notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
drc: rework smc checks again
[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 struct ht_entry hash_table[65536];
246 static struct block_info *blocks[PAGE_COUNT];
247 static struct jump_info *jumps[PAGE_COUNT];
248 static u_int start;
249 static u_int *source;
250 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
251 static uint64_t gte_rt[MAXBLOCK];
252 static uint64_t gte_unneeded[MAXBLOCK];
253 static u_int smrv[32]; // speculated MIPS register values
254 static u_int smrv_strong; // mask or regs that are likely to have correct values
255 static u_int smrv_weak; // same, but somewhat less likely
256 static u_int smrv_strong_next; // same, but after current insn executes
257 static u_int smrv_weak_next;
258 static int imm[MAXBLOCK];
259 static u_int ba[MAXBLOCK];
260 static uint64_t unneeded_reg[MAXBLOCK];
261 static uint64_t branch_unneeded_reg[MAXBLOCK];
262 // see 'struct regstat' for a description
263 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
264 // contains 'real' consts at [i] insn, but may differ from what's actually
265 // loaded in host reg as 'final' value is always loaded, see get_final_value()
266 static uint32_t current_constmap[HOST_REGS];
267 static uint32_t constmap[MAXBLOCK][HOST_REGS];
268 static struct regstat regs[MAXBLOCK];
269 static struct regstat branch_regs[MAXBLOCK];
270 static signed char minimum_free_regs[MAXBLOCK];
271 static int ccadj[MAXBLOCK];
272 static int slen;
273 static void *instr_addr[MAXBLOCK];
274 static struct link_entry link_addr[MAXBLOCK];
275 static int linkcount;
276 static struct code_stub stubs[MAXBLOCK*3];
277 static int stubcount;
278 static u_int literals[1024][2];
279 static int literalcount;
280 static int is_delayslot;
281 static char shadow[1048576] __attribute__((aligned(16)));
282 static void *copy;
283 static u_int expirep;
284 static u_int stop_after_jal;
285 static u_int f1_hack;
286#ifdef STAT_PRINT
287 static int stat_bc_direct;
288 static int stat_bc_pre;
289 static int stat_bc_restore;
290 static int stat_ht_lookups;
291 static int stat_jump_in_lookups;
292 static int stat_restore_tries;
293 static int stat_restore_compares;
294 static int stat_inv_addr_calls;
295 static int stat_inv_hits;
296 static int stat_blocks;
297 static int stat_links;
298 #define stat_inc(s) s++
299 #define stat_dec(s) s--
300 #define stat_clear(s) s = 0
301#else
302 #define stat_inc(s)
303 #define stat_dec(s)
304 #define stat_clear(s)
305#endif
306
307 int new_dynarec_hacks;
308 int new_dynarec_hacks_pergame;
309 int new_dynarec_hacks_old;
310 int new_dynarec_did_compile;
311
312 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
313
314 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
315 extern int last_count; // last absolute target, often = next_interupt
316 extern int pcaddr;
317 extern int pending_exception;
318 extern int branch_target;
319 extern uintptr_t ram_offset;
320 extern uintptr_t mini_ht[32][2];
321
322 /* registers that may be allocated */
323 /* 1-31 gpr */
324#define LOREG 32 // lo
325#define HIREG 33 // hi
326//#define FSREG 34 // FPU status (FCSR)
327#define CSREG 35 // Coprocessor status
328#define CCREG 36 // Cycle count
329#define INVCP 37 // Pointer to invalid_code
330//#define MMREG 38 // Pointer to memory_map
331#define ROREG 39 // ram offset (if rdram!=0x80000000)
332#define TEMPREG 40
333#define FTEMP 40 // FPU temporary register
334#define PTEMP 41 // Prefetch temporary register
335//#define TLREG 42 // TLB mapping offset
336#define RHASH 43 // Return address hash
337#define RHTBL 44 // Return address hash table address
338#define RTEMP 45 // JR/JALR address register
339#define MAXREG 45
340#define AGEN1 46 // Address generation temporary register
341//#define AGEN2 47 // Address generation temporary register
342//#define MGEN1 48 // Maptable address generation temporary register
343//#define MGEN2 49 // Maptable address generation temporary register
344#define BTREG 50 // Branch target temporary register
345
346 /* instruction types */
347#define NOP 0 // No operation
348#define LOAD 1 // Load
349#define STORE 2 // Store
350#define LOADLR 3 // Unaligned load
351#define STORELR 4 // Unaligned store
352#define MOV 5 // Move
353#define ALU 6 // Arithmetic/logic
354#define MULTDIV 7 // Multiply/divide
355#define SHIFT 8 // Shift by register
356#define SHIFTIMM 9// Shift by immediate
357#define IMM16 10 // 16-bit immediate
358#define RJUMP 11 // Unconditional jump to register
359#define UJUMP 12 // Unconditional jump
360#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
361#define SJUMP 14 // Conditional branch (regimm format)
362#define COP0 15 // Coprocessor 0
363#define COP1 16 // Coprocessor 1
364#define C1LS 17 // Coprocessor 1 load/store
365//#define FJUMP 18 // Conditional branch (floating point)
366//#define FLOAT 19 // Floating point unit
367//#define FCONV 20 // Convert integer to float
368//#define FCOMP 21 // Floating point compare (sets FSREG)
369#define SYSCALL 22// SYSCALL,BREAK
370#define OTHER 23 // Other
371//#define SPAN 24 // Branch/delay slot spans 2 pages
372#define NI 25 // Not implemented
373#define HLECALL 26// PCSX fake opcodes for HLE
374#define COP2 27 // Coprocessor 2 move
375#define C2LS 28 // Coprocessor 2 load/store
376#define C2OP 29 // Coprocessor 2 operation
377#define INTCALL 30// Call interpreter to handle rare corner cases
378
379 /* branch codes */
380#define TAKEN 1
381#define NOTTAKEN 2
382#define NULLDS 3
383
384#define DJT_1 (void *)1l // no function, just a label in assem_debug log
385#define DJT_2 (void *)2l
386
387// asm linkage
388void dyna_linker();
389void cc_interrupt();
390void fp_exception();
391void fp_exception_ds();
392void jump_syscall (u_int u0, u_int u1, u_int pc);
393void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
394void jump_break (u_int u0, u_int u1, u_int pc);
395void jump_break_ds(u_int u0, u_int u1, u_int pc);
396void jump_to_new_pc();
397void call_gteStall();
398void new_dyna_leave();
399
400void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile);
401void *ndrc_get_addr_ht(u_int vaddr);
402void ndrc_add_jump_out(u_int vaddr, void *src);
403void ndrc_write_invalidate_one(u_int addr);
404static void ndrc_write_invalidate_many(u_int addr, u_int end);
405
406static int new_recompile_block(u_int addr);
407static void invalidate_block(struct block_info *block);
408
409// Needed by assembler
410static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
411static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
412static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
413static void load_all_regs(const signed char i_regmap[]);
414static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
415static void load_regs_entry(int t);
416static void load_all_consts(const signed char regmap[], u_int dirty, int i);
417static u_int get_host_reglist(const signed char *regmap);
418
419static int get_final_value(int hr, int i, int *value);
420static void add_stub(enum stub_type type, void *addr, void *retaddr,
421 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
422static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
423 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
424static void add_to_linker(void *addr, u_int target, int ext);
425static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
426 int addr, int *offset_reg, int *addr_reg_override);
427static void *get_direct_memhandler(void *table, u_int addr,
428 enum stub_type type, uintptr_t *addr_host);
429static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
430static void pass_args(int a0, int a1);
431static void emit_far_jump(const void *f);
432static void emit_far_call(const void *f);
433
434#ifdef VITA
435#include <psp2/kernel/sysmem.h>
436static int sceBlock;
437// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
438extern int getVMBlock();
439int _newlib_vm_size_user = sizeof(*ndrc);
440#endif
441
442static void mprotect_w_x(void *start, void *end, int is_x)
443{
444#ifdef NO_WRITE_EXEC
445 #if defined(VITA)
446 // *Open* enables write on all memory that was
447 // allocated by sceKernelAllocMemBlockForVM()?
448 if (is_x)
449 sceKernelCloseVMDomain();
450 else
451 sceKernelOpenVMDomain();
452 #elif defined(HAVE_LIBNX)
453 Result rc;
454 // check to avoid the full flush in jitTransitionToExecutable()
455 if (g_jit.type != JitType_CodeMemory) {
456 if (is_x)
457 rc = jitTransitionToExecutable(&g_jit);
458 else
459 rc = jitTransitionToWritable(&g_jit);
460 if (R_FAILED(rc))
461 ;//SysPrintf("jitTransition %d %08x\n", is_x, rc);
462 }
463 #elif defined(TC_WRITE_OFFSET)
464 // separated rx and rw areas are always available
465 #else
466 u_long mstart = (u_long)start & ~4095ul;
467 u_long mend = (u_long)end;
468 if (mprotect((void *)mstart, mend - mstart,
469 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
470 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
471 #endif
472#endif
473}
474
475static void start_tcache_write(void *start, void *end)
476{
477 mprotect_w_x(start, end, 0);
478}
479
480static void end_tcache_write(void *start, void *end)
481{
482#if defined(__arm__) || defined(__aarch64__)
483 size_t len = (char *)end - (char *)start;
484 #if defined(__BLACKBERRY_QNX__)
485 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
486 #elif defined(__MACH__)
487 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
488 #elif defined(VITA)
489 sceKernelSyncVMDomain(sceBlock, start, len);
490 #elif defined(_3DS)
491 ctr_flush_invalidate_cache();
492 #elif defined(HAVE_LIBNX)
493 if (g_jit.type == JitType_CodeMemory) {
494 armDCacheClean(start, len);
495 armICacheInvalidate((char *)start - ndrc_write_ofs, len);
496 // as of v4.2.1 libnx lacks isb
497 __asm__ volatile("isb" ::: "memory");
498 }
499 #elif defined(__aarch64__)
500 // as of 2021, __clear_cache() is still broken on arm64
501 // so here is a custom one :(
502 clear_cache_arm64(start, end);
503 #else
504 __clear_cache(start, end);
505 #endif
506 (void)len;
507#endif
508
509 mprotect_w_x(start, end, 1);
510}
511
512static void *start_block(void)
513{
514 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
515 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
516 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
517 start_tcache_write(NDRC_WRITE_OFFSET(out), NDRC_WRITE_OFFSET(end));
518 return out;
519}
520
521static void end_block(void *start)
522{
523 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(out));
524}
525
526#ifdef NDRC_CACHE_FLUSH_ALL
527
528static int needs_clear_cache;
529
530static void mark_clear_cache(void *target)
531{
532 if (!needs_clear_cache) {
533 start_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
534 needs_clear_cache = 1;
535 }
536}
537
538static void do_clear_cache(void)
539{
540 if (needs_clear_cache) {
541 end_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
542 needs_clear_cache = 0;
543 }
544}
545
546#else
547
548// also takes care of w^x mappings when patching code
549static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
550
551static void mark_clear_cache(void *target)
552{
553 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
554 u_int mask = 1u << ((offset >> 12) & 31);
555 if (!(needs_clear_cache[offset >> 17] & mask)) {
556 char *start = (char *)NDRC_WRITE_OFFSET((uintptr_t)target & ~4095l);
557 start_tcache_write(start, start + 4095);
558 needs_clear_cache[offset >> 17] |= mask;
559 }
560}
561
562// Clearing the cache is rather slow on ARM Linux, so mark the areas
563// that need to be cleared, and then only clear these areas once.
564static void do_clear_cache(void)
565{
566 int i, j;
567 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
568 {
569 u_int bitmap = needs_clear_cache[i];
570 if (!bitmap)
571 continue;
572 for (j = 0; j < 32; j++)
573 {
574 u_char *start, *end;
575 if (!(bitmap & (1u << j)))
576 continue;
577
578 start = ndrc->translation_cache + i*131072 + j*4096;
579 end = start + 4095;
580 for (j++; j < 32; j++) {
581 if (!(bitmap & (1u << j)))
582 break;
583 end += 4096;
584 }
585 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(end));
586 }
587 needs_clear_cache[i] = 0;
588 }
589}
590
591#endif // NDRC_CACHE_FLUSH_ALL
592
593#define NO_CYCLE_PENALTY_THR 12
594
595int cycle_multiplier = CYCLE_MULT_DEFAULT; // 100 for 1.0
596int cycle_multiplier_override;
597int cycle_multiplier_old;
598static int cycle_multiplier_active;
599
600static int CLOCK_ADJUST(int x)
601{
602 int m = cycle_multiplier_active;
603 int s = (x >> 31) | 1;
604 return (x * m + s * 50) / 100;
605}
606
607static int ds_writes_rjump_rs(int i)
608{
609 return dops[i].rs1 != 0 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2);
610}
611
612// psx addr mirror masking (for invalidation)
613static u_int pmmask(u_int vaddr)
614{
615 vaddr &= ~0xe0000000;
616 if (vaddr < 0x01000000)
617 vaddr &= ~0x00e00000; // RAM mirrors
618 return vaddr;
619}
620
621static u_int get_page(u_int vaddr)
622{
623 u_int page = pmmask(vaddr) >> 12;
624 if (page >= PAGE_COUNT / 2)
625 page = PAGE_COUNT / 2 + (page & (PAGE_COUNT / 2 - 1));
626 return page;
627}
628
629// get a page for looking for a block that has vaddr
630// (needed because the block may start in previous page)
631static u_int get_page_prev(u_int vaddr)
632{
633 assert(MAXBLOCK <= (1 << 12));
634 u_int page = get_page(vaddr);
635 if (page & 511)
636 page--;
637 return page;
638}
639
640static struct ht_entry *hash_table_get(u_int vaddr)
641{
642 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
643}
644
645static void hash_table_add(u_int vaddr, void *tcaddr)
646{
647 struct ht_entry *ht_bin = hash_table_get(vaddr);
648 assert(tcaddr);
649 ht_bin->vaddr[1] = ht_bin->vaddr[0];
650 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
651 ht_bin->vaddr[0] = vaddr;
652 ht_bin->tcaddr[0] = tcaddr;
653}
654
655static void hash_table_remove(int vaddr)
656{
657 //printf("remove hash: %x\n",vaddr);
658 struct ht_entry *ht_bin = hash_table_get(vaddr);
659 if (ht_bin->vaddr[1] == vaddr) {
660 ht_bin->vaddr[1] = -1;
661 ht_bin->tcaddr[1] = NULL;
662 }
663 if (ht_bin->vaddr[0] == vaddr) {
664 ht_bin->vaddr[0] = ht_bin->vaddr[1];
665 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
666 ht_bin->vaddr[1] = -1;
667 ht_bin->tcaddr[1] = NULL;
668 }
669}
670
671static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
672{
673 u_int vaddr_m = vaddr & 0x1fffffff;
674 u_int i, j;
675 for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
676 // ram mirrors, but should not hurt bios
677 for (j = 0; j < 0x800000; j += 0x200000) {
678 invalid_code[(i|j) >> 12] =
679 invalid_code[(i|j|0x80000000u) >> 12] =
680 invalid_code[(i|j|0xa0000000u) >> 12] = invalid;
681 }
682 }
683 if (!invalid && vaddr + len > inv_code_start && vaddr <= inv_code_end)
684 inv_code_start = inv_code_end = ~0;
685}
686
687static int doesnt_expire_soon(u_char *tcaddr)
688{
689 u_int diff = (u_int)(tcaddr - out) & ((1u << TARGET_SIZE_2) - 1u);
690 return diff > EXPIRITY_OFFSET + MAX_OUTPUT_BLOCK_SIZE;
691}
692
693static unused void check_for_block_changes(u_int start, u_int end)
694{
695 u_int start_page = get_page_prev(start);
696 u_int end_page = get_page(end - 1);
697 u_int page;
698
699 for (page = start_page; page <= end_page; page++) {
700 struct block_info *block;
701 for (block = blocks[page]; block != NULL; block = block->next) {
702 if (block->is_dirty)
703 continue;
704 if (memcmp(block->source, block->copy, block->len)) {
705 printf("bad block %08x-%08x %016llx %016llx @%08x\n",
706 block->start, block->start + block->len,
707 *(long long *)block->source, *(long long *)block->copy, psxRegs.pc);
708 fflush(stdout);
709 abort();
710 }
711 }
712 }
713}
714
715static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
716{
717 void *found_clean = NULL;
718 u_int i, page;
719
720 stat_inc(stat_restore_tries);
721 for (page = start_page; page <= end_page; page++) {
722 struct block_info *block;
723 for (block = blocks[page]; block != NULL; block = block->next) {
724 if (vaddr < block->start)
725 break;
726 if (!block->is_dirty || vaddr >= block->start + block->len)
727 continue;
728 for (i = 0; i < block->jump_in_cnt; i++)
729 if (block->jump_in[i].vaddr == vaddr)
730 break;
731 if (i == block->jump_in_cnt)
732 continue;
733 assert(block->source && block->copy);
734 stat_inc(stat_restore_compares);
735 if (memcmp(block->source, block->copy, block->len))
736 continue;
737
738 block->is_dirty = block->inv_near_misses = 0;
739 found_clean = block->jump_in[i].addr;
740 hash_table_add(vaddr, found_clean);
741 mark_invalid_code(block->start, block->len, 0);
742 stat_inc(stat_bc_restore);
743 inv_debug("INV: restored %08x %p (%d)\n", vaddr, found_clean, block->jump_in_cnt);
744 return found_clean;
745 }
746 }
747 return NULL;
748}
749
750// Get address from virtual address
751// This is called from the recompiled JR/JALR instructions
752static void noinline *get_addr(u_int vaddr, int can_compile)
753{
754 u_int start_page = get_page_prev(vaddr);
755 u_int i, page, end_page = get_page(vaddr);
756 void *found_clean = NULL;
757
758 stat_inc(stat_jump_in_lookups);
759 for (page = start_page; page <= end_page; page++) {
760 const struct block_info *block;
761 for (block = blocks[page]; block != NULL; block = block->next) {
762 if (vaddr < block->start)
763 break;
764 if (block->is_dirty || vaddr >= block->start + block->len)
765 continue;
766 for (i = 0; i < block->jump_in_cnt; i++)
767 if (block->jump_in[i].vaddr == vaddr)
768 break;
769 if (i == block->jump_in_cnt)
770 continue;
771 found_clean = block->jump_in[i].addr;
772 hash_table_add(vaddr, found_clean);
773 return found_clean;
774 }
775 }
776 found_clean = try_restore_block(vaddr, start_page, end_page);
777 if (found_clean)
778 return found_clean;
779
780 if (!can_compile)
781 return NULL;
782
783 int r = new_recompile_block(vaddr);
784 if (r == 0)
785 return ndrc_get_addr_ht(vaddr);
786
787 // generate an address error
788 Status|=2;
789 Cause=(vaddr<<31)|(4<<2);
790 EPC=(vaddr&1)?vaddr-5:vaddr;
791 BadVAddr=(vaddr&~1);
792 return ndrc_get_addr_ht(0x80000080);
793}
794
795// Look up address in hash table first
796void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile)
797{
798 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
799 const struct ht_entry *ht_bin = hash_table_get(vaddr);
800 stat_inc(stat_ht_lookups);
801 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
802 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
803 return get_addr(vaddr, can_compile);
804}
805
806void *ndrc_get_addr_ht(u_int vaddr)
807{
808 return ndrc_get_addr_ht_param(vaddr, 1);
809}
810
811static void clear_all_regs(signed char regmap[])
812{
813 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
814}
815
816// get_reg: get allocated host reg from mips reg
817// returns -1 if no such mips reg was allocated
818#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
819
820extern signed char get_reg(const signed char regmap[], signed char r);
821
822#else
823
824static signed char get_reg(const signed char regmap[], signed char r)
825{
826 int hr;
827 for (hr = 0; hr < HOST_REGS; hr++) {
828 if (hr == EXCLUDE_REG)
829 continue;
830 if (regmap[hr] == r)
831 return hr;
832 }
833 return -1;
834}
835
836#endif
837
838// get reg as mask bit (1 << hr)
839static u_int get_regm(const signed char regmap[], signed char r)
840{
841 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
842}
843
844static signed char get_reg_temp(const signed char regmap[])
845{
846 int hr;
847 for (hr = 0; hr < HOST_REGS; hr++) {
848 if (hr == EXCLUDE_REG)
849 continue;
850 if (regmap[hr] == (signed char)-1)
851 return hr;
852 }
853 return -1;
854}
855
856// Find a register that is available for two consecutive cycles
857static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
858{
859 int hr;
860 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
861 return -1;
862}
863
864// reverse reg map: mips -> host
865#define RRMAP_SIZE 64
866static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
867 u_int *regs_can_change)
868{
869 u_int r, hr, hr_can_change = 0;
870 memset(rrmap, -1, RRMAP_SIZE);
871 for (hr = 0; hr < HOST_REGS; )
872 {
873 r = regmap[hr];
874 rrmap[r & (RRMAP_SIZE - 1)] = hr;
875 // only add mips $1-$31+$lo, others shifted out
876 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
877 hr++;
878 if (hr == EXCLUDE_REG)
879 hr++;
880 }
881 hr_can_change |= 1u << (rrmap[33] & 31);
882 hr_can_change |= 1u << (rrmap[CCREG] & 31);
883 hr_can_change &= ~(1u << 31);
884 *regs_can_change = hr_can_change;
885}
886
887// same as get_reg, but takes rrmap
888static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
889{
890 assert(0 <= r && r < RRMAP_SIZE);
891 return rrmap[r];
892}
893
894static int count_free_regs(const signed char regmap[])
895{
896 int count=0;
897 int hr;
898 for(hr=0;hr<HOST_REGS;hr++)
899 {
900 if(hr!=EXCLUDE_REG) {
901 if(regmap[hr]<0) count++;
902 }
903 }
904 return count;
905}
906
907static void dirty_reg(struct regstat *cur, signed char reg)
908{
909 int hr;
910 if (!reg) return;
911 hr = get_reg(cur->regmap, reg);
912 if (hr >= 0)
913 cur->dirty |= 1<<hr;
914}
915
916static void set_const(struct regstat *cur, signed char reg, uint32_t value)
917{
918 int hr;
919 if (!reg) return;
920 hr = get_reg(cur->regmap, reg);
921 if (hr >= 0) {
922 cur->isconst |= 1<<hr;
923 current_constmap[hr] = value;
924 }
925}
926
927static void clear_const(struct regstat *cur, signed char reg)
928{
929 int hr;
930 if (!reg) return;
931 hr = get_reg(cur->regmap, reg);
932 if (hr >= 0)
933 cur->isconst &= ~(1<<hr);
934}
935
936static int is_const(const struct regstat *cur, signed char reg)
937{
938 int hr;
939 if (reg < 0) return 0;
940 if (!reg) return 1;
941 hr = get_reg(cur->regmap, reg);
942 if (hr >= 0)
943 return (cur->isconst>>hr)&1;
944 return 0;
945}
946
947static uint32_t get_const(const struct regstat *cur, signed char reg)
948{
949 int hr;
950 if (!reg) return 0;
951 hr = get_reg(cur->regmap, reg);
952 if (hr >= 0)
953 return current_constmap[hr];
954
955 SysPrintf("Unknown constant in r%d\n", reg);
956 abort();
957}
958
959// Least soon needed registers
960// Look at the next ten instructions and see which registers
961// will be used. Try not to reallocate these.
962static void lsn(u_char hsn[], int i, int *preferred_reg)
963{
964 int j;
965 int b=-1;
966 for(j=0;j<9;j++)
967 {
968 if(i+j>=slen) {
969 j=slen-i-1;
970 break;
971 }
972 if (dops[i+j].is_ujump)
973 {
974 // Don't go past an unconditonal jump
975 j++;
976 break;
977 }
978 }
979 for(;j>=0;j--)
980 {
981 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
982 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
983 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
984 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
985 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
986 // Stores can allocate zero
987 hsn[dops[i+j].rs1]=j;
988 hsn[dops[i+j].rs2]=j;
989 }
990 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
991 hsn[ROREG] = j;
992 // On some architectures stores need invc_ptr
993 #if defined(HOST_IMM8)
994 if (dops[i+j].is_store)
995 hsn[INVCP] = j;
996 #endif
997 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
998 {
999 hsn[CCREG]=j;
1000 b=j;
1001 }
1002 }
1003 if(b>=0)
1004 {
1005 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
1006 {
1007 // Follow first branch
1008 int t=(ba[i+b]-start)>>2;
1009 j=7-b;if(t+j>=slen) j=slen-t-1;
1010 for(;j>=0;j--)
1011 {
1012 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1013 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1014 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1015 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1016 }
1017 }
1018 // TODO: preferred register based on backward branch
1019 }
1020 // Delay slot should preferably not overwrite branch conditions or cycle count
1021 if (i > 0 && dops[i-1].is_jump) {
1022 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1023 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1024 hsn[CCREG]=1;
1025 // ...or hash tables
1026 hsn[RHASH]=1;
1027 hsn[RHTBL]=1;
1028 }
1029 // Coprocessor load/store needs FTEMP, even if not declared
1030 if(dops[i].itype==C2LS) {
1031 hsn[FTEMP]=0;
1032 }
1033 // Load L/R also uses FTEMP as a temporary register
1034 if(dops[i].itype==LOADLR) {
1035 hsn[FTEMP]=0;
1036 }
1037 // Also SWL/SWR/SDL/SDR
1038 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
1039 hsn[FTEMP]=0;
1040 }
1041 // Don't remove the miniht registers
1042 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1043 {
1044 hsn[RHASH]=0;
1045 hsn[RHTBL]=0;
1046 }
1047}
1048
1049// We only want to allocate registers if we're going to use them again soon
1050static int needed_again(int r, int i)
1051{
1052 int j;
1053 int b=-1;
1054 int rn=10;
1055
1056 if (i > 0 && dops[i-1].is_ujump)
1057 {
1058 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
1059 return 0; // Don't need any registers if exiting the block
1060 }
1061 for(j=0;j<9;j++)
1062 {
1063 if(i+j>=slen) {
1064 j=slen-i-1;
1065 break;
1066 }
1067 if (dops[i+j].is_ujump)
1068 {
1069 // Don't go past an unconditonal jump
1070 j++;
1071 break;
1072 }
1073 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
1074 {
1075 break;
1076 }
1077 }
1078 for(;j>=1;j--)
1079 {
1080 if(dops[i+j].rs1==r) rn=j;
1081 if(dops[i+j].rs2==r) rn=j;
1082 if((unneeded_reg[i+j]>>r)&1) rn=10;
1083 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1084 {
1085 b=j;
1086 }
1087 }
1088 if(rn<10) return 1;
1089 (void)b;
1090 return 0;
1091}
1092
1093// Try to match register allocations at the end of a loop with those
1094// at the beginning
1095static int loop_reg(int i, int r, int hr)
1096{
1097 int j,k;
1098 for(j=0;j<9;j++)
1099 {
1100 if(i+j>=slen) {
1101 j=slen-i-1;
1102 break;
1103 }
1104 if (dops[i+j].is_ujump)
1105 {
1106 // Don't go past an unconditonal jump
1107 j++;
1108 break;
1109 }
1110 }
1111 k=0;
1112 if(i>0){
1113 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1114 k--;
1115 }
1116 for(;k<j;k++)
1117 {
1118 assert(r < 64);
1119 if((unneeded_reg[i+k]>>r)&1) return hr;
1120 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1121 {
1122 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
1123 {
1124 int t=(ba[i+k]-start)>>2;
1125 int reg=get_reg(regs[t].regmap_entry,r);
1126 if(reg>=0) return reg;
1127 //reg=get_reg(regs[t+1].regmap_entry,r);
1128 //if(reg>=0) return reg;
1129 }
1130 }
1131 }
1132 return hr;
1133}
1134
1135
1136// Allocate every register, preserving source/target regs
1137static void alloc_all(struct regstat *cur,int i)
1138{
1139 int hr;
1140
1141 for(hr=0;hr<HOST_REGS;hr++) {
1142 if(hr!=EXCLUDE_REG) {
1143 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1144 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1145 {
1146 cur->regmap[hr]=-1;
1147 cur->dirty&=~(1<<hr);
1148 }
1149 // Don't need zeros
1150 if(cur->regmap[hr]==0)
1151 {
1152 cur->regmap[hr]=-1;
1153 cur->dirty&=~(1<<hr);
1154 }
1155 }
1156 }
1157}
1158
1159#ifndef NDEBUG
1160static int host_tempreg_in_use;
1161
1162static void host_tempreg_acquire(void)
1163{
1164 assert(!host_tempreg_in_use);
1165 host_tempreg_in_use = 1;
1166}
1167
1168static void host_tempreg_release(void)
1169{
1170 host_tempreg_in_use = 0;
1171}
1172#else
1173static void host_tempreg_acquire(void) {}
1174static void host_tempreg_release(void) {}
1175#endif
1176
1177#ifdef ASSEM_PRINT
1178extern void gen_interupt();
1179extern void do_insn_cmp();
1180#define FUNCNAME(f) { f, " " #f }
1181static const struct {
1182 void *addr;
1183 const char *name;
1184} function_names[] = {
1185 FUNCNAME(cc_interrupt),
1186 FUNCNAME(gen_interupt),
1187 FUNCNAME(ndrc_get_addr_ht),
1188 FUNCNAME(jump_handler_read8),
1189 FUNCNAME(jump_handler_read16),
1190 FUNCNAME(jump_handler_read32),
1191 FUNCNAME(jump_handler_write8),
1192 FUNCNAME(jump_handler_write16),
1193 FUNCNAME(jump_handler_write32),
1194 FUNCNAME(ndrc_write_invalidate_one),
1195 FUNCNAME(ndrc_write_invalidate_many),
1196 FUNCNAME(jump_to_new_pc),
1197 FUNCNAME(jump_break),
1198 FUNCNAME(jump_break_ds),
1199 FUNCNAME(jump_syscall),
1200 FUNCNAME(jump_syscall_ds),
1201 FUNCNAME(call_gteStall),
1202 FUNCNAME(new_dyna_leave),
1203 FUNCNAME(pcsx_mtc0),
1204 FUNCNAME(pcsx_mtc0_ds),
1205#ifdef DRC_DBG
1206 FUNCNAME(do_insn_cmp),
1207#endif
1208};
1209
1210static const char *func_name(const void *a)
1211{
1212 int i;
1213 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1214 if (function_names[i].addr == a)
1215 return function_names[i].name;
1216 return "";
1217}
1218#else
1219#define func_name(x) ""
1220#endif
1221
1222#ifdef __i386__
1223#include "assem_x86.c"
1224#endif
1225#ifdef __x86_64__
1226#include "assem_x64.c"
1227#endif
1228#ifdef __arm__
1229#include "assem_arm.c"
1230#endif
1231#ifdef __aarch64__
1232#include "assem_arm64.c"
1233#endif
1234
1235static void *get_trampoline(const void *f)
1236{
1237 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1238 size_t i;
1239
1240 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1241 if (tramp->f[i] == f || tramp->f[i] == NULL)
1242 break;
1243 }
1244 if (i == ARRAY_SIZE(tramp->f)) {
1245 SysPrintf("trampoline table is full, last func %p\n", f);
1246 abort();
1247 }
1248 if (tramp->f[i] == NULL) {
1249 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1250 tramp->f[i] = f;
1251 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1252#ifdef HAVE_LIBNX
1253 // invalidate the RX mirror (unsure if necessary, but just in case...)
1254 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1255#endif
1256 }
1257 return &ndrc->tramp.ops[i];
1258}
1259
1260static void emit_far_jump(const void *f)
1261{
1262 if (can_jump_or_call(f)) {
1263 emit_jmp(f);
1264 return;
1265 }
1266
1267 f = get_trampoline(f);
1268 emit_jmp(f);
1269}
1270
1271static void emit_far_call(const void *f)
1272{
1273 if (can_jump_or_call(f)) {
1274 emit_call(f);
1275 return;
1276 }
1277
1278 f = get_trampoline(f);
1279 emit_call(f);
1280}
1281
1282// Check if an address is already compiled
1283// but don't return addresses which are about to expire from the cache
1284static void *check_addr(u_int vaddr)
1285{
1286 struct ht_entry *ht_bin = hash_table_get(vaddr);
1287 size_t i;
1288 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1289 if (ht_bin->vaddr[i] == vaddr)
1290 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1291 return ht_bin->tcaddr[i];
1292 }
1293
1294 // refactor to get_addr_nocompile?
1295 u_int start_page = get_page_prev(vaddr);
1296 u_int page, end_page = get_page(vaddr);
1297
1298 stat_inc(stat_jump_in_lookups);
1299 for (page = start_page; page <= end_page; page++) {
1300 const struct block_info *block;
1301 for (block = blocks[page]; block != NULL; block = block->next) {
1302 if (vaddr < block->start)
1303 break;
1304 if (block->is_dirty || vaddr >= block->start + block->len)
1305 continue;
1306 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1307 continue;
1308 for (i = 0; i < block->jump_in_cnt; i++)
1309 if (block->jump_in[i].vaddr == vaddr)
1310 break;
1311 if (i == block->jump_in_cnt)
1312 continue;
1313
1314 // Update existing entry with current address
1315 void *addr = block->jump_in[i].addr;
1316 if (ht_bin->vaddr[0] == vaddr) {
1317 ht_bin->tcaddr[0] = addr;
1318 return addr;
1319 }
1320 if (ht_bin->vaddr[1] == vaddr) {
1321 ht_bin->tcaddr[1] = addr;
1322 return addr;
1323 }
1324 // Insert into hash table with low priority.
1325 // Don't evict existing entries, as they are probably
1326 // addresses that are being accessed frequently.
