notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
drc: handle regs-not-in-psxRegs case better
[pcsx_rearmed.git] / libpcsxcore / new_dynarec / new_dynarec.c
... / ...
CommitLineData
1/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
2 * Mupen64plus - new_dynarec.c *
3 * Copyright (C) 2009-2011 Ari64 *
4 * *
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. *
9 * *
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. *
14 * *
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 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
20
21#include <stdlib.h>
22#include <stdint.h> //include for uint64_t
23#include <assert.h>
24#include <errno.h>
25#include <sys/mman.h>
26#ifdef __MACH__
27#include <libkern/OSCacheControl.h>
28#endif
29#ifdef _3DS
30#include <3ds_utils.h>
31#endif
32#ifdef HAVE_LIBNX
33#include <switch.h>
34static Jit g_jit;
35#endif
36
37#include "new_dynarec_config.h"
38#include "../psxhle.h"
39#include "../psxinterpreter.h"
40#include "../gte.h"
41#include "emu_if.h" // emulator interface
42#include "arm_features.h"
43
44#define unused __attribute__((unused))
45#ifdef __clang__
46#define noinline __attribute__((noinline))
47#else
48#define noinline __attribute__((noinline,noclone))
49#endif
50#ifndef ARRAY_SIZE
51#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
52#endif
53#ifndef min
54#define min(a, b) ((b) < (a) ? (b) : (a))
55#endif
56#ifndef max
57#define max(a, b) ((b) > (a) ? (b) : (a))
58#endif
59
60//#define DISASM
61//#define ASSEM_PRINT
62//#define INV_DEBUG_W
63//#define STAT_PRINT
64
65#ifdef ASSEM_PRINT
66#define assem_debug printf
67#else
68#define assem_debug(...)
69#endif
70//#define inv_debug printf
71#define inv_debug(...)
72
73#ifdef __i386__
74#include "assem_x86.h"
75#endif
76#ifdef __x86_64__
77#include "assem_x64.h"
78#endif
79#ifdef __arm__
80#include "assem_arm.h"
81#endif
82#ifdef __aarch64__
83#include "assem_arm64.h"
84#endif
85
86#define RAM_SIZE 0x200000
87#define MAXBLOCK 4096
88#define MAX_OUTPUT_BLOCK_SIZE 262144
89#define EXPIRITY_OFFSET (MAX_OUTPUT_BLOCK_SIZE * 2)
90#define PAGE_COUNT 1024
91
92#if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
93#define INVALIDATE_USE_COND_CALL
94#endif
95
96#ifdef VITA
97// apparently Vita has a 16MB limit, so either we cut tc in half,
98// or use this hack (it's a hack because tc size was designed to be power-of-2)
99#define TC_REDUCE_BYTES 4096
100#else
101#define TC_REDUCE_BYTES 0
102#endif
103
104struct ndrc_tramp
105{
106 struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
107 const void *f[2048 / sizeof(void *)];
108};
109
110struct ndrc_mem
111{
112 u_char translation_cache[(1 << TARGET_SIZE_2) - TC_REDUCE_BYTES];
113 struct ndrc_tramp tramp;
114};
115
116#ifdef BASE_ADDR_DYNAMIC
117static struct ndrc_mem *ndrc;
118#else
119static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
120static struct ndrc_mem *ndrc = &ndrc_;
121#endif
122#ifdef TC_WRITE_OFFSET
123# ifdef __GLIBC__
124# include <sys/types.h>
125# include <sys/stat.h>
126# include <fcntl.h>
127# include <unistd.h>
128# endif
129static long ndrc_write_ofs;
130#define NDRC_WRITE_OFFSET(x) (void *)((char *)(x) + ndrc_write_ofs)
131#else
132#define NDRC_WRITE_OFFSET(x) (x)
133#endif
134
135// stubs
136enum stub_type {
137 CC_STUB = 1,
138 FP_STUB = 2,
139 LOADB_STUB = 3,
140 LOADH_STUB = 4,
141 LOADW_STUB = 5,
142 LOADD_STUB = 6,
143 LOADBU_STUB = 7,
144 LOADHU_STUB = 8,
145 STOREB_STUB = 9,
146 STOREH_STUB = 10,
147 STOREW_STUB = 11,
148 STORED_STUB = 12,
149 STORELR_STUB = 13,
150 INVCODE_STUB = 14,
151};
152
153// regmap_pre[i] - regs before [i] insn starts; dirty things here that
154// don't match .regmap will be written back
155// [i].regmap_entry - regs that must be set up if someone jumps here
156// [i].regmap - regs [i] insn will read/(over)write
157// branch_regs[i].* - same as above but for branches, takes delay slot into account
158struct regstat
159{
160 signed char regmap_entry[HOST_REGS];
161 signed char regmap[HOST_REGS];
162 uint64_t wasdirty;
163 uint64_t dirty;
164 uint64_t u;
165 u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
166 u_int isconst; // ... but isconst is false when r2 is known
167 u_int loadedconst; // host regs that have constants loaded
168 //u_int waswritten; // MIPS regs that were used as store base before
169};
170
171struct ht_entry
172{
173 u_int vaddr[2];
174 void *tcaddr[2];
175};
176
177struct code_stub
178{
179 enum stub_type type;
180 void *addr;
181 void *retaddr;
182 u_int a;
183 uintptr_t b;
184 uintptr_t c;
185 u_int d;
186 u_int e;
187};
188
189struct link_entry
190{
191 void *addr;
192 u_int target;
193 u_int internal;
194};
195
196struct block_info
197{
198 struct block_info *next;
199 const void *source;
200 const void *copy;
201 u_int start; // vaddr of the block start
202 u_int len; // of the whole block source
203 u_int tc_offs;
204 //u_int tc_len;
205 u_int reg_sv_flags;
206 u_char is_dirty;
207 u_char inv_near_misses;
208 u_short jump_in_cnt;
209 struct {
210 u_int vaddr;
211 void *addr;
212 } jump_in[0];
213};
214
215struct jump_info
216{
217 int alloc;
218 int count;
219 struct {
220 u_int target_vaddr;
221 void *stub;
222 } e[0];
223};
224
225static struct decoded_insn
226{
227 u_char itype;
228 u_char opcode;
229 u_char opcode2;
230 u_char rs1;
231 u_char rs2;
232 u_char rt1;
233 u_char rt2;
234 u_char use_lt1:1;
235 u_char bt:1;
236 u_char ooo:1;
237 u_char is_ds:1;
238 u_char is_jump:1;
239 u_char is_ujump:1;
240 u_char is_load:1;
241 u_char is_store:1;
242} dops[MAXBLOCK];
243
244 static u_char *out;
245 static char invalid_code[0x100000];
246 static struct ht_entry hash_table[65536];
247 static struct block_info *blocks[PAGE_COUNT];
248 static struct jump_info *jumps[PAGE_COUNT];
249 static u_int start;
250 static u_int *source;
251 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
252 static uint64_t gte_rt[MAXBLOCK];
253 static uint64_t gte_unneeded[MAXBLOCK];
254 static u_int smrv[32]; // speculated MIPS register values
255 static u_int smrv_strong; // mask or regs that are likely to have correct values
256 static u_int smrv_weak; // same, but somewhat less likely
257 static u_int smrv_strong_next; // same, but after current insn executes
258 static u_int smrv_weak_next;
259 static int imm[MAXBLOCK];
260 static u_int ba[MAXBLOCK];
261 static uint64_t unneeded_reg[MAXBLOCK];
262 static uint64_t branch_unneeded_reg[MAXBLOCK];
263 // see 'struct regstat' for a description
264 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
265 // contains 'real' consts at [i] insn, but may differ from what's actually
266 // loaded in host reg as 'final' value is always loaded, see get_final_value()
267 static uint32_t current_constmap[HOST_REGS];
268 static uint32_t constmap[MAXBLOCK][HOST_REGS];
269 static struct regstat regs[MAXBLOCK];
270 static struct regstat branch_regs[MAXBLOCK];
271 static signed char minimum_free_regs[MAXBLOCK];
272 static int ccadj[MAXBLOCK];
273 static int slen;
274 static void *instr_addr[MAXBLOCK];
275 static struct link_entry link_addr[MAXBLOCK];
276 static int linkcount;
277 static struct code_stub stubs[MAXBLOCK*3];
278 static int stubcount;
279 static u_int literals[1024][2];
280 static int literalcount;
281 static int is_delayslot;
282 static char shadow[1048576] __attribute__((aligned(16)));
283 static void *copy;
284 static u_int expirep;
285 static u_int stop_after_jal;
286 static u_int f1_hack;
287#ifdef STAT_PRINT
288 static int stat_bc_direct;
289 static int stat_bc_pre;
290 static int stat_bc_restore;
291 static int stat_ht_lookups;
292 static int stat_jump_in_lookups;
293 static int stat_restore_tries;
294 static int stat_restore_compares;
295 static int stat_inv_addr_calls;
296 static int stat_inv_hits;
297 static int stat_blocks;
298 static int stat_links;
299 #define stat_inc(s) s++
300 #define stat_dec(s) s--
301 #define stat_clear(s) s = 0
302#else
303 #define stat_inc(s)
304 #define stat_dec(s)
305 #define stat_clear(s)
306#endif
307
308 int new_dynarec_hacks;
309 int new_dynarec_hacks_pergame;
310 int new_dynarec_hacks_old;
311 int new_dynarec_did_compile;
312
313 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
314
315 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
316 extern int last_count; // last absolute target, often = next_interupt
317 extern int pcaddr;
318 extern int pending_exception;
319 extern int branch_target;
320 extern uintptr_t ram_offset;
321 extern uintptr_t mini_ht[32][2];
322
323 /* registers that may be allocated */
324 /* 1-31 gpr */
325#define LOREG 32 // lo
326#define HIREG 33 // hi
327//#define FSREG 34 // FPU status (FCSR)
328#define CSREG 35 // Coprocessor status
329#define CCREG 36 // Cycle count
330#define INVCP 37 // Pointer to invalid_code
331//#define MMREG 38 // Pointer to memory_map
332#define ROREG 39 // ram offset (if rdram!=0x80000000)
333#define TEMPREG 40
334#define FTEMP 40 // FPU temporary register
335#define PTEMP 41 // Prefetch temporary register
336//#define TLREG 42 // TLB mapping offset
337#define RHASH 43 // Return address hash
338#define RHTBL 44 // Return address hash table address
339#define RTEMP 45 // JR/JALR address register
340#define MAXREG 45
341#define AGEN1 46 // Address generation temporary register
342//#define AGEN2 47 // Address generation temporary register
343//#define MGEN1 48 // Maptable address generation temporary register
344//#define MGEN2 49 // Maptable address generation temporary register
345#define BTREG 50 // Branch target temporary register
346
347 /* instruction types */
348#define NOP 0 // No operation
349#define LOAD 1 // Load
350#define STORE 2 // Store
351#define LOADLR 3 // Unaligned load
352#define STORELR 4 // Unaligned store
353#define MOV 5 // Move
354#define ALU 6 // Arithmetic/logic
355#define MULTDIV 7 // Multiply/divide
356#define SHIFT 8 // Shift by register
357#define SHIFTIMM 9// Shift by immediate
358#define IMM16 10 // 16-bit immediate
359#define RJUMP 11 // Unconditional jump to register
360#define UJUMP 12 // Unconditional jump
361#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
362#define SJUMP 14 // Conditional branch (regimm format)
363#define COP0 15 // Coprocessor 0
364#define COP1 16 // Coprocessor 1
365#define C1LS 17 // Coprocessor 1 load/store
366//#define FJUMP 18 // Conditional branch (floating point)
367//#define FLOAT 19 // Floating point unit
368//#define FCONV 20 // Convert integer to float
369//#define FCOMP 21 // Floating point compare (sets FSREG)
370#define SYSCALL 22// SYSCALL,BREAK
371#define OTHER 23 // Other
372//#define SPAN 24 // Branch/delay slot spans 2 pages
373#define NI 25 // Not implemented
374#define HLECALL 26// PCSX fake opcodes for HLE
375#define COP2 27 // Coprocessor 2 move
376#define C2LS 28 // Coprocessor 2 load/store
377#define C2OP 29 // Coprocessor 2 operation
378#define INTCALL 30// Call interpreter to handle rare corner cases
379
380 /* branch codes */
381#define TAKEN 1
382#define NOTTAKEN 2
383#define NULLDS 3
384
385#define DJT_1 (void *)1l // no function, just a label in assem_debug log
386#define DJT_2 (void *)2l
387
388// asm linkage
389void dyna_linker();
390void cc_interrupt();
391void fp_exception();
392void fp_exception_ds();
393void jump_syscall (u_int u0, u_int u1, u_int pc);
394void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
395void jump_break (u_int u0, u_int u1, u_int pc);
396void jump_break_ds(u_int u0, u_int u1, u_int pc);
397void jump_to_new_pc();
398void call_gteStall();
399void new_dyna_leave();
400
401void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile);
402void *ndrc_get_addr_ht(u_int vaddr);
403void ndrc_add_jump_out(u_int vaddr, void *src);
404void ndrc_write_invalidate_one(u_int addr);
405static void ndrc_write_invalidate_many(u_int addr, u_int end);
406
407static int new_recompile_block(u_int addr);
408static void invalidate_block(struct block_info *block);
409
410// Needed by assembler
411static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
412static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
413static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
414static void load_all_regs(const signed char i_regmap[]);
415static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
416static void load_regs_entry(int t);
417static void load_all_consts(const signed char regmap[], u_int dirty, int i);
418static u_int get_host_reglist(const signed char *regmap);
419
420static int get_final_value(int hr, int i, int *value);
421static void add_stub(enum stub_type type, void *addr, void *retaddr,
422 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
423static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
424 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
425static void add_to_linker(void *addr, u_int target, int ext);
426static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
427 int addr, int *offset_reg, int *addr_reg_override);
428static void *get_direct_memhandler(void *table, u_int addr,
429 enum stub_type type, uintptr_t *addr_host);
430static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
431static void pass_args(int a0, int a1);
432static void emit_far_jump(const void *f);
433static void emit_far_call(const void *f);
434
435#ifdef VITA
436#include <psp2/kernel/sysmem.h>
437static int sceBlock;
438// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
439extern int getVMBlock();
440int _newlib_vm_size_user = sizeof(*ndrc);
441#endif
442
443static void mprotect_w_x(void *start, void *end, int is_x)
444{
445#ifdef NO_WRITE_EXEC
446 #if defined(VITA)
447 // *Open* enables write on all memory that was
448 // allocated by sceKernelAllocMemBlockForVM()?
449 if (is_x)
450 sceKernelCloseVMDomain();
451 else
452 sceKernelOpenVMDomain();
453 #elif defined(HAVE_LIBNX)
454 Result rc;
455 // check to avoid the full flush in jitTransitionToExecutable()
456 if (g_jit.type != JitType_CodeMemory) {
457 if (is_x)
458 rc = jitTransitionToExecutable(&g_jit);
459 else
460 rc = jitTransitionToWritable(&g_jit);
461 if (R_FAILED(rc))
462 ;//SysPrintf("jitTransition %d %08x\n", is_x, rc);
463 }
464 #elif defined(TC_WRITE_OFFSET)
465 // separated rx and rw areas are always available
466 #else
467 u_long mstart = (u_long)start & ~4095ul;
468 u_long mend = (u_long)end;
469 if (mprotect((void *)mstart, mend - mstart,
470 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
471 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
472 #endif
473#endif
474}
475
476static void start_tcache_write(void *start, void *end)
477{
478 mprotect_w_x(start, end, 0);
479}
480
481static void end_tcache_write(void *start, void *end)
482{
483#if defined(__arm__) || defined(__aarch64__)
484 size_t len = (char *)end - (char *)start;
485 #if defined(__BLACKBERRY_QNX__)
486 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
487 #elif defined(__MACH__)
488 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
489 #elif defined(VITA)
490 sceKernelSyncVMDomain(sceBlock, start, len);
491 #elif defined(_3DS)
492 ctr_flush_invalidate_cache();
493 #elif defined(HAVE_LIBNX)
494 if (g_jit.type == JitType_CodeMemory) {
495 armDCacheClean(start, len);
496 armICacheInvalidate((char *)start - ndrc_write_ofs, len);
497 // as of v4.2.1 libnx lacks isb
498 __asm__ volatile("isb" ::: "memory");
499 }
500 #elif defined(__aarch64__)
501 // as of 2021, __clear_cache() is still broken on arm64
502 // so here is a custom one :(
503 clear_cache_arm64(start, end);
504 #else
505 __clear_cache(start, end);
506 #endif
507 (void)len;
508#endif
509
510 mprotect_w_x(start, end, 1);
511}
512
513static void *start_block(void)
514{
515 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
516 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
517 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
518 start_tcache_write(NDRC_WRITE_OFFSET(out), NDRC_WRITE_OFFSET(end));
519 return out;
520}
521
522static void end_block(void *start)
523{
524 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(out));
525}
526
527#ifdef NDRC_CACHE_FLUSH_ALL
528
529static int needs_clear_cache;
530
531static void mark_clear_cache(void *target)
532{
533 if (!needs_clear_cache) {
534 start_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
535 needs_clear_cache = 1;
536 }
537}
538
539static void do_clear_cache(void)
540{
541 if (needs_clear_cache) {
542 end_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
543 needs_clear_cache = 0;
544 }
545}
546
547#else
548
549// also takes care of w^x mappings when patching code
550static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
551
552static void mark_clear_cache(void *target)
553{
554 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
555 u_int mask = 1u << ((offset >> 12) & 31);
556 if (!(needs_clear_cache[offset >> 17] & mask)) {
557 char *start = (char *)NDRC_WRITE_OFFSET((uintptr_t)target & ~4095l);
558 start_tcache_write(start, start + 4095);
559 needs_clear_cache[offset >> 17] |= mask;
560 }
561}
562
563// Clearing the cache is rather slow on ARM Linux, so mark the areas
564// that need to be cleared, and then only clear these areas once.
565static void do_clear_cache(void)
566{
567 int i, j;
568 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
569 {
570 u_int bitmap = needs_clear_cache[i];
571 if (!bitmap)
572 continue;
573 for (j = 0; j < 32; j++)
574 {
575 u_char *start, *end;
576 if (!(bitmap & (1u << j)))
577 continue;
578
579 start = ndrc->translation_cache + i*131072 + j*4096;
580 end = start + 4095;
581 for (j++; j < 32; j++) {
582 if (!(bitmap & (1u << j)))
583 break;
584 end += 4096;
585 }
586 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(end));
587 }
588 needs_clear_cache[i] = 0;
589 }
590}
591
592#endif // NDRC_CACHE_FLUSH_ALL
593
594#define NO_CYCLE_PENALTY_THR 12
595
596int cycle_multiplier_old;
597static int cycle_multiplier_active;
598
599static int CLOCK_ADJUST(int x)
600{
601 int m = cycle_multiplier_active;
602 int s = (x >> 31) | 1;
603 return (x * m + s * 50) / 100;
604}
605
606static int ds_writes_rjump_rs(int i)
607{
608 return dops[i].rs1 != 0 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2);
609}
610
611// psx addr mirror masking (for invalidation)
612static u_int pmmask(u_int vaddr)
613{
614 vaddr &= ~0xe0000000;
615 if (vaddr < 0x01000000)
616 vaddr &= ~0x00e00000; // RAM mirrors
617 return vaddr;
618}
619
620static u_int get_page(u_int vaddr)
621{
622 u_int page = pmmask(vaddr) >> 12;
623 if (page >= PAGE_COUNT / 2)
624 page = PAGE_COUNT / 2 + (page & (PAGE_COUNT / 2 - 1));
625 return page;
626}
627
628// get a page for looking for a block that has vaddr
629// (needed because the block may start in previous page)
630static u_int get_page_prev(u_int vaddr)
631{
632 assert(MAXBLOCK <= (1 << 12));
633 u_int page = get_page(vaddr);
634 if (page & 511)
635 page--;
636 return page;
637}
638
639static struct ht_entry *hash_table_get(u_int vaddr)
640{
641 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
642}
643
644static void hash_table_add(u_int vaddr, void *tcaddr)
645{
646 struct ht_entry *ht_bin = hash_table_get(vaddr);
647 assert(tcaddr);
648 ht_bin->vaddr[1] = ht_bin->vaddr[0];
649 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
650 ht_bin->vaddr[0] = vaddr;
651 ht_bin->tcaddr[0] = tcaddr;
652}
653
654static void hash_table_remove(int vaddr)
655{
656 //printf("remove hash: %x\n",vaddr);
657 struct ht_entry *ht_bin = hash_table_get(vaddr);
658 if (ht_bin->vaddr[1] == vaddr) {
659 ht_bin->vaddr[1] = -1;
660 ht_bin->tcaddr[1] = NULL;
661 }
662 if (ht_bin->vaddr[0] == vaddr) {
663 ht_bin->vaddr[0] = ht_bin->vaddr[1];
664 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
665 ht_bin->vaddr[1] = -1;
666 ht_bin->tcaddr[1] = NULL;
667 }
668}
669
670static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
671{
672 u_int vaddr_m = vaddr & 0x1fffffff;
673 u_int i, j;
674 for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
675 // ram mirrors, but should not hurt bios
676 for (j = 0; j < 0x800000; j += 0x200000) {
677 invalid_code[(i|j) >> 12] =
678 invalid_code[(i|j|0x80000000u) >> 12] =
679 invalid_code[(i|j|0xa0000000u) >> 12] = invalid;
680 }
681 }
682 if (!invalid && vaddr + len > inv_code_start && vaddr <= inv_code_end)
683 inv_code_start = inv_code_end = ~0;
684}
685
686static int doesnt_expire_soon(u_char *tcaddr)
687{
688 u_int diff = (u_int)(tcaddr - out) & ((1u << TARGET_SIZE_2) - 1u);
689 return diff > EXPIRITY_OFFSET + MAX_OUTPUT_BLOCK_SIZE;
690}
691
692static unused void check_for_block_changes(u_int start, u_int end)
693{
694 u_int start_page = get_page_prev(start);
695 u_int end_page = get_page(end - 1);
696 u_int page;
697
698 for (page = start_page; page <= end_page; page++) {
699 struct block_info *block;
700 for (block = blocks[page]; block != NULL; block = block->next) {
701 if (block->is_dirty)
702 continue;
703 if (memcmp(block->source, block->copy, block->len)) {
704 printf("bad block %08x-%08x %016llx %016llx @%08x\n",
705 block->start, block->start + block->len,
706 *(long long *)block->source, *(long long *)block->copy, psxRegs.pc);
707 fflush(stdout);
708 abort();
709 }
710 }
711 }
712}
713
714static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
715{
716 void *found_clean = NULL;
717 u_int i, page;
718
719 stat_inc(stat_restore_tries);
720 for (page = start_page; page <= end_page; page++) {
721 struct block_info *block;
722 for (block = blocks[page]; block != NULL; block = block->next) {
723 if (vaddr < block->start)
724 break;
725 if (!block->is_dirty || vaddr >= block->start + block->len)
726 continue;
727 for (i = 0; i < block->jump_in_cnt; i++)
728 if (block->jump_in[i].vaddr == vaddr)
729 break;
730 if (i == block->jump_in_cnt)
731 continue;
732 assert(block->source && block->copy);
733 stat_inc(stat_restore_compares);
734 if (memcmp(block->source, block->copy, block->len))
735 continue;
736
737 block->is_dirty = block->inv_near_misses = 0;
738 found_clean = block->jump_in[i].addr;
739 hash_table_add(vaddr, found_clean);
740 mark_invalid_code(block->start, block->len, 0);
741 stat_inc(stat_bc_restore);
742 inv_debug("INV: restored %08x %p (%d)\n", vaddr, found_clean, block->jump_in_cnt);
743 return found_clean;
744 }
745 }
746 return NULL;
747}
748
749// Get address from virtual address
750// This is called from the recompiled JR/JALR instructions
751static void noinline *get_addr(u_int vaddr, int can_compile)
752{
753 u_int start_page = get_page_prev(vaddr);
754 u_int i, page, end_page = get_page(vaddr);
755 void *found_clean = NULL;
756
757 stat_inc(stat_jump_in_lookups);
758 for (page = start_page; page <= end_page; page++) {
759 const struct block_info *block;
760 for (block = blocks[page]; block != NULL; block = block->next) {
761 if (vaddr < block->start)
762 break;
763 if (block->is_dirty || vaddr >= block->start + block->len)
764 continue;
765 for (i = 0; i < block->jump_in_cnt; i++)
766 if (block->jump_in[i].vaddr == vaddr)
767 break;
768 if (i == block->jump_in_cnt)
769 continue;
770 found_clean = block->jump_in[i].addr;
771 hash_table_add(vaddr, found_clean);
772 return found_clean;
773 }
774 }
775 found_clean = try_restore_block(vaddr, start_page, end_page);
776 if (found_clean)
777 return found_clean;
778
779 if (!can_compile)
780 return NULL;
781
782 int r = new_recompile_block(vaddr);
783 if (r == 0)
784 return ndrc_get_addr_ht(vaddr);
785
786 // generate an address error
787 psxRegs.CP0.n.Status |= 2;
788 psxRegs.CP0.n.Cause = (vaddr<<31) | (4<<2);
789 psxRegs.CP0.n.EPC = (vaddr&1) ? vaddr-5 : vaddr;
790 psxRegs.CP0.n.BadVAddr = vaddr & ~1;
791 return ndrc_get_addr_ht(0x80000080);
792}
793
794// Look up address in hash table first
795void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile)
796{
797 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
798 const struct ht_entry *ht_bin = hash_table_get(vaddr);
799 stat_inc(stat_ht_lookups);
800 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
801 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
802 return get_addr(vaddr, can_compile);
803}
804
805void *ndrc_get_addr_ht(u_int vaddr)
806{
807 return ndrc_get_addr_ht_param(vaddr, 1);
808}
809
810static void clear_all_regs(signed char regmap[])
811{
812 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
813}
814
815// get_reg: get allocated host reg from mips reg
816// returns -1 if no such mips reg was allocated
817#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
818
819extern signed char get_reg(const signed char regmap[], signed char r);
820
821#else
822
823static signed char get_reg(const signed char regmap[], signed char r)
824{
825 int hr;
826 for (hr = 0; hr < HOST_REGS; hr++) {
827 if (hr == EXCLUDE_REG)
828 continue;
829 if (regmap[hr] == r)
830 return hr;
831 }
832 return -1;
833}
834
835#endif
836
837// get reg as mask bit (1 << hr)
838static u_int get_regm(const signed char regmap[], signed char r)
839{
840 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
841}
842
843static signed char get_reg_temp(const signed char regmap[])
844{
845 int hr;
846 for (hr = 0; hr < HOST_REGS; hr++) {
847 if (hr == EXCLUDE_REG)
848 continue;
849 if (regmap[hr] == (signed char)-1)
850 return hr;
851 }
852 return -1;
853}
854
855// Find a register that is available for two consecutive cycles
856static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
857{
858 int hr;
859 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
860 return -1;
861}
862
863// reverse reg map: mips -> host
864#define RRMAP_SIZE 64
865static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
866 u_int *regs_can_change)
867{
868 u_int r, hr, hr_can_change = 0;
869 memset(rrmap, -1, RRMAP_SIZE);
870 for (hr = 0; hr < HOST_REGS; )
871 {
872 r = regmap[hr];
873 rrmap[r & (RRMAP_SIZE - 1)] = hr;
874 // only add mips $1-$31+$lo, others shifted out
875 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
876 hr++;
877 if (hr == EXCLUDE_REG)
878 hr++;
879 }
880 hr_can_change |= 1u << (rrmap[33] & 31);
881 hr_can_change |= 1u << (rrmap[CCREG] & 31);
882 hr_can_change &= ~(1u << 31);
883 *regs_can_change = hr_can_change;
884}
885
886// same as get_reg, but takes rrmap
887static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
888{
889 assert(0 <= r && r < RRMAP_SIZE);
890 return rrmap[r];
891}
892
893static int count_free_regs(const signed char regmap[])
894{
895 int count=0;
896 int hr;
897 for(hr=0;hr<HOST_REGS;hr++)
898 {
899 if(hr!=EXCLUDE_REG) {
900 if(regmap[hr]<0) count++;
901 }
902 }
903 return count;
904}
905
906static void dirty_reg(struct regstat *cur, signed char reg)
907{
908 int hr;
909 if (!reg) return;
910 hr = get_reg(cur->regmap, reg);
911 if (hr >= 0)
912 cur->dirty |= 1<<hr;
913}
914
915static void set_const(struct regstat *cur, signed char reg, uint32_t value)
916{
917 int hr;
918 if (!reg) return;
919 hr = get_reg(cur->regmap, reg);
920 if (hr >= 0) {
921 cur->isconst |= 1<<hr;
922 current_constmap[hr] = value;
923 }
924}
925
926static void clear_const(struct regstat *cur, signed char reg)
927{
928 int hr;
929 if (!reg) return;
930 hr = get_reg(cur->regmap, reg);
931 if (hr >= 0)
932 cur->isconst &= ~(1<<hr);
933}
934
935static int is_const(const struct regstat *cur, signed char reg)
936{
937 int hr;
938 if (reg < 0) return 0;
939 if (!reg) return 1;
940 hr = get_reg(cur->regmap, reg);
941 if (hr >= 0)
942 return (cur->isconst>>hr)&1;
943 return 0;
944}
945
946static uint32_t get_const(const struct regstat *cur, signed char reg)
947{
948 int hr;
949 if (!reg) return 0;
950 hr = get_reg(cur->regmap, reg);
951 if (hr >= 0)
952 return current_constmap[hr];
953
954 SysPrintf("Unknown constant in r%d\n", reg);
955 abort();
956}
957
958// Least soon needed registers
959// Look at the next ten instructions and see which registers
960// will be used. Try not to reallocate these.
961static void lsn(u_char hsn[], int i, int *preferred_reg)
962{
963 int j;
964 int b=-1;
965 for(j=0;j<9;j++)
966 {
967 if(i+j>=slen) {
968 j=slen-i-1;
969 break;
970 }
971 if (dops[i+j].is_ujump)
972 {
973 // Don't go past an unconditonal jump
974 j++;
975 break;
976 }
977 }
978 for(;j>=0;j--)
979 {
980 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
981 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
982 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
983 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
984 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
985 // Stores can allocate zero
986 hsn[dops[i+j].rs1]=j;
987 hsn[dops[i+j].rs2]=j;
988 }
989 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
990 hsn[ROREG] = j;
991 // On some architectures stores need invc_ptr
992 #if defined(HOST_IMM8)
993 if (dops[i+j].is_store)
994 hsn[INVCP] = j;
995 #endif
996 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
997 {
998 hsn[CCREG]=j;
999 b=j;
1000 }
1001 }
1002 if(b>=0)
1003 {
1004 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
1005 {
1006 // Follow first branch
1007 int t=(ba[i+b]-start)>>2;
1008 j=7-b;if(t+j>=slen) j=slen-t-1;
1009 for(;j>=0;j--)
1010 {
1011 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1012 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1013 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1014 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1015 }
1016 }
1017 // TODO: preferred register based on backward branch
1018 }
1019 // Delay slot should preferably not overwrite branch conditions or cycle count
1020 if (i > 0 && dops[i-1].is_jump) {
1021 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1022 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1023 hsn[CCREG]=1;
1024 // ...or hash tables
1025 hsn[RHASH]=1;
1026 hsn[RHTBL]=1;
1027 }
1028 // Coprocessor load/store needs FTEMP, even if not declared
1029 if(dops[i].itype==C2LS) {
1030 hsn[FTEMP]=0;
1031 }
1032 // Load L/R also uses FTEMP as a temporary register
1033 if(dops[i].itype==LOADLR) {
1034 hsn[FTEMP]=0;
1035 }
1036 // Also SWL/SWR/SDL/SDR
1037 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
1038 hsn[FTEMP]=0;
1039 }
1040 // Don't remove the miniht registers
1041 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1042 {
1043 hsn[RHASH]=0;
1044 hsn[RHTBL]=0;
1045 }
1046}
1047
1048// We only want to allocate registers if we're going to use them again soon
1049static int needed_again(int r, int i)
1050{
1051 int j;
1052 int b=-1;
1053 int rn=10;
1054
1055 if (i > 0 && dops[i-1].is_ujump)
1056 {
1057 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
1058 return 0; // Don't need any registers if exiting the block
1059 }
1060 for(j=0;j<9;j++)
1061 {
1062 if(i+j>=slen) {
1063 j=slen-i-1;
1064 break;
1065 }
1066 if (dops[i+j].is_ujump)
1067 {
1068 // Don't go past an unconditonal jump
1069 j++;
1070 break;
1071 }
1072 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
1073 {
1074 break;
1075 }
1076 }
1077 for(;j>=1;j--)
1078 {
1079 if(dops[i+j].rs1==r) rn=j;
1080 if(dops[i+j].rs2==r) rn=j;
1081 if((unneeded_reg[i+j]>>r)&1) rn=10;
1082 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1083 {
1084 b=j;
1085 }
1086 }
1087 if(rn<10) return 1;
1088 (void)b;
1089 return 0;
1090}
1091
1092// Try to match register allocations at the end of a loop with those
1093// at the beginning
1094static int loop_reg(int i, int r, int hr)
1095{
1096 int j,k;
1097 for(j=0;j<9;j++)
1098 {
1099 if(i+j>=slen) {
1100 j=slen-i-1;
1101 break;
1102 }
1103 if (dops[i+j].is_ujump)
1104 {
1105 // Don't go past an unconditonal jump
1106 j++;
1107 break;
1108 }
1109 }
1110 k=0;
1111 if(i>0){
1112 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1113 k--;
1114 }
1115 for(;k<j;k++)
1116 {
1117 assert(r < 64);
1118 if((unneeded_reg[i+k]>>r)&1) return hr;
1119 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1120 {
1121 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
1122 {
1123 int t=(ba[i+k]-start)>>2;
1124 int reg=get_reg(regs[t].regmap_entry,r);
1125 if(reg>=0) return reg;
1126 //reg=get_reg(regs[t+1].regmap_entry,r);
1127 //if(reg>=0) return reg;
1128 }
1129 }
1130 }
1131 return hr;
1132}
1133
1134
1135// Allocate every register, preserving source/target regs
1136static void alloc_all(struct regstat *cur,int i)
1137{
1138 int hr;
1139
1140 for(hr=0;hr<HOST_REGS;hr++) {
1141 if(hr!=EXCLUDE_REG) {
1142 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1143 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1144 {
1145 cur->regmap[hr]=-1;
1146 cur->dirty&=~(1<<hr);
1147 }
1148 // Don't need zeros
1149 if(cur->regmap[hr]==0)
1150 {
1151 cur->regmap[hr]=-1;
1152 cur->dirty&=~(1<<hr);
1153 }
1154 }
1155 }
1156}
1157
1158#ifndef NDEBUG
1159static int host_tempreg_in_use;
1160
1161static void host_tempreg_acquire(void)
1162{
1163 assert(!host_tempreg_in_use);
1164 host_tempreg_in_use = 1;
1165}
1166
1167static void host_tempreg_release(void)
1168{
1169 host_tempreg_in_use = 0;
1170}
1171#else
1172static void host_tempreg_acquire(void) {}
1173static void host_tempreg_release(void) {}
1174#endif
1175
1176#ifdef ASSEM_PRINT
1177extern void gen_interupt();
1178extern void do_insn_cmp();
1179#define FUNCNAME(f) { f, " " #f }
1180static const struct {
1181 void *addr;
1182 const char *name;
1183} function_names[] = {
1184 FUNCNAME(cc_interrupt),
1185 FUNCNAME(gen_interupt),
1186 FUNCNAME(ndrc_get_addr_ht),
1187 FUNCNAME(jump_handler_read8),
1188 FUNCNAME(jump_handler_read16),
1189 FUNCNAME(jump_handler_read32),
1190 FUNCNAME(jump_handler_write8),
1191 FUNCNAME(jump_handler_write16),
1192 FUNCNAME(jump_handler_write32),
1193 FUNCNAME(ndrc_write_invalidate_one),
1194 FUNCNAME(ndrc_write_invalidate_many),
1195 FUNCNAME(jump_to_new_pc),
1196 FUNCNAME(jump_break),
1197 FUNCNAME(jump_break_ds),
1198 FUNCNAME(jump_syscall),
1199 FUNCNAME(jump_syscall_ds),
1200 FUNCNAME(call_gteStall),
1201 FUNCNAME(new_dyna_leave),
1202 FUNCNAME(pcsx_mtc0),
1203 FUNCNAME(pcsx_mtc0_ds),
1204#ifdef DRC_DBG
1205 FUNCNAME(do_insn_cmp),
1206#endif
1207};
1208
1209static const char *func_name(const void *a)
1210{
1211 int i;
1212 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1213 if (function_names[i].addr == a)
1214 return function_names[i].name;
1215 return "";
1216}
1217#else
1218#define func_name(x) ""
1219#endif
1220
1221#ifdef __i386__
1222#include "assem_x86.c"
1223#endif
1224#ifdef __x86_64__
1225#include "assem_x64.c"
1226#endif
1227#ifdef __arm__
1228#include "assem_arm.c"
1229#endif
1230#ifdef __aarch64__
1231#include "assem_arm64.c"
1232#endif
1233
1234static void *get_trampoline(const void *f)
1235{
1236 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1237 size_t i;
1238
1239 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1240 if (tramp->f[i] == f || tramp->f[i] == NULL)
1241 break;
1242 }
1243 if (i == ARRAY_SIZE(tramp->f)) {
1244 SysPrintf("trampoline table is full, last func %p\n", f);
1245 abort();
1246 }
1247 if (tramp->f[i] == NULL) {
1248 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1249 tramp->f[i] = f;
1250 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1251#ifdef HAVE_LIBNX
1252 // invalidate the RX mirror (unsure if necessary, but just in case...)
1253 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1254#endif
1255 }
1256 return &ndrc->tramp.ops[i];
1257}
1258
1259static void emit_far_jump(const void *f)
1260{
1261 if (can_jump_or_call(f)) {
1262 emit_jmp(f);
1263 return;
1264 }
1265
1266 f = get_trampoline(f);
1267 emit_jmp(f);
1268}
1269
1270static void emit_far_call(const void *f)
1271{
1272 if (can_jump_or_call(f)) {
1273 emit_call(f);
1274 return;
1275 }
1276
1277 f = get_trampoline(f);
1278 emit_call(f);
1279}
1280
1281// Check if an address is already compiled
1282// but don't return addresses which are about to expire from the cache
1283static void *check_addr(u_int vaddr)
1284{
1285 struct ht_entry *ht_bin = hash_table_get(vaddr);
1286 size_t i;
1287 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1288 if (ht_bin->vaddr[i] == vaddr)
1289 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1290 return ht_bin->tcaddr[i];
1291 }
1292
1293 // refactor to get_addr_nocompile?
1294 u_int start_page = get_page_prev(vaddr);
1295 u_int page, end_page = get_page(vaddr);
1296
1297 stat_inc(stat_jump_in_lookups);
1298 for (page = start_page; page <= end_page; page++) {
1299 const struct block_info *block;
1300 for (block = blocks[page]; block != NULL; block = block->next) {
1301 if (vaddr < block->start)
1302 break;
1303 if (block->is_dirty || vaddr >= block->start + block->len)
1304 continue;
1305 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1306 continue;
1307 for (i = 0; i < block->jump_in_cnt; i++)
1308 if (block->jump_in[i].vaddr == vaddr)
1309 break;
1310 if (i == block->jump_in_cnt)
1311 continue;
1312
1313 // Update existing entry with current address
1314 void *addr = block->jump_in[i].addr;
1315 if (ht_bin->vaddr[0] == vaddr) {
1316 ht_bin->tcaddr[0] = addr;
1317 return addr;
1318 }
1319 if (ht_bin->vaddr[1] == vaddr) {
1320 ht_bin->tcaddr[1] = addr;
1321 return addr;
1322 }
1323 // Insert into hash table with low priority.
1324 // Don't evict existing entries, as they are probably
1325 // addresses that are being accessed frequently.
1326 if (ht_bin->vaddr[0] == -1) {
1327 ht_bin->vaddr[0] = vaddr;
1328 ht_bin->tcaddr[0] = addr;
1329 }
1330 else if (ht_bin->vaddr[1] == -1) {
1331 ht_bin->vaddr[1] = vaddr;
1332 ht_bin->tcaddr[1] = addr;
1333 }
1334 return addr;
1335 }
1336 }
1337 return NULL;
1338}
1339
1340static void blocks_clear(struct block_info **head)
1341{
1342 struct block_info *cur, *next;
1343
1344 if ((cur = *head)) {
1345 *head = NULL;
1346 while (cur) {
1347 next = cur->next;
1348 free(cur);
1349 cur = next;
1350 }
1351 }
1352}
1353
1354static int blocks_remove_matching_addrs(struct block_info **head,
1355 u_int base_offs, int shift)
1356{
1357 struct block_info *next;
1358 int hit = 0;
1359 while (*head) {
1360 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1361 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1362 invalidate_block(*head);
1363 next = (*head)->next;
1364 free(*head);
1365 *head = next;
1366 stat_dec(stat_blocks);
1367 hit = 1;
1368 }
1369 else
1370 {
1371 head = &((*head)->next);
1372 }
1373 }
1374 return hit;
1375}
1376
1377// This is called when we write to a compiled block (see do_invstub)
1378static void unlink_jumps_vaddr_range(u_int start, u_int end)
1379{
1380 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1381 int i;
1382
1383 for (page = start_page; page <= end_page; page++) {
1384 struct jump_info *ji = jumps[page];
1385 if (ji == NULL)
1386 continue;
1387 for (i = 0; i < ji->count; ) {
1388 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1389 i++;
1390 continue;
1391 }
1392
1393 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1394 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1395 void *host_addr = find_extjump_insn(ji->e[i].stub);
1396 mark_clear_cache(host_addr);
1397 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1398
1399 stat_dec(stat_links);
1400 ji->count--;
1401 if (i < ji->count) {
1402 ji->e[i] = ji->e[ji->count];
1403 continue;
1404 }
1405 i++;
1406 }
1407 }
1408}
1409
1410static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1411{
1412 int i;
1413 if (ji == NULL)
1414 return;
1415 for (i = 0; i < ji->count; ) {
1416 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1417 if (((tc_offs ^ base_offs) >> shift) != 0) {
1418 i++;
1419 continue;
1420 }
1421
1422 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1423 stat_dec(stat_links);
1424 ji->count--;
1425 if (i < ji->count) {
1426 ji->e[i] = ji->e[ji->count];
1427 continue;
1428 }
1429 i++;
1430 }
1431}
1432
1433static void invalidate_block(struct block_info *block)
1434{
1435 u_int i;
1436
1437 block->is_dirty = 1;
1438 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1439 for (i = 0; i < block->jump_in_cnt; i++)
1440 hash_table_remove(block->jump_in[i].vaddr);
1441}
1442
1443static int invalidate_range(u_int start, u_int end,
1444 u32 *inv_start_ret, u32 *inv_end_ret)
1445{
1446 struct block_info *last_block = NULL;
1447 u_int start_page = get_page_prev(start);
1448 u_int end_page = get_page(end - 1);
1449 u_int start_m = pmmask(start);
1450 u_int end_m = pmmask(end - 1);
1451 u_int inv_start, inv_end;
1452 u_int blk_start_m, blk_end_m;
1453 u_int page;
1454 int hit = 0;
1455
1456 // additional area without code (to supplement invalid_code[]), [start, end)
1457 // avoids excessive ndrc_write_invalidate*() calls
1458 inv_start = start_m & ~0xfff;
1459 inv_end = end_m | 0xfff;
1460
1461 for (page = start_page; page <= end_page; page++) {
1462 struct block_info *block;
1463 for (block = blocks[page]; block != NULL; block = block->next) {
1464 if (block->is_dirty)
1465 continue;
1466 last_block = block;
1467 blk_end_m = pmmask(block->start + block->len);
1468 if (blk_end_m <= start_m) {
1469 inv_start = max(inv_start, blk_end_m);
1470 continue;
1471 }
1472 blk_start_m = pmmask(block->start);
1473 if (end_m <= blk_start_m) {
1474 inv_end = min(inv_end, blk_start_m - 1);
1475 continue;
1476 }
1477 if (!block->source) // "hack" block - leave it alone
1478 continue;
1479
1480 hit++;
1481 invalidate_block(block);
1482 stat_inc(stat_inv_hits);
1483 }
1484 }
1485
1486 if (!hit && last_block && last_block->source) {
1487 // could be some leftover unused block, uselessly trapping writes
1488 last_block->inv_near_misses++;
1489 if (last_block->inv_near_misses > 128) {
1490 invalidate_block(last_block);
1491 stat_inc(stat_inv_hits);
1492 hit++;
1493 }
1494 }
1495 if (hit) {
1496 do_clear_cache();
1497#ifdef USE_MINI_HT
1498 memset(mini_ht, -1, sizeof(mini_ht));
1499#endif
1500 }
1501
1502 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1503 // the whole page is empty now
1504 mark_invalid_code(start, 1, 1);
1505
1506 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1507 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1508 return hit;
1509}
1510
1511void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1512{
1513 invalidate_range(start, end, NULL, NULL);
1514}
1515
1516static void ndrc_write_invalidate_many(u_int start, u_int end)
1517{
1518 // this check is done by the caller
1519 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1520 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1521#ifdef INV_DEBUG_W
1522 int invc = invalid_code[start >> 12];
1523 u_int len = end - start;
1524 if (ret)
1525 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1526 else
1527 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1528 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1529 check_for_block_changes(start, end);
1530#endif
1531 stat_inc(stat_inv_addr_calls);
1532 (void)ret;
1533}
1534
1535void ndrc_write_invalidate_one(u_int addr)
1536{
1537 ndrc_write_invalidate_many(addr, addr + 4);
1538}
1539
1540// This is called when loading a save state.
1541// Anything could have changed, so invalidate everything.
1542void new_dynarec_invalidate_all_pages(void)
1543{
1544 struct block_info *block;
1545 u_int page;
1546 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1547 for (block = blocks[page]; block != NULL; block = block->next) {
1548 if (block->is_dirty)
1549 continue;
1550 if (!block->source) // hack block?
1551 continue;
1552 invalidate_block(block);
1553 }
1554 }
1555
1556 #ifdef USE_MINI_HT
1557 memset(mini_ht, -1, sizeof(mini_ht));
1558 #endif
1559 do_clear_cache();
1560}
1561
1562// Add an entry to jump_out after making a link
1563// src should point to code by emit_extjump()
1564void ndrc_add_jump_out(u_int vaddr, void *src)
1565{
1566 inv_debug("ndrc_add_jump_out: %p -> %x\n", src, vaddr);
1567 u_int page = get_page(vaddr);
1568 struct jump_info *ji;
1569
1570 stat_inc(stat_links);
1571 check_extjump2(src);
1572 ji = jumps[page];
1573 if (ji == NULL) {
1574 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1575 ji->alloc = 16;
1576 ji->count = 0;
1577 }
1578 else if (ji->count >= ji->alloc) {
1579 ji->alloc += 16;
1580 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1581 }
1582 jumps[page] = ji;
1583 ji->e[ji->count].target_vaddr = vaddr;
1584 ji->e[ji->count].stub = src;
1585 ji->count++;
1586}
1587
1588/* Register allocation */
1589
1590// Note: registers are allocated clean (unmodified state)
1591// if you intend to modify the register, you must call dirty_reg().
1592static void alloc_reg(struct regstat *cur,int i,signed char reg)
1593{
1594 int r,hr;
1595 int preferred_reg = PREFERRED_REG_FIRST
1596 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1597 if (reg == CCREG) preferred_reg = HOST_CCREG;
1598 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1599 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1600 assert(reg >= 0);
1601
1602 // Don't allocate unused registers
1603 if((cur->u>>reg)&1) return;
1604
1605 // see if it's already allocated
1606 if (get_reg(cur->regmap, reg) >= 0)
1607 return;
1608
1609 // Keep the same mapping if the register was already allocated in a loop
1610 preferred_reg = loop_reg(i,reg,preferred_reg);
1611
1612 // Try to allocate the preferred register
1613 if(cur->regmap[preferred_reg]==-1) {
1614 cur->regmap[preferred_reg]=reg;
1615 cur->dirty&=~(1<<preferred_reg);
1616 cur->isconst&=~(1<<preferred_reg);
1617 return;
1618 }
1619 r=cur->regmap[preferred_reg];
1620 assert(r < 64);
1621 if((cur->u>>r)&1) {
1622 cur->regmap[preferred_reg]=reg;
1623 cur->dirty&=~(1<<preferred_reg);
1624 cur->isconst&=~(1<<preferred_reg);
1625 return;
1626 }
1627
1628 // Clear any unneeded registers
1629 // We try to keep the mapping consistent, if possible, because it
1630 // makes branches easier (especially loops). So we try to allocate
1631 // first (see above) before removing old mappings. If this is not
1632 // possible then go ahead and clear out the registers that are no
1633 // longer needed.
1634 for(hr=0;hr<HOST_REGS;hr++)
1635 {
1636 r=cur->regmap[hr];
1637 if(r>=0) {
1638 assert(r < 64);
1639 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1640 }
1641 }
1642
1643 // Try to allocate any available register, but prefer
1644 // registers that have not been used recently.
1645 if (i > 0) {
1646 for (hr = PREFERRED_REG_FIRST; ; ) {
1647 if (cur->regmap[hr] < 0) {
1648 int oldreg = regs[i-1].regmap[hr];
1649 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1650 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1651 {
1652 cur->regmap[hr]=reg;
1653 cur->dirty&=~(1<<hr);
1654 cur->isconst&=~(1<<hr);
1655 return;
1656 }
1657 }
1658 hr++;
1659 if (hr == EXCLUDE_REG)
1660 hr++;
1661 if (hr == HOST_REGS)
1662 hr = 0;
1663 if (hr == PREFERRED_REG_FIRST)
1664 break;
1665 }
1666 }
1667
1668 // Try to allocate any available register
1669 for (hr = PREFERRED_REG_FIRST; ; ) {
1670 if (cur->regmap[hr] < 0) {
1671 cur->regmap[hr]=reg;
1672 cur->dirty&=~(1<<hr);
1673 cur->isconst&=~(1<<hr);
1674 return;
1675 }
1676 hr++;
1677 if (hr == EXCLUDE_REG)
1678 hr++;
1679 if (hr == HOST_REGS)
1680 hr = 0;
1681 if (hr == PREFERRED_REG_FIRST)
1682 break;
1683 }
1684
1685 // Ok, now we have to evict someone
1686 // Pick a register we hopefully won't need soon
1687 u_char hsn[MAXREG+1];
1688 memset(hsn,10,sizeof(hsn));
1689 int j;
1690 lsn(hsn,i,&preferred_reg);
1691 //printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",cur->regmap[0],cur->regmap[1],cur->regmap[2],cur->regmap[3],cur->regmap[5],cur->regmap[6],cur->regmap[7]);
1692 //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]);
1693 if(i>0) {
1694 // Don't evict the cycle count at entry points, otherwise the entry
1695 // stub will have to write it.
1696 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1697 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1698 for(j=10;j>=3;j--)
1699 {
1700 // Alloc preferred register if available
1701 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1702 for(hr=0;hr<HOST_REGS;hr++) {
1703 // Evict both parts of a 64-bit register
1704 if(cur->regmap[hr]==r) {
1705 cur->regmap[hr]=-1;
1706 cur->dirty&=~(1<<hr);
1707 cur->isconst&=~(1<<hr);
1708 }
1709 }
1710 cur->regmap[preferred_reg]=reg;
1711 return;
1712 }
1713 for(r=1;r<=MAXREG;r++)
1714 {
1715 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1716 for(hr=0;hr<HOST_REGS;hr++) {
1717 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1718 if(cur->regmap[hr]==r) {
1719 cur->regmap[hr]=reg;
1720 cur->dirty&=~(1<<hr);
1721 cur->isconst&=~(1<<hr);
1722 return;
1723 }
1724 }
1725 }
1726 }
1727 }
1728 }
1729 }
1730 for(j=10;j>=0;j--)
1731 {
1732 for(r=1;r<=MAXREG;r++)
1733 {
1734 if(hsn[r]==j) {
1735 for(hr=0;hr<HOST_REGS;hr++) {
1736 if(cur->regmap[hr]==r) {
1737 cur->regmap[hr]=reg;
1738 cur->dirty&=~(1<<hr);
1739 cur->isconst&=~(1<<hr);
1740 return;
1741 }
1742 }
1743 }
1744 }
1745 }
1746 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1747}
1748
1749// Allocate a temporary register. This is done without regard to
1750// dirty status or whether the register we request is on the unneeded list
1751// Note: This will only allocate one register, even if called multiple times
1752static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1753{
1754 int r,hr;
1755 int preferred_reg = -1;
1756
1757 // see if it's already allocated
1758 for(hr=0;hr<HOST_REGS;hr++)
1759 {
1760 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1761 }
1762
1763 // Try to allocate any available register
1764 for(hr=HOST_REGS-1;hr>=0;hr--) {
1765 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1766 cur->regmap[hr]=reg;
1767 cur->dirty&=~(1<<hr);
1768 cur->isconst&=~(1<<hr);
1769 return;
1770 }
1771 }
1772
1773 // Find an unneeded register
1774 for(hr=HOST_REGS-1;hr>=0;hr--)
1775 {
1776 r=cur->regmap[hr];
1777 if(r>=0) {
1778 assert(r < 64);
1779 if((cur->u>>r)&1) {
1780 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1781 cur->regmap[hr]=reg;
1782 cur->dirty&=~(1<<hr);
1783 cur->isconst&=~(1<<hr);
1784 return;
1785 }
1786 }
1787 }
1788 }
1789
1790 // Ok, now we have to evict someone
1791 // Pick a register we hopefully won't need soon
1792 // TODO: we might want to follow unconditional jumps here
1793 // TODO: get rid of dupe code and make this into a function
1794 u_char hsn[MAXREG+1];
1795 memset(hsn,10,sizeof(hsn));
1796 int j;
1797 lsn(hsn,i,&preferred_reg);
1798 //printf("hsn: %d %d %d %d %d %d %d\n",hsn[cur->regmap[0]&63],hsn[cur->regmap[1]&63],hsn[cur->regmap[2]&63],hsn[cur->regmap[3]&63],hsn[cur->regmap[5]&63],hsn[cur->regmap[6]&63],hsn[cur->regmap[7]&63]);
1799 if(i>0) {
1800 // Don't evict the cycle count at entry points, otherwise the entry
1801 // stub will have to write it.
1802 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1803 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1804 for(j=10;j>=3;j--)
1805 {
1806 for(r=1;r<=MAXREG;r++)
1807 {
1808 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1809 for(hr=0;hr<HOST_REGS;hr++) {
1810 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1811 if(cur->regmap[hr]==r) {
1812 cur->regmap[hr]=reg;
1813 cur->dirty&=~(1<<hr);
1814 cur->isconst&=~(1<<hr);
1815 return;
1816 }
1817 }
1818 }
1819 }
1820 }
1821 }
1822 }
1823 for(j=10;j>=0;j--)
1824 {
1825 for(r=1;r<=MAXREG;r++)
1826 {
1827 if(hsn[r]==j) {
1828 for(hr=0;hr<HOST_REGS;hr++) {
1829 if(cur->regmap[hr]==r) {
1830 cur->regmap[hr]=reg;
1831 cur->dirty&=~(1<<hr);
1832 cur->isconst&=~(1<<hr);
1833 return;
1834 }
1835 }
1836 }
1837 }
1838 }
1839 SysPrintf("This shouldn't happen");abort();
1840}
1841
1842static void mov_alloc(struct regstat *current,int i)
1843{
1844 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1845 alloc_cc(current,i); // for stalls
1846 dirty_reg(current,CCREG);
1847 }
1848
1849 // Note: Don't need to actually alloc the source registers
1850 //alloc_reg(current,i,dops[i].rs1);
1851 alloc_reg(current,i,dops[i].rt1);
1852
1853 clear_const(current,dops[i].rs1);
1854 clear_const(current,dops[i].rt1);
1855 dirty_reg(current,dops[i].rt1);
1856}
1857
1858static void shiftimm_alloc(struct regstat *current,int i)
1859{
1860 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1861 {
1862 if(dops[i].rt1) {
1863 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1864 else dops[i].use_lt1=!!dops[i].rs1;
1865 alloc_reg(current,i,dops[i].rt1);
1866 dirty_reg(current,dops[i].rt1);
1867 if(is_const(current,dops[i].rs1)) {
1868 int v=get_const(current,dops[i].rs1);
1869 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1870 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1871 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1872 }
1873 else clear_const(current,dops[i].rt1);
1874 }
1875 }
1876 else
1877 {
1878 clear_const(current,dops[i].rs1);
1879 clear_const(current,dops[i].rt1);
1880 }
1881
1882 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1883 {
1884 assert(0);
1885 }
1886 if(dops[i].opcode2==0x3c) // DSLL32
1887 {
1888 assert(0);
1889 }
1890 if(dops[i].opcode2==0x3e) // DSRL32
1891 {
1892 assert(0);
1893 }
1894 if(dops[i].opcode2==0x3f) // DSRA32
1895 {
1896 assert(0);
1897 }
1898}
1899
1900static void shift_alloc(struct regstat *current,int i)
1901{
1902 if(dops[i].rt1) {
1903 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1904 {
1905 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1906 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1907 alloc_reg(current,i,dops[i].rt1);
1908 if(dops[i].rt1==dops[i].rs2) {
1909 alloc_reg_temp(current,i,-1);
1910 minimum_free_regs[i]=1;
1911 }
1912 } else { // DSLLV/DSRLV/DSRAV
1913 assert(0);
1914 }
1915 clear_const(current,dops[i].rs1);
1916 clear_const(current,dops[i].rs2);
1917 clear_const(current,dops[i].rt1);
1918 dirty_reg(current,dops[i].rt1);
1919 }
1920}
1921
1922static void alu_alloc(struct regstat *current,int i)
1923{
1924 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1925 if(dops[i].rt1) {
1926 if(dops[i].rs1&&dops[i].rs2) {
1927 alloc_reg(current,i,dops[i].rs1);
1928 alloc_reg(current,i,dops[i].rs2);
1929 }
1930 else {
1931 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1932 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1933 }
1934 alloc_reg(current,i,dops[i].rt1);
1935 }
1936 }
1937 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1938 if(dops[i].rt1) {
1939 alloc_reg(current,i,dops[i].rs1);
1940 alloc_reg(current,i,dops[i].rs2);
1941 alloc_reg(current,i,dops[i].rt1);
1942 }
1943 }
1944 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1945 if(dops[i].rt1) {
1946 if(dops[i].rs1&&dops[i].rs2) {
1947 alloc_reg(current,i,dops[i].rs1);
1948 alloc_reg(current,i,dops[i].rs2);
1949 }
1950 else
1951 {
1952 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1953 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1954 }
1955 alloc_reg(current,i,dops[i].rt1);
1956 }
1957 }
1958 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1959 assert(0);
1960 }
1961 clear_const(current,dops[i].rs1);
1962 clear_const(current,dops[i].rs2);
1963 clear_const(current,dops[i].rt1);
1964 dirty_reg(current,dops[i].rt1);
1965}
1966
1967static void imm16_alloc(struct regstat *current,int i)
1968{
1969 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1970 else dops[i].use_lt1=!!dops[i].rs1;
1971 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1972 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1973 assert(0);
1974 }
1975 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1976 clear_const(current,dops[i].rs1);
1977 clear_const(current,dops[i].rt1);
1978 }
1979 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1980 if(is_const(current,dops[i].rs1)) {
1981 int v=get_const(current,dops[i].rs1);
1982 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1983 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1984 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1985 }
1986 else clear_const(current,dops[i].rt1);
1987 }
1988 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1989 if(is_const(current,dops[i].rs1)) {
1990 int v=get_const(current,dops[i].rs1);
1991 set_const(current,dops[i].rt1,v+imm[i]);
1992 }
1993 else clear_const(current,dops[i].rt1);
1994 }
1995 else {
1996 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1997 }
1998 dirty_reg(current,dops[i].rt1);
1999}
2000
2001static void load_alloc(struct regstat *current,int i)
2002{
2003 clear_const(current,dops[i].rt1);
2004 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2005 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2006 if (needed_again(dops[i].rs1, i))
2007 alloc_reg(current, i, dops[i].rs1);
2008 if (ram_offset)
2009 alloc_reg(current, i, ROREG);
2010 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2011 alloc_reg(current,i,dops[i].rt1);
2012 assert(get_reg(current->regmap,dops[i].rt1)>=0);
2013 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
2014 {
2015 assert(0);
2016 }
2017 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
2018 {
2019 assert(0);
2020 }
2021 dirty_reg(current,dops[i].rt1);
2022 // LWL/LWR need a temporary register for the old value
2023 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2024 {
2025 alloc_reg(current,i,FTEMP);
2026 alloc_reg_temp(current,i,-1);
2027 minimum_free_regs[i]=1;
2028 }
2029 }
2030 else
2031 {
2032 // Load to r0 or unneeded register (dummy load)
2033 // but we still need a register to calculate the address
2034 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2035 {
2036 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2037 }
2038 alloc_reg_temp(current,i,-1);
2039 minimum_free_regs[i]=1;
2040 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
2041 {
2042 assert(0);
2043 }
2044 }
2045}
2046
2047static void store_alloc(struct regstat *current,int i)
2048{
2049 clear_const(current,dops[i].rs2);
2050 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2051 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2052 alloc_reg(current,i,dops[i].rs2);
2053 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
2054 assert(0);
2055 }
2056 if (ram_offset)
2057 alloc_reg(current, i, ROREG);
2058 #if defined(HOST_IMM8)
2059 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2060 alloc_reg(current, i, INVCP);
2061 #endif
2062 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
2063 alloc_reg(current,i,FTEMP);
2064 }
2065 // We need a temporary register for address generation
2066 alloc_reg_temp(current,i,-1);
2067 minimum_free_regs[i]=1;
2068}
2069
2070static void c1ls_alloc(struct regstat *current,int i)
2071{
2072 clear_const(current,dops[i].rt1);
2073 alloc_reg(current,i,CSREG); // Status
2074}
2075
2076static void c2ls_alloc(struct regstat *current,int i)
2077{
2078 clear_const(current,dops[i].rt1);
2079 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2080 alloc_reg(current,i,FTEMP);
2081 if (ram_offset)
2082 alloc_reg(current, i, ROREG);
2083 #if defined(HOST_IMM8)
2084 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2085 if (dops[i].opcode == 0x3a) // SWC2
2086 alloc_reg(current,i,INVCP);
2087 #endif
2088 // We need a temporary register for address generation
2089 alloc_reg_temp(current,i,-1);
2090 minimum_free_regs[i]=1;
2091}
2092
2093#ifndef multdiv_alloc
2094static void multdiv_alloc(struct regstat *current,int i)
2095{
2096 // case 0x18: MULT
2097 // case 0x19: MULTU
2098 // case 0x1A: DIV
2099 // case 0x1B: DIVU
2100 // case 0x1C: DMULT
2101 // case 0x1D: DMULTU
2102 // case 0x1E: DDIV
2103 // case 0x1F: DDIVU
2104 clear_const(current,dops[i].rs1);
2105 clear_const(current,dops[i].rs2);
2106 alloc_cc(current,i); // for stalls
2107 if(dops[i].rs1&&dops[i].rs2)
2108 {
2109 if((dops[i].opcode2&4)==0) // 32-bit
2110 {
2111 current->u&=~(1LL<<HIREG);
2112 current->u&=~(1LL<<LOREG);
2113 alloc_reg(current,i,HIREG);
2114 alloc_reg(current,i,LOREG);
2115 alloc_reg(current,i,dops[i].rs1);
2116 alloc_reg(current,i,dops[i].rs2);
2117 dirty_reg(current,HIREG);
2118 dirty_reg(current,LOREG);
2119 }
2120 else // 64-bit
2121 {
2122 assert(0);
2123 }
2124 }
2125 else
2126 {
2127 // Multiply by zero is zero.
2128 // MIPS does not have a divide by zero exception.
2129 // The result is undefined, we return zero.
2130 alloc_reg(current,i,HIREG);
2131 alloc_reg(current,i,LOREG);
2132 dirty_reg(current,HIREG);
2133 dirty_reg(current,LOREG);
2134 }
2135}
2136#endif
2137
2138static void cop0_alloc(struct regstat *current,int i)
2139{
2140 if(dops[i].opcode2==0) // MFC0
2141 {
2142 if(dops[i].rt1) {
2143 clear_const(current,dops[i].rt1);
2144 alloc_all(current,i);
2145 alloc_reg(current,i,dops[i].rt1);
2146 dirty_reg(current,dops[i].rt1);
2147 }
2148 }
2149 else if(dops[i].opcode2==4) // MTC0
2150 {
2151 if(dops[i].rs1){
2152 clear_const(current,dops[i].rs1);
2153 alloc_reg(current,i,dops[i].rs1);
2154 alloc_all(current,i);
2155 }
2156 else {
2157 alloc_all(current,i); // FIXME: Keep r0
2158 current->u&=~1LL;
2159 alloc_reg(current,i,0);
2160 }
2161 }
2162 else
2163 {
2164 // RFE
2165 assert(dops[i].opcode2==0x10);
2166 alloc_all(current,i);
2167 }
2168 minimum_free_regs[i]=HOST_REGS;
2169}
2170
2171static void cop2_alloc(struct regstat *current,int i)
2172{
2173 if (dops[i].opcode2 < 3) // MFC2/CFC2
2174 {
2175 alloc_cc(current,i); // for stalls
2176 dirty_reg(current,CCREG);
2177 if(dops[i].rt1){
2178 clear_const(current,dops[i].rt1);
2179 alloc_reg(current,i,dops[i].rt1);
2180 dirty_reg(current,dops[i].rt1);
2181 }
2182 }
2183 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2184 {
2185 if(dops[i].rs1){
2186 clear_const(current,dops[i].rs1);
2187 alloc_reg(current,i,dops[i].rs1);
2188 }
2189 else {
2190 current->u&=~1LL;
2191 alloc_reg(current,i,0);
2192 }
2193 }
2194 alloc_reg_temp(current,i,-1);
2195 minimum_free_regs[i]=1;
2196}
2197
2198static void c2op_alloc(struct regstat *current,int i)
2199{
2200 alloc_cc(current,i); // for stalls
2201 dirty_reg(current,CCREG);
2202 alloc_reg_temp(current,i,-1);
2203}
2204
2205static void syscall_alloc(struct regstat *current,int i)
2206{
2207 alloc_cc(current,i);
2208 dirty_reg(current,CCREG);
2209 alloc_all(current,i);
2210 minimum_free_regs[i]=HOST_REGS;
2211 current->isconst=0;
2212}
2213
2214static void delayslot_alloc(struct regstat *current,int i)
2215{
2216 switch(dops[i].itype) {
2217 case UJUMP:
2218 case CJUMP:
2219 case SJUMP:
2220 case RJUMP:
2221 case SYSCALL:
2222 case HLECALL:
2223 case IMM16:
2224 imm16_alloc(current,i);
2225 break;
2226 case LOAD:
2227 case LOADLR:
2228 load_alloc(current,i);
2229 break;
2230 case STORE:
2231 case STORELR:
2232 store_alloc(current,i);
2233 break;
2234 case ALU:
2235 alu_alloc(current,i);
2236 break;
2237 case SHIFT:
2238 shift_alloc(current,i);
2239 break;
2240 case MULTDIV:
2241 multdiv_alloc(current,i);
2242 break;
2243 case SHIFTIMM:
2244 shiftimm_alloc(current,i);
2245 break;
2246 case MOV:
2247 mov_alloc(current,i);
2248 break;
2249 case COP0:
2250 cop0_alloc(current,i);
2251 break;
2252 case COP1:
2253 break;
2254 case COP2:
2255 cop2_alloc(current,i);
2256 break;
2257 case C1LS:
2258 c1ls_alloc(current,i);
2259 break;
2260 case C2LS:
2261 c2ls_alloc(current,i);
2262 break;
2263 case C2OP:
2264 c2op_alloc(current,i);
2265 break;
2266 }
2267}
2268
2269static void add_stub(enum stub_type type, void *addr, void *retaddr,
2270 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2271{
2272 assert(stubcount < ARRAY_SIZE(stubs));
2273 stubs[stubcount].type = type;
2274 stubs[stubcount].addr = addr;
2275 stubs[stubcount].retaddr = retaddr;
2276 stubs[stubcount].a = a;
2277 stubs[stubcount].b = b;
2278 stubs[stubcount].c = c;
2279 stubs[stubcount].d = d;
2280 stubs[stubcount].e = e;
2281 stubcount++;
2282}
2283
2284static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2285 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2286{
2287 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2288}
2289
2290// Write out a single register
2291static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2292{
2293 int hr;
2294 for(hr=0;hr<HOST_REGS;hr++) {
2295 if(hr!=EXCLUDE_REG) {
2296 if(regmap[hr]==r) {
2297 if((dirty>>hr)&1) {
2298 assert(regmap[hr]<64);
2299 emit_storereg(r,hr);
2300 }
2301 }
2302 }
2303 }
2304}
2305
2306static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2307{
2308 //if(dirty_pre==dirty) return;
2309 int hr, r;
2310 for (hr = 0; hr < HOST_REGS; hr++) {
2311 r = pre[hr];
2312 if (r < 1 || r > 33 || ((u >> r) & 1))
2313 continue;
2314 if (((dirty_pre & ~dirty) >> hr) & 1)
2315 emit_storereg(r, hr);
2316 }
2317}
2318
2319// trashes r2
2320static void pass_args(int a0, int a1)
2321{
2322 if(a0==1&&a1==0) {
2323 // must swap
2324 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2325 }
2326 else if(a0!=0&&a1==0) {
2327 emit_mov(a1,1);
2328 if (a0>=0) emit_mov(a0,0);
2329 }
2330 else {
2331 if(a0>=0&&a0!=0) emit_mov(a0,0);
2332 if(a1>=0&&a1!=1) emit_mov(a1,1);
2333 }
2334}
2335
2336static void alu_assemble(int i, const struct regstat *i_regs)
2337{
2338 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2339 if(dops[i].rt1) {
2340 signed char s1,s2,t;
2341 t=get_reg(i_regs->regmap,dops[i].rt1);
2342 if(t>=0) {
2343 s1=get_reg(i_regs->regmap,dops[i].rs1);
2344 s2=get_reg(i_regs->regmap,dops[i].rs2);
2345 if(dops[i].rs1&&dops[i].rs2) {
2346 assert(s1>=0);
2347 assert(s2>=0);
2348 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2349 else emit_add(s1,s2,t);
2350 }
2351 else if(dops[i].rs1) {
2352 if(s1>=0) emit_mov(s1,t);
2353 else emit_loadreg(dops[i].rs1,t);
2354 }
2355 else if(dops[i].rs2) {
2356 if(s2>=0) {
2357 if(dops[i].opcode2&2) emit_neg(s2,t);
2358 else emit_mov(s2,t);
2359 }
2360 else {
2361 emit_loadreg(dops[i].rs2,t);
2362 if(dops[i].opcode2&2) emit_neg(t,t);
2363 }
2364 }
2365 else emit_zeroreg(t);
2366 }
2367 }
2368 }
2369 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2370 assert(0);
2371 }
2372 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2373 if(dops[i].rt1) {
2374 signed char s1l,s2l,t;
2375 {
2376 t=get_reg(i_regs->regmap,dops[i].rt1);
2377 //assert(t>=0);
2378 if(t>=0) {
2379 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2380 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2381 if(dops[i].rs2==0) // rx<r0
2382 {
2383 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2384 assert(s1l>=0);
2385 emit_shrimm(s1l,31,t);
2386 }
2387 else // SLTU (unsigned can not be less than zero, 0<0)
2388 emit_zeroreg(t);
2389 }
2390 else if(dops[i].rs1==0) // r0<rx
2391 {
2392 assert(s2l>=0);
2393 if(dops[i].opcode2==0x2a) // SLT
2394 emit_set_gz32(s2l,t);
2395 else // SLTU (set if not zero)
2396 emit_set_nz32(s2l,t);
2397 }
2398 else{
2399 assert(s1l>=0);assert(s2l>=0);
2400 if(dops[i].opcode2==0x2a) // SLT
2401 emit_set_if_less32(s1l,s2l,t);
2402 else // SLTU
2403 emit_set_if_carry32(s1l,s2l,t);
2404 }
2405 }
2406 }
2407 }
2408 }
2409 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2410 if(dops[i].rt1) {
2411 signed char s1l,s2l,tl;
2412 tl=get_reg(i_regs->regmap,dops[i].rt1);
2413 {
2414 if(tl>=0) {
2415 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2416 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2417 if(dops[i].rs1&&dops[i].rs2) {
2418 assert(s1l>=0);
2419 assert(s2l>=0);
2420 if(dops[i].opcode2==0x24) { // AND
2421 emit_and(s1l,s2l,tl);
2422 } else
2423 if(dops[i].opcode2==0x25) { // OR
2424 emit_or(s1l,s2l,tl);
2425 } else
2426 if(dops[i].opcode2==0x26) { // XOR
2427 emit_xor(s1l,s2l,tl);
2428 } else
2429 if(dops[i].opcode2==0x27) { // NOR
2430 emit_or(s1l,s2l,tl);
2431 emit_not(tl,tl);
2432 }
2433 }
2434 else
2435 {
2436 if(dops[i].opcode2==0x24) { // AND
2437 emit_zeroreg(tl);
2438 } else
2439 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2440 if(dops[i].rs1){
2441 if(s1l>=0) emit_mov(s1l,tl);
2442 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2443 }
2444 else
2445 if(dops[i].rs2){
2446 if(s2l>=0) emit_mov(s2l,tl);
2447 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2448 }
2449 else emit_zeroreg(tl);
2450 } else
2451 if(dops[i].opcode2==0x27) { // NOR
2452 if(dops[i].rs1){
2453 if(s1l>=0) emit_not(s1l,tl);
2454 else {
2455 emit_loadreg(dops[i].rs1,tl);
2456 emit_not(tl,tl);
2457 }
2458 }
2459 else
2460 if(dops[i].rs2){
2461 if(s2l>=0) emit_not(s2l,tl);
2462 else {
2463 emit_loadreg(dops[i].rs2,tl);
2464 emit_not(tl,tl);
2465 }
2466 }
2467 else emit_movimm(-1,tl);
2468 }
2469 }
2470 }
2471 }
2472 }
2473 }
2474}
2475
2476static void imm16_assemble(int i, const struct regstat *i_regs)
2477{
2478 if (dops[i].opcode==0x0f) { // LUI
2479 if(dops[i].rt1) {
2480 signed char t;
2481 t=get_reg(i_regs->regmap,dops[i].rt1);
2482 //assert(t>=0);
2483 if(t>=0) {
2484 if(!((i_regs->isconst>>t)&1))
2485 emit_movimm(imm[i]<<16,t);
2486 }
2487 }
2488 }
2489 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2490 if(dops[i].rt1) {
2491 signed char s,t;
2492 t=get_reg(i_regs->regmap,dops[i].rt1);
2493 s=get_reg(i_regs->regmap,dops[i].rs1);
2494 if(dops[i].rs1) {
2495 //assert(t>=0);
2496 //assert(s>=0);
2497 if(t>=0) {
2498 if(!((i_regs->isconst>>t)&1)) {
2499 if(s<0) {
2500 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2501 emit_addimm(t,imm[i],t);
2502 }else{
2503 if(!((i_regs->wasconst>>s)&1))
2504 emit_addimm(s,imm[i],t);
2505 else
2506 emit_movimm(constmap[i][s]+imm[i],t);
2507 }
2508 }
2509 }
2510 } else {
2511 if(t>=0) {
2512 if(!((i_regs->isconst>>t)&1))
2513 emit_movimm(imm[i],t);
2514 }
2515 }
2516 }
2517 }
2518 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2519 if(dops[i].rt1) {
2520 signed char sl,tl;
2521 tl=get_reg(i_regs->regmap,dops[i].rt1);
2522 sl=get_reg(i_regs->regmap,dops[i].rs1);
2523 if(tl>=0) {
2524 if(dops[i].rs1) {
2525 assert(sl>=0);
2526 emit_addimm(sl,imm[i],tl);
2527 } else {
2528 emit_movimm(imm[i],tl);
2529 }
2530 }
2531 }
2532 }
2533 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2534 if(dops[i].rt1) {
2535 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2536 signed char sl,t;
2537 t=get_reg(i_regs->regmap,dops[i].rt1);
2538 sl=get_reg(i_regs->regmap,dops[i].rs1);
2539 //assert(t>=0);
2540 if(t>=0) {
2541 if(dops[i].rs1>0) {
2542 if(dops[i].opcode==0x0a) { // SLTI
2543 if(sl<0) {
2544 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2545 emit_slti32(t,imm[i],t);
2546 }else{
2547 emit_slti32(sl,imm[i],t);
2548 }
2549 }
2550 else { // SLTIU
2551 if(sl<0) {
2552 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2553 emit_sltiu32(t,imm[i],t);
2554 }else{
2555 emit_sltiu32(sl,imm[i],t);
2556 }
2557 }
2558 }else{
2559 // SLTI(U) with r0 is just stupid,
2560 // nonetheless examples can be found
2561 if(dops[i].opcode==0x0a) // SLTI
2562 if(0<imm[i]) emit_movimm(1,t);
2563 else emit_zeroreg(t);
2564 else // SLTIU
2565 {
2566 if(imm[i]) emit_movimm(1,t);
2567 else emit_zeroreg(t);
2568 }
2569 }
2570 }
2571 }
2572 }
2573 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2574 if(dops[i].rt1) {
2575 signed char sl,tl;
2576 tl=get_reg(i_regs->regmap,dops[i].rt1);
2577 sl=get_reg(i_regs->regmap,dops[i].rs1);
2578 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2579 if(dops[i].opcode==0x0c) //ANDI
2580 {
2581 if(dops[i].rs1) {
2582 if(sl<0) {
2583 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2584 emit_andimm(tl,imm[i],tl);
2585 }else{
2586 if(!((i_regs->wasconst>>sl)&1))
2587 emit_andimm(sl,imm[i],tl);
2588 else
2589 emit_movimm(constmap[i][sl]&imm[i],tl);
2590 }
2591 }
2592 else
2593 emit_zeroreg(tl);
2594 }
2595 else
2596 {
2597 if(dops[i].rs1) {
2598 if(sl<0) {
2599 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2600 }
2601 if(dops[i].opcode==0x0d) { // ORI
2602 if(sl<0) {
2603 emit_orimm(tl,imm[i],tl);
2604 }else{
2605 if(!((i_regs->wasconst>>sl)&1))
2606 emit_orimm(sl,imm[i],tl);
2607 else
2608 emit_movimm(constmap[i][sl]|imm[i],tl);
2609 }
2610 }
2611 if(dops[i].opcode==0x0e) { // XORI
2612 if(sl<0) {
2613 emit_xorimm(tl,imm[i],tl);
2614 }else{
2615 if(!((i_regs->wasconst>>sl)&1))
2616 emit_xorimm(sl,imm[i],tl);
2617 else
2618 emit_movimm(constmap[i][sl]^imm[i],tl);
2619 }
2620 }
2621 }
2622 else {
2623 emit_movimm(imm[i],tl);
2624 }
2625 }
2626 }
2627 }
2628 }
2629}
2630
2631static void shiftimm_assemble(int i, const struct regstat *i_regs)
2632{
2633 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2634 {
2635 if(dops[i].rt1) {
2636 signed char s,t;
2637 t=get_reg(i_regs->regmap,dops[i].rt1);
2638 s=get_reg(i_regs->regmap,dops[i].rs1);
2639 //assert(t>=0);
2640 if(t>=0&&!((i_regs->isconst>>t)&1)){
2641 if(dops[i].rs1==0)
2642 {
2643 emit_zeroreg(t);
2644 }
2645 else
2646 {
2647 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2648 if(imm[i]) {
2649 if(dops[i].opcode2==0) // SLL
2650 {
2651 emit_shlimm(s<0?t:s,imm[i],t);
2652 }
2653 if(dops[i].opcode2==2) // SRL
2654 {
2655 emit_shrimm(s<0?t:s,imm[i],t);
2656 }
2657 if(dops[i].opcode2==3) // SRA
2658 {
2659 emit_sarimm(s<0?t:s,imm[i],t);
2660 }
2661 }else{
2662 // Shift by zero
2663 if(s>=0 && s!=t) emit_mov(s,t);
2664 }
2665 }
2666 }
2667 //emit_storereg(dops[i].rt1,t); //DEBUG
2668 }
2669 }
2670 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2671 {
2672 assert(0);
2673 }
2674 if(dops[i].opcode2==0x3c) // DSLL32
2675 {
2676 assert(0);
2677 }
2678 if(dops[i].opcode2==0x3e) // DSRL32
2679 {
2680 assert(0);
2681 }
2682 if(dops[i].opcode2==0x3f) // DSRA32
2683 {
2684 assert(0);
2685 }
2686}
2687
2688#ifndef shift_assemble
2689static void shift_assemble(int i, const struct regstat *i_regs)
2690{
2691 signed char s,t,shift;
2692 if (dops[i].rt1 == 0)
2693 return;
2694 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2695 t = get_reg(i_regs->regmap, dops[i].rt1);
2696 s = get_reg(i_regs->regmap, dops[i].rs1);
2697 shift = get_reg(i_regs->regmap, dops[i].rs2);
2698 if (t < 0)
2699 return;
2700
2701 if(dops[i].rs1==0)
2702 emit_zeroreg(t);
2703 else if(dops[i].rs2==0) {
2704 assert(s>=0);
2705 if(s!=t) emit_mov(s,t);
2706 }
2707 else {
2708 host_tempreg_acquire();
2709 emit_andimm(shift,31,HOST_TEMPREG);
2710 switch(dops[i].opcode2) {
2711 case 4: // SLLV
2712 emit_shl(s,HOST_TEMPREG,t);
2713 break;
2714 case 6: // SRLV
2715 emit_shr(s,HOST_TEMPREG,t);
2716 break;
2717 case 7: // SRAV
2718 emit_sar(s,HOST_TEMPREG,t);
2719 break;
2720 default:
2721 assert(0);
2722 }
2723 host_tempreg_release();
2724 }
2725}
2726
2727#endif
2728
2729enum {
2730 MTYPE_8000 = 0,
2731 MTYPE_8020,
2732 MTYPE_0000,
2733 MTYPE_A000,
2734 MTYPE_1F80,
2735};
2736
2737static int get_ptr_mem_type(u_int a)
2738{
2739 if(a < 0x00200000) {
2740 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2741 // return wrong, must use memhandler for BIOS self-test to pass
2742 // 007 does similar stuff from a00 mirror, weird stuff
2743 return MTYPE_8000;
2744 return MTYPE_0000;
2745 }
2746 if(0x1f800000 <= a && a < 0x1f801000)
2747 return MTYPE_1F80;
2748 if(0x80200000 <= a && a < 0x80800000)
2749 return MTYPE_8020;
2750 if(0xa0000000 <= a && a < 0xa0200000)
2751 return MTYPE_A000;
2752 return MTYPE_8000;
2753}
2754
2755static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2756{
2757 int r = get_reg(i_regs->regmap, ROREG);
2758 if (r < 0 && host_tempreg_free) {
2759 host_tempreg_acquire();
2760 emit_loadreg(ROREG, r = HOST_TEMPREG);
2761 }
2762 if (r < 0)
2763 abort();
2764 return r;
2765}
2766
2767static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2768 int addr, int *offset_reg, int *addr_reg_override)
2769{
2770 void *jaddr = NULL;
2771 int type = 0;
2772 int mr = dops[i].rs1;
2773 *offset_reg = -1;
2774 if(((smrv_strong|smrv_weak)>>mr)&1) {
2775 type=get_ptr_mem_type(smrv[mr]);
2776 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2777 }
2778 else {
2779 // use the mirror we are running on
2780 type=get_ptr_mem_type(start);
2781 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2782 }
2783
2784 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2785 host_tempreg_acquire();
2786 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2787 addr=*addr_reg_override=HOST_TEMPREG;
2788 type=0;
2789 }
2790 else if(type==MTYPE_0000) { // RAM 0 mirror
2791 host_tempreg_acquire();
2792 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2793 addr=*addr_reg_override=HOST_TEMPREG;
2794 type=0;
2795 }
2796 else if(type==MTYPE_A000) { // RAM A mirror
2797 host_tempreg_acquire();
2798 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2799 addr=*addr_reg_override=HOST_TEMPREG;
2800 type=0;
2801 }
2802 else if(type==MTYPE_1F80) { // scratchpad
2803 if (psxH == (void *)0x1f800000) {
2804 host_tempreg_acquire();
2805 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2806 emit_cmpimm(HOST_TEMPREG,0x1000);
2807 host_tempreg_release();
2808 jaddr=out;
2809 emit_jc(0);
2810 }
2811 else {
2812 // do the usual RAM check, jump will go to the right handler
2813 type=0;
2814 }
2815 }
2816
2817 if (type == 0) // need ram check
2818 {
2819 emit_cmpimm(addr,RAM_SIZE);
2820 jaddr = out;
2821 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2822 // Hint to branch predictor that the branch is unlikely to be taken
2823 if (dops[i].rs1 >= 28)
2824 emit_jno_unlikely(0);
2825 else
2826 #endif
2827 emit_jno(0);
2828 if (ram_offset != 0)
2829 *offset_reg = get_ro_reg(i_regs, 0);
2830 }
2831
2832 return jaddr;
2833}
2834
2835// return memhandler, or get directly accessable address and return 0
2836static void *get_direct_memhandler(void *table, u_int addr,
2837 enum stub_type type, uintptr_t *addr_host)
2838{
2839 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2840 uintptr_t l1, l2 = 0;
2841 l1 = ((uintptr_t *)table)[addr>>12];
2842 if (!(l1 & msb)) {
2843 uintptr_t v = l1 << 1;
2844 *addr_host = v + addr;
2845 return NULL;
2846 }
2847 else {
2848 l1 <<= 1;
2849 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2850 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2851 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2852 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2853 else
2854 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2855 if (!(l2 & msb)) {
2856 uintptr_t v = l2 << 1;
2857 *addr_host = v + (addr&0xfff);
2858 return NULL;
2859 }
2860 return (void *)(l2 << 1);
2861 }
2862}
2863
2864static u_int get_host_reglist(const signed char *regmap)
2865{
2866 u_int reglist = 0, hr;
2867 for (hr = 0; hr < HOST_REGS; hr++) {
2868 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2869 reglist |= 1 << hr;
2870 }
2871 return reglist;
2872}
2873
2874static u_int reglist_exclude(u_int reglist, int r1, int r2)
2875{
2876 if (r1 >= 0)
2877 reglist &= ~(1u << r1);
2878 if (r2 >= 0)
2879 reglist &= ~(1u << r2);
2880 return reglist;
2881}
2882
2883// find a temp caller-saved register not in reglist (so assumed to be free)
2884static int reglist_find_free(u_int reglist)
2885{
2886 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2887 if (free_regs == 0)
2888 return -1;
2889 return __builtin_ctz(free_regs);
2890}
2891
2892static void do_load_word(int a, int rt, int offset_reg)
2893{
2894 if (offset_reg >= 0)
2895 emit_ldr_dualindexed(offset_reg, a, rt);
2896 else
2897 emit_readword_indexed(0, a, rt);
2898}
2899
2900static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2901{
2902 if (offset_reg < 0) {
2903 emit_writeword_indexed(rt, ofs, a);
2904 return;
2905 }
2906 if (ofs != 0)
2907 emit_addimm(a, ofs, a);
2908 emit_str_dualindexed(offset_reg, a, rt);
2909 if (ofs != 0 && preseve_a)
2910 emit_addimm(a, -ofs, a);
2911}
2912
2913static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2914{
2915 if (offset_reg < 0) {
2916 emit_writehword_indexed(rt, ofs, a);
2917 return;
2918 }
2919 if (ofs != 0)
2920 emit_addimm(a, ofs, a);
2921 emit_strh_dualindexed(offset_reg, a, rt);
2922 if (ofs != 0 && preseve_a)
2923 emit_addimm(a, -ofs, a);
2924}
2925
2926static void do_store_byte(int a, int rt, int offset_reg)
2927{
2928 if (offset_reg >= 0)
2929 emit_strb_dualindexed(offset_reg, a, rt);
2930 else
2931 emit_writebyte_indexed(rt, 0, a);
2932}
2933
2934static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2935{
2936 int s,tl,addr;
2937 int offset;
2938 void *jaddr=0;
2939 int memtarget=0,c=0;
2940 int offset_reg = -1;
2941 int fastio_reg_override = -1;
2942 u_int reglist=get_host_reglist(i_regs->regmap);
2943 tl=get_reg(i_regs->regmap,dops[i].rt1);
2944 s=get_reg(i_regs->regmap,dops[i].rs1);
2945 offset=imm[i];
2946 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2947 if(s>=0) {
2948 c=(i_regs->wasconst>>s)&1;
2949 if (c) {
2950 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2951 }
2952 }
2953 //printf("load_assemble: c=%d\n",c);
2954 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2955 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2956 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2957 ||dops[i].rt1==0) {
2958 // could be FIFO, must perform the read
2959 // ||dummy read
2960 assem_debug("(forced read)\n");
2961 tl=get_reg_temp(i_regs->regmap);
2962 assert(tl>=0);
2963 }
2964 if(offset||s<0||c) addr=tl;
2965 else addr=s;
2966 //if(tl<0) tl=get_reg_temp(i_regs->regmap);
2967 if(tl>=0) {
2968 //printf("load_assemble: c=%d\n",c);
2969 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2970 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2971 reglist&=~(1<<tl);
2972 if(!c) {
2973 #ifdef R29_HACK
2974 // Strmnnrmn's speed hack
2975 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2976 #endif
2977 {
2978 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
2979 &offset_reg, &fastio_reg_override);
2980 }
2981 }
2982 else if (ram_offset && memtarget) {
2983 offset_reg = get_ro_reg(i_regs, 0);
2984 }
2985 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2986 switch (dops[i].opcode) {
2987 case 0x20: // LB
2988 if(!c||memtarget) {
2989 if(!dummy) {
2990 int a = tl;
2991 if (!c) a = addr;
2992 if (fastio_reg_override >= 0)
2993 a = fastio_reg_override;
2994
2995 if (offset_reg >= 0)
2996 emit_ldrsb_dualindexed(offset_reg, a, tl);
2997 else
2998 emit_movsbl_indexed(0, a, tl);
2999 }
3000 if(jaddr)
3001 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3002 }
3003 else
3004 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3005 break;
3006 case 0x21: // LH
3007 if(!c||memtarget) {
3008 if(!dummy) {
3009 int a = tl;
3010 if (!c) a = addr;
3011 if (fastio_reg_override >= 0)
3012 a = fastio_reg_override;
3013 if (offset_reg >= 0)
3014 emit_ldrsh_dualindexed(offset_reg, a, tl);
3015 else
3016 emit_movswl_indexed(0, a, tl);
3017 }
3018 if(jaddr)
3019 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3020 }
3021 else
3022 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3023 break;
3024 case 0x23: // LW
3025 if(!c||memtarget) {
3026 if(!dummy) {
3027 int a = addr;
3028 if (fastio_reg_override >= 0)
3029 a = fastio_reg_override;
3030 do_load_word(a, tl, offset_reg);
3031 }
3032 if(jaddr)
3033 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3034 }
3035 else
3036 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3037 break;
3038 case 0x24: // LBU
3039 if(!c||memtarget) {
3040 if(!dummy) {
3041 int a = tl;
3042 if (!c) a = addr;
3043 if (fastio_reg_override >= 0)
3044 a = fastio_reg_override;
3045
3046 if (offset_reg >= 0)
3047 emit_ldrb_dualindexed(offset_reg, a, tl);
3048 else
3049 emit_movzbl_indexed(0, a, tl);
3050 }
3051 if(jaddr)
3052 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3053 }
3054 else
3055 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3056 break;
3057 case 0x25: // LHU
3058 if(!c||memtarget) {
3059 if(!dummy) {
3060 int a = tl;
3061 if(!c) a = addr;
3062 if (fastio_reg_override >= 0)
3063 a = fastio_reg_override;
3064 if (offset_reg >= 0)
3065 emit_ldrh_dualindexed(offset_reg, a, tl);
3066 else
3067 emit_movzwl_indexed(0, a, tl);
3068 }
3069 if(jaddr)
3070 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3071 }
3072 else
3073 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3074 break;
3075 case 0x27: // LWU
3076 case 0x37: // LD
3077 default:
3078 assert(0);
3079 }
3080 }
3081 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3082 host_tempreg_release();
3083}
3084
3085#ifndef loadlr_assemble
3086static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3087{
3088 int s,tl,temp,temp2,addr;
3089 int offset;
3090 void *jaddr=0;
3091 int memtarget=0,c=0;
3092 int offset_reg = -1;
3093 int fastio_reg_override = -1;
3094 u_int reglist=get_host_reglist(i_regs->regmap);
3095 tl=get_reg(i_regs->regmap,dops[i].rt1);
3096 s=get_reg(i_regs->regmap,dops[i].rs1);
3097 temp=get_reg_temp(i_regs->regmap);
3098 temp2=get_reg(i_regs->regmap,FTEMP);
3099 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
3100 assert(addr<0);
3101 offset=imm[i];
3102 reglist|=1<<temp;
3103 if(offset||s<0||c) addr=temp2;
3104 else addr=s;
3105 if(s>=0) {
3106 c=(i_regs->wasconst>>s)&1;
3107 if(c) {
3108 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3109 }
3110 }
3111 if(!c) {
3112 emit_shlimm(addr,3,temp);
3113 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3114 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3115 }else{
3116 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3117 }
3118 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3119 &offset_reg, &fastio_reg_override);
3120 }
3121 else {
3122 if (ram_offset && memtarget) {
3123 offset_reg = get_ro_reg(i_regs, 0);
3124 }
3125 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3126 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3127 }else{
3128 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3129 }
3130 }
3131 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3132 if(!c||memtarget) {
3133 int a = temp2;
3134 if (fastio_reg_override >= 0)
3135 a = fastio_reg_override;
3136 do_load_word(a, temp2, offset_reg);
3137 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3138 host_tempreg_release();
3139 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3140 }
3141 else
3142 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3143 if(dops[i].rt1) {
3144 assert(tl>=0);
3145 emit_andimm(temp,24,temp);
3146 if (dops[i].opcode==0x22) // LWL
3147 emit_xorimm(temp,24,temp);
3148 host_tempreg_acquire();
3149 emit_movimm(-1,HOST_TEMPREG);
3150 if (dops[i].opcode==0x26) {
3151 emit_shr(temp2,temp,temp2);
3152 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3153 }else{
3154 emit_shl(temp2,temp,temp2);
3155 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3156 }
3157 host_tempreg_release();
3158 emit_or(temp2,tl,tl);
3159 }
3160 //emit_storereg(dops[i].rt1,tl); // DEBUG
3161 }
3162 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3163 assert(0);
3164 }
3165}
3166#endif
3167
3168static void do_invstub(int n)
3169{
3170 literal_pool(20);
3171 assem_debug("do_invstub\n");
3172 u_int reglist = stubs[n].a;
3173 u_int addrr = stubs[n].b;
3174 int ofs_start = stubs[n].c;
3175 int ofs_end = stubs[n].d;
3176 int len = ofs_end - ofs_start;
3177 u_int rightr = 0;
3178
3179 set_jump_target(stubs[n].addr, out);
3180 save_regs(reglist);
3181 if (addrr != 0 || ofs_start != 0)
3182 emit_addimm(addrr, ofs_start, 0);
3183 emit_readword(&inv_code_start, 2);
3184 emit_readword(&inv_code_end, 3);
3185 if (len != 0)
3186 emit_addimm(0, len + 4, (rightr = 1));
3187 emit_cmp(0, 2);
3188 emit_cmpcs(3, rightr);
3189 void *jaddr = out;
3190 emit_jc(0);
3191 void *func = (len != 0)
3192 ? (void *)ndrc_write_invalidate_many
3193 : (void *)ndrc_write_invalidate_one;
3194 emit_far_call(func);
3195 set_jump_target(jaddr, out);
3196 restore_regs(reglist);
3197 emit_jmp(stubs[n].retaddr);
3198}
3199
3200static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3201{
3202 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3203 return;
3204 // this can't be used any more since we started to check exact
3205 // block boundaries in invalidate_range()
3206 //if (i_regs->waswritten & (1<<dops[i].rs1))
3207 // return;
3208 // (naively) assume nobody will run code from stack
3209 if (dops[i].rs1 == 29)
3210 return;
3211
3212 int j, imm_maxdiff = 32, imm_min = imm[i], imm_max = imm[i], count = 1;
3213 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3214 && abs(imm[i+1] - imm[i]) <= imm_maxdiff)
3215 return;
3216 for (j = i - 1; j >= 0; j--) {
3217 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3218 || abs(imm[j] - imm[j+1]) > imm_maxdiff)
3219 break;
3220 count++;
3221 if (imm_min > imm[j])
3222 imm_min = imm[j];
3223 if (imm_max < imm[j])
3224 imm_max = imm[j];
3225 }
3226#if defined(HOST_IMM8)
3227 int ir = get_reg(i_regs->regmap, INVCP);
3228 assert(ir >= 0);
3229 host_tempreg_acquire();
3230 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3231#else
3232 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3233 #error not handled
3234#endif
3235#ifdef INVALIDATE_USE_COND_CALL
3236 if (count == 1) {
3237 emit_cmpimm(HOST_TEMPREG, 1);
3238 emit_callne(invalidate_addr_reg[addr]);
3239 host_tempreg_release();
3240 return;
3241 }
3242#endif
3243 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3244 host_tempreg_release();
3245 imm_min -= imm[i];
3246 imm_max -= imm[i];
3247 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3248 addr, imm_min, imm_max, 0);
3249}
3250
3251static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3252{
3253 int s,tl;
3254 int addr,temp;
3255 int offset;
3256 void *jaddr=0;
3257 enum stub_type type=0;
3258 int memtarget=0,c=0;
3259 int agr=AGEN1+(i&1);
3260 int offset_reg = -1;
3261 int fastio_reg_override = -1;
3262 u_int reglist=get_host_reglist(i_regs->regmap);
3263 tl=get_reg(i_regs->regmap,dops[i].rs2);
3264 s=get_reg(i_regs->regmap,dops[i].rs1);
3265 temp=get_reg(i_regs->regmap,agr);
3266 if(temp<0) temp=get_reg_temp(i_regs->regmap);
3267 offset=imm[i];
3268 if(s>=0) {
3269 c=(i_regs->wasconst>>s)&1;
3270 if(c) {
3271 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3272 }
3273 }
3274 assert(tl>=0);
3275 assert(temp>=0);
3276 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3277 if(offset||s<0||c) addr=temp;
3278 else addr=s;
3279 if (!c) {
3280 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3281 &offset_reg, &fastio_reg_override);
3282 }
3283 else if (ram_offset && memtarget) {
3284 offset_reg = get_ro_reg(i_regs, 0);
3285 }
3286
3287 switch (dops[i].opcode) {
3288 case 0x28: // SB
3289 if(!c||memtarget) {
3290 int a = temp;
3291 if (!c) a = addr;
3292 if (fastio_reg_override >= 0)
3293 a = fastio_reg_override;
3294 do_store_byte(a, tl, offset_reg);
3295 }
3296 type = STOREB_STUB;
3297 break;
3298 case 0x29: // SH
3299 if(!c||memtarget) {
3300 int a = temp;
3301 if (!c) a = addr;
3302 if (fastio_reg_override >= 0)
3303 a = fastio_reg_override;
3304 do_store_hword(a, 0, tl, offset_reg, 1);
3305 }
3306 type = STOREH_STUB;
3307 break;
3308 case 0x2B: // SW
3309 if(!c||memtarget) {
3310 int a = addr;
3311 if (fastio_reg_override >= 0)
3312 a = fastio_reg_override;
3313 do_store_word(a, 0, tl, offset_reg, 1);
3314 }
3315 type = STOREW_STUB;
3316 break;
3317 case 0x3F: // SD
3318 default:
3319 assert(0);
3320 }
3321 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3322 host_tempreg_release();
3323 if(jaddr) {
3324 // PCSX store handlers don't check invcode again
3325 reglist|=1<<addr;
3326 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3327 jaddr=0;
3328 }
3329 {
3330 if(!c||memtarget) {
3331 #ifdef DESTRUCTIVE_SHIFT
3332 // The x86 shift operation is 'destructive'; it overwrites the
3333 // source register, so we need to make a copy first and use that.
3334 addr=temp;
3335 #endif
3336 do_store_smc_check(i, i_regs, reglist, addr);
3337 }
3338 }
3339 u_int addr_val=constmap[i][s]+offset;
3340 if(jaddr) {
3341 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3342 } else if(c&&!memtarget) {
3343 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3344 }
3345 // basic current block modification detection..
3346 // not looking back as that should be in mips cache already
3347 // (see Spyro2 title->attract mode)
3348 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3349 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3350 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3351 if(i_regs->regmap==regs[i].regmap) {
3352 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3353 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3354 emit_movimm(start+i*4+4,0);
3355 emit_writeword(0,&pcaddr);
3356 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3357 emit_far_call(ndrc_get_addr_ht);
3358 emit_jmpreg(0);
3359 }
3360 }
3361}
3362
3363static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3364{
3365 int s,tl;
3366 int temp;
3367 int offset;
3368 void *jaddr=0;
3369 void *case1, *case23, *case3;
3370 void *done0, *done1, *done2;
3371 int memtarget=0,c=0;
3372 int agr=AGEN1+(i&1);
3373 int offset_reg = -1;
3374 u_int reglist=get_host_reglist(i_regs->regmap);
3375 tl=get_reg(i_regs->regmap,dops[i].rs2);
3376 s=get_reg(i_regs->regmap,dops[i].rs1);
3377 temp=get_reg(i_regs->regmap,agr);
3378 if(temp<0) temp=get_reg_temp(i_regs->regmap);
3379 offset=imm[i];
3380 if(s>=0) {
3381 c=(i_regs->isconst>>s)&1;
3382 if(c) {
3383 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3384 }
3385 }
3386 assert(tl>=0);
3387 assert(temp>=0);
3388 if(!c) {
3389 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3390 if(!offset&&s!=temp) emit_mov(s,temp);
3391 jaddr=out;
3392 emit_jno(0);
3393 }
3394 else
3395 {
3396 if(!memtarget||!dops[i].rs1) {
3397 jaddr=out;
3398 emit_jmp(0);
3399 }
3400 }
3401 if (ram_offset)
3402 offset_reg = get_ro_reg(i_regs, 0);
3403
3404 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3405 assert(0);
3406 }
3407
3408 emit_testimm(temp,2);
3409 case23=out;
3410 emit_jne(0);
3411 emit_testimm(temp,1);
3412 case1=out;
3413 emit_jne(0);
3414 // 0
3415 if (dops[i].opcode == 0x2A) { // SWL
3416 // Write msb into least significant byte
3417 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3418 do_store_byte(temp, tl, offset_reg);
3419 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3420 }
3421 else if (dops[i].opcode == 0x2E) { // SWR
3422 // Write entire word
3423 do_store_word(temp, 0, tl, offset_reg, 1);
3424 }
3425 done0 = out;
3426 emit_jmp(0);
3427 // 1
3428 set_jump_target(case1, out);
3429 if (dops[i].opcode == 0x2A) { // SWL
3430 // Write two msb into two least significant bytes
3431 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3432 do_store_hword(temp, -1, tl, offset_reg, 0);
3433 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3434 }
3435 else if (dops[i].opcode == 0x2E) { // SWR
3436 // Write 3 lsb into three most significant bytes
3437 do_store_byte(temp, tl, offset_reg);
3438 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3439 do_store_hword(temp, 1, tl, offset_reg, 0);
3440 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3441 }
3442 done1=out;
3443 emit_jmp(0);
3444 // 2,3
3445 set_jump_target(case23, out);
3446 emit_testimm(temp,1);
3447 case3 = out;
3448 emit_jne(0);
3449 // 2
3450 if (dops[i].opcode==0x2A) { // SWL
3451 // Write 3 msb into three least significant bytes
3452 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3453 do_store_hword(temp, -2, tl, offset_reg, 1);
3454 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3455 do_store_byte(temp, tl, offset_reg);
3456 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3457 }
3458 else if (dops[i].opcode == 0x2E) { // SWR
3459 // Write two lsb into two most significant bytes
3460 do_store_hword(temp, 0, tl, offset_reg, 1);
3461 }
3462 done2 = out;
3463 emit_jmp(0);
3464 // 3
3465 set_jump_target(case3, out);
3466 if (dops[i].opcode == 0x2A) { // SWL
3467 do_store_word(temp, -3, tl, offset_reg, 0);
3468 }
3469 else if (dops[i].opcode == 0x2E) { // SWR
3470 do_store_byte(temp, tl, offset_reg);
3471 }
3472 set_jump_target(done0, out);
3473 set_jump_target(done1, out);
3474 set_jump_target(done2, out);
3475 if (offset_reg == HOST_TEMPREG)
3476 host_tempreg_release();
3477 if(!c||!memtarget)
3478 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj_,reglist);
3479 do_store_smc_check(i, i_regs, reglist, temp);
3480}
3481
3482static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3483{
3484 if(dops[i].opcode2==0) // MFC0
3485 {
3486 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3487 u_int copr=(source[i]>>11)&0x1f;
3488 //assert(t>=0); // Why does this happen? OOT is weird
3489 if(t>=0&&dops[i].rt1!=0) {
3490 emit_readword(&reg_cop0[copr],t);
3491 }
3492 }
3493 else if(dops[i].opcode2==4) // MTC0
3494 {
3495 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3496 char copr=(source[i]>>11)&0x1f;
3497 assert(s>=0);
3498 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3499 if(copr==9||copr==11||copr==12||copr==13) {
3500 emit_readword(&last_count,HOST_TEMPREG);
3501 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3502 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3503 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3504 emit_writeword(HOST_CCREG,&psxRegs.cycle);
3505 }
3506 // What a mess. The status register (12) can enable interrupts,
3507 // so needs a special case to handle a pending interrupt.
3508 // The interrupt must be taken immediately, because a subsequent
3509 // instruction might disable interrupts again.
3510 if(copr==12||copr==13) {
3511 if (is_delayslot) {
3512 // burn cycles to cause cc_interrupt, which will
3513 // reschedule next_interupt. Relies on CCREG from above.
3514 assem_debug("MTC0 DS %d\n", copr);
3515 emit_writeword(HOST_CCREG,&last_count);
3516 emit_movimm(0,HOST_CCREG);
3517 emit_storereg(CCREG,HOST_CCREG);
3518 emit_loadreg(dops[i].rs1,1);
3519 emit_movimm(copr,0);
3520 emit_far_call(pcsx_mtc0_ds);
3521 emit_loadreg(dops[i].rs1,s);
3522 return;
3523 }
3524 emit_movimm(start+i*4+4,HOST_TEMPREG);
3525 emit_writeword(HOST_TEMPREG,&pcaddr);
3526 emit_movimm(0,HOST_TEMPREG);
3527 emit_writeword(HOST_TEMPREG,&pending_exception);
3528 }
3529 if(s==HOST_CCREG)
3530 emit_loadreg(dops[i].rs1,1);
3531 else if(s!=1)
3532 emit_mov(s,1);
3533 emit_movimm(copr,0);
3534 emit_far_call(pcsx_mtc0);
3535 if(copr==9||copr==11||copr==12||copr==13) {
3536 emit_readword(&psxRegs.cycle,HOST_CCREG);
3537 emit_readword(&next_interupt,HOST_TEMPREG);
3538 emit_addimm(HOST_CCREG,-ccadj_,HOST_CCREG);
3539 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3540 emit_writeword(HOST_TEMPREG,&last_count);
3541 emit_storereg(CCREG,HOST_CCREG);
3542 }
3543 if(copr==12||copr==13) {
3544 assert(!is_delayslot);
3545 emit_readword(&pending_exception,HOST_TEMPREG);
3546 emit_test(HOST_TEMPREG,HOST_TEMPREG);
3547 void *jaddr = out;
3548 emit_jeq(0);
3549 emit_readword(&pcaddr, 0);
3550 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3551 emit_far_call(ndrc_get_addr_ht);
3552 emit_jmpreg(0);
3553 set_jump_target(jaddr, out);
3554 }
3555 emit_loadreg(dops[i].rs1,s);
3556 }
3557 else
3558 {
3559 assert(dops[i].opcode2==0x10);
3560 //if((source[i]&0x3f)==0x10) // RFE
3561 {
3562 emit_readword(&psxRegs.CP0.n.Status,0);
3563 emit_andimm(0,0x3c,1);
3564 emit_andimm(0,~0xf,0);
3565 emit_orrshr_imm(1,2,0);
3566 emit_writeword(0,&psxRegs.CP0.n.Status);
3567 }
3568 }
3569}
3570
3571static void cop1_unusable(int i, const struct regstat *i_regs)
3572{
3573 // XXX: should just just do the exception instead
3574 //if(!cop1_usable)
3575 {
3576 void *jaddr=out;
3577 emit_jmp(0);
3578 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3579 }
3580}
3581
3582static void cop1_assemble(int i, const struct regstat *i_regs)
3583{
3584 cop1_unusable(i, i_regs);
3585}
3586
3587static void c1ls_assemble(int i, const struct regstat *i_regs)
3588{
3589 cop1_unusable(i, i_regs);
3590}
3591
3592// FP_STUB
3593static void do_cop1stub(int n)
3594{
3595 literal_pool(256);
3596 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3597 set_jump_target(stubs[n].addr, out);
3598 int i=stubs[n].a;
3599// int rs=stubs[n].b;
3600 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3601 int ds=stubs[n].d;
3602 if(!ds) {
3603 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3604 //if(i_regs!=&regs[i]) printf("oops: regs[i]=%x i_regs=%x",(int)&regs[i],(int)i_regs);
3605 }
3606 //else {printf("fp exception in delay slot\n");}
3607 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3608 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3609 emit_movimm(start+(i-ds)*4,0); // Get PC
3610 emit_addimm(HOST_CCREG,ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3611 emit_far_jump(ds?fp_exception_ds:fp_exception);
3612}
3613
3614static int cop2_is_stalling_op(int i, int *cycles)
3615{
3616 if (dops[i].opcode == 0x3a) { // SWC2
3617 *cycles = 0;
3618 return 1;
3619 }
3620 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3621 *cycles = 0;
3622 return 1;
3623 }
3624 if (dops[i].itype == C2OP) {
3625 *cycles = gte_cycletab[source[i] & 0x3f];
3626 return 1;
3627 }
3628 // ... what about MTC2/CTC2/LWC2?
3629 return 0;
3630}
3631
3632#if 0
3633static void log_gte_stall(int stall, u_int cycle)
3634{
3635 if ((u_int)stall <= 44)
3636 printf("x stall %2d %u\n", stall, cycle + last_count);
3637}
3638
3639static void emit_log_gte_stall(int i, int stall, u_int reglist)
3640{
3641 save_regs(reglist);
3642 if (stall > 0)
3643 emit_movimm(stall, 0);
3644 else
3645 emit_mov(HOST_TEMPREG, 0);
3646 emit_addimm(HOST_CCREG, ccadj[i], 1);
3647 emit_far_call(log_gte_stall);
3648 restore_regs(reglist);
3649}
3650#endif
3651
3652static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3653{
3654 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3655 int rtmp = reglist_find_free(reglist);
3656
3657 if (HACK_ENABLED(NDHACK_NO_STALLS))
3658 return;
3659 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3660 // happens occasionally... cc evicted? Don't bother then
3661 //printf("no cc %08x\n", start + i*4);
3662 return;
3663 }
3664 if (!dops[i].bt) {
3665 for (j = i - 1; j >= 0; j--) {
3666 //if (dops[j].is_ds) break;
3667 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3668 break;
3669 if (j > 0 && ccadj[j - 1] > ccadj[j])
3670 break;
3671 }
3672 j = max(j, 0);
3673 }
3674 cycles_passed = ccadj[i] - ccadj[j];
3675 if (other_gte_op_cycles >= 0)
3676 stall = other_gte_op_cycles - cycles_passed;
3677 else if (cycles_passed >= 44)
3678 stall = 0; // can't stall
3679 if (stall == -MAXBLOCK && rtmp >= 0) {
3680 // unknown stall, do the expensive runtime check
3681 assem_debug("; cop2_do_stall_check\n");
3682#if 0 // too slow
3683 save_regs(reglist);
3684 emit_movimm(gte_cycletab[op], 0);
3685 emit_addimm(HOST_CCREG, ccadj[i], 1);
3686 emit_far_call(call_gteStall);
3687 restore_regs(reglist);
3688#else
3689 host_tempreg_acquire();
3690 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3691 emit_addimm(rtmp, -ccadj[i], rtmp);
3692 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3693 emit_cmpimm(HOST_TEMPREG, 44);
3694 emit_cmovb_reg(rtmp, HOST_CCREG);
3695 //emit_log_gte_stall(i, 0, reglist);
3696 host_tempreg_release();
3697#endif
3698 }
3699 else if (stall > 0) {
3700 //emit_log_gte_stall(i, stall, reglist);
3701 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3702 }
3703
3704 // save gteBusyCycle, if needed
3705 if (gte_cycletab[op] == 0)
3706 return;
3707 other_gte_op_cycles = -1;
3708 for (j = i + 1; j < slen; j++) {
3709 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3710 break;
3711 if (dops[j].is_jump) {
3712 // check ds
3713 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3714 j++;
3715 break;
3716 }
3717 }
3718 if (other_gte_op_cycles >= 0)
3719 // will handle stall when assembling that op
3720 return;
3721 cycles_passed = ccadj[min(j, slen -1)] - ccadj[i];
3722 if (cycles_passed >= 44)
3723 return;
3724 assem_debug("; save gteBusyCycle\n");
3725 host_tempreg_acquire();
3726#if 0
3727 emit_readword(&last_count, HOST_TEMPREG);
3728 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3729 emit_addimm(HOST_TEMPREG, ccadj[i], HOST_TEMPREG);
3730 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3731 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3732#else
3733 emit_addimm(HOST_CCREG, ccadj[i] + gte_cycletab[op], HOST_TEMPREG);
3734 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3735#endif
3736 host_tempreg_release();
3737}
3738
3739static int is_mflohi(int i)
3740{
3741 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3742}
3743
3744static int check_multdiv(int i, int *cycles)
3745{
3746 if (dops[i].itype != MULTDIV)
3747 return 0;
3748 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3749 *cycles = 11; // approx from 7 11 14
3750 else
3751 *cycles = 37;
3752 return 1;
3753}
3754
3755static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3756{
3757 int j, found = 0, c = 0;
3758 if (HACK_ENABLED(NDHACK_NO_STALLS))
3759 return;
3760 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3761 // happens occasionally... cc evicted? Don't bother then
3762 return;
3763 }
3764 for (j = i + 1; j < slen; j++) {
3765 if (dops[j].bt)
3766 break;
3767 if ((found = is_mflohi(j)))
3768 break;
3769 if (dops[j].is_jump) {
3770 // check ds
3771 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3772 j++;
3773 break;
3774 }
3775 }
3776 if (found)
3777 // handle all in multdiv_do_stall()
3778 return;
3779 check_multdiv(i, &c);
3780 assert(c > 0);
3781 assem_debug("; muldiv prepare stall %d\n", c);
3782 host_tempreg_acquire();
3783 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3784 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3785 host_tempreg_release();
3786}
3787
3788static void multdiv_do_stall(int i, const struct regstat *i_regs)
3789{
3790 int j, known_cycles = 0;
3791 u_int reglist = get_host_reglist(i_regs->regmap);
3792 int rtmp = get_reg_temp(i_regs->regmap);
3793 if (rtmp < 0)
3794 rtmp = reglist_find_free(reglist);
3795 if (HACK_ENABLED(NDHACK_NO_STALLS))
3796 return;
3797 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3798 // happens occasionally... cc evicted? Don't bother then
3799 //printf("no cc/rtmp %08x\n", start + i*4);
3800 return;
3801 }
3802 if (!dops[i].bt) {
3803 for (j = i - 1; j >= 0; j--) {
3804 if (dops[j].is_ds) break;
3805 if (check_multdiv(j, &known_cycles))
3806 break;
3807 if (is_mflohi(j))
3808 // already handled by this op
3809 return;
3810 if (dops[j].bt || (j > 0 && ccadj[j - 1] > ccadj[j]))
3811 break;
3812 }
3813 j = max(j, 0);
3814 }
3815 if (known_cycles > 0) {
3816 known_cycles -= ccadj[i] - ccadj[j];
3817 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3818 if (known_cycles > 0)
3819 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3820 return;
3821 }
3822 assem_debug("; muldiv stall unresolved\n");
3823 host_tempreg_acquire();
3824 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3825 emit_addimm(rtmp, -ccadj[i], rtmp);
3826 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3827 emit_cmpimm(HOST_TEMPREG, 37);
3828 emit_cmovb_reg(rtmp, HOST_CCREG);
3829 //emit_log_gte_stall(i, 0, reglist);
3830 host_tempreg_release();
3831}
3832
3833static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3834{
3835 switch (copr) {
3836 case 1:
3837 case 3:
3838 case 5:
3839 case 8:
3840 case 9:
3841 case 10:
3842 case 11:
3843 emit_readword(&reg_cop2d[copr],tl);
3844 emit_signextend16(tl,tl);
3845 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3846 break;
3847 case 7:
3848 case 16:
3849 case 17:
3850 case 18:
3851 case 19:
3852 emit_readword(&reg_cop2d[copr],tl);
3853 emit_andimm(tl,0xffff,tl);
3854 emit_writeword(tl,&reg_cop2d[copr]);
3855 break;
3856 case 15:
3857 emit_readword(&reg_cop2d[14],tl); // SXY2
3858 emit_writeword(tl,&reg_cop2d[copr]);
3859 break;
3860 case 28:
3861 case 29:
3862 c2op_mfc2_29_assemble(tl,temp);
3863 break;
3864 default:
3865 emit_readword(&reg_cop2d[copr],tl);
3866 break;
3867 }
3868}
3869
3870static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3871{
3872 switch (copr) {
3873 case 15:
3874 emit_readword(&reg_cop2d[13],temp); // SXY1
3875 emit_writeword(sl,&reg_cop2d[copr]);
3876 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3877 emit_readword(&reg_cop2d[14],temp); // SXY2
3878 emit_writeword(sl,&reg_cop2d[14]);
3879 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3880 break;
3881 case 28:
3882 emit_andimm(sl,0x001f,temp);
3883 emit_shlimm(temp,7,temp);
3884 emit_writeword(temp,&reg_cop2d[9]);
3885 emit_andimm(sl,0x03e0,temp);
3886 emit_shlimm(temp,2,temp);
3887 emit_writeword(temp,&reg_cop2d[10]);
3888 emit_andimm(sl,0x7c00,temp);
3889 emit_shrimm(temp,3,temp);
3890 emit_writeword(temp,&reg_cop2d[11]);
3891 emit_writeword(sl,&reg_cop2d[28]);
3892 break;
3893 case 30:
3894 emit_xorsar_imm(sl,sl,31,temp);
3895#if defined(HAVE_ARMV5) || defined(__aarch64__)
3896 emit_clz(temp,temp);
3897#else
3898 emit_movs(temp,HOST_TEMPREG);
3899 emit_movimm(0,temp);
3900 emit_jeq((int)out+4*4);
3901 emit_addpl_imm(temp,1,temp);
3902 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3903 emit_jns((int)out-2*4);
3904#endif
3905 emit_writeword(sl,&reg_cop2d[30]);
3906 emit_writeword(temp,&reg_cop2d[31]);
3907 break;
3908 case 31:
3909 break;
3910 default:
3911 emit_writeword(sl,&reg_cop2d[copr]);
3912 break;
3913 }
3914}
3915
3916static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3917{
3918 int s,tl;
3919 int ar;
3920 int offset;
3921 int memtarget=0,c=0;
3922 void *jaddr2=NULL;
3923 enum stub_type type;
3924 int agr=AGEN1+(i&1);
3925 int offset_reg = -1;
3926 int fastio_reg_override = -1;
3927 u_int reglist=get_host_reglist(i_regs->regmap);
3928 u_int copr=(source[i]>>16)&0x1f;
3929 s=get_reg(i_regs->regmap,dops[i].rs1);
3930 tl=get_reg(i_regs->regmap,FTEMP);
3931 offset=imm[i];
3932 assert(dops[i].rs1>0);
3933 assert(tl>=0);
3934
3935 if(i_regs->regmap[HOST_CCREG]==CCREG)
3936 reglist&=~(1<<HOST_CCREG);
3937
3938 // get the address
3939 if (dops[i].opcode==0x3a) { // SWC2
3940 ar=get_reg(i_regs->regmap,agr);
3941 if(ar<0) ar=get_reg_temp(i_regs->regmap);
3942 reglist|=1<<ar;
3943 } else { // LWC2
3944 ar=tl;
3945 }
3946 if(s>=0) c=(i_regs->wasconst>>s)&1;
3947 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3948 if (!offset&&!c&&s>=0) ar=s;
3949 assert(ar>=0);
3950
3951 cop2_do_stall_check(0, i, i_regs, reglist);
3952
3953 if (dops[i].opcode==0x3a) { // SWC2
3954 cop2_get_dreg(copr,tl,-1);
3955 type=STOREW_STUB;
3956 }
3957 else
3958 type=LOADW_STUB;
3959
3960 if(c&&!memtarget) {
3961 jaddr2=out;
3962 emit_jmp(0); // inline_readstub/inline_writestub?
3963 }
3964 else {
3965 if(!c) {
3966 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3967 &offset_reg, &fastio_reg_override);
3968 }
3969 else if (ram_offset && memtarget) {
3970 offset_reg = get_ro_reg(i_regs, 0);
3971 }
3972 switch (dops[i].opcode) {
3973 case 0x32: { // LWC2
3974 int a = ar;
3975 if (fastio_reg_override >= 0)
3976 a = fastio_reg_override;
3977 do_load_word(a, tl, offset_reg);
3978 break;
3979 }
3980 case 0x3a: { // SWC2
3981 #ifdef DESTRUCTIVE_SHIFT
3982 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3983 #endif
3984 int a = ar;
3985 if (fastio_reg_override >= 0)
3986 a = fastio_reg_override;
3987 do_store_word(a, 0, tl, offset_reg, 1);
3988 break;
3989 }
3990 default:
3991 assert(0);
3992 }
3993 }
3994 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3995 host_tempreg_release();
3996 if(jaddr2)
3997 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3998 if(dops[i].opcode==0x3a) // SWC2
3999 do_store_smc_check(i, i_regs, reglist, ar);
4000 if (dops[i].opcode==0x32) { // LWC2
4001 host_tempreg_acquire();
4002 cop2_put_dreg(copr,tl,HOST_TEMPREG);
4003 host_tempreg_release();
4004 }
4005}
4006
4007static void cop2_assemble(int i, const struct regstat *i_regs)
4008{
4009 u_int copr = (source[i]>>11) & 0x1f;
4010 signed char temp = get_reg_temp(i_regs->regmap);
4011
4012 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
4013 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
4014 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
4015 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
4016 reglist = reglist_exclude(reglist, tl, -1);
4017 }
4018 cop2_do_stall_check(0, i, i_regs, reglist);
4019 }
4020 if (dops[i].opcode2==0) { // MFC2
4021 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
4022 if(tl>=0&&dops[i].rt1!=0)
4023 cop2_get_dreg(copr,tl,temp);
4024 }
4025 else if (dops[i].opcode2==4) { // MTC2
4026 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4027 cop2_put_dreg(copr,sl,temp);
4028 }
4029 else if (dops[i].opcode2==2) // CFC2
4030 {
4031 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
4032 if(tl>=0&&dops[i].rt1!=0)
4033 emit_readword(&reg_cop2c[copr],tl);
4034 }
4035 else if (dops[i].opcode2==6) // CTC2
4036 {
4037 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4038 switch(copr) {
4039 case 4:
4040 case 12:
4041 case 20:
4042 case 26:
4043 case 27:
4044 case 29:
4045 case 30:
4046 emit_signextend16(sl,temp);
4047 break;
4048 case 31:
4049 c2op_ctc2_31_assemble(sl,temp);
4050 break;
4051 default:
4052 temp=sl;
4053 break;
4054 }
4055 emit_writeword(temp,&reg_cop2c[copr]);
4056 assert(sl>=0);
4057 }
4058}
4059
4060static void do_unalignedwritestub(int n)
4061{
4062 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4063 literal_pool(256);
4064 set_jump_target(stubs[n].addr, out);
4065
4066 int i=stubs[n].a;
4067 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4068 int addr=stubs[n].b;
4069 u_int reglist=stubs[n].e;
4070 signed char *i_regmap=i_regs->regmap;
4071 int temp2=get_reg(i_regmap,FTEMP);
4072 int rt;
4073 rt=get_reg(i_regmap,dops[i].rs2);
4074 assert(rt>=0);
4075 assert(addr>=0);
4076 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4077 reglist|=(1<<addr);
4078 reglist&=~(1<<temp2);
4079
4080 // don't bother with it and call write handler
4081 save_regs(reglist);
4082 pass_args(addr,rt);
4083 int cc=get_reg(i_regmap,CCREG);
4084 if(cc<0)
4085 emit_loadreg(CCREG,2);
4086 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4087 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4088 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4089 if(cc<0)
4090 emit_storereg(CCREG,2);
4091 restore_regs(reglist);
4092 emit_jmp(stubs[n].retaddr); // return address
4093}
4094
4095#ifndef multdiv_assemble
4096void multdiv_assemble(int i,struct regstat *i_regs)
4097{
4098 printf("Need multdiv_assemble for this architecture.\n");
4099 abort();
4100}
4101#endif
4102
4103static void mov_assemble(int i, const struct regstat *i_regs)
4104{
4105 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
4106 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
4107 if(dops[i].rt1) {
4108 signed char sl,tl;
4109 tl=get_reg(i_regs->regmap,dops[i].rt1);
4110 //assert(tl>=0);
4111 if(tl>=0) {
4112 sl=get_reg(i_regs->regmap,dops[i].rs1);
4113 if(sl>=0) emit_mov(sl,tl);
4114 else emit_loadreg(dops[i].rs1,tl);
4115 }
4116 }
4117 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
4118 multdiv_do_stall(i, i_regs);
4119}
4120
4121// call interpreter, exception handler, things that change pc/regs/cycles ...
4122static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
4123{
4124 signed char ccreg=get_reg(i_regs->regmap,CCREG);
4125 assert(ccreg==HOST_CCREG);
4126 assert(!is_delayslot);
4127 (void)ccreg;
4128
4129 emit_movimm(pc,3); // Get PC
4130 emit_readword(&last_count,2);
4131 emit_writeword(3,&psxRegs.pc);
4132 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
4133 emit_add(2,HOST_CCREG,2);
4134 emit_writeword(2,&psxRegs.cycle);
4135 emit_far_call(func);
4136 emit_far_jump(jump_to_new_pc);
4137}
4138
4139static void syscall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4140{
4141 // 'break' tends to be littered around to catch things like
4142 // division by 0 and is almost never executed, so don't emit much code here
4143 void *func = (dops[i].opcode2 == 0x0C)
4144 ? (is_delayslot ? jump_syscall_ds : jump_syscall)
4145 : (is_delayslot ? jump_break_ds : jump_break);
4146 assert(get_reg(i_regs->regmap, CCREG) == HOST_CCREG);
4147 emit_movimm(start + i*4, 2); // pc
4148 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4149 emit_far_jump(func);
4150}
4151
4152static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4153{
4154 void *hlefunc = gteNULL;
4155 uint32_t hleCode = source[i] & 0x03ffffff;
4156 if (hleCode < ARRAY_SIZE(psxHLEt))
4157 hlefunc = psxHLEt[hleCode];
4158
4159 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4160}
4161
4162static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4163{
4164 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4165}
4166
4167static void speculate_mov(int rs,int rt)
4168{
4169 if(rt!=0) {
4170 smrv_strong_next|=1<<rt;
4171 smrv[rt]=smrv[rs];
4172 }
4173}
4174
4175static void speculate_mov_weak(int rs,int rt)
4176{
4177 if(rt!=0) {
4178 smrv_weak_next|=1<<rt;
4179 smrv[rt]=smrv[rs];
4180 }
4181}
4182
4183static void speculate_register_values(int i)
4184{
4185 if(i==0) {
4186 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4187 // gp,sp are likely to stay the same throughout the block
4188 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4189 smrv_weak_next=~smrv_strong_next;
4190 //printf(" llr %08x\n", smrv[4]);
4191 }
4192 smrv_strong=smrv_strong_next;
4193 smrv_weak=smrv_weak_next;
4194 switch(dops[i].itype) {
4195 case ALU:
4196 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4197 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4198 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4199 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4200 else {
4201 smrv_strong_next&=~(1<<dops[i].rt1);
4202 smrv_weak_next&=~(1<<dops[i].rt1);
4203 }
4204 break;
4205 case SHIFTIMM:
4206 smrv_strong_next&=~(1<<dops[i].rt1);
4207 smrv_weak_next&=~(1<<dops[i].rt1);
4208 // fallthrough
4209 case IMM16:
4210 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4211 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
4212 if(hr>=0) {
4213 if(get_final_value(hr,i,&value))
4214 smrv[dops[i].rt1]=value;
4215 else smrv[dops[i].rt1]=constmap[i][hr];
4216 smrv_strong_next|=1<<dops[i].rt1;
4217 }
4218 }
4219 else {
4220 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4221 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4222 }
4223 break;
4224 case LOAD:
4225 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4226 // special case for BIOS
4227 smrv[dops[i].rt1]=0xa0000000;
4228 smrv_strong_next|=1<<dops[i].rt1;
4229 break;
4230 }
4231 // fallthrough
4232 case SHIFT:
4233 case LOADLR:
4234 case MOV:
4235 smrv_strong_next&=~(1<<dops[i].rt1);
4236 smrv_weak_next&=~(1<<dops[i].rt1);
4237 break;
4238 case COP0:
4239 case COP2:
4240 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4241 smrv_strong_next&=~(1<<dops[i].rt1);
4242 smrv_weak_next&=~(1<<dops[i].rt1);
4243 }
4244 break;
4245 case C2LS:
4246 if (dops[i].opcode==0x32) { // LWC2
4247 smrv_strong_next&=~(1<<dops[i].rt1);
4248 smrv_weak_next&=~(1<<dops[i].rt1);
4249 }
4250 break;
4251 }
4252#if 0
4253 int r=4;
4254 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4255 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4256#endif
4257}
4258
4259static void ujump_assemble(int i, const struct regstat *i_regs);
4260static void rjump_assemble(int i, const struct regstat *i_regs);
4261static void cjump_assemble(int i, const struct regstat *i_regs);
4262static void sjump_assemble(int i, const struct regstat *i_regs);
4263
4264static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4265{
4266 int ds = 0;
4267 switch (dops[i].itype) {
4268 case ALU:
4269 alu_assemble(i, i_regs);
4270 break;
4271 case IMM16:
4272 imm16_assemble(i, i_regs);
4273 break;
4274 case SHIFT:
4275 shift_assemble(i, i_regs);
4276 break;
4277 case SHIFTIMM:
4278 shiftimm_assemble(i, i_regs);
4279 break;
4280 case LOAD:
4281 load_assemble(i, i_regs, ccadj_);
4282 break;
4283 case LOADLR:
4284 loadlr_assemble(i, i_regs, ccadj_);
4285 break;
4286 case STORE:
4287 store_assemble(i, i_regs, ccadj_);
4288 break;
4289 case STORELR:
4290 storelr_assemble(i, i_regs, ccadj_);
4291 break;
4292 case COP0:
4293 cop0_assemble(i, i_regs, ccadj_);
4294 break;
4295 case COP1:
4296 cop1_assemble(i, i_regs);
4297 break;
4298 case C1LS:
4299 c1ls_assemble(i, i_regs);
4300 break;
4301 case COP2:
4302 cop2_assemble(i, i_regs);
4303 break;
4304 case C2LS:
4305 c2ls_assemble(i, i_regs, ccadj_);
4306 break;
4307 case C2OP:
4308 c2op_assemble(i, i_regs);
4309 break;
4310 case MULTDIV:
4311 multdiv_assemble(i, i_regs);
4312 multdiv_prepare_stall(i, i_regs, ccadj_);
4313 break;
4314 case MOV:
4315 mov_assemble(i, i_regs);
4316 break;
4317 case SYSCALL:
4318 syscall_assemble(i, i_regs, ccadj_);
4319 break;
4320 case HLECALL:
4321 hlecall_assemble(i, i_regs, ccadj_);
4322 break;
4323 case INTCALL:
4324 intcall_assemble(i, i_regs, ccadj_);
4325 break;
4326 case UJUMP:
4327 ujump_assemble(i, i_regs);
4328 ds = 1;
4329 break;
4330 case RJUMP:
4331 rjump_assemble(i, i_regs);
4332 ds = 1;
4333 break;
4334 case CJUMP:
4335 cjump_assemble(i, i_regs);
4336 ds = 1;
4337 break;
4338 case SJUMP:
4339 sjump_assemble(i, i_regs);
4340 ds = 1;
4341 break;
4342 case NOP:
4343 case OTHER:
4344 case NI:
4345 // not handled, just skip
4346 break;
4347 default:
4348 assert(0);
4349 }
4350 return ds;
4351}
4352
4353static void ds_assemble(int i, const struct regstat *i_regs)
4354{
4355 speculate_register_values(i);
4356 is_delayslot = 1;
4357 switch (dops[i].itype) {
4358 case SYSCALL:
4359 case HLECALL:
4360 case INTCALL:
4361 case UJUMP:
4362 case RJUMP:
4363 case CJUMP:
4364 case SJUMP:
4365 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4366 break;
4367 default:
4368 assemble(i, i_regs, ccadj[i]);
4369 }
4370 is_delayslot = 0;
4371}
4372
4373// Is the branch target a valid internal jump?
4374static int internal_branch(int addr)
4375{
4376 if(addr&1) return 0; // Indirect (register) jump
4377 if(addr>=start && addr<start+slen*4-4)
4378 {
4379 return 1;
4380 }
4381 return 0;
4382}
4383
4384static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4385{
4386 int hr;
4387 for(hr=0;hr<HOST_REGS;hr++) {
4388 if(hr!=EXCLUDE_REG) {
4389 if(pre[hr]!=entry[hr]) {
4390 if(pre[hr]>=0) {
4391 if((dirty>>hr)&1) {
4392 if(get_reg(entry,pre[hr])<0) {
4393 assert(pre[hr]<64);
4394 if(!((u>>pre[hr])&1))
4395 emit_storereg(pre[hr],hr);
4396 }
4397 }
4398 }
4399 }
4400 }
4401 }
4402 // Move from one register to another (no writeback)
4403 for(hr=0;hr<HOST_REGS;hr++) {
4404 if(hr!=EXCLUDE_REG) {
4405 if(pre[hr]!=entry[hr]) {
4406 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4407 int nr;
4408 if((nr=get_reg(entry,pre[hr]))>=0) {
4409 emit_mov(hr,nr);
4410 }
4411 }
4412 }
4413 }
4414 }
4415}
4416
4417// Load the specified registers
4418// This only loads the registers given as arguments because
4419// we don't want to load things that will be overwritten
4420static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4421{
4422 int hr = get_reg(regmap, rs);
4423 if (hr >= 0 && entry[hr] != regmap[hr])
4424 emit_loadreg(regmap[hr], hr);
4425}
4426
4427static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4428{
4429 load_reg(entry, regmap, rs1);
4430 if (rs1 != rs2)
4431 load_reg(entry, regmap, rs2);
4432}
4433
4434// Load registers prior to the start of a loop
4435// so that they are not loaded within the loop
4436static void loop_preload(signed char pre[],signed char entry[])
4437{
4438 int hr;
4439 for (hr = 0; hr < HOST_REGS; hr++) {
4440 int r = entry[hr];
4441 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4442 assem_debug("loop preload:\n");
4443 if (r < TEMPREG)
4444 emit_loadreg(r, hr);
4445 }
4446 }
4447}
4448
4449// Generate address for load/store instruction
4450// goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4451static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4452{
4453 if (dops[i].is_load || dops[i].is_store) {
4454 int ra=-1;
4455 int agr=AGEN1+(i&1);
4456 if(dops[i].itype==LOAD) {
4457 ra=get_reg(i_regs->regmap,dops[i].rt1);
4458 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4459 assert(ra>=0);
4460 }
4461 if(dops[i].itype==LOADLR) {
4462 ra=get_reg(i_regs->regmap,FTEMP);
4463 }
4464 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4465 ra=get_reg(i_regs->regmap,agr);
4466 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4467 }
4468 if(dops[i].itype==C2LS) {
4469 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4470 ra=get_reg(i_regs->regmap,FTEMP);
4471 else { // SWC1/SDC1/SWC2/SDC2
4472 ra=get_reg(i_regs->regmap,agr);
4473 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4474 }
4475 }
4476 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4477 if(ra>=0) {
4478 int offset=imm[i];
4479 int c=(i_regs->wasconst>>rs)&1;
4480 if(dops[i].rs1==0) {
4481 // Using r0 as a base address
4482 if(!entry||entry[ra]!=agr) {
4483 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4484 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4485 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4486 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4487 }else{
4488 emit_movimm(offset,ra);
4489 }
4490 } // else did it in the previous cycle
4491 }
4492 else if(rs<0) {
4493 if(!entry||entry[ra]!=dops[i].rs1)
4494 emit_loadreg(dops[i].rs1,ra);
4495 //if(!entry||entry[ra]!=dops[i].rs1)
4496 // printf("poor load scheduling!\n");
4497 }
4498 else if(c) {
4499 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4500 if(!entry||entry[ra]!=agr) {
4501 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4502 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4503 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4504 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4505 }else{
4506 emit_movimm(constmap[i][rs]+offset,ra);
4507 regs[i].loadedconst|=1<<ra;
4508 }
4509 } // else did it in the previous cycle
4510 } // else load_consts already did it
4511 }
4512 if(offset&&!c&&dops[i].rs1) {
4513 if(rs>=0) {
4514 emit_addimm(rs,offset,ra);
4515 }else{
4516 emit_addimm(ra,offset,ra);
4517 }
4518 }
4519 }
4520 }
4521 // Preload constants for next instruction
4522 if (dops[i+1].is_load || dops[i+1].is_store) {
4523 int agr,ra;
4524 // Actual address
4525 agr=AGEN1+((i+1)&1);
4526 ra=get_reg(i_regs->regmap,agr);
4527 if(ra>=0) {
4528 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4529 int offset=imm[i+1];
4530 int c=(regs[i+1].wasconst>>rs)&1;
4531 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4532 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4533 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4534 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4535 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4536 }else{
4537 emit_movimm(constmap[i+1][rs]+offset,ra);
4538 regs[i+1].loadedconst|=1<<ra;
4539 }
4540 }
4541 else if(dops[i+1].rs1==0) {
4542 // Using r0 as a base address
4543 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4544 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4545 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4546 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4547 }else{
4548 emit_movimm(offset,ra);
4549 }
4550 }
4551 }
4552 }
4553}
4554
4555static int get_final_value(int hr, int i, int *value)
4556{
4557 int reg=regs[i].regmap[hr];
4558 while(i<slen-1) {
4559 if(regs[i+1].regmap[hr]!=reg) break;
4560 if(!((regs[i+1].isconst>>hr)&1)) break;
4561 if(dops[i+1].bt) break;
4562 i++;
4563 }
4564 if(i<slen-1) {
4565 if (dops[i].is_jump) {
4566 *value=constmap[i][hr];
4567 return 1;
4568 }
4569 if(!dops[i+1].bt) {
4570 if (dops[i+1].is_jump) {
4571 // Load in delay slot, out-of-order execution
4572 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4573 {
4574 // Precompute load address
4575 *value=constmap[i][hr]+imm[i+2];
4576 return 1;
4577 }
4578 }
4579 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4580 {
4581 // Precompute load address
4582 *value=constmap[i][hr]+imm[i+1];
4583 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4584 return 1;
4585 }
4586 }
4587 }
4588 *value=constmap[i][hr];
4589 //printf("c=%lx\n",(long)constmap[i][hr]);
4590 if(i==slen-1) return 1;
4591 assert(reg < 64);
4592 return !((unneeded_reg[i+1]>>reg)&1);
4593}
4594
4595// Load registers with known constants
4596static void load_consts(signed char pre[],signed char regmap[],int i)
4597{
4598 int hr,hr2;
4599 // propagate loaded constant flags
4600 if(i==0||dops[i].bt)
4601 regs[i].loadedconst=0;
4602 else {
4603 for(hr=0;hr<HOST_REGS;hr++) {
4604 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4605 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4606 {
4607 regs[i].loadedconst|=1<<hr;
4608 }
4609 }
4610 }
4611 // Load 32-bit regs
4612 for(hr=0;hr<HOST_REGS;hr++) {
4613 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4614 //if(entry[hr]!=regmap[hr]) {
4615 if(!((regs[i].loadedconst>>hr)&1)) {
4616 assert(regmap[hr]<64);
4617 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4618 int value,similar=0;
4619 if(get_final_value(hr,i,&value)) {
4620 // see if some other register has similar value
4621 for(hr2=0;hr2<HOST_REGS;hr2++) {
4622 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4623 if(is_similar_value(value,constmap[i][hr2])) {
4624 similar=1;
4625 break;
4626 }
4627 }
4628 }
4629 if(similar) {
4630 int value2;
4631 if(get_final_value(hr2,i,&value2)) // is this needed?
4632 emit_movimm_from(value2,hr2,value,hr);
4633 else
4634 emit_movimm(value,hr);
4635 }
4636 else if(value==0) {
4637 emit_zeroreg(hr);
4638 }
4639 else {
4640 emit_movimm(value,hr);
4641 }
4642 }
4643 regs[i].loadedconst|=1<<hr;
4644 }
4645 }
4646 }
4647 }
4648}
4649
4650static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4651{
4652 int hr;
4653 // Load 32-bit regs
4654 for(hr=0;hr<HOST_REGS;hr++) {
4655 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4656 assert(regmap[hr] < 64);
4657 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4658 int value=constmap[i][hr];
4659 if(value==0) {
4660 emit_zeroreg(hr);
4661 }
4662 else {
4663 emit_movimm(value,hr);
4664 }
4665 }
4666 }
4667 }
4668}
4669
4670// Write out all dirty registers (except cycle count)
4671static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4672{
4673 int hr;
4674 for(hr=0;hr<HOST_REGS;hr++) {
4675 if(hr!=EXCLUDE_REG) {
4676 if(i_regmap[hr]>0) {
4677 if(i_regmap[hr]!=CCREG) {
4678 if((i_dirty>>hr)&1) {
4679 assert(i_regmap[hr]<64);
4680 emit_storereg(i_regmap[hr],hr);
4681 }
4682 }
4683 }
4684 }
4685 }
4686}
4687
4688// Write out dirty registers that we need to reload (pair with load_needed_regs)
4689// This writes the registers not written by store_regs_bt
4690static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4691{
4692 int hr;
4693 int t=(addr-start)>>2;
4694 for(hr=0;hr<HOST_REGS;hr++) {
4695 if(hr!=EXCLUDE_REG) {
4696 if(i_regmap[hr]>0) {
4697 if(i_regmap[hr]!=CCREG) {
4698 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4699 if((i_dirty>>hr)&1) {
4700 assert(i_regmap[hr]<64);
4701 emit_storereg(i_regmap[hr],hr);
4702 }
4703 }
4704 }
4705 }
4706 }
4707 }
4708}
4709
4710// Load all registers (except cycle count)
4711static void load_all_regs(const signed char i_regmap[])
4712{
4713 int hr;
4714 for(hr=0;hr<HOST_REGS;hr++) {
4715 if(hr!=EXCLUDE_REG) {
4716 if(i_regmap[hr]==0) {
4717 emit_zeroreg(hr);
4718 }
4719 else
4720 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4721 {
4722 emit_loadreg(i_regmap[hr],hr);
4723 }
4724 }
4725 }
4726}
4727
4728// Load all current registers also needed by next instruction
4729static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4730{
4731 int hr;
4732 for(hr=0;hr<HOST_REGS;hr++) {
4733 if(hr!=EXCLUDE_REG) {
4734 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4735 if(i_regmap[hr]==0) {
4736 emit_zeroreg(hr);
4737 }
4738 else
4739 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4740 {
4741 emit_loadreg(i_regmap[hr],hr);
4742 }
4743 }
4744 }
4745 }
4746}
4747
4748// Load all regs, storing cycle count if necessary
4749static void load_regs_entry(int t)
4750{
4751 int hr;
4752 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4753 else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t],HOST_CCREG);
4754 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4755 emit_storereg(CCREG,HOST_CCREG);
4756 }
4757 // Load 32-bit regs
4758 for(hr=0;hr<HOST_REGS;hr++) {
4759 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4760 if(regs[t].regmap_entry[hr]==0) {
4761 emit_zeroreg(hr);
4762 }
4763 else if(regs[t].regmap_entry[hr]!=CCREG)
4764 {
4765 emit_loadreg(regs[t].regmap_entry[hr],hr);
4766 }
4767 }
4768 }
4769}
4770
4771// Store dirty registers prior to branch
4772static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4773{
4774 if(internal_branch(addr))
4775 {
4776 int t=(addr-start)>>2;
4777 int hr;
4778 for(hr=0;hr<HOST_REGS;hr++) {
4779 if(hr!=EXCLUDE_REG) {
4780 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4781 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4782 if((i_dirty>>hr)&1) {
4783 assert(i_regmap[hr]<64);
4784 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4785 emit_storereg(i_regmap[hr],hr);
4786 }
4787 }
4788 }
4789 }
4790 }
4791 }
4792 else
4793 {
4794 // Branch out of this block, write out all dirty regs
4795 wb_dirtys(i_regmap,i_dirty);
4796 }
4797}
4798
4799// Load all needed registers for branch target
4800static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4801{
4802 //if(addr>=start && addr<(start+slen*4))
4803 if(internal_branch(addr))
4804 {
4805 int t=(addr-start)>>2;
4806 int hr;
4807 // Store the cycle count before loading something else
4808 if(i_regmap[HOST_CCREG]!=CCREG) {
4809 assert(i_regmap[HOST_CCREG]==-1);
4810 }
4811 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4812 emit_storereg(CCREG,HOST_CCREG);
4813 }
4814 // Load 32-bit regs
4815 for(hr=0;hr<HOST_REGS;hr++) {
4816 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4817 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4818 if(regs[t].regmap_entry[hr]==0) {
4819 emit_zeroreg(hr);
4820 }
4821 else if(regs[t].regmap_entry[hr]!=CCREG)
4822 {
4823 emit_loadreg(regs[t].regmap_entry[hr],hr);
4824 }
4825 }
4826 }
4827 }
4828 }
4829}
4830
4831static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4832{
4833 if(addr>=start && addr<start+slen*4-4)
4834 {
4835 int t=(addr-start)>>2;
4836 int hr;
4837 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4838 for(hr=0;hr<HOST_REGS;hr++)
4839 {
4840 if(hr!=EXCLUDE_REG)
4841 {
4842 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4843 {
4844 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4845 {
4846 return 0;
4847 }
4848 else
4849 if((i_dirty>>hr)&1)
4850 {
4851 if(i_regmap[hr]<TEMPREG)
4852 {
4853 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4854 return 0;
4855 }
4856 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4857 {
4858 assert(0);
4859 }
4860 }
4861 }
4862 else // Same register but is it 32-bit or dirty?
4863 if(i_regmap[hr]>=0)
4864 {
4865 if(!((regs[t].dirty>>hr)&1))
4866 {
4867 if((i_dirty>>hr)&1)
4868 {
4869 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4870 {
4871 //printf("%x: dirty no match\n",addr);
4872 return 0;
4873 }
4874 }
4875 }
4876 }
4877 }
4878 }
4879 // Delay slots are not valid branch targets
4880 //if(t>0&&(dops[t-1].is_jump) return 0;
4881 // Delay slots require additional processing, so do not match
4882 if(dops[t].is_ds) return 0;
4883 }
4884 else
4885 {
4886 int hr;
4887 for(hr=0;hr<HOST_REGS;hr++)
4888 {
4889 if(hr!=EXCLUDE_REG)
4890 {
4891 if(i_regmap[hr]>=0)
4892 {
4893 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4894 {
4895 if((i_dirty>>hr)&1)
4896 {
4897 return 0;
4898 }
4899 }
4900 }
4901 }
4902 }
4903 }
4904 return 1;
4905}
4906
4907#ifdef DRC_DBG
4908static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4909{
4910 extern void do_insn_cmp();
4911 //extern int cycle;
4912 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4913
4914 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4915 save_regs(reglist);
4916 // write out changed consts to match the interpreter
4917 if (i > 0 && !dops[i].bt) {
4918 for (hr = 0; hr < HOST_REGS; hr++) {
4919 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4920 if (hr == EXCLUDE_REG || reg < 0)
4921 continue;
4922 if (!((regs[i-1].isconst >> hr) & 1))
4923 continue;
4924 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4925 continue;
4926 emit_movimm(constmap[i-1][hr],0);
4927 emit_storereg(reg, 0);
4928 }
4929 }
4930 emit_movimm(start+i*4,0);
4931 emit_writeword(0,&pcaddr);
4932 int cc = get_reg(regs[i].regmap_entry, CCREG);
4933 if (cc < 0)
4934 emit_loadreg(CCREG, cc = 0);
4935 emit_addimm(cc, ccadj_, 0);
4936 emit_writeword(0, &psxRegs.cycle);
4937 emit_far_call(do_insn_cmp);
4938 //emit_readword(&cycle,0);
4939 //emit_addimm(0,2,0);
4940 //emit_writeword(0,&cycle);
4941 (void)get_reg2;
4942 restore_regs(reglist);
4943 assem_debug("\\\\do_insn_cmp\n");
4944}
4945#else
4946#define drc_dbg_emit_do_cmp(x,y)
4947#endif
4948
4949// Used when a branch jumps into the delay slot of another branch
4950static void ds_assemble_entry(int i)
4951{
4952 int t = (ba[i] - start) >> 2;
4953 int ccadj_ = -CLOCK_ADJUST(1);
4954 if (!instr_addr[t])
4955 instr_addr[t] = out;
4956 assem_debug("Assemble delay slot at %x\n",ba[i]);
4957 assem_debug("<->\n");
4958 drc_dbg_emit_do_cmp(t, ccadj_);
4959 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4960 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4961 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4962 address_generation(t,&regs[t],regs[t].regmap_entry);
4963 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4964 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
4965 if (dops[t].is_store)
4966 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
4967 is_delayslot=0;
4968 switch (dops[t].itype) {
4969 case SYSCALL:
4970 case HLECALL:
4971 case INTCALL:
4972 case UJUMP:
4973 case RJUMP:
4974 case CJUMP:
4975 case SJUMP:
4976 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4977 break;
4978 default:
4979 assemble(t, &regs[t], ccadj_);
4980 }
4981 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4982 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4983 if(internal_branch(ba[i]+4))
4984 assem_debug("branch: internal\n");
4985 else
4986 assem_debug("branch: external\n");
4987 assert(internal_branch(ba[i]+4));
4988 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4989 emit_jmp(0);
4990}
4991
4992// Load 2 immediates optimizing for small code size
4993static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4994{
4995 emit_movimm(imm1,rt1);
4996 emit_movimm_from(imm1,rt1,imm2,rt2);
4997}
4998
4999static void do_cc(int i, const signed char i_regmap[], int *adj,
5000 int addr, int taken, int invert)
5001{
5002 int count, count_plus2;
5003 void *jaddr;
5004 void *idle=NULL;
5005 int t=0;
5006 if(dops[i].itype==RJUMP)
5007 {
5008 *adj=0;
5009 }
5010 //if(ba[i]>=start && ba[i]<(start+slen*4))
5011 if(internal_branch(ba[i]))
5012 {
5013 t=(ba[i]-start)>>2;
5014 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5015 else *adj=ccadj[t];
5016 }
5017 else
5018 {
5019 *adj=0;
5020 }
5021 count = ccadj[i];
5022 count_plus2 = count + CLOCK_ADJUST(2);
5023 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
5024 // Idle loop
5025 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5026 idle=out;
5027 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5028 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5029 jaddr=out;
5030 emit_jmp(0);
5031 }
5032 else if(*adj==0||invert) {
5033 int cycles = count_plus2;
5034 // faster loop HACK
5035#if 0
5036 if (t&&*adj) {
5037 int rel=t-i;
5038 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5039 cycles=*adj+count+2-*adj;
5040 }
5041#endif
5042 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5043 jaddr = out;
5044 emit_jns(0);
5045 }
5046 else
5047 {
5048 emit_cmpimm(HOST_CCREG, -count_plus2);
5049 jaddr = out;
5050 emit_jns(0);
5051 }
5052 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5053}
5054
5055static void do_ccstub(int n)
5056{
5057 literal_pool(256);
5058 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5059 set_jump_target(stubs[n].addr, out);
5060 int i=stubs[n].b;
5061 if(stubs[n].d==NULLDS) {
5062 // Delay slot instruction is nullified ("likely" branch)
5063 wb_dirtys(regs[i].regmap,regs[i].dirty);
5064 }
5065 else if(stubs[n].d!=TAKEN) {
5066 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5067 }
5068 else {
5069 if(internal_branch(ba[i]))
5070 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5071 }
5072 if(stubs[n].c!=-1)
5073 {
5074 // Save PC as return address
5075 emit_movimm(stubs[n].c,0);
5076 emit_writeword(0,&pcaddr);
5077 }
5078 else
5079 {
5080 // Return address depends on which way the branch goes
5081 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5082 {
5083 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5084 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5085 if(dops[i].rs1==0)
5086 {
5087 s1l=s2l;
5088 s2l=-1;
5089 }
5090 else if(dops[i].rs2==0)
5091 {
5092 s2l=-1;
5093 }
5094 assert(s1l>=0);
5095 #ifdef DESTRUCTIVE_WRITEBACK
5096 if(dops[i].rs1) {
5097 if((branch_regs[i].dirty>>s1l)&&1)
5098 emit_loadreg(dops[i].rs1,s1l);
5099 }
5100 else {
5101 if((branch_regs[i].dirty>>s1l)&1)
5102 emit_loadreg(dops[i].rs2,s1l);
5103 }
5104 if(s2l>=0)
5105 if((branch_regs[i].dirty>>s2l)&1)
5106 emit_loadreg(dops[i].rs2,s2l);
5107 #endif
5108 int hr=0;
5109 int addr=-1,alt=-1,ntaddr=-1;
5110 while(hr<HOST_REGS)
5111 {
5112 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5113 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5114 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5115 {
5116 addr=hr++;break;
5117 }
5118 hr++;
5119 }
5120 while(hr<HOST_REGS)
5121 {
5122 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5123 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5124 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5125 {
5126 alt=hr++;break;
5127 }
5128 hr++;
5129 }
5130 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5131 {
5132 while(hr<HOST_REGS)
5133 {
5134 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5135 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5136 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5137 {
5138 ntaddr=hr;break;
5139 }
5140 hr++;
5141 }
5142 assert(hr<HOST_REGS);
5143 }
5144 if((dops[i].opcode&0x2f)==4) // BEQ
5145 {
5146 #ifdef HAVE_CMOV_IMM
5147 if(s2l>=0) emit_cmp(s1l,s2l);
5148 else emit_test(s1l,s1l);
5149 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5150 #else
5151 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5152 if(s2l>=0) emit_cmp(s1l,s2l);
5153 else emit_test(s1l,s1l);
5154 emit_cmovne_reg(alt,addr);
5155 #endif
5156 }
5157 if((dops[i].opcode&0x2f)==5) // BNE
5158 {
5159 #ifdef HAVE_CMOV_IMM
5160 if(s2l>=0) emit_cmp(s1l,s2l);
5161 else emit_test(s1l,s1l);
5162 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5163 #else
5164 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5165 if(s2l>=0) emit_cmp(s1l,s2l);
5166 else emit_test(s1l,s1l);
5167 emit_cmovne_reg(alt,addr);
5168 #endif
5169 }
5170 if((dops[i].opcode&0x2f)==6) // BLEZ
5171 {
5172 //emit_movimm(ba[i],alt);
5173 //emit_movimm(start+i*4+8,addr);
5174 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5175 emit_cmpimm(s1l,1);
5176 emit_cmovl_reg(alt,addr);
5177 }
5178 if((dops[i].opcode&0x2f)==7) // BGTZ
5179 {
5180 //emit_movimm(ba[i],addr);
5181 //emit_movimm(start+i*4+8,ntaddr);
5182 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5183 emit_cmpimm(s1l,1);
5184 emit_cmovl_reg(ntaddr,addr);
5185 }
5186 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
5187 {
5188 //emit_movimm(ba[i],alt);
5189 //emit_movimm(start+i*4+8,addr);
5190 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5191 emit_test(s1l,s1l);
5192 emit_cmovs_reg(alt,addr);
5193 }
5194 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
5195 {
5196 //emit_movimm(ba[i],addr);
5197 //emit_movimm(start+i*4+8,alt);
5198 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5199 emit_test(s1l,s1l);
5200 emit_cmovs_reg(alt,addr);
5201 }
5202 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5203 if(source[i]&0x10000) // BC1T
5204 {
5205 //emit_movimm(ba[i],alt);
5206 //emit_movimm(start+i*4+8,addr);
5207 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5208 emit_testimm(s1l,0x800000);
5209 emit_cmovne_reg(alt,addr);
5210 }
5211 else // BC1F
5212 {
5213 //emit_movimm(ba[i],addr);
5214 //emit_movimm(start+i*4+8,alt);
5215 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5216 emit_testimm(s1l,0x800000);
5217 emit_cmovne_reg(alt,addr);
5218 }
5219 }
5220 emit_writeword(addr,&pcaddr);
5221 }
5222 else
5223 if(dops[i].itype==RJUMP)
5224 {
5225 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5226 if (ds_writes_rjump_rs(i)) {
5227 r=get_reg(branch_regs[i].regmap,RTEMP);
5228 }
5229 emit_writeword(r,&pcaddr);
5230 }
5231 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5232 }
5233 // Update cycle count
5234 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5235 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5236 emit_far_call(cc_interrupt);
5237 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5238 if(stubs[n].d==TAKEN) {
5239 if(internal_branch(ba[i]))
5240 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
5241 else if(dops[i].itype==RJUMP) {
5242 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5243 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5244 else
5245 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5246 }
5247 }else if(stubs[n].d==NOTTAKEN) {
5248 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5249 else load_all_regs(branch_regs[i].regmap);
5250 }else if(stubs[n].d==NULLDS) {
5251 // Delay slot instruction is nullified ("likely" branch)
5252 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5253 else load_all_regs(regs[i].regmap);
5254 }else{
5255 load_all_regs(branch_regs[i].regmap);
5256 }
5257 if (stubs[n].retaddr)
5258 emit_jmp(stubs[n].retaddr);
5259 else
5260 do_jump_vaddr(stubs[n].e);
5261}
5262
5263static void add_to_linker(void *addr, u_int target, int is_internal)
5264{
5265 assert(linkcount < ARRAY_SIZE(link_addr));
5266 link_addr[linkcount].addr = addr;
5267 link_addr[linkcount].target = target;
5268 link_addr[linkcount].internal = is_internal;
5269 linkcount++;
5270}
5271
5272static void ujump_assemble_write_ra(int i)
5273{
5274 int rt;
5275 unsigned int return_address;
5276 rt=get_reg(branch_regs[i].regmap,31);
5277 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]);
5278 //assert(rt>=0);
5279 return_address=start+i*4+8;
5280 if(rt>=0) {
5281 #ifdef USE_MINI_HT
5282 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5283 int temp=-1; // note: must be ds-safe
5284 #ifdef HOST_TEMPREG
5285 temp=HOST_TEMPREG;
5286 #endif
5287 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5288 else emit_movimm(return_address,rt);
5289 }
5290 else
5291 #endif
5292 {
5293 #ifdef REG_PREFETCH
5294 if(temp>=0)
5295 {
5296 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5297 }
5298 #endif
5299 emit_movimm(return_address,rt); // PC into link register
5300 #ifdef IMM_PREFETCH
5301 emit_prefetch(hash_table_get(return_address));
5302 #endif
5303 }
5304 }
5305}
5306
5307static void ujump_assemble(int i, const struct regstat *i_regs)
5308{
5309 int ra_done=0;
5310 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5311 address_generation(i+1,i_regs,regs[i].regmap_entry);
5312 #ifdef REG_PREFETCH
5313 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5314 if(dops[i].rt1==31&&temp>=0)
5315 {
5316 signed char *i_regmap=i_regs->regmap;
5317 int return_address=start+i*4+8;
5318 if(get_reg(branch_regs[i].regmap,31)>0)
5319 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5320 }
5321 #endif
5322 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5323 ujump_assemble_write_ra(i); // writeback ra for DS
5324 ra_done=1;
5325 }
5326 ds_assemble(i+1,i_regs);
5327 uint64_t bc_unneeded=branch_regs[i].u;
5328 bc_unneeded|=1|(1LL<<dops[i].rt1);
5329 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5330 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5331 if(!ra_done&&dops[i].rt1==31)
5332 ujump_assemble_write_ra(i);
5333 int cc,adj;
5334 cc=get_reg(branch_regs[i].regmap,CCREG);
5335 assert(cc==HOST_CCREG);
5336 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5337 #ifdef REG_PREFETCH
5338 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5339 #endif
5340 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5341 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5342 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5343 if(internal_branch(ba[i]))
5344 assem_debug("branch: internal\n");
5345 else
5346 assem_debug("branch: external\n");
5347 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5348 ds_assemble_entry(i);
5349 }
5350 else {
5351 add_to_linker(out,ba[i],internal_branch(ba[i]));
5352 emit_jmp(0);
5353 }
5354}
5355
5356static void rjump_assemble_write_ra(int i)
5357{
5358 int rt,return_address;
5359 assert(dops[i+1].rt1!=dops[i].rt1);
5360 assert(dops[i+1].rt2!=dops[i].rt1);
5361 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5362 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]);
5363 assert(rt>=0);
5364 return_address=start+i*4+8;
5365 #ifdef REG_PREFETCH
5366 if(temp>=0)
5367 {
5368 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5369 }
5370 #endif
5371 emit_movimm(return_address,rt); // PC into link register
5372 #ifdef IMM_PREFETCH
5373 emit_prefetch(hash_table_get(return_address));
5374 #endif
5375}
5376
5377static void rjump_assemble(int i, const struct regstat *i_regs)
5378{
5379 int temp;
5380 int rs,cc;
5381 int ra_done=0;
5382 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5383 assert(rs>=0);
5384 if (ds_writes_rjump_rs(i)) {
5385 // Delay slot abuse, make a copy of the branch address register
5386 temp=get_reg(branch_regs[i].regmap,RTEMP);
5387 assert(temp>=0);
5388 assert(regs[i].regmap[temp]==RTEMP);
5389 emit_mov(rs,temp);
5390 rs=temp;
5391 }
5392 address_generation(i+1,i_regs,regs[i].regmap_entry);
5393 #ifdef REG_PREFETCH
5394 if(dops[i].rt1==31)
5395 {
5396 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5397 signed char *i_regmap=i_regs->regmap;
5398 int return_address=start+i*4+8;
5399 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5400 }
5401 }
5402 #endif
5403 #ifdef USE_MINI_HT
5404 if(dops[i].rs1==31) {
5405 int rh=get_reg(regs[i].regmap,RHASH);
5406 if(rh>=0) do_preload_rhash(rh);
5407 }
5408 #endif
5409 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5410 rjump_assemble_write_ra(i);
5411 ra_done=1;
5412 }
5413 ds_assemble(i+1,i_regs);
5414 uint64_t bc_unneeded=branch_regs[i].u;
5415 bc_unneeded|=1|(1LL<<dops[i].rt1);
5416 bc_unneeded&=~(1LL<<dops[i].rs1);
5417 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5418 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5419 if(!ra_done&&dops[i].rt1!=0)
5420 rjump_assemble_write_ra(i);
5421 cc=get_reg(branch_regs[i].regmap,CCREG);
5422 assert(cc==HOST_CCREG);
5423 (void)cc;
5424 #ifdef USE_MINI_HT
5425 int rh=get_reg(branch_regs[i].regmap,RHASH);
5426 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5427 if(dops[i].rs1==31) {
5428 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5429 do_preload_rhtbl(ht);
5430 do_rhash(rs,rh);
5431 }
5432 #endif
5433 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5434 #ifdef DESTRUCTIVE_WRITEBACK
5435 if((branch_regs[i].dirty>>rs)&1) {
5436 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5437 emit_loadreg(dops[i].rs1,rs);
5438 }
5439 }
5440 #endif
5441 #ifdef REG_PREFETCH
5442 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5443 #endif
5444 #ifdef USE_MINI_HT
5445 if(dops[i].rs1==31) {
5446 do_miniht_load(ht,rh);
5447 }
5448 #endif
5449 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5450 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5451 //assert(adj==0);
5452 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5453 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5454 if(dops[i+1].itype==COP0 && dops[i+1].opcode2==0x10)
5455 // special case for RFE
5456 emit_jmp(0);
5457 else
5458 emit_jns(0);
5459 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5460 #ifdef USE_MINI_HT
5461 if(dops[i].rs1==31) {
5462 do_miniht_jump(rs,rh,ht);
5463 }
5464 else
5465 #endif
5466 {
5467 do_jump_vaddr(rs);
5468 }
5469 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5470 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5471 #endif
5472}
5473
5474static void cjump_assemble(int i, const struct regstat *i_regs)
5475{
5476 const signed char *i_regmap = i_regs->regmap;
5477 int cc;
5478 int match;
5479 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5480 assem_debug("match=%d\n",match);
5481 int s1l,s2l;
5482 int unconditional=0,nop=0;
5483 int invert=0;
5484 int internal=internal_branch(ba[i]);
5485 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5486 if(!match) invert=1;
5487 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5488 if(i>(ba[i]-start)>>2) invert=1;
5489 #endif
5490 #ifdef __aarch64__
5491 invert=1; // because of near cond. branches
5492 #endif
5493
5494 if(dops[i].ooo) {
5495 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5496 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5497 }
5498 else {
5499 s1l=get_reg(i_regmap,dops[i].rs1);
5500 s2l=get_reg(i_regmap,dops[i].rs2);
5501 }
5502 if(dops[i].rs1==0&&dops[i].rs2==0)
5503 {
5504 if(dops[i].opcode&1) nop=1;
5505 else unconditional=1;
5506 //assert(dops[i].opcode!=5);
5507 //assert(dops[i].opcode!=7);
5508 //assert(dops[i].opcode!=0x15);
5509 //assert(dops[i].opcode!=0x17);
5510 }
5511 else if(dops[i].rs1==0)
5512 {
5513 s1l=s2l;
5514 s2l=-1;
5515 }
5516 else if(dops[i].rs2==0)
5517 {
5518 s2l=-1;
5519 }
5520
5521 if(dops[i].ooo) {
5522 // Out of order execution (delay slot first)
5523 //printf("OOOE\n");
5524 address_generation(i+1,i_regs,regs[i].regmap_entry);
5525 ds_assemble(i+1,i_regs);
5526 int adj;
5527 uint64_t bc_unneeded=branch_regs[i].u;
5528 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5529 bc_unneeded|=1;
5530 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5531 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5532 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5533 cc=get_reg(branch_regs[i].regmap,CCREG);
5534 assert(cc==HOST_CCREG);
5535 if(unconditional)
5536 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5537 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5538 //assem_debug("cycle count (adj)\n");
5539 if(unconditional) {
5540 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5541 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5542 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5543 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5544 if(internal)
5545 assem_debug("branch: internal\n");
5546 else
5547 assem_debug("branch: external\n");
5548 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5549 ds_assemble_entry(i);
5550 }
5551 else {
5552 add_to_linker(out,ba[i],internal);
5553 emit_jmp(0);
5554 }
5555 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5556 if(((u_int)out)&7) emit_addnop(0);
5557 #endif
5558 }
5559 }
5560 else if(nop) {
5561 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5562 void *jaddr=out;
5563 emit_jns(0);
5564 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5565 }
5566 else {
5567 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5568 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5569 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5570
5571 //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]);
5572 assert(s1l>=0);
5573 if(dops[i].opcode==4) // BEQ
5574 {
5575 if(s2l>=0) emit_cmp(s1l,s2l);
5576 else emit_test(s1l,s1l);
5577 if(invert){
5578 nottaken=out;
5579 emit_jne(DJT_1);
5580 }else{
5581 add_to_linker(out,ba[i],internal);
5582 emit_jeq(0);
5583 }
5584 }
5585 if(dops[i].opcode==5) // BNE
5586 {
5587 if(s2l>=0) emit_cmp(s1l,s2l);
5588 else emit_test(s1l,s1l);
5589 if(invert){
5590 nottaken=out;
5591 emit_jeq(DJT_1);
5592 }else{
5593 add_to_linker(out,ba[i],internal);
5594 emit_jne(0);
5595 }
5596 }
5597 if(dops[i].opcode==6) // BLEZ
5598 {
5599 emit_cmpimm(s1l,1);
5600 if(invert){
5601 nottaken=out;
5602 emit_jge(DJT_1);
5603 }else{
5604 add_to_linker(out,ba[i],internal);
5605 emit_jl(0);
5606 }
5607 }
5608 if(dops[i].opcode==7) // BGTZ
5609 {
5610 emit_cmpimm(s1l,1);
5611 if(invert){
5612 nottaken=out;
5613 emit_jl(DJT_1);
5614 }else{
5615 add_to_linker(out,ba[i],internal);
5616 emit_jge(0);
5617 }
5618 }
5619 if(invert) {
5620 if(taken) set_jump_target(taken, out);
5621 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5622 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5623 if(adj) {
5624 emit_addimm(cc,-adj,cc);
5625 add_to_linker(out,ba[i],internal);
5626 }else{
5627 emit_addnop(13);
5628 add_to_linker(out,ba[i],internal*2);
5629 }
5630 emit_jmp(0);
5631 }else
5632 #endif
5633 {
5634 if(adj) emit_addimm(cc,-adj,cc);
5635 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5636 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5637 if(internal)
5638 assem_debug("branch: internal\n");
5639 else
5640 assem_debug("branch: external\n");
5641 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5642 ds_assemble_entry(i);
5643 }
5644 else {
5645 add_to_linker(out,ba[i],internal);
5646 emit_jmp(0);
5647 }
5648 }
5649 set_jump_target(nottaken, out);
5650 }
5651
5652 if(nottaken1) set_jump_target(nottaken1, out);
5653 if(adj) {
5654 if(!invert) emit_addimm(cc,adj,cc);
5655 }
5656 } // (!unconditional)
5657 } // if(ooo)
5658 else
5659 {
5660 // In-order execution (branch first)
5661 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5662 if(!unconditional&&!nop) {
5663 //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]);
5664 assert(s1l>=0);
5665 if((dops[i].opcode&0x2f)==4) // BEQ
5666 {
5667 if(s2l>=0) emit_cmp(s1l,s2l);
5668 else emit_test(s1l,s1l);
5669 nottaken=out;
5670 emit_jne(DJT_2);
5671 }
5672 if((dops[i].opcode&0x2f)==5) // BNE
5673 {
5674 if(s2l>=0) emit_cmp(s1l,s2l);
5675 else emit_test(s1l,s1l);
5676 nottaken=out;
5677 emit_jeq(DJT_2);
5678 }
5679 if((dops[i].opcode&0x2f)==6) // BLEZ
5680 {
5681 emit_cmpimm(s1l,1);
5682 nottaken=out;
5683 emit_jge(DJT_2);
5684 }
5685 if((dops[i].opcode&0x2f)==7) // BGTZ
5686 {
5687 emit_cmpimm(s1l,1);
5688 nottaken=out;
5689 emit_jl(DJT_2);
5690 }
5691 } // if(!unconditional)
5692 int adj;
5693 uint64_t ds_unneeded=branch_regs[i].u;
5694 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5695 ds_unneeded|=1;
5696 // branch taken
5697 if(!nop) {
5698 if(taken) set_jump_target(taken, out);
5699 assem_debug("1:\n");
5700 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5701 // load regs
5702 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5703 address_generation(i+1,&branch_regs[i],0);
5704 if (ram_offset)
5705 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5706 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5707 ds_assemble(i+1,&branch_regs[i]);
5708 cc=get_reg(branch_regs[i].regmap,CCREG);
5709 if(cc==-1) {
5710 emit_loadreg(CCREG,cc=HOST_CCREG);
5711 // CHECK: Is the following instruction (fall thru) allocated ok?
5712 }
5713 assert(cc==HOST_CCREG);
5714 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5715 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5716 assem_debug("cycle count (adj)\n");
5717 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5718 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5719 if(internal)
5720 assem_debug("branch: internal\n");
5721 else
5722 assem_debug("branch: external\n");
5723 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5724 ds_assemble_entry(i);
5725 }
5726 else {
5727 add_to_linker(out,ba[i],internal);
5728 emit_jmp(0);
5729 }
5730 }
5731 // branch not taken
5732 if(!unconditional) {
5733 if(nottaken1) set_jump_target(nottaken1, out);
5734 set_jump_target(nottaken, out);
5735 assem_debug("2:\n");
5736 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5737 // load regs
5738 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5739 address_generation(i+1,&branch_regs[i],0);
5740 if (ram_offset)
5741 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5742 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5743 ds_assemble(i+1,&branch_regs[i]);
5744 cc=get_reg(branch_regs[i].regmap,CCREG);
5745 if (cc == -1) {
5746 // Cycle count isn't in a register, temporarily load it then write it out
5747 emit_loadreg(CCREG,HOST_CCREG);
5748 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5749 void *jaddr=out;
5750 emit_jns(0);
5751 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5752 emit_storereg(CCREG,HOST_CCREG);
5753 }
5754 else{
5755 cc=get_reg(i_regmap,CCREG);
5756 assert(cc==HOST_CCREG);
5757 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5758 void *jaddr=out;
5759 emit_jns(0);
5760 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5761 }
5762 }
5763 }
5764}
5765
5766static void sjump_assemble(int i, const struct regstat *i_regs)
5767{
5768 const signed char *i_regmap = i_regs->regmap;
5769 int cc;
5770 int match;
5771 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5772 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5773 int s1l;
5774 int unconditional=0,nevertaken=0;
5775 int invert=0;
5776 int internal=internal_branch(ba[i]);
5777 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5778 if(!match) invert=1;
5779 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5780 if(i>(ba[i]-start)>>2) invert=1;
5781 #endif
5782 #ifdef __aarch64__
5783 invert=1; // because of near cond. branches
5784 #endif
5785
5786 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5787 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5788
5789 if(dops[i].ooo) {
5790 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5791 }
5792 else {
5793 s1l=get_reg(i_regmap,dops[i].rs1);
5794 }
5795 if(dops[i].rs1==0)
5796 {
5797 if(dops[i].opcode2&1) unconditional=1;
5798 else nevertaken=1;
5799 // These are never taken (r0 is never less than zero)
5800 //assert(dops[i].opcode2!=0);
5801 //assert(dops[i].opcode2!=2);
5802 //assert(dops[i].opcode2!=0x10);
5803 //assert(dops[i].opcode2!=0x12);
5804 }
5805
5806 if(dops[i].ooo) {
5807 // Out of order execution (delay slot first)
5808 //printf("OOOE\n");
5809 address_generation(i+1,i_regs,regs[i].regmap_entry);
5810 ds_assemble(i+1,i_regs);
5811 int adj;
5812 uint64_t bc_unneeded=branch_regs[i].u;
5813 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5814 bc_unneeded|=1;
5815 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5816 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5817 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5818 if(dops[i].rt1==31) {
5819 int rt,return_address;
5820 rt=get_reg(branch_regs[i].regmap,31);
5821 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]);
5822 if(rt>=0) {
5823 // Save the PC even if the branch is not taken
5824 return_address=start+i*4+8;
5825 emit_movimm(return_address,rt); // PC into link register
5826 #ifdef IMM_PREFETCH
5827 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5828 #endif
5829 }
5830 }
5831 cc=get_reg(branch_regs[i].regmap,CCREG);
5832 assert(cc==HOST_CCREG);
5833 if(unconditional)
5834 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5835 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5836 assem_debug("cycle count (adj)\n");
5837 if(unconditional) {
5838 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5839 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5840 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5841 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5842 if(internal)
5843 assem_debug("branch: internal\n");
5844 else
5845 assem_debug("branch: external\n");
5846 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5847 ds_assemble_entry(i);
5848 }
5849 else {
5850 add_to_linker(out,ba[i],internal);
5851 emit_jmp(0);
5852 }
5853 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5854 if(((u_int)out)&7) emit_addnop(0);
5855 #endif
5856 }
5857 }
5858 else if(nevertaken) {
5859 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5860 void *jaddr=out;
5861 emit_jns(0);
5862 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5863 }
5864 else {
5865 void *nottaken = NULL;
5866 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5867 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5868 {
5869 assert(s1l>=0);
5870 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5871 {
5872 emit_test(s1l,s1l);
5873 if(invert){
5874 nottaken=out;
5875 emit_jns(DJT_1);
5876 }else{
5877 add_to_linker(out,ba[i],internal);
5878 emit_js(0);
5879 }
5880 }
5881 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5882 {
5883 emit_test(s1l,s1l);
5884 if(invert){
5885 nottaken=out;
5886 emit_js(DJT_1);
5887 }else{
5888 add_to_linker(out,ba[i],internal);
5889 emit_jns(0);
5890 }
5891 }
5892 }
5893
5894 if(invert) {
5895 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5896 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5897 if(adj) {
5898 emit_addimm(cc,-adj,cc);
5899 add_to_linker(out,ba[i],internal);
5900 }else{
5901 emit_addnop(13);
5902 add_to_linker(out,ba[i],internal*2);
5903 }
5904 emit_jmp(0);
5905 }else
5906 #endif
5907 {
5908 if(adj) emit_addimm(cc,-adj,cc);
5909 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5910 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5911 if(internal)
5912 assem_debug("branch: internal\n");
5913 else
5914 assem_debug("branch: external\n");
5915 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5916 ds_assemble_entry(i);
5917 }
5918 else {
5919 add_to_linker(out,ba[i],internal);
5920 emit_jmp(0);
5921 }
5922 }
5923 set_jump_target(nottaken, out);
5924 }
5925
5926 if(adj) {
5927 if(!invert) emit_addimm(cc,adj,cc);
5928 }
5929 } // (!unconditional)
5930 } // if(ooo)
5931 else
5932 {
5933 // In-order execution (branch first)
5934 //printf("IOE\n");
5935 void *nottaken = NULL;
5936 if(dops[i].rt1==31) {
5937 int rt,return_address;
5938 rt=get_reg(branch_regs[i].regmap,31);
5939 if(rt>=0) {
5940 // Save the PC even if the branch is not taken
5941 return_address=start+i*4+8;
5942 emit_movimm(return_address,rt); // PC into link register
5943 #ifdef IMM_PREFETCH
5944 emit_prefetch(hash_table_get(return_address));
5945 #endif
5946 }
5947 }
5948 if(!unconditional) {
5949 //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]);
5950 assert(s1l>=0);
5951 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5952 {
5953 emit_test(s1l,s1l);
5954 nottaken=out;
5955 emit_jns(DJT_1);
5956 }
5957 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5958 {
5959 emit_test(s1l,s1l);
5960 nottaken=out;
5961 emit_js(DJT_1);
5962 }
5963 } // if(!unconditional)
5964 int adj;
5965 uint64_t ds_unneeded=branch_regs[i].u;
5966 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5967 ds_unneeded|=1;
5968 // branch taken
5969 if(!nevertaken) {
5970 //assem_debug("1:\n");
5971 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5972 // load regs
5973 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5974 address_generation(i+1,&branch_regs[i],0);
5975 if (ram_offset)
5976 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5977 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5978 ds_assemble(i+1,&branch_regs[i]);
5979 cc=get_reg(branch_regs[i].regmap,CCREG);
5980 if(cc==-1) {
5981 emit_loadreg(CCREG,cc=HOST_CCREG);
5982 // CHECK: Is the following instruction (fall thru) allocated ok?
5983 }
5984 assert(cc==HOST_CCREG);
5985 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5986 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5987 assem_debug("cycle count (adj)\n");
5988 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5989 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5990 if(internal)
5991 assem_debug("branch: internal\n");
5992 else
5993 assem_debug("branch: external\n");
5994 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5995 ds_assemble_entry(i);
5996 }
5997 else {
5998 add_to_linker(out,ba[i],internal);
5999 emit_jmp(0);
6000 }
6001 }
6002 // branch not taken
6003 if(!unconditional) {
6004 set_jump_target(nottaken, out);
6005 assem_debug("1:\n");
6006 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6007 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6008 address_generation(i+1,&branch_regs[i],0);
6009 if (ram_offset)
6010 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6011 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6012 ds_assemble(i+1,&branch_regs[i]);
6013 cc=get_reg(branch_regs[i].regmap,CCREG);
6014 if (cc == -1) {
6015 // Cycle count isn't in a register, temporarily load it then write it out
6016 emit_loadreg(CCREG,HOST_CCREG);
6017 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
6018 void *jaddr=out;
6019 emit_jns(0);
6020 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6021 emit_storereg(CCREG,HOST_CCREG);
6022 }
6023 else{
6024 cc=get_reg(i_regmap,CCREG);
6025 assert(cc==HOST_CCREG);
6026 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
6027 void *jaddr=out;
6028 emit_jns(0);
6029 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6030 }
6031 }
6032 }
6033}
6034
6035static void check_regmap(signed char *regmap)
6036{
6037#ifndef NDEBUG
6038 int i,j;
6039 for (i = 0; i < HOST_REGS; i++) {
6040 if (regmap[i] < 0)
6041 continue;
6042 for (j = i + 1; j < HOST_REGS; j++)
6043 assert(regmap[i] != regmap[j]);
6044 }
6045#endif
6046}
6047
6048#ifdef DISASM
6049#include <inttypes.h>
6050static char insn[MAXBLOCK][10];
6051
6052#define set_mnemonic(i_, n_) \
6053 strcpy(insn[i_], n_)
6054
6055void print_regmap(const char *name, const signed char *regmap)
6056{
6057 char buf[5];
6058 int i, l;
6059 fputs(name, stdout);
6060 for (i = 0; i < HOST_REGS; i++) {
6061 l = 0;
6062 if (regmap[i] >= 0)
6063 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6064 for (; l < 3; l++)
6065 buf[l] = ' ';
6066 buf[l] = 0;
6067 printf(" r%d=%s", i, buf);
6068 }
6069 fputs("\n", stdout);
6070}
6071
6072 /* disassembly */
6073void disassemble_inst(int i)
6074{
6075 if (dops[i].bt) printf("*"); else printf(" ");
6076 switch(dops[i].itype) {
6077 case UJUMP:
6078 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6079 case CJUMP:
6080 printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):*ba);break;
6081 case SJUMP:
6082 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;
6083 case RJUMP:
6084 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6085 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6086 else
6087 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6088 break;
6089 case IMM16:
6090 if(dops[i].opcode==0xf) //LUI
6091 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6092 else
6093 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6094 break;
6095 case LOAD:
6096 case LOADLR:
6097 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6098 break;
6099 case STORE:
6100 case STORELR:
6101 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6102 break;
6103 case ALU:
6104 case SHIFT:
6105 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6106 break;
6107 case MULTDIV:
6108 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6109 break;
6110 case SHIFTIMM:
6111 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6112 break;
6113 case MOV:
6114 if((dops[i].opcode2&0x1d)==0x10)
6115 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6116 else if((dops[i].opcode2&0x1d)==0x11)
6117 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6118 else
6119 printf (" %x: %s\n",start+i*4,insn[i]);
6120 break;
6121 case COP0:
6122 if(dops[i].opcode2==0)
6123 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6124 else if(dops[i].opcode2==4)
6125 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6126 else printf (" %x: %s\n",start+i*4,insn[i]);
6127 break;
6128 case COP1:
6129 if(dops[i].opcode2<3)
6130 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6131 else if(dops[i].opcode2>3)
6132 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6133 else printf (" %x: %s\n",start+i*4,insn[i]);
6134 break;
6135 case COP2:
6136 if(dops[i].opcode2<3)
6137 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6138 else if(dops[i].opcode2>3)
6139 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6140 else printf (" %x: %s\n",start+i*4,insn[i]);
6141 break;
6142 case C1LS:
6143 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6144 break;
6145 case C2LS:
6146 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6147 break;
6148 case INTCALL:
6149 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6150 break;
6151 default:
6152 //printf (" %s %8x\n",insn[i],source[i]);
6153 printf (" %x: %s\n",start+i*4,insn[i]);
6154 }
6155 return;
6156 printf("D: %"PRIu64" WD: %"PRIu64" U: %"PRIu64"\n",
6157 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i]);
6158 print_regmap("pre: ", regmap_pre[i]);
6159 print_regmap("entry: ", regs[i].regmap_entry);
6160 print_regmap("map: ", regs[i].regmap);
6161 if (dops[i].is_jump) {
6162 print_regmap("bentry:", branch_regs[i].regmap_entry);
6163 print_regmap("bmap: ", branch_regs[i].regmap);
6164 }
6165}
6166#else
6167#define set_mnemonic(i_, n_)
6168static void disassemble_inst(int i) {}
6169#endif // DISASM
6170
6171#define DRC_TEST_VAL 0x74657374
6172
6173static noinline void new_dynarec_test(void)
6174{
6175 int (*testfunc)(void);
6176 void *beginning;
6177 int ret[2];
6178 size_t i;
6179
6180 // check structure linkage
6181 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6182 {
6183 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6184 }
6185
6186 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6187 new_dynarec_test, out);
6188 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6189
6190 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6191 out = ndrc->translation_cache;
6192 beginning = start_block();
6193 emit_movimm(DRC_TEST_VAL + i, 0); // test
6194 emit_ret();
6195 literal_pool(0);
6196 end_block(beginning);
6197 testfunc = beginning;
6198 ret[i] = testfunc();
6199 }
6200
6201 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6202 SysPrintf("test passed.\n");
6203 else
6204 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6205 out = ndrc->translation_cache;
6206}
6207
6208// clear the state completely, instead of just marking
6209// things invalid like invalidate_all_pages() does
6210void new_dynarec_clear_full(void)
6211{
6212 int n;
6213 out = ndrc->translation_cache;
6214 memset(invalid_code,1,sizeof(invalid_code));
6215 memset(hash_table,0xff,sizeof(hash_table));
6216 memset(mini_ht,-1,sizeof(mini_ht));
6217 memset(shadow,0,sizeof(shadow));
6218 copy=shadow;
6219 expirep = EXPIRITY_OFFSET;
6220 pending_exception=0;
6221 literalcount=0;
6222 stop_after_jal=0;
6223 inv_code_start=inv_code_end=~0;
6224 hack_addr=0;
6225 f1_hack=0;
6226 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6227 blocks_clear(&blocks[n]);
6228 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6229 free(jumps[n]);
6230 jumps[n] = NULL;
6231 }
6232 stat_clear(stat_blocks);
6233 stat_clear(stat_links);
6234
6235 cycle_multiplier_old = Config.cycle_multiplier;
6236 new_dynarec_hacks_old = new_dynarec_hacks;
6237}
6238
6239void new_dynarec_init(void)
6240{
6241 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6242
6243#ifdef _3DS
6244 check_rosalina();
6245#endif
6246#ifdef BASE_ADDR_DYNAMIC
6247 #ifdef VITA
6248 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6249 if (sceBlock <= 0)
6250 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6251 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6252 if (ret < 0)
6253 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6254 sceKernelOpenVMDomain();
6255 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6256 #elif defined(_MSC_VER)
6257 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6258 PAGE_EXECUTE_READWRITE);
6259 #elif defined(HAVE_LIBNX)
6260 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6261 if (R_FAILED(rc))
6262 SysPrintf("jitCreate failed: %08x\n", rc);
6263 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6264 jitTransitionToWritable(&g_jit);
6265 ndrc = g_jit.rx_addr;
6266 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6267 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6268 #else
6269 uintptr_t desired_addr = 0;
6270 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6271 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6272 int fd = -1;
6273 #ifdef __ELF__
6274 extern char _end;
6275 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6276 #endif
6277 #ifdef TC_WRITE_OFFSET
6278 // mostly for testing
6279 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6280 ftruncate(fd, sizeof(*ndrc));
6281 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6282 (flags = MAP_SHARED), fd, 0);
6283 assert(mw != MAP_FAILED);
6284 prot = PROT_READ | PROT_EXEC;
6285 #endif
6286 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6287 if (ndrc == MAP_FAILED) {
6288 SysPrintf("mmap() failed: %s\n", strerror(errno));
6289 abort();
6290 }
6291 #ifdef TC_WRITE_OFFSET
6292 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6293 #endif
6294 #endif
6295#else
6296 #ifndef NO_WRITE_EXEC
6297 // not all systems allow execute in data segment by default
6298 // size must be 4K aligned for 3DS?
6299 if (mprotect(ndrc, sizeof(*ndrc),
6300 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6301 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6302 #endif
6303#endif
6304 out = ndrc->translation_cache;
6305 new_dynarec_clear_full();
6306#ifdef HOST_IMM8
6307 // Copy this into local area so we don't have to put it in every literal pool
6308 invc_ptr=invalid_code;
6309#endif
6310 arch_init();
6311 new_dynarec_test();
6312 ram_offset=(uintptr_t)rdram-0x80000000;
6313 if (ram_offset!=0)
6314 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6315 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6316 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6317}
6318
6319void new_dynarec_cleanup(void)
6320{
6321 int n;
6322#ifdef BASE_ADDR_DYNAMIC
6323 #ifdef VITA
6324 // sceBlock is managed by retroarch's bootstrap code
6325 //sceKernelFreeMemBlock(sceBlock);
6326 //sceBlock = -1;
6327 #elif defined(HAVE_LIBNX)
6328 jitClose(&g_jit);
6329 ndrc = NULL;
6330 #else
6331 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6332 SysPrintf("munmap() failed\n");
6333 ndrc = NULL;
6334 #endif
6335#endif
6336 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6337 blocks_clear(&blocks[n]);
6338 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6339 free(jumps[n]);
6340 jumps[n] = NULL;
6341 }
6342 stat_clear(stat_blocks);
6343 stat_clear(stat_links);
6344 new_dynarec_print_stats();
6345}
6346
6347static u_int *get_source_start(u_int addr, u_int *limit)
6348{
6349 if (addr < 0x00200000 ||
6350 (0xa0000000 <= addr && addr < 0xa0200000))
6351 {
6352 // used for BIOS calls mostly?
6353 *limit = (addr&0xa0000000)|0x00200000;
6354 return (u_int *)(rdram + (addr&0x1fffff));
6355 }
6356 else if (!Config.HLE && (
6357 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6358 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6359 {
6360 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6361 // but timings in PCSX are too tied to the interpreter's 2-per-insn assumption
6362 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6363 cycle_multiplier_active = 200;
6364
6365 *limit = (addr & 0xfff00000) | 0x80000;
6366 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6367 }
6368 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6369 *limit = (addr & 0x80600000) + 0x00200000;
6370 return (u_int *)(rdram + (addr&0x1fffff));
6371 }
6372 return NULL;
6373}
6374
6375static u_int scan_for_ret(u_int addr)
6376{
6377 u_int limit = 0;
6378 u_int *mem;
6379
6380 mem = get_source_start(addr, &limit);
6381 if (mem == NULL)
6382 return addr;
6383
6384 if (limit > addr + 0x1000)
6385 limit = addr + 0x1000;
6386 for (; addr < limit; addr += 4, mem++) {
6387 if (*mem == 0x03e00008) // jr $ra
6388 return addr + 8;
6389 }
6390 return addr;
6391}
6392
6393struct savestate_block {
6394 uint32_t addr;
6395 uint32_t regflags;
6396};
6397
6398static int addr_cmp(const void *p1_, const void *p2_)
6399{
6400 const struct savestate_block *p1 = p1_, *p2 = p2_;
6401 return p1->addr - p2->addr;
6402}
6403
6404int new_dynarec_save_blocks(void *save, int size)
6405{
6406 struct savestate_block *sblocks = save;
6407 int maxcount = size / sizeof(sblocks[0]);
6408 struct savestate_block tmp_blocks[1024];
6409 struct block_info *block;
6410 int p, s, d, o, bcnt;
6411 u_int addr;
6412
6413 o = 0;
6414 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6415 bcnt = 0;
6416 for (block = blocks[p]; block != NULL; block = block->next) {
6417 if (block->is_dirty)
6418 continue;
6419 tmp_blocks[bcnt].addr = block->start;
6420 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6421 bcnt++;
6422 }
6423 if (bcnt < 1)
6424 continue;
6425 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6426
6427 addr = tmp_blocks[0].addr;
6428 for (s = d = 0; s < bcnt; s++) {
6429 if (tmp_blocks[s].addr < addr)
6430 continue;
6431 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6432 tmp_blocks[d++] = tmp_blocks[s];
6433 addr = scan_for_ret(tmp_blocks[s].addr);
6434 }
6435
6436 if (o + d > maxcount)
6437 d = maxcount - o;
6438 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6439 o += d;
6440 }
6441
6442 return o * sizeof(sblocks[0]);
6443}
6444
6445void new_dynarec_load_blocks(const void *save, int size)
6446{
6447 const struct savestate_block *sblocks = save;
6448 int count = size / sizeof(sblocks[0]);
6449 struct block_info *block;
6450 u_int regs_save[32];
6451 u_int page;
6452 uint32_t f;
6453 int i, b;
6454
6455 // restore clean blocks, if any
6456 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6457 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6458 if (!block->is_dirty)
6459 continue;
6460 assert(block->source && block->copy);
6461 if (memcmp(block->source, block->copy, block->len))
6462 continue;
6463
6464 // see try_restore_block
6465 block->is_dirty = 0;
6466 mark_invalid_code(block->start, block->len, 0);
6467 i++;
6468 }
6469 }
6470 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6471
6472 // change GPRs for speculation to at least partially work..
6473 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6474 for (i = 1; i < 32; i++)
6475 psxRegs.GPR.r[i] = 0x80000000;
6476
6477 for (b = 0; b < count; b++) {
6478 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6479 if (f & 1)
6480 psxRegs.GPR.r[i] = 0x1f800000;
6481 }
6482
6483 ndrc_get_addr_ht(sblocks[b].addr);
6484
6485 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6486 if (f & 1)
6487 psxRegs.GPR.r[i] = 0x80000000;
6488 }
6489 }
6490
6491 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6492}
6493
6494void new_dynarec_print_stats(void)
6495{
6496#ifdef STAT_PRINT
6497 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6498 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6499 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6500 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6501 out - ndrc->translation_cache, stat_blocks, stat_links);
6502 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6503 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6504 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6505#endif
6506}
6507
6508static int apply_hacks(void)
6509{
6510 int i;
6511 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6512 return 0;
6513 /* special hack(s) */
6514 for (i = 0; i < slen - 4; i++)
6515 {
6516 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6517 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6518 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6519 && imm[i+3] == 0x6e40 && dops[i+3].rs1 == 2)
6520 {
6521 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6522 dops[i + 3].itype = NOP;
6523 }
6524 }
6525 i = slen;
6526 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6527 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6528 && dops[i-7].itype == STORE)
6529 {
6530 i = i-8;
6531 if (dops[i].itype == IMM16)
6532 i--;
6533 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6534 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6535 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6536 {
6537 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6538 f1_hack = 1;
6539 return 1;
6540 }
6541 }
6542 return 0;
6543}
6544
6545static noinline void pass1_disassemble(u_int pagelimit)
6546{
6547 int i, j, done = 0, ni_count = 0;
6548 unsigned int type,op,op2;
6549
6550 for (i = 0; !done; i++)
6551 {
6552 memset(&dops[i], 0, sizeof(dops[i]));
6553 op2=0;
6554 minimum_free_regs[i]=0;
6555 dops[i].opcode=op=source[i]>>26;
6556 switch(op)
6557 {
6558 case 0x00: set_mnemonic(i, "special"); type=NI;
6559 op2=source[i]&0x3f;
6560 switch(op2)
6561 {
6562 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6563 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6564 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6565 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6566 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6567 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6568 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6569 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6570 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6571 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6572 case 0x0F: set_mnemonic(i, "SYNC"); type=OTHER; break;
6573 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6574 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6575 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6576 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6577 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6578 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6579 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6580 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6581 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6582 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6583 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6584 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6585 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6586 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6587 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6588 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6589 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6590 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6591 case 0x30: set_mnemonic(i, "TGE"); type=NI; break;
6592 case 0x31: set_mnemonic(i, "TGEU"); type=NI; break;
6593 case 0x32: set_mnemonic(i, "TLT"); type=NI; break;
6594 case 0x33: set_mnemonic(i, "TLTU"); type=NI; break;
6595 case 0x34: set_mnemonic(i, "TEQ"); type=NI; break;
6596 case 0x36: set_mnemonic(i, "TNE"); type=NI; break;
6597#if 0
6598 case 0x14: set_mnemonic(i, "DSLLV"); type=SHIFT; break;
6599 case 0x16: set_mnemonic(i, "DSRLV"); type=SHIFT; break;
6600 case 0x17: set_mnemonic(i, "DSRAV"); type=SHIFT; break;
6601 case 0x1C: set_mnemonic(i, "DMULT"); type=MULTDIV; break;
6602 case 0x1D: set_mnemonic(i, "DMULTU"); type=MULTDIV; break;
6603 case 0x1E: set_mnemonic(i, "DDIV"); type=MULTDIV; break;
6604 case 0x1F: set_mnemonic(i, "DDIVU"); type=MULTDIV; break;
6605 case 0x2C: set_mnemonic(i, "DADD"); type=ALU; break;
6606 case 0x2D: set_mnemonic(i, "DADDU"); type=ALU; break;
6607 case 0x2E: set_mnemonic(i, "DSUB"); type=ALU; break;
6608 case 0x2F: set_mnemonic(i, "DSUBU"); type=ALU; break;
6609 case 0x38: set_mnemonic(i, "DSLL"); type=SHIFTIMM; break;
6610 case 0x3A: set_mnemonic(i, "DSRL"); type=SHIFTIMM; break;
6611 case 0x3B: set_mnemonic(i, "DSRA"); type=SHIFTIMM; break;
6612 case 0x3C: set_mnemonic(i, "DSLL32"); type=SHIFTIMM; break;
6613 case 0x3E: set_mnemonic(i, "DSRL32"); type=SHIFTIMM; break;
6614 case 0x3F: set_mnemonic(i, "DSRA32"); type=SHIFTIMM; break;
6615#endif
6616 }
6617 break;
6618 case 0x01: set_mnemonic(i, "regimm"); type=NI;
6619 op2=(source[i]>>16)&0x1f;
6620 switch(op2)
6621 {
6622 case 0x00: set_mnemonic(i, "BLTZ"); type=SJUMP; break;
6623 case 0x01: set_mnemonic(i, "BGEZ"); type=SJUMP; break;
6624 //case 0x02: set_mnemonic(i, "BLTZL"); type=SJUMP; break;
6625 //case 0x03: set_mnemonic(i, "BGEZL"); type=SJUMP; break;
6626 //case 0x08: set_mnemonic(i, "TGEI"); type=NI; break;
6627 //case 0x09: set_mnemonic(i, "TGEIU"); type=NI; break;
6628 //case 0x0A: set_mnemonic(i, "TLTI"); type=NI; break;
6629 //case 0x0B: set_mnemonic(i, "TLTIU"); type=NI; break;
6630 //case 0x0C: set_mnemonic(i, "TEQI"); type=NI; break;
6631 //case 0x0E: set_mnemonic(i, "TNEI"); type=NI; break;
6632 case 0x10: set_mnemonic(i, "BLTZAL"); type=SJUMP; break;
6633 case 0x11: set_mnemonic(i, "BGEZAL"); type=SJUMP; break;
6634 //case 0x12: set_mnemonic(i, "BLTZALL"); type=SJUMP; break;
6635 //case 0x13: set_mnemonic(i, "BGEZALL"); type=SJUMP; break;
6636 }
6637 break;
6638 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6639 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6640 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6641 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6642 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6643 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6644 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6645 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6646 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6647 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6648 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6649 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6650 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6651 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6652 case 0x10: set_mnemonic(i, "cop0"); type=NI;
6653 op2=(source[i]>>21)&0x1f;
6654 switch(op2)
6655 {
6656 case 0x00: set_mnemonic(i, "MFC0"); type=COP0; break;
6657 case 0x02: set_mnemonic(i, "CFC0"); type=COP0; break;
6658 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6659 case 0x06: set_mnemonic(i, "CTC0"); type=COP0; break;
6660 case 0x10: set_mnemonic(i, "RFE"); type=COP0; break;
6661 }
6662 break;
6663 case 0x11: set_mnemonic(i, "cop1"); type=COP1;
6664 op2=(source[i]>>21)&0x1f;
6665 break;
6666#if 0
6667 case 0x14: set_mnemonic(i, "BEQL"); type=CJUMP; break;
6668 case 0x15: set_mnemonic(i, "BNEL"); type=CJUMP; break;
6669 case 0x16: set_mnemonic(i, "BLEZL"); type=CJUMP; break;
6670 case 0x17: set_mnemonic(i, "BGTZL"); type=CJUMP; break;
6671 case 0x18: set_mnemonic(i, "DADDI"); type=IMM16; break;
6672 case 0x19: set_mnemonic(i, "DADDIU"); type=IMM16; break;
6673 case 0x1A: set_mnemonic(i, "LDL"); type=LOADLR; break;
6674 case 0x1B: set_mnemonic(i, "LDR"); type=LOADLR; break;
6675#endif
6676 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6677 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6678 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6679 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6680 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6681 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6682 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6683#if 0
6684 case 0x27: set_mnemonic(i, "LWU"); type=LOAD; break;
6685#endif
6686 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6687 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6688 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6689 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6690#if 0
6691 case 0x2C: set_mnemonic(i, "SDL"); type=STORELR; break;
6692 case 0x2D: set_mnemonic(i, "SDR"); type=STORELR; break;
6693#endif
6694 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6695 case 0x2F: set_mnemonic(i, "CACHE"); type=NOP; break;
6696 case 0x30: set_mnemonic(i, "LL"); type=NI; break;
6697 case 0x31: set_mnemonic(i, "LWC1"); type=C1LS; break;
6698#if 0
6699 case 0x34: set_mnemonic(i, "LLD"); type=NI; break;
6700 case 0x35: set_mnemonic(i, "LDC1"); type=C1LS; break;
6701 case 0x37: set_mnemonic(i, "LD"); type=LOAD; break;
6702#endif
6703 case 0x38: set_mnemonic(i, "SC"); type=NI; break;
6704 case 0x39: set_mnemonic(i, "SWC1"); type=C1LS; break;
6705#if 0
6706 case 0x3C: set_mnemonic(i, "SCD"); type=NI; break;
6707 case 0x3D: set_mnemonic(i, "SDC1"); type=C1LS; break;
6708 case 0x3F: set_mnemonic(i, "SD"); type=STORE; break;
6709#endif
6710 case 0x12: set_mnemonic(i, "COP2"); type=NI;
6711 op2=(source[i]>>21)&0x1f;
6712 //if (op2 & 0x10)
6713 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
6714 if (gte_handlers[source[i]&0x3f]!=NULL) {
6715#ifdef DISASM
6716 if (gte_regnames[source[i]&0x3f]!=NULL)
6717 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
6718 else
6719 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
6720#endif
6721 type=C2OP;
6722 }
6723 }
6724 else switch(op2)
6725 {
6726 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6727 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6728 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6729 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6730 }
6731 break;
6732 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6733 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6734 case 0x3B: set_mnemonic(i, "HLECALL"); type=HLECALL; break;
6735 default: set_mnemonic(i, "???"); type=NI;
6736 SysPrintf("NI %08x @%08x (%08x)\n", source[i], start + i*4, start);
6737 break;
6738 }
6739 dops[i].itype=type;
6740 dops[i].opcode2=op2;
6741 /* Get registers/immediates */
6742 dops[i].use_lt1=0;
6743 gte_rs[i]=gte_rt[i]=0;
6744 switch(type) {
6745 case LOAD:
6746 dops[i].rs1=(source[i]>>21)&0x1f;
6747 dops[i].rs2=0;
6748 dops[i].rt1=(source[i]>>16)&0x1f;
6749 dops[i].rt2=0;
6750 imm[i]=(short)source[i];
6751 break;
6752 case STORE:
6753 case STORELR:
6754 dops[i].rs1=(source[i]>>21)&0x1f;
6755 dops[i].rs2=(source[i]>>16)&0x1f;
6756 dops[i].rt1=0;
6757 dops[i].rt2=0;
6758 imm[i]=(short)source[i];
6759 break;
6760 case LOADLR:
6761 // LWL/LWR only load part of the register,
6762 // therefore the target register must be treated as a source too
6763 dops[i].rs1=(source[i]>>21)&0x1f;
6764 dops[i].rs2=(source[i]>>16)&0x1f;
6765 dops[i].rt1=(source[i]>>16)&0x1f;
6766 dops[i].rt2=0;
6767 imm[i]=(short)source[i];
6768 break;
6769 case IMM16:
6770 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6771 else dops[i].rs1=(source[i]>>21)&0x1f;
6772 dops[i].rs2=0;
6773 dops[i].rt1=(source[i]>>16)&0x1f;
6774 dops[i].rt2=0;
6775 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6776 imm[i]=(unsigned short)source[i];
6777 }else{
6778 imm[i]=(short)source[i];
6779 }
6780 break;
6781 case UJUMP:
6782 dops[i].rs1=0;
6783 dops[i].rs2=0;
6784 dops[i].rt1=0;
6785 dops[i].rt2=0;
6786 // The JAL instruction writes to r31.
6787 if (op&1) {
6788 dops[i].rt1=31;
6789 }
6790 dops[i].rs2=CCREG;
6791 break;
6792 case RJUMP:
6793 dops[i].rs1=(source[i]>>21)&0x1f;
6794 dops[i].rs2=0;
6795 dops[i].rt1=0;
6796 dops[i].rt2=0;
6797 // The JALR instruction writes to rd.
6798 if (op2&1) {
6799 dops[i].rt1=(source[i]>>11)&0x1f;
6800 }
6801 dops[i].rs2=CCREG;
6802 break;
6803 case CJUMP:
6804 dops[i].rs1=(source[i]>>21)&0x1f;
6805 dops[i].rs2=(source[i]>>16)&0x1f;
6806 dops[i].rt1=0;
6807 dops[i].rt2=0;
6808 if(op&2) { // BGTZ/BLEZ
6809 dops[i].rs2=0;
6810 }
6811 break;
6812 case SJUMP:
6813 dops[i].rs1=(source[i]>>21)&0x1f;
6814 dops[i].rs2=CCREG;
6815 dops[i].rt1=0;
6816 dops[i].rt2=0;
6817 if(op2&0x10) { // BxxAL
6818 dops[i].rt1=31;
6819 // NOTE: If the branch is not taken, r31 is still overwritten
6820 }
6821 break;
6822 case ALU:
6823 dops[i].rs1=(source[i]>>21)&0x1f; // source
6824 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
6825 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6826 dops[i].rt2=0;
6827 break;
6828 case MULTDIV:
6829 dops[i].rs1=(source[i]>>21)&0x1f; // source
6830 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
6831 dops[i].rt1=HIREG;
6832 dops[i].rt2=LOREG;
6833 break;
6834 case MOV:
6835 dops[i].rs1=0;
6836 dops[i].rs2=0;
6837 dops[i].rt1=0;
6838 dops[i].rt2=0;
6839 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6840 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6841 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6842 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6843 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
6844 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
6845 break;
6846 case SHIFT:
6847 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
6848 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
6849 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6850 dops[i].rt2=0;
6851 break;
6852 case SHIFTIMM:
6853 dops[i].rs1=(source[i]>>16)&0x1f;
6854 dops[i].rs2=0;
6855 dops[i].rt1=(source[i]>>11)&0x1f;
6856 dops[i].rt2=0;
6857 imm[i]=(source[i]>>6)&0x1f;
6858 // DSxx32 instructions
6859 if(op2>=0x3c) imm[i]|=0x20;
6860 break;
6861 case COP0:
6862 dops[i].rs1=0;
6863 dops[i].rs2=0;
6864 dops[i].rt1=0;
6865 dops[i].rt2=0;
6866 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
6867 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
6868 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
6869 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
6870 break;
6871 case COP1:
6872 dops[i].rs1=0;
6873 dops[i].rs2=0;
6874 dops[i].rt1=0;
6875 dops[i].rt2=0;
6876 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
6877 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
6878 dops[i].rs2=CSREG;
6879 break;
6880 case COP2:
6881 dops[i].rs1=0;
6882 dops[i].rs2=0;
6883 dops[i].rt1=0;
6884 dops[i].rt2=0;
6885 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
6886 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
6887 dops[i].rs2=CSREG;
6888 int gr=(source[i]>>11)&0x1F;
6889 switch(op2)
6890 {
6891 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6892 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6893 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6894 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6895 }
6896 break;
6897 case C1LS:
6898 dops[i].rs1=(source[i]>>21)&0x1F;
6899 dops[i].rs2=CSREG;
6900 dops[i].rt1=0;
6901 dops[i].rt2=0;
6902 imm[i]=(short)source[i];
6903 break;
6904 case C2LS:
6905 dops[i].rs1=(source[i]>>21)&0x1F;
6906 dops[i].rs2=0;
6907 dops[i].rt1=0;
6908 dops[i].rt2=0;
6909 imm[i]=(short)source[i];
6910 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
6911 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
6912 break;
6913 case C2OP:
6914 dops[i].rs1=0;
6915 dops[i].rs2=0;
6916 dops[i].rt1=0;
6917 dops[i].rt2=0;
6918 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
6919 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
6920 gte_rt[i]|=1ll<<63; // every op changes flags
6921 if((source[i]&0x3f)==GTE_MVMVA) {
6922 int v = (source[i] >> 15) & 3;
6923 gte_rs[i]&=~0xe3fll;
6924 if(v==3) gte_rs[i]|=0xe00ll;
6925 else gte_rs[i]|=3ll<<(v*2);
6926 }
6927 break;
6928 case SYSCALL:
6929 case HLECALL:
6930 case INTCALL:
6931 dops[i].rs1=CCREG;
6932 dops[i].rs2=0;
6933 dops[i].rt1=0;
6934 dops[i].rt2=0;
6935 break;
6936 default:
6937 dops[i].rs1=0;
6938 dops[i].rs2=0;
6939 dops[i].rt1=0;
6940 dops[i].rt2=0;
6941 }
6942 /* Calculate branch target addresses */
6943 if(type==UJUMP)
6944 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6945 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
6946 ba[i]=start+i*4+8; // Ignore never taken branch
6947 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
6948 ba[i]=start+i*4+8; // Ignore never taken branch
6949 else if(type==CJUMP||type==SJUMP)
6950 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6951 else ba[i]=-1;
6952
6953 /* simplify always (not)taken branches */
6954 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
6955 dops[i].rs1 = dops[i].rs2 = 0;
6956 if (!(op & 1)) {
6957 dops[i].itype = type = UJUMP;
6958 dops[i].rs2 = CCREG;
6959 }
6960 }
6961 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
6962 dops[i].itype = type = UJUMP;
6963
6964 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
6965 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
6966 dops[i].is_load = (dops[i].itype == LOAD || dops[i].itype == LOADLR || op == 0x32); // LWC2
6967 dops[i].is_store = (dops[i].itype == STORE || dops[i].itype == STORELR || op == 0x3a); // SWC2
6968
6969 /* messy cases to just pass over to the interpreter */
6970 if (i > 0 && dops[i-1].is_jump) {
6971 int do_in_intrp=0;
6972 // branch in delay slot?
6973 if (dops[i].is_jump) {
6974 // don't handle first branch and call interpreter if it's hit
6975 SysPrintf("branch in delay slot @%08x (%08x)\n", start + i*4, start);
6976 do_in_intrp=1;
6977 }
6978 // basic load delay detection
6979 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
6980 int t=(ba[i-1]-start)/4;
6981 if(0 <= t && t < i &&(dops[i].rt1==dops[t].rs1||dops[i].rt1==dops[t].rs2)&&dops[t].itype!=CJUMP&&dops[t].itype!=SJUMP) {
6982 // jump target wants DS result - potential load delay effect
6983 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
6984 do_in_intrp=1;
6985 dops[t+1].bt=1; // expected return from interpreter
6986 }
6987 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&&
6988 !(i>=3&&dops[i-3].is_jump)) {
6989 // v0 overwrite like this is a sign of trouble, bail out
6990 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
6991 do_in_intrp=1;
6992 }
6993 }
6994 if (do_in_intrp) {
6995 memset(&dops[i-1], 0, sizeof(dops[i-1]));
6996 dops[i-1].itype = INTCALL;
6997 dops[i-1].rs1 = CCREG;
6998 ba[i-1] = -1;
6999 done = 2;
7000 i--; // don't compile the DS
7001 }
7002 }
7003
7004 /* Is this the end of the block? */
7005 if (i > 0 && dops[i-1].is_ujump) {
7006 if (dops[i-1].rt1 == 0) { // not jal
7007 int found_bbranch = 0, t = (ba[i-1] - start) / 4;
7008 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
7009 // scan for a branch back to i+1
7010 for (j = t; j < t + 64; j++) {
7011 int tmpop = source[j] >> 26;
7012 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
7013 int t2 = j + 1 + (int)(signed short)source[j];
7014 if (t2 == i + 1) {
7015 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
7016 found_bbranch = 1;
7017 break;
7018 }
7019 }
7020 }
7021 }
7022 if (!found_bbranch)
7023 done = 2;
7024 }
7025 else {
7026 if(stop_after_jal) done=1;
7027 // Stop on BREAK
7028 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7029 }
7030 // Don't recompile stuff that's already compiled
7031 if(check_addr(start+i*4+4)) done=1;
7032 // Don't get too close to the limit
7033 if(i>MAXBLOCK/2) done=1;
7034 }
7035 if (dops[i].itype == SYSCALL || dops[i].itype == HLECALL || dops[i].itype == INTCALL)
7036 done = stop_after_jal ? 1 : 2;
7037 if (done == 2) {
7038 // Does the block continue due to a branch?
7039 for(j=i-1;j>=0;j--)
7040 {
7041 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7042 if(ba[j]==start+i*4+4) done=j=0;
7043 if(ba[j]==start+i*4+8) done=j=0;
7044 }
7045 }
7046 //assert(i<MAXBLOCK-1);
7047 if(start+i*4==pagelimit-4) done=1;
7048 assert(start+i*4<pagelimit);
7049 if (i==MAXBLOCK-1) done=1;
7050 // Stop if we're compiling junk
7051 if(dops[i].itype == NI && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7052 done=stop_after_jal=1;
7053 SysPrintf("Disabled speculative precompilation\n");
7054 }
7055 }
7056 while (i > 0 && dops[i-1].is_jump)
7057 i--;
7058 assert(i > 0);
7059 assert(!dops[i-1].is_jump);
7060 slen = i;
7061}
7062
7063// Basic liveness analysis for MIPS registers
7064static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7065{
7066 int i;
7067 uint64_t u,gte_u,b,gte_b;
7068 uint64_t temp_u,temp_gte_u=0;
7069 uint64_t gte_u_unknown=0;
7070 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7071 gte_u_unknown=~0ll;
7072 if(iend==slen-1) {
7073 u=1;
7074 gte_u=gte_u_unknown;
7075 }else{
7076 //u=unneeded_reg[iend+1];
7077 u=1;
7078 gte_u=gte_unneeded[iend+1];
7079 }
7080
7081 for (i=iend;i>=istart;i--)
7082 {
7083 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7084 if(dops[i].is_jump)
7085 {
7086 // If subroutine call, flag return address as a possible branch target
7087 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7088
7089 if(ba[i]<start || ba[i]>=(start+slen*4))
7090 {
7091 // Branch out of this block, flush all regs
7092 u=1;
7093 gte_u=gte_u_unknown;
7094 branch_unneeded_reg[i]=u;
7095 // Merge in delay slot
7096 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7097 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7098 u|=1;
7099 gte_u|=gte_rt[i+1];
7100 gte_u&=~gte_rs[i+1];
7101 }
7102 else
7103 {
7104 // Internal branch, flag target
7105 dops[(ba[i]-start)>>2].bt=1;
7106 if(ba[i]<=start+i*4) {
7107 // Backward branch
7108 if(dops[i].is_ujump)
7109 {
7110 // Unconditional branch
7111 temp_u=1;
7112 temp_gte_u=0;
7113 } else {
7114 // Conditional branch (not taken case)
7115 temp_u=unneeded_reg[i+2];
7116 temp_gte_u&=gte_unneeded[i+2];
7117 }
7118 // Merge in delay slot
7119 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7120 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7121 temp_u|=1;
7122 temp_gte_u|=gte_rt[i+1];
7123 temp_gte_u&=~gte_rs[i+1];
7124 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7125 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7126 temp_u|=1;
7127 temp_gte_u|=gte_rt[i];
7128 temp_gte_u&=~gte_rs[i];
7129 unneeded_reg[i]=temp_u;
7130 gte_unneeded[i]=temp_gte_u;
7131 // Only go three levels deep. This recursion can take an
7132 // excessive amount of time if there are a lot of nested loops.
7133 if(r<2) {
7134 pass2_unneeded_regs((ba[i]-start)>>2,i-1,r+1);
7135 }else{
7136 unneeded_reg[(ba[i]-start)>>2]=1;
7137 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
7138 }
7139 } /*else*/ if(1) {
7140 if (dops[i].is_ujump)
7141 {
7142 // Unconditional branch
7143 u=unneeded_reg[(ba[i]-start)>>2];
7144 gte_u=gte_unneeded[(ba[i]-start)>>2];
7145 branch_unneeded_reg[i]=u;
7146 // Merge in delay slot
7147 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7148 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7149 u|=1;
7150 gte_u|=gte_rt[i+1];
7151 gte_u&=~gte_rs[i+1];
7152 } else {
7153 // Conditional branch
7154 b=unneeded_reg[(ba[i]-start)>>2];
7155 gte_b=gte_unneeded[(ba[i]-start)>>2];
7156 branch_unneeded_reg[i]=b;
7157 // Branch delay slot
7158 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7159 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7160 b|=1;
7161 gte_b|=gte_rt[i+1];
7162 gte_b&=~gte_rs[i+1];
7163 u&=b;
7164 gte_u&=gte_b;
7165 if(i<slen-1) {
7166 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7167 } else {
7168 branch_unneeded_reg[i]=1;
7169 }
7170 }
7171 }
7172 }
7173 }
7174 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
7175 {
7176 // SYSCALL instruction (software interrupt)
7177 u=1;
7178 }
7179 else if(dops[i].itype==COP0 && dops[i].opcode2==0x10)
7180 {
7181 // RFE
7182 u=1;
7183 }
7184 //u=1; // DEBUG
7185 // Written registers are unneeded
7186 u|=1LL<<dops[i].rt1;
7187 u|=1LL<<dops[i].rt2;
7188 gte_u|=gte_rt[i];
7189 // Accessed registers are needed
7190 u&=~(1LL<<dops[i].rs1);
7191 u&=~(1LL<<dops[i].rs2);
7192 gte_u&=~gte_rs[i];
7193 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7194 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7195 // Source-target dependencies
7196 // R0 is always unneeded
7197 u|=1;
7198 // Save it
7199 unneeded_reg[i]=u;
7200 gte_unneeded[i]=gte_u;
7201 /*
7202 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7203 printf("U:");
7204 int r;
7205 for(r=1;r<=CCREG;r++) {
7206 if((unneeded_reg[i]>>r)&1) {
7207 if(r==HIREG) printf(" HI");
7208 else if(r==LOREG) printf(" LO");
7209 else printf(" r%d",r);
7210 }
7211 }
7212 printf("\n");
7213 */
7214 }
7215}
7216
7217static noinline void pass3_register_alloc(u_int addr)
7218{
7219 struct regstat current; // Current register allocations/status
7220 clear_all_regs(current.regmap_entry);
7221 clear_all_regs(current.regmap);
7222 current.wasdirty = current.dirty = 0;
7223 current.u = unneeded_reg[0];
7224 alloc_reg(&current, 0, CCREG);
7225 dirty_reg(&current, CCREG);
7226 current.wasconst = 0;
7227 current.isconst = 0;
7228 current.loadedconst = 0;
7229 //current.waswritten = 0;
7230 int ds=0;
7231 int cc=0;
7232 int hr;
7233 int i, j;
7234
7235 if (addr & 1) {
7236 // First instruction is delay slot
7237 cc=-1;
7238 dops[1].bt=1;
7239 ds=1;
7240 unneeded_reg[0]=1;
7241 current.regmap[HOST_BTREG]=BTREG;
7242 }
7243
7244 for(i=0;i<slen;i++)
7245 {
7246 if(dops[i].bt)
7247 {
7248 for(hr=0;hr<HOST_REGS;hr++)
7249 {
7250 // Is this really necessary?
7251 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7252 }
7253 current.isconst=0;
7254 //current.waswritten=0;
7255 }
7256
7257 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7258 regs[i].wasconst=current.isconst;
7259 regs[i].wasdirty=current.dirty;
7260 regs[i].dirty=0;
7261 regs[i].u=0;
7262 regs[i].isconst=0;
7263 regs[i].loadedconst=0;
7264 if (!dops[i].is_jump) {
7265 if(i+1<slen) {
7266 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7267 current.u|=1;
7268 } else {
7269 current.u=1;
7270 }
7271 } else {
7272 if(i+1<slen) {
7273 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7274 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7275 current.u|=1;
7276 } else {
7277 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7278 abort();
7279 }
7280 }
7281 dops[i].is_ds=ds;
7282 if(ds) {
7283 ds=0; // Skip delay slot, already allocated as part of branch
7284 // ...but we need to alloc it in case something jumps here
7285 if(i+1<slen) {
7286 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7287 }else{
7288 current.u=branch_unneeded_reg[i-1];
7289 }
7290 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7291 current.u|=1;
7292 struct regstat temp;
7293 memcpy(&temp,&current,sizeof(current));
7294 temp.wasdirty=temp.dirty;
7295 // TODO: Take into account unconditional branches, as below
7296 delayslot_alloc(&temp,i);
7297 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7298 regs[i].wasdirty=temp.wasdirty;
7299 regs[i].dirty=temp.dirty;
7300 regs[i].isconst=0;
7301 regs[i].wasconst=0;
7302 current.isconst=0;
7303 // Create entry (branch target) regmap
7304 for(hr=0;hr<HOST_REGS;hr++)
7305 {
7306 int r=temp.regmap[hr];
7307 if(r>=0) {
7308 if(r!=regmap_pre[i][hr]) {
7309 regs[i].regmap_entry[hr]=-1;
7310 }
7311 else
7312 {
7313 assert(r < 64);
7314 if((current.u>>r)&1) {
7315 regs[i].regmap_entry[hr]=-1;
7316 regs[i].regmap[hr]=-1;
7317 //Don't clear regs in the delay slot as the branch might need them
7318 //current.regmap[hr]=-1;
7319 }else
7320 regs[i].regmap_entry[hr]=r;
7321 }
7322 } else {
7323 // First instruction expects CCREG to be allocated
7324 if(i==0&&hr==HOST_CCREG)
7325 regs[i].regmap_entry[hr]=CCREG;
7326 else
7327 regs[i].regmap_entry[hr]=-1;
7328 }
7329 }
7330 }
7331 else { // Not delay slot
7332 switch(dops[i].itype) {
7333 case UJUMP:
7334 //current.isconst=0; // DEBUG
7335 //current.wasconst=0; // DEBUG
7336 //regs[i].wasconst=0; // DEBUG
7337 clear_const(&current,dops[i].rt1);
7338 alloc_cc(&current,i);
7339 dirty_reg(&current,CCREG);
7340 if (dops[i].rt1==31) {
7341 alloc_reg(&current,i,31);
7342 dirty_reg(&current,31);
7343 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7344 //assert(dops[i+1].rt1!=dops[i].rt1);
7345 #ifdef REG_PREFETCH
7346 alloc_reg(&current,i,PTEMP);
7347 #endif
7348 }
7349 dops[i].ooo=1;
7350 delayslot_alloc(&current,i+1);
7351 //current.isconst=0; // DEBUG
7352 ds=1;
7353 //printf("i=%d, isconst=%x\n",i,current.isconst);
7354 break;
7355 case RJUMP:
7356 //current.isconst=0;
7357 //current.wasconst=0;
7358 //regs[i].wasconst=0;
7359 clear_const(&current,dops[i].rs1);
7360 clear_const(&current,dops[i].rt1);
7361 alloc_cc(&current,i);
7362 dirty_reg(&current,CCREG);
7363 if (!ds_writes_rjump_rs(i)) {
7364 alloc_reg(&current,i,dops[i].rs1);
7365 if (dops[i].rt1!=0) {
7366 alloc_reg(&current,i,dops[i].rt1);
7367 dirty_reg(&current,dops[i].rt1);
7368 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7369 assert(dops[i+1].rt1!=dops[i].rt1);
7370 #ifdef REG_PREFETCH
7371 alloc_reg(&current,i,PTEMP);
7372 #endif
7373 }
7374 #ifdef USE_MINI_HT
7375 if(dops[i].rs1==31) { // JALR
7376 alloc_reg(&current,i,RHASH);
7377 alloc_reg(&current,i,RHTBL);
7378 }
7379 #endif
7380 delayslot_alloc(&current,i+1);
7381 } else {
7382 // The delay slot overwrites our source register,
7383 // allocate a temporary register to hold the old value.
7384 current.isconst=0;
7385 current.wasconst=0;
7386 regs[i].wasconst=0;
7387 delayslot_alloc(&current,i+1);
7388 current.isconst=0;
7389 alloc_reg(&current,i,RTEMP);
7390 }
7391 //current.isconst=0; // DEBUG
7392 dops[i].ooo=1;
7393 ds=1;
7394 break;
7395 case CJUMP:
7396 //current.isconst=0;
7397 //current.wasconst=0;
7398 //regs[i].wasconst=0;
7399 clear_const(&current,dops[i].rs1);
7400 clear_const(&current,dops[i].rs2);
7401 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7402 {
7403 alloc_cc(&current,i);
7404 dirty_reg(&current,CCREG);
7405 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7406 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7407 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7408 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7409 // The delay slot overwrites one of our conditions.
7410 // Allocate the branch condition registers instead.
7411 current.isconst=0;
7412 current.wasconst=0;
7413 regs[i].wasconst=0;
7414 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7415 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7416 }
7417 else
7418 {
7419 dops[i].ooo=1;
7420 delayslot_alloc(&current,i+1);
7421 }
7422 }
7423 else
7424 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7425 {
7426 alloc_cc(&current,i);
7427 dirty_reg(&current,CCREG);
7428 alloc_reg(&current,i,dops[i].rs1);
7429 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7430 // The delay slot overwrites one of our conditions.
7431 // Allocate the branch condition registers instead.
7432 current.isconst=0;
7433 current.wasconst=0;
7434 regs[i].wasconst=0;
7435 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7436 }
7437 else
7438 {
7439 dops[i].ooo=1;
7440 delayslot_alloc(&current,i+1);
7441 }
7442 }
7443 else
7444 // Don't alloc the delay slot yet because we might not execute it
7445 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7446 {
7447 current.isconst=0;
7448 current.wasconst=0;
7449 regs[i].wasconst=0;
7450 alloc_cc(&current,i);
7451 dirty_reg(&current,CCREG);
7452 alloc_reg(&current,i,dops[i].rs1);
7453 alloc_reg(&current,i,dops[i].rs2);
7454 }
7455 else
7456 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7457 {
7458 current.isconst=0;
7459 current.wasconst=0;
7460 regs[i].wasconst=0;
7461 alloc_cc(&current,i);
7462 dirty_reg(&current,CCREG);
7463 alloc_reg(&current,i,dops[i].rs1);
7464 }
7465 ds=1;
7466 //current.isconst=0;
7467 break;
7468 case SJUMP:
7469 //current.isconst=0;
7470 //current.wasconst=0;
7471 //regs[i].wasconst=0;
7472 clear_const(&current,dops[i].rs1);
7473 clear_const(&current,dops[i].rt1);
7474 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7475 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7476 {
7477 alloc_cc(&current,i);
7478 dirty_reg(&current,CCREG);
7479 alloc_reg(&current,i,dops[i].rs1);
7480 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7481 alloc_reg(&current,i,31);
7482 dirty_reg(&current,31);
7483 //#ifdef REG_PREFETCH
7484 //alloc_reg(&current,i,PTEMP);
7485 //#endif
7486 }
7487 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7488 ||(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
7489 // Allocate the branch condition registers instead.
7490 current.isconst=0;
7491 current.wasconst=0;
7492 regs[i].wasconst=0;
7493 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7494 }
7495 else
7496 {
7497 dops[i].ooo=1;
7498 delayslot_alloc(&current,i+1);
7499 }
7500 }
7501 else
7502 // Don't alloc the delay slot yet because we might not execute it
7503 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7504 {
7505 current.isconst=0;
7506 current.wasconst=0;
7507 regs[i].wasconst=0;
7508 alloc_cc(&current,i);
7509 dirty_reg(&current,CCREG);
7510 alloc_reg(&current,i,dops[i].rs1);
7511 }
7512 ds=1;
7513 //current.isconst=0;
7514 break;
7515 case IMM16:
7516 imm16_alloc(&current,i);
7517 break;
7518 case LOAD:
7519 case LOADLR:
7520 load_alloc(&current,i);
7521 break;
7522 case STORE:
7523 case STORELR:
7524 store_alloc(&current,i);
7525 break;
7526 case ALU:
7527 alu_alloc(&current,i);
7528 break;
7529 case SHIFT:
7530 shift_alloc(&current,i);
7531 break;
7532 case MULTDIV:
7533 multdiv_alloc(&current,i);
7534 break;
7535 case SHIFTIMM:
7536 shiftimm_alloc(&current,i);
7537 break;
7538 case MOV:
7539 mov_alloc(&current,i);
7540 break;
7541 case COP0:
7542 cop0_alloc(&current,i);
7543 break;
7544 case COP1:
7545 break;
7546 case COP2:
7547 cop2_alloc(&current,i);
7548 break;
7549 case C1LS:
7550 c1ls_alloc(&current,i);
7551 break;
7552 case C2LS:
7553 c2ls_alloc(&current,i);
7554 break;
7555 case C2OP:
7556 c2op_alloc(&current,i);
7557 break;
7558 case SYSCALL:
7559 case HLECALL:
7560 case INTCALL:
7561 syscall_alloc(&current,i);
7562 break;
7563 }
7564
7565 // Create entry (branch target) regmap
7566 for(hr=0;hr<HOST_REGS;hr++)
7567 {
7568 int r,or;
7569 r=current.regmap[hr];
7570 if(r>=0) {
7571 if(r!=regmap_pre[i][hr]) {
7572 // TODO: delay slot (?)
7573 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7574 if(or<0||r>=TEMPREG){
7575 regs[i].regmap_entry[hr]=-1;
7576 }
7577 else
7578 {
7579 // Just move it to a different register
7580 regs[i].regmap_entry[hr]=r;
7581 // If it was dirty before, it's still dirty
7582 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7583 }
7584 }
7585 else
7586 {
7587 // Unneeded
7588 if(r==0){
7589 regs[i].regmap_entry[hr]=0;
7590 }
7591 else
7592 {
7593 assert(r<64);
7594 if((current.u>>r)&1) {
7595 regs[i].regmap_entry[hr]=-1;
7596 //regs[i].regmap[hr]=-1;
7597 current.regmap[hr]=-1;
7598 }else
7599 regs[i].regmap_entry[hr]=r;
7600 }
7601 }
7602 } else {
7603 // Branches expect CCREG to be allocated at the target
7604 if(regmap_pre[i][hr]==CCREG)
7605 regs[i].regmap_entry[hr]=CCREG;
7606 else
7607 regs[i].regmap_entry[hr]=-1;
7608 }
7609 }
7610 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7611 }
7612
7613#if 0 // see do_store_smc_check()
7614 if(i>0&&(dops[i-1].itype==STORE||dops[i-1].itype==STORELR||(dops[i-1].itype==C2LS&&dops[i-1].opcode==0x3a))&&(u_int)imm[i-1]<0x800)
7615 current.waswritten|=1<<dops[i-1].rs1;
7616 current.waswritten&=~(1<<dops[i].rt1);
7617 current.waswritten&=~(1<<dops[i].rt2);
7618 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
7619 current.waswritten&=~(1<<dops[i].rs1);
7620#endif
7621
7622 /* Branch post-alloc */
7623 if(i>0)
7624 {
7625 current.wasdirty=current.dirty;
7626 switch(dops[i-1].itype) {
7627 case UJUMP:
7628 memcpy(&branch_regs[i-1],&current,sizeof(current));
7629 branch_regs[i-1].isconst=0;
7630 branch_regs[i-1].wasconst=0;
7631 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7632 alloc_cc(&branch_regs[i-1],i-1);
7633 dirty_reg(&branch_regs[i-1],CCREG);
7634 if(dops[i-1].rt1==31) { // JAL
7635 alloc_reg(&branch_regs[i-1],i-1,31);
7636 dirty_reg(&branch_regs[i-1],31);
7637 }
7638 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7639 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7640 break;
7641 case RJUMP:
7642 memcpy(&branch_regs[i-1],&current,sizeof(current));
7643 branch_regs[i-1].isconst=0;
7644 branch_regs[i-1].wasconst=0;
7645 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7646 alloc_cc(&branch_regs[i-1],i-1);
7647 dirty_reg(&branch_regs[i-1],CCREG);
7648 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7649 if(dops[i-1].rt1!=0) { // JALR
7650 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7651 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7652 }
7653 #ifdef USE_MINI_HT
7654 if(dops[i-1].rs1==31) { // JALR
7655 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7656 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7657 }
7658 #endif
7659 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7660 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7661 break;
7662 case CJUMP:
7663 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7664 {
7665 alloc_cc(&current,i-1);
7666 dirty_reg(&current,CCREG);
7667 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7668 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7669 // The delay slot overwrote one of our conditions
7670 // Delay slot goes after the test (in order)
7671 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7672 current.u|=1;
7673 delayslot_alloc(&current,i);
7674 current.isconst=0;
7675 }
7676 else
7677 {
7678 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7679 // Alloc the branch condition registers
7680 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7681 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7682 }
7683 memcpy(&branch_regs[i-1],&current,sizeof(current));
7684 branch_regs[i-1].isconst=0;
7685 branch_regs[i-1].wasconst=0;
7686 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7687 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7688 }
7689 else
7690 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7691 {
7692 alloc_cc(&current,i-1);
7693 dirty_reg(&current,CCREG);
7694 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7695 // The delay slot overwrote the branch condition
7696 // Delay slot goes after the test (in order)
7697 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7698 current.u|=1;
7699 delayslot_alloc(&current,i);
7700 current.isconst=0;
7701 }
7702 else
7703 {
7704 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7705 // Alloc the branch condition register
7706 alloc_reg(&current,i-1,dops[i-1].rs1);
7707 }
7708 memcpy(&branch_regs[i-1],&current,sizeof(current));
7709 branch_regs[i-1].isconst=0;
7710 branch_regs[i-1].wasconst=0;
7711 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7712 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7713 }
7714 else
7715 // Alloc the delay slot in case the branch is taken
7716 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
7717 {
7718 memcpy(&branch_regs[i-1],&current,sizeof(current));
7719 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
7720 alloc_cc(&branch_regs[i-1],i);
7721 dirty_reg(&branch_regs[i-1],CCREG);
7722 delayslot_alloc(&branch_regs[i-1],i);
7723 branch_regs[i-1].isconst=0;
7724 alloc_reg(&current,i,CCREG); // Not taken path
7725 dirty_reg(&current,CCREG);
7726 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7727 }
7728 else
7729 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
7730 {
7731 memcpy(&branch_regs[i-1],&current,sizeof(current));
7732 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
7733 alloc_cc(&branch_regs[i-1],i);
7734 dirty_reg(&branch_regs[i-1],CCREG);
7735 delayslot_alloc(&branch_regs[i-1],i);
7736 branch_regs[i-1].isconst=0;
7737 alloc_reg(&current,i,CCREG); // Not taken path
7738 dirty_reg(&current,CCREG);
7739 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7740 }
7741 break;
7742 case SJUMP:
7743 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
7744 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
7745 {
7746 alloc_cc(&current,i-1);
7747 dirty_reg(&current,CCREG);
7748 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7749 // The delay slot overwrote the branch condition
7750 // Delay slot goes after the test (in order)
7751 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7752 current.u|=1;
7753 delayslot_alloc(&current,i);
7754 current.isconst=0;
7755 }
7756 else
7757 {
7758 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7759 // Alloc the branch condition register
7760 alloc_reg(&current,i-1,dops[i-1].rs1);
7761 }
7762 memcpy(&branch_regs[i-1],&current,sizeof(current));
7763 branch_regs[i-1].isconst=0;
7764 branch_regs[i-1].wasconst=0;
7765 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7766 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7767 }
7768 else
7769 // Alloc the delay slot in case the branch is taken
7770 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
7771 {
7772 memcpy(&branch_regs[i-1],&current,sizeof(current));
7773 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
7774 alloc_cc(&branch_regs[i-1],i);
7775 dirty_reg(&branch_regs[i-1],CCREG);
7776 delayslot_alloc(&branch_regs[i-1],i);
7777 branch_regs[i-1].isconst=0;
7778 alloc_reg(&current,i,CCREG); // Not taken path
7779 dirty_reg(&current,CCREG);
7780 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7781 }
7782 // FIXME: BLTZAL/BGEZAL
7783 if(dops[i-1].opcode2&0x10) { // BxxZAL
7784 alloc_reg(&branch_regs[i-1],i-1,31);
7785 dirty_reg(&branch_regs[i-1],31);
7786 }
7787 break;
7788 }
7789
7790 if (dops[i-1].is_ujump)
7791 {
7792 if(dops[i-1].rt1==31) // JAL/JALR
7793 {
7794 // Subroutine call will return here, don't alloc any registers
7795 current.dirty=0;
7796 clear_all_regs(current.regmap);
7797 alloc_reg(&current,i,CCREG);
7798 dirty_reg(&current,CCREG);
7799 }
7800 else if(i+1<slen)
7801 {
7802 // Internal branch will jump here, match registers to caller
7803 current.dirty=0;
7804 clear_all_regs(current.regmap);
7805 alloc_reg(&current,i,CCREG);
7806 dirty_reg(&current,CCREG);
7807 for(j=i-1;j>=0;j--)
7808 {
7809 if(ba[j]==start+i*4+4) {
7810 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7811 current.dirty=branch_regs[j].dirty;
7812 break;
7813 }
7814 }
7815 while(j>=0) {
7816 if(ba[j]==start+i*4+4) {
7817 for(hr=0;hr<HOST_REGS;hr++) {
7818 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7819 current.regmap[hr]=-1;
7820 }
7821 current.dirty&=branch_regs[j].dirty;
7822 }
7823 }
7824 j--;
7825 }
7826 }
7827 }
7828 }
7829
7830 // Count cycles in between branches
7831 ccadj[i] = CLOCK_ADJUST(cc);
7832 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
7833 {
7834 cc=0;
7835 }
7836#if !defined(DRC_DBG)
7837 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7838 {
7839 // this should really be removed since the real stalls have been implemented,
7840 // but doing so causes sizeable perf regression against the older version
7841 u_int gtec = gte_cycletab[source[i] & 0x3f];
7842 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7843 }
7844 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7845 {
7846 cc+=4;
7847 }
7848 else if(dops[i].itype==C2LS)
7849 {
7850 // same as with C2OP
7851 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7852 }
7853#endif
7854 else
7855 {
7856 cc++;
7857 }
7858
7859 if(!dops[i].is_ds) {
7860 regs[i].dirty=current.dirty;
7861 regs[i].isconst=current.isconst;
7862 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7863 }
7864 for(hr=0;hr<HOST_REGS;hr++) {
7865 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7866 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7867 regs[i].wasconst&=~(1<<hr);
7868 }
7869 }
7870 }
7871 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7872 //regs[i].waswritten=current.waswritten;
7873 }
7874}
7875
7876static noinline void pass4_cull_unused_regs(void)
7877{
7878 u_int last_needed_regs[4] = {0,0,0,0};
7879 u_int nr=0;
7880 int i;
7881
7882 for (i=slen-1;i>=0;i--)
7883 {
7884 int hr;
7885 __builtin_prefetch(regs[i-2].regmap);
7886 if(dops[i].is_jump)
7887 {
7888 if(ba[i]<start || ba[i]>=(start+slen*4))
7889 {
7890 // Branch out of this block, don't need anything
7891 nr=0;
7892 }
7893 else
7894 {
7895 // Internal branch
7896 // Need whatever matches the target
7897 nr=0;
7898 int t=(ba[i]-start)>>2;
7899 for(hr=0;hr<HOST_REGS;hr++)
7900 {
7901 if(regs[i].regmap_entry[hr]>=0) {
7902 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7903 }
7904 }
7905 }
7906 // Conditional branch may need registers for following instructions
7907 if (!dops[i].is_ujump)
7908 {
7909 if(i<slen-2) {
7910 nr |= last_needed_regs[(i+2) & 3];
7911 for(hr=0;hr<HOST_REGS;hr++)
7912 {
7913 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7914 //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]);
7915 }
7916 }
7917 }
7918 // Don't need stuff which is overwritten
7919 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7920 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7921 // Merge in delay slot
7922 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
7923 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
7924 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
7925 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
7926 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
7927 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
7928 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
7929 nr |= get_regm(regmap_pre[i], ROREG);
7930 nr |= get_regm(regs[i].regmap_entry, ROREG);
7931 }
7932 if (dops[i+1].is_store) {
7933 nr |= get_regm(regmap_pre[i], INVCP);
7934 nr |= get_regm(regs[i].regmap_entry, INVCP);
7935 }
7936 }
7937 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
7938 {
7939 // SYSCALL instruction (software interrupt)
7940 nr=0;
7941 }
7942 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
7943 {
7944 // ERET instruction (return from interrupt)
7945 nr=0;
7946 }
7947 else // Non-branch
7948 {
7949 if(i<slen-1) {
7950 for(hr=0;hr<HOST_REGS;hr++) {
7951 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7952 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7953 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7954 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7955 }
7956 }
7957 }
7958 // Overwritten registers are not needed
7959 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
7960 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
7961 nr &= ~get_regm(regs[i].regmap, FTEMP);
7962 // Source registers are needed
7963 nr |= get_regm(regmap_pre[i], dops[i].rs1);
7964 nr |= get_regm(regmap_pre[i], dops[i].rs2);
7965 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
7966 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
7967 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
7968 nr |= get_regm(regmap_pre[i], ROREG);
7969 nr |= get_regm(regs[i].regmap_entry, ROREG);
7970 }
7971 if (dops[i].is_store) {
7972 nr |= get_regm(regmap_pre[i], INVCP);
7973 nr |= get_regm(regs[i].regmap_entry, INVCP);
7974 }
7975
7976 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
7977 for(hr=0;hr<HOST_REGS;hr++)
7978 {
7979 // Don't store a register immediately after writing it,
7980 // may prevent dual-issue.
7981 // But do so if this is a branch target, otherwise we
7982 // might have to load the register before the branch.
7983 if((regs[i].wasdirty>>hr)&1) {
7984 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7985 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
7986 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
7987 }
7988 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7989 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
7990 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
7991 }
7992 }
7993 }
7994 // Cycle count is needed at branches. Assume it is needed at the target too.
7995 if(i==0||dops[i].bt||dops[i].itype==CJUMP) {
7996 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7997 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7998 }
7999 // Save it
8000 last_needed_regs[i & 3] = nr;
8001
8002 // Deallocate unneeded registers
8003 for(hr=0;hr<HOST_REGS;hr++)
8004 {
8005 if(!((nr>>hr)&1)) {
8006 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8007 if(dops[i].is_jump)
8008 {
8009 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
8010 if (dops[i+1].is_load || dops[i+1].is_store)
8011 map1 = ROREG;
8012 if (dops[i+1].is_store)
8013 map2 = INVCP;
8014 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
8015 temp = FTEMP;
8016 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8017 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8018 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
8019 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8020 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
8021 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8022 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8023 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
8024 {
8025 regs[i].regmap[hr]=-1;
8026 regs[i].isconst&=~(1<<hr);
8027 regs[i].dirty&=~(1<<hr);
8028 regs[i+1].wasdirty&=~(1<<hr);
8029 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
8030 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
8031 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
8032 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8033 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8034 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8035 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8036 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8037 {
8038 branch_regs[i].regmap[hr]=-1;
8039 branch_regs[i].regmap_entry[hr]=-1;
8040 if (!dops[i].is_ujump)
8041 {
8042 if (i < slen-2) {
8043 regmap_pre[i+2][hr]=-1;
8044 regs[i+2].wasconst&=~(1<<hr);
8045 }
8046 }
8047 }
8048 }
8049 }
8050 else
8051 {
8052 // Non-branch
8053 if(i>0)
8054 {
8055 int map1 = -1, map2 = -1, temp=-1;
8056 if (dops[i].is_load || dops[i].is_store)
8057 map1 = ROREG;
8058 if (dops[i].is_store)
8059 map2 = INVCP;
8060 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8061 temp = FTEMP;
8062 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8063 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8064 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8065 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8066 regs[i].regmap[hr] != CCREG)
8067 {
8068 if(i<slen-1&&!dops[i].is_ds) {
8069 assert(regs[i].regmap[hr]<64);
8070 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8071 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8072 {
8073 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8074 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8075 }
8076 regmap_pre[i+1][hr]=-1;
8077 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8078 regs[i+1].wasconst&=~(1<<hr);
8079 }
8080 regs[i].regmap[hr]=-1;
8081 regs[i].isconst&=~(1<<hr);
8082 regs[i].dirty&=~(1<<hr);
8083 regs[i+1].wasdirty&=~(1<<hr);
8084 }
8085 }
8086 }
8087 } // if needed
8088 } // for hr
8089 }
8090}
8091
8092// If a register is allocated during a loop, try to allocate it for the
8093// entire loop, if possible. This avoids loading/storing registers
8094// inside of the loop.
8095static noinline void pass5a_preallocate1(void)
8096{
8097 int i, j, hr;
8098 signed char f_regmap[HOST_REGS];
8099 clear_all_regs(f_regmap);
8100 for(i=0;i<slen-1;i++)
8101 {
8102 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8103 {
8104 if(ba[i]>=start && ba[i]<(start+i*4))
8105 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8106 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8107 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8108 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8109 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8110 {
8111 int t=(ba[i]-start)>>2;
8112 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8113 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8114 for(hr=0;hr<HOST_REGS;hr++)
8115 {
8116 if(regs[i].regmap[hr]>=0) {
8117 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8118 // dealloc old register
8119 int n;
8120 for(n=0;n<HOST_REGS;n++)
8121 {
8122 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8123 }
8124 // and alloc new one
8125 f_regmap[hr]=regs[i].regmap[hr];
8126 }
8127 }
8128 if(branch_regs[i].regmap[hr]>=0) {
8129 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8130 // dealloc old register
8131 int n;
8132 for(n=0;n<HOST_REGS;n++)
8133 {
8134 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8135 }
8136 // and alloc new one
8137 f_regmap[hr]=branch_regs[i].regmap[hr];
8138 }
8139 }
8140 if(dops[i].ooo) {
8141 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8142 f_regmap[hr]=branch_regs[i].regmap[hr];
8143 }else{
8144 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8145 f_regmap[hr]=branch_regs[i].regmap[hr];
8146 }
8147 // Avoid dirty->clean transition
8148 #ifdef DESTRUCTIVE_WRITEBACK
8149 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;
8150 #endif
8151 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8152 // case above, however it's always a good idea. We can't hoist the
8153 // load if the register was already allocated, so there's no point
8154 // wasting time analyzing most of these cases. It only "succeeds"
8155 // when the mapping was different and the load can be replaced with
8156 // a mov, which is of negligible benefit. So such cases are
8157 // skipped below.
8158 if(f_regmap[hr]>0) {
8159 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8160 int r=f_regmap[hr];
8161 for(j=t;j<=i;j++)
8162 {
8163 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8164 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8165 assert(r < 64);
8166 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8167 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8168 int k;
8169 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8170 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8171 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8172 k=i;
8173 while(k>1&&regs[k-1].regmap[hr]==-1) {
8174 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8175 //printf("no free regs for store %x\n",start+(k-1)*4);
8176 break;
8177 }
8178 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8179 //printf("no-match due to different register\n");
8180 break;
8181 }
8182 if (dops[k-2].is_jump) {
8183 //printf("no-match due to branch\n");
8184 break;
8185 }
8186 // call/ret fast path assumes no registers allocated
8187 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8188 break;
8189 }
8190 k--;
8191 }
8192 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8193 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8194 while(k<i) {
8195 regs[k].regmap_entry[hr]=f_regmap[hr];
8196 regs[k].regmap[hr]=f_regmap[hr];
8197 regmap_pre[k+1][hr]=f_regmap[hr];
8198 regs[k].wasdirty&=~(1<<hr);
8199 regs[k].dirty&=~(1<<hr);
8200 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8201 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8202 regs[k].wasconst&=~(1<<hr);
8203 regs[k].isconst&=~(1<<hr);
8204 k++;
8205 }
8206 }
8207 else {
8208 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8209 break;
8210 }
8211 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8212 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8213 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8214 regs[i].regmap_entry[hr]=f_regmap[hr];
8215 regs[i].regmap[hr]=f_regmap[hr];
8216 regs[i].wasdirty&=~(1<<hr);
8217 regs[i].dirty&=~(1<<hr);
8218 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8219 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8220 regs[i].wasconst&=~(1<<hr);
8221 regs[i].isconst&=~(1<<hr);
8222 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8223 branch_regs[i].wasdirty&=~(1<<hr);
8224 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8225 branch_regs[i].regmap[hr]=f_regmap[hr];
8226 branch_regs[i].dirty&=~(1<<hr);
8227 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8228 branch_regs[i].wasconst&=~(1<<hr);
8229 branch_regs[i].isconst&=~(1<<hr);
8230 if (!dops[i].is_ujump) {
8231 regmap_pre[i+2][hr]=f_regmap[hr];
8232 regs[i+2].wasdirty&=~(1<<hr);
8233 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8234 }
8235 }
8236 }
8237 for(k=t;k<j;k++) {
8238 // Alloc register clean at beginning of loop,
8239 // but may dirty it in pass 6
8240 regs[k].regmap_entry[hr]=f_regmap[hr];
8241 regs[k].regmap[hr]=f_regmap[hr];
8242 regs[k].dirty&=~(1<<hr);
8243 regs[k].wasconst&=~(1<<hr);
8244 regs[k].isconst&=~(1<<hr);
8245 if (dops[k].is_jump) {
8246 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8247 branch_regs[k].regmap[hr]=f_regmap[hr];
8248 branch_regs[k].dirty&=~(1<<hr);
8249 branch_regs[k].wasconst&=~(1<<hr);
8250 branch_regs[k].isconst&=~(1<<hr);
8251 if (!dops[k].is_ujump) {
8252 regmap_pre[k+2][hr]=f_regmap[hr];
8253 regs[k+2].wasdirty&=~(1<<hr);
8254 }
8255 }
8256 else
8257 {
8258 regmap_pre[k+1][hr]=f_regmap[hr];
8259 regs[k+1].wasdirty&=~(1<<hr);
8260 }
8261 }
8262 if(regs[j].regmap[hr]==f_regmap[hr])
8263 regs[j].regmap_entry[hr]=f_regmap[hr];
8264 break;
8265 }
8266 if(j==i) break;
8267 if(regs[j].regmap[hr]>=0)
8268 break;
8269 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8270 //printf("no-match due to different register\n");
8271 break;
8272 }
8273 if (dops[j].is_ujump)
8274 {
8275 // Stop on unconditional branch
8276 break;
8277 }
8278 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8279 {
8280 if(dops[j].ooo) {
8281 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8282 break;
8283 }else{
8284 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8285 break;
8286 }
8287 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8288 //printf("no-match due to different register (branch)\n");
8289 break;
8290 }
8291 }
8292 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8293 //printf("No free regs for store %x\n",start+j*4);
8294 break;
8295 }
8296 assert(f_regmap[hr]<64);
8297 }
8298 }
8299 }
8300 }
8301 }
8302 }else{
8303 // Non branch or undetermined branch target
8304 for(hr=0;hr<HOST_REGS;hr++)
8305 {
8306 if(hr!=EXCLUDE_REG) {
8307 if(regs[i].regmap[hr]>=0) {
8308 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8309 // dealloc old register
8310 int n;
8311 for(n=0;n<HOST_REGS;n++)
8312 {
8313 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8314 }
8315 // and alloc new one
8316 f_regmap[hr]=regs[i].regmap[hr];
8317 }
8318 }
8319 }
8320 }
8321 // Try to restore cycle count at branch targets
8322 if(dops[i].bt) {
8323 for(j=i;j<slen-1;j++) {
8324 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8325 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8326 //printf("no free regs for store %x\n",start+j*4);
8327 break;
8328 }
8329 }
8330 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8331 int k=i;
8332 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8333 while(k<j) {
8334 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8335 regs[k].regmap[HOST_CCREG]=CCREG;
8336 regmap_pre[k+1][HOST_CCREG]=CCREG;
8337 regs[k+1].wasdirty|=1<<HOST_CCREG;
8338 regs[k].dirty|=1<<HOST_CCREG;
8339 regs[k].wasconst&=~(1<<HOST_CCREG);
8340 regs[k].isconst&=~(1<<HOST_CCREG);
8341 k++;
8342 }
8343 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8344 }
8345 // Work backwards from the branch target
8346 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8347 {
8348 //printf("Extend backwards\n");
8349 int k;
8350 k=i;
8351 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8352 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8353 //printf("no free regs for store %x\n",start+(k-1)*4);
8354 break;
8355 }
8356 k--;
8357 }
8358 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8359 //printf("Extend CC, %x ->\n",start+k*4);
8360 while(k<=i) {
8361 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8362 regs[k].regmap[HOST_CCREG]=CCREG;
8363 regmap_pre[k+1][HOST_CCREG]=CCREG;
8364 regs[k+1].wasdirty|=1<<HOST_CCREG;
8365 regs[k].dirty|=1<<HOST_CCREG;
8366 regs[k].wasconst&=~(1<<HOST_CCREG);
8367 regs[k].isconst&=~(1<<HOST_CCREG);
8368 k++;
8369 }
8370 }
8371 else {
8372 //printf("Fail Extend CC, %x ->\n",start+k*4);
8373 }
8374 }
8375 }
8376 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8377 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8378 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8379 {
8380 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8381 }
8382 }
8383 }
8384}
8385
8386// This allocates registers (if possible) one instruction prior
8387// to use, which can avoid a load-use penalty on certain CPUs.
8388static noinline void pass5b_preallocate2(void)
8389{
8390 int i, hr;
8391 for(i=0;i<slen-1;i++)
8392 {
8393 if (!i || !dops[i-1].is_jump)
8394 {
8395 if(!dops[i+1].bt)
8396 {
8397 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8398 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8399 {
8400 if(dops[i+1].rs1) {
8401 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8402 {
8403 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8404 {
8405 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8406 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8407 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8408 regs[i].isconst&=~(1<<hr);
8409 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8410 constmap[i][hr]=constmap[i+1][hr];
8411 regs[i+1].wasdirty&=~(1<<hr);
8412 regs[i].dirty&=~(1<<hr);
8413 }
8414 }
8415 }
8416 if(dops[i+1].rs2) {
8417 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8418 {
8419 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8420 {
8421 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8422 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8423 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8424 regs[i].isconst&=~(1<<hr);
8425 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8426 constmap[i][hr]=constmap[i+1][hr];
8427 regs[i+1].wasdirty&=~(1<<hr);
8428 regs[i].dirty&=~(1<<hr);
8429 }
8430 }
8431 }
8432 // Preload target address for load instruction (non-constant)
8433 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8434 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8435 {
8436 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8437 {
8438 regs[i].regmap[hr]=dops[i+1].rs1;
8439 regmap_pre[i+1][hr]=dops[i+1].rs1;
8440 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8441 regs[i].isconst&=~(1<<hr);
8442 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8443 constmap[i][hr]=constmap[i+1][hr];
8444 regs[i+1].wasdirty&=~(1<<hr);
8445 regs[i].dirty&=~(1<<hr);
8446 }
8447 }
8448 }
8449 // Load source into target register
8450 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8451 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8452 {
8453 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8454 {
8455 regs[i].regmap[hr]=dops[i+1].rs1;
8456 regmap_pre[i+1][hr]=dops[i+1].rs1;
8457 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8458 regs[i].isconst&=~(1<<hr);
8459 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8460 constmap[i][hr]=constmap[i+1][hr];
8461 regs[i+1].wasdirty&=~(1<<hr);
8462 regs[i].dirty&=~(1<<hr);
8463 }
8464 }
8465 }
8466 // Address for store instruction (non-constant)
8467 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8468 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8469 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8470 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8471 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8472 else {
8473 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8474 regs[i+1].isconst&=~(1<<hr);
8475 }
8476 assert(hr>=0);
8477 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8478 {
8479 regs[i].regmap[hr]=dops[i+1].rs1;
8480 regmap_pre[i+1][hr]=dops[i+1].rs1;
8481 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8482 regs[i].isconst&=~(1<<hr);
8483 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8484 constmap[i][hr]=constmap[i+1][hr];
8485 regs[i+1].wasdirty&=~(1<<hr);
8486 regs[i].dirty&=~(1<<hr);
8487 }
8488 }
8489 }
8490 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8491 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8492 int nr;
8493 hr=get_reg(regs[i+1].regmap,FTEMP);
8494 assert(hr>=0);
8495 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8496 {
8497 regs[i].regmap[hr]=dops[i+1].rs1;
8498 regmap_pre[i+1][hr]=dops[i+1].rs1;
8499 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8500 regs[i].isconst&=~(1<<hr);
8501 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8502 constmap[i][hr]=constmap[i+1][hr];
8503 regs[i+1].wasdirty&=~(1<<hr);
8504 regs[i].dirty&=~(1<<hr);
8505 }
8506 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8507 {
8508 // move it to another register
8509 regs[i+1].regmap[hr]=-1;
8510 regmap_pre[i+2][hr]=-1;
8511 regs[i+1].regmap[nr]=FTEMP;
8512 regmap_pre[i+2][nr]=FTEMP;
8513 regs[i].regmap[nr]=dops[i+1].rs1;
8514 regmap_pre[i+1][nr]=dops[i+1].rs1;
8515 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8516 regs[i].isconst&=~(1<<nr);
8517 regs[i+1].isconst&=~(1<<nr);
8518 regs[i].dirty&=~(1<<nr);
8519 regs[i+1].wasdirty&=~(1<<nr);
8520 regs[i+1].dirty&=~(1<<nr);
8521 regs[i+2].wasdirty&=~(1<<nr);
8522 }
8523 }
8524 }
8525 if(dops[i+1].itype==LOAD||dops[i+1].itype==LOADLR||dops[i+1].itype==STORE||dops[i+1].itype==STORELR/*||dops[i+1].itype==C1LS||||dops[i+1].itype==C2LS*/) {
8526 hr = -1;
8527 if(dops[i+1].itype==LOAD)
8528 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8529 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8530 hr=get_reg(regs[i+1].regmap,FTEMP);
8531 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
8532 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8533 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8534 }
8535 if(hr>=0&&regs[i].regmap[hr]<0) {
8536 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8537 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8538 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8539 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8540 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8541 regs[i].isconst&=~(1<<hr);
8542 regs[i+1].wasdirty&=~(1<<hr);
8543 regs[i].dirty&=~(1<<hr);
8544 }
8545 }
8546 }
8547 }
8548 }
8549 }
8550 }
8551}
8552
8553// Write back dirty registers as soon as we will no longer modify them,
8554// so that we don't end up with lots of writes at the branches.
8555static noinline void pass6_clean_registers(int istart, int iend, int wr)
8556{
8557 static u_int wont_dirty[MAXBLOCK];
8558 static u_int will_dirty[MAXBLOCK];
8559 int i;
8560 int r;
8561 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8562 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8563 if(iend==slen-1) {
8564 will_dirty_i=will_dirty_next=0;
8565 wont_dirty_i=wont_dirty_next=0;
8566 }else{
8567 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8568 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8569 }
8570 for (i=iend;i>=istart;i--)
8571 {
8572 signed char rregmap_i[RRMAP_SIZE];
8573 u_int hr_candirty = 0;
8574 assert(HOST_REGS < 32);
8575 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8576 __builtin_prefetch(regs[i-1].regmap);
8577 if(dops[i].is_jump)
8578 {
8579 signed char branch_rregmap_i[RRMAP_SIZE];
8580 u_int branch_hr_candirty = 0;
8581 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8582 if(ba[i]<start || ba[i]>=(start+slen*4))
8583 {
8584 // Branch out of this block, flush all regs
8585 will_dirty_i = 0;
8586 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8587 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8588 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8589 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8590 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8591 will_dirty_i &= branch_hr_candirty;
8592 if (dops[i].is_ujump)
8593 {
8594 // Unconditional branch
8595 wont_dirty_i = 0;
8596 // Merge in delay slot (will dirty)
8597 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8598 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8599 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8600 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8601 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8602 will_dirty_i &= hr_candirty;
8603 }
8604 else
8605 {
8606 // Conditional branch
8607 wont_dirty_i = wont_dirty_next;
8608 // Merge in delay slot (will dirty)
8609 // (the original code had no explanation why these 2 are commented out)
8610 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8611 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8612 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8613 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8614 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8615 will_dirty_i &= hr_candirty;
8616 }
8617 // Merge in delay slot (wont dirty)
8618 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8619 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8620 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8621 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8622 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8623 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8624 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8625 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8626 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8627 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8628 wont_dirty_i &= ~(1u << 31);
8629 if(wr) {
8630 #ifndef DESTRUCTIVE_WRITEBACK
8631 branch_regs[i].dirty&=wont_dirty_i;
8632 #endif
8633 branch_regs[i].dirty|=will_dirty_i;
8634 }
8635 }
8636 else
8637 {
8638 // Internal branch
8639 if(ba[i]<=start+i*4) {
8640 // Backward branch
8641 if (dops[i].is_ujump)
8642 {
8643 // Unconditional branch
8644 temp_will_dirty=0;
8645 temp_wont_dirty=0;
8646 // Merge in delay slot (will dirty)
8647 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8648 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8649 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8650 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8651 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8652 temp_will_dirty &= branch_hr_candirty;
8653 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8654 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8655 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8656 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8657 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8658 temp_will_dirty &= hr_candirty;
8659 } else {
8660 // Conditional branch (not taken case)
8661 temp_will_dirty=will_dirty_next;
8662 temp_wont_dirty=wont_dirty_next;
8663 // Merge in delay slot (will dirty)
8664 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8665 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8666 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8667 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8668 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8669 temp_will_dirty &= branch_hr_candirty;
8670 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8671 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8672 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8673 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8674 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8675 temp_will_dirty &= hr_candirty;
8676 }
8677 // Merge in delay slot (wont dirty)
8678 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8679 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8680 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8681 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8682 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8683 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8684 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8685 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8686 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8687 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8688 temp_wont_dirty &= ~(1u << 31);
8689 // Deal with changed mappings
8690 if(i<iend) {
8691 for(r=0;r<HOST_REGS;r++) {
8692 if(r!=EXCLUDE_REG) {
8693 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8694 temp_will_dirty&=~(1<<r);
8695 temp_wont_dirty&=~(1<<r);
8696 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8697 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8698 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8699 } else {
8700 temp_will_dirty|=1<<r;
8701 temp_wont_dirty|=1<<r;
8702 }
8703 }
8704 }
8705 }
8706 }
8707 if(wr) {
8708 will_dirty[i]=temp_will_dirty;
8709 wont_dirty[i]=temp_wont_dirty;
8710 pass6_clean_registers((ba[i]-start)>>2,i-1,0);
8711 }else{
8712 // Limit recursion. It can take an excessive amount
8713 // of time if there are a lot of nested loops.
8714 will_dirty[(ba[i]-start)>>2]=0;
8715 wont_dirty[(ba[i]-start)>>2]=-1;
8716 }
8717 }
8718 /*else*/ if(1)
8719 {
8720 if (dops[i].is_ujump)
8721 {
8722 // Unconditional branch
8723 will_dirty_i=0;
8724 wont_dirty_i=0;
8725 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
8726 for(r=0;r<HOST_REGS;r++) {
8727 if(r!=EXCLUDE_REG) {
8728 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
8729 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
8730 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
8731 }
8732 if(branch_regs[i].regmap[r]>=0) {
8733 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8734 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8735 }
8736 }
8737 }
8738 //}
8739 // Merge in delay slot
8740 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8741 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8742 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8743 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8744 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8745 will_dirty_i &= branch_hr_candirty;
8746 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8747 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8748 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8749 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8750 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8751 will_dirty_i &= hr_candirty;
8752 } else {
8753 // Conditional branch
8754 will_dirty_i=will_dirty_next;
8755 wont_dirty_i=wont_dirty_next;
8756 //if(ba[i]>start+i*4) // Disable recursion (for debugging)
8757 for(r=0;r<HOST_REGS;r++) {
8758 if(r!=EXCLUDE_REG) {
8759 signed char target_reg=branch_regs[i].regmap[r];
8760 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
8761 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
8762 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
8763 }
8764 else if(target_reg>=0) {
8765 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>target_reg)&1)<<r;
8766 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>target_reg)&1)<<r;
8767 }
8768 }
8769 }
8770 // Merge in delay slot
8771 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8772 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8773 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8774 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8775 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8776 will_dirty_i &= branch_hr_candirty;
8777 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8778 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8779 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8780 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8781 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8782 will_dirty_i &= hr_candirty;
8783 }
8784 // Merge in delay slot (won't dirty)
8785 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8786 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8787 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8788 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8789 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8790 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8791 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8792 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8793 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8794 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8795 wont_dirty_i &= ~(1u << 31);
8796 if(wr) {
8797 #ifndef DESTRUCTIVE_WRITEBACK
8798 branch_regs[i].dirty&=wont_dirty_i;
8799 #endif
8800 branch_regs[i].dirty|=will_dirty_i;
8801 }
8802 }
8803 }
8804 }
8805 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8806 {
8807 // SYSCALL instruction (software interrupt)
8808 will_dirty_i=0;
8809 wont_dirty_i=0;
8810 }
8811 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8812 {
8813 // ERET instruction (return from interrupt)
8814 will_dirty_i=0;
8815 wont_dirty_i=0;
8816 }
8817 will_dirty_next=will_dirty_i;
8818 wont_dirty_next=wont_dirty_i;
8819 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8820 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8821 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8822 will_dirty_i &= hr_candirty;
8823 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8824 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8825 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8826 wont_dirty_i &= ~(1u << 31);
8827 if (i > istart && !dops[i].is_jump) {
8828 // Don't store a register immediately after writing it,
8829 // may prevent dual-issue.
8830 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8831 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8832 }
8833 // Save it
8834 will_dirty[i]=will_dirty_i;
8835 wont_dirty[i]=wont_dirty_i;
8836 // Mark registers that won't be dirtied as not dirty
8837 if(wr) {
8838 regs[i].dirty|=will_dirty_i;
8839 #ifndef DESTRUCTIVE_WRITEBACK
8840 regs[i].dirty&=wont_dirty_i;
8841 if(dops[i].is_jump)
8842 {
8843 if (i < iend-1 && !dops[i].is_ujump) {
8844 for(r=0;r<HOST_REGS;r++) {
8845 if(r!=EXCLUDE_REG) {
8846 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8847 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8848 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8849 }
8850 }
8851 }
8852 }
8853 else
8854 {
8855 if(i<iend) {
8856 for(r=0;r<HOST_REGS;r++) {
8857 if(r!=EXCLUDE_REG) {
8858 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8859 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8860 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8861 }
8862 }
8863 }
8864 }
8865 #endif
8866 }
8867 // Deal with changed mappings
8868 temp_will_dirty=will_dirty_i;
8869 temp_wont_dirty=wont_dirty_i;
8870 for(r=0;r<HOST_REGS;r++) {
8871 if(r!=EXCLUDE_REG) {
8872 int nr;
8873 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8874 if(wr) {
8875 #ifndef DESTRUCTIVE_WRITEBACK
8876 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8877 #endif
8878 regs[i].wasdirty|=will_dirty_i&(1<<r);
8879 }
8880 }
8881 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8882 // Register moved to a different register
8883 will_dirty_i&=~(1<<r);
8884 wont_dirty_i&=~(1<<r);
8885 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8886 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8887 if(wr) {
8888 #ifndef DESTRUCTIVE_WRITEBACK
8889 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8890 #endif
8891 regs[i].wasdirty|=will_dirty_i&(1<<r);
8892 }
8893 }
8894 else {
8895 will_dirty_i&=~(1<<r);
8896 wont_dirty_i&=~(1<<r);
8897 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8898 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8899 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8900 } else {
8901 wont_dirty_i|=1<<r;
8902 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8903 }
8904 }
8905 }
8906 }
8907 }
8908}
8909
8910static noinline void pass10_expire_blocks(void)
8911{
8912 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
8913 // not sizeof(ndrc->translation_cache) due to vita hack
8914 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
8915 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
8916 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
8917 int hit;
8918
8919 for (; expirep != end; expirep = ((expirep + step) & step_mask))
8920 {
8921 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
8922 u_int block_i = expirep / step & (PAGE_COUNT - 1);
8923 u_int phase = (expirep >> (base_shift - 1)) & 1u;
8924 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
8925 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
8926 (long)(out - ndrc->translation_cache), phase);
8927 }
8928
8929 if (!phase) {
8930 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
8931 if (hit) {
8932 do_clear_cache();
8933 #ifdef USE_MINI_HT
8934 memset(mini_ht, -1, sizeof(mini_ht));
8935 #endif
8936 }
8937 }
8938 else
8939 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
8940 }
8941}
8942
8943static struct block_info *new_block_info(u_int start, u_int len,
8944 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
8945{
8946 struct block_info **b_pptr;
8947 struct block_info *block;
8948 u_int page = get_page(start);
8949
8950 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
8951 assert(block);
8952 assert(jump_in_count > 0);
8953 block->source = source;
8954 block->copy = copy;
8955 block->start = start;
8956 block->len = len;
8957 block->reg_sv_flags = 0;
8958 block->tc_offs = beginning - ndrc->translation_cache;
8959 //block->tc_len = out - beginning;
8960 block->is_dirty = 0;
8961 block->inv_near_misses = 0;
8962 block->jump_in_cnt = jump_in_count;
8963
8964 // insert sorted by start mirror-unmasked vaddr
8965 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
8966 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
8967 block->next = *b_pptr;
8968 *b_pptr = block;
8969 break;
8970 }
8971 }
8972 stat_inc(stat_blocks);
8973 return block;
8974}
8975
8976static int new_recompile_block(u_int addr)
8977{
8978 u_int pagelimit = 0;
8979 u_int state_rflags = 0;
8980 int i;
8981
8982 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
8983
8984 // this is just for speculation
8985 for (i = 1; i < 32; i++) {
8986 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
8987 state_rflags |= 1 << i;
8988 }
8989
8990 assert(!(addr & 3));
8991 start = addr & ~3;
8992 new_dynarec_did_compile=1;
8993 if (Config.HLE && start == 0x80001000) // hlecall
8994 {
8995 // XXX: is this enough? Maybe check hleSoftCall?
8996 void *beginning = start_block();
8997
8998 emit_movimm(start,0);
8999 emit_writeword(0,&pcaddr);
9000 emit_far_jump(new_dyna_leave);
9001 literal_pool(0);
9002 end_block(beginning);
9003 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9004 block->jump_in[0].vaddr = start;
9005 block->jump_in[0].addr = beginning;
9006 return 0;
9007 }
9008 else if (f1_hack && hack_addr == 0) {
9009 void *beginning = start_block();
9010 emit_movimm(start, 0);
9011 emit_writeword(0, &hack_addr);
9012 emit_readword(&psxRegs.GPR.n.sp, 0);
9013 emit_readptr(&mem_rtab, 1);
9014 emit_shrimm(0, 12, 2);
9015 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
9016 emit_addimm(0, 0x18, 0);
9017 emit_adds_ptr(1, 1, 1);
9018 emit_ldr_dualindexed(1, 0, 0);
9019 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
9020 emit_far_call(ndrc_get_addr_ht);
9021 emit_jmpreg(0); // jr k0
9022 literal_pool(0);
9023 end_block(beginning);
9024
9025 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9026 block->jump_in[0].vaddr = start;
9027 block->jump_in[0].addr = beginning;
9028 SysPrintf("F1 hack to %08x\n", start);
9029 return 0;
9030 }
9031
9032 cycle_multiplier_active = Config.cycle_multiplier_override && Config.cycle_multiplier == CYCLE_MULT_DEFAULT
9033 ? Config.cycle_multiplier_override : Config.cycle_multiplier;
9034
9035 source = get_source_start(start, &pagelimit);
9036 if (source == NULL) {
9037 if (addr != hack_addr) {
9038 SysPrintf("Compile at bogus memory address: %08x\n", addr);
9039 hack_addr = addr;
9040 }
9041 //abort();
9042 return -1;
9043 }
9044
9045 /* Pass 1: disassemble */
9046 /* Pass 2: register dependencies, branch targets */
9047 /* Pass 3: register allocation */
9048 /* Pass 4: branch dependencies */
9049 /* Pass 5: pre-alloc */
9050 /* Pass 6: optimize clean/dirty state */
9051 /* Pass 7: flag 32-bit registers */
9052 /* Pass 8: assembly */
9053 /* Pass 9: linker */
9054 /* Pass 10: garbage collection / free memory */
9055
9056 /* Pass 1 disassembly */
9057
9058 pass1_disassemble(pagelimit);
9059
9060 int clear_hack_addr = apply_hacks();
9061
9062 /* Pass 2 - Register dependencies and branch targets */
9063
9064 pass2_unneeded_regs(0,slen-1,0);
9065
9066 /* Pass 3 - Register allocation */
9067
9068 pass3_register_alloc(addr);
9069
9070 /* Pass 4 - Cull unused host registers */
9071
9072 pass4_cull_unused_regs();
9073
9074 /* Pass 5 - Pre-allocate registers */
9075
9076 pass5a_preallocate1();
9077 pass5b_preallocate2();
9078
9079 /* Pass 6 - Optimize clean/dirty state */
9080 pass6_clean_registers(0, slen-1, 1);
9081
9082 /* Pass 7 - Identify 32-bit registers */
9083 for (i=slen-1;i>=0;i--)
9084 {
9085 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9086 {
9087 // Conditional branch
9088 if((source[i]>>16)!=0x1000&&i<slen-2) {
9089 // Mark this address as a branch target since it may be called
9090 // upon return from interrupt
9091 dops[i+2].bt=1;
9092 }
9093 }
9094 }
9095
9096 /* Pass 8 - Assembly */
9097 linkcount=0;stubcount=0;
9098 is_delayslot=0;
9099 u_int dirty_pre=0;
9100 void *beginning=start_block();
9101 void *instr_addr0_override = NULL;
9102 int ds = 0;
9103
9104 if (start == 0x80030000) {
9105 // nasty hack for the fastbios thing
9106 // override block entry to this code
9107 instr_addr0_override = out;
9108 emit_movimm(start,0);
9109 // abuse io address var as a flag that we
9110 // have already returned here once
9111 emit_readword(&address,1);
9112 emit_writeword(0,&pcaddr);
9113 emit_writeword(0,&address);
9114 emit_cmp(0,1);
9115 #ifdef __aarch64__
9116 emit_jeq(out + 4*2);
9117 emit_far_jump(new_dyna_leave);
9118 #else
9119 emit_jne(new_dyna_leave);
9120 #endif
9121 }
9122 for(i=0;i<slen;i++)
9123 {
9124 __builtin_prefetch(regs[i+1].regmap);
9125 check_regmap(regmap_pre[i]);
9126 check_regmap(regs[i].regmap_entry);
9127 check_regmap(regs[i].regmap);
9128 //if(ds) printf("ds: ");
9129 disassemble_inst(i);
9130 if(ds) {
9131 ds=0; // Skip delay slot
9132 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9133 instr_addr[i] = NULL;
9134 } else {
9135 speculate_register_values(i);
9136 #ifndef DESTRUCTIVE_WRITEBACK
9137 if (i < 2 || !dops[i-2].is_ujump)
9138 {
9139 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9140 }
9141 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9142 dirty_pre=branch_regs[i].dirty;
9143 }else{
9144 dirty_pre=regs[i].dirty;
9145 }
9146 #endif
9147 // write back
9148 if (i < 2 || !dops[i-2].is_ujump)
9149 {
9150 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9151 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9152 }
9153 // branch target entry point
9154 instr_addr[i] = out;
9155 assem_debug("<->\n");
9156 drc_dbg_emit_do_cmp(i, ccadj[i]);
9157 if (clear_hack_addr) {
9158 emit_movimm(0, 0);
9159 emit_writeword(0, &hack_addr);
9160 clear_hack_addr = 0;
9161 }
9162
9163 // load regs
9164 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9165 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9166 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9167 address_generation(i,&regs[i],regs[i].regmap_entry);
9168 load_consts(regmap_pre[i],regs[i].regmap,i);
9169 if(dops[i].is_jump)
9170 {
9171 // Load the delay slot registers if necessary
9172 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))
9173 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9174 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))
9175 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9176 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9177 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9178 if (dops[i+1].is_store)
9179 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9180 }
9181 else if(i+1<slen)
9182 {
9183 // Preload registers for following instruction
9184 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9185 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9186 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9187 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9188 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9189 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9190 }
9191 // TODO: if(is_ooo(i)) address_generation(i+1);
9192 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9193 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9194 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9195 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9196 if (dops[i].is_store)
9197 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9198
9199 ds = assemble(i, &regs[i], ccadj[i]);
9200
9201 if (dops[i].is_ujump)
9202 literal_pool(1024);
9203 else
9204 literal_pool_jumpover(256);
9205 }
9206 }
9207
9208 assert(slen > 0);
9209 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9210 // no ending needed for this block since INTCALL never returns
9211 }
9212 // If the block did not end with an unconditional branch,
9213 // add a jump to the next instruction.
9214 else if (i > 1) {
9215 if (!dops[i-2].is_ujump) {
9216 assert(!dops[i-1].is_jump);
9217 assert(i==slen);
9218 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9219 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9220 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9221 emit_loadreg(CCREG,HOST_CCREG);
9222 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9223 }
9224 else
9225 {
9226 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9227 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9228 }
9229 add_to_linker(out,start+i*4,0);
9230 emit_jmp(0);
9231 }
9232 }
9233 else
9234 {
9235 assert(i>0);
9236 assert(!dops[i-1].is_jump);
9237 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9238 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9239 emit_loadreg(CCREG,HOST_CCREG);
9240 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9241 add_to_linker(out,start+i*4,0);
9242 emit_jmp(0);
9243 }
9244
9245 // TODO: delay slot stubs?
9246 // Stubs
9247 for(i=0;i<stubcount;i++)
9248 {
9249 switch(stubs[i].type)
9250 {
9251 case LOADB_STUB:
9252 case LOADH_STUB:
9253 case LOADW_STUB:
9254 case LOADD_STUB:
9255 case LOADBU_STUB:
9256 case LOADHU_STUB:
9257 do_readstub(i);break;
9258 case STOREB_STUB:
9259 case STOREH_STUB:
9260 case STOREW_STUB:
9261 case STORED_STUB:
9262 do_writestub(i);break;
9263 case CC_STUB:
9264 do_ccstub(i);break;
9265 case INVCODE_STUB:
9266 do_invstub(i);break;
9267 case FP_STUB:
9268 do_cop1stub(i);break;
9269 case STORELR_STUB:
9270 do_unalignedwritestub(i);break;
9271 }
9272 }
9273
9274 if (instr_addr0_override)
9275 instr_addr[0] = instr_addr0_override;
9276
9277#if 0
9278 /* check for improper expiration */
9279 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9280 int j;
9281 if (!jumps[i])
9282 continue;
9283 for (j = 0; j < jumps[i]->count; j++)
9284 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9285 }
9286#endif
9287
9288 /* Pass 9 - Linker */
9289 for(i=0;i<linkcount;i++)
9290 {
9291 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9292 literal_pool(64);
9293 if (!link_addr[i].internal)
9294 {
9295 void *stub = out;
9296 void *addr = check_addr(link_addr[i].target);
9297 emit_extjump(link_addr[i].addr, link_addr[i].target);
9298 if (addr) {
9299 set_jump_target(link_addr[i].addr, addr);
9300 ndrc_add_jump_out(link_addr[i].target,stub);
9301 }
9302 else
9303 set_jump_target(link_addr[i].addr, stub);
9304 }
9305 else
9306 {
9307 // Internal branch
9308 int target=(link_addr[i].target-start)>>2;
9309 assert(target>=0&&target<slen);
9310 assert(instr_addr[target]);
9311 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9312 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9313 //#else
9314 set_jump_target(link_addr[i].addr, instr_addr[target]);
9315 //#endif
9316 }
9317 }
9318
9319 u_int source_len = slen*4;
9320 if (dops[slen-1].itype == INTCALL && source_len > 4)
9321 // no need to treat the last instruction as compiled
9322 // as interpreter fully handles it
9323 source_len -= 4;
9324
9325 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9326 copy = shadow;
9327
9328 // External Branch Targets (jump_in)
9329 int jump_in_count = 1;
9330 assert(instr_addr[0]);
9331 for (i = 1; i < slen; i++)
9332 {
9333 if (dops[i].bt && instr_addr[i])
9334 jump_in_count++;
9335 }
9336
9337 struct block_info *block =
9338 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9339 block->reg_sv_flags = state_rflags;
9340
9341 int jump_in_i = 0;
9342 for (i = 0; i < slen; i++)
9343 {
9344 if ((i == 0 || dops[i].bt) && instr_addr[i])
9345 {
9346 assem_debug("%p (%d) <- %8x\n", instr_addr[i], i, start + i*4);
9347 u_int vaddr = start + i*4;
9348
9349 literal_pool(256);
9350 void *entry = out;
9351 load_regs_entry(i);
9352 if (entry == out)
9353 entry = instr_addr[i];
9354 else
9355 emit_jmp(instr_addr[i]);
9356
9357 block->jump_in[jump_in_i].vaddr = vaddr;
9358 block->jump_in[jump_in_i].addr = entry;
9359 jump_in_i++;
9360 }
9361 }
9362 assert(jump_in_i == jump_in_count);
9363 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9364 // Write out the literal pool if necessary
9365 literal_pool(0);
9366 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9367 // Align code
9368 if(((u_int)out)&7) emit_addnop(13);
9369 #endif
9370 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9371 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9372 memcpy(copy, source, source_len);
9373 copy += source_len;
9374
9375 end_block(beginning);
9376
9377 // If we're within 256K of the end of the buffer,
9378 // start over from the beginning. (Is 256K enough?)
9379 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9380 out = ndrc->translation_cache;
9381
9382 // Trap writes to any of the pages we compiled
9383 mark_invalid_code(start, slen*4, 0);
9384
9385 /* Pass 10 - Free memory by expiring oldest blocks */
9386
9387 pass10_expire_blocks();
9388
9389#ifdef ASSEM_PRINT
9390 fflush(stdout);
9391#endif
9392 stat_inc(stat_bc_direct);
9393 return 0;
9394}
9395
9396// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork