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
28 #include <libkern/OSCacheControl.h>
31 #include <3ds_utils.h>
38 #include "new_dynarec_config.h"
39 #include "../psxhle.h"
40 #include "../psxinterpreter.h"
41 #include "../psxcounters.h"
43 #include "emu_if.h" // emulator interface
44 #include "linkage_offsets.h"
45 #include "compiler_features.h"
46 #include "arm_features.h"
49 #define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
52 #define min(a, b) ((b) < (a) ? (b) : (a))
55 #define max(a, b) ((b) > (a) ? (b) : (a))
60 //#define ASSEM_PRINT_ADDRS
61 //#define REGMAP_PRINT // with DISASM only
66 #define assem_debug printf
68 #define assem_debug(...)
70 #ifdef ASSEM_PRINT_ADDRS
71 #define log_addr(a) (a)
73 // for diff-able output
74 #define log_addr(a) ((u_long)(a) <= 1024u ? (void *)(a) : (void *)0xadd0l)
76 //#define inv_debug printf
77 #define inv_debug(...)
79 #define SysPrintf_lim(...) do { \
80 if (err_print_count++ < 64u) \
81 SysPrintf(__VA_ARGS__); \
85 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0 (CCREG)
86 extern int last_count; // last absolute target, often = next_interupt
88 extern int reg_cop2d[], reg_cop2c[];
90 extern void *hash_table_ptr;
91 extern uintptr_t ram_offset;
92 extern uintptr_t mini_ht[32][2];
95 #include "assem_x86.h"
98 #include "assem_x64.h"
101 #include "assem_arm.h"
104 #include "assem_arm64.h"
107 #define RAM_SIZE 0x200000
108 #define MAX_OUTPUT_BLOCK_SIZE 262144
109 #define EXPIRITY_OFFSET (MAX_OUTPUT_BLOCK_SIZE * 2)
110 #define PAGE_COUNT 1024
112 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
113 #define INVALIDATE_USE_COND_CALL
117 // apparently Vita has a 16MB limit, so either we cut tc in half,
118 // or use this hack (it's a hack because tc size was designed to be power-of-2)
119 #define TC_REDUCE_BYTES 4096
121 #define TC_REDUCE_BYTES 0
124 struct ndrc_globals ndrc_g;
128 struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
129 const void *f[2048 / sizeof(void *)];
134 u_char translation_cache[(1 << TARGET_SIZE_2) - TC_REDUCE_BYTES];
135 struct ndrc_tramp tramp;
138 static struct ndrc_mem *ndrc;
139 #ifndef BASE_ADDR_DYNAMIC
140 // reserve .bss space with upto 64k page size in mind
141 static char ndrc_bss[((sizeof(*ndrc) + 65535) & ~65535) + 65536];
143 #ifdef TC_WRITE_OFFSET
145 # include <sys/types.h>
146 # include <sys/stat.h>
149 static long ndrc_write_ofs;
150 #define NDRC_WRITE_OFFSET(x) (void *)((char *)(x) + ndrc_write_ofs)
152 #define NDRC_WRITE_OFFSET(x) (x)
175 // regmap_pre[i] - regs before [i] insn starts; dirty things here that
176 // don't match .regmap will be written back
177 // [i].regmap_entry - regs that must be set up if someone jumps here
178 // [i].regmap - regs [i] insn will read/(over)write
179 // branch_regs[i].* - same as above but for branches, takes delay slot into account
182 signed char regmap_entry[HOST_REGS];
183 signed char regmap[HOST_REGS];
186 u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
187 u_int isconst; // ... but isconst is false when r2 is known (hr)
188 u_int loadedconst; // host regs that have constants loaded
189 u_int noevict; // can't evict this hr (alloced by current op)
190 //u_int waswritten; // MIPS regs that were used as store base before
221 struct block_info *next;
224 u_int start; // vaddr of the block start
225 u_int len; // of the whole block source
230 u_char inv_near_misses;
248 static struct decoded_insn
251 u_char opcode; // bits 31-26
252 u_char opcode2; // (depends on opcode)
265 u_char is_delay_load:1; // is_load + MFC/CFC
266 u_char is_exception:1; // unconditional, also interp. fallback
267 u_char may_except:1; // might generate an exception
268 u_char ls_type:2; // load/store type (ls_width_type LS_*)
272 LS_8 = 0, LS_16, LS_32, LS_LR
275 static struct compile_info
280 signed char min_free_regs;
282 signed char reserved[2];
286 static char invalid_code[0x100000];
287 static struct ht_entry hash_table[65536];
288 static struct block_info *blocks[PAGE_COUNT];
289 static struct jump_info *jumps[PAGE_COUNT];
291 static u_int *source;
292 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
293 static uint64_t gte_rt[MAXBLOCK];
294 static uint64_t gte_unneeded[MAXBLOCK];
295 unsigned int ndrc_smrv_regs[32]; // speculated MIPS register values
296 static u_int smrv_strong; // mask or regs that are likely to have correct values
297 static u_int smrv_weak; // same, but somewhat less likely
298 static u_int smrv_strong_next; // same, but after current insn executes
299 static u_int smrv_weak_next;
300 static uint64_t unneeded_reg[MAXBLOCK];
301 static uint64_t branch_unneeded_reg[MAXBLOCK];
302 // see 'struct regstat' for a description
303 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
304 // contains 'real' consts at [i] insn, but may differ from what's actually
305 // loaded in host reg as 'final' value is always loaded, see get_final_value()
306 static uint32_t current_constmap[HOST_REGS];
307 static uint32_t constmap[MAXBLOCK][HOST_REGS];
308 static struct regstat regs[MAXBLOCK];
309 static struct regstat branch_regs[MAXBLOCK];
311 static void *instr_addr[MAXBLOCK];
312 static struct link_entry link_addr[MAXBLOCK];
313 static int linkcount;
314 static struct code_stub stubs[MAXBLOCK*3];
315 static int stubcount;
316 static u_int literals[1024][2];
317 static int literalcount;
318 static int is_delayslot;
319 static char shadow[1048576] __attribute__((aligned(16)));
321 static u_int expirep;
322 static u_int stop_after_jal;
323 static u_int ni_count;
324 static u_int err_print_count;
325 static u_int f1_hack;
326 static u_int vsync_hack;
328 static int stat_bc_direct;
329 static int stat_bc_pre;
330 static int stat_bc_restore;
331 static int stat_ht_lookups;
332 static int stat_jump_in_lookups;
333 static int stat_restore_tries;
334 static int stat_restore_compares;
335 static int stat_inv_addr_calls;
336 static int stat_inv_hits;
337 static int stat_blocks;
338 static int stat_links;
339 #define stat_inc(s) s++
340 #define stat_dec(s) s--
341 #define stat_clear(s) s = 0
345 #define stat_clear(s)
348 #define HACK_ENABLED(x) ((ndrc_g.hacks | ndrc_g.hacks_pergame) & (x))
350 /* registers that may be allocated */
352 #define LOREG 32 // lo
353 #define HIREG 33 // hi
354 //#define FSREG 34 // FPU status (FCSR)
355 //#define CSREG 35 // Coprocessor status
356 #define CCREG 36 // Cycle count
357 #define INVCP 37 // Pointer to invalid_code
358 //#define MMREG 38 // Pointer to memory_map
359 #define ROREG 39 // ram offset (if psxM != 0x80000000)
361 #define FTEMP 40 // Load/store temporary register (was fpu)
362 #define PTEMP 41 // Prefetch temporary register
363 //#define TLREG 42 // TLB mapping offset
364 #define RHASH 43 // Return address hash
365 #define RHTBL 44 // Return address hash table address
366 #define RTEMP 45 // JR/JALR address register
368 #define AGEN1 46 // Address generation temporary register (pass5b_preallocate2)
369 //#define AGEN2 47 // Address generation temporary register
371 /* instruction types */
372 #define NOP 0 // No operation
373 #define LOAD 1 // Load
374 #define STORE 2 // Store
375 #define LOADLR 3 // Unaligned load
376 #define STORELR 4 // Unaligned store
377 #define MOV 5 // Move (hi/lo only)
378 #define ALU 6 // Arithmetic/logic
379 #define MULTDIV 7 // Multiply/divide
380 #define SHIFT 8 // Shift by register
381 #define SHIFTIMM 9// Shift by immediate
382 #define IMM16 10 // 16-bit immediate
383 #define RJUMP 11 // Unconditional jump to register
384 #define UJUMP 12 // Unconditional jump
385 #define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
386 #define SJUMP 14 // Conditional branch (regimm format)
387 #define COP0 15 // Coprocessor 0
389 #define SYSCALL 22// SYSCALL,BREAK
390 #define OTHER 23 // Other/unknown - do nothing
391 #define HLECALL 26// PCSX fake opcodes for HLE
392 #define COP2 27 // Coprocessor 2 move
393 #define C2LS 28 // Coprocessor 2 load/store
394 #define C2OP 29 // Coprocessor 2 operation
395 #define INTCALL 30// Call interpreter to handle rare corner cases
401 #define DJT_1 (void *)1l // no function, just a label in assem_debug log
402 #define DJT_2 (void *)2l
407 void jump_syscall (u_int u0, u_int u1, u_int pc);
408 void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
409 void jump_break (u_int u0, u_int u1, u_int pc);
410 void jump_break_ds(u_int u0, u_int u1, u_int pc);
411 void jump_overflow (u_int u0, u_int u1, u_int pc);
412 void jump_overflow_ds(u_int u0, u_int u1, u_int pc);
413 void jump_addrerror (u_int cause, u_int addr, u_int pc);
414 void jump_addrerror_ds(u_int cause, u_int addr, u_int pc);
415 void jump_to_new_pc();
416 void new_dyna_leave();
418 void *ndrc_get_addr_ht(u_int vaddr, struct ht_entry *ht);
419 void ndrc_write_invalidate_one(u_int addr);
420 static void ndrc_write_invalidate_many(u_int addr, u_int end);
422 static int new_recompile_block(u_int addr);
423 static void invalidate_block(struct block_info *block);
424 static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_);
426 // Needed by assembler
427 static void wb_register(signed char r, const signed char regmap[], u_int dirty);
428 static void wb_dirtys(const signed char i_regmap[], u_int i_dirty);
429 static void wb_needed_dirtys(const signed char i_regmap[], u_int i_dirty, int addr);
430 static void load_all_regs(const signed char i_regmap[]);
431 static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
432 static void load_regs_entry(int t);
433 static void load_all_consts(const signed char regmap[], u_int dirty, int i);
434 static u_int get_host_reglist(const signed char *regmap);
436 static int get_final_value(int hr, int i, u_int *value);
437 static void add_stub(enum stub_type type, void *addr, void *retaddr,
438 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
439 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
440 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
441 static void add_to_linker(void *addr, u_int target, int ext);
442 static void *get_direct_memhandler(void *table, u_int addr,
443 enum stub_type type, uintptr_t *addr_host);
444 static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
445 static void pass_args(int a0, int a1);
446 static void emit_far_jump(const void *f);
447 static void emit_far_call(const void *f);
450 #include <psp2/kernel/sysmem.h>
452 // note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
453 extern int getVMBlock();
454 int _newlib_vm_size_user = sizeof(*ndrc);
457 static void mprotect_w_x(void *start, void *end, int is_x)
461 // *Open* enables write on all memory that was
462 // allocated by sceKernelAllocMemBlockForVM()?
464 sceKernelCloseVMDomain();
466 sceKernelOpenVMDomain();
467 #elif defined(HAVE_LIBNX)
469 // check to avoid the full flush in jitTransitionToExecutable()
470 if (g_jit.type != JitType_CodeMemory) {
472 rc = jitTransitionToExecutable(&g_jit);
474 rc = jitTransitionToWritable(&g_jit);
476 ;//SysPrintf("jitTransition %d %08x\n", is_x, rc);
478 #elif defined(TC_WRITE_OFFSET)
479 // separated rx and rw areas are always available
481 u_long mstart = (u_long)start & ~4095ul;
482 u_long mend = (u_long)end;
483 if (mprotect((void *)mstart, mend - mstart,
484 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
485 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
490 void new_dyna_clear_cache(void *start, void *end)
492 #if defined(__arm__) || defined(__aarch64__)
493 size_t len = (char *)end - (char *)start;
494 #if defined(__BLACKBERRY_QNX__)
495 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
496 #elif defined(__MACH__)
497 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
499 sceKernelSyncVMDomain(sceBlock, start, len);
501 // tuned for old3ds' 16k:16k cache (in it's mostly clean state...)
502 if ((char *)end - (char *)start <= 2*1024)
503 ctr_clear_cache_range(start, end);
506 #elif defined(HAVE_LIBNX)
507 if (g_jit.type == JitType_CodeMemory) {
508 armDCacheClean(start, len);
509 armICacheInvalidate((char *)start - ndrc_write_ofs, len);
510 // as of v4.2.1 libnx lacks isb
511 __asm__ volatile("isb" ::: "memory");
513 #elif defined(__aarch64__)
514 // __clear_cache() doesn't handle differing cacheline sizes on big.LITTLE and
515 // leaves it to the kernel to virtualize ctr_el0, which some old kernels don't do
516 clear_cache_arm64(start, end);
518 __clear_cache(start, end);
524 static void start_tcache_write(void *start, void *end)
526 mprotect_w_x(start, end, 0);
529 static void end_tcache_write(void *start, void *end)
532 if (!ndrc_g.thread.dirty_start || (size_t)ndrc_g.thread.dirty_start > (size_t)start)
533 ndrc_g.thread.dirty_start = start;
534 if ((size_t)ndrc_g.thread.dirty_end < (size_t)end)
535 ndrc_g.thread.dirty_end = end;
537 new_dyna_clear_cache(start, end);
539 mprotect_w_x(start, end, 1);
542 static void *start_block(void)
544 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
545 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
546 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
547 start_tcache_write(NDRC_WRITE_OFFSET(out), NDRC_WRITE_OFFSET(end));
551 static void end_block(void *start)
553 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(out));
556 #ifdef NDRC_CACHE_FLUSH_ALL
558 static int needs_clear_cache;
560 static void mark_clear_cache(void *target)
562 if (!needs_clear_cache) {
563 start_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
564 needs_clear_cache = 1;
568 static void do_clear_cache(void)
570 if (needs_clear_cache) {
571 end_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
572 needs_clear_cache = 0;
578 // also takes care of w^x mappings when patching code
579 static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
581 static void mark_clear_cache(void *target)
583 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
584 u_int mask = 1u << ((offset >> 12) & 31);
585 if (!(needs_clear_cache[offset >> 17] & mask)) {
586 char *start = (char *)NDRC_WRITE_OFFSET((uintptr_t)target & ~4095l);
587 start_tcache_write(start, start + 4095);
588 needs_clear_cache[offset >> 17] |= mask;
592 // Clearing the cache is rather slow on ARM Linux, so mark the areas
593 // that need to be cleared, and then only clear these areas once.
594 static void do_clear_cache(void)
597 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
599 u_int bitmap = needs_clear_cache[i];
602 for (j = 0; j < 32; j++)
605 if (!(bitmap & (1u << j)))
608 start = ndrc->translation_cache + i*131072 + j*4096;
610 for (j++; j < 32; j++) {
611 if (!(bitmap & (1u << j)))
615 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(end));
617 needs_clear_cache[i] = 0;
621 #endif // NDRC_CACHE_FLUSH_ALL
623 #define NO_CYCLE_PENALTY_THR 12
625 static int cycle_multiplier_active;
627 static int CLOCK_ADJUST(int x)
629 int m = cycle_multiplier_active;
630 int s = (x >> 31) | 1;
631 return (x * m + s * 50) / 100;
634 static int ds_writes_rjump_rs(int i)
636 return dops[i].rs1 != 0
637 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2
638 || dops[i].rs1 == dops[i].rt1); // overwrites itself - same effect
641 // psx addr mirror masking (for invalidation)
642 static u_int pmmask(u_int vaddr)
644 vaddr &= ~0xe0000000;
645 if (vaddr < 0x01000000)
646 vaddr &= ~0x00e00000; // RAM mirrors
650 static u_int get_page(u_int vaddr)
652 u_int page = pmmask(vaddr) >> 12;
653 if (page >= PAGE_COUNT / 2)
654 page = PAGE_COUNT / 2 + (page & (PAGE_COUNT / 2 - 1));
658 // get a page for looking for a block that has vaddr
659 // (needed because the block may start in previous page)
660 static u_int get_page_prev(u_int vaddr)
662 assert(MAXBLOCK <= (1 << 12));
663 u_int page = get_page(vaddr);
669 static struct ht_entry *hash_table_get_p(struct ht_entry *ht, u_int vaddr)
671 return &ht[((vaddr >> 16) ^ vaddr) & 0xFFFF];
674 static struct ht_entry *hash_table_get(u_int vaddr)
676 return hash_table_get_p(hash_table, vaddr);
679 #define HASH_TABLE_BAD 0xbac
681 static void hash_table_clear(void)
683 struct ht_entry *ht_bin;
685 for (i = 0; i < ARRAY_SIZE(hash_table); i++) {
686 for (j = 0; j < ARRAY_SIZE(hash_table[i].vaddr); j++) {
687 hash_table[i].vaddr[j] = ~0;
688 hash_table[i].tcaddr[j] = (void *)(uintptr_t)HASH_TABLE_BAD;
691 // don't allow ~0 to hit
692 ht_bin = hash_table_get(~0);
693 for (j = 0; j < ARRAY_SIZE(ht_bin->vaddr); j++)
694 ht_bin->vaddr[j] = 1;
697 static void hash_table_add(u_int vaddr, void *tcaddr)
699 struct ht_entry *ht_bin = hash_table_get(vaddr);
701 ht_bin->vaddr[1] = ht_bin->vaddr[0];
702 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
703 ht_bin->vaddr[0] = vaddr;
704 ht_bin->tcaddr[0] = tcaddr;
707 static void hash_table_remove(int vaddr)
709 //printf("remove hash: %x\n",vaddr);
710 struct ht_entry *ht_bin = hash_table_get(vaddr);
711 if (ht_bin->vaddr[1] == vaddr) {
712 ht_bin->vaddr[1] = ~0;
713 ht_bin->tcaddr[1] = (void *)(uintptr_t)HASH_TABLE_BAD;
715 if (ht_bin->vaddr[0] == vaddr) {
716 ht_bin->vaddr[0] = ht_bin->vaddr[1];
717 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
718 ht_bin->vaddr[1] = ~0;
719 ht_bin->tcaddr[1] = (void *)(uintptr_t)HASH_TABLE_BAD;
723 static void mini_ht_clear(void)
727 for (i = 0; i < ARRAY_SIZE(mini_ht) - 1; i++) {
729 mini_ht[i][1] = HASH_TABLE_BAD;
732 mini_ht[i][1] = HASH_TABLE_BAD;
736 static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
738 u_int vaddr_m = vaddr & 0x1fffffff;
740 for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
741 // ram mirrors, but should not hurt bios
742 for (j = 0; j < 0x800000; j += 0x200000) {
743 invalid_code[(i|j) >> 12] =
744 invalid_code[(i|j|0x80000000u) >> 12] =
745 invalid_code[(i|j|0xa0000000u) >> 12] = invalid;
748 if (!invalid && vaddr + len > inv_code_start && vaddr <= inv_code_end)
749 inv_code_start = inv_code_end = ~0;
752 static int doesnt_expire_soon(u_char *tcaddr)
754 u_int diff = (u_int)(tcaddr - out) & ((1u << TARGET_SIZE_2) - 1u);
755 return diff > EXPIRITY_OFFSET + MAX_OUTPUT_BLOCK_SIZE;
758 static attr_unused void check_for_block_changes(u_int start, u_int end)
760 u_int start_page = get_page_prev(start);
761 u_int end_page = get_page(end - 1);
764 for (page = start_page; page <= end_page; page++) {
765 struct block_info *block;
766 for (block = blocks[page]; block != NULL; block = block->next) {
769 if (memcmp(block->source, block->copy, block->len)) {
770 printf("bad block %08x-%08x %016llx %016llx @%08x\n",
771 block->start, block->start + block->len,
772 *(long long *)block->source, *(long long *)block->copy, psxRegs.pc);
780 static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
782 void *found_clean = NULL;
785 stat_inc(stat_restore_tries);
786 for (page = start_page; page <= end_page; page++) {
787 struct block_info *block;
788 for (block = blocks[page]; block != NULL; block = block->next) {
789 if (vaddr < block->start)
791 if (!block->is_dirty || vaddr >= block->start + block->len)
793 for (i = 0; i < block->jump_in_cnt; i++)
794 if (block->jump_in[i].vaddr == vaddr)
796 if (i == block->jump_in_cnt)
798 assert(block->source && block->copy);
799 stat_inc(stat_restore_compares);
800 if (memcmp(block->source, block->copy, block->len))
803 block->is_dirty = block->inv_near_misses = 0;
804 found_clean = block->jump_in[i].addr;
805 hash_table_add(vaddr, found_clean);
806 mark_invalid_code(block->start, block->len, 0);
807 stat_inc(stat_bc_restore);
808 inv_debug("INV: restored %08x %p (%d)\n", vaddr, found_clean, block->jump_in_cnt);
815 // this doesn't normally happen
816 static noinline u_int generate_exception(u_int pc)
818 //if (execBreakCheck(&psxRegs, pc))
819 // return psxRegs.pc;
821 // generate an address or bus error
822 psxRegs.CP0.n.Cause &= 0x300;
823 psxRegs.CP0.n.EPC = pc;
825 psxRegs.CP0.n.Cause |= R3000E_AdEL << 2;
826 psxRegs.CP0.n.BadVAddr = pc;
831 psxRegs.CP0.n.Cause |= R3000E_IBE << 2;
832 return (psxRegs.pc = 0x80000080);
835 // Get address from virtual address
836 // This is called from the recompiled JR/JALR instructions
837 static void noinline *get_addr(struct ht_entry *ht, const u_int vaddr,
838 enum ndrc_compile_mode compile_mode)
840 u_int start_page = get_page_prev(vaddr);
841 u_int i, page, end_page = get_page(vaddr);
842 void *found_clean = NULL;
844 stat_inc(stat_jump_in_lookups);
845 for (page = start_page; page <= end_page; page++) {
846 const struct block_info *block;
847 for (block = blocks[page]; block != NULL; block = block->next) {
848 if (vaddr < block->start)
850 if (block->is_dirty || vaddr >= block->start + block->len)
852 for (i = 0; i < block->jump_in_cnt; i++)
853 if (block->jump_in[i].vaddr == vaddr)
855 if (i == block->jump_in_cnt)
857 found_clean = block->jump_in[i].addr;
858 hash_table_add(vaddr, found_clean);
862 found_clean = try_restore_block(vaddr, start_page, end_page);
866 if (compile_mode == ndrc_cm_no_compile)
869 if (ndrc_g.thread.handle && compile_mode == ndrc_cm_compile_live) {
871 return new_dyna_leave;
873 if (!ndrc_g.thread.handle)
875 memcpy(ndrc_smrv_regs, psxRegs.GPR.r, sizeof(ndrc_smrv_regs));
877 int r = new_recompile_block(vaddr);
879 return ndrc_get_addr_ht(vaddr, ht);
881 if (compile_mode == ndrc_cm_compile_live)
882 return ndrc_get_addr_ht(generate_exception(vaddr), ht);
887 // Look up address in hash table first
888 void *ndrc_get_addr_ht_param(struct ht_entry *ht, unsigned int vaddr,
889 enum ndrc_compile_mode compile_mode)
891 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
892 const struct ht_entry *ht_bin = hash_table_get_p(ht, vaddr);
893 u_int vaddr_a = vaddr & ~3;
894 stat_inc(stat_ht_lookups);
895 if (ht_bin->vaddr[0] == vaddr_a) return ht_bin->tcaddr[0];
896 if (ht_bin->vaddr[1] == vaddr_a) return ht_bin->tcaddr[1];
897 return get_addr(ht, vaddr, compile_mode);
900 // "usual" addr lookup for indirect branches, etc
901 // to be used by currently running code only
902 void *ndrc_get_addr_ht(u_int vaddr, struct ht_entry *ht)
904 return ndrc_get_addr_ht_param(ht, vaddr, ndrc_cm_compile_live);
907 static void clear_all_regs(signed char regmap[])
909 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
912 // get_reg: get allocated host reg from mips reg
913 // returns -1 if no such mips reg was allocated
914 #if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
916 extern signed char get_reg(const signed char regmap[], signed char r);
920 static signed char get_reg(const signed char regmap[], signed char r)
923 for (hr = 0; hr < HOST_REGS; hr++) {
924 if (hr == EXCLUDE_REG)
934 // get reg suitable for writing
935 static signed char get_reg_w(const signed char regmap[], signed char r)
937 return r == 0 ? -1 : get_reg(regmap, r);
940 // get reg as mask bit (1 << hr)
941 static u_int get_regm(const signed char regmap[], signed char r)
943 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
946 static signed char get_reg_temp(const signed char regmap[])
949 for (hr = 0; hr < HOST_REGS; hr++) {
950 if (hr == EXCLUDE_REG)
952 if (regmap[hr] == (signed char)-1)
958 // Find a register that is available for two consecutive cycles
959 static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
962 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map1[hr]==r&®map2[hr]==r) return hr;
966 // reverse reg map: mips -> host
967 #define RRMAP_SIZE 64
968 static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
969 u_int *regs_can_change)
971 u_int r, hr, hr_can_change = 0;
972 memset(rrmap, -1, RRMAP_SIZE);
973 for (hr = 0; hr < HOST_REGS; )
976 rrmap[r & (RRMAP_SIZE - 1)] = hr;
977 // only add mips $1-$31+$lo, others shifted out
978 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
980 if (hr == EXCLUDE_REG)
983 hr_can_change |= 1u << (rrmap[33] & 31);
984 hr_can_change |= 1u << (rrmap[CCREG] & 31);
985 hr_can_change &= ~(1u << 31);
986 *regs_can_change = hr_can_change;
989 // same as get_reg, but takes rrmap
990 static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
992 assert(0 <= r && r < RRMAP_SIZE);
996 static int count_free_regs(const signed char regmap[])
1000 for(hr=0;hr<HOST_REGS;hr++)
1002 if(hr!=EXCLUDE_REG) {
1003 if(regmap[hr]<0) count++;
1009 static void dirty_reg(struct regstat *cur, signed char reg)
1013 hr = get_reg(cur->regmap, reg);
1015 cur->dirty |= 1<<hr;
1018 static void set_const(struct regstat *cur, signed char reg, uint32_t value)
1022 hr = get_reg(cur->regmap, reg);
1024 cur->isconst |= 1<<hr;
1025 current_constmap[hr] = value;
1029 static void clear_const(struct regstat *cur, signed char reg)
1033 hr = get_reg(cur->regmap, reg);
1035 cur->isconst &= ~(1<<hr);
1038 static int is_const(const struct regstat *cur, signed char reg)
1041 if (reg < 0) return 0;
1043 hr = get_reg(cur->regmap, reg);
1045 return (cur->isconst>>hr)&1;
1049 static uint32_t get_const(const struct regstat *cur, signed char reg)
1053 hr = get_reg(cur->regmap, reg);
1055 return current_constmap[hr];
1057 SysPrintf("Unknown constant in r%d\n", reg);
1061 // Least soon needed registers
1062 // Look at the next ten instructions and see which registers
1063 // will be used. Try not to reallocate these.
1064 static void lsn(u_char hsn[], int i)
1074 if (dops[i+j].is_ujump)
1076 // Don't go past an unconditonal jump
1083 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
1084 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
1085 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
1086 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
1087 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
1088 // Stores can allocate zero
1089 hsn[dops[i+j].rs1]=j;
1090 hsn[dops[i+j].rs2]=j;
1092 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
1094 // On some architectures stores need invc_ptr
1095 #if defined(HOST_IMM8)
1096 if (dops[i+j].is_store)
1099 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1107 if(cinfo[i+b].ba>=start && cinfo[i+b].ba<(start+slen*4))
1109 // Follow first branch
1110 int t=(cinfo[i+b].ba-start)>>2;
1111 j=7-b;if(t+j>=slen) j=slen-t-1;
1114 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1115 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1116 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1117 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1120 // TODO: preferred register based on backward branch
1122 // Delay slot should preferably not overwrite branch conditions or cycle count
1123 if (i > 0 && dops[i-1].is_jump) {
1124 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1125 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1127 // ...or hash tables
1131 // Coprocessor load/store needs FTEMP, even if not declared
1132 if(dops[i].itype==C2LS) {
1135 // Load/store L/R also uses FTEMP as a temporary register
1136 if (dops[i].itype == LOADLR || dops[i].itype == STORELR) {
1139 // Don't remove the miniht registers
1140 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1147 // We only want to allocate registers if we're going to use them again soon
1148 static int needed_again(int r, int i)
1154 if (i > 0 && dops[i-1].is_ujump)
1156 if(cinfo[i-1].ba<start || cinfo[i-1].ba>start+slen*4-4)
1157 return 0; // Don't need any registers if exiting the block
1165 if (dops[i+j].is_ujump)
1167 // Don't go past an unconditonal jump
1171 if (dops[i+j].is_exception)
1178 if(dops[i+j].rs1==r) rn=j;
1179 if(dops[i+j].rs2==r) rn=j;
1180 if((unneeded_reg[i+j]>>r)&1) rn=10;
1181 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1191 // Try to match register allocations at the end of a loop with those
1193 static int loop_reg(int i, int r, int hr)
1202 if (dops[i+j].is_ujump)
1204 // Don't go past an unconditonal jump
1211 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1217 if((unneeded_reg[i+k]>>r)&1) return hr;
1218 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1220 if(cinfo[i+k].ba>=start && cinfo[i+k].ba<(start+i*4))
1222 int t=(cinfo[i+k].ba-start)>>2;
1223 int reg=get_reg(regs[t].regmap_entry,r);
1224 if(reg>=0) return reg;
1225 //reg=get_reg(regs[t+1].regmap_entry,r);
1226 //if(reg>=0) return reg;
1234 // Allocate every register, preserving source/target regs
1235 static void alloc_all(struct regstat *cur,int i)
1239 for(hr=0;hr<HOST_REGS;hr++) {
1240 if(hr!=EXCLUDE_REG) {
1241 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1242 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1245 cur->dirty&=~(1<<hr);
1248 if(cur->regmap[hr]==0)
1251 cur->dirty&=~(1<<hr);
1258 static int host_tempreg_in_use;
1260 static void host_tempreg_acquire(void)
1262 assert(!host_tempreg_in_use);
1263 host_tempreg_in_use = 1;
1266 static void host_tempreg_release(void)
1268 host_tempreg_in_use = 0;
1271 static void host_tempreg_acquire(void) {}
1272 static void host_tempreg_release(void) {}
1276 extern void gen_interupt();
1277 extern void do_insn_cmp();
1278 #define FUNCNAME(f) { f, " " #f }
1279 static const struct {
1282 } function_names[] = {
1283 FUNCNAME(cc_interrupt),
1284 FUNCNAME(gen_interupt),
1285 FUNCNAME(ndrc_get_addr_ht),
1286 FUNCNAME(ndrc_get_addr_ht_param),
1287 FUNCNAME(jump_handler_read8),
1288 FUNCNAME(jump_handler_read16),
1289 FUNCNAME(jump_handler_read32),
1290 FUNCNAME(jump_handler_write8),
1291 FUNCNAME(jump_handler_write16),
1292 FUNCNAME(jump_handler_write32),
1293 FUNCNAME(ndrc_write_invalidate_one),
1294 FUNCNAME(ndrc_write_invalidate_many),
1295 FUNCNAME(jump_to_new_pc),
1296 FUNCNAME(jump_break),
1297 FUNCNAME(jump_break_ds),
1298 FUNCNAME(jump_syscall),
1299 FUNCNAME(jump_syscall_ds),
1300 FUNCNAME(jump_overflow),
1301 FUNCNAME(jump_overflow_ds),
1302 FUNCNAME(jump_addrerror),
1303 FUNCNAME(jump_addrerror_ds),
1304 FUNCNAME(new_dyna_leave),
1305 FUNCNAME(pcsx_mtc0),
1306 FUNCNAME(pcsx_mtc0_ds),
1309 FUNCNAME(do_memhandler_pre),
1310 FUNCNAME(do_memhandler_post),
1314 FUNCNAME(do_insn_cmp_arm64),
1316 FUNCNAME(do_insn_cmp),
1321 static const char *func_name(const void *a)
1324 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1325 if (function_names[i].addr == a)
1326 return function_names[i].name;
1330 static const char *fpofs_name(u_int ofs)
1332 u_int *p = (u_int *)&dynarec_local + ofs/sizeof(u_int);
1333 static char buf[64];
1335 #define ofscase(x) case LO_##x: return " ; " #x
1336 ofscase(next_interupt);
1337 ofscase(cycle_count);
1338 ofscase(last_count);
1349 ofscase(ram_offset);
1350 ofscase(hash_table_ptr);
1354 if (psxRegs.GPR.r <= p && p < &psxRegs.GPR.r[32])
1355 snprintf(buf, sizeof(buf), " ; r%d", (int)(p - psxRegs.GPR.r));
1356 else if (psxRegs.CP0.r <= p && p < &psxRegs.CP0.r[32])
1357 snprintf(buf, sizeof(buf), " ; cp0 $%d", (int)(p - psxRegs.CP0.r));
1358 else if (psxRegs.CP2D.r <= p && p < &psxRegs.CP2D.r[32])
1359 snprintf(buf, sizeof(buf), " ; cp2d $%d", (int)(p - psxRegs.CP2D.r));
1360 else if (psxRegs.CP2C.r <= p && p < &psxRegs.CP2C.r[32])
1361 snprintf(buf, sizeof(buf), " ; cp2c $%d", (int)(p - psxRegs.CP2C.r));
1365 #define func_name(x) ""
1366 #define fpofs_name(x) ""
1370 #include "assem_x86.c"
1373 #include "assem_x64.c"
1376 #include "assem_arm.c"
1379 #include "assem_arm64.c"
1382 static void *get_trampoline(const void *f)
1384 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1387 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1388 if (tramp->f[i] == f || tramp->f[i] == NULL)
1391 if (i == ARRAY_SIZE(tramp->f)) {
1392 SysPrintf("trampoline table is full, last func %p\n", f);
1395 if (tramp->f[i] == NULL) {
1396 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1398 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1400 // invalidate the RX mirror (unsure if necessary, but just in case...)
1401 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1404 return &ndrc->tramp.ops[i];
1407 static void emit_far_jump(const void *f)
1409 if (can_jump_or_call(f)) {
1414 f = get_trampoline(f);
1418 static void emit_far_call(const void *f)
1420 if (can_jump_or_call(f)) {
1425 f = get_trampoline(f);
1429 // Check if an address is already compiled
1430 // but don't return addresses which are about to expire from the cache
1431 static void *check_addr(u_int vaddr)
1433 struct ht_entry *ht_bin = hash_table_get(vaddr);
1435 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1436 if (ht_bin->vaddr[i] == vaddr)
1437 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1438 return ht_bin->tcaddr[i];
1441 // refactor to get_addr_nocompile?
1442 u_int start_page = get_page_prev(vaddr);
1443 u_int page, end_page = get_page(vaddr);
1445 stat_inc(stat_jump_in_lookups);
1446 for (page = start_page; page <= end_page; page++) {
1447 const struct block_info *block;
1448 for (block = blocks[page]; block != NULL; block = block->next) {
1449 if (vaddr < block->start)
1451 if (block->is_dirty || vaddr >= block->start + block->len)
1453 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1455 for (i = 0; i < block->jump_in_cnt; i++)
1456 if (block->jump_in[i].vaddr == vaddr)
1458 if (i == block->jump_in_cnt)
1461 // Update existing entry with current address
1462 void *addr = block->jump_in[i].addr;
1463 if (ht_bin->vaddr[0] == vaddr) {
1464 ht_bin->tcaddr[0] = addr;
1467 if (ht_bin->vaddr[1] == vaddr) {
1468 ht_bin->tcaddr[1] = addr;
1471 // Insert into hash table with low priority.
1472 // Don't evict existing entries, as they are probably
1473 // addresses that are being accessed frequently.
1474 if (ht_bin->vaddr[0] == -1) {
1475 ht_bin->vaddr[0] = vaddr;
1476 ht_bin->tcaddr[0] = addr;
1478 else if (ht_bin->vaddr[1] == -1) {
1479 ht_bin->vaddr[1] = vaddr;
1480 ht_bin->tcaddr[1] = addr;
1488 static void blocks_clear(struct block_info **head)
1490 struct block_info *cur, *next;
1492 if ((cur = *head)) {
1502 static int blocks_remove_matching_addrs(struct block_info **head,
1503 u_int base_offs, int shift)
1505 struct block_info *next;
1508 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1509 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1510 invalidate_block(*head);
1511 next = (*head)->next;
1514 stat_dec(stat_blocks);
1519 head = &((*head)->next);
1525 // This is called when we write to a compiled block (see do_invstub)
1526 static void unlink_jumps_vaddr_range(u_int start, u_int end)
1528 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1531 for (page = start_page; page <= end_page; page++) {
1532 struct jump_info *ji = jumps[page];
1535 for (i = 0; i < ji->count; ) {
1536 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1541 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1542 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1543 void *host_addr = find_extjump_insn(ji->e[i].stub);
1544 mark_clear_cache(host_addr);
1545 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1547 stat_dec(stat_links);
1549 if (i < ji->count) {
1550 ji->e[i] = ji->e[ji->count];
1558 static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1563 for (i = 0; i < ji->count; ) {
1564 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1565 if (((tc_offs ^ base_offs) >> shift) != 0) {
1570 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1571 stat_dec(stat_links);
1573 if (i < ji->count) {
1574 ji->e[i] = ji->e[ji->count];
1581 static void invalidate_block(struct block_info *block)
1585 block->is_dirty = 1;
1586 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1587 for (i = 0; i < block->jump_in_cnt; i++)
1588 hash_table_remove(block->jump_in[i].vaddr);
1591 static int invalidate_range(u_int start, u_int end,
1592 u32 *inv_start_ret, u32 *inv_end_ret)
1594 struct block_info *last_block = NULL;
1595 u_int start_page = get_page_prev(start);
1596 u_int end_page = get_page(end - 1);
1597 u_int start_m = pmmask(start);
1598 u_int end_m = pmmask(end - 1);
1599 u_int inv_start, inv_end;
1600 u_int blk_start_m, blk_end_m;
1604 // additional area without code (to supplement invalid_code[]), [start, end)
1605 // avoids excessive ndrc_write_invalidate*() calls
1606 inv_start = start_m & ~0xfff;
1607 inv_end = end_m | 0xfff;
1609 for (page = start_page; page <= end_page; page++) {
1610 struct block_info *block;
1611 for (block = blocks[page]; block != NULL; block = block->next) {
1612 if (block->is_dirty)
1615 blk_end_m = pmmask(block->start + block->len);
1616 if (blk_end_m <= start_m) {
1617 inv_start = max(inv_start, blk_end_m);
1620 blk_start_m = pmmask(block->start);
1621 if (end_m <= blk_start_m) {
1622 inv_end = min(inv_end, blk_start_m - 1);
1625 if (!block->source) // "hack" block - leave it alone
1629 invalidate_block(block);
1630 stat_inc(stat_inv_hits);
1634 if (!hit && last_block && last_block->source) {
1635 // could be some leftover unused block, uselessly trapping writes
1636 last_block->inv_near_misses++;
1637 if (last_block->inv_near_misses > 128) {
1638 invalidate_block(last_block);
1639 stat_inc(stat_inv_hits);
1648 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1649 // the whole page is empty now
1650 mark_invalid_code(start, 1, 1);
1652 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1653 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1657 void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1659 invalidate_range(start, end, NULL, NULL);
1662 // check if the range may need invalidation (must be thread-safe)
1663 int new_dynarec_quick_check_range(unsigned int start, unsigned int end)
1665 u_int start_page = get_page_prev(start);
1666 u_int end_page = get_page(end - 1);
1669 if (inv_code_start <= start && end <= inv_code_end)
1671 for (page = start_page; page <= end_page; page++) {
1673 //SysPrintf("quick hit %x-%x\n", start, end);
1680 static void ndrc_write_invalidate_many(u_int start, u_int end)
1682 // this check is done by the caller
1683 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1684 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1686 int invc = invalid_code[start >> 12];
1687 u_int len = end - start;
1689 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1691 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1692 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1693 check_for_block_changes(start, end);
1695 stat_inc(stat_inv_addr_calls);
1699 void ndrc_write_invalidate_one(u_int addr)
1701 ndrc_write_invalidate_many(addr, addr + 4);
1704 // This is called when loading a save state.
1705 // Anything could have changed, so invalidate everything.
1706 void new_dynarec_invalidate_all_pages(void)
1708 struct block_info *block;
1710 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1711 for (block = blocks[page]; block != NULL; block = block->next) {
1712 if (block->is_dirty)
1714 if (!block->source) // hack block?
1716 invalidate_block(block);
1724 // Add an entry to jump_out after making a link
1725 // stub should point to stub code by emit_extjump()
1726 static void ndrc_add_jump_out(u_int vaddr, void *stub)
1728 inv_debug("ndrc_add_jump_out: %p -> %x\n", stub, vaddr);
1729 u_int page = get_page(vaddr);
1730 struct jump_info *ji;
1732 stat_inc(stat_links);
1733 check_extjump2(stub);
1736 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1740 else if (ji->count >= ji->alloc) {
1742 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1745 ji->e[ji->count].target_vaddr = vaddr;
1746 ji->e[ji->count].stub = stub;
1750 void ndrc_patch_link(u_int vaddr, void *insn, void *stub, void *target)
1752 void *insn_end = (char *)insn + 4;
1754 //start_tcache_write(insn, insn_end);
1755 mprotect_w_x(insn, insn_end, 0);
1757 assert(target != stub);
1758 set_jump_target_far1(insn, target);
1759 ndrc_add_jump_out(vaddr, stub);
1761 #if defined(__aarch64__) || defined(NO_WRITE_EXEC)
1762 // arm64: no syscall concerns, dyna_linker lacks stale detection
1763 // w^x: have to do costly permission switching anyway
1764 new_dyna_clear_cache(NDRC_WRITE_OFFSET(insn), NDRC_WRITE_OFFSET(insn_end));
1766 //end_tcache_write(insn, insn_end);
1767 mprotect_w_x(insn, insn_end, 1);
1770 /* Register allocation */
1772 static void alloc_set(struct regstat *cur, int reg, int hr)
1774 cur->regmap[hr] = reg;
1775 cur->dirty &= ~(1u << hr);
1776 cur->isconst &= ~(1u << hr);
1777 cur->noevict |= 1u << hr;
1780 static void evict_alloc_reg(struct regstat *cur, int i, int reg, int preferred_hr)
1782 u_char hsn[MAXREG+1];
1784 memset(hsn, 10, sizeof(hsn));
1786 //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]);
1788 // Don't evict the cycle count at entry points, otherwise the entry
1789 // stub will have to write it.
1790 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1791 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1794 // Alloc preferred register if available
1795 if (!((cur->noevict >> preferred_hr) & 1)
1796 && hsn[cur->regmap[preferred_hr]] == j)
1798 alloc_set(cur, reg, preferred_hr);
1801 for(r=1;r<=MAXREG;r++)
1803 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1804 for(hr=0;hr<HOST_REGS;hr++) {
1805 if (hr == EXCLUDE_REG || ((cur->noevict >> hr) & 1))
1807 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1808 if(cur->regmap[hr]==r) {
1809 alloc_set(cur, reg, hr);
1820 for(r=1;r<=MAXREG;r++)
1823 for(hr=0;hr<HOST_REGS;hr++) {
1824 if (hr == EXCLUDE_REG || ((cur->noevict >> hr) & 1))
1826 if(cur->regmap[hr]==r) {
1827 alloc_set(cur, reg, hr);
1834 SysPrintf("This shouldn't happen (evict_alloc_reg)\n");
1838 // Note: registers are allocated clean (unmodified state)
1839 // if you intend to modify the register, you must call dirty_reg().
1840 static void alloc_reg(struct regstat *cur,int i,signed char reg)
1843 int preferred_reg = PREFERRED_REG_FIRST
1844 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1845 if (reg == CCREG) preferred_reg = HOST_CCREG;
1846 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1847 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1850 // Don't allocate unused registers
1851 if((cur->u>>reg)&1) return;
1853 // see if it's already allocated
1854 if ((hr = get_reg(cur->regmap, reg)) >= 0) {
1855 cur->noevict |= 1u << hr;
1859 // Keep the same mapping if the register was already allocated in a loop
1860 preferred_reg = loop_reg(i,reg,preferred_reg);
1862 // Try to allocate the preferred register
1863 if (cur->regmap[preferred_reg] == -1) {
1864 alloc_set(cur, reg, preferred_reg);
1867 r=cur->regmap[preferred_reg];
1870 alloc_set(cur, reg, preferred_reg);
1874 // Clear any unneeded registers
1875 // We try to keep the mapping consistent, if possible, because it
1876 // makes branches easier (especially loops). So we try to allocate
1877 // first (see above) before removing old mappings. If this is not
1878 // possible then go ahead and clear out the registers that are no
1880 for(hr=0;hr<HOST_REGS;hr++)
1885 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1889 // Try to allocate any available register, but prefer
1890 // registers that have not been used recently.
1892 for (hr = PREFERRED_REG_FIRST; ; ) {
1893 if (cur->regmap[hr] < 0) {
1894 int oldreg = regs[i-1].regmap[hr];
1895 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1896 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1898 alloc_set(cur, reg, hr);
1903 if (hr == EXCLUDE_REG)
1905 if (hr == HOST_REGS)
1907 if (hr == PREFERRED_REG_FIRST)
1912 // Try to allocate any available register
1913 for (hr = PREFERRED_REG_FIRST; ; ) {
1914 if (cur->regmap[hr] < 0) {
1915 alloc_set(cur, reg, hr);
1919 if (hr == EXCLUDE_REG)
1921 if (hr == HOST_REGS)
1923 if (hr == PREFERRED_REG_FIRST)
1927 // Ok, now we have to evict someone
1928 // Pick a register we hopefully won't need soon
1929 evict_alloc_reg(cur, i, reg, preferred_reg);
1932 // Allocate a temporary register. This is done without regard to
1933 // dirty status or whether the register we request is on the unneeded list
1934 // Note: This will only allocate one register, even if called multiple times
1935 static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1939 // see if it's already allocated
1940 for (hr = 0; hr < HOST_REGS; hr++)
1942 if (hr != EXCLUDE_REG && cur->regmap[hr] == reg) {
1943 cur->noevict |= 1u << hr;
1948 // Try to allocate any available register
1949 for(hr=HOST_REGS-1;hr>=0;hr--) {
1950 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1951 alloc_set(cur, reg, hr);
1956 // Find an unneeded register
1957 for(hr=HOST_REGS-1;hr>=0;hr--)
1963 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1964 alloc_set(cur, reg, hr);
1971 // Ok, now we have to evict someone
1972 // Pick a register we hopefully won't need soon
1973 evict_alloc_reg(cur, i, reg, 0);
1976 static void mov_alloc(struct regstat *current,int i)
1978 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1979 alloc_cc(current,i); // for stalls
1980 dirty_reg(current,CCREG);
1983 // Note: Don't need to actually alloc the source registers
1984 //alloc_reg(current,i,dops[i].rs1);
1985 alloc_reg(current,i,dops[i].rt1);
1987 clear_const(current,dops[i].rs1);
1988 clear_const(current,dops[i].rt1);
1989 dirty_reg(current,dops[i].rt1);
1992 static void shiftimm_alloc(struct regstat *current,int i)
1994 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1997 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1998 else dops[i].use_lt1=!!dops[i].rs1;
1999 alloc_reg(current,i,dops[i].rt1);
2000 dirty_reg(current,dops[i].rt1);
2001 if(is_const(current,dops[i].rs1)) {
2002 int v=get_const(current,dops[i].rs1);
2003 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<cinfo[i].imm);
2004 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>cinfo[i].imm);
2005 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>cinfo[i].imm);
2007 else clear_const(current,dops[i].rt1);
2012 clear_const(current,dops[i].rs1);
2013 clear_const(current,dops[i].rt1);
2016 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2020 if(dops[i].opcode2==0x3c) // DSLL32
2024 if(dops[i].opcode2==0x3e) // DSRL32
2028 if(dops[i].opcode2==0x3f) // DSRA32
2034 static void shift_alloc(struct regstat *current,int i)
2037 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2038 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
2039 alloc_reg(current,i,dops[i].rt1);
2040 if(dops[i].rt1==dops[i].rs2) {
2041 alloc_reg_temp(current,i,-1);
2042 cinfo[i].min_free_regs=1;
2044 clear_const(current,dops[i].rs1);
2045 clear_const(current,dops[i].rs2);
2046 clear_const(current,dops[i].rt1);
2047 dirty_reg(current,dops[i].rt1);
2051 static void alu_alloc(struct regstat *current,int i)
2053 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2055 if(dops[i].rs1&&dops[i].rs2) {
2056 alloc_reg(current,i,dops[i].rs1);
2057 alloc_reg(current,i,dops[i].rs2);
2060 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2061 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
2063 alloc_reg(current,i,dops[i].rt1);
2065 if (dops[i].may_except) {
2066 alloc_cc_optional(current, i); // for exceptions
2067 alloc_reg_temp(current, i, -1);
2068 cinfo[i].min_free_regs = 1;
2071 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2073 alloc_reg(current,i,dops[i].rs1);
2074 alloc_reg(current,i,dops[i].rs2);
2075 alloc_reg(current,i,dops[i].rt1);
2078 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2080 if(dops[i].rs1&&dops[i].rs2) {
2081 alloc_reg(current,i,dops[i].rs1);
2082 alloc_reg(current,i,dops[i].rs2);
2086 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2087 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
2089 alloc_reg(current,i,dops[i].rt1);
2092 clear_const(current,dops[i].rs1);
2093 clear_const(current,dops[i].rs2);
2094 clear_const(current,dops[i].rt1);
2095 dirty_reg(current,dops[i].rt1);
2098 static void imm16_alloc(struct regstat *current,int i)
2100 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2101 else dops[i].use_lt1=!!dops[i].rs1;
2102 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
2103 if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2104 clear_const(current,dops[i].rs1);
2105 clear_const(current,dops[i].rt1);
2107 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2108 if(is_const(current,dops[i].rs1)) {
2109 int v=get_const(current,dops[i].rs1);
2110 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&cinfo[i].imm);
2111 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|cinfo[i].imm);
2112 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^cinfo[i].imm);
2114 else clear_const(current,dops[i].rt1);
2116 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2117 if(is_const(current,dops[i].rs1)) {
2118 int v=get_const(current,dops[i].rs1);
2119 set_const(current,dops[i].rt1,v+cinfo[i].imm);
2121 else clear_const(current,dops[i].rt1);
2122 if (dops[i].may_except) {
2123 alloc_cc_optional(current, i); // for exceptions
2124 alloc_reg_temp(current, i, -1);
2125 cinfo[i].min_free_regs = 1;
2129 set_const(current,dops[i].rt1,cinfo[i].imm<<16); // LUI
2131 dirty_reg(current,dops[i].rt1);
2134 static void load_alloc(struct regstat *current,int i)
2137 clear_const(current,dops[i].rt1);
2138 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2139 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2140 if (needed_again(dops[i].rs1, i))
2141 alloc_reg(current, i, dops[i].rs1);
2143 alloc_reg(current, i, ROREG);
2144 if (dops[i].may_except) {
2145 alloc_cc_optional(current, i); // for exceptions
2148 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2149 alloc_reg(current,i,dops[i].rt1);
2150 assert(get_reg_w(current->regmap, dops[i].rt1)>=0);
2151 dirty_reg(current,dops[i].rt1);
2152 // LWL/LWR need a temporary register for the old value
2153 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2155 alloc_reg(current,i,FTEMP);
2161 // Load to r0 or unneeded register (dummy load)
2162 // but we still need a register to calculate the address
2163 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2164 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2168 alloc_reg_temp(current, i, -1);
2169 cinfo[i].min_free_regs = 1;
2173 // this may eat up to 7 registers
2174 static void store_alloc(struct regstat *current, int i)
2176 clear_const(current,dops[i].rs2);
2177 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2178 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2179 alloc_reg(current,i,dops[i].rs2);
2181 alloc_reg(current, i, ROREG);
2182 #if defined(HOST_IMM8)
2183 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2184 alloc_reg(current, i, INVCP);
2186 if (dops[i].opcode == 0x2a || dops[i].opcode == 0x2e) { // SWL/SWL
2187 alloc_reg(current,i,FTEMP);
2189 if (dops[i].may_except)
2190 alloc_cc_optional(current, i); // for exceptions
2191 // We need a temporary register for address generation
2192 alloc_reg_temp(current,i,-1);
2193 cinfo[i].min_free_regs=1;
2196 static void c2ls_alloc(struct regstat *current, int i)
2198 clear_const(current,dops[i].rt1);
2199 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2200 alloc_reg(current,i,FTEMP);
2202 alloc_reg(current, i, ROREG);
2203 #if defined(HOST_IMM8)
2204 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2205 if (dops[i].opcode == 0x3a) // SWC2
2206 alloc_reg(current,i,INVCP);
2208 if (dops[i].may_except)
2209 alloc_cc_optional(current, i); // for exceptions
2210 // We need a temporary register for address generation
2211 alloc_reg_temp(current,i,-1);
2212 cinfo[i].min_free_regs=1;
2215 #ifndef multdiv_alloc
2216 static void multdiv_alloc(struct regstat *current,int i)
2222 clear_const(current,dops[i].rs1);
2223 clear_const(current,dops[i].rs2);
2224 alloc_cc(current,i); // for stalls
2225 dirty_reg(current,CCREG);
2226 current->u &= ~(1ull << HIREG);
2227 current->u &= ~(1ull << LOREG);
2228 alloc_reg(current, i, HIREG);
2229 alloc_reg(current, i, LOREG);
2230 dirty_reg(current, HIREG);
2231 dirty_reg(current, LOREG);
2232 if ((dops[i].opcode2 & 0x3e) == 0x1a || (dops[i].rs1 && dops[i].rs2)) // div(u)
2234 alloc_reg(current, i, dops[i].rs1);
2235 alloc_reg(current, i, dops[i].rs2);
2237 // else multiply by zero is zero
2241 static void cop0_alloc(struct regstat *current,int i)
2243 if(dops[i].opcode2==0) // MFC0
2246 clear_const(current,dops[i].rt1);
2247 alloc_reg(current,i,dops[i].rt1);
2248 dirty_reg(current,dops[i].rt1);
2251 else if(dops[i].opcode2==4) // MTC0
2253 if (((source[i]>>11)&0x1e) == 12) {
2254 alloc_cc(current, i);
2255 dirty_reg(current, CCREG);
2258 clear_const(current,dops[i].rs1);
2259 alloc_reg(current,i,dops[i].rs1);
2260 alloc_all(current,i);
2263 alloc_all(current,i); // FIXME: Keep r0
2265 alloc_reg(current,i,0);
2267 cinfo[i].min_free_regs = HOST_REGS;
2271 static void rfe_alloc(struct regstat *current, int i)
2273 alloc_all(current, i);
2274 cinfo[i].min_free_regs = HOST_REGS;
2277 static void cop2_alloc(struct regstat *current,int i)
2279 if (dops[i].opcode2 < 3) // MFC2/CFC2
2281 alloc_cc(current,i); // for stalls
2282 dirty_reg(current,CCREG);
2284 clear_const(current,dops[i].rt1);
2285 alloc_reg(current,i,dops[i].rt1);
2286 dirty_reg(current,dops[i].rt1);
2289 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2292 clear_const(current,dops[i].rs1);
2293 alloc_reg(current,i,dops[i].rs1);
2297 alloc_reg(current,i,0);
2300 alloc_reg_temp(current,i,-1);
2301 cinfo[i].min_free_regs=1;
2304 static void c2op_alloc(struct regstat *current,int i)
2306 alloc_cc(current,i); // for stalls
2307 dirty_reg(current,CCREG);
2308 alloc_reg_temp(current,i,-1);
2311 static void syscall_alloc(struct regstat *current,int i)
2313 alloc_cc(current,i);
2314 dirty_reg(current,CCREG);
2315 alloc_all(current,i);
2316 cinfo[i].min_free_regs=HOST_REGS;
2320 static void delayslot_alloc(struct regstat *current,int i)
2322 switch(dops[i].itype) {
2330 imm16_alloc(current,i);
2334 load_alloc(current,i);
2338 store_alloc(current,i);
2341 alu_alloc(current,i);
2344 shift_alloc(current,i);
2347 multdiv_alloc(current,i);
2350 shiftimm_alloc(current,i);
2353 mov_alloc(current,i);
2356 cop0_alloc(current,i);
2359 rfe_alloc(current,i);
2362 cop2_alloc(current,i);
2365 c2ls_alloc(current,i);
2368 c2op_alloc(current,i);
2373 static void add_stub(enum stub_type type, void *addr, void *retaddr,
2374 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2376 assert(stubcount < ARRAY_SIZE(stubs));
2377 stubs[stubcount].type = type;
2378 stubs[stubcount].addr = addr;
2379 stubs[stubcount].retaddr = retaddr;
2380 stubs[stubcount].a = a;
2381 stubs[stubcount].b = b;
2382 stubs[stubcount].c = c;
2383 stubs[stubcount].d = d;
2384 stubs[stubcount].e = e;
2388 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2389 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2391 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2394 // Write out a single register
2395 static void wb_register(signed char r, const signed char regmap[], u_int dirty)
2398 for(hr=0;hr<HOST_REGS;hr++) {
2399 if(hr!=EXCLUDE_REG) {
2402 assert(regmap[hr]<64);
2403 emit_storereg(r,hr);
2411 static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2413 //if(dirty_pre==dirty) return;
2415 for (hr = 0; hr < HOST_REGS; hr++) {
2417 if (r < 1 || r > 33 || ((u >> r) & 1))
2419 if (((dirty_pre & ~dirty) >> hr) & 1)
2420 emit_storereg(r, hr);
2425 static void pass_args(int a0, int a1)
2429 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2431 else if(a0!=0&&a1==0) {
2433 if (a0>=0) emit_mov(a0,0);
2436 if(a0>=0&&a0!=0) emit_mov(a0,0);
2437 if(a1>=0&&a1!=1) emit_mov(a1,1);
2441 static void alu_assemble(int i, const struct regstat *i_regs, int ccadj_)
2443 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2444 int do_oflow = dops[i].may_except; // ADD/SUB with exceptions enabled
2445 if (dops[i].rt1 || do_oflow) {
2446 int do_exception_check = 0;
2447 signed char s1, s2, t, tmp;
2448 t = get_reg_w(i_regs->regmap, dops[i].rt1);
2449 tmp = get_reg_temp(i_regs->regmap);
2452 if (t < 0 && do_oflow)
2455 s1 = get_reg(i_regs->regmap, dops[i].rs1);
2456 s2 = get_reg(i_regs->regmap, dops[i].rs2);
2457 if (dops[i].rs1 && dops[i].rs2) {
2460 if (dops[i].opcode2 & 2) {
2462 emit_subs(s1, s2, tmp);
2463 do_exception_check = 1;
2470 emit_adds(s1, s2, tmp);
2471 do_exception_check = 1;
2477 else if(dops[i].rs1) {
2478 if(s1>=0) emit_mov(s1,t);
2479 else emit_loadreg(dops[i].rs1,t);
2481 else if(dops[i].rs2) {
2483 emit_loadreg(dops[i].rs2, t);
2486 if (dops[i].opcode2 & 2) {
2489 do_exception_check = 1;
2500 if (do_exception_check) {
2503 if (t >= 0 && tmp != t)
2505 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2509 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2511 signed char s1l,s2l,t;
2513 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2516 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2517 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2518 if(dops[i].rs2==0) // rx<r0
2520 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2522 emit_shrimm(s1l,31,t);
2524 else // SLTU (unsigned can not be less than zero, 0<0)
2527 else if(dops[i].rs1==0) // r0<rx
2530 if(dops[i].opcode2==0x2a) // SLT
2531 emit_set_gz32(s2l,t);
2532 else // SLTU (set if not zero)
2533 emit_set_nz32(s2l,t);
2536 assert(s1l>=0);assert(s2l>=0);
2537 if(dops[i].opcode2==0x2a) // SLT
2538 emit_set_if_less32(s1l,s2l,t);
2540 emit_set_if_carry32(s1l,s2l,t);
2546 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2548 signed char s1l,s2l,tl;
2549 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2552 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2553 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2554 if(dops[i].rs1&&dops[i].rs2) {
2557 if(dops[i].opcode2==0x24) { // AND
2558 emit_and(s1l,s2l,tl);
2560 if(dops[i].opcode2==0x25) { // OR
2561 emit_or(s1l,s2l,tl);
2563 if(dops[i].opcode2==0x26) { // XOR
2564 emit_xor(s1l,s2l,tl);
2566 if(dops[i].opcode2==0x27) { // NOR
2567 emit_or(s1l,s2l,tl);
2573 if(dops[i].opcode2==0x24) { // AND
2576 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2578 if(s1l>=0) emit_mov(s1l,tl);
2579 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2583 if(s2l>=0) emit_mov(s2l,tl);
2584 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2586 else emit_zeroreg(tl);
2588 if(dops[i].opcode2==0x27) { // NOR
2590 if(s1l>=0) emit_not(s1l,tl);
2592 emit_loadreg(dops[i].rs1,tl);
2598 if(s2l>=0) emit_not(s2l,tl);
2600 emit_loadreg(dops[i].rs2,tl);
2604 else emit_movimm(-1,tl);
2613 static void imm16_assemble(int i, const struct regstat *i_regs, int ccadj_)
2615 if (dops[i].opcode==0x0f) { // LUI
2618 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2621 if(!((i_regs->isconst>>t)&1))
2622 emit_movimm(cinfo[i].imm<<16,t);
2626 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2627 int is_addi = dops[i].may_except;
2628 if (dops[i].rt1 || is_addi) {
2629 signed char s, t, tmp;
2630 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2631 s=get_reg(i_regs->regmap,dops[i].rs1);
2633 tmp = get_reg_temp(i_regs->regmap);
2639 if(!((i_regs->isconst>>t)&1)) {
2640 int sum, do_exception_check = 0;
2642 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2644 emit_addimm_and_set_flags3(t, cinfo[i].imm, tmp);
2645 do_exception_check = 1;
2648 emit_addimm(t, cinfo[i].imm, t);
2650 if (!((i_regs->wasconst >> s) & 1)) {
2652 emit_addimm_and_set_flags3(s, cinfo[i].imm, tmp);
2653 do_exception_check = 1;
2656 emit_addimm(s, cinfo[i].imm, t);
2659 int oflow = add_overflow(constmap[i][s], cinfo[i].imm, sum);
2660 if (is_addi && oflow)
2661 do_exception_check = 2;
2663 emit_movimm(sum, t);
2666 if (do_exception_check) {
2668 if (do_exception_check == 2)
2675 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2681 if(!((i_regs->isconst>>t)&1))
2682 emit_movimm(cinfo[i].imm,t);
2687 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2689 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2691 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2692 sl=get_reg(i_regs->regmap,dops[i].rs1);
2696 if(dops[i].opcode==0x0a) { // SLTI
2698 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2699 emit_slti32(t,cinfo[i].imm,t);
2701 emit_slti32(sl,cinfo[i].imm,t);
2706 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2707 emit_sltiu32(t,cinfo[i].imm,t);
2709 emit_sltiu32(sl,cinfo[i].imm,t);
2713 // SLTI(U) with r0 is just stupid,
2714 // nonetheless examples can be found
2715 if(dops[i].opcode==0x0a) // SLTI
2716 if(0<cinfo[i].imm) emit_movimm(1,t);
2717 else emit_zeroreg(t);
2720 if(cinfo[i].imm) emit_movimm(1,t);
2721 else emit_zeroreg(t);
2727 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2730 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2731 sl=get_reg(i_regs->regmap,dops[i].rs1);
2732 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2733 if(dops[i].opcode==0x0c) //ANDI
2737 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2738 emit_andimm(tl,cinfo[i].imm,tl);
2740 if(!((i_regs->wasconst>>sl)&1))
2741 emit_andimm(sl,cinfo[i].imm,tl);
2743 emit_movimm(constmap[i][sl]&cinfo[i].imm,tl);
2753 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2755 if(dops[i].opcode==0x0d) { // ORI
2757 emit_orimm(tl,cinfo[i].imm,tl);
2759 if(!((i_regs->wasconst>>sl)&1))
2760 emit_orimm(sl,cinfo[i].imm,tl);
2762 emit_movimm(constmap[i][sl]|cinfo[i].imm,tl);
2765 if(dops[i].opcode==0x0e) { // XORI
2767 emit_xorimm(tl,cinfo[i].imm,tl);
2769 if(!((i_regs->wasconst>>sl)&1))
2770 emit_xorimm(sl,cinfo[i].imm,tl);
2772 emit_movimm(constmap[i][sl]^cinfo[i].imm,tl);
2777 emit_movimm(cinfo[i].imm,tl);
2785 static void shiftimm_assemble(int i, const struct regstat *i_regs)
2787 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2791 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2792 s=get_reg(i_regs->regmap,dops[i].rs1);
2794 if(t>=0&&!((i_regs->isconst>>t)&1)){
2801 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2803 if(dops[i].opcode2==0) // SLL
2805 emit_shlimm(s<0?t:s,cinfo[i].imm,t);
2807 if(dops[i].opcode2==2) // SRL
2809 emit_shrimm(s<0?t:s,cinfo[i].imm,t);
2811 if(dops[i].opcode2==3) // SRA
2813 emit_sarimm(s<0?t:s,cinfo[i].imm,t);
2817 if(s>=0 && s!=t) emit_mov(s,t);
2821 //emit_storereg(dops[i].rt1,t); //DEBUG
2824 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2828 if(dops[i].opcode2==0x3c) // DSLL32
2832 if(dops[i].opcode2==0x3e) // DSRL32
2836 if(dops[i].opcode2==0x3f) // DSRA32
2842 #ifndef shift_assemble
2843 static void shift_assemble(int i, const struct regstat *i_regs)
2845 signed char s,t,shift;
2846 if (dops[i].rt1 == 0)
2848 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2849 t = get_reg(i_regs->regmap, dops[i].rt1);
2850 s = get_reg(i_regs->regmap, dops[i].rs1);
2851 shift = get_reg(i_regs->regmap, dops[i].rs2);
2857 else if(dops[i].rs2==0) {
2859 if(s!=t) emit_mov(s,t);
2862 host_tempreg_acquire();
2863 emit_andimm(shift,31,HOST_TEMPREG);
2864 switch(dops[i].opcode2) {
2866 emit_shl(s,HOST_TEMPREG,t);
2869 emit_shr(s,HOST_TEMPREG,t);
2872 emit_sar(s,HOST_TEMPREG,t);
2877 host_tempreg_release();
2891 static int get_ptr_mem_type(u_int a)
2893 if(a < 0x00200000) {
2894 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2895 // return wrong, must use memhandler for BIOS self-test to pass
2896 // 007 does similar stuff from a00 mirror, weird stuff
2900 if(0x1f800000 <= a && a < 0x1f801000)
2902 if(0x80200000 <= a && a < 0x80800000)
2904 if(0xa0000000 <= a && a < 0xa0200000)
2909 static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2911 int r = get_reg(i_regs->regmap, ROREG);
2912 if (r < 0 && host_tempreg_free) {
2913 host_tempreg_acquire();
2914 emit_loadreg(ROREG, r = HOST_TEMPREG);
2921 static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2922 int addr, int *offset_reg, int *addr_reg_override, int ccadj_)
2926 int mr = dops[i].rs1;
2929 if(((smrv_strong|smrv_weak)>>mr)&1) {
2930 type=get_ptr_mem_type(ndrc_smrv_regs[mr]);
2931 //printf("set %08x @%08x r%d %d\n", ndrc_smrv_regs[mr], start+i*4, mr, type);
2934 // use the mirror we are running on
2935 type=get_ptr_mem_type(start);
2936 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2939 if (dops[i].may_except) {
2941 u_int op = dops[i].opcode;
2942 int mask = ((op & 0x37) == 0x21 || op == 0x25) ? 1 : 3; // LH/SH/LHU
2944 emit_testimm(addr, mask);
2947 add_stub_r(ALIGNMENT_STUB, jaddr2, out, i, addr, i_regs, ccadj_, 0);
2950 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2951 host_tempreg_acquire();
2952 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2953 addr=*addr_reg_override=HOST_TEMPREG;
2956 else if(type==MTYPE_0000) { // RAM 0 mirror
2957 host_tempreg_acquire();
2958 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2959 addr=*addr_reg_override=HOST_TEMPREG;
2962 else if(type==MTYPE_A000) { // RAM A mirror
2963 host_tempreg_acquire();
2964 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2965 addr=*addr_reg_override=HOST_TEMPREG;
2968 else if(type==MTYPE_1F80) { // scratchpad
2969 if (psxH == (void *)0x1f800000) {
2970 host_tempreg_acquire();
2971 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2972 emit_cmpimm(HOST_TEMPREG,0x1000);
2973 host_tempreg_release();
2978 // do the usual RAM check, jump will go to the right handler
2983 if (type == 0) // need ram check
2985 emit_cmpimm(addr,RAM_SIZE);
2987 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2988 // Hint to branch predictor that the branch is unlikely to be taken
2989 if (dops[i].rs1 >= 28)
2990 emit_jno_unlikely(0);
2994 if (ram_offset != 0)
2995 *offset_reg = get_ro_reg(i_regs, 0);
3001 // return memhandler, or get directly accessable address and return 0
3002 static void *get_direct_memhandler(void *table, u_int addr,
3003 enum stub_type type, uintptr_t *addr_host)
3005 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
3006 uintptr_t l1, l2 = 0;
3007 l1 = ((uintptr_t *)table)[addr>>12];
3009 uintptr_t v = l1 << 1;
3010 *addr_host = v + addr;
3015 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
3016 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
3017 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
3018 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
3020 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
3022 uintptr_t v = l2 << 1;
3023 *addr_host = v + (addr&0xfff);
3026 return (void *)(l2 << 1);
3030 static u_int get_host_reglist(const signed char *regmap)
3032 u_int reglist = 0, hr;
3033 for (hr = 0; hr < HOST_REGS; hr++) {
3034 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
3040 static u_int reglist_exclude(u_int reglist, int r1, int r2)
3043 reglist &= ~(1u << r1);
3045 reglist &= ~(1u << r2);
3049 // find a temp caller-saved register not in reglist (so assumed to be free)
3050 static int reglist_find_free(u_int reglist)
3052 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
3055 return __builtin_ctz(free_regs);
3058 static void do_load_word(int a, int rt, int offset_reg)
3060 if (offset_reg >= 0)
3061 emit_ldr_dualindexed(offset_reg, a, rt);
3063 emit_readword_indexed(0, a, rt);
3066 static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
3068 if (offset_reg < 0) {
3069 emit_writeword_indexed(rt, ofs, a);
3073 emit_addimm(a, ofs, a);
3074 emit_str_dualindexed(offset_reg, a, rt);
3075 if (ofs != 0 && preseve_a)
3076 emit_addimm(a, -ofs, a);
3079 static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
3081 if (offset_reg < 0) {
3082 emit_writehword_indexed(rt, ofs, a);
3086 emit_addimm(a, ofs, a);
3087 emit_strh_dualindexed(offset_reg, a, rt);
3088 if (ofs != 0 && preseve_a)
3089 emit_addimm(a, -ofs, a);
3092 static void do_store_byte(int a, int rt, int offset_reg)
3094 if (offset_reg >= 0)
3095 emit_strb_dualindexed(offset_reg, a, rt);
3097 emit_writebyte_indexed(rt, 0, a);
3100 static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
3102 int addr = cinfo[i].addr;
3106 int memtarget=0,c=0;
3107 int offset_reg = -1;
3108 int fastio_reg_override = -1;
3109 u_int reglist=get_host_reglist(i_regs->regmap);
3110 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3111 s=get_reg(i_regs->regmap,dops[i].rs1);
3112 offset=cinfo[i].imm;
3113 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3115 c=(i_regs->wasconst>>s)&1;
3117 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3120 //printf("load_assemble: c=%d\n",c);
3121 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3122 if(tl<0 && ((!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80) || dops[i].rt1==0)) {
3123 // could be FIFO, must perform the read
3125 assem_debug("(forced read)\n");
3126 tl = get_reg_temp(i_regs->regmap); // may be == addr
3131 //printf("load_assemble: c=%d\n",c);
3132 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3136 // Strmnnrmn's speed hack
3137 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3140 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3141 &offset_reg, &fastio_reg_override, ccadj_);
3144 else if (ram_offset && memtarget) {
3145 offset_reg = get_ro_reg(i_regs, 0);
3147 int dummy=(dops[i].rt1==0)||(tl!=get_reg_w(i_regs->regmap, dops[i].rt1)); // ignore loads to r0 and unneeded reg
3148 switch (dops[i].opcode) {
3153 if (fastio_reg_override >= 0)
3154 a = fastio_reg_override;
3156 if (offset_reg >= 0)
3157 emit_ldrsb_dualindexed(offset_reg, a, tl);
3159 emit_movsbl_indexed(0, a, tl);
3162 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3165 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3171 if (fastio_reg_override >= 0)
3172 a = fastio_reg_override;
3173 if (offset_reg >= 0)
3174 emit_ldrsh_dualindexed(offset_reg, a, tl);
3176 emit_movswl_indexed(0, a, tl);
3179 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3182 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3188 if (fastio_reg_override >= 0)
3189 a = fastio_reg_override;
3190 do_load_word(a, tl, offset_reg);
3193 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3196 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3202 if (fastio_reg_override >= 0)
3203 a = fastio_reg_override;
3205 if (offset_reg >= 0)
3206 emit_ldrb_dualindexed(offset_reg, a, tl);
3208 emit_movzbl_indexed(0, a, tl);
3211 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3214 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3220 if (fastio_reg_override >= 0)
3221 a = fastio_reg_override;
3222 if (offset_reg >= 0)
3223 emit_ldrh_dualindexed(offset_reg, a, tl);
3225 emit_movzwl_indexed(0, a, tl);
3228 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3231 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3237 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3238 host_tempreg_release();
3241 #ifndef loadlr_assemble
3242 static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3244 int addr = cinfo[i].addr;
3245 int s,tl,temp,temp2;
3248 int memtarget=0,c=0;
3249 int offset_reg = -1;
3250 int fastio_reg_override = -1;
3251 u_int reglist=get_host_reglist(i_regs->regmap);
3252 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3253 s=get_reg(i_regs->regmap,dops[i].rs1);
3254 temp=get_reg_temp(i_regs->regmap);
3255 temp2=get_reg(i_regs->regmap,FTEMP);
3256 offset=cinfo[i].imm;
3260 c=(i_regs->wasconst>>s)&1;
3262 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3266 emit_shlimm(addr,3,temp);
3267 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3268 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3270 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3272 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3273 &offset_reg, &fastio_reg_override, ccadj_);
3276 if (ram_offset && memtarget) {
3277 offset_reg = get_ro_reg(i_regs, 0);
3279 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3280 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3282 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3285 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3288 if (fastio_reg_override >= 0)
3289 a = fastio_reg_override;
3290 do_load_word(a, temp2, offset_reg);
3291 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3292 host_tempreg_release();
3293 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3296 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3299 emit_andimm(temp,24,temp);
3300 if (dops[i].opcode==0x22) // LWL
3301 emit_xorimm(temp,24,temp);
3302 host_tempreg_acquire();
3303 emit_movimm(-1,HOST_TEMPREG);
3304 if (dops[i].opcode==0x26) {
3305 emit_shr(temp2,temp,temp2);
3306 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3308 emit_shl(temp2,temp,temp2);
3309 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3311 host_tempreg_release();
3312 emit_or(temp2,tl,tl);
3314 //emit_storereg(dops[i].rt1,tl); // DEBUG
3316 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3322 static void do_invstub(int n)
3325 assem_debug("do_invstub %x\n", start + stubs[n].e*4);
3326 u_int reglist = stubs[n].a;
3327 u_int addrr = stubs[n].b;
3328 int ofs_start = stubs[n].c;
3329 int ofs_end = stubs[n].d;
3330 int len = ofs_end - ofs_start;
3333 set_jump_target(stubs[n].addr, out);
3335 if (addrr != 0 || ofs_start != 0)
3336 emit_addimm(addrr, ofs_start, 0);
3337 emit_readword(&inv_code_start, 2);
3338 emit_readword(&inv_code_end, 3);
3340 emit_addimm(0, len + 4, (rightr = 1));
3342 emit_cmpcs(3, rightr);
3345 void *func = (len != 0)
3346 ? (void *)ndrc_write_invalidate_many
3347 : (void *)ndrc_write_invalidate_one;
3348 emit_far_call(func);
3349 set_jump_target(jaddr, out);
3350 restore_regs(reglist);
3351 emit_jmp(stubs[n].retaddr);
3354 static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3356 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3358 // this can't be used any more since we started to check exact
3359 // block boundaries in invalidate_range()
3360 //if (i_regs->waswritten & (1<<dops[i].rs1))
3362 // (naively) assume nobody will run code from stack
3363 if (dops[i].rs1 == 29)
3366 int j, imm_maxdiff = 32, imm_min = cinfo[i].imm, imm_max = cinfo[i].imm, count = 1;
3367 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3368 && abs(cinfo[i+1].imm - cinfo[i].imm) <= imm_maxdiff)
3370 for (j = i - 1; j >= 0; j--) {
3371 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3372 || abs(cinfo[j].imm - cinfo[j+1].imm) > imm_maxdiff)
3375 if (imm_min > cinfo[j].imm)
3376 imm_min = cinfo[j].imm;
3377 if (imm_max < cinfo[j].imm)
3378 imm_max = cinfo[j].imm;
3380 #if defined(HOST_IMM8)
3381 int ir = get_reg(i_regs->regmap, INVCP);
3383 host_tempreg_acquire();
3384 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3386 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3390 #ifdef INVALIDATE_USE_COND_CALL
3392 emit_cmpimm(HOST_TEMPREG, 1);
3393 emit_callne(invalidate_addr_reg[addr]);
3394 host_tempreg_release();
3398 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3399 host_tempreg_release();
3400 imm_min -= cinfo[i].imm;
3401 imm_max -= cinfo[i].imm;
3402 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3403 addr, imm_min, imm_max, i);
3406 // determines if code overwrite checking is needed only
3407 // (also true non-existent 0x20000000 mirror that shouldn't matter)
3408 #define is_ram_addr(a) !((a) & 0x5f800000)
3410 static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3413 int addr = cinfo[i].addr;
3416 enum stub_type type=0;
3417 int memtarget=0,c=0;
3418 int offset_reg = -1;
3419 int fastio_reg_override = -1;
3420 u_int addr_const = ~0;
3421 u_int reglist=get_host_reglist(i_regs->regmap);
3422 tl=get_reg(i_regs->regmap,dops[i].rs2);
3423 s=get_reg(i_regs->regmap,dops[i].rs1);
3424 offset=cinfo[i].imm;
3426 c=(i_regs->wasconst>>s)&1;
3428 addr_const = constmap[i][s] + offset;
3429 memtarget = ((signed int)addr_const) < (signed int)(0x80000000 + RAM_SIZE);
3434 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3435 reglist |= 1u << addr;
3437 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3438 &offset_reg, &fastio_reg_override, ccadj_);
3440 else if (ram_offset && memtarget) {
3441 offset_reg = get_ro_reg(i_regs, 0);
3444 switch (dops[i].opcode) {
3448 if (fastio_reg_override >= 0)
3449 a = fastio_reg_override;
3450 do_store_byte(a, tl, offset_reg);
3457 if (fastio_reg_override >= 0)
3458 a = fastio_reg_override;
3459 do_store_hword(a, 0, tl, offset_reg, 1);
3466 if (fastio_reg_override >= 0)
3467 a = fastio_reg_override;
3468 do_store_word(a, 0, tl, offset_reg, 1);
3475 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3476 host_tempreg_release();
3478 // PCSX store handlers don't check invcode again
3479 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3481 if (!c || is_ram_addr(addr_const))
3482 do_store_smc_check(i, i_regs, reglist, addr);
3483 if (c && !memtarget)
3484 inline_writestub(type, i, addr_const, i_regs->regmap, dops[i].rs2, ccadj_, reglist);
3485 // basic current block modification detection..
3486 // not looking back as that should be in mips cache already
3487 // (see Spyro2 title->attract mode)
3488 if (start + i*4 < addr_const && addr_const < start + slen*4) {
3489 SysPrintf_lim("write to %08x hits block %08x, pc=%08x\n",
3490 addr_const, start, start+i*4);
3491 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3492 if(i_regs->regmap==regs[i].regmap) {
3493 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3494 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3495 emit_readptr(&hash_table_ptr, 1);
3496 emit_movimm(start+i*4+4, 0);
3497 emit_writeword(0, &psxRegs.pc);
3498 emit_addimm(HOST_CCREG, 2, HOST_CCREG);
3499 emit_far_call(ndrc_get_addr_ht);
3505 static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3507 int addr = cinfo[i].addr;
3511 void *case1, *case23, *case3;
3512 void *done0, *done1, *done2;
3513 int memtarget=0,c=0;
3514 int offset_reg = -1;
3515 u_int addr_const = ~0;
3516 u_int reglist = get_host_reglist(i_regs->regmap);
3517 tl=get_reg(i_regs->regmap,dops[i].rs2);
3518 s=get_reg(i_regs->regmap,dops[i].rs1);
3519 offset=cinfo[i].imm;
3521 c = (i_regs->isconst >> s) & 1;
3523 addr_const = constmap[i][s] + offset;
3524 memtarget = ((signed int)addr_const) < (signed int)(0x80000000 + RAM_SIZE);
3529 reglist |= 1u << addr;
3531 emit_cmpimm(addr, RAM_SIZE);
3537 if(!memtarget||!dops[i].rs1) {
3543 offset_reg = get_ro_reg(i_regs, 0);
3545 emit_testimm(addr,2);
3548 emit_testimm(addr,1);
3552 if (dops[i].opcode == 0x2A) { // SWL
3553 // Write msb into least significant byte
3554 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3555 do_store_byte(addr, tl, offset_reg);
3556 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3558 else if (dops[i].opcode == 0x2E) { // SWR
3559 // Write entire word
3560 do_store_word(addr, 0, tl, offset_reg, 1);
3565 set_jump_target(case1, out);
3566 if (dops[i].opcode == 0x2A) { // SWL
3567 // Write two msb into two least significant bytes
3568 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3569 do_store_hword(addr, -1, tl, offset_reg, 1);
3570 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3572 else if (dops[i].opcode == 0x2E) { // SWR
3573 // Write 3 lsb into three most significant bytes
3574 do_store_byte(addr, tl, offset_reg);
3575 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3576 do_store_hword(addr, 1, tl, offset_reg, 1);
3577 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3582 set_jump_target(case23, out);
3583 emit_testimm(addr,1);
3587 if (dops[i].opcode==0x2A) { // SWL
3588 // Write 3 msb into three least significant bytes
3589 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3590 do_store_hword(addr, -2, tl, offset_reg, 1);
3591 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3592 do_store_byte(addr, tl, offset_reg);
3593 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3595 else if (dops[i].opcode == 0x2E) { // SWR
3596 // Write two lsb into two most significant bytes
3597 do_store_hword(addr, 0, tl, offset_reg, 1);
3602 set_jump_target(case3, out);
3603 if (dops[i].opcode == 0x2A) { // SWL
3604 do_store_word(addr, -3, tl, offset_reg, 1);
3606 else if (dops[i].opcode == 0x2E) { // SWR
3607 do_store_byte(addr, tl, offset_reg);
3609 set_jump_target(done0, out);
3610 set_jump_target(done1, out);
3611 set_jump_target(done2, out);
3612 if (offset_reg == HOST_TEMPREG)
3613 host_tempreg_release();
3614 if (!c || !memtarget)
3615 add_stub_r(STORELR_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3616 if (!c || is_ram_addr(addr_const))
3617 do_store_smc_check(i, i_regs, reglist, addr);
3620 static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3622 if(dops[i].opcode2==0) // MFC0
3624 signed char t=get_reg_w(i_regs->regmap, dops[i].rt1);
3625 u_int copr=(source[i]>>11)&0x1f;
3626 if(t>=0&&dops[i].rt1!=0) {
3627 emit_readword(&psxRegs.CP0.r[copr],t);
3630 else if(dops[i].opcode2==4) // MTC0
3632 int s = get_reg(i_regs->regmap, dops[i].rs1);
3633 int cc = get_reg(i_regs->regmap, CCREG);
3634 char copr=(source[i]>>11)&0x1f;
3636 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3637 if (copr == 12 || copr == 13) {
3638 emit_readword(&last_count,HOST_TEMPREG);
3639 if (cc != HOST_CCREG)
3640 emit_loadreg(CCREG, HOST_CCREG);
3641 emit_add(HOST_CCREG, HOST_TEMPREG, HOST_CCREG);
3642 emit_addimm(HOST_CCREG, ccadj_ + 2, HOST_CCREG);
3643 emit_writeword(HOST_CCREG, &psxRegs.cycle);
3645 // burn cycles to cause cc_interrupt, which will
3646 // reschedule next_interupt. Relies on CCREG from above.
3647 assem_debug("MTC0 DS %d\n", copr);
3648 emit_writeword(HOST_CCREG,&last_count);
3649 emit_movimm(0,HOST_CCREG);
3650 emit_storereg(CCREG,HOST_CCREG);
3651 emit_loadreg(dops[i].rs1, 2);
3652 emit_movimm(copr, 1);
3653 emit_addimm_ptr(FP, (u_char *)&psxRegs - (u_char *)&dynarec_local, 0);
3654 emit_far_call(pcsx_mtc0_ds);
3655 emit_loadreg(dops[i].rs1,s);
3658 emit_movimm(start+i*4+4,HOST_TEMPREG);
3659 emit_writeword(HOST_TEMPREG,&psxRegs.pc);
3663 emit_movimm(copr, 1);
3664 emit_addimm_ptr(FP, (u_char *)&psxRegs - (u_char *)&dynarec_local, 0);
3665 emit_far_call(pcsx_mtc0);
3666 if (copr == 12 || copr == 13) {
3667 emit_readword(&psxRegs.cycle,HOST_CCREG);
3668 emit_readword(&last_count,HOST_TEMPREG);
3669 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3670 //emit_writeword(HOST_TEMPREG,&last_count);
3671 assert(!is_delayslot);
3672 emit_readword(&psxRegs.pc, 0);
3673 emit_movimm(start+i*4+4, HOST_TEMPREG);
3674 emit_cmp(HOST_TEMPREG, 0);
3677 emit_readptr(&hash_table_ptr, 1);
3678 emit_far_call(ndrc_get_addr_ht);
3680 set_jump_target(jaddr, out);
3681 emit_addimm(HOST_CCREG, -ccadj_ - 2, HOST_CCREG);
3682 if (cc != HOST_CCREG)
3683 emit_storereg(CCREG, HOST_CCREG);
3685 emit_loadreg(dops[i].rs1,s);
3689 static void rfe_assemble(int i, const struct regstat *i_regs)
3691 emit_readword(&psxRegs.CP0.n.SR, 0);
3692 emit_andimm(0, 0x3c, 1);
3693 emit_andimm(0, ~0xf, 0);
3694 emit_orrshr_imm(1, 2, 0);
3695 emit_writeword(0, &psxRegs.CP0.n.SR);
3698 static int cop2_is_stalling_op(int i, int *cycles)
3700 if (dops[i].itype == COP2 || dops[i].itype == C2LS) {
3704 if (dops[i].itype == C2OP) {
3705 *cycles = gte_cycletab[source[i] & 0x3f];
3708 // ... what about MTC2/CTC2/LWC2?
3713 static void log_gte_stall(int stall, u_int cycle)
3715 if ((u_int)stall <= 44)
3716 printf("x stall %2d %u\n", stall, cycle + last_count);
3719 static void emit_log_gte_stall(int i, int stall, u_int reglist)
3723 emit_movimm(stall, 0);
3725 emit_mov(HOST_TEMPREG, 0);
3726 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3727 emit_far_call(log_gte_stall);
3728 restore_regs(reglist);
3732 static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3734 int j = i, cycles, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3735 int rtmp = reglist_find_free(reglist);
3737 if (HACK_ENABLED(NDHACK_NO_STALLS))
3739 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3740 // happens occasionally... cc evicted? Don't bother then
3741 //printf("no cc %08x\n", start + i*4);
3745 for (j = i - 1; j >= 0; j--) {
3746 //if (dops[j].is_ds) break;
3747 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3749 if (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj)
3754 cycles_passed = cinfo[i].ccadj - cinfo[j].ccadj;
3755 if (other_gte_op_cycles >= 0)
3756 stall = other_gte_op_cycles - cycles_passed;
3757 else if (cycles_passed >= 44)
3758 stall = 0; // can't possibly stall
3759 if (stall == -MAXBLOCK && rtmp >= 0) {
3760 // unknown stall, do the expensive runtime check
3761 assem_debug("; cop2_do_stall_check\n");
3762 // busy - (cc + adj) -> busy - adj - cc
3763 host_tempreg_acquire();
3764 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3765 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3766 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3767 emit_cmpimm(HOST_TEMPREG, 44);
3768 emit_cmovb_reg(rtmp, HOST_CCREG);
3769 //emit_log_gte_stall(i, 0, reglist);
3770 host_tempreg_release();
3772 else if (stall > 0) {
3773 //emit_log_gte_stall(i, stall, reglist);
3774 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3777 // save gteBusyCycle, if needed
3778 cycles = gte_cycletab[op];
3781 other_gte_op_cycles = -1;
3782 for (j = i + 1; j < slen; j++) {
3783 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3785 if (dops[j].is_jump) {
3787 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3792 if (other_gte_op_cycles >= 0)
3793 // will handle stall when assembling that op
3795 cycles_passed = cinfo[min(j, slen -1)].ccadj - cinfo[i].ccadj;
3796 if (cycles_passed >= cycles)
3798 assem_debug("; save gteBusyCycle\n");
3799 host_tempreg_acquire();
3800 emit_addimm(HOST_CCREG, cinfo[i].ccadj + cycles, HOST_TEMPREG);
3801 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3802 host_tempreg_release();
3805 static int is_mflohi(int i)
3807 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3810 static int check_multdiv(int i, int *cycles)
3812 if (dops[i].itype != MULTDIV)
3814 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3815 *cycles = 11; // approx from 7 11 14
3821 static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3823 int j, found = 0, c = 0;
3824 if (HACK_ENABLED(NDHACK_NO_STALLS))
3826 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3827 // happens occasionally... cc evicted? Don't bother then
3830 for (j = i + 1; j < slen; j++) {
3833 if ((found = is_mflohi(j)))
3835 if (dops[j].is_jump) {
3837 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3843 // handle all in multdiv_do_stall()
3845 check_multdiv(i, &c);
3847 assem_debug("; muldiv prepare stall %d\n", c);
3848 host_tempreg_acquire();
3849 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3850 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3851 host_tempreg_release();
3854 static void multdiv_do_stall(int i, const struct regstat *i_regs)
3856 int j, known_cycles = 0;
3857 u_int reglist = get_host_reglist(i_regs->regmap);
3858 int rtmp = get_reg_temp(i_regs->regmap);
3860 rtmp = reglist_find_free(reglist);
3861 if (HACK_ENABLED(NDHACK_NO_STALLS))
3863 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3864 // happens occasionally... cc evicted? Don't bother then
3865 //printf("no cc/rtmp %08x\n", start + i*4);
3869 for (j = i - 1; j >= 0; j--) {
3870 if (dops[j].is_ds) break;
3871 if (check_multdiv(j, &known_cycles))
3874 // already handled by this op
3876 if (dops[j].bt || (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj))
3881 if (known_cycles > 0) {
3882 known_cycles -= cinfo[i].ccadj - cinfo[j].ccadj;
3883 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3884 if (known_cycles > 0)
3885 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3888 assem_debug("; muldiv stall unresolved\n");
3889 host_tempreg_acquire();
3890 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3891 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3892 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3893 emit_cmpimm(HOST_TEMPREG, 37);
3894 emit_cmovb_reg(rtmp, HOST_CCREG);
3895 //emit_log_gte_stall(i, 0, reglist);
3896 host_tempreg_release();
3899 static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3909 emit_readword(®_cop2d[copr],tl);
3910 emit_signextend16(tl,tl);
3911 emit_writeword(tl,®_cop2d[copr]); // hmh
3918 emit_readword(®_cop2d[copr],tl);
3919 emit_andimm(tl,0xffff,tl);
3920 emit_writeword(tl,®_cop2d[copr]);
3923 emit_readword(®_cop2d[14],tl); // SXY2
3924 emit_writeword(tl,®_cop2d[copr]);
3928 c2op_mfc2_29_assemble(tl,temp);
3931 emit_readword(®_cop2d[copr],tl);
3936 static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3940 emit_readword(®_cop2d[13],temp); // SXY1
3941 emit_writeword(sl,®_cop2d[copr]);
3942 emit_writeword(temp,®_cop2d[12]); // SXY0
3943 emit_readword(®_cop2d[14],temp); // SXY2
3944 emit_writeword(sl,®_cop2d[14]);
3945 emit_writeword(temp,®_cop2d[13]); // SXY1
3948 emit_andimm(sl,0x001f,temp);
3949 emit_shlimm(temp,7,temp);
3950 emit_writeword(temp,®_cop2d[9]);
3951 emit_andimm(sl,0x03e0,temp);
3952 emit_shlimm(temp,2,temp);
3953 emit_writeword(temp,®_cop2d[10]);
3954 emit_andimm(sl,0x7c00,temp);
3955 emit_shrimm(temp,3,temp);
3956 emit_writeword(temp,®_cop2d[11]);
3957 emit_writeword(sl,®_cop2d[28]);
3960 emit_xorsar_imm(sl,sl,31,temp);
3961 #if defined(HAVE_ARMV5) || defined(__aarch64__)
3962 emit_clz(temp,temp);
3964 emit_movs(temp,HOST_TEMPREG);
3965 emit_movimm(0,temp);
3966 emit_jeq((int)out+4*4);
3967 emit_addpl_imm(temp,1,temp);
3968 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3969 emit_jns((int)out-2*4);
3971 emit_writeword(sl,®_cop2d[30]);
3972 emit_writeword(temp,®_cop2d[31]);
3977 emit_writeword(sl,®_cop2d[copr]);
3982 static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3987 int memtarget=0,c=0;
3989 enum stub_type type;
3990 int offset_reg = -1;
3991 int fastio_reg_override = -1;
3992 u_int addr_const = ~0;
3993 u_int reglist=get_host_reglist(i_regs->regmap);
3994 u_int copr=(source[i]>>16)&0x1f;
3995 s=get_reg(i_regs->regmap,dops[i].rs1);
3996 tl=get_reg(i_regs->regmap,FTEMP);
3997 offset=cinfo[i].imm;
4000 if(i_regs->regmap[HOST_CCREG]==CCREG)
4001 reglist&=~(1<<HOST_CCREG);
4006 if (dops[i].opcode==0x3a) { // SWC2
4010 c = (i_regs->isconst >> s) & 1;
4012 addr_const = constmap[i][s] + offset;
4013 memtarget = ((signed int)addr_const) < (signed int)(0x80000000 + RAM_SIZE);
4017 cop2_do_stall_check(0, i, i_regs, reglist);
4019 if (dops[i].opcode==0x3a) { // SWC2
4020 cop2_get_dreg(copr,tl,-1);
4028 emit_jmp(0); // inline_readstub/inline_writestub?
4032 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
4033 &offset_reg, &fastio_reg_override, ccadj_);
4035 else if (ram_offset && memtarget) {
4036 offset_reg = get_ro_reg(i_regs, 0);
4038 switch (dops[i].opcode) {
4039 case 0x32: { // LWC2
4041 if (fastio_reg_override >= 0)
4042 a = fastio_reg_override;
4043 do_load_word(a, tl, offset_reg);
4046 case 0x3a: { // SWC2
4047 #ifdef DESTRUCTIVE_SHIFT
4048 if(!offset&&!c&&s>=0) emit_mov(s,ar);
4051 if (fastio_reg_override >= 0)
4052 a = fastio_reg_override;
4053 do_store_word(a, 0, tl, offset_reg, 1);
4060 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
4061 host_tempreg_release();
4063 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
4064 if (dops[i].opcode == 0x3a && (!c || is_ram_addr(addr_const))) // SWC2
4065 do_store_smc_check(i, i_regs, reglist, ar);
4066 if (dops[i].opcode == 0x32) { // LWC2
4067 host_tempreg_acquire();
4068 cop2_put_dreg(copr,tl,HOST_TEMPREG);
4069 host_tempreg_release();
4073 static void cop2_assemble(int i, const struct regstat *i_regs)
4075 u_int copr = (source[i]>>11) & 0x1f;
4076 signed char temp = get_reg_temp(i_regs->regmap);
4078 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
4079 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
4080 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
4081 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
4082 reglist = reglist_exclude(reglist, tl, -1);
4084 cop2_do_stall_check(0, i, i_regs, reglist);
4086 if (dops[i].opcode2==0) { // MFC2
4087 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4088 if(tl>=0&&dops[i].rt1!=0)
4089 cop2_get_dreg(copr,tl,temp);
4091 else if (dops[i].opcode2==4) { // MTC2
4092 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4093 cop2_put_dreg(copr,sl,temp);
4095 else if (dops[i].opcode2==2) // CFC2
4097 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4098 if(tl>=0&&dops[i].rt1!=0)
4099 emit_readword(®_cop2c[copr],tl);
4101 else if (dops[i].opcode2==6) // CTC2
4103 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4112 emit_signextend16(sl,temp);
4115 c2op_ctc2_31_assemble(sl,temp);
4121 emit_writeword(temp,®_cop2c[copr]);
4126 static void do_unalignedwritestub(int n)
4128 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4130 set_jump_target(stubs[n].addr, out);
4133 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4134 int addr=stubs[n].b;
4135 u_int reglist=stubs[n].e;
4136 signed char *i_regmap=i_regs->regmap;
4137 int temp2=get_reg(i_regmap,FTEMP);
4139 rt=get_reg(i_regmap,dops[i].rs2);
4142 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4144 reglist&=~(1<<temp2);
4146 // don't bother with it and call write handler
4149 int cc=get_reg(i_regmap,CCREG);
4151 emit_loadreg(CCREG,2);
4152 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4153 emit_movimm(start + i*4,3);
4154 emit_writeword(3,&psxRegs.pc);
4155 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4156 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4158 emit_storereg(CCREG,2);
4159 restore_regs(reglist);
4160 emit_jmp(stubs[n].retaddr); // return address
4163 static void do_overflowstub(int n)
4165 assem_debug("do_overflowstub %x\n", start + (u_int)stubs[n].a * 4);
4168 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4169 int ccadj = stubs[n].d;
4170 set_jump_target(stubs[n].addr, out);
4171 wb_dirtys(regs[i].regmap, regs[i].dirty);
4172 exception_assemble(i, i_regs, ccadj);
4175 static void do_alignmentstub(int n)
4177 assem_debug("do_alignmentstub %x\n", start + (u_int)stubs[n].a * 4);
4180 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4181 int ccadj = stubs[n].d;
4182 int is_store = dops[i].itype == STORE || dops[i].opcode == 0x3A; // SWC2
4183 int cause = (dops[i].opcode & 3) << 28;
4184 cause |= is_store ? (R3000E_AdES << 2) : (R3000E_AdEL << 2);
4185 set_jump_target(stubs[n].addr, out);
4186 wb_dirtys(regs[i].regmap, regs[i].dirty);
4187 if (stubs[n].b != 1)
4188 emit_mov(stubs[n].b, 1); // faulting address
4189 emit_movimm(cause, 0);
4190 exception_assemble(i, i_regs, ccadj);
4193 #ifndef multdiv_assemble
4194 void multdiv_assemble(int i,struct regstat *i_regs)
4196 printf("Need multdiv_assemble for this architecture.\n");
4201 static void mov_assemble(int i, const struct regstat *i_regs)
4203 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
4204 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
4207 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4210 sl=get_reg(i_regs->regmap,dops[i].rs1);
4211 if(sl>=0) emit_mov(sl,tl);
4212 else emit_loadreg(dops[i].rs1,tl);
4215 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
4216 multdiv_do_stall(i, i_regs);
4219 // call interpreter, exception handler, things that change pc/regs/cycles ...
4220 static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
4222 signed char ccreg=get_reg(i_regs->regmap,CCREG);
4223 assert(ccreg==HOST_CCREG);
4224 assert(!is_delayslot);
4227 emit_movimm(pc,3); // Get PC
4228 emit_readword(&last_count,2);
4229 emit_writeword(3,&psxRegs.pc);
4230 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
4231 emit_add(2,HOST_CCREG,2);
4232 emit_writeword(2,&psxRegs.cycle);
4233 emit_addimm_ptr(FP,(u_char *)&psxRegs - (u_char *)&dynarec_local,0);
4234 emit_far_call(func);
4235 emit_far_jump(jump_to_new_pc);
4238 static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_)
4240 // 'break' tends to be littered around to catch things like
4241 // division by 0 and is almost never executed, so don't emit much code here
4243 if (dops[i].itype == ALU || dops[i].itype == IMM16)
4244 func = is_delayslot ? jump_overflow_ds : jump_overflow;
4245 else if (dops[i].itype == LOAD || dops[i].itype == STORE)
4246 func = is_delayslot ? jump_addrerror_ds : jump_addrerror;
4247 else if (dops[i].opcode2 == 0x0C)
4248 func = is_delayslot ? jump_syscall_ds : jump_syscall;
4250 func = is_delayslot ? jump_break_ds : jump_break;
4251 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) // evicted
4252 emit_loadreg(CCREG, HOST_CCREG);
4253 emit_movimm(start + i*4, 2); // pc
4254 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4255 emit_far_jump(func);
4258 static void hlecall_bad()
4263 static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4265 void *hlefunc = hlecall_bad;
4266 uint32_t hleCode = source[i] & 0x03ffffff;
4267 if (hleCode < ARRAY_SIZE(psxHLEt))
4268 hlefunc = psxHLEt[hleCode];
4270 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4273 static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4275 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4278 static void speculate_mov(int rs,int rt)
4281 smrv_strong_next |= 1 << rt;
4282 ndrc_smrv_regs[rt] = ndrc_smrv_regs[rs];
4286 static void speculate_mov_weak(int rs,int rt)
4289 smrv_weak_next |= 1 << rt;
4290 ndrc_smrv_regs[rt] = ndrc_smrv_regs[rs];
4294 static void speculate_register_values(int i)
4297 // gp,sp are likely to stay the same throughout the block
4298 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4299 smrv_weak_next=~smrv_strong_next;
4300 //printf(" llr %08x\n", ndrc_smrv_regs[4]);
4302 smrv_strong=smrv_strong_next;
4303 smrv_weak=smrv_weak_next;
4304 switch(dops[i].itype) {
4306 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4307 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4308 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4309 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4311 smrv_strong_next&=~(1<<dops[i].rt1);
4312 smrv_weak_next&=~(1<<dops[i].rt1);
4316 smrv_strong_next&=~(1<<dops[i].rt1);
4317 smrv_weak_next&=~(1<<dops[i].rt1);
4320 if(dops[i].rt1&&is_const(®s[i],dops[i].rt1)) {
4321 int hr = get_reg_w(regs[i].regmap, dops[i].rt1);
4324 if(get_final_value(hr,i,&value))
4325 ndrc_smrv_regs[dops[i].rt1]=value;
4326 else ndrc_smrv_regs[dops[i].rt1]=constmap[i][hr];
4327 smrv_strong_next|=1<<dops[i].rt1;
4331 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4332 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4336 if(start<0x2000&&(dops[i].rt1==26||(ndrc_smrv_regs[dops[i].rt1]>>24)==0xa0)) {
4337 // special case for BIOS
4338 ndrc_smrv_regs[dops[i].rt1]=0xa0000000;
4339 smrv_strong_next|=1<<dops[i].rt1;
4346 smrv_strong_next&=~(1<<dops[i].rt1);
4347 smrv_weak_next&=~(1<<dops[i].rt1);
4351 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4352 smrv_strong_next&=~(1<<dops[i].rt1);
4353 smrv_weak_next&=~(1<<dops[i].rt1);
4357 if (dops[i].opcode==0x32) { // LWC2
4358 smrv_strong_next&=~(1<<dops[i].rt1);
4359 smrv_weak_next&=~(1<<dops[i].rt1);
4365 printf("x %08x %08x %d %d c %08x %08x\n",ndrc_smrv_regs[r],start+i*4,
4366 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4370 static void ujump_assemble(int i, const struct regstat *i_regs);
4371 static void rjump_assemble(int i, const struct regstat *i_regs);
4372 static void cjump_assemble(int i, const struct regstat *i_regs);
4373 static void sjump_assemble(int i, const struct regstat *i_regs);
4375 static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4378 switch (dops[i].itype) {
4380 alu_assemble(i, i_regs, ccadj_);
4383 imm16_assemble(i, i_regs, ccadj_);
4386 shift_assemble(i, i_regs);
4389 shiftimm_assemble(i, i_regs);
4392 load_assemble(i, i_regs, ccadj_);
4395 loadlr_assemble(i, i_regs, ccadj_);
4398 store_assemble(i, i_regs, ccadj_);
4401 storelr_assemble(i, i_regs, ccadj_);
4404 cop0_assemble(i, i_regs, ccadj_);
4407 rfe_assemble(i, i_regs);
4410 cop2_assemble(i, i_regs);
4413 c2ls_assemble(i, i_regs, ccadj_);
4416 c2op_assemble(i, i_regs);
4419 multdiv_assemble(i, i_regs);
4420 multdiv_prepare_stall(i, i_regs, ccadj_);
4423 mov_assemble(i, i_regs);
4426 exception_assemble(i, i_regs, ccadj_);
4429 hlecall_assemble(i, i_regs, ccadj_);
4432 intcall_assemble(i, i_regs, ccadj_);
4435 ujump_assemble(i, i_regs);
4439 rjump_assemble(i, i_regs);
4443 cjump_assemble(i, i_regs);
4447 sjump_assemble(i, i_regs);
4452 // not handled, just skip
4460 static void ds_assemble(int i, const struct regstat *i_regs)
4462 speculate_register_values(i);
4464 switch (dops[i].itype) {
4472 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4475 assemble(i, i_regs, cinfo[i].ccadj);
4480 // Is the branch target a valid internal jump?
4481 static int internal_branch(int addr)
4483 if(addr&1) return 0; // Indirect (register) jump
4484 if(addr>=start && addr<start+slen*4-4)
4491 static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4494 for(hr=0;hr<HOST_REGS;hr++) {
4495 if(hr!=EXCLUDE_REG) {
4496 if(pre[hr]!=entry[hr]) {
4499 if(get_reg(entry,pre[hr])<0) {
4501 if(!((u>>pre[hr])&1))
4502 emit_storereg(pre[hr],hr);
4509 // Move from one register to another (no writeback)
4510 for(hr=0;hr<HOST_REGS;hr++) {
4511 if(hr!=EXCLUDE_REG) {
4512 if(pre[hr]!=entry[hr]) {
4513 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4515 if((nr=get_reg(entry,pre[hr]))>=0) {
4524 // Load the specified registers
4525 // This only loads the registers given as arguments because
4526 // we don't want to load things that will be overwritten
4527 static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4529 int hr = get_reg(regmap, rs);
4530 if (hr >= 0 && entry[hr] != regmap[hr])
4531 emit_loadreg(regmap[hr], hr);
4534 static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4536 load_reg(entry, regmap, rs1);
4538 load_reg(entry, regmap, rs2);
4541 // Load registers prior to the start of a loop
4542 // so that they are not loaded within the loop
4543 static void loop_preload(signed char pre[],signed char entry[])
4546 for (hr = 0; hr < HOST_REGS; hr++) {
4548 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4549 assem_debug("loop preload:\n");
4551 emit_loadreg(r, hr);
4556 // Generate address for load/store instruction
4557 // goes to AGEN (or temp) for writes, FTEMP for LOADLR and cop1/2 loads
4558 // AGEN is assigned by pass5b_preallocate2
4559 static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4561 if (dops[i].is_load || dops[i].is_store) {
4563 int agr = AGEN1 + (i&1);
4564 if(dops[i].itype==LOAD) {
4565 if (!dops[i].may_except)
4566 ra = get_reg_w(i_regs->regmap, dops[i].rt1); // reuse dest for agen
4568 ra = get_reg_temp(i_regs->regmap);
4570 if(dops[i].itype==LOADLR) {
4571 ra=get_reg(i_regs->regmap,FTEMP);
4573 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4574 ra=get_reg(i_regs->regmap,agr);
4575 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4577 if(dops[i].itype==C2LS) {
4578 if (dops[i].opcode == 0x32) // LWC2
4579 ra=get_reg(i_regs->regmap,FTEMP);
4581 ra=get_reg(i_regs->regmap,agr);
4582 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4585 int rs = get_reg(i_regs->regmap, dops[i].rs1);
4588 int offset = cinfo[i].imm;
4589 int add_offset = offset != 0;
4590 int c = rs >= 0 && ((i_regs->wasconst >> rs) & 1);
4591 if(dops[i].rs1==0) {
4592 // Using r0 as a base address
4594 if(!entry||entry[ra]!=agr) {
4595 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4596 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4598 emit_movimm(offset,ra);
4600 } // else did it in the previous cycle
4606 if (!entry || entry[ra] != dops[i].rs1)
4607 emit_loadreg(dops[i].rs1, ra);
4609 //if(!entry||entry[ra]!=dops[i].rs1)
4610 // printf("poor load scheduling!\n");
4613 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4615 if(!entry||entry[ra]!=agr) {
4616 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4617 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4619 emit_movimm(constmap[i][rs]+offset,ra);
4620 regs[i].loadedconst|=1<<ra;
4622 } // else did it in the previous cycle
4625 else // else load_consts already did it
4634 emit_addimm(rs,offset,ra);
4636 emit_addimm(ra,offset,ra);
4641 assert(cinfo[i].addr >= 0);
4643 // Preload constants for next instruction
4644 if (dops[i+1].is_load || dops[i+1].is_store) {
4647 agr=AGEN1+((i+1)&1);
4648 ra=get_reg(i_regs->regmap,agr);
4650 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4651 int offset=cinfo[i+1].imm;
4652 int c=(regs[i+1].wasconst>>rs)&1;
4653 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4654 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4655 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4656 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4657 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4659 emit_movimm(constmap[i+1][rs]+offset,ra);
4660 regs[i+1].loadedconst|=1<<ra;
4663 else if(dops[i+1].rs1==0) {
4664 // Using r0 as a base address
4665 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4666 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4667 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4668 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4670 emit_movimm(offset,ra);
4677 static int get_final_value(int hr, int i, u_int *value)
4679 int reg=regs[i].regmap[hr];
4681 if(regs[i+1].regmap[hr]!=reg) break;
4682 if(!((regs[i+1].isconst>>hr)&1)) break;
4683 if(dops[i+1].bt) break;
4687 if (dops[i].is_jump) {
4688 *value=constmap[i][hr];
4692 if (dops[i+1].is_jump) {
4693 // Load in delay slot, out-of-order execution
4694 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4696 // Precompute load address
4697 *value=constmap[i][hr]+cinfo[i+2].imm;
4701 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4703 // Precompute load address
4704 *value=constmap[i][hr]+cinfo[i+1].imm;
4705 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],cinfo[i+1].imm);
4710 *value=constmap[i][hr];
4711 //printf("c=%lx\n",(long)constmap[i][hr]);
4712 if(i==slen-1) return 1;
4714 return !((unneeded_reg[i+1]>>reg)&1);
4717 // Load registers with known constants
4718 static void load_consts(signed char pre[],signed char regmap[],int i)
4721 // propagate loaded constant flags
4722 if(i==0||dops[i].bt)
4723 regs[i].loadedconst=0;
4725 for (hr = 0; hr < HOST_REGS; hr++) {
4726 if (hr == EXCLUDE_REG || regmap[hr] < 0 || pre[hr] != regmap[hr])
4728 if ((((regs[i-1].isconst & regs[i-1].loadedconst) >> hr) & 1)
4729 && regmap[hr] == regs[i-1].regmap[hr])
4731 regs[i].loadedconst |= 1u << hr;
4736 for(hr=0;hr<HOST_REGS;hr++) {
4737 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
4738 //if(entry[hr]!=regmap[hr]) {
4739 if(!((regs[i].loadedconst>>hr)&1)) {
4740 assert(regmap[hr]<64);
4741 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
4742 u_int value, similar=0;
4743 if(get_final_value(hr,i,&value)) {
4744 // see if some other register has similar value
4745 for(hr2=0;hr2<HOST_REGS;hr2++) {
4746 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4747 if(is_similar_value(value,constmap[i][hr2])) {
4755 if(get_final_value(hr2,i,&value2)) // is this needed?
4756 emit_movimm_from(value2,hr2,value,hr);
4758 emit_movimm(value,hr);
4764 emit_movimm(value,hr);
4767 regs[i].loadedconst|=1<<hr;
4774 static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4778 for(hr=0;hr<HOST_REGS;hr++) {
4779 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
4780 assert(regmap[hr] < 64);
4781 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
4782 int value=constmap[i][hr];
4787 emit_movimm(value,hr);
4794 // Write out all dirty registers (except cycle count)
4796 static void wb_dirtys(const signed char i_regmap[], u_int i_dirty)
4799 for(hr=0;hr<HOST_REGS;hr++) {
4800 if(hr!=EXCLUDE_REG) {
4801 if(i_regmap[hr]>0) {
4802 if(i_regmap[hr]!=CCREG) {
4803 if((i_dirty>>hr)&1) {
4804 assert(i_regmap[hr]<64);
4805 emit_storereg(i_regmap[hr],hr);
4814 // Write out dirty registers that we need to reload (pair with load_needed_regs)
4815 // This writes the registers not written by store_regs_bt
4816 static void wb_needed_dirtys(const signed char i_regmap[], u_int i_dirty, int addr)
4819 int t=(addr-start)>>2;
4820 for(hr=0;hr<HOST_REGS;hr++) {
4821 if(hr!=EXCLUDE_REG) {
4822 if(i_regmap[hr]>0) {
4823 if(i_regmap[hr]!=CCREG) {
4824 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4825 if((i_dirty>>hr)&1) {
4826 assert(i_regmap[hr]<64);
4827 emit_storereg(i_regmap[hr],hr);
4836 // Load all registers (except cycle count)
4837 #ifndef load_all_regs
4838 static void load_all_regs(const signed char i_regmap[])
4841 for(hr=0;hr<HOST_REGS;hr++) {
4842 if(hr!=EXCLUDE_REG) {
4843 if(i_regmap[hr]==0) {
4847 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4849 emit_loadreg(i_regmap[hr],hr);
4856 // Load all current registers also needed by next instruction
4857 static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4859 signed char regmap_sel[HOST_REGS];
4861 for (hr = 0; hr < HOST_REGS; hr++) {
4862 regmap_sel[hr] = -1;
4863 if (hr != EXCLUDE_REG)
4864 if (next_regmap[hr] == i_regmap[hr] || get_reg(next_regmap, i_regmap[hr]) >= 0)
4865 regmap_sel[hr] = i_regmap[hr];
4867 load_all_regs(regmap_sel);
4870 // Load all regs, storing cycle count if necessary
4871 static void load_regs_entry(int t)
4873 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4874 else if(cinfo[t].ccadj) emit_addimm(HOST_CCREG,-cinfo[t].ccadj,HOST_CCREG);
4875 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4876 emit_storereg(CCREG,HOST_CCREG);
4878 load_all_regs(regs[t].regmap_entry);
4881 // Store dirty registers prior to branch
4882 static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4884 if(internal_branch(addr))
4886 int t=(addr-start)>>2;
4888 for(hr=0;hr<HOST_REGS;hr++) {
4889 if(hr!=EXCLUDE_REG) {
4890 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4891 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4892 if((i_dirty>>hr)&1) {
4893 assert(i_regmap[hr]<64);
4894 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4895 emit_storereg(i_regmap[hr],hr);
4904 // Branch out of this block, write out all dirty regs
4905 wb_dirtys(i_regmap,i_dirty);
4909 // Load all needed registers for branch target
4910 static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4912 //if(addr>=start && addr<(start+slen*4))
4913 if(internal_branch(addr))
4915 int t=(addr-start)>>2;
4917 // Store the cycle count before loading something else
4918 if(i_regmap[HOST_CCREG]!=CCREG) {
4919 assert(i_regmap[HOST_CCREG]==-1);
4921 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4922 emit_storereg(CCREG,HOST_CCREG);
4925 for(hr=0;hr<HOST_REGS;hr++) {
4926 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4927 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4928 if(regs[t].regmap_entry[hr]==0) {
4931 else if(regs[t].regmap_entry[hr]!=CCREG)
4933 emit_loadreg(regs[t].regmap_entry[hr],hr);
4941 static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4943 if(addr>=start && addr<start+slen*4-4)
4945 int t=(addr-start)>>2;
4947 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4948 for(hr=0;hr<HOST_REGS;hr++)
4952 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4954 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4961 if(i_regmap[hr]<TEMPREG)
4963 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4966 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4972 else // Same register but is it 32-bit or dirty?
4975 if(!((regs[t].dirty>>hr)&1))
4979 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4981 //printf("%x: dirty no match\n",addr);
4989 // Delay slots are not valid branch targets
4990 //if(t>0&&(dops[t-1].is_jump) return 0;
4991 // Delay slots require additional processing, so do not match
4992 if(dops[t].is_ds) return 0;
4997 for(hr=0;hr<HOST_REGS;hr++)
5003 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
5018 static void drc_dbg_emit_do_cmp(int i, int ccadj_)
5020 extern void do_insn_cmp();
5022 u_int hr, reglist = get_host_reglist(regs[i].regmap);
5023 reglist |= get_host_reglist(regs[i].regmap_entry);
5024 reglist &= DRC_DBG_REGMASK;
5026 assem_debug("//do_insn_cmp %08x\n", start+i*4);
5028 // write out changed consts to match the interpreter
5029 if (i > 0 && !dops[i].bt) {
5030 for (hr = 0; hr < HOST_REGS; hr++) {
5031 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
5032 if (hr == EXCLUDE_REG || reg <= 0)
5034 if (!((regs[i-1].isconst >> hr) & 1))
5036 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
5038 emit_movimm(constmap[i-1][hr],0);
5039 emit_storereg(reg, 0);
5042 if (dops[i].opcode == 0x0f) { // LUI
5043 emit_movimm(cinfo[i].imm << 16, 0);
5044 emit_storereg(dops[i].rt1, 0);
5046 emit_movimm(start+i*4,0);
5047 emit_writeword(0,&psxRegs.pc);
5048 int cc = get_reg(regs[i].regmap_entry, CCREG);
5050 emit_loadreg(CCREG, cc = 0);
5051 emit_addimm(cc, ccadj_, 0);
5052 emit_writeword(0, &psxRegs.cycle);
5053 emit_far_call(do_insn_cmp);
5054 //emit_readword(&cycle,0);
5055 //emit_addimm(0,2,0);
5056 //emit_writeword(0,&cycle);
5058 restore_regs(reglist);
5059 assem_debug("\\\\do_insn_cmp\n");
5061 static void drc_dbg_emit_wb_dirtys(int i, const struct regstat *i_regs)
5063 // write-out non-consts, consts are likely different because of get_final_value()
5064 if (i_regs->dirty & ~i_regs->loadedconst) {
5065 assem_debug("/ drc_dbg_wb\n");
5066 wb_dirtys(i_regs->regmap, i_regs->dirty & ~i_regs->loadedconst);
5067 assem_debug("\\ drc_dbg_wb\n");
5071 #define drc_dbg_emit_do_cmp(x,y)
5072 #define drc_dbg_emit_wb_dirtys(x,y)
5075 // Used when a branch jumps into the delay slot of another branch
5076 static void ds_assemble_entry(int i)
5078 int t = (cinfo[i].ba - start) >> 2;
5079 int ccadj_ = -CLOCK_ADJUST(1);
5081 instr_addr[t] = out;
5082 assem_debug("Assemble delay slot at %x\n",cinfo[i].ba);
5083 assem_debug("<->\n");
5084 drc_dbg_emit_do_cmp(t, ccadj_);
5085 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&®s[t].regmap[HOST_CCREG]!=CCREG)
5086 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
5087 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
5088 address_generation(t,®s[t],regs[t].regmap_entry);
5089 if (ram_offset && (dops[t].is_load || dops[t].is_store))
5090 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
5091 if (dops[t].is_store)
5092 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
5094 switch (dops[t].itype) {
5102 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5105 assemble(t, ®s[t], ccadj_);
5107 store_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5108 load_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5109 if(internal_branch(cinfo[i].ba+4))
5110 assem_debug("branch: internal\n");
5112 assem_debug("branch: external\n");
5113 assert(internal_branch(cinfo[i].ba+4));
5114 add_to_linker(out,cinfo[i].ba+4,internal_branch(cinfo[i].ba+4));
5118 // Load 2 immediates optimizing for small code size
5119 static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
5121 emit_movimm(imm1,rt1);
5122 emit_movimm_from(imm1,rt1,imm2,rt2);
5125 static void do_cc(int i, const signed char i_regmap[], int *adj,
5126 int addr, int taken, int invert)
5128 int count, count_plus2;
5132 if(dops[i].itype==RJUMP)
5136 //if(cinfo[i].ba>=start && cinfo[i].ba<(start+slen*4))
5137 if(internal_branch(cinfo[i].ba))
5139 t=(cinfo[i].ba-start)>>2;
5140 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5141 else *adj=cinfo[t].ccadj;
5147 count = cinfo[i].ccadj;
5148 count_plus2 = count + CLOCK_ADJUST(2);
5149 if(taken==TAKEN && i==(cinfo[i].ba-start)>>2 && source[i+1]==0) {
5151 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5153 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5154 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5158 else if(*adj==0||invert) {
5159 int cycles = count_plus2;
5164 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5165 cycles=*adj+count+2-*adj;
5168 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5174 emit_cmpimm(HOST_CCREG, -count_plus2);
5178 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5181 static void do_ccstub(int n)
5184 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5185 set_jump_target(stubs[n].addr, out);
5188 if (stubs[n].d != TAKEN) {
5189 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5192 if(internal_branch(cinfo[i].ba))
5193 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5197 // Save PC as return address
5198 emit_movimm(stubs[n].c, (r_pc = 0));
5202 // Return address depends on which way the branch goes
5203 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5205 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5206 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5212 else if(dops[i].rs2==0)
5217 #ifdef DESTRUCTIVE_WRITEBACK
5219 if((branch_regs[i].dirty>>s1l)&&1)
5220 emit_loadreg(dops[i].rs1,s1l);
5223 if((branch_regs[i].dirty>>s1l)&1)
5224 emit_loadreg(dops[i].rs2,s1l);
5227 if((branch_regs[i].dirty>>s2l)&1)
5228 emit_loadreg(dops[i].rs2,s2l);
5231 int addr=-1,alt=-1,ntaddr=-1;
5234 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5235 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5236 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5244 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5245 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5246 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5252 if ((dops[i].opcode & 0x3e) == 6) // BLEZ/BGTZ needs another register
5256 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5257 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5258 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5264 assert(hr<HOST_REGS);
5266 if (dops[i].opcode == 4) // BEQ
5268 #ifdef HAVE_CMOV_IMM
5269 if(s2l>=0) emit_cmp(s1l,s2l);
5270 else emit_test(s1l,s1l);
5271 emit_cmov2imm_e_ne_compact(cinfo[i].ba,start+i*4+8,addr);
5273 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,alt);
5274 if(s2l>=0) emit_cmp(s1l,s2l);
5275 else emit_test(s1l,s1l);
5276 emit_cmovne_reg(alt,addr);
5279 else if (dops[i].opcode == 5) // BNE
5281 #ifdef HAVE_CMOV_IMM
5282 if(s2l>=0) emit_cmp(s1l,s2l);
5283 else emit_test(s1l,s1l);
5284 emit_cmov2imm_e_ne_compact(start+i*4+8,cinfo[i].ba,addr);
5286 emit_mov2imm_compact(start+i*4+8,addr,cinfo[i].ba,alt);
5287 if(s2l>=0) emit_cmp(s1l,s2l);
5288 else emit_test(s1l,s1l);
5289 emit_cmovne_reg(alt,addr);
5292 else if (dops[i].opcode == 6) // BLEZ
5294 //emit_movimm(cinfo[i].ba,alt);
5295 //emit_movimm(start+i*4+8,addr);
5296 emit_mov2imm_compact(cinfo[i].ba,alt,start+i*4+8,addr);
5298 emit_cmovl_reg(alt,addr);
5300 else if (dops[i].opcode == 7) // BGTZ
5302 //emit_movimm(cinfo[i].ba,addr);
5303 //emit_movimm(start+i*4+8,ntaddr);
5304 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,ntaddr);
5306 emit_cmovl_reg(ntaddr,addr);
5308 else if (dops[i].itype == SJUMP) // BLTZ/BGEZ
5310 //emit_movimm(cinfo[i].ba,alt);
5311 //emit_movimm(start+i*4+8,addr);
5313 emit_mov2imm_compact(cinfo[i].ba,
5314 (dops[i].opcode2 & 1) ? addr : alt, start + i*4 + 8,
5315 (dops[i].opcode2 & 1) ? alt : addr);
5317 emit_cmovs_reg(alt,addr);
5320 emit_movimm((dops[i].opcode2 & 1) ? cinfo[i].ba : start + i*4 + 8, addr);
5325 if(dops[i].itype==RJUMP)
5327 r_pc = get_reg(branch_regs[i].regmap, dops[i].rs1);
5328 if (ds_writes_rjump_rs(i)) {
5329 r_pc = get_reg(branch_regs[i].regmap, RTEMP);
5332 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5334 emit_writeword(r_pc, &psxRegs.pc);
5335 // Update cycle count
5336 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5337 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5338 emit_far_call(cc_interrupt);
5339 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5340 if(stubs[n].d==TAKEN) {
5341 if(internal_branch(cinfo[i].ba))
5342 load_needed_regs(branch_regs[i].regmap,regs[(cinfo[i].ba-start)>>2].regmap_entry);
5343 else if(dops[i].itype==RJUMP) {
5344 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5345 emit_readword(&psxRegs.pc,get_reg(branch_regs[i].regmap,RTEMP));
5347 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5349 }else if(stubs[n].d==NOTTAKEN) {
5350 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5351 else load_all_regs(branch_regs[i].regmap);
5353 load_all_regs(branch_regs[i].regmap);
5355 if (stubs[n].retaddr)
5356 emit_jmp(stubs[n].retaddr);
5358 do_jump_vaddr(stubs[n].e);
5361 static void add_to_linker(void *addr, u_int target, int is_internal)
5363 assert(linkcount < ARRAY_SIZE(link_addr));
5364 link_addr[linkcount].addr = addr;
5365 link_addr[linkcount].target = target;
5366 link_addr[linkcount].internal = is_internal;
5370 static void ujump_assemble_write_ra(int i)
5373 unsigned int return_address;
5374 rt=get_reg(branch_regs[i].regmap,31);
5375 //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]);
5377 return_address=start+i*4+8;
5380 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5381 int temp=-1; // note: must be ds-safe
5385 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5386 else emit_movimm(return_address,rt);
5394 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5397 if (!((regs[i].loadedconst >> rt) & 1))
5398 emit_movimm(return_address, rt); // PC into link register
5400 emit_prefetch(hash_table_get(return_address));
5406 static void ujump_assemble(int i, const struct regstat *i_regs)
5408 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5409 address_generation(i+1,i_regs,regs[i].regmap_entry);
5411 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5412 if(dops[i].rt1==31&&temp>=0)
5414 signed char *i_regmap=i_regs->regmap;
5415 int return_address=start+i*4+8;
5416 if(get_reg(branch_regs[i].regmap,31)>0)
5417 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5420 if (dops[i].rt1 == 31)
5421 ujump_assemble_write_ra(i); // writeback ra for DS
5422 ds_assemble(i+1,i_regs);
5423 uint64_t bc_unneeded=branch_regs[i].u;
5424 bc_unneeded|=1|(1LL<<dops[i].rt1);
5425 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5426 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5428 cc=get_reg(branch_regs[i].regmap,CCREG);
5429 assert(cc==HOST_CCREG);
5430 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5432 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5434 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5435 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5436 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5437 if(internal_branch(cinfo[i].ba))
5438 assem_debug("branch: internal\n");
5440 assem_debug("branch: external\n");
5441 if (internal_branch(cinfo[i].ba) && dops[(cinfo[i].ba-start)>>2].is_ds) {
5442 ds_assemble_entry(i);
5445 add_to_linker(out,cinfo[i].ba,internal_branch(cinfo[i].ba));
5450 static void rjump_assemble_write_ra(int i)
5452 int rt,return_address;
5453 rt=get_reg_w(branch_regs[i].regmap, dops[i].rt1);
5454 //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]);
5456 return_address=start+i*4+8;
5460 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5463 if (!((regs[i].loadedconst >> rt) & 1))
5464 emit_movimm(return_address, rt); // PC into link register
5466 emit_prefetch(hash_table_get(return_address));
5470 static void rjump_assemble(int i, const struct regstat *i_regs)
5474 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5476 if (ds_writes_rjump_rs(i)) {
5477 // Delay slot abuse, make a copy of the branch address register
5478 temp=get_reg(branch_regs[i].regmap,RTEMP);
5480 assert(regs[i].regmap[temp]==RTEMP);
5484 address_generation(i+1,i_regs,regs[i].regmap_entry);
5488 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5489 signed char *i_regmap=i_regs->regmap;
5490 int return_address=start+i*4+8;
5491 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5496 if(dops[i].rs1==31) {
5497 int rh=get_reg(regs[i].regmap,RHASH);
5498 if(rh>=0) do_preload_rhash(rh);
5501 if (dops[i].rt1 != 0)
5502 rjump_assemble_write_ra(i);
5503 ds_assemble(i+1,i_regs);
5504 uint64_t bc_unneeded=branch_regs[i].u;
5505 bc_unneeded|=1|(1LL<<dops[i].rt1);
5506 bc_unneeded&=~(1LL<<dops[i].rs1);
5507 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5508 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5509 cc=get_reg(branch_regs[i].regmap,CCREG);
5510 assert(cc==HOST_CCREG);
5513 int rh=get_reg(branch_regs[i].regmap,RHASH);
5514 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5515 if(dops[i].rs1==31) {
5516 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5517 do_preload_rhtbl(ht);
5521 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5522 #ifdef DESTRUCTIVE_WRITEBACK
5523 if((branch_regs[i].dirty>>rs)&1) {
5524 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5525 emit_loadreg(dops[i].rs1,rs);
5530 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5533 if(dops[i].rs1==31) {
5534 do_miniht_load(ht,rh);
5537 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5538 //if(adj) emit_addimm(cc,2*(cinfo[i].ccadj+2-adj),cc); // ??? - Shouldn't happen
5540 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5541 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5542 if (dops[i+1].itype == RFE)
5543 // special case for RFE
5547 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5549 if(dops[i].rs1==31) {
5550 do_miniht_jump(rs,rh,ht);
5557 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5558 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5562 static void vsync_hack_assemble(int i, int ld_ofs, int cc)
5564 int sp = get_reg(branch_regs[i].regmap, 29);
5565 int ro = get_reg(branch_regs[i].regmap, ROREG);
5566 int cycles = CLOCK_ADJUST(9+5) * 16;
5567 void *t_exit[3], *loop_target, *t_loop_break;
5569 if (sp < 0 || (ram_offset && ro < 0))
5571 assem_debug("; vsync hack\n");
5572 host_tempreg_acquire();
5573 emit_cmpimm(cc, -cycles);
5576 emit_cmpimm(sp, RAM_SIZE);
5580 emit_addimm(sp, ld_ofs, HOST_TEMPREG);
5581 emit_ldr_dualindexed(ro, HOST_TEMPREG, HOST_TEMPREG);
5584 emit_readword_indexed(ld_ofs, sp, HOST_TEMPREG);
5585 emit_cmpimm(HOST_TEMPREG, 17);
5589 assem_debug("1:\n");
5591 emit_addimm(HOST_TEMPREG, -16, HOST_TEMPREG);
5592 emit_addimm(cc, cycles, cc);
5593 emit_cmpimm(HOST_TEMPREG, 17);
5596 emit_cmpimm(cc, -cycles);
5597 emit_jl(loop_target);
5599 assem_debug("2:\n");
5600 set_jump_target(t_loop_break, out);
5601 do_store_word(sp, ld_ofs, HOST_TEMPREG, ro, 1);
5603 for (j = 0; j < ARRAY_SIZE(t_exit); j++)
5604 set_jump_target(t_exit[j], out);
5605 host_tempreg_release();
5608 static void cjump_assemble(int i, const struct regstat *i_regs)
5610 const signed char *i_regmap = i_regs->regmap;
5613 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5614 assem_debug("match=%d\n",match);
5616 int unconditional=0,nop=0;
5618 int internal=internal_branch(cinfo[i].ba);
5619 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5620 if(!match) invert=1;
5621 if (vsync_hack && (vsync_hack >> 16) == i) invert=1;
5622 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5623 if(i>(cinfo[i].ba-start)>>2) invert=1;
5626 invert=1; // because of near cond. branches
5630 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5631 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5634 s1l=get_reg(i_regmap,dops[i].rs1);
5635 s2l=get_reg(i_regmap,dops[i].rs2);
5637 if(dops[i].rs1==0&&dops[i].rs2==0)
5639 if(dops[i].opcode&1) nop=1;
5640 else unconditional=1;
5641 //assert(dops[i].opcode!=5);
5642 //assert(dops[i].opcode!=7);
5643 //assert(dops[i].opcode!=0x15);
5644 //assert(dops[i].opcode!=0x17);
5646 else if(dops[i].rs1==0)
5651 else if(dops[i].rs2==0)
5657 // Out of order execution (delay slot first)
5659 address_generation(i+1,i_regs,regs[i].regmap_entry);
5660 ds_assemble(i+1,i_regs);
5662 uint64_t bc_unneeded=branch_regs[i].u;
5663 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5665 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5666 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5667 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5668 cc=get_reg(branch_regs[i].regmap,CCREG);
5669 assert(cc==HOST_CCREG);
5671 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5672 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5673 //assem_debug("cycle count (adj)\n");
5675 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5676 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5677 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5678 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5680 assem_debug("branch: internal\n");
5682 assem_debug("branch: external\n");
5683 if (internal && dops[(cinfo[i].ba-start)>>2].is_ds) {
5684 ds_assemble_entry(i);
5687 add_to_linker(out,cinfo[i].ba,internal);
5690 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5691 if(((u_int)out)&7) emit_addnop(0);
5696 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5699 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5702 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5703 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5704 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5706 //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]);
5708 if(dops[i].opcode==4) // BEQ
5710 if(s2l>=0) emit_cmp(s1l,s2l);
5711 else emit_test(s1l,s1l);
5716 add_to_linker(out,cinfo[i].ba,internal);
5720 if(dops[i].opcode==5) // BNE
5722 if(s2l>=0) emit_cmp(s1l,s2l);
5723 else emit_test(s1l,s1l);
5728 add_to_linker(out,cinfo[i].ba,internal);
5732 if(dops[i].opcode==6) // BLEZ
5739 add_to_linker(out,cinfo[i].ba,internal);
5743 if(dops[i].opcode==7) // BGTZ
5750 add_to_linker(out,cinfo[i].ba,internal);
5755 if(taken) set_jump_target(taken, out);
5756 if (vsync_hack && (vsync_hack >> 16) == i)
5757 vsync_hack_assemble(i, vsync_hack & 0xffff, cc);
5758 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5759 if (match && (!internal || !dops[(cinfo[i].ba-start)>>2].is_ds)) {
5761 emit_addimm(cc,-adj,cc);
5762 add_to_linker(out,cinfo[i].ba,internal);
5765 add_to_linker(out,cinfo[i].ba,internal*2);
5771 if(adj) emit_addimm(cc,-adj,cc);
5772 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5773 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5775 assem_debug("branch: internal\n");
5777 assem_debug("branch: external\n");
5778 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5779 ds_assemble_entry(i);
5782 add_to_linker(out,cinfo[i].ba,internal);
5786 set_jump_target(nottaken, out);
5789 if(nottaken1) set_jump_target(nottaken1, out);
5791 if(!invert) emit_addimm(cc,adj,cc);
5793 } // (!unconditional)
5797 // In-order execution (branch first)
5798 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5799 if(!unconditional&&!nop) {
5800 //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]);
5802 if((dops[i].opcode&0x2f)==4) // BEQ
5804 if(s2l>=0) emit_cmp(s1l,s2l);
5805 else emit_test(s1l,s1l);
5809 if((dops[i].opcode&0x2f)==5) // BNE
5811 if(s2l>=0) emit_cmp(s1l,s2l);
5812 else emit_test(s1l,s1l);
5816 if((dops[i].opcode&0x2f)==6) // BLEZ
5822 if((dops[i].opcode&0x2f)==7) // BGTZ
5828 } // if(!unconditional)
5830 uint64_t ds_unneeded=branch_regs[i].u;
5831 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5835 if(taken) set_jump_target(taken, out);
5836 assem_debug("1:\n");
5837 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5839 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5840 address_generation(i+1,&branch_regs[i],0);
5842 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5843 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5844 ds_assemble(i+1,&branch_regs[i]);
5845 drc_dbg_emit_wb_dirtys(i+1, &branch_regs[i]);
5846 cc=get_reg(branch_regs[i].regmap,CCREG);
5848 emit_loadreg(CCREG,cc=HOST_CCREG);
5849 // CHECK: Is the following instruction (fall thru) allocated ok?
5851 assert(cc==HOST_CCREG);
5852 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5853 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5854 assem_debug("cycle count (adj)\n");
5855 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5856 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5858 assem_debug("branch: internal\n");
5860 assem_debug("branch: external\n");
5861 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5862 ds_assemble_entry(i);
5865 add_to_linker(out,cinfo[i].ba,internal);
5870 if(!unconditional) {
5871 if(nottaken1) set_jump_target(nottaken1, out);
5872 set_jump_target(nottaken, out);
5873 assem_debug("2:\n");
5874 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5876 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5877 address_generation(i+1,&branch_regs[i],0);
5879 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5880 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5881 ds_assemble(i+1,&branch_regs[i]);
5882 cc=get_reg(branch_regs[i].regmap,CCREG);
5884 // Cycle count isn't in a register, temporarily load it then write it out
5885 emit_loadreg(CCREG,HOST_CCREG);
5886 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5889 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5890 emit_storereg(CCREG,HOST_CCREG);
5893 cc=get_reg(i_regmap,CCREG);
5894 assert(cc==HOST_CCREG);
5895 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5898 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5904 static void sjump_assemble(int i, const struct regstat *i_regs)
5906 const signed char *i_regmap = i_regs->regmap;
5909 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5910 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5912 int unconditional=0,nevertaken=0;
5914 int internal=internal_branch(cinfo[i].ba);
5915 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5916 if(!match) invert=1;
5917 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5918 if(i>(cinfo[i].ba-start)>>2) invert=1;
5921 invert=1; // because of near cond. branches
5924 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5925 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5928 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5931 s1l=get_reg(i_regmap,dops[i].rs1);
5935 if(dops[i].opcode2&1) unconditional=1;
5937 // These are never taken (r0 is never less than zero)
5938 //assert(dops[i].opcode2!=0);
5939 //assert(dops[i].opcode2!=2);
5940 //assert(dops[i].opcode2!=0x10);
5941 //assert(dops[i].opcode2!=0x12);
5945 // Out of order execution (delay slot first)
5947 address_generation(i+1,i_regs,regs[i].regmap_entry);
5948 ds_assemble(i+1,i_regs);
5950 uint64_t bc_unneeded=branch_regs[i].u;
5951 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5953 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5954 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5955 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5956 if(dops[i].rt1==31) {
5957 int rt,return_address;
5958 rt=get_reg(branch_regs[i].regmap,31);
5959 //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]);
5961 // Save the PC even if the branch is not taken
5962 return_address=start+i*4+8;
5963 emit_movimm(return_address,rt); // PC into link register
5965 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5969 cc=get_reg(branch_regs[i].regmap,CCREG);
5970 assert(cc==HOST_CCREG);
5972 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5973 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5974 assem_debug("cycle count (adj)\n");
5976 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5977 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5978 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5979 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5981 assem_debug("branch: internal\n");
5983 assem_debug("branch: external\n");
5984 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5985 ds_assemble_entry(i);
5988 add_to_linker(out,cinfo[i].ba,internal);
5991 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5992 if(((u_int)out)&7) emit_addnop(0);
5996 else if(nevertaken) {
5997 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
6000 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6003 void *nottaken = NULL;
6004 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
6005 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
6008 if ((dops[i].opcode2 & 1) == 0) // BLTZ/BLTZAL
6015 add_to_linker(out,cinfo[i].ba,internal);
6026 add_to_linker(out,cinfo[i].ba,internal);
6033 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
6034 if (match && (!internal || !dops[(cinfo[i].ba - start) >> 2].is_ds)) {
6036 emit_addimm(cc,-adj,cc);
6037 add_to_linker(out,cinfo[i].ba,internal);
6040 add_to_linker(out,cinfo[i].ba,internal*2);
6046 if(adj) emit_addimm(cc,-adj,cc);
6047 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6048 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6050 assem_debug("branch: internal\n");
6052 assem_debug("branch: external\n");
6053 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
6054 ds_assemble_entry(i);
6057 add_to_linker(out,cinfo[i].ba,internal);
6061 set_jump_target(nottaken, out);
6065 if(!invert) emit_addimm(cc,adj,cc);
6067 } // (!unconditional)
6071 // In-order execution (branch first)
6073 void *nottaken = NULL;
6074 if (!unconditional && !nevertaken) {
6076 emit_test(s1l, s1l);
6078 if (dops[i].rt1 == 31) {
6079 int rt, return_address;
6080 rt = get_reg(branch_regs[i].regmap,31);
6082 // Save the PC even if the branch is not taken
6083 return_address = start + i*4+8;
6084 emit_movimm(return_address, rt); // PC into link register
6086 emit_prefetch(hash_table_get(return_address));
6090 if (!unconditional && !nevertaken) {
6092 if (!(dops[i].opcode2 & 1)) // BLTZ/BLTZAL
6098 uint64_t ds_unneeded=branch_regs[i].u;
6099 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6103 //assem_debug("1:\n");
6104 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6106 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6107 address_generation(i+1,&branch_regs[i],0);
6109 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6110 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6111 ds_assemble(i+1,&branch_regs[i]);
6112 cc=get_reg(branch_regs[i].regmap,CCREG);
6114 emit_loadreg(CCREG,cc=HOST_CCREG);
6115 // CHECK: Is the following instruction (fall thru) allocated ok?
6117 assert(cc==HOST_CCREG);
6118 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6119 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
6120 assem_debug("cycle count (adj)\n");
6121 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
6122 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6124 assem_debug("branch: internal\n");
6126 assem_debug("branch: external\n");
6127 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
6128 ds_assemble_entry(i);
6131 add_to_linker(out,cinfo[i].ba,internal);
6136 if(!unconditional) {
6139 set_jump_target(nottaken, out);
6141 assem_debug("1:\n");
6142 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6143 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6144 address_generation(i+1,&branch_regs[i],0);
6146 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6147 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6148 ds_assemble(i+1,&branch_regs[i]);
6149 cc=get_reg(branch_regs[i].regmap,CCREG);
6151 // Cycle count isn't in a register, temporarily load it then write it out
6152 emit_loadreg(CCREG,HOST_CCREG);
6153 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
6156 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6157 emit_storereg(CCREG,HOST_CCREG);
6160 cc=get_reg(i_regmap,CCREG);
6161 assert(cc==HOST_CCREG);
6162 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
6165 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6171 static void check_regmap(signed char *regmap)
6175 for (i = 0; i < HOST_REGS; i++) {
6178 for (j = i + 1; j < HOST_REGS; j++)
6179 assert(regmap[i] != regmap[j]);
6185 #include <inttypes.h>
6186 static char insn[MAXBLOCK][10];
6188 #define set_mnemonic(i_, n_) \
6189 strcpy(insn[i_], n_)
6191 void print_regmap(const char *name, const signed char *regmap)
6195 fputs(name, stdout);
6196 for (i = 0; i < HOST_REGS; i++) {
6199 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6203 printf(" r%d=%s", i, buf);
6205 fputs("\n", stdout);
6209 void disassemble_inst(int i)
6211 if (dops[i].bt) printf("*"); else printf(" ");
6212 switch(dops[i].itype) {
6214 printf (" %x: %s %8x\n",start+i*4,insn[i],cinfo[i].ba);break;
6216 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):cinfo[i].ba);break;
6218 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;
6220 if (dops[i].opcode2 == 9 && dops[i].rt1 != 31)
6221 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6223 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6226 if(dops[i].opcode==0xf) //LUI
6227 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,cinfo[i].imm&0xffff);
6229 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6233 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6237 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,cinfo[i].imm);
6241 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6244 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6247 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6250 if((dops[i].opcode2&0x1d)==0x10)
6251 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6252 else if((dops[i].opcode2&0x1d)==0x11)
6253 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6255 printf (" %x: %s\n",start+i*4,insn[i]);
6258 if(dops[i].opcode2==0)
6259 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6260 else if(dops[i].opcode2==4)
6261 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6262 else printf (" %x: %s\n",start+i*4,insn[i]);
6265 if(dops[i].opcode2<3)
6266 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6267 else if(dops[i].opcode2>3)
6268 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6269 else printf (" %x: %s\n",start+i*4,insn[i]);
6272 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,cinfo[i].imm);
6275 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6278 //printf (" %s %8x\n",insn[i],source[i]);
6279 printf (" %x: %s\n",start+i*4,insn[i]);
6281 #ifndef REGMAP_PRINT
6284 printf("D: %x WD: %x U: %"PRIx64" hC: %x hWC: %x hLC: %x\n",
6285 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i],
6286 regs[i].isconst, regs[i].wasconst, regs[i].loadedconst);
6287 print_regmap("pre: ", regmap_pre[i]);
6288 print_regmap("entry: ", regs[i].regmap_entry);
6289 print_regmap("map: ", regs[i].regmap);
6290 if (dops[i].is_jump) {
6291 print_regmap("bentry:", branch_regs[i].regmap_entry);
6292 print_regmap("bmap: ", branch_regs[i].regmap);
6296 #define set_mnemonic(i_, n_)
6297 static void disassemble_inst(int i) {}
6300 #define DRC_TEST_VAL 0x74657374
6302 static noinline void new_dynarec_test(void)
6304 int (*testfunc)(void);
6309 // check structure linkage
6310 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6312 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6315 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6316 new_dynarec_test, out);
6318 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6319 out = ndrc->translation_cache;
6320 beginning = start_block();
6321 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6322 emit_movimm(DRC_TEST_VAL + i, 0); // test
6325 end_block(beginning);
6326 testfunc = beginning;
6327 ret[i] = testfunc();
6330 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6331 SysPrintf("test passed.\n");
6333 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6334 out = ndrc->translation_cache;
6337 static int get_cycle_multiplier(void)
6339 return Config.cycle_multiplier_override && Config.cycle_multiplier == CYCLE_MULT_DEFAULT
6340 ? Config.cycle_multiplier_override : Config.cycle_multiplier;
6343 // clear the state completely, instead of just marking
6344 // things invalid like invalidate_all_pages() does
6345 void new_dynarec_clear_full(void)
6348 out = ndrc->translation_cache;
6349 memset(invalid_code,1,sizeof(invalid_code));
6350 memset(shadow,0,sizeof(shadow));
6354 expirep = EXPIRITY_OFFSET;
6359 inv_code_start=inv_code_end=~0;
6362 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6363 blocks_clear(&blocks[n]);
6364 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6368 stat_clear(stat_blocks);
6369 stat_clear(stat_links);
6371 if (ndrc_g.cycle_multiplier_old != Config.cycle_multiplier
6372 || ndrc_g.hacks_old != (ndrc_g.hacks | ndrc_g.hacks_pergame))
6374 SysPrintf("ndrc config: mul=%d, ha=%x, pex=%d\n",
6375 get_cycle_multiplier(), ndrc_g.hacks, Config.PreciseExceptions);
6377 ndrc_g.cycle_multiplier_old = Config.cycle_multiplier;
6378 ndrc_g.hacks_old = ndrc_g.hacks | ndrc_g.hacks_pergame;
6381 static int pgsize(void)
6385 ret = sysconf(_SC_PAGESIZE);
6392 void new_dynarec_init(void)
6394 int align = pgsize() - 1;
6395 SysPrintf("Init new dynarec, ndrc size %x, pgsize %d\n",
6396 (int)sizeof(*ndrc), align + 1);
6398 #ifdef BASE_ADDR_DYNAMIC
6400 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6402 SysPrintf("getVMBlock failed: %x\n", sceBlock);
6403 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6405 SysPrintf("sceKernelGetMemBlockBase: %x\n", ret);
6406 ret = sceKernelOpenVMDomain();
6408 SysPrintf("sceKernelOpenVMDomain: %x\n", ret);
6409 #elif defined(_MSC_VER)
6410 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6411 PAGE_EXECUTE_READWRITE);
6412 #elif defined(HAVE_LIBNX)
6413 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6415 SysPrintf("jitCreate failed: %08x\n", rc);
6416 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6417 jitTransitionToWritable(&g_jit);
6418 ndrc = g_jit.rx_addr;
6419 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6420 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6422 uintptr_t desired_addr = 0;
6423 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6424 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6428 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6430 #ifdef TC_WRITE_OFFSET
6431 // mostly for testing
6432 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6433 ftruncate(fd, sizeof(*ndrc));
6434 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6435 (flags = MAP_SHARED), fd, 0);
6436 assert(mw != MAP_FAILED);
6438 #if defined(NO_WRITE_EXEC) || defined(TC_WRITE_OFFSET)
6439 prot = PROT_READ | PROT_EXEC;
6441 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6442 if (ndrc == MAP_FAILED) {
6443 SysPrintf("mmap() failed: %s\n", strerror(errno));
6446 #ifdef TC_WRITE_OFFSET
6447 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6451 ndrc = (struct ndrc_mem *)((size_t)(ndrc_bss + align) & ~align);
6452 #ifndef NO_WRITE_EXEC
6453 // not all systems allow execute in data segment by default
6454 // size must be 4K aligned for 3DS?
6455 if (mprotect(ndrc, sizeof(*ndrc),
6456 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6457 SysPrintf("mprotect(%p) failed: %s\n", ndrc, strerror(errno));
6459 #ifdef TC_WRITE_OFFSET
6460 #error "misconfiguration detected"
6463 out = ndrc->translation_cache;
6464 new_dynarec_clear_full();
6465 hash_table_ptr = hash_table;
6467 // Copy this into local area so we don't have to put it in every literal pool
6468 invc_ptr=invalid_code;
6472 ram_offset = (uintptr_t)psxM - 0x80000000;
6473 if (ram_offset != 0)
6474 SysPrintf("RAM is not directly mapped\n");
6475 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6476 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6479 void new_dynarec_cleanup(void)
6482 #ifdef BASE_ADDR_DYNAMIC
6484 // sceBlock is managed by retroarch's bootstrap code
6485 //sceKernelFreeMemBlock(sceBlock);
6487 #elif defined(HAVE_LIBNX)
6491 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6492 SysPrintf("munmap() failed\n");
6496 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6497 blocks_clear(&blocks[n]);
6498 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6502 stat_clear(stat_blocks);
6503 stat_clear(stat_links);
6504 new_dynarec_print_stats();
6507 static u_int *get_source_start(u_int addr, u_int *limit)
6509 if (addr < 0x00800000u
6510 || (0x80000000u <= addr && addr < 0x80800000u)
6511 || (0xa0000000u <= addr && addr < 0xa0800000u))
6513 // used for BIOS calls mostly?
6514 *limit = (addr & 0xa0600000) + 0x00200000;
6515 return (u_int *)(psxM + (addr & 0x1fffff));
6518 (0x9fc00000u <= addr && addr < 0x9fc80000u) ||
6519 (0xbfc00000u <= addr && addr < 0xbfc80000u))
6521 // BIOS. The multiplier should be much higher as it's uncached 8bit mem
6522 // XXX: disabled as this introduces differences from the interpreter
6523 // and lightrec multipliers making emu variations act inconsistently
6524 //if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6525 // cycle_multiplier_active = 200;
6527 *limit = (addr & 0xfff00000) | 0x80000;
6528 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6533 static u_int scan_for_ret(u_int addr)
6538 mem = get_source_start(addr, &limit);
6542 if (limit > addr + 0x1000)
6543 limit = addr + 0x1000;
6544 for (; addr < limit; addr += 4, mem++) {
6545 if (*mem == 0x03e00008) // jr $ra
6551 struct savestate_block {
6556 static int addr_cmp(const void *p1_, const void *p2_)
6558 const struct savestate_block *p1 = p1_, *p2 = p2_;
6559 return p1->addr - p2->addr;
6562 int new_dynarec_save_blocks(void *save, int size)
6564 struct savestate_block *sblocks = save;
6565 int maxcount = size / sizeof(sblocks[0]);
6566 struct savestate_block tmp_blocks[1024];
6567 struct block_info *block;
6568 int p, s, d, o, bcnt;
6572 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6574 for (block = blocks[p]; block != NULL; block = block->next) {
6575 if (block->is_dirty)
6577 tmp_blocks[bcnt].addr = block->start;
6578 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6583 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6585 addr = tmp_blocks[0].addr;
6586 for (s = d = 0; s < bcnt; s++) {
6587 if (tmp_blocks[s].addr < addr)
6589 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6590 tmp_blocks[d++] = tmp_blocks[s];
6591 addr = scan_for_ret(tmp_blocks[s].addr);
6594 if (o + d > maxcount)
6596 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6600 return o * sizeof(sblocks[0]);
6603 void new_dynarec_load_blocks(const void *save, int size)
6605 const struct savestate_block *sblocks = save;
6606 int count = size / sizeof(sblocks[0]);
6607 struct block_info *block;
6608 u_int regs_save[32];
6613 // restore clean blocks, if any
6614 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6615 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6616 if (!block->is_dirty)
6618 assert(block->source && block->copy);
6619 if (memcmp(block->source, block->copy, block->len))
6622 // see try_restore_block
6623 block->is_dirty = 0;
6624 mark_invalid_code(block->start, block->len, 0);
6628 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6630 // change GPRs for speculation to at least partially work..
6631 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6632 for (i = 1; i < 32; i++)
6633 psxRegs.GPR.r[i] = 0x80000000;
6635 for (b = 0; b < count; b++) {
6636 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6638 psxRegs.GPR.r[i] = 0x1f800000;
6641 ndrc_get_addr_ht_param(hash_table, sblocks[b].addr, ndrc_cm_compile_offline);
6643 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6645 psxRegs.GPR.r[i] = 0x80000000;
6649 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6652 void new_dynarec_print_stats(void)
6655 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6656 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6657 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6658 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6659 out - ndrc->translation_cache, stat_blocks, stat_links);
6660 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6661 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6662 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6666 static void force_intcall(int i)
6668 memset(&dops[i], 0, sizeof(dops[i]));
6669 dops[i].itype = INTCALL;
6670 dops[i].rs1 = CCREG;
6671 dops[i].is_exception = dops[i].may_except = 1;
6675 static noinline void do_vsync(int i)
6677 // lui a0, x; addiu a0, x; jal puts
6678 u32 addr = (cinfo[i].imm << 16) + (signed short)cinfo[i+1].imm;
6680 int j, t, jals_cnt = 0;
6682 if (!is_ram_addr(addr))
6684 str = (char *)psxM + (addr & 0x1fffff);
6685 if (!str || strncmp(str, "VSync: timeout", 14))
6687 // jal clearPad, jal clearRCnt; j return; nop
6688 for (j = i+2; j < slen; j++) {
6689 if (dops[j].itype == SHIFTIMM || dops[j].itype == IMM16 || dops[j].itype == ALU)
6691 if (dops[j].opcode == 0x03) {
6692 jals_cnt++; continue;
6696 if (j >= slen || jals_cnt != 3 || dops[j++].opcode != 0x02)
6698 for (; j < slen; j++)
6699 if (dops[j].itype != SHIFTIMM && dops[j].itype != IMM16)
6701 if (j >= slen || dops[j].opcode != 0x23) // lw x, condition
6704 if (dops[j].opcode != 0 || dops[j].opcode2 != 0x2A) // slt x, y
6706 if (dops[++j].opcode != 0x05) // bnez x, loop
6708 t = (cinfo[j].ba - start) / 4;
6709 if (t < 0 || t >= slen)
6712 if (dops[t].opcode != 0x23 || dops[t].rs1 != 29 || (u32)cinfo[t].imm >= 1024)
6714 if (dops[t+2].opcode != 0x09 || cinfo[t+2].imm != -1) // addiu x, -1
6716 SysPrintf("vsync @%08x\n", start + t*4);
6717 vsync_hack = (j << 16) | (cinfo[t].imm & 0xffff);
6720 static int apply_hacks(void)
6724 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6726 /* special hack(s) */
6727 for (i = 0; i < slen - 4; i++)
6729 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6730 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6731 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6732 && cinfo[i+3].imm == 0x6e40 && dops[i+3].rs1 == 2)
6734 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6735 dops[i + 3].itype = NOP;
6737 // see also: psxBiosCheckExe()
6738 if (i > 1 && dops[i].opcode == 0x0f && dops[i].rt1 == 4
6739 && dops[i+1].opcode == 0x09 && dops[i+1].rt1 == 4 && dops[i+1].rs1 == 4
6740 && dops[i+2].opcode == 0x03)
6745 if (source[0] == 0x3c05edb8 && source[1] == 0x34a58320)
6747 // lui a1, 0xEDB8; ori a1, 0x8320
6748 SysPrintf("F1 2000 hack @%08x\n", start);
6749 cycle_multiplier_active = 100;
6752 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6753 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6754 && dops[i-7].itype == STORE)
6757 if (dops[i].itype == IMM16)
6759 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6760 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6761 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6763 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6768 #if 0 // alt vsync, not used
6771 if (start <= psxRegs.biosBranchCheck && psxRegs.biosBranchCheck < start + i*4)
6773 i = (psxRegs.biosBranchCheck - start) / 4u + 23;
6774 if (dops[i].is_jump && !dops[i+1].bt)
6777 dops[i+1].is_ds = 0;
6785 static int is_ld_use_hazard(const struct decoded_insn *op_ld,
6786 const struct decoded_insn *op)
6788 if (op_ld->rt1 == 0 || (op_ld->rt1 != op->rs1 && op_ld->rt1 != op->rs2))
6790 if (op_ld->itype == LOADLR && op->itype == LOADLR)
6791 return op_ld->rt1 == op_ld->rs1;
6792 return op->itype != CJUMP && op->itype != SJUMP;
6795 static void disassemble_one(int i, u_int src)
6797 unsigned int type, op, op2, op3;
6798 enum ls_width_type ls_type = LS_32;
6799 memset(&dops[i], 0, sizeof(dops[i]));
6800 memset(&cinfo[i], 0, sizeof(cinfo[i]));
6803 dops[i].opcode = op = src >> 26;
6806 set_mnemonic(i, "???");
6809 case 0x00: set_mnemonic(i, "special");
6813 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6814 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6815 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6816 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6817 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6818 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6819 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6820 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6821 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6822 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6823 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6824 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6825 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6826 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6827 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6828 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6829 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6830 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6831 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6832 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6833 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6834 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6835 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6836 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6837 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6838 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6839 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6840 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6843 case 0x01: set_mnemonic(i, "regimm");
6845 op2 = (src >> 16) & 0x1f;
6848 case 0x10: set_mnemonic(i, "BLTZAL"); break;
6849 case 0x11: set_mnemonic(i, "BGEZAL"); break;
6852 set_mnemonic(i, "BGEZ");
6854 set_mnemonic(i, "BLTZ");
6857 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6858 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6859 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6860 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6861 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6862 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6863 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6864 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6865 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6866 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6867 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6868 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6869 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6870 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6871 case 0x10: set_mnemonic(i, "COP0");
6872 op2 = (src >> 21) & 0x1f;
6877 case 0x01: case 0x02: case 0x06: case 0x08: type = INTCALL; break;
6878 case 0x10: set_mnemonic(i, "RFE"); type=RFE; break;
6879 default: type = OTHER; break;
6887 set_mnemonic(i, "MFC0");
6888 rd = (src >> 11) & 0x1F;
6889 if (!(0x00000417u & (1u << rd)))
6892 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6894 case 0x06: type = INTCALL; break;
6895 default: type = OTHER; break;
6898 case 0x11: set_mnemonic(i, "COP1");
6899 op2 = (src >> 21) & 0x1f;
6901 case 0x12: set_mnemonic(i, "COP2");
6902 op2 = (src >> 21) & 0x1f;
6905 if (gte_handlers[src & 0x3f] != NULL) {
6907 if (gte_regnames[src & 0x3f] != NULL)
6908 strcpy(insn[i], gte_regnames[src & 0x3f]);
6910 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", src & 0x3f);
6917 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6918 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6919 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6920 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6923 case 0x13: set_mnemonic(i, "COP3");
6924 op2 = (src >> 21) & 0x1f;
6926 case 0x20: set_mnemonic(i, "LB"); type=LOAD; ls_type = LS_8; break;
6927 case 0x21: set_mnemonic(i, "LH"); type=LOAD; ls_type = LS_16; break;
6928 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; ls_type = LS_LR; break;
6929 case 0x23: set_mnemonic(i, "LW"); type=LOAD; ls_type = LS_32; break;
6930 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; ls_type = LS_8; break;
6931 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; ls_type = LS_16; break;
6932 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; ls_type = LS_LR; break;
6933 case 0x28: set_mnemonic(i, "SB"); type=STORE; ls_type = LS_8; break;
6934 case 0x29: set_mnemonic(i, "SH"); type=STORE; ls_type = LS_16; break;
6935 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; ls_type = LS_LR; break;
6936 case 0x2B: set_mnemonic(i, "SW"); type=STORE; ls_type = LS_32; break;
6937 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; ls_type = LS_LR; break;
6938 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; ls_type = LS_32; break;
6939 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; ls_type = LS_32; break;
6941 if (Config.HLE && (src & 0x03ffffff) < ARRAY_SIZE(psxHLEt)) {
6942 set_mnemonic(i, "HLECALL");
6949 if (type == INTCALL)
6950 SysPrintf_lim("NI %08x @%08x (%08x)\n", src, start + i*4, start);
6951 dops[i].itype = type;
6952 dops[i].opcode2 = op2;
6953 dops[i].ls_type = ls_type;
6954 /* Get registers/immediates */
6956 gte_rs[i]=gte_rt[i]=0;
6963 dops[i].rs1 = (src >> 21) & 0x1f;
6964 dops[i].rt1 = (src >> 16) & 0x1f;
6965 cinfo[i].imm = (short)src;
6969 dops[i].rs1 = (src >> 21) & 0x1f;
6970 dops[i].rs2 = (src >> 16) & 0x1f;
6971 cinfo[i].imm = (short)src;
6974 // LWL/LWR only load part of the register,
6975 // therefore the target register must be treated as a source too
6976 dops[i].rs1 = (src >> 21) & 0x1f;
6977 dops[i].rs2 = (src >> 16) & 0x1f;
6978 dops[i].rt1 = (src >> 16) & 0x1f;
6979 cinfo[i].imm = (short)src;
6982 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6983 else dops[i].rs1 = (src >> 21) & 0x1f;
6985 dops[i].rt1 = (src >> 16) & 0x1f;
6986 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6987 cinfo[i].imm = (unsigned short)src;
6989 cinfo[i].imm = (short)src;
6993 // The JAL instruction writes to r31.
7000 dops[i].rs1 = (src >> 21) & 0x1f;
7001 // The JALR instruction writes to rd.
7003 dops[i].rt1 = (src >> 11) & 0x1f;
7008 dops[i].rs1 = (src >> 21) & 0x1f;
7009 dops[i].rs2 = (src >> 16) & 0x1f;
7010 if(op&2) { // BGTZ/BLEZ
7015 dops[i].rs1 = (src >> 21) & 0x1f;
7016 dops[i].rs2 = CCREG;
7017 if (op2 == 0x10 || op2 == 0x11) { // BxxAL
7019 // NOTE: If the branch is not taken, r31 is still overwritten
7023 dops[i].rs1=(src>>21)&0x1f; // source
7024 dops[i].rs2=(src>>16)&0x1f; // subtract amount
7025 dops[i].rt1=(src>>11)&0x1f; // destination
7028 dops[i].rs1=(src>>21)&0x1f; // source
7029 dops[i].rs2=(src>>16)&0x1f; // divisor
7034 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
7035 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
7036 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
7037 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
7038 if((op2&0x1d)==0x10) dops[i].rt1=(src>>11)&0x1f; // MFxx
7039 if((op2&0x1d)==0x11) dops[i].rs1=(src>>21)&0x1f; // MTxx
7042 dops[i].rs1=(src>>16)&0x1f; // target of shift
7043 dops[i].rs2=(src>>21)&0x1f; // shift amount
7044 dops[i].rt1=(src>>11)&0x1f; // destination
7047 dops[i].rs1=(src>>16)&0x1f;
7049 dops[i].rt1=(src>>11)&0x1f;
7050 cinfo[i].imm=(src>>6)&0x1f;
7053 if(op2==0) dops[i].rt1=(src>>16)&0x1F; // MFC0
7054 if(op2==4) dops[i].rs1=(src>>16)&0x1F; // MTC0
7055 if(op2==4&&((src>>11)&0x1e)==12) dops[i].rs2=CCREG;
7058 if(op2<3) dops[i].rt1=(src>>16)&0x1F; // MFC2/CFC2
7059 if(op2>3) dops[i].rs1=(src>>16)&0x1F; // MTC2/CTC2
7060 int gr=(src>>11)&0x1F;
7063 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
7064 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
7065 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
7066 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
7070 dops[i].rs1=(src>>21)&0x1F;
7071 cinfo[i].imm=(short)src;
7072 if(op==0x32) gte_rt[i]=1ll<<((src>>16)&0x1F); // LWC2
7073 else gte_rs[i]=1ll<<((src>>16)&0x1F); // SWC2
7076 gte_rs[i]=gte_reg_reads[src&0x3f];
7077 gte_rt[i]=gte_reg_writes[src&0x3f];
7078 gte_rt[i]|=1ll<<63; // every op changes flags
7079 if((src&0x3f)==GTE_MVMVA) {
7080 int v = (src >> 15) & 3;
7081 gte_rs[i]&=~0xe3fll;
7082 if(v==3) gte_rs[i]|=0xe00ll;
7083 else gte_rs[i]|=3ll<<(v*2);
7096 static noinline void pass1a_disassemble(u_int pagelimit)
7101 for (i = 0; !done; i++)
7103 int force_j_to_interpreter = 0;
7104 unsigned int type, op, op2;
7106 disassemble_one(i, source[i]);
7107 dops[i].is_ds = ds_next; ds_next = 0;
7108 type = dops[i].itype;
7109 op = dops[i].opcode;
7110 op2 = dops[i].opcode2;
7112 /* Calculate branch target addresses */
7114 cinfo[i].ba=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7115 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
7116 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
7117 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
7118 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
7119 else if(type==CJUMP||type==SJUMP)
7120 cinfo[i].ba=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7122 /* simplify always (not)taken branches */
7123 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
7124 dops[i].rs1 = dops[i].rs2 = 0;
7126 dops[i].itype = type = UJUMP;
7127 dops[i].rs2 = CCREG;
7130 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
7131 dops[i].itype = type = UJUMP;
7133 dops[i].is_jump = type == RJUMP || type == UJUMP || type == CJUMP || type == SJUMP;
7134 dops[i].is_ujump = type == RJUMP || type == UJUMP;
7135 dops[i].is_load = type == LOAD || type == LOADLR || op == 0x32; // LWC2
7136 dops[i].is_delay_load = (dops[i].is_load || (source[i] & 0xf3d00000) == 0x40000000); // MFC/CFC
7137 dops[i].is_store = type == STORE || type == STORELR || op == 0x3a; // SWC2
7138 dops[i].is_exception = type == SYSCALL || type == HLECALL || type == INTCALL;
7139 dops[i].may_except = dops[i].is_exception || (type == ALU && (op2 == 0x20 || op2 == 0x22)) || op == 8;
7140 ds_next = dops[i].is_jump;
7142 if (((op & 0x37) == 0x21 || op == 0x25) // LH/SH/LHU
7143 && ((cinfo[i].imm & 1) || Config.PreciseExceptions))
7144 dops[i].may_except = 1;
7145 if (((op & 0x37) == 0x23 || (op & 0x37) == 0x32) // LW/SW/LWC2/SWC2
7146 && ((cinfo[i].imm & 3) || Config.PreciseExceptions))
7147 dops[i].may_except = 1;
7149 /* rare messy cases to just pass over to the interpreter */
7150 if (i > 0 && dops[i-1].is_jump) {
7152 // branch in delay slot?
7153 if (dops[i].is_jump) {
7154 // don't handle first branch and call interpreter if it's hit
7155 SysPrintf_lim("branch in DS @%08x (%08x)\n", start + i*4, start);
7156 force_j_to_interpreter = 1;
7158 // load delay detection through a branch
7159 else if (dops[i].is_delay_load && dops[i].rt1 != 0) {
7160 const struct decoded_insn *dop = NULL;
7162 if (cinfo[i-1].ba != -1) {
7163 t = (cinfo[i-1].ba - start) / 4;
7164 if (t < 0 || t > i) {
7166 u_int *mem = get_source_start(cinfo[i-1].ba, &limit);
7168 disassemble_one(MAXBLOCK - 1, mem[0]);
7169 dop = &dops[MAXBLOCK - 1];
7175 if ((dop && is_ld_use_hazard(&dops[i], dop))
7176 || (!dop && Config.PreciseExceptions)) {
7177 // jump target wants DS result - potential load delay effect
7178 SysPrintf_lim("load delay in DS @%08x (%08x)\n", start + i*4, start);
7179 force_j_to_interpreter = 1;
7180 if (0 <= t && t < i)
7181 dops[t + 1].bt = 1; // expected return from interpreter
7183 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&&
7184 !(i>=3&&dops[i-3].is_jump)) {
7185 // v0 overwrite like this is a sign of trouble, bail out
7186 SysPrintf_lim("v0 overwrite @%08x (%08x)\n", start + i*4, start);
7187 force_j_to_interpreter = 1;
7191 else if (i > 0 && dops[i-1].is_delay_load
7192 && is_ld_use_hazard(&dops[i-1], &dops[i])
7193 && (i < 2 || !dops[i-2].is_ujump)) {
7194 SysPrintf_lim("load delay @%08x (%08x)\n", start + i*4, start);
7195 for (j = i - 1; j > 0 && dops[j-1].is_delay_load; j--)
7196 if (dops[j-1].rt1 != dops[i-1].rt1)
7198 force_j_to_interpreter = 1;
7200 if (force_j_to_interpreter) {
7203 i = j; // don't compile the problematic branch/load/etc
7205 if (dops[i].is_exception && i > 0 && dops[i-1].is_jump) {
7206 SysPrintf_lim("exception in DS @%08x (%08x)\n", start + i*4, start);
7212 if ((source[i-2] & 0xffe0f800) == 0x40806000 // MTC0 $12
7213 || (dops[i-2].is_jump && dops[i-2].rt1 == 31)) // call
7216 if (i >= 1 && (source[i-1] & 0xffe0f800) == 0x40806800) // MTC0 $13
7219 /* Is this the end of the block? */
7220 if (i > 0 && dops[i-1].is_ujump) {
7221 // Don't recompile stuff that's already compiled
7222 if (check_addr(start + i*4+4)) {
7226 // Don't get too close to the limit
7227 if (i > MAXBLOCK - 64)
7229 if (dops[i-1].opcode2 == 0x08 || dops[i-1].rs1 == 31) // JR; JALR x, lr
7231 else if (dops[i-1].itype != RJUMP && dops[i-1].rt1 == 0) { // not JAL(R)
7232 int found_bbranch = 0, t = (cinfo[i-1].ba - start) / 4;
7233 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
7234 // scan for a branch back to i+1
7235 for (j = t; j < t + 64; j++) {
7236 int tmpop = source[j] >> 26;
7237 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
7238 int t2 = j + 1 + (int)(signed short)source[j];
7240 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
7251 // jal(r) - continue or perf may suffer for platforms without
7252 // runtime block linking (like in crash3)
7257 if (dops[i].itype == HLECALL)
7259 else if (dops[i].itype == INTCALL) {
7263 else if (dops[i].is_exception)
7266 // Does the block continue due to a branch?
7267 for (j = i-1; j >= 0; j--) {
7268 if (cinfo[j].ba == start+i*4) done=j=0; // Branch into delay slot
7269 if (cinfo[j].ba == start+i*4+4) done=j=0;
7270 if (cinfo[j].ba == start+i*4+8) done=j=0;
7273 //assert(i<MAXBLOCK-1);
7274 if(start+i*4==pagelimit-4) done=1;
7275 assert(start+i*4<pagelimit);
7276 if (i == MAXBLOCK - 2)
7279 if (ni_count > 32 && !stop_after_jal) {
7281 SysPrintf("Disabled speculative precompilation\n");
7283 while (i > 0 && dops[i-1].is_jump)
7286 assert(!dops[i-1].is_jump);
7290 static noinline void pass1b_bt(void)
7293 for (i = 0; i < slen; i++)
7294 if (dops[i].is_jump && start <= cinfo[i].ba && cinfo[i].ba < start+slen*4)
7295 // Internal branch, flag target
7296 dops[(cinfo[i].ba - start) >> 2].bt = 1;
7299 // Basic liveness analysis for MIPS registers
7300 static noinline void pass2b_unneeded_regs(int istart, int iend, int r)
7303 uint64_t u,gte_u,b,gte_b;
7304 uint64_t temp_u,temp_gte_u=0;
7305 uint64_t gte_u_unknown=0;
7306 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7310 gte_u=gte_u_unknown;
7312 //u=unneeded_reg[iend+1];
7314 gte_u=gte_unneeded[iend+1];
7317 for (i=iend;i>=istart;i--)
7319 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7322 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7324 // Branch out of this block, flush all regs
7326 gte_u=gte_u_unknown;
7327 branch_unneeded_reg[i]=u;
7328 // Merge in delay slot
7329 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7330 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7333 gte_u&=~gte_rs[i+1];
7337 if(cinfo[i].ba<=start+i*4) {
7339 if(dops[i].is_ujump)
7341 // Unconditional branch
7345 // Conditional branch (not taken case)
7346 temp_u=unneeded_reg[i+2];
7347 temp_gte_u&=gte_unneeded[i+2];
7349 // Merge in delay slot
7350 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7351 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7353 temp_gte_u|=gte_rt[i+1];
7354 temp_gte_u&=~gte_rs[i+1];
7355 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7356 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7358 temp_gte_u|=gte_rt[i];
7359 temp_gte_u&=~gte_rs[i];
7360 unneeded_reg[i]=temp_u;
7361 gte_unneeded[i]=temp_gte_u;
7362 // Only go three levels deep. This recursion can take an
7363 // excessive amount of time if there are a lot of nested loops.
7365 pass2b_unneeded_regs((cinfo[i].ba-start)>>2, i-1, r+1);
7367 unneeded_reg[(cinfo[i].ba-start)>>2]=1;
7368 gte_unneeded[(cinfo[i].ba-start)>>2]=gte_u_unknown;
7371 if (dops[i].is_ujump)
7373 // Unconditional branch
7374 u=unneeded_reg[(cinfo[i].ba-start)>>2];
7375 gte_u=gte_unneeded[(cinfo[i].ba-start)>>2];
7376 branch_unneeded_reg[i]=u;
7377 // Merge in delay slot
7378 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7379 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7382 gte_u&=~gte_rs[i+1];
7384 // Conditional branch
7385 b=unneeded_reg[(cinfo[i].ba-start)>>2];
7386 gte_b=gte_unneeded[(cinfo[i].ba-start)>>2];
7387 branch_unneeded_reg[i]=b;
7388 // Branch delay slot
7389 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7390 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7393 gte_b&=~gte_rs[i+1];
7397 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7399 branch_unneeded_reg[i]=1;
7406 // Written registers are unneeded
7407 u|=1LL<<dops[i].rt1;
7408 u|=1LL<<dops[i].rt2;
7410 // Accessed registers are needed
7411 u&=~(1LL<<dops[i].rs1);
7412 u&=~(1LL<<dops[i].rs2);
7414 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7415 gte_u|=gte_rs[i]>e_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7416 if (dops[i].may_except || dops[i].itype == RFE)
7418 // SYSCALL instruction, etc or conditional exception
7421 // Source-target dependencies
7422 // R0 is always unneeded
7426 gte_unneeded[i]=gte_u;
7428 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7431 for(r=1;r<=CCREG;r++) {
7432 if((unneeded_reg[i]>>r)&1) {
7433 if(r==HIREG) printf(" HI");
7434 else if(r==LOREG) printf(" LO");
7435 else printf(" r%d",r);
7443 static noinline void pass2a_unneeded(void)
7446 for (i = 0; i < slen; i++)
7448 // remove redundant alignment checks
7449 if (dops[i].may_except && (dops[i].is_load || dops[i].is_store)
7450 && dops[i].rt1 != dops[i].rs1 && !dops[i].is_ds)
7452 int base = dops[i].rs1, lsb = cinfo[i].imm, ls_type = dops[i].ls_type;
7453 int mask = ls_type == LS_32 ? 3 : 1;
7455 for (j = i + 1; j < slen; j++) {
7456 if (dops[j].bt || dops[j].is_jump)
7458 if ((dops[j].is_load || dops[j].is_store) && dops[j].rs1 == base
7459 && dops[j].ls_type == ls_type && (cinfo[j].imm & mask) == lsb)
7460 dops[j].may_except = 0;
7461 if (dops[j].rt1 == base)
7465 // rm redundant stack loads (unoptimized code, assuming no io mem access through sp)
7466 if (i > 0 && dops[i].is_load && dops[i].rs1 == 29 && dops[i].ls_type == LS_32
7467 && dops[i-1].is_store && dops[i-1].rs1 == 29 && dops[i-1].ls_type == LS_32
7468 && dops[i-1].rs2 == dops[i].rt1 && !dops[i-1].is_ds && i < slen - 1
7469 && dops[i+1].rs1 != dops[i].rt1 && dops[i+1].rs2 != dops[i].rt1
7470 && !dops[i].bt && cinfo[i].imm == cinfo[i-1].imm)
7473 memset(&dops[i], 0, sizeof(dops[i]));
7474 dops[i].itype = NOP;
7479 static noinline void pass3_register_alloc(u_int addr)
7481 struct regstat current; // Current register allocations/status
7482 clear_all_regs(current.regmap_entry);
7483 clear_all_regs(current.regmap);
7484 current.wasdirty = current.dirty = 0;
7485 current.u = unneeded_reg[0];
7486 alloc_reg(¤t, 0, CCREG);
7487 dirty_reg(¤t, CCREG);
7488 current.wasconst = 0;
7489 current.isconst = 0;
7490 current.loadedconst = 0;
7491 current.noevict = 0;
7492 //current.waswritten = 0;
7499 // First instruction is delay slot
7510 for(hr=0;hr<HOST_REGS;hr++)
7512 // Is this really necessary?
7513 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7516 //current.waswritten=0;
7519 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7520 regs[i].wasconst=current.isconst;
7521 regs[i].wasdirty=current.dirty;
7525 regs[i].loadedconst=0;
7526 if (!dops[i].is_jump) {
7528 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7535 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7536 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7539 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7543 assert(dops[i].is_ds == ds);
7545 ds=0; // Skip delay slot, already allocated as part of branch
7546 // ...but we need to alloc it in case something jumps here
7548 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7550 current.u=branch_unneeded_reg[i-1];
7552 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7554 struct regstat temp;
7555 memcpy(&temp,¤t,sizeof(current));
7556 temp.wasdirty=temp.dirty;
7557 // TODO: Take into account unconditional branches, as below
7558 delayslot_alloc(&temp,i);
7559 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7560 regs[i].wasdirty=temp.wasdirty;
7561 regs[i].dirty=temp.dirty;
7565 // Create entry (branch target) regmap
7566 for(hr=0;hr<HOST_REGS;hr++)
7568 int r=temp.regmap[hr];
7570 if(r!=regmap_pre[i][hr]) {
7571 regs[i].regmap_entry[hr]=-1;
7576 if((current.u>>r)&1) {
7577 regs[i].regmap_entry[hr]=-1;
7578 regs[i].regmap[hr]=-1;
7579 //Don't clear regs in the delay slot as the branch might need them
7580 //current.regmap[hr]=-1;
7582 regs[i].regmap_entry[hr]=r;
7585 // First instruction expects CCREG to be allocated
7586 if(i==0&&hr==HOST_CCREG)
7587 regs[i].regmap_entry[hr]=CCREG;
7589 regs[i].regmap_entry[hr]=-1;
7593 else { // Not delay slot
7594 current.noevict = 0;
7595 switch(dops[i].itype) {
7597 //current.isconst=0; // DEBUG
7598 //current.wasconst=0; // DEBUG
7599 //regs[i].wasconst=0; // DEBUG
7600 clear_const(¤t,dops[i].rt1);
7601 alloc_cc(¤t,i);
7602 dirty_reg(¤t,CCREG);
7603 if (dops[i].rt1==31) {
7604 alloc_reg(¤t,i,31);
7605 dirty_reg(¤t,31);
7606 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7607 //assert(dops[i+1].rt1!=dops[i].rt1);
7609 alloc_reg(¤t,i,PTEMP);
7613 delayslot_alloc(¤t,i+1);
7614 //current.isconst=0; // DEBUG
7618 //current.isconst=0;
7619 //current.wasconst=0;
7620 //regs[i].wasconst=0;
7621 clear_const(¤t,dops[i].rs1);
7622 clear_const(¤t,dops[i].rt1);
7623 alloc_cc(¤t,i);
7624 dirty_reg(¤t,CCREG);
7625 if (!ds_writes_rjump_rs(i)) {
7626 alloc_reg(¤t,i,dops[i].rs1);
7627 if (dops[i].rt1!=0) {
7628 alloc_reg(¤t,i,dops[i].rt1);
7629 dirty_reg(¤t,dops[i].rt1);
7631 alloc_reg(¤t,i,PTEMP);
7635 if(dops[i].rs1==31) { // JALR
7636 alloc_reg(¤t,i,RHASH);
7637 alloc_reg(¤t,i,RHTBL);
7640 delayslot_alloc(¤t,i+1);
7642 // The delay slot overwrites our source register,
7643 // allocate a temporary register to hold the old value.
7647 delayslot_alloc(¤t,i+1);
7649 alloc_reg(¤t,i,RTEMP);
7651 //current.isconst=0; // DEBUG
7656 //current.isconst=0;
7657 //current.wasconst=0;
7658 //regs[i].wasconst=0;
7659 clear_const(¤t,dops[i].rs1);
7660 clear_const(¤t,dops[i].rs2);
7661 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7663 alloc_cc(¤t,i);
7664 dirty_reg(¤t,CCREG);
7665 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7666 if(dops[i].rs2) alloc_reg(¤t,i,dops[i].rs2);
7667 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7668 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7669 // The delay slot overwrites one of our conditions.
7670 // Allocate the branch condition registers instead.
7674 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7675 if(dops[i].rs2) alloc_reg(¤t,i,dops[i].rs2);
7680 delayslot_alloc(¤t,i+1);
7684 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7686 alloc_cc(¤t,i);
7687 dirty_reg(¤t,CCREG);
7688 alloc_reg(¤t,i,dops[i].rs1);
7689 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7690 // The delay slot overwrites one of our conditions.
7691 // Allocate the branch condition registers instead.
7695 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7700 delayslot_alloc(¤t,i+1);
7704 // Don't alloc the delay slot yet because we might not execute it
7705 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7710 alloc_cc(¤t,i);
7711 dirty_reg(¤t,CCREG);
7712 alloc_reg(¤t,i,dops[i].rs1);
7713 alloc_reg(¤t,i,dops[i].rs2);
7716 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7721 alloc_cc(¤t,i);
7722 dirty_reg(¤t,CCREG);
7723 alloc_reg(¤t,i,dops[i].rs1);
7726 //current.isconst=0;
7729 clear_const(¤t,dops[i].rs1);
7730 clear_const(¤t,dops[i].rt1);
7732 alloc_cc(¤t,i);
7733 dirty_reg(¤t,CCREG);
7734 alloc_reg(¤t,i,dops[i].rs1);
7735 if (dops[i].rt1 == 31) { // BLTZAL/BGEZAL
7736 alloc_reg(¤t,i,31);
7737 dirty_reg(¤t,31);
7740 (dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7741 ||(dops[i].rt1 == 31 && dops[i].rs1 == 31) // overwrites it's own condition
7742 ||(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
7743 // Allocate the branch condition registers instead.
7747 if(dops[i].rs1) alloc_reg(¤t,i,dops[i].rs1);
7752 delayslot_alloc(¤t,i+1);
7756 //current.isconst=0;
7759 imm16_alloc(¤t,i);
7763 load_alloc(¤t,i);
7767 store_alloc(¤t,i);
7770 alu_alloc(¤t,i);
7773 shift_alloc(¤t,i);
7776 multdiv_alloc(¤t,i);
7779 shiftimm_alloc(¤t,i);
7782 mov_alloc(¤t,i);
7785 cop0_alloc(¤t,i);
7788 rfe_alloc(¤t,i);
7791 cop2_alloc(¤t,i);
7794 c2ls_alloc(¤t,i);
7797 c2op_alloc(¤t,i);
7802 syscall_alloc(¤t,i);
7806 // Create entry (branch target) regmap
7807 for(hr=0;hr<HOST_REGS;hr++)
7810 r=current.regmap[hr];
7812 if(r!=regmap_pre[i][hr]) {
7813 // TODO: delay slot (?)
7814 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7815 if(or<0||r>=TEMPREG){
7816 regs[i].regmap_entry[hr]=-1;
7820 // Just move it to a different register
7821 regs[i].regmap_entry[hr]=r;
7822 // If it was dirty before, it's still dirty
7823 if((regs[i].wasdirty>>or)&1) dirty_reg(¤t,r);
7830 regs[i].regmap_entry[hr]=0;
7835 if((current.u>>r)&1) {
7836 regs[i].regmap_entry[hr]=-1;
7837 //regs[i].regmap[hr]=-1;
7838 current.regmap[hr]=-1;
7840 regs[i].regmap_entry[hr]=r;
7844 // Branches expect CCREG to be allocated at the target
7845 if(regmap_pre[i][hr]==CCREG)
7846 regs[i].regmap_entry[hr]=CCREG;
7848 regs[i].regmap_entry[hr]=-1;
7851 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7854 #if 0 // see do_store_smc_check()
7855 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)cinfo[i-1].imm<0x800)
7856 current.waswritten|=1<<dops[i-1].rs1;
7857 current.waswritten&=~(1<<dops[i].rt1);
7858 current.waswritten&=~(1<<dops[i].rt2);
7859 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)cinfo[i].imm>=0x800)
7860 current.waswritten&=~(1<<dops[i].rs1);
7863 /* Branch post-alloc */
7866 current.wasdirty=current.dirty;
7867 switch(dops[i-1].itype) {
7869 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7870 branch_regs[i-1].isconst=0;
7871 branch_regs[i-1].wasconst=0;
7872 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7873 alloc_cc(&branch_regs[i-1],i-1);
7874 dirty_reg(&branch_regs[i-1],CCREG);
7875 if(dops[i-1].rt1==31) { // JAL
7876 alloc_reg(&branch_regs[i-1],i-1,31);
7877 dirty_reg(&branch_regs[i-1],31);
7879 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7880 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7883 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7884 branch_regs[i-1].isconst=0;
7885 branch_regs[i-1].wasconst=0;
7886 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7887 alloc_cc(&branch_regs[i-1],i-1);
7888 dirty_reg(&branch_regs[i-1],CCREG);
7889 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7890 if(dops[i-1].rt1!=0) { // JALR
7891 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7892 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7895 if(dops[i-1].rs1==31) { // JALR
7896 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7897 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7900 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7901 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7904 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7906 alloc_cc(¤t,i-1);
7907 dirty_reg(¤t,CCREG);
7908 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7909 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7910 // The delay slot overwrote one of our conditions
7911 // Delay slot goes after the test (in order)
7912 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7914 delayslot_alloc(¤t,i);
7919 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7920 // Alloc the branch condition registers
7921 if(dops[i-1].rs1) alloc_reg(¤t,i-1,dops[i-1].rs1);
7922 if(dops[i-1].rs2) alloc_reg(¤t,i-1,dops[i-1].rs2);
7924 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7925 branch_regs[i-1].isconst=0;
7926 branch_regs[i-1].wasconst=0;
7927 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7928 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7931 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7933 alloc_cc(¤t,i-1);
7934 dirty_reg(¤t,CCREG);
7935 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7936 // The delay slot overwrote the branch condition
7937 // Delay slot goes after the test (in order)
7938 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7940 delayslot_alloc(¤t,i);
7945 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7946 // Alloc the branch condition register
7947 alloc_reg(¤t,i-1,dops[i-1].rs1);
7949 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7950 branch_regs[i-1].isconst=0;
7951 branch_regs[i-1].wasconst=0;
7952 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7953 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7958 alloc_cc(¤t,i-1);
7959 dirty_reg(¤t,CCREG);
7960 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7961 // The delay slot overwrote the branch condition
7962 // Delay slot goes after the test (in order)
7963 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7965 delayslot_alloc(¤t,i);
7970 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7971 // Alloc the branch condition register
7972 alloc_reg(¤t,i-1,dops[i-1].rs1);
7974 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7975 branch_regs[i-1].isconst=0;
7976 branch_regs[i-1].wasconst=0;
7977 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7978 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7983 if (dops[i-1].is_ujump)
7985 if(dops[i-1].rt1==31) // JAL/JALR
7987 // Subroutine call will return here, don't alloc any registers
7989 clear_all_regs(current.regmap);
7990 alloc_reg(¤t,i,CCREG);
7991 dirty_reg(¤t,CCREG);
7995 // Internal branch will jump here, match registers to caller
7997 clear_all_regs(current.regmap);
7998 alloc_reg(¤t,i,CCREG);
7999 dirty_reg(¤t,CCREG);
8002 if(cinfo[j].ba==start+i*4+4) {
8003 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
8004 current.dirty=branch_regs[j].dirty;
8009 if(cinfo[j].ba==start+i*4+4) {
8010 for(hr=0;hr<HOST_REGS;hr++) {
8011 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
8012 current.regmap[hr]=-1;
8014 current.dirty&=branch_regs[j].dirty;
8023 // Count cycles in between branches
8024 cinfo[i].ccadj = CLOCK_ADJUST(cc);
8025 if (i > 0 && (dops[i-1].is_jump || dops[i].is_exception))
8029 #if !defined(DRC_DBG)
8030 else if(dops[i].itype==C2OP&>e_cycletab[source[i]&0x3f]>2)
8032 // this should really be removed since the real stalls have been implemented,
8033 // but doing so causes sizeable perf regression against the older version
8034 u_int gtec = gte_cycletab[source[i] & 0x3f];
8035 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
8037 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
8041 else if(dops[i].itype==C2LS)
8043 // same as with C2OP
8044 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
8052 if(!dops[i].is_ds) {
8053 regs[i].dirty=current.dirty;
8054 regs[i].isconst=current.isconst;
8055 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
8057 for(hr=0;hr<HOST_REGS;hr++) {
8058 if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
8059 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
8060 regs[i].wasconst&=~(1<<hr);
8064 //regs[i].waswritten=current.waswritten;
8068 static noinline void pass4_cull_unused_regs(void)
8070 u_int last_needed_regs[4] = {0,0,0,0};
8074 for (i=slen-1;i>=0;i--)
8077 __builtin_prefetch(regs[i-2].regmap);
8080 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
8082 // Branch out of this block, don't need anything
8088 // Need whatever matches the target
8090 int t=(cinfo[i].ba-start)>>2;
8091 for(hr=0;hr<HOST_REGS;hr++)
8093 if(regs[i].regmap_entry[hr]>=0) {
8094 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
8098 // Conditional branch may need registers for following instructions
8099 if (!dops[i].is_ujump)
8102 nr |= last_needed_regs[(i+2) & 3];
8103 for(hr=0;hr<HOST_REGS;hr++)
8105 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
8106 //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]);
8110 // Don't need stuff which is overwritten
8111 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8112 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8113 // Merge in delay slot
8114 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
8115 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
8116 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
8117 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
8118 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
8119 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
8120 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
8121 nr |= get_regm(regmap_pre[i], ROREG);
8122 nr |= get_regm(regs[i].regmap_entry, ROREG);
8124 if (dops[i+1].is_store) {
8125 nr |= get_regm(regmap_pre[i], INVCP);
8126 nr |= get_regm(regs[i].regmap_entry, INVCP);
8129 else if (dops[i].is_exception)
8131 // SYSCALL instruction, etc
8137 for(hr=0;hr<HOST_REGS;hr++) {
8138 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8139 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8140 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8141 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8145 // Overwritten registers are not needed
8146 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
8147 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
8148 nr &= ~get_regm(regs[i].regmap, FTEMP);
8149 // Source registers are needed
8150 nr |= get_regm(regmap_pre[i], dops[i].rs1);
8151 nr |= get_regm(regmap_pre[i], dops[i].rs2);
8152 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
8153 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
8154 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
8155 nr |= get_regm(regmap_pre[i], ROREG);
8156 nr |= get_regm(regs[i].regmap_entry, ROREG);
8158 if (dops[i].is_store) {
8159 nr |= get_regm(regmap_pre[i], INVCP);
8160 nr |= get_regm(regs[i].regmap_entry, INVCP);
8163 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
8164 for(hr=0;hr<HOST_REGS;hr++)
8166 // Don't store a register immediately after writing it,
8167 // may prevent dual-issue.
8168 // But do so if this is a branch target, otherwise we
8169 // might have to load the register before the branch.
8170 if((regs[i].wasdirty>>hr)&1) {
8171 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
8172 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
8173 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
8175 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
8176 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8177 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8181 // Cycle count is needed at branches. Assume it is needed at the target too.
8182 if (i == 0 || dops[i].bt || dops[i].may_except || dops[i].itype == CJUMP) {
8183 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8184 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8187 last_needed_regs[i & 3] = nr;
8189 // Deallocate unneeded registers
8190 for(hr=0;hr<HOST_REGS;hr++)
8193 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8196 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
8197 if (dops[i+1].is_load || dops[i+1].is_store)
8199 if (dops[i+1].is_store)
8201 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
8203 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8204 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8205 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
8206 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8207 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
8208 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8209 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8210 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
8212 regs[i].regmap[hr]=-1;
8213 regs[i].isconst&=~(1<<hr);
8214 regs[i].dirty&=~(1<<hr);
8215 regs[i+1].wasdirty&=~(1<<hr);
8216 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
8217 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
8218 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
8219 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8220 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8221 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8222 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8223 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8225 branch_regs[i].regmap[hr]=-1;
8226 branch_regs[i].regmap_entry[hr]=-1;
8227 if (!dops[i].is_ujump)
8230 regmap_pre[i+2][hr]=-1;
8231 regs[i+2].wasconst&=~(1<<hr);
8242 int map1 = -1, map2 = -1, temp=-1;
8243 if (dops[i].is_load || dops[i].is_store)
8245 if (dops[i].is_store)
8247 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8249 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8250 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8251 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8252 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8253 regs[i].regmap[hr] != CCREG)
8255 if(i<slen-1&&!dops[i].is_ds) {
8256 assert(regs[i].regmap[hr]<64);
8257 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8258 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8260 SysPrintf_lim("fail: %x (%d %d!=%d)\n",
8261 start+i*4, hr, regmap_pre[i+1][hr], regs[i].regmap[hr]);
8262 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8264 regmap_pre[i+1][hr]=-1;
8265 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8266 regs[i+1].wasconst&=~(1<<hr);
8268 regs[i].regmap[hr]=-1;
8269 regs[i].isconst&=~(1<<hr);
8270 regs[i].dirty&=~(1<<hr);
8271 regs[i+1].wasdirty&=~(1<<hr);
8280 // If a register is allocated during a loop, try to allocate it for the
8281 // entire loop, if possible. This avoids loading/storing registers
8282 // inside of the loop.
8283 static noinline void pass5a_preallocate1(void)
8286 signed char f_regmap[HOST_REGS];
8287 clear_all_regs(f_regmap);
8288 for(i=0;i<slen-1;i++)
8290 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8292 if(cinfo[i].ba>=start && cinfo[i].ba<(start+i*4))
8293 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8294 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8295 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8296 ||dops[i+1].itype==SHIFT
8297 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8299 int t=(cinfo[i].ba-start)>>2;
8300 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8301 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8302 for(hr=0;hr<HOST_REGS;hr++)
8304 if(regs[i].regmap[hr]>=0) {
8305 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8306 // dealloc old register
8308 for(n=0;n<HOST_REGS;n++)
8310 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8312 // and alloc new one
8313 f_regmap[hr]=regs[i].regmap[hr];
8316 if(branch_regs[i].regmap[hr]>=0) {
8317 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8318 // dealloc old register
8320 for(n=0;n<HOST_REGS;n++)
8322 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8324 // and alloc new one
8325 f_regmap[hr]=branch_regs[i].regmap[hr];
8329 if(count_free_regs(regs[i].regmap)<=cinfo[i+1].min_free_regs)
8330 f_regmap[hr]=branch_regs[i].regmap[hr];
8332 if(count_free_regs(branch_regs[i].regmap)<=cinfo[i+1].min_free_regs)
8333 f_regmap[hr]=branch_regs[i].regmap[hr];
8335 // Avoid dirty->clean transition
8336 #ifdef DESTRUCTIVE_WRITEBACK
8337 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;
8339 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8340 // case above, however it's always a good idea. We can't hoist the
8341 // load if the register was already allocated, so there's no point
8342 // wasting time analyzing most of these cases. It only "succeeds"
8343 // when the mapping was different and the load can be replaced with
8344 // a mov, which is of negligible benefit. So such cases are
8346 if(f_regmap[hr]>0) {
8347 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8351 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8352 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8354 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8355 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8357 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8358 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8359 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8361 while(k>1&®s[k-1].regmap[hr]==-1) {
8362 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8363 //printf("no free regs for store %x\n",start+(k-1)*4);
8366 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8367 //printf("no-match due to different register\n");
8370 if (dops[k-2].is_jump) {
8371 //printf("no-match due to branch\n");
8374 // call/ret fast path assumes no registers allocated
8375 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8380 if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
8381 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8383 regs[k].regmap_entry[hr]=f_regmap[hr];
8384 regs[k].regmap[hr]=f_regmap[hr];
8385 regmap_pre[k+1][hr]=f_regmap[hr];
8386 regs[k].wasdirty&=~(1<<hr);
8387 regs[k].dirty&=~(1<<hr);
8388 regs[k].wasdirty|=(1<<hr)®s[k-1].dirty;
8389 regs[k].dirty|=(1<<hr)®s[k].wasdirty;
8390 regs[k].wasconst&=~(1<<hr);
8391 regs[k].isconst&=~(1<<hr);
8396 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8399 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8400 if(regs[i-1].regmap[hr]==f_regmap[hr]&®map_pre[i][hr]==f_regmap[hr]) {
8401 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8402 regs[i].regmap_entry[hr]=f_regmap[hr];
8403 regs[i].regmap[hr]=f_regmap[hr];
8404 regs[i].wasdirty&=~(1<<hr);
8405 regs[i].dirty&=~(1<<hr);
8406 regs[i].wasdirty|=(1<<hr)®s[i-1].dirty;
8407 regs[i].dirty|=(1<<hr)®s[i-1].dirty;
8408 regs[i].wasconst&=~(1<<hr);
8409 regs[i].isconst&=~(1<<hr);
8410 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8411 branch_regs[i].wasdirty&=~(1<<hr);
8412 branch_regs[i].wasdirty|=(1<<hr)®s[i].dirty;
8413 branch_regs[i].regmap[hr]=f_regmap[hr];
8414 branch_regs[i].dirty&=~(1<<hr);
8415 branch_regs[i].dirty|=(1<<hr)®s[i].dirty;
8416 branch_regs[i].wasconst&=~(1<<hr);
8417 branch_regs[i].isconst&=~(1<<hr);
8418 if (!dops[i].is_ujump) {
8419 regmap_pre[i+2][hr]=f_regmap[hr];
8420 regs[i+2].wasdirty&=~(1<<hr);
8421 regs[i+2].wasdirty|=(1<<hr)®s[i].dirty;
8426 // Alloc register clean at beginning of loop,
8427 // but may dirty it in pass 6
8428 regs[k].regmap_entry[hr]=f_regmap[hr];
8429 regs[k].regmap[hr]=f_regmap[hr];
8430 regs[k].dirty&=~(1<<hr);
8431 regs[k].wasconst&=~(1<<hr);
8432 regs[k].isconst&=~(1<<hr);
8433 if (dops[k].is_jump) {
8434 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8435 branch_regs[k].regmap[hr]=f_regmap[hr];
8436 branch_regs[k].dirty&=~(1<<hr);
8437 branch_regs[k].wasconst&=~(1<<hr);
8438 branch_regs[k].isconst&=~(1<<hr);
8439 if (!dops[k].is_ujump) {
8440 regmap_pre[k+2][hr]=f_regmap[hr];
8441 regs[k+2].wasdirty&=~(1<<hr);
8446 regmap_pre[k+1][hr]=f_regmap[hr];
8447 regs[k+1].wasdirty&=~(1<<hr);
8450 if(regs[j].regmap[hr]==f_regmap[hr])
8451 regs[j].regmap_entry[hr]=f_regmap[hr];
8455 if(regs[j].regmap[hr]>=0)
8457 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8458 //printf("no-match due to different register\n");
8461 if (dops[j].is_ujump)
8463 // Stop on unconditional branch
8466 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8469 if(count_free_regs(regs[j].regmap)<=cinfo[j+1].min_free_regs)
8472 if(count_free_regs(branch_regs[j].regmap)<=cinfo[j+1].min_free_regs)
8475 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8476 //printf("no-match due to different register (branch)\n");
8480 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8481 //printf("No free regs for store %x\n",start+j*4);
8484 assert(f_regmap[hr]<64);
8491 // Non branch or undetermined branch target
8492 for(hr=0;hr<HOST_REGS;hr++)
8494 if(hr!=EXCLUDE_REG) {
8495 if(regs[i].regmap[hr]>=0) {
8496 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8497 // dealloc old register
8499 for(n=0;n<HOST_REGS;n++)
8501 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8503 // and alloc new one
8504 f_regmap[hr]=regs[i].regmap[hr];
8509 // Try to restore cycle count at branch targets
8511 for(j=i;j<slen-1;j++) {
8512 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8513 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8514 //printf("no free regs for store %x\n",start+j*4);
8518 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8520 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8522 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8523 regs[k].regmap[HOST_CCREG]=CCREG;
8524 regmap_pre[k+1][HOST_CCREG]=CCREG;
8525 regs[k+1].wasdirty|=1<<HOST_CCREG;
8526 regs[k].dirty|=1<<HOST_CCREG;
8527 regs[k].wasconst&=~(1<<HOST_CCREG);
8528 regs[k].isconst&=~(1<<HOST_CCREG);
8531 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8533 // Work backwards from the branch target
8534 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8536 //printf("Extend backwards\n");
8539 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8540 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8541 //printf("no free regs for store %x\n",start+(k-1)*4);
8546 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8547 //printf("Extend CC, %x ->\n",start+k*4);
8549 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8550 regs[k].regmap[HOST_CCREG]=CCREG;
8551 regmap_pre[k+1][HOST_CCREG]=CCREG;
8552 regs[k+1].wasdirty|=1<<HOST_CCREG;
8553 regs[k].dirty|=1<<HOST_CCREG;
8554 regs[k].wasconst&=~(1<<HOST_CCREG);
8555 regs[k].isconst&=~(1<<HOST_CCREG);
8560 //printf("Fail Extend CC, %x ->\n",start+k*4);
8564 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=SHIFT&&
8565 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8566 dops[i].itype!=IMM16&&dops[i].itype!=LOAD)
8568 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8574 // This allocates registers (if possible) one instruction prior
8575 // to use, which can avoid a load-use penalty on certain CPUs.
8576 static noinline void pass5b_preallocate2(void)
8578 int i, hr, limit = min(slen - 1, MAXBLOCK - 2);
8579 for (i = 0; i < limit; i++)
8581 if (!i || !dops[i-1].is_jump)
8585 int j, can_steal = 1;
8586 for (j = i; j < i + 2; j++) {
8588 if (cinfo[j].min_free_regs == 0)
8590 for (hr = 0; hr < HOST_REGS; hr++)
8591 if (hr != EXCLUDE_REG && regs[j].regmap[hr] < 0)
8593 if (free_regs <= cinfo[j].min_free_regs) {
8600 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8601 ||(dops[i].itype==COP2&&dops[i].opcode2<3))
8604 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8606 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8608 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8609 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8610 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8611 regs[i].isconst&=~(1<<hr);
8612 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8613 constmap[i][hr]=constmap[i+1][hr];
8614 regs[i+1].wasdirty&=~(1<<hr);
8615 regs[i].dirty&=~(1<<hr);
8620 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8622 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8624 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8625 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8626 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8627 regs[i].isconst&=~(1<<hr);
8628 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8629 constmap[i][hr]=constmap[i+1][hr];
8630 regs[i+1].wasdirty&=~(1<<hr);
8631 regs[i].dirty&=~(1<<hr);
8635 // Preload target address for load instruction (non-constant)
8636 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8637 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8639 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8641 regs[i].regmap[hr]=dops[i+1].rs1;
8642 regmap_pre[i+1][hr]=dops[i+1].rs1;
8643 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8644 regs[i].isconst&=~(1<<hr);
8645 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8646 constmap[i][hr]=constmap[i+1][hr];
8647 regs[i+1].wasdirty&=~(1<<hr);
8648 regs[i].dirty&=~(1<<hr);
8652 // Load source into target register
8653 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8654 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8656 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8658 regs[i].regmap[hr]=dops[i+1].rs1;
8659 regmap_pre[i+1][hr]=dops[i+1].rs1;
8660 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8661 regs[i].isconst&=~(1<<hr);
8662 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8663 constmap[i][hr]=constmap[i+1][hr];
8664 regs[i+1].wasdirty&=~(1<<hr);
8665 regs[i].dirty&=~(1<<hr);
8669 // Address for store instruction (non-constant)
8670 if (dops[i+1].is_store) { // SB/SH/SW/SWC2
8671 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8672 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8673 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8675 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8676 regs[i+1].isconst&=~(1<<hr);
8677 regs[i+1].dirty&=~(1<<hr);
8678 regs[i+2].wasdirty&=~(1<<hr);
8681 #if 0 // what is this for? double allocs $0 in ps1_rom.bin
8682 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8684 regs[i].regmap[hr]=dops[i+1].rs1;
8685 regmap_pre[i+1][hr]=dops[i+1].rs1;
8686 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8687 regs[i].isconst&=~(1<<hr);
8688 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8689 constmap[i][hr]=constmap[i+1][hr];
8690 regs[i+1].wasdirty&=~(1<<hr);
8691 regs[i].dirty&=~(1<<hr);
8696 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) { // LWC2
8697 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8699 hr=get_reg(regs[i+1].regmap,FTEMP);
8701 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8703 regs[i].regmap[hr]=dops[i+1].rs1;
8704 regmap_pre[i+1][hr]=dops[i+1].rs1;
8705 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8706 regs[i].isconst&=~(1<<hr);
8707 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8708 constmap[i][hr]=constmap[i+1][hr];
8709 regs[i+1].wasdirty&=~(1<<hr);
8710 regs[i].dirty&=~(1<<hr);
8712 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8714 // move it to another register
8715 regs[i+1].regmap[hr]=-1;
8716 regmap_pre[i+2][hr]=-1;
8717 regs[i+1].regmap[nr]=FTEMP;
8718 regmap_pre[i+2][nr]=FTEMP;
8719 regs[i].regmap[nr]=dops[i+1].rs1;
8720 regmap_pre[i+1][nr]=dops[i+1].rs1;
8721 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8722 regs[i].isconst&=~(1<<nr);
8723 regs[i+1].isconst&=~(1<<nr);
8724 regs[i].dirty&=~(1<<nr);
8725 regs[i+1].wasdirty&=~(1<<nr);
8726 regs[i+1].dirty&=~(1<<nr);
8727 regs[i+2].wasdirty&=~(1<<nr);
8731 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==C2LS*/) {
8733 if(dops[i+1].itype==LOAD)
8734 hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1);
8735 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) // LWC2
8736 hr=get_reg(regs[i+1].regmap,FTEMP);
8737 if (dops[i+1].is_store) {
8738 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8739 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8741 if(hr>=0&®s[i].regmap[hr]<0) {
8742 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8743 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8744 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8745 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8746 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8747 regs[i].isconst&=~(1<<hr);
8748 regs[i+1].wasdirty&=~(1<<hr);
8749 regs[i].dirty&=~(1<<hr);
8759 // Write back dirty registers as soon as we will no longer modify them,
8760 // so that we don't end up with lots of writes at the branches.
8761 static noinline void pass6_clean_registers(int istart, int iend, int wr)
8763 static u_int wont_dirty[MAXBLOCK];
8764 static u_int will_dirty[MAXBLOCK];
8767 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8768 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8770 will_dirty_i=will_dirty_next=0;
8771 wont_dirty_i=wont_dirty_next=0;
8773 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8774 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8776 for (i=iend;i>=istart;i--)
8778 signed char rregmap_i[RRMAP_SIZE];
8779 u_int hr_candirty = 0;
8780 assert(HOST_REGS < 32);
8781 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8782 __builtin_prefetch(regs[i-1].regmap);
8785 signed char branch_rregmap_i[RRMAP_SIZE];
8786 u_int branch_hr_candirty = 0;
8787 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8788 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
8790 // Branch out of this block, flush all regs
8792 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8793 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8794 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8795 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8796 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8797 will_dirty_i &= branch_hr_candirty;
8798 if (dops[i].is_ujump)
8800 // Unconditional branch
8802 // Merge in delay slot (will dirty)
8803 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8804 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8805 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8806 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8807 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8808 will_dirty_i &= hr_candirty;
8812 // Conditional branch
8813 wont_dirty_i = wont_dirty_next;
8814 // Merge in delay slot (will dirty)
8815 // (the original code had no explanation why these 2 are commented out)
8816 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8817 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8818 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8819 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8820 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8821 will_dirty_i &= hr_candirty;
8823 // Merge in delay slot (wont dirty)
8824 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8825 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8826 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8827 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8828 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8829 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8830 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8831 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8832 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8833 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8834 wont_dirty_i &= ~(1u << 31);
8836 #ifndef DESTRUCTIVE_WRITEBACK
8837 branch_regs[i].dirty&=wont_dirty_i;
8839 branch_regs[i].dirty|=will_dirty_i;
8845 if(cinfo[i].ba<=start+i*4) {
8847 if (dops[i].is_ujump)
8849 // Unconditional branch
8852 // Merge in delay slot (will dirty)
8853 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8854 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8855 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8856 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8857 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8858 temp_will_dirty &= branch_hr_candirty;
8859 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8860 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8861 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8862 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8863 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8864 temp_will_dirty &= hr_candirty;
8866 // Conditional branch (not taken case)
8867 temp_will_dirty=will_dirty_next;
8868 temp_wont_dirty=wont_dirty_next;
8869 // Merge in delay slot (will dirty)
8870 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8871 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8872 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8873 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8874 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8875 temp_will_dirty &= branch_hr_candirty;
8876 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8877 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8878 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8879 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8880 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8881 temp_will_dirty &= hr_candirty;
8883 // Merge in delay slot (wont dirty)
8884 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8885 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8886 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8887 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8888 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8889 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8890 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8891 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8892 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8893 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8894 temp_wont_dirty &= ~(1u << 31);
8895 // Deal with changed mappings
8897 for(r=0;r<HOST_REGS;r++) {
8898 if(r!=EXCLUDE_REG) {
8899 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8900 temp_will_dirty&=~(1<<r);
8901 temp_wont_dirty&=~(1<<r);
8902 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8903 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8904 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8906 temp_will_dirty|=1<<r;
8907 temp_wont_dirty|=1<<r;
8914 will_dirty[i]=temp_will_dirty;
8915 wont_dirty[i]=temp_wont_dirty;
8916 pass6_clean_registers((cinfo[i].ba-start)>>2,i-1,0);
8918 // Limit recursion. It can take an excessive amount
8919 // of time if there are a lot of nested loops.
8920 will_dirty[(cinfo[i].ba-start)>>2]=0;
8921 wont_dirty[(cinfo[i].ba-start)>>2]=-1;
8926 if (dops[i].is_ujump)
8928 // Unconditional branch
8931 //if(cinfo[i].ba>start+i*4) { // Disable recursion (for debugging)
8932 for(r=0;r<HOST_REGS;r++) {
8933 if(r!=EXCLUDE_REG) {
8934 if(branch_regs[i].regmap[r]==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8935 will_dirty_i|=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8936 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8938 if(branch_regs[i].regmap[r]>=0) {
8939 will_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8940 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8945 // Merge in delay slot
8946 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8947 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8948 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8949 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8950 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8951 will_dirty_i &= branch_hr_candirty;
8952 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8953 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8954 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8955 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8956 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8957 will_dirty_i &= hr_candirty;
8959 // Conditional branch
8960 will_dirty_i=will_dirty_next;
8961 wont_dirty_i=wont_dirty_next;
8962 //if(cinfo[i].ba>start+i*4) // Disable recursion (for debugging)
8963 for(r=0;r<HOST_REGS;r++) {
8964 if(r!=EXCLUDE_REG) {
8965 signed char target_reg=branch_regs[i].regmap[r];
8966 if(target_reg==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8967 will_dirty_i&=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8968 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8970 else if(target_reg>=0) {
8971 will_dirty_i&=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8972 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8976 // Merge in delay slot
8977 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8978 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8979 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8980 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8981 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8982 will_dirty_i &= branch_hr_candirty;
8983 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8984 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8985 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8986 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8987 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8988 will_dirty_i &= hr_candirty;
8990 // Merge in delay slot (won't dirty)
8991 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8992 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8993 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8994 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8995 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8996 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8997 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8998 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8999 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
9000 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
9001 wont_dirty_i &= ~(1u << 31);
9003 #ifndef DESTRUCTIVE_WRITEBACK
9004 branch_regs[i].dirty&=wont_dirty_i;
9006 branch_regs[i].dirty|=will_dirty_i;
9011 else if (dops[i].is_exception)
9013 // SYSCALL instruction, etc
9017 will_dirty_next=will_dirty_i;
9018 wont_dirty_next=wont_dirty_i;
9019 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
9020 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
9021 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
9022 will_dirty_i &= hr_candirty;
9023 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
9024 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
9025 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
9026 wont_dirty_i &= ~(1u << 31);
9027 if (i > istart && !dops[i].is_jump) {
9028 // Don't store a register immediately after writing it,
9029 // may prevent dual-issue.
9030 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
9031 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
9034 will_dirty[i]=will_dirty_i;
9035 wont_dirty[i]=wont_dirty_i;
9036 // Mark registers that won't be dirtied as not dirty
9038 regs[i].dirty|=will_dirty_i;
9039 #ifndef DESTRUCTIVE_WRITEBACK
9040 regs[i].dirty&=wont_dirty_i;
9043 if (i < iend-1 && !dops[i].is_ujump) {
9044 for(r=0;r<HOST_REGS;r++) {
9045 if(r!=EXCLUDE_REG) {
9046 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
9047 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
9048 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
9056 for(r=0;r<HOST_REGS;r++) {
9057 if(r!=EXCLUDE_REG) {
9058 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
9059 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
9060 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
9067 // Deal with changed mappings
9068 temp_will_dirty=will_dirty_i;
9069 temp_wont_dirty=wont_dirty_i;
9070 for(r=0;r<HOST_REGS;r++) {
9071 if(r!=EXCLUDE_REG) {
9073 if(regs[i].regmap[r]==regmap_pre[i][r]) {
9075 #ifndef DESTRUCTIVE_WRITEBACK
9076 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
9078 regs[i].wasdirty|=will_dirty_i&(1<<r);
9081 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
9082 // Register moved to a different register
9083 will_dirty_i&=~(1<<r);
9084 wont_dirty_i&=~(1<<r);
9085 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
9086 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
9088 #ifndef DESTRUCTIVE_WRITEBACK
9089 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
9091 regs[i].wasdirty|=will_dirty_i&(1<<r);
9095 will_dirty_i&=~(1<<r);
9096 wont_dirty_i&=~(1<<r);
9097 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
9098 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
9099 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
9102 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
9110 static noinline void pass10_expire_blocks(void)
9112 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
9113 // not sizeof(ndrc->translation_cache) due to vita hack
9114 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
9115 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
9116 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
9119 for (; expirep != end; expirep = ((expirep + step) & step_mask))
9121 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
9122 u_int block_i = expirep / step & (PAGE_COUNT - 1);
9123 u_int phase = (expirep >> (base_shift - 1)) & 1u;
9124 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
9125 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
9126 (long)(out - ndrc->translation_cache), phase);
9130 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
9137 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
9141 static struct block_info *new_block_info(u_int start, u_int len,
9142 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
9144 struct block_info **b_pptr;
9145 struct block_info *block;
9146 u_int page = get_page(start);
9148 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
9150 assert(jump_in_count > 0);
9151 block->source = source;
9153 block->start = start;
9155 block->reg_sv_flags = 0;
9156 block->tc_offs = beginning - ndrc->translation_cache;
9157 //block->tc_len = out - beginning;
9158 block->is_dirty = 0;
9159 block->inv_near_misses = 0;
9160 block->jump_in_cnt = jump_in_count;
9162 // insert sorted by start mirror-unmasked vaddr
9163 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
9164 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
9165 block->next = *b_pptr;
9170 stat_inc(stat_blocks);
9174 static int noinline new_recompile_block(u_int addr)
9176 u_int pagelimit = 0;
9177 u_int state_rflags = 0;
9180 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, log_addr(out));
9183 if (addr != hack_addr) {
9184 SysPrintf_lim("game crash @%08x, ra=%08x\n", addr, psxRegs.GPR.n.ra);
9190 // this is just for speculation
9191 for (i = 1; i < 32; i++) {
9192 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
9193 state_rflags |= 1 << i;
9197 ndrc_g.did_compile++;
9198 if (Config.HLE && start == 0x80001000) // hlecall
9200 void *beginning = start_block();
9202 emit_movimm(start,0);
9203 emit_writeword(0,&psxRegs.pc);
9204 emit_far_jump(new_dyna_leave);
9206 end_block(beginning);
9207 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9208 block->jump_in[0].vaddr = start;
9209 block->jump_in[0].addr = beginning;
9212 else if (f1_hack && hack_addr == 0) {
9213 void *beginning = start_block();
9214 emit_movimm(start, 0);
9215 emit_writeword(0, &hack_addr);
9216 emit_readword(&psxRegs.GPR.n.sp, 0);
9217 emit_readptr(&mem_rtab, 1);
9218 emit_shrimm(0, 12, 2);
9219 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
9220 emit_addimm(0, 0x18, 0);
9221 emit_adds_ptr(1, 1, 1);
9222 emit_ldr_dualindexed(1, 0, 0);
9223 emit_readptr(&hash_table_ptr, 1);
9224 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
9225 emit_far_call(ndrc_get_addr_ht);
9226 emit_jmpreg(0); // jr k0
9228 end_block(beginning);
9230 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9231 block->jump_in[0].vaddr = start;
9232 block->jump_in[0].addr = beginning;
9233 SysPrintf("F1 hack to %08x\n", start);
9237 cycle_multiplier_active = get_cycle_multiplier();
9239 source = get_source_start(start, &pagelimit);
9240 if (source == NULL) {
9241 if (addr != hack_addr) {
9242 SysPrintf_lim("Compile at bogus memory address: %08x, ra=%x\n",
9243 addr, psxRegs.GPR.n.ra);
9250 /* Pass 1: disassemble */
9251 /* Pass 2: register dependencies, branch targets */
9252 /* Pass 3: register allocation */
9253 /* Pass 4: branch dependencies */
9254 /* Pass 5: pre-alloc */
9255 /* Pass 6: optimize clean/dirty state */
9256 /* Pass 7: flag 32-bit registers */
9257 /* Pass 8: assembly */
9258 /* Pass 9: linker */
9259 /* Pass 10: garbage collection / free memory */
9261 /* Pass 1 disassembly */
9263 pass1a_disassemble(pagelimit);
9266 int clear_hack_addr = apply_hacks();
9268 /* Pass 2 - unneeded, register dependencies */
9271 pass2b_unneeded_regs(0, slen-1, 0);
9273 /* Pass 3 - Register allocation */
9275 pass3_register_alloc(addr);
9277 /* Pass 4 - Cull unused host registers */
9279 pass4_cull_unused_regs();
9281 /* Pass 5 - Pre-allocate registers */
9283 pass5a_preallocate1();
9284 pass5b_preallocate2();
9286 /* Pass 6 - Optimize clean/dirty state */
9287 pass6_clean_registers(0, slen-1, 1);
9290 for (i=slen-1;i>=0;i--)
9292 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9294 // Conditional branch
9295 if((source[i]>>16)!=0x1000&&i<slen-2) {
9296 // Mark this address as a branch target since it may be called
9297 // upon return from interrupt
9303 /* Pass 8 - Assembly */
9304 linkcount=0;stubcount=0;
9307 void *beginning=start_block();
9308 void *instr_addr0_override = NULL;
9311 if ((Config.HLE && start == 0x80000080) || start == 0x80030000) {
9312 instr_addr0_override = out;
9313 emit_movimm(start, 0);
9314 if (start == 0x80030000) {
9315 // for BiosBootBypass() to work
9316 // io address var abused as a "already been here" flag
9317 emit_readword(&address, 1);
9318 emit_writeword(0, &psxRegs.pc);
9319 emit_writeword(0, &address);
9323 emit_readword(&psxRegs.cpuInRecursion, 1);
9324 emit_writeword(0, &psxRegs.pc);
9328 emit_jeq(out + 4*2);
9329 emit_far_jump(new_dyna_leave);
9331 emit_jne(new_dyna_leave);
9336 __builtin_prefetch(regs[i+1].regmap);
9337 check_regmap(regmap_pre[i]);
9338 check_regmap(regs[i].regmap_entry);
9339 check_regmap(regs[i].regmap);
9340 //if(ds) printf("ds: ");
9341 disassemble_inst(i);
9343 ds=0; // Skip delay slot
9344 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9345 instr_addr[i] = NULL;
9347 speculate_register_values(i);
9348 #ifndef DESTRUCTIVE_WRITEBACK
9349 if (i < 2 || !dops[i-2].is_ujump)
9351 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9353 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9354 dirty_pre=branch_regs[i].dirty;
9356 dirty_pre=regs[i].dirty;
9360 if (i < 2 || !dops[i-2].is_ujump)
9362 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9363 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9365 // branch target entry point
9366 instr_addr[i] = out;
9367 assem_debug("<->\n");
9368 drc_dbg_emit_do_cmp(i, cinfo[i].ccadj);
9369 if (clear_hack_addr) {
9371 emit_writeword(0, &hack_addr);
9372 clear_hack_addr = 0;
9376 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
9377 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9378 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9379 address_generation(i,®s[i],regs[i].regmap_entry);
9380 load_consts(regmap_pre[i],regs[i].regmap,i);
9383 // Load the delay slot registers if necessary
9384 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))
9385 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9386 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))
9387 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9388 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9389 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9390 if (dops[i+1].is_store)
9391 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9395 // Preload registers for following instruction
9396 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9397 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9398 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9399 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9400 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9401 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9403 // TODO: if(is_ooo(i)) address_generation(i+1);
9404 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9405 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9406 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9407 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9408 if (dops[i].is_store)
9409 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9411 ds = assemble(i, ®s[i], cinfo[i].ccadj);
9413 drc_dbg_emit_wb_dirtys(i, ®s[i]);
9414 if (dops[i].is_ujump)
9417 literal_pool_jumpover(256);
9422 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9423 // no ending needed for this block since INTCALL never returns
9425 // If the block did not end with an unconditional branch,
9426 // add a jump to the next instruction.
9428 if (!dops[i-2].is_ujump) {
9429 assert(!dops[i-1].is_jump);
9431 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9432 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9433 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9434 emit_loadreg(CCREG,HOST_CCREG);
9435 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9439 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9440 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9442 add_to_linker(out,start+i*4,0);
9449 assert(!dops[i-1].is_jump);
9450 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9451 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9452 emit_loadreg(CCREG,HOST_CCREG);
9453 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9454 add_to_linker(out,start+i*4,0);
9459 for(i = 0; i < stubcount; i++)
9461 switch(stubs[i].type)
9468 do_readstub(i);break;
9472 do_writestub(i);break;
9476 do_invstub(i);break;
9478 do_unalignedwritestub(i);break;
9480 do_overflowstub(i); break;
9481 case ALIGNMENT_STUB:
9482 do_alignmentstub(i); break;
9488 if (instr_addr0_override)
9489 instr_addr[0] = instr_addr0_override;
9492 /* check for improper expiration */
9493 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9497 for (j = 0; j < jumps[i]->count; j++)
9498 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9502 /* Pass 9 - Linker */
9503 for(i=0;i<linkcount;i++)
9505 assem_debug("link: %p -> %08x\n",
9506 log_addr(link_addr[i].addr), link_addr[i].target);
9508 if (!link_addr[i].internal)
9511 void *addr = check_addr(link_addr[i].target);
9512 emit_extjump(link_addr[i].addr, link_addr[i].target);
9514 set_jump_target(link_addr[i].addr, addr);
9515 ndrc_add_jump_out(link_addr[i].target,stub);
9518 set_jump_target(link_addr[i].addr, stub);
9523 int target=(link_addr[i].target-start)>>2;
9524 assert(target>=0&&target<slen);
9525 assert(instr_addr[target]);
9526 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9527 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9529 set_jump_target(link_addr[i].addr, instr_addr[target]);
9534 u_int source_len = slen*4;
9535 if (dops[slen-1].itype == INTCALL && source_len > 4)
9536 // no need to treat the last instruction as compiled
9537 // as interpreter fully handles it
9540 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9543 // External Branch Targets (jump_in)
9544 int jump_in_count = 1;
9545 assert(instr_addr[0]);
9546 for (i = 1; i < slen; i++)
9548 if (dops[i].bt && instr_addr[i])
9552 struct block_info *block =
9553 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9554 block->reg_sv_flags = state_rflags;
9557 for (i = 0; i < slen; i++)
9559 if ((i == 0 || dops[i].bt) && instr_addr[i])
9561 assem_debug("%p (%d) <- %8x\n", log_addr(instr_addr[i]), i, start + i*4);
9562 u_int vaddr = start + i*4;
9568 entry = instr_addr[i];
9570 emit_jmp(instr_addr[i]);
9572 block->jump_in[jump_in_i].vaddr = vaddr;
9573 block->jump_in[jump_in_i].addr = entry;
9577 assert(jump_in_i == jump_in_count);
9578 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9579 // Write out the literal pool if necessary
9581 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9583 if(((u_int)out)&7) emit_addnop(13);
9585 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9586 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9587 memcpy(copy, source, source_len);
9590 end_block(beginning);
9592 // If we're within 256K of the end of the buffer,
9593 // start over from the beginning. (Is 256K enough?)
9594 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9595 out = ndrc->translation_cache;
9597 // Trap writes to any of the pages we compiled
9598 mark_invalid_code(start, slen*4, 0);
9600 /* Pass 10 - Free memory by expiring oldest blocks */
9602 pass10_expire_blocks();
9607 stat_inc(stat_bc_direct);
9611 // vim:shiftwidth=2:expandtab