1327 if (ht_bin->vaddr[0] == -1) {
1328 ht_bin->vaddr[0] = vaddr;
1329 ht_bin->tcaddr[0] = addr;
1330 }
1331 else if (ht_bin->vaddr[1] == -1) {
1332 ht_bin->vaddr[1] = vaddr;
1333 ht_bin->tcaddr[1] = addr;
1334 }
1335 return addr;
1336 }
1337 }
1338 return NULL;
1339}
1340
1341static void blocks_clear(struct block_info **head)
1342{
1343 struct block_info *cur, *next;
1344
1345 if ((cur = *head)) {
1346 *head = NULL;
1347 while (cur) {
1348 next = cur->next;
1349 free(cur);
1350 cur = next;
1351 }
1352 }
1353}
1354
1355static int blocks_remove_matching_addrs(struct block_info **head,
1356 u_int base_offs, int shift)
1357{
1358 struct block_info *next;
1359 int hit = 0;
1360 while (*head) {
1361 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1362 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1363 invalidate_block(*head);
1364 next = (*head)->next;
1365 free(*head);
1366 *head = next;
1367 stat_dec(stat_blocks);
1368 hit = 1;
1369 }
1370 else
1371 {
1372 head = &((*head)->next);
1373 }
1374 }
1375 return hit;
1376}
1377
1378// This is called when we write to a compiled block (see do_invstub)
1379static void unlink_jumps_vaddr_range(u_int start, u_int end)
1380{
1381 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1382 int i;
1383
1384 for (page = start_page; page <= end_page; page++) {
1385 struct jump_info *ji = jumps[page];
1386 if (ji == NULL)
1387 continue;
1388 for (i = 0; i < ji->count; ) {
1389 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1390 i++;
1391 continue;
1392 }
1393
1394 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1395 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1396 void *host_addr = find_extjump_insn(ji->e[i].stub);
1397 mark_clear_cache(host_addr);
1398 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1399
1400 stat_dec(stat_links);
1401 ji->count--;
1402 if (i < ji->count) {
1403 ji->e[i] = ji->e[ji->count];
1404 continue;
1405 }
1406 i++;
1407 }
1408 }
1409}
1410
1411static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1412{
1413 int i;
1414 if (ji == NULL)
1415 return;
1416 for (i = 0; i < ji->count; ) {
1417 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1418 if (((tc_offs ^ base_offs) >> shift) != 0) {
1419 i++;
1420 continue;
1421 }
1422
1423 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1424 stat_dec(stat_links);
1425 ji->count--;
1426 if (i < ji->count) {
1427 ji->e[i] = ji->e[ji->count];
1428 continue;
1429 }
1430 i++;
1431 }
1432}
1433
1434static void invalidate_block(struct block_info *block)
1435{
1436 u_int i;
1437
1438 block->is_dirty = 1;
1439 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1440 for (i = 0; i < block->jump_in_cnt; i++)
1441 hash_table_remove(block->jump_in[i].vaddr);
1442}
1443
1444static int invalidate_range(u_int start, u_int end,
1445 u32 *inv_start_ret, u32 *inv_end_ret)
1446{
1447 struct block_info *last_block = NULL;
1448 u_int start_page = get_page_prev(start);
1449 u_int end_page = get_page(end - 1);
1450 u_int start_m = pmmask(start);
1451 u_int end_m = pmmask(end - 1);
1452 u_int inv_start, inv_end;
1453 u_int blk_start_m, blk_end_m;
1454 u_int page;
1455 int hit = 0;
1456
1457 // additional area without code (to supplement invalid_code[]), [start, end)
1458 // avoids excessive ndrc_write_invalidate*() calls
1459 inv_start = start_m & ~0xfff;
1460 inv_end = end_m | 0xfff;
1461
1462 for (page = start_page; page <= end_page; page++) {
1463 struct block_info *block;
1464 for (block = blocks[page]; block != NULL; block = block->next) {
1465 if (block->is_dirty)
1466 continue;
1467 last_block = block;
1468 blk_end_m = pmmask(block->start + block->len);
1469 if (blk_end_m <= start_m) {
1470 inv_start = max(inv_start, blk_end_m);
1471 continue;
1472 }
1473 blk_start_m = pmmask(block->start);
1474 if (end_m <= blk_start_m) {
1475 inv_end = min(inv_end, blk_start_m - 1);
1476 continue;
1477 }
1478 if (!block->source) // "hack" block - leave it alone
1479 continue;
1480
1481 hit++;
1482 invalidate_block(block);
1483 stat_inc(stat_inv_hits);
1484 }
1485 }
1486
1487 if (!hit && last_block && last_block->source) {
1488 // could be some leftover unused block, uselessly trapping writes
1489 last_block->inv_near_misses++;
1490 if (last_block->inv_near_misses > 128) {
1491 invalidate_block(last_block);
1492 stat_inc(stat_inv_hits);
1493 hit++;
1494 }
1495 }
1496 if (hit) {
1497 do_clear_cache();
1498#ifdef USE_MINI_HT
1499 memset(mini_ht, -1, sizeof(mini_ht));
1500#endif
1501 }
1502
1503 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1504 // the whole page is empty now
1505 mark_invalid_code(start, 1, 1);
1506
1507 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1508 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1509 return hit;
1510}
1511
1512void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1513{
1514 invalidate_range(start, end, NULL, NULL);
1515}
1516
1517static void ndrc_write_invalidate_many(u_int start, u_int end)
1518{
1519 // this check is done by the caller
1520 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1521 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1522#ifdef INV_DEBUG_W
1523 int invc = invalid_code[start >> 12];
1524 u_int len = end - start;
1525 if (ret)
1526 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1527 else
1528 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1529 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1530 check_for_block_changes(start, end);
1531#endif
1532 stat_inc(stat_inv_addr_calls);
1533 (void)ret;
1534}
1535
1536void ndrc_write_invalidate_one(u_int addr)
1537{
1538 ndrc_write_invalidate_many(addr, addr + 4);
1539}
1540
1541// This is called when loading a save state.
1542// Anything could have changed, so invalidate everything.
1543void new_dynarec_invalidate_all_pages(void)
1544{
1545 struct block_info *block;
1546 u_int page;
1547 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1548 for (block = blocks[page]; block != NULL; block = block->next) {
1549 if (block->is_dirty)
1550 continue;
1551 if (!block->source) // hack block?
1552 continue;
1553 invalidate_block(block);
1554 }
1555 }
1556
1557 #ifdef USE_MINI_HT
1558 memset(mini_ht, -1, sizeof(mini_ht));
1559 #endif
1560 do_clear_cache();
1561}
1562
1563// Add an entry to jump_out after making a link
1564// src should point to code by emit_extjump()
1565void ndrc_add_jump_out(u_int vaddr, void *src)
1566{
1567 inv_debug("ndrc_add_jump_out: %p -> %x\n", src, vaddr);
1568 u_int page = get_page(vaddr);
1569 struct jump_info *ji;
1570
1571 stat_inc(stat_links);
1572 check_extjump2(src);
1573 ji = jumps[page];
1574 if (ji == NULL) {
1575 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1576 ji->alloc = 16;
1577 ji->count = 0;
1578 }
1579 else if (ji->count >= ji->alloc) {
1580 ji->alloc += 16;
1581 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1582 }
1583 jumps[page] = ji;
1584 ji->e[ji->count].target_vaddr = vaddr;
1585 ji->e[ji->count].stub = src;
1586 ji->count++;
1587}
1588
1589/* Register allocation */
1590
1591// Note: registers are allocated clean (unmodified state)
1592// if you intend to modify the register, you must call dirty_reg().
1593static void alloc_reg(struct regstat *cur,int i,signed char reg)
1594{
1595 int r,hr;
1596 int preferred_reg = PREFERRED_REG_FIRST
1597 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1598 if (reg == CCREG) preferred_reg = HOST_CCREG;
1599 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1600 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1601 assert(reg >= 0);
1602
1603 // Don't allocate unused registers
1604 if((cur->u>>reg)&1) return;
1605
1606 // see if it's already allocated
1607 if (get_reg(cur->regmap, reg) >= 0)
1608 return;
1609
1610 // Keep the same mapping if the register was already allocated in a loop
1611 preferred_reg = loop_reg(i,reg,preferred_reg);
1612
1613 // Try to allocate the preferred register
1614 if(cur->regmap[preferred_reg]==-1) {
1615 cur->regmap[preferred_reg]=reg;
1616 cur->dirty&=~(1<<preferred_reg);
1617 cur->isconst&=~(1<<preferred_reg);
1618 return;
1619 }
1620 r=cur->regmap[preferred_reg];
1621 assert(r < 64);
1622 if((cur->u>>r)&1) {
1623 cur->regmap[preferred_reg]=reg;
1624 cur->dirty&=~(1<<preferred_reg);
1625 cur->isconst&=~(1<<preferred_reg);
1626 return;
1627 }
1628
1629 // Clear any unneeded registers
1630 // We try to keep the mapping consistent, if possible, because it
1631 // makes branches easier (especially loops). So we try to allocate
1632 // first (see above) before removing old mappings. If this is not
1633 // possible then go ahead and clear out the registers that are no
1634 // longer needed.
1635 for(hr=0;hr<HOST_REGS;hr++)
1636 {
1637 r=cur->regmap[hr];
1638 if(r>=0) {
1639 assert(r < 64);
1640 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1641 }
1642 }
1643
1644 // Try to allocate any available register, but prefer
1645 // registers that have not been used recently.
1646 if (i > 0) {
1647 for (hr = PREFERRED_REG_FIRST; ; ) {
1648 if (cur->regmap[hr] < 0) {
1649 int oldreg = regs[i-1].regmap[hr];
1650 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1651 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1652 {
1653 cur->regmap[hr]=reg;
1654 cur->dirty&=~(1<<hr);
1655 cur->isconst&=~(1<<hr);
1656 return;
1657 }
1658 }
1659 hr++;
1660 if (hr == EXCLUDE_REG)
1661 hr++;
1662 if (hr == HOST_REGS)
1663 hr = 0;
1664 if (hr == PREFERRED_REG_FIRST)
1665 break;
1666 }
1667 }
1668
1669 // Try to allocate any available register
1670 for (hr = PREFERRED_REG_FIRST; ; ) {
1671 if (cur->regmap[hr] < 0) {
1672 cur->regmap[hr]=reg;
1673 cur->dirty&=~(1<<hr);
1674 cur->isconst&=~(1<<hr);
1675 return;
1676 }
1677 hr++;
1678 if (hr == EXCLUDE_REG)
1679 hr++;
1680 if (hr == HOST_REGS)
1681 hr = 0;
1682 if (hr == PREFERRED_REG_FIRST)
1683 break;
1684 }
1685
1686 // Ok, now we have to evict someone
1687 // Pick a register we hopefully won't need soon
1688 u_char hsn[MAXREG+1];
1689 memset(hsn,10,sizeof(hsn));
1690 int j;
1691 lsn(hsn,i,&preferred_reg);
1692 //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]);
1693 //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]);
1694 if(i>0) {
1695 // Don't evict the cycle count at entry points, otherwise the entry
1696 // stub will have to write it.
1697 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1698 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1699 for(j=10;j>=3;j--)
1700 {
1701 // Alloc preferred register if available
1702 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1703 for(hr=0;hr<HOST_REGS;hr++) {
1704 // Evict both parts of a 64-bit register
1705 if(cur->regmap[hr]==r) {
1706 cur->regmap[hr]=-1;
1707 cur->dirty&=~(1<<hr);
1708 cur->isconst&=~(1<<hr);
1709 }
1710 }
1711 cur->regmap[preferred_reg]=reg;
1712 return;
1713 }
1714 for(r=1;r<=MAXREG;r++)
1715 {
1716 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1717 for(hr=0;hr<HOST_REGS;hr++) {
1718 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1719 if(cur->regmap[hr]==r) {
1720 cur->regmap[hr]=reg;
1721 cur->dirty&=~(1<<hr);
1722 cur->isconst&=~(1<<hr);
1723 return;
1724 }
1725 }
1726 }
1727 }
1728 }
1729 }
1730 }
1731 for(j=10;j>=0;j--)
1732 {
1733 for(r=1;r<=MAXREG;r++)
1734 {
1735 if(hsn[r]==j) {
1736 for(hr=0;hr<HOST_REGS;hr++) {
1737 if(cur->regmap[hr]==r) {
1738 cur->regmap[hr]=reg;
1739 cur->dirty&=~(1<<hr);
1740 cur->isconst&=~(1<<hr);
1741 return;
1742 }
1743 }
1744 }
1745 }
1746 }
1747 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1748}
1749
1750// Allocate a temporary register. This is done without regard to
1751// dirty status or whether the register we request is on the unneeded list
1752// Note: This will only allocate one register, even if called multiple times
1753static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1754{
1755 int r,hr;
1756 int preferred_reg = -1;
1757
1758 // see if it's already allocated
1759 for(hr=0;hr<HOST_REGS;hr++)
1760 {
1761 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1762 }
1763
1764 // Try to allocate any available register
1765 for(hr=HOST_REGS-1;hr>=0;hr--) {
1766 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1767 cur->regmap[hr]=reg;
1768 cur->dirty&=~(1<<hr);
1769 cur->isconst&=~(1<<hr);
1770 return;
1771 }
1772 }
1773
1774 // Find an unneeded register
1775 for(hr=HOST_REGS-1;hr>=0;hr--)
1776 {
1777 r=cur->regmap[hr];
1778 if(r>=0) {
1779 assert(r < 64);
1780 if((cur->u>>r)&1) {
1781 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1782 cur->regmap[hr]=reg;
1783 cur->dirty&=~(1<<hr);
1784 cur->isconst&=~(1<<hr);
1785 return;
1786 }
1787 }
1788 }
1789 }
1790
1791 // Ok, now we have to evict someone
1792 // Pick a register we hopefully won't need soon
1793 // TODO: we might want to follow unconditional jumps here
1794 // TODO: get rid of dupe code and make this into a function
1795 u_char hsn[MAXREG+1];
1796 memset(hsn,10,sizeof(hsn));
1797 int j;
1798 lsn(hsn,i,&preferred_reg);
1799 //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]);
1800 if(i>0) {
1801 // Don't evict the cycle count at entry points, otherwise the entry
1802 // stub will have to write it.
1803 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1804 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1805 for(j=10;j>=3;j--)
1806 {
1807 for(r=1;r<=MAXREG;r++)
1808 {
1809 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1810 for(hr=0;hr<HOST_REGS;hr++) {
1811 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1812 if(cur->regmap[hr]==r) {
1813 cur->regmap[hr]=reg;
1814 cur->dirty&=~(1<<hr);
1815 cur->isconst&=~(1<<hr);
1816 return;
1817 }
1818 }
1819 }
1820 }
1821 }
1822 }
1823 }
1824 for(j=10;j>=0;j--)
1825 {
1826 for(r=1;r<=MAXREG;r++)
1827 {
1828 if(hsn[r]==j) {
1829 for(hr=0;hr<HOST_REGS;hr++) {
1830 if(cur->regmap[hr]==r) {
1831 cur->regmap[hr]=reg;
1832 cur->dirty&=~(1<<hr);
1833 cur->isconst&=~(1<<hr);
1834 return;
1835 }
1836 }
1837 }
1838 }
1839 }
1840 SysPrintf("This shouldn't happen");abort();
1841}
1842
1843static void mov_alloc(struct regstat *current,int i)
1844{
1845 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1846 alloc_cc(current,i); // for stalls
1847 dirty_reg(current,CCREG);
1848 }
1849
1850 // Note: Don't need to actually alloc the source registers
1851 //alloc_reg(current,i,dops[i].rs1);
1852 alloc_reg(current,i,dops[i].rt1);
1853
1854 clear_const(current,dops[i].rs1);
1855 clear_const(current,dops[i].rt1);
1856 dirty_reg(current,dops[i].rt1);
1857}
1858
1859static void shiftimm_alloc(struct regstat *current,int i)
1860{
1861 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1862 {
1863 if(dops[i].rt1) {
1864 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1865 else dops[i].use_lt1=!!dops[i].rs1;
1866 alloc_reg(current,i,dops[i].rt1);
1867 dirty_reg(current,dops[i].rt1);
1868 if(is_const(current,dops[i].rs1)) {
1869 int v=get_const(current,dops[i].rs1);
1870 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1871 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1872 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1873 }
1874 else clear_const(current,dops[i].rt1);
1875 }
1876 }
1877 else
1878 {
1879 clear_const(current,dops[i].rs1);
1880 clear_const(current,dops[i].rt1);
1881 }
1882
1883 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1884 {
1885 assert(0);
1886 }
1887 if(dops[i].opcode2==0x3c) // DSLL32
1888 {
1889 assert(0);
1890 }
1891 if(dops[i].opcode2==0x3e) // DSRL32
1892 {
1893 assert(0);
1894 }
1895 if(dops[i].opcode2==0x3f) // DSRA32
1896 {
1897 assert(0);
1898 }
1899}
1900
1901static void shift_alloc(struct regstat *current,int i)
1902{
1903 if(dops[i].rt1) {
1904 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1905 {
1906 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1907 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1908 alloc_reg(current,i,dops[i].rt1);
1909 if(dops[i].rt1==dops[i].rs2) {
1910 alloc_reg_temp(current,i,-1);
1911 minimum_free_regs[i]=1;
1912 }
1913 } else { // DSLLV/DSRLV/DSRAV
1914 assert(0);
1915 }
1916 clear_const(current,dops[i].rs1);
1917 clear_const(current,dops[i].rs2);
1918 clear_const(current,dops[i].rt1);
1919 dirty_reg(current,dops[i].rt1);
1920 }
1921}
1922
1923static void alu_alloc(struct regstat *current,int i)
1924{
1925 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1926 if(dops[i].rt1) {
1927 if(dops[i].rs1&&dops[i].rs2) {
1928 alloc_reg(current,i,dops[i].rs1);
1929 alloc_reg(current,i,dops[i].rs2);
1930 }
1931 else {
1932 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1933 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1934 }
1935 alloc_reg(current,i,dops[i].rt1);
1936 }
1937 }
1938 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1939 if(dops[i].rt1) {
1940 alloc_reg(current,i,dops[i].rs1);
1941 alloc_reg(current,i,dops[i].rs2);
1942 alloc_reg(current,i,dops[i].rt1);
1943 }
1944 }
1945 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1946 if(dops[i].rt1) {
1947 if(dops[i].rs1&&dops[i].rs2) {
1948 alloc_reg(current,i,dops[i].rs1);
1949 alloc_reg(current,i,dops[i].rs2);
1950 }
1951 else
1952 {
1953 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1954 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1955 }
1956 alloc_reg(current,i,dops[i].rt1);
1957 }
1958 }
1959 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1960 assert(0);
1961 }
1962 clear_const(current,dops[i].rs1);
1963 clear_const(current,dops[i].rs2);
1964 clear_const(current,dops[i].rt1);
1965 dirty_reg(current,dops[i].rt1);
1966}
1967
1968static void imm16_alloc(struct regstat *current,int i)
1969{
1970 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1971 else dops[i].use_lt1=!!dops[i].rs1;
1972 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1973 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1974 assert(0);
1975 }
1976 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1977 clear_const(current,dops[i].rs1);
1978 clear_const(current,dops[i].rt1);
1979 }
1980 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1981 if(is_const(current,dops[i].rs1)) {
1982 int v=get_const(current,dops[i].rs1);
1983 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1984 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1985 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1986 }
1987 else clear_const(current,dops[i].rt1);
1988 }
1989 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1990 if(is_const(current,dops[i].rs1)) {
1991 int v=get_const(current,dops[i].rs1);
1992 set_const(current,dops[i].rt1,v+imm[i]);
1993 }
1994 else clear_const(current,dops[i].rt1);
1995 }
1996 else {
1997 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1998 }
1999 dirty_reg(current,dops[i].rt1);
2000}
2001
2002static void load_alloc(struct regstat *current,int i)
2003{
2004 clear_const(current,dops[i].rt1);
2005 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2006 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2007 if (needed_again(dops[i].rs1, i))
2008 alloc_reg(current, i, dops[i].rs1);
2009 if (ram_offset)
2010 alloc_reg(current, i, ROREG);
2011 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2012 alloc_reg(current,i,dops[i].rt1);
2013 assert(get_reg(current->regmap,dops[i].rt1)>=0);
2014 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
2015 {
2016 assert(0);
2017 }
2018 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
2019 {
2020 assert(0);
2021 }
2022 dirty_reg(current,dops[i].rt1);
2023 // LWL/LWR need a temporary register for the old value
2024 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2025 {
2026 alloc_reg(current,i,FTEMP);
2027 alloc_reg_temp(current,i,-1);
2028 minimum_free_regs[i]=1;
2029 }
2030 }
2031 else
2032 {
2033 // Load to r0 or unneeded register (dummy load)
2034 // but we still need a register to calculate the address
2035 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2036 {
2037 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2038 }
2039 alloc_reg_temp(current,i,-1);
2040 minimum_free_regs[i]=1;
2041 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
2042 {
2043 assert(0);
2044 }
2045 }
2046}
2047
2048static void store_alloc(struct regstat *current,int i)
2049{
2050 clear_const(current,dops[i].rs2);
2051 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2052 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2053 alloc_reg(current,i,dops[i].rs2);
2054 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
2055 assert(0);
2056 }
2057 if (ram_offset)
2058 alloc_reg(current, i, ROREG);
2059 #if defined(HOST_IMM8)
2060 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2061 alloc_reg(current, i, INVCP);
2062 #endif
2063 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
2064 alloc_reg(current,i,FTEMP);
2065 }
2066 // We need a temporary register for address generation
2067 alloc_reg_temp(current,i,-1);
2068 minimum_free_regs[i]=1;
2069}
2070
2071static void c1ls_alloc(struct regstat *current,int i)
2072{
2073 clear_const(current,dops[i].rt1);
2074 alloc_reg(current,i,CSREG); // Status
2075}
2076
2077static void c2ls_alloc(struct regstat *current,int i)
2078{
2079 clear_const(current,dops[i].rt1);
2080 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2081 alloc_reg(current,i,FTEMP);
2082 if (ram_offset)
2083 alloc_reg(current, i, ROREG);
2084 #if defined(HOST_IMM8)
2085 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2086 if (dops[i].opcode == 0x3a) // SWC2
2087 alloc_reg(current,i,INVCP);
2088 #endif
2089 // We need a temporary register for address generation
2090 alloc_reg_temp(current,i,-1);
2091 minimum_free_regs[i]=1;
2092}
2093
2094#ifndef multdiv_alloc
2095static void multdiv_alloc(struct regstat *current,int i)
2096{
2097 // case 0x18: MULT
2098 // case 0x19: MULTU
2099 // case 0x1A: DIV
2100 // case 0x1B: DIVU
2101 // case 0x1C: DMULT
2102 // case 0x1D: DMULTU
2103 // case 0x1E: DDIV
2104 // case 0x1F: DDIVU
2105 clear_const(current,dops[i].rs1);
2106 clear_const(current,dops[i].rs2);
2107 alloc_cc(current,i); // for stalls
2108 if(dops[i].rs1&&dops[i].rs2)
2109 {
2110 if((dops[i].opcode2&4)==0) // 32-bit
2111 {
2112 current->u&=~(1LL<<HIREG);
2113 current->u&=~(1LL<<LOREG);
2114 alloc_reg(current,i,HIREG);
2115 alloc_reg(current,i,LOREG);
2116 alloc_reg(current,i,dops[i].rs1);
2117 alloc_reg(current,i,dops[i].rs2);
2118 dirty_reg(current,HIREG);
2119 dirty_reg(current,LOREG);
2120 }
2121 else // 64-bit
2122 {
2123 assert(0);
2124 }
2125 }
2126 else
2127 {
2128 // Multiply by zero is zero.
2129 // MIPS does not have a divide by zero exception.
2130 // The result is undefined, we return zero.
2131 alloc_reg(current,i,HIREG);
2132 alloc_reg(current,i,LOREG);
2133 dirty_reg(current,HIREG);
2134 dirty_reg(current,LOREG);
2135 }
2136}
2137#endif
2138
2139static void cop0_alloc(struct regstat *current,int i)
2140{
2141 if(dops[i].opcode2==0) // MFC0
2142 {
2143 if(dops[i].rt1) {
2144 clear_const(current,dops[i].rt1);
2145 alloc_all(current,i);
2146 alloc_reg(current,i,dops[i].rt1);
2147 dirty_reg(current,dops[i].rt1);
2148 }
2149 }
2150 else if(dops[i].opcode2==4) // MTC0
2151 {
2152 if(dops[i].rs1){
2153 clear_const(current,dops[i].rs1);
2154 alloc_reg(current,i,dops[i].rs1);
2155 alloc_all(current,i);
2156 }
2157 else {
2158 alloc_all(current,i); // FIXME: Keep r0
2159 current->u&=~1LL;
2160 alloc_reg(current,i,0);
2161 }
2162 }
2163 else
2164 {
2165 // RFE
2166 assert(dops[i].opcode2==0x10);
2167 alloc_all(current,i);
2168 }
2169 minimum_free_regs[i]=HOST_REGS;
2170}
2171
2172static void cop2_alloc(struct regstat *current,int i)
2173{
2174 if (dops[i].opcode2 < 3) // MFC2/CFC2
2175 {
2176 alloc_cc(current,i); // for stalls
2177 dirty_reg(current,CCREG);
2178 if(dops[i].rt1){
2179 clear_const(current,dops[i].rt1);
2180 alloc_reg(current,i,dops[i].rt1);
2181 dirty_reg(current,dops[i].rt1);
2182 }
2183 }
2184 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2185 {
2186 if(dops[i].rs1){
2187 clear_const(current,dops[i].rs1);
2188 alloc_reg(current,i,dops[i].rs1);
2189 }
2190 else {
2191 current->u&=~1LL;
2192 alloc_reg(current,i,0);
2193 }
2194 }
2195 alloc_reg_temp(current,i,-1);
2196 minimum_free_regs[i]=1;
2197}
2198
2199static void c2op_alloc(struct regstat *current,int i)
2200{
2201 alloc_cc(current,i); // for stalls
2202 dirty_reg(current,CCREG);
2203 alloc_reg_temp(current,i,-1);
2204}
2205
2206static void syscall_alloc(struct regstat *current,int i)
2207{
2208 alloc_cc(current,i);
2209 dirty_reg(current,CCREG);
2210 alloc_all(current,i);
2211 minimum_free_regs[i]=HOST_REGS;
2212 current->isconst=0;
2213}
2214
2215static void delayslot_alloc(struct regstat *current,int i)
2216{
2217 switch(dops[i].itype) {
2218 case UJUMP:
2219 case CJUMP:
2220 case SJUMP:
2221 case RJUMP:
2222 case SYSCALL:
2223 case HLECALL:
2224 case IMM16:
2225 imm16_alloc(current,i);
2226 break;
2227 case LOAD:
2228 case LOADLR:
2229 load_alloc(current,i);
2230 break;
2231 case STORE:
2232 case STORELR:
2233 store_alloc(current,i);
2234 break;
2235 case ALU:
2236 alu_alloc(current,i);
2237 break;
2238 case SHIFT:
2239 shift_alloc(current,i);
2240 break;
2241 case MULTDIV:
2242 multdiv_alloc(current,i);
2243 break;
2244 case SHIFTIMM:
2245 shiftimm_alloc(current,i);
2246 break;
2247 case MOV:
2248 mov_alloc(current,i);
2249 break;
2250 case COP0:
2251 cop0_alloc(current,i);
2252 break;
2253 case COP1:
2254 break;
2255 case COP2:
2256 cop2_alloc(current,i);
2257 break;
2258 case C1LS:
2259 c1ls_alloc(current,i);
2260 break;
2261 case C2LS:
2262 c2ls_alloc(current,i);
2263 break;
2264 case C2OP:
2265 c2op_alloc(current,i);
2266 break;
2267 }
2268}
2269
2270static void add_stub(enum stub_type type, void *addr, void *retaddr,
2271 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2272{
2273 assert(stubcount < ARRAY_SIZE(stubs));
2274 stubs[stubcount].type = type;
2275 stubs[stubcount].addr = addr;
2276 stubs[stubcount].retaddr = retaddr;
2277 stubs[stubcount].a = a;
2278 stubs[stubcount].b = b;
2279 stubs[stubcount].c = c;
2280 stubs[stubcount].d = d;
2281 stubs[stubcount].e = e;
2282 stubcount++;
2283}
2284
2285static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2286 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2287{
2288 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2289}
2290
2291// Write out a single register
2292static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2293{
2294 int hr;
2295 for(hr=0;hr<HOST_REGS;hr++) {
2296 if(hr!=EXCLUDE_REG) {
2297 if(regmap[hr]==r) {
2298 if((dirty>>hr)&1) {
2299 assert(regmap[hr]<64);
2300 emit_storereg(r,hr);
2301 }
2302 }
2303 }
2304 }
2305}
2306
2307static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2308{
2309 //if(dirty_pre==dirty) return;
2310 int hr, r;
2311 for (hr = 0; hr < HOST_REGS; hr++) {
2312 r = pre[hr];
2313 if (r < 1 || r > 33 || ((u >> r) & 1))
2314 continue;
2315 if (((dirty_pre & ~dirty) >> hr) & 1)
2316 emit_storereg(r, hr);
2317 }
2318}
2319
2320// trashes r2
2321static void pass_args(int a0, int a1)
2322{
2323 if(a0==1&&a1==0) {
2324 // must swap
2325 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2326 }
2327 else if(a0!=0&&a1==0) {
2328 emit_mov(a1,1);
2329 if (a0>=0) emit_mov(a0,0);
2330 }
2331 else {
2332 if(a0>=0&&a0!=0) emit_mov(a0,0);
2333 if(a1>=0&&a1!=1) emit_mov(a1,1);
2334 }
2335}
2336
2337static void alu_assemble(int i, const struct regstat *i_regs)
2338{
2339 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2340 if(dops[i].rt1) {
2341 signed char s1,s2,t;
2342 t=get_reg(i_regs->regmap,dops[i].rt1);
2343 if(t>=0) {
2344 s1=get_reg(i_regs->regmap,dops[i].rs1);
2345 s2=get_reg(i_regs->regmap,dops[i].rs2);
2346 if(dops[i].rs1&&dops[i].rs2) {
2347 assert(s1>=0);
2348 assert(s2>=0);
2349 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2350 else emit_add(s1,s2,t);
2351 }
2352 else if(dops[i].rs1) {
2353 if(s1>=0) emit_mov(s1,t);
2354 else emit_loadreg(dops[i].rs1,t);
2355 }
2356 else if(dops[i].rs2) {
2357 if(s2>=0) {
2358 if(dops[i].opcode2&2) emit_neg(s2,t);
2359 else emit_mov(s2,t);
2360 }
2361 else {
2362 emit_loadreg(dops[i].rs2,t);
2363 if(dops[i].opcode2&2) emit_neg(t,t);
2364 }
2365 }
2366 else emit_zeroreg(t);
2367 }
2368 }
2369 }
2370 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2371 assert(0);
2372 }
2373 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2374 if(dops[i].rt1) {
2375 signed char s1l,s2l,t;
2376 {
2377 t=get_reg(i_regs->regmap,dops[i].rt1);
2378 //assert(t>=0);
2379 if(t>=0) {
2380 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2381 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2382 if(dops[i].rs2==0) // rx<r0
2383 {
2384 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2385 assert(s1l>=0);
2386 emit_shrimm(s1l,31,t);
2387 }
2388 else // SLTU (unsigned can not be less than zero, 0<0)
2389 emit_zeroreg(t);
2390 }
2391 else if(dops[i].rs1==0) // r0<rx
2392 {
2393 assert(s2l>=0);
2394 if(dops[i].opcode2==0x2a) // SLT
2395 emit_set_gz32(s2l,t);
2396 else // SLTU (set if not zero)
2397 emit_set_nz32(s2l,t);
2398 }
2399 else{
2400 assert(s1l>=0);assert(s2l>=0);
2401 if(dops[i].opcode2==0x2a) // SLT
2402 emit_set_if_less32(s1l,s2l,t);
2403 else // SLTU
2404 emit_set_if_carry32(s1l,s2l,t);
2405 }
2406 }
2407 }
2408 }
2409 }
2410 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2411 if(dops[i].rt1) {
2412 signed char s1l,s2l,tl;
2413 tl=get_reg(i_regs->regmap,dops[i].rt1);
2414 {
2415 if(tl>=0) {
2416 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2417 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2418 if(dops[i].rs1&&dops[i].rs2) {
2419 assert(s1l>=0);
2420 assert(s2l>=0);
2421 if(dops[i].opcode2==0x24) { // AND
2422 emit_and(s1l,s2l,tl);
2423 } else
2424 if(dops[i].opcode2==0x25) { // OR
2425 emit_or(s1l,s2l,tl);
2426 } else
2427 if(dops[i].opcode2==0x26) { // XOR
2428 emit_xor(s1l,s2l,tl);
2429 } else
2430 if(dops[i].opcode2==0x27) { // NOR
2431 emit_or(s1l,s2l,tl);
2432 emit_not(tl,tl);
2433 }
2434 }
2435 else
2436 {
2437 if(dops[i].opcode2==0x24) { // AND
2438 emit_zeroreg(tl);
2439 } else
2440 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2441 if(dops[i].rs1){
2442 if(s1l>=0) emit_mov(s1l,tl);
2443 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2444 }
2445 else
2446 if(dops[i].rs2){
2447 if(s2l>=0) emit_mov(s2l,tl);
2448 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2449 }
2450 else emit_zeroreg(tl);
2451 } else
2452 if(dops[i].opcode2==0x27) { // NOR
2453 if(dops[i].rs1){
2454 if(s1l>=0) emit_not(s1l,tl);
2455 else {
2456 emit_loadreg(dops[i].rs1,tl);
2457 emit_not(tl,tl);
2458 }
2459 }
2460 else
2461 if(dops[i].rs2){
2462 if(s2l>=0) emit_not(s2l,tl);
2463 else {
2464 emit_loadreg(dops[i].rs2,tl);
2465 emit_not(tl,tl);
2466 }
2467 }
2468 else emit_movimm(-1,tl);
2469 }
2470 }
2471 }
2472 }
2473 }
2474 }
2475}
2476
2477static void imm16_assemble(int i, const struct regstat *i_regs)
2478{
2479 if (dops[i].opcode==0x0f) { // LUI
2480 if(dops[i].rt1) {
2481 signed char t;
2482 t=get_reg(i_regs->regmap,dops[i].rt1);
2483 //assert(t>=0);
2484 if(t>=0) {
2485 if(!((i_regs->isconst>>t)&1))
2486 emit_movimm(imm[i]<<16,t);
2487 }
2488 }
2489 }
2490 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2491 if(dops[i].rt1) {
2492 signed char s,t;
2493 t=get_reg(i_regs->regmap,dops[i].rt1);
2494 s=get_reg(i_regs->regmap,dops[i].rs1);
2495 if(dops[i].rs1) {
2496 //assert(t>=0);
2497 //assert(s>=0);
2498 if(t>=0) {
2499 if(!((i_regs->isconst>>t)&1)) {
2500 if(s<0) {
2501 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2502 emit_addimm(t,imm[i],t);
2503 }else{
2504 if(!((i_regs->wasconst>>s)&1))
2505 emit_addimm(s,imm[i],t);
2506 else
2507 emit_movimm(constmap[i][s]+imm[i],t);
2508 }
2509 }
2510 }
2511 } else {
2512 if(t>=0) {
2513 if(!((i_regs->isconst>>t)&1))
2514 emit_movimm(imm[i],t);
2515 }
2516 }
2517 }
2518 }
2519 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2520 if(dops[i].rt1) {
2521 signed char sl,tl;
2522 tl=get_reg(i_regs->regmap,dops[i].rt1);
2523 sl=get_reg(i_regs->regmap,dops[i].rs1);
2524 if(tl>=0) {
2525 if(dops[i].rs1) {
2526 assert(sl>=0);
2527 emit_addimm(sl,imm[i],tl);
2528 } else {
2529 emit_movimm(imm[i],tl);
2530 }
2531 }
2532 }
2533 }
2534 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2535 if(dops[i].rt1) {
2536 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2537 signed char sl,t;
2538 t=get_reg(i_regs->regmap,dops[i].rt1);
2539 sl=get_reg(i_regs->regmap,dops[i].rs1);
2540 //assert(t>=0);
2541 if(t>=0) {
2542 if(dops[i].rs1>0) {
2543 if(dops[i].opcode==0x0a) { // SLTI
2544 if(sl<0) {
2545 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2546 emit_slti32(t,imm[i],t);
2547 }else{
2548 emit_slti32(sl,imm[i],t);
2549 }
2550 }
2551 else { // SLTIU
2552 if(sl<0) {
2553 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2554 emit_sltiu32(t,imm[i],t);
2555 }else{
2556 emit_sltiu32(sl,imm[i],t);
2557 }
2558 }
2559 }else{
2560 // SLTI(U) with r0 is just stupid,
2561 // nonetheless examples can be found
2562 if(dops[i].opcode==0x0a) // SLTI
2563 if(0<imm[i]) emit_movimm(1,t);
2564 else emit_zeroreg(t);
2565 else // SLTIU
2566 {
2567 if(imm[i]) emit_movimm(1,t);
2568 else emit_zeroreg(t);
2569 }
2570 }
2571 }
2572 }
2573 }
2574 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2575 if(dops[i].rt1) {
2576 signed char sl,tl;
2577 tl=get_reg(i_regs->regmap,dops[i].rt1);
2578 sl=get_reg(i_regs->regmap,dops[i].rs1);
2579 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2580 if(dops[i].opcode==0x0c) //ANDI
2581 {
2582 if(dops[i].rs1) {
2583 if(sl<0) {
2584 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2585 emit_andimm(tl,imm[i],tl);
2586 }else{
2587 if(!((i_regs->wasconst>>sl)&1))
2588 emit_andimm(sl,imm[i],tl);
2589 else
2590 emit_movimm(constmap[i][sl]&imm[i],tl);
2591 }
2592 }
2593 else
2594 emit_zeroreg(tl);
2595 }
2596 else
2597 {
2598 if(dops[i].rs1) {
2599 if(sl<0) {
2600 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2601 }
2602 if(dops[i].opcode==0x0d) { // ORI
2603 if(sl<0) {
2604 emit_orimm(tl,imm[i],tl);
2605 }else{
2606 if(!((i_regs->wasconst>>sl)&1))
2607 emit_orimm(sl,imm[i],tl);
2608 else
2609 emit_movimm(constmap[i][sl]|imm[i],tl);
2610 }
2611 }
2612 if(dops[i].opcode==0x0e) { // XORI
2613 if(sl<0) {
2614 emit_xorimm(tl,imm[i],tl);
2615 }else{
2616 if(!((i_regs->wasconst>>sl)&1))
2617 emit_xorimm(sl,imm[i],tl);
2618 else
2619 emit_movimm(constmap[i][sl]^imm[i],tl);
2620 }
2621 }
2622 }
2623 else {
2624 emit_movimm(imm[i],tl);
2625 }
2626 }
2627 }
2628 }
2629 }
2630}
2631
2632static void shiftimm_assemble(int i, const struct regstat *i_regs)
2633{
2634 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2635 {
2636 if(dops[i].rt1) {
2637 signed char s,t;
2638 t=get_reg(i_regs->regmap,dops[i].rt1);
2639 s=get_reg(i_regs->regmap,dops[i].rs1);
2640 //assert(t>=0);
2641 if(t>=0&&!((i_regs->isconst>>t)&1)){
2642 if(dops[i].rs1==0)
2643 {
2644 emit_zeroreg(t);
2645 }
2646 else
2647 {
2648 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2649 if(imm[i]) {
2650 if(dops[i].opcode2==0) // SLL
2651 {
2652 emit_shlimm(s<0?t:s,imm[i],t);
2653 }
2654 if(dops[i].opcode2==2) // SRL
2655 {
2656 emit_shrimm(s<0?t:s,imm[i],t);
2657 }
2658 if(dops[i].opcode2==3) // SRA
2659 {
2660 emit_sarimm(s<0?t:s,imm[i],t);
2661 }
2662 }else{
2663 // Shift by zero
2664 if(s>=0 && s!=t) emit_mov(s,t);
2665 }
2666 }
2667 }
2668 //emit_storereg(dops[i].rt1,t); //DEBUG
2669 }
2670 }
2671 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2672 {
2673 assert(0);
2674 }
2675 if(dops[i].opcode2==0x3c) // DSLL32
2676 {
2677 assert(0);
2678 }
2679 if(dops[i].opcode2==0x3e) // DSRL32
2680 {
2681 assert(0);
2682 }
2683 if(dops[i].opcode2==0x3f) // DSRA32
2684 {
2685 assert(0);
2686 }
2687}
2688
2689#ifndef shift_assemble
2690static void shift_assemble(int i, const struct regstat *i_regs)
2691{
2692 signed char s,t,shift;
2693 if (dops[i].rt1 == 0)
2694 return;
2695 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2696 t = get_reg(i_regs->regmap, dops[i].rt1);
2697 s = get_reg(i_regs->regmap, dops[i].rs1);
2698 shift = get_reg(i_regs->regmap, dops[i].rs2);
2699 if (t < 0)
2700 return;
2701
2702 if(dops[i].rs1==0)
2703 emit_zeroreg(t);
2704 else if(dops[i].rs2==0) {
2705 assert(s>=0);
2706 if(s!=t) emit_mov(s,t);
2707 }
2708 else {
2709 host_tempreg_acquire();
2710 emit_andimm(shift,31,HOST_TEMPREG);
2711 switch(dops[i].opcode2) {
2712 case 4: // SLLV
2713 emit_shl(s,HOST_TEMPREG,t);
2714 break;
2715 case 6: // SRLV
2716 emit_shr(s,HOST_TEMPREG,t);
2717 break;
2718 case 7: // SRAV
2719 emit_sar(s,HOST_TEMPREG,t);
2720 break;
2721 default:
2722 assert(0);
2723 }
2724 host_tempreg_release();
2725 }
2726}
2727
2728#endif
2729
2730enum {
2731 MTYPE_8000 = 0,
2732 MTYPE_8020,
2733 MTYPE_0000,
2734 MTYPE_A000,
2735 MTYPE_1F80,
2736};
2737
2738static int get_ptr_mem_type(u_int a)
2739{
2740 if(a < 0x00200000) {
2741 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2742 // return wrong, must use memhandler for BIOS self-test to pass
2743 // 007 does similar stuff from a00 mirror, weird stuff
2744 return MTYPE_8000;
2745 return MTYPE_0000;
2746 }
2747 if(0x1f800000 <= a && a < 0x1f801000)
2748 return MTYPE_1F80;
2749 if(0x80200000 <= a && a < 0x80800000)
2750 return MTYPE_8020;
2751 if(0xa0000000 <= a && a < 0xa0200000)
2752 return MTYPE_A000;
2753 return MTYPE_8000;
2754}
2755
2756static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2757{
2758 int r = get_reg(i_regs->regmap, ROREG);
2759 if (r < 0 && host_tempreg_free) {
2760 host_tempreg_acquire();
2761 emit_loadreg(ROREG, r = HOST_TEMPREG);
2762 }
2763 if (r < 0)
2764 abort();
2765 return r;
2766}
2767
2768static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2769 int addr, int *offset_reg, int *addr_reg_override)
2770{
2771 void *jaddr = NULL;
2772 int type = 0;
2773 int mr = dops[i].rs1;
2774 *offset_reg = -1;
2775 if(((smrv_strong|smrv_weak)>>mr)&1) {
2776 type=get_ptr_mem_type(smrv[mr]);
2777 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2778 }
2779 else {
2780 // use the mirror we are running on
2781 type=get_ptr_mem_type(start);
2782 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2783 }
2784
2785 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2786 host_tempreg_acquire();
2787 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2788 addr=*addr_reg_override=HOST_TEMPREG;
2789 type=0;
2790 }
2791 else if(type==MTYPE_0000) { // RAM 0 mirror
2792 host_tempreg_acquire();
2793 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2794 addr=*addr_reg_override=HOST_TEMPREG;
2795 type=0;
2796 }
2797 else if(type==MTYPE_A000) { // RAM A mirror
2798 host_tempreg_acquire();
2799 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2800 addr=*addr_reg_override=HOST_TEMPREG;
2801 type=0;
2802 }
2803 else if(type==MTYPE_1F80) { // scratchpad
2804 if (psxH == (void *)0x1f800000) {
2805 host_tempreg_acquire();
2806 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2807 emit_cmpimm(HOST_TEMPREG,0x1000);
2808 host_tempreg_release();
2809 jaddr=out;
2810 emit_jc(0);
2811 }
2812 else {
2813 // do the usual RAM check, jump will go to the right handler
2814 type=0;
2815 }
2816 }
2817
2818 if (type == 0) // need ram check
2819 {
2820 emit_cmpimm(addr,RAM_SIZE);
2821 jaddr = out;
2822 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2823 // Hint to branch predictor that the branch is unlikely to be taken
2824 if (dops[i].rs1 >= 28)
2825 emit_jno_unlikely(0);
2826 else
2827 #endif
2828 emit_jno(0);
2829 if (ram_offset != 0)
2830 *offset_reg = get_ro_reg(i_regs, 0);
2831 }
2832
2833 return jaddr;
2834}
2835
2836// return memhandler, or get directly accessable address and return 0
2837static void *get_direct_memhandler(void *table, u_int addr,
2838 enum stub_type type, uintptr_t *addr_host)
2839{
2840 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2841 uintptr_t l1, l2 = 0;
2842 l1 = ((uintptr_t *)table)[addr>>12];
2843 if (!(l1 & msb)) {
2844 uintptr_t v = l1 << 1;
2845 *addr_host = v + addr;
2846 return NULL;
2847 }
2848 else {
2849 l1 <<= 1;
2850 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2851 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2852 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2853 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2854 else
2855 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2856 if (!(l2 & msb)) {
2857 uintptr_t v = l2 << 1;
2858 *addr_host = v + (addr&0xfff);
2859 return NULL;
2860 }
2861 return (void *)(l2 << 1);
2862 }
2863}
2864
2865static u_int get_host_reglist(const signed char *regmap)
2866{
2867 u_int reglist = 0, hr;
2868 for (hr = 0; hr < HOST_REGS; hr++) {
2869 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2870 reglist |= 1 << hr;
2871 }
2872 return reglist;
2873}
2874
2875static u_int reglist_exclude(u_int reglist, int r1, int r2)
2876{
2877 if (r1 >= 0)
2878 reglist &= ~(1u << r1);
2879 if (r2 >= 0)
2880 reglist &= ~(1u << r2);
2881 return reglist;
2882}
2883
2884// find a temp caller-saved register not in reglist (so assumed to be free)
2885static int reglist_find_free(u_int reglist)
2886{
2887 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2888 if (free_regs == 0)
2889 return -1;
2890 return __builtin_ctz(free_regs);
2891}
2892
2893static void do_load_word(int a, int rt, int offset_reg)
2894{
2895 if (offset_reg >= 0)
2896 emit_ldr_dualindexed(offset_reg, a, rt);
2897 else
2898 emit_readword_indexed(0, a, rt);
2899}
2900
2901static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2902{
2903 if (offset_reg < 0) {
2904 emit_writeword_indexed(rt, ofs, a);
2905 return;
2906 }
2907 if (ofs != 0)
2908 emit_addimm(a, ofs, a);
2909 emit_str_dualindexed(offset_reg, a, rt);
2910 if (ofs != 0 && preseve_a)
2911 emit_addimm(a, -ofs, a);
2912}
2913
2914static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2915{
2916 if (offset_reg < 0) {
2917 emit_writehword_indexed(rt, ofs, a);
2918 return;
2919 }
2920 if (ofs != 0)
2921 emit_addimm(a, ofs, a);
2922 emit_strh_dualindexed(offset_reg, a, rt);
2923 if (ofs != 0 && preseve_a)
2924 emit_addimm(a, -ofs, a);
2925}
2926
2927static void do_store_byte(int a, int rt, int offset_reg)
2928{
2929 if (offset_reg >= 0)
2930 emit_strb_dualindexed(offset_reg, a, rt);
2931 else
2932 emit_writebyte_indexed(rt, 0, a);
2933}
2934
2935static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2936{
2937 int s,tl,addr;
2938 int offset;
2939 void *jaddr=0;
2940 int memtarget=0,c=0;
2941 int offset_reg = -1;
2942 int fastio_reg_override = -1;
2943 u_int reglist=get_host_reglist(i_regs->regmap);
2944 tl=get_reg(i_regs->regmap,dops[i].rt1);
2945 s=get_reg(i_regs->regmap,dops[i].rs1);
2946 offset=imm[i];
2947 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2948 if(s>=0) {
2949 c=(i_regs->wasconst>>s)&1;
2950 if (c) {
2951 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2952 }
2953 }
2954 //printf("load_assemble: c=%d\n",c);
2955 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2956 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2957 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2958 ||dops[i].rt1==0) {
2959 // could be FIFO, must perform the read
2960 // ||dummy read
2961 assem_debug("(forced read)\n");
2962 tl=get_reg_temp(i_regs->regmap);
2963 assert(tl>=0);
2964 }
2965 if(offset||s<0||c) addr=tl;
2966 else addr=s;
2967 //if(tl<0) tl=get_reg_temp(i_regs->regmap);
2968 if(tl>=0) {
2969 //printf("load_assemble: c=%d\n",c);
2970 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2971 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2972 reglist&=~(1<<tl);
2973 if(!c) {
2974 #ifdef R29_HACK
2975 // Strmnnrmn's speed hack
2976 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2977 #endif
2978 {
2979 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
2980 &offset_reg, &fastio_reg_override);
2981 }
2982 }
2983 else if (ram_offset && memtarget) {
2984 offset_reg = get_ro_reg(i_regs, 0);
2985 }
2986 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2987 switch (dops[i].opcode) {
2988 case 0x20: // LB
2989 if(!c||memtarget) {
2990 if(!dummy) {
2991 int a = tl;
2992 if (!c) a = addr;
2993 if (fastio_reg_override >= 0)
2994 a = fastio_reg_override;
2995
2996 if (offset_reg >= 0)
2997 emit_ldrsb_dualindexed(offset_reg, a, tl);
2998 else
2999 emit_movsbl_indexed(0, a, tl);
3000 }
3001 if(jaddr)
3002 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3003 }
3004 else
3005 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3006 break;
3007 case 0x21: // LH
3008 if(!c||memtarget) {
3009 if(!dummy) {
3010 int a = tl;
3011 if (!c) a = addr;
3012 if (fastio_reg_override >= 0)
3013 a = fastio_reg_override;
3014 if (offset_reg >= 0)
3015 emit_ldrsh_dualindexed(offset_reg, a, tl);
3016 else
3017 emit_movswl_indexed(0, a, tl);
3018 }
3019 if(jaddr)
3020 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3021 }
3022 else
3023 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3024 break;
3025 case 0x23: // LW
3026 if(!c||memtarget) {
3027 if(!dummy) {
3028 int a = addr;
3029 if (fastio_reg_override >= 0)
3030 a = fastio_reg_override;
3031 do_load_word(a, tl, offset_reg);
3032 }
3033 if(jaddr)
3034 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3035 }
3036 else
3037 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3038 break;
3039 case 0x24: // LBU
3040 if(!c||memtarget) {
3041 if(!dummy) {
3042 int a = tl;
3043 if (!c) a = addr;
3044 if (fastio_reg_override >= 0)
3045 a = fastio_reg_override;
3046
3047 if (offset_reg >= 0)
3048 emit_ldrb_dualindexed(offset_reg, a, tl);
3049 else
3050 emit_movzbl_indexed(0, a, tl);
3051 }
3052 if(jaddr)
3053 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3054 }
3055 else
3056 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3057 break;
3058 case 0x25: // LHU
3059 if(!c||memtarget) {
3060 if(!dummy) {
3061 int a = tl;
3062 if(!c) a = addr;
3063 if (fastio_reg_override >= 0)
3064 a = fastio_reg_override;
3065 if (offset_reg >= 0)
3066 emit_ldrh_dualindexed(offset_reg, a, tl);
3067 else
3068 emit_movzwl_indexed(0, a, tl);
3069 }
3070 if(jaddr)
3071 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3072 }
3073 else
3074 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3075 break;
3076 case 0x27: // LWU
3077 case 0x37: // LD
3078 default:
3079 assert(0);
3080 }
3081 }
3082 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3083 host_tempreg_release();
3084}
3085
3086#ifndef loadlr_assemble
3087static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3088{
3089 int s,tl,temp,temp2,addr;
3090 int offset;
3091 void *jaddr=0;
3092 int memtarget=0,c=0;
3093 int offset_reg = -1;
3094 int fastio_reg_override = -1;
3095 u_int reglist=get_host_reglist(i_regs->regmap);
3096 tl=get_reg(i_regs->regmap,dops[i].rt1);
3097 s=get_reg(i_regs->regmap,dops[i].rs1);
3098 temp=get_reg_temp(i_regs->regmap);
3099 temp2=get_reg(i_regs->regmap,FTEMP);
3100 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
3101 assert(addr<0);
3102 offset=imm[i];
3103 reglist|=1<<temp;
3104 if(offset||s<0||c) addr=temp2;
3105 else addr=s;
3106 if(s>=0) {
3107 c=(i_regs->wasconst>>s)&1;
3108 if(c) {
3109 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3110 }
3111 }
3112 if(!c) {
3113 emit_shlimm(addr,3,temp);
3114 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3115 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3116 }else{
3117 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3118 }
3119 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3120 &offset_reg, &fastio_reg_override);
3121 }
3122 else {
3123 if (ram_offset && memtarget) {
3124 offset_reg = get_ro_reg(i_regs, 0);
3125 }
3126 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3127 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3128 }else{
3129 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3130 }
3131 }
3132 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3133 if(!c||memtarget) {
3134 int a = temp2;
3135 if (fastio_reg_override >= 0)
3136 a = fastio_reg_override;
3137 do_load_word(a, temp2, offset_reg);
3138 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3139 host_tempreg_release();
3140 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3141 }
3142 else
3143 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3144 if(dops[i].rt1) {
3145 assert(tl>=0);
3146 emit_andimm(temp,24,temp);
3147 if (dops[i].opcode==0x22) // LWL
3148 emit_xorimm(temp,24,temp);
3149 host_tempreg_acquire();
3150 emit_movimm(-1,HOST_TEMPREG);
3151 if (dops[i].opcode==0x26) {
3152 emit_shr(temp2,temp,temp2);
3153 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3154 }else{
3155 emit_shl(temp2,temp,temp2);
3156 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3157 }
3158 host_tempreg_release();
3159 emit_or(temp2,tl,tl);
3160 }
3161 //emit_storereg(dops[i].rt1,tl); // DEBUG
3162 }
3163 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3164 assert(0);
3165 }
3166}
3167#endif
3168
3169static void do_invstub(int n)
3170{
3171 literal_pool(20);
3172 assem_debug("do_invstub\n");
3173 u_int reglist = stubs[n].a;
3174 u_int addrr = stubs[n].b;
3175 int ofs_start = stubs[n].c;
3176 int ofs_end = stubs[n].d;
3177 int len = ofs_end - ofs_start;
3178 u_int rightr = 0;
3179
3180 set_jump_target(stubs[n].addr, out);
3181 save_regs(reglist);
3182 emit_readword(&inv_code_start, 2);
3183 emit_readword(&inv_code_end, 3);
3184 if (addrr != 0 || ofs_start != 0)
3185 emit_addimm(addrr, ofs_start, 0);
3186 if (len != 0)
3187 emit_addimm(0, len + 4, (rightr = 1));
3188 emit_cmp(0, 2);
3189 emit_cmpcs(3, rightr);
3190 void *jaddr = out;
3191 emit_jc(0);
3192 void *func = (len != 0)
3193 ? (void *)ndrc_write_invalidate_many
3194 : (void *)ndrc_write_invalidate_one;
3195 emit_far_call(func);
3196 set_jump_target(jaddr, out);
3197 restore_regs(reglist);
3198 emit_jmp(stubs[n].retaddr);
3199}
3200
3201static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3202{
3203 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3204 return;
3205 // this can't be used any more since we started to check exact
3206 // block boundaries in invalidate_range()
3207 //if (i_regs->waswritten & (1<<dops[i].rs1))
3208 // return;
3209 // (naively) assume nobody will run code from stack
3210 if (dops[i].rs1 == 29)
3211 return;
3212
3213 int j, imm_maxdiff = 32, imm_min = imm[i], imm_max = imm[i], count = 1;
3214 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3215 && abs(imm[i+1] - imm[i]) <= imm_maxdiff)
3216 return;
3217 for (j = i - 1; j >= 0; j--) {
3218 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3219 || abs(imm[j] - imm[j+1]) > imm_maxdiff)
3220 break;
3221 count++;
3222 if (imm_min > imm[j])
3223 imm_min = imm[j];
3224 if (imm_max < imm[j])
3225 imm_max = imm[j];
3226 }
3227#if defined(HOST_IMM8)
3228 int ir = get_reg(i_regs->regmap, INVCP);
3229 assert(ir >= 0);
3230 host_tempreg_acquire();
3231 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3232#else
3233 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3234 #error not handled
3235#endif
3236#ifdef INVALIDATE_USE_COND_CALL
3237 if (count == 1) {
3238 emit_cmpimm(HOST_TEMPREG, 1);
3239 emit_callne(invalidate_addr_reg[addr]);
3240 host_tempreg_release();
3241 return;
3242 }
3243#endif
3244 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3245 host_tempreg_release();
3246 imm_min -= imm[i];
3247 imm_max -= imm[i];
3248 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3249 addr, imm_min, imm_max, 0);
3250}
3251
3252static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3253{
3254 int s,tl;
3255 int addr,temp;
3256 int offset;
3257 void *jaddr=0;
3258 enum stub_type type=0;
3259 int memtarget=0,c=0;
3260 int agr=AGEN1+(i&1);
3261 int offset_reg = -1;
3262 int fastio_reg_override = -1;
3263 u_int reglist=get_host_reglist(i_regs->regmap);
3264 tl=get_reg(i_regs->regmap,dops[i].rs2);
3265 s=get_reg(i_regs->regmap,dops[i].rs1);
3266 temp=get_reg(i_regs->regmap,agr);
3267 if(temp<0) temp=get_reg_temp(i_regs->regmap);
3268 offset=imm[i];
3269 if(s>=0) {
3270 c=(i_regs->wasconst>>s)&1;
3271 if(c) {
3272 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3273 }
3274 }
3275 assert(tl>=0);
3276 assert(temp>=0);
3277 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3278 if(offset||s<0||c) addr=temp;
3279 else addr=s;
3280 if (!c) {
3281 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3282 &offset_reg, &fastio_reg_override);
3283 }
3284 else if (ram_offset && memtarget) {
3285 offset_reg = get_ro_reg(i_regs, 0);
3286 }
3287
3288 switch (dops[i].opcode) {
3289 case 0x28: // SB
3290 if(!c||memtarget) {
3291 int a = temp;
3292 if (!c) a = addr;
3293 if (fastio_reg_override >= 0)
3294 a = fastio_reg_override;
3295 do_store_byte(a, tl, offset_reg);
3296 }
3297 type = STOREB_STUB;
3298 break;
3299 case 0x29: // SH
3300 if(!c||memtarget) {
3301 int a = temp;
3302 if (!c) a = addr;
3303 if (fastio_reg_override >= 0)
3304 a = fastio_reg_override;
3305 do_store_hword(a, 0, tl, offset_reg, 1);
3306 }
3307 type = STOREH_STUB;
3308 break;
3309 case 0x2B: // SW
3310 if(!c||memtarget) {
3311 int a = addr;
3312 if (fastio_reg_override >= 0)
3313 a = fastio_reg_override;
3314 do_store_word(a, 0, tl, offset_reg, 1);
3315 }
3316 type = STOREW_STUB;
3317 break;
3318 case 0x3F: // SD
3319 default:
3320 assert(0);
3321 }
3322 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3323 host_tempreg_release();
3324 if(jaddr) {
3325 // PCSX store handlers don't check invcode again
3326 reglist|=1<<addr;
3327 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3328 jaddr=0;
3329 }
3330 {
3331 if(!c||memtarget) {
3332 #ifdef DESTRUCTIVE_SHIFT
3333 // The x86 shift operation is 'destructive'; it overwrites the
3334 // source register, so we need to make a copy first and use that.
3335 addr=temp;
3336 #endif
3337 do_store_smc_check(i, i_regs, reglist, addr);
3338 }
3339 }
3340 u_int addr_val=constmap[i][s]+offset;
3341 if(jaddr) {
3342 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3343 } else if(c&&!memtarget) {
3344 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3345 }
3346 // basic current block modification detection..
3347 // not looking back as that should be in mips cache already
3348 // (see Spyro2 title->attract mode)
3349 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3350 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3351 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3352 if(i_regs->regmap==regs[i].regmap) {
3353 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3354 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3355 emit_movimm(start+i*4+4,0);
3356 emit_writeword(0,&pcaddr);
3357 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3358 emit_far_call(ndrc_get_addr_ht);
3359 emit_jmpreg(0);
3360 }
3361 }
3362}
3363
3364static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3365{
3366 int s,tl;
3367 int temp;
3368 int offset;
3369 void *jaddr=0;
3370 void *case1, *case23, *case3;
3371 void *done0, *done1, *done2;
3372 int memtarget=0,c=0;
3373 int agr=AGEN1+(i&1);
3374 int offset_reg = -1;
3375 u_int reglist=get_host_reglist(i_regs->regmap);
3376 tl=get_reg(i_regs->regmap,dops[i].rs2);
3377 s=get_reg(i_regs->regmap,dops[i].rs1);
3378 temp=get_reg(i_regs->regmap,agr);
3379 if(temp<0) temp=get_reg_temp(i_regs->regmap);
3380 offset=imm[i];
3381 if(s>=0) {
3382 c=(i_regs->isconst>>s)&1;
3383 if(c) {
3384 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3385 }
3386 }
3387 assert(tl>=0);
3388 assert(temp>=0);
3389 if(!c) {
3390 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3391 if(!offset&&s!=temp) emit_mov(s,temp);
3392 jaddr=out;
3393 emit_jno(0);
3394 }
3395 else
3396 {
3397 if(!memtarget||!dops[i].rs1) {
3398 jaddr=out;
3399 emit_jmp(0);
3400 }
3401 }
3402 if (ram_offset)
3403 offset_reg = get_ro_reg(i_regs, 0);
3404
3405 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3406 assert(0);
3407 }
3408
3409 emit_testimm(temp,2);
3410 case23=out;
3411 emit_jne(0);
3412 emit_testimm(temp,1);
3413 case1=out;
3414 emit_jne(0);
3415 // 0
3416 if (dops[i].opcode == 0x2A) { // SWL
3417 // Write msb into least significant byte
3418 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3419 do_store_byte(temp, tl, offset_reg);
3420 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3421 }
3422 else if (dops[i].opcode == 0x2E) { // SWR
3423 // Write entire word
3424 do_store_word(temp, 0, tl, offset_reg, 1);
3425 }
3426 done0 = out;
3427 emit_jmp(0);
3428 // 1
3429 set_jump_target(case1, out);
3430 if (dops[i].opcode == 0x2A) { // SWL
3431 // Write two msb into two least significant bytes
3432 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3433 do_store_hword(temp, -1, tl, offset_reg, 0);
3434 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3435 }
3436 else if (dops[i].opcode == 0x2E) { // SWR
3437 // Write 3 lsb into three most significant bytes
3438 do_store_byte(temp, tl, offset_reg);
3439 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3440 do_store_hword(temp, 1, tl, offset_reg, 0);
3441 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3442 }
3443 done1=out;
3444 emit_jmp(0);
3445 // 2,3
3446 set_jump_target(case23, out);
3447 emit_testimm(temp,1);
3448 case3 = out;
3449 emit_jne(0);
3450 // 2
3451 if (dops[i].opcode==0x2A) { // SWL
3452 // Write 3 msb into three least significant bytes
3453 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3454 do_store_hword(temp, -2, tl, offset_reg, 1);
3455 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3456 do_store_byte(temp, tl, offset_reg);
3457 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3458 }
3459 else if (dops[i].opcode == 0x2E) { // SWR
3460 // Write two lsb into two most significant bytes
3461 do_store_hword(temp, 0, tl, offset_reg, 1);
3462 }
3463 done2 = out;
3464 emit_jmp(0);
3465 // 3
3466 set_jump_target(case3, out);
3467 if (dops[i].opcode == 0x2A) { // SWL
3468 do_store_word(temp, -3, tl, offset_reg, 0);
3469 }
3470 else if (dops[i].opcode == 0x2E) { // SWR
3471 do_store_byte(temp, tl, offset_reg);
3472 }
3473 set_jump_target(done0, out);
3474 set_jump_target(done1, out);
3475 set_jump_target(done2, out);
3476 if (offset_reg == HOST_TEMPREG)
3477 host_tempreg_release();
3478 if(!c||!memtarget)
3479 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj_,reglist);
3480 do_store_smc_check(i, i_regs, reglist, temp);
3481}
3482
3483static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3484{
3485 if(dops[i].opcode2==0) // MFC0
3486 {
3487 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3488 u_int copr=(source[i]>>11)&0x1f;
3489 //assert(t>=0); // Why does this happen? OOT is weird
3490 if(t>=0&&dops[i].rt1!=0) {
3491 emit_readword(&reg_cop0[copr],t);
3492 }
3493 }
3494 else if(dops[i].opcode2==4) // MTC0
3495 {
3496 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3497 char copr=(source[i]>>11)&0x1f;
3498 assert(s>=0);
3499 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3500 if(copr==9||copr==11||copr==12||copr==13) {
3501 emit_readword(&last_count,HOST_TEMPREG);
3502 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3503 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3504 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3505 emit_writeword(HOST_CCREG,&Count);
3506 }
3507 // What a mess. The status register (12) can enable interrupts,
3508 // so needs a special case to handle a pending interrupt.
3509 // The interrupt must be taken immediately, because a subsequent
3510 // instruction might disable interrupts again.
3511 if(copr==12||copr==13) {
3512 if (is_delayslot) {
3513 // burn cycles to cause cc_interrupt, which will
3514 // reschedule next_interupt. Relies on CCREG from above.
3515 assem_debug("MTC0 DS %d\n", copr);
3516 emit_writeword(HOST_CCREG,&last_count);
3517 emit_movimm(0,HOST_CCREG);
3518 emit_storereg(CCREG,HOST_CCREG);
3519 emit_loadreg(dops[i].rs1,1);
3520 emit_movimm(copr,0);
3521 emit_far_call(pcsx_mtc0_ds);
3522 emit_loadreg(dops[i].rs1,s);
3523 return;
3524 }
3525 emit_movimm(start+i*4+4,HOST_TEMPREG);
3526 emit_writeword(HOST_TEMPREG,&pcaddr);
3527 emit_movimm(0,HOST_TEMPREG);
3528 emit_writeword(HOST_TEMPREG,&pending_exception);
3529 }
3530 if(s==HOST_CCREG)
3531 emit_loadreg(dops[i].rs1,1);
3532 else if(s!=1)
3533 emit_mov(s,1);
3534 emit_movimm(copr,0);
3535 emit_far_call(pcsx_mtc0);
3536 if(copr==9||copr==11||copr==12||copr==13) {
3537 emit_readword(&Count,HOST_CCREG);
3538 emit_readword(&next_interupt,HOST_TEMPREG);
3539 emit_addimm(HOST_CCREG,-ccadj_,HOST_CCREG);
3540 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3541 emit_writeword(HOST_TEMPREG,&last_count);
3542 emit_storereg(CCREG,HOST_CCREG);
3543 }
3544 if(copr==12||copr==13) {
3545 assert(!is_delayslot);
3546 emit_readword(&pending_exception,HOST_TEMPREG);
3547 emit_test(HOST_TEMPREG,HOST_TEMPREG);
3548 void *jaddr = out;
3549 emit_jeq(0);
3550 emit_readword(&pcaddr, 0);
3551 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3552 emit_far_call(ndrc_get_addr_ht);
3553 emit_jmpreg(0);
3554 set_jump_target(jaddr, out);
3555 }
3556 emit_loadreg(dops[i].rs1,s);
3557 }
3558 else
3559 {
3560 assert(dops[i].opcode2==0x10);
3561 //if((source[i]&0x3f)==0x10) // RFE
3562 {
3563 emit_readword(&Status,0);
3564 emit_andimm(0,0x3c,1);
3565 emit_andimm(0,~0xf,0);
3566 emit_orrshr_imm(1,2,0);
3567 emit_writeword(0,&Status);
3568 }
3569 }
3570}
3571
3572static void cop1_unusable(int i, const struct regstat *i_regs)
3573{
3574 // XXX: should just just do the exception instead
3575 //if(!cop1_usable)
3576 {
3577 void *jaddr=out;
3578 emit_jmp(0);
3579 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3580 }
3581}
3582
3583static void cop1_assemble(int i, const struct regstat *i_regs)
3584{
3585 cop1_unusable(i, i_regs);
3586}
3587
3588static void c1ls_assemble(int i, const struct regstat *i_regs)
3589{
3590 cop1_unusable(i, i_regs);
3591}
3592
3593// FP_STUB
3594static void do_cop1stub(int n)
3595{
3596 literal_pool(256);
3597 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3598 set_jump_target(stubs[n].addr, out);
3599 int i=stubs[n].a;
3600// int rs=stubs[n].b;
3601 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3602 int ds=stubs[n].d;
3603 if(!ds) {
3604 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3605 //if(i_regs!=&regs[i]) printf("oops: regs[i]=%x i_regs=%x",(int)&regs[i],(int)i_regs);
3606 }
3607 //else {printf("fp exception in delay slot\n");}
3608 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3609 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3610 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3611 emit_addimm(HOST_CCREG,ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3612 emit_far_jump(ds?fp_exception_ds:fp_exception);
3613}
3614
3615static int cop2_is_stalling_op(int i, int *cycles)
3616{
3617 if (dops[i].opcode == 0x3a) { // SWC2
3618 *cycles = 0;
3619 return 1;
3620 }
3621 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3622 *cycles = 0;
3623 return 1;
3624 }
3625 if (dops[i].itype == C2OP) {
3626 *cycles = gte_cycletab[source[i] & 0x3f];
3627 return 1;
3628 }
3629 // ... what about MTC2/CTC2/LWC2?
3630 return 0;
3631}
3632
3633#if 0
3634static void log_gte_stall(int stall, u_int cycle)
3635{
3636 if ((u_int)stall <= 44)
3637 printf("x stall %2d %u\n", stall, cycle + last_count);
3638}
3639
3640static void emit_log_gte_stall(int i, int stall, u_int reglist)
3641{
3642 save_regs(reglist);
3643 if (stall > 0)
3644 emit_movimm(stall, 0);
3645 else
3646 emit_mov(HOST_TEMPREG, 0);
3647 emit_addimm(HOST_CCREG, ccadj[i], 1);
3648 emit_far_call(log_gte_stall);
3649 restore_regs(reglist);
3650}
3651#endif
3652
3653static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3654{
3655 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3656 int rtmp = reglist_find_free(reglist);
3657
3658 if (HACK_ENABLED(NDHACK_NO_STALLS))
3659 return;
3660 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3661 // happens occasionally... cc evicted? Don't bother then
3662 //printf("no cc %08x\n", start + i*4);
3663 return;
3664 }
3665 if (!dops[i].bt) {
3666 for (j = i - 1; j >= 0; j--) {
3667 //if (dops[j].is_ds) break;
3668 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3669 break;
3670 if (j > 0 && ccadj[j - 1] > ccadj[j])
3671 break;
3672 }
3673 j = max(j, 0);
3674 }
3675 cycles_passed = ccadj[i] - ccadj[j];
3676 if (other_gte_op_cycles >= 0)
3677 stall = other_gte_op_cycles - cycles_passed;
3678 else if (cycles_passed >= 44)
3679 stall = 0; // can't stall
3680 if (stall == -MAXBLOCK && rtmp >= 0) {
3681 // unknown stall, do the expensive runtime check
3682 assem_debug("; cop2_do_stall_check\n");
3683#if 0 // too slow
3684 save_regs(reglist);
3685 emit_movimm(gte_cycletab[op], 0);
3686 emit_addimm(HOST_CCREG, ccadj[i], 1);
3687 emit_far_call(call_gteStall);
3688 restore_regs(reglist);
3689#else
3690 host_tempreg_acquire();
3691 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3692 emit_addimm(rtmp, -ccadj[i], rtmp);
3693 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3694 emit_cmpimm(HOST_TEMPREG, 44);
3695 emit_cmovb_reg(rtmp, HOST_CCREG);
3696 //emit_log_gte_stall(i, 0, reglist);
3697 host_tempreg_release();
3698#endif
3699 }
3700 else if (stall > 0) {
3701 //emit_log_gte_stall(i, stall, reglist);
3702 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3703 }
3704
3705 // save gteBusyCycle, if needed
3706 if (gte_cycletab[op] == 0)
3707 return;
3708 other_gte_op_cycles = -1;
3709 for (j = i + 1; j < slen; j++) {
3710 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3711 break;
3712 if (dops[j].is_jump) {
3713 // check ds
3714 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3715 j++;
3716 break;
3717 }
3718 }
3719 if (other_gte_op_cycles >= 0)
3720 // will handle stall when assembling that op
3721 return;
3722 cycles_passed = ccadj[min(j, slen -1)] - ccadj[i];
3723 if (cycles_passed >= 44)
3724 return;
3725 assem_debug("; save gteBusyCycle\n");
3726 host_tempreg_acquire();
3727#if 0
3728 emit_readword(&last_count, HOST_TEMPREG);
3729 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3730 emit_addimm(HOST_TEMPREG, ccadj[i], HOST_TEMPREG);
3731 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3732 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3733#else
3734 emit_addimm(HOST_CCREG, ccadj[i] + gte_cycletab[op], HOST_TEMPREG);
3735 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3736#endif
3737 host_tempreg_release();
3738}
3739
3740static int is_mflohi(int i)
3741{
3742 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3743}
3744
3745static int check_multdiv(int i, int *cycles)
3746{
3747 if (dops[i].itype != MULTDIV)
3748 return 0;
3749 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3750 *cycles = 11; // approx from 7 11 14
3751 else
3752 *cycles = 37;
3753 return 1;
3754}
3755
3756static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3757{
3758 int j, found = 0, c = 0;
3759 if (HACK_ENABLED(NDHACK_NO_STALLS))
3760 return;
3761 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3762 // happens occasionally... cc evicted? Don't bother then
3763 return;
3764 }
3765 for (j = i + 1; j < slen; j++) {
3766 if (dops[j].bt)
3767 break;
3768 if ((found = is_mflohi(j)))
3769 break;
3770 if (dops[j].is_jump) {
3771 // check ds
3772 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3773 j++;
3774 break;
3775 }
3776 }
3777 if (found)
3778 // handle all in multdiv_do_stall()
3779 return;
3780 check_multdiv(i, &c);
3781 assert(c > 0);
3782 assem_debug("; muldiv prepare stall %d\n", c);
3783 host_tempreg_acquire();
3784 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3785 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3786 host_tempreg_release();
3787}
3788
3789static void multdiv_do_stall(int i, const struct regstat *i_regs)
3790{
3791 int j, known_cycles = 0;
3792 u_int reglist = get_host_reglist(i_regs->regmap);
3793 int rtmp = get_reg_temp(i_regs->regmap);
3794 if (rtmp < 0)
3795 rtmp = reglist_find_free(reglist);
3796 if (HACK_ENABLED(NDHACK_NO_STALLS))
3797 return;
3798 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3799 // happens occasionally... cc evicted? Don't bother then
3800 //printf("no cc/rtmp %08x\n", start + i*4);
3801 return;
3802 }
3803 if (!dops[i].bt) {
3804 for (j = i - 1; j >= 0; j--) {
3805 if (dops[j].is_ds) break;
3806 if (check_multdiv(j, &known_cycles))
3807 break;
3808 if (is_mflohi(j))
3809 // already handled by this op
3810 return;
3811 if (dops[j].bt || (j > 0 && ccadj[j - 1] > ccadj[j]))
3812 break;
3813 }
3814 j = max(j, 0);
3815 }
3816 if (known_cycles > 0) {
3817 known_cycles -= ccadj[i] - ccadj[j];
3818 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3819 if (known_cycles > 0)
3820 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3821 return;
3822 }
3823 assem_debug("; muldiv stall unresolved\n");
3824 host_tempreg_acquire();
3825 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3826 emit_addimm(rtmp, -ccadj[i], rtmp);
3827 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3828 emit_cmpimm(HOST_TEMPREG, 37);
3829 emit_cmovb_reg(rtmp, HOST_CCREG);
3830 //emit_log_gte_stall(i, 0, reglist);
3831 host_tempreg_release();
3832}
3833
3834static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3835{
3836 switch (copr) {
3837 case 1:
3838 case 3:
3839 case 5:
3840 case 8:
3841 case 9:
3842 case 10:
3843 case 11:
3844 emit_readword(&reg_cop2d[copr],tl);
3845 emit_signextend16(tl,tl);
3846 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3847 break;
3848 case 7:
3849 case 16:
3850 case 17:
3851 case 18:
3852 case 19:
3853 emit_readword(&reg_cop2d[copr],tl);
3854 emit_andimm(tl,0xffff,tl);
3855 emit_writeword(tl,&reg_cop2d[copr]);
3856 break;
3857 case 15:
3858 emit_readword(&reg_cop2d[14],tl); // SXY2
3859 emit_writeword(tl,&reg_cop2d[copr]);
3860 break;
3861 case 28:
3862 case 29:
3863 c2op_mfc2_29_assemble(tl,temp);
3864 break;
3865 default:
3866 emit_readword(&reg_cop2d[copr],tl);
3867 break;
3868 }
3869}
3870
3871static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3872{
3873 switch (copr) {
3874 case 15:
3875 emit_readword(&reg_cop2d[13],temp); // SXY1
3876 emit_writeword(sl,&reg_cop2d[copr]);
3877 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3878 emit_readword(&reg_cop2d[14],temp); // SXY2
3879 emit_writeword(sl,&reg_cop2d[14]);
3880 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3881 break;
3882 case 28:
3883 emit_andimm(sl,0x001f,temp);
3884 emit_shlimm(temp,7,temp);
3885 emit_writeword(temp,&reg_cop2d[9]);
3886 emit_andimm(sl,0x03e0,temp);
3887 emit_shlimm(temp,2,temp);
3888 emit_writeword(temp,&reg_cop2d[10]);
3889 emit_andimm(sl,0x7c00,temp);
3890 emit_shrimm(temp,3,temp);
3891 emit_writeword(temp,&reg_cop2d[11]);
3892 emit_writeword(sl,&reg_cop2d[28]);
3893 break;
3894 case 30:
3895 emit_xorsar_imm(sl,sl,31,temp);
3896#if defined(HAVE_ARMV5) || defined(__aarch64__)
3897 emit_clz(temp,temp);
3898#else
3899 emit_movs(temp,HOST_TEMPREG);
3900 emit_movimm(0,temp);
3901 emit_jeq((int)out+4*4);
3902 emit_addpl_imm(temp,1,temp);
3903 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3904 emit_jns((int)out-2*4);
3905#endif
3906 emit_writeword(sl,&reg_cop2d[30]);
3907 emit_writeword(temp,&reg_cop2d[31]);
3908 break;
3909 case 31:
3910 break;
3911 default:
3912 emit_writeword(sl,&reg_cop2d[copr]);
3913 break;
3914 }
3915}
3916
3917static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3918{
3919 int s,tl;
3920 int ar;
3921 int offset;
3922 int memtarget=0,c=0;
3923 void *jaddr2=NULL;
3924 enum stub_type type;
3925 int agr=AGEN1+(i&1);
3926 int offset_reg = -1;
3927 int fastio_reg_override = -1;
3928 u_int reglist=get_host_reglist(i_regs->regmap);
3929 u_int copr=(source[i]>>16)&0x1f;
3930 s=get_reg(i_regs->regmap,dops[i].rs1);
3931 tl=get_reg(i_regs->regmap,FTEMP);
3932 offset=imm[i];
3933 assert(dops[i].rs1>0);
3934 assert(tl>=0);
3935
3936 if(i_regs->regmap[HOST_CCREG]==CCREG)
3937 reglist&=~(1<<HOST_CCREG);
3938
3939 // get the address
3940 if (dops[i].opcode==0x3a) { // SWC2
3941 ar=get_reg(i_regs->regmap,agr);
3942 if(ar<0) ar=get_reg_temp(i_regs->regmap);
3943 reglist|=1<<ar;
3944 } else { // LWC2
3945 ar=tl;
3946 }
3947 if(s>=0) c=(i_regs->wasconst>>s)&1;
3948 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3949 if (!offset&&!c&&s>=0) ar=s;
3950 assert(ar>=0);
3951
3952 cop2_do_stall_check(0, i, i_regs, reglist);
3953
3954 if (dops[i].opcode==0x3a) { // SWC2
3955 cop2_get_dreg(copr,tl,-1);
3956 type=STOREW_STUB;
3957 }
3958 else
3959 type=LOADW_STUB;
3960
3961 if(c&&!memtarget) {
3962 jaddr2=out;
3963 emit_jmp(0); // inline_readstub/inline_writestub?
3964 }
3965 else {
3966 if(!c) {
3967 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3968 &offset_reg, &fastio_reg_override);
3969 }
3970 else if (ram_offset && memtarget) {
3971 offset_reg = get_ro_reg(i_regs, 0);
3972 }
3973 switch (dops[i].opcode) {
3974 case 0x32: { // LWC2
3975 int a = ar;
3976 if (fastio_reg_override >= 0)
3977 a = fastio_reg_override;
3978 do_load_word(a, tl, offset_reg);
3979 break;
3980 }
3981 case 0x3a: { // SWC2
3982 #ifdef DESTRUCTIVE_SHIFT
3983 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3984 #endif
3985 int a = ar;
3986 if (fastio_reg_override >= 0)
3987 a = fastio_reg_override;
3988 do_store_word(a, 0, tl, offset_reg, 1);
3989 break;
3990 }
3991 default:
3992 assert(0);
3993 }
3994 }
3995 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3996 host_tempreg_release();
3997 if(jaddr2)
3998 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3999 if(dops[i].opcode==0x3a) // SWC2
4000 do_store_smc_check(i, i_regs, reglist, ar);
4001 if (dops[i].opcode==0x32) { // LWC2
4002 host_tempreg_acquire();
4003 cop2_put_dreg(copr,tl,HOST_TEMPREG);
4004 host_tempreg_release();
4005 }
4006}
4007
4008static void cop2_assemble(int i, const struct regstat *i_regs)
4009{
4010 u_int copr = (source[i]>>11) & 0x1f;
4011 signed char temp = get_reg_temp(i_regs->regmap);
4012
4013 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
4014 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
4015 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
4016 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
4017 reglist = reglist_exclude(reglist, tl, -1);
4018 }
4019 cop2_do_stall_check(0, i, i_regs, reglist);
4020 }
4021 if (dops[i].opcode2==0) { // MFC2
4022 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
4023 if(tl>=0&&dops[i].rt1!=0)
4024 cop2_get_dreg(copr,tl,temp);
4025 }
4026 else if (dops[i].opcode2==4) { // MTC2
4027 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4028 cop2_put_dreg(copr,sl,temp);
4029 }
4030 else if (dops[i].opcode2==2) // CFC2
4031 {
4032 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
4033 if(tl>=0&&dops[i].rt1!=0)
4034 emit_readword(&reg_cop2c[copr],tl);
4035 }
4036 else if (dops[i].opcode2==6) // CTC2
4037 {
4038 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4039 switch(copr) {
4040 case 4:
4041 case 12:
4042 case 20:
4043 case 26:
4044 case 27:
4045 case 29:
4046 case 30:
4047 emit_signextend16(sl,temp);
4048 break;
4049 case 31:
4050 c2op_ctc2_31_assemble(sl,temp);
4051 break;
4052 default:
4053 temp=sl;
4054 break;
4055 }
4056 emit_writeword(temp,&reg_cop2c[copr]);
4057 assert(sl>=0);
4058 }
4059}
4060
4061static void do_unalignedwritestub(int n)
4062{
4063 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4064 literal_pool(256);
4065 set_jump_target(stubs[n].addr, out);
4066
4067 int i=stubs[n].a;
4068 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4069 int addr=stubs[n].b;
4070 u_int reglist=stubs[n].e;
4071 signed char *i_regmap=i_regs->regmap;
4072 int temp2=get_reg(i_regmap,FTEMP);
4073 int rt;
4074 rt=get_reg(i_regmap,dops[i].rs2);
4075 assert(rt>=0);
4076 assert(addr>=0);
4077 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4078 reglist|=(1<<addr);
4079 reglist&=~(1<<temp2);
4080
4081 // don't bother with it and call write handler
4082 save_regs(reglist);
4083 pass_args(addr,rt);
4084 int cc=get_reg(i_regmap,CCREG);
4085 if(cc<0)
4086 emit_loadreg(CCREG,2);
4087 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4088 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4089 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4090 if(cc<0)
4091 emit_storereg(CCREG,2);
4092 restore_regs(reglist);
4093 emit_jmp(stubs[n].retaddr); // return address
4094}
4095
4096#ifndef multdiv_assemble
4097void multdiv_assemble(int i,struct regstat *i_regs)
4098{
4099 printf("Need multdiv_assemble for this architecture.\n");
4100 abort();
4101}
4102#endif
4103
4104static void mov_assemble(int i, const struct regstat *i_regs)
4105{
4106 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
4107 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
4108 if(dops[i].rt1) {
4109 signed char sl,tl;
4110 tl=get_reg(i_regs->regmap,dops[i].rt1);
4111 //assert(tl>=0);
4112 if(tl>=0) {
4113 sl=get_reg(i_regs->regmap,dops[i].rs1);
4114 if(sl>=0) emit_mov(sl,tl);
4115 else emit_loadreg(dops[i].rs1,tl);
4116 }
4117 }
4118 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
4119 multdiv_do_stall(i, i_regs);
4120}
4121
4122// call interpreter, exception handler, things that change pc/regs/cycles ...
4123static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
4124{
4125 signed char ccreg=get_reg(i_regs->regmap,CCREG);
4126 assert(ccreg==HOST_CCREG);
4127 assert(!is_delayslot);
4128 (void)ccreg;
4129
4130 emit_movimm(pc,3); // Get PC
4131 emit_readword(&last_count,2);
4132 emit_writeword(3,&psxRegs.pc);
4133 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
4134 emit_add(2,HOST_CCREG,2);
4135 emit_writeword(2,&psxRegs.cycle);
4136 emit_far_call(func);
4137 emit_far_jump(jump_to_new_pc);
4138}
4139
4140static void syscall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4141{
4142 // 'break' tends to be littered around to catch things like
4143 // division by 0 and is almost never executed, so don't emit much code here
4144 void *func = (dops[i].opcode2 == 0x0C)
4145 ? (is_delayslot ? jump_syscall_ds : jump_syscall)
4146 : (is_delayslot ? jump_break_ds : jump_break);
4147 assert(get_reg(i_regs->regmap, CCREG) == HOST_CCREG);
4148 emit_movimm(start + i*4, 2); // pc
4149 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4150 emit_far_jump(func);
4151}
4152
4153static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4154{
4155 void *hlefunc = psxNULL;
4156 uint32_t hleCode = source[i] & 0x03ffffff;
4157 if (hleCode < ARRAY_SIZE(psxHLEt))
4158 hlefunc = psxHLEt[hleCode];
4159
4160 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4161}
4162
4163static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4164{
4165 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4166}
4167
4168static void speculate_mov(int rs,int rt)
4169{
4170 if(rt!=0) {
4171 smrv_strong_next|=1<<rt;
4172 smrv[rt]=smrv[rs];
4173 }
4174}
4175
4176static void speculate_mov_weak(int rs,int rt)
4177{
4178 if(rt!=0) {
4179 smrv_weak_next|=1<<rt;
4180 smrv[rt]=smrv[rs];
4181 }
4182}
4183
4184static void speculate_register_values(int i)
4185{
4186 if(i==0) {
4187 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4188 // gp,sp are likely to stay the same throughout the block
4189 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4190 smrv_weak_next=~smrv_strong_next;
4191 //printf(" llr %08x\n", smrv[4]);
4192 }
4193 smrv_strong=smrv_strong_next;
4194 smrv_weak=smrv_weak_next;
4195 switch(dops[i].itype) {
4196 case ALU:
4197 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4198 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4199 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4200 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4201 else {
4202 smrv_strong_next&=~(1<<dops[i].rt1);
4203 smrv_weak_next&=~(1<<dops[i].rt1);
4204 }
4205 break;
4206 case SHIFTIMM:
4207 smrv_strong_next&=~(1<<dops[i].rt1);
4208 smrv_weak_next&=~(1<<dops[i].rt1);
4209 // fallthrough
4210 case IMM16:
4211 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4212 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
4213 if(hr>=0) {
4214 if(get_final_value(hr,i,&value))
4215 smrv[dops[i].rt1]=value;
4216 else smrv[dops[i].rt1]=constmap[i][hr];
4217 smrv_strong_next|=1<<dops[i].rt1;
4218 }
4219 }
4220 else {
4221 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4222 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4223 }
4224 break;
4225 case LOAD:
4226 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4227 // special case for BIOS
4228 smrv[dops[i].rt1]=0xa0000000;
4229 smrv_strong_next|=1<<dops[i].rt1;
4230 break;
4231 }
4232 // fallthrough
4233 case SHIFT:
4234 case LOADLR:
4235 case MOV:
4236 smrv_strong_next&=~(1<<dops[i].rt1);
4237 smrv_weak_next&=~(1<<dops[i].rt1);
4238 break;
4239 case COP0:
4240 case COP2:
4241 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4242 smrv_strong_next&=~(1<<dops[i].rt1);
4243 smrv_weak_next&=~(1<<dops[i].rt1);
4244 }
4245 break;
4246 case C2LS:
4247 if (dops[i].opcode==0x32) { // LWC2
4248 smrv_strong_next&=~(1<<dops[i].rt1);
4249 smrv_weak_next&=~(1<<dops[i].rt1);
4250 }
4251 break;
4252 }
4253#if 0
4254 int r=4;
4255 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4256 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4257#endif
4258}
4259
4260static void ujump_assemble(int i, const struct regstat *i_regs);
4261static void rjump_assemble(int i, const struct regstat *i_regs);
4262static void cjump_assemble(int i, const struct regstat *i_regs);
4263static void sjump_assemble(int i, const struct regstat *i_regs);
4264
4265static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4266{
4267 int ds = 0;
4268 switch (dops[i].itype) {
4269 case ALU:
4270 alu_assemble(i, i_regs);
4271 break;
4272 case IMM16:
4273 imm16_assemble(i, i_regs);
4274 break;
4275 case SHIFT:
4276 shift_assemble(i, i_regs);
4277 break;
4278 case SHIFTIMM:
4279 shiftimm_assemble(i, i_regs);
4280 break;
4281 case LOAD:
4282 load_assemble(i, i_regs, ccadj_);
4283 break;
4284 case LOADLR:
4285 loadlr_assemble(i, i_regs, ccadj_);
4286 break;
4287 case STORE:
4288 store_assemble(i, i_regs, ccadj_);
4289 break;
4290 case STORELR:
4291 storelr_assemble(i, i_regs, ccadj_);
4292 break;
4293 case COP0:
4294 cop0_assemble(i, i_regs, ccadj_);
4295 break;
4296 case COP1:
4297 cop1_assemble(i, i_regs);
4298 break;
4299 case C1LS:
4300 c1ls_assemble(i, i_regs);
4301 break;
4302 case COP2:
4303 cop2_assemble(i, i_regs);
4304 break;
4305 case C2LS:
4306 c2ls_assemble(i, i_regs, ccadj_);
4307 break;
4308 case C2OP:
4309 c2op_assemble(i, i_regs);
4310 break;
4311 case MULTDIV:
4312 multdiv_assemble(i, i_regs);
4313 multdiv_prepare_stall(i, i_regs, ccadj_);
4314 break;
4315 case MOV:
4316 mov_assemble(i, i_regs);
4317 break;
4318 case SYSCALL:
4319 syscall_assemble(i, i_regs, ccadj_);
4320 break;
4321 case HLECALL:
4322 hlecall_assemble(i, i_regs, ccadj_);
4323 break;
4324 case INTCALL:
4325 intcall_assemble(i, i_regs, ccadj_);
4326 break;
4327 case UJUMP:
4328 ujump_assemble(i, i_regs);
4329 ds = 1;
4330 break;
4331 case RJUMP:
4332 rjump_assemble(i, i_regs);
4333 ds = 1;
4334 break;
4335 case CJUMP:
4336 cjump_assemble(i, i_regs);
4337 ds = 1;
4338 break;
4339 case SJUMP:
4340 sjump_assemble(i, i_regs);
4341 ds = 1;
4342 break;
4343 case NOP:
4344 case OTHER:
4345 case NI:
4346 // not handled, just skip
4347 break;
4348 default:
4349 assert(0);
4350 }
4351 return ds;
4352}
4353
4354static void ds_assemble(int i, const struct regstat *i_regs)
4355{
4356 speculate_register_values(i);
4357 is_delayslot = 1;
4358 switch (dops[i].itype) {
4359 case SYSCALL:
4360 case HLECALL:
4361 case INTCALL:
4362 case UJUMP:
4363 case RJUMP:
4364 case CJUMP:
4365 case SJUMP:
4366 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4367 break;
4368 default:
4369 assemble(i, i_regs, ccadj[i]);
4370 }
4371 is_delayslot = 0;
4372}
4373
4374// Is the branch target a valid internal jump?
4375static int internal_branch(int addr)
4376{
4377 if(addr&1) return 0; // Indirect (register) jump
4378 if(addr>=start && addr<start+slen*4-4)
4379 {
4380 return 1;
4381 }
4382 return 0;
4383}
4384
4385static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4386{
4387 int hr;
4388 for(hr=0;hr<HOST_REGS;hr++) {
4389 if(hr!=EXCLUDE_REG) {
4390 if(pre[hr]!=entry[hr]) {
4391 if(pre[hr]>=0) {
4392 if((dirty>>hr)&1) {
4393 if(get_reg(entry,pre[hr])<0) {
4394 assert(pre[hr]<64);
4395 if(!((u>>pre[hr])&1))
4396 emit_storereg(pre[hr],hr);
4397 }
4398 }
4399 }
4400 }
4401 }
4402 }
4403 // Move from one register to another (no writeback)
4404 for(hr=0;hr<HOST_REGS;hr++) {
4405 if(hr!=EXCLUDE_REG) {
4406 if(pre[hr]!=entry[hr]) {
4407 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4408 int nr;
4409 if((nr=get_reg(entry,pre[hr]))>=0) {
4410 emit_mov(hr,nr);
4411 }
4412 }
4413 }
4414 }
4415 }
4416}
4417
4418// Load the specified registers
4419// This only loads the registers given as arguments because
4420// we don't want to load things that will be overwritten
4421static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4422{
4423 int hr = get_reg(regmap, rs);
4424 if (hr >= 0 && entry[hr] != regmap[hr])
4425 emit_loadreg(regmap[hr], hr);
4426}
4427
4428static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4429{
4430 load_reg(entry, regmap, rs1);
4431 if (rs1 != rs2)
4432 load_reg(entry, regmap, rs2);
4433}
4434
4435// Load registers prior to the start of a loop
4436// so that they are not loaded within the loop
4437static void loop_preload(signed char pre[],signed char entry[])
4438{
4439 int hr;
4440 for (hr = 0; hr < HOST_REGS; hr++) {
4441 int r = entry[hr];
4442 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4443 assem_debug("loop preload:\n");
4444 if (r < TEMPREG)
4445 emit_loadreg(r, hr);
4446 }
4447 }
4448}
4449
4450// Generate address for load/store instruction
4451// goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4452static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4453{
4454 if (dops[i].is_load || dops[i].is_store) {
4455 int ra=-1;
4456 int agr=AGEN1+(i&1);
4457 if(dops[i].itype==LOAD) {
4458 ra=get_reg(i_regs->regmap,dops[i].rt1);
4459 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4460 assert(ra>=0);
4461 }
4462 if(dops[i].itype==LOADLR) {
4463 ra=get_reg(i_regs->regmap,FTEMP);
4464 }
4465 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4466 ra=get_reg(i_regs->regmap,agr);
4467 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4468 }
4469 if(dops[i].itype==C2LS) {
4470 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4471 ra=get_reg(i_regs->regmap,FTEMP);
4472 else { // SWC1/SDC1/SWC2/SDC2
4473 ra=get_reg(i_regs->regmap,agr);
4474 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4475 }
4476 }
4477 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4478 if(ra>=0) {
4479 int offset=imm[i];
4480 int c=(i_regs->wasconst>>rs)&1;
4481 if(dops[i].rs1==0) {
4482 // Using r0 as a base address
4483 if(!entry||entry[ra]!=agr) {
4484 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4485 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4486 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4487 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4488 }else{
4489 emit_movimm(offset,ra);
4490 }
4491 } // else did it in the previous cycle
4492 }
4493 else if(rs<0) {
4494 if(!entry||entry[ra]!=dops[i].rs1)
4495 emit_loadreg(dops[i].rs1,ra);
4496 //if(!entry||entry[ra]!=dops[i].rs1)
4497 // printf("poor load scheduling!\n");
4498 }
4499 else if(c) {
4500 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4501 if(!entry||entry[ra]!=agr) {
4502 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4503 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4504 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4505 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4506 }else{
4507 emit_movimm(constmap[i][rs]+offset,ra);
4508 regs[i].loadedconst|=1<<ra;
4509 }
4510 } // else did it in the previous cycle
4511 } // else load_consts already did it
4512 }
4513 if(offset&&!c&&dops[i].rs1) {
4514 if(rs>=0) {
4515 emit_addimm(rs,offset,ra);
4516 }else{
4517 emit_addimm(ra,offset,ra);
4518 }
4519 }
4520 }
4521 }
4522 // Preload constants for next instruction
4523 if (dops[i+1].is_load || dops[i+1].is_store) {
4524 int agr,ra;
4525 // Actual address
4526 agr=AGEN1+((i+1)&1);
4527 ra=get_reg(i_regs->regmap,agr);
4528 if(ra>=0) {
4529 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4530 int offset=imm[i+1];
4531 int c=(regs[i+1].wasconst>>rs)&1;
4532 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4533 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4534 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4535 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4536 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4537 }else{
4538 emit_movimm(constmap[i+1][rs]+offset,ra);
4539 regs[i+1].loadedconst|=1<<ra;
4540 }
4541 }
4542 else if(dops[i+1].rs1==0) {
4543 // Using r0 as a base address
4544 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4545 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4546 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4547 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4548 }else{
4549 emit_movimm(offset,ra);
4550 }
4551 }
4552 }
4553 }
4554}
4555
4556static int get_final_value(int hr, int i, int *value)
4557{
4558 int reg=regs[i].regmap[hr];
4559 while(i<slen-1) {
4560 if(regs[i+1].regmap[hr]!=reg) break;
4561 if(!((regs[i+1].isconst>>hr)&1)) break;
4562 if(dops[i+1].bt) break;
4563 i++;
4564 }
4565 if(i<slen-1) {
4566 if (dops[i].is_jump) {
4567 *value=constmap[i][hr];
4568 return 1;
4569 }
4570 if(!dops[i+1].bt) {
4571 if (dops[i+1].is_jump) {
4572 // Load in delay slot, out-of-order execution
4573 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4574 {
4575 // Precompute load address
4576 *value=constmap[i][hr]+imm[i+2];
4577 return 1;
4578 }
4579 }
4580 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4581 {
4582 // Precompute load address
4583 *value=constmap[i][hr]+imm[i+1];
4584 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4585 return 1;
4586 }
4587 }
4588 }
4589 *value=constmap[i][hr];
4590 //printf("c=%lx\n",(long)constmap[i][hr]);
4591 if(i==slen-1) return 1;
4592 assert(reg < 64);
4593 return !((unneeded_reg[i+1]>>reg)&1);
4594}
4595
4596// Load registers with known constants
4597static void load_consts(signed char pre[],signed char regmap[],int i)
4598{
4599 int hr,hr2;
4600 // propagate loaded constant flags
4601 if(i==0||dops[i].bt)
4602 regs[i].loadedconst=0;
4603 else {
4604 for(hr=0;hr<HOST_REGS;hr++) {
4605 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4606 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4607 {
4608 regs[i].loadedconst|=1<<hr;
4609 }
4610 }
4611 }
4612 // Load 32-bit regs
4613 for(hr=0;hr<HOST_REGS;hr++) {
4614 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4615 //if(entry[hr]!=regmap[hr]) {
4616 if(!((regs[i].loadedconst>>hr)&1)) {
4617 assert(regmap[hr]<64);
4618 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4619 int value,similar=0;
4620 if(get_final_value(hr,i,&value)) {
4621 // see if some other register has similar value
4622 for(hr2=0;hr2<HOST_REGS;hr2++) {
4623 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4624 if(is_similar_value(value,constmap[i][hr2])) {
4625 similar=1;
4626 break;
4627 }
4628 }
4629 }
4630 if(similar) {
4631 int value2;
4632 if(get_final_value(hr2,i,&value2)) // is this needed?
4633 emit_movimm_from(value2,hr2,value,hr);
4634 else
4635 emit_movimm(value,hr);
4636 }
4637 else if(value==0) {
4638 emit_zeroreg(hr);
4639 }
4640 else {
4641 emit_movimm(value,hr);
4642 }
4643 }
4644 regs[i].loadedconst|=1<<hr;
4645 }
4646 }
4647 }
4648 }
4649}
4650
4651static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4652{
4653 int hr;
4654 // Load 32-bit regs
4655 for(hr=0;hr<HOST_REGS;hr++) {
4656 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4657 assert(regmap[hr] < 64);
4658 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4659 int value=constmap[i][hr];
4660 if(value==0) {
4661 emit_zeroreg(hr);
4662 }
4663 else {
4664 emit_movimm(value,hr);
4665 }
4666 }
4667 }
4668 }
4669}
4670
4671// Write out all dirty registers (except cycle count)
4672static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4673{
4674 int hr;
4675 for(hr=0;hr<HOST_REGS;hr++) {
4676 if(hr!=EXCLUDE_REG) {
4677 if(i_regmap[hr]>0) {
4678 if(i_regmap[hr]!=CCREG) {
4679 if((i_dirty>>hr)&1) {
4680 assert(i_regmap[hr]<64);
4681 emit_storereg(i_regmap[hr],hr);
4682 }
4683 }
4684 }
4685 }
4686 }
4687}
4688
4689// Write out dirty registers that we need to reload (pair with load_needed_regs)
4690// This writes the registers not written by store_regs_bt
4691static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4692{
4693 int hr;
4694 int t=(addr-start)>>2;
4695 for(hr=0;hr<HOST_REGS;hr++) {
4696 if(hr!=EXCLUDE_REG) {
4697 if(i_regmap[hr]>0) {
4698 if(i_regmap[hr]!=CCREG) {
4699 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4700 if((i_dirty>>hr)&1) {
4701 assert(i_regmap[hr]<64);
4702 emit_storereg(i_regmap[hr],hr);
4703 }
4704 }
4705 }
4706 }
4707 }
4708 }
4709}
4710
4711// Load all registers (except cycle count)
4712static void load_all_regs(const signed char i_regmap[])
4713{
4714 int hr;
4715 for(hr=0;hr<HOST_REGS;hr++) {
4716 if(hr!=EXCLUDE_REG) {
4717 if(i_regmap[hr]==0) {
4718 emit_zeroreg(hr);
4719 }
4720 else
4721 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4722 {
4723 emit_loadreg(i_regmap[hr],hr);
4724 }
4725 }
4726 }
4727}
4728
4729// Load all current registers also needed by next instruction
4730static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4731{
4732 int hr;
4733 for(hr=0;hr<HOST_REGS;hr++) {
4734 if(hr!=EXCLUDE_REG) {
4735 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4736 if(i_regmap[hr]==0) {
4737 emit_zeroreg(hr);
4738 }
4739 else
4740 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4741 {
4742 emit_loadreg(i_regmap[hr],hr);
4743 }
4744 }
4745 }
4746 }
4747}
4748
4749// Load all regs, storing cycle count if necessary
4750static void load_regs_entry(int t)
4751{
4752 int hr;
4753 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4754 else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t],HOST_CCREG);
4755 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4756 emit_storereg(CCREG,HOST_CCREG);
4757 }
4758 // Load 32-bit regs
4759 for(hr=0;hr<HOST_REGS;hr++) {
4760 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4761 if(regs[t].regmap_entry[hr]==0) {
4762 emit_zeroreg(hr);
4763 }
4764 else if(regs[t].regmap_entry[hr]!=CCREG)
4765 {
4766 emit_loadreg(regs[t].regmap_entry[hr],hr);
4767 }
4768 }
4769 }
4770}
4771
4772// Store dirty registers prior to branch
4773static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4774{
4775 if(internal_branch(addr))
4776 {
4777 int t=(addr-start)>>2;
4778 int hr;
4779 for(hr=0;hr<HOST_REGS;hr++) {
4780 if(hr!=EXCLUDE_REG) {
4781 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4782 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4783 if((i_dirty>>hr)&1) {
4784 assert(i_regmap[hr]<64);
4785 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4786 emit_storereg(i_regmap[hr],hr);
4787 }
4788 }
4789 }
4790 }
4791 }
4792 }
4793 else
4794 {
4795 // Branch out of this block, write out all dirty regs
4796 wb_dirtys(i_regmap,i_dirty);
4797 }
4798}
4799
4800// Load all needed registers for branch target
4801static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4802{
4803 //if(addr>=start && addr<(start+slen*4))
4804 if(internal_branch(addr))
4805 {
4806 int t=(addr-start)>>2;
4807 int hr;
4808 // Store the cycle count before loading something else
4809 if(i_regmap[HOST_CCREG]!=CCREG) {
4810 assert(i_regmap[HOST_CCREG]==-1);
4811 }
4812 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4813 emit_storereg(CCREG,HOST_CCREG);
4814 }
4815 // Load 32-bit regs
4816 for(hr=0;hr<HOST_REGS;hr++) {
4817 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4818 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4819 if(regs[t].regmap_entry[hr]==0) {
4820 emit_zeroreg(hr);
4821 }
4822 else if(regs[t].regmap_entry[hr]!=CCREG)
4823 {
4824 emit_loadreg(regs[t].regmap_entry[hr],hr);
4825 }
4826 }
4827 }
4828 }
4829 }
4830}
4831
4832static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4833{
4834 if(addr>=start && addr<start+slen*4-4)
4835 {
4836 int t=(addr-start)>>2;
4837 int hr;
4838 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4839 for(hr=0;hr<HOST_REGS;hr++)
4840 {
4841 if(hr!=EXCLUDE_REG)
4842 {
4843 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4844 {
4845 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4846 {
4847 return 0;
4848 }
4849 else
4850 if((i_dirty>>hr)&1)
4851 {
4852 if(i_regmap[hr]<TEMPREG)
4853 {
4854 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4855 return 0;
4856 }
4857 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4858 {
4859 assert(0);
4860 }
4861 }
4862 }
4863 else // Same register but is it 32-bit or dirty?
4864 if(i_regmap[hr]>=0)
4865 {
4866 if(!((regs[t].dirty>>hr)&1))
4867 {
4868 if((i_dirty>>hr)&1)
4869 {
4870 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4871 {
4872 //printf("%x: dirty no match\n",addr);
4873 return 0;
4874 }
4875 }
4876 }
4877 }
4878 }
4879 }
4880 // Delay slots are not valid branch targets
4881 //if(t>0&&(dops[t-1].is_jump) return 0;
4882 // Delay slots require additional processing, so do not match
4883 if(dops[t].is_ds) return 0;
4884 }
4885 else
4886 {
4887 int hr;
4888 for(hr=0;hr<HOST_REGS;hr++)
4889 {
4890 if(hr!=EXCLUDE_REG)
4891 {
4892 if(i_regmap[hr]>=0)
4893 {
4894 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4895 {
4896 if((i_dirty>>hr)&1)
4897 {
4898 return 0;
4899 }
4900 }
4901 }
4902 }
4903 }
4904 }
4905 return 1;
4906}
4907
4908#ifdef DRC_DBG
4909static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4910{
4911 extern void do_insn_cmp();
4912 //extern int cycle;
4913 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4914
4915 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4916 save_regs(reglist);
4917 // write out changed consts to match the interpreter
4918 if (i > 0 && !dops[i].bt) {
4919 for (hr = 0; hr < HOST_REGS; hr++) {
4920 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4921 if (hr == EXCLUDE_REG || reg < 0)
4922 continue;
4923 if (!((regs[i-1].isconst >> hr) & 1))
4924 continue;
4925 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4926 continue;
4927 emit_movimm(constmap[i-1][hr],0);
4928 emit_storereg(reg, 0);
4929 }
4930 }
4931 emit_movimm(start+i*4,0);
4932 emit_writeword(0,&pcaddr);
4933 int cc = get_reg(regs[i].regmap_entry, CCREG);
4934 if (cc < 0)
4935 emit_loadreg(CCREG, cc = 0);
4936 emit_addimm(cc, ccadj_, 0);
4937 emit_writeword(0, &psxRegs.cycle);
4938 emit_far_call(do_insn_cmp);
4939 //emit_readword(&cycle,0);
4940 //emit_addimm(0,2,0);
4941 //emit_writeword(0,&cycle);
4942 (void)get_reg2;
4943 restore_regs(reglist);
4944 assem_debug("\\\\do_insn_cmp\n");
4945}
4946#else
4947#define drc_dbg_emit_do_cmp(x,y)
4948#endif
4949
4950// Used when a branch jumps into the delay slot of another branch
4951static void ds_assemble_entry(int i)
4952{
4953 int t = (ba[i] - start) >> 2;
4954 int ccadj_ = -CLOCK_ADJUST(1);
4955 if (!instr_addr[t])
4956 instr_addr[t] = out;
4957 assem_debug("Assemble delay slot at %x\n",ba[i]);
4958 assem_debug("<->\n");
4959 drc_dbg_emit_do_cmp(t, ccadj_);
4960 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4961 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4962 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4963 address_generation(t,&regs[t],regs[t].regmap_entry);
4964 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4965 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
4966 if (dops[t].is_store)
4967 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
4968 is_delayslot=0;
4969 switch (dops[t].itype) {
4970 case SYSCALL:
4971 case HLECALL:
4972 case INTCALL:
4973 case UJUMP:
4974 case RJUMP:
4975 case CJUMP:
4976 case SJUMP:
4977 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4978 break;
4979 default:
4980 assemble(t, &regs[t], ccadj_);
4981 }
4982 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4983 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4984 if(internal_branch(ba[i]+4))
4985 assem_debug("branch: internal\n");
4986 else
4987 assem_debug("branch: external\n");
4988 assert(internal_branch(ba[i]+4));
4989 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4990 emit_jmp(0);
4991}
4992
4993// Load 2 immediates optimizing for small code size
4994static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4995{
4996 emit_movimm(imm1,rt1);
4997 emit_movimm_from(imm1,rt1,imm2,rt2);
4998}
4999
5000static void do_cc(int i, const signed char i_regmap[], int *adj,
5001 int addr, int taken, int invert)
5002{
5003 int count, count_plus2;
5004 void *jaddr;
5005 void *idle=NULL;
5006 int t=0;
5007 if(dops[i].itype==RJUMP)
5008 {
5009 *adj=0;
5010 }
5011 //if(ba[i]>=start && ba[i]<(start+slen*4))
5012 if(internal_branch(ba[i]))
5013 {
5014 t=(ba[i]-start)>>2;
5015 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5016 else *adj=ccadj[t];
5017 }
5018 else
5019 {
5020 *adj=0;
5021 }
5022 count = ccadj[i];
5023 count_plus2 = count + CLOCK_ADJUST(2);
5024 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
5025 // Idle loop
5026 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5027 idle=out;
5028 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5029 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5030 jaddr=out;
5031 emit_jmp(0);
5032 }
5033 else if(*adj==0||invert) {
5034 int cycles = count_plus2;
5035 // faster loop HACK
5036#if 0
5037 if (t&&*adj) {
5038 int rel=t-i;
5039 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5040 cycles=*adj+count+2-*adj;
5041 }
5042#endif
5043 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5044 jaddr = out;
5045 emit_jns(0);
5046 }
5047 else
5048 {
5049 emit_cmpimm(HOST_CCREG, -count_plus2);
5050 jaddr = out;
5051 emit_jns(0);
5052 }
5053 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5054}
5055
5056static void do_ccstub(int n)
5057{
5058 literal_pool(256);
5059 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5060 set_jump_target(stubs[n].addr, out);
5061 int i=stubs[n].b;
5062 if(stubs[n].d==NULLDS) {
5063 // Delay slot instruction is nullified ("likely" branch)
5064 wb_dirtys(regs[i].regmap,regs[i].dirty);
5065 }
5066 else if(stubs[n].d!=TAKEN) {
5067 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5068 }
5069 else {
5070 if(internal_branch(ba[i]))
5071 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5072 }
5073 if(stubs[n].c!=-1)
5074 {
5075 // Save PC as return address
5076 emit_movimm(stubs[n].c,EAX);
5077 emit_writeword(EAX,&pcaddr);
5078 }
5079 else
5080 {
5081 // Return address depends on which way the branch goes
5082 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5083 {
5084 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5085 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5086 if(dops[i].rs1==0)
5087 {
5088 s1l=s2l;
5089 s2l=-1;
5090 }
5091 else if(dops[i].rs2==0)
5092 {
5093 s2l=-1;
5094 }
5095 assert(s1l>=0);
5096 #ifdef DESTRUCTIVE_WRITEBACK
5097 if(dops[i].rs1) {
5098 if((branch_regs[i].dirty>>s1l)&&1)
5099 emit_loadreg(dops[i].rs1,s1l);
5100 }
5101 else {
5102 if((branch_regs[i].dirty>>s1l)&1)
5103 emit_loadreg(dops[i].rs2,s1l);
5104 }
5105 if(s2l>=0)
5106 if((branch_regs[i].dirty>>s2l)&1)
5107 emit_loadreg(dops[i].rs2,s2l);
5108 #endif
5109 int hr=0;
5110 int addr=-1,alt=-1,ntaddr=-1;
5111 while(hr<HOST_REGS)
5112 {
5113 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5114 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5115 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5116 {
5117 addr=hr++;break;
5118 }
5119 hr++;
5120 }
5121 while(hr<HOST_REGS)
5122 {
5123 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5124 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5125 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5126 {
5127 alt=hr++;break;
5128 }
5129 hr++;
5130 }
5131 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5132 {
5133 while(hr<HOST_REGS)
5134 {
5135 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5136 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5137 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5138 {
5139 ntaddr=hr;break;
5140 }
5141 hr++;
5142 }
5143 assert(hr<HOST_REGS);
5144 }
5145 if((dops[i].opcode&0x2f)==4) // BEQ
5146 {
5147 #ifdef HAVE_CMOV_IMM
5148 if(s2l>=0) emit_cmp(s1l,s2l);
5149 else emit_test(s1l,s1l);
5150 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5151 #else
5152 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5153 if(s2l>=0) emit_cmp(s1l,s2l);
5154 else emit_test(s1l,s1l);
5155 emit_cmovne_reg(alt,addr);
5156 #endif
5157 }
5158 if((dops[i].opcode&0x2f)==5) // BNE
5159 {
5160 #ifdef HAVE_CMOV_IMM
5161 if(s2l>=0) emit_cmp(s1l,s2l);
5162 else emit_test(s1l,s1l);
5163 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5164 #else
5165 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5166 if(s2l>=0) emit_cmp(s1l,s2l);
5167 else emit_test(s1l,s1l);
5168 emit_cmovne_reg(alt,addr);
5169 #endif
5170 }
5171 if((dops[i].opcode&0x2f)==6) // BLEZ
5172 {
5173 //emit_movimm(ba[i],alt);
5174 //emit_movimm(start+i*4+8,addr);
5175 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5176 emit_cmpimm(s1l,1);
5177 emit_cmovl_reg(alt,addr);
5178 }
5179 if((dops[i].opcode&0x2f)==7) // BGTZ
5180 {
5181 //emit_movimm(ba[i],addr);
5182 //emit_movimm(start+i*4+8,ntaddr);
5183 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5184 emit_cmpimm(s1l,1);
5185 emit_cmovl_reg(ntaddr,addr);
5186 }
5187 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
5188 {
5189 //emit_movimm(ba[i],alt);
5190 //emit_movimm(start+i*4+8,addr);
5191 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5192 emit_test(s1l,s1l);
5193 emit_cmovs_reg(alt,addr);
5194 }
5195 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
5196 {
5197 //emit_movimm(ba[i],addr);
5198 //emit_movimm(start+i*4+8,alt);
5199 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5200 emit_test(s1l,s1l);
5201 emit_cmovs_reg(alt,addr);
5202 }
5203 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5204 if(source[i]&0x10000) // BC1T
5205 {
5206 //emit_movimm(ba[i],alt);
5207 //emit_movimm(start+i*4+8,addr);
5208 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5209 emit_testimm(s1l,0x800000);
5210 emit_cmovne_reg(alt,addr);
5211 }
5212 else // BC1F
5213 {
5214 //emit_movimm(ba[i],addr);
5215 //emit_movimm(start+i*4+8,alt);
5216 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5217 emit_testimm(s1l,0x800000);
5218 emit_cmovne_reg(alt,addr);
5219 }
5220 }
5221 emit_writeword(addr,&pcaddr);
5222 }
5223 else
5224 if(dops[i].itype==RJUMP)
5225 {
5226 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5227 if (ds_writes_rjump_rs(i)) {
5228 r=get_reg(branch_regs[i].regmap,RTEMP);
5229 }
5230 emit_writeword(r,&pcaddr);
5231 }
5232 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5233 }
5234 // Update cycle count
5235 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5236 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5237 emit_far_call(cc_interrupt);
5238 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5239 if(stubs[n].d==TAKEN) {
5240 if(internal_branch(ba[i]))
5241 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
5242 else if(dops[i].itype==RJUMP) {
5243 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5244 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5245 else
5246 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5247 }
5248 }else if(stubs[n].d==NOTTAKEN) {
5249 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5250 else load_all_regs(branch_regs[i].regmap);
5251 }else if(stubs[n].d==NULLDS) {
5252 // Delay slot instruction is nullified ("likely" branch)
5253 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5254 else load_all_regs(regs[i].regmap);
5255 }else{
5256 load_all_regs(branch_regs[i].regmap);
5257 }
5258 if (stubs[n].retaddr)
5259 emit_jmp(stubs[n].retaddr);
5260 else
5261 do_jump_vaddr(stubs[n].e);
5262}
5263
5264static void add_to_linker(void *addr, u_int target, int is_internal)
5265{
5266 assert(linkcount < ARRAY_SIZE(link_addr));
5267 link_addr[linkcount].addr = addr;
5268 link_addr[linkcount].target = target;
5269 link_addr[linkcount].internal = is_internal;
5270 linkcount++;
5271}
5272
5273static void ujump_assemble_write_ra(int i)
5274{
5275 int rt;
5276 unsigned int return_address;
5277 rt=get_reg(branch_regs[i].regmap,31);
5278 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]);
5279 //assert(rt>=0);
5280 return_address=start+i*4+8;
5281 if(rt>=0) {
5282 #ifdef USE_MINI_HT
5283 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5284 int temp=-1; // note: must be ds-safe
5285 #ifdef HOST_TEMPREG
5286 temp=HOST_TEMPREG;
5287 #endif
5288 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5289 else emit_movimm(return_address,rt);
5290 }
5291 else
5292 #endif
5293 {
5294 #ifdef REG_PREFETCH
5295 if(temp>=0)
5296 {
5297 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5298 }
5299 #endif
5300 emit_movimm(return_address,rt); // PC into link register
5301 #ifdef IMM_PREFETCH
5302 emit_prefetch(hash_table_get(return_address));
5303 #endif
5304 }
5305 }
5306}
5307
5308static void ujump_assemble(int i, const struct regstat *i_regs)
5309{
5310 int ra_done=0;
5311 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5312 address_generation(i+1,i_regs,regs[i].regmap_entry);
5313 #ifdef REG_PREFETCH
5314 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5315 if(dops[i].rt1==31&&temp>=0)
5316 {
5317 signed char *i_regmap=i_regs->regmap;
5318 int return_address=start+i*4+8;
5319 if(get_reg(branch_regs[i].regmap,31)>0)
5320 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5321 }
5322 #endif
5323 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5324 ujump_assemble_write_ra(i); // writeback ra for DS
5325 ra_done=1;
5326 }
5327 ds_assemble(i+1,i_regs);
5328 uint64_t bc_unneeded=branch_regs[i].u;
5329 bc_unneeded|=1|(1LL<<dops[i].rt1);
5330 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5331 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5332 if(!ra_done&&dops[i].rt1==31)
5333 ujump_assemble_write_ra(i);
5334 int cc,adj;
5335 cc=get_reg(branch_regs[i].regmap,CCREG);
5336 assert(cc==HOST_CCREG);
5337 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5338 #ifdef REG_PREFETCH
5339 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5340 #endif
5341 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5342 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5343 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5344 if(internal_branch(ba[i]))
5345 assem_debug("branch: internal\n");
5346 else
5347 assem_debug("branch: external\n");
5348 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5349 ds_assemble_entry(i);
5350 }
5351 else {
5352 add_to_linker(out,ba[i],internal_branch(ba[i]));
5353 emit_jmp(0);
5354 }
5355}
5356
5357static void rjump_assemble_write_ra(int i)
5358{
5359 int rt,return_address;
5360 assert(dops[i+1].rt1!=dops[i].rt1);
5361 assert(dops[i+1].rt2!=dops[i].rt1);
5362 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5363 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]);
5364 assert(rt>=0);
5365 return_address=start+i*4+8;
5366 #ifdef REG_PREFETCH
5367 if(temp>=0)
5368 {
5369 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5370 }
5371 #endif
5372 emit_movimm(return_address,rt); // PC into link register
5373 #ifdef IMM_PREFETCH
5374 emit_prefetch(hash_table_get(return_address));
5375 #endif
5376}
5377
5378static void rjump_assemble(int i, const struct regstat *i_regs)
5379{
5380 int temp;
5381 int rs,cc;
5382 int ra_done=0;
5383 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5384 assert(rs>=0);
5385 if (ds_writes_rjump_rs(i)) {
5386 // Delay slot abuse, make a copy of the branch address register
5387 temp=get_reg(branch_regs[i].regmap,RTEMP);
5388 assert(temp>=0);
5389 assert(regs[i].regmap[temp]==RTEMP);
5390 emit_mov(rs,temp);
5391 rs=temp;
5392 }
5393 address_generation(i+1,i_regs,regs[i].regmap_entry);
5394 #ifdef REG_PREFETCH
5395 if(dops[i].rt1==31)
5396 {
5397 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5398 signed char *i_regmap=i_regs->regmap;
5399 int return_address=start+i*4+8;
5400 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5401 }
5402 }
5403 #endif
5404 #ifdef USE_MINI_HT
5405 if(dops[i].rs1==31) {
5406 int rh=get_reg(regs[i].regmap,RHASH);
5407 if(rh>=0) do_preload_rhash(rh);
5408 }
5409 #endif
5410 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5411 rjump_assemble_write_ra(i);
5412 ra_done=1;
5413 }
5414 ds_assemble(i+1,i_regs);
5415 uint64_t bc_unneeded=branch_regs[i].u;
5416 bc_unneeded|=1|(1LL<<dops[i].rt1);
5417 bc_unneeded&=~(1LL<<dops[i].rs1);
5418 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5419 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5420 if(!ra_done&&dops[i].rt1!=0)
5421 rjump_assemble_write_ra(i);
5422 cc=get_reg(branch_regs[i].regmap,CCREG);
5423 assert(cc==HOST_CCREG);
5424 (void)cc;
5425 #ifdef USE_MINI_HT
5426 int rh=get_reg(branch_regs[i].regmap,RHASH);
5427 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5428 if(dops[i].rs1==31) {
5429 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5430 do_preload_rhtbl(ht);
5431 do_rhash(rs,rh);
5432 }
5433 #endif
5434 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5435 #ifdef DESTRUCTIVE_WRITEBACK
5436 if((branch_regs[i].dirty>>rs)&1) {
5437 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5438 emit_loadreg(dops[i].rs1,rs);
5439 }
5440 }
5441 #endif
5442 #ifdef REG_PREFETCH
5443 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5444 #endif
5445 #ifdef USE_MINI_HT
5446 if(dops[i].rs1==31) {
5447 do_miniht_load(ht,rh);
5448 }
5449 #endif
5450 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5451 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5452 //assert(adj==0);
5453 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5454 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5455 if(dops[i+1].itype==COP0 && dops[i+1].opcode2==0x10)
5456 // special case for RFE
5457 emit_jmp(0);
5458 else
5459 emit_jns(0);
5460 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5461 #ifdef USE_MINI_HT
5462 if(dops[i].rs1==31) {
5463 do_miniht_jump(rs,rh,ht);
5464 }
5465 else
5466 #endif
5467 {
5468 do_jump_vaddr(rs);
5469 }
5470 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5471 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5472 #endif
5473}
5474
5475static void cjump_assemble(int i, const struct regstat *i_regs)
5476{
5477 const signed char *i_regmap = i_regs->regmap;
5478 int cc;
5479 int match;
5480 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5481 assem_debug("match=%d\n",match);
5482 int s1l,s2l;
5483 int unconditional=0,nop=0;
5484 int invert=0;
5485 int internal=internal_branch(ba[i]);
5486 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5487 if(!match) invert=1;
5488 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5489 if(i>(ba[i]-start)>>2) invert=1;
5490 #endif
5491 #ifdef __aarch64__
5492 invert=1; // because of near cond. branches
5493 #endif
5494
5495 if(dops[i].ooo) {
5496 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5497 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5498 }
5499 else {
5500 s1l=get_reg(i_regmap,dops[i].rs1);
5501 s2l=get_reg(i_regmap,dops[i].rs2);
5502 }
5503 if(dops[i].rs1==0&&dops[i].rs2==0)
5504 {
5505 if(dops[i].opcode&1) nop=1;
5506 else unconditional=1;
5507 //assert(dops[i].opcode!=5);
5508 //assert(dops[i].opcode!=7);
5509 //assert(dops[i].opcode!=0x15);
5510 //assert(dops[i].opcode!=0x17);
5511 }
5512 else if(dops[i].rs1==0)
5513 {
5514 s1l=s2l;
5515 s2l=-1;
5516 }
5517 else if(dops[i].rs2==0)
5518 {
5519 s2l=-1;
5520 }
5521
5522 if(dops[i].ooo) {
5523 // Out of order execution (delay slot first)
5524 //printf("OOOE\n");
5525 address_generation(i+1,i_regs,regs[i].regmap_entry);
5526 ds_assemble(i+1,i_regs);
5527 int adj;
5528 uint64_t bc_unneeded=branch_regs[i].u;
5529 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5530 bc_unneeded|=1;
5531 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5532 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5533 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5534 cc=get_reg(branch_regs[i].regmap,CCREG);
5535 assert(cc==HOST_CCREG);
5536 if(unconditional)
5537 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5538 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5539 //assem_debug("cycle count (adj)\n");
5540 if(unconditional) {
5541 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5542 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5543 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5544 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5545 if(internal)
5546 assem_debug("branch: internal\n");
5547 else
5548 assem_debug("branch: external\n");
5549 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5550 ds_assemble_entry(i);
5551 }
5552 else {
5553 add_to_linker(out,ba[i],internal);
5554 emit_jmp(0);
5555 }
5556 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5557 if(((u_int)out)&7) emit_addnop(0);
5558 #endif
5559 }
5560 }
5561 else if(nop) {
5562 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5563 void *jaddr=out;
5564 emit_jns(0);
5565 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5566 }
5567 else {
5568 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5569 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5570 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5571
5572 //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]);
5573 assert(s1l>=0);
5574 if(dops[i].opcode==4) // BEQ
5575 {
5576 if(s2l>=0) emit_cmp(s1l,s2l);
5577 else emit_test(s1l,s1l);
5578 if(invert){
5579 nottaken=out;
5580 emit_jne(DJT_1);
5581 }else{
5582 add_to_linker(out,ba[i],internal);
5583 emit_jeq(0);
5584 }
5585 }
5586 if(dops[i].opcode==5) // BNE
5587 {
5588 if(s2l>=0) emit_cmp(s1l,s2l);
5589 else emit_test(s1l,s1l);
5590 if(invert){
5591 nottaken=out;
5592 emit_jeq(DJT_1);
5593 }else{
5594 add_to_linker(out,ba[i],internal);
5595 emit_jne(0);
5596 }
5597 }
5598 if(dops[i].opcode==6) // BLEZ
5599 {
5600 emit_cmpimm(s1l,1);
5601 if(invert){
5602 nottaken=out;
5603 emit_jge(DJT_1);
5604 }else{
5605 add_to_linker(out,ba[i],internal);
5606 emit_jl(0);
5607 }
5608 }
5609 if(dops[i].opcode==7) // BGTZ
5610 {
5611 emit_cmpimm(s1l,1);
5612 if(invert){
5613 nottaken=out;
5614 emit_jl(DJT_1);
5615 }else{
5616 add_to_linker(out,ba[i],internal);
5617 emit_jge(0);
5618 }
5619 }
5620 if(invert) {
5621 if(taken) set_jump_target(taken, out);
5622 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5623 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5624 if(adj) {
5625 emit_addimm(cc,-adj,cc);
5626 add_to_linker(out,ba[i],internal);
5627 }else{
5628 emit_addnop(13);
5629 add_to_linker(out,ba[i],internal*2);
5630 }
5631 emit_jmp(0);
5632 }else
5633 #endif
5634 {
5635 if(adj) emit_addimm(cc,-adj,cc);
5636 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5637 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5638 if(internal)
5639 assem_debug("branch: internal\n");
5640 else
5641 assem_debug("branch: external\n");
5642 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5643 ds_assemble_entry(i);
5644 }
5645 else {
5646 add_to_linker(out,ba[i],internal);
5647 emit_jmp(0);
5648 }
5649 }
5650 set_jump_target(nottaken, out);
5651 }
5652
5653 if(nottaken1) set_jump_target(nottaken1, out);
5654 if(adj) {
5655 if(!invert) emit_addimm(cc,adj,cc);
5656 }
5657 } // (!unconditional)
5658 } // if(ooo)
5659 else
5660 {
5661 // In-order execution (branch first)
5662 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5663 if(!unconditional&&!nop) {
5664 //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]);
5665 assert(s1l>=0);
5666 if((dops[i].opcode&0x2f)==4) // BEQ
5667 {
5668 if(s2l>=0) emit_cmp(s1l,s2l);
5669 else emit_test(s1l,s1l);
5670 nottaken=out;
5671 emit_jne(DJT_2);
5672 }
5673 if((dops[i].opcode&0x2f)==5) // BNE
5674 {
5675 if(s2l>=0) emit_cmp(s1l,s2l);
5676 else emit_test(s1l,s1l);
5677 nottaken=out;
5678 emit_jeq(DJT_2);
5679 }
5680 if((dops[i].opcode&0x2f)==6) // BLEZ
5681 {
5682 emit_cmpimm(s1l,1);
5683 nottaken=out;
5684 emit_jge(DJT_2);
5685 }
5686 if((dops[i].opcode&0x2f)==7) // BGTZ
5687 {
5688 emit_cmpimm(s1l,1);
5689 nottaken=out;
5690 emit_jl(DJT_2);
5691 }
5692 } // if(!unconditional)
5693 int adj;
5694 uint64_t ds_unneeded=branch_regs[i].u;
5695 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5696 ds_unneeded|=1;
5697 // branch taken
5698 if(!nop) {
5699 if(taken) set_jump_target(taken, out);
5700 assem_debug("1:\n");
5701 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5702 // load regs
5703 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5704 address_generation(i+1,&branch_regs[i],0);
5705 if (ram_offset)
5706 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5707 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5708 ds_assemble(i+1,&branch_regs[i]);
5709 cc=get_reg(branch_regs[i].regmap,CCREG);
5710 if(cc==-1) {
5711 emit_loadreg(CCREG,cc=HOST_CCREG);
5712 // CHECK: Is the following instruction (fall thru) allocated ok?
5713 }
5714 assert(cc==HOST_CCREG);
5715 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5716 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5717 assem_debug("cycle count (adj)\n");
5718 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5719 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5720 if(internal)
5721 assem_debug("branch: internal\n");
5722 else
5723 assem_debug("branch: external\n");
5724 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5725 ds_assemble_entry(i);
5726 }
5727 else {
5728 add_to_linker(out,ba[i],internal);
5729 emit_jmp(0);
5730 }
5731 }
5732 // branch not taken
5733 if(!unconditional) {
5734 if(nottaken1) set_jump_target(nottaken1, out);
5735 set_jump_target(nottaken, out);
5736 assem_debug("2:\n");
5737 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5738 // load regs
5739 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5740 address_generation(i+1,&branch_regs[i],0);
5741 if (ram_offset)
5742 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5743 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5744 ds_assemble(i+1,&branch_regs[i]);
5745 cc=get_reg(branch_regs[i].regmap,CCREG);
5746 if (cc == -1) {
5747 // Cycle count isn't in a register, temporarily load it then write it out
5748 emit_loadreg(CCREG,HOST_CCREG);
5749 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5750 void *jaddr=out;
5751 emit_jns(0);
5752 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5753 emit_storereg(CCREG,HOST_CCREG);
5754 }
5755 else{
5756 cc=get_reg(i_regmap,CCREG);
5757 assert(cc==HOST_CCREG);
5758 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5759 void *jaddr=out;
5760 emit_jns(0);
5761 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5762 }
5763 }
5764 }
5765}
5766
5767static void sjump_assemble(int i, const struct regstat *i_regs)
5768{
5769 const signed char *i_regmap = i_regs->regmap;
5770 int cc;
5771 int match;
5772 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5773 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5774 int s1l;
5775 int unconditional=0,nevertaken=0;
5776 int invert=0;
5777 int internal=internal_branch(ba[i]);
5778 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5779 if(!match) invert=1;
5780 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5781 if(i>(ba[i]-start)>>2) invert=1;
5782 #endif
5783 #ifdef __aarch64__
5784 invert=1; // because of near cond. branches
5785 #endif
5786
5787 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5788 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5789
5790 if(dops[i].ooo) {
5791 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5792 }
5793 else {
5794 s1l=get_reg(i_regmap,dops[i].rs1);
5795 }
5796 if(dops[i].rs1==0)
5797 {
5798 if(dops[i].opcode2&1) unconditional=1;
5799 else nevertaken=1;
5800 // These are never taken (r0 is never less than zero)
5801 //assert(dops[i].opcode2!=0);
5802 //assert(dops[i].opcode2!=2);
5803 //assert(dops[i].opcode2!=0x10);
5804 //assert(dops[i].opcode2!=0x12);
5805 }
5806
5807 if(dops[i].ooo) {
5808 // Out of order execution (delay slot first)
5809 //printf("OOOE\n");
5810 address_generation(i+1,i_regs,regs[i].regmap_entry);
5811 ds_assemble(i+1,i_regs);
5812 int adj;
5813 uint64_t bc_unneeded=branch_regs[i].u;
5814 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5815 bc_unneeded|=1;
5816 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5817 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5818 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5819 if(dops[i].rt1==31) {
5820 int rt,return_address;
5821 rt=get_reg(branch_regs[i].regmap,31);
5822 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]);
5823 if(rt>=0) {
5824 // Save the PC even if the branch is not taken
5825 return_address=start+i*4+8;
5826 emit_movimm(return_address,rt); // PC into link register
5827 #ifdef IMM_PREFETCH
5828 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5829 #endif
5830 }
5831 }
5832 cc=get_reg(branch_regs[i].regmap,CCREG);
5833 assert(cc==HOST_CCREG);
5834 if(unconditional)
5835 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5836 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5837 assem_debug("cycle count (adj)\n");
5838 if(unconditional) {
5839 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5840 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5841 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5842 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5843 if(internal)
5844 assem_debug("branch: internal\n");
5845 else
5846 assem_debug("branch: external\n");
5847 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5848 ds_assemble_entry(i);
5849 }
5850 else {
5851 add_to_linker(out,ba[i],internal);
5852 emit_jmp(0);
5853 }
5854 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5855 if(((u_int)out)&7) emit_addnop(0);
5856 #endif
5857 }
5858 }
5859 else if(nevertaken) {
5860 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5861 void *jaddr=out;
5862 emit_jns(0);
5863 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5864 }
5865 else {
5866 void *nottaken = NULL;
5867 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5868 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5869 {
5870 assert(s1l>=0);
5871 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5872 {
5873 emit_test(s1l,s1l);
5874 if(invert){
5875 nottaken=out;
5876 emit_jns(DJT_1);
5877 }else{
5878 add_to_linker(out,ba[i],internal);
5879 emit_js(0);
5880 }
5881 }
5882 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5883 {
5884 emit_test(s1l,s1l);
5885 if(invert){
5886 nottaken=out;
5887 emit_js(DJT_1);
5888 }else{
5889 add_to_linker(out,ba[i],internal);
5890 emit_jns(0);
5891 }
5892 }
5893 }
5894
5895 if(invert) {
5896 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5897 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5898 if(adj) {
5899 emit_addimm(cc,-adj,cc);
5900 add_to_linker(out,ba[i],internal);
5901 }else{
5902 emit_addnop(13);
5903 add_to_linker(out,ba[i],internal*2);
5904 }
5905 emit_jmp(0);
5906 }else
5907 #endif
5908 {
5909 if(adj) emit_addimm(cc,-adj,cc);
5910 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5911 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5912 if(internal)
5913 assem_debug("branch: internal\n");
5914 else
5915 assem_debug("branch: external\n");
5916 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5917 ds_assemble_entry(i);
5918 }
5919 else {
5920 add_to_linker(out,ba[i],internal);
5921 emit_jmp(0);
5922 }
5923 }
5924 set_jump_target(nottaken, out);
5925 }
5926
5927 if(adj) {
5928 if(!invert) emit_addimm(cc,adj,cc);
5929 }
5930 } // (!unconditional)
5931 } // if(ooo)
5932 else
5933 {
5934 // In-order execution (branch first)
5935 //printf("IOE\n");
5936 void *nottaken = NULL;
5937 if(dops[i].rt1==31) {
5938 int rt,return_address;
5939 rt=get_reg(branch_regs[i].regmap,31);
5940 if(rt>=0) {
5941 // Save the PC even if the branch is not taken
5942 return_address=start+i*4+8;
5943 emit_movimm(return_address,rt); // PC into link register
5944 #ifdef IMM_PREFETCH
5945 emit_prefetch(hash_table_get(return_address));
5946 #endif
5947 }
5948 }
5949 if(!unconditional) {
5950 //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]);
5951 assert(s1l>=0);
5952 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5953 {
5954 emit_test(s1l,s1l);
5955 nottaken=out;
5956 emit_jns(DJT_1);
5957 }
5958 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5959 {
5960 emit_test(s1l,s1l);
5961 nottaken=out;
5962 emit_js(DJT_1);
5963 }
5964 } // if(!unconditional)
5965 int adj;
5966 uint64_t ds_unneeded=branch_regs[i].u;
5967 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5968 ds_unneeded|=1;
5969 // branch taken
5970 if(!nevertaken) {
5971 //assem_debug("1:\n");
5972 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5973 // load regs
5974 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5975 address_generation(i+1,&branch_regs[i],0);
5976 if (ram_offset)
5977 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5978 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5979 ds_assemble(i+1,&branch_regs[i]);
5980 cc=get_reg(branch_regs[i].regmap,CCREG);
5981 if(cc==-1) {
5982 emit_loadreg(CCREG,cc=HOST_CCREG);
5983 // CHECK: Is the following instruction (fall thru) allocated ok?
5984 }
5985 assert(cc==HOST_CCREG);
5986 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5987 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5988 assem_debug("cycle count (adj)\n");
5989 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5990 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5991 if(internal)
5992 assem_debug("branch: internal\n");
5993 else
5994 assem_debug("branch: external\n");
5995 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5996 ds_assemble_entry(i);
5997 }
5998 else {
5999 add_to_linker(out,ba[i],internal);
6000 emit_jmp(0);
6001 }
6002 }
6003 // branch not taken
6004 if(!unconditional) {
6005 set_jump_target(nottaken, out);
6006 assem_debug("1:\n");
6007 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6008 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6009 address_generation(i+1,&branch_regs[i],0);
6010 if (ram_offset)
6011 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6012 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6013 ds_assemble(i+1,&branch_regs[i]);
6014 cc=get_reg(branch_regs[i].regmap,CCREG);
6015 if (cc == -1) {
6016 // Cycle count isn't in a register, temporarily load it then write it out
6017 emit_loadreg(CCREG,HOST_CCREG);
6018 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
6019 void *jaddr=out;
6020 emit_jns(0);
6021 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6022 emit_storereg(CCREG,HOST_CCREG);
6023 }
6024 else{
6025 cc=get_reg(i_regmap,CCREG);
6026 assert(cc==HOST_CCREG);
6027 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
6028 void *jaddr=out;
6029 emit_jns(0);
6030 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6031 }
6032 }
6033 }
6034}
6035
6036static void check_regmap(signed char *regmap)
6037{
6038#ifndef NDEBUG
6039 int i,j;
6040 for (i = 0; i < HOST_REGS; i++) {
6041 if (regmap[i] < 0)
6042 continue;
6043 for (j = i + 1; j < HOST_REGS; j++)
6044 assert(regmap[i] != regmap[j]);
6045 }
6046#endif
6047}
6048
6049#ifdef DISASM
6050#include <inttypes.h>
6051static char insn[MAXBLOCK][10];
6052
6053#define set_mnemonic(i_, n_) \
6054 strcpy(insn[i_], n_)
6055
6056void print_regmap(const char *name, const signed char *regmap)
6057{
6058 char buf[5];
6059 int i, l;
6060 fputs(name, stdout);
6061 for (i = 0; i < HOST_REGS; i++) {
6062 l = 0;
6063 if (regmap[i] >= 0)
6064 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6065 for (; l < 3; l++)
6066 buf[l] = ' ';
6067 buf[l] = 0;
6068 printf(" r%d=%s", i, buf);
6069 }
6070 fputs("\n", stdout);
6071}
6072
6073 /* disassembly */
6074void disassemble_inst(int i)
6075{
6076 if (dops[i].bt) printf("*"); else printf(" ");
6077 switch(dops[i].itype) {
6078 case UJUMP:
6079 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6080 case CJUMP:
6081 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;
6082 case SJUMP:
6083 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;
6084 case RJUMP:
6085 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6086 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6087 else
6088 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6089 break;
6090 case IMM16:
6091 if(dops[i].opcode==0xf) //LUI
6092 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6093 else
6094 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6095 break;
6096 case LOAD:
6097 case LOADLR:
6098 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6099 break;
6100 case STORE:
6101 case STORELR:
6102 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6103 break;
6104 case ALU:
6105 case SHIFT:
6106 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6107 break;
6108 case MULTDIV:
6109 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6110 break;
6111 case SHIFTIMM:
6112 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6113 break;
6114 case MOV:
6115 if((dops[i].opcode2&0x1d)==0x10)
6116 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6117 else if((dops[i].opcode2&0x1d)==0x11)
6118 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6119 else
6120 printf (" %x: %s\n",start+i*4,insn[i]);
6121 break;
6122 case COP0:
6123 if(dops[i].opcode2==0)
6124 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6125 else if(dops[i].opcode2==4)
6126 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6127 else printf (" %x: %s\n",start+i*4,insn[i]);
6128 break;
6129 case COP1:
6130 if(dops[i].opcode2<3)
6131 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6132 else if(dops[i].opcode2>3)
6133 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6134 else printf (" %x: %s\n",start+i*4,insn[i]);
6135 break;
6136 case COP2:
6137 if(dops[i].opcode2<3)
6138 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6139 else if(dops[i].opcode2>3)
6140 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6141 else printf (" %x: %s\n",start+i*4,insn[i]);
6142 break;
6143 case C1LS:
6144 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6145 break;
6146 case C2LS:
6147 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6148 break;
6149 case INTCALL:
6150 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6151 break;
6152 default:
6153 //printf (" %s %8x\n",insn[i],source[i]);
6154 printf (" %x: %s\n",start+i*4,insn[i]);
6155 }
6156 return;
6157 printf("D: %"PRIu64" WD: %"PRIu64" U: %"PRIu64"\n",
6158 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i]);
6159 print_regmap("pre: ", regmap_pre[i]);
6160 print_regmap("entry: ", regs[i].regmap_entry);
6161 print_regmap("map: ", regs[i].regmap);
6162 if (dops[i].is_jump) {
6163 print_regmap("bentry:", branch_regs[i].regmap_entry);
6164 print_regmap("bmap: ", branch_regs[i].regmap);
6165 }
6166}
6167#else
6168#define set_mnemonic(i_, n_)
6169static void disassemble_inst(int i) {}
6170#endif // DISASM
6171
6172#define DRC_TEST_VAL 0x74657374
6173
6174static noinline void new_dynarec_test(void)
6175{
6176 int (*testfunc)(void);
6177 void *beginning;
6178 int ret[2];
6179 size_t i;
6180
6181 // check structure linkage
6182 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6183 {
6184 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6185 }
6186
6187 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6188 new_dynarec_test, out);
6189 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6190
6191 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6192 out = ndrc->translation_cache;
6193 beginning = start_block();
6194 emit_movimm(DRC_TEST_VAL + i, 0); // test
6195 emit_ret();
6196 literal_pool(0);
6197 end_block(beginning);
6198 testfunc = beginning;
6199 ret[i] = testfunc();
6200 }
6201
6202 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6203 SysPrintf("test passed.\n");
6204 else
6205 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6206 out = ndrc->translation_cache;
6207}
6208
6209// clear the state completely, instead of just marking
6210// things invalid like invalidate_all_pages() does
6211void new_dynarec_clear_full(void)
6212{
6213 int n;
6214 out = ndrc->translation_cache;
6215 memset(invalid_code,1,sizeof(invalid_code));
6216 memset(hash_table,0xff,sizeof(hash_table));
6217 memset(mini_ht,-1,sizeof(mini_ht));
6218 memset(shadow,0,sizeof(shadow));
6219 copy=shadow;
6220 expirep = EXPIRITY_OFFSET;
6221 pending_exception=0;
6222 literalcount=0;
6223 stop_after_jal=0;
6224 inv_code_start=inv_code_end=~0;
6225 hack_addr=0;
6226 f1_hack=0;
6227 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6228 blocks_clear(&blocks[n]);
6229 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6230 free(jumps[n]);
6231 jumps[n] = NULL;
6232 }
6233 stat_clear(stat_blocks);
6234 stat_clear(stat_links);
6235
6236 cycle_multiplier_old = cycle_multiplier;
6237 new_dynarec_hacks_old = new_dynarec_hacks;
6238}
6239
6240void new_dynarec_init(void)
6241{
6242 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6243
6244#ifdef _3DS
6245 check_rosalina();
6246#endif
6247#ifdef BASE_ADDR_DYNAMIC
6248 #ifdef VITA
6249 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6250 if (sceBlock <= 0)
6251 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6252 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6253 if (ret < 0)
6254 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6255 sceKernelOpenVMDomain();
6256 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6257 #elif defined(_MSC_VER)
6258 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6259 PAGE_EXECUTE_READWRITE);
6260 #elif defined(HAVE_LIBNX)
6261 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6262 if (R_FAILED(rc))
6263 SysPrintf("jitCreate failed: %08x\n", rc);
6264 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6265 jitTransitionToWritable(&g_jit);
6266 ndrc = g_jit.rx_addr;
6267 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6268 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6269 #else
6270 uintptr_t desired_addr = 0;
6271 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6272 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6273 int fd = -1;
6274 #ifdef __ELF__
6275 extern char _end;
6276 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6277 #endif
6278 #ifdef TC_WRITE_OFFSET
6279 // mostly for testing
6280 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6281 ftruncate(fd, sizeof(*ndrc));
6282 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6283 (flags = MAP_SHARED), fd, 0);
6284 assert(mw != MAP_FAILED);
6285 prot = PROT_READ | PROT_EXEC;
6286 #endif
6287 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6288 if (ndrc == MAP_FAILED) {
6289 SysPrintf("mmap() failed: %s\n", strerror(errno));
6290 abort();
6291 }
6292 #ifdef TC_WRITE_OFFSET
6293 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6294 #endif
6295 #endif
6296#else
6297 #ifndef NO_WRITE_EXEC
6298 // not all systems allow execute in data segment by default
6299 // size must be 4K aligned for 3DS?
6300 if (mprotect(ndrc, sizeof(*ndrc),
6301 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6302 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6303 #endif
6304#endif
6305 out = ndrc->translation_cache;
6306 cycle_multiplier=200;
6307 new_dynarec_clear_full();
6308#ifdef HOST_IMM8
6309 // Copy this into local area so we don't have to put it in every literal pool
6310 invc_ptr=invalid_code;
6311#endif
6312 arch_init();
6313 new_dynarec_test();
6314 ram_offset=(uintptr_t)rdram-0x80000000;
6315 if (ram_offset!=0)
6316 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6317 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6318 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6319}
6320
6321void new_dynarec_cleanup(void)
6322{
6323 int n;
6324#ifdef BASE_ADDR_DYNAMIC
6325 #ifdef VITA
6326 // sceBlock is managed by retroarch's bootstrap code
6327 //sceKernelFreeMemBlock(sceBlock);
6328 //sceBlock = -1;
6329 #elif defined(HAVE_LIBNX)
6330 jitClose(&g_jit);
6331 ndrc = NULL;
6332 #else
6333 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6334 SysPrintf("munmap() failed\n");
6335 ndrc = NULL;
6336 #endif
6337#endif
6338 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6339 blocks_clear(&blocks[n]);
6340 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6341 free(jumps[n]);
6342 jumps[n] = NULL;
6343 }
6344 stat_clear(stat_blocks);
6345 stat_clear(stat_links);
6346 new_dynarec_print_stats();
6347}
6348
6349static u_int *get_source_start(u_int addr, u_int *limit)
6350{
6351 if (addr < 0x00200000 ||
6352 (0xa0000000 <= addr && addr < 0xa0200000))
6353 {
6354 // used for BIOS calls mostly?
6355 *limit = (addr&0xa0000000)|0x00200000;
6356 return (u_int *)(rdram + (addr&0x1fffff));
6357 }
6358 else if (!Config.HLE && (
6359 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6360 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6361 {
6362 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6363 // but timings in PCSX are too tied to the interpreter's BIAS
6364 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6365 cycle_multiplier_active = 200;
6366
6367 *limit = (addr & 0xfff00000) | 0x80000;
6368 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6369 }
6370 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6371 *limit = (addr & 0x80600000) + 0x00200000;
6372 return (u_int *)(rdram + (addr&0x1fffff));
6373 }
6374 return NULL;
6375}
6376
6377static u_int scan_for_ret(u_int addr)
6378{
6379 u_int limit = 0;
6380 u_int *mem;
6381
6382 mem = get_source_start(addr, &limit);
6383 if (mem == NULL)
6384 return addr;
6385
6386 if (limit > addr + 0x1000)
6387 limit = addr + 0x1000;
6388 for (; addr < limit; addr += 4, mem++) {
6389 if (*mem == 0x03e00008) // jr $ra
6390 return addr + 8;
6391 }
6392 return addr;
6393}
6394
6395struct savestate_block {
6396 uint32_t addr;
6397 uint32_t regflags;
6398};
6399
6400static int addr_cmp(const void *p1_, const void *p2_)
6401{
6402 const struct savestate_block *p1 = p1_, *p2 = p2_;
6403 return p1->addr - p2->addr;
6404}
6405
6406int new_dynarec_save_blocks(void *save, int size)
6407{
6408 struct savestate_block *sblocks = save;
6409 int maxcount = size / sizeof(sblocks[0]);
6410 struct savestate_block tmp_blocks[1024];
6411 struct block_info *block;
6412 int p, s, d, o, bcnt;
6413 u_int addr;
6414
6415 o = 0;
6416 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6417 bcnt = 0;
6418 for (block = blocks[p]; block != NULL; block = block->next) {
6419 if (block->is_dirty)
6420 continue;
6421 tmp_blocks[bcnt].addr = block->start;
6422 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6423 bcnt++;
6424 }
6425 if (bcnt < 1)
6426 continue;
6427 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6428
6429 addr = tmp_blocks[0].addr;
6430 for (s = d = 0; s < bcnt; s++) {
6431 if (tmp_blocks[s].addr < addr)
6432 continue;
6433 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6434 tmp_blocks[d++] = tmp_blocks[s];
6435 addr = scan_for_ret(tmp_blocks[s].addr);
6436 }
6437
6438 if (o + d > maxcount)
6439 d = maxcount - o;
6440 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6441 o += d;
6442 }
6443
6444 return o * sizeof(sblocks[0]);
6445}
6446
6447void new_dynarec_load_blocks(const void *save, int size)
6448{
6449 const struct savestate_block *sblocks = save;
6450 int count = size / sizeof(sblocks[0]);
6451 struct block_info *block;
6452 u_int regs_save[32];
6453 u_int page;
6454 uint32_t f;
6455 int i, b;
6456
6457 // restore clean blocks, if any
6458 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6459 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6460 if (!block->is_dirty)
6461 continue;
6462 assert(block->source && block->copy);
6463 if (memcmp(block->source, block->copy, block->len))
6464 continue;
6465
6466 // see try_restore_block
6467 block->is_dirty = 0;
6468 mark_invalid_code(block->start, block->len, 0);
6469 i++;
6470 }
6471 }
6472 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6473
6474 // change GPRs for speculation to at least partially work..
6475 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6476 for (i = 1; i < 32; i++)
6477 psxRegs.GPR.r[i] = 0x80000000;
6478
6479 for (b = 0; b < count; b++) {
6480 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6481 if (f & 1)
6482 psxRegs.GPR.r[i] = 0x1f800000;
6483 }
6484
6485 ndrc_get_addr_ht(sblocks[b].addr);
6486
6487 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6488 if (f & 1)
6489 psxRegs.GPR.r[i] = 0x80000000;
6490 }
6491 }
6492
6493 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6494}
6495
6496void new_dynarec_print_stats(void)
6497{
6498#ifdef STAT_PRINT
6499 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6500 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6501 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6502 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6503 out - ndrc->translation_cache, stat_blocks, stat_links);
6504 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6505 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6506 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6507#endif
6508}
6509
6510static int apply_hacks(void)
6511{
6512 int i;
6513 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6514 return 0;
6515 /* special hack(s) */
6516 for (i = 0; i < slen - 4; i++)
6517 {
6518 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6519 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6520 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6521 && imm[i+3] == 0x6e40 && dops[i+3].rs1 == 2)
6522 {
6523 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6524 dops[i + 3].itype = NOP;
6525 }
6526 }
6527 i = slen;
6528 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6529 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6530 && dops[i-7].itype == STORE)
6531 {
6532 i = i-8;
6533 if (dops[i].itype == IMM16)
6534 i--;
6535 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6536 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6537 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6538 {
6539 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6540 f1_hack = 1;
6541 return 1;
6542 }
6543 }
6544 return 0;
6545}
6546
6547static noinline void pass1_disassemble(u_int pagelimit)
6548{
6549 int i, j, done = 0, ni_count = 0;
6550 unsigned int type,op,op2;
6551
6552 for (i = 0; !done; i++)
6553 {
6554 memset(&dops[i], 0, sizeof(dops[i]));
6555 op2=0;
6556 minimum_free_regs[i]=0;
6557 dops[i].opcode=op=source[i]>>26;
6558 switch(op)
6559 {
6560 case 0x00: set_mnemonic(i, "special"); type=NI;
6561 op2=source[i]&0x3f;
6562 switch(op2)
6563 {
6564 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6565 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6566 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6567 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6568 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6569 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6570 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6571 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6572 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6573 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6574 case 0x0F: set_mnemonic(i, "SYNC"); type=OTHER; break;
6575 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6576 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6577 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6578 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6579 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6580 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6581 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6582 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6583 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6584 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6585 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6586 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6587 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6588 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6589 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6590 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6591 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6592 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6593 case 0x30: set_mnemonic(i, "TGE"); type=NI; break;
6594 case 0x31: set_mnemonic(i, "TGEU"); type=NI; break;
6595 case 0x32: set_mnemonic(i, "TLT"); type=NI; break;
6596 case 0x33: set_mnemonic(i, "TLTU"); type=NI; break;
6597 case 0x34: set_mnemonic(i, "TEQ"); type=NI; break;
6598 case 0x36: set_mnemonic(i, "TNE"); type=NI; break;
6599#if 0
6600 case 0x14: set_mnemonic(i, "DSLLV"); type=SHIFT; break;
6601 case 0x16: set_mnemonic(i, "DSRLV"); type=SHIFT; break;
6602 case 0x17: set_mnemonic(i, "DSRAV"); type=SHIFT; break;
6603 case 0x1C: set_mnemonic(i, "DMULT"); type=MULTDIV; break;
6604 case 0x1D: set_mnemonic(i, "DMULTU"); type=MULTDIV; break;
6605 case 0x1E: set_mnemonic(i, "DDIV"); type=MULTDIV; break;
6606 case 0x1F: set_mnemonic(i, "DDIVU"); type=MULTDIV; break;
6607 case 0x2C: set_mnemonic(i, "DADD"); type=ALU; break;
6608 case 0x2D: set_mnemonic(i, "DADDU"); type=ALU; break;
6609 case 0x2E: set_mnemonic(i, "DSUB"); type=ALU; break;
6610 case 0x2F: set_mnemonic(i, "DSUBU"); type=ALU; break;
6611 case 0x38: set_mnemonic(i, "DSLL"); type=SHIFTIMM; break;
6612 case 0x3A: set_mnemonic(i, "DSRL"); type=SHIFTIMM; break;
6613 case 0x3B: set_mnemonic(i, "DSRA"); type=SHIFTIMM; break;
6614 case 0x3C: set_mnemonic(i, "DSLL32"); type=SHIFTIMM; break;
6615 case 0x3E: set_mnemonic(i, "DSRL32"); type=SHIFTIMM; break;
6616 case 0x3F: set_mnemonic(i, "DSRA32"); type=SHIFTIMM; break;
6617#endif
6618 }
6619 break;
6620 case 0x01: set_mnemonic(i, "regimm"); type=NI;
6621 op2=(source[i]>>16)&0x1f;
6622 switch(op2)
6623 {
6624 case 0x00: set_mnemonic(i, "BLTZ"); type=SJUMP; break;
6625 case 0x01: set_mnemonic(i, "BGEZ"); type=SJUMP; break;
6626 //case 0x02: set_mnemonic(i, "BLTZL"); type=SJUMP; break;
6627 //case 0x03: set_mnemonic(i, "BGEZL"); type=SJUMP; break;
6628 //case 0x08: set_mnemonic(i, "TGEI"); type=NI; break;
6629 //case 0x09: set_mnemonic(i, "TGEIU"); type=NI; break;
6630 //case 0x0A: set_mnemonic(i, "TLTI"); type=NI; break;
6631 //case 0x0B: set_mnemonic(i, "TLTIU"); type=NI; break;
6632 //case 0x0C: set_mnemonic(i, "TEQI"); type=NI; break;
6633 //case 0x0E: set_mnemonic(i, "TNEI"); type=NI; break;
6634 case 0x10: set_mnemonic(i, "BLTZAL"); type=SJUMP; break;
6635 case 0x11: set_mnemonic(i, "BGEZAL"); type=SJUMP; break;
6636 //case 0x12: set_mnemonic(i, "BLTZALL"); type=SJUMP; break;
6637 //case 0x13: set_mnemonic(i, "BGEZALL"); type=SJUMP; break;
6638 }
6639 break;
6640 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6641 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6642 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6643 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6644 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6645 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6646 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6647 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6648 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6649 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6650 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6651 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6652 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6653 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6654 case 0x10: set_mnemonic(i, "cop0"); type=NI;
6655 op2=(source[i]>>21)&0x1f;
6656 switch(op2)
6657 {
6658 case 0x00: set_mnemonic(i, "MFC0"); type=COP0; break;
6659 case 0x02: set_mnemonic(i, "CFC0"); type=COP0; break;
6660 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6661 case 0x06: set_mnemonic(i, "CTC0"); type=COP0; break;
6662 case 0x10: set_mnemonic(i, "RFE"); type=COP0; break;
6663 }
6664 break;
6665 case 0x11: set_mnemonic(i, "cop1"); type=COP1;
6666 op2=(source[i]>>21)&0x1f;
6667 break;
6668#if 0
6669 case 0x14: set_mnemonic(i, "BEQL"); type=CJUMP; break;
6670 case 0x15: set_mnemonic(i, "BNEL"); type=CJUMP; break;
6671 case 0x16: set_mnemonic(i, "BLEZL"); type=CJUMP; break;
6672 case 0x17: set_mnemonic(i, "BGTZL"); type=CJUMP; break;
6673 case 0x18: set_mnemonic(i, "DADDI"); type=IMM16; break;
6674 case 0x19: set_mnemonic(i, "DADDIU"); type=IMM16; break;
6675 case 0x1A: set_mnemonic(i, "LDL"); type=LOADLR; break;
6676 case 0x1B: set_mnemonic(i, "LDR"); type=LOADLR; break;
6677#endif
6678 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6679 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6680 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6681 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6682 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6683 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6684 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6685#if 0
6686 case 0x27: set_mnemonic(i, "LWU"); type=LOAD; break;
6687#endif
6688 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6689 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6690 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6691 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6692#if 0
6693 case 0x2C: set_mnemonic(i, "SDL"); type=STORELR; break;
6694 case 0x2D: set_mnemonic(i, "SDR"); type=STORELR; break;
6695#endif
6696 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6697 case 0x2F: set_mnemonic(i, "CACHE"); type=NOP; break;
6698 case 0x30: set_mnemonic(i, "LL"); type=NI; break;
6699 case 0x31: set_mnemonic(i, "LWC1"); type=C1LS; break;
6700#if 0
6701 case 0x34: set_mnemonic(i, "LLD"); type=NI; break;
6702 case 0x35: set_mnemonic(i, "LDC1"); type=C1LS; break;
6703 case 0x37: set_mnemonic(i, "LD"); type=LOAD; break;
6704#endif
6705 case 0x38: set_mnemonic(i, "SC"); type=NI; break;
6706 case 0x39: set_mnemonic(i, "SWC1"); type=C1LS; break;
6707#if 0
6708 case 0x3C: set_mnemonic(i, "SCD"); type=NI; break;
6709 case 0x3D: set_mnemonic(i, "SDC1"); type=C1LS; break;
6710 case 0x3F: set_mnemonic(i, "SD"); type=STORE; break;
6711#endif
6712 case 0x12: set_mnemonic(i, "COP2"); type=NI;
6713 op2=(source[i]>>21)&0x1f;
6714 //if (op2 & 0x10)
6715 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
6716 if (gte_handlers[source[i]&0x3f]!=NULL) {
6717#ifdef DISASM
6718 if (gte_regnames[source[i]&0x3f]!=NULL)
6719 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
6720 else
6721 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
6722#endif
6723 type=C2OP;
6724 }
6725 }
6726 else switch(op2)
6727 {
6728 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6729 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6730 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6731 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6732 }
6733 break;
6734 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6735 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6736 case 0x3B: set_mnemonic(i, "HLECALL"); type=HLECALL; break;
6737 default: set_mnemonic(i, "???"); type=NI;
6738 SysPrintf("NI %08x @%08x (%08x)\n", source[i], start + i*4, start);
6739 break;
6740 }
6741 dops[i].itype=type;
6742 dops[i].opcode2=op2;
6743 /* Get registers/immediates */
6744 dops[i].use_lt1=0;
6745 gte_rs[i]=gte_rt[i]=0;
6746 switch(type) {
6747 case LOAD:
6748 dops[i].rs1=(source[i]>>21)&0x1f;
6749 dops[i].rs2=0;
6750 dops[i].rt1=(source[i]>>16)&0x1f;
6751 dops[i].rt2=0;
6752 imm[i]=(short)source[i];
6753 break;
6754 case STORE:
6755 case STORELR:
6756 dops[i].rs1=(source[i]>>21)&0x1f;
6757 dops[i].rs2=(source[i]>>16)&0x1f;
6758 dops[i].rt1=0;
6759 dops[i].rt2=0;
6760 imm[i]=(short)source[i];
6761 break;
6762 case LOADLR:
6763 // LWL/LWR only load part of the register,
6764 // therefore the target register must be treated as a source too
6765 dops[i].rs1=(source[i]>>21)&0x1f;
6766 dops[i].rs2=(source[i]>>16)&0x1f;
6767 dops[i].rt1=(source[i]>>16)&0x1f;
6768 dops[i].rt2=0;
6769 imm[i]=(short)source[i];
6770 break;
6771 case IMM16:
6772 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6773 else dops[i].rs1=(source[i]>>21)&0x1f;
6774 dops[i].rs2=0;
6775 dops[i].rt1=(source[i]>>16)&0x1f;
6776 dops[i].rt2=0;
6777 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6778 imm[i]=(unsigned short)source[i];
6779 }else{
6780 imm[i]=(short)source[i];
6781 }
6782 break;
6783 case UJUMP:
6784 dops[i].rs1=0;
6785 dops[i].rs2=0;
6786 dops[i].rt1=0;
6787 dops[i].rt2=0;
6788 // The JAL instruction writes to r31.
6789 if (op&1) {
6790 dops[i].rt1=31;
6791 }
6792 dops[i].rs2=CCREG;
6793 break;
6794 case RJUMP:
6795 dops[i].rs1=(source[i]>>21)&0x1f;
6796 dops[i].rs2=0;
6797 dops[i].rt1=0;
6798 dops[i].rt2=0;
6799 // The JALR instruction writes to rd.
6800 if (op2&1) {
6801 dops[i].rt1=(source[i]>>11)&0x1f;
6802 }
6803 dops[i].rs2=CCREG;
6804 break;
6805 case CJUMP:
6806 dops[i].rs1=(source[i]>>21)&0x1f;
6807 dops[i].rs2=(source[i]>>16)&0x1f;
6808 dops[i].rt1=0;
6809 dops[i].rt2=0;
6810 if(op&2) { // BGTZ/BLEZ
6811 dops[i].rs2=0;
6812 }
6813 break;
6814 case SJUMP:
6815 dops[i].rs1=(source[i]>>21)&0x1f;
6816 dops[i].rs2=CCREG;
6817 dops[i].rt1=0;
6818 dops[i].rt2=0;
6819 if(op2&0x10) { // BxxAL
6820 dops[i].rt1=31;
6821 // NOTE: If the branch is not taken, r31 is still overwritten
6822 }
6823 break;
6824 case ALU:
6825 dops[i].rs1=(source[i]>>21)&0x1f; // source
6826 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
6827 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6828 dops[i].rt2=0;
6829 break;
6830 case MULTDIV:
6831 dops[i].rs1=(source[i]>>21)&0x1f; // source
6832 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
6833 dops[i].rt1=HIREG;
6834 dops[i].rt2=LOREG;
6835 break;
6836 case MOV:
6837 dops[i].rs1=0;
6838 dops[i].rs2=0;
6839 dops[i].rt1=0;
6840 dops[i].rt2=0;
6841 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6842 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6843 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6844 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6845 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
6846 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
6847 break;
6848 case SHIFT:
6849 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
6850 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
6851 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6852 dops[i].rt2=0;
6853 break;
6854 case SHIFTIMM:
6855 dops[i].rs1=(source[i]>>16)&0x1f;
6856 dops[i].rs2=0;
6857 dops[i].rt1=(source[i]>>11)&0x1f;
6858 dops[i].rt2=0;
6859 imm[i]=(source[i]>>6)&0x1f;
6860 // DSxx32 instructions
6861 if(op2>=0x3c) imm[i]|=0x20;
6862 break;
6863 case COP0:
6864 dops[i].rs1=0;
6865 dops[i].rs2=0;
6866 dops[i].rt1=0;
6867 dops[i].rt2=0;
6868 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
6869 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
6870 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
6871 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
6872 break;
6873 case COP1:
6874 dops[i].rs1=0;
6875 dops[i].rs2=0;
6876 dops[i].rt1=0;
6877 dops[i].rt2=0;
6878 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
6879 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
6880 dops[i].rs2=CSREG;
6881 break;
6882 case COP2:
6883 dops[i].rs1=0;
6884 dops[i].rs2=0;
6885 dops[i].rt1=0;
6886 dops[i].rt2=0;
6887 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
6888 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
6889 dops[i].rs2=CSREG;
6890 int gr=(source[i]>>11)&0x1F;
6891 switch(op2)
6892 {
6893 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6894 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6895 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6896 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6897 }
6898 break;
6899 case C1LS:
6900 dops[i].rs1=(source[i]>>21)&0x1F;
6901 dops[i].rs2=CSREG;
6902 dops[i].rt1=0;
6903 dops[i].rt2=0;
6904 imm[i]=(short)source[i];
6905 break;
6906 case C2LS:
6907 dops[i].rs1=(source[i]>>21)&0x1F;
6908 dops[i].rs2=0;
6909 dops[i].rt1=0;
6910 dops[i].rt2=0;
6911 imm[i]=(short)source[i];
6912 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
6913 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
6914 break;
6915 case C2OP:
6916 dops[i].rs1=0;
6917 dops[i].rs2=0;
6918 dops[i].rt1=0;
6919 dops[i].rt2=0;
6920 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
6921 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
6922 gte_rt[i]|=1ll<<63; // every op changes flags
6923 if((source[i]&0x3f)==GTE_MVMVA) {
6924 int v = (source[i] >> 15) & 3;
6925 gte_rs[i]&=~0xe3fll;
6926 if(v==3) gte_rs[i]|=0xe00ll;
6927 else gte_rs[i]|=3ll<<(v*2);
6928 }
6929 break;
6930 case SYSCALL:
6931 case HLECALL:
6932 case INTCALL:
6933 dops[i].rs1=CCREG;
6934 dops[i].rs2=0;
6935 dops[i].rt1=0;
6936 dops[i].rt2=0;
6937 break;
6938 default:
6939 dops[i].rs1=0;
6940 dops[i].rs2=0;
6941 dops[i].rt1=0;
6942 dops[i].rt2=0;
6943 }
6944 /* Calculate branch target addresses */
6945 if(type==UJUMP)
6946 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6947 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
6948 ba[i]=start+i*4+8; // Ignore never taken branch
6949 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
6950 ba[i]=start+i*4+8; // Ignore never taken branch
6951 else if(type==CJUMP||type==SJUMP)
6952 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6953 else ba[i]=-1;
6954
6955 /* simplify always (not)taken branches */
6956 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
6957 dops[i].rs1 = dops[i].rs2 = 0;
6958 if (!(op & 1)) {
6959 dops[i].itype = type = UJUMP;
6960 dops[i].rs2 = CCREG;
6961 }
6962 }
6963 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
6964 dops[i].itype = type = UJUMP;
6965
6966 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
6967 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
6968 dops[i].is_load = (dops[i].itype == LOAD || dops[i].itype == LOADLR || op == 0x32); // LWC2
6969 dops[i].is_store = (dops[i].itype == STORE || dops[i].itype == STORELR || op == 0x3a); // SWC2
6970
6971 /* messy cases to just pass over to the interpreter */
6972 if (i > 0 && dops[i-1].is_jump) {
6973 int do_in_intrp=0;
6974 // branch in delay slot?
6975 if (dops[i].is_jump) {
6976 // don't handle first branch and call interpreter if it's hit
6977 SysPrintf("branch in delay slot @%08x (%08x)\n", start + i*4, start);
6978 do_in_intrp=1;
6979 }
6980 // basic load delay detection
6981 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
6982 int t=(ba[i-1]-start)/4;
6983 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) {
6984 // jump target wants DS result - potential load delay effect
6985 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
6986 do_in_intrp=1;
6987 dops[t+1].bt=1; // expected return from interpreter
6988 }
6989 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&&
6990 !(i>=3&&dops[i-3].is_jump)) {
6991 // v0 overwrite like this is a sign of trouble, bail out
6992 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
6993 do_in_intrp=1;
6994 }
6995 }
6996 if (do_in_intrp) {
6997 memset(&dops[i-1], 0, sizeof(dops[i-1]));
6998 dops[i-1].itype = INTCALL;
6999 dops[i-1].rs1 = CCREG;
7000 ba[i-1] = -1;
7001 done = 2;
7002 i--; // don't compile the DS
7003 }
7004 }
7005
7006 /* Is this the end of the block? */
7007 if (i > 0 && dops[i-1].is_ujump) {
7008 if (dops[i-1].rt1 == 0) { // not jal
7009 int found_bbranch = 0, t = (ba[i-1] - start) / 4;
7010 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
7011 // scan for a branch back to i+1
7012 for (j = t; j < t + 64; j++) {
7013 int tmpop = source[j] >> 26;
7014 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
7015 int t2 = j + 1 + (int)(signed short)source[j];
7016 if (t2 == i + 1) {
7017 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
7018 found_bbranch = 1;
7019 break;
7020 }
7021 }
7022 }
7023 }
7024 if (!found_bbranch)
7025 done = 2;
7026 }
7027 else {
7028 if(stop_after_jal) done=1;
7029 // Stop on BREAK
7030 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7031 }
7032 // Don't recompile stuff that's already compiled
7033 if(check_addr(start+i*4+4)) done=1;
7034 // Don't get too close to the limit
7035 if(i>MAXBLOCK/2) done=1;
7036 }
7037 if (dops[i].itype == SYSCALL || dops[i].itype == HLECALL || dops[i].itype == INTCALL)
7038 done = stop_after_jal ? 1 : 2;
7039 if (done == 2) {
7040 // Does the block continue due to a branch?
7041 for(j=i-1;j>=0;j--)
7042 {
7043 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7044 if(ba[j]==start+i*4+4) done=j=0;
7045 if(ba[j]==start+i*4+8) done=j=0;
7046 }
7047 }
7048 //assert(i<MAXBLOCK-1);
7049 if(start+i*4==pagelimit-4) done=1;
7050 assert(start+i*4<pagelimit);
7051 if (i==MAXBLOCK-1) done=1;
7052 // Stop if we're compiling junk
7053 if(dops[i].itype == NI && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7054 done=stop_after_jal=1;
7055 SysPrintf("Disabled speculative precompilation\n");
7056 }
7057 }
7058 while (i > 0 && dops[i-1].is_jump)
7059 i--;
7060 assert(i > 0);
7061 assert(!dops[i-1].is_jump);
7062 slen = i;
7063}
7064
7065// Basic liveness analysis for MIPS registers
7066static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7067{
7068 int i;
7069 uint64_t u,gte_u,b,gte_b;
7070 uint64_t temp_u,temp_gte_u=0;
7071 uint64_t gte_u_unknown=0;
7072 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7073 gte_u_unknown=~0ll;
7074 if(iend==slen-1) {
7075 u=1;
7076 gte_u=gte_u_unknown;
7077 }else{
7078 //u=unneeded_reg[iend+1];
7079 u=1;
7080 gte_u=gte_unneeded[iend+1];
7081 }
7082
7083 for (i=iend;i>=istart;i--)
7084 {
7085 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7086 if(dops[i].is_jump)
7087 {
7088 // If subroutine call, flag return address as a possible branch target
7089 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7090
7091 if(ba[i]<start || ba[i]>=(start+slen*4))
7092 {
7093 // Branch out of this block, flush all regs
7094 u=1;
7095 gte_u=gte_u_unknown;
7096 branch_unneeded_reg[i]=u;
7097 // Merge in delay slot
7098 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7099 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7100 u|=1;
7101 gte_u|=gte_rt[i+1];
7102 gte_u&=~gte_rs[i+1];
7103 }
7104 else
7105 {
7106 // Internal branch, flag target
7107 dops[(ba[i]-start)>>2].bt=1;
7108 if(ba[i]<=start+i*4) {
7109 // Backward branch
7110 if(dops[i].is_ujump)
7111 {
7112 // Unconditional branch
7113 temp_u=1;
7114 temp_gte_u=0;
7115 } else {
7116 // Conditional branch (not taken case)
7117 temp_u=unneeded_reg[i+2];
7118 temp_gte_u&=gte_unneeded[i+2];
7119 }
7120 // Merge in delay slot
7121 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7122 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7123 temp_u|=1;
7124 temp_gte_u|=gte_rt[i+1];
7125 temp_gte_u&=~gte_rs[i+1];
7126 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7127 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7128 temp_u|=1;
7129 temp_gte_u|=gte_rt[i];
7130 temp_gte_u&=~gte_rs[i];
7131 unneeded_reg[i]=temp_u;
7132 gte_unneeded[i]=temp_gte_u;
7133 // Only go three levels deep. This recursion can take an
7134 // excessive amount of time if there are a lot of nested loops.
7135 if(r<2) {
7136 pass2_unneeded_regs((ba[i]-start)>>2,i-1,r+1);
7137 }else{
7138 unneeded_reg[(ba[i]-start)>>2]=1;
7139 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
7140 }
7141 } /*else*/ if(1) {
7142 if (dops[i].is_ujump)
7143 {
7144 // Unconditional branch
7145 u=unneeded_reg[(ba[i]-start)>>2];
7146 gte_u=gte_unneeded[(ba[i]-start)>>2];
7147 branch_unneeded_reg[i]=u;
7148 // Merge in delay slot
7149 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7150 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7151 u|=1;
7152 gte_u|=gte_rt[i+1];
7153 gte_u&=~gte_rs[i+1];
7154 } else {
7155 // Conditional branch
7156 b=unneeded_reg[(ba[i]-start)>>2];
7157 gte_b=gte_unneeded[(ba[i]-start)>>2];
7158 branch_unneeded_reg[i]=b;
7159 // Branch delay slot
7160 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7161 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7162 b|=1;
7163 gte_b|=gte_rt[i+1];
7164 gte_b&=~gte_rs[i+1];
7165 u&=b;
7166 gte_u&=gte_b;
7167 if(i<slen-1) {
7168 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7169 } else {
7170 branch_unneeded_reg[i]=1;
7171 }
7172 }
7173 }
7174 }
7175 }
7176 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
7177 {
7178 // SYSCALL instruction (software interrupt)
7179 u=1;
7180 }
7181 else if(dops[i].itype==COP0 && dops[i].opcode2==0x10)
7182 {
7183 // RFE
7184 u=1;
7185 }
7186 //u=1; // DEBUG
7187 // Written registers are unneeded
7188 u|=1LL<<dops[i].rt1;
7189 u|=1LL<<dops[i].rt2;
7190 gte_u|=gte_rt[i];
7191 // Accessed registers are needed
7192 u&=~(1LL<<dops[i].rs1);
7193 u&=~(1LL<<dops[i].rs2);
7194 gte_u&=~gte_rs[i];
7195 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7196 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7197 // Source-target dependencies
7198 // R0 is always unneeded
7199 u|=1;
7200 // Save it
7201 unneeded_reg[i]=u;
7202 gte_unneeded[i]=gte_u;
7203 /*
7204 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7205 printf("U:");
7206 int r;
7207 for(r=1;r<=CCREG;r++) {
7208 if((unneeded_reg[i]>>r)&1) {
7209 if(r==HIREG) printf(" HI");
7210 else if(r==LOREG) printf(" LO");
7211 else printf(" r%d",r);
7212 }
7213 }
7214 printf("\n");
7215 */
7216 }
7217}
7218
7219static noinline void pass3_register_alloc(u_int addr)
7220{
7221 struct regstat current; // Current register allocations/status
7222 clear_all_regs(current.regmap_entry);
7223 clear_all_regs(current.regmap);
7224 current.wasdirty = current.dirty = 0;
7225 current.u = unneeded_reg[0];
7226 alloc_reg(&current, 0, CCREG);
7227 dirty_reg(&current, CCREG);
7228 current.wasconst = 0;
7229 current.isconst = 0;
7230 current.loadedconst = 0;
7231 //current.waswritten = 0;
7232 int ds=0;
7233 int cc=0;
7234 int hr;
7235 int i, j;
7236
7237 if (addr & 1) {
7238 // First instruction is delay slot
7239 cc=-1;
7240 dops[1].bt=1;
7241 ds=1;
7242 unneeded_reg[0]=1;
7243 current.regmap[HOST_BTREG]=BTREG;
7244 }
7245
7246 for(i=0;i<slen;i++)
7247 {
7248 if(dops[i].bt)
7249 {
7250 for(hr=0;hr<HOST_REGS;hr++)
7251 {
7252 // Is this really necessary?
7253 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7254 }
7255 current.isconst=0;
7256 //current.waswritten=0;
7257 }
7258
7259 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7260 regs[i].wasconst=current.isconst;
7261 regs[i].wasdirty=current.dirty;
7262 regs[i].dirty=0;
7263 regs[i].u=0;
7264 regs[i].isconst=0;
7265 regs[i].loadedconst=0;
7266 if (!dops[i].is_jump) {
7267 if(i+1<slen) {
7268 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7269 current.u|=1;
7270 } else {
7271 current.u=1;
7272 }
7273 } else {
7274 if(i+1<slen) {
7275 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7276 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7277 current.u|=1;
7278 } else {
7279 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7280 abort();
7281 }
7282 }
7283 dops[i].is_ds=ds;
7284 if(ds) {
7285 ds=0; // Skip delay slot, already allocated as part of branch
7286 // ...but we need to alloc it in case something jumps here
7287 if(i+1<slen) {
7288 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7289 }else{
7290 current.u=branch_unneeded_reg[i-1];
7291 }
7292 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7293 current.u|=1;
7294 struct regstat temp;
7295 memcpy(&temp,&current,sizeof(current));
7296 temp.wasdirty=temp.dirty;
7297 // TODO: Take into account unconditional branches, as below
7298 delayslot_alloc(&temp,i);
7299 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7300 regs[i].wasdirty=temp.wasdirty;
7301 regs[i].dirty=temp.dirty;
7302 regs[i].isconst=0;
7303 regs[i].wasconst=0;
7304 current.isconst=0;
7305 // Create entry (branch target) regmap
7306 for(hr=0;hr<HOST_REGS;hr++)
7307 {
7308 int r=temp.regmap[hr];
7309 if(r>=0) {
7310 if(r!=regmap_pre[i][hr]) {
7311 regs[i].regmap_entry[hr]=-1;
7312 }
7313 else
7314 {
7315 assert(r < 64);
7316 if((current.u>>r)&1) {
7317 regs[i].regmap_entry[hr]=-1;
7318 regs[i].regmap[hr]=-1;
7319 //Don't clear regs in the delay slot as the branch might need them
7320 //current.regmap[hr]=-1;
7321 }else
7322 regs[i].regmap_entry[hr]=r;
7323 }
7324 } else {
7325 // First instruction expects CCREG to be allocated
7326 if(i==0&&hr==HOST_CCREG)
7327 regs[i].regmap_entry[hr]=CCREG;
7328 else
7329 regs[i].regmap_entry[hr]=-1;
7330 }
7331 }
7332 }
7333 else { // Not delay slot
7334 switch(dops[i].itype) {
7335 case UJUMP:
7336 //current.isconst=0; // DEBUG
7337 //current.wasconst=0; // DEBUG
7338 //regs[i].wasconst=0; // DEBUG
7339 clear_const(&current,dops[i].rt1);
7340 alloc_cc(&current,i);
7341 dirty_reg(&current,CCREG);
7342 if (dops[i].rt1==31) {
7343 alloc_reg(&current,i,31);
7344 dirty_reg(&current,31);
7345 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7346 //assert(dops[i+1].rt1!=dops[i].rt1);
7347 #ifdef REG_PREFETCH
7348 alloc_reg(&current,i,PTEMP);
7349 #endif
7350 }
7351 dops[i].ooo=1;
7352 delayslot_alloc(&current,i+1);
7353 //current.isconst=0; // DEBUG
7354 ds=1;
7355 //printf("i=%d, isconst=%x\n",i,current.isconst);
7356 break;
7357 case RJUMP:
7358 //current.isconst=0;
7359 //current.wasconst=0;
7360 //regs[i].wasconst=0;
7361 clear_const(&current,dops[i].rs1);
7362 clear_const(&current,dops[i].rt1);
7363 alloc_cc(&current,i);
7364 dirty_reg(&current,CCREG);
7365 if (!ds_writes_rjump_rs(i)) {
7366 alloc_reg(&current,i,dops[i].rs1);
7367 if (dops[i].rt1!=0) {
7368 alloc_reg(&current,i,dops[i].rt1);
7369 dirty_reg(&current,dops[i].rt1);
7370 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7371 assert(dops[i+1].rt1!=dops[i].rt1);
7372 #ifdef REG_PREFETCH
7373 alloc_reg(&current,i,PTEMP);
7374 #endif
7375 }
7376 #ifdef USE_MINI_HT
7377 if(dops[i].rs1==31) { // JALR
7378 alloc_reg(&current,i,RHASH);
7379 alloc_reg(&current,i,RHTBL);
7380 }
7381 #endif
7382 delayslot_alloc(&current,i+1);
7383 } else {
7384 // The delay slot overwrites our source register,
7385 // allocate a temporary register to hold the old value.
7386 current.isconst=0;
7387 current.wasconst=0;
7388 regs[i].wasconst=0;
7389 delayslot_alloc(&current,i+1);
7390 current.isconst=0;
7391 alloc_reg(&current,i,RTEMP);
7392 }
7393 //current.isconst=0; // DEBUG
7394 dops[i].ooo=1;
7395 ds=1;
7396 break;
7397 case CJUMP:
7398 //current.isconst=0;
7399 //current.wasconst=0;
7400 //regs[i].wasconst=0;
7401 clear_const(&current,dops[i].rs1);
7402 clear_const(&current,dops[i].rs2);
7403 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7404 {
7405 alloc_cc(&current,i);
7406 dirty_reg(&current,CCREG);
7407 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7408 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7409 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7410 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7411 // The delay slot overwrites one of our conditions.
7412 // Allocate the branch condition registers instead.
7413 current.isconst=0;
7414 current.wasconst=0;
7415 regs[i].wasconst=0;
7416 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7417 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7418 }
7419 else
7420 {
7421 dops[i].ooo=1;
7422 delayslot_alloc(&current,i+1);
7423 }
7424 }
7425 else
7426 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7427 {
7428 alloc_cc(&current,i);
7429 dirty_reg(&current,CCREG);
7430 alloc_reg(&current,i,dops[i].rs1);
7431 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7432 // The delay slot overwrites one of our conditions.
7433 // Allocate the branch condition registers instead.
7434 current.isconst=0;
7435 current.wasconst=0;
7436 regs[i].wasconst=0;
7437 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7438 }
7439 else
7440 {
7441 dops[i].ooo=1;
7442 delayslot_alloc(&current,i+1);
7443 }
7444 }
7445 else
7446 // Don't alloc the delay slot yet because we might not execute it
7447 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7448 {
7449 current.isconst=0;
7450 current.wasconst=0;
7451 regs[i].wasconst=0;
7452 alloc_cc(&current,i);
7453 dirty_reg(&current,CCREG);
7454 alloc_reg(&current,i,dops[i].rs1);
7455 alloc_reg(&current,i,dops[i].rs2);
7456 }
7457 else
7458 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7459 {
7460 current.isconst=0;
7461 current.wasconst=0;
7462 regs[i].wasconst=0;
7463 alloc_cc(&current,i);
7464 dirty_reg(&current,CCREG);
7465 alloc_reg(&current,i,dops[i].rs1);
7466 }
7467 ds=1;
7468 //current.isconst=0;
7469 break;
7470 case SJUMP:
7471 //current.isconst=0;
7472 //current.wasconst=0;
7473 //regs[i].wasconst=0;
7474 clear_const(&current,dops[i].rs1);
7475 clear_const(&current,dops[i].rt1);
7476 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7477 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7478 {
7479 alloc_cc(&current,i);
7480 dirty_reg(&current,CCREG);
7481 alloc_reg(&current,i,dops[i].rs1);
7482 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7483 alloc_reg(&current,i,31);
7484 dirty_reg(&current,31);
7485 //#ifdef REG_PREFETCH
7486 //alloc_reg(&current,i,PTEMP);
7487 //#endif
7488 }
7489 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.
7490 ||(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
7491 // Allocate the branch condition registers instead.
7492 current.isconst=0;
7493 current.wasconst=0;
7494 regs[i].wasconst=0;
7495 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7496 }
7497 else
7498 {
7499 dops[i].ooo=1;
7500 delayslot_alloc(&current,i+1);
7501 }
7502 }
7503 else
7504 // Don't alloc the delay slot yet because we might not execute it
7505 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7506 {
7507 current.isconst=0;
7508 current.wasconst=0;
7509 regs[i].wasconst=0;
7510 alloc_cc(&current,i);
7511 dirty_reg(&current,CCREG);
7512 alloc_reg(&current,i,dops[i].rs1);
7513 }
7514 ds=1;
7515 //current.isconst=0;
7516 break;
7517 case IMM16:
7518 imm16_alloc(&current,i);
7519 break;
7520 case LOAD:
7521 case LOADLR:
7522 load_alloc(&current,i);
7523 break;
7524 case STORE:
7525 case STORELR:
7526 store_alloc(&current,i);
7527 break;
7528 case ALU:
7529 alu_alloc(&current,i);
7530 break;
7531 case SHIFT:
7532 shift_alloc(&current,i);
7533 break;
7534 case MULTDIV:
7535 multdiv_alloc(&current,i);
7536 break;
7537 case SHIFTIMM:
7538 shiftimm_alloc(&current,i);
7539 break;
7540 case MOV:
7541 mov_alloc(&current,i);
7542 break;
7543 case COP0:
7544 cop0_alloc(&current,i);
7545 break;
7546 case COP1:
7547 break;
7548 case COP2:
7549 cop2_alloc(&current,i);
7550 break;
7551 case C1LS:
7552 c1ls_alloc(&current,i);
7553 break;
7554 case C2LS:
7555 c2ls_alloc(&current,i);
7556 break;
7557 case C2OP:
7558 c2op_alloc(&current,i);
7559 break;
7560 case SYSCALL:
7561 case HLECALL:
7562 case INTCALL:
7563 syscall_alloc(&current,i);
7564 break;
7565 }
7566
7567 // Create entry (branch target) regmap
7568 for(hr=0;hr<HOST_REGS;hr++)
7569 {
7570 int r,or;
7571 r=current.regmap[hr];
7572 if(r>=0) {
7573 if(r!=regmap_pre[i][hr]) {
7574 // TODO: delay slot (?)
7575 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7576 if(or<0||r>=TEMPREG){
7577 regs[i].regmap_entry[hr]=-1;
7578 }
7579 else
7580 {
7581 // Just move it to a different register
7582 regs[i].regmap_entry[hr]=r;
7583 // If it was dirty before, it's still dirty
7584 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7585 }
7586 }
7587 else
7588 {
7589 // Unneeded
7590 if(r==0){
7591 regs[i].regmap_entry[hr]=0;
7592 }
7593 else
7594 {
7595 assert(r<64);
7596 if((current.u>>r)&1) {
7597 regs[i].regmap_entry[hr]=-1;
7598 //regs[i].regmap[hr]=-1;
7599 current.regmap[hr]=-1;
7600 }else
7601 regs[i].regmap_entry[hr]=r;
7602 }
7603 }
7604 } else {
7605 // Branches expect CCREG to be allocated at the target
7606 if(regmap_pre[i][hr]==CCREG)
7607 regs[i].regmap_entry[hr]=CCREG;
7608 else
7609 regs[i].regmap_entry[hr]=-1;
7610 }
7611 }
7612 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7613 }
7614
7615#if 0 // see do_store_smc_check()
7616 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)
7617 current.waswritten|=1<<dops[i-1].rs1;
7618 current.waswritten&=~(1<<dops[i].rt1);
7619 current.waswritten&=~(1<<dops[i].rt2);
7620 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
7621 current.waswritten&=~(1<<dops[i].rs1);
7622#endif
7623
7624 /* Branch post-alloc */
7625 if(i>0)
7626 {
7627 current.wasdirty=current.dirty;
7628 switch(dops[i-1].itype) {
7629 case UJUMP:
7630 memcpy(&branch_regs[i-1],&current,sizeof(current));
7631 branch_regs[i-1].isconst=0;
7632 branch_regs[i-1].wasconst=0;
7633 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7634 alloc_cc(&branch_regs[i-1],i-1);
7635 dirty_reg(&branch_regs[i-1],CCREG);
7636 if(dops[i-1].rt1==31) { // JAL
7637 alloc_reg(&branch_regs[i-1],i-1,31);
7638 dirty_reg(&branch_regs[i-1],31);
7639 }
7640 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7641 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7642 break;
7643 case RJUMP:
7644 memcpy(&branch_regs[i-1],&current,sizeof(current));
7645 branch_regs[i-1].isconst=0;
7646 branch_regs[i-1].wasconst=0;
7647 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7648 alloc_cc(&branch_regs[i-1],i-1);
7649 dirty_reg(&branch_regs[i-1],CCREG);
7650 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7651 if(dops[i-1].rt1!=0) { // JALR
7652 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7653 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7654 }
7655 #ifdef USE_MINI_HT
7656 if(dops[i-1].rs1==31) { // JALR
7657 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7658 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7659 }
7660 #endif
7661 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7662 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7663 break;
7664 case CJUMP:
7665 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7666 {
7667 alloc_cc(&current,i-1);
7668 dirty_reg(&current,CCREG);
7669 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7670 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7671 // The delay slot overwrote one of our conditions
7672 // Delay slot goes after the test (in order)
7673 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7674 current.u|=1;
7675 delayslot_alloc(&current,i);
7676 current.isconst=0;
7677 }
7678 else
7679 {
7680 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7681 // Alloc the branch condition registers
7682 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7683 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7684 }
7685 memcpy(&branch_regs[i-1],&current,sizeof(current));
7686 branch_regs[i-1].isconst=0;
7687 branch_regs[i-1].wasconst=0;
7688 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7689 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7690 }
7691 else
7692 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7693 {
7694 alloc_cc(&current,i-1);
7695 dirty_reg(&current,CCREG);
7696 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7697 // The delay slot overwrote the branch condition
7698 // Delay slot goes after the test (in order)
7699 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7700 current.u|=1;
7701 delayslot_alloc(&current,i);
7702 current.isconst=0;
7703 }
7704 else
7705 {
7706 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7707 // Alloc the branch condition register
7708 alloc_reg(&current,i-1,dops[i-1].rs1);
7709 }
7710 memcpy(&branch_regs[i-1],&current,sizeof(current));
7711 branch_regs[i-1].isconst=0;
7712 branch_regs[i-1].wasconst=0;
7713 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7714 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7715 }
7716 else
7717 // Alloc the delay slot in case the branch is taken
7718 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
7719 {
7720 memcpy(&branch_regs[i-1],&current,sizeof(current));
7721 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;
7722 alloc_cc(&branch_regs[i-1],i);
7723 dirty_reg(&branch_regs[i-1],CCREG);
7724 delayslot_alloc(&branch_regs[i-1],i);
7725 branch_regs[i-1].isconst=0;
7726 alloc_reg(&current,i,CCREG); // Not taken path
7727 dirty_reg(&current,CCREG);
7728 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7729 }
7730 else
7731 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
7732 {
7733 memcpy(&branch_regs[i-1],&current,sizeof(current));
7734 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;
7735 alloc_cc(&branch_regs[i-1],i);
7736 dirty_reg(&branch_regs[i-1],CCREG);
7737 delayslot_alloc(&branch_regs[i-1],i);
7738 branch_regs[i-1].isconst=0;
7739 alloc_reg(&current,i,CCREG); // Not taken path
7740 dirty_reg(&current,CCREG);
7741 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7742 }
7743 break;
7744 case SJUMP:
7745 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
7746 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
7747 {
7748 alloc_cc(&current,i-1);
7749 dirty_reg(&current,CCREG);
7750 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7751 // The delay slot overwrote the branch condition
7752 // Delay slot goes after the test (in order)
7753 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7754 current.u|=1;
7755 delayslot_alloc(&current,i);
7756 current.isconst=0;
7757 }
7758 else
7759 {
7760 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7761 // Alloc the branch condition register
7762 alloc_reg(&current,i-1,dops[i-1].rs1);
7763 }
7764 memcpy(&branch_regs[i-1],&current,sizeof(current));
7765 branch_regs[i-1].isconst=0;
7766 branch_regs[i-1].wasconst=0;
7767 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7768 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7769 }
7770 else
7771 // Alloc the delay slot in case the branch is taken
7772 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
7773 {
7774 memcpy(&branch_regs[i-1],&current,sizeof(current));
7775 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;
7776 alloc_cc(&branch_regs[i-1],i);
7777 dirty_reg(&branch_regs[i-1],CCREG);
7778 delayslot_alloc(&branch_regs[i-1],i);
7779 branch_regs[i-1].isconst=0;
7780 alloc_reg(&current,i,CCREG); // Not taken path
7781 dirty_reg(&current,CCREG);
7782 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7783 }
7784 // FIXME: BLTZAL/BGEZAL
7785 if(dops[i-1].opcode2&0x10) { // BxxZAL
7786 alloc_reg(&branch_regs[i-1],i-1,31);
7787 dirty_reg(&branch_regs[i-1],31);
7788 }
7789 break;
7790 }
7791
7792 if (dops[i-1].is_ujump)
7793 {
7794 if(dops[i-1].rt1==31) // JAL/JALR
7795 {
7796 // Subroutine call will return here, don't alloc any registers
7797 current.dirty=0;
7798 clear_all_regs(current.regmap);
7799 alloc_reg(&current,i,CCREG);
7800 dirty_reg(&current,CCREG);
7801 }
7802 else if(i+1<slen)
7803 {
7804 // Internal branch will jump here, match registers to caller
7805 current.dirty=0;
7806 clear_all_regs(current.regmap);
7807 alloc_reg(&current,i,CCREG);
7808 dirty_reg(&current,CCREG);
7809 for(j=i-1;j>=0;j--)
7810 {
7811 if(ba[j]==start+i*4+4) {
7812 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7813 current.dirty=branch_regs[j].dirty;
7814 break;
7815 }
7816 }
7817 while(j>=0) {
7818 if(ba[j]==start+i*4+4) {
7819 for(hr=0;hr<HOST_REGS;hr++) {
7820 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7821 current.regmap[hr]=-1;
7822 }
7823 current.dirty&=branch_regs[j].dirty;
7824 }
7825 }
7826 j--;
7827 }
7828 }
7829 }
7830 }
7831
7832 // Count cycles in between branches
7833 ccadj[i] = CLOCK_ADJUST(cc);
7834 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
7835 {
7836 cc=0;
7837 }
7838#if !defined(DRC_DBG)
7839 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7840 {
7841 // this should really be removed since the real stalls have been implemented,
7842 // but doing so causes sizeable perf regression against the older version
7843 u_int gtec = gte_cycletab[source[i] & 0x3f];
7844 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7845 }
7846 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7847 {
7848 cc+=4;
7849 }
7850 else if(dops[i].itype==C2LS)
7851 {
7852 // same as with C2OP
7853 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7854 }
7855#endif
7856 else
7857 {
7858 cc++;
7859 }
7860
7861 if(!dops[i].is_ds) {
7862 regs[i].dirty=current.dirty;
7863 regs[i].isconst=current.isconst;
7864 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7865 }
7866 for(hr=0;hr<HOST_REGS;hr++) {
7867 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7868 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7869 regs[i].wasconst&=~(1<<hr);
7870 }
7871 }
7872 }
7873 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7874 //regs[i].waswritten=current.waswritten;
7875 }
7876}
7877
7878static noinline void pass4_cull_unused_regs(void)
7879{
7880 u_int last_needed_regs[4] = {0,0,0,0};
7881 u_int nr=0;
7882 int i;
7883
7884 for (i=slen-1;i>=0;i--)
7885 {
7886 int hr;
7887 __builtin_prefetch(regs[i-2].regmap);
7888 if(dops[i].is_jump)
7889 {
7890 if(ba[i]<start || ba[i]>=(start+slen*4))
7891 {
7892 // Branch out of this block, don't need anything
7893 nr=0;
7894 }
7895 else
7896 {
7897 // Internal branch
7898 // Need whatever matches the target
7899 nr=0;
7900 int t=(ba[i]-start)>>2;
7901 for(hr=0;hr<HOST_REGS;hr++)
7902 {
7903 if(regs[i].regmap_entry[hr]>=0) {
7904 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7905 }
7906 }
7907 }
7908 // Conditional branch may need registers for following instructions
7909 if (!dops[i].is_ujump)
7910 {
7911 if(i<slen-2) {
7912 nr |= last_needed_regs[(i+2) & 3];
7913 for(hr=0;hr<HOST_REGS;hr++)
7914 {
7915 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7916 //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]);
7917 }
7918 }
7919 }
7920 // Don't need stuff which is overwritten
7921 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7922 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7923 // Merge in delay slot
7924 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
7925 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
7926 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
7927 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
7928 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
7929 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
7930 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
7931 nr |= get_regm(regmap_pre[i], ROREG);
7932 nr |= get_regm(regs[i].regmap_entry, ROREG);
7933 }
7934 if (dops[i+1].is_store) {
7935 nr |= get_regm(regmap_pre[i], INVCP);
7936 nr |= get_regm(regs[i].regmap_entry, INVCP);
7937 }
7938 }
7939 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
7940 {
7941 // SYSCALL instruction (software interrupt)
7942 nr=0;
7943 }
7944 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
7945 {
7946 // ERET instruction (return from interrupt)
7947 nr=0;
7948 }
7949 else // Non-branch
7950 {
7951 if(i<slen-1) {
7952 for(hr=0;hr<HOST_REGS;hr++) {
7953 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7954 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7955 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7956 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7957 }
7958 }
7959 }
7960 // Overwritten registers are not needed
7961 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
7962 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
7963 nr &= ~get_regm(regs[i].regmap, FTEMP);
7964 // Source registers are needed
7965 nr |= get_regm(regmap_pre[i], dops[i].rs1);
7966 nr |= get_regm(regmap_pre[i], dops[i].rs2);
7967 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
7968 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
7969 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
7970 nr |= get_regm(regmap_pre[i], ROREG);
7971 nr |= get_regm(regs[i].regmap_entry, ROREG);
7972 }
7973 if (dops[i].is_store) {
7974 nr |= get_regm(regmap_pre[i], INVCP);
7975 nr |= get_regm(regs[i].regmap_entry, INVCP);
7976 }
7977
7978 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
7979 for(hr=0;hr<HOST_REGS;hr++)
7980 {
7981 // Don't store a register immediately after writing it,
7982 // may prevent dual-issue.
7983 // But do so if this is a branch target, otherwise we
7984 // might have to load the register before the branch.
7985 if((regs[i].wasdirty>>hr)&1) {
7986 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7987 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
7988 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
7989 }
7990 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7991 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
7992 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
7993 }
7994 }
7995 }
7996 // Cycle count is needed at branches. Assume it is needed at the target too.
7997 if(i==0||dops[i].bt||dops[i].itype==CJUMP) {
7998 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7999 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8000 }
8001 // Save it
8002 last_needed_regs[i & 3] = nr;
8003
8004 // Deallocate unneeded registers
8005 for(hr=0;hr<HOST_REGS;hr++)
8006 {
8007 if(!((nr>>hr)&1)) {
8008 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8009 if(dops[i].is_jump)
8010 {
8011 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
8012 if (dops[i+1].is_load || dops[i+1].is_store)
8013 map1 = ROREG;
8014 if (dops[i+1].is_store)
8015 map2 = INVCP;
8016 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
8017 temp = FTEMP;
8018 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8019 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8020 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
8021 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8022 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
8023 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8024 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8025 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
8026 {
8027 regs[i].regmap[hr]=-1;
8028 regs[i].isconst&=~(1<<hr);
8029 regs[i].dirty&=~(1<<hr);
8030 regs[i+1].wasdirty&=~(1<<hr);
8031 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
8032 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
8033 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
8034 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8035 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8036 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8037 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8038 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8039 {
8040 branch_regs[i].regmap[hr]=-1;
8041 branch_regs[i].regmap_entry[hr]=-1;
8042 if (!dops[i].is_ujump)
8043 {
8044 if (i < slen-2) {
8045 regmap_pre[i+2][hr]=-1;
8046 regs[i+2].wasconst&=~(1<<hr);
8047 }
8048 }
8049 }
8050 }
8051 }
8052 else
8053 {
8054 // Non-branch
8055 if(i>0)
8056 {
8057 int map1 = -1, map2 = -1, temp=-1;
8058 if (dops[i].is_load || dops[i].is_store)
8059 map1 = ROREG;
8060 if (dops[i].is_store)
8061 map2 = INVCP;
8062 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8063 temp = FTEMP;
8064 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8065 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8066 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8067 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8068 regs[i].regmap[hr] != CCREG)
8069 {
8070 if(i<slen-1&&!dops[i].is_ds) {
8071 assert(regs[i].regmap[hr]<64);
8072 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8073 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8074 {
8075 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8076 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8077 }
8078 regmap_pre[i+1][hr]=-1;
8079 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8080 regs[i+1].wasconst&=~(1<<hr);
8081 }
8082 regs[i].regmap[hr]=-1;
8083 regs[i].isconst&=~(1<<hr);
8084 regs[i].dirty&=~(1<<hr);
8085 regs[i+1].wasdirty&=~(1<<hr);
8086 }
8087 }
8088 }
8089 } // if needed
8090 } // for hr
8091 }
8092}
8093
8094// If a register is allocated during a loop, try to allocate it for the
8095// entire loop, if possible. This avoids loading/storing registers
8096// inside of the loop.
8097static noinline void pass5a_preallocate1(void)
8098{
8099 int i, j, hr;
8100 signed char f_regmap[HOST_REGS];
8101 clear_all_regs(f_regmap);
8102 for(i=0;i<slen-1;i++)
8103 {
8104 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8105 {
8106 if(ba[i]>=start && ba[i]<(start+i*4))
8107 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8108 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8109 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8110 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8111 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8112 {
8113 int t=(ba[i]-start)>>2;
8114 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8115 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8116 for(hr=0;hr<HOST_REGS;hr++)
8117 {
8118 if(regs[i].regmap[hr]>=0) {
8119 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8120 // dealloc old register
8121 int n;
8122 for(n=0;n<HOST_REGS;n++)
8123 {
8124 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8125 }
8126 // and alloc new one
8127 f_regmap[hr]=regs[i].regmap[hr];
8128 }
8129 }
8130 if(branch_regs[i].regmap[hr]>=0) {
8131 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8132 // dealloc old register
8133 int n;
8134 for(n=0;n<HOST_REGS;n++)
8135 {
8136 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8137 }
8138 // and alloc new one
8139 f_regmap[hr]=branch_regs[i].regmap[hr];
8140 }
8141 }
8142 if(dops[i].ooo) {
8143 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8144 f_regmap[hr]=branch_regs[i].regmap[hr];
8145 }else{
8146 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8147 f_regmap[hr]=branch_regs[i].regmap[hr];
8148 }
8149 // Avoid dirty->clean transition
8150 #ifdef DESTRUCTIVE_WRITEBACK
8151 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;
8152 #endif
8153 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8154 // case above, however it's always a good idea. We can't hoist the
8155 // load if the register was already allocated, so there's no point
8156 // wasting time analyzing most of these cases. It only "succeeds"
8157 // when the mapping was different and the load can be replaced with
8158 // a mov, which is of negligible benefit. So such cases are
8159 // skipped below.
8160 if(f_regmap[hr]>0) {
8161 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8162 int r=f_regmap[hr];
8163 for(j=t;j<=i;j++)
8164 {
8165 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8166 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8167 assert(r < 64);
8168 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8169 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8170 int k;
8171 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8172 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8173 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8174 k=i;
8175 while(k>1&&regs[k-1].regmap[hr]==-1) {
8176 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8177 //printf("no free regs for store %x\n",start+(k-1)*4);
8178 break;
8179 }
8180 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8181 //printf("no-match due to different register\n");
8182 break;
8183 }
8184 if (dops[k-2].is_jump) {
8185 //printf("no-match due to branch\n");
8186 break;
8187 }
8188 // call/ret fast path assumes no registers allocated
8189 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8190 break;
8191 }
8192 k--;
8193 }
8194 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8195 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8196 while(k<i) {
8197 regs[k].regmap_entry[hr]=f_regmap[hr];
8198 regs[k].regmap[hr]=f_regmap[hr];
8199 regmap_pre[k+1][hr]=f_regmap[hr];
8200 regs[k].wasdirty&=~(1<<hr);
8201 regs[k].dirty&=~(1<<hr);
8202 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8203 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8204 regs[k].wasconst&=~(1<<hr);
8205 regs[k].isconst&=~(1<<hr);
8206 k++;
8207 }
8208 }
8209 else {
8210 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8211 break;
8212 }
8213 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8214 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8215 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8216 regs[i].regmap_entry[hr]=f_regmap[hr];
8217 regs[i].regmap[hr]=f_regmap[hr];
8218 regs[i].wasdirty&=~(1<<hr);
8219 regs[i].dirty&=~(1<<hr);
8220 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8221 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8222 regs[i].wasconst&=~(1<<hr);
8223 regs[i].isconst&=~(1<<hr);
8224 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8225 branch_regs[i].wasdirty&=~(1<<hr);
8226 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8227 branch_regs[i].regmap[hr]=f_regmap[hr];
8228 branch_regs[i].dirty&=~(1<<hr);
8229 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8230 branch_regs[i].wasconst&=~(1<<hr);
8231 branch_regs[i].isconst&=~(1<<hr);
8232 if (!dops[i].is_ujump) {
8233 regmap_pre[i+2][hr]=f_regmap[hr];
8234 regs[i+2].wasdirty&=~(1<<hr);
8235 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8236 }
8237 }
8238 }
8239 for(k=t;k<j;k++) {
8240 // Alloc register clean at beginning of loop,
8241 // but may dirty it in pass 6
8242 regs[k].regmap_entry[hr]=f_regmap[hr];
8243 regs[k].regmap[hr]=f_regmap[hr];
8244 regs[k].dirty&=~(1<<hr);
8245 regs[k].wasconst&=~(1<<hr);
8246 regs[k].isconst&=~(1<<hr);
8247 if (dops[k].is_jump) {
8248 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8249 branch_regs[k].regmap[hr]=f_regmap[hr];
8250 branch_regs[k].dirty&=~(1<<hr);
8251 branch_regs[k].wasconst&=~(1<<hr);
8252 branch_regs[k].isconst&=~(1<<hr);
8253 if (!dops[k].is_ujump) {
8254 regmap_pre[k+2][hr]=f_regmap[hr];
8255 regs[k+2].wasdirty&=~(1<<hr);
8256 }
8257 }
8258 else
8259 {
8260 regmap_pre[k+1][hr]=f_regmap[hr];
8261 regs[k+1].wasdirty&=~(1<<hr);
8262 }
8263 }
8264 if(regs[j].regmap[hr]==f_regmap[hr])
8265 regs[j].regmap_entry[hr]=f_regmap[hr];
8266 break;
8267 }
8268 if(j==i) break;
8269 if(regs[j].regmap[hr]>=0)
8270 break;
8271 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8272 //printf("no-match due to different register\n");
8273 break;
8274 }
8275 if (dops[j].is_ujump)
8276 {
8277 // Stop on unconditional branch
8278 break;
8279 }
8280 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8281 {
8282 if(dops[j].ooo) {
8283 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8284 break;
8285 }else{
8286 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8287 break;
8288 }
8289 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8290 //printf("no-match due to different register (branch)\n");
8291 break;
8292 }
8293 }
8294 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8295 //printf("No free regs for store %x\n",start+j*4);
8296 break;
8297 }
8298 assert(f_regmap[hr]<64);
8299 }
8300 }
8301 }
8302 }
8303 }
8304 }else{
8305 // Non branch or undetermined branch target
8306 for(hr=0;hr<HOST_REGS;hr++)
8307 {
8308 if(hr!=EXCLUDE_REG) {
8309 if(regs[i].regmap[hr]>=0) {
8310 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8311 // dealloc old register
8312 int n;
8313 for(n=0;n<HOST_REGS;n++)
8314 {
8315 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8316 }
8317 // and alloc new one
8318 f_regmap[hr]=regs[i].regmap[hr];
8319 }
8320 }
8321 }
8322 }
8323 // Try to restore cycle count at branch targets
8324 if(dops[i].bt) {
8325 for(j=i;j<slen-1;j++) {
8326 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8327 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8328 //printf("no free regs for store %x\n",start+j*4);
8329 break;
8330 }
8331 }
8332 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8333 int k=i;
8334 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8335 while(k<j) {
8336 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8337 regs[k].regmap[HOST_CCREG]=CCREG;
8338 regmap_pre[k+1][HOST_CCREG]=CCREG;
8339 regs[k+1].wasdirty|=1<<HOST_CCREG;
8340 regs[k].dirty|=1<<HOST_CCREG;
8341 regs[k].wasconst&=~(1<<HOST_CCREG);
8342 regs[k].isconst&=~(1<<HOST_CCREG);
8343 k++;
8344 }
8345 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8346 }
8347 // Work backwards from the branch target
8348 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8349 {
8350 //printf("Extend backwards\n");
8351 int k;
8352 k=i;
8353 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8354 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8355 //printf("no free regs for store %x\n",start+(k-1)*4);
8356 break;
8357 }
8358 k--;
8359 }
8360 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8361 //printf("Extend CC, %x ->\n",start+k*4);
8362 while(k<=i) {
8363 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8364 regs[k].regmap[HOST_CCREG]=CCREG;
8365 regmap_pre[k+1][HOST_CCREG]=CCREG;
8366 regs[k+1].wasdirty|=1<<HOST_CCREG;
8367 regs[k].dirty|=1<<HOST_CCREG;
8368 regs[k].wasconst&=~(1<<HOST_CCREG);
8369 regs[k].isconst&=~(1<<HOST_CCREG);
8370 k++;
8371 }
8372 }
8373 else {
8374 //printf("Fail Extend CC, %x ->\n",start+k*4);
8375 }
8376 }
8377 }
8378 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8379 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8380 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8381 {
8382 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8383 }
8384 }
8385 }
8386}
8387
8388// This allocates registers (if possible) one instruction prior
8389// to use, which can avoid a load-use penalty on certain CPUs.
8390static noinline void pass5b_preallocate2(void)
8391{
8392 int i, hr;
8393 for(i=0;i<slen-1;i++)
8394 {
8395 if (!i || !dops[i-1].is_jump)
8396 {
8397 if(!dops[i+1].bt)
8398 {
8399 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8400 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8401 {
8402 if(dops[i+1].rs1) {
8403 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8404 {
8405 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8406 {
8407 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8408 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8409 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8410 regs[i].isconst&=~(1<<hr);
8411 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8412 constmap[i][hr]=constmap[i+1][hr];
8413 regs[i+1].wasdirty&=~(1<<hr);
8414 regs[i].dirty&=~(1<<hr);
8415 }
8416 }
8417 }
8418 if(dops[i+1].rs2) {
8419 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8420 {
8421 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8422 {
8423 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8424 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8425 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8426 regs[i].isconst&=~(1<<hr);
8427 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8428 constmap[i][hr]=constmap[i+1][hr];
8429 regs[i+1].wasdirty&=~(1<<hr);
8430 regs[i].dirty&=~(1<<hr);
8431 }
8432 }
8433 }
8434 // Preload target address for load instruction (non-constant)
8435 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8436 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8437 {
8438 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8439 {
8440 regs[i].regmap[hr]=dops[i+1].rs1;
8441 regmap_pre[i+1][hr]=dops[i+1].rs1;
8442 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8443 regs[i].isconst&=~(1<<hr);
8444 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8445 constmap[i][hr]=constmap[i+1][hr];
8446 regs[i+1].wasdirty&=~(1<<hr);
8447 regs[i].dirty&=~(1<<hr);
8448 }
8449 }
8450 }
8451 // Load source into target register
8452 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8453 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8454 {
8455 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8456 {
8457 regs[i].regmap[hr]=dops[i+1].rs1;
8458 regmap_pre[i+1][hr]=dops[i+1].rs1;
8459 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8460 regs[i].isconst&=~(1<<hr);
8461 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8462 constmap[i][hr]=constmap[i+1][hr];
8463 regs[i+1].wasdirty&=~(1<<hr);
8464 regs[i].dirty&=~(1<<hr);
8465 }
8466 }
8467 }
8468 // Address for store instruction (non-constant)
8469 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8470 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8471 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8472 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8473 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8474 else {
8475 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8476 regs[i+1].isconst&=~(1<<hr);
8477 }
8478 assert(hr>=0);
8479 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8480 {
8481 regs[i].regmap[hr]=dops[i+1].rs1;
8482 regmap_pre[i+1][hr]=dops[i+1].rs1;
8483 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8484 regs[i].isconst&=~(1<<hr);
8485 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8486 constmap[i][hr]=constmap[i+1][hr];
8487 regs[i+1].wasdirty&=~(1<<hr);
8488 regs[i].dirty&=~(1<<hr);
8489 }
8490 }
8491 }
8492 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8493 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8494 int nr;
8495 hr=get_reg(regs[i+1].regmap,FTEMP);
8496 assert(hr>=0);
8497 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8498 {
8499 regs[i].regmap[hr]=dops[i+1].rs1;
8500 regmap_pre[i+1][hr]=dops[i+1].rs1;
8501 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8502 regs[i].isconst&=~(1<<hr);
8503 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8504 constmap[i][hr]=constmap[i+1][hr];
8505 regs[i+1].wasdirty&=~(1<<hr);
8506 regs[i].dirty&=~(1<<hr);
8507 }
8508 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8509 {
8510 // move it to another register
8511 regs[i+1].regmap[hr]=-1;
8512 regmap_pre[i+2][hr]=-1;
8513 regs[i+1].regmap[nr]=FTEMP;
8514 regmap_pre[i+2][nr]=FTEMP;
8515 regs[i].regmap[nr]=dops[i+1].rs1;
8516 regmap_pre[i+1][nr]=dops[i+1].rs1;
8517 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8518 regs[i].isconst&=~(1<<nr);
8519 regs[i+1].isconst&=~(1<<nr);
8520 regs[i].dirty&=~(1<<nr);
8521 regs[i+1].wasdirty&=~(1<<nr);
8522 regs[i+1].dirty&=~(1<<nr);
8523 regs[i+2].wasdirty&=~(1<<nr);
8524 }
8525 }
8526 }
8527 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*/) {
8528 hr = -1;
8529 if(dops[i+1].itype==LOAD)
8530 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8531 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8532 hr=get_reg(regs[i+1].regmap,FTEMP);
8533 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
8534 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8535 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8536 }
8537 if(hr>=0&&regs[i].regmap[hr]<0) {
8538 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8539 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8540 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8541 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8542 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8543 regs[i].isconst&=~(1<<hr);
8544 regs[i+1].wasdirty&=~(1<<hr);
8545 regs[i].dirty&=~(1<<hr);
8546 }
8547 }
8548 }
8549 }
8550 }
8551 }
8552 }
8553}
8554
8555// Write back dirty registers as soon as we will no longer modify them,
8556// so that we don't end up with lots of writes at the branches.
8557static noinline void pass6_clean_registers(int istart, int iend, int wr)
8558{
8559 static u_int wont_dirty[MAXBLOCK];
8560 static u_int will_dirty[MAXBLOCK];
8561 int i;
8562 int r;
8563 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8564 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8565 if(iend==slen-1) {
8566 will_dirty_i=will_dirty_next=0;
8567 wont_dirty_i=wont_dirty_next=0;
8568 }else{
8569 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8570 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8571 }
8572 for (i=iend;i>=istart;i--)
8573 {
8574 signed char rregmap_i[RRMAP_SIZE];
8575 u_int hr_candirty = 0;
8576 assert(HOST_REGS < 32);
8577 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8578 __builtin_prefetch(regs[i-1].regmap);
8579 if(dops[i].is_jump)
8580 {
8581 signed char branch_rregmap_i[RRMAP_SIZE];
8582 u_int branch_hr_candirty = 0;
8583 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8584 if(ba[i]<start || ba[i]>=(start+slen*4))
8585 {
8586 // Branch out of this block, flush all regs
8587 will_dirty_i = 0;
8588 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8589 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8590 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8591 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8592 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8593 will_dirty_i &= branch_hr_candirty;
8594 if (dops[i].is_ujump)
8595 {
8596 // Unconditional branch
8597 wont_dirty_i = 0;
8598 // Merge in delay slot (will dirty)
8599 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8600 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8601 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8602 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8603 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8604 will_dirty_i &= hr_candirty;
8605 }
8606 else
8607 {
8608 // Conditional branch
8609 wont_dirty_i = wont_dirty_next;
8610 // Merge in delay slot (will dirty)
8611 // (the original code had no explanation why these 2 are commented out)
8612 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8613 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8614 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8615 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8616 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8617 will_dirty_i &= hr_candirty;
8618 }
8619 // Merge in delay slot (wont dirty)
8620 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8621 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8622 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8623 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8624 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8625 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8626 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8627 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8628 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8629 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8630 wont_dirty_i &= ~(1u << 31);
8631 if(wr) {
8632 #ifndef DESTRUCTIVE_WRITEBACK
8633 branch_regs[i].dirty&=wont_dirty_i;
8634 #endif
8635 branch_regs[i].dirty|=will_dirty_i;
8636 }
8637 }
8638 else
8639 {
8640 // Internal branch
8641 if(ba[i]<=start+i*4) {
8642 // Backward branch
8643 if (dops[i].is_ujump)
8644 {
8645 // Unconditional branch
8646 temp_will_dirty=0;
8647 temp_wont_dirty=0;
8648 // Merge in delay slot (will dirty)
8649 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8650 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8651 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8652 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8653 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8654 temp_will_dirty &= branch_hr_candirty;
8655 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8656 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8657 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8658 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8659 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8660 temp_will_dirty &= hr_candirty;
8661 } else {
8662 // Conditional branch (not taken case)
8663 temp_will_dirty=will_dirty_next;
8664 temp_wont_dirty=wont_dirty_next;
8665 // Merge in delay slot (will dirty)
8666 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8667 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8668 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8669 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8670 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8671 temp_will_dirty &= branch_hr_candirty;
8672 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8673 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8674 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8675 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8676 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8677 temp_will_dirty &= hr_candirty;
8678 }
8679 // Merge in delay slot (wont dirty)
8680 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8681 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8682 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8683 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8684 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8685 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8686 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8687 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8688 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8689 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8690 temp_wont_dirty &= ~(1u << 31);
8691 // Deal with changed mappings
8692 if(i<iend) {
8693 for(r=0;r<HOST_REGS;r++) {
8694 if(r!=EXCLUDE_REG) {
8695 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8696 temp_will_dirty&=~(1<<r);
8697 temp_wont_dirty&=~(1<<r);
8698 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8699 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8700 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8701 } else {
8702 temp_will_dirty|=1<<r;
8703 temp_wont_dirty|=1<<r;
8704 }
8705 }
8706 }
8707 }
8708 }
8709 if(wr) {
8710 will_dirty[i]=temp_will_dirty;
8711 wont_dirty[i]=temp_wont_dirty;
8712 pass6_clean_registers((ba[i]-start)>>2,i-1,0);
8713 }else{
8714 // Limit recursion. It can take an excessive amount
8715 // of time if there are a lot of nested loops.
8716 will_dirty[(ba[i]-start)>>2]=0;
8717 wont_dirty[(ba[i]-start)>>2]=-1;
8718 }
8719 }
8720 /*else*/ if(1)
8721 {
8722 if (dops[i].is_ujump)
8723 {
8724 // Unconditional branch
8725 will_dirty_i=0;
8726 wont_dirty_i=0;
8727 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
8728 for(r=0;r<HOST_REGS;r++) {
8729 if(r!=EXCLUDE_REG) {
8730 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
8731 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
8732 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
8733 }
8734 if(branch_regs[i].regmap[r]>=0) {
8735 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8736 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8737 }
8738 }
8739 }
8740 //}
8741 // Merge in delay slot
8742 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8743 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8744 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8745 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8746 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8747 will_dirty_i &= branch_hr_candirty;
8748 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8749 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8750 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8751 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8752 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8753 will_dirty_i &= hr_candirty;
8754 } else {
8755 // Conditional branch
8756 will_dirty_i=will_dirty_next;
8757 wont_dirty_i=wont_dirty_next;
8758 //if(ba[i]>start+i*4) // Disable recursion (for debugging)
8759 for(r=0;r<HOST_REGS;r++) {
8760 if(r!=EXCLUDE_REG) {
8761 signed char target_reg=branch_regs[i].regmap[r];
8762 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
8763 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
8764 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
8765 }
8766 else if(target_reg>=0) {
8767 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>target_reg)&1)<<r;
8768 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>target_reg)&1)<<r;
8769 }
8770 }
8771 }
8772 // Merge in delay slot
8773 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8774 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8775 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8776 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8777 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8778 will_dirty_i &= branch_hr_candirty;
8779 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8780 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8781 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8782 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8783 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8784 will_dirty_i &= hr_candirty;
8785 }
8786 // Merge in delay slot (won't dirty)
8787 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8788 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8789 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8790 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8791 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8792 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8793 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8794 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8795 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8796 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8797 wont_dirty_i &= ~(1u << 31);
8798 if(wr) {
8799 #ifndef DESTRUCTIVE_WRITEBACK
8800 branch_regs[i].dirty&=wont_dirty_i;
8801 #endif
8802 branch_regs[i].dirty|=will_dirty_i;
8803 }
8804 }
8805 }
8806 }
8807 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8808 {
8809 // SYSCALL instruction (software interrupt)
8810 will_dirty_i=0;
8811 wont_dirty_i=0;
8812 }
8813 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8814 {
8815 // ERET instruction (return from interrupt)
8816 will_dirty_i=0;
8817 wont_dirty_i=0;
8818 }
8819 will_dirty_next=will_dirty_i;
8820 wont_dirty_next=wont_dirty_i;
8821 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8822 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8823 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8824 will_dirty_i &= hr_candirty;
8825 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8826 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8827 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8828 wont_dirty_i &= ~(1u << 31);
8829 if (i > istart && !dops[i].is_jump) {
8830 // Don't store a register immediately after writing it,
8831 // may prevent dual-issue.
8832 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8833 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8834 }
8835 // Save it
8836 will_dirty[i]=will_dirty_i;
8837 wont_dirty[i]=wont_dirty_i;
8838 // Mark registers that won't be dirtied as not dirty
8839 if(wr) {
8840 regs[i].dirty|=will_dirty_i;
8841 #ifndef DESTRUCTIVE_WRITEBACK
8842 regs[i].dirty&=wont_dirty_i;
8843 if(dops[i].is_jump)
8844 {
8845 if (i < iend-1 && !dops[i].is_ujump) {
8846 for(r=0;r<HOST_REGS;r++) {
8847 if(r!=EXCLUDE_REG) {
8848 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8849 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8850 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8851 }
8852 }
8853 }
8854 }
8855 else
8856 {
8857 if(i<iend) {
8858 for(r=0;r<HOST_REGS;r++) {
8859 if(r!=EXCLUDE_REG) {
8860 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8861 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8862 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8863 }
8864 }
8865 }
8866 }
8867 #endif
8868 }
8869 // Deal with changed mappings
8870 temp_will_dirty=will_dirty_i;
8871 temp_wont_dirty=wont_dirty_i;
8872 for(r=0;r<HOST_REGS;r++) {
8873 if(r!=EXCLUDE_REG) {
8874 int nr;
8875 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8876 if(wr) {
8877 #ifndef DESTRUCTIVE_WRITEBACK
8878 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8879 #endif
8880 regs[i].wasdirty|=will_dirty_i&(1<<r);
8881 }
8882 }
8883 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8884 // Register moved to a different register
8885 will_dirty_i&=~(1<<r);
8886 wont_dirty_i&=~(1<<r);
8887 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8888 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8889 if(wr) {
8890 #ifndef DESTRUCTIVE_WRITEBACK
8891 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8892 #endif
8893 regs[i].wasdirty|=will_dirty_i&(1<<r);
8894 }
8895 }
8896 else {
8897 will_dirty_i&=~(1<<r);
8898 wont_dirty_i&=~(1<<r);
8899 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8900 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8901 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8902 } else {
8903 wont_dirty_i|=1<<r;
8904 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8905 }
8906 }
8907 }
8908 }
8909 }
8910}
8911
8912static noinline void pass10_expire_blocks(void)
8913{
8914 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
8915 // not sizeof(ndrc->translation_cache) due to vita hack
8916 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
8917 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
8918 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
8919 int hit;
8920
8921 for (; expirep != end; expirep = ((expirep + step) & step_mask))
8922 {
8923 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
8924 u_int block_i = expirep / step & (PAGE_COUNT - 1);
8925 u_int phase = (expirep >> (base_shift - 1)) & 1u;
8926 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
8927 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
8928 (long)(out - ndrc->translation_cache), phase);
8929 }
8930
8931 if (!phase) {
8932 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
8933 if (hit) {
8934 do_clear_cache();
8935 #ifdef USE_MINI_HT
8936 memset(mini_ht, -1, sizeof(mini_ht));
8937 #endif
8938 }
8939 }
8940 else
8941 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
8942 }
8943}
8944
8945static struct block_info *new_block_info(u_int start, u_int len,
8946 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
8947{
8948 struct block_info **b_pptr;
8949 struct block_info *block;
8950 u_int page = get_page(start);
8951
8952 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
8953 assert(block);
8954 assert(jump_in_count > 0);
8955 block->source = source;
8956 block->copy = copy;
8957 block->start = start;
8958 block->len = len;
8959 block->reg_sv_flags = 0;
8960 block->tc_offs = beginning - ndrc->translation_cache;
8961 //block->tc_len = out - beginning;
8962 block->is_dirty = 0;
8963 block->inv_near_misses = 0;
8964 block->jump_in_cnt = jump_in_count;
8965
8966 // insert sorted by start mirror-unmasked vaddr
8967 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
8968 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
8969 block->next = *b_pptr;
8970 *b_pptr = block;
8971 break;
8972 }
8973 }
8974 stat_inc(stat_blocks);
8975 return block;
8976}
8977
8978static int new_recompile_block(u_int addr)
8979{
8980 u_int pagelimit = 0;
8981 u_int state_rflags = 0;
8982 int i;
8983
8984 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
8985
8986 // this is just for speculation
8987 for (i = 1; i < 32; i++) {
8988 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
8989 state_rflags |= 1 << i;
8990 }
8991
8992 assert(!(addr & 3));
8993 start = addr & ~3;
8994 new_dynarec_did_compile=1;
8995 if (Config.HLE && start == 0x80001000) // hlecall
8996 {
8997 // XXX: is this enough? Maybe check hleSoftCall?
8998 void *beginning = start_block();
8999
9000 emit_movimm(start,0);
9001 emit_writeword(0,&pcaddr);
9002 emit_far_jump(new_dyna_leave);
9003 literal_pool(0);
9004 end_block(beginning);
9005 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9006 block->jump_in[0].vaddr = start;
9007 block->jump_in[0].addr = beginning;
9008 return 0;
9009 }
9010 else if (f1_hack && hack_addr == 0) {
9011 void *beginning = start_block();
9012 emit_movimm(start, 0);
9013 emit_writeword(0, &hack_addr);
9014 emit_readword(&psxRegs.GPR.n.sp, 0);
9015 emit_readptr(&mem_rtab, 1);
9016 emit_shrimm(0, 12, 2);
9017 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
9018 emit_addimm(0, 0x18, 0);
9019 emit_adds_ptr(1, 1, 1);
9020 emit_ldr_dualindexed(1, 0, 0);
9021 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
9022 emit_far_call(ndrc_get_addr_ht);
9023 emit_jmpreg(0); // jr k0
9024 literal_pool(0);
9025 end_block(beginning);
9026
9027 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9028 block->jump_in[0].vaddr = start;
9029 block->jump_in[0].addr = beginning;
9030 SysPrintf("F1 hack to %08x\n", start);
9031 return 0;
9032 }
9033
9034 cycle_multiplier_active = cycle_multiplier_override && cycle_multiplier == CYCLE_MULT_DEFAULT
9035 ? cycle_multiplier_override : cycle_multiplier;
9036
9037 source = get_source_start(start, &pagelimit);
9038 if (source == NULL) {
9039 if (addr != hack_addr) {
9040 SysPrintf("Compile at bogus memory address: %08x\n", addr);
9041 hack_addr = addr;
9042 }
9043 //abort();
9044 return -1;
9045 }
9046
9047 /* Pass 1: disassemble */
9048 /* Pass 2: register dependencies, branch targets */
9049 /* Pass 3: register allocation */
9050 /* Pass 4: branch dependencies */
9051 /* Pass 5: pre-alloc */
9052 /* Pass 6: optimize clean/dirty state */
9053 /* Pass 7: flag 32-bit registers */
9054 /* Pass 8: assembly */
9055 /* Pass 9: linker */
9056 /* Pass 10: garbage collection / free memory */
9057
9058 /* Pass 1 disassembly */
9059
9060 pass1_disassemble(pagelimit);
9061
9062 int clear_hack_addr = apply_hacks();
9063
9064 /* Pass 2 - Register dependencies and branch targets */
9065
9066 pass2_unneeded_regs(0,slen-1,0);
9067
9068 /* Pass 3 - Register allocation */
9069
9070 pass3_register_alloc(addr);
9071
9072 /* Pass 4 - Cull unused host registers */
9073
9074 pass4_cull_unused_regs();
9075
9076 /* Pass 5 - Pre-allocate registers */
9077
9078 pass5a_preallocate1();
9079 pass5b_preallocate2();
9080
9081 /* Pass 6 - Optimize clean/dirty state */
9082 pass6_clean_registers(0, slen-1, 1);
9083
9084 /* Pass 7 - Identify 32-bit registers */
9085 for (i=slen-1;i>=0;i--)
9086 {
9087 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9088 {
9089 // Conditional branch
9090 if((source[i]>>16)!=0x1000&&i<slen-2) {
9091 // Mark this address as a branch target since it may be called
9092 // upon return from interrupt
9093 dops[i+2].bt=1;
9094 }
9095 }
9096 }
9097
9098 /* Pass 8 - Assembly */
9099 linkcount=0;stubcount=0;
9100 is_delayslot=0;
9101 u_int dirty_pre=0;
9102 void *beginning=start_block();
9103 void *instr_addr0_override = NULL;
9104 int ds = 0;
9105
9106 if (start == 0x80030000) {
9107 // nasty hack for the fastbios thing
9108 // override block entry to this code
9109 instr_addr0_override = out;
9110 emit_movimm(start,0);
9111 // abuse io address var as a flag that we
9112 // have already returned here once
9113 emit_readword(&address,1);
9114 emit_writeword(0,&pcaddr);
9115 emit_writeword(0,&address);
9116 emit_cmp(0,1);
9117 #ifdef __aarch64__
9118 emit_jeq(out + 4*2);
9119 emit_far_jump(new_dyna_leave);
9120 #else
9121 emit_jne(new_dyna_leave);
9122 #endif
9123 }
9124 for(i=0;i<slen;i++)
9125 {
9126 __builtin_prefetch(regs[i+1].regmap);
9127 check_regmap(regmap_pre[i]);
9128 check_regmap(regs[i].regmap_entry);
9129 check_regmap(regs[i].regmap);
9130 //if(ds) printf("ds: ");
9131 disassemble_inst(i);
9132 if(ds) {
9133 ds=0; // Skip delay slot
9134 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9135 instr_addr[i] = NULL;
9136 } else {
9137 speculate_register_values(i);
9138 #ifndef DESTRUCTIVE_WRITEBACK
9139 if (i < 2 || !dops[i-2].is_ujump)
9140 {
9141 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9142 }
9143 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9144 dirty_pre=branch_regs[i].dirty;
9145 }else{
9146 dirty_pre=regs[i].dirty;
9147 }
9148 #endif
9149 // write back
9150 if (i < 2 || !dops[i-2].is_ujump)
9151 {
9152 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9153 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9154 }
9155 // branch target entry point
9156 instr_addr[i] = out;
9157 assem_debug("<->\n");
9158 drc_dbg_emit_do_cmp(i, ccadj[i]);
9159 if (clear_hack_addr) {
9160 emit_movimm(0, 0);
9161 emit_writeword(0, &hack_addr);
9162 clear_hack_addr = 0;
9163 }
9164
9165 // load regs
9166 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9167 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9168 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9169 address_generation(i,&regs[i],regs[i].regmap_entry);
9170 load_consts(regmap_pre[i],regs[i].regmap,i);
9171 if(dops[i].is_jump)
9172 {
9173 // Load the delay slot registers if necessary
9174 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))
9175 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9176 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))
9177 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9178 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9179 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9180 if (dops[i+1].is_store)
9181 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9182 }
9183 else if(i+1<slen)
9184 {
9185 // Preload registers for following instruction
9186 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9187 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9188 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9189 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9190 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9191 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9192 }
9193 // TODO: if(is_ooo(i)) address_generation(i+1);
9194 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9195 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9196 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9197 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9198 if (dops[i].is_store)
9199 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9200
9201 ds = assemble(i, &regs[i], ccadj[i]);
9202
9203 if (dops[i].is_ujump)
9204 literal_pool(1024);
9205 else
9206 literal_pool_jumpover(256);
9207 }
9208 }
9209
9210 assert(slen > 0);
9211 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9212 // no ending needed for this block since INTCALL never returns
9213 }
9214 // If the block did not end with an unconditional branch,
9215 // add a jump to the next instruction.
9216 else if (i > 1) {
9217 if (!dops[i-2].is_ujump) {
9218 assert(!dops[i-1].is_jump);
9219 assert(i==slen);
9220 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9221 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9222 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9223 emit_loadreg(CCREG,HOST_CCREG);
9224 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9225 }
9226 else
9227 {
9228 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9229 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9230 }
9231 add_to_linker(out,start+i*4,0);
9232 emit_jmp(0);
9233 }
9234 }
9235 else
9236 {
9237 assert(i>0);
9238 assert(!dops[i-1].is_jump);
9239 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9240 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9241 emit_loadreg(CCREG,HOST_CCREG);
9242 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9243 add_to_linker(out,start+i*4,0);
9244 emit_jmp(0);
9245 }
9246
9247 // TODO: delay slot stubs?
9248 // Stubs
9249 for(i=0;i<stubcount;i++)
9250 {
9251 switch(stubs[i].type)
9252 {
9253 case LOADB_STUB:
9254 case LOADH_STUB:
9255 case LOADW_STUB:
9256 case LOADD_STUB:
9257 case LOADBU_STUB:
9258 case LOADHU_STUB:
9259 do_readstub(i);break;
9260 case STOREB_STUB:
9261 case STOREH_STUB:
9262 case STOREW_STUB:
9263 case STORED_STUB:
9264 do_writestub(i);break;
9265 case CC_STUB:
9266 do_ccstub(i);break;
9267 case INVCODE_STUB:
9268 do_invstub(i);break;
9269 case FP_STUB:
9270 do_cop1stub(i);break;
9271 case STORELR_STUB:
9272 do_unalignedwritestub(i);break;
9273 }
9274 }
9275
9276 if (instr_addr0_override)
9277 instr_addr[0] = instr_addr0_override;
9278
9279#if 0
9280 /* check for improper expiration */
9281 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9282 int j;
9283 if (!jumps[i])
9284 continue;
9285 for (j = 0; j < jumps[i]->count; j++)
9286 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9287 }
9288#endif
9289
9290 /* Pass 9 - Linker */
9291 for(i=0;i<linkcount;i++)
9292 {
9293 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9294 literal_pool(64);
9295 if (!link_addr[i].internal)
9296 {
9297 void *stub = out;
9298 void *addr = check_addr(link_addr[i].target);
9299 emit_extjump(link_addr[i].addr, link_addr[i].target);
9300 if (addr) {
9301 set_jump_target(link_addr[i].addr, addr);
9302 ndrc_add_jump_out(link_addr[i].target,stub);
9303 }
9304 else
9305 set_jump_target(link_addr[i].addr, stub);
9306 }
9307 else
9308 {
9309 // Internal branch
9310 int target=(link_addr[i].target-start)>>2;
9311 assert(target>=0&&target<slen);
9312 assert(instr_addr[target]);
9313 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9314 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9315 //#else
9316 set_jump_target(link_addr[i].addr, instr_addr[target]);
9317 //#endif
9318 }
9319 }
9320
9321 u_int source_len = slen*4;
9322 if (dops[slen-1].itype == INTCALL && source_len > 4)
9323 // no need to treat the last instruction as compiled
9324 // as interpreter fully handles it
9325 source_len -= 4;
9326
9327 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9328 copy = shadow;
9329
9330 // External Branch Targets (jump_in)
9331 int jump_in_count = 1;
9332 assert(instr_addr[0]);
9333 for (i = 1; i < slen; i++)
9334 {
9335 if (dops[i].bt && instr_addr[i])
9336 jump_in_count++;
9337 }
9338
9339 struct block_info *block =
9340 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9341 block->reg_sv_flags = state_rflags;
9342
9343 int jump_in_i = 0;
9344 for (i = 0; i < slen; i++)
9345 {
9346 if ((i == 0 || dops[i].bt) && instr_addr[i])
9347 {
9348 assem_debug("%p (%d) <- %8x\n", instr_addr[i], i, start + i*4);
9349 u_int vaddr = start + i*4;
9350
9351 literal_pool(256);
9352 void *entry = out;
9353 load_regs_entry(i);
9354 if (entry == out)
9355 entry = instr_addr[i];
9356 else
9357 emit_jmp(instr_addr[i]);
9358
9359 block->jump_in[jump_in_i].vaddr = vaddr;
9360 block->jump_in[jump_in_i].addr = entry;
9361 jump_in_i++;
9362 }
9363 }
9364 assert(jump_in_i == jump_in_count);
9365 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9366 // Write out the literal pool if necessary
9367 literal_pool(0);
9368 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9369 // Align code
9370 if(((u_int)out)&7) emit_addnop(13);
9371 #endif
9372 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9373 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9374 memcpy(copy, source, source_len);
9375 copy += source_len;
9376
9377 end_block(beginning);
9378
9379 // If we're within 256K of the end of the buffer,
9380 // start over from the beginning. (Is 256K enough?)
9381 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9382 out = ndrc->translation_cache;
9383
9384 // Trap writes to any of the pages we compiled
9385 mark_invalid_code(start, slen*4, 0);
9386
9387 /* Pass 10 - Free memory by expiring oldest blocks */
9388
9389 pass10_expire_blocks();
9390
9391#ifdef ASSEM_PRINT
9392 fflush(stdout);
9393#endif
9394 stat_inc(stat_bc_direct);
9395 return 0;
9396}
9397
9398// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork