294 if((regs[i].regmap[hr]&63)!=dops[i].rs1 && (regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8295 (regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8296 (regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8297 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8298 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
8299 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8300 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8301 regs[i].regmap[hr]!=map )
8303 regs[i].regmap[hr]=-1;
8304 regs[i].isconst&=~(1<<hr);
8305 if((branch_regs[i].regmap[hr]&63)!=dops[i].rs1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8306 (branch_regs[i].regmap[hr]&63)!=dops[i].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8307 (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8308 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8309 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8310 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8311 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8312 branch_regs[i].regmap[hr]!=map)
8314 branch_regs[i].regmap[hr]=-1;
8315 branch_regs[i].regmap_entry[hr]=-1;
8316 if (!dops[i].is_ujump)
8319 regmap_pre[i+2][hr]=-1;
8320 regs[i+2].wasconst&=~(1<<hr);
8332 if(dops[i].itype==STORE || dops[i].itype==STORELR ||
8333 (dops[i].opcode&0x3b)==0x39 || (dops[i].opcode&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
8336 if(dops[i].itype==LOADLR || dops[i].itype==STORELR ||
8337 dops[i].itype==C1LS || dops[i].itype==C2LS)
8339 if((regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8340 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8341 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map &&
8342 (dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG))
8344 if(i<slen-1&&!dops[i].is_ds) {
8345 assert(regs[i].regmap[hr]<64);
8346 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8347 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8349 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8350 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8352 regmap_pre[i+1][hr]=-1;
8353 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8354 regs[i+1].wasconst&=~(1<<hr);
8356 regs[i].regmap[hr]=-1;
8357 regs[i].isconst&=~(1<<hr);
8365 /* Pass 5 - Pre-allocate registers */
8367 // If a register is allocated during a loop, try to allocate it for the
8368 // entire loop, if possible. This avoids loading/storing registers
8369 // inside of the loop.
8371 signed char f_regmap[HOST_REGS];
8372 clear_all_regs(f_regmap);
8373 for(i=0;i<slen-1;i++)
8375 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8377 if(ba[i]>=start && ba[i]<(start+i*4))
8378 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8379 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8380 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8381 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8382 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8384 int t=(ba[i]-start)>>2;
8385 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8386 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8387 for(hr=0;hr<HOST_REGS;hr++)
8389 if(regs[i].regmap[hr]>=0) {
8390 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8391 // dealloc old register
8393 for(n=0;n<HOST_REGS;n++)
8395 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8397 // and alloc new one
8398 f_regmap[hr]=regs[i].regmap[hr];
8401 if(branch_regs[i].regmap[hr]>=0) {
8402 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8403 // dealloc old register
8405 for(n=0;n<HOST_REGS;n++)
8407 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8409 // and alloc new one
8410 f_regmap[hr]=branch_regs[i].regmap[hr];
8414 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8415 f_regmap[hr]=branch_regs[i].regmap[hr];
8417 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8418 f_regmap[hr]=branch_regs[i].regmap[hr];
8420 // Avoid dirty->clean transition
8421 #ifdef DESTRUCTIVE_WRITEBACK
8422 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;
8424 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8425 // case above, however it's always a good idea. We can't hoist the
8426 // load if the register was already allocated, so there's no point
8427 // wasting time analyzing most of these cases. It only "succeeds"
8428 // when the mapping was different and the load can be replaced with
8429 // a mov, which is of negligible benefit. So such cases are
8431 if(f_regmap[hr]>0) {
8432 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8436 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8437 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8439 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
8440 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8442 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8443 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8445 if(get_reg(regs[i].regmap,r&63)<0) break;
8446 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
8449 while(k>1&®s[k-1].regmap[hr]==-1) {
8450 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8451 //printf("no free regs for store %x\n",start+(k-1)*4);
8454 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8455 //printf("no-match due to different register\n");
8458 if (dops[k-2].is_jump) {
8459 //printf("no-match due to branch\n");
8462 // call/ret fast path assumes no registers allocated
8463 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8469 if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
8470 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8472 regs[k].regmap_entry[hr]=f_regmap[hr];
8473 regs[k].regmap[hr]=f_regmap[hr];
8474 regmap_pre[k+1][hr]=f_regmap[hr];
8475 regs[k].wasdirty&=~(1<<hr);
8476 regs[k].dirty&=~(1<<hr);
8477 regs[k].wasdirty|=(1<<hr)®s[k-1].dirty;
8478 regs[k].dirty|=(1<<hr)®s[k].wasdirty;
8479 regs[k].wasconst&=~(1<<hr);
8480 regs[k].isconst&=~(1<<hr);
8485 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8488 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8489 if(regs[i-1].regmap[hr]==f_regmap[hr]&®map_pre[i][hr]==f_regmap[hr]) {
8490 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8491 regs[i].regmap_entry[hr]=f_regmap[hr];
8492 regs[i].regmap[hr]=f_regmap[hr];
8493 regs[i].wasdirty&=~(1<<hr);
8494 regs[i].dirty&=~(1<<hr);
8495 regs[i].wasdirty|=(1<<hr)®s[i-1].dirty;
8496 regs[i].dirty|=(1<<hr)®s[i-1].dirty;
8497 regs[i].wasconst&=~(1<<hr);
8498 regs[i].isconst&=~(1<<hr);
8499 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8500 branch_regs[i].wasdirty&=~(1<<hr);
8501 branch_regs[i].wasdirty|=(1<<hr)®s[i].dirty;
8502 branch_regs[i].regmap[hr]=f_regmap[hr];
8503 branch_regs[i].dirty&=~(1<<hr);
8504 branch_regs[i].dirty|=(1<<hr)®s[i].dirty;
8505 branch_regs[i].wasconst&=~(1<<hr);
8506 branch_regs[i].isconst&=~(1<<hr);
8507 if (!dops[i].is_ujump) {
8508 regmap_pre[i+2][hr]=f_regmap[hr];
8509 regs[i+2].wasdirty&=~(1<<hr);
8510 regs[i+2].wasdirty|=(1<<hr)®s[i].dirty;
8515 // Alloc register clean at beginning of loop,
8516 // but may dirty it in pass 6
8517 regs[k].regmap_entry[hr]=f_regmap[hr];
8518 regs[k].regmap[hr]=f_regmap[hr];
8519 regs[k].dirty&=~(1<<hr);
8520 regs[k].wasconst&=~(1<<hr);
8521 regs[k].isconst&=~(1<<hr);
8522 if (dops[k].is_jump) {
8523 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8524 branch_regs[k].regmap[hr]=f_regmap[hr];
8525 branch_regs[k].dirty&=~(1<<hr);
8526 branch_regs[k].wasconst&=~(1<<hr);
8527 branch_regs[k].isconst&=~(1<<hr);
8528 if (!dops[k].is_ujump) {
8529 regmap_pre[k+2][hr]=f_regmap[hr];
8530 regs[k+2].wasdirty&=~(1<<hr);
8535 regmap_pre[k+1][hr]=f_regmap[hr];
8536 regs[k+1].wasdirty&=~(1<<hr);
8539 if(regs[j].regmap[hr]==f_regmap[hr])
8540 regs[j].regmap_entry[hr]=f_regmap[hr];
8544 if(regs[j].regmap[hr]>=0)
8546 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8547 //printf("no-match due to different register\n");
8550 if (dops[j].is_ujump)
8552 // Stop on unconditional branch
8555 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8558 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8561 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8564 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8565 //printf("no-match due to different register (branch)\n");
8569 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8570 //printf("No free regs for store %x\n",start+j*4);
8573 assert(f_regmap[hr]<64);
8580 // Non branch or undetermined branch target
8581 for(hr=0;hr<HOST_REGS;hr++)
8583 if(hr!=EXCLUDE_REG) {
8584 if(regs[i].regmap[hr]>=0) {
8585 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8586 // dealloc old register
8588 for(n=0;n<HOST_REGS;n++)
8590 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8592 // and alloc new one
8593 f_regmap[hr]=regs[i].regmap[hr];
8598 // Try to restore cycle count at branch targets
8600 for(j=i;j<slen-1;j++) {
8601 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8602 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8603 //printf("no free regs for store %x\n",start+j*4);
8607 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8609 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8611 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8612 regs[k].regmap[HOST_CCREG]=CCREG;
8613 regmap_pre[k+1][HOST_CCREG]=CCREG;
8614 regs[k+1].wasdirty|=1<<HOST_CCREG;
8615 regs[k].dirty|=1<<HOST_CCREG;
8616 regs[k].wasconst&=~(1<<HOST_CCREG);
8617 regs[k].isconst&=~(1<<HOST_CCREG);
8620 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8622 // Work backwards from the branch target
8623 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8625 //printf("Extend backwards\n");
8628 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8629 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8630 //printf("no free regs for store %x\n",start+(k-1)*4);
8635 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8636 //printf("Extend CC, %x ->\n",start+k*4);
8638 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8639 regs[k].regmap[HOST_CCREG]=CCREG;
8640 regmap_pre[k+1][HOST_CCREG]=CCREG;
8641 regs[k+1].wasdirty|=1<<HOST_CCREG;
8642 regs[k].dirty|=1<<HOST_CCREG;
8643 regs[k].wasconst&=~(1<<HOST_CCREG);
8644 regs[k].isconst&=~(1<<HOST_CCREG);
8649 //printf("Fail Extend CC, %x ->\n",start+k*4);
8653 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8654 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8655 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8657 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8662 // This allocates registers (if possible) one instruction prior
8663 // to use, which can avoid a load-use penalty on certain CPUs.
8664 for(i=0;i<slen-1;i++)
8666 if (!i || !dops[i-1].is_jump)
8670 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8671 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8674 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8676 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8678 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8679 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8680 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8681 regs[i].isconst&=~(1<<hr);
8682 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8683 constmap[i][hr]=constmap[i+1][hr];
8684 regs[i+1].wasdirty&=~(1<<hr);
8685 regs[i].dirty&=~(1<<hr);
8690 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8692 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8694 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8695 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8696 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8697 regs[i].isconst&=~(1<<hr);
8698 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8699 constmap[i][hr]=constmap[i+1][hr];
8700 regs[i+1].wasdirty&=~(1<<hr);
8701 regs[i].dirty&=~(1<<hr);
8705 // Preload target address for load instruction (non-constant)
8706 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8707 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8709 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8711 regs[i].regmap[hr]=dops[i+1].rs1;
8712 regmap_pre[i+1][hr]=dops[i+1].rs1;
8713 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8714 regs[i].isconst&=~(1<<hr);
8715 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8716 constmap[i][hr]=constmap[i+1][hr];
8717 regs[i+1].wasdirty&=~(1<<hr);
8718 regs[i].dirty&=~(1<<hr);
8722 // Load source into target register
8723 if(dops[i+1].lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8724 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8726 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8728 regs[i].regmap[hr]=dops[i+1].rs1;
8729 regmap_pre[i+1][hr]=dops[i+1].rs1;
8730 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8731 regs[i].isconst&=~(1<<hr);
8732 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8733 constmap[i][hr]=constmap[i+1][hr];
8734 regs[i+1].wasdirty&=~(1<<hr);
8735 regs[i].dirty&=~(1<<hr);
8739 // Address for store instruction (non-constant)
8740 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8741 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8742 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8743 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8744 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8745 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
8747 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8749 regs[i].regmap[hr]=dops[i+1].rs1;
8750 regmap_pre[i+1][hr]=dops[i+1].rs1;
8751 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8752 regs[i].isconst&=~(1<<hr);
8753 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8754 constmap[i][hr]=constmap[i+1][hr];
8755 regs[i+1].wasdirty&=~(1<<hr);
8756 regs[i].dirty&=~(1<<hr);
8760 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8761 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8763 hr=get_reg(regs[i+1].regmap,FTEMP);
8765 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8767 regs[i].regmap[hr]=dops[i+1].rs1;
8768 regmap_pre[i+1][hr]=dops[i+1].rs1;
8769 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8770 regs[i].isconst&=~(1<<hr);
8771 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8772 constmap[i][hr]=constmap[i+1][hr];
8773 regs[i+1].wasdirty&=~(1<<hr);
8774 regs[i].dirty&=~(1<<hr);
8776 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8778 // move it to another register
8779 regs[i+1].regmap[hr]=-1;
8780 regmap_pre[i+2][hr]=-1;
8781 regs[i+1].regmap[nr]=FTEMP;
8782 regmap_pre[i+2][nr]=FTEMP;
8783 regs[i].regmap[nr]=dops[i+1].rs1;
8784 regmap_pre[i+1][nr]=dops[i+1].rs1;
8785 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8786 regs[i].isconst&=~(1<<nr);
8787 regs[i+1].isconst&=~(1<<nr);
8788 regs[i].dirty&=~(1<<nr);
8789 regs[i+1].wasdirty&=~(1<<nr);
8790 regs[i+1].dirty&=~(1<<nr);
8791 regs[i+2].wasdirty&=~(1<<nr);
8795 if(dops[i+1].itype==LOAD||dops[i+1].itype==LOADLR||dops[i+1].itype==STORE||dops[i+1].itype==STORELR/*||dops[i+1].itype==C1LS||||dops[i+1].itype==C2LS*/) {
8796 if(dops[i+1].itype==LOAD)
8797 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8798 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8799 hr=get_reg(regs[i+1].regmap,FTEMP);
8800 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
8801 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8802 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8804 if(hr>=0&®s[i].regmap[hr]<0) {
8805 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8806 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8807 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8808 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8809 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8810 regs[i].isconst&=~(1<<hr);
8811 regs[i+1].wasdirty&=~(1<<hr);
8812 regs[i].dirty&=~(1<<hr);
8821 /* Pass 6 - Optimize clean/dirty state */
8822 clean_registers(0,slen-1,1);
8824 /* Pass 7 - Identify 32-bit registers */
8825 for (i=slen-1;i>=0;i--)
8827 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8829 // Conditional branch
8830 if((source[i]>>16)!=0x1000&&i<slen-2) {
8831 // Mark this address as a branch target since it may be called
8832 // upon return from interrupt
8838 if(dops[slen-1].itype==SPAN) {
8839 dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
8843 /* Debug/disassembly */
8848 for(r=1;r<=CCREG;r++) {
8849 if((unneeded_reg[i]>>r)&1) {
8850 if(r==HIREG) printf(" HI");
8851 else if(r==LOREG) printf(" LO");
8852 else printf(" r%d",r);
8856 #if defined(__i386__) || defined(__x86_64__)
8857 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]);
8860 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]);
8862 #if defined(__i386__) || defined(__x86_64__)
8864 if(needed_reg[i]&1) printf("eax ");
8865 if((needed_reg[i]>>1)&1) printf("ecx ");
8866 if((needed_reg[i]>>2)&1) printf("edx ");
8867 if((needed_reg[i]>>3)&1) printf("ebx ");
8868 if((needed_reg[i]>>5)&1) printf("ebp ");
8869 if((needed_reg[i]>>6)&1) printf("esi ");
8870 if((needed_reg[i]>>7)&1) printf("edi ");
8872 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]);
8874 if(regs[i].wasdirty&1) printf("eax ");
8875 if((regs[i].wasdirty>>1)&1) printf("ecx ");
8876 if((regs[i].wasdirty>>2)&1) printf("edx ");
8877 if((regs[i].wasdirty>>3)&1) printf("ebx ");
8878 if((regs[i].wasdirty>>5)&1) printf("ebp ");
8879 if((regs[i].wasdirty>>6)&1) printf("esi ");
8880 if((regs[i].wasdirty>>7)&1) printf("edi ");
8883 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]);
8885 if(regs[i].wasdirty&1) printf("r0 ");
8886 if((regs[i].wasdirty>>1)&1) printf("r1 ");
8887 if((regs[i].wasdirty>>2)&1) printf("r2 ");
8888 if((regs[i].wasdirty>>3)&1) printf("r3 ");
8889 if((regs[i].wasdirty>>4)&1) printf("r4 ");
8890 if((regs[i].wasdirty>>5)&1) printf("r5 ");
8891 if((regs[i].wasdirty>>6)&1) printf("r6 ");
8892 if((regs[i].wasdirty>>7)&1) printf("r7 ");
8893 if((regs[i].wasdirty>>8)&1) printf("r8 ");
8894 if((regs[i].wasdirty>>9)&1) printf("r9 ");
8895 if((regs[i].wasdirty>>10)&1) printf("r10 ");
8896 if((regs[i].wasdirty>>12)&1) printf("r12 ");
8899 disassemble_inst(i);
8900 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
8901 #if defined(__i386__) || defined(__x86_64__)
8902 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]);
8903 if(regs[i].dirty&1) printf("eax ");
8904 if((regs[i].dirty>>1)&1) printf("ecx ");
8905 if((regs[i].dirty>>2)&1) printf("edx ");
8906 if((regs[i].dirty>>3)&1) printf("ebx ");
8907 if((regs[i].dirty>>5)&1) printf("ebp ");
8908 if((regs[i].dirty>>6)&1) printf("esi ");
8909 if((regs[i].dirty>>7)&1) printf("edi ");
8912 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]);
8913 if(regs[i].dirty&1) printf("r0 ");
8914 if((regs[i].dirty>>1)&1) printf("r1 ");
8915 if((regs[i].dirty>>2)&1) printf("r2 ");
8916 if((regs[i].dirty>>3)&1) printf("r3 ");
8917 if((regs[i].dirty>>4)&1) printf("r4 ");
8918 if((regs[i].dirty>>5)&1) printf("r5 ");
8919 if((regs[i].dirty>>6)&1) printf("r6 ");
8920 if((regs[i].dirty>>7)&1) printf("r7 ");
8921 if((regs[i].dirty>>8)&1) printf("r8 ");
8922 if((regs[i].dirty>>9)&1) printf("r9 ");
8923 if((regs[i].dirty>>10)&1) printf("r10 ");
8924 if((regs[i].dirty>>12)&1) printf("r12 ");
8927 if(regs[i].isconst) {
8928 printf("constants: ");
8929 #if defined(__i386__) || defined(__x86_64__)
8930 if(regs[i].isconst&1) printf("eax=%x ",(u_int)constmap[i][0]);
8931 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(u_int)constmap[i][1]);
8932 if((regs[i].isconst>>2)&1) printf("edx=%x ",(u_int)constmap[i][2]);
8933 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(u_int)constmap[i][3]);
8934 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(u_int)constmap[i][5]);
8935 if((regs[i].isconst>>6)&1) printf("esi=%x ",(u_int)constmap[i][6]);
8936 if((regs[i].isconst>>7)&1) printf("edi=%x ",(u_int)constmap[i][7]);
8938 #if defined(__arm__) || defined(__aarch64__)
8940 for (r = 0; r < ARRAY_SIZE(constmap[i]); r++)
8941 if ((regs[i].isconst >> r) & 1)
8942 printf(" r%d=%x", r, (u_int)constmap[i][r]);
8946 if(dops[i].is_jump) {
8947 #if defined(__i386__) || defined(__x86_64__)
8948 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]);
8949 if(branch_regs[i].dirty&1) printf("eax ");
8950 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
8951 if((branch_regs[i].dirty>>2)&1) printf("edx ");
8952 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
8953 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
8954 if((branch_regs[i].dirty>>6)&1) printf("esi ");
8955 if((branch_regs[i].dirty>>7)&1) printf("edi ");
8958 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]);
8959 if(branch_regs[i].dirty&1) printf("r0 ");
8960 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
8961 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
8962 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
8963 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
8964 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
8965 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
8966 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
8967 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
8968 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
8969 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
8970 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
8976 /* Pass 8 - Assembly */
8977 linkcount=0;stubcount=0;
8978 ds=0;is_delayslot=0;
8980 void *beginning=start_block();
8985 void *instr_addr0_override = NULL;
8987 if (start == 0x80030000) {
8988 // nasty hack for the fastbios thing
8989 // override block entry to this code
8990 instr_addr0_override = out;
8991 emit_movimm(start,0);
8992 // abuse io address var as a flag that we
8993 // have already returned here once
8994 emit_readword(&address,1);
8995 emit_writeword(0,&pcaddr);
8996 emit_writeword(0,&address);
8999 emit_jeq(out + 4*2);
9000 emit_far_jump(new_dyna_leave);
9002 emit_jne(new_dyna_leave);
9007 //if(ds) printf("ds: ");
9008 disassemble_inst(i);
9010 ds=0; // Skip delay slot
9011 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9012 instr_addr[i] = NULL;
9014 speculate_register_values(i);
9015 #ifndef DESTRUCTIVE_WRITEBACK
9016 if (i < 2 || !dops[i-2].is_ujump)
9018 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9020 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9021 dirty_pre=branch_regs[i].dirty;
9023 dirty_pre=regs[i].dirty;
9027 if (i < 2 || !dops[i-2].is_ujump)
9029 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9030 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9032 // branch target entry point
9033 instr_addr[i] = out;
9034 assem_debug("<->\n");
9035 drc_dbg_emit_do_cmp(i);
9038 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
9039 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9040 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9041 address_generation(i,®s[i],regs[i].regmap_entry);
9042 load_consts(regmap_pre[i],regs[i].regmap,i);
9045 // Load the delay slot registers if necessary
9046 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2&&(dops[i+1].rs1!=dops[i].rt1||dops[i].rt1==0))
9047 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9048 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2&&(dops[i+1].rs2!=dops[i].rt1||dops[i].rt1==0))
9049 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9050 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a)
9051 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9055 // Preload registers for following instruction
9056 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9057 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9058 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9059 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9060 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9061 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9063 // TODO: if(is_ooo(i)) address_generation(i+1);
9064 if(dops[i].itype==CJUMP)
9065 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
9066 if(dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].opcode&0x3b)==0x39||(dops[i].opcode&0x3b)==0x3a)
9067 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9069 switch(dops[i].itype) {
9071 alu_assemble(i,®s[i]);break;
9073 imm16_assemble(i,®s[i]);break;
9075 shift_assemble(i,®s[i]);break;
9077 shiftimm_assemble(i,®s[i]);break;
9079 load_assemble(i,®s[i]);break;
9081 loadlr_assemble(i,®s[i]);break;
9083 store_assemble(i,®s[i]);break;
9085 storelr_assemble(i,®s[i]);break;
9087 cop0_assemble(i,®s[i]);break;
9089 cop1_assemble(i,®s[i]);break;
9091 c1ls_assemble(i,®s[i]);break;
9093 cop2_assemble(i,®s[i]);break;
9095 c2ls_assemble(i,®s[i]);break;
9097 c2op_assemble(i,®s[i]);break;
9099 multdiv_assemble(i,®s[i]);
9100 multdiv_prepare_stall(i,®s[i]);
9103 mov_assemble(i,®s[i]);break;
9105 syscall_assemble(i,®s[i]);break;
9107 hlecall_assemble(i,®s[i]);break;
9109 intcall_assemble(i,®s[i]);break;
9111 ujump_assemble(i,®s[i]);ds=1;break;
9113 rjump_assemble(i,®s[i]);ds=1;break;
9115 cjump_assemble(i,®s[i]);ds=1;break;
9117 sjump_assemble(i,®s[i]);ds=1;break;
9119 pagespan_assemble(i,®s[i]);break;
9121 if (dops[i].is_ujump)
9124 literal_pool_jumpover(256);
9129 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9130 // no ending needed for this block since INTCALL never returns
9132 // If the block did not end with an unconditional branch,
9133 // add a jump to the next instruction.
9135 if (!dops[i-2].is_ujump && dops[i-1].itype != SPAN) {
9136 assert(!dops[i-1].is_jump);
9138 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9139 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9140 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9141 emit_loadreg(CCREG,HOST_CCREG);
9142 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
9146 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9147 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9149 add_to_linker(out,start+i*4,0);
9156 assert(!dops[i-1].is_jump);
9157 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9158 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9159 emit_loadreg(CCREG,HOST_CCREG);
9160 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
9161 add_to_linker(out,start+i*4,0);
9165 // TODO: delay slot stubs?
9167 for(i=0;i<stubcount;i++)
9169 switch(stubs[i].type)
9177 do_readstub(i);break;
9182 do_writestub(i);break;
9186 do_invstub(i);break;
9188 do_cop1stub(i);break;
9190 do_unalignedwritestub(i);break;
9194 if (instr_addr0_override)
9195 instr_addr[0] = instr_addr0_override;
9197 /* Pass 9 - Linker */
9198 for(i=0;i<linkcount;i++)
9200 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9202 if (!link_addr[i].ext)
9205 void *addr = check_addr(link_addr[i].target);
9206 emit_extjump(link_addr[i].addr, link_addr[i].target);
9208 set_jump_target(link_addr[i].addr, addr);
9209 add_jump_out(link_addr[i].target,stub);
9212 set_jump_target(link_addr[i].addr, stub);
9217 int target=(link_addr[i].target-start)>>2;
9218 assert(target>=0&&target<slen);
9219 assert(instr_addr[target]);
9220 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9221 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9223 set_jump_target(link_addr[i].addr, instr_addr[target]);
9228 u_int source_len = slen*4;
9229 if (dops[slen-1].itype == INTCALL && source_len > 4)
9230 // no need to treat the last instruction as compiled
9231 // as interpreter fully handles it
9234 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9237 // External Branch Targets (jump_in)
9240 if(dops[i].bt||i==0)
9242 if(instr_addr[i]) // TODO - delay slots (=null)
9244 u_int vaddr=start+i*4;
9245 u_int page=get_page(vaddr);
9246 u_int vpage=get_vpage(vaddr);
9249 assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
9250 assem_debug("jump_in: %x\n",start+i*4);
9251 ll_add(jump_dirty+vpage,vaddr,out);
9252 void *entry_point = do_dirty_stub(i, source_len);
9253 ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
9254 // If there was an existing entry in the hash table,
9255 // replace it with the new address.
9256 // Don't add new entries. We'll insert the
9257 // ones that actually get used in check_addr().
9258 struct ht_entry *ht_bin = hash_table_get(vaddr);
9259 if (ht_bin->vaddr[0] == vaddr)
9260 ht_bin->tcaddr[0] = entry_point;
9261 if (ht_bin->vaddr[1] == vaddr)
9262 ht_bin->tcaddr[1] = entry_point;
9267 // Write out the literal pool if necessary
9269 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9271 if(((u_int)out)&7) emit_addnop(13);
9273 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9274 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9275 memcpy(copy, source, source_len);
9278 end_block(beginning);
9280 // If we're within 256K of the end of the buffer,
9281 // start over from the beginning. (Is 256K enough?)
9282 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9283 out = ndrc->translation_cache;
9285 // Trap writes to any of the pages we compiled
9286 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
9289 inv_code_start=inv_code_end=~0;
9291 // for PCSX we need to mark all mirrors too
9292 if(get_page(start)<(RAM_SIZE>>12))
9293 for(i=start>>12;i<=(start+slen*4)>>12;i++)
9294 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
9295 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
9296 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
9298 /* Pass 10 - Free memory by expiring oldest blocks */
9300 int end=(((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
9303 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
9304 uintptr_t base_offs = ((uintptr_t)(expirep >> 13) << shift); // Base offset of this block
9305 uintptr_t base_offs_s = base_offs >> shift;
9306 inv_debug("EXP: Phase %d\n",expirep);
9307 switch((expirep>>11)&3)
9310 // Clear jump_in and jump_dirty
9311 ll_remove_matching_addrs(jump_in+(expirep&2047),base_offs_s,shift);
9312 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base_offs_s,shift);
9313 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base_offs_s,shift);
9314 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base_offs_s,shift);
9318 ll_kill_pointers(jump_out[expirep&2047],base_offs_s,shift);
9319 ll_kill_pointers(jump_out[(expirep&2047)+2048],base_offs_s,shift);
9324 struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
9325 uintptr_t o1 = (u_char *)ht_bin->tcaddr[1] - ndrc->translation_cache;
9326 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9327 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9328 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
9329 ht_bin->vaddr[1] = -1;
9330 ht_bin->tcaddr[1] = NULL;
9332 o1 = (u_char *)ht_bin->tcaddr[0] - ndrc->translation_cache;
9333 o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9334 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9335 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
9336 ht_bin->vaddr[0] = ht_bin->vaddr[1];
9337 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
9338 ht_bin->vaddr[1] = -1;
9339 ht_bin->tcaddr[1] = NULL;
9345 if((expirep&2047)==0)
9347 ll_remove_matching_addrs(jump_out+(expirep&2047),base_offs_s,shift);
9348 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base_offs_s,shift);
9351 expirep=(expirep+1)&65535;
9356 // vim:shiftwidth=2:expandtab