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]; // pre-insn + loop preloaded 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
185 struct ht_entry hash_table[65536] __attribute__((aligned(16)));
186 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
187 struct ll_entry *jump_dirty[4096];
189 static struct ll_entry *jump_out[4096];
191 static u_int *source;
192 static char insn[MAXBLOCK][10];
193 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
194 static uint64_t gte_rt[MAXBLOCK];
195 static uint64_t gte_unneeded[MAXBLOCK];
196 static u_int smrv[32]; // speculated MIPS register values
197 static u_int smrv_strong; // mask or regs that are likely to have correct values
198 static u_int smrv_weak; // same, but somewhat less likely
199 static u_int smrv_strong_next; // same, but after current insn executes
200 static u_int smrv_weak_next;
201 static int imm[MAXBLOCK];
202 static u_int ba[MAXBLOCK];
203 static uint64_t unneeded_reg[MAXBLOCK];
204 static uint64_t branch_unneeded_reg[MAXBLOCK];
205 // pre-instruction [i], excluding loop-preload regs?
206 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
207 // contains 'real' consts at [i] insn, but may differ from what's actually
208 // loaded in host reg as 'final' value is always loaded, see get_final_value()
209 static uint32_t current_constmap[HOST_REGS];
210 static uint32_t constmap[MAXBLOCK][HOST_REGS];
211 static struct regstat regs[MAXBLOCK];
212 static struct regstat branch_regs[MAXBLOCK];
213 static signed char minimum_free_regs[MAXBLOCK];
214 static u_int needed_reg[MAXBLOCK];
215 static u_int wont_dirty[MAXBLOCK];
216 static u_int will_dirty[MAXBLOCK];
217 static int ccadj[MAXBLOCK];
219 static void *instr_addr[MAXBLOCK];
220 static struct link_entry link_addr[MAXBLOCK];
221 static int linkcount;
222 static struct code_stub stubs[MAXBLOCK*3];
223 static int stubcount;
224 static u_int literals[1024][2];
225 static int literalcount;
226 static int is_delayslot;
227 static char shadow[1048576] __attribute__((aligned(16)));
230 static u_int stop_after_jal;
231 static u_int f1_hack; // 0 - off, ~0 - capture address, else addr
233 int new_dynarec_hacks;
234 int new_dynarec_hacks_pergame;
235 int new_dynarec_hacks_old;
236 int new_dynarec_did_compile;
238 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
240 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
241 extern int last_count; // last absolute target, often = next_interupt
243 extern int pending_exception;
244 extern int branch_target;
245 extern uintptr_t ram_offset;
246 extern uintptr_t mini_ht[32][2];
247 extern u_char restore_candidate[512];
249 /* registers that may be allocated */
251 #define LOREG 32 // lo
252 #define HIREG 33 // hi
253 //#define FSREG 34 // FPU status (FCSR)
254 #define CSREG 35 // Coprocessor status
255 #define CCREG 36 // Cycle count
256 #define INVCP 37 // Pointer to invalid_code
257 //#define MMREG 38 // Pointer to memory_map
258 #define ROREG 39 // ram offset (if rdram!=0x80000000)
260 #define FTEMP 40 // FPU temporary register
261 #define PTEMP 41 // Prefetch temporary register
262 //#define TLREG 42 // TLB mapping offset
263 #define RHASH 43 // Return address hash
264 #define RHTBL 44 // Return address hash table address
265 #define RTEMP 45 // JR/JALR address register
267 #define AGEN1 46 // Address generation temporary register
268 //#define AGEN2 47 // Address generation temporary register
269 //#define MGEN1 48 // Maptable address generation temporary register
270 //#define MGEN2 49 // Maptable address generation temporary register
271 #define BTREG 50 // Branch target temporary register
273 /* instruction types */
274 #define NOP 0 // No operation
275 #define LOAD 1 // Load
276 #define STORE 2 // Store
277 #define LOADLR 3 // Unaligned load
278 #define STORELR 4 // Unaligned store
279 #define MOV 5 // Move
280 #define ALU 6 // Arithmetic/logic
281 #define MULTDIV 7 // Multiply/divide
282 #define SHIFT 8 // Shift by register
283 #define SHIFTIMM 9// Shift by immediate
284 #define IMM16 10 // 16-bit immediate
285 #define RJUMP 11 // Unconditional jump to register
286 #define UJUMP 12 // Unconditional jump
287 #define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
288 #define SJUMP 14 // Conditional branch (regimm format)
289 #define COP0 15 // Coprocessor 0
290 #define COP1 16 // Coprocessor 1
291 #define C1LS 17 // Coprocessor 1 load/store
292 //#define FJUMP 18 // Conditional branch (floating point)
293 //#define FLOAT 19 // Floating point unit
294 //#define FCONV 20 // Convert integer to float
295 //#define FCOMP 21 // Floating point compare (sets FSREG)
296 #define SYSCALL 22// SYSCALL
297 #define OTHER 23 // Other
298 #define SPAN 24 // Branch/delay slot spans 2 pages
299 #define NI 25 // Not implemented
300 #define HLECALL 26// PCSX fake opcodes for HLE
301 #define COP2 27 // Coprocessor 2 move
302 #define C2LS 28 // Coprocessor 2 load/store
303 #define C2OP 29 // Coprocessor 2 operation
304 #define INTCALL 30// Call interpreter to handle rare corner cases
311 #define DJT_1 (void *)1l // no function, just a label in assem_debug log
312 #define DJT_2 (void *)2l
315 int new_recompile_block(u_int addr);
316 void *get_addr_ht(u_int vaddr);
317 void invalidate_block(u_int block);
318 void invalidate_addr(u_int addr);
319 void remove_hash(int vaddr);
321 void dyna_linker_ds();
323 void verify_code_ds();
326 void fp_exception_ds();
327 void jump_to_new_pc();
328 void call_gteStall();
329 void new_dyna_leave();
331 // Needed by assembler
332 static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
333 static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
334 static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
335 static void load_all_regs(const signed char i_regmap[]);
336 static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
337 static void load_regs_entry(int t);
338 static void load_all_consts(const signed char regmap[], u_int dirty, int i);
339 static u_int get_host_reglist(const signed char *regmap);
341 static int verify_dirty(const u_int *ptr);
342 static int get_final_value(int hr, int i, int *value);
343 static void add_stub(enum stub_type type, void *addr, void *retaddr,
344 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
345 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
346 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
347 static void add_to_linker(void *addr, u_int target, int ext);
348 static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
349 int addr, int *offset_reg, 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 = CYCLE_MULT_DEFAULT; // 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 && cycle_multiplier == CYCLE_MULT_DEFAULT
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 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
712 // On some architectures stores need invc_ptr
713 #if defined(HOST_IMM8)
714 if (dops[i+j].is_store)
717 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
725 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
727 // Follow first branch
728 int t=(ba[i+b]-start)>>2;
729 j=7-b;if(t+j>=slen) j=slen-t-1;
732 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
733 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
734 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
735 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
738 // TODO: preferred register based on backward branch
740 // Delay slot should preferably not overwrite branch conditions or cycle count
741 if (i > 0 && dops[i-1].is_jump) {
742 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
743 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
749 // Coprocessor load/store needs FTEMP, even if not declared
750 if(dops[i].itype==C2LS) {
753 // Load L/R also uses FTEMP as a temporary register
754 if(dops[i].itype==LOADLR) {
757 // Also SWL/SWR/SDL/SDR
758 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
761 // Don't remove the miniht registers
762 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
769 // We only want to allocate registers if we're going to use them again soon
770 int needed_again(int r, int i)
776 if (i > 0 && dops[i-1].is_ujump)
778 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
779 return 0; // Don't need any registers if exiting the block
787 if (dops[i+j].is_ujump)
789 // Don't go past an unconditonal jump
793 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
800 if(dops[i+j].rs1==r) rn=j;
801 if(dops[i+j].rs2==r) rn=j;
802 if((unneeded_reg[i+j]>>r)&1) rn=10;
803 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
813 // Try to match register allocations at the end of a loop with those
815 int loop_reg(int i, int r, int hr)
824 if (dops[i+j].is_ujump)
826 // Don't go past an unconditonal jump
833 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
839 if((unneeded_reg[i+k]>>r)&1) return hr;
840 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
842 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
844 int t=(ba[i+k]-start)>>2;
845 int reg=get_reg(regs[t].regmap_entry,r);
846 if(reg>=0) return reg;
847 //reg=get_reg(regs[t+1].regmap_entry,r);
848 //if(reg>=0) return reg;
856 // Allocate every register, preserving source/target regs
857 void alloc_all(struct regstat *cur,int i)
861 for(hr=0;hr<HOST_REGS;hr++) {
862 if(hr!=EXCLUDE_REG) {
863 if(((cur->regmap[hr]&63)!=dops[i].rs1)&&((cur->regmap[hr]&63)!=dops[i].rs2)&&
864 ((cur->regmap[hr]&63)!=dops[i].rt1)&&((cur->regmap[hr]&63)!=dops[i].rt2))
867 cur->dirty&=~(1<<hr);
870 if((cur->regmap[hr]&63)==0)
873 cur->dirty&=~(1<<hr);
880 static int host_tempreg_in_use;
882 static void host_tempreg_acquire(void)
884 assert(!host_tempreg_in_use);
885 host_tempreg_in_use = 1;
888 static void host_tempreg_release(void)
890 host_tempreg_in_use = 0;
893 static void host_tempreg_acquire(void) {}
894 static void host_tempreg_release(void) {}
898 extern void gen_interupt();
899 extern void do_insn_cmp();
900 #define FUNCNAME(f) { f, " " #f }
901 static const struct {
904 } function_names[] = {
905 FUNCNAME(cc_interrupt),
906 FUNCNAME(gen_interupt),
907 FUNCNAME(get_addr_ht),
909 FUNCNAME(jump_handler_read8),
910 FUNCNAME(jump_handler_read16),
911 FUNCNAME(jump_handler_read32),
912 FUNCNAME(jump_handler_write8),
913 FUNCNAME(jump_handler_write16),
914 FUNCNAME(jump_handler_write32),
915 FUNCNAME(invalidate_addr),
916 FUNCNAME(jump_to_new_pc),
917 FUNCNAME(call_gteStall),
918 FUNCNAME(new_dyna_leave),
920 FUNCNAME(pcsx_mtc0_ds),
922 FUNCNAME(do_insn_cmp),
925 FUNCNAME(verify_code),
929 static const char *func_name(const void *a)
932 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
933 if (function_names[i].addr == a)
934 return function_names[i].name;
938 #define func_name(x) ""
942 #include "assem_x86.c"
945 #include "assem_x64.c"
948 #include "assem_arm.c"
951 #include "assem_arm64.c"
954 static void *get_trampoline(const void *f)
958 for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
959 if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
962 if (i == ARRAY_SIZE(ndrc->tramp.f)) {
963 SysPrintf("trampoline table is full, last func %p\n", f);
966 if (ndrc->tramp.f[i] == NULL) {
967 start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
968 ndrc->tramp.f[i] = f;
969 end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
971 return &ndrc->tramp.ops[i];
974 static void emit_far_jump(const void *f)
976 if (can_jump_or_call(f)) {
981 f = get_trampoline(f);
985 static void emit_far_call(const void *f)
987 if (can_jump_or_call(f)) {
992 f = get_trampoline(f);
996 // Add virtual address mapping to linked list
997 void ll_add(struct ll_entry **head,int vaddr,void *addr)
999 struct ll_entry *new_entry;
1000 new_entry=malloc(sizeof(struct ll_entry));
1001 assert(new_entry!=NULL);
1002 new_entry->vaddr=vaddr;
1003 new_entry->reg_sv_flags=0;
1004 new_entry->addr=addr;
1005 new_entry->next=*head;
1009 void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
1011 ll_add(head,vaddr,addr);
1012 (*head)->reg_sv_flags=reg_sv_flags;
1015 // Check if an address is already compiled
1016 // but don't return addresses which are about to expire from the cache
1017 void *check_addr(u_int vaddr)
1019 struct ht_entry *ht_bin = hash_table_get(vaddr);
1021 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1022 if (ht_bin->vaddr[i] == vaddr)
1023 if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
1024 if (isclean(ht_bin->tcaddr[i]))
1025 return ht_bin->tcaddr[i];
1027 u_int page=get_page(vaddr);
1028 struct ll_entry *head;
1030 while (head != NULL) {
1031 if (head->vaddr == vaddr) {
1032 if (doesnt_expire_soon(head->addr)) {
1033 // Update existing entry with current address
1034 if (ht_bin->vaddr[0] == vaddr) {
1035 ht_bin->tcaddr[0] = head->addr;
1038 if (ht_bin->vaddr[1] == vaddr) {
1039 ht_bin->tcaddr[1] = head->addr;
1042 // Insert into hash table with low priority.
1043 // Don't evict existing entries, as they are probably
1044 // addresses that are being accessed frequently.
1045 if (ht_bin->vaddr[0] == -1) {
1046 ht_bin->vaddr[0] = vaddr;
1047 ht_bin->tcaddr[0] = head->addr;
1049 else if (ht_bin->vaddr[1] == -1) {
1050 ht_bin->vaddr[1] = vaddr;
1051 ht_bin->tcaddr[1] = head->addr;
1061 void remove_hash(int vaddr)
1063 //printf("remove hash: %x\n",vaddr);
1064 struct ht_entry *ht_bin = hash_table_get(vaddr);
1065 if (ht_bin->vaddr[1] == vaddr) {
1066 ht_bin->vaddr[1] = -1;
1067 ht_bin->tcaddr[1] = NULL;
1069 if (ht_bin->vaddr[0] == vaddr) {
1070 ht_bin->vaddr[0] = ht_bin->vaddr[1];
1071 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
1072 ht_bin->vaddr[1] = -1;
1073 ht_bin->tcaddr[1] = NULL;
1077 static void ll_remove_matching_addrs(struct ll_entry **head,
1078 uintptr_t base_offs_s, int shift)
1080 struct ll_entry *next;
1082 uintptr_t o1 = (u_char *)(*head)->addr - ndrc->translation_cache;
1083 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1084 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1086 inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
1087 remove_hash((*head)->vaddr);
1094 head=&((*head)->next);
1099 // Remove all entries from linked list
1100 void ll_clear(struct ll_entry **head)
1102 struct ll_entry *cur;
1103 struct ll_entry *next;
1114 // Dereference the pointers and remove if it matches
1115 static void ll_kill_pointers(struct ll_entry *head,
1116 uintptr_t base_offs_s, int shift)
1119 u_char *ptr = get_pointer(head->addr);
1120 uintptr_t o1 = ptr - ndrc->translation_cache;
1121 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1122 inv_debug("EXP: Lookup pointer to %p at %p (%x)\n",ptr,head->addr,head->vaddr);
1123 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1125 inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
1126 void *host_addr=find_extjump_insn(head->addr);
1127 mark_clear_cache(host_addr);
1128 set_jump_target(host_addr, head->addr);
1134 // This is called when we write to a compiled block (see do_invstub)
1135 static void invalidate_page(u_int page)
1137 struct ll_entry *head;
1138 struct ll_entry *next;
1142 inv_debug("INVALIDATE: %x\n",head->vaddr);
1143 remove_hash(head->vaddr);
1148 head=jump_out[page];
1151 inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
1152 void *host_addr=find_extjump_insn(head->addr);
1153 mark_clear_cache(host_addr);
1154 set_jump_target(host_addr, head->addr); // point back to dyna_linker
1161 static void invalidate_block_range(u_int block, u_int first, u_int last)
1163 u_int page=get_page(block<<12);
1164 //printf("first=%d last=%d\n",first,last);
1165 invalidate_page(page);
1166 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1167 assert(last<page+5);
1168 // Invalidate the adjacent pages if a block crosses a 4K boundary
1170 invalidate_page(first);
1173 for(first=page+1;first<last;first++) {
1174 invalidate_page(first);
1178 // Don't trap writes
1179 invalid_code[block]=1;
1182 memset(mini_ht,-1,sizeof(mini_ht));
1186 void invalidate_block(u_int block)
1188 u_int page=get_page(block<<12);
1189 u_int vpage=get_vpage(block<<12);
1190 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1191 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1194 struct ll_entry *head;
1195 head=jump_dirty[vpage];
1196 //printf("page=%d vpage=%d\n",page,vpage);
1198 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1199 u_char *start, *end;
1200 get_bounds(head->addr, &start, &end);
1201 //printf("start: %p end: %p\n", start, end);
1202 if (page < 2048 && start >= rdram && end < rdram+RAM_SIZE) {
1203 if (((start-rdram)>>12) <= page && ((end-1-rdram)>>12) >= page) {
1204 if ((((start-rdram)>>12)&2047) < first) first = ((start-rdram)>>12)&2047;
1205 if ((((end-1-rdram)>>12)&2047) > last) last = ((end-1-rdram)>>12)&2047;
1211 invalidate_block_range(block,first,last);
1214 void invalidate_addr(u_int addr)
1217 // this check is done by the caller
1218 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1219 u_int page=get_vpage(addr);
1220 if(page<2048) { // RAM
1221 struct ll_entry *head;
1222 u_int addr_min=~0, addr_max=0;
1223 u_int mask=RAM_SIZE-1;
1224 u_int addr_main=0x80000000|(addr&mask);
1226 inv_code_start=addr_main&~0xfff;
1227 inv_code_end=addr_main|0xfff;
1230 // must check previous page too because of spans..
1232 inv_code_start-=0x1000;
1234 for(;pg1<=page;pg1++) {
1235 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1236 u_char *start_h, *end_h;
1238 get_bounds(head->addr, &start_h, &end_h);
1239 start = (uintptr_t)start_h - ram_offset;
1240 end = (uintptr_t)end_h - ram_offset;
1241 if(start<=addr_main&&addr_main<end) {
1242 if(start<addr_min) addr_min=start;
1243 if(end>addr_max) addr_max=end;
1245 else if(addr_main<start) {
1246 if(start<inv_code_end)
1247 inv_code_end=start-1;
1250 if(end>inv_code_start)
1256 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1257 inv_code_start=inv_code_end=~0;
1258 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1262 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1263 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1264 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1268 invalidate_block(addr>>12);
1271 // This is called when loading a save state.
1272 // Anything could have changed, so invalidate everything.
1273 void invalidate_all_pages(void)
1276 for(page=0;page<4096;page++)
1277 invalidate_page(page);
1278 for(page=0;page<1048576;page++)
1279 if(!invalid_code[page]) {
1280 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1281 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1284 memset(mini_ht,-1,sizeof(mini_ht));
1289 static void do_invstub(int n)
1292 u_int reglist=stubs[n].a;
1293 set_jump_target(stubs[n].addr, out);
1295 if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
1296 emit_far_call(invalidate_addr);
1297 restore_regs(reglist);
1298 emit_jmp(stubs[n].retaddr); // return address
1301 // Add an entry to jump_out after making a link
1302 // src should point to code by emit_extjump2()
1303 void add_jump_out(u_int vaddr,void *src)
1305 u_int page=get_page(vaddr);
1306 inv_debug("add_jump_out: %p -> %x (%d)\n",src,vaddr,page);
1307 check_extjump2(src);
1308 ll_add(jump_out+page,vaddr,src);
1309 //inv_debug("add_jump_out: to %p\n",get_pointer(src));
1312 // If a code block was found to be unmodified (bit was set in
1313 // restore_candidate) and it remains unmodified (bit is clear
1314 // in invalid_code) then move the entries for that 4K page from
1315 // the dirty list to the clean list.
1316 void clean_blocks(u_int page)
1318 struct ll_entry *head;
1319 inv_debug("INV: clean_blocks page=%d\n",page);
1320 head=jump_dirty[page];
1322 if(!invalid_code[head->vaddr>>12]) {
1323 // Don't restore blocks which are about to expire from the cache
1324 if (doesnt_expire_soon(head->addr)) {
1325 if(verify_dirty(head->addr)) {
1326 u_char *start, *end;
1327 //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
1330 get_bounds(head->addr, &start, &end);
1331 if (start - rdram < RAM_SIZE) {
1332 for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
1333 inv|=invalid_code[i];
1336 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1340 void *clean_addr = get_clean_addr(head->addr);
1341 if (doesnt_expire_soon(clean_addr)) {
1343 inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, clean_addr);
1344 //printf("page=%x, addr=%x\n",page,head->vaddr);
1345 //assert(head->vaddr>>12==(page|0x80000));
1346 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1347 struct ht_entry *ht_bin = hash_table_get(head->vaddr);
1348 if (ht_bin->vaddr[0] == head->vaddr)
1349 ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
1350 if (ht_bin->vaddr[1] == head->vaddr)
1351 ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
1361 /* Register allocation */
1363 // Note: registers are allocated clean (unmodified state)
1364 // if you intend to modify the register, you must call dirty_reg().
1365 static void alloc_reg(struct regstat *cur,int i,signed char reg)
1368 int preferred_reg = PREFERRED_REG_FIRST
1369 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1370 if (reg == CCREG) preferred_reg = HOST_CCREG;
1371 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1372 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1374 // Don't allocate unused registers
1375 if((cur->u>>reg)&1) return;
1377 // see if it's already allocated
1378 for(hr=0;hr<HOST_REGS;hr++)
1380 if(cur->regmap[hr]==reg) return;
1383 // Keep the same mapping if the register was already allocated in a loop
1384 preferred_reg = loop_reg(i,reg,preferred_reg);
1386 // Try to allocate the preferred register
1387 if(cur->regmap[preferred_reg]==-1) {
1388 cur->regmap[preferred_reg]=reg;
1389 cur->dirty&=~(1<<preferred_reg);
1390 cur->isconst&=~(1<<preferred_reg);
1393 r=cur->regmap[preferred_reg];
1396 cur->regmap[preferred_reg]=reg;
1397 cur->dirty&=~(1<<preferred_reg);
1398 cur->isconst&=~(1<<preferred_reg);
1402 // Clear any unneeded registers
1403 // We try to keep the mapping consistent, if possible, because it
1404 // makes branches easier (especially loops). So we try to allocate
1405 // first (see above) before removing old mappings. If this is not
1406 // possible then go ahead and clear out the registers that are no
1408 for(hr=0;hr<HOST_REGS;hr++)
1413 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1417 // Try to allocate any available register, but prefer
1418 // registers that have not been used recently.
1420 for (hr = PREFERRED_REG_FIRST; ; ) {
1421 if (cur->regmap[hr] < 0) {
1422 int oldreg = regs[i-1].regmap[hr];
1423 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1424 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1426 cur->regmap[hr]=reg;
1427 cur->dirty&=~(1<<hr);
1428 cur->isconst&=~(1<<hr);
1433 if (hr == EXCLUDE_REG)
1435 if (hr == HOST_REGS)
1437 if (hr == PREFERRED_REG_FIRST)
1442 // Try to allocate any available register
1443 for (hr = PREFERRED_REG_FIRST; ; ) {
1444 if (cur->regmap[hr] < 0) {
1445 cur->regmap[hr]=reg;
1446 cur->dirty&=~(1<<hr);
1447 cur->isconst&=~(1<<hr);
1451 if (hr == EXCLUDE_REG)
1453 if (hr == HOST_REGS)
1455 if (hr == PREFERRED_REG_FIRST)
1459 // Ok, now we have to evict someone
1460 // Pick a register we hopefully won't need soon
1461 u_char hsn[MAXREG+1];
1462 memset(hsn,10,sizeof(hsn));
1464 lsn(hsn,i,&preferred_reg);
1465 //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]);
1466 //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]);
1468 // Don't evict the cycle count at entry points, otherwise the entry
1469 // stub will have to write it.
1470 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1471 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1474 // Alloc preferred register if available
1475 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1476 for(hr=0;hr<HOST_REGS;hr++) {
1477 // Evict both parts of a 64-bit register
1478 if((cur->regmap[hr]&63)==r) {
1480 cur->dirty&=~(1<<hr);
1481 cur->isconst&=~(1<<hr);
1484 cur->regmap[preferred_reg]=reg;
1487 for(r=1;r<=MAXREG;r++)
1489 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1490 for(hr=0;hr<HOST_REGS;hr++) {
1491 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1492 if(cur->regmap[hr]==r) {
1493 cur->regmap[hr]=reg;
1494 cur->dirty&=~(1<<hr);
1495 cur->isconst&=~(1<<hr);
1506 for(r=1;r<=MAXREG;r++)
1509 for(hr=0;hr<HOST_REGS;hr++) {
1510 if(cur->regmap[hr]==r) {
1511 cur->regmap[hr]=reg;
1512 cur->dirty&=~(1<<hr);
1513 cur->isconst&=~(1<<hr);
1520 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1523 // Allocate a temporary register. This is done without regard to
1524 // dirty status or whether the register we request is on the unneeded list
1525 // Note: This will only allocate one register, even if called multiple times
1526 static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1529 int preferred_reg = -1;
1531 // see if it's already allocated
1532 for(hr=0;hr<HOST_REGS;hr++)
1534 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1537 // Try to allocate any available register
1538 for(hr=HOST_REGS-1;hr>=0;hr--) {
1539 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1540 cur->regmap[hr]=reg;
1541 cur->dirty&=~(1<<hr);
1542 cur->isconst&=~(1<<hr);
1547 // Find an unneeded register
1548 for(hr=HOST_REGS-1;hr>=0;hr--)
1554 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1555 cur->regmap[hr]=reg;
1556 cur->dirty&=~(1<<hr);
1557 cur->isconst&=~(1<<hr);
1564 // Ok, now we have to evict someone
1565 // Pick a register we hopefully won't need soon
1566 // TODO: we might want to follow unconditional jumps here
1567 // TODO: get rid of dupe code and make this into a function
1568 u_char hsn[MAXREG+1];
1569 memset(hsn,10,sizeof(hsn));
1571 lsn(hsn,i,&preferred_reg);
1572 //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]);
1574 // Don't evict the cycle count at entry points, otherwise the entry
1575 // stub will have to write it.
1576 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1577 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1580 for(r=1;r<=MAXREG;r++)
1582 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1583 for(hr=0;hr<HOST_REGS;hr++) {
1584 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1585 if(cur->regmap[hr]==r) {
1586 cur->regmap[hr]=reg;
1587 cur->dirty&=~(1<<hr);
1588 cur->isconst&=~(1<<hr);
1599 for(r=1;r<=MAXREG;r++)
1602 for(hr=0;hr<HOST_REGS;hr++) {
1603 if(cur->regmap[hr]==r) {
1604 cur->regmap[hr]=reg;
1605 cur->dirty&=~(1<<hr);
1606 cur->isconst&=~(1<<hr);
1613 SysPrintf("This shouldn't happen");abort();
1616 static void mov_alloc(struct regstat *current,int i)
1618 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1619 alloc_cc(current,i); // for stalls
1620 dirty_reg(current,CCREG);
1623 // Note: Don't need to actually alloc the source registers
1624 //alloc_reg(current,i,dops[i].rs1);
1625 alloc_reg(current,i,dops[i].rt1);
1627 clear_const(current,dops[i].rs1);
1628 clear_const(current,dops[i].rt1);
1629 dirty_reg(current,dops[i].rt1);
1632 static void shiftimm_alloc(struct regstat *current,int i)
1634 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1637 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1638 else dops[i].lt1=dops[i].rs1;
1639 alloc_reg(current,i,dops[i].rt1);
1640 dirty_reg(current,dops[i].rt1);
1641 if(is_const(current,dops[i].rs1)) {
1642 int v=get_const(current,dops[i].rs1);
1643 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1644 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1645 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1647 else clear_const(current,dops[i].rt1);
1652 clear_const(current,dops[i].rs1);
1653 clear_const(current,dops[i].rt1);
1656 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1660 if(dops[i].opcode2==0x3c) // DSLL32
1664 if(dops[i].opcode2==0x3e) // DSRL32
1668 if(dops[i].opcode2==0x3f) // DSRA32
1674 static void shift_alloc(struct regstat *current,int i)
1677 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1679 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1680 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1681 alloc_reg(current,i,dops[i].rt1);
1682 if(dops[i].rt1==dops[i].rs2) {
1683 alloc_reg_temp(current,i,-1);
1684 minimum_free_regs[i]=1;
1686 } else { // DSLLV/DSRLV/DSRAV
1689 clear_const(current,dops[i].rs1);
1690 clear_const(current,dops[i].rs2);
1691 clear_const(current,dops[i].rt1);
1692 dirty_reg(current,dops[i].rt1);
1696 static void alu_alloc(struct regstat *current,int i)
1698 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1700 if(dops[i].rs1&&dops[i].rs2) {
1701 alloc_reg(current,i,dops[i].rs1);
1702 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==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1713 alloc_reg(current,i,dops[i].rs1);
1714 alloc_reg(current,i,dops[i].rs2);
1715 alloc_reg(current,i,dops[i].rt1);
1718 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1720 if(dops[i].rs1&&dops[i].rs2) {
1721 alloc_reg(current,i,dops[i].rs1);
1722 alloc_reg(current,i,dops[i].rs2);
1726 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1727 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1729 alloc_reg(current,i,dops[i].rt1);
1732 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1735 clear_const(current,dops[i].rs1);
1736 clear_const(current,dops[i].rs2);
1737 clear_const(current,dops[i].rt1);
1738 dirty_reg(current,dops[i].rt1);
1741 static void imm16_alloc(struct regstat *current,int i)
1743 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1744 else dops[i].lt1=dops[i].rs1;
1745 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1746 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1749 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1750 clear_const(current,dops[i].rs1);
1751 clear_const(current,dops[i].rt1);
1753 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1754 if(is_const(current,dops[i].rs1)) {
1755 int v=get_const(current,dops[i].rs1);
1756 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1757 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1758 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1760 else clear_const(current,dops[i].rt1);
1762 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1763 if(is_const(current,dops[i].rs1)) {
1764 int v=get_const(current,dops[i].rs1);
1765 set_const(current,dops[i].rt1,v+imm[i]);
1767 else clear_const(current,dops[i].rt1);
1770 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1772 dirty_reg(current,dops[i].rt1);
1775 static void load_alloc(struct regstat *current,int i)
1777 clear_const(current,dops[i].rt1);
1778 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
1779 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
1780 if (needed_again(dops[i].rs1, i))
1781 alloc_reg(current, i, dops[i].rs1);
1783 alloc_reg(current, i, ROREG);
1784 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
1785 alloc_reg(current,i,dops[i].rt1);
1786 assert(get_reg(current->regmap,dops[i].rt1)>=0);
1787 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
1791 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1795 dirty_reg(current,dops[i].rt1);
1796 // LWL/LWR need a temporary register for the old value
1797 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1799 alloc_reg(current,i,FTEMP);
1800 alloc_reg_temp(current,i,-1);
1801 minimum_free_regs[i]=1;
1806 // Load to r0 or unneeded register (dummy load)
1807 // but we still need a register to calculate the address
1808 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1810 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1812 alloc_reg_temp(current,i,-1);
1813 minimum_free_regs[i]=1;
1814 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1821 void store_alloc(struct regstat *current,int i)
1823 clear_const(current,dops[i].rs2);
1824 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
1825 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1826 alloc_reg(current,i,dops[i].rs2);
1827 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
1831 alloc_reg(current, i, ROREG);
1832 #if defined(HOST_IMM8)
1833 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1834 alloc_reg(current, i, INVCP);
1836 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
1837 alloc_reg(current,i,FTEMP);
1839 // We need a temporary register for address generation
1840 alloc_reg_temp(current,i,-1);
1841 minimum_free_regs[i]=1;
1844 void c1ls_alloc(struct regstat *current,int i)
1846 clear_const(current,dops[i].rt1);
1847 alloc_reg(current,i,CSREG); // Status
1850 void c2ls_alloc(struct regstat *current,int i)
1852 clear_const(current,dops[i].rt1);
1853 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1854 alloc_reg(current,i,FTEMP);
1856 alloc_reg(current, i, ROREG);
1857 #if defined(HOST_IMM8)
1858 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1859 if (dops[i].opcode == 0x3a) // SWC2
1860 alloc_reg(current,i,INVCP);
1862 // We need a temporary register for address generation
1863 alloc_reg_temp(current,i,-1);
1864 minimum_free_regs[i]=1;
1867 #ifndef multdiv_alloc
1868 void multdiv_alloc(struct regstat *current,int i)
1875 // case 0x1D: DMULTU
1878 clear_const(current,dops[i].rs1);
1879 clear_const(current,dops[i].rs2);
1880 alloc_cc(current,i); // for stalls
1881 if(dops[i].rs1&&dops[i].rs2)
1883 if((dops[i].opcode2&4)==0) // 32-bit
1885 current->u&=~(1LL<<HIREG);
1886 current->u&=~(1LL<<LOREG);
1887 alloc_reg(current,i,HIREG);
1888 alloc_reg(current,i,LOREG);
1889 alloc_reg(current,i,dops[i].rs1);
1890 alloc_reg(current,i,dops[i].rs2);
1891 dirty_reg(current,HIREG);
1892 dirty_reg(current,LOREG);
1901 // Multiply by zero is zero.
1902 // MIPS does not have a divide by zero exception.
1903 // The result is undefined, we return zero.
1904 alloc_reg(current,i,HIREG);
1905 alloc_reg(current,i,LOREG);
1906 dirty_reg(current,HIREG);
1907 dirty_reg(current,LOREG);
1912 void cop0_alloc(struct regstat *current,int i)
1914 if(dops[i].opcode2==0) // MFC0
1917 clear_const(current,dops[i].rt1);
1918 alloc_all(current,i);
1919 alloc_reg(current,i,dops[i].rt1);
1920 dirty_reg(current,dops[i].rt1);
1923 else if(dops[i].opcode2==4) // MTC0
1926 clear_const(current,dops[i].rs1);
1927 alloc_reg(current,i,dops[i].rs1);
1928 alloc_all(current,i);
1931 alloc_all(current,i); // FIXME: Keep r0
1933 alloc_reg(current,i,0);
1938 // TLBR/TLBWI/TLBWR/TLBP/ERET
1939 assert(dops[i].opcode2==0x10);
1940 alloc_all(current,i);
1942 minimum_free_regs[i]=HOST_REGS;
1945 static void cop2_alloc(struct regstat *current,int i)
1947 if (dops[i].opcode2 < 3) // MFC2/CFC2
1949 alloc_cc(current,i); // for stalls
1950 dirty_reg(current,CCREG);
1952 clear_const(current,dops[i].rt1);
1953 alloc_reg(current,i,dops[i].rt1);
1954 dirty_reg(current,dops[i].rt1);
1957 else if (dops[i].opcode2 > 3) // MTC2/CTC2
1960 clear_const(current,dops[i].rs1);
1961 alloc_reg(current,i,dops[i].rs1);
1965 alloc_reg(current,i,0);
1968 alloc_reg_temp(current,i,-1);
1969 minimum_free_regs[i]=1;
1972 void c2op_alloc(struct regstat *current,int i)
1974 alloc_cc(current,i); // for stalls
1975 dirty_reg(current,CCREG);
1976 alloc_reg_temp(current,i,-1);
1979 void syscall_alloc(struct regstat *current,int i)
1981 alloc_cc(current,i);
1982 dirty_reg(current,CCREG);
1983 alloc_all(current,i);
1984 minimum_free_regs[i]=HOST_REGS;
1988 void delayslot_alloc(struct regstat *current,int i)
1990 switch(dops[i].itype) {
1998 assem_debug("jump in the delay slot. this shouldn't happen.\n");//abort();
1999 SysPrintf("Disabled speculative precompilation\n");
2003 imm16_alloc(current,i);
2007 load_alloc(current,i);
2011 store_alloc(current,i);
2014 alu_alloc(current,i);
2017 shift_alloc(current,i);
2020 multdiv_alloc(current,i);
2023 shiftimm_alloc(current,i);
2026 mov_alloc(current,i);
2029 cop0_alloc(current,i);
2034 cop2_alloc(current,i);
2037 c1ls_alloc(current,i);
2040 c2ls_alloc(current,i);
2043 c2op_alloc(current,i);
2048 // Special case where a branch and delay slot span two pages in virtual memory
2049 static void pagespan_alloc(struct regstat *current,int i)
2052 current->wasconst=0;
2054 minimum_free_regs[i]=HOST_REGS;
2055 alloc_all(current,i);
2056 alloc_cc(current,i);
2057 dirty_reg(current,CCREG);
2058 if(dops[i].opcode==3) // JAL
2060 alloc_reg(current,i,31);
2061 dirty_reg(current,31);
2063 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
2065 alloc_reg(current,i,dops[i].rs1);
2066 if (dops[i].rt1!=0) {
2067 alloc_reg(current,i,dops[i].rt1);
2068 dirty_reg(current,dops[i].rt1);
2071 if((dops[i].opcode&0x2E)==4) // BEQ/BNE/BEQL/BNEL
2073 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2074 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
2077 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
2079 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2084 static void add_stub(enum stub_type type, void *addr, void *retaddr,
2085 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2087 assert(stubcount < ARRAY_SIZE(stubs));
2088 stubs[stubcount].type = type;
2089 stubs[stubcount].addr = addr;
2090 stubs[stubcount].retaddr = retaddr;
2091 stubs[stubcount].a = a;
2092 stubs[stubcount].b = b;
2093 stubs[stubcount].c = c;
2094 stubs[stubcount].d = d;
2095 stubs[stubcount].e = e;
2099 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2100 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2102 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2105 // Write out a single register
2106 static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2109 for(hr=0;hr<HOST_REGS;hr++) {
2110 if(hr!=EXCLUDE_REG) {
2111 if((regmap[hr]&63)==r) {
2113 assert(regmap[hr]<64);
2114 emit_storereg(r,hr);
2121 static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2123 //if(dirty_pre==dirty) return;
2125 for(hr=0;hr<HOST_REGS;hr++) {
2126 if(hr!=EXCLUDE_REG) {
2128 if(((~u)>>(reg&63))&1) {
2130 if(((dirty_pre&~dirty)>>hr)&1) {
2132 emit_storereg(reg,hr);
2145 static void pass_args(int a0, int a1)
2149 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2151 else if(a0!=0&&a1==0) {
2153 if (a0>=0) emit_mov(a0,0);
2156 if(a0>=0&&a0!=0) emit_mov(a0,0);
2157 if(a1>=0&&a1!=1) emit_mov(a1,1);
2161 static void alu_assemble(int i, const struct regstat *i_regs)
2163 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2165 signed char s1,s2,t;
2166 t=get_reg(i_regs->regmap,dops[i].rt1);
2168 s1=get_reg(i_regs->regmap,dops[i].rs1);
2169 s2=get_reg(i_regs->regmap,dops[i].rs2);
2170 if(dops[i].rs1&&dops[i].rs2) {
2173 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2174 else emit_add(s1,s2,t);
2176 else if(dops[i].rs1) {
2177 if(s1>=0) emit_mov(s1,t);
2178 else emit_loadreg(dops[i].rs1,t);
2180 else if(dops[i].rs2) {
2182 if(dops[i].opcode2&2) emit_neg(s2,t);
2183 else emit_mov(s2,t);
2186 emit_loadreg(dops[i].rs2,t);
2187 if(dops[i].opcode2&2) emit_neg(t,t);
2190 else emit_zeroreg(t);
2194 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2197 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2199 signed char s1l,s2l,t;
2201 t=get_reg(i_regs->regmap,dops[i].rt1);
2204 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2205 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2206 if(dops[i].rs2==0) // rx<r0
2208 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2210 emit_shrimm(s1l,31,t);
2212 else // SLTU (unsigned can not be less than zero, 0<0)
2215 else if(dops[i].rs1==0) // r0<rx
2218 if(dops[i].opcode2==0x2a) // SLT
2219 emit_set_gz32(s2l,t);
2220 else // SLTU (set if not zero)
2221 emit_set_nz32(s2l,t);
2224 assert(s1l>=0);assert(s2l>=0);
2225 if(dops[i].opcode2==0x2a) // SLT
2226 emit_set_if_less32(s1l,s2l,t);
2228 emit_set_if_carry32(s1l,s2l,t);
2234 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2236 signed char s1l,s2l,tl;
2237 tl=get_reg(i_regs->regmap,dops[i].rt1);
2240 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2241 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2242 if(dops[i].rs1&&dops[i].rs2) {
2245 if(dops[i].opcode2==0x24) { // AND
2246 emit_and(s1l,s2l,tl);
2248 if(dops[i].opcode2==0x25) { // OR
2249 emit_or(s1l,s2l,tl);
2251 if(dops[i].opcode2==0x26) { // XOR
2252 emit_xor(s1l,s2l,tl);
2254 if(dops[i].opcode2==0x27) { // NOR
2255 emit_or(s1l,s2l,tl);
2261 if(dops[i].opcode2==0x24) { // AND
2264 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2266 if(s1l>=0) emit_mov(s1l,tl);
2267 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2271 if(s2l>=0) emit_mov(s2l,tl);
2272 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2274 else emit_zeroreg(tl);
2276 if(dops[i].opcode2==0x27) { // NOR
2278 if(s1l>=0) emit_not(s1l,tl);
2280 emit_loadreg(dops[i].rs1,tl);
2286 if(s2l>=0) emit_not(s2l,tl);
2288 emit_loadreg(dops[i].rs2,tl);
2292 else emit_movimm(-1,tl);
2301 static void imm16_assemble(int i, const struct regstat *i_regs)
2303 if (dops[i].opcode==0x0f) { // LUI
2306 t=get_reg(i_regs->regmap,dops[i].rt1);
2309 if(!((i_regs->isconst>>t)&1))
2310 emit_movimm(imm[i]<<16,t);
2314 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2317 t=get_reg(i_regs->regmap,dops[i].rt1);
2318 s=get_reg(i_regs->regmap,dops[i].rs1);
2323 if(!((i_regs->isconst>>t)&1)) {
2325 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2326 emit_addimm(t,imm[i],t);
2328 if(!((i_regs->wasconst>>s)&1))
2329 emit_addimm(s,imm[i],t);
2331 emit_movimm(constmap[i][s]+imm[i],t);
2337 if(!((i_regs->isconst>>t)&1))
2338 emit_movimm(imm[i],t);
2343 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2346 tl=get_reg(i_regs->regmap,dops[i].rt1);
2347 sl=get_reg(i_regs->regmap,dops[i].rs1);
2351 emit_addimm(sl,imm[i],tl);
2353 emit_movimm(imm[i],tl);
2358 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2360 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2362 t=get_reg(i_regs->regmap,dops[i].rt1);
2363 sl=get_reg(i_regs->regmap,dops[i].rs1);
2367 if(dops[i].opcode==0x0a) { // SLTI
2369 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2370 emit_slti32(t,imm[i],t);
2372 emit_slti32(sl,imm[i],t);
2377 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2378 emit_sltiu32(t,imm[i],t);
2380 emit_sltiu32(sl,imm[i],t);
2384 // SLTI(U) with r0 is just stupid,
2385 // nonetheless examples can be found
2386 if(dops[i].opcode==0x0a) // SLTI
2387 if(0<imm[i]) emit_movimm(1,t);
2388 else emit_zeroreg(t);
2391 if(imm[i]) emit_movimm(1,t);
2392 else emit_zeroreg(t);
2398 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2401 tl=get_reg(i_regs->regmap,dops[i].rt1);
2402 sl=get_reg(i_regs->regmap,dops[i].rs1);
2403 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2404 if(dops[i].opcode==0x0c) //ANDI
2408 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2409 emit_andimm(tl,imm[i],tl);
2411 if(!((i_regs->wasconst>>sl)&1))
2412 emit_andimm(sl,imm[i],tl);
2414 emit_movimm(constmap[i][sl]&imm[i],tl);
2424 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2426 if(dops[i].opcode==0x0d) { // ORI
2428 emit_orimm(tl,imm[i],tl);
2430 if(!((i_regs->wasconst>>sl)&1))
2431 emit_orimm(sl,imm[i],tl);
2433 emit_movimm(constmap[i][sl]|imm[i],tl);
2436 if(dops[i].opcode==0x0e) { // XORI
2438 emit_xorimm(tl,imm[i],tl);
2440 if(!((i_regs->wasconst>>sl)&1))
2441 emit_xorimm(sl,imm[i],tl);
2443 emit_movimm(constmap[i][sl]^imm[i],tl);
2448 emit_movimm(imm[i],tl);
2456 static void shiftimm_assemble(int i, const struct regstat *i_regs)
2458 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2462 t=get_reg(i_regs->regmap,dops[i].rt1);
2463 s=get_reg(i_regs->regmap,dops[i].rs1);
2465 if(t>=0&&!((i_regs->isconst>>t)&1)){
2472 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2474 if(dops[i].opcode2==0) // SLL
2476 emit_shlimm(s<0?t:s,imm[i],t);
2478 if(dops[i].opcode2==2) // SRL
2480 emit_shrimm(s<0?t:s,imm[i],t);
2482 if(dops[i].opcode2==3) // SRA
2484 emit_sarimm(s<0?t:s,imm[i],t);
2488 if(s>=0 && s!=t) emit_mov(s,t);
2492 //emit_storereg(dops[i].rt1,t); //DEBUG
2495 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2499 if(dops[i].opcode2==0x3c) // DSLL32
2503 if(dops[i].opcode2==0x3e) // DSRL32
2507 if(dops[i].opcode2==0x3f) // DSRA32
2513 #ifndef shift_assemble
2514 static void shift_assemble(int i, const struct regstat *i_regs)
2516 signed char s,t,shift;
2517 if (dops[i].rt1 == 0)
2519 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2520 t = get_reg(i_regs->regmap, dops[i].rt1);
2521 s = get_reg(i_regs->regmap, dops[i].rs1);
2522 shift = get_reg(i_regs->regmap, dops[i].rs2);
2528 else if(dops[i].rs2==0) {
2530 if(s!=t) emit_mov(s,t);
2533 host_tempreg_acquire();
2534 emit_andimm(shift,31,HOST_TEMPREG);
2535 switch(dops[i].opcode2) {
2537 emit_shl(s,HOST_TEMPREG,t);
2540 emit_shr(s,HOST_TEMPREG,t);
2543 emit_sar(s,HOST_TEMPREG,t);
2548 host_tempreg_release();
2562 static int get_ptr_mem_type(u_int a)
2564 if(a < 0x00200000) {
2565 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2566 // return wrong, must use memhandler for BIOS self-test to pass
2567 // 007 does similar stuff from a00 mirror, weird stuff
2571 if(0x1f800000 <= a && a < 0x1f801000)
2573 if(0x80200000 <= a && a < 0x80800000)
2575 if(0xa0000000 <= a && a < 0xa0200000)
2580 static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2582 int r = get_reg(i_regs->regmap, ROREG);
2583 if (r < 0 && host_tempreg_free) {
2584 host_tempreg_acquire();
2585 emit_loadreg(ROREG, r = HOST_TEMPREG);
2592 static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2593 int addr, int *offset_reg, int *addr_reg_override)
2597 int mr = dops[i].rs1;
2599 if(((smrv_strong|smrv_weak)>>mr)&1) {
2600 type=get_ptr_mem_type(smrv[mr]);
2601 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2604 // use the mirror we are running on
2605 type=get_ptr_mem_type(start);
2606 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2609 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2610 host_tempreg_acquire();
2611 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2612 addr=*addr_reg_override=HOST_TEMPREG;
2615 else if(type==MTYPE_0000) { // RAM 0 mirror
2616 host_tempreg_acquire();
2617 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2618 addr=*addr_reg_override=HOST_TEMPREG;
2621 else if(type==MTYPE_A000) { // RAM A mirror
2622 host_tempreg_acquire();
2623 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2624 addr=*addr_reg_override=HOST_TEMPREG;
2627 else if(type==MTYPE_1F80) { // scratchpad
2628 if (psxH == (void *)0x1f800000) {
2629 host_tempreg_acquire();
2630 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2631 emit_cmpimm(HOST_TEMPREG,0x1000);
2632 host_tempreg_release();
2637 // do the usual RAM check, jump will go to the right handler
2642 if (type == 0) // need ram check
2644 emit_cmpimm(addr,RAM_SIZE);
2646 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2647 // Hint to branch predictor that the branch is unlikely to be taken
2648 if (dops[i].rs1 >= 28)
2649 emit_jno_unlikely(0);
2653 if (ram_offset != 0)
2654 *offset_reg = get_ro_reg(i_regs, 0);
2660 // return memhandler, or get directly accessable address and return 0
2661 static void *get_direct_memhandler(void *table, u_int addr,
2662 enum stub_type type, uintptr_t *addr_host)
2664 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2665 uintptr_t l1, l2 = 0;
2666 l1 = ((uintptr_t *)table)[addr>>12];
2668 uintptr_t v = l1 << 1;
2669 *addr_host = v + addr;
2674 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2675 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2676 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2677 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2679 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2681 uintptr_t v = l2 << 1;
2682 *addr_host = v + (addr&0xfff);
2685 return (void *)(l2 << 1);
2689 static u_int get_host_reglist(const signed char *regmap)
2691 u_int reglist = 0, hr;
2692 for (hr = 0; hr < HOST_REGS; hr++) {
2693 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2699 static u_int reglist_exclude(u_int reglist, int r1, int r2)
2702 reglist &= ~(1u << r1);
2704 reglist &= ~(1u << r2);
2708 // find a temp caller-saved register not in reglist (so assumed to be free)
2709 static int reglist_find_free(u_int reglist)
2711 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2714 return __builtin_ctz(free_regs);
2717 static void do_load_word(int a, int rt, int offset_reg)
2719 if (offset_reg >= 0)
2720 emit_ldr_dualindexed(offset_reg, a, rt);
2722 emit_readword_indexed(0, a, rt);
2725 static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2727 if (offset_reg < 0) {
2728 emit_writeword_indexed(rt, ofs, a);
2732 emit_addimm(a, ofs, a);
2733 emit_str_dualindexed(offset_reg, a, rt);
2734 if (ofs != 0 && preseve_a)
2735 emit_addimm(a, -ofs, a);
2738 static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2740 if (offset_reg < 0) {
2741 emit_writehword_indexed(rt, ofs, a);
2745 emit_addimm(a, ofs, a);
2746 emit_strh_dualindexed(offset_reg, a, rt);
2747 if (ofs != 0 && preseve_a)
2748 emit_addimm(a, -ofs, a);
2751 static void do_store_byte(int a, int rt, int offset_reg)
2753 if (offset_reg >= 0)
2754 emit_strb_dualindexed(offset_reg, a, rt);
2756 emit_writebyte_indexed(rt, 0, a);
2759 static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2764 int memtarget=0,c=0;
2765 int offset_reg = -1;
2766 int fastio_reg_override = -1;
2767 u_int reglist=get_host_reglist(i_regs->regmap);
2768 tl=get_reg(i_regs->regmap,dops[i].rt1);
2769 s=get_reg(i_regs->regmap,dops[i].rs1);
2771 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2773 c=(i_regs->wasconst>>s)&1;
2775 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2778 //printf("load_assemble: c=%d\n",c);
2779 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2780 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2781 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2783 // could be FIFO, must perform the read
2785 assem_debug("(forced read)\n");
2786 tl=get_reg(i_regs->regmap,-1);
2789 if(offset||s<0||c) addr=tl;
2791 //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2793 //printf("load_assemble: c=%d\n",c);
2794 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2795 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2799 // Strmnnrmn's speed hack
2800 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2803 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
2804 &offset_reg, &fastio_reg_override);
2807 else if (ram_offset && memtarget) {
2808 offset_reg = get_ro_reg(i_regs, 0);
2810 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2811 switch (dops[i].opcode) {
2817 if (fastio_reg_override >= 0)
2818 a = fastio_reg_override;
2820 if (offset_reg >= 0)
2821 emit_ldrsb_dualindexed(offset_reg, a, tl);
2823 emit_movsbl_indexed(0, a, tl);
2826 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2829 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2836 if (fastio_reg_override >= 0)
2837 a = fastio_reg_override;
2838 if (offset_reg >= 0)
2839 emit_ldrsh_dualindexed(offset_reg, a, tl);
2841 emit_movswl_indexed(0, a, tl);
2844 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2847 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2853 if (fastio_reg_override >= 0)
2854 a = fastio_reg_override;
2855 do_load_word(a, tl, offset_reg);
2858 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2861 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2868 if (fastio_reg_override >= 0)
2869 a = fastio_reg_override;
2871 if (offset_reg >= 0)
2872 emit_ldrb_dualindexed(offset_reg, a, tl);
2874 emit_movzbl_indexed(0, a, tl);
2877 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2880 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2887 if (fastio_reg_override >= 0)
2888 a = fastio_reg_override;
2889 if (offset_reg >= 0)
2890 emit_ldrh_dualindexed(offset_reg, a, tl);
2892 emit_movzwl_indexed(0, a, tl);
2895 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2898 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2906 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
2907 host_tempreg_release();
2910 #ifndef loadlr_assemble
2911 static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
2913 int s,tl,temp,temp2,addr;
2916 int memtarget=0,c=0;
2917 int offset_reg = -1;
2918 int fastio_reg_override = -1;
2919 u_int reglist=get_host_reglist(i_regs->regmap);
2920 tl=get_reg(i_regs->regmap,dops[i].rt1);
2921 s=get_reg(i_regs->regmap,dops[i].rs1);
2922 temp=get_reg(i_regs->regmap,-1);
2923 temp2=get_reg(i_regs->regmap,FTEMP);
2924 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
2928 if(offset||s<0||c) addr=temp2;
2931 c=(i_regs->wasconst>>s)&1;
2933 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2937 emit_shlimm(addr,3,temp);
2938 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2939 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
2941 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
2943 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
2944 &offset_reg, &fastio_reg_override);
2947 if (ram_offset && memtarget) {
2948 offset_reg = get_ro_reg(i_regs, 0);
2950 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2951 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
2953 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
2956 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
2959 if (fastio_reg_override >= 0)
2960 a = fastio_reg_override;
2961 do_load_word(a, temp2, offset_reg);
2962 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
2963 host_tempreg_release();
2964 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
2967 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
2970 emit_andimm(temp,24,temp);
2971 if (dops[i].opcode==0x22) // LWL
2972 emit_xorimm(temp,24,temp);
2973 host_tempreg_acquire();
2974 emit_movimm(-1,HOST_TEMPREG);
2975 if (dops[i].opcode==0x26) {
2976 emit_shr(temp2,temp,temp2);
2977 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
2979 emit_shl(temp2,temp,temp2);
2980 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
2982 host_tempreg_release();
2983 emit_or(temp2,tl,tl);
2985 //emit_storereg(dops[i].rt1,tl); // DEBUG
2987 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
2993 static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
2999 enum stub_type type=0;
3000 int memtarget=0,c=0;
3001 int agr=AGEN1+(i&1);
3002 int offset_reg = -1;
3003 int fastio_reg_override = -1;
3004 u_int reglist=get_host_reglist(i_regs->regmap);
3005 tl=get_reg(i_regs->regmap,dops[i].rs2);
3006 s=get_reg(i_regs->regmap,dops[i].rs1);
3007 temp=get_reg(i_regs->regmap,agr);
3008 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3011 c=(i_regs->wasconst>>s)&1;
3013 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3018 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3019 if(offset||s<0||c) addr=temp;
3022 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3023 &offset_reg, &fastio_reg_override);
3025 else if (ram_offset && memtarget) {
3026 offset_reg = get_ro_reg(i_regs, 0);
3029 switch (dops[i].opcode) {
3034 if (fastio_reg_override >= 0)
3035 a = fastio_reg_override;
3036 do_store_byte(a, tl, offset_reg);
3044 if (fastio_reg_override >= 0)
3045 a = fastio_reg_override;
3046 do_store_hword(a, 0, tl, offset_reg, 1);
3053 if (fastio_reg_override >= 0)
3054 a = fastio_reg_override;
3055 do_store_word(a, 0, tl, offset_reg, 1);
3063 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3064 host_tempreg_release();
3066 // PCSX store handlers don't check invcode again
3068 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3071 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3073 #ifdef DESTRUCTIVE_SHIFT
3074 // The x86 shift operation is 'destructive'; it overwrites the
3075 // source register, so we need to make a copy first and use that.
3078 #if defined(HOST_IMM8)
3079 int ir=get_reg(i_regs->regmap,INVCP);
3081 emit_cmpmem_indexedsr12_reg(ir,addr,1);
3083 emit_cmpmem_indexedsr12_imm(invalid_code,addr,1);
3085 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3086 emit_callne(invalidate_addr_reg[addr]);
3090 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3094 u_int addr_val=constmap[i][s]+offset;
3096 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3097 } else if(c&&!memtarget) {
3098 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3100 // basic current block modification detection..
3101 // not looking back as that should be in mips cache already
3102 // (see Spyro2 title->attract mode)
3103 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3104 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3105 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3106 if(i_regs->regmap==regs[i].regmap) {
3107 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3108 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3109 emit_movimm(start+i*4+4,0);
3110 emit_writeword(0,&pcaddr);
3111 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3112 emit_far_call(get_addr_ht);
3118 static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3124 void *case1, *case23, *case3;
3125 void *done0, *done1, *done2;
3126 int memtarget=0,c=0;
3127 int agr=AGEN1+(i&1);
3128 int offset_reg = -1;
3129 u_int reglist=get_host_reglist(i_regs->regmap);
3130 tl=get_reg(i_regs->regmap,dops[i].rs2);
3131 s=get_reg(i_regs->regmap,dops[i].rs1);
3132 temp=get_reg(i_regs->regmap,agr);
3133 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3136 c=(i_regs->isconst>>s)&1;
3138 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3144 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3145 if(!offset&&s!=temp) emit_mov(s,temp);
3151 if(!memtarget||!dops[i].rs1) {
3157 offset_reg = get_ro_reg(i_regs, 0);
3159 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3163 emit_testimm(temp,2);
3166 emit_testimm(temp,1);
3170 if (dops[i].opcode == 0x2A) { // SWL
3171 // Write msb into least significant byte
3172 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3173 do_store_byte(temp, tl, offset_reg);
3174 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3176 else if (dops[i].opcode == 0x2E) { // SWR
3177 // Write entire word
3178 do_store_word(temp, 0, tl, offset_reg, 1);
3183 set_jump_target(case1, out);
3184 if (dops[i].opcode == 0x2A) { // SWL
3185 // Write two msb into two least significant bytes
3186 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3187 do_store_hword(temp, -1, tl, offset_reg, 0);
3188 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3190 else if (dops[i].opcode == 0x2E) { // SWR
3191 // Write 3 lsb into three most significant bytes
3192 do_store_byte(temp, tl, offset_reg);
3193 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3194 do_store_hword(temp, 1, tl, offset_reg, 0);
3195 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3200 set_jump_target(case23, out);
3201 emit_testimm(temp,1);
3205 if (dops[i].opcode==0x2A) { // SWL
3206 // Write 3 msb into three least significant bytes
3207 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3208 do_store_hword(temp, -2, tl, offset_reg, 1);
3209 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3210 do_store_byte(temp, tl, offset_reg);
3211 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3213 else if (dops[i].opcode == 0x2E) { // SWR
3214 // Write two lsb into two most significant bytes
3215 do_store_hword(temp, 0, tl, offset_reg, 1);
3220 set_jump_target(case3, out);
3221 if (dops[i].opcode == 0x2A) { // SWL
3222 do_store_word(temp, -3, tl, offset_reg, 0);
3224 else if (dops[i].opcode == 0x2E) { // SWR
3225 do_store_byte(temp, tl, offset_reg);
3227 set_jump_target(done0, out);
3228 set_jump_target(done1, out);
3229 set_jump_target(done2, out);
3230 if (offset_reg == HOST_TEMPREG)
3231 host_tempreg_release();
3233 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj_,reglist);
3234 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3235 #if defined(HOST_IMM8)
3236 int ir=get_reg(i_regs->regmap,INVCP);
3238 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3240 emit_cmpmem_indexedsr12_imm(invalid_code,temp,1);
3242 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3243 emit_callne(invalidate_addr_reg[temp]);
3247 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3252 static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3254 if(dops[i].opcode2==0) // MFC0
3256 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3257 u_int copr=(source[i]>>11)&0x1f;
3258 //assert(t>=0); // Why does this happen? OOT is weird
3259 if(t>=0&&dops[i].rt1!=0) {
3260 emit_readword(®_cop0[copr],t);
3263 else if(dops[i].opcode2==4) // MTC0
3265 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3266 char copr=(source[i]>>11)&0x1f;
3268 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3269 if(copr==9||copr==11||copr==12||copr==13) {
3270 emit_readword(&last_count,HOST_TEMPREG);
3271 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3272 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3273 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3274 emit_writeword(HOST_CCREG,&Count);
3276 // What a mess. The status register (12) can enable interrupts,
3277 // so needs a special case to handle a pending interrupt.
3278 // The interrupt must be taken immediately, because a subsequent
3279 // instruction might disable interrupts again.
3280 if(copr==12||copr==13) {
3282 // burn cycles to cause cc_interrupt, which will
3283 // reschedule next_interupt. Relies on CCREG from above.
3284 assem_debug("MTC0 DS %d\n", copr);
3285 emit_writeword(HOST_CCREG,&last_count);
3286 emit_movimm(0,HOST_CCREG);
3287 emit_storereg(CCREG,HOST_CCREG);
3288 emit_loadreg(dops[i].rs1,1);
3289 emit_movimm(copr,0);
3290 emit_far_call(pcsx_mtc0_ds);
3291 emit_loadreg(dops[i].rs1,s);
3294 emit_movimm(start+i*4+4,HOST_TEMPREG);
3295 emit_writeword(HOST_TEMPREG,&pcaddr);
3296 emit_movimm(0,HOST_TEMPREG);
3297 emit_writeword(HOST_TEMPREG,&pending_exception);
3300 emit_loadreg(dops[i].rs1,1);
3303 emit_movimm(copr,0);
3304 emit_far_call(pcsx_mtc0);
3305 if(copr==9||copr==11||copr==12||copr==13) {
3306 emit_readword(&Count,HOST_CCREG);
3307 emit_readword(&next_interupt,HOST_TEMPREG);
3308 emit_addimm(HOST_CCREG,-ccadj_,HOST_CCREG);
3309 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3310 emit_writeword(HOST_TEMPREG,&last_count);
3311 emit_storereg(CCREG,HOST_CCREG);
3313 if(copr==12||copr==13) {
3314 assert(!is_delayslot);
3315 emit_readword(&pending_exception,14);
3319 emit_readword(&pcaddr, 0);
3320 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3321 emit_far_call(get_addr_ht);
3323 set_jump_target(jaddr, out);
3325 emit_loadreg(dops[i].rs1,s);
3329 assert(dops[i].opcode2==0x10);
3330 //if((source[i]&0x3f)==0x10) // RFE
3332 emit_readword(&Status,0);
3333 emit_andimm(0,0x3c,1);
3334 emit_andimm(0,~0xf,0);
3335 emit_orrshr_imm(1,2,0);
3336 emit_writeword(0,&Status);
3341 static void cop1_unusable(int i, const struct regstat *i_regs)
3343 // XXX: should just just do the exception instead
3348 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3352 static void cop1_assemble(int i, const struct regstat *i_regs)
3354 cop1_unusable(i, i_regs);
3357 static void c1ls_assemble(int i, const struct regstat *i_regs)
3359 cop1_unusable(i, i_regs);
3363 static void do_cop1stub(int n)
3366 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3367 set_jump_target(stubs[n].addr, out);
3369 // int rs=stubs[n].b;
3370 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3373 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3374 //if(i_regs!=®s[i]) printf("oops: regs[i]=%x i_regs=%x",(int)®s[i],(int)i_regs);
3376 //else {printf("fp exception in delay slot\n");}
3377 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3378 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3379 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3380 emit_addimm(HOST_CCREG,ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3381 emit_far_jump(ds?fp_exception_ds:fp_exception);
3384 static int cop2_is_stalling_op(int i, int *cycles)
3386 if (dops[i].opcode == 0x3a) { // SWC2
3390 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3394 if (dops[i].itype == C2OP) {
3395 *cycles = gte_cycletab[source[i] & 0x3f];
3398 // ... what about MTC2/CTC2/LWC2?
3403 static void log_gte_stall(int stall, u_int cycle)
3405 if ((u_int)stall <= 44)
3406 printf("x stall %2d %u\n", stall, cycle + last_count);
3409 static void emit_log_gte_stall(int i, int stall, u_int reglist)
3413 emit_movimm(stall, 0);
3415 emit_mov(HOST_TEMPREG, 0);
3416 emit_addimm(HOST_CCREG, ccadj[i], 1);
3417 emit_far_call(log_gte_stall);
3418 restore_regs(reglist);
3422 static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3424 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3425 int rtmp = reglist_find_free(reglist);
3427 if (HACK_ENABLED(NDHACK_NO_STALLS))
3429 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3430 // happens occasionally... cc evicted? Don't bother then
3431 //printf("no cc %08x\n", start + i*4);
3435 for (j = i - 1; j >= 0; j--) {
3436 //if (dops[j].is_ds) break;
3437 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3439 if (j > 0 && ccadj[j - 1] > ccadj[j])
3444 cycles_passed = ccadj[i] - ccadj[j];
3445 if (other_gte_op_cycles >= 0)
3446 stall = other_gte_op_cycles - cycles_passed;
3447 else if (cycles_passed >= 44)
3448 stall = 0; // can't stall
3449 if (stall == -MAXBLOCK && rtmp >= 0) {
3450 // unknown stall, do the expensive runtime check
3451 assem_debug("; cop2_do_stall_check\n");
3454 emit_movimm(gte_cycletab[op], 0);
3455 emit_addimm(HOST_CCREG, ccadj[i], 1);
3456 emit_far_call(call_gteStall);
3457 restore_regs(reglist);
3459 host_tempreg_acquire();
3460 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3461 emit_addimm(rtmp, -ccadj[i], rtmp);
3462 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3463 emit_cmpimm(HOST_TEMPREG, 44);
3464 emit_cmovb_reg(rtmp, HOST_CCREG);
3465 //emit_log_gte_stall(i, 0, reglist);
3466 host_tempreg_release();
3469 else if (stall > 0) {
3470 //emit_log_gte_stall(i, stall, reglist);
3471 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3474 // save gteBusyCycle, if needed
3475 if (gte_cycletab[op] == 0)
3477 other_gte_op_cycles = -1;
3478 for (j = i + 1; j < slen; j++) {
3479 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3481 if (dops[j].is_jump) {
3483 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3488 if (other_gte_op_cycles >= 0)
3489 // will handle stall when assembling that op
3491 cycles_passed = ccadj[min(j, slen -1)] - ccadj[i];
3492 if (cycles_passed >= 44)
3494 assem_debug("; save gteBusyCycle\n");
3495 host_tempreg_acquire();
3497 emit_readword(&last_count, HOST_TEMPREG);
3498 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3499 emit_addimm(HOST_TEMPREG, ccadj[i], HOST_TEMPREG);
3500 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3501 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3503 emit_addimm(HOST_CCREG, ccadj[i] + gte_cycletab[op], HOST_TEMPREG);
3504 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3506 host_tempreg_release();
3509 static int is_mflohi(int i)
3511 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3514 static int check_multdiv(int i, int *cycles)
3516 if (dops[i].itype != MULTDIV)
3518 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3519 *cycles = 11; // approx from 7 11 14
3525 static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3527 int j, found = 0, c = 0;
3528 if (HACK_ENABLED(NDHACK_NO_STALLS))
3530 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3531 // happens occasionally... cc evicted? Don't bother then
3534 for (j = i + 1; j < slen; j++) {
3537 if ((found = is_mflohi(j)))
3539 if (dops[j].is_jump) {
3541 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3547 // handle all in multdiv_do_stall()
3549 check_multdiv(i, &c);
3551 assem_debug("; muldiv prepare stall %d\n", c);
3552 host_tempreg_acquire();
3553 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3554 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3555 host_tempreg_release();
3558 static void multdiv_do_stall(int i, const struct regstat *i_regs)
3560 int j, known_cycles = 0;
3561 u_int reglist = get_host_reglist(i_regs->regmap);
3562 int rtmp = get_reg(i_regs->regmap, -1);
3564 rtmp = reglist_find_free(reglist);
3565 if (HACK_ENABLED(NDHACK_NO_STALLS))
3567 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3568 // happens occasionally... cc evicted? Don't bother then
3569 //printf("no cc/rtmp %08x\n", start + i*4);
3573 for (j = i - 1; j >= 0; j--) {
3574 if (dops[j].is_ds) break;
3575 if (check_multdiv(j, &known_cycles))
3578 // already handled by this op
3580 if (dops[j].bt || (j > 0 && ccadj[j - 1] > ccadj[j]))
3585 if (known_cycles > 0) {
3586 known_cycles -= ccadj[i] - ccadj[j];
3587 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3588 if (known_cycles > 0)
3589 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3592 assem_debug("; muldiv stall unresolved\n");
3593 host_tempreg_acquire();
3594 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3595 emit_addimm(rtmp, -ccadj[i], rtmp);
3596 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3597 emit_cmpimm(HOST_TEMPREG, 37);
3598 emit_cmovb_reg(rtmp, HOST_CCREG);
3599 //emit_log_gte_stall(i, 0, reglist);
3600 host_tempreg_release();
3603 static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3613 emit_readword(®_cop2d[copr],tl);
3614 emit_signextend16(tl,tl);
3615 emit_writeword(tl,®_cop2d[copr]); // hmh
3622 emit_readword(®_cop2d[copr],tl);
3623 emit_andimm(tl,0xffff,tl);
3624 emit_writeword(tl,®_cop2d[copr]);
3627 emit_readword(®_cop2d[14],tl); // SXY2
3628 emit_writeword(tl,®_cop2d[copr]);
3632 c2op_mfc2_29_assemble(tl,temp);
3635 emit_readword(®_cop2d[copr],tl);
3640 static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3644 emit_readword(®_cop2d[13],temp); // SXY1
3645 emit_writeword(sl,®_cop2d[copr]);
3646 emit_writeword(temp,®_cop2d[12]); // SXY0
3647 emit_readword(®_cop2d[14],temp); // SXY2
3648 emit_writeword(sl,®_cop2d[14]);
3649 emit_writeword(temp,®_cop2d[13]); // SXY1
3652 emit_andimm(sl,0x001f,temp);
3653 emit_shlimm(temp,7,temp);
3654 emit_writeword(temp,®_cop2d[9]);
3655 emit_andimm(sl,0x03e0,temp);
3656 emit_shlimm(temp,2,temp);
3657 emit_writeword(temp,®_cop2d[10]);
3658 emit_andimm(sl,0x7c00,temp);
3659 emit_shrimm(temp,3,temp);
3660 emit_writeword(temp,®_cop2d[11]);
3661 emit_writeword(sl,®_cop2d[28]);
3664 emit_xorsar_imm(sl,sl,31,temp);
3665 #if defined(HAVE_ARMV5) || defined(__aarch64__)
3666 emit_clz(temp,temp);
3668 emit_movs(temp,HOST_TEMPREG);
3669 emit_movimm(0,temp);
3670 emit_jeq((int)out+4*4);
3671 emit_addpl_imm(temp,1,temp);
3672 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3673 emit_jns((int)out-2*4);
3675 emit_writeword(sl,®_cop2d[30]);
3676 emit_writeword(temp,®_cop2d[31]);
3681 emit_writeword(sl,®_cop2d[copr]);
3686 static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3691 int memtarget=0,c=0;
3693 enum stub_type type;
3694 int agr=AGEN1+(i&1);
3695 int offset_reg = -1;
3696 int fastio_reg_override = -1;
3697 u_int reglist=get_host_reglist(i_regs->regmap);
3698 u_int copr=(source[i]>>16)&0x1f;
3699 s=get_reg(i_regs->regmap,dops[i].rs1);
3700 tl=get_reg(i_regs->regmap,FTEMP);
3702 assert(dops[i].rs1>0);
3705 if(i_regs->regmap[HOST_CCREG]==CCREG)
3706 reglist&=~(1<<HOST_CCREG);
3709 if (dops[i].opcode==0x3a) { // SWC2
3710 ar=get_reg(i_regs->regmap,agr);
3711 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3716 if(s>=0) c=(i_regs->wasconst>>s)&1;
3717 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3718 if (!offset&&!c&&s>=0) ar=s;
3721 cop2_do_stall_check(0, i, i_regs, reglist);
3723 if (dops[i].opcode==0x3a) { // SWC2
3724 cop2_get_dreg(copr,tl,-1);
3732 emit_jmp(0); // inline_readstub/inline_writestub?
3736 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3737 &offset_reg, &fastio_reg_override);
3739 else if (ram_offset && memtarget) {
3740 offset_reg = get_ro_reg(i_regs, 0);
3742 switch (dops[i].opcode) {
3743 case 0x32: { // LWC2
3745 if (fastio_reg_override >= 0)
3746 a = fastio_reg_override;
3747 do_load_word(a, tl, offset_reg);
3750 case 0x3a: { // SWC2
3751 #ifdef DESTRUCTIVE_SHIFT
3752 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3755 if (fastio_reg_override >= 0)
3756 a = fastio_reg_override;
3757 do_store_word(a, 0, tl, offset_reg, 1);
3764 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3765 host_tempreg_release();
3767 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3768 if(dops[i].opcode==0x3a) // SWC2
3769 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3770 #if defined(HOST_IMM8)
3771 int ir=get_reg(i_regs->regmap,INVCP);
3773 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3775 emit_cmpmem_indexedsr12_imm(invalid_code,ar,1);
3777 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3778 emit_callne(invalidate_addr_reg[ar]);
3782 add_stub(INVCODE_STUB,jaddr3,out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3785 if (dops[i].opcode==0x32) { // LWC2
3786 host_tempreg_acquire();
3787 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3788 host_tempreg_release();
3792 static void cop2_assemble(int i, const struct regstat *i_regs)
3794 u_int copr = (source[i]>>11) & 0x1f;
3795 signed char temp = get_reg(i_regs->regmap, -1);
3797 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3798 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3799 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3800 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3801 reglist = reglist_exclude(reglist, tl, -1);
3803 cop2_do_stall_check(0, i, i_regs, reglist);
3805 if (dops[i].opcode2==0) { // MFC2
3806 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3807 if(tl>=0&&dops[i].rt1!=0)
3808 cop2_get_dreg(copr,tl,temp);
3810 else if (dops[i].opcode2==4) { // MTC2
3811 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3812 cop2_put_dreg(copr,sl,temp);
3814 else if (dops[i].opcode2==2) // CFC2
3816 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3817 if(tl>=0&&dops[i].rt1!=0)
3818 emit_readword(®_cop2c[copr],tl);
3820 else if (dops[i].opcode2==6) // CTC2
3822 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3831 emit_signextend16(sl,temp);
3834 c2op_ctc2_31_assemble(sl,temp);
3840 emit_writeword(temp,®_cop2c[copr]);
3845 static void do_unalignedwritestub(int n)
3847 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
3849 set_jump_target(stubs[n].addr, out);
3852 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3853 int addr=stubs[n].b;
3854 u_int reglist=stubs[n].e;
3855 signed char *i_regmap=i_regs->regmap;
3856 int temp2=get_reg(i_regmap,FTEMP);
3858 rt=get_reg(i_regmap,dops[i].rs2);
3861 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
3863 reglist&=~(1<<temp2);
3865 // don't bother with it and call write handler
3868 int cc=get_reg(i_regmap,CCREG);
3870 emit_loadreg(CCREG,2);
3871 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
3872 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
3873 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
3875 emit_storereg(CCREG,2);
3876 restore_regs(reglist);
3877 emit_jmp(stubs[n].retaddr); // return address
3880 #ifndef multdiv_assemble
3881 void multdiv_assemble(int i,struct regstat *i_regs)
3883 printf("Need multdiv_assemble for this architecture.\n");
3888 static void mov_assemble(int i, const struct regstat *i_regs)
3890 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
3891 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
3894 tl=get_reg(i_regs->regmap,dops[i].rt1);
3897 sl=get_reg(i_regs->regmap,dops[i].rs1);
3898 if(sl>=0) emit_mov(sl,tl);
3899 else emit_loadreg(dops[i].rs1,tl);
3902 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
3903 multdiv_do_stall(i, i_regs);
3906 // call interpreter, exception handler, things that change pc/regs/cycles ...
3907 static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
3909 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3910 assert(ccreg==HOST_CCREG);
3911 assert(!is_delayslot);
3914 emit_movimm(pc,3); // Get PC
3915 emit_readword(&last_count,2);
3916 emit_writeword(3,&psxRegs.pc);
3917 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3918 emit_add(2,HOST_CCREG,2);
3919 emit_writeword(2,&psxRegs.cycle);
3920 emit_far_call(func);
3921 emit_far_jump(jump_to_new_pc);
3924 static void syscall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3926 emit_movimm(0x20,0); // cause code
3927 emit_movimm(0,1); // not in delay slot
3928 call_c_cpu_handler(i, i_regs, ccadj_, start+i*4, psxException);
3931 static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3933 void *hlefunc = psxNULL;
3934 uint32_t hleCode = source[i] & 0x03ffffff;
3935 if (hleCode < ARRAY_SIZE(psxHLEt))
3936 hlefunc = psxHLEt[hleCode];
3938 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
3941 static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3943 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
3946 static void speculate_mov(int rs,int rt)
3949 smrv_strong_next|=1<<rt;
3954 static void speculate_mov_weak(int rs,int rt)
3957 smrv_weak_next|=1<<rt;
3962 static void speculate_register_values(int i)
3965 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
3966 // gp,sp are likely to stay the same throughout the block
3967 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
3968 smrv_weak_next=~smrv_strong_next;
3969 //printf(" llr %08x\n", smrv[4]);
3971 smrv_strong=smrv_strong_next;
3972 smrv_weak=smrv_weak_next;
3973 switch(dops[i].itype) {
3975 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
3976 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
3977 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
3978 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
3980 smrv_strong_next&=~(1<<dops[i].rt1);
3981 smrv_weak_next&=~(1<<dops[i].rt1);
3985 smrv_strong_next&=~(1<<dops[i].rt1);
3986 smrv_weak_next&=~(1<<dops[i].rt1);
3989 if(dops[i].rt1&&is_const(®s[i],dops[i].rt1)) {
3990 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
3992 if(get_final_value(hr,i,&value))
3993 smrv[dops[i].rt1]=value;
3994 else smrv[dops[i].rt1]=constmap[i][hr];
3995 smrv_strong_next|=1<<dops[i].rt1;
3999 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4000 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4004 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4005 // special case for BIOS
4006 smrv[dops[i].rt1]=0xa0000000;
4007 smrv_strong_next|=1<<dops[i].rt1;
4014 smrv_strong_next&=~(1<<dops[i].rt1);
4015 smrv_weak_next&=~(1<<dops[i].rt1);
4019 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4020 smrv_strong_next&=~(1<<dops[i].rt1);
4021 smrv_weak_next&=~(1<<dops[i].rt1);
4025 if (dops[i].opcode==0x32) { // LWC2
4026 smrv_strong_next&=~(1<<dops[i].rt1);
4027 smrv_weak_next&=~(1<<dops[i].rt1);
4033 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4034 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4038 static void ujump_assemble(int i, const struct regstat *i_regs);
4039 static void rjump_assemble(int i, const struct regstat *i_regs);
4040 static void cjump_assemble(int i, const struct regstat *i_regs);
4041 static void sjump_assemble(int i, const struct regstat *i_regs);
4042 static void pagespan_assemble(int i, const struct regstat *i_regs);
4044 static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4047 switch (dops[i].itype) {
4049 alu_assemble(i, i_regs);
4052 imm16_assemble(i, i_regs);
4055 shift_assemble(i, i_regs);
4058 shiftimm_assemble(i, i_regs);
4061 load_assemble(i, i_regs, ccadj_);
4064 loadlr_assemble(i, i_regs, ccadj_);
4067 store_assemble(i, i_regs, ccadj_);
4070 storelr_assemble(i, i_regs, ccadj_);
4073 cop0_assemble(i, i_regs, ccadj_);
4076 cop1_assemble(i, i_regs);
4079 c1ls_assemble(i, i_regs);
4082 cop2_assemble(i, i_regs);
4085 c2ls_assemble(i, i_regs, ccadj_);
4088 c2op_assemble(i, i_regs);
4091 multdiv_assemble(i, i_regs);
4092 multdiv_prepare_stall(i, i_regs, ccadj_);
4095 mov_assemble(i, i_regs);
4098 syscall_assemble(i, i_regs, ccadj_);
4101 hlecall_assemble(i, i_regs, ccadj_);
4104 intcall_assemble(i, i_regs, ccadj_);
4107 ujump_assemble(i, i_regs);
4111 rjump_assemble(i, i_regs);
4115 cjump_assemble(i, i_regs);
4119 sjump_assemble(i, i_regs);
4123 pagespan_assemble(i, i_regs);
4127 // not handled, just skip
4135 static void ds_assemble(int i, const struct regstat *i_regs)
4137 speculate_register_values(i);
4139 switch (dops[i].itype) {
4148 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4151 assemble(i, i_regs, ccadj[i]);
4156 // Is the branch target a valid internal jump?
4157 static int internal_branch(int addr)
4159 if(addr&1) return 0; // Indirect (register) jump
4160 if(addr>=start && addr<start+slen*4-4)
4167 static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4170 for(hr=0;hr<HOST_REGS;hr++) {
4171 if(hr!=EXCLUDE_REG) {
4172 if(pre[hr]!=entry[hr]) {
4175 if(get_reg(entry,pre[hr])<0) {
4177 if(!((u>>pre[hr])&1))
4178 emit_storereg(pre[hr],hr);
4185 // Move from one register to another (no writeback)
4186 for(hr=0;hr<HOST_REGS;hr++) {
4187 if(hr!=EXCLUDE_REG) {
4188 if(pre[hr]!=entry[hr]) {
4189 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
4191 if((nr=get_reg(entry,pre[hr]))>=0) {
4200 // Load the specified registers
4201 // This only loads the registers given as arguments because
4202 // we don't want to load things that will be overwritten
4203 static void load_regs(signed char entry[],signed char regmap[],int rs1,int rs2)
4207 for(hr=0;hr<HOST_REGS;hr++) {
4208 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4209 if(entry[hr]!=regmap[hr]) {
4210 if(regmap[hr]==rs1||regmap[hr]==rs2)
4217 emit_loadreg(regmap[hr],hr);
4225 // Load registers prior to the start of a loop
4226 // so that they are not loaded within the loop
4227 static void loop_preload(signed char pre[],signed char entry[])
4230 for(hr=0;hr<HOST_REGS;hr++) {
4231 if(hr!=EXCLUDE_REG) {
4232 if(pre[hr]!=entry[hr]) {
4234 if(get_reg(pre,entry[hr])<0) {
4235 assem_debug("loop preload:\n");
4236 //printf("loop preload: %d\n",hr);
4240 else if(entry[hr]<TEMPREG)
4242 emit_loadreg(entry[hr],hr);
4244 else if(entry[hr]-64<TEMPREG)
4246 emit_loadreg(entry[hr],hr);
4255 // Generate address for load/store instruction
4256 // goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4257 void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4259 if (dops[i].is_load || dops[i].is_store) {
4261 int agr=AGEN1+(i&1);
4262 if(dops[i].itype==LOAD) {
4263 ra=get_reg(i_regs->regmap,dops[i].rt1);
4264 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4267 if(dops[i].itype==LOADLR) {
4268 ra=get_reg(i_regs->regmap,FTEMP);
4270 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4271 ra=get_reg(i_regs->regmap,agr);
4272 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4274 if(dops[i].itype==C2LS) {
4275 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4276 ra=get_reg(i_regs->regmap,FTEMP);
4277 else { // SWC1/SDC1/SWC2/SDC2
4278 ra=get_reg(i_regs->regmap,agr);
4279 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4282 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4285 int c=(i_regs->wasconst>>rs)&1;
4286 if(dops[i].rs1==0) {
4287 // Using r0 as a base address
4288 if(!entry||entry[ra]!=agr) {
4289 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4290 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4291 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4292 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4294 emit_movimm(offset,ra);
4296 } // else did it in the previous cycle
4299 if(!entry||entry[ra]!=dops[i].rs1)
4300 emit_loadreg(dops[i].rs1,ra);
4301 //if(!entry||entry[ra]!=dops[i].rs1)
4302 // printf("poor load scheduling!\n");
4305 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4306 if(!entry||entry[ra]!=agr) {
4307 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4308 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4309 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4310 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4312 emit_movimm(constmap[i][rs]+offset,ra);
4313 regs[i].loadedconst|=1<<ra;
4315 } // else did it in the previous cycle
4316 } // else load_consts already did it
4318 if(offset&&!c&&dops[i].rs1) {
4320 emit_addimm(rs,offset,ra);
4322 emit_addimm(ra,offset,ra);
4327 // Preload constants for next instruction
4328 if (dops[i+1].is_load || dops[i+1].is_store) {
4331 agr=AGEN1+((i+1)&1);
4332 ra=get_reg(i_regs->regmap,agr);
4334 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4335 int offset=imm[i+1];
4336 int c=(regs[i+1].wasconst>>rs)&1;
4337 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4338 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4339 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4340 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4341 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4343 emit_movimm(constmap[i+1][rs]+offset,ra);
4344 regs[i+1].loadedconst|=1<<ra;
4347 else if(dops[i+1].rs1==0) {
4348 // Using r0 as a base address
4349 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4350 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4351 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4352 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4354 emit_movimm(offset,ra);
4361 static int get_final_value(int hr, int i, int *value)
4363 int reg=regs[i].regmap[hr];
4365 if(regs[i+1].regmap[hr]!=reg) break;
4366 if(!((regs[i+1].isconst>>hr)&1)) break;
4367 if(dops[i+1].bt) break;
4371 if (dops[i].is_jump) {
4372 *value=constmap[i][hr];
4376 if (dops[i+1].is_jump) {
4377 // Load in delay slot, out-of-order execution
4378 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4380 // Precompute load address
4381 *value=constmap[i][hr]+imm[i+2];
4385 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4387 // Precompute load address
4388 *value=constmap[i][hr]+imm[i+1];
4389 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4394 *value=constmap[i][hr];
4395 //printf("c=%lx\n",(long)constmap[i][hr]);
4396 if(i==slen-1) return 1;
4398 return !((unneeded_reg[i+1]>>reg)&1);
4401 // Load registers with known constants
4402 static void load_consts(signed char pre[],signed char regmap[],int i)
4405 // propagate loaded constant flags
4406 if(i==0||dops[i].bt)
4407 regs[i].loadedconst=0;
4409 for(hr=0;hr<HOST_REGS;hr++) {
4410 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4411 &®map[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4413 regs[i].loadedconst|=1<<hr;
4418 for(hr=0;hr<HOST_REGS;hr++) {
4419 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4420 //if(entry[hr]!=regmap[hr]) {
4421 if(!((regs[i].loadedconst>>hr)&1)) {
4422 assert(regmap[hr]<64);
4423 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
4424 int value,similar=0;
4425 if(get_final_value(hr,i,&value)) {
4426 // see if some other register has similar value
4427 for(hr2=0;hr2<HOST_REGS;hr2++) {
4428 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4429 if(is_similar_value(value,constmap[i][hr2])) {
4437 if(get_final_value(hr2,i,&value2)) // is this needed?
4438 emit_movimm_from(value2,hr2,value,hr);
4440 emit_movimm(value,hr);
4446 emit_movimm(value,hr);
4449 regs[i].loadedconst|=1<<hr;
4456 static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4460 for(hr=0;hr<HOST_REGS;hr++) {
4461 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
4462 assert(regmap[hr] < 64);
4463 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
4464 int value=constmap[i][hr];
4469 emit_movimm(value,hr);
4476 // Write out all dirty registers (except cycle count)
4477 static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4480 for(hr=0;hr<HOST_REGS;hr++) {
4481 if(hr!=EXCLUDE_REG) {
4482 if(i_regmap[hr]>0) {
4483 if(i_regmap[hr]!=CCREG) {
4484 if((i_dirty>>hr)&1) {
4485 assert(i_regmap[hr]<64);
4486 emit_storereg(i_regmap[hr],hr);
4494 // Write out dirty registers that we need to reload (pair with load_needed_regs)
4495 // This writes the registers not written by store_regs_bt
4496 static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4499 int t=(addr-start)>>2;
4500 for(hr=0;hr<HOST_REGS;hr++) {
4501 if(hr!=EXCLUDE_REG) {
4502 if(i_regmap[hr]>0) {
4503 if(i_regmap[hr]!=CCREG) {
4504 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4505 if((i_dirty>>hr)&1) {
4506 assert(i_regmap[hr]<64);
4507 emit_storereg(i_regmap[hr],hr);
4516 // Load all registers (except cycle count)
4517 static void load_all_regs(const signed char i_regmap[])
4520 for(hr=0;hr<HOST_REGS;hr++) {
4521 if(hr!=EXCLUDE_REG) {
4522 if(i_regmap[hr]==0) {
4526 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4528 emit_loadreg(i_regmap[hr],hr);
4534 // Load all current registers also needed by next instruction
4535 static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4538 for(hr=0;hr<HOST_REGS;hr++) {
4539 if(hr!=EXCLUDE_REG) {
4540 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4541 if(i_regmap[hr]==0) {
4545 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4547 emit_loadreg(i_regmap[hr],hr);
4554 // Load all regs, storing cycle count if necessary
4555 static void load_regs_entry(int t)
4558 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4559 else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t],HOST_CCREG);
4560 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4561 emit_storereg(CCREG,HOST_CCREG);
4564 for(hr=0;hr<HOST_REGS;hr++) {
4565 if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4566 if(regs[t].regmap_entry[hr]==0) {
4569 else if(regs[t].regmap_entry[hr]!=CCREG)
4571 emit_loadreg(regs[t].regmap_entry[hr],hr);
4577 // Store dirty registers prior to branch
4578 void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4580 if(internal_branch(addr))
4582 int t=(addr-start)>>2;
4584 for(hr=0;hr<HOST_REGS;hr++) {
4585 if(hr!=EXCLUDE_REG) {
4586 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4587 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4588 if((i_dirty>>hr)&1) {
4589 assert(i_regmap[hr]<64);
4590 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4591 emit_storereg(i_regmap[hr],hr);
4600 // Branch out of this block, write out all dirty regs
4601 wb_dirtys(i_regmap,i_dirty);
4605 // Load all needed registers for branch target
4606 static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4608 //if(addr>=start && addr<(start+slen*4))
4609 if(internal_branch(addr))
4611 int t=(addr-start)>>2;
4613 // Store the cycle count before loading something else
4614 if(i_regmap[HOST_CCREG]!=CCREG) {
4615 assert(i_regmap[HOST_CCREG]==-1);
4617 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4618 emit_storereg(CCREG,HOST_CCREG);
4621 for(hr=0;hr<HOST_REGS;hr++) {
4622 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4623 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4624 if(regs[t].regmap_entry[hr]==0) {
4627 else if(regs[t].regmap_entry[hr]!=CCREG)
4629 emit_loadreg(regs[t].regmap_entry[hr],hr);
4637 static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4639 if(addr>=start && addr<start+slen*4-4)
4641 int t=(addr-start)>>2;
4643 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4644 for(hr=0;hr<HOST_REGS;hr++)
4648 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4650 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4657 if(i_regmap[hr]<TEMPREG)
4659 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4662 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4668 else // Same register but is it 32-bit or dirty?
4671 if(!((regs[t].dirty>>hr)&1))
4675 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4677 //printf("%x: dirty no match\n",addr);
4685 // Delay slots are not valid branch targets
4686 //if(t>0&&(dops[t-1].is_jump) return 0;
4687 // Delay slots require additional processing, so do not match
4688 if(dops[t].is_ds) return 0;
4693 for(hr=0;hr<HOST_REGS;hr++)
4699 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4714 static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4716 extern void do_insn_cmp();
4718 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4720 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4722 // write out changed consts to match the interpreter
4723 if (i > 0 && !dops[i].bt) {
4724 for (hr = 0; hr < HOST_REGS; hr++) {
4725 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4726 if (hr == EXCLUDE_REG || reg < 0)
4728 if (!((regs[i-1].isconst >> hr) & 1))
4730 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4732 emit_movimm(constmap[i-1][hr],0);
4733 emit_storereg(reg, 0);
4736 emit_movimm(start+i*4,0);
4737 emit_writeword(0,&pcaddr);
4738 int cc = get_reg(regs[i].regmap_entry, CCREG);
4740 emit_loadreg(CCREG, cc = 0);
4741 emit_addimm(cc, ccadj_, 0);
4742 emit_writeword(0, &psxRegs.cycle);
4743 emit_far_call(do_insn_cmp);
4744 //emit_readword(&cycle,0);
4745 //emit_addimm(0,2,0);
4746 //emit_writeword(0,&cycle);
4748 restore_regs(reglist);
4749 assem_debug("\\\\do_insn_cmp\n");
4752 #define drc_dbg_emit_do_cmp(x,y)
4755 // Used when a branch jumps into the delay slot of another branch
4756 static void ds_assemble_entry(int i)
4758 int t = (ba[i] - start) >> 2;
4759 int ccadj_ = -CLOCK_ADJUST(1);
4761 instr_addr[t] = out;
4762 assem_debug("Assemble delay slot at %x\n",ba[i]);
4763 assem_debug("<->\n");
4764 drc_dbg_emit_do_cmp(t, ccadj_);
4765 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&®s[t].regmap[HOST_CCREG]!=CCREG)
4766 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4767 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4768 address_generation(t,®s[t],regs[t].regmap_entry);
4769 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4770 load_regs(regs[t].regmap_entry,regs[t].regmap,ROREG,ROREG);
4771 if (dops[t].is_store)
4772 load_regs(regs[t].regmap_entry,regs[t].regmap,INVCP,INVCP);
4774 switch (dops[t].itype) {
4783 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4786 assemble(t, ®s[t], ccadj_);
4788 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4789 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4790 if(internal_branch(ba[i]+4))
4791 assem_debug("branch: internal\n");
4793 assem_debug("branch: external\n");
4794 assert(internal_branch(ba[i]+4));
4795 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4799 static void emit_extjump(void *addr, u_int target)
4801 emit_extjump2(addr, target, dyna_linker);
4804 static void emit_extjump_ds(void *addr, u_int target)
4806 emit_extjump2(addr, target, dyna_linker_ds);
4809 // Load 2 immediates optimizing for small code size
4810 static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4812 emit_movimm(imm1,rt1);
4813 emit_movimm_from(imm1,rt1,imm2,rt2);
4816 static void do_cc(int i, const signed char i_regmap[], int *adj,
4817 int addr, int taken, int invert)
4819 int count, count_plus2;
4823 if(dops[i].itype==RJUMP)
4827 //if(ba[i]>=start && ba[i]<(start+slen*4))
4828 if(internal_branch(ba[i]))
4831 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
4839 count_plus2 = count + CLOCK_ADJUST(2);
4840 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4842 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4844 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4845 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4849 else if(*adj==0||invert) {
4850 int cycles = count_plus2;
4855 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4856 cycles=*adj+count+2-*adj;
4859 emit_addimm_and_set_flags(cycles, HOST_CCREG);
4865 emit_cmpimm(HOST_CCREG, -count_plus2);
4869 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
4872 static void do_ccstub(int n)
4875 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
4876 set_jump_target(stubs[n].addr, out);
4878 if(stubs[n].d==NULLDS) {
4879 // Delay slot instruction is nullified ("likely" branch)
4880 wb_dirtys(regs[i].regmap,regs[i].dirty);
4882 else if(stubs[n].d!=TAKEN) {
4883 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
4886 if(internal_branch(ba[i]))
4887 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4891 // Save PC as return address
4892 emit_movimm(stubs[n].c,EAX);
4893 emit_writeword(EAX,&pcaddr);
4897 // Return address depends on which way the branch goes
4898 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
4900 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
4901 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
4907 else if(dops[i].rs2==0)
4912 #ifdef DESTRUCTIVE_WRITEBACK
4914 if((branch_regs[i].dirty>>s1l)&&1)
4915 emit_loadreg(dops[i].rs1,s1l);
4918 if((branch_regs[i].dirty>>s1l)&1)
4919 emit_loadreg(dops[i].rs2,s1l);
4922 if((branch_regs[i].dirty>>s2l)&1)
4923 emit_loadreg(dops[i].rs2,s2l);
4926 int addr=-1,alt=-1,ntaddr=-1;
4929 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4930 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4931 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4939 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4940 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4941 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4947 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
4951 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4952 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4953 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4959 assert(hr<HOST_REGS);
4961 if((dops[i].opcode&0x2f)==4) // BEQ
4963 #ifdef HAVE_CMOV_IMM
4964 if(s2l>=0) emit_cmp(s1l,s2l);
4965 else emit_test(s1l,s1l);
4966 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4968 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4969 if(s2l>=0) emit_cmp(s1l,s2l);
4970 else emit_test(s1l,s1l);
4971 emit_cmovne_reg(alt,addr);
4974 if((dops[i].opcode&0x2f)==5) // BNE
4976 #ifdef HAVE_CMOV_IMM
4977 if(s2l>=0) emit_cmp(s1l,s2l);
4978 else emit_test(s1l,s1l);
4979 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4981 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4982 if(s2l>=0) emit_cmp(s1l,s2l);
4983 else emit_test(s1l,s1l);
4984 emit_cmovne_reg(alt,addr);
4987 if((dops[i].opcode&0x2f)==6) // BLEZ
4989 //emit_movimm(ba[i],alt);
4990 //emit_movimm(start+i*4+8,addr);
4991 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4993 emit_cmovl_reg(alt,addr);
4995 if((dops[i].opcode&0x2f)==7) // BGTZ
4997 //emit_movimm(ba[i],addr);
4998 //emit_movimm(start+i*4+8,ntaddr);
4999 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5001 emit_cmovl_reg(ntaddr,addr);
5003 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
5005 //emit_movimm(ba[i],alt);
5006 //emit_movimm(start+i*4+8,addr);
5007 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5009 emit_cmovs_reg(alt,addr);
5011 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
5013 //emit_movimm(ba[i],addr);
5014 //emit_movimm(start+i*4+8,alt);
5015 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5017 emit_cmovs_reg(alt,addr);
5019 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5020 if(source[i]&0x10000) // BC1T
5022 //emit_movimm(ba[i],alt);
5023 //emit_movimm(start+i*4+8,addr);
5024 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5025 emit_testimm(s1l,0x800000);
5026 emit_cmovne_reg(alt,addr);
5030 //emit_movimm(ba[i],addr);
5031 //emit_movimm(start+i*4+8,alt);
5032 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5033 emit_testimm(s1l,0x800000);
5034 emit_cmovne_reg(alt,addr);
5037 emit_writeword(addr,&pcaddr);
5040 if(dops[i].itype==RJUMP)
5042 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5043 if (ds_writes_rjump_rs(i)) {
5044 r=get_reg(branch_regs[i].regmap,RTEMP);
5046 emit_writeword(r,&pcaddr);
5048 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5050 // Update cycle count
5051 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5052 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5053 emit_far_call(cc_interrupt);
5054 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5055 if(stubs[n].d==TAKEN) {
5056 if(internal_branch(ba[i]))
5057 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
5058 else if(dops[i].itype==RJUMP) {
5059 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5060 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5062 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5064 }else if(stubs[n].d==NOTTAKEN) {
5065 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5066 else load_all_regs(branch_regs[i].regmap);
5067 }else if(stubs[n].d==NULLDS) {
5068 // Delay slot instruction is nullified ("likely" branch)
5069 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5070 else load_all_regs(regs[i].regmap);
5072 load_all_regs(branch_regs[i].regmap);
5074 if (stubs[n].retaddr)
5075 emit_jmp(stubs[n].retaddr);
5077 do_jump_vaddr(stubs[n].e);
5080 static void add_to_linker(void *addr, u_int target, int ext)
5082 assert(linkcount < ARRAY_SIZE(link_addr));
5083 link_addr[linkcount].addr = addr;
5084 link_addr[linkcount].target = target;
5085 link_addr[linkcount].ext = ext;
5089 static void ujump_assemble_write_ra(int i)
5092 unsigned int return_address;
5093 rt=get_reg(branch_regs[i].regmap,31);
5094 assem_debug("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
5096 return_address=start+i*4+8;
5099 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5100 int temp=-1; // note: must be ds-safe
5104 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5105 else emit_movimm(return_address,rt);
5113 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5116 emit_movimm(return_address,rt); // PC into link register
5118 emit_prefetch(hash_table_get(return_address));
5124 static void ujump_assemble(int i, const struct regstat *i_regs)
5127 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5128 address_generation(i+1,i_regs,regs[i].regmap_entry);
5130 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5131 if(dops[i].rt1==31&&temp>=0)
5133 signed char *i_regmap=i_regs->regmap;
5134 int return_address=start+i*4+8;
5135 if(get_reg(branch_regs[i].regmap,31)>0)
5136 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5139 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5140 ujump_assemble_write_ra(i); // writeback ra for DS
5143 ds_assemble(i+1,i_regs);
5144 uint64_t bc_unneeded=branch_regs[i].u;
5145 bc_unneeded|=1|(1LL<<dops[i].rt1);
5146 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5147 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5148 if(!ra_done&&dops[i].rt1==31)
5149 ujump_assemble_write_ra(i);
5151 cc=get_reg(branch_regs[i].regmap,CCREG);
5152 assert(cc==HOST_CCREG);
5153 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5155 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5157 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5158 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5159 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5160 if(internal_branch(ba[i]))
5161 assem_debug("branch: internal\n");
5163 assem_debug("branch: external\n");
5164 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5165 ds_assemble_entry(i);
5168 add_to_linker(out,ba[i],internal_branch(ba[i]));
5173 static void rjump_assemble_write_ra(int i)
5175 int rt,return_address;
5176 assert(dops[i+1].rt1!=dops[i].rt1);
5177 assert(dops[i+1].rt2!=dops[i].rt1);
5178 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5179 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]);
5181 return_address=start+i*4+8;
5185 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5188 emit_movimm(return_address,rt); // PC into link register
5190 emit_prefetch(hash_table_get(return_address));
5194 static void rjump_assemble(int i, const struct regstat *i_regs)
5199 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5201 if (ds_writes_rjump_rs(i)) {
5202 // Delay slot abuse, make a copy of the branch address register
5203 temp=get_reg(branch_regs[i].regmap,RTEMP);
5205 assert(regs[i].regmap[temp]==RTEMP);
5209 address_generation(i+1,i_regs,regs[i].regmap_entry);
5213 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5214 signed char *i_regmap=i_regs->regmap;
5215 int return_address=start+i*4+8;
5216 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5221 if(dops[i].rs1==31) {
5222 int rh=get_reg(regs[i].regmap,RHASH);
5223 if(rh>=0) do_preload_rhash(rh);
5226 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5227 rjump_assemble_write_ra(i);
5230 ds_assemble(i+1,i_regs);
5231 uint64_t bc_unneeded=branch_regs[i].u;
5232 bc_unneeded|=1|(1LL<<dops[i].rt1);
5233 bc_unneeded&=~(1LL<<dops[i].rs1);
5234 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5235 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5236 if(!ra_done&&dops[i].rt1!=0)
5237 rjump_assemble_write_ra(i);
5238 cc=get_reg(branch_regs[i].regmap,CCREG);
5239 assert(cc==HOST_CCREG);
5242 int rh=get_reg(branch_regs[i].regmap,RHASH);
5243 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5244 if(dops[i].rs1==31) {
5245 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5246 do_preload_rhtbl(ht);
5250 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5251 #ifdef DESTRUCTIVE_WRITEBACK
5252 if((branch_regs[i].dirty>>rs)&1) {
5253 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5254 emit_loadreg(dops[i].rs1,rs);
5259 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5262 if(dops[i].rs1==31) {
5263 do_miniht_load(ht,rh);
5266 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5267 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5269 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5270 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5271 if(dops[i+1].itype==COP0&&(source[i+1]&0x3f)==0x10)
5272 // special case for RFE
5276 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5278 if(dops[i].rs1==31) {
5279 do_miniht_jump(rs,rh,ht);
5286 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5287 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5291 static void cjump_assemble(int i, const struct regstat *i_regs)
5293 const signed char *i_regmap = i_regs->regmap;
5296 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5297 assem_debug("match=%d\n",match);
5299 int unconditional=0,nop=0;
5301 int internal=internal_branch(ba[i]);
5302 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5303 if(!match) invert=1;
5304 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5305 if(i>(ba[i]-start)>>2) invert=1;
5308 invert=1; // because of near cond. branches
5312 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5313 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5316 s1l=get_reg(i_regmap,dops[i].rs1);
5317 s2l=get_reg(i_regmap,dops[i].rs2);
5319 if(dops[i].rs1==0&&dops[i].rs2==0)
5321 if(dops[i].opcode&1) nop=1;
5322 else unconditional=1;
5323 //assert(dops[i].opcode!=5);
5324 //assert(dops[i].opcode!=7);
5325 //assert(dops[i].opcode!=0x15);
5326 //assert(dops[i].opcode!=0x17);
5328 else if(dops[i].rs1==0)
5333 else if(dops[i].rs2==0)
5339 // Out of order execution (delay slot first)
5341 address_generation(i+1,i_regs,regs[i].regmap_entry);
5342 ds_assemble(i+1,i_regs);
5344 uint64_t bc_unneeded=branch_regs[i].u;
5345 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5347 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5348 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5349 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5350 cc=get_reg(branch_regs[i].regmap,CCREG);
5351 assert(cc==HOST_CCREG);
5353 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5354 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5355 //assem_debug("cycle count (adj)\n");
5357 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5358 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5359 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5360 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5362 assem_debug("branch: internal\n");
5364 assem_debug("branch: external\n");
5365 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5366 ds_assemble_entry(i);
5369 add_to_linker(out,ba[i],internal);
5372 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5373 if(((u_int)out)&7) emit_addnop(0);
5378 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5381 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5384 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5385 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5386 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5388 //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]);
5390 if(dops[i].opcode==4) // BEQ
5392 if(s2l>=0) emit_cmp(s1l,s2l);
5393 else emit_test(s1l,s1l);
5398 add_to_linker(out,ba[i],internal);
5402 if(dops[i].opcode==5) // BNE
5404 if(s2l>=0) emit_cmp(s1l,s2l);
5405 else emit_test(s1l,s1l);
5410 add_to_linker(out,ba[i],internal);
5414 if(dops[i].opcode==6) // BLEZ
5421 add_to_linker(out,ba[i],internal);
5425 if(dops[i].opcode==7) // BGTZ
5432 add_to_linker(out,ba[i],internal);
5437 if(taken) set_jump_target(taken, out);
5438 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5439 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5441 emit_addimm(cc,-adj,cc);
5442 add_to_linker(out,ba[i],internal);
5445 add_to_linker(out,ba[i],internal*2);
5451 if(adj) emit_addimm(cc,-adj,cc);
5452 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5453 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5455 assem_debug("branch: internal\n");
5457 assem_debug("branch: external\n");
5458 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5459 ds_assemble_entry(i);
5462 add_to_linker(out,ba[i],internal);
5466 set_jump_target(nottaken, out);
5469 if(nottaken1) set_jump_target(nottaken1, out);
5471 if(!invert) emit_addimm(cc,adj,cc);
5473 } // (!unconditional)
5477 // In-order execution (branch first)
5478 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5479 if(!unconditional&&!nop) {
5480 //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]);
5482 if((dops[i].opcode&0x2f)==4) // BEQ
5484 if(s2l>=0) emit_cmp(s1l,s2l);
5485 else emit_test(s1l,s1l);
5489 if((dops[i].opcode&0x2f)==5) // BNE
5491 if(s2l>=0) emit_cmp(s1l,s2l);
5492 else emit_test(s1l,s1l);
5496 if((dops[i].opcode&0x2f)==6) // BLEZ
5502 if((dops[i].opcode&0x2f)==7) // BGTZ
5508 } // if(!unconditional)
5510 uint64_t ds_unneeded=branch_regs[i].u;
5511 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5515 if(taken) set_jump_target(taken, out);
5516 assem_debug("1:\n");
5517 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5519 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5520 address_generation(i+1,&branch_regs[i],0);
5522 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5523 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5524 ds_assemble(i+1,&branch_regs[i]);
5525 cc=get_reg(branch_regs[i].regmap,CCREG);
5527 emit_loadreg(CCREG,cc=HOST_CCREG);
5528 // CHECK: Is the following instruction (fall thru) allocated ok?
5530 assert(cc==HOST_CCREG);
5531 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5532 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5533 assem_debug("cycle count (adj)\n");
5534 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5535 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5537 assem_debug("branch: internal\n");
5539 assem_debug("branch: external\n");
5540 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5541 ds_assemble_entry(i);
5544 add_to_linker(out,ba[i],internal);
5549 if(!unconditional) {
5550 if(nottaken1) set_jump_target(nottaken1, out);
5551 set_jump_target(nottaken, out);
5552 assem_debug("2:\n");
5553 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5555 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5556 address_generation(i+1,&branch_regs[i],0);
5558 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5559 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5560 ds_assemble(i+1,&branch_regs[i]);
5561 cc=get_reg(branch_regs[i].regmap,CCREG);
5563 // Cycle count isn't in a register, temporarily load it then write it out
5564 emit_loadreg(CCREG,HOST_CCREG);
5565 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5568 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5569 emit_storereg(CCREG,HOST_CCREG);
5572 cc=get_reg(i_regmap,CCREG);
5573 assert(cc==HOST_CCREG);
5574 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5577 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5583 static void sjump_assemble(int i, const struct regstat *i_regs)
5585 const signed char *i_regmap = i_regs->regmap;
5588 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5589 assem_debug("smatch=%d\n",match);
5591 int unconditional=0,nevertaken=0;
5593 int internal=internal_branch(ba[i]);
5594 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5595 if(!match) invert=1;
5596 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5597 if(i>(ba[i]-start)>>2) invert=1;
5600 invert=1; // because of near cond. branches
5603 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5604 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5607 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5610 s1l=get_reg(i_regmap,dops[i].rs1);
5614 if(dops[i].opcode2&1) unconditional=1;
5616 // These are never taken (r0 is never less than zero)
5617 //assert(dops[i].opcode2!=0);
5618 //assert(dops[i].opcode2!=2);
5619 //assert(dops[i].opcode2!=0x10);
5620 //assert(dops[i].opcode2!=0x12);
5624 // Out of order execution (delay slot first)
5626 address_generation(i+1,i_regs,regs[i].regmap_entry);
5627 ds_assemble(i+1,i_regs);
5629 uint64_t bc_unneeded=branch_regs[i].u;
5630 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5632 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5633 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5634 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5635 if(dops[i].rt1==31) {
5636 int rt,return_address;
5637 rt=get_reg(branch_regs[i].regmap,31);
5638 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]);
5640 // Save the PC even if the branch is not taken
5641 return_address=start+i*4+8;
5642 emit_movimm(return_address,rt); // PC into link register
5644 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5648 cc=get_reg(branch_regs[i].regmap,CCREG);
5649 assert(cc==HOST_CCREG);
5651 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5652 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5653 assem_debug("cycle count (adj)\n");
5655 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5656 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5657 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5658 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5660 assem_debug("branch: internal\n");
5662 assem_debug("branch: external\n");
5663 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5664 ds_assemble_entry(i);
5667 add_to_linker(out,ba[i],internal);
5670 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5671 if(((u_int)out)&7) emit_addnop(0);
5675 else if(nevertaken) {
5676 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5679 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5682 void *nottaken = NULL;
5683 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5684 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5687 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5694 add_to_linker(out,ba[i],internal);
5698 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5705 add_to_linker(out,ba[i],internal);
5712 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5713 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5715 emit_addimm(cc,-adj,cc);
5716 add_to_linker(out,ba[i],internal);
5719 add_to_linker(out,ba[i],internal*2);
5725 if(adj) emit_addimm(cc,-adj,cc);
5726 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5727 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5729 assem_debug("branch: internal\n");
5731 assem_debug("branch: external\n");
5732 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5733 ds_assemble_entry(i);
5736 add_to_linker(out,ba[i],internal);
5740 set_jump_target(nottaken, out);
5744 if(!invert) emit_addimm(cc,adj,cc);
5746 } // (!unconditional)
5750 // In-order execution (branch first)
5752 void *nottaken = NULL;
5753 if(dops[i].rt1==31) {
5754 int rt,return_address;
5755 rt=get_reg(branch_regs[i].regmap,31);
5757 // Save the PC even if the branch is not taken
5758 return_address=start+i*4+8;
5759 emit_movimm(return_address,rt); // PC into link register
5761 emit_prefetch(hash_table_get(return_address));
5765 if(!unconditional) {
5766 //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]);
5768 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5774 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5780 } // if(!unconditional)
5782 uint64_t ds_unneeded=branch_regs[i].u;
5783 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5787 //assem_debug("1:\n");
5788 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5790 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5791 address_generation(i+1,&branch_regs[i],0);
5793 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5794 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5795 ds_assemble(i+1,&branch_regs[i]);
5796 cc=get_reg(branch_regs[i].regmap,CCREG);
5798 emit_loadreg(CCREG,cc=HOST_CCREG);
5799 // CHECK: Is the following instruction (fall thru) allocated ok?
5801 assert(cc==HOST_CCREG);
5802 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5803 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5804 assem_debug("cycle count (adj)\n");
5805 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5806 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5808 assem_debug("branch: internal\n");
5810 assem_debug("branch: external\n");
5811 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5812 ds_assemble_entry(i);
5815 add_to_linker(out,ba[i],internal);
5820 if(!unconditional) {
5821 set_jump_target(nottaken, out);
5822 assem_debug("1:\n");
5823 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5824 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5825 address_generation(i+1,&branch_regs[i],0);
5826 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5827 ds_assemble(i+1,&branch_regs[i]);
5828 cc=get_reg(branch_regs[i].regmap,CCREG);
5830 // Cycle count isn't in a register, temporarily load it then write it out
5831 emit_loadreg(CCREG,HOST_CCREG);
5832 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5835 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5836 emit_storereg(CCREG,HOST_CCREG);
5839 cc=get_reg(i_regmap,CCREG);
5840 assert(cc==HOST_CCREG);
5841 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5844 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5850 static void pagespan_assemble(int i, const struct regstat *i_regs)
5852 int s1l=get_reg(i_regs->regmap,dops[i].rs1);
5853 int s2l=get_reg(i_regs->regmap,dops[i].rs2);
5855 void *nottaken = NULL;
5856 int unconditional=0;
5862 else if(dops[i].rs2==0)
5867 int addr=-1,alt=-1,ntaddr=-1;
5868 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5872 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5873 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5874 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5883 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5884 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5885 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5891 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5895 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5896 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5897 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5904 assert(hr<HOST_REGS);
5905 if((dops[i].opcode&0x2e)==4||dops[i].opcode==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5906 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
5908 emit_addimm(HOST_CCREG, ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5909 if(dops[i].opcode==2) // J
5913 if(dops[i].opcode==3) // JAL
5916 int rt=get_reg(i_regs->regmap,31);
5917 emit_movimm(start+i*4+8,rt);
5920 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
5923 if(dops[i].opcode2==9) // JALR
5925 int rt=get_reg(i_regs->regmap,dops[i].rt1);
5926 emit_movimm(start+i*4+8,rt);
5929 if((dops[i].opcode&0x3f)==4) // BEQ
5931 if(dops[i].rs1==dops[i].rs2)
5936 #ifdef HAVE_CMOV_IMM
5938 if(s2l>=0) emit_cmp(s1l,s2l);
5939 else emit_test(s1l,s1l);
5940 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5946 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5947 if(s2l>=0) emit_cmp(s1l,s2l);
5948 else emit_test(s1l,s1l);
5949 emit_cmovne_reg(alt,addr);
5952 if((dops[i].opcode&0x3f)==5) // BNE
5954 #ifdef HAVE_CMOV_IMM
5955 if(s2l>=0) emit_cmp(s1l,s2l);
5956 else emit_test(s1l,s1l);
5957 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5960 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5961 if(s2l>=0) emit_cmp(s1l,s2l);
5962 else emit_test(s1l,s1l);
5963 emit_cmovne_reg(alt,addr);
5966 if((dops[i].opcode&0x3f)==0x14) // BEQL
5968 if(s2l>=0) emit_cmp(s1l,s2l);
5969 else emit_test(s1l,s1l);
5970 if(nottaken) set_jump_target(nottaken, out);
5974 if((dops[i].opcode&0x3f)==0x15) // BNEL
5976 if(s2l>=0) emit_cmp(s1l,s2l);
5977 else emit_test(s1l,s1l);
5980 if(taken) set_jump_target(taken, out);
5982 if((dops[i].opcode&0x3f)==6) // BLEZ
5984 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5986 emit_cmovl_reg(alt,addr);
5988 if((dops[i].opcode&0x3f)==7) // BGTZ
5990 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5992 emit_cmovl_reg(ntaddr,addr);
5994 if((dops[i].opcode&0x3f)==0x16) // BLEZL
5996 assert((dops[i].opcode&0x3f)!=0x16);
5998 if((dops[i].opcode&0x3f)==0x17) // BGTZL
6000 assert((dops[i].opcode&0x3f)!=0x17);
6002 assert(dops[i].opcode!=1); // BLTZ/BGEZ
6004 //FIXME: Check CSREG
6005 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
6006 if((source[i]&0x30000)==0) // BC1F
6008 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6009 emit_testimm(s1l,0x800000);
6010 emit_cmovne_reg(alt,addr);
6012 if((source[i]&0x30000)==0x10000) // BC1T
6014 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6015 emit_testimm(s1l,0x800000);
6016 emit_cmovne_reg(alt,addr);
6018 if((source[i]&0x30000)==0x20000) // BC1FL
6020 emit_testimm(s1l,0x800000);
6024 if((source[i]&0x30000)==0x30000) // BC1TL
6026 emit_testimm(s1l,0x800000);
6032 assert(i_regs->regmap[HOST_CCREG]==CCREG);
6033 wb_dirtys(regs[i].regmap,regs[i].dirty);
6036 emit_movimm(ba[i],HOST_BTREG);
6038 else if(addr!=HOST_BTREG)
6040 emit_mov(addr,HOST_BTREG);
6042 void *branch_addr=out;
6044 int target_addr=start+i*4+5;
6046 void *compiled_target_addr=check_addr(target_addr);
6047 emit_extjump_ds(branch_addr, target_addr);
6048 if(compiled_target_addr) {
6049 set_jump_target(branch_addr, compiled_target_addr);
6050 add_jump_out(target_addr,stub);
6052 else set_jump_target(branch_addr, stub);
6055 // Assemble the delay slot for the above
6056 static void pagespan_ds()
6058 assem_debug("initial delay slot:\n");
6059 u_int vaddr=start+1;
6060 u_int page=get_page(vaddr);
6061 u_int vpage=get_vpage(vaddr);
6062 ll_add(jump_dirty+vpage,vaddr,(void *)out);
6063 do_dirty_stub_ds(slen*4);
6064 ll_add(jump_in+page,vaddr,(void *)out);
6065 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
6066 if(regs[0].regmap[HOST_CCREG]!=CCREG)
6067 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
6068 if(regs[0].regmap[HOST_BTREG]!=BTREG)
6069 emit_writeword(HOST_BTREG,&branch_target);
6070 load_regs(regs[0].regmap_entry,regs[0].regmap,dops[0].rs1,dops[0].rs2);
6071 address_generation(0,®s[0],regs[0].regmap_entry);
6072 if (ram_offset && (dops[0].is_load || dops[0].is_store))
6073 load_regs(regs[0].regmap_entry,regs[0].regmap,ROREG,ROREG);
6074 if (dops[0].is_store)
6075 load_regs(regs[0].regmap_entry,regs[0].regmap,INVCP,INVCP);
6077 switch (dops[0].itype) {
6086 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
6089 assemble(0, ®s[0], 0);
6091 int btaddr=get_reg(regs[0].regmap,BTREG);
6093 btaddr=get_reg(regs[0].regmap,-1);
6094 emit_readword(&branch_target,btaddr);
6096 assert(btaddr!=HOST_CCREG);
6097 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
6099 host_tempreg_acquire();
6100 emit_movimm(start+4,HOST_TEMPREG);
6101 emit_cmp(btaddr,HOST_TEMPREG);
6102 host_tempreg_release();
6104 emit_cmpimm(btaddr,start+4);
6108 store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
6109 do_jump_vaddr(btaddr);
6110 set_jump_target(branch, out);
6111 store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
6112 load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
6115 // Basic liveness analysis for MIPS registers
6116 void unneeded_registers(int istart,int iend,int r)
6119 uint64_t u,gte_u,b,gte_b;
6120 uint64_t temp_u,temp_gte_u=0;
6121 uint64_t gte_u_unknown=0;
6122 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
6126 gte_u=gte_u_unknown;
6128 //u=unneeded_reg[iend+1];
6130 gte_u=gte_unneeded[iend+1];
6133 for (i=iend;i>=istart;i--)
6135 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6138 // If subroutine call, flag return address as a possible branch target
6139 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
6141 if(ba[i]<start || ba[i]>=(start+slen*4))
6143 // Branch out of this block, flush all regs
6145 gte_u=gte_u_unknown;
6146 branch_unneeded_reg[i]=u;
6147 // Merge in delay slot
6148 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6149 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6152 gte_u&=~gte_rs[i+1];
6156 // Internal branch, flag target
6157 dops[(ba[i]-start)>>2].bt=1;
6158 if(ba[i]<=start+i*4) {
6160 if(dops[i].is_ujump)
6162 // Unconditional branch
6166 // Conditional branch (not taken case)
6167 temp_u=unneeded_reg[i+2];
6168 temp_gte_u&=gte_unneeded[i+2];
6170 // Merge in delay slot
6171 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6172 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6174 temp_gte_u|=gte_rt[i+1];
6175 temp_gte_u&=~gte_rs[i+1];
6176 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
6177 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
6179 temp_gte_u|=gte_rt[i];
6180 temp_gte_u&=~gte_rs[i];
6181 unneeded_reg[i]=temp_u;
6182 gte_unneeded[i]=temp_gte_u;
6183 // Only go three levels deep. This recursion can take an
6184 // excessive amount of time if there are a lot of nested loops.
6186 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6188 unneeded_reg[(ba[i]-start)>>2]=1;
6189 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
6192 if (dops[i].is_ujump)
6194 // Unconditional branch
6195 u=unneeded_reg[(ba[i]-start)>>2];
6196 gte_u=gte_unneeded[(ba[i]-start)>>2];
6197 branch_unneeded_reg[i]=u;
6198 // Merge in delay slot
6199 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6200 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6203 gte_u&=~gte_rs[i+1];
6205 // Conditional branch
6206 b=unneeded_reg[(ba[i]-start)>>2];
6207 gte_b=gte_unneeded[(ba[i]-start)>>2];
6208 branch_unneeded_reg[i]=b;
6209 // Branch delay slot
6210 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6211 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6214 gte_b&=~gte_rs[i+1];
6218 branch_unneeded_reg[i]&=unneeded_reg[i+2];
6220 branch_unneeded_reg[i]=1;
6226 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6228 // SYSCALL instruction (software interrupt)
6231 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6233 // ERET instruction (return from interrupt)
6237 // Written registers are unneeded
6238 u|=1LL<<dops[i].rt1;
6239 u|=1LL<<dops[i].rt2;
6241 // Accessed registers are needed
6242 u&=~(1LL<<dops[i].rs1);
6243 u&=~(1LL<<dops[i].rs2);
6245 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
6246 gte_u|=gte_rs[i]>e_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
6247 // Source-target dependencies
6248 // R0 is always unneeded
6252 gte_unneeded[i]=gte_u;
6254 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6257 for(r=1;r<=CCREG;r++) {
6258 if((unneeded_reg[i]>>r)&1) {
6259 if(r==HIREG) printf(" HI");
6260 else if(r==LOREG) printf(" LO");
6261 else printf(" r%d",r);
6269 // Write back dirty registers as soon as we will no longer modify them,
6270 // so that we don't end up with lots of writes at the branches.
6271 void clean_registers(int istart,int iend,int wr)
6275 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
6276 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
6278 will_dirty_i=will_dirty_next=0;
6279 wont_dirty_i=wont_dirty_next=0;
6281 will_dirty_i=will_dirty_next=will_dirty[iend+1];
6282 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
6284 for (i=iend;i>=istart;i--)
6288 if(ba[i]<start || ba[i]>=(start+slen*4))
6290 // Branch out of this block, flush all regs
6291 if (dops[i].is_ujump)
6293 // Unconditional branch
6296 // Merge in delay slot (will dirty)
6297 for(r=0;r<HOST_REGS;r++) {
6298 if(r!=EXCLUDE_REG) {
6299 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6300 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6301 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6302 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6303 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6304 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6305 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6306 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6307 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6308 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6309 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6310 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6311 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6312 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6318 // Conditional branch
6320 wont_dirty_i=wont_dirty_next;
6321 // Merge in delay slot (will dirty)
6322 for(r=0;r<HOST_REGS;r++) {
6323 if(r!=EXCLUDE_REG) {
6324 if (1) { // !dops[i].likely) {
6325 // Might not dirty if likely branch is not taken
6326 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6327 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6328 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6329 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6330 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6331 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
6332 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6333 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6334 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6335 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6336 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6337 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6338 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6339 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6344 // Merge in delay slot (wont dirty)
6345 for(r=0;r<HOST_REGS;r++) {
6346 if(r!=EXCLUDE_REG) {
6347 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6348 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6349 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6350 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6351 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6352 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6353 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6354 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6355 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6356 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6360 #ifndef DESTRUCTIVE_WRITEBACK
6361 branch_regs[i].dirty&=wont_dirty_i;
6363 branch_regs[i].dirty|=will_dirty_i;
6369 if(ba[i]<=start+i*4) {
6371 if (dops[i].is_ujump)
6373 // Unconditional branch
6376 // Merge in delay slot (will dirty)
6377 for(r=0;r<HOST_REGS;r++) {
6378 if(r!=EXCLUDE_REG) {
6379 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6380 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6381 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6382 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6383 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6384 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6385 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6386 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6387 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6388 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6389 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6390 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6391 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6392 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6396 // Conditional branch (not taken case)
6397 temp_will_dirty=will_dirty_next;
6398 temp_wont_dirty=wont_dirty_next;
6399 // Merge in delay slot (will dirty)
6400 for(r=0;r<HOST_REGS;r++) {
6401 if(r!=EXCLUDE_REG) {
6402 if (1) { // !dops[i].likely) {
6403 // Will not dirty if likely branch is not taken
6404 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6405 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6406 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6407 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6408 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6409 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
6410 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6411 //if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6412 //if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6413 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6414 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6415 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6416 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6417 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6422 // Merge in delay slot (wont dirty)
6423 for(r=0;r<HOST_REGS;r++) {
6424 if(r!=EXCLUDE_REG) {
6425 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6426 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6427 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6428 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6429 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6430 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6431 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6432 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6433 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6434 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6437 // Deal with changed mappings
6439 for(r=0;r<HOST_REGS;r++) {
6440 if(r!=EXCLUDE_REG) {
6441 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
6442 temp_will_dirty&=~(1<<r);
6443 temp_wont_dirty&=~(1<<r);
6444 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6445 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6446 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6448 temp_will_dirty|=1<<r;
6449 temp_wont_dirty|=1<<r;
6456 will_dirty[i]=temp_will_dirty;
6457 wont_dirty[i]=temp_wont_dirty;
6458 clean_registers((ba[i]-start)>>2,i-1,0);
6460 // Limit recursion. It can take an excessive amount
6461 // of time if there are a lot of nested loops.
6462 will_dirty[(ba[i]-start)>>2]=0;
6463 wont_dirty[(ba[i]-start)>>2]=-1;
6468 if (dops[i].is_ujump)
6470 // Unconditional branch
6473 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6474 for(r=0;r<HOST_REGS;r++) {
6475 if(r!=EXCLUDE_REG) {
6476 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6477 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
6478 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6480 if(branch_regs[i].regmap[r]>=0) {
6481 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6482 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6487 // Merge in delay slot
6488 for(r=0;r<HOST_REGS;r++) {
6489 if(r!=EXCLUDE_REG) {
6490 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6491 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6492 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6493 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6494 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6495 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6496 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6497 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6498 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6499 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6500 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6501 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6502 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6503 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6507 // Conditional branch
6508 will_dirty_i=will_dirty_next;
6509 wont_dirty_i=wont_dirty_next;
6510 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6511 for(r=0;r<HOST_REGS;r++) {
6512 if(r!=EXCLUDE_REG) {
6513 signed char target_reg=branch_regs[i].regmap[r];
6514 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6515 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
6516 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6518 else if(target_reg>=0) {
6519 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6520 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6525 // Merge in delay slot
6526 for(r=0;r<HOST_REGS;r++) {
6527 if(r!=EXCLUDE_REG) {
6528 if (1) { // !dops[i].likely) {
6529 // Might not dirty if likely branch is not taken
6530 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6531 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6532 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6533 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6534 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6535 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6536 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6537 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6538 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6539 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6540 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6541 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6542 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6543 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6548 // Merge in delay slot (won't dirty)
6549 for(r=0;r<HOST_REGS;r++) {
6550 if(r!=EXCLUDE_REG) {
6551 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6552 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6553 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6554 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6555 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6556 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6557 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6558 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6559 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6560 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6564 #ifndef DESTRUCTIVE_WRITEBACK
6565 branch_regs[i].dirty&=wont_dirty_i;
6567 branch_regs[i].dirty|=will_dirty_i;
6572 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6574 // SYSCALL instruction (software interrupt)
6578 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6580 // ERET instruction (return from interrupt)
6584 will_dirty_next=will_dirty_i;
6585 wont_dirty_next=wont_dirty_i;
6586 for(r=0;r<HOST_REGS;r++) {
6587 if(r!=EXCLUDE_REG) {
6588 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6589 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6590 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6591 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6592 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6593 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6594 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6595 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6597 if (!dops[i].is_jump)
6599 // Don't store a register immediately after writing it,
6600 // may prevent dual-issue.
6601 if((regs[i].regmap[r]&63)==dops[i-1].rt1) wont_dirty_i|=1<<r;
6602 if((regs[i].regmap[r]&63)==dops[i-1].rt2) wont_dirty_i|=1<<r;
6608 will_dirty[i]=will_dirty_i;
6609 wont_dirty[i]=wont_dirty_i;
6610 // Mark registers that won't be dirtied as not dirty
6612 regs[i].dirty|=will_dirty_i;
6613 #ifndef DESTRUCTIVE_WRITEBACK
6614 regs[i].dirty&=wont_dirty_i;
6617 if (i < iend-1 && !dops[i].is_ujump) {
6618 for(r=0;r<HOST_REGS;r++) {
6619 if(r!=EXCLUDE_REG) {
6620 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
6621 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
6622 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6630 for(r=0;r<HOST_REGS;r++) {
6631 if(r!=EXCLUDE_REG) {
6632 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
6633 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
6634 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6642 // Deal with changed mappings
6643 temp_will_dirty=will_dirty_i;
6644 temp_wont_dirty=wont_dirty_i;
6645 for(r=0;r<HOST_REGS;r++) {
6646 if(r!=EXCLUDE_REG) {
6648 if(regs[i].regmap[r]==regmap_pre[i][r]) {
6650 #ifndef DESTRUCTIVE_WRITEBACK
6651 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6653 regs[i].wasdirty|=will_dirty_i&(1<<r);
6656 else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
6657 // Register moved to a different register
6658 will_dirty_i&=~(1<<r);
6659 wont_dirty_i&=~(1<<r);
6660 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
6661 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
6663 #ifndef DESTRUCTIVE_WRITEBACK
6664 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6666 regs[i].wasdirty|=will_dirty_i&(1<<r);
6670 will_dirty_i&=~(1<<r);
6671 wont_dirty_i&=~(1<<r);
6672 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6673 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6674 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6677 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
6687 void disassemble_inst(int i)
6689 if (dops[i].bt) printf("*"); else printf(" ");
6690 switch(dops[i].itype) {
6692 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6694 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;
6696 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;
6698 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6699 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6701 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6704 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,ba[i]);break;
6706 if(dops[i].opcode==0xf) //LUI
6707 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6709 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6713 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6717 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6721 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6724 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6727 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6730 if((dops[i].opcode2&0x1d)==0x10)
6731 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6732 else if((dops[i].opcode2&0x1d)==0x11)
6733 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6735 printf (" %x: %s\n",start+i*4,insn[i]);
6738 if(dops[i].opcode2==0)
6739 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6740 else if(dops[i].opcode2==4)
6741 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6742 else printf (" %x: %s\n",start+i*4,insn[i]);
6745 if(dops[i].opcode2<3)
6746 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6747 else if(dops[i].opcode2>3)
6748 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6749 else printf (" %x: %s\n",start+i*4,insn[i]);
6752 if(dops[i].opcode2<3)
6753 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6754 else if(dops[i].opcode2>3)
6755 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6756 else printf (" %x: %s\n",start+i*4,insn[i]);
6759 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6762 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6765 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6768 //printf (" %s %8x\n",insn[i],source[i]);
6769 printf (" %x: %s\n",start+i*4,insn[i]);
6773 static void disassemble_inst(int i) {}
6776 #define DRC_TEST_VAL 0x74657374
6778 static void new_dynarec_test(void)
6780 int (*testfunc)(void);
6785 // check structure linkage
6786 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6788 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6791 SysPrintf("testing if we can run recompiled code...\n");
6792 ((volatile u_int *)out)[0]++; // make cache dirty
6794 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6795 out = ndrc->translation_cache;
6796 beginning = start_block();
6797 emit_movimm(DRC_TEST_VAL + i, 0); // test
6800 end_block(beginning);
6801 testfunc = beginning;
6802 ret[i] = testfunc();
6805 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6806 SysPrintf("test passed.\n");
6808 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6809 out = ndrc->translation_cache;
6812 // clear the state completely, instead of just marking
6813 // things invalid like invalidate_all_pages() does
6814 void new_dynarec_clear_full(void)
6817 out = ndrc->translation_cache;
6818 memset(invalid_code,1,sizeof(invalid_code));
6819 memset(hash_table,0xff,sizeof(hash_table));
6820 memset(mini_ht,-1,sizeof(mini_ht));
6821 memset(restore_candidate,0,sizeof(restore_candidate));
6822 memset(shadow,0,sizeof(shadow));
6824 expirep=16384; // Expiry pointer, +2 blocks
6825 pending_exception=0;
6828 inv_code_start=inv_code_end=~0;
6831 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6832 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6833 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6835 cycle_multiplier_old = cycle_multiplier;
6836 new_dynarec_hacks_old = new_dynarec_hacks;
6839 void new_dynarec_init(void)
6841 SysPrintf("Init new dynarec\n");
6843 #ifdef BASE_ADDR_DYNAMIC
6845 sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
6847 SysPrintf("sceKernelAllocMemBlockForVM failed\n");
6848 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6850 SysPrintf("sceKernelGetMemBlockBase failed\n");
6852 uintptr_t desired_addr = 0;
6855 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6857 ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
6858 PROT_READ | PROT_WRITE | PROT_EXEC,
6859 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
6860 if (ndrc == MAP_FAILED) {
6861 SysPrintf("mmap() failed: %s\n", strerror(errno));
6866 #ifndef NO_WRITE_EXEC
6867 // not all systems allow execute in data segment by default
6868 if (mprotect(ndrc, sizeof(ndrc->translation_cache) + sizeof(ndrc->tramp.ops),
6869 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6870 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6873 out = ndrc->translation_cache;
6874 cycle_multiplier=200;
6875 new_dynarec_clear_full();
6877 // Copy this into local area so we don't have to put it in every literal pool
6878 invc_ptr=invalid_code;
6882 ram_offset=(uintptr_t)rdram-0x80000000;
6884 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6887 void new_dynarec_cleanup(void)
6890 #ifdef BASE_ADDR_DYNAMIC
6892 sceKernelFreeMemBlock(sceBlock);
6895 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6896 SysPrintf("munmap() failed\n");
6899 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6900 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6901 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6903 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
6907 static u_int *get_source_start(u_int addr, u_int *limit)
6909 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6910 cycle_multiplier_override = 0;
6912 if (addr < 0x00200000 ||
6913 (0xa0000000 <= addr && addr < 0xa0200000))
6915 // used for BIOS calls mostly?
6916 *limit = (addr&0xa0000000)|0x00200000;
6917 return (u_int *)(rdram + (addr&0x1fffff));
6919 else if (!Config.HLE && (
6920 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6921 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6923 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6924 // but timings in PCSX are too tied to the interpreter's BIAS
6925 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6926 cycle_multiplier_override = 200;
6928 *limit = (addr & 0xfff00000) | 0x80000;
6929 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6931 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6932 *limit = (addr & 0x80600000) + 0x00200000;
6933 return (u_int *)(rdram + (addr&0x1fffff));
6938 static u_int scan_for_ret(u_int addr)
6943 mem = get_source_start(addr, &limit);
6947 if (limit > addr + 0x1000)
6948 limit = addr + 0x1000;
6949 for (; addr < limit; addr += 4, mem++) {
6950 if (*mem == 0x03e00008) // jr $ra
6956 struct savestate_block {
6961 static int addr_cmp(const void *p1_, const void *p2_)
6963 const struct savestate_block *p1 = p1_, *p2 = p2_;
6964 return p1->addr - p2->addr;
6967 int new_dynarec_save_blocks(void *save, int size)
6969 struct savestate_block *blocks = save;
6970 int maxcount = size / sizeof(blocks[0]);
6971 struct savestate_block tmp_blocks[1024];
6972 struct ll_entry *head;
6973 int p, s, d, o, bcnt;
6977 for (p = 0; p < ARRAY_SIZE(jump_in); p++) {
6979 for (head = jump_in[p]; head != NULL; head = head->next) {
6980 tmp_blocks[bcnt].addr = head->vaddr;
6981 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
6986 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6988 addr = tmp_blocks[0].addr;
6989 for (s = d = 0; s < bcnt; s++) {
6990 if (tmp_blocks[s].addr < addr)
6992 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6993 tmp_blocks[d++] = tmp_blocks[s];
6994 addr = scan_for_ret(tmp_blocks[s].addr);
6997 if (o + d > maxcount)
6999 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
7003 return o * sizeof(blocks[0]);
7006 void new_dynarec_load_blocks(const void *save, int size)
7008 const struct savestate_block *blocks = save;
7009 int count = size / sizeof(blocks[0]);
7010 u_int regs_save[32];
7014 get_addr(psxRegs.pc);
7016 // change GPRs for speculation to at least partially work..
7017 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
7018 for (i = 1; i < 32; i++)
7019 psxRegs.GPR.r[i] = 0x80000000;
7021 for (b = 0; b < count; b++) {
7022 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
7024 psxRegs.GPR.r[i] = 0x1f800000;
7027 get_addr(blocks[b].addr);
7029 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
7031 psxRegs.GPR.r[i] = 0x80000000;
7035 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
7038 int new_recompile_block(u_int addr)
7040 u_int pagelimit = 0;
7041 u_int state_rflags = 0;
7044 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
7045 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
7047 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
7049 // this is just for speculation
7050 for (i = 1; i < 32; i++) {
7051 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
7052 state_rflags |= 1 << i;
7055 start = (u_int)addr&~3;
7056 //assert(((u_int)addr&1)==0); // start-in-delay-slot flag
7057 new_dynarec_did_compile=1;
7058 if (Config.HLE && start == 0x80001000) // hlecall
7060 // XXX: is this enough? Maybe check hleSoftCall?
7061 void *beginning=start_block();
7062 u_int page=get_page(start);
7064 invalid_code[start>>12]=0;
7065 emit_movimm(start,0);
7066 emit_writeword(0,&pcaddr);
7067 emit_far_jump(new_dyna_leave);
7069 end_block(beginning);
7070 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
7073 else if (f1_hack == ~0u || (f1_hack != 0 && start == f1_hack)) {
7074 void *beginning = start_block();
7075 u_int page = get_page(start);
7076 emit_readword(&psxRegs.GPR.n.sp, 0);
7077 emit_readptr(&mem_rtab, 1);
7078 emit_shrimm(0, 12, 2);
7079 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
7080 emit_addimm(0, 0x18, 0);
7081 emit_adds_ptr(1, 1, 1);
7082 emit_ldr_dualindexed(1, 0, 0);
7083 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
7084 emit_far_call(get_addr_ht);
7085 emit_jmpreg(0); // jr k0
7087 end_block(beginning);
7089 ll_add_flags(jump_in + page, start, state_rflags, beginning);
7090 SysPrintf("F1 hack to %08x\n", start);
7095 source = get_source_start(start, &pagelimit);
7096 if (source == NULL) {
7097 SysPrintf("Compile at bogus memory address: %08x\n", addr);
7101 /* Pass 1: disassemble */
7102 /* Pass 2: register dependencies, branch targets */
7103 /* Pass 3: register allocation */
7104 /* Pass 4: branch dependencies */
7105 /* Pass 5: pre-alloc */
7106 /* Pass 6: optimize clean/dirty state */
7107 /* Pass 7: flag 32-bit registers */
7108 /* Pass 8: assembly */
7109 /* Pass 9: linker */
7110 /* Pass 10: garbage collection / free memory */
7114 unsigned int type,op,op2;
7116 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
7118 /* Pass 1 disassembly */
7120 for(i=0;!done;i++) {
7124 minimum_free_regs[i]=0;
7125 dops[i].opcode=op=source[i]>>26;
7128 case 0x00: strcpy(insn[i],"special"); type=NI;
7132 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
7133 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
7134 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
7135 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
7136 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
7137 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
7138 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
7139 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
7140 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
7141 case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
7142 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
7143 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
7144 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
7145 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
7146 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
7147 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
7148 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
7149 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
7150 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
7151 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
7152 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
7153 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
7154 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
7155 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
7156 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
7157 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
7158 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
7159 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
7160 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
7161 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
7162 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
7163 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
7164 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
7165 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
7166 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
7168 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
7169 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
7170 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
7171 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
7172 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
7173 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
7174 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
7175 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
7176 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
7177 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
7178 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
7179 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
7180 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
7181 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
7182 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
7183 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
7184 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
7188 case 0x01: strcpy(insn[i],"regimm"); type=NI;
7189 op2=(source[i]>>16)&0x1f;
7192 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
7193 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
7194 //case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
7195 //case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
7196 //case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
7197 //case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
7198 //case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
7199 //case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
7200 //case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
7201 //case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
7202 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
7203 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
7204 //case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
7205 //case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
7208 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
7209 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
7210 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
7211 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
7212 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
7213 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
7214 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
7215 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
7216 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
7217 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
7218 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
7219 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
7220 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
7221 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
7222 case 0x10: strcpy(insn[i],"cop0"); type=NI;
7223 op2=(source[i]>>21)&0x1f;
7226 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
7227 case 0x02: strcpy(insn[i],"CFC0"); type=COP0; break;
7228 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
7229 case 0x06: strcpy(insn[i],"CTC0"); type=COP0; break;
7230 case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
7233 case 0x11: strcpy(insn[i],"cop1"); type=COP1;
7234 op2=(source[i]>>21)&0x1f;
7237 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
7238 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
7239 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
7240 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
7241 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
7242 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
7243 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
7244 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
7246 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
7247 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
7248 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
7249 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
7250 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
7251 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
7252 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
7254 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
7256 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
7257 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
7258 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
7259 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
7261 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
7262 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
7264 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
7265 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
7266 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
7267 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
7269 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
7270 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
7271 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
7273 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
7274 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
7276 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
7277 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
7278 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
7280 case 0x12: strcpy(insn[i],"COP2"); type=NI;
7281 op2=(source[i]>>21)&0x1f;
7283 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
7284 if (gte_handlers[source[i]&0x3f]!=NULL) {
7285 if (gte_regnames[source[i]&0x3f]!=NULL)
7286 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
7288 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
7294 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
7295 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
7296 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
7297 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
7300 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
7301 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
7302 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
7303 default: strcpy(insn[i],"???"); type=NI;
7304 SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
7308 dops[i].opcode2=op2;
7309 /* Get registers/immediates */
7311 gte_rs[i]=gte_rt[i]=0;
7314 dops[i].rs1=(source[i]>>21)&0x1f;
7316 dops[i].rt1=(source[i]>>16)&0x1f;
7318 imm[i]=(short)source[i];
7322 dops[i].rs1=(source[i]>>21)&0x1f;
7323 dops[i].rs2=(source[i]>>16)&0x1f;
7326 imm[i]=(short)source[i];
7329 // LWL/LWR only load part of the register,
7330 // therefore the target register must be treated as a source too
7331 dops[i].rs1=(source[i]>>21)&0x1f;
7332 dops[i].rs2=(source[i]>>16)&0x1f;
7333 dops[i].rt1=(source[i]>>16)&0x1f;
7335 imm[i]=(short)source[i];
7338 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
7339 else dops[i].rs1=(source[i]>>21)&0x1f;
7341 dops[i].rt1=(source[i]>>16)&0x1f;
7343 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
7344 imm[i]=(unsigned short)source[i];
7346 imm[i]=(short)source[i];
7354 // The JAL instruction writes to r31.
7361 dops[i].rs1=(source[i]>>21)&0x1f;
7365 // The JALR instruction writes to rd.
7367 dops[i].rt1=(source[i]>>11)&0x1f;
7372 dops[i].rs1=(source[i]>>21)&0x1f;
7373 dops[i].rs2=(source[i]>>16)&0x1f;
7376 if(op&2) { // BGTZ/BLEZ
7381 dops[i].rs1=(source[i]>>21)&0x1f;
7385 if(op2&0x10) { // BxxAL
7387 // NOTE: If the branch is not taken, r31 is still overwritten
7391 dops[i].rs1=(source[i]>>21)&0x1f; // source
7392 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
7393 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7397 dops[i].rs1=(source[i]>>21)&0x1f; // source
7398 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
7407 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
7408 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
7409 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
7410 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
7411 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
7412 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
7415 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
7416 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
7417 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7421 dops[i].rs1=(source[i]>>16)&0x1f;
7423 dops[i].rt1=(source[i]>>11)&0x1f;
7425 imm[i]=(source[i]>>6)&0x1f;
7426 // DSxx32 instructions
7427 if(op2>=0x3c) imm[i]|=0x20;
7434 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
7435 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
7436 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
7437 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
7444 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
7445 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
7453 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
7454 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
7456 int gr=(source[i]>>11)&0x1F;
7459 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
7460 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
7461 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
7462 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
7466 dops[i].rs1=(source[i]>>21)&0x1F;
7470 imm[i]=(short)source[i];
7473 dops[i].rs1=(source[i]>>21)&0x1F;
7477 imm[i]=(short)source[i];
7478 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
7479 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
7486 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
7487 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
7488 gte_rt[i]|=1ll<<63; // every op changes flags
7489 if((source[i]&0x3f)==GTE_MVMVA) {
7490 int v = (source[i] >> 15) & 3;
7491 gte_rs[i]&=~0xe3fll;
7492 if(v==3) gte_rs[i]|=0xe00ll;
7493 else gte_rs[i]|=3ll<<(v*2);
7510 /* Calculate branch target addresses */
7512 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7513 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
7514 ba[i]=start+i*4+8; // Ignore never taken branch
7515 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
7516 ba[i]=start+i*4+8; // Ignore never taken branch
7517 else if(type==CJUMP||type==SJUMP)
7518 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7521 /* simplify always (not)taken branches */
7522 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
7523 dops[i].rs1 = dops[i].rs2 = 0;
7525 dops[i].itype = type = UJUMP;
7526 dops[i].rs2 = CCREG;
7529 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
7530 dops[i].itype = type = UJUMP;
7532 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
7533 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
7534 dops[i].is_load = (dops[i].itype == LOAD || dops[i].itype == LOADLR || op == 0x32); // LWC2
7535 dops[i].is_store = (dops[i].itype == STORE || dops[i].itype == STORELR || op == 0x3a); // SWC2
7537 /* messy cases to just pass over to the interpreter */
7538 if (i > 0 && dops[i-1].is_jump) {
7540 // branch in delay slot?
7541 if (dops[i].is_jump) {
7542 // don't handle first branch and call interpreter if it's hit
7543 SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
7546 // basic load delay detection
7547 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
7548 int t=(ba[i-1]-start)/4;
7549 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) {
7550 // jump target wants DS result - potential load delay effect
7551 SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
7553 dops[t+1].bt=1; // expected return from interpreter
7555 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&&
7556 !(i>=3&&dops[i-3].is_jump)) {
7557 // v0 overwrite like this is a sign of trouble, bail out
7558 SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
7563 dops[i-1].rs1=CCREG;
7564 dops[i-1].rs2=dops[i-1].rt1=dops[i-1].rt2=0;
7566 dops[i-1].itype=INTCALL;
7568 i--; // don't compile the DS
7572 /* Is this the end of the block? */
7573 if (i > 0 && dops[i-1].is_ujump) {
7574 if(dops[i-1].rt1==0) { // Continue past subroutine call (JAL)
7578 if(stop_after_jal) done=1;
7580 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7582 // Don't recompile stuff that's already compiled
7583 if(check_addr(start+i*4+4)) done=1;
7584 // Don't get too close to the limit
7585 if(i>MAXBLOCK/2) done=1;
7587 if(dops[i].itype==SYSCALL&&stop_after_jal) done=1;
7588 if(dops[i].itype==HLECALL||dops[i].itype==INTCALL) done=2;
7590 // Does the block continue due to a branch?
7593 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7594 if(ba[j]==start+i*4+4) done=j=0;
7595 if(ba[j]==start+i*4+8) done=j=0;
7598 //assert(i<MAXBLOCK-1);
7599 if(start+i*4==pagelimit-4) done=1;
7600 assert(start+i*4<pagelimit);
7601 if (i==MAXBLOCK-1) done=1;
7602 // Stop if we're compiling junk
7603 if(dops[i].itype==NI&&dops[i].opcode==0x11) {
7604 done=stop_after_jal=1;
7605 SysPrintf("Disabled speculative precompilation\n");
7609 if (dops[i-1].is_jump) {
7610 if(start+i*4==pagelimit) {
7611 dops[i-1].itype=SPAN;
7616 /* spacial hack(s) */
7617 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
7618 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
7619 && dops[i-7].itype == STORE)
7622 if (dops[i].itype == IMM16)
7624 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
7625 if (dops[i].itype == STORELR && dops[i].rs1 == 6
7626 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
7628 SysPrintf("F1 hack from %08x\n", start);
7634 /* Pass 2 - Register dependencies and branch targets */
7636 unneeded_registers(0,slen-1,0);
7638 /* Pass 3 - Register allocation */
7640 struct regstat current; // Current register allocations/status
7642 current.u=unneeded_reg[0];
7643 clear_all_regs(current.regmap);
7644 alloc_reg(¤t,0,CCREG);
7645 dirty_reg(¤t,CCREG);
7648 current.waswritten=0;
7654 // First instruction is delay slot
7659 current.regmap[HOST_BTREG]=BTREG;
7667 for(hr=0;hr<HOST_REGS;hr++)
7669 // Is this really necessary?
7670 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7673 current.waswritten=0;
7676 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7677 regs[i].wasconst=current.isconst;
7678 regs[i].wasdirty=current.dirty;
7679 regs[i].loadedconst=0;
7680 if (!dops[i].is_jump) {
7682 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7689 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7690 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7692 } else { SysPrintf("oops, branch at end of block with no delay slot\n");abort(); }
7696 ds=0; // Skip delay slot, already allocated as part of branch
7697 // ...but we need to alloc it in case something jumps here
7699 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7701 current.u=branch_unneeded_reg[i-1];
7703 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7705 struct regstat temp;
7706 memcpy(&temp,¤t,sizeof(current));
7707 temp.wasdirty=temp.dirty;
7708 // TODO: Take into account unconditional branches, as below
7709 delayslot_alloc(&temp,i);
7710 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7711 regs[i].wasdirty=temp.wasdirty;
7712 regs[i].dirty=temp.dirty;
7716 // Create entry (branch target) regmap
7717 for(hr=0;hr<HOST_REGS;hr++)
7719 int r=temp.regmap[hr];
7721 if(r!=regmap_pre[i][hr]) {
7722 regs[i].regmap_entry[hr]=-1;
7727 if((current.u>>r)&1) {
7728 regs[i].regmap_entry[hr]=-1;
7729 regs[i].regmap[hr]=-1;
7730 //Don't clear regs in the delay slot as the branch might need them
7731 //current.regmap[hr]=-1;
7733 regs[i].regmap_entry[hr]=r;
7736 // First instruction expects CCREG to be allocated
7737 if(i==0&&hr==HOST_CCREG)
7738 regs[i].regmap_entry[hr]=CCREG;
7740 regs[i].regmap_entry[hr]=-1;
7744 else { // Not delay slot
7745 switch(dops[i].itype) {
7747 //current.isconst=0; // DEBUG
7748 //current.wasconst=0; // DEBUG
7749 //regs[i].wasconst=0; // DEBUG
7750 clear_const(¤t,dops[i].rt1);
7751 alloc_cc(¤t,i);
7752 dirty_reg(¤t,CCREG);
7753 if (dops[i].rt1==31) {
7754 alloc_reg(¤t,i,31);
7755 dirty_reg(¤t,31);
7756 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7757 //assert(dops[i+1].rt1!=dops[i].rt1);
7759 alloc_reg(¤t,i,PTEMP);
7763 delayslot_alloc(¤t,i+1);
7764 //current.isconst=0; // DEBUG
7766 //printf("i=%d, isconst=%x\n",i,current.isconst);
7769 //current.isconst=0;
7770 //current.wasconst=0;
7771 //regs[i].wasconst=0;
7772 clear_const(¤t,dops[i].rs1);
7773 clear_const(¤t,dops[i].rt1);
7774 alloc_cc(¤t,i);
7775 dirty_reg(¤t,CCREG);
7776 if (!ds_writes_rjump_rs(i)) {
7777 alloc_reg(¤t,i,dops[i].rs1);
7778 if (dops[i].rt1!=0) {
7779 alloc_reg(¤t,i,dops[i].rt1);
7780 dirty_reg(¤t,dops[i].rt1);
7781 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7782 assert(dops[i+1].rt1!=dops[i].rt1);
7784 alloc_reg(¤t,i,PTEMP);
7788 if(dops[i].rs1==31) { // JALR
7789 alloc_reg(¤t,i,RHASH);
7790 alloc_reg(¤t,i,RHTBL);
7793 delayslot_alloc(¤t,i+1);
7795 // The delay slot overwrites our source register,
7796 // allocate a temporary register to hold the old value.
7800 delayslot_alloc(¤t,i+1);
7802 alloc_reg(¤t,i,RTEMP);
7804 //current.isconst=0; // DEBUG
7809 //current.isconst=0;
7810 //current.wasconst=0;
7811 //regs[i].wasconst=0;
7812 clear_const(¤t,dops[i].rs1);
7813 clear_const(¤t,dops[i].rs2);
7814 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7816 alloc_cc(¤t,i);
7817 dirty_reg(¤t,CCREG);
7818 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7819 if(dops[i].rs2) alloc_reg(¤t,i,dops[i].rs2);
7820 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7821 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7822 // The delay slot overwrites one of our conditions.
7823 // Allocate the branch condition registers instead.
7827 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7828 if(dops[i].rs2) alloc_reg(¤t,i,dops[i].rs2);
7833 delayslot_alloc(¤t,i+1);
7837 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7839 alloc_cc(¤t,i);
7840 dirty_reg(¤t,CCREG);
7841 alloc_reg(¤t,i,dops[i].rs1);
7842 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7843 // The delay slot overwrites one of our conditions.
7844 // Allocate the branch condition registers instead.
7848 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7853 delayslot_alloc(¤t,i+1);
7857 // Don't alloc the delay slot yet because we might not execute it
7858 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7863 alloc_cc(¤t,i);
7864 dirty_reg(¤t,CCREG);
7865 alloc_reg(¤t,i,dops[i].rs1);
7866 alloc_reg(¤t,i,dops[i].rs2);
7869 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7874 alloc_cc(¤t,i);
7875 dirty_reg(¤t,CCREG);
7876 alloc_reg(¤t,i,dops[i].rs1);
7879 //current.isconst=0;
7882 //current.isconst=0;
7883 //current.wasconst=0;
7884 //regs[i].wasconst=0;
7885 clear_const(¤t,dops[i].rs1);
7886 clear_const(¤t,dops[i].rt1);
7887 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7888 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7890 alloc_cc(¤t,i);
7891 dirty_reg(¤t,CCREG);
7892 alloc_reg(¤t,i,dops[i].rs1);
7893 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7894 alloc_reg(¤t,i,31);
7895 dirty_reg(¤t,31);
7896 //#ifdef REG_PREFETCH
7897 //alloc_reg(¤t,i,PTEMP);
7900 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.
7901 ||(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
7902 // Allocate the branch condition registers instead.
7906 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7911 delayslot_alloc(¤t,i+1);
7915 // Don't alloc the delay slot yet because we might not execute it
7916 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7921 alloc_cc(¤t,i);
7922 dirty_reg(¤t,CCREG);
7923 alloc_reg(¤t,i,dops[i].rs1);
7926 //current.isconst=0;
7929 imm16_alloc(¤t,i);
7933 load_alloc(¤t,i);
7937 store_alloc(¤t,i);
7940 alu_alloc(¤t,i);
7943 shift_alloc(¤t,i);
7946 multdiv_alloc(¤t,i);
7949 shiftimm_alloc(¤t,i);
7952 mov_alloc(¤t,i);
7955 cop0_alloc(¤t,i);
7960 cop2_alloc(¤t,i);
7963 c1ls_alloc(¤t,i);
7966 c2ls_alloc(¤t,i);
7969 c2op_alloc(¤t,i);
7974 syscall_alloc(¤t,i);
7977 pagespan_alloc(¤t,i);
7981 // Create entry (branch target) regmap
7982 for(hr=0;hr<HOST_REGS;hr++)
7985 r=current.regmap[hr];
7987 if(r!=regmap_pre[i][hr]) {
7988 // TODO: delay slot (?)
7989 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7990 if(or<0||(r&63)>=TEMPREG){
7991 regs[i].regmap_entry[hr]=-1;
7995 // Just move it to a different register
7996 regs[i].regmap_entry[hr]=r;
7997 // If it was dirty before, it's still dirty
7998 if((regs[i].wasdirty>>or)&1) dirty_reg(¤t,r&63);
8005 regs[i].regmap_entry[hr]=0;
8010 if((current.u>>r)&1) {
8011 regs[i].regmap_entry[hr]=-1;
8012 //regs[i].regmap[hr]=-1;
8013 current.regmap[hr]=-1;
8015 regs[i].regmap_entry[hr]=r;
8019 // Branches expect CCREG to be allocated at the target
8020 if(regmap_pre[i][hr]==CCREG)
8021 regs[i].regmap_entry[hr]=CCREG;
8023 regs[i].regmap_entry[hr]=-1;
8026 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
8029 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)
8030 current.waswritten|=1<<dops[i-1].rs1;
8031 current.waswritten&=~(1<<dops[i].rt1);
8032 current.waswritten&=~(1<<dops[i].rt2);
8033 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
8034 current.waswritten&=~(1<<dops[i].rs1);
8036 /* Branch post-alloc */
8039 current.wasdirty=current.dirty;
8040 switch(dops[i-1].itype) {
8042 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8043 branch_regs[i-1].isconst=0;
8044 branch_regs[i-1].wasconst=0;
8045 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
8046 alloc_cc(&branch_regs[i-1],i-1);
8047 dirty_reg(&branch_regs[i-1],CCREG);
8048 if(dops[i-1].rt1==31) { // JAL
8049 alloc_reg(&branch_regs[i-1],i-1,31);
8050 dirty_reg(&branch_regs[i-1],31);
8052 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8053 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8056 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8057 branch_regs[i-1].isconst=0;
8058 branch_regs[i-1].wasconst=0;
8059 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
8060 alloc_cc(&branch_regs[i-1],i-1);
8061 dirty_reg(&branch_regs[i-1],CCREG);
8062 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
8063 if(dops[i-1].rt1!=0) { // JALR
8064 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
8065 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
8068 if(dops[i-1].rs1==31) { // JALR
8069 alloc_reg(&branch_regs[i-1],i-1,RHASH);
8070 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
8073 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8074 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8077 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
8079 alloc_cc(¤t,i-1);
8080 dirty_reg(¤t,CCREG);
8081 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
8082 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
8083 // The delay slot overwrote one of our conditions
8084 // Delay slot goes after the test (in order)
8085 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8087 delayslot_alloc(¤t,i);
8092 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
8093 // Alloc the branch condition registers
8094 if(dops[i-1].rs1) alloc_reg(¤t,i-1,dops[i-1].rs1);
8095 if(dops[i-1].rs2) alloc_reg(¤t,i-1,dops[i-1].rs2);
8097 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8098 branch_regs[i-1].isconst=0;
8099 branch_regs[i-1].wasconst=0;
8100 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8101 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8104 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
8106 alloc_cc(¤t,i-1);
8107 dirty_reg(¤t,CCREG);
8108 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
8109 // The delay slot overwrote the branch condition
8110 // Delay slot goes after the test (in order)
8111 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8113 delayslot_alloc(¤t,i);
8118 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
8119 // Alloc the branch condition register
8120 alloc_reg(¤t,i-1,dops[i-1].rs1);
8122 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8123 branch_regs[i-1].isconst=0;
8124 branch_regs[i-1].wasconst=0;
8125 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8126 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8129 // Alloc the delay slot in case the branch is taken
8130 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
8132 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8133 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;
8134 alloc_cc(&branch_regs[i-1],i);
8135 dirty_reg(&branch_regs[i-1],CCREG);
8136 delayslot_alloc(&branch_regs[i-1],i);
8137 branch_regs[i-1].isconst=0;
8138 alloc_reg(¤t,i,CCREG); // Not taken path
8139 dirty_reg(¤t,CCREG);
8140 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8143 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
8145 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8146 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;
8147 alloc_cc(&branch_regs[i-1],i);
8148 dirty_reg(&branch_regs[i-1],CCREG);
8149 delayslot_alloc(&branch_regs[i-1],i);
8150 branch_regs[i-1].isconst=0;
8151 alloc_reg(¤t,i,CCREG); // Not taken path
8152 dirty_reg(¤t,CCREG);
8153 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8157 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
8158 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
8160 alloc_cc(¤t,i-1);
8161 dirty_reg(¤t,CCREG);
8162 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
8163 // The delay slot overwrote the branch condition
8164 // Delay slot goes after the test (in order)
8165 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8167 delayslot_alloc(¤t,i);
8172 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
8173 // Alloc the branch condition register
8174 alloc_reg(¤t,i-1,dops[i-1].rs1);
8176 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8177 branch_regs[i-1].isconst=0;
8178 branch_regs[i-1].wasconst=0;
8179 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
8180 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8183 // Alloc the delay slot in case the branch is taken
8184 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
8186 memcpy(&branch_regs[i-1],¤t,sizeof(current));
8187 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;
8188 alloc_cc(&branch_regs[i-1],i);
8189 dirty_reg(&branch_regs[i-1],CCREG);
8190 delayslot_alloc(&branch_regs[i-1],i);
8191 branch_regs[i-1].isconst=0;
8192 alloc_reg(¤t,i,CCREG); // Not taken path
8193 dirty_reg(¤t,CCREG);
8194 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8196 // FIXME: BLTZAL/BGEZAL
8197 if(dops[i-1].opcode2&0x10) { // BxxZAL
8198 alloc_reg(&branch_regs[i-1],i-1,31);
8199 dirty_reg(&branch_regs[i-1],31);
8204 if (dops[i-1].is_ujump)
8206 if(dops[i-1].rt1==31) // JAL/JALR
8208 // Subroutine call will return here, don't alloc any registers
8210 clear_all_regs(current.regmap);
8211 alloc_reg(¤t,i,CCREG);
8212 dirty_reg(¤t,CCREG);
8216 // Internal branch will jump here, match registers to caller
8218 clear_all_regs(current.regmap);
8219 alloc_reg(¤t,i,CCREG);
8220 dirty_reg(¤t,CCREG);
8223 if(ba[j]==start+i*4+4) {
8224 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
8225 current.dirty=branch_regs[j].dirty;
8230 if(ba[j]==start+i*4+4) {
8231 for(hr=0;hr<HOST_REGS;hr++) {
8232 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
8233 current.regmap[hr]=-1;
8235 current.dirty&=branch_regs[j].dirty;
8244 // Count cycles in between branches
8245 ccadj[i] = CLOCK_ADJUST(cc);
8246 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
8250 #if !defined(DRC_DBG)
8251 else if(dops[i].itype==C2OP&>e_cycletab[source[i]&0x3f]>2)
8253 // this should really be removed since the real stalls have been implemented,
8254 // but doing so causes sizeable perf regression against the older version
8255 u_int gtec = gte_cycletab[source[i] & 0x3f];
8256 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
8258 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
8262 else if(dops[i].itype==C2LS)
8264 // same as with C2OP
8265 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
8273 if(!dops[i].is_ds) {
8274 regs[i].dirty=current.dirty;
8275 regs[i].isconst=current.isconst;
8276 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
8278 for(hr=0;hr<HOST_REGS;hr++) {
8279 if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
8280 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
8281 regs[i].wasconst&=~(1<<hr);
8285 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
8286 regs[i].waswritten=current.waswritten;
8289 /* Pass 4 - Cull unused host registers */
8293 for (i=slen-1;i>=0;i--)
8298 if(ba[i]<start || ba[i]>=(start+slen*4))
8300 // Branch out of this block, don't need anything
8306 // Need whatever matches the target
8308 int t=(ba[i]-start)>>2;
8309 for(hr=0;hr<HOST_REGS;hr++)
8311 if(regs[i].regmap_entry[hr]>=0) {
8312 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
8316 // Conditional branch may need registers for following instructions
8317 if (!dops[i].is_ujump)
8320 nr|=needed_reg[i+2];
8321 for(hr=0;hr<HOST_REGS;hr++)
8323 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
8324 //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]);
8328 // Don't need stuff which is overwritten
8329 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8330 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8331 // Merge in delay slot
8332 for(hr=0;hr<HOST_REGS;hr++)
8334 if(dops[i+1].rt1&&dops[i+1].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8335 if(dops[i+1].rt2&&dops[i+1].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8336 if(dops[i+1].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8337 if(dops[i+1].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8338 if(dops[i+1].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8339 if(dops[i+1].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8340 if(ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
8341 if(regmap_pre[i][hr]==ROREG) nr|=1<<hr;
8342 if(regs[i].regmap_entry[hr]==ROREG) nr|=1<<hr;
8344 if(dops[i+1].is_store) {
8345 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8346 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8350 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8352 // SYSCALL instruction (software interrupt)
8355 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8357 // ERET instruction (return from interrupt)
8363 for(hr=0;hr<HOST_REGS;hr++) {
8364 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8365 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8366 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8367 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8371 for(hr=0;hr<HOST_REGS;hr++)
8373 // Overwritten registers are not needed
8374 if(dops[i].rt1&&dops[i].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8375 if(dops[i].rt2&&dops[i].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8376 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8377 // Source registers are needed
8378 if(dops[i].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8379 if(dops[i].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8380 if(dops[i].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8381 if(dops[i].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8382 if(ram_offset && (dops[i].is_load || dops[i].is_store)) {
8383 if(regmap_pre[i][hr]==ROREG) nr|=1<<hr;
8384 if(regs[i].regmap_entry[hr]==ROREG) nr|=1<<hr;
8386 if(dops[i].is_store) {
8387 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8388 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8390 // Don't store a register immediately after writing it,
8391 // may prevent dual-issue.
8392 // But do so if this is a branch target, otherwise we
8393 // might have to load the register before the branch.
8394 if(i>0&&!dops[i].bt&&((regs[i].wasdirty>>hr)&1)) {
8395 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
8396 if(dops[i-1].rt1==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8397 if(dops[i-1].rt2==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8399 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
8400 if(dops[i-1].rt1==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8401 if(dops[i-1].rt2==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8405 // Cycle count is needed at branches. Assume it is needed at the target too.
8406 if(i==0||dops[i].bt||dops[i].itype==CJUMP||dops[i].itype==SPAN) {
8407 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8408 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8413 // Deallocate unneeded registers
8414 for(hr=0;hr<HOST_REGS;hr++)
8417 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8420 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
8421 if (dops[i+1].is_load || dops[i+1].is_store)
8423 if (dops[i+1].is_store)
8425 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
8427 if((regs[i].regmap[hr]&63)!=dops[i].rs1 && (regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8428 (regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8429 (regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8430 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8431 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
8432 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8433 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8434 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
8436 regs[i].regmap[hr]=-1;
8437 regs[i].isconst&=~(1<<hr);
8438 if((branch_regs[i].regmap[hr]&63)!=dops[i].rs1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8439 (branch_regs[i].regmap[hr]&63)!=dops[i].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8440 (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8441 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8442 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8443 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8444 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8445 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8447 branch_regs[i].regmap[hr]=-1;
8448 branch_regs[i].regmap_entry[hr]=-1;
8449 if (!dops[i].is_ujump)
8452 regmap_pre[i+2][hr]=-1;
8453 regs[i+2].wasconst&=~(1<<hr);
8464 int map1 = -1, map2 = -1, temp=-1;
8465 if (dops[i].is_load || dops[i].is_store)
8467 if (dops[i].is_store)
8469 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8471 if((regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8472 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8473 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8474 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8475 regs[i].regmap[hr] != CCREG)
8477 if(i<slen-1&&!dops[i].is_ds) {
8478 assert(regs[i].regmap[hr]<64);
8479 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8480 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8482 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8483 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8485 regmap_pre[i+1][hr]=-1;
8486 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8487 regs[i+1].wasconst&=~(1<<hr);
8489 regs[i].regmap[hr]=-1;
8490 regs[i].isconst&=~(1<<hr);
8498 /* Pass 5 - Pre-allocate registers */
8500 // If a register is allocated during a loop, try to allocate it for the
8501 // entire loop, if possible. This avoids loading/storing registers
8502 // inside of the loop.
8504 signed char f_regmap[HOST_REGS];
8505 clear_all_regs(f_regmap);
8506 for(i=0;i<slen-1;i++)
8508 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8510 if(ba[i]>=start && ba[i]<(start+i*4))
8511 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8512 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8513 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8514 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8515 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8517 int t=(ba[i]-start)>>2;
8518 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8519 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8520 for(hr=0;hr<HOST_REGS;hr++)
8522 if(regs[i].regmap[hr]>=0) {
8523 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8524 // dealloc old register
8526 for(n=0;n<HOST_REGS;n++)
8528 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8530 // and alloc new one
8531 f_regmap[hr]=regs[i].regmap[hr];
8534 if(branch_regs[i].regmap[hr]>=0) {
8535 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8536 // dealloc old register
8538 for(n=0;n<HOST_REGS;n++)
8540 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8542 // and alloc new one
8543 f_regmap[hr]=branch_regs[i].regmap[hr];
8547 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8548 f_regmap[hr]=branch_regs[i].regmap[hr];
8550 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8551 f_regmap[hr]=branch_regs[i].regmap[hr];
8553 // Avoid dirty->clean transition
8554 #ifdef DESTRUCTIVE_WRITEBACK
8555 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;
8557 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8558 // case above, however it's always a good idea. We can't hoist the
8559 // load if the register was already allocated, so there's no point
8560 // wasting time analyzing most of these cases. It only "succeeds"
8561 // when the mapping was different and the load can be replaced with
8562 // a mov, which is of negligible benefit. So such cases are
8564 if(f_regmap[hr]>0) {
8565 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8569 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8570 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8572 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
8573 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8575 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8576 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8578 if(get_reg(regs[i].regmap,r&63)<0) break;
8579 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
8582 while(k>1&®s[k-1].regmap[hr]==-1) {
8583 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8584 //printf("no free regs for store %x\n",start+(k-1)*4);
8587 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8588 //printf("no-match due to different register\n");
8591 if (dops[k-2].is_jump) {
8592 //printf("no-match due to branch\n");
8595 // call/ret fast path assumes no registers allocated
8596 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8602 if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
8603 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8605 regs[k].regmap_entry[hr]=f_regmap[hr];
8606 regs[k].regmap[hr]=f_regmap[hr];
8607 regmap_pre[k+1][hr]=f_regmap[hr];
8608 regs[k].wasdirty&=~(1<<hr);
8609 regs[k].dirty&=~(1<<hr);
8610 regs[k].wasdirty|=(1<<hr)®s[k-1].dirty;
8611 regs[k].dirty|=(1<<hr)®s[k].wasdirty;
8612 regs[k].wasconst&=~(1<<hr);
8613 regs[k].isconst&=~(1<<hr);
8618 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8621 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8622 if(regs[i-1].regmap[hr]==f_regmap[hr]&®map_pre[i][hr]==f_regmap[hr]) {
8623 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8624 regs[i].regmap_entry[hr]=f_regmap[hr];
8625 regs[i].regmap[hr]=f_regmap[hr];
8626 regs[i].wasdirty&=~(1<<hr);
8627 regs[i].dirty&=~(1<<hr);
8628 regs[i].wasdirty|=(1<<hr)®s[i-1].dirty;
8629 regs[i].dirty|=(1<<hr)®s[i-1].dirty;
8630 regs[i].wasconst&=~(1<<hr);
8631 regs[i].isconst&=~(1<<hr);
8632 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8633 branch_regs[i].wasdirty&=~(1<<hr);
8634 branch_regs[i].wasdirty|=(1<<hr)®s[i].dirty;
8635 branch_regs[i].regmap[hr]=f_regmap[hr];
8636 branch_regs[i].dirty&=~(1<<hr);
8637 branch_regs[i].dirty|=(1<<hr)®s[i].dirty;
8638 branch_regs[i].wasconst&=~(1<<hr);
8639 branch_regs[i].isconst&=~(1<<hr);
8640 if (!dops[i].is_ujump) {
8641 regmap_pre[i+2][hr]=f_regmap[hr];
8642 regs[i+2].wasdirty&=~(1<<hr);
8643 regs[i+2].wasdirty|=(1<<hr)®s[i].dirty;
8648 // Alloc register clean at beginning of loop,
8649 // but may dirty it in pass 6
8650 regs[k].regmap_entry[hr]=f_regmap[hr];
8651 regs[k].regmap[hr]=f_regmap[hr];
8652 regs[k].dirty&=~(1<<hr);
8653 regs[k].wasconst&=~(1<<hr);
8654 regs[k].isconst&=~(1<<hr);
8655 if (dops[k].is_jump) {
8656 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8657 branch_regs[k].regmap[hr]=f_regmap[hr];
8658 branch_regs[k].dirty&=~(1<<hr);
8659 branch_regs[k].wasconst&=~(1<<hr);
8660 branch_regs[k].isconst&=~(1<<hr);
8661 if (!dops[k].is_ujump) {
8662 regmap_pre[k+2][hr]=f_regmap[hr];
8663 regs[k+2].wasdirty&=~(1<<hr);
8668 regmap_pre[k+1][hr]=f_regmap[hr];
8669 regs[k+1].wasdirty&=~(1<<hr);
8672 if(regs[j].regmap[hr]==f_regmap[hr])
8673 regs[j].regmap_entry[hr]=f_regmap[hr];
8677 if(regs[j].regmap[hr]>=0)
8679 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8680 //printf("no-match due to different register\n");
8683 if (dops[j].is_ujump)
8685 // Stop on unconditional branch
8688 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8691 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8694 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8697 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8698 //printf("no-match due to different register (branch)\n");
8702 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8703 //printf("No free regs for store %x\n",start+j*4);
8706 assert(f_regmap[hr]<64);
8713 // Non branch or undetermined branch target
8714 for(hr=0;hr<HOST_REGS;hr++)
8716 if(hr!=EXCLUDE_REG) {
8717 if(regs[i].regmap[hr]>=0) {
8718 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8719 // dealloc old register
8721 for(n=0;n<HOST_REGS;n++)
8723 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8725 // and alloc new one
8726 f_regmap[hr]=regs[i].regmap[hr];
8731 // Try to restore cycle count at branch targets
8733 for(j=i;j<slen-1;j++) {
8734 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8735 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8736 //printf("no free regs for store %x\n",start+j*4);
8740 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8742 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8744 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8745 regs[k].regmap[HOST_CCREG]=CCREG;
8746 regmap_pre[k+1][HOST_CCREG]=CCREG;
8747 regs[k+1].wasdirty|=1<<HOST_CCREG;
8748 regs[k].dirty|=1<<HOST_CCREG;
8749 regs[k].wasconst&=~(1<<HOST_CCREG);
8750 regs[k].isconst&=~(1<<HOST_CCREG);
8753 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8755 // Work backwards from the branch target
8756 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8758 //printf("Extend backwards\n");
8761 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8762 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8763 //printf("no free regs for store %x\n",start+(k-1)*4);
8768 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8769 //printf("Extend CC, %x ->\n",start+k*4);
8771 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8772 regs[k].regmap[HOST_CCREG]=CCREG;
8773 regmap_pre[k+1][HOST_CCREG]=CCREG;
8774 regs[k+1].wasdirty|=1<<HOST_CCREG;
8775 regs[k].dirty|=1<<HOST_CCREG;
8776 regs[k].wasconst&=~(1<<HOST_CCREG);
8777 regs[k].isconst&=~(1<<HOST_CCREG);
8782 //printf("Fail Extend CC, %x ->\n",start+k*4);
8786 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8787 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8788 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8790 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8795 // This allocates registers (if possible) one instruction prior
8796 // to use, which can avoid a load-use penalty on certain CPUs.
8797 for(i=0;i<slen-1;i++)
8799 if (!i || !dops[i-1].is_jump)
8803 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8804 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8807 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8809 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8811 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8812 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8813 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8814 regs[i].isconst&=~(1<<hr);
8815 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8816 constmap[i][hr]=constmap[i+1][hr];
8817 regs[i+1].wasdirty&=~(1<<hr);
8818 regs[i].dirty&=~(1<<hr);
8823 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8825 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8827 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8828 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8829 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8830 regs[i].isconst&=~(1<<hr);
8831 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8832 constmap[i][hr]=constmap[i+1][hr];
8833 regs[i+1].wasdirty&=~(1<<hr);
8834 regs[i].dirty&=~(1<<hr);
8838 // Preload target address for load instruction (non-constant)
8839 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8840 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8842 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8844 regs[i].regmap[hr]=dops[i+1].rs1;
8845 regmap_pre[i+1][hr]=dops[i+1].rs1;
8846 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8847 regs[i].isconst&=~(1<<hr);
8848 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8849 constmap[i][hr]=constmap[i+1][hr];
8850 regs[i+1].wasdirty&=~(1<<hr);
8851 regs[i].dirty&=~(1<<hr);
8855 // Load source into target register
8856 if(dops[i+1].lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8857 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8859 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8861 regs[i].regmap[hr]=dops[i+1].rs1;
8862 regmap_pre[i+1][hr]=dops[i+1].rs1;
8863 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8864 regs[i].isconst&=~(1<<hr);
8865 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8866 constmap[i][hr]=constmap[i+1][hr];
8867 regs[i+1].wasdirty&=~(1<<hr);
8868 regs[i].dirty&=~(1<<hr);
8872 // Address for store instruction (non-constant)
8873 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8874 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8875 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8876 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8877 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8878 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
8880 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8882 regs[i].regmap[hr]=dops[i+1].rs1;
8883 regmap_pre[i+1][hr]=dops[i+1].rs1;
8884 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8885 regs[i].isconst&=~(1<<hr);
8886 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8887 constmap[i][hr]=constmap[i+1][hr];
8888 regs[i+1].wasdirty&=~(1<<hr);
8889 regs[i].dirty&=~(1<<hr);
8893 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8894 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8896 hr=get_reg(regs[i+1].regmap,FTEMP);
8898 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8900 regs[i].regmap[hr]=dops[i+1].rs1;
8901 regmap_pre[i+1][hr]=dops[i+1].rs1;
8902 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8903 regs[i].isconst&=~(1<<hr);
8904 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8905 constmap[i][hr]=constmap[i+1][hr];
8906 regs[i+1].wasdirty&=~(1<<hr);
8907 regs[i].dirty&=~(1<<hr);
8909 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8911 // move it to another register
8912 regs[i+1].regmap[hr]=-1;
8913 regmap_pre[i+2][hr]=-1;
8914 regs[i+1].regmap[nr]=FTEMP;
8915 regmap_pre[i+2][nr]=FTEMP;
8916 regs[i].regmap[nr]=dops[i+1].rs1;
8917 regmap_pre[i+1][nr]=dops[i+1].rs1;
8918 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8919 regs[i].isconst&=~(1<<nr);
8920 regs[i+1].isconst&=~(1<<nr);
8921 regs[i].dirty&=~(1<<nr);
8922 regs[i+1].wasdirty&=~(1<<nr);
8923 regs[i+1].dirty&=~(1<<nr);
8924 regs[i+2].wasdirty&=~(1<<nr);
8928 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*/) {
8929 if(dops[i+1].itype==LOAD)
8930 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8931 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8932 hr=get_reg(regs[i+1].regmap,FTEMP);
8933 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
8934 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8935 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8937 if(hr>=0&®s[i].regmap[hr]<0) {
8938 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8939 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8940 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8941 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8942 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8943 regs[i].isconst&=~(1<<hr);
8944 regs[i+1].wasdirty&=~(1<<hr);
8945 regs[i].dirty&=~(1<<hr);
8954 /* Pass 6 - Optimize clean/dirty state */
8955 clean_registers(0,slen-1,1);
8957 /* Pass 7 - Identify 32-bit registers */
8958 for (i=slen-1;i>=0;i--)
8960 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8962 // Conditional branch
8963 if((source[i]>>16)!=0x1000&&i<slen-2) {
8964 // Mark this address as a branch target since it may be called
8965 // upon return from interrupt
8971 if(dops[slen-1].itype==SPAN) {
8972 dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
8976 /* Debug/disassembly */
8981 for(r=1;r<=CCREG;r++) {
8982 if((unneeded_reg[i]>>r)&1) {
8983 if(r==HIREG) printf(" HI");
8984 else if(r==LOREG) printf(" LO");
8985 else printf(" r%d",r);
8989 #if defined(__i386__) || defined(__x86_64__)
8990 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]);
8993 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]);
8995 #if defined(__i386__) || defined(__x86_64__)
8997 if(needed_reg[i]&1) printf("eax ");
8998 if((needed_reg[i]>>1)&1) printf("ecx ");
8999 if((needed_reg[i]>>2)&1) printf("edx ");
9000 if((needed_reg[i]>>3)&1) printf("ebx ");
9001 if((needed_reg[i]>>5)&1) printf("ebp ");
9002 if((needed_reg[i]>>6)&1) printf("esi ");
9003 if((needed_reg[i]>>7)&1) printf("edi ");
9005 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]);
9007 if(regs[i].wasdirty&1) printf("eax ");
9008 if((regs[i].wasdirty>>1)&1) printf("ecx ");
9009 if((regs[i].wasdirty>>2)&1) printf("edx ");
9010 if((regs[i].wasdirty>>3)&1) printf("ebx ");
9011 if((regs[i].wasdirty>>5)&1) printf("ebp ");
9012 if((regs[i].wasdirty>>6)&1) printf("esi ");
9013 if((regs[i].wasdirty>>7)&1) printf("edi ");
9016 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]);
9018 if(regs[i].wasdirty&1) printf("r0 ");
9019 if((regs[i].wasdirty>>1)&1) printf("r1 ");
9020 if((regs[i].wasdirty>>2)&1) printf("r2 ");
9021 if((regs[i].wasdirty>>3)&1) printf("r3 ");
9022 if((regs[i].wasdirty>>4)&1) printf("r4 ");
9023 if((regs[i].wasdirty>>5)&1) printf("r5 ");
9024 if((regs[i].wasdirty>>6)&1) printf("r6 ");
9025 if((regs[i].wasdirty>>7)&1) printf("r7 ");
9026 if((regs[i].wasdirty>>8)&1) printf("r8 ");
9027 if((regs[i].wasdirty>>9)&1) printf("r9 ");
9028 if((regs[i].wasdirty>>10)&1) printf("r10 ");
9029 if((regs[i].wasdirty>>12)&1) printf("r12 ");
9032 disassemble_inst(i);
9033 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
9034 #if defined(__i386__) || defined(__x86_64__)
9035 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]);
9036 if(regs[i].dirty&1) printf("eax ");
9037 if((regs[i].dirty>>1)&1) printf("ecx ");
9038 if((regs[i].dirty>>2)&1) printf("edx ");
9039 if((regs[i].dirty>>3)&1) printf("ebx ");
9040 if((regs[i].dirty>>5)&1) printf("ebp ");
9041 if((regs[i].dirty>>6)&1) printf("esi ");
9042 if((regs[i].dirty>>7)&1) printf("edi ");
9045 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]);
9046 if(regs[i].dirty&1) printf("r0 ");
9047 if((regs[i].dirty>>1)&1) printf("r1 ");
9048 if((regs[i].dirty>>2)&1) printf("r2 ");
9049 if((regs[i].dirty>>3)&1) printf("r3 ");
9050 if((regs[i].dirty>>4)&1) printf("r4 ");
9051 if((regs[i].dirty>>5)&1) printf("r5 ");
9052 if((regs[i].dirty>>6)&1) printf("r6 ");
9053 if((regs[i].dirty>>7)&1) printf("r7 ");
9054 if((regs[i].dirty>>8)&1) printf("r8 ");
9055 if((regs[i].dirty>>9)&1) printf("r9 ");
9056 if((regs[i].dirty>>10)&1) printf("r10 ");
9057 if((regs[i].dirty>>12)&1) printf("r12 ");
9060 if(regs[i].isconst) {
9061 printf("constants: ");
9062 #if defined(__i386__) || defined(__x86_64__)
9063 if(regs[i].isconst&1) printf("eax=%x ",(u_int)constmap[i][0]);
9064 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(u_int)constmap[i][1]);
9065 if((regs[i].isconst>>2)&1) printf("edx=%x ",(u_int)constmap[i][2]);
9066 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(u_int)constmap[i][3]);
9067 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(u_int)constmap[i][5]);
9068 if((regs[i].isconst>>6)&1) printf("esi=%x ",(u_int)constmap[i][6]);
9069 if((regs[i].isconst>>7)&1) printf("edi=%x ",(u_int)constmap[i][7]);
9071 #if defined(__arm__) || defined(__aarch64__)
9073 for (r = 0; r < ARRAY_SIZE(constmap[i]); r++)
9074 if ((regs[i].isconst >> r) & 1)
9075 printf(" r%d=%x", r, (u_int)constmap[i][r]);
9079 if(dops[i].is_jump) {
9080 #if defined(__i386__) || defined(__x86_64__)
9081 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]);
9082 if(branch_regs[i].dirty&1) printf("eax ");
9083 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
9084 if((branch_regs[i].dirty>>2)&1) printf("edx ");
9085 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
9086 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
9087 if((branch_regs[i].dirty>>6)&1) printf("esi ");
9088 if((branch_regs[i].dirty>>7)&1) printf("edi ");
9091 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]);
9092 if(branch_regs[i].dirty&1) printf("r0 ");
9093 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
9094 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
9095 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
9096 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
9097 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
9098 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
9099 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
9100 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
9101 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
9102 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
9103 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
9109 /* Pass 8 - Assembly */
9110 linkcount=0;stubcount=0;
9111 ds=0;is_delayslot=0;
9113 void *beginning=start_block();
9118 void *instr_addr0_override = NULL;
9120 if (start == 0x80030000) {
9121 // nasty hack for the fastbios thing
9122 // override block entry to this code
9123 instr_addr0_override = out;
9124 emit_movimm(start,0);
9125 // abuse io address var as a flag that we
9126 // have already returned here once
9127 emit_readword(&address,1);
9128 emit_writeword(0,&pcaddr);
9129 emit_writeword(0,&address);
9132 emit_jeq(out + 4*2);
9133 emit_far_jump(new_dyna_leave);
9135 emit_jne(new_dyna_leave);
9140 //if(ds) printf("ds: ");
9141 disassemble_inst(i);
9143 ds=0; // Skip delay slot
9144 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9145 instr_addr[i] = NULL;
9147 speculate_register_values(i);
9148 #ifndef DESTRUCTIVE_WRITEBACK
9149 if (i < 2 || !dops[i-2].is_ujump)
9151 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9153 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9154 dirty_pre=branch_regs[i].dirty;
9156 dirty_pre=regs[i].dirty;
9160 if (i < 2 || !dops[i-2].is_ujump)
9162 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9163 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9165 // branch target entry point
9166 instr_addr[i] = out;
9167 assem_debug("<->\n");
9168 drc_dbg_emit_do_cmp(i, ccadj[i]);
9171 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
9172 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9173 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9174 address_generation(i,®s[i],regs[i].regmap_entry);
9175 load_consts(regmap_pre[i],regs[i].regmap,i);
9178 // Load the delay slot registers if necessary
9179 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))
9180 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9181 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))
9182 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9183 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9184 load_regs(regs[i].regmap_entry,regs[i].regmap,ROREG,ROREG);
9185 if (dops[i+1].is_store)
9186 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9190 // Preload registers for following instruction
9191 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9192 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9193 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9194 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9195 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9196 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9198 // TODO: if(is_ooo(i)) address_generation(i+1);
9199 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9200 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
9201 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9202 load_regs(regs[i].regmap_entry,regs[i].regmap,ROREG,ROREG);
9203 if (dops[i].is_store)
9204 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9206 ds = assemble(i, ®s[i], ccadj[i]);
9208 if (dops[i].is_ujump)
9211 literal_pool_jumpover(256);
9216 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9217 // no ending needed for this block since INTCALL never returns
9219 // If the block did not end with an unconditional branch,
9220 // add a jump to the next instruction.
9222 if (!dops[i-2].is_ujump && dops[i-1].itype != SPAN) {
9223 assert(!dops[i-1].is_jump);
9225 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9226 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9227 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9228 emit_loadreg(CCREG,HOST_CCREG);
9229 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9233 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9234 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9236 add_to_linker(out,start+i*4,0);
9243 assert(!dops[i-1].is_jump);
9244 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9245 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9246 emit_loadreg(CCREG,HOST_CCREG);
9247 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9248 add_to_linker(out,start+i*4,0);
9252 // TODO: delay slot stubs?
9254 for(i=0;i<stubcount;i++)
9256 switch(stubs[i].type)
9264 do_readstub(i);break;
9269 do_writestub(i);break;
9273 do_invstub(i);break;
9275 do_cop1stub(i);break;
9277 do_unalignedwritestub(i);break;
9281 if (instr_addr0_override)
9282 instr_addr[0] = instr_addr0_override;
9284 /* Pass 9 - Linker */
9285 for(i=0;i<linkcount;i++)
9287 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9289 if (!link_addr[i].ext)
9292 void *addr = check_addr(link_addr[i].target);
9293 emit_extjump(link_addr[i].addr, link_addr[i].target);
9295 set_jump_target(link_addr[i].addr, addr);
9296 add_jump_out(link_addr[i].target,stub);
9299 set_jump_target(link_addr[i].addr, stub);
9304 int target=(link_addr[i].target-start)>>2;
9305 assert(target>=0&&target<slen);
9306 assert(instr_addr[target]);
9307 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9308 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9310 set_jump_target(link_addr[i].addr, instr_addr[target]);
9315 u_int source_len = slen*4;
9316 if (dops[slen-1].itype == INTCALL && source_len > 4)
9317 // no need to treat the last instruction as compiled
9318 // as interpreter fully handles it
9321 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9324 // External Branch Targets (jump_in)
9327 if(dops[i].bt||i==0)
9329 if(instr_addr[i]) // TODO - delay slots (=null)
9331 u_int vaddr=start+i*4;
9332 u_int page=get_page(vaddr);
9333 u_int vpage=get_vpage(vaddr);
9336 assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
9337 assem_debug("jump_in: %x\n",start+i*4);
9338 ll_add(jump_dirty+vpage,vaddr,out);
9339 void *entry_point = do_dirty_stub(i, source_len);
9340 ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
9341 // If there was an existing entry in the hash table,
9342 // replace it with the new address.
9343 // Don't add new entries. We'll insert the
9344 // ones that actually get used in check_addr().
9345 struct ht_entry *ht_bin = hash_table_get(vaddr);
9346 if (ht_bin->vaddr[0] == vaddr)
9347 ht_bin->tcaddr[0] = entry_point;
9348 if (ht_bin->vaddr[1] == vaddr)
9349 ht_bin->tcaddr[1] = entry_point;
9354 // Write out the literal pool if necessary
9356 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9358 if(((u_int)out)&7) emit_addnop(13);
9360 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9361 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9362 memcpy(copy, source, source_len);
9365 end_block(beginning);
9367 // If we're within 256K of the end of the buffer,
9368 // start over from the beginning. (Is 256K enough?)
9369 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9370 out = ndrc->translation_cache;
9372 // Trap writes to any of the pages we compiled
9373 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
9376 inv_code_start=inv_code_end=~0;
9378 // for PCSX we need to mark all mirrors too
9379 if(get_page(start)<(RAM_SIZE>>12))
9380 for(i=start>>12;i<=(start+slen*4)>>12;i++)
9381 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
9382 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
9383 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
9385 /* Pass 10 - Free memory by expiring oldest blocks */
9387 int end=(((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
9390 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
9391 uintptr_t base_offs = ((uintptr_t)(expirep >> 13) << shift); // Base offset of this block
9392 uintptr_t base_offs_s = base_offs >> shift;
9393 inv_debug("EXP: Phase %d\n",expirep);
9394 switch((expirep>>11)&3)
9397 // Clear jump_in and jump_dirty
9398 ll_remove_matching_addrs(jump_in+(expirep&2047),base_offs_s,shift);
9399 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base_offs_s,shift);
9400 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base_offs_s,shift);
9401 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base_offs_s,shift);
9405 ll_kill_pointers(jump_out[expirep&2047],base_offs_s,shift);
9406 ll_kill_pointers(jump_out[(expirep&2047)+2048],base_offs_s,shift);
9411 struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
9412 uintptr_t o1 = (u_char *)ht_bin->tcaddr[1] - ndrc->translation_cache;
9413 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9414 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9415 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
9416 ht_bin->vaddr[1] = -1;
9417 ht_bin->tcaddr[1] = NULL;
9419 o1 = (u_char *)ht_bin->tcaddr[0] - ndrc->translation_cache;
9420 o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9421 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9422 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
9423 ht_bin->vaddr[0] = ht_bin->vaddr[1];
9424 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
9425 ht_bin->vaddr[1] = -1;
9426 ht_bin->tcaddr[1] = NULL;
9432 if((expirep&2047)==0)
9434 ll_remove_matching_addrs(jump_out+(expirep&2047),base_offs_s,shift);
9435 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base_offs_s,shift);
9438 expirep=(expirep+1)&65535;
9446 // vim:shiftwidth=2:expandtab