notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
dma: don't copy out of range
[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 "linkage_offsets.h"
43#include "compiler_features.h"
44#include "arm_features.h"
45
46#ifndef ARRAY_SIZE
47#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
48#endif
49#ifndef min
50#define min(a, b) ((b) < (a) ? (b) : (a))
51#endif
52#ifndef max
53#define max(a, b) ((b) > (a) ? (b) : (a))
54#endif
55
56//#define DISASM
57//#define ASSEM_PRINT
58//#define REGMAP_PRINT // with DISASM only
59//#define INV_DEBUG_W
60//#define STAT_PRINT
61
62#ifdef ASSEM_PRINT
63#define assem_debug printf
64#else
65#define assem_debug(...)
66#endif
67//#define inv_debug printf
68#define inv_debug(...)
69
70#ifdef __i386__
71#include "assem_x86.h"
72#endif
73#ifdef __x86_64__
74#include "assem_x64.h"
75#endif
76#ifdef __arm__
77#include "assem_arm.h"
78#endif
79#ifdef __aarch64__
80#include "assem_arm64.h"
81#endif
82
83#define RAM_SIZE 0x200000
84#define MAXBLOCK 4096
85#define MAX_OUTPUT_BLOCK_SIZE 262144
86#define EXPIRITY_OFFSET (MAX_OUTPUT_BLOCK_SIZE * 2)
87#define PAGE_COUNT 1024
88
89#if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
90#define INVALIDATE_USE_COND_CALL
91#endif
92
93#ifdef VITA
94// apparently Vita has a 16MB limit, so either we cut tc in half,
95// or use this hack (it's a hack because tc size was designed to be power-of-2)
96#define TC_REDUCE_BYTES 4096
97#else
98#define TC_REDUCE_BYTES 0
99#endif
100
101struct ndrc_tramp
102{
103 struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
104 const void *f[2048 / sizeof(void *)];
105};
106
107struct ndrc_mem
108{
109 u_char translation_cache[(1 << TARGET_SIZE_2) - TC_REDUCE_BYTES];
110 struct ndrc_tramp tramp;
111};
112
113#ifdef BASE_ADDR_DYNAMIC
114static struct ndrc_mem *ndrc;
115#else
116static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
117static struct ndrc_mem *ndrc = &ndrc_;
118#endif
119#ifdef TC_WRITE_OFFSET
120# ifdef __GLIBC__
121# include <sys/types.h>
122# include <sys/stat.h>
123# include <fcntl.h>
124# include <unistd.h>
125# endif
126static long ndrc_write_ofs;
127#define NDRC_WRITE_OFFSET(x) (void *)((char *)(x) + ndrc_write_ofs)
128#else
129#define NDRC_WRITE_OFFSET(x) (x)
130#endif
131
132// stubs
133enum stub_type {
134 CC_STUB = 1,
135 //FP_STUB = 2,
136 LOADB_STUB = 3,
137 LOADH_STUB = 4,
138 LOADW_STUB = 5,
139 //LOADD_STUB = 6,
140 LOADBU_STUB = 7,
141 LOADHU_STUB = 8,
142 STOREB_STUB = 9,
143 STOREH_STUB = 10,
144 STOREW_STUB = 11,
145 //STORED_STUB = 12,
146 STORELR_STUB = 13,
147 INVCODE_STUB = 14,
148 OVERFLOW_STUB = 15,
149 ALIGNMENT_STUB = 16,
150};
151
152// regmap_pre[i] - regs before [i] insn starts; dirty things here that
153// don't match .regmap will be written back
154// [i].regmap_entry - regs that must be set up if someone jumps here
155// [i].regmap - regs [i] insn will read/(over)write
156// branch_regs[i].* - same as above but for branches, takes delay slot into account
157struct regstat
158{
159 signed char regmap_entry[HOST_REGS];
160 signed char regmap[HOST_REGS];
161 uint64_t wasdirty;
162 uint64_t dirty;
163 uint64_t u;
164 u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
165 u_int isconst; // ... but isconst is false when r2 is known (hr)
166 u_int loadedconst; // host regs that have constants loaded
167 u_int noevict; // can't evict this hr (alloced by current op)
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; // bits 31-26
229 u_char opcode2; // (depends on opcode)
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 u_char is_delay_load:1; // is_load + MFC/CFC
243 u_char is_exception:1; // unconditional, also interp. fallback
244 u_char may_except:1; // might generate an exception
245} dops[MAXBLOCK];
246
247static struct compile_info
248{
249 int imm;
250 u_int ba;
251 int ccadj;
252 signed char min_free_regs;
253 signed char addr;
254 signed char reserved[2];
255} cinfo[MAXBLOCK];
256
257 static u_char *out;
258 static char invalid_code[0x100000];
259 static struct ht_entry hash_table[65536];
260 static struct block_info *blocks[PAGE_COUNT];
261 static struct jump_info *jumps[PAGE_COUNT];
262 static u_int start;
263 static u_int *source;
264 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
265 static uint64_t gte_rt[MAXBLOCK];
266 static uint64_t gte_unneeded[MAXBLOCK];
267 static u_int smrv[32]; // speculated MIPS register values
268 static u_int smrv_strong; // mask or regs that are likely to have correct values
269 static u_int smrv_weak; // same, but somewhat less likely
270 static u_int smrv_strong_next; // same, but after current insn executes
271 static u_int smrv_weak_next;
272 static uint64_t unneeded_reg[MAXBLOCK];
273 static uint64_t branch_unneeded_reg[MAXBLOCK];
274 // see 'struct regstat' for a description
275 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
276 // contains 'real' consts at [i] insn, but may differ from what's actually
277 // loaded in host reg as 'final' value is always loaded, see get_final_value()
278 static uint32_t current_constmap[HOST_REGS];
279 static uint32_t constmap[MAXBLOCK][HOST_REGS];
280 static struct regstat regs[MAXBLOCK];
281 static struct regstat branch_regs[MAXBLOCK];
282 static int slen;
283 static void *instr_addr[MAXBLOCK];
284 static struct link_entry link_addr[MAXBLOCK];
285 static int linkcount;
286 static struct code_stub stubs[MAXBLOCK*3];
287 static int stubcount;
288 static u_int literals[1024][2];
289 static int literalcount;
290 static int is_delayslot;
291 static char shadow[1048576] __attribute__((aligned(16)));
292 static void *copy;
293 static u_int expirep;
294 static u_int stop_after_jal;
295 static u_int f1_hack;
296#ifdef STAT_PRINT
297 static int stat_bc_direct;
298 static int stat_bc_pre;
299 static int stat_bc_restore;
300 static int stat_ht_lookups;
301 static int stat_jump_in_lookups;
302 static int stat_restore_tries;
303 static int stat_restore_compares;
304 static int stat_inv_addr_calls;
305 static int stat_inv_hits;
306 static int stat_blocks;
307 static int stat_links;
308 #define stat_inc(s) s++
309 #define stat_dec(s) s--
310 #define stat_clear(s) s = 0
311#else
312 #define stat_inc(s)
313 #define stat_dec(s)
314 #define stat_clear(s)
315#endif
316
317 int new_dynarec_hacks;
318 int new_dynarec_hacks_pergame;
319 int new_dynarec_hacks_old;
320 int new_dynarec_did_compile;
321
322 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
323
324 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0 (CCREG)
325 extern int last_count; // last absolute target, often = next_interupt
326 extern int pcaddr;
327 extern int pending_exception;
328 extern int branch_target;
329 extern uintptr_t ram_offset;
330 extern uintptr_t mini_ht[32][2];
331
332 /* registers that may be allocated */
333 /* 1-31 gpr */
334#define LOREG 32 // lo
335#define HIREG 33 // hi
336//#define FSREG 34 // FPU status (FCSR)
337//#define CSREG 35 // Coprocessor status
338#define CCREG 36 // Cycle count
339#define INVCP 37 // Pointer to invalid_code
340//#define MMREG 38 // Pointer to memory_map
341#define ROREG 39 // ram offset (if rdram!=0x80000000)
342#define TEMPREG 40
343#define FTEMP 40 // FPU temporary register
344#define PTEMP 41 // Prefetch temporary register
345//#define TLREG 42 // TLB mapping offset
346#define RHASH 43 // Return address hash
347#define RHTBL 44 // Return address hash table address
348#define RTEMP 45 // JR/JALR address register
349#define MAXREG 45
350#define AGEN1 46 // Address generation temporary register (pass5b_preallocate2)
351//#define AGEN2 47 // Address generation temporary register
352#define BTREG 50 // Branch target temporary register
353
354 /* instruction types */
355#define NOP 0 // No operation
356#define LOAD 1 // Load
357#define STORE 2 // Store
358#define LOADLR 3 // Unaligned load
359#define STORELR 4 // Unaligned store
360#define MOV 5 // Move (hi/lo only)
361#define ALU 6 // Arithmetic/logic
362#define MULTDIV 7 // Multiply/divide
363#define SHIFT 8 // Shift by register
364#define SHIFTIMM 9// Shift by immediate
365#define IMM16 10 // 16-bit immediate
366#define RJUMP 11 // Unconditional jump to register
367#define UJUMP 12 // Unconditional jump
368#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
369#define SJUMP 14 // Conditional branch (regimm format)
370#define COP0 15 // Coprocessor 0
371#define RFE 16
372#define SYSCALL 22// SYSCALL,BREAK
373#define OTHER 23 // Other/unknown - do nothing
374#define HLECALL 26// PCSX fake opcodes for HLE
375#define COP2 27 // Coprocessor 2 move
376#define C2LS 28 // Coprocessor 2 load/store
377#define C2OP 29 // Coprocessor 2 operation
378#define INTCALL 30// Call interpreter to handle rare corner cases
379
380 /* branch codes */
381#define TAKEN 1
382#define NOTTAKEN 2
383#define NULLDS 3
384
385#define DJT_1 (void *)1l // no function, just a label in assem_debug log
386#define DJT_2 (void *)2l
387
388// asm linkage
389void dyna_linker();
390void cc_interrupt();
391void jump_syscall (u_int u0, u_int u1, u_int pc);
392void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
393void jump_break (u_int u0, u_int u1, u_int pc);
394void jump_break_ds(u_int u0, u_int u1, u_int pc);
395void jump_overflow (u_int u0, u_int u1, u_int pc);
396void jump_overflow_ds(u_int u0, u_int u1, u_int pc);
397void jump_addrerror (u_int cause, u_int addr, u_int pc);
398void jump_addrerror_ds(u_int cause, u_int addr, u_int pc);
399void jump_to_new_pc();
400void call_gteStall();
401void new_dyna_leave();
402
403void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile);
404void *ndrc_get_addr_ht(u_int vaddr);
405void ndrc_add_jump_out(u_int vaddr, void *src);
406void ndrc_write_invalidate_one(u_int addr);
407static void ndrc_write_invalidate_many(u_int addr, u_int end);
408
409static int new_recompile_block(u_int addr);
410static void invalidate_block(struct block_info *block);
411static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_);
412
413// Needed by assembler
414static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
415static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
416static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
417static void load_all_regs(const signed char i_regmap[]);
418static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
419static void load_regs_entry(int t);
420static void load_all_consts(const signed char regmap[], u_int dirty, int i);
421static u_int get_host_reglist(const signed char *regmap);
422
423static int get_final_value(int hr, int i, u_int *value);
424static void add_stub(enum stub_type type, void *addr, void *retaddr,
425 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
426static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
427 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
428static void add_to_linker(void *addr, u_int target, int ext);
429static void *get_direct_memhandler(void *table, u_int addr,
430 enum stub_type type, uintptr_t *addr_host);
431static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
432static void pass_args(int a0, int a1);
433static void emit_far_jump(const void *f);
434static void emit_far_call(const void *f);
435
436#ifdef VITA
437#include <psp2/kernel/sysmem.h>
438static int sceBlock;
439// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
440extern int getVMBlock();
441int _newlib_vm_size_user = sizeof(*ndrc);
442#endif
443
444static void mprotect_w_x(void *start, void *end, int is_x)
445{
446#ifdef NO_WRITE_EXEC
447 #if defined(VITA)
448 // *Open* enables write on all memory that was
449 // allocated by sceKernelAllocMemBlockForVM()?
450 if (is_x)
451 sceKernelCloseVMDomain();
452 else
453 sceKernelOpenVMDomain();
454 #elif defined(HAVE_LIBNX)
455 Result rc;
456 // check to avoid the full flush in jitTransitionToExecutable()
457 if (g_jit.type != JitType_CodeMemory) {
458 if (is_x)
459 rc = jitTransitionToExecutable(&g_jit);
460 else
461 rc = jitTransitionToWritable(&g_jit);
462 if (R_FAILED(rc))
463 ;//SysPrintf("jitTransition %d %08x\n", is_x, rc);
464 }
465 #elif defined(TC_WRITE_OFFSET)
466 // separated rx and rw areas are always available
467 #else
468 u_long mstart = (u_long)start & ~4095ul;
469 u_long mend = (u_long)end;
470 if (mprotect((void *)mstart, mend - mstart,
471 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
472 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
473 #endif
474#endif
475}
476
477static void start_tcache_write(void *start, void *end)
478{
479 mprotect_w_x(start, end, 0);
480}
481
482static void end_tcache_write(void *start, void *end)
483{
484#if defined(__arm__) || defined(__aarch64__)
485 size_t len = (char *)end - (char *)start;
486 #if defined(__BLACKBERRY_QNX__)
487 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
488 #elif defined(__MACH__)
489 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
490 #elif defined(VITA)
491 sceKernelSyncVMDomain(sceBlock, start, len);
492 #elif defined(_3DS)
493 ctr_flush_invalidate_cache();
494 #elif defined(HAVE_LIBNX)
495 if (g_jit.type == JitType_CodeMemory) {
496 armDCacheClean(start, len);
497 armICacheInvalidate((char *)start - ndrc_write_ofs, len);
498 // as of v4.2.1 libnx lacks isb
499 __asm__ volatile("isb" ::: "memory");
500 }
501 #elif defined(__aarch64__)
502 // as of 2021, __clear_cache() is still broken on arm64
503 // so here is a custom one :(
504 clear_cache_arm64(start, end);
505 #else
506 __clear_cache(start, end);
507 #endif
508 (void)len;
509#endif
510
511 mprotect_w_x(start, end, 1);
512}
513
514static void *start_block(void)
515{
516 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
517 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
518 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
519 start_tcache_write(NDRC_WRITE_OFFSET(out), NDRC_WRITE_OFFSET(end));
520 return out;
521}
522
523static void end_block(void *start)
524{
525 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(out));
526}
527
528#ifdef NDRC_CACHE_FLUSH_ALL
529
530static int needs_clear_cache;
531
532static void mark_clear_cache(void *target)
533{
534 if (!needs_clear_cache) {
535 start_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
536 needs_clear_cache = 1;
537 }
538}
539
540static void do_clear_cache(void)
541{
542 if (needs_clear_cache) {
543 end_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
544 needs_clear_cache = 0;
545 }
546}
547
548#else
549
550// also takes care of w^x mappings when patching code
551static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
552
553static void mark_clear_cache(void *target)
554{
555 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
556 u_int mask = 1u << ((offset >> 12) & 31);
557 if (!(needs_clear_cache[offset >> 17] & mask)) {
558 char *start = (char *)NDRC_WRITE_OFFSET((uintptr_t)target & ~4095l);
559 start_tcache_write(start, start + 4095);
560 needs_clear_cache[offset >> 17] |= mask;
561 }
562}
563
564// Clearing the cache is rather slow on ARM Linux, so mark the areas
565// that need to be cleared, and then only clear these areas once.
566static void do_clear_cache(void)
567{
568 int i, j;
569 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
570 {
571 u_int bitmap = needs_clear_cache[i];
572 if (!bitmap)
573 continue;
574 for (j = 0; j < 32; j++)
575 {
576 u_char *start, *end;
577 if (!(bitmap & (1u << j)))
578 continue;
579
580 start = ndrc->translation_cache + i*131072 + j*4096;
581 end = start + 4095;
582 for (j++; j < 32; j++) {
583 if (!(bitmap & (1u << j)))
584 break;
585 end += 4096;
586 }
587 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(end));
588 }
589 needs_clear_cache[i] = 0;
590 }
591}
592
593#endif // NDRC_CACHE_FLUSH_ALL
594
595#define NO_CYCLE_PENALTY_THR 12
596
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
610 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2
611 || dops[i].rs1 == dops[i].rt1); // overwrites itself - same effect
612}
613
614// psx addr mirror masking (for invalidation)
615static u_int pmmask(u_int vaddr)
616{
617 vaddr &= ~0xe0000000;
618 if (vaddr < 0x01000000)
619 vaddr &= ~0x00e00000; // RAM mirrors
620 return vaddr;
621}
622
623static u_int get_page(u_int vaddr)
624{
625 u_int page = pmmask(vaddr) >> 12;
626 if (page >= PAGE_COUNT / 2)
627 page = PAGE_COUNT / 2 + (page & (PAGE_COUNT / 2 - 1));
628 return page;
629}
630
631// get a page for looking for a block that has vaddr
632// (needed because the block may start in previous page)
633static u_int get_page_prev(u_int vaddr)
634{
635 assert(MAXBLOCK <= (1 << 12));
636 u_int page = get_page(vaddr);
637 if (page & 511)
638 page--;
639 return page;
640}
641
642static struct ht_entry *hash_table_get(u_int vaddr)
643{
644 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
645}
646
647static void hash_table_add(u_int vaddr, void *tcaddr)
648{
649 struct ht_entry *ht_bin = hash_table_get(vaddr);
650 assert(tcaddr);
651 ht_bin->vaddr[1] = ht_bin->vaddr[0];
652 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
653 ht_bin->vaddr[0] = vaddr;
654 ht_bin->tcaddr[0] = tcaddr;
655}
656
657static void hash_table_remove(int vaddr)
658{
659 //printf("remove hash: %x\n",vaddr);
660 struct ht_entry *ht_bin = hash_table_get(vaddr);
661 if (ht_bin->vaddr[1] == vaddr) {
662 ht_bin->vaddr[1] = -1;
663 ht_bin->tcaddr[1] = NULL;
664 }
665 if (ht_bin->vaddr[0] == vaddr) {
666 ht_bin->vaddr[0] = ht_bin->vaddr[1];
667 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
668 ht_bin->vaddr[1] = -1;
669 ht_bin->tcaddr[1] = NULL;
670 }
671}
672
673static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
674{
675 u_int vaddr_m = vaddr & 0x1fffffff;
676 u_int i, j;
677 for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
678 // ram mirrors, but should not hurt bios
679 for (j = 0; j < 0x800000; j += 0x200000) {
680 invalid_code[(i|j) >> 12] =
681 invalid_code[(i|j|0x80000000u) >> 12] =
682 invalid_code[(i|j|0xa0000000u) >> 12] = invalid;
683 }
684 }
685 if (!invalid && vaddr + len > inv_code_start && vaddr <= inv_code_end)
686 inv_code_start = inv_code_end = ~0;
687}
688
689static int doesnt_expire_soon(u_char *tcaddr)
690{
691 u_int diff = (u_int)(tcaddr - out) & ((1u << TARGET_SIZE_2) - 1u);
692 return diff > EXPIRITY_OFFSET + MAX_OUTPUT_BLOCK_SIZE;
693}
694
695static unused void check_for_block_changes(u_int start, u_int end)
696{
697 u_int start_page = get_page_prev(start);
698 u_int end_page = get_page(end - 1);
699 u_int page;
700
701 for (page = start_page; page <= end_page; page++) {
702 struct block_info *block;
703 for (block = blocks[page]; block != NULL; block = block->next) {
704 if (block->is_dirty)
705 continue;
706 if (memcmp(block->source, block->copy, block->len)) {
707 printf("bad block %08x-%08x %016llx %016llx @%08x\n",
708 block->start, block->start + block->len,
709 *(long long *)block->source, *(long long *)block->copy, psxRegs.pc);
710 fflush(stdout);
711 abort();
712 }
713 }
714 }
715}
716
717static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
718{
719 void *found_clean = NULL;
720 u_int i, page;
721
722 stat_inc(stat_restore_tries);
723 for (page = start_page; page <= end_page; page++) {
724 struct block_info *block;
725 for (block = blocks[page]; block != NULL; block = block->next) {
726 if (vaddr < block->start)
727 break;
728 if (!block->is_dirty || vaddr >= block->start + block->len)
729 continue;
730 for (i = 0; i < block->jump_in_cnt; i++)
731 if (block->jump_in[i].vaddr == vaddr)
732 break;
733 if (i == block->jump_in_cnt)
734 continue;
735 assert(block->source && block->copy);
736 stat_inc(stat_restore_compares);
737 if (memcmp(block->source, block->copy, block->len))
738 continue;
739
740 block->is_dirty = block->inv_near_misses = 0;
741 found_clean = block->jump_in[i].addr;
742 hash_table_add(vaddr, found_clean);
743 mark_invalid_code(block->start, block->len, 0);
744 stat_inc(stat_bc_restore);
745 inv_debug("INV: restored %08x %p (%d)\n", vaddr, found_clean, block->jump_in_cnt);
746 return found_clean;
747 }
748 }
749 return NULL;
750}
751
752// this doesn't normally happen
753static noinline u_int generate_exception(u_int pc)
754{
755 //if (execBreakCheck(&psxRegs, pc))
756 // return psxRegs.pc;
757
758 // generate an address or bus error
759 psxRegs.CP0.n.Cause &= 0x300;
760 psxRegs.CP0.n.EPC = pc;
761 if (pc & 3) {
762 psxRegs.CP0.n.Cause |= R3000E_AdEL << 2;
763 psxRegs.CP0.n.BadVAddr = pc;
764#ifdef DRC_DBG
765 last_count -= 2;
766#endif
767 } else
768 psxRegs.CP0.n.Cause |= R3000E_IBE << 2;
769 return (psxRegs.pc = 0x80000080);
770}
771
772// Get address from virtual address
773// This is called from the recompiled JR/JALR instructions
774static void noinline *get_addr(u_int vaddr, int can_compile)
775{
776 u_int start_page = get_page_prev(vaddr);
777 u_int i, page, end_page = get_page(vaddr);
778 void *found_clean = NULL;
779
780 stat_inc(stat_jump_in_lookups);
781 for (page = start_page; page <= end_page; page++) {
782 const struct block_info *block;
783 for (block = blocks[page]; block != NULL; block = block->next) {
784 if (vaddr < block->start)
785 break;
786 if (block->is_dirty || vaddr >= block->start + block->len)
787 continue;
788 for (i = 0; i < block->jump_in_cnt; i++)
789 if (block->jump_in[i].vaddr == vaddr)
790 break;
791 if (i == block->jump_in_cnt)
792 continue;
793 found_clean = block->jump_in[i].addr;
794 hash_table_add(vaddr, found_clean);
795 return found_clean;
796 }
797 }
798 found_clean = try_restore_block(vaddr, start_page, end_page);
799 if (found_clean)
800 return found_clean;
801
802 if (!can_compile)
803 return NULL;
804
805 int r = new_recompile_block(vaddr);
806 if (likely(r == 0))
807 return ndrc_get_addr_ht(vaddr);
808
809 return ndrc_get_addr_ht(generate_exception(vaddr));
810}
811
812// Look up address in hash table first
813void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile)
814{
815 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
816 const struct ht_entry *ht_bin = hash_table_get(vaddr);
817 u_int vaddr_a = vaddr & ~3;
818 stat_inc(stat_ht_lookups);
819 if (ht_bin->vaddr[0] == vaddr_a) return ht_bin->tcaddr[0];
820 if (ht_bin->vaddr[1] == vaddr_a) return ht_bin->tcaddr[1];
821 return get_addr(vaddr, can_compile);
822}
823
824void *ndrc_get_addr_ht(u_int vaddr)
825{
826 return ndrc_get_addr_ht_param(vaddr, 1);
827}
828
829static void clear_all_regs(signed char regmap[])
830{
831 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
832}
833
834// get_reg: get allocated host reg from mips reg
835// returns -1 if no such mips reg was allocated
836#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
837
838extern signed char get_reg(const signed char regmap[], signed char r);
839
840#else
841
842static signed char get_reg(const signed char regmap[], signed char r)
843{
844 int hr;
845 for (hr = 0; hr < HOST_REGS; hr++) {
846 if (hr == EXCLUDE_REG)
847 continue;
848 if (regmap[hr] == r)
849 return hr;
850 }
851 return -1;
852}
853
854#endif
855
856// get reg suitable for writing
857static signed char get_reg_w(const signed char regmap[], signed char r)
858{
859 return r == 0 ? -1 : get_reg(regmap, r);
860}
861
862// get reg as mask bit (1 << hr)
863static u_int get_regm(const signed char regmap[], signed char r)
864{
865 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
866}
867
868static signed char get_reg_temp(const signed char regmap[])
869{
870 int hr;
871 for (hr = 0; hr < HOST_REGS; hr++) {
872 if (hr == EXCLUDE_REG)
873 continue;
874 if (regmap[hr] == (signed char)-1)
875 return hr;
876 }
877 return -1;
878}
879
880// Find a register that is available for two consecutive cycles
881static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
882{
883 int hr;
884 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
885 return -1;
886}
887
888// reverse reg map: mips -> host
889#define RRMAP_SIZE 64
890static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
891 u_int *regs_can_change)
892{
893 u_int r, hr, hr_can_change = 0;
894 memset(rrmap, -1, RRMAP_SIZE);
895 for (hr = 0; hr < HOST_REGS; )
896 {
897 r = regmap[hr];
898 rrmap[r & (RRMAP_SIZE - 1)] = hr;
899 // only add mips $1-$31+$lo, others shifted out
900 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
901 hr++;
902 if (hr == EXCLUDE_REG)
903 hr++;
904 }
905 hr_can_change |= 1u << (rrmap[33] & 31);
906 hr_can_change |= 1u << (rrmap[CCREG] & 31);
907 hr_can_change &= ~(1u << 31);
908 *regs_can_change = hr_can_change;
909}
910
911// same as get_reg, but takes rrmap
912static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
913{
914 assert(0 <= r && r < RRMAP_SIZE);
915 return rrmap[r];
916}
917
918static int count_free_regs(const signed char regmap[])
919{
920 int count=0;
921 int hr;
922 for(hr=0;hr<HOST_REGS;hr++)
923 {
924 if(hr!=EXCLUDE_REG) {
925 if(regmap[hr]<0) count++;
926 }
927 }
928 return count;
929}
930
931static void dirty_reg(struct regstat *cur, signed char reg)
932{
933 int hr;
934 if (!reg) return;
935 hr = get_reg(cur->regmap, reg);
936 if (hr >= 0)
937 cur->dirty |= 1<<hr;
938}
939
940static void set_const(struct regstat *cur, signed char reg, uint32_t value)
941{
942 int hr;
943 if (!reg) return;
944 hr = get_reg(cur->regmap, reg);
945 if (hr >= 0) {
946 cur->isconst |= 1<<hr;
947 current_constmap[hr] = value;
948 }
949}
950
951static void clear_const(struct regstat *cur, signed char reg)
952{
953 int hr;
954 if (!reg) return;
955 hr = get_reg(cur->regmap, reg);
956 if (hr >= 0)
957 cur->isconst &= ~(1<<hr);
958}
959
960static int is_const(const struct regstat *cur, signed char reg)
961{
962 int hr;
963 if (reg < 0) return 0;
964 if (!reg) return 1;
965 hr = get_reg(cur->regmap, reg);
966 if (hr >= 0)
967 return (cur->isconst>>hr)&1;
968 return 0;
969}
970
971static uint32_t get_const(const struct regstat *cur, signed char reg)
972{
973 int hr;
974 if (!reg) return 0;
975 hr = get_reg(cur->regmap, reg);
976 if (hr >= 0)
977 return current_constmap[hr];
978
979 SysPrintf("Unknown constant in r%d\n", reg);
980 abort();
981}
982
983// Least soon needed registers
984// Look at the next ten instructions and see which registers
985// will be used. Try not to reallocate these.
986static void lsn(u_char hsn[], int i)
987{
988 int j;
989 int b=-1;
990 for(j=0;j<9;j++)
991 {
992 if(i+j>=slen) {
993 j=slen-i-1;
994 break;
995 }
996 if (dops[i+j].is_ujump)
997 {
998 // Don't go past an unconditonal jump
999 j++;
1000 break;
1001 }
1002 }
1003 for(;j>=0;j--)
1004 {
1005 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
1006 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
1007 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
1008 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
1009 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
1010 // Stores can allocate zero
1011 hsn[dops[i+j].rs1]=j;
1012 hsn[dops[i+j].rs2]=j;
1013 }
1014 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
1015 hsn[ROREG] = j;
1016 // On some architectures stores need invc_ptr
1017 #if defined(HOST_IMM8)
1018 if (dops[i+j].is_store)
1019 hsn[INVCP] = j;
1020 #endif
1021 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1022 {
1023 hsn[CCREG]=j;
1024 b=j;
1025 }
1026 }
1027 if(b>=0)
1028 {
1029 if(cinfo[i+b].ba>=start && cinfo[i+b].ba<(start+slen*4))
1030 {
1031 // Follow first branch
1032 int t=(cinfo[i+b].ba-start)>>2;
1033 j=7-b;if(t+j>=slen) j=slen-t-1;
1034 for(;j>=0;j--)
1035 {
1036 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1037 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1038 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1039 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1040 }
1041 }
1042 // TODO: preferred register based on backward branch
1043 }
1044 // Delay slot should preferably not overwrite branch conditions or cycle count
1045 if (i > 0 && dops[i-1].is_jump) {
1046 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1047 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1048 hsn[CCREG]=1;
1049 // ...or hash tables
1050 hsn[RHASH]=1;
1051 hsn[RHTBL]=1;
1052 }
1053 // Coprocessor load/store needs FTEMP, even if not declared
1054 if(dops[i].itype==C2LS) {
1055 hsn[FTEMP]=0;
1056 }
1057 // Load L/R also uses FTEMP as a temporary register
1058 if(dops[i].itype==LOADLR) {
1059 hsn[FTEMP]=0;
1060 }
1061 // Also SWL/SWR/SDL/SDR
1062 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
1063 hsn[FTEMP]=0;
1064 }
1065 // Don't remove the miniht registers
1066 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1067 {
1068 hsn[RHASH]=0;
1069 hsn[RHTBL]=0;
1070 }
1071}
1072
1073// We only want to allocate registers if we're going to use them again soon
1074static int needed_again(int r, int i)
1075{
1076 int j;
1077 int b=-1;
1078 int rn=10;
1079
1080 if (i > 0 && dops[i-1].is_ujump)
1081 {
1082 if(cinfo[i-1].ba<start || cinfo[i-1].ba>start+slen*4-4)
1083 return 0; // Don't need any registers if exiting the block
1084 }
1085 for(j=0;j<9;j++)
1086 {
1087 if(i+j>=slen) {
1088 j=slen-i-1;
1089 break;
1090 }
1091 if (dops[i+j].is_ujump)
1092 {
1093 // Don't go past an unconditonal jump
1094 j++;
1095 break;
1096 }
1097 if (dops[i+j].is_exception)
1098 {
1099 break;
1100 }
1101 }
1102 for(;j>=1;j--)
1103 {
1104 if(dops[i+j].rs1==r) rn=j;
1105 if(dops[i+j].rs2==r) rn=j;
1106 if((unneeded_reg[i+j]>>r)&1) rn=10;
1107 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1108 {
1109 b=j;
1110 }
1111 }
1112 if(rn<10) return 1;
1113 (void)b;
1114 return 0;
1115}
1116
1117// Try to match register allocations at the end of a loop with those
1118// at the beginning
1119static int loop_reg(int i, int r, int hr)
1120{
1121 int j,k;
1122 for(j=0;j<9;j++)
1123 {
1124 if(i+j>=slen) {
1125 j=slen-i-1;
1126 break;
1127 }
1128 if (dops[i+j].is_ujump)
1129 {
1130 // Don't go past an unconditonal jump
1131 j++;
1132 break;
1133 }
1134 }
1135 k=0;
1136 if(i>0){
1137 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1138 k--;
1139 }
1140 for(;k<j;k++)
1141 {
1142 assert(r < 64);
1143 if((unneeded_reg[i+k]>>r)&1) return hr;
1144 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1145 {
1146 if(cinfo[i+k].ba>=start && cinfo[i+k].ba<(start+i*4))
1147 {
1148 int t=(cinfo[i+k].ba-start)>>2;
1149 int reg=get_reg(regs[t].regmap_entry,r);
1150 if(reg>=0) return reg;
1151 //reg=get_reg(regs[t+1].regmap_entry,r);
1152 //if(reg>=0) return reg;
1153 }
1154 }
1155 }
1156 return hr;
1157}
1158
1159
1160// Allocate every register, preserving source/target regs
1161static void alloc_all(struct regstat *cur,int i)
1162{
1163 int hr;
1164
1165 for(hr=0;hr<HOST_REGS;hr++) {
1166 if(hr!=EXCLUDE_REG) {
1167 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1168 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1169 {
1170 cur->regmap[hr]=-1;
1171 cur->dirty&=~(1<<hr);
1172 }
1173 // Don't need zeros
1174 if(cur->regmap[hr]==0)
1175 {
1176 cur->regmap[hr]=-1;
1177 cur->dirty&=~(1<<hr);
1178 }
1179 }
1180 }
1181}
1182
1183#ifndef NDEBUG
1184static int host_tempreg_in_use;
1185
1186static void host_tempreg_acquire(void)
1187{
1188 assert(!host_tempreg_in_use);
1189 host_tempreg_in_use = 1;
1190}
1191
1192static void host_tempreg_release(void)
1193{
1194 host_tempreg_in_use = 0;
1195}
1196#else
1197static void host_tempreg_acquire(void) {}
1198static void host_tempreg_release(void) {}
1199#endif
1200
1201#ifdef ASSEM_PRINT
1202extern void gen_interupt();
1203extern void do_insn_cmp();
1204#define FUNCNAME(f) { f, " " #f }
1205static const struct {
1206 void *addr;
1207 const char *name;
1208} function_names[] = {
1209 FUNCNAME(cc_interrupt),
1210 FUNCNAME(gen_interupt),
1211 FUNCNAME(ndrc_get_addr_ht),
1212 FUNCNAME(jump_handler_read8),
1213 FUNCNAME(jump_handler_read16),
1214 FUNCNAME(jump_handler_read32),
1215 FUNCNAME(jump_handler_write8),
1216 FUNCNAME(jump_handler_write16),
1217 FUNCNAME(jump_handler_write32),
1218 FUNCNAME(ndrc_write_invalidate_one),
1219 FUNCNAME(ndrc_write_invalidate_many),
1220 FUNCNAME(jump_to_new_pc),
1221 FUNCNAME(jump_break),
1222 FUNCNAME(jump_break_ds),
1223 FUNCNAME(jump_syscall),
1224 FUNCNAME(jump_syscall_ds),
1225 FUNCNAME(jump_overflow),
1226 FUNCNAME(jump_overflow_ds),
1227 FUNCNAME(jump_addrerror),
1228 FUNCNAME(jump_addrerror_ds),
1229 FUNCNAME(call_gteStall),
1230 FUNCNAME(new_dyna_leave),
1231 FUNCNAME(pcsx_mtc0),
1232 FUNCNAME(pcsx_mtc0_ds),
1233 FUNCNAME(execI),
1234#ifdef __aarch64__
1235 FUNCNAME(do_memhandler_pre),
1236 FUNCNAME(do_memhandler_post),
1237#endif
1238#ifdef DRC_DBG
1239 FUNCNAME(do_insn_cmp),
1240#endif
1241};
1242
1243static const char *func_name(const void *a)
1244{
1245 int i;
1246 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1247 if (function_names[i].addr == a)
1248 return function_names[i].name;
1249 return "";
1250}
1251
1252static const char *fpofs_name(u_int ofs)
1253{
1254 u_int *p = (u_int *)&dynarec_local + ofs/sizeof(u_int);
1255 static char buf[64];
1256 switch (ofs) {
1257 #define ofscase(x) case LO_##x: return " ; " #x
1258 ofscase(next_interupt);
1259 ofscase(cycle_count);
1260 ofscase(last_count);
1261 ofscase(pending_exception);
1262 ofscase(stop);
1263 ofscase(address);
1264 ofscase(lo);
1265 ofscase(hi);
1266 ofscase(PC);
1267 ofscase(cycle);
1268 ofscase(mem_rtab);
1269 ofscase(mem_wtab);
1270 ofscase(psxH_ptr);
1271 ofscase(invc_ptr);
1272 ofscase(ram_offset);
1273 #undef ofscase
1274 }
1275 buf[0] = 0;
1276 if (psxRegs.GPR.r <= p && p < &psxRegs.GPR.r[32])
1277 snprintf(buf, sizeof(buf), " ; r%d", (int)(p - psxRegs.GPR.r));
1278 else if (psxRegs.CP0.r <= p && p < &psxRegs.CP0.r[32])
1279 snprintf(buf, sizeof(buf), " ; cp0 $%d", (int)(p - psxRegs.CP0.r));
1280 else if (psxRegs.CP2D.r <= p && p < &psxRegs.CP2D.r[32])
1281 snprintf(buf, sizeof(buf), " ; cp2d $%d", (int)(p - psxRegs.CP2D.r));
1282 else if (psxRegs.CP2C.r <= p && p < &psxRegs.CP2C.r[32])
1283 snprintf(buf, sizeof(buf), " ; cp2c $%d", (int)(p - psxRegs.CP2C.r));
1284 return buf;
1285}
1286#else
1287#define func_name(x) ""
1288#define fpofs_name(x) ""
1289#endif
1290
1291#ifdef __i386__
1292#include "assem_x86.c"
1293#endif
1294#ifdef __x86_64__
1295#include "assem_x64.c"
1296#endif
1297#ifdef __arm__
1298#include "assem_arm.c"
1299#endif
1300#ifdef __aarch64__
1301#include "assem_arm64.c"
1302#endif
1303
1304static void *get_trampoline(const void *f)
1305{
1306 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1307 size_t i;
1308
1309 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1310 if (tramp->f[i] == f || tramp->f[i] == NULL)
1311 break;
1312 }
1313 if (i == ARRAY_SIZE(tramp->f)) {
1314 SysPrintf("trampoline table is full, last func %p\n", f);
1315 abort();
1316 }
1317 if (tramp->f[i] == NULL) {
1318 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1319 tramp->f[i] = f;
1320 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1321#ifdef HAVE_LIBNX
1322 // invalidate the RX mirror (unsure if necessary, but just in case...)
1323 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1324#endif
1325 }
1326 return &ndrc->tramp.ops[i];
1327}
1328
1329static void emit_far_jump(const void *f)
1330{
1331 if (can_jump_or_call(f)) {
1332 emit_jmp(f);
1333 return;
1334 }
1335
1336 f = get_trampoline(f);
1337 emit_jmp(f);
1338}
1339
1340static void emit_far_call(const void *f)
1341{
1342 if (can_jump_or_call(f)) {
1343 emit_call(f);
1344 return;
1345 }
1346
1347 f = get_trampoline(f);
1348 emit_call(f);
1349}
1350
1351// Check if an address is already compiled
1352// but don't return addresses which are about to expire from the cache
1353static void *check_addr(u_int vaddr)
1354{
1355 struct ht_entry *ht_bin = hash_table_get(vaddr);
1356 size_t i;
1357 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1358 if (ht_bin->vaddr[i] == vaddr)
1359 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1360 return ht_bin->tcaddr[i];
1361 }
1362
1363 // refactor to get_addr_nocompile?
1364 u_int start_page = get_page_prev(vaddr);
1365 u_int page, end_page = get_page(vaddr);
1366
1367 stat_inc(stat_jump_in_lookups);
1368 for (page = start_page; page <= end_page; page++) {
1369 const struct block_info *block;
1370 for (block = blocks[page]; block != NULL; block = block->next) {
1371 if (vaddr < block->start)
1372 break;
1373 if (block->is_dirty || vaddr >= block->start + block->len)
1374 continue;
1375 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1376 continue;
1377 for (i = 0; i < block->jump_in_cnt; i++)
1378 if (block->jump_in[i].vaddr == vaddr)
1379 break;
1380 if (i == block->jump_in_cnt)
1381 continue;
1382
1383 // Update existing entry with current address
1384 void *addr = block->jump_in[i].addr;
1385 if (ht_bin->vaddr[0] == vaddr) {
1386 ht_bin->tcaddr[0] = addr;
1387 return addr;
1388 }
1389 if (ht_bin->vaddr[1] == vaddr) {
1390 ht_bin->tcaddr[1] = addr;
1391 return addr;
1392 }
1393 // Insert into hash table with low priority.
1394 // Don't evict existing entries, as they are probably
1395 // addresses that are being accessed frequently.
1396 if (ht_bin->vaddr[0] == -1) {
1397 ht_bin->vaddr[0] = vaddr;
1398 ht_bin->tcaddr[0] = addr;
1399 }
1400 else if (ht_bin->vaddr[1] == -1) {
1401 ht_bin->vaddr[1] = vaddr;
1402 ht_bin->tcaddr[1] = addr;
1403 }
1404 return addr;
1405 }
1406 }
1407 return NULL;
1408}
1409
1410static void blocks_clear(struct block_info **head)
1411{
1412 struct block_info *cur, *next;
1413
1414 if ((cur = *head)) {
1415 *head = NULL;
1416 while (cur) {
1417 next = cur->next;
1418 free(cur);
1419 cur = next;
1420 }
1421 }
1422}
1423
1424static int blocks_remove_matching_addrs(struct block_info **head,
1425 u_int base_offs, int shift)
1426{
1427 struct block_info *next;
1428 int hit = 0;
1429 while (*head) {
1430 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1431 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1432 invalidate_block(*head);
1433 next = (*head)->next;
1434 free(*head);
1435 *head = next;
1436 stat_dec(stat_blocks);
1437 hit = 1;
1438 }
1439 else
1440 {
1441 head = &((*head)->next);
1442 }
1443 }
1444 return hit;
1445}
1446
1447// This is called when we write to a compiled block (see do_invstub)
1448static void unlink_jumps_vaddr_range(u_int start, u_int end)
1449{
1450 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1451 int i;
1452
1453 for (page = start_page; page <= end_page; page++) {
1454 struct jump_info *ji = jumps[page];
1455 if (ji == NULL)
1456 continue;
1457 for (i = 0; i < ji->count; ) {
1458 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1459 i++;
1460 continue;
1461 }
1462
1463 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1464 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1465 void *host_addr = find_extjump_insn(ji->e[i].stub);
1466 mark_clear_cache(host_addr);
1467 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1468
1469 stat_dec(stat_links);
1470 ji->count--;
1471 if (i < ji->count) {
1472 ji->e[i] = ji->e[ji->count];
1473 continue;
1474 }
1475 i++;
1476 }
1477 }
1478}
1479
1480static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1481{
1482 int i;
1483 if (ji == NULL)
1484 return;
1485 for (i = 0; i < ji->count; ) {
1486 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1487 if (((tc_offs ^ base_offs) >> shift) != 0) {
1488 i++;
1489 continue;
1490 }
1491
1492 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1493 stat_dec(stat_links);
1494 ji->count--;
1495 if (i < ji->count) {
1496 ji->e[i] = ji->e[ji->count];
1497 continue;
1498 }
1499 i++;
1500 }
1501}
1502
1503static void invalidate_block(struct block_info *block)
1504{
1505 u_int i;
1506
1507 block->is_dirty = 1;
1508 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1509 for (i = 0; i < block->jump_in_cnt; i++)
1510 hash_table_remove(block->jump_in[i].vaddr);
1511}
1512
1513static int invalidate_range(u_int start, u_int end,
1514 u32 *inv_start_ret, u32 *inv_end_ret)
1515{
1516 struct block_info *last_block = NULL;
1517 u_int start_page = get_page_prev(start);
1518 u_int end_page = get_page(end - 1);
1519 u_int start_m = pmmask(start);
1520 u_int end_m = pmmask(end - 1);
1521 u_int inv_start, inv_end;
1522 u_int blk_start_m, blk_end_m;
1523 u_int page;
1524 int hit = 0;
1525
1526 // additional area without code (to supplement invalid_code[]), [start, end)
1527 // avoids excessive ndrc_write_invalidate*() calls
1528 inv_start = start_m & ~0xfff;
1529 inv_end = end_m | 0xfff;
1530
1531 for (page = start_page; page <= end_page; page++) {
1532 struct block_info *block;
1533 for (block = blocks[page]; block != NULL; block = block->next) {
1534 if (block->is_dirty)
1535 continue;
1536 last_block = block;
1537 blk_end_m = pmmask(block->start + block->len);
1538 if (blk_end_m <= start_m) {
1539 inv_start = max(inv_start, blk_end_m);
1540 continue;
1541 }
1542 blk_start_m = pmmask(block->start);
1543 if (end_m <= blk_start_m) {
1544 inv_end = min(inv_end, blk_start_m - 1);
1545 continue;
1546 }
1547 if (!block->source) // "hack" block - leave it alone
1548 continue;
1549
1550 hit++;
1551 invalidate_block(block);
1552 stat_inc(stat_inv_hits);
1553 }
1554 }
1555
1556 if (!hit && last_block && last_block->source) {
1557 // could be some leftover unused block, uselessly trapping writes
1558 last_block->inv_near_misses++;
1559 if (last_block->inv_near_misses > 128) {
1560 invalidate_block(last_block);
1561 stat_inc(stat_inv_hits);
1562 hit++;
1563 }
1564 }
1565 if (hit) {
1566 do_clear_cache();
1567#ifdef USE_MINI_HT
1568 memset(mini_ht, -1, sizeof(mini_ht));
1569#endif
1570 }
1571
1572 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1573 // the whole page is empty now
1574 mark_invalid_code(start, 1, 1);
1575
1576 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1577 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1578 return hit;
1579}
1580
1581void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1582{
1583 invalidate_range(start, end, NULL, NULL);
1584}
1585
1586static void ndrc_write_invalidate_many(u_int start, u_int end)
1587{
1588 // this check is done by the caller
1589 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1590 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1591#ifdef INV_DEBUG_W
1592 int invc = invalid_code[start >> 12];
1593 u_int len = end - start;
1594 if (ret)
1595 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1596 else
1597 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1598 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1599 check_for_block_changes(start, end);
1600#endif
1601 stat_inc(stat_inv_addr_calls);
1602 (void)ret;
1603}
1604
1605void ndrc_write_invalidate_one(u_int addr)
1606{
1607 ndrc_write_invalidate_many(addr, addr + 4);
1608}
1609
1610// This is called when loading a save state.
1611// Anything could have changed, so invalidate everything.
1612void new_dynarec_invalidate_all_pages(void)
1613{
1614 struct block_info *block;
1615 u_int page;
1616 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1617 for (block = blocks[page]; block != NULL; block = block->next) {
1618 if (block->is_dirty)
1619 continue;
1620 if (!block->source) // hack block?
1621 continue;
1622 invalidate_block(block);
1623 }
1624 }
1625
1626 #ifdef USE_MINI_HT
1627 memset(mini_ht, -1, sizeof(mini_ht));
1628 #endif
1629 do_clear_cache();
1630}
1631
1632// Add an entry to jump_out after making a link
1633// src should point to code by emit_extjump()
1634void ndrc_add_jump_out(u_int vaddr, void *src)
1635{
1636 inv_debug("ndrc_add_jump_out: %p -> %x\n", src, vaddr);
1637 u_int page = get_page(vaddr);
1638 struct jump_info *ji;
1639
1640 stat_inc(stat_links);
1641 check_extjump2(src);
1642 ji = jumps[page];
1643 if (ji == NULL) {
1644 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1645 ji->alloc = 16;
1646 ji->count = 0;
1647 }
1648 else if (ji->count >= ji->alloc) {
1649 ji->alloc += 16;
1650 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1651 }
1652 jumps[page] = ji;
1653 ji->e[ji->count].target_vaddr = vaddr;
1654 ji->e[ji->count].stub = src;
1655 ji->count++;
1656}
1657
1658/* Register allocation */
1659
1660static void alloc_set(struct regstat *cur, int reg, int hr)
1661{
1662 cur->regmap[hr] = reg;
1663 cur->dirty &= ~(1u << hr);
1664 cur->isconst &= ~(1u << hr);
1665 cur->noevict |= 1u << hr;
1666}
1667
1668static void evict_alloc_reg(struct regstat *cur, int i, int reg, int preferred_hr)
1669{
1670 u_char hsn[MAXREG+1];
1671 int j, r, hr;
1672 memset(hsn, 10, sizeof(hsn));
1673 lsn(hsn, i);
1674 //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]);
1675 if(i>0) {
1676 // Don't evict the cycle count at entry points, otherwise the entry
1677 // stub will have to write it.
1678 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1679 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1680 for(j=10;j>=3;j--)
1681 {
1682 // Alloc preferred register if available
1683 if (!((cur->noevict >> preferred_hr) & 1)
1684 && hsn[cur->regmap[preferred_hr]] == j)
1685 {
1686 alloc_set(cur, reg, preferred_hr);
1687 return;
1688 }
1689 for(r=1;r<=MAXREG;r++)
1690 {
1691 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1692 for(hr=0;hr<HOST_REGS;hr++) {
1693 if (hr == EXCLUDE_REG || ((cur->noevict >> hr) & 1))
1694 continue;
1695 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1696 if(cur->regmap[hr]==r) {
1697 alloc_set(cur, reg, hr);
1698 return;
1699 }
1700 }
1701 }
1702 }
1703 }
1704 }
1705 }
1706 for(j=10;j>=0;j--)
1707 {
1708 for(r=1;r<=MAXREG;r++)
1709 {
1710 if(hsn[r]==j) {
1711 for(hr=0;hr<HOST_REGS;hr++) {
1712 if (hr == EXCLUDE_REG || ((cur->noevict >> hr) & 1))
1713 continue;
1714 if(cur->regmap[hr]==r) {
1715 alloc_set(cur, reg, hr);
1716 return;
1717 }
1718 }
1719 }
1720 }
1721 }
1722 SysPrintf("This shouldn't happen (evict_alloc_reg)\n");
1723 abort();
1724}
1725
1726// Note: registers are allocated clean (unmodified state)
1727// if you intend to modify the register, you must call dirty_reg().
1728static void alloc_reg(struct regstat *cur,int i,signed char reg)
1729{
1730 int r,hr;
1731 int preferred_reg = PREFERRED_REG_FIRST
1732 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1733 if (reg == CCREG) preferred_reg = HOST_CCREG;
1734 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1735 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1736 assert(reg >= 0);
1737
1738 // Don't allocate unused registers
1739 if((cur->u>>reg)&1) return;
1740
1741 // see if it's already allocated
1742 if ((hr = get_reg(cur->regmap, reg)) >= 0) {
1743 cur->noevict |= 1u << hr;
1744 return;
1745 }
1746
1747 // Keep the same mapping if the register was already allocated in a loop
1748 preferred_reg = loop_reg(i,reg,preferred_reg);
1749
1750 // Try to allocate the preferred register
1751 if (cur->regmap[preferred_reg] == -1) {
1752 alloc_set(cur, reg, preferred_reg);
1753 return;
1754 }
1755 r=cur->regmap[preferred_reg];
1756 assert(r < 64);
1757 if((cur->u>>r)&1) {
1758 alloc_set(cur, reg, preferred_reg);
1759 return;
1760 }
1761
1762 // Clear any unneeded registers
1763 // We try to keep the mapping consistent, if possible, because it
1764 // makes branches easier (especially loops). So we try to allocate
1765 // first (see above) before removing old mappings. If this is not
1766 // possible then go ahead and clear out the registers that are no
1767 // longer needed.
1768 for(hr=0;hr<HOST_REGS;hr++)
1769 {
1770 r=cur->regmap[hr];
1771 if(r>=0) {
1772 assert(r < 64);
1773 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1774 }
1775 }
1776
1777 // Try to allocate any available register, but prefer
1778 // registers that have not been used recently.
1779 if (i > 0) {
1780 for (hr = PREFERRED_REG_FIRST; ; ) {
1781 if (cur->regmap[hr] < 0) {
1782 int oldreg = regs[i-1].regmap[hr];
1783 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1784 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1785 {
1786 alloc_set(cur, reg, hr);
1787 return;
1788 }
1789 }
1790 hr++;
1791 if (hr == EXCLUDE_REG)
1792 hr++;
1793 if (hr == HOST_REGS)
1794 hr = 0;
1795 if (hr == PREFERRED_REG_FIRST)
1796 break;
1797 }
1798 }
1799
1800 // Try to allocate any available register
1801 for (hr = PREFERRED_REG_FIRST; ; ) {
1802 if (cur->regmap[hr] < 0) {
1803 alloc_set(cur, reg, hr);
1804 return;
1805 }
1806 hr++;
1807 if (hr == EXCLUDE_REG)
1808 hr++;
1809 if (hr == HOST_REGS)
1810 hr = 0;
1811 if (hr == PREFERRED_REG_FIRST)
1812 break;
1813 }
1814
1815 // Ok, now we have to evict someone
1816 // Pick a register we hopefully won't need soon
1817 evict_alloc_reg(cur, i, reg, preferred_reg);
1818}
1819
1820// Allocate a temporary register. This is done without regard to
1821// dirty status or whether the register we request is on the unneeded list
1822// Note: This will only allocate one register, even if called multiple times
1823static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1824{
1825 int r,hr;
1826
1827 // see if it's already allocated
1828 for (hr = 0; hr < HOST_REGS; hr++)
1829 {
1830 if (hr != EXCLUDE_REG && cur->regmap[hr] == reg) {
1831 cur->noevict |= 1u << hr;
1832 return;
1833 }
1834 }
1835
1836 // Try to allocate any available register
1837 for(hr=HOST_REGS-1;hr>=0;hr--) {
1838 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1839 alloc_set(cur, reg, hr);
1840 return;
1841 }
1842 }
1843
1844 // Find an unneeded register
1845 for(hr=HOST_REGS-1;hr>=0;hr--)
1846 {
1847 r=cur->regmap[hr];
1848 if(r>=0) {
1849 assert(r < 64);
1850 if((cur->u>>r)&1) {
1851 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1852 alloc_set(cur, reg, hr);
1853 return;
1854 }
1855 }
1856 }
1857 }
1858
1859 // Ok, now we have to evict someone
1860 // Pick a register we hopefully won't need soon
1861 evict_alloc_reg(cur, i, reg, 0);
1862}
1863
1864static void mov_alloc(struct regstat *current,int i)
1865{
1866 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1867 alloc_cc(current,i); // for stalls
1868 dirty_reg(current,CCREG);
1869 }
1870
1871 // Note: Don't need to actually alloc the source registers
1872 //alloc_reg(current,i,dops[i].rs1);
1873 alloc_reg(current,i,dops[i].rt1);
1874
1875 clear_const(current,dops[i].rs1);
1876 clear_const(current,dops[i].rt1);
1877 dirty_reg(current,dops[i].rt1);
1878}
1879
1880static void shiftimm_alloc(struct regstat *current,int i)
1881{
1882 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1883 {
1884 if(dops[i].rt1) {
1885 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1886 else dops[i].use_lt1=!!dops[i].rs1;
1887 alloc_reg(current,i,dops[i].rt1);
1888 dirty_reg(current,dops[i].rt1);
1889 if(is_const(current,dops[i].rs1)) {
1890 int v=get_const(current,dops[i].rs1);
1891 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<cinfo[i].imm);
1892 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>cinfo[i].imm);
1893 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>cinfo[i].imm);
1894 }
1895 else clear_const(current,dops[i].rt1);
1896 }
1897 }
1898 else
1899 {
1900 clear_const(current,dops[i].rs1);
1901 clear_const(current,dops[i].rt1);
1902 }
1903
1904 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1905 {
1906 assert(0);
1907 }
1908 if(dops[i].opcode2==0x3c) // DSLL32
1909 {
1910 assert(0);
1911 }
1912 if(dops[i].opcode2==0x3e) // DSRL32
1913 {
1914 assert(0);
1915 }
1916 if(dops[i].opcode2==0x3f) // DSRA32
1917 {
1918 assert(0);
1919 }
1920}
1921
1922static void shift_alloc(struct regstat *current,int i)
1923{
1924 if(dops[i].rt1) {
1925 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1926 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1927 alloc_reg(current,i,dops[i].rt1);
1928 if(dops[i].rt1==dops[i].rs2) {
1929 alloc_reg_temp(current,i,-1);
1930 cinfo[i].min_free_regs=1;
1931 }
1932 clear_const(current,dops[i].rs1);
1933 clear_const(current,dops[i].rs2);
1934 clear_const(current,dops[i].rt1);
1935 dirty_reg(current,dops[i].rt1);
1936 }
1937}
1938
1939static void alu_alloc(struct regstat *current,int i)
1940{
1941 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1942 if(dops[i].rt1) {
1943 if(dops[i].rs1&&dops[i].rs2) {
1944 alloc_reg(current,i,dops[i].rs1);
1945 alloc_reg(current,i,dops[i].rs2);
1946 }
1947 else {
1948 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1949 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1950 }
1951 alloc_reg(current,i,dops[i].rt1);
1952 }
1953 if (dops[i].may_except) {
1954 alloc_cc_optional(current, i); // for exceptions
1955 alloc_reg_temp(current, i, -1);
1956 cinfo[i].min_free_regs = 1;
1957 }
1958 }
1959 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1960 if(dops[i].rt1) {
1961 alloc_reg(current,i,dops[i].rs1);
1962 alloc_reg(current,i,dops[i].rs2);
1963 alloc_reg(current,i,dops[i].rt1);
1964 }
1965 }
1966 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1967 if(dops[i].rt1) {
1968 if(dops[i].rs1&&dops[i].rs2) {
1969 alloc_reg(current,i,dops[i].rs1);
1970 alloc_reg(current,i,dops[i].rs2);
1971 }
1972 else
1973 {
1974 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1975 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1976 }
1977 alloc_reg(current,i,dops[i].rt1);
1978 }
1979 }
1980 clear_const(current,dops[i].rs1);
1981 clear_const(current,dops[i].rs2);
1982 clear_const(current,dops[i].rt1);
1983 dirty_reg(current,dops[i].rt1);
1984}
1985
1986static void imm16_alloc(struct regstat *current,int i)
1987{
1988 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1989 else dops[i].use_lt1=!!dops[i].rs1;
1990 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1991 if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1992 clear_const(current,dops[i].rs1);
1993 clear_const(current,dops[i].rt1);
1994 }
1995 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1996 if(is_const(current,dops[i].rs1)) {
1997 int v=get_const(current,dops[i].rs1);
1998 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&cinfo[i].imm);
1999 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|cinfo[i].imm);
2000 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^cinfo[i].imm);
2001 }
2002 else clear_const(current,dops[i].rt1);
2003 }
2004 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2005 if(is_const(current,dops[i].rs1)) {
2006 int v=get_const(current,dops[i].rs1);
2007 set_const(current,dops[i].rt1,v+cinfo[i].imm);
2008 }
2009 else clear_const(current,dops[i].rt1);
2010 if (dops[i].may_except) {
2011 alloc_cc_optional(current, i); // for exceptions
2012 alloc_reg_temp(current, i, -1);
2013 cinfo[i].min_free_regs = 1;
2014 }
2015 }
2016 else {
2017 set_const(current,dops[i].rt1,cinfo[i].imm<<16); // LUI
2018 }
2019 dirty_reg(current,dops[i].rt1);
2020}
2021
2022static void load_alloc(struct regstat *current,int i)
2023{
2024 int need_temp = 0;
2025 clear_const(current,dops[i].rt1);
2026 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2027 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2028 if (needed_again(dops[i].rs1, i))
2029 alloc_reg(current, i, dops[i].rs1);
2030 if (ram_offset)
2031 alloc_reg(current, i, ROREG);
2032 if (dops[i].may_except) {
2033 alloc_cc_optional(current, i); // for exceptions
2034 need_temp = 1;
2035 }
2036 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2037 alloc_reg(current,i,dops[i].rt1);
2038 assert(get_reg_w(current->regmap, dops[i].rt1)>=0);
2039 dirty_reg(current,dops[i].rt1);
2040 // LWL/LWR need a temporary register for the old value
2041 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2042 {
2043 alloc_reg(current,i,FTEMP);
2044 need_temp = 1;
2045 }
2046 }
2047 else
2048 {
2049 // Load to r0 or unneeded register (dummy load)
2050 // but we still need a register to calculate the address
2051 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2052 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2053 need_temp = 1;
2054 }
2055 if (need_temp) {
2056 alloc_reg_temp(current, i, -1);
2057 cinfo[i].min_free_regs = 1;
2058 }
2059}
2060
2061// this may eat up to 7 registers
2062static void store_alloc(struct regstat *current, int i)
2063{
2064 clear_const(current,dops[i].rs2);
2065 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2066 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2067 alloc_reg(current,i,dops[i].rs2);
2068 if (ram_offset)
2069 alloc_reg(current, i, ROREG);
2070 #if defined(HOST_IMM8)
2071 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2072 alloc_reg(current, i, INVCP);
2073 #endif
2074 if (dops[i].opcode == 0x2a || dops[i].opcode == 0x2e) { // SWL/SWL
2075 alloc_reg(current,i,FTEMP);
2076 }
2077 if (dops[i].may_except)
2078 alloc_cc_optional(current, i); // for exceptions
2079 // We need a temporary register for address generation
2080 alloc_reg_temp(current,i,-1);
2081 cinfo[i].min_free_regs=1;
2082}
2083
2084static void c2ls_alloc(struct regstat *current, int i)
2085{
2086 clear_const(current,dops[i].rt1);
2087 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2088 alloc_reg(current,i,FTEMP);
2089 if (ram_offset)
2090 alloc_reg(current, i, ROREG);
2091 #if defined(HOST_IMM8)
2092 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2093 if (dops[i].opcode == 0x3a) // SWC2
2094 alloc_reg(current,i,INVCP);
2095 #endif
2096 if (dops[i].may_except)
2097 alloc_cc_optional(current, i); // for exceptions
2098 // We need a temporary register for address generation
2099 alloc_reg_temp(current,i,-1);
2100 cinfo[i].min_free_regs=1;
2101}
2102
2103#ifndef multdiv_alloc
2104static void multdiv_alloc(struct regstat *current,int i)
2105{
2106 // case 0x18: MULT
2107 // case 0x19: MULTU
2108 // case 0x1A: DIV
2109 // case 0x1B: DIVU
2110 clear_const(current,dops[i].rs1);
2111 clear_const(current,dops[i].rs2);
2112 alloc_cc(current,i); // for stalls
2113 dirty_reg(current,CCREG);
2114 if(dops[i].rs1&&dops[i].rs2)
2115 {
2116 current->u&=~(1LL<<HIREG);
2117 current->u&=~(1LL<<LOREG);
2118 alloc_reg(current,i,HIREG);
2119 alloc_reg(current,i,LOREG);
2120 alloc_reg(current,i,dops[i].rs1);
2121 alloc_reg(current,i,dops[i].rs2);
2122 dirty_reg(current,HIREG);
2123 dirty_reg(current,LOREG);
2124 }
2125 else
2126 {
2127 // Multiply by zero is zero.
2128 // MIPS does not have a divide by zero exception.
2129 alloc_reg(current,i,HIREG);
2130 alloc_reg(current,i,LOREG);
2131 dirty_reg(current,HIREG);
2132 dirty_reg(current,LOREG);
2133 if (dops[i].rs1 && ((dops[i].opcode2 & 0x3e) == 0x1a)) // div(u) 0
2134 alloc_reg(current, i, dops[i].rs1);
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_reg(current,i,dops[i].rt1);
2146 dirty_reg(current,dops[i].rt1);
2147 }
2148 }
2149 else if(dops[i].opcode2==4) // MTC0
2150 {
2151 if (((source[i]>>11)&0x1e) == 12) {
2152 alloc_cc(current, i);
2153 dirty_reg(current, CCREG);
2154 }
2155 if(dops[i].rs1){
2156 clear_const(current,dops[i].rs1);
2157 alloc_reg(current,i,dops[i].rs1);
2158 alloc_all(current,i);
2159 }
2160 else {
2161 alloc_all(current,i); // FIXME: Keep r0
2162 current->u&=~1LL;
2163 alloc_reg(current,i,0);
2164 }
2165 cinfo[i].min_free_regs = HOST_REGS;
2166 }
2167}
2168
2169static void rfe_alloc(struct regstat *current, int i)
2170{
2171 alloc_all(current, i);
2172 cinfo[i].min_free_regs = HOST_REGS;
2173}
2174
2175static void cop2_alloc(struct regstat *current,int i)
2176{
2177 if (dops[i].opcode2 < 3) // MFC2/CFC2
2178 {
2179 alloc_cc(current,i); // for stalls
2180 dirty_reg(current,CCREG);
2181 if(dops[i].rt1){
2182 clear_const(current,dops[i].rt1);
2183 alloc_reg(current,i,dops[i].rt1);
2184 dirty_reg(current,dops[i].rt1);
2185 }
2186 }
2187 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2188 {
2189 if(dops[i].rs1){
2190 clear_const(current,dops[i].rs1);
2191 alloc_reg(current,i,dops[i].rs1);
2192 }
2193 else {
2194 current->u&=~1LL;
2195 alloc_reg(current,i,0);
2196 }
2197 }
2198 alloc_reg_temp(current,i,-1);
2199 cinfo[i].min_free_regs=1;
2200}
2201
2202static void c2op_alloc(struct regstat *current,int i)
2203{
2204 alloc_cc(current,i); // for stalls
2205 dirty_reg(current,CCREG);
2206 alloc_reg_temp(current,i,-1);
2207}
2208
2209static void syscall_alloc(struct regstat *current,int i)
2210{
2211 alloc_cc(current,i);
2212 dirty_reg(current,CCREG);
2213 alloc_all(current,i);
2214 cinfo[i].min_free_regs=HOST_REGS;
2215 current->isconst=0;
2216}
2217
2218static void delayslot_alloc(struct regstat *current,int i)
2219{
2220 switch(dops[i].itype) {
2221 case UJUMP:
2222 case CJUMP:
2223 case SJUMP:
2224 case RJUMP:
2225 case SYSCALL:
2226 case HLECALL:
2227 case IMM16:
2228 imm16_alloc(current,i);
2229 break;
2230 case LOAD:
2231 case LOADLR:
2232 load_alloc(current,i);
2233 break;
2234 case STORE:
2235 case STORELR:
2236 store_alloc(current,i);
2237 break;
2238 case ALU:
2239 alu_alloc(current,i);
2240 break;
2241 case SHIFT:
2242 shift_alloc(current,i);
2243 break;
2244 case MULTDIV:
2245 multdiv_alloc(current,i);
2246 break;
2247 case SHIFTIMM:
2248 shiftimm_alloc(current,i);
2249 break;
2250 case MOV:
2251 mov_alloc(current,i);
2252 break;
2253 case COP0:
2254 cop0_alloc(current,i);
2255 break;
2256 case RFE:
2257 rfe_alloc(current,i);
2258 break;
2259 case COP2:
2260 cop2_alloc(current,i);
2261 break;
2262 case C2LS:
2263 c2ls_alloc(current,i);
2264 break;
2265 case C2OP:
2266 c2op_alloc(current,i);
2267 break;
2268 }
2269}
2270
2271static void add_stub(enum stub_type type, void *addr, void *retaddr,
2272 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2273{
2274 assert(stubcount < ARRAY_SIZE(stubs));
2275 stubs[stubcount].type = type;
2276 stubs[stubcount].addr = addr;
2277 stubs[stubcount].retaddr = retaddr;
2278 stubs[stubcount].a = a;
2279 stubs[stubcount].b = b;
2280 stubs[stubcount].c = c;
2281 stubs[stubcount].d = d;
2282 stubs[stubcount].e = e;
2283 stubcount++;
2284}
2285
2286static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2287 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2288{
2289 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2290}
2291
2292// Write out a single register
2293static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2294{
2295 int hr;
2296 for(hr=0;hr<HOST_REGS;hr++) {
2297 if(hr!=EXCLUDE_REG) {
2298 if(regmap[hr]==r) {
2299 if((dirty>>hr)&1) {
2300 assert(regmap[hr]<64);
2301 emit_storereg(r,hr);
2302 }
2303 }
2304 }
2305 }
2306}
2307
2308static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2309{
2310 //if(dirty_pre==dirty) return;
2311 int hr, r;
2312 for (hr = 0; hr < HOST_REGS; hr++) {
2313 r = pre[hr];
2314 if (r < 1 || r > 33 || ((u >> r) & 1))
2315 continue;
2316 if (((dirty_pre & ~dirty) >> hr) & 1)
2317 emit_storereg(r, hr);
2318 }
2319}
2320
2321// trashes r2
2322static void pass_args(int a0, int a1)
2323{
2324 if(a0==1&&a1==0) {
2325 // must swap
2326 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2327 }
2328 else if(a0!=0&&a1==0) {
2329 emit_mov(a1,1);
2330 if (a0>=0) emit_mov(a0,0);
2331 }
2332 else {
2333 if(a0>=0&&a0!=0) emit_mov(a0,0);
2334 if(a1>=0&&a1!=1) emit_mov(a1,1);
2335 }
2336}
2337
2338static void alu_assemble(int i, const struct regstat *i_regs, int ccadj_)
2339{
2340 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2341 int do_oflow = dops[i].may_except; // ADD/SUB with exceptions enabled
2342 if (dops[i].rt1 || do_oflow) {
2343 int do_exception_check = 0;
2344 signed char s1, s2, t, tmp;
2345 t = get_reg_w(i_regs->regmap, dops[i].rt1);
2346 tmp = get_reg_temp(i_regs->regmap);
2347 if (do_oflow)
2348 assert(tmp >= 0);
2349 if (t < 0 && do_oflow)
2350 t = tmp;
2351 if (t >= 0) {
2352 s1 = get_reg(i_regs->regmap, dops[i].rs1);
2353 s2 = get_reg(i_regs->regmap, dops[i].rs2);
2354 if (dops[i].rs1 && dops[i].rs2) {
2355 assert(s1>=0);
2356 assert(s2>=0);
2357 if (dops[i].opcode2 & 2) {
2358 if (do_oflow) {
2359 emit_subs(s1, s2, tmp);
2360 do_exception_check = 1;
2361 }
2362 else
2363 emit_sub(s1,s2,t);
2364 }
2365 else {
2366 if (do_oflow) {
2367 emit_adds(s1, s2, tmp);
2368 do_exception_check = 1;
2369 }
2370 else
2371 emit_add(s1,s2,t);
2372 }
2373 }
2374 else if(dops[i].rs1) {
2375 if(s1>=0) emit_mov(s1,t);
2376 else emit_loadreg(dops[i].rs1,t);
2377 }
2378 else if(dops[i].rs2) {
2379 if (s2 < 0) {
2380 emit_loadreg(dops[i].rs2, t);
2381 s2 = t;
2382 }
2383 if (dops[i].opcode2 & 2) {
2384 if (do_oflow) {
2385 emit_negs(s2, tmp);
2386 do_exception_check = 1;
2387 }
2388 else
2389 emit_neg(s2, t);
2390 }
2391 else if (s2 != t)
2392 emit_mov(s2, t);
2393 }
2394 else
2395 emit_zeroreg(t);
2396 }
2397 if (do_exception_check) {
2398 void *jaddr = out;
2399 emit_jo(0);
2400 if (t >= 0 && tmp != t)
2401 emit_mov(tmp, t);
2402 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2403 }
2404 }
2405 }
2406 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2407 if(dops[i].rt1) {
2408 signed char s1l,s2l,t;
2409 {
2410 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2411 //assert(t>=0);
2412 if(t>=0) {
2413 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2414 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2415 if(dops[i].rs2==0) // rx<r0
2416 {
2417 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2418 assert(s1l>=0);
2419 emit_shrimm(s1l,31,t);
2420 }
2421 else // SLTU (unsigned can not be less than zero, 0<0)
2422 emit_zeroreg(t);
2423 }
2424 else if(dops[i].rs1==0) // r0<rx
2425 {
2426 assert(s2l>=0);
2427 if(dops[i].opcode2==0x2a) // SLT
2428 emit_set_gz32(s2l,t);
2429 else // SLTU (set if not zero)
2430 emit_set_nz32(s2l,t);
2431 }
2432 else{
2433 assert(s1l>=0);assert(s2l>=0);
2434 if(dops[i].opcode2==0x2a) // SLT
2435 emit_set_if_less32(s1l,s2l,t);
2436 else // SLTU
2437 emit_set_if_carry32(s1l,s2l,t);
2438 }
2439 }
2440 }
2441 }
2442 }
2443 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2444 if(dops[i].rt1) {
2445 signed char s1l,s2l,tl;
2446 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2447 {
2448 if(tl>=0) {
2449 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2450 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2451 if(dops[i].rs1&&dops[i].rs2) {
2452 assert(s1l>=0);
2453 assert(s2l>=0);
2454 if(dops[i].opcode2==0x24) { // AND
2455 emit_and(s1l,s2l,tl);
2456 } else
2457 if(dops[i].opcode2==0x25) { // OR
2458 emit_or(s1l,s2l,tl);
2459 } else
2460 if(dops[i].opcode2==0x26) { // XOR
2461 emit_xor(s1l,s2l,tl);
2462 } else
2463 if(dops[i].opcode2==0x27) { // NOR
2464 emit_or(s1l,s2l,tl);
2465 emit_not(tl,tl);
2466 }
2467 }
2468 else
2469 {
2470 if(dops[i].opcode2==0x24) { // AND
2471 emit_zeroreg(tl);
2472 } else
2473 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2474 if(dops[i].rs1){
2475 if(s1l>=0) emit_mov(s1l,tl);
2476 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2477 }
2478 else
2479 if(dops[i].rs2){
2480 if(s2l>=0) emit_mov(s2l,tl);
2481 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2482 }
2483 else emit_zeroreg(tl);
2484 } else
2485 if(dops[i].opcode2==0x27) { // NOR
2486 if(dops[i].rs1){
2487 if(s1l>=0) emit_not(s1l,tl);
2488 else {
2489 emit_loadreg(dops[i].rs1,tl);
2490 emit_not(tl,tl);
2491 }
2492 }
2493 else
2494 if(dops[i].rs2){
2495 if(s2l>=0) emit_not(s2l,tl);
2496 else {
2497 emit_loadreg(dops[i].rs2,tl);
2498 emit_not(tl,tl);
2499 }
2500 }
2501 else emit_movimm(-1,tl);
2502 }
2503 }
2504 }
2505 }
2506 }
2507 }
2508}
2509
2510static void imm16_assemble(int i, const struct regstat *i_regs, int ccadj_)
2511{
2512 if (dops[i].opcode==0x0f) { // LUI
2513 if(dops[i].rt1) {
2514 signed char t;
2515 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2516 //assert(t>=0);
2517 if(t>=0) {
2518 if(!((i_regs->isconst>>t)&1))
2519 emit_movimm(cinfo[i].imm<<16,t);
2520 }
2521 }
2522 }
2523 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2524 int is_addi = dops[i].may_except;
2525 if (dops[i].rt1 || is_addi) {
2526 signed char s, t, tmp;
2527 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2528 s=get_reg(i_regs->regmap,dops[i].rs1);
2529 if(dops[i].rs1) {
2530 tmp = get_reg_temp(i_regs->regmap);
2531 if (is_addi) {
2532 assert(tmp >= 0);
2533 if (t < 0) t = tmp;
2534 }
2535 if(t>=0) {
2536 if(!((i_regs->isconst>>t)&1)) {
2537 int sum, do_exception_check = 0;
2538 if (s < 0) {
2539 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2540 if (is_addi) {
2541 emit_addimm_and_set_flags3(t, cinfo[i].imm, tmp);
2542 do_exception_check = 1;
2543 }
2544 else
2545 emit_addimm(t, cinfo[i].imm, t);
2546 } else {
2547 if (!((i_regs->wasconst >> s) & 1)) {
2548 if (is_addi) {
2549 emit_addimm_and_set_flags3(s, cinfo[i].imm, tmp);
2550 do_exception_check = 1;
2551 }
2552 else
2553 emit_addimm(s, cinfo[i].imm, t);
2554 }
2555 else {
2556 int oflow = add_overflow(constmap[i][s], cinfo[i].imm, sum);
2557 if (is_addi && oflow)
2558 do_exception_check = 2;
2559 else
2560 emit_movimm(sum, t);
2561 }
2562 }
2563 if (do_exception_check) {
2564 void *jaddr = out;
2565 if (do_exception_check == 2)
2566 emit_jmp(0);
2567 else {
2568 emit_jo(0);
2569 if (tmp != t)
2570 emit_mov(tmp, t);
2571 }
2572 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2573 }
2574 }
2575 }
2576 } else {
2577 if(t>=0) {
2578 if(!((i_regs->isconst>>t)&1))
2579 emit_movimm(cinfo[i].imm,t);
2580 }
2581 }
2582 }
2583 }
2584 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2585 if(dops[i].rt1) {
2586 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2587 signed char sl,t;
2588 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2589 sl=get_reg(i_regs->regmap,dops[i].rs1);
2590 //assert(t>=0);
2591 if(t>=0) {
2592 if(dops[i].rs1>0) {
2593 if(dops[i].opcode==0x0a) { // SLTI
2594 if(sl<0) {
2595 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2596 emit_slti32(t,cinfo[i].imm,t);
2597 }else{
2598 emit_slti32(sl,cinfo[i].imm,t);
2599 }
2600 }
2601 else { // SLTIU
2602 if(sl<0) {
2603 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2604 emit_sltiu32(t,cinfo[i].imm,t);
2605 }else{
2606 emit_sltiu32(sl,cinfo[i].imm,t);
2607 }
2608 }
2609 }else{
2610 // SLTI(U) with r0 is just stupid,
2611 // nonetheless examples can be found
2612 if(dops[i].opcode==0x0a) // SLTI
2613 if(0<cinfo[i].imm) emit_movimm(1,t);
2614 else emit_zeroreg(t);
2615 else // SLTIU
2616 {
2617 if(cinfo[i].imm) emit_movimm(1,t);
2618 else emit_zeroreg(t);
2619 }
2620 }
2621 }
2622 }
2623 }
2624 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2625 if(dops[i].rt1) {
2626 signed char sl,tl;
2627 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2628 sl=get_reg(i_regs->regmap,dops[i].rs1);
2629 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2630 if(dops[i].opcode==0x0c) //ANDI
2631 {
2632 if(dops[i].rs1) {
2633 if(sl<0) {
2634 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2635 emit_andimm(tl,cinfo[i].imm,tl);
2636 }else{
2637 if(!((i_regs->wasconst>>sl)&1))
2638 emit_andimm(sl,cinfo[i].imm,tl);
2639 else
2640 emit_movimm(constmap[i][sl]&cinfo[i].imm,tl);
2641 }
2642 }
2643 else
2644 emit_zeroreg(tl);
2645 }
2646 else
2647 {
2648 if(dops[i].rs1) {
2649 if(sl<0) {
2650 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2651 }
2652 if(dops[i].opcode==0x0d) { // ORI
2653 if(sl<0) {
2654 emit_orimm(tl,cinfo[i].imm,tl);
2655 }else{
2656 if(!((i_regs->wasconst>>sl)&1))
2657 emit_orimm(sl,cinfo[i].imm,tl);
2658 else
2659 emit_movimm(constmap[i][sl]|cinfo[i].imm,tl);
2660 }
2661 }
2662 if(dops[i].opcode==0x0e) { // XORI
2663 if(sl<0) {
2664 emit_xorimm(tl,cinfo[i].imm,tl);
2665 }else{
2666 if(!((i_regs->wasconst>>sl)&1))
2667 emit_xorimm(sl,cinfo[i].imm,tl);
2668 else
2669 emit_movimm(constmap[i][sl]^cinfo[i].imm,tl);
2670 }
2671 }
2672 }
2673 else {
2674 emit_movimm(cinfo[i].imm,tl);
2675 }
2676 }
2677 }
2678 }
2679 }
2680}
2681
2682static void shiftimm_assemble(int i, const struct regstat *i_regs)
2683{
2684 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2685 {
2686 if(dops[i].rt1) {
2687 signed char s,t;
2688 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2689 s=get_reg(i_regs->regmap,dops[i].rs1);
2690 //assert(t>=0);
2691 if(t>=0&&!((i_regs->isconst>>t)&1)){
2692 if(dops[i].rs1==0)
2693 {
2694 emit_zeroreg(t);
2695 }
2696 else
2697 {
2698 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2699 if(cinfo[i].imm) {
2700 if(dops[i].opcode2==0) // SLL
2701 {
2702 emit_shlimm(s<0?t:s,cinfo[i].imm,t);
2703 }
2704 if(dops[i].opcode2==2) // SRL
2705 {
2706 emit_shrimm(s<0?t:s,cinfo[i].imm,t);
2707 }
2708 if(dops[i].opcode2==3) // SRA
2709 {
2710 emit_sarimm(s<0?t:s,cinfo[i].imm,t);
2711 }
2712 }else{
2713 // Shift by zero
2714 if(s>=0 && s!=t) emit_mov(s,t);
2715 }
2716 }
2717 }
2718 //emit_storereg(dops[i].rt1,t); //DEBUG
2719 }
2720 }
2721 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2722 {
2723 assert(0);
2724 }
2725 if(dops[i].opcode2==0x3c) // DSLL32
2726 {
2727 assert(0);
2728 }
2729 if(dops[i].opcode2==0x3e) // DSRL32
2730 {
2731 assert(0);
2732 }
2733 if(dops[i].opcode2==0x3f) // DSRA32
2734 {
2735 assert(0);
2736 }
2737}
2738
2739#ifndef shift_assemble
2740static void shift_assemble(int i, const struct regstat *i_regs)
2741{
2742 signed char s,t,shift;
2743 if (dops[i].rt1 == 0)
2744 return;
2745 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2746 t = get_reg(i_regs->regmap, dops[i].rt1);
2747 s = get_reg(i_regs->regmap, dops[i].rs1);
2748 shift = get_reg(i_regs->regmap, dops[i].rs2);
2749 if (t < 0)
2750 return;
2751
2752 if(dops[i].rs1==0)
2753 emit_zeroreg(t);
2754 else if(dops[i].rs2==0) {
2755 assert(s>=0);
2756 if(s!=t) emit_mov(s,t);
2757 }
2758 else {
2759 host_tempreg_acquire();
2760 emit_andimm(shift,31,HOST_TEMPREG);
2761 switch(dops[i].opcode2) {
2762 case 4: // SLLV
2763 emit_shl(s,HOST_TEMPREG,t);
2764 break;
2765 case 6: // SRLV
2766 emit_shr(s,HOST_TEMPREG,t);
2767 break;
2768 case 7: // SRAV
2769 emit_sar(s,HOST_TEMPREG,t);
2770 break;
2771 default:
2772 assert(0);
2773 }
2774 host_tempreg_release();
2775 }
2776}
2777
2778#endif
2779
2780enum {
2781 MTYPE_8000 = 0,
2782 MTYPE_8020,
2783 MTYPE_0000,
2784 MTYPE_A000,
2785 MTYPE_1F80,
2786};
2787
2788static int get_ptr_mem_type(u_int a)
2789{
2790 if(a < 0x00200000) {
2791 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2792 // return wrong, must use memhandler for BIOS self-test to pass
2793 // 007 does similar stuff from a00 mirror, weird stuff
2794 return MTYPE_8000;
2795 return MTYPE_0000;
2796 }
2797 if(0x1f800000 <= a && a < 0x1f801000)
2798 return MTYPE_1F80;
2799 if(0x80200000 <= a && a < 0x80800000)
2800 return MTYPE_8020;
2801 if(0xa0000000 <= a && a < 0xa0200000)
2802 return MTYPE_A000;
2803 return MTYPE_8000;
2804}
2805
2806static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2807{
2808 int r = get_reg(i_regs->regmap, ROREG);
2809 if (r < 0 && host_tempreg_free) {
2810 host_tempreg_acquire();
2811 emit_loadreg(ROREG, r = HOST_TEMPREG);
2812 }
2813 if (r < 0)
2814 abort();
2815 return r;
2816}
2817
2818static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2819 int addr, int *offset_reg, int *addr_reg_override, int ccadj_)
2820{
2821 void *jaddr = NULL;
2822 int type = 0;
2823 int mr = dops[i].rs1;
2824 assert(addr >= 0);
2825 *offset_reg = -1;
2826 if(((smrv_strong|smrv_weak)>>mr)&1) {
2827 type=get_ptr_mem_type(smrv[mr]);
2828 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2829 }
2830 else {
2831 // use the mirror we are running on
2832 type=get_ptr_mem_type(start);
2833 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2834 }
2835
2836 if (dops[i].may_except) {
2837 // alignment check
2838 u_int op = dops[i].opcode;
2839 int mask = ((op & 0x37) == 0x21 || op == 0x25) ? 1 : 3; // LH/SH/LHU
2840 void *jaddr;
2841 emit_testimm(addr, mask);
2842 jaddr = out;
2843 emit_jne(0);
2844 add_stub_r(ALIGNMENT_STUB, jaddr, out, i, addr, i_regs, ccadj_, 0);
2845 }
2846
2847 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2848 host_tempreg_acquire();
2849 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2850 addr=*addr_reg_override=HOST_TEMPREG;
2851 type=0;
2852 }
2853 else if(type==MTYPE_0000) { // RAM 0 mirror
2854 host_tempreg_acquire();
2855 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2856 addr=*addr_reg_override=HOST_TEMPREG;
2857 type=0;
2858 }
2859 else if(type==MTYPE_A000) { // RAM A mirror
2860 host_tempreg_acquire();
2861 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2862 addr=*addr_reg_override=HOST_TEMPREG;
2863 type=0;
2864 }
2865 else if(type==MTYPE_1F80) { // scratchpad
2866 if (psxH == (void *)0x1f800000) {
2867 host_tempreg_acquire();
2868 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2869 emit_cmpimm(HOST_TEMPREG,0x1000);
2870 host_tempreg_release();
2871 jaddr=out;
2872 emit_jc(0);
2873 }
2874 else {
2875 // do the usual RAM check, jump will go to the right handler
2876 type=0;
2877 }
2878 }
2879
2880 if (type == 0) // need ram check
2881 {
2882 emit_cmpimm(addr,RAM_SIZE);
2883 jaddr = out;
2884 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2885 // Hint to branch predictor that the branch is unlikely to be taken
2886 if (dops[i].rs1 >= 28)
2887 emit_jno_unlikely(0);
2888 else
2889 #endif
2890 emit_jno(0);
2891 if (ram_offset != 0)
2892 *offset_reg = get_ro_reg(i_regs, 0);
2893 }
2894
2895 return jaddr;
2896}
2897
2898// return memhandler, or get directly accessable address and return 0
2899static void *get_direct_memhandler(void *table, u_int addr,
2900 enum stub_type type, uintptr_t *addr_host)
2901{
2902 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2903 uintptr_t l1, l2 = 0;
2904 l1 = ((uintptr_t *)table)[addr>>12];
2905 if (!(l1 & msb)) {
2906 uintptr_t v = l1 << 1;
2907 *addr_host = v + addr;
2908 return NULL;
2909 }
2910 else {
2911 l1 <<= 1;
2912 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2913 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2914 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2915 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2916 else
2917 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2918 if (!(l2 & msb)) {
2919 uintptr_t v = l2 << 1;
2920 *addr_host = v + (addr&0xfff);
2921 return NULL;
2922 }
2923 return (void *)(l2 << 1);
2924 }
2925}
2926
2927static u_int get_host_reglist(const signed char *regmap)
2928{
2929 u_int reglist = 0, hr;
2930 for (hr = 0; hr < HOST_REGS; hr++) {
2931 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2932 reglist |= 1 << hr;
2933 }
2934 return reglist;
2935}
2936
2937static u_int reglist_exclude(u_int reglist, int r1, int r2)
2938{
2939 if (r1 >= 0)
2940 reglist &= ~(1u << r1);
2941 if (r2 >= 0)
2942 reglist &= ~(1u << r2);
2943 return reglist;
2944}
2945
2946// find a temp caller-saved register not in reglist (so assumed to be free)
2947static int reglist_find_free(u_int reglist)
2948{
2949 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2950 if (free_regs == 0)
2951 return -1;
2952 return __builtin_ctz(free_regs);
2953}
2954
2955static void do_load_word(int a, int rt, int offset_reg)
2956{
2957 if (offset_reg >= 0)
2958 emit_ldr_dualindexed(offset_reg, a, rt);
2959 else
2960 emit_readword_indexed(0, a, rt);
2961}
2962
2963static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2964{
2965 if (offset_reg < 0) {
2966 emit_writeword_indexed(rt, ofs, a);
2967 return;
2968 }
2969 if (ofs != 0)
2970 emit_addimm(a, ofs, a);
2971 emit_str_dualindexed(offset_reg, a, rt);
2972 if (ofs != 0 && preseve_a)
2973 emit_addimm(a, -ofs, a);
2974}
2975
2976static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2977{
2978 if (offset_reg < 0) {
2979 emit_writehword_indexed(rt, ofs, a);
2980 return;
2981 }
2982 if (ofs != 0)
2983 emit_addimm(a, ofs, a);
2984 emit_strh_dualindexed(offset_reg, a, rt);
2985 if (ofs != 0 && preseve_a)
2986 emit_addimm(a, -ofs, a);
2987}
2988
2989static void do_store_byte(int a, int rt, int offset_reg)
2990{
2991 if (offset_reg >= 0)
2992 emit_strb_dualindexed(offset_reg, a, rt);
2993 else
2994 emit_writebyte_indexed(rt, 0, a);
2995}
2996
2997static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2998{
2999 int addr = cinfo[i].addr;
3000 int s,tl;
3001 int offset;
3002 void *jaddr=0;
3003 int memtarget=0,c=0;
3004 int offset_reg = -1;
3005 int fastio_reg_override = -1;
3006 u_int reglist=get_host_reglist(i_regs->regmap);
3007 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3008 s=get_reg(i_regs->regmap,dops[i].rs1);
3009 offset=cinfo[i].imm;
3010 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3011 if(s>=0) {
3012 c=(i_regs->wasconst>>s)&1;
3013 if (c) {
3014 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3015 }
3016 }
3017 //printf("load_assemble: c=%d\n",c);
3018 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3019 if(tl<0 && ((!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80) || dops[i].rt1==0)) {
3020 // could be FIFO, must perform the read
3021 // ||dummy read
3022 assem_debug("(forced read)\n");
3023 tl = get_reg_temp(i_regs->regmap); // may be == addr
3024 assert(tl>=0);
3025 }
3026 assert(addr >= 0);
3027 if(tl>=0) {
3028 //printf("load_assemble: c=%d\n",c);
3029 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3030 reglist&=~(1<<tl);
3031 if(!c) {
3032 #ifdef R29_HACK
3033 // Strmnnrmn's speed hack
3034 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3035 #endif
3036 {
3037 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3038 &offset_reg, &fastio_reg_override, ccadj_);
3039 }
3040 }
3041 else if (ram_offset && memtarget) {
3042 offset_reg = get_ro_reg(i_regs, 0);
3043 }
3044 int dummy=(dops[i].rt1==0)||(tl!=get_reg_w(i_regs->regmap, dops[i].rt1)); // ignore loads to r0 and unneeded reg
3045 switch (dops[i].opcode) {
3046 case 0x20: // LB
3047 if(!c||memtarget) {
3048 if(!dummy) {
3049 int a = addr;
3050 if (fastio_reg_override >= 0)
3051 a = fastio_reg_override;
3052
3053 if (offset_reg >= 0)
3054 emit_ldrsb_dualindexed(offset_reg, a, tl);
3055 else
3056 emit_movsbl_indexed(0, a, tl);
3057 }
3058 if(jaddr)
3059 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3060 }
3061 else
3062 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3063 break;
3064 case 0x21: // LH
3065 if(!c||memtarget) {
3066 if(!dummy) {
3067 int a = addr;
3068 if (fastio_reg_override >= 0)
3069 a = fastio_reg_override;
3070 if (offset_reg >= 0)
3071 emit_ldrsh_dualindexed(offset_reg, a, tl);
3072 else
3073 emit_movswl_indexed(0, a, tl);
3074 }
3075 if(jaddr)
3076 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3077 }
3078 else
3079 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3080 break;
3081 case 0x23: // LW
3082 if(!c||memtarget) {
3083 if(!dummy) {
3084 int a = addr;
3085 if (fastio_reg_override >= 0)
3086 a = fastio_reg_override;
3087 do_load_word(a, tl, offset_reg);
3088 }
3089 if(jaddr)
3090 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3091 }
3092 else
3093 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3094 break;
3095 case 0x24: // LBU
3096 if(!c||memtarget) {
3097 if(!dummy) {
3098 int a = addr;
3099 if (fastio_reg_override >= 0)
3100 a = fastio_reg_override;
3101
3102 if (offset_reg >= 0)
3103 emit_ldrb_dualindexed(offset_reg, a, tl);
3104 else
3105 emit_movzbl_indexed(0, a, tl);
3106 }
3107 if(jaddr)
3108 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3109 }
3110 else
3111 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3112 break;
3113 case 0x25: // LHU
3114 if(!c||memtarget) {
3115 if(!dummy) {
3116 int a = addr;
3117 if (fastio_reg_override >= 0)
3118 a = fastio_reg_override;
3119 if (offset_reg >= 0)
3120 emit_ldrh_dualindexed(offset_reg, a, tl);
3121 else
3122 emit_movzwl_indexed(0, a, tl);
3123 }
3124 if(jaddr)
3125 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3126 }
3127 else
3128 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3129 break;
3130 default:
3131 assert(0);
3132 }
3133 } // tl >= 0
3134 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3135 host_tempreg_release();
3136}
3137
3138#ifndef loadlr_assemble
3139static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3140{
3141 int addr = cinfo[i].addr;
3142 int s,tl,temp,temp2;
3143 int offset;
3144 void *jaddr=0;
3145 int memtarget=0,c=0;
3146 int offset_reg = -1;
3147 int fastio_reg_override = -1;
3148 u_int reglist=get_host_reglist(i_regs->regmap);
3149 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3150 s=get_reg(i_regs->regmap,dops[i].rs1);
3151 temp=get_reg_temp(i_regs->regmap);
3152 temp2=get_reg(i_regs->regmap,FTEMP);
3153 offset=cinfo[i].imm;
3154 reglist|=1<<temp;
3155 assert(addr >= 0);
3156 if(s>=0) {
3157 c=(i_regs->wasconst>>s)&1;
3158 if(c) {
3159 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3160 }
3161 }
3162 if(!c) {
3163 emit_shlimm(addr,3,temp);
3164 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3165 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3166 }else{
3167 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3168 }
3169 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3170 &offset_reg, &fastio_reg_override, ccadj_);
3171 }
3172 else {
3173 if (ram_offset && memtarget) {
3174 offset_reg = get_ro_reg(i_regs, 0);
3175 }
3176 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3177 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3178 }else{
3179 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3180 }
3181 }
3182 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3183 if(!c||memtarget) {
3184 int a = temp2;
3185 if (fastio_reg_override >= 0)
3186 a = fastio_reg_override;
3187 do_load_word(a, temp2, offset_reg);
3188 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3189 host_tempreg_release();
3190 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3191 }
3192 else
3193 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3194 if(dops[i].rt1) {
3195 assert(tl>=0);
3196 emit_andimm(temp,24,temp);
3197 if (dops[i].opcode==0x22) // LWL
3198 emit_xorimm(temp,24,temp);
3199 host_tempreg_acquire();
3200 emit_movimm(-1,HOST_TEMPREG);
3201 if (dops[i].opcode==0x26) {
3202 emit_shr(temp2,temp,temp2);
3203 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3204 }else{
3205 emit_shl(temp2,temp,temp2);
3206 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3207 }
3208 host_tempreg_release();
3209 emit_or(temp2,tl,tl);
3210 }
3211 //emit_storereg(dops[i].rt1,tl); // DEBUG
3212 }
3213 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3214 assert(0);
3215 }
3216}
3217#endif
3218
3219static void do_invstub(int n)
3220{
3221 literal_pool(20);
3222 assem_debug("do_invstub\n");
3223 u_int reglist = stubs[n].a;
3224 u_int addrr = stubs[n].b;
3225 int ofs_start = stubs[n].c;
3226 int ofs_end = stubs[n].d;
3227 int len = ofs_end - ofs_start;
3228 u_int rightr = 0;
3229
3230 set_jump_target(stubs[n].addr, out);
3231 save_regs(reglist);
3232 if (addrr != 0 || ofs_start != 0)
3233 emit_addimm(addrr, ofs_start, 0);
3234 emit_readword(&inv_code_start, 2);
3235 emit_readword(&inv_code_end, 3);
3236 if (len != 0)
3237 emit_addimm(0, len + 4, (rightr = 1));
3238 emit_cmp(0, 2);
3239 emit_cmpcs(3, rightr);
3240 void *jaddr = out;
3241 emit_jc(0);
3242 void *func = (len != 0)
3243 ? (void *)ndrc_write_invalidate_many
3244 : (void *)ndrc_write_invalidate_one;
3245 emit_far_call(func);
3246 set_jump_target(jaddr, out);
3247 restore_regs(reglist);
3248 emit_jmp(stubs[n].retaddr);
3249}
3250
3251static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3252{
3253 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3254 return;
3255 // this can't be used any more since we started to check exact
3256 // block boundaries in invalidate_range()
3257 //if (i_regs->waswritten & (1<<dops[i].rs1))
3258 // return;
3259 // (naively) assume nobody will run code from stack
3260 if (dops[i].rs1 == 29)
3261 return;
3262
3263 int j, imm_maxdiff = 32, imm_min = cinfo[i].imm, imm_max = cinfo[i].imm, count = 1;
3264 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3265 && abs(cinfo[i+1].imm - cinfo[i].imm) <= imm_maxdiff)
3266 return;
3267 for (j = i - 1; j >= 0; j--) {
3268 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3269 || abs(cinfo[j].imm - cinfo[j+1].imm) > imm_maxdiff)
3270 break;
3271 count++;
3272 if (imm_min > cinfo[j].imm)
3273 imm_min = cinfo[j].imm;
3274 if (imm_max < cinfo[j].imm)
3275 imm_max = cinfo[j].imm;
3276 }
3277#if defined(HOST_IMM8)
3278 int ir = get_reg(i_regs->regmap, INVCP);
3279 assert(ir >= 0);
3280 host_tempreg_acquire();
3281 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3282#else
3283 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3284 #error not handled
3285#endif
3286#ifdef INVALIDATE_USE_COND_CALL
3287 if (count == 1) {
3288 emit_cmpimm(HOST_TEMPREG, 1);
3289 emit_callne(invalidate_addr_reg[addr]);
3290 host_tempreg_release();
3291 return;
3292 }
3293#endif
3294 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3295 host_tempreg_release();
3296 imm_min -= cinfo[i].imm;
3297 imm_max -= cinfo[i].imm;
3298 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3299 addr, imm_min, imm_max, 0);
3300}
3301
3302static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3303{
3304 int s,tl;
3305 int addr = cinfo[i].addr;
3306 int offset;
3307 void *jaddr=0;
3308 enum stub_type type=0;
3309 int memtarget=0,c=0;
3310 int offset_reg = -1;
3311 int fastio_reg_override = -1;
3312 u_int reglist=get_host_reglist(i_regs->regmap);
3313 tl=get_reg(i_regs->regmap,dops[i].rs2);
3314 s=get_reg(i_regs->regmap,dops[i].rs1);
3315 offset=cinfo[i].imm;
3316 if(s>=0) {
3317 c=(i_regs->wasconst>>s)&1;
3318 if(c) {
3319 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3320 }
3321 }
3322 assert(tl>=0);
3323 assert(addr >= 0);
3324 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3325 if (!c) {
3326 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3327 &offset_reg, &fastio_reg_override, ccadj_);
3328 }
3329 else if (ram_offset && memtarget) {
3330 offset_reg = get_ro_reg(i_regs, 0);
3331 }
3332
3333 switch (dops[i].opcode) {
3334 case 0x28: // SB
3335 if(!c||memtarget) {
3336 int a = addr;
3337 if (fastio_reg_override >= 0)
3338 a = fastio_reg_override;
3339 do_store_byte(a, tl, offset_reg);
3340 }
3341 type = STOREB_STUB;
3342 break;
3343 case 0x29: // SH
3344 if(!c||memtarget) {
3345 int a = addr;
3346 if (fastio_reg_override >= 0)
3347 a = fastio_reg_override;
3348 do_store_hword(a, 0, tl, offset_reg, 1);
3349 }
3350 type = STOREH_STUB;
3351 break;
3352 case 0x2B: // SW
3353 if(!c||memtarget) {
3354 int a = addr;
3355 if (fastio_reg_override >= 0)
3356 a = fastio_reg_override;
3357 do_store_word(a, 0, tl, offset_reg, 1);
3358 }
3359 type = STOREW_STUB;
3360 break;
3361 default:
3362 assert(0);
3363 }
3364 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3365 host_tempreg_release();
3366 if(jaddr) {
3367 // PCSX store handlers don't check invcode again
3368 reglist|=1<<addr;
3369 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3370 jaddr=0;
3371 }
3372 {
3373 if(!c||memtarget) {
3374 do_store_smc_check(i, i_regs, reglist, addr);
3375 }
3376 }
3377 u_int addr_val=constmap[i][s]+offset;
3378 if(jaddr) {
3379 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3380 } else if(c&&!memtarget) {
3381 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3382 }
3383 // basic current block modification detection..
3384 // not looking back as that should be in mips cache already
3385 // (see Spyro2 title->attract mode)
3386 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3387 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3388 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3389 if(i_regs->regmap==regs[i].regmap) {
3390 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3391 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3392 emit_movimm(start+i*4+4,0);
3393 emit_writeword(0,&pcaddr);
3394 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3395 emit_far_call(ndrc_get_addr_ht);
3396 emit_jmpreg(0);
3397 }
3398 }
3399}
3400
3401static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3402{
3403 int addr = cinfo[i].addr;
3404 int s,tl;
3405 int offset;
3406 void *jaddr=0;
3407 void *case1, *case23, *case3;
3408 void *done0, *done1, *done2;
3409 int memtarget=0,c=0;
3410 int offset_reg = -1;
3411 u_int reglist=get_host_reglist(i_regs->regmap);
3412 tl=get_reg(i_regs->regmap,dops[i].rs2);
3413 s=get_reg(i_regs->regmap,dops[i].rs1);
3414 offset=cinfo[i].imm;
3415 if(s>=0) {
3416 c=(i_regs->isconst>>s)&1;
3417 if(c) {
3418 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3419 }
3420 }
3421 assert(tl>=0);
3422 assert(addr >= 0);
3423 if(!c) {
3424 emit_cmpimm(addr, RAM_SIZE);
3425 jaddr=out;
3426 emit_jno(0);
3427 }
3428 else
3429 {
3430 if(!memtarget||!dops[i].rs1) {
3431 jaddr=out;
3432 emit_jmp(0);
3433 }
3434 }
3435 if (ram_offset)
3436 offset_reg = get_ro_reg(i_regs, 0);
3437
3438 emit_testimm(addr,2);
3439 case23=out;
3440 emit_jne(0);
3441 emit_testimm(addr,1);
3442 case1=out;
3443 emit_jne(0);
3444 // 0
3445 if (dops[i].opcode == 0x2A) { // SWL
3446 // Write msb into least significant byte
3447 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3448 do_store_byte(addr, tl, offset_reg);
3449 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3450 }
3451 else if (dops[i].opcode == 0x2E) { // SWR
3452 // Write entire word
3453 do_store_word(addr, 0, tl, offset_reg, 1);
3454 }
3455 done0 = out;
3456 emit_jmp(0);
3457 // 1
3458 set_jump_target(case1, out);
3459 if (dops[i].opcode == 0x2A) { // SWL
3460 // Write two msb into two least significant bytes
3461 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3462 do_store_hword(addr, -1, tl, offset_reg, 0);
3463 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3464 }
3465 else if (dops[i].opcode == 0x2E) { // SWR
3466 // Write 3 lsb into three most significant bytes
3467 do_store_byte(addr, tl, offset_reg);
3468 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3469 do_store_hword(addr, 1, tl, offset_reg, 0);
3470 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3471 }
3472 done1=out;
3473 emit_jmp(0);
3474 // 2,3
3475 set_jump_target(case23, out);
3476 emit_testimm(addr,1);
3477 case3 = out;
3478 emit_jne(0);
3479 // 2
3480 if (dops[i].opcode==0x2A) { // SWL
3481 // Write 3 msb into three least significant bytes
3482 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3483 do_store_hword(addr, -2, tl, offset_reg, 1);
3484 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3485 do_store_byte(addr, tl, offset_reg);
3486 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3487 }
3488 else if (dops[i].opcode == 0x2E) { // SWR
3489 // Write two lsb into two most significant bytes
3490 do_store_hword(addr, 0, tl, offset_reg, 1);
3491 }
3492 done2 = out;
3493 emit_jmp(0);
3494 // 3
3495 set_jump_target(case3, out);
3496 if (dops[i].opcode == 0x2A) { // SWL
3497 do_store_word(addr, -3, tl, offset_reg, 0);
3498 }
3499 else if (dops[i].opcode == 0x2E) { // SWR
3500 do_store_byte(addr, tl, offset_reg);
3501 }
3502 set_jump_target(done0, out);
3503 set_jump_target(done1, out);
3504 set_jump_target(done2, out);
3505 if (offset_reg == HOST_TEMPREG)
3506 host_tempreg_release();
3507 if(!c||!memtarget)
3508 add_stub_r(STORELR_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3509 do_store_smc_check(i, i_regs, reglist, addr);
3510}
3511
3512static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3513{
3514 if(dops[i].opcode2==0) // MFC0
3515 {
3516 signed char t=get_reg_w(i_regs->regmap, dops[i].rt1);
3517 u_int copr=(source[i]>>11)&0x1f;
3518 if(t>=0&&dops[i].rt1!=0) {
3519 emit_readword(&reg_cop0[copr],t);
3520 }
3521 }
3522 else if(dops[i].opcode2==4) // MTC0
3523 {
3524 int s = get_reg(i_regs->regmap, dops[i].rs1);
3525 int cc = get_reg(i_regs->regmap, CCREG);
3526 char copr=(source[i]>>11)&0x1f;
3527 assert(s>=0);
3528 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3529 if (copr == 12 || copr == 13) {
3530 emit_readword(&last_count,HOST_TEMPREG);
3531 if (cc != HOST_CCREG)
3532 emit_loadreg(CCREG, HOST_CCREG);
3533 emit_add(HOST_CCREG, HOST_TEMPREG, HOST_CCREG);
3534 emit_addimm(HOST_CCREG, ccadj_ + 2, HOST_CCREG);
3535 emit_writeword(HOST_CCREG, &psxRegs.cycle);
3536 if (is_delayslot) {
3537 // burn cycles to cause cc_interrupt, which will
3538 // reschedule next_interupt. Relies on CCREG from above.
3539 assem_debug("MTC0 DS %d\n", copr);
3540 emit_writeword(HOST_CCREG,&last_count);
3541 emit_movimm(0,HOST_CCREG);
3542 emit_storereg(CCREG,HOST_CCREG);
3543 emit_loadreg(dops[i].rs1,1);
3544 emit_movimm(copr,0);
3545 emit_far_call(pcsx_mtc0_ds);
3546 emit_loadreg(dops[i].rs1,s);
3547 return;
3548 }
3549 emit_movimm(start+i*4+4,HOST_TEMPREG);
3550 emit_writeword(HOST_TEMPREG,&pcaddr);
3551 emit_movimm(0,HOST_TEMPREG);
3552 emit_writeword(HOST_TEMPREG,&pending_exception);
3553 }
3554 if( s != 1)
3555 emit_mov(s, 1);
3556 emit_movimm(copr, 0);
3557 emit_far_call(pcsx_mtc0);
3558 if (copr == 12 || copr == 13) {
3559 emit_readword(&psxRegs.cycle,HOST_CCREG);
3560 emit_readword(&last_count,HOST_TEMPREG);
3561 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3562 //emit_writeword(HOST_TEMPREG,&last_count);
3563 assert(!is_delayslot);
3564 emit_readword(&pending_exception,HOST_TEMPREG);
3565 emit_test(HOST_TEMPREG,HOST_TEMPREG);
3566 void *jaddr = out;
3567 emit_jeq(0);
3568 emit_readword(&pcaddr, 0);
3569 emit_far_call(ndrc_get_addr_ht);
3570 emit_jmpreg(0);
3571 set_jump_target(jaddr, out);
3572 emit_addimm(HOST_CCREG, -ccadj_ - 2, HOST_CCREG);
3573 if (cc != HOST_CCREG)
3574 emit_storereg(CCREG, HOST_CCREG);
3575 }
3576 emit_loadreg(dops[i].rs1,s);
3577 }
3578}
3579
3580static void rfe_assemble(int i, const struct regstat *i_regs)
3581{
3582 emit_readword(&psxRegs.CP0.n.SR, 0);
3583 emit_andimm(0, 0x3c, 1);
3584 emit_andimm(0, ~0xf, 0);
3585 emit_orrshr_imm(1, 2, 0);
3586 emit_writeword(0, &psxRegs.CP0.n.SR);
3587}
3588
3589static int cop2_is_stalling_op(int i, int *cycles)
3590{
3591 if (dops[i].opcode == 0x3a) { // SWC2
3592 *cycles = 0;
3593 return 1;
3594 }
3595 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3596 *cycles = 0;
3597 return 1;
3598 }
3599 if (dops[i].itype == C2OP) {
3600 *cycles = gte_cycletab[source[i] & 0x3f];
3601 return 1;
3602 }
3603 // ... what about MTC2/CTC2/LWC2?
3604 return 0;
3605}
3606
3607#if 0
3608static void log_gte_stall(int stall, u_int cycle)
3609{
3610 if ((u_int)stall <= 44)
3611 printf("x stall %2d %u\n", stall, cycle + last_count);
3612}
3613
3614static void emit_log_gte_stall(int i, int stall, u_int reglist)
3615{
3616 save_regs(reglist);
3617 if (stall > 0)
3618 emit_movimm(stall, 0);
3619 else
3620 emit_mov(HOST_TEMPREG, 0);
3621 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3622 emit_far_call(log_gte_stall);
3623 restore_regs(reglist);
3624}
3625#endif
3626
3627static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3628{
3629 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3630 int rtmp = reglist_find_free(reglist);
3631
3632 if (HACK_ENABLED(NDHACK_NO_STALLS))
3633 return;
3634 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3635 // happens occasionally... cc evicted? Don't bother then
3636 //printf("no cc %08x\n", start + i*4);
3637 return;
3638 }
3639 if (!dops[i].bt) {
3640 for (j = i - 1; j >= 0; j--) {
3641 //if (dops[j].is_ds) break;
3642 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3643 break;
3644 if (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj)
3645 break;
3646 }
3647 j = max(j, 0);
3648 }
3649 cycles_passed = cinfo[i].ccadj - cinfo[j].ccadj;
3650 if (other_gte_op_cycles >= 0)
3651 stall = other_gte_op_cycles - cycles_passed;
3652 else if (cycles_passed >= 44)
3653 stall = 0; // can't stall
3654 if (stall == -MAXBLOCK && rtmp >= 0) {
3655 // unknown stall, do the expensive runtime check
3656 assem_debug("; cop2_do_stall_check\n");
3657#if 0 // too slow
3658 save_regs(reglist);
3659 emit_movimm(gte_cycletab[op], 0);
3660 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3661 emit_far_call(call_gteStall);
3662 restore_regs(reglist);
3663#else
3664 host_tempreg_acquire();
3665 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3666 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3667 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3668 emit_cmpimm(HOST_TEMPREG, 44);
3669 emit_cmovb_reg(rtmp, HOST_CCREG);
3670 //emit_log_gte_stall(i, 0, reglist);
3671 host_tempreg_release();
3672#endif
3673 }
3674 else if (stall > 0) {
3675 //emit_log_gte_stall(i, stall, reglist);
3676 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3677 }
3678
3679 // save gteBusyCycle, if needed
3680 if (gte_cycletab[op] == 0)
3681 return;
3682 other_gte_op_cycles = -1;
3683 for (j = i + 1; j < slen; j++) {
3684 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3685 break;
3686 if (dops[j].is_jump) {
3687 // check ds
3688 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3689 j++;
3690 break;
3691 }
3692 }
3693 if (other_gte_op_cycles >= 0)
3694 // will handle stall when assembling that op
3695 return;
3696 cycles_passed = cinfo[min(j, slen -1)].ccadj - cinfo[i].ccadj;
3697 if (cycles_passed >= 44)
3698 return;
3699 assem_debug("; save gteBusyCycle\n");
3700 host_tempreg_acquire();
3701#if 0
3702 emit_readword(&last_count, HOST_TEMPREG);
3703 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3704 emit_addimm(HOST_TEMPREG, cinfo[i].ccadj, HOST_TEMPREG);
3705 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3706 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3707#else
3708 emit_addimm(HOST_CCREG, cinfo[i].ccadj + gte_cycletab[op], HOST_TEMPREG);
3709 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3710#endif
3711 host_tempreg_release();
3712}
3713
3714static int is_mflohi(int i)
3715{
3716 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3717}
3718
3719static int check_multdiv(int i, int *cycles)
3720{
3721 if (dops[i].itype != MULTDIV)
3722 return 0;
3723 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3724 *cycles = 11; // approx from 7 11 14
3725 else
3726 *cycles = 37;
3727 return 1;
3728}
3729
3730static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3731{
3732 int j, found = 0, c = 0;
3733 if (HACK_ENABLED(NDHACK_NO_STALLS))
3734 return;
3735 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3736 // happens occasionally... cc evicted? Don't bother then
3737 return;
3738 }
3739 for (j = i + 1; j < slen; j++) {
3740 if (dops[j].bt)
3741 break;
3742 if ((found = is_mflohi(j)))
3743 break;
3744 if (dops[j].is_jump) {
3745 // check ds
3746 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3747 j++;
3748 break;
3749 }
3750 }
3751 if (found)
3752 // handle all in multdiv_do_stall()
3753 return;
3754 check_multdiv(i, &c);
3755 assert(c > 0);
3756 assem_debug("; muldiv prepare stall %d\n", c);
3757 host_tempreg_acquire();
3758 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3759 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3760 host_tempreg_release();
3761}
3762
3763static void multdiv_do_stall(int i, const struct regstat *i_regs)
3764{
3765 int j, known_cycles = 0;
3766 u_int reglist = get_host_reglist(i_regs->regmap);
3767 int rtmp = get_reg_temp(i_regs->regmap);
3768 if (rtmp < 0)
3769 rtmp = reglist_find_free(reglist);
3770 if (HACK_ENABLED(NDHACK_NO_STALLS))
3771 return;
3772 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3773 // happens occasionally... cc evicted? Don't bother then
3774 //printf("no cc/rtmp %08x\n", start + i*4);
3775 return;
3776 }
3777 if (!dops[i].bt) {
3778 for (j = i - 1; j >= 0; j--) {
3779 if (dops[j].is_ds) break;
3780 if (check_multdiv(j, &known_cycles))
3781 break;
3782 if (is_mflohi(j))
3783 // already handled by this op
3784 return;
3785 if (dops[j].bt || (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj))
3786 break;
3787 }
3788 j = max(j, 0);
3789 }
3790 if (known_cycles > 0) {
3791 known_cycles -= cinfo[i].ccadj - cinfo[j].ccadj;
3792 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3793 if (known_cycles > 0)
3794 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3795 return;
3796 }
3797 assem_debug("; muldiv stall unresolved\n");
3798 host_tempreg_acquire();
3799 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3800 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3801 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3802 emit_cmpimm(HOST_TEMPREG, 37);
3803 emit_cmovb_reg(rtmp, HOST_CCREG);
3804 //emit_log_gte_stall(i, 0, reglist);
3805 host_tempreg_release();
3806}
3807
3808static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3809{
3810 switch (copr) {
3811 case 1:
3812 case 3:
3813 case 5:
3814 case 8:
3815 case 9:
3816 case 10:
3817 case 11:
3818 emit_readword(&reg_cop2d[copr],tl);
3819 emit_signextend16(tl,tl);
3820 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3821 break;
3822 case 7:
3823 case 16:
3824 case 17:
3825 case 18:
3826 case 19:
3827 emit_readword(&reg_cop2d[copr],tl);
3828 emit_andimm(tl,0xffff,tl);
3829 emit_writeword(tl,&reg_cop2d[copr]);
3830 break;
3831 case 15:
3832 emit_readword(&reg_cop2d[14],tl); // SXY2
3833 emit_writeword(tl,&reg_cop2d[copr]);
3834 break;
3835 case 28:
3836 case 29:
3837 c2op_mfc2_29_assemble(tl,temp);
3838 break;
3839 default:
3840 emit_readword(&reg_cop2d[copr],tl);
3841 break;
3842 }
3843}
3844
3845static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3846{
3847 switch (copr) {
3848 case 15:
3849 emit_readword(&reg_cop2d[13],temp); // SXY1
3850 emit_writeword(sl,&reg_cop2d[copr]);
3851 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3852 emit_readword(&reg_cop2d[14],temp); // SXY2
3853 emit_writeword(sl,&reg_cop2d[14]);
3854 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3855 break;
3856 case 28:
3857 emit_andimm(sl,0x001f,temp);
3858 emit_shlimm(temp,7,temp);
3859 emit_writeword(temp,&reg_cop2d[9]);
3860 emit_andimm(sl,0x03e0,temp);
3861 emit_shlimm(temp,2,temp);
3862 emit_writeword(temp,&reg_cop2d[10]);
3863 emit_andimm(sl,0x7c00,temp);
3864 emit_shrimm(temp,3,temp);
3865 emit_writeword(temp,&reg_cop2d[11]);
3866 emit_writeword(sl,&reg_cop2d[28]);
3867 break;
3868 case 30:
3869 emit_xorsar_imm(sl,sl,31,temp);
3870#if defined(HAVE_ARMV5) || defined(__aarch64__)
3871 emit_clz(temp,temp);
3872#else
3873 emit_movs(temp,HOST_TEMPREG);
3874 emit_movimm(0,temp);
3875 emit_jeq((int)out+4*4);
3876 emit_addpl_imm(temp,1,temp);
3877 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3878 emit_jns((int)out-2*4);
3879#endif
3880 emit_writeword(sl,&reg_cop2d[30]);
3881 emit_writeword(temp,&reg_cop2d[31]);
3882 break;
3883 case 31:
3884 break;
3885 default:
3886 emit_writeword(sl,&reg_cop2d[copr]);
3887 break;
3888 }
3889}
3890
3891static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3892{
3893 int s,tl;
3894 int ar;
3895 int offset;
3896 int memtarget=0,c=0;
3897 void *jaddr2=NULL;
3898 enum stub_type type;
3899 int offset_reg = -1;
3900 int fastio_reg_override = -1;
3901 u_int reglist=get_host_reglist(i_regs->regmap);
3902 u_int copr=(source[i]>>16)&0x1f;
3903 s=get_reg(i_regs->regmap,dops[i].rs1);
3904 tl=get_reg(i_regs->regmap,FTEMP);
3905 offset=cinfo[i].imm;
3906 assert(tl>=0);
3907
3908 if(i_regs->regmap[HOST_CCREG]==CCREG)
3909 reglist&=~(1<<HOST_CCREG);
3910
3911 // get the address
3912 ar = cinfo[i].addr;
3913 assert(ar >= 0);
3914 if (dops[i].opcode==0x3a) { // SWC2
3915 reglist |= 1<<ar;
3916 }
3917 if(s>=0) c=(i_regs->wasconst>>s)&1;
3918 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3919
3920 cop2_do_stall_check(0, i, i_regs, reglist);
3921
3922 if (dops[i].opcode==0x3a) { // SWC2
3923 cop2_get_dreg(copr,tl,-1);
3924 type=STOREW_STUB;
3925 }
3926 else
3927 type=LOADW_STUB;
3928
3929 if(c&&!memtarget) {
3930 jaddr2=out;
3931 emit_jmp(0); // inline_readstub/inline_writestub?
3932 }
3933 else {
3934 if(!c) {
3935 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3936 &offset_reg, &fastio_reg_override, ccadj_);
3937 }
3938 else if (ram_offset && memtarget) {
3939 offset_reg = get_ro_reg(i_regs, 0);
3940 }
3941 switch (dops[i].opcode) {
3942 case 0x32: { // LWC2
3943 int a = ar;
3944 if (fastio_reg_override >= 0)
3945 a = fastio_reg_override;
3946 do_load_word(a, tl, offset_reg);
3947 break;
3948 }
3949 case 0x3a: { // SWC2
3950 #ifdef DESTRUCTIVE_SHIFT
3951 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3952 #endif
3953 int a = ar;
3954 if (fastio_reg_override >= 0)
3955 a = fastio_reg_override;
3956 do_store_word(a, 0, tl, offset_reg, 1);
3957 break;
3958 }
3959 default:
3960 assert(0);
3961 }
3962 }
3963 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3964 host_tempreg_release();
3965 if(jaddr2)
3966 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3967 if(dops[i].opcode==0x3a) // SWC2
3968 do_store_smc_check(i, i_regs, reglist, ar);
3969 if (dops[i].opcode==0x32) { // LWC2
3970 host_tempreg_acquire();
3971 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3972 host_tempreg_release();
3973 }
3974}
3975
3976static void cop2_assemble(int i, const struct regstat *i_regs)
3977{
3978 u_int copr = (source[i]>>11) & 0x1f;
3979 signed char temp = get_reg_temp(i_regs->regmap);
3980
3981 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3982 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3983 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3984 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3985 reglist = reglist_exclude(reglist, tl, -1);
3986 }
3987 cop2_do_stall_check(0, i, i_regs, reglist);
3988 }
3989 if (dops[i].opcode2==0) { // MFC2
3990 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3991 if(tl>=0&&dops[i].rt1!=0)
3992 cop2_get_dreg(copr,tl,temp);
3993 }
3994 else if (dops[i].opcode2==4) { // MTC2
3995 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3996 cop2_put_dreg(copr,sl,temp);
3997 }
3998 else if (dops[i].opcode2==2) // CFC2
3999 {
4000 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4001 if(tl>=0&&dops[i].rt1!=0)
4002 emit_readword(&reg_cop2c[copr],tl);
4003 }
4004 else if (dops[i].opcode2==6) // CTC2
4005 {
4006 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4007 switch(copr) {
4008 case 4:
4009 case 12:
4010 case 20:
4011 case 26:
4012 case 27:
4013 case 29:
4014 case 30:
4015 emit_signextend16(sl,temp);
4016 break;
4017 case 31:
4018 c2op_ctc2_31_assemble(sl,temp);
4019 break;
4020 default:
4021 temp=sl;
4022 break;
4023 }
4024 emit_writeword(temp,&reg_cop2c[copr]);
4025 assert(sl>=0);
4026 }
4027}
4028
4029static void do_unalignedwritestub(int n)
4030{
4031 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4032 literal_pool(256);
4033 set_jump_target(stubs[n].addr, out);
4034
4035 int i=stubs[n].a;
4036 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4037 int addr=stubs[n].b;
4038 u_int reglist=stubs[n].e;
4039 signed char *i_regmap=i_regs->regmap;
4040 int temp2=get_reg(i_regmap,FTEMP);
4041 int rt;
4042 rt=get_reg(i_regmap,dops[i].rs2);
4043 assert(rt>=0);
4044 assert(addr>=0);
4045 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4046 reglist|=(1<<addr);
4047 reglist&=~(1<<temp2);
4048
4049 // don't bother with it and call write handler
4050 save_regs(reglist);
4051 pass_args(addr,rt);
4052 int cc=get_reg(i_regmap,CCREG);
4053 if(cc<0)
4054 emit_loadreg(CCREG,2);
4055 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4056 emit_movimm(start + i*4,3);
4057 emit_writeword(3,&psxRegs.pc);
4058 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4059 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4060 if(cc<0)
4061 emit_storereg(CCREG,2);
4062 restore_regs(reglist);
4063 emit_jmp(stubs[n].retaddr); // return address
4064}
4065
4066static void do_overflowstub(int n)
4067{
4068 assem_debug("do_overflowstub %x\n", start + (u_int)stubs[n].a * 4);
4069 literal_pool(24);
4070 int i = stubs[n].a;
4071 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4072 int ccadj = stubs[n].d;
4073 set_jump_target(stubs[n].addr, out);
4074 wb_dirtys(regs[i].regmap, regs[i].dirty);
4075 exception_assemble(i, i_regs, ccadj);
4076}
4077
4078static void do_alignmentstub(int n)
4079{
4080 assem_debug("do_alignmentstub %x\n", start + (u_int)stubs[n].a * 4);
4081 literal_pool(24);
4082 int i = stubs[n].a;
4083 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4084 int ccadj = stubs[n].d;
4085 int is_store = dops[i].itype == STORE || dops[i].opcode == 0x3A; // SWC2
4086 int cause = (dops[i].opcode & 3) << 28;
4087 cause |= is_store ? (R3000E_AdES << 2) : (R3000E_AdEL << 2);
4088 set_jump_target(stubs[n].addr, out);
4089 wb_dirtys(regs[i].regmap, regs[i].dirty);
4090 if (stubs[n].b != 1)
4091 emit_mov(stubs[n].b, 1); // faulting address
4092 emit_movimm(cause, 0);
4093 exception_assemble(i, i_regs, ccadj);
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_w(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_addimm_ptr(FP,(u_char *)&psxRegs - (u_char *)&dynarec_local,0);
4137 emit_far_call(func);
4138 emit_far_jump(jump_to_new_pc);
4139}
4140
4141static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_)
4142{
4143 // 'break' tends to be littered around to catch things like
4144 // division by 0 and is almost never executed, so don't emit much code here
4145 void *func;
4146 if (dops[i].itype == ALU || dops[i].itype == IMM16)
4147 func = is_delayslot ? jump_overflow_ds : jump_overflow;
4148 else if (dops[i].itype == LOAD || dops[i].itype == STORE)
4149 func = is_delayslot ? jump_addrerror_ds : jump_addrerror;
4150 else if (dops[i].opcode2 == 0x0C)
4151 func = is_delayslot ? jump_syscall_ds : jump_syscall;
4152 else
4153 func = is_delayslot ? jump_break_ds : jump_break;
4154 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) // evicted
4155 emit_loadreg(CCREG, HOST_CCREG);
4156 emit_movimm(start + i*4, 2); // pc
4157 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4158 emit_far_jump(func);
4159}
4160
4161static void hlecall_bad()
4162{
4163 assert(0);
4164}
4165
4166static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4167{
4168 void *hlefunc = hlecall_bad;
4169 uint32_t hleCode = source[i] & 0x03ffffff;
4170 if (hleCode < ARRAY_SIZE(psxHLEt))
4171 hlefunc = psxHLEt[hleCode];
4172
4173 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4174}
4175
4176static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4177{
4178 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4179}
4180
4181static void speculate_mov(int rs,int rt)
4182{
4183 if(rt!=0) {
4184 smrv_strong_next|=1<<rt;
4185 smrv[rt]=smrv[rs];
4186 }
4187}
4188
4189static void speculate_mov_weak(int rs,int rt)
4190{
4191 if(rt!=0) {
4192 smrv_weak_next|=1<<rt;
4193 smrv[rt]=smrv[rs];
4194 }
4195}
4196
4197static void speculate_register_values(int i)
4198{
4199 if(i==0) {
4200 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4201 // gp,sp are likely to stay the same throughout the block
4202 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4203 smrv_weak_next=~smrv_strong_next;
4204 //printf(" llr %08x\n", smrv[4]);
4205 }
4206 smrv_strong=smrv_strong_next;
4207 smrv_weak=smrv_weak_next;
4208 switch(dops[i].itype) {
4209 case ALU:
4210 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4211 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4212 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4213 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4214 else {
4215 smrv_strong_next&=~(1<<dops[i].rt1);
4216 smrv_weak_next&=~(1<<dops[i].rt1);
4217 }
4218 break;
4219 case SHIFTIMM:
4220 smrv_strong_next&=~(1<<dops[i].rt1);
4221 smrv_weak_next&=~(1<<dops[i].rt1);
4222 // fallthrough
4223 case IMM16:
4224 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4225 int hr = get_reg_w(regs[i].regmap, dops[i].rt1);
4226 u_int value;
4227 if(hr>=0) {
4228 if(get_final_value(hr,i,&value))
4229 smrv[dops[i].rt1]=value;
4230 else smrv[dops[i].rt1]=constmap[i][hr];
4231 smrv_strong_next|=1<<dops[i].rt1;
4232 }
4233 }
4234 else {
4235 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4236 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4237 }
4238 break;
4239 case LOAD:
4240 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4241 // special case for BIOS
4242 smrv[dops[i].rt1]=0xa0000000;
4243 smrv_strong_next|=1<<dops[i].rt1;
4244 break;
4245 }
4246 // fallthrough
4247 case SHIFT:
4248 case LOADLR:
4249 case MOV:
4250 smrv_strong_next&=~(1<<dops[i].rt1);
4251 smrv_weak_next&=~(1<<dops[i].rt1);
4252 break;
4253 case COP0:
4254 case COP2:
4255 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4256 smrv_strong_next&=~(1<<dops[i].rt1);
4257 smrv_weak_next&=~(1<<dops[i].rt1);
4258 }
4259 break;
4260 case C2LS:
4261 if (dops[i].opcode==0x32) { // LWC2
4262 smrv_strong_next&=~(1<<dops[i].rt1);
4263 smrv_weak_next&=~(1<<dops[i].rt1);
4264 }
4265 break;
4266 }
4267#if 0
4268 int r=4;
4269 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4270 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4271#endif
4272}
4273
4274static void ujump_assemble(int i, const struct regstat *i_regs);
4275static void rjump_assemble(int i, const struct regstat *i_regs);
4276static void cjump_assemble(int i, const struct regstat *i_regs);
4277static void sjump_assemble(int i, const struct regstat *i_regs);
4278
4279static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4280{
4281 int ds = 0;
4282 switch (dops[i].itype) {
4283 case ALU:
4284 alu_assemble(i, i_regs, ccadj_);
4285 break;
4286 case IMM16:
4287 imm16_assemble(i, i_regs, ccadj_);
4288 break;
4289 case SHIFT:
4290 shift_assemble(i, i_regs);
4291 break;
4292 case SHIFTIMM:
4293 shiftimm_assemble(i, i_regs);
4294 break;
4295 case LOAD:
4296 load_assemble(i, i_regs, ccadj_);
4297 break;
4298 case LOADLR:
4299 loadlr_assemble(i, i_regs, ccadj_);
4300 break;
4301 case STORE:
4302 store_assemble(i, i_regs, ccadj_);
4303 break;
4304 case STORELR:
4305 storelr_assemble(i, i_regs, ccadj_);
4306 break;
4307 case COP0:
4308 cop0_assemble(i, i_regs, ccadj_);
4309 break;
4310 case RFE:
4311 rfe_assemble(i, i_regs);
4312 break;
4313 case COP2:
4314 cop2_assemble(i, i_regs);
4315 break;
4316 case C2LS:
4317 c2ls_assemble(i, i_regs, ccadj_);
4318 break;
4319 case C2OP:
4320 c2op_assemble(i, i_regs);
4321 break;
4322 case MULTDIV:
4323 multdiv_assemble(i, i_regs);
4324 multdiv_prepare_stall(i, i_regs, ccadj_);
4325 break;
4326 case MOV:
4327 mov_assemble(i, i_regs);
4328 break;
4329 case SYSCALL:
4330 exception_assemble(i, i_regs, ccadj_);
4331 break;
4332 case HLECALL:
4333 hlecall_assemble(i, i_regs, ccadj_);
4334 break;
4335 case INTCALL:
4336 intcall_assemble(i, i_regs, ccadj_);
4337 break;
4338 case UJUMP:
4339 ujump_assemble(i, i_regs);
4340 ds = 1;
4341 break;
4342 case RJUMP:
4343 rjump_assemble(i, i_regs);
4344 ds = 1;
4345 break;
4346 case CJUMP:
4347 cjump_assemble(i, i_regs);
4348 ds = 1;
4349 break;
4350 case SJUMP:
4351 sjump_assemble(i, i_regs);
4352 ds = 1;
4353 break;
4354 case NOP:
4355 case OTHER:
4356 // not handled, just skip
4357 break;
4358 default:
4359 assert(0);
4360 }
4361 return ds;
4362}
4363
4364static void ds_assemble(int i, const struct regstat *i_regs)
4365{
4366 speculate_register_values(i);
4367 is_delayslot = 1;
4368 switch (dops[i].itype) {
4369 case SYSCALL:
4370 case HLECALL:
4371 case INTCALL:
4372 case UJUMP:
4373 case RJUMP:
4374 case CJUMP:
4375 case SJUMP:
4376 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4377 break;
4378 default:
4379 assemble(i, i_regs, cinfo[i].ccadj);
4380 }
4381 is_delayslot = 0;
4382}
4383
4384// Is the branch target a valid internal jump?
4385static int internal_branch(int addr)
4386{
4387 if(addr&1) return 0; // Indirect (register) jump
4388 if(addr>=start && addr<start+slen*4-4)
4389 {
4390 return 1;
4391 }
4392 return 0;
4393}
4394
4395static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4396{
4397 int hr;
4398 for(hr=0;hr<HOST_REGS;hr++) {
4399 if(hr!=EXCLUDE_REG) {
4400 if(pre[hr]!=entry[hr]) {
4401 if(pre[hr]>=0) {
4402 if((dirty>>hr)&1) {
4403 if(get_reg(entry,pre[hr])<0) {
4404 assert(pre[hr]<64);
4405 if(!((u>>pre[hr])&1))
4406 emit_storereg(pre[hr],hr);
4407 }
4408 }
4409 }
4410 }
4411 }
4412 }
4413 // Move from one register to another (no writeback)
4414 for(hr=0;hr<HOST_REGS;hr++) {
4415 if(hr!=EXCLUDE_REG) {
4416 if(pre[hr]!=entry[hr]) {
4417 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4418 int nr;
4419 if((nr=get_reg(entry,pre[hr]))>=0) {
4420 emit_mov(hr,nr);
4421 }
4422 }
4423 }
4424 }
4425 }
4426}
4427
4428// Load the specified registers
4429// This only loads the registers given as arguments because
4430// we don't want to load things that will be overwritten
4431static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4432{
4433 int hr = get_reg(regmap, rs);
4434 if (hr >= 0 && entry[hr] != regmap[hr])
4435 emit_loadreg(regmap[hr], hr);
4436}
4437
4438static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4439{
4440 load_reg(entry, regmap, rs1);
4441 if (rs1 != rs2)
4442 load_reg(entry, regmap, rs2);
4443}
4444
4445// Load registers prior to the start of a loop
4446// so that they are not loaded within the loop
4447static void loop_preload(signed char pre[],signed char entry[])
4448{
4449 int hr;
4450 for (hr = 0; hr < HOST_REGS; hr++) {
4451 int r = entry[hr];
4452 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4453 assem_debug("loop preload:\n");
4454 if (r < TEMPREG)
4455 emit_loadreg(r, hr);
4456 }
4457 }
4458}
4459
4460// Generate address for load/store instruction
4461// goes to AGEN (or temp) for writes, FTEMP for LOADLR and cop1/2 loads
4462// AGEN is assigned by pass5b_preallocate2
4463static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4464{
4465 if (dops[i].is_load || dops[i].is_store) {
4466 int ra = -1;
4467 int agr = AGEN1 + (i&1);
4468 if(dops[i].itype==LOAD) {
4469 if (!dops[i].may_except)
4470 ra = get_reg_w(i_regs->regmap, dops[i].rt1); // reuse dest for agen
4471 if (ra < 0)
4472 ra = get_reg_temp(i_regs->regmap);
4473 }
4474 if(dops[i].itype==LOADLR) {
4475 ra=get_reg(i_regs->regmap,FTEMP);
4476 }
4477 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4478 ra=get_reg(i_regs->regmap,agr);
4479 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4480 }
4481 if(dops[i].itype==C2LS) {
4482 if (dops[i].opcode == 0x32) // LWC2
4483 ra=get_reg(i_regs->regmap,FTEMP);
4484 else { // SWC2
4485 ra=get_reg(i_regs->regmap,agr);
4486 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4487 }
4488 }
4489 int rs = get_reg(i_regs->regmap, dops[i].rs1);
4490 //if(ra>=0)
4491 {
4492 int offset = cinfo[i].imm;
4493 int add_offset = offset != 0;
4494 int c=(i_regs->wasconst>>rs)&1;
4495 if(dops[i].rs1==0) {
4496 // Using r0 as a base address
4497 assert(ra >= 0);
4498 if(!entry||entry[ra]!=agr) {
4499 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4500 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4501 }else{
4502 emit_movimm(offset,ra);
4503 }
4504 } // else did it in the previous cycle
4505 cinfo[i].addr = ra;
4506 add_offset = 0;
4507 }
4508 else if (rs < 0) {
4509 assert(ra >= 0);
4510 if (!entry || entry[ra] != dops[i].rs1)
4511 emit_loadreg(dops[i].rs1, ra);
4512 cinfo[i].addr = ra;
4513 //if(!entry||entry[ra]!=dops[i].rs1)
4514 // printf("poor load scheduling!\n");
4515 }
4516 else if(c) {
4517 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4518 assert(ra >= 0);
4519 if(!entry||entry[ra]!=agr) {
4520 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4521 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4522 }else{
4523 emit_movimm(constmap[i][rs]+offset,ra);
4524 regs[i].loadedconst|=1<<ra;
4525 }
4526 } // else did it in the previous cycle
4527 cinfo[i].addr = ra;
4528 }
4529 else // else load_consts already did it
4530 cinfo[i].addr = rs;
4531 add_offset = 0;
4532 }
4533 else if (dops[i].itype == STORELR) { // overwrites addr
4534 assert(ra >= 0);
4535 assert(rs != ra);
4536 emit_mov(rs, ra);
4537 cinfo[i].addr = ra;
4538 }
4539 else
4540 cinfo[i].addr = rs;
4541 if (add_offset) {
4542 assert(ra >= 0);
4543 if(rs>=0) {
4544 emit_addimm(rs,offset,ra);
4545 }else{
4546 emit_addimm(ra,offset,ra);
4547 }
4548 cinfo[i].addr = ra;
4549 }
4550 }
4551 assert(cinfo[i].addr >= 0);
4552 }
4553 // Preload constants for next instruction
4554 if (dops[i+1].is_load || dops[i+1].is_store) {
4555 int agr,ra;
4556 // Actual address
4557 agr=AGEN1+((i+1)&1);
4558 ra=get_reg(i_regs->regmap,agr);
4559 if(ra>=0) {
4560 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4561 int offset=cinfo[i+1].imm;
4562 int c=(regs[i+1].wasconst>>rs)&1;
4563 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4564 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4565 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4566 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4567 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4568 }else{
4569 emit_movimm(constmap[i+1][rs]+offset,ra);
4570 regs[i+1].loadedconst|=1<<ra;
4571 }
4572 }
4573 else if(dops[i+1].rs1==0) {
4574 // Using r0 as a base address
4575 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4576 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4577 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4578 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4579 }else{
4580 emit_movimm(offset,ra);
4581 }
4582 }
4583 }
4584 }
4585}
4586
4587static int get_final_value(int hr, int i, u_int *value)
4588{
4589 int reg=regs[i].regmap[hr];
4590 while(i<slen-1) {
4591 if(regs[i+1].regmap[hr]!=reg) break;
4592 if(!((regs[i+1].isconst>>hr)&1)) break;
4593 if(dops[i+1].bt) break;
4594 i++;
4595 }
4596 if(i<slen-1) {
4597 if (dops[i].is_jump) {
4598 *value=constmap[i][hr];
4599 return 1;
4600 }
4601 if(!dops[i+1].bt) {
4602 if (dops[i+1].is_jump) {
4603 // Load in delay slot, out-of-order execution
4604 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4605 {
4606 // Precompute load address
4607 *value=constmap[i][hr]+cinfo[i+2].imm;
4608 return 1;
4609 }
4610 }
4611 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4612 {
4613 // Precompute load address
4614 *value=constmap[i][hr]+cinfo[i+1].imm;
4615 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],cinfo[i+1].imm);
4616 return 1;
4617 }
4618 }
4619 }
4620 *value=constmap[i][hr];
4621 //printf("c=%lx\n",(long)constmap[i][hr]);
4622 if(i==slen-1) return 1;
4623 assert(reg < 64);
4624 return !((unneeded_reg[i+1]>>reg)&1);
4625}
4626
4627// Load registers with known constants
4628static void load_consts(signed char pre[],signed char regmap[],int i)
4629{
4630 int hr,hr2;
4631 // propagate loaded constant flags
4632 if(i==0||dops[i].bt)
4633 regs[i].loadedconst=0;
4634 else {
4635 for(hr=0;hr<HOST_REGS;hr++) {
4636 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4637 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4638 {
4639 regs[i].loadedconst|=1<<hr;
4640 }
4641 }
4642 }
4643 // Load 32-bit regs
4644 for(hr=0;hr<HOST_REGS;hr++) {
4645 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4646 //if(entry[hr]!=regmap[hr]) {
4647 if(!((regs[i].loadedconst>>hr)&1)) {
4648 assert(regmap[hr]<64);
4649 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4650 u_int value, similar=0;
4651 if(get_final_value(hr,i,&value)) {
4652 // see if some other register has similar value
4653 for(hr2=0;hr2<HOST_REGS;hr2++) {
4654 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4655 if(is_similar_value(value,constmap[i][hr2])) {
4656 similar=1;
4657 break;
4658 }
4659 }
4660 }
4661 if(similar) {
4662 u_int value2;
4663 if(get_final_value(hr2,i,&value2)) // is this needed?
4664 emit_movimm_from(value2,hr2,value,hr);
4665 else
4666 emit_movimm(value,hr);
4667 }
4668 else if(value==0) {
4669 emit_zeroreg(hr);
4670 }
4671 else {
4672 emit_movimm(value,hr);
4673 }
4674 }
4675 regs[i].loadedconst|=1<<hr;
4676 }
4677 }
4678 }
4679 }
4680}
4681
4682static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4683{
4684 int hr;
4685 // Load 32-bit regs
4686 for(hr=0;hr<HOST_REGS;hr++) {
4687 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4688 assert(regmap[hr] < 64);
4689 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4690 int value=constmap[i][hr];
4691 if(value==0) {
4692 emit_zeroreg(hr);
4693 }
4694 else {
4695 emit_movimm(value,hr);
4696 }
4697 }
4698 }
4699 }
4700}
4701
4702// Write out all dirty registers (except cycle count)
4703static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4704{
4705 int hr;
4706 for(hr=0;hr<HOST_REGS;hr++) {
4707 if(hr!=EXCLUDE_REG) {
4708 if(i_regmap[hr]>0) {
4709 if(i_regmap[hr]!=CCREG) {
4710 if((i_dirty>>hr)&1) {
4711 assert(i_regmap[hr]<64);
4712 emit_storereg(i_regmap[hr],hr);
4713 }
4714 }
4715 }
4716 }
4717 }
4718}
4719
4720// Write out dirty registers that we need to reload (pair with load_needed_regs)
4721// This writes the registers not written by store_regs_bt
4722static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4723{
4724 int hr;
4725 int t=(addr-start)>>2;
4726 for(hr=0;hr<HOST_REGS;hr++) {
4727 if(hr!=EXCLUDE_REG) {
4728 if(i_regmap[hr]>0) {
4729 if(i_regmap[hr]!=CCREG) {
4730 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4731 if((i_dirty>>hr)&1) {
4732 assert(i_regmap[hr]<64);
4733 emit_storereg(i_regmap[hr],hr);
4734 }
4735 }
4736 }
4737 }
4738 }
4739 }
4740}
4741
4742// Load all registers (except cycle count)
4743static void load_all_regs(const signed char i_regmap[])
4744{
4745 int hr;
4746 for(hr=0;hr<HOST_REGS;hr++) {
4747 if(hr!=EXCLUDE_REG) {
4748 if(i_regmap[hr]==0) {
4749 emit_zeroreg(hr);
4750 }
4751 else
4752 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4753 {
4754 emit_loadreg(i_regmap[hr],hr);
4755 }
4756 }
4757 }
4758}
4759
4760// Load all current registers also needed by next instruction
4761static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4762{
4763 int hr;
4764 for(hr=0;hr<HOST_REGS;hr++) {
4765 if(hr!=EXCLUDE_REG) {
4766 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4767 if(i_regmap[hr]==0) {
4768 emit_zeroreg(hr);
4769 }
4770 else
4771 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4772 {
4773 emit_loadreg(i_regmap[hr],hr);
4774 }
4775 }
4776 }
4777 }
4778}
4779
4780// Load all regs, storing cycle count if necessary
4781static void load_regs_entry(int t)
4782{
4783 int hr;
4784 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4785 else if(cinfo[t].ccadj) emit_addimm(HOST_CCREG,-cinfo[t].ccadj,HOST_CCREG);
4786 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4787 emit_storereg(CCREG,HOST_CCREG);
4788 }
4789 // Load 32-bit regs
4790 for(hr=0;hr<HOST_REGS;hr++) {
4791 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4792 if(regs[t].regmap_entry[hr]==0) {
4793 emit_zeroreg(hr);
4794 }
4795 else if(regs[t].regmap_entry[hr]!=CCREG)
4796 {
4797 emit_loadreg(regs[t].regmap_entry[hr],hr);
4798 }
4799 }
4800 }
4801}
4802
4803// Store dirty registers prior to branch
4804static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4805{
4806 if(internal_branch(addr))
4807 {
4808 int t=(addr-start)>>2;
4809 int hr;
4810 for(hr=0;hr<HOST_REGS;hr++) {
4811 if(hr!=EXCLUDE_REG) {
4812 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4813 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4814 if((i_dirty>>hr)&1) {
4815 assert(i_regmap[hr]<64);
4816 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4817 emit_storereg(i_regmap[hr],hr);
4818 }
4819 }
4820 }
4821 }
4822 }
4823 }
4824 else
4825 {
4826 // Branch out of this block, write out all dirty regs
4827 wb_dirtys(i_regmap,i_dirty);
4828 }
4829}
4830
4831// Load all needed registers for branch target
4832static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4833{
4834 //if(addr>=start && addr<(start+slen*4))
4835 if(internal_branch(addr))
4836 {
4837 int t=(addr-start)>>2;
4838 int hr;
4839 // Store the cycle count before loading something else
4840 if(i_regmap[HOST_CCREG]!=CCREG) {
4841 assert(i_regmap[HOST_CCREG]==-1);
4842 }
4843 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4844 emit_storereg(CCREG,HOST_CCREG);
4845 }
4846 // Load 32-bit regs
4847 for(hr=0;hr<HOST_REGS;hr++) {
4848 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4849 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4850 if(regs[t].regmap_entry[hr]==0) {
4851 emit_zeroreg(hr);
4852 }
4853 else if(regs[t].regmap_entry[hr]!=CCREG)
4854 {
4855 emit_loadreg(regs[t].regmap_entry[hr],hr);
4856 }
4857 }
4858 }
4859 }
4860 }
4861}
4862
4863static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4864{
4865 if(addr>=start && addr<start+slen*4-4)
4866 {
4867 int t=(addr-start)>>2;
4868 int hr;
4869 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4870 for(hr=0;hr<HOST_REGS;hr++)
4871 {
4872 if(hr!=EXCLUDE_REG)
4873 {
4874 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4875 {
4876 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4877 {
4878 return 0;
4879 }
4880 else
4881 if((i_dirty>>hr)&1)
4882 {
4883 if(i_regmap[hr]<TEMPREG)
4884 {
4885 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4886 return 0;
4887 }
4888 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4889 {
4890 assert(0);
4891 }
4892 }
4893 }
4894 else // Same register but is it 32-bit or dirty?
4895 if(i_regmap[hr]>=0)
4896 {
4897 if(!((regs[t].dirty>>hr)&1))
4898 {
4899 if((i_dirty>>hr)&1)
4900 {
4901 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4902 {
4903 //printf("%x: dirty no match\n",addr);
4904 return 0;
4905 }
4906 }
4907 }
4908 }
4909 }
4910 }
4911 // Delay slots are not valid branch targets
4912 //if(t>0&&(dops[t-1].is_jump) return 0;
4913 // Delay slots require additional processing, so do not match
4914 if(dops[t].is_ds) return 0;
4915 }
4916 else
4917 {
4918 int hr;
4919 for(hr=0;hr<HOST_REGS;hr++)
4920 {
4921 if(hr!=EXCLUDE_REG)
4922 {
4923 if(i_regmap[hr]>=0)
4924 {
4925 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4926 {
4927 if((i_dirty>>hr)&1)
4928 {
4929 return 0;
4930 }
4931 }
4932 }
4933 }
4934 }
4935 }
4936 return 1;
4937}
4938
4939#ifdef DRC_DBG
4940static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4941{
4942 extern void do_insn_cmp();
4943 //extern int cycle;
4944 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4945
4946 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4947 save_regs(reglist);
4948 // write out changed consts to match the interpreter
4949 if (i > 0 && !dops[i].bt) {
4950 for (hr = 0; hr < HOST_REGS; hr++) {
4951 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4952 if (hr == EXCLUDE_REG || reg <= 0)
4953 continue;
4954 if (!((regs[i-1].isconst >> hr) & 1))
4955 continue;
4956 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4957 continue;
4958 emit_movimm(constmap[i-1][hr],0);
4959 emit_storereg(reg, 0);
4960 }
4961 }
4962 emit_movimm(start+i*4,0);
4963 emit_writeword(0,&pcaddr);
4964 int cc = get_reg(regs[i].regmap_entry, CCREG);
4965 if (cc < 0)
4966 emit_loadreg(CCREG, cc = 0);
4967 emit_addimm(cc, ccadj_, 0);
4968 emit_writeword(0, &psxRegs.cycle);
4969 emit_far_call(do_insn_cmp);
4970 //emit_readword(&cycle,0);
4971 //emit_addimm(0,2,0);
4972 //emit_writeword(0,&cycle);
4973 (void)get_reg2;
4974 restore_regs(reglist);
4975 assem_debug("\\\\do_insn_cmp\n");
4976}
4977#else
4978#define drc_dbg_emit_do_cmp(x,y)
4979#endif
4980
4981// Used when a branch jumps into the delay slot of another branch
4982static void ds_assemble_entry(int i)
4983{
4984 int t = (cinfo[i].ba - start) >> 2;
4985 int ccadj_ = -CLOCK_ADJUST(1);
4986 if (!instr_addr[t])
4987 instr_addr[t] = out;
4988 assem_debug("Assemble delay slot at %x\n",cinfo[i].ba);
4989 assem_debug("<->\n");
4990 drc_dbg_emit_do_cmp(t, ccadj_);
4991 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4992 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4993 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4994 address_generation(t,&regs[t],regs[t].regmap_entry);
4995 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4996 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
4997 if (dops[t].is_store)
4998 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
4999 is_delayslot=0;
5000 switch (dops[t].itype) {
5001 case SYSCALL:
5002 case HLECALL:
5003 case INTCALL:
5004 case UJUMP:
5005 case RJUMP:
5006 case CJUMP:
5007 case SJUMP:
5008 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5009 break;
5010 default:
5011 assemble(t, &regs[t], ccadj_);
5012 }
5013 store_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5014 load_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5015 if(internal_branch(cinfo[i].ba+4))
5016 assem_debug("branch: internal\n");
5017 else
5018 assem_debug("branch: external\n");
5019 assert(internal_branch(cinfo[i].ba+4));
5020 add_to_linker(out,cinfo[i].ba+4,internal_branch(cinfo[i].ba+4));
5021 emit_jmp(0);
5022}
5023
5024// Load 2 immediates optimizing for small code size
5025static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
5026{
5027 emit_movimm(imm1,rt1);
5028 emit_movimm_from(imm1,rt1,imm2,rt2);
5029}
5030
5031static void do_cc(int i, const signed char i_regmap[], int *adj,
5032 int addr, int taken, int invert)
5033{
5034 int count, count_plus2;
5035 void *jaddr;
5036 void *idle=NULL;
5037 int t=0;
5038 if(dops[i].itype==RJUMP)
5039 {
5040 *adj=0;
5041 }
5042 //if(cinfo[i].ba>=start && cinfo[i].ba<(start+slen*4))
5043 if(internal_branch(cinfo[i].ba))
5044 {
5045 t=(cinfo[i].ba-start)>>2;
5046 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5047 else *adj=cinfo[t].ccadj;
5048 }
5049 else
5050 {
5051 *adj=0;
5052 }
5053 count = cinfo[i].ccadj;
5054 count_plus2 = count + CLOCK_ADJUST(2);
5055 if(taken==TAKEN && i==(cinfo[i].ba-start)>>2 && source[i+1]==0) {
5056 // Idle loop
5057 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5058 idle=out;
5059 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5060 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5061 jaddr=out;
5062 emit_jmp(0);
5063 }
5064 else if(*adj==0||invert) {
5065 int cycles = count_plus2;
5066 // faster loop HACK
5067#if 0
5068 if (t&&*adj) {
5069 int rel=t-i;
5070 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5071 cycles=*adj+count+2-*adj;
5072 }
5073#endif
5074 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5075 jaddr = out;
5076 emit_jns(0);
5077 }
5078 else
5079 {
5080 emit_cmpimm(HOST_CCREG, -count_plus2);
5081 jaddr = out;
5082 emit_jns(0);
5083 }
5084 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5085}
5086
5087static void do_ccstub(int n)
5088{
5089 literal_pool(256);
5090 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5091 set_jump_target(stubs[n].addr, out);
5092 int i=stubs[n].b;
5093 if(stubs[n].d==NULLDS) {
5094 // Delay slot instruction is nullified ("likely" branch)
5095 wb_dirtys(regs[i].regmap,regs[i].dirty);
5096 }
5097 else if(stubs[n].d!=TAKEN) {
5098 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5099 }
5100 else {
5101 if(internal_branch(cinfo[i].ba))
5102 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5103 }
5104 if(stubs[n].c!=-1)
5105 {
5106 // Save PC as return address
5107 emit_movimm(stubs[n].c,0);
5108 emit_writeword(0,&pcaddr);
5109 }
5110 else
5111 {
5112 // Return address depends on which way the branch goes
5113 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5114 {
5115 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5116 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5117 if(dops[i].rs1==0)
5118 {
5119 s1l=s2l;
5120 s2l=-1;
5121 }
5122 else if(dops[i].rs2==0)
5123 {
5124 s2l=-1;
5125 }
5126 assert(s1l>=0);
5127 #ifdef DESTRUCTIVE_WRITEBACK
5128 if(dops[i].rs1) {
5129 if((branch_regs[i].dirty>>s1l)&&1)
5130 emit_loadreg(dops[i].rs1,s1l);
5131 }
5132 else {
5133 if((branch_regs[i].dirty>>s1l)&1)
5134 emit_loadreg(dops[i].rs2,s1l);
5135 }
5136 if(s2l>=0)
5137 if((branch_regs[i].dirty>>s2l)&1)
5138 emit_loadreg(dops[i].rs2,s2l);
5139 #endif
5140 int hr=0;
5141 int addr=-1,alt=-1,ntaddr=-1;
5142 while(hr<HOST_REGS)
5143 {
5144 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5145 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5146 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5147 {
5148 addr=hr++;break;
5149 }
5150 hr++;
5151 }
5152 while(hr<HOST_REGS)
5153 {
5154 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5155 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5156 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5157 {
5158 alt=hr++;break;
5159 }
5160 hr++;
5161 }
5162 if ((dops[i].opcode & 0x3e) == 6) // BLEZ/BGTZ needs another register
5163 {
5164 while(hr<HOST_REGS)
5165 {
5166 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5167 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5168 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5169 {
5170 ntaddr=hr;break;
5171 }
5172 hr++;
5173 }
5174 assert(hr<HOST_REGS);
5175 }
5176 if (dops[i].opcode == 4) // BEQ
5177 {
5178 #ifdef HAVE_CMOV_IMM
5179 if(s2l>=0) emit_cmp(s1l,s2l);
5180 else emit_test(s1l,s1l);
5181 emit_cmov2imm_e_ne_compact(cinfo[i].ba,start+i*4+8,addr);
5182 #else
5183 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,alt);
5184 if(s2l>=0) emit_cmp(s1l,s2l);
5185 else emit_test(s1l,s1l);
5186 emit_cmovne_reg(alt,addr);
5187 #endif
5188 }
5189 else if (dops[i].opcode == 5) // BNE
5190 {
5191 #ifdef HAVE_CMOV_IMM
5192 if(s2l>=0) emit_cmp(s1l,s2l);
5193 else emit_test(s1l,s1l);
5194 emit_cmov2imm_e_ne_compact(start+i*4+8,cinfo[i].ba,addr);
5195 #else
5196 emit_mov2imm_compact(start+i*4+8,addr,cinfo[i].ba,alt);
5197 if(s2l>=0) emit_cmp(s1l,s2l);
5198 else emit_test(s1l,s1l);
5199 emit_cmovne_reg(alt,addr);
5200 #endif
5201 }
5202 else if (dops[i].opcode == 6) // BLEZ
5203 {
5204 //emit_movimm(cinfo[i].ba,alt);
5205 //emit_movimm(start+i*4+8,addr);
5206 emit_mov2imm_compact(cinfo[i].ba,alt,start+i*4+8,addr);
5207 emit_cmpimm(s1l,1);
5208 emit_cmovl_reg(alt,addr);
5209 }
5210 else if (dops[i].opcode == 7) // BGTZ
5211 {
5212 //emit_movimm(cinfo[i].ba,addr);
5213 //emit_movimm(start+i*4+8,ntaddr);
5214 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,ntaddr);
5215 emit_cmpimm(s1l,1);
5216 emit_cmovl_reg(ntaddr,addr);
5217 }
5218 else if (dops[i].itype == SJUMP) // BLTZ/BGEZ
5219 {
5220 //emit_movimm(cinfo[i].ba,alt);
5221 //emit_movimm(start+i*4+8,addr);
5222 if (dops[i].rs1) {
5223 emit_mov2imm_compact(cinfo[i].ba,
5224 (dops[i].opcode2 & 1) ? addr : alt, start + i*4 + 8,
5225 (dops[i].opcode2 & 1) ? alt : addr);
5226 emit_test(s1l,s1l);
5227 emit_cmovs_reg(alt,addr);
5228 }
5229 else
5230 emit_movimm((dops[i].opcode2 & 1) ? cinfo[i].ba : start + i*4 + 8, addr);
5231 }
5232 emit_writeword(addr, &pcaddr);
5233 }
5234 else
5235 if(dops[i].itype==RJUMP)
5236 {
5237 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5238 if (ds_writes_rjump_rs(i)) {
5239 r=get_reg(branch_regs[i].regmap,RTEMP);
5240 }
5241 emit_writeword(r,&pcaddr);
5242 }
5243 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5244 }
5245 // Update cycle count
5246 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5247 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5248 emit_far_call(cc_interrupt);
5249 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5250 if(stubs[n].d==TAKEN) {
5251 if(internal_branch(cinfo[i].ba))
5252 load_needed_regs(branch_regs[i].regmap,regs[(cinfo[i].ba-start)>>2].regmap_entry);
5253 else if(dops[i].itype==RJUMP) {
5254 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5255 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5256 else
5257 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5258 }
5259 }else if(stubs[n].d==NOTTAKEN) {
5260 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5261 else load_all_regs(branch_regs[i].regmap);
5262 }else if(stubs[n].d==NULLDS) {
5263 // Delay slot instruction is nullified ("likely" branch)
5264 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5265 else load_all_regs(regs[i].regmap);
5266 }else{
5267 load_all_regs(branch_regs[i].regmap);
5268 }
5269 if (stubs[n].retaddr)
5270 emit_jmp(stubs[n].retaddr);
5271 else
5272 do_jump_vaddr(stubs[n].e);
5273}
5274
5275static void add_to_linker(void *addr, u_int target, int is_internal)
5276{
5277 assert(linkcount < ARRAY_SIZE(link_addr));
5278 link_addr[linkcount].addr = addr;
5279 link_addr[linkcount].target = target;
5280 link_addr[linkcount].internal = is_internal;
5281 linkcount++;
5282}
5283
5284static void ujump_assemble_write_ra(int i)
5285{
5286 int rt;
5287 unsigned int return_address;
5288 rt=get_reg(branch_regs[i].regmap,31);
5289 //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]);
5290 //assert(rt>=0);
5291 return_address=start+i*4+8;
5292 if(rt>=0) {
5293 #ifdef USE_MINI_HT
5294 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5295 int temp=-1; // note: must be ds-safe
5296 #ifdef HOST_TEMPREG
5297 temp=HOST_TEMPREG;
5298 #endif
5299 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5300 else emit_movimm(return_address,rt);
5301 }
5302 else
5303 #endif
5304 {
5305 #ifdef REG_PREFETCH
5306 if(temp>=0)
5307 {
5308 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5309 }
5310 #endif
5311 if (!((regs[i].loadedconst >> rt) & 1))
5312 emit_movimm(return_address, rt); // PC into link register
5313 #ifdef IMM_PREFETCH
5314 emit_prefetch(hash_table_get(return_address));
5315 #endif
5316 }
5317 }
5318}
5319
5320static void ujump_assemble(int i, const struct regstat *i_regs)
5321{
5322 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5323 address_generation(i+1,i_regs,regs[i].regmap_entry);
5324 #ifdef REG_PREFETCH
5325 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5326 if(dops[i].rt1==31&&temp>=0)
5327 {
5328 signed char *i_regmap=i_regs->regmap;
5329 int return_address=start+i*4+8;
5330 if(get_reg(branch_regs[i].regmap,31)>0)
5331 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5332 }
5333 #endif
5334 if (dops[i].rt1 == 31)
5335 ujump_assemble_write_ra(i); // writeback ra for DS
5336 ds_assemble(i+1,i_regs);
5337 uint64_t bc_unneeded=branch_regs[i].u;
5338 bc_unneeded|=1|(1LL<<dops[i].rt1);
5339 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5340 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5341 int cc,adj;
5342 cc=get_reg(branch_regs[i].regmap,CCREG);
5343 assert(cc==HOST_CCREG);
5344 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5345 #ifdef REG_PREFETCH
5346 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5347 #endif
5348 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5349 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5350 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5351 if(internal_branch(cinfo[i].ba))
5352 assem_debug("branch: internal\n");
5353 else
5354 assem_debug("branch: external\n");
5355 if (internal_branch(cinfo[i].ba) && dops[(cinfo[i].ba-start)>>2].is_ds) {
5356 ds_assemble_entry(i);
5357 }
5358 else {
5359 add_to_linker(out,cinfo[i].ba,internal_branch(cinfo[i].ba));
5360 emit_jmp(0);
5361 }
5362}
5363
5364static void rjump_assemble_write_ra(int i)
5365{
5366 int rt,return_address;
5367 rt=get_reg_w(branch_regs[i].regmap, dops[i].rt1);
5368 //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]);
5369 assert(rt>=0);
5370 return_address=start+i*4+8;
5371 #ifdef REG_PREFETCH
5372 if(temp>=0)
5373 {
5374 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5375 }
5376 #endif
5377 if (!((regs[i].loadedconst >> rt) & 1))
5378 emit_movimm(return_address, rt); // PC into link register
5379 #ifdef IMM_PREFETCH
5380 emit_prefetch(hash_table_get(return_address));
5381 #endif
5382}
5383
5384static void rjump_assemble(int i, const struct regstat *i_regs)
5385{
5386 int temp;
5387 int rs,cc;
5388 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5389 assert(rs>=0);
5390 if (ds_writes_rjump_rs(i)) {
5391 // Delay slot abuse, make a copy of the branch address register
5392 temp=get_reg(branch_regs[i].regmap,RTEMP);
5393 assert(temp>=0);
5394 assert(regs[i].regmap[temp]==RTEMP);
5395 emit_mov(rs,temp);
5396 rs=temp;
5397 }
5398 address_generation(i+1,i_regs,regs[i].regmap_entry);
5399 #ifdef REG_PREFETCH
5400 if(dops[i].rt1==31)
5401 {
5402 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5403 signed char *i_regmap=i_regs->regmap;
5404 int return_address=start+i*4+8;
5405 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5406 }
5407 }
5408 #endif
5409 #ifdef USE_MINI_HT
5410 if(dops[i].rs1==31) {
5411 int rh=get_reg(regs[i].regmap,RHASH);
5412 if(rh>=0) do_preload_rhash(rh);
5413 }
5414 #endif
5415 if (dops[i].rt1 != 0)
5416 rjump_assemble_write_ra(i);
5417 ds_assemble(i+1,i_regs);
5418 uint64_t bc_unneeded=branch_regs[i].u;
5419 bc_unneeded|=1|(1LL<<dops[i].rt1);
5420 bc_unneeded&=~(1LL<<dops[i].rs1);
5421 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5422 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5423 cc=get_reg(branch_regs[i].regmap,CCREG);
5424 assert(cc==HOST_CCREG);
5425 (void)cc;
5426 #ifdef USE_MINI_HT
5427 int rh=get_reg(branch_regs[i].regmap,RHASH);
5428 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5429 if(dops[i].rs1==31) {
5430 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5431 do_preload_rhtbl(ht);
5432 do_rhash(rs,rh);
5433 }
5434 #endif
5435 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5436 #ifdef DESTRUCTIVE_WRITEBACK
5437 if((branch_regs[i].dirty>>rs)&1) {
5438 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5439 emit_loadreg(dops[i].rs1,rs);
5440 }
5441 }
5442 #endif
5443 #ifdef REG_PREFETCH
5444 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5445 #endif
5446 #ifdef USE_MINI_HT
5447 if(dops[i].rs1==31) {
5448 do_miniht_load(ht,rh);
5449 }
5450 #endif
5451 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5452 //if(adj) emit_addimm(cc,2*(cinfo[i].ccadj+2-adj),cc); // ??? - Shouldn't happen
5453 //assert(adj==0);
5454 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5455 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5456 if (dops[i+1].itype == RFE)
5457 // special case for RFE
5458 emit_jmp(0);
5459 else
5460 emit_jns(0);
5461 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5462 #ifdef USE_MINI_HT
5463 if(dops[i].rs1==31) {
5464 do_miniht_jump(rs,rh,ht);
5465 }
5466 else
5467 #endif
5468 {
5469 do_jump_vaddr(rs);
5470 }
5471 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5472 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5473 #endif
5474}
5475
5476static void cjump_assemble(int i, const struct regstat *i_regs)
5477{
5478 const signed char *i_regmap = i_regs->regmap;
5479 int cc;
5480 int match;
5481 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5482 assem_debug("match=%d\n",match);
5483 int s1l,s2l;
5484 int unconditional=0,nop=0;
5485 int invert=0;
5486 int internal=internal_branch(cinfo[i].ba);
5487 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5488 if(!match) invert=1;
5489 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5490 if(i>(cinfo[i].ba-start)>>2) invert=1;
5491 #endif
5492 #ifdef __aarch64__
5493 invert=1; // because of near cond. branches
5494 #endif
5495
5496 if(dops[i].ooo) {
5497 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5498 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5499 }
5500 else {
5501 s1l=get_reg(i_regmap,dops[i].rs1);
5502 s2l=get_reg(i_regmap,dops[i].rs2);
5503 }
5504 if(dops[i].rs1==0&&dops[i].rs2==0)
5505 {
5506 if(dops[i].opcode&1) nop=1;
5507 else unconditional=1;
5508 //assert(dops[i].opcode!=5);
5509 //assert(dops[i].opcode!=7);
5510 //assert(dops[i].opcode!=0x15);
5511 //assert(dops[i].opcode!=0x17);
5512 }
5513 else if(dops[i].rs1==0)
5514 {
5515 s1l=s2l;
5516 s2l=-1;
5517 }
5518 else if(dops[i].rs2==0)
5519 {
5520 s2l=-1;
5521 }
5522
5523 if(dops[i].ooo) {
5524 // Out of order execution (delay slot first)
5525 //printf("OOOE\n");
5526 address_generation(i+1,i_regs,regs[i].regmap_entry);
5527 ds_assemble(i+1,i_regs);
5528 int adj;
5529 uint64_t bc_unneeded=branch_regs[i].u;
5530 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5531 bc_unneeded|=1;
5532 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5533 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5534 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5535 cc=get_reg(branch_regs[i].regmap,CCREG);
5536 assert(cc==HOST_CCREG);
5537 if(unconditional)
5538 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5539 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5540 //assem_debug("cycle count (adj)\n");
5541 if(unconditional) {
5542 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5543 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5544 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5545 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5546 if(internal)
5547 assem_debug("branch: internal\n");
5548 else
5549 assem_debug("branch: external\n");
5550 if (internal && dops[(cinfo[i].ba-start)>>2].is_ds) {
5551 ds_assemble_entry(i);
5552 }
5553 else {
5554 add_to_linker(out,cinfo[i].ba,internal);
5555 emit_jmp(0);
5556 }
5557 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5558 if(((u_int)out)&7) emit_addnop(0);
5559 #endif
5560 }
5561 }
5562 else if(nop) {
5563 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5564 void *jaddr=out;
5565 emit_jns(0);
5566 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5567 }
5568 else {
5569 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5570 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5571 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5572
5573 //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]);
5574 assert(s1l>=0);
5575 if(dops[i].opcode==4) // BEQ
5576 {
5577 if(s2l>=0) emit_cmp(s1l,s2l);
5578 else emit_test(s1l,s1l);
5579 if(invert){
5580 nottaken=out;
5581 emit_jne(DJT_1);
5582 }else{
5583 add_to_linker(out,cinfo[i].ba,internal);
5584 emit_jeq(0);
5585 }
5586 }
5587 if(dops[i].opcode==5) // BNE
5588 {
5589 if(s2l>=0) emit_cmp(s1l,s2l);
5590 else emit_test(s1l,s1l);
5591 if(invert){
5592 nottaken=out;
5593 emit_jeq(DJT_1);
5594 }else{
5595 add_to_linker(out,cinfo[i].ba,internal);
5596 emit_jne(0);
5597 }
5598 }
5599 if(dops[i].opcode==6) // BLEZ
5600 {
5601 emit_cmpimm(s1l,1);
5602 if(invert){
5603 nottaken=out;
5604 emit_jge(DJT_1);
5605 }else{
5606 add_to_linker(out,cinfo[i].ba,internal);
5607 emit_jl(0);
5608 }
5609 }
5610 if(dops[i].opcode==7) // BGTZ
5611 {
5612 emit_cmpimm(s1l,1);
5613 if(invert){
5614 nottaken=out;
5615 emit_jl(DJT_1);
5616 }else{
5617 add_to_linker(out,cinfo[i].ba,internal);
5618 emit_jge(0);
5619 }
5620 }
5621 if(invert) {
5622 if(taken) set_jump_target(taken, out);
5623 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5624 if (match && (!internal || !dops[(cinfo[i].ba-start)>>2].is_ds)) {
5625 if(adj) {
5626 emit_addimm(cc,-adj,cc);
5627 add_to_linker(out,cinfo[i].ba,internal);
5628 }else{
5629 emit_addnop(13);
5630 add_to_linker(out,cinfo[i].ba,internal*2);
5631 }
5632 emit_jmp(0);
5633 }else
5634 #endif
5635 {
5636 if(adj) emit_addimm(cc,-adj,cc);
5637 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5638 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5639 if(internal)
5640 assem_debug("branch: internal\n");
5641 else
5642 assem_debug("branch: external\n");
5643 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5644 ds_assemble_entry(i);
5645 }
5646 else {
5647 add_to_linker(out,cinfo[i].ba,internal);
5648 emit_jmp(0);
5649 }
5650 }
5651 set_jump_target(nottaken, out);
5652 }
5653
5654 if(nottaken1) set_jump_target(nottaken1, out);
5655 if(adj) {
5656 if(!invert) emit_addimm(cc,adj,cc);
5657 }
5658 } // (!unconditional)
5659 } // if(ooo)
5660 else
5661 {
5662 // In-order execution (branch first)
5663 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5664 if(!unconditional&&!nop) {
5665 //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]);
5666 assert(s1l>=0);
5667 if((dops[i].opcode&0x2f)==4) // BEQ
5668 {
5669 if(s2l>=0) emit_cmp(s1l,s2l);
5670 else emit_test(s1l,s1l);
5671 nottaken=out;
5672 emit_jne(DJT_2);
5673 }
5674 if((dops[i].opcode&0x2f)==5) // BNE
5675 {
5676 if(s2l>=0) emit_cmp(s1l,s2l);
5677 else emit_test(s1l,s1l);
5678 nottaken=out;
5679 emit_jeq(DJT_2);
5680 }
5681 if((dops[i].opcode&0x2f)==6) // BLEZ
5682 {
5683 emit_cmpimm(s1l,1);
5684 nottaken=out;
5685 emit_jge(DJT_2);
5686 }
5687 if((dops[i].opcode&0x2f)==7) // BGTZ
5688 {
5689 emit_cmpimm(s1l,1);
5690 nottaken=out;
5691 emit_jl(DJT_2);
5692 }
5693 } // if(!unconditional)
5694 int adj;
5695 uint64_t ds_unneeded=branch_regs[i].u;
5696 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5697 ds_unneeded|=1;
5698 // branch taken
5699 if(!nop) {
5700 if(taken) set_jump_target(taken, out);
5701 assem_debug("1:\n");
5702 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5703 // load regs
5704 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5705 address_generation(i+1,&branch_regs[i],0);
5706 if (ram_offset)
5707 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5708 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5709 ds_assemble(i+1,&branch_regs[i]);
5710 cc=get_reg(branch_regs[i].regmap,CCREG);
5711 if(cc==-1) {
5712 emit_loadreg(CCREG,cc=HOST_CCREG);
5713 // CHECK: Is the following instruction (fall thru) allocated ok?
5714 }
5715 assert(cc==HOST_CCREG);
5716 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5717 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5718 assem_debug("cycle count (adj)\n");
5719 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5720 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5721 if(internal)
5722 assem_debug("branch: internal\n");
5723 else
5724 assem_debug("branch: external\n");
5725 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5726 ds_assemble_entry(i);
5727 }
5728 else {
5729 add_to_linker(out,cinfo[i].ba,internal);
5730 emit_jmp(0);
5731 }
5732 }
5733 // branch not taken
5734 if(!unconditional) {
5735 if(nottaken1) set_jump_target(nottaken1, out);
5736 set_jump_target(nottaken, out);
5737 assem_debug("2:\n");
5738 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5739 // load regs
5740 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5741 address_generation(i+1,&branch_regs[i],0);
5742 if (ram_offset)
5743 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5744 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5745 ds_assemble(i+1,&branch_regs[i]);
5746 cc=get_reg(branch_regs[i].regmap,CCREG);
5747 if (cc == -1) {
5748 // Cycle count isn't in a register, temporarily load it then write it out
5749 emit_loadreg(CCREG,HOST_CCREG);
5750 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5751 void *jaddr=out;
5752 emit_jns(0);
5753 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5754 emit_storereg(CCREG,HOST_CCREG);
5755 }
5756 else{
5757 cc=get_reg(i_regmap,CCREG);
5758 assert(cc==HOST_CCREG);
5759 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5760 void *jaddr=out;
5761 emit_jns(0);
5762 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5763 }
5764 }
5765 }
5766}
5767
5768static void sjump_assemble(int i, const struct regstat *i_regs)
5769{
5770 const signed char *i_regmap = i_regs->regmap;
5771 int cc;
5772 int match;
5773 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5774 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5775 int s1l;
5776 int unconditional=0,nevertaken=0;
5777 int invert=0;
5778 int internal=internal_branch(cinfo[i].ba);
5779 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5780 if(!match) invert=1;
5781 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5782 if(i>(cinfo[i].ba-start)>>2) invert=1;
5783 #endif
5784 #ifdef __aarch64__
5785 invert=1; // because of near cond. branches
5786 #endif
5787
5788 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5789 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5790
5791 if(dops[i].ooo) {
5792 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5793 }
5794 else {
5795 s1l=get_reg(i_regmap,dops[i].rs1);
5796 }
5797 if(dops[i].rs1==0)
5798 {
5799 if(dops[i].opcode2&1) unconditional=1;
5800 else nevertaken=1;
5801 // These are never taken (r0 is never less than zero)
5802 //assert(dops[i].opcode2!=0);
5803 //assert(dops[i].opcode2!=2);
5804 //assert(dops[i].opcode2!=0x10);
5805 //assert(dops[i].opcode2!=0x12);
5806 }
5807
5808 if(dops[i].ooo) {
5809 // Out of order execution (delay slot first)
5810 //printf("OOOE\n");
5811 address_generation(i+1,i_regs,regs[i].regmap_entry);
5812 ds_assemble(i+1,i_regs);
5813 int adj;
5814 uint64_t bc_unneeded=branch_regs[i].u;
5815 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5816 bc_unneeded|=1;
5817 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5818 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5819 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5820 if(dops[i].rt1==31) {
5821 int rt,return_address;
5822 rt=get_reg(branch_regs[i].regmap,31);
5823 //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]);
5824 if(rt>=0) {
5825 // Save the PC even if the branch is not taken
5826 return_address=start+i*4+8;
5827 emit_movimm(return_address,rt); // PC into link register
5828 #ifdef IMM_PREFETCH
5829 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5830 #endif
5831 }
5832 }
5833 cc=get_reg(branch_regs[i].regmap,CCREG);
5834 assert(cc==HOST_CCREG);
5835 if(unconditional)
5836 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5837 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5838 assem_debug("cycle count (adj)\n");
5839 if(unconditional) {
5840 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5841 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5842 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5843 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5844 if(internal)
5845 assem_debug("branch: internal\n");
5846 else
5847 assem_debug("branch: external\n");
5848 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5849 ds_assemble_entry(i);
5850 }
5851 else {
5852 add_to_linker(out,cinfo[i].ba,internal);
5853 emit_jmp(0);
5854 }
5855 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5856 if(((u_int)out)&7) emit_addnop(0);
5857 #endif
5858 }
5859 }
5860 else if(nevertaken) {
5861 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5862 void *jaddr=out;
5863 emit_jns(0);
5864 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5865 }
5866 else {
5867 void *nottaken = NULL;
5868 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5869 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5870 {
5871 assert(s1l>=0);
5872 if ((dops[i].opcode2 & 1) == 0) // BLTZ/BLTZAL
5873 {
5874 emit_test(s1l,s1l);
5875 if(invert){
5876 nottaken=out;
5877 emit_jns(DJT_1);
5878 }else{
5879 add_to_linker(out,cinfo[i].ba,internal);
5880 emit_js(0);
5881 }
5882 }
5883 else // BGEZ/BGEZAL
5884 {
5885 emit_test(s1l,s1l);
5886 if(invert){
5887 nottaken=out;
5888 emit_js(DJT_1);
5889 }else{
5890 add_to_linker(out,cinfo[i].ba,internal);
5891 emit_jns(0);
5892 }
5893 }
5894 }
5895
5896 if(invert) {
5897 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5898 if (match && (!internal || !dops[(cinfo[i].ba - start) >> 2].is_ds)) {
5899 if(adj) {
5900 emit_addimm(cc,-adj,cc);
5901 add_to_linker(out,cinfo[i].ba,internal);
5902 }else{
5903 emit_addnop(13);
5904 add_to_linker(out,cinfo[i].ba,internal*2);
5905 }
5906 emit_jmp(0);
5907 }else
5908 #endif
5909 {
5910 if(adj) emit_addimm(cc,-adj,cc);
5911 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5912 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5913 if(internal)
5914 assem_debug("branch: internal\n");
5915 else
5916 assem_debug("branch: external\n");
5917 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5918 ds_assemble_entry(i);
5919 }
5920 else {
5921 add_to_linker(out,cinfo[i].ba,internal);
5922 emit_jmp(0);
5923 }
5924 }
5925 set_jump_target(nottaken, out);
5926 }
5927
5928 if(adj) {
5929 if(!invert) emit_addimm(cc,adj,cc);
5930 }
5931 } // (!unconditional)
5932 } // if(ooo)
5933 else
5934 {
5935 // In-order execution (branch first)
5936 //printf("IOE\n");
5937 void *nottaken = NULL;
5938 if (!unconditional && !nevertaken) {
5939 assert(s1l >= 0);
5940 emit_test(s1l, s1l);
5941 }
5942 if (dops[i].rt1 == 31) {
5943 int rt, return_address;
5944 rt = get_reg(branch_regs[i].regmap,31);
5945 if(rt >= 0) {
5946 // Save the PC even if the branch is not taken
5947 return_address = start + i*4+8;
5948 emit_movimm(return_address, rt); // PC into link register
5949 #ifdef IMM_PREFETCH
5950 emit_prefetch(hash_table_get(return_address));
5951 #endif
5952 }
5953 }
5954 if (!unconditional && !nevertaken) {
5955 nottaken = out;
5956 if (!(dops[i].opcode2 & 1)) // BLTZ/BLTZAL
5957 emit_jns(DJT_1);
5958 else // BGEZ/BGEZAL
5959 emit_js(DJT_1);
5960 }
5961 int adj;
5962 uint64_t ds_unneeded=branch_regs[i].u;
5963 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5964 ds_unneeded|=1;
5965 // branch taken
5966 if(!nevertaken) {
5967 //assem_debug("1:\n");
5968 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5969 // load regs
5970 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5971 address_generation(i+1,&branch_regs[i],0);
5972 if (ram_offset)
5973 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5974 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5975 ds_assemble(i+1,&branch_regs[i]);
5976 cc=get_reg(branch_regs[i].regmap,CCREG);
5977 if(cc==-1) {
5978 emit_loadreg(CCREG,cc=HOST_CCREG);
5979 // CHECK: Is the following instruction (fall thru) allocated ok?
5980 }
5981 assert(cc==HOST_CCREG);
5982 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5983 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5984 assem_debug("cycle count (adj)\n");
5985 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5986 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5987 if(internal)
5988 assem_debug("branch: internal\n");
5989 else
5990 assem_debug("branch: external\n");
5991 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5992 ds_assemble_entry(i);
5993 }
5994 else {
5995 add_to_linker(out,cinfo[i].ba,internal);
5996 emit_jmp(0);
5997 }
5998 }
5999 // branch not taken
6000 if(!unconditional) {
6001 if (!nevertaken) {
6002 assert(nottaken);
6003 set_jump_target(nottaken, out);
6004 }
6005 assem_debug("1:\n");
6006 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6007 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6008 address_generation(i+1,&branch_regs[i],0);
6009 if (ram_offset)
6010 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6011 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6012 ds_assemble(i+1,&branch_regs[i]);
6013 cc=get_reg(branch_regs[i].regmap,CCREG);
6014 if (cc == -1) {
6015 // Cycle count isn't in a register, temporarily load it then write it out
6016 emit_loadreg(CCREG,HOST_CCREG);
6017 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
6018 void *jaddr=out;
6019 emit_jns(0);
6020 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6021 emit_storereg(CCREG,HOST_CCREG);
6022 }
6023 else{
6024 cc=get_reg(i_regmap,CCREG);
6025 assert(cc==HOST_CCREG);
6026 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
6027 void *jaddr=out;
6028 emit_jns(0);
6029 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6030 }
6031 }
6032 }
6033}
6034
6035static void check_regmap(signed char *regmap)
6036{
6037#ifndef NDEBUG
6038 int i,j;
6039 for (i = 0; i < HOST_REGS; i++) {
6040 if (regmap[i] < 0)
6041 continue;
6042 for (j = i + 1; j < HOST_REGS; j++)
6043 assert(regmap[i] != regmap[j]);
6044 }
6045#endif
6046}
6047
6048#ifdef DISASM
6049#include <inttypes.h>
6050static char insn[MAXBLOCK][10];
6051
6052#define set_mnemonic(i_, n_) \
6053 strcpy(insn[i_], n_)
6054
6055void print_regmap(const char *name, const signed char *regmap)
6056{
6057 char buf[5];
6058 int i, l;
6059 fputs(name, stdout);
6060 for (i = 0; i < HOST_REGS; i++) {
6061 l = 0;
6062 if (regmap[i] >= 0)
6063 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6064 for (; l < 3; l++)
6065 buf[l] = ' ';
6066 buf[l] = 0;
6067 printf(" r%d=%s", i, buf);
6068 }
6069 fputs("\n", stdout);
6070}
6071
6072 /* disassembly */
6073void disassemble_inst(int i)
6074{
6075 if (dops[i].bt) printf("*"); else printf(" ");
6076 switch(dops[i].itype) {
6077 case UJUMP:
6078 printf (" %x: %s %8x\n",start+i*4,insn[i],cinfo[i].ba);break;
6079 case CJUMP:
6080 printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):cinfo[i].ba);break;
6081 case SJUMP:
6082 printf (" %x: %s r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14));break;
6083 case RJUMP:
6084 if (dops[i].opcode2 == 9 && dops[i].rt1 != 31)
6085 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6086 else
6087 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6088 break;
6089 case IMM16:
6090 if(dops[i].opcode==0xf) //LUI
6091 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,cinfo[i].imm&0xffff);
6092 else
6093 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6094 break;
6095 case LOAD:
6096 case LOADLR:
6097 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6098 break;
6099 case STORE:
6100 case STORELR:
6101 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,cinfo[i].imm);
6102 break;
6103 case ALU:
6104 case SHIFT:
6105 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6106 break;
6107 case MULTDIV:
6108 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6109 break;
6110 case SHIFTIMM:
6111 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6112 break;
6113 case MOV:
6114 if((dops[i].opcode2&0x1d)==0x10)
6115 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6116 else if((dops[i].opcode2&0x1d)==0x11)
6117 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6118 else
6119 printf (" %x: %s\n",start+i*4,insn[i]);
6120 break;
6121 case COP0:
6122 if(dops[i].opcode2==0)
6123 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6124 else if(dops[i].opcode2==4)
6125 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6126 else printf (" %x: %s\n",start+i*4,insn[i]);
6127 break;
6128 case COP2:
6129 if(dops[i].opcode2<3)
6130 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6131 else if(dops[i].opcode2>3)
6132 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6133 else printf (" %x: %s\n",start+i*4,insn[i]);
6134 break;
6135 case C2LS:
6136 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,cinfo[i].imm);
6137 break;
6138 case INTCALL:
6139 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6140 break;
6141 default:
6142 //printf (" %s %8x\n",insn[i],source[i]);
6143 printf (" %x: %s\n",start+i*4,insn[i]);
6144 }
6145 #ifndef REGMAP_PRINT
6146 return;
6147 #endif
6148 printf("D: %"PRIx64" WD: %"PRIx64" U: %"PRIx64" hC: %x hWC: %x hLC: %x\n",
6149 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i],
6150 regs[i].isconst, regs[i].wasconst, regs[i].loadedconst);
6151 print_regmap("pre: ", regmap_pre[i]);
6152 print_regmap("entry: ", regs[i].regmap_entry);
6153 print_regmap("map: ", regs[i].regmap);
6154 if (dops[i].is_jump) {
6155 print_regmap("bentry:", branch_regs[i].regmap_entry);
6156 print_regmap("bmap: ", branch_regs[i].regmap);
6157 }
6158}
6159#else
6160#define set_mnemonic(i_, n_)
6161static void disassemble_inst(int i) {}
6162#endif // DISASM
6163
6164#define DRC_TEST_VAL 0x74657374
6165
6166static noinline void new_dynarec_test(void)
6167{
6168 int (*testfunc)(void);
6169 void *beginning;
6170 int ret[2];
6171 size_t i;
6172
6173 // check structure linkage
6174 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6175 {
6176 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6177 }
6178
6179 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6180 new_dynarec_test, out);
6181 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6182
6183 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6184 out = ndrc->translation_cache;
6185 beginning = start_block();
6186 emit_movimm(DRC_TEST_VAL + i, 0); // test
6187 emit_ret();
6188 literal_pool(0);
6189 end_block(beginning);
6190 testfunc = beginning;
6191 ret[i] = testfunc();
6192 }
6193
6194 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6195 SysPrintf("test passed.\n");
6196 else
6197 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6198 out = ndrc->translation_cache;
6199}
6200
6201// clear the state completely, instead of just marking
6202// things invalid like invalidate_all_pages() does
6203void new_dynarec_clear_full(void)
6204{
6205 int n;
6206 out = ndrc->translation_cache;
6207 memset(invalid_code,1,sizeof(invalid_code));
6208 memset(hash_table,0xff,sizeof(hash_table));
6209 memset(mini_ht,-1,sizeof(mini_ht));
6210 memset(shadow,0,sizeof(shadow));
6211 copy=shadow;
6212 expirep = EXPIRITY_OFFSET;
6213 pending_exception=0;
6214 literalcount=0;
6215 stop_after_jal=0;
6216 inv_code_start=inv_code_end=~0;
6217 hack_addr=0;
6218 f1_hack=0;
6219 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6220 blocks_clear(&blocks[n]);
6221 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6222 free(jumps[n]);
6223 jumps[n] = NULL;
6224 }
6225 stat_clear(stat_blocks);
6226 stat_clear(stat_links);
6227
6228 cycle_multiplier_old = Config.cycle_multiplier;
6229 new_dynarec_hacks_old = new_dynarec_hacks;
6230}
6231
6232void new_dynarec_init(void)
6233{
6234 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6235
6236#ifdef _3DS
6237 check_rosalina();
6238#endif
6239#ifdef BASE_ADDR_DYNAMIC
6240 #ifdef VITA
6241 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6242 if (sceBlock <= 0)
6243 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6244 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6245 if (ret < 0)
6246 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6247 sceKernelOpenVMDomain();
6248 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6249 #elif defined(_MSC_VER)
6250 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6251 PAGE_EXECUTE_READWRITE);
6252 #elif defined(HAVE_LIBNX)
6253 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6254 if (R_FAILED(rc))
6255 SysPrintf("jitCreate failed: %08x\n", rc);
6256 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6257 jitTransitionToWritable(&g_jit);
6258 ndrc = g_jit.rx_addr;
6259 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6260 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6261 #else
6262 uintptr_t desired_addr = 0;
6263 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6264 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6265 int fd = -1;
6266 #ifdef __ELF__
6267 extern char _end;
6268 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6269 #endif
6270 #ifdef TC_WRITE_OFFSET
6271 // mostly for testing
6272 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6273 ftruncate(fd, sizeof(*ndrc));
6274 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6275 (flags = MAP_SHARED), fd, 0);
6276 assert(mw != MAP_FAILED);
6277 prot = PROT_READ | PROT_EXEC;
6278 #endif
6279 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6280 if (ndrc == MAP_FAILED) {
6281 SysPrintf("mmap() failed: %s\n", strerror(errno));
6282 abort();
6283 }
6284 #ifdef TC_WRITE_OFFSET
6285 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6286 #endif
6287 #endif
6288#else
6289 #ifndef NO_WRITE_EXEC
6290 // not all systems allow execute in data segment by default
6291 // size must be 4K aligned for 3DS?
6292 if (mprotect(ndrc, sizeof(*ndrc),
6293 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6294 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6295 #endif
6296#endif
6297 out = ndrc->translation_cache;
6298 new_dynarec_clear_full();
6299#ifdef HOST_IMM8
6300 // Copy this into local area so we don't have to put it in every literal pool
6301 invc_ptr=invalid_code;
6302#endif
6303 arch_init();
6304 new_dynarec_test();
6305 ram_offset=(uintptr_t)rdram-0x80000000;
6306 if (ram_offset!=0)
6307 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6308 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6309 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6310}
6311
6312void new_dynarec_cleanup(void)
6313{
6314 int n;
6315#ifdef BASE_ADDR_DYNAMIC
6316 #ifdef VITA
6317 // sceBlock is managed by retroarch's bootstrap code
6318 //sceKernelFreeMemBlock(sceBlock);
6319 //sceBlock = -1;
6320 #elif defined(HAVE_LIBNX)
6321 jitClose(&g_jit);
6322 ndrc = NULL;
6323 #else
6324 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6325 SysPrintf("munmap() failed\n");
6326 ndrc = NULL;
6327 #endif
6328#endif
6329 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6330 blocks_clear(&blocks[n]);
6331 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6332 free(jumps[n]);
6333 jumps[n] = NULL;
6334 }
6335 stat_clear(stat_blocks);
6336 stat_clear(stat_links);
6337 new_dynarec_print_stats();
6338}
6339
6340static u_int *get_source_start(u_int addr, u_int *limit)
6341{
6342 if (addr < 0x00200000 ||
6343 (0xa0000000 <= addr && addr < 0xa0200000))
6344 {
6345 // used for BIOS calls mostly?
6346 *limit = (addr&0xa0000000)|0x00200000;
6347 return (u_int *)(rdram + (addr&0x1fffff));
6348 }
6349 else if (!Config.HLE && (
6350 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6351 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6352 {
6353 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6354 // but timings in PCSX are too tied to the interpreter's 2-per-insn assumption
6355 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6356 cycle_multiplier_active = 200;
6357
6358 *limit = (addr & 0xfff00000) | 0x80000;
6359 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6360 }
6361 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6362 *limit = (addr & 0x80600000) + 0x00200000;
6363 return (u_int *)(rdram + (addr&0x1fffff));
6364 }
6365 return NULL;
6366}
6367
6368static u_int scan_for_ret(u_int addr)
6369{
6370 u_int limit = 0;
6371 u_int *mem;
6372
6373 mem = get_source_start(addr, &limit);
6374 if (mem == NULL)
6375 return addr;
6376
6377 if (limit > addr + 0x1000)
6378 limit = addr + 0x1000;
6379 for (; addr < limit; addr += 4, mem++) {
6380 if (*mem == 0x03e00008) // jr $ra
6381 return addr + 8;
6382 }
6383 return addr;
6384}
6385
6386struct savestate_block {
6387 uint32_t addr;
6388 uint32_t regflags;
6389};
6390
6391static int addr_cmp(const void *p1_, const void *p2_)
6392{
6393 const struct savestate_block *p1 = p1_, *p2 = p2_;
6394 return p1->addr - p2->addr;
6395}
6396
6397int new_dynarec_save_blocks(void *save, int size)
6398{
6399 struct savestate_block *sblocks = save;
6400 int maxcount = size / sizeof(sblocks[0]);
6401 struct savestate_block tmp_blocks[1024];
6402 struct block_info *block;
6403 int p, s, d, o, bcnt;
6404 u_int addr;
6405
6406 o = 0;
6407 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6408 bcnt = 0;
6409 for (block = blocks[p]; block != NULL; block = block->next) {
6410 if (block->is_dirty)
6411 continue;
6412 tmp_blocks[bcnt].addr = block->start;
6413 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6414 bcnt++;
6415 }
6416 if (bcnt < 1)
6417 continue;
6418 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6419
6420 addr = tmp_blocks[0].addr;
6421 for (s = d = 0; s < bcnt; s++) {
6422 if (tmp_blocks[s].addr < addr)
6423 continue;
6424 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6425 tmp_blocks[d++] = tmp_blocks[s];
6426 addr = scan_for_ret(tmp_blocks[s].addr);
6427 }
6428
6429 if (o + d > maxcount)
6430 d = maxcount - o;
6431 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6432 o += d;
6433 }
6434
6435 return o * sizeof(sblocks[0]);
6436}
6437
6438void new_dynarec_load_blocks(const void *save, int size)
6439{
6440 const struct savestate_block *sblocks = save;
6441 int count = size / sizeof(sblocks[0]);
6442 struct block_info *block;
6443 u_int regs_save[32];
6444 u_int page;
6445 uint32_t f;
6446 int i, b;
6447
6448 // restore clean blocks, if any
6449 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6450 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6451 if (!block->is_dirty)
6452 continue;
6453 assert(block->source && block->copy);
6454 if (memcmp(block->source, block->copy, block->len))
6455 continue;
6456
6457 // see try_restore_block
6458 block->is_dirty = 0;
6459 mark_invalid_code(block->start, block->len, 0);
6460 i++;
6461 }
6462 }
6463 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6464
6465 // change GPRs for speculation to at least partially work..
6466 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6467 for (i = 1; i < 32; i++)
6468 psxRegs.GPR.r[i] = 0x80000000;
6469
6470 for (b = 0; b < count; b++) {
6471 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6472 if (f & 1)
6473 psxRegs.GPR.r[i] = 0x1f800000;
6474 }
6475
6476 ndrc_get_addr_ht(sblocks[b].addr);
6477
6478 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6479 if (f & 1)
6480 psxRegs.GPR.r[i] = 0x80000000;
6481 }
6482 }
6483
6484 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6485}
6486
6487void new_dynarec_print_stats(void)
6488{
6489#ifdef STAT_PRINT
6490 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6491 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6492 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6493 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6494 out - ndrc->translation_cache, stat_blocks, stat_links);
6495 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6496 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6497 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6498#endif
6499}
6500
6501static int apply_hacks(void)
6502{
6503 int i;
6504 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6505 return 0;
6506 /* special hack(s) */
6507 for (i = 0; i < slen - 4; i++)
6508 {
6509 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6510 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6511 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6512 && cinfo[i+3].imm == 0x6e40 && dops[i+3].rs1 == 2)
6513 {
6514 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6515 dops[i + 3].itype = NOP;
6516 }
6517 }
6518 i = slen;
6519 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6520 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6521 && dops[i-7].itype == STORE)
6522 {
6523 i = i-8;
6524 if (dops[i].itype == IMM16)
6525 i--;
6526 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6527 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6528 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6529 {
6530 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6531 f1_hack = 1;
6532 return 1;
6533 }
6534 }
6535 return 0;
6536}
6537
6538static int is_ld_use_hazard(int ld_rt, const struct decoded_insn *op)
6539{
6540 return ld_rt != 0 && (ld_rt == op->rs1 || ld_rt == op->rs2)
6541 && op->itype != LOADLR && op->itype != CJUMP && op->itype != SJUMP;
6542}
6543
6544static void force_intcall(int i)
6545{
6546 memset(&dops[i], 0, sizeof(dops[i]));
6547 dops[i].itype = INTCALL;
6548 dops[i].rs1 = CCREG;
6549 dops[i].is_exception = 1;
6550 cinfo[i].ba = -1;
6551}
6552
6553static void disassemble_one(int i, u_int src)
6554{
6555 unsigned int type, op, op2, op3;
6556 memset(&dops[i], 0, sizeof(dops[i]));
6557 memset(&cinfo[i], 0, sizeof(cinfo[i]));
6558 cinfo[i].ba = -1;
6559 cinfo[i].addr = -1;
6560 dops[i].opcode = op = src >> 26;
6561 op2 = 0;
6562 type = INTCALL;
6563 set_mnemonic(i, "???");
6564 switch(op)
6565 {
6566 case 0x00: set_mnemonic(i, "special");
6567 op2 = src & 0x3f;
6568 switch(op2)
6569 {
6570 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6571 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6572 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6573 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6574 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6575 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6576 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6577 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6578 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6579 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6580 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6581 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6582 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6583 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6584 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6585 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6586 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6587 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6588 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6589 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6590 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6591 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6592 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6593 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6594 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6595 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6596 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6597 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6598 }
6599 break;
6600 case 0x01: set_mnemonic(i, "regimm");
6601 type = SJUMP;
6602 op2 = (src >> 16) & 0x1f;
6603 switch(op2)
6604 {
6605 case 0x10: set_mnemonic(i, "BLTZAL"); break;
6606 case 0x11: set_mnemonic(i, "BGEZAL"); break;
6607 default:
6608 if (op2 & 1)
6609 set_mnemonic(i, "BGEZ");
6610 else
6611 set_mnemonic(i, "BLTZ");
6612 }
6613 break;
6614 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6615 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6616 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6617 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6618 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6619 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6620 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6621 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6622 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6623 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6624 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6625 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6626 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6627 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6628 case 0x10: set_mnemonic(i, "COP0");
6629 op2 = (src >> 21) & 0x1f;
6630 if (op2 & 0x10) {
6631 op3 = src & 0x1f;
6632 switch (op3)
6633 {
6634 case 0x01: case 0x02: case 0x06: case 0x08: type = INTCALL; break;
6635 case 0x10: set_mnemonic(i, "RFE"); type=RFE; break;
6636 default: type = OTHER; break;
6637 }
6638 break;
6639 }
6640 switch(op2)
6641 {
6642 u32 rd;
6643 case 0x00:
6644 set_mnemonic(i, "MFC0");
6645 rd = (src >> 11) & 0x1F;
6646 if (!(0x00000417u & (1u << rd)))
6647 type = COP0;
6648 break;
6649 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6650 case 0x02:
6651 case 0x06: type = INTCALL; break;
6652 default: type = OTHER; break;
6653 }
6654 break;
6655 case 0x11: set_mnemonic(i, "COP1");
6656 op2 = (src >> 21) & 0x1f;
6657 break;
6658 case 0x12: set_mnemonic(i, "COP2");
6659 op2 = (src >> 21) & 0x1f;
6660 if (op2 & 0x10) {
6661 type = OTHER;
6662 if (gte_handlers[src & 0x3f] != NULL) {
6663#ifdef DISASM
6664 if (gte_regnames[src & 0x3f] != NULL)
6665 strcpy(insn[i], gte_regnames[src & 0x3f]);
6666 else
6667 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", src & 0x3f);
6668#endif
6669 type = C2OP;
6670 }
6671 }
6672 else switch(op2)
6673 {
6674 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6675 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6676 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6677 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6678 }
6679 break;
6680 case 0x13: set_mnemonic(i, "COP3");
6681 op2 = (src >> 21) & 0x1f;
6682 break;
6683 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6684 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6685 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6686 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6687 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6688 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6689 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6690 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6691 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6692 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6693 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6694 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6695 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6696 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6697 case 0x3B:
6698 if (Config.HLE && (src & 0x03ffffff) < ARRAY_SIZE(psxHLEt)) {
6699 set_mnemonic(i, "HLECALL");
6700 type = HLECALL;
6701 }
6702 break;
6703 default:
6704 break;
6705 }
6706 if (type == INTCALL)
6707 SysPrintf("NI %08x @%08x (%08x)\n", src, start + i*4, start);
6708 dops[i].itype=type;
6709 dops[i].opcode2=op2;
6710 /* Get registers/immediates */
6711 dops[i].use_lt1=0;
6712 gte_rs[i]=gte_rt[i]=0;
6713 dops[i].rs1 = 0;
6714 dops[i].rs2 = 0;
6715 dops[i].rt1 = 0;
6716 dops[i].rt2 = 0;
6717 switch(type) {
6718 case LOAD:
6719 dops[i].rs1 = (src >> 21) & 0x1f;
6720 dops[i].rt1 = (src >> 16) & 0x1f;
6721 cinfo[i].imm = (short)src;
6722 break;
6723 case STORE:
6724 case STORELR:
6725 dops[i].rs1 = (src >> 21) & 0x1f;
6726 dops[i].rs2 = (src >> 16) & 0x1f;
6727 cinfo[i].imm = (short)src;
6728 break;
6729 case LOADLR:
6730 // LWL/LWR only load part of the register,
6731 // therefore the target register must be treated as a source too
6732 dops[i].rs1 = (src >> 21) & 0x1f;
6733 dops[i].rs2 = (src >> 16) & 0x1f;
6734 dops[i].rt1 = (src >> 16) & 0x1f;
6735 cinfo[i].imm = (short)src;
6736 break;
6737 case IMM16:
6738 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6739 else dops[i].rs1 = (src >> 21) & 0x1f;
6740 dops[i].rs2 = 0;
6741 dops[i].rt1 = (src >> 16) & 0x1f;
6742 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6743 cinfo[i].imm = (unsigned short)src;
6744 }else{
6745 cinfo[i].imm = (short)src;
6746 }
6747 break;
6748 case UJUMP:
6749 // The JAL instruction writes to r31.
6750 if (op&1) {
6751 dops[i].rt1=31;
6752 }
6753 dops[i].rs2=CCREG;
6754 break;
6755 case RJUMP:
6756 dops[i].rs1 = (src >> 21) & 0x1f;
6757 // The JALR instruction writes to rd.
6758 if (op2&1) {
6759 dops[i].rt1 = (src >> 11) & 0x1f;
6760 }
6761 dops[i].rs2=CCREG;
6762 break;
6763 case CJUMP:
6764 dops[i].rs1 = (src >> 21) & 0x1f;
6765 dops[i].rs2 = (src >> 16) & 0x1f;
6766 if(op&2) { // BGTZ/BLEZ
6767 dops[i].rs2=0;
6768 }
6769 break;
6770 case SJUMP:
6771 dops[i].rs1 = (src >> 21) & 0x1f;
6772 dops[i].rs2 = CCREG;
6773 if (op2 == 0x10 || op2 == 0x11) { // BxxAL
6774 dops[i].rt1 = 31;
6775 // NOTE: If the branch is not taken, r31 is still overwritten
6776 }
6777 break;
6778 case ALU:
6779 dops[i].rs1=(src>>21)&0x1f; // source
6780 dops[i].rs2=(src>>16)&0x1f; // subtract amount
6781 dops[i].rt1=(src>>11)&0x1f; // destination
6782 break;
6783 case MULTDIV:
6784 dops[i].rs1=(src>>21)&0x1f; // source
6785 dops[i].rs2=(src>>16)&0x1f; // divisor
6786 dops[i].rt1=HIREG;
6787 dops[i].rt2=LOREG;
6788 break;
6789 case MOV:
6790 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6791 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6792 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6793 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6794 if((op2&0x1d)==0x10) dops[i].rt1=(src>>11)&0x1f; // MFxx
6795 if((op2&0x1d)==0x11) dops[i].rs1=(src>>21)&0x1f; // MTxx
6796 break;
6797 case SHIFT:
6798 dops[i].rs1=(src>>16)&0x1f; // target of shift
6799 dops[i].rs2=(src>>21)&0x1f; // shift amount
6800 dops[i].rt1=(src>>11)&0x1f; // destination
6801 break;
6802 case SHIFTIMM:
6803 dops[i].rs1=(src>>16)&0x1f;
6804 dops[i].rs2=0;
6805 dops[i].rt1=(src>>11)&0x1f;
6806 cinfo[i].imm=(src>>6)&0x1f;
6807 break;
6808 case COP0:
6809 if(op2==0) dops[i].rt1=(src>>16)&0x1F; // MFC0
6810 if(op2==4) dops[i].rs1=(src>>16)&0x1F; // MTC0
6811 if(op2==4&&((src>>11)&0x1e)==12) dops[i].rs2=CCREG;
6812 break;
6813 case COP2:
6814 if(op2<3) dops[i].rt1=(src>>16)&0x1F; // MFC2/CFC2
6815 if(op2>3) dops[i].rs1=(src>>16)&0x1F; // MTC2/CTC2
6816 int gr=(src>>11)&0x1F;
6817 switch(op2)
6818 {
6819 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6820 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6821 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6822 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6823 }
6824 break;
6825 case C2LS:
6826 dops[i].rs1=(src>>21)&0x1F;
6827 cinfo[i].imm=(short)src;
6828 if(op==0x32) gte_rt[i]=1ll<<((src>>16)&0x1F); // LWC2
6829 else gte_rs[i]=1ll<<((src>>16)&0x1F); // SWC2
6830 break;
6831 case C2OP:
6832 gte_rs[i]=gte_reg_reads[src&0x3f];
6833 gte_rt[i]=gte_reg_writes[src&0x3f];
6834 gte_rt[i]|=1ll<<63; // every op changes flags
6835 if((src&0x3f)==GTE_MVMVA) {
6836 int v = (src >> 15) & 3;
6837 gte_rs[i]&=~0xe3fll;
6838 if(v==3) gte_rs[i]|=0xe00ll;
6839 else gte_rs[i]|=3ll<<(v*2);
6840 }
6841 break;
6842 case SYSCALL:
6843 case HLECALL:
6844 case INTCALL:
6845 dops[i].rs1=CCREG;
6846 break;
6847 default:
6848 break;
6849 }
6850}
6851
6852static noinline void pass1_disassemble(u_int pagelimit)
6853{
6854 int i, j, done = 0, ni_count = 0;
6855
6856 for (i = 0; !done; i++)
6857 {
6858 int force_j_to_interpreter = 0;
6859 unsigned int type, op, op2;
6860
6861 disassemble_one(i, source[i]);
6862 type = dops[i].itype;
6863 op = dops[i].opcode;
6864 op2 = dops[i].opcode2;
6865
6866 /* Calculate branch target addresses */
6867 if(type==UJUMP)
6868 cinfo[i].ba=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6869 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
6870 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6871 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
6872 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6873 else if(type==CJUMP||type==SJUMP)
6874 cinfo[i].ba=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6875
6876 /* simplify always (not)taken branches */
6877 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
6878 dops[i].rs1 = dops[i].rs2 = 0;
6879 if (!(op & 1)) {
6880 dops[i].itype = type = UJUMP;
6881 dops[i].rs2 = CCREG;
6882 }
6883 }
6884 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
6885 dops[i].itype = type = UJUMP;
6886
6887 dops[i].is_jump = type == RJUMP || type == UJUMP || type == CJUMP || type == SJUMP;
6888 dops[i].is_ujump = type == RJUMP || type == UJUMP;
6889 dops[i].is_load = type == LOAD || type == LOADLR || op == 0x32; // LWC2
6890 dops[i].is_delay_load = (dops[i].is_load || (source[i] & 0xf3d00000) == 0x40000000); // MFC/CFC
6891 dops[i].is_store = type == STORE || type == STORELR || op == 0x3a; // SWC2
6892 dops[i].is_exception = type == SYSCALL || type == HLECALL || type == INTCALL;
6893 dops[i].may_except = dops[i].is_exception || (type == ALU && (op2 == 0x20 || op2 == 0x22)) || op == 8;
6894
6895 if (((op & 0x37) == 0x21 || op == 0x25) // LH/SH/LHU
6896 && ((cinfo[i].imm & 1) || Config.PreciseExceptions))
6897 dops[i].may_except = 1;
6898 if (((op & 0x37) == 0x23 || (op & 0x37) == 0x32) // LW/SW/LWC2/SWC2
6899 && ((cinfo[i].imm & 3) || Config.PreciseExceptions))
6900 dops[i].may_except = 1;
6901
6902 /* rare messy cases to just pass over to the interpreter */
6903 if (i > 0 && dops[i-1].is_jump) {
6904 j = i - 1;
6905 // branch in delay slot?
6906 if (dops[i].is_jump) {
6907 // don't handle first branch and call interpreter if it's hit
6908 SysPrintf("branch in DS @%08x (%08x)\n", start + i*4, start);
6909 force_j_to_interpreter = 1;
6910 }
6911 // load delay detection through a branch
6912 else if (dops[i].is_delay_load && dops[i].rt1 != 0) {
6913 const struct decoded_insn *dop = NULL;
6914 int t = -1;
6915 if (cinfo[i-1].ba != -1) {
6916 t = (cinfo[i-1].ba - start) / 4;
6917 if (t < 0 || t > i) {
6918 u_int limit = 0;
6919 u_int *mem = get_source_start(cinfo[i-1].ba, &limit);
6920 if (mem != NULL) {
6921 disassemble_one(MAXBLOCK - 1, mem[0]);
6922 dop = &dops[MAXBLOCK - 1];
6923 }
6924 }
6925 else
6926 dop = &dops[t];
6927 }
6928 if ((dop && is_ld_use_hazard(dops[i].rt1, dop))
6929 || (!dop && Config.PreciseExceptions)) {
6930 // jump target wants DS result - potential load delay effect
6931 SysPrintf("load delay in DS @%08x (%08x)\n", start + i*4, start);
6932 force_j_to_interpreter = 1;
6933 if (0 <= t && t < i)
6934 dops[t + 1].bt = 1; // expected return from interpreter
6935 }
6936 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&&
6937 !(i>=3&&dops[i-3].is_jump)) {
6938 // v0 overwrite like this is a sign of trouble, bail out
6939 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
6940 force_j_to_interpreter = 1;
6941 }
6942 }
6943 }
6944 else if (i > 0 && dops[i-1].is_delay_load
6945 && is_ld_use_hazard(dops[i-1].rt1, &dops[i])
6946 && (i < 2 || !dops[i-2].is_ujump)) {
6947 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
6948 for (j = i - 1; j > 0 && dops[j-1].is_delay_load; j--)
6949 if (dops[j-1].rt1 != dops[i-1].rt1)
6950 break;
6951 force_j_to_interpreter = 1;
6952 }
6953 if (force_j_to_interpreter) {
6954 force_intcall(j);
6955 done = 2;
6956 i = j; // don't compile the problematic branch/load/etc
6957 }
6958 if (dops[i].is_exception && i > 0 && dops[i-1].is_jump) {
6959 SysPrintf("exception in DS @%08x (%08x)\n", start + i*4, start);
6960 i--;
6961 force_intcall(i);
6962 done = 2;
6963 }
6964 if (i >= 2 && (source[i-2] & 0xffe0f800) == 0x40806000) // MTC0 $12
6965 dops[i].bt = 1;
6966 if (i >= 1 && (source[i-1] & 0xffe0f800) == 0x40806800) // MTC0 $13
6967 dops[i].bt = 1;
6968
6969 /* Is this the end of the block? */
6970 if (i > 0 && dops[i-1].is_ujump) {
6971 if (dops[i-1].rt1 == 0) { // not jal
6972 int found_bbranch = 0, t = (cinfo[i-1].ba - start) / 4;
6973 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
6974 // scan for a branch back to i+1
6975 for (j = t; j < t + 64; j++) {
6976 int tmpop = source[j] >> 26;
6977 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
6978 int t2 = j + 1 + (int)(signed short)source[j];
6979 if (t2 == i + 1) {
6980 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
6981 found_bbranch = 1;
6982 break;
6983 }
6984 }
6985 }
6986 }
6987 if (!found_bbranch)
6988 done = 2;
6989 }
6990 else {
6991 if(stop_after_jal) done=1;
6992 // Stop on BREAK
6993 if((source[i+1]&0xfc00003f)==0x0d) done=1;
6994 }
6995 // Don't recompile stuff that's already compiled
6996 if(check_addr(start+i*4+4)) done=1;
6997 // Don't get too close to the limit
6998 if(i>MAXBLOCK/2) done=1;
6999 }
7000 if (dops[i].itype == HLECALL)
7001 stop = 1;
7002 else if (dops[i].itype == INTCALL)
7003 stop = 2;
7004 else if (dops[i].is_exception)
7005 done = stop_after_jal ? 1 : 2;
7006 if (done == 2) {
7007 // Does the block continue due to a branch?
7008 for(j=i-1;j>=0;j--)
7009 {
7010 if(cinfo[j].ba==start+i*4) done=j=0; // Branch into delay slot
7011 if(cinfo[j].ba==start+i*4+4) done=j=0;
7012 if(cinfo[j].ba==start+i*4+8) done=j=0;
7013 }
7014 }
7015 //assert(i<MAXBLOCK-1);
7016 if(start+i*4==pagelimit-4) done=1;
7017 assert(start+i*4<pagelimit);
7018 if (i==MAXBLOCK-1) done=1;
7019 // Stop if we're compiling junk
7020 if (dops[i].itype == INTCALL && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7021 done=stop_after_jal=1;
7022 SysPrintf("Disabled speculative precompilation\n");
7023 }
7024 }
7025 while (i > 0 && dops[i-1].is_jump)
7026 i--;
7027 assert(i > 0);
7028 assert(!dops[i-1].is_jump);
7029 slen = i;
7030}
7031
7032// Basic liveness analysis for MIPS registers
7033static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7034{
7035 int i;
7036 uint64_t u,gte_u,b,gte_b;
7037 uint64_t temp_u,temp_gte_u=0;
7038 uint64_t gte_u_unknown=0;
7039 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7040 gte_u_unknown=~0ll;
7041 if(iend==slen-1) {
7042 u=1;
7043 gte_u=gte_u_unknown;
7044 }else{
7045 //u=unneeded_reg[iend+1];
7046 u=1;
7047 gte_u=gte_unneeded[iend+1];
7048 }
7049
7050 for (i=iend;i>=istart;i--)
7051 {
7052 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7053 if(dops[i].is_jump)
7054 {
7055 // If subroutine call, flag return address as a possible branch target
7056 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7057
7058 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7059 {
7060 // Branch out of this block, flush all regs
7061 u=1;
7062 gte_u=gte_u_unknown;
7063 branch_unneeded_reg[i]=u;
7064 // Merge in delay slot
7065 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7066 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7067 u|=1;
7068 gte_u|=gte_rt[i+1];
7069 gte_u&=~gte_rs[i+1];
7070 }
7071 else
7072 {
7073 // Internal branch, flag target
7074 dops[(cinfo[i].ba-start)>>2].bt=1;
7075 if(cinfo[i].ba<=start+i*4) {
7076 // Backward branch
7077 if(dops[i].is_ujump)
7078 {
7079 // Unconditional branch
7080 temp_u=1;
7081 temp_gte_u=0;
7082 } else {
7083 // Conditional branch (not taken case)
7084 temp_u=unneeded_reg[i+2];
7085 temp_gte_u&=gte_unneeded[i+2];
7086 }
7087 // Merge in delay slot
7088 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7089 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7090 temp_u|=1;
7091 temp_gte_u|=gte_rt[i+1];
7092 temp_gte_u&=~gte_rs[i+1];
7093 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7094 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7095 temp_u|=1;
7096 temp_gte_u|=gte_rt[i];
7097 temp_gte_u&=~gte_rs[i];
7098 unneeded_reg[i]=temp_u;
7099 gte_unneeded[i]=temp_gte_u;
7100 // Only go three levels deep. This recursion can take an
7101 // excessive amount of time if there are a lot of nested loops.
7102 if(r<2) {
7103 pass2_unneeded_regs((cinfo[i].ba-start)>>2,i-1,r+1);
7104 }else{
7105 unneeded_reg[(cinfo[i].ba-start)>>2]=1;
7106 gte_unneeded[(cinfo[i].ba-start)>>2]=gte_u_unknown;
7107 }
7108 } /*else*/ if(1) {
7109 if (dops[i].is_ujump)
7110 {
7111 // Unconditional branch
7112 u=unneeded_reg[(cinfo[i].ba-start)>>2];
7113 gte_u=gte_unneeded[(cinfo[i].ba-start)>>2];
7114 branch_unneeded_reg[i]=u;
7115 // Merge in delay slot
7116 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7117 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7118 u|=1;
7119 gte_u|=gte_rt[i+1];
7120 gte_u&=~gte_rs[i+1];
7121 } else {
7122 // Conditional branch
7123 b=unneeded_reg[(cinfo[i].ba-start)>>2];
7124 gte_b=gte_unneeded[(cinfo[i].ba-start)>>2];
7125 branch_unneeded_reg[i]=b;
7126 // Branch delay slot
7127 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7128 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7129 b|=1;
7130 gte_b|=gte_rt[i+1];
7131 gte_b&=~gte_rs[i+1];
7132 u&=b;
7133 gte_u&=gte_b;
7134 if(i<slen-1) {
7135 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7136 } else {
7137 branch_unneeded_reg[i]=1;
7138 }
7139 }
7140 }
7141 }
7142 }
7143 //u=1; // DEBUG
7144 // Written registers are unneeded
7145 u|=1LL<<dops[i].rt1;
7146 u|=1LL<<dops[i].rt2;
7147 gte_u|=gte_rt[i];
7148 // Accessed registers are needed
7149 u&=~(1LL<<dops[i].rs1);
7150 u&=~(1LL<<dops[i].rs2);
7151 gte_u&=~gte_rs[i];
7152 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7153 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7154 if (dops[i].may_except || dops[i].itype == RFE)
7155 {
7156 // SYSCALL instruction, etc or conditional exception
7157 u=1;
7158 }
7159 // Source-target dependencies
7160 // R0 is always unneeded
7161 u|=1;
7162 // Save it
7163 unneeded_reg[i]=u;
7164 gte_unneeded[i]=gte_u;
7165 /*
7166 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7167 printf("U:");
7168 int r;
7169 for(r=1;r<=CCREG;r++) {
7170 if((unneeded_reg[i]>>r)&1) {
7171 if(r==HIREG) printf(" HI");
7172 else if(r==LOREG) printf(" LO");
7173 else printf(" r%d",r);
7174 }
7175 }
7176 printf("\n");
7177 */
7178 }
7179}
7180
7181static noinline void pass3_register_alloc(u_int addr)
7182{
7183 struct regstat current; // Current register allocations/status
7184 clear_all_regs(current.regmap_entry);
7185 clear_all_regs(current.regmap);
7186 current.wasdirty = current.dirty = 0;
7187 current.u = unneeded_reg[0];
7188 alloc_reg(&current, 0, CCREG);
7189 dirty_reg(&current, CCREG);
7190 current.wasconst = 0;
7191 current.isconst = 0;
7192 current.loadedconst = 0;
7193 current.noevict = 0;
7194 //current.waswritten = 0;
7195 int ds=0;
7196 int cc=0;
7197 int hr;
7198 int i, j;
7199
7200 if (addr & 1) {
7201 // First instruction is delay slot
7202 cc=-1;
7203 dops[1].bt=1;
7204 ds=1;
7205 unneeded_reg[0]=1;
7206 current.regmap[HOST_BTREG]=BTREG;
7207 }
7208
7209 for(i=0;i<slen;i++)
7210 {
7211 if(dops[i].bt)
7212 {
7213 for(hr=0;hr<HOST_REGS;hr++)
7214 {
7215 // Is this really necessary?
7216 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7217 }
7218 current.isconst=0;
7219 //current.waswritten=0;
7220 }
7221
7222 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7223 regs[i].wasconst=current.isconst;
7224 regs[i].wasdirty=current.dirty;
7225 regs[i].dirty=0;
7226 regs[i].u=0;
7227 regs[i].isconst=0;
7228 regs[i].loadedconst=0;
7229 if (!dops[i].is_jump) {
7230 if(i+1<slen) {
7231 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7232 current.u|=1;
7233 } else {
7234 current.u=1;
7235 }
7236 } else {
7237 if(i+1<slen) {
7238 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7239 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7240 current.u|=1;
7241 } else {
7242 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7243 abort();
7244 }
7245 }
7246 dops[i].is_ds=ds;
7247 if(ds) {
7248 ds=0; // Skip delay slot, already allocated as part of branch
7249 // ...but we need to alloc it in case something jumps here
7250 if(i+1<slen) {
7251 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7252 }else{
7253 current.u=branch_unneeded_reg[i-1];
7254 }
7255 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7256 current.u|=1;
7257 struct regstat temp;
7258 memcpy(&temp,&current,sizeof(current));
7259 temp.wasdirty=temp.dirty;
7260 // TODO: Take into account unconditional branches, as below
7261 delayslot_alloc(&temp,i);
7262 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7263 regs[i].wasdirty=temp.wasdirty;
7264 regs[i].dirty=temp.dirty;
7265 regs[i].isconst=0;
7266 regs[i].wasconst=0;
7267 current.isconst=0;
7268 // Create entry (branch target) regmap
7269 for(hr=0;hr<HOST_REGS;hr++)
7270 {
7271 int r=temp.regmap[hr];
7272 if(r>=0) {
7273 if(r!=regmap_pre[i][hr]) {
7274 regs[i].regmap_entry[hr]=-1;
7275 }
7276 else
7277 {
7278 assert(r < 64);
7279 if((current.u>>r)&1) {
7280 regs[i].regmap_entry[hr]=-1;
7281 regs[i].regmap[hr]=-1;
7282 //Don't clear regs in the delay slot as the branch might need them
7283 //current.regmap[hr]=-1;
7284 }else
7285 regs[i].regmap_entry[hr]=r;
7286 }
7287 } else {
7288 // First instruction expects CCREG to be allocated
7289 if(i==0&&hr==HOST_CCREG)
7290 regs[i].regmap_entry[hr]=CCREG;
7291 else
7292 regs[i].regmap_entry[hr]=-1;
7293 }
7294 }
7295 }
7296 else { // Not delay slot
7297 current.noevict = 0;
7298 switch(dops[i].itype) {
7299 case UJUMP:
7300 //current.isconst=0; // DEBUG
7301 //current.wasconst=0; // DEBUG
7302 //regs[i].wasconst=0; // DEBUG
7303 clear_const(&current,dops[i].rt1);
7304 alloc_cc(&current,i);
7305 dirty_reg(&current,CCREG);
7306 if (dops[i].rt1==31) {
7307 alloc_reg(&current,i,31);
7308 dirty_reg(&current,31);
7309 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7310 //assert(dops[i+1].rt1!=dops[i].rt1);
7311 #ifdef REG_PREFETCH
7312 alloc_reg(&current,i,PTEMP);
7313 #endif
7314 }
7315 dops[i].ooo=1;
7316 delayslot_alloc(&current,i+1);
7317 //current.isconst=0; // DEBUG
7318 ds=1;
7319 break;
7320 case RJUMP:
7321 //current.isconst=0;
7322 //current.wasconst=0;
7323 //regs[i].wasconst=0;
7324 clear_const(&current,dops[i].rs1);
7325 clear_const(&current,dops[i].rt1);
7326 alloc_cc(&current,i);
7327 dirty_reg(&current,CCREG);
7328 if (!ds_writes_rjump_rs(i)) {
7329 alloc_reg(&current,i,dops[i].rs1);
7330 if (dops[i].rt1!=0) {
7331 alloc_reg(&current,i,dops[i].rt1);
7332 dirty_reg(&current,dops[i].rt1);
7333 #ifdef REG_PREFETCH
7334 alloc_reg(&current,i,PTEMP);
7335 #endif
7336 }
7337 #ifdef USE_MINI_HT
7338 if(dops[i].rs1==31) { // JALR
7339 alloc_reg(&current,i,RHASH);
7340 alloc_reg(&current,i,RHTBL);
7341 }
7342 #endif
7343 delayslot_alloc(&current,i+1);
7344 } else {
7345 // The delay slot overwrites our source register,
7346 // allocate a temporary register to hold the old value.
7347 current.isconst=0;
7348 current.wasconst=0;
7349 regs[i].wasconst=0;
7350 delayslot_alloc(&current,i+1);
7351 current.isconst=0;
7352 alloc_reg(&current,i,RTEMP);
7353 }
7354 //current.isconst=0; // DEBUG
7355 dops[i].ooo=1;
7356 ds=1;
7357 break;
7358 case CJUMP:
7359 //current.isconst=0;
7360 //current.wasconst=0;
7361 //regs[i].wasconst=0;
7362 clear_const(&current,dops[i].rs1);
7363 clear_const(&current,dops[i].rs2);
7364 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7365 {
7366 alloc_cc(&current,i);
7367 dirty_reg(&current,CCREG);
7368 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7369 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7370 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7371 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7372 // The delay slot overwrites one of our conditions.
7373 // Allocate the branch condition registers instead.
7374 current.isconst=0;
7375 current.wasconst=0;
7376 regs[i].wasconst=0;
7377 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7378 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7379 }
7380 else
7381 {
7382 dops[i].ooo=1;
7383 delayslot_alloc(&current,i+1);
7384 }
7385 }
7386 else
7387 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7388 {
7389 alloc_cc(&current,i);
7390 dirty_reg(&current,CCREG);
7391 alloc_reg(&current,i,dops[i].rs1);
7392 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7393 // The delay slot overwrites one of our conditions.
7394 // Allocate the branch condition registers instead.
7395 current.isconst=0;
7396 current.wasconst=0;
7397 regs[i].wasconst=0;
7398 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7399 }
7400 else
7401 {
7402 dops[i].ooo=1;
7403 delayslot_alloc(&current,i+1);
7404 }
7405 }
7406 else
7407 // Don't alloc the delay slot yet because we might not execute it
7408 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7409 {
7410 current.isconst=0;
7411 current.wasconst=0;
7412 regs[i].wasconst=0;
7413 alloc_cc(&current,i);
7414 dirty_reg(&current,CCREG);
7415 alloc_reg(&current,i,dops[i].rs1);
7416 alloc_reg(&current,i,dops[i].rs2);
7417 }
7418 else
7419 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7420 {
7421 current.isconst=0;
7422 current.wasconst=0;
7423 regs[i].wasconst=0;
7424 alloc_cc(&current,i);
7425 dirty_reg(&current,CCREG);
7426 alloc_reg(&current,i,dops[i].rs1);
7427 }
7428 ds=1;
7429 //current.isconst=0;
7430 break;
7431 case SJUMP:
7432 clear_const(&current,dops[i].rs1);
7433 clear_const(&current,dops[i].rt1);
7434 {
7435 alloc_cc(&current,i);
7436 dirty_reg(&current,CCREG);
7437 alloc_reg(&current,i,dops[i].rs1);
7438 if (dops[i].rt1 == 31) { // BLTZAL/BGEZAL
7439 alloc_reg(&current,i,31);
7440 dirty_reg(&current,31);
7441 }
7442 if ((dops[i].rs1 &&
7443 (dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7444 ||(dops[i].rt1 == 31 && dops[i].rs1 == 31) // overwrites it's own condition
7445 ||(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
7446 // Allocate the branch condition registers instead.
7447 current.isconst=0;
7448 current.wasconst=0;
7449 regs[i].wasconst=0;
7450 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7451 }
7452 else
7453 {
7454 dops[i].ooo=1;
7455 delayslot_alloc(&current,i+1);
7456 }
7457 }
7458 ds=1;
7459 //current.isconst=0;
7460 break;
7461 case IMM16:
7462 imm16_alloc(&current,i);
7463 break;
7464 case LOAD:
7465 case LOADLR:
7466 load_alloc(&current,i);
7467 break;
7468 case STORE:
7469 case STORELR:
7470 store_alloc(&current,i);
7471 break;
7472 case ALU:
7473 alu_alloc(&current,i);
7474 break;
7475 case SHIFT:
7476 shift_alloc(&current,i);
7477 break;
7478 case MULTDIV:
7479 multdiv_alloc(&current,i);
7480 break;
7481 case SHIFTIMM:
7482 shiftimm_alloc(&current,i);
7483 break;
7484 case MOV:
7485 mov_alloc(&current,i);
7486 break;
7487 case COP0:
7488 cop0_alloc(&current,i);
7489 break;
7490 case RFE:
7491 rfe_alloc(&current,i);
7492 break;
7493 case COP2:
7494 cop2_alloc(&current,i);
7495 break;
7496 case C2LS:
7497 c2ls_alloc(&current,i);
7498 break;
7499 case C2OP:
7500 c2op_alloc(&current,i);
7501 break;
7502 case SYSCALL:
7503 case HLECALL:
7504 case INTCALL:
7505 syscall_alloc(&current,i);
7506 break;
7507 }
7508
7509 // Create entry (branch target) regmap
7510 for(hr=0;hr<HOST_REGS;hr++)
7511 {
7512 int r,or;
7513 r=current.regmap[hr];
7514 if(r>=0) {
7515 if(r!=regmap_pre[i][hr]) {
7516 // TODO: delay slot (?)
7517 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7518 if(or<0||r>=TEMPREG){
7519 regs[i].regmap_entry[hr]=-1;
7520 }
7521 else
7522 {
7523 // Just move it to a different register
7524 regs[i].regmap_entry[hr]=r;
7525 // If it was dirty before, it's still dirty
7526 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7527 }
7528 }
7529 else
7530 {
7531 // Unneeded
7532 if(r==0){
7533 regs[i].regmap_entry[hr]=0;
7534 }
7535 else
7536 {
7537 assert(r<64);
7538 if((current.u>>r)&1) {
7539 regs[i].regmap_entry[hr]=-1;
7540 //regs[i].regmap[hr]=-1;
7541 current.regmap[hr]=-1;
7542 }else
7543 regs[i].regmap_entry[hr]=r;
7544 }
7545 }
7546 } else {
7547 // Branches expect CCREG to be allocated at the target
7548 if(regmap_pre[i][hr]==CCREG)
7549 regs[i].regmap_entry[hr]=CCREG;
7550 else
7551 regs[i].regmap_entry[hr]=-1;
7552 }
7553 }
7554 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7555 }
7556
7557#if 0 // see do_store_smc_check()
7558 if(i>0&&(dops[i-1].itype==STORE||dops[i-1].itype==STORELR||(dops[i-1].itype==C2LS&&dops[i-1].opcode==0x3a))&&(u_int)cinfo[i-1].imm<0x800)
7559 current.waswritten|=1<<dops[i-1].rs1;
7560 current.waswritten&=~(1<<dops[i].rt1);
7561 current.waswritten&=~(1<<dops[i].rt2);
7562 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)cinfo[i].imm>=0x800)
7563 current.waswritten&=~(1<<dops[i].rs1);
7564#endif
7565
7566 /* Branch post-alloc */
7567 if(i>0)
7568 {
7569 current.wasdirty=current.dirty;
7570 switch(dops[i-1].itype) {
7571 case UJUMP:
7572 memcpy(&branch_regs[i-1],&current,sizeof(current));
7573 branch_regs[i-1].isconst=0;
7574 branch_regs[i-1].wasconst=0;
7575 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7576 alloc_cc(&branch_regs[i-1],i-1);
7577 dirty_reg(&branch_regs[i-1],CCREG);
7578 if(dops[i-1].rt1==31) { // JAL
7579 alloc_reg(&branch_regs[i-1],i-1,31);
7580 dirty_reg(&branch_regs[i-1],31);
7581 }
7582 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7583 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7584 break;
7585 case RJUMP:
7586 memcpy(&branch_regs[i-1],&current,sizeof(current));
7587 branch_regs[i-1].isconst=0;
7588 branch_regs[i-1].wasconst=0;
7589 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7590 alloc_cc(&branch_regs[i-1],i-1);
7591 dirty_reg(&branch_regs[i-1],CCREG);
7592 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7593 if(dops[i-1].rt1!=0) { // JALR
7594 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7595 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7596 }
7597 #ifdef USE_MINI_HT
7598 if(dops[i-1].rs1==31) { // JALR
7599 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7600 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7601 }
7602 #endif
7603 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7604 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7605 break;
7606 case CJUMP:
7607 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7608 {
7609 alloc_cc(&current,i-1);
7610 dirty_reg(&current,CCREG);
7611 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7612 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7613 // The delay slot overwrote one of our conditions
7614 // Delay slot goes after the test (in order)
7615 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7616 current.u|=1;
7617 delayslot_alloc(&current,i);
7618 current.isconst=0;
7619 }
7620 else
7621 {
7622 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7623 // Alloc the branch condition registers
7624 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7625 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7626 }
7627 memcpy(&branch_regs[i-1],&current,sizeof(current));
7628 branch_regs[i-1].isconst=0;
7629 branch_regs[i-1].wasconst=0;
7630 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7631 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7632 }
7633 else
7634 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7635 {
7636 alloc_cc(&current,i-1);
7637 dirty_reg(&current,CCREG);
7638 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7639 // The delay slot overwrote the branch condition
7640 // Delay slot goes after the test (in order)
7641 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7642 current.u|=1;
7643 delayslot_alloc(&current,i);
7644 current.isconst=0;
7645 }
7646 else
7647 {
7648 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7649 // Alloc the branch condition register
7650 alloc_reg(&current,i-1,dops[i-1].rs1);
7651 }
7652 memcpy(&branch_regs[i-1],&current,sizeof(current));
7653 branch_regs[i-1].isconst=0;
7654 branch_regs[i-1].wasconst=0;
7655 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7656 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7657 }
7658 break;
7659 case SJUMP:
7660 {
7661 alloc_cc(&current,i-1);
7662 dirty_reg(&current,CCREG);
7663 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7664 // The delay slot overwrote the branch condition
7665 // Delay slot goes after the test (in order)
7666 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7667 current.u|=1;
7668 delayslot_alloc(&current,i);
7669 current.isconst=0;
7670 }
7671 else
7672 {
7673 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7674 // Alloc the branch condition register
7675 alloc_reg(&current,i-1,dops[i-1].rs1);
7676 }
7677 memcpy(&branch_regs[i-1],&current,sizeof(current));
7678 branch_regs[i-1].isconst=0;
7679 branch_regs[i-1].wasconst=0;
7680 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7681 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7682 }
7683 break;
7684 }
7685
7686 if (dops[i-1].is_ujump)
7687 {
7688 if(dops[i-1].rt1==31) // JAL/JALR
7689 {
7690 // Subroutine call will return here, don't alloc any registers
7691 current.dirty=0;
7692 clear_all_regs(current.regmap);
7693 alloc_reg(&current,i,CCREG);
7694 dirty_reg(&current,CCREG);
7695 }
7696 else if(i+1<slen)
7697 {
7698 // Internal branch will jump here, match registers to caller
7699 current.dirty=0;
7700 clear_all_regs(current.regmap);
7701 alloc_reg(&current,i,CCREG);
7702 dirty_reg(&current,CCREG);
7703 for(j=i-1;j>=0;j--)
7704 {
7705 if(cinfo[j].ba==start+i*4+4) {
7706 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7707 current.dirty=branch_regs[j].dirty;
7708 break;
7709 }
7710 }
7711 while(j>=0) {
7712 if(cinfo[j].ba==start+i*4+4) {
7713 for(hr=0;hr<HOST_REGS;hr++) {
7714 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7715 current.regmap[hr]=-1;
7716 }
7717 current.dirty&=branch_regs[j].dirty;
7718 }
7719 }
7720 j--;
7721 }
7722 }
7723 }
7724 }
7725
7726 // Count cycles in between branches
7727 cinfo[i].ccadj = CLOCK_ADJUST(cc);
7728 if (i > 0 && (dops[i-1].is_jump || dops[i].is_exception))
7729 {
7730 cc=0;
7731 }
7732#if !defined(DRC_DBG)
7733 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7734 {
7735 // this should really be removed since the real stalls have been implemented,
7736 // but doing so causes sizeable perf regression against the older version
7737 u_int gtec = gte_cycletab[source[i] & 0x3f];
7738 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7739 }
7740 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7741 {
7742 cc+=4;
7743 }
7744 else if(dops[i].itype==C2LS)
7745 {
7746 // same as with C2OP
7747 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7748 }
7749#endif
7750 else
7751 {
7752 cc++;
7753 }
7754
7755 if(!dops[i].is_ds) {
7756 regs[i].dirty=current.dirty;
7757 regs[i].isconst=current.isconst;
7758 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7759 }
7760 for(hr=0;hr<HOST_REGS;hr++) {
7761 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7762 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7763 regs[i].wasconst&=~(1<<hr);
7764 }
7765 }
7766 }
7767 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7768 //regs[i].waswritten=current.waswritten;
7769 }
7770}
7771
7772static noinline void pass4_cull_unused_regs(void)
7773{
7774 u_int last_needed_regs[4] = {0,0,0,0};
7775 u_int nr=0;
7776 int i;
7777
7778 for (i=slen-1;i>=0;i--)
7779 {
7780 int hr;
7781 __builtin_prefetch(regs[i-2].regmap);
7782 if(dops[i].is_jump)
7783 {
7784 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7785 {
7786 // Branch out of this block, don't need anything
7787 nr=0;
7788 }
7789 else
7790 {
7791 // Internal branch
7792 // Need whatever matches the target
7793 nr=0;
7794 int t=(cinfo[i].ba-start)>>2;
7795 for(hr=0;hr<HOST_REGS;hr++)
7796 {
7797 if(regs[i].regmap_entry[hr]>=0) {
7798 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7799 }
7800 }
7801 }
7802 // Conditional branch may need registers for following instructions
7803 if (!dops[i].is_ujump)
7804 {
7805 if(i<slen-2) {
7806 nr |= last_needed_regs[(i+2) & 3];
7807 for(hr=0;hr<HOST_REGS;hr++)
7808 {
7809 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7810 //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]);
7811 }
7812 }
7813 }
7814 // Don't need stuff which is overwritten
7815 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7816 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7817 // Merge in delay slot
7818 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
7819 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
7820 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
7821 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
7822 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
7823 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
7824 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
7825 nr |= get_regm(regmap_pre[i], ROREG);
7826 nr |= get_regm(regs[i].regmap_entry, ROREG);
7827 }
7828 if (dops[i+1].is_store) {
7829 nr |= get_regm(regmap_pre[i], INVCP);
7830 nr |= get_regm(regs[i].regmap_entry, INVCP);
7831 }
7832 }
7833 else if (dops[i].is_exception)
7834 {
7835 // SYSCALL instruction, etc
7836 nr=0;
7837 }
7838 else // Non-branch
7839 {
7840 if(i<slen-1) {
7841 for(hr=0;hr<HOST_REGS;hr++) {
7842 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7843 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7844 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7845 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7846 }
7847 }
7848 }
7849 // Overwritten registers are not needed
7850 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
7851 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
7852 nr &= ~get_regm(regs[i].regmap, FTEMP);
7853 // Source registers are needed
7854 nr |= get_regm(regmap_pre[i], dops[i].rs1);
7855 nr |= get_regm(regmap_pre[i], dops[i].rs2);
7856 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
7857 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
7858 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
7859 nr |= get_regm(regmap_pre[i], ROREG);
7860 nr |= get_regm(regs[i].regmap_entry, ROREG);
7861 }
7862 if (dops[i].is_store) {
7863 nr |= get_regm(regmap_pre[i], INVCP);
7864 nr |= get_regm(regs[i].regmap_entry, INVCP);
7865 }
7866
7867 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
7868 for(hr=0;hr<HOST_REGS;hr++)
7869 {
7870 // Don't store a register immediately after writing it,
7871 // may prevent dual-issue.
7872 // But do so if this is a branch target, otherwise we
7873 // might have to load the register before the branch.
7874 if((regs[i].wasdirty>>hr)&1) {
7875 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7876 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
7877 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
7878 }
7879 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7880 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
7881 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
7882 }
7883 }
7884 }
7885 // Cycle count is needed at branches. Assume it is needed at the target too.
7886 if (i == 0 || dops[i].bt || dops[i].may_except || dops[i].itype == CJUMP) {
7887 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7888 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7889 }
7890 // Save it
7891 last_needed_regs[i & 3] = nr;
7892
7893 // Deallocate unneeded registers
7894 for(hr=0;hr<HOST_REGS;hr++)
7895 {
7896 if(!((nr>>hr)&1)) {
7897 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
7898 if(dops[i].is_jump)
7899 {
7900 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
7901 if (dops[i+1].is_load || dops[i+1].is_store)
7902 map1 = ROREG;
7903 if (dops[i+1].is_store)
7904 map2 = INVCP;
7905 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
7906 temp = FTEMP;
7907 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
7908 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
7909 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
7910 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
7911 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
7912 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
7913 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
7914 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
7915 {
7916 regs[i].regmap[hr]=-1;
7917 regs[i].isconst&=~(1<<hr);
7918 regs[i].dirty&=~(1<<hr);
7919 regs[i+1].wasdirty&=~(1<<hr);
7920 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
7921 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
7922 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
7923 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
7924 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
7925 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
7926 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
7927 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
7928 {
7929 branch_regs[i].regmap[hr]=-1;
7930 branch_regs[i].regmap_entry[hr]=-1;
7931 if (!dops[i].is_ujump)
7932 {
7933 if (i < slen-2) {
7934 regmap_pre[i+2][hr]=-1;
7935 regs[i+2].wasconst&=~(1<<hr);
7936 }
7937 }
7938 }
7939 }
7940 }
7941 else
7942 {
7943 // Non-branch
7944 if(i>0)
7945 {
7946 int map1 = -1, map2 = -1, temp=-1;
7947 if (dops[i].is_load || dops[i].is_store)
7948 map1 = ROREG;
7949 if (dops[i].is_store)
7950 map2 = INVCP;
7951 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
7952 temp = FTEMP;
7953 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
7954 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
7955 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
7956 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
7957 regs[i].regmap[hr] != CCREG)
7958 {
7959 if(i<slen-1&&!dops[i].is_ds) {
7960 assert(regs[i].regmap[hr]<64);
7961 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
7962 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
7963 {
7964 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
7965 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
7966 }
7967 regmap_pre[i+1][hr]=-1;
7968 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
7969 regs[i+1].wasconst&=~(1<<hr);
7970 }
7971 regs[i].regmap[hr]=-1;
7972 regs[i].isconst&=~(1<<hr);
7973 regs[i].dirty&=~(1<<hr);
7974 regs[i+1].wasdirty&=~(1<<hr);
7975 }
7976 }
7977 }
7978 } // if needed
7979 } // for hr
7980 }
7981}
7982
7983// If a register is allocated during a loop, try to allocate it for the
7984// entire loop, if possible. This avoids loading/storing registers
7985// inside of the loop.
7986static noinline void pass5a_preallocate1(void)
7987{
7988 int i, j, hr;
7989 signed char f_regmap[HOST_REGS];
7990 clear_all_regs(f_regmap);
7991 for(i=0;i<slen-1;i++)
7992 {
7993 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
7994 {
7995 if(cinfo[i].ba>=start && cinfo[i].ba<(start+i*4))
7996 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
7997 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
7998 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR
7999 ||dops[i+1].itype==SHIFT
8000 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8001 {
8002 int t=(cinfo[i].ba-start)>>2;
8003 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8004 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8005 for(hr=0;hr<HOST_REGS;hr++)
8006 {
8007 if(regs[i].regmap[hr]>=0) {
8008 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8009 // dealloc old register
8010 int n;
8011 for(n=0;n<HOST_REGS;n++)
8012 {
8013 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8014 }
8015 // and alloc new one
8016 f_regmap[hr]=regs[i].regmap[hr];
8017 }
8018 }
8019 if(branch_regs[i].regmap[hr]>=0) {
8020 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8021 // dealloc old register
8022 int n;
8023 for(n=0;n<HOST_REGS;n++)
8024 {
8025 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8026 }
8027 // and alloc new one
8028 f_regmap[hr]=branch_regs[i].regmap[hr];
8029 }
8030 }
8031 if(dops[i].ooo) {
8032 if(count_free_regs(regs[i].regmap)<=cinfo[i+1].min_free_regs)
8033 f_regmap[hr]=branch_regs[i].regmap[hr];
8034 }else{
8035 if(count_free_regs(branch_regs[i].regmap)<=cinfo[i+1].min_free_regs)
8036 f_regmap[hr]=branch_regs[i].regmap[hr];
8037 }
8038 // Avoid dirty->clean transition
8039 #ifdef DESTRUCTIVE_WRITEBACK
8040 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;
8041 #endif
8042 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8043 // case above, however it's always a good idea. We can't hoist the
8044 // load if the register was already allocated, so there's no point
8045 // wasting time analyzing most of these cases. It only "succeeds"
8046 // when the mapping was different and the load can be replaced with
8047 // a mov, which is of negligible benefit. So such cases are
8048 // skipped below.
8049 if(f_regmap[hr]>0) {
8050 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8051 int r=f_regmap[hr];
8052 for(j=t;j<=i;j++)
8053 {
8054 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8055 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8056 assert(r < 64);
8057 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8058 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8059 int k;
8060 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8061 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8062 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8063 k=i;
8064 while(k>1&&regs[k-1].regmap[hr]==-1) {
8065 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8066 //printf("no free regs for store %x\n",start+(k-1)*4);
8067 break;
8068 }
8069 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8070 //printf("no-match due to different register\n");
8071 break;
8072 }
8073 if (dops[k-2].is_jump) {
8074 //printf("no-match due to branch\n");
8075 break;
8076 }
8077 // call/ret fast path assumes no registers allocated
8078 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8079 break;
8080 }
8081 k--;
8082 }
8083 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8084 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8085 while(k<i) {
8086 regs[k].regmap_entry[hr]=f_regmap[hr];
8087 regs[k].regmap[hr]=f_regmap[hr];
8088 regmap_pre[k+1][hr]=f_regmap[hr];
8089 regs[k].wasdirty&=~(1<<hr);
8090 regs[k].dirty&=~(1<<hr);
8091 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8092 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8093 regs[k].wasconst&=~(1<<hr);
8094 regs[k].isconst&=~(1<<hr);
8095 k++;
8096 }
8097 }
8098 else {
8099 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8100 break;
8101 }
8102 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8103 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8104 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8105 regs[i].regmap_entry[hr]=f_regmap[hr];
8106 regs[i].regmap[hr]=f_regmap[hr];
8107 regs[i].wasdirty&=~(1<<hr);
8108 regs[i].dirty&=~(1<<hr);
8109 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8110 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8111 regs[i].wasconst&=~(1<<hr);
8112 regs[i].isconst&=~(1<<hr);
8113 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8114 branch_regs[i].wasdirty&=~(1<<hr);
8115 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8116 branch_regs[i].regmap[hr]=f_regmap[hr];
8117 branch_regs[i].dirty&=~(1<<hr);
8118 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8119 branch_regs[i].wasconst&=~(1<<hr);
8120 branch_regs[i].isconst&=~(1<<hr);
8121 if (!dops[i].is_ujump) {
8122 regmap_pre[i+2][hr]=f_regmap[hr];
8123 regs[i+2].wasdirty&=~(1<<hr);
8124 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8125 }
8126 }
8127 }
8128 for(k=t;k<j;k++) {
8129 // Alloc register clean at beginning of loop,
8130 // but may dirty it in pass 6
8131 regs[k].regmap_entry[hr]=f_regmap[hr];
8132 regs[k].regmap[hr]=f_regmap[hr];
8133 regs[k].dirty&=~(1<<hr);
8134 regs[k].wasconst&=~(1<<hr);
8135 regs[k].isconst&=~(1<<hr);
8136 if (dops[k].is_jump) {
8137 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8138 branch_regs[k].regmap[hr]=f_regmap[hr];
8139 branch_regs[k].dirty&=~(1<<hr);
8140 branch_regs[k].wasconst&=~(1<<hr);
8141 branch_regs[k].isconst&=~(1<<hr);
8142 if (!dops[k].is_ujump) {
8143 regmap_pre[k+2][hr]=f_regmap[hr];
8144 regs[k+2].wasdirty&=~(1<<hr);
8145 }
8146 }
8147 else
8148 {
8149 regmap_pre[k+1][hr]=f_regmap[hr];
8150 regs[k+1].wasdirty&=~(1<<hr);
8151 }
8152 }
8153 if(regs[j].regmap[hr]==f_regmap[hr])
8154 regs[j].regmap_entry[hr]=f_regmap[hr];
8155 break;
8156 }
8157 if(j==i) break;
8158 if(regs[j].regmap[hr]>=0)
8159 break;
8160 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8161 //printf("no-match due to different register\n");
8162 break;
8163 }
8164 if (dops[j].is_ujump)
8165 {
8166 // Stop on unconditional branch
8167 break;
8168 }
8169 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8170 {
8171 if(dops[j].ooo) {
8172 if(count_free_regs(regs[j].regmap)<=cinfo[j+1].min_free_regs)
8173 break;
8174 }else{
8175 if(count_free_regs(branch_regs[j].regmap)<=cinfo[j+1].min_free_regs)
8176 break;
8177 }
8178 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8179 //printf("no-match due to different register (branch)\n");
8180 break;
8181 }
8182 }
8183 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8184 //printf("No free regs for store %x\n",start+j*4);
8185 break;
8186 }
8187 assert(f_regmap[hr]<64);
8188 }
8189 }
8190 }
8191 }
8192 }
8193 }else{
8194 // Non branch or undetermined branch target
8195 for(hr=0;hr<HOST_REGS;hr++)
8196 {
8197 if(hr!=EXCLUDE_REG) {
8198 if(regs[i].regmap[hr]>=0) {
8199 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8200 // dealloc old register
8201 int n;
8202 for(n=0;n<HOST_REGS;n++)
8203 {
8204 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8205 }
8206 // and alloc new one
8207 f_regmap[hr]=regs[i].regmap[hr];
8208 }
8209 }
8210 }
8211 }
8212 // Try to restore cycle count at branch targets
8213 if(dops[i].bt) {
8214 for(j=i;j<slen-1;j++) {
8215 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8216 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8217 //printf("no free regs for store %x\n",start+j*4);
8218 break;
8219 }
8220 }
8221 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8222 int k=i;
8223 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8224 while(k<j) {
8225 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8226 regs[k].regmap[HOST_CCREG]=CCREG;
8227 regmap_pre[k+1][HOST_CCREG]=CCREG;
8228 regs[k+1].wasdirty|=1<<HOST_CCREG;
8229 regs[k].dirty|=1<<HOST_CCREG;
8230 regs[k].wasconst&=~(1<<HOST_CCREG);
8231 regs[k].isconst&=~(1<<HOST_CCREG);
8232 k++;
8233 }
8234 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8235 }
8236 // Work backwards from the branch target
8237 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8238 {
8239 //printf("Extend backwards\n");
8240 int k;
8241 k=i;
8242 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8243 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8244 //printf("no free regs for store %x\n",start+(k-1)*4);
8245 break;
8246 }
8247 k--;
8248 }
8249 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8250 //printf("Extend CC, %x ->\n",start+k*4);
8251 while(k<=i) {
8252 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8253 regs[k].regmap[HOST_CCREG]=CCREG;
8254 regmap_pre[k+1][HOST_CCREG]=CCREG;
8255 regs[k+1].wasdirty|=1<<HOST_CCREG;
8256 regs[k].dirty|=1<<HOST_CCREG;
8257 regs[k].wasconst&=~(1<<HOST_CCREG);
8258 regs[k].isconst&=~(1<<HOST_CCREG);
8259 k++;
8260 }
8261 }
8262 else {
8263 //printf("Fail Extend CC, %x ->\n",start+k*4);
8264 }
8265 }
8266 }
8267 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=SHIFT&&
8268 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8269 dops[i].itype!=IMM16&&dops[i].itype!=LOAD)
8270 {
8271 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8272 }
8273 }
8274 }
8275}
8276
8277// This allocates registers (if possible) one instruction prior
8278// to use, which can avoid a load-use penalty on certain CPUs.
8279static noinline void pass5b_preallocate2(void)
8280{
8281 int i, hr;
8282 for(i=0;i<slen-1;i++)
8283 {
8284 if (!i || !dops[i-1].is_jump)
8285 {
8286 if(!dops[i+1].bt)
8287 {
8288 int j, can_steal = 1;
8289 for (j = i; j < i + 2; j++) {
8290 int free_regs = 0;
8291 if (cinfo[j].min_free_regs == 0)
8292 continue;
8293 for (hr = 0; hr < HOST_REGS; hr++)
8294 if (hr != EXCLUDE_REG && regs[j].regmap[hr] < 0)
8295 free_regs++;
8296 if (free_regs <= cinfo[j].min_free_regs) {
8297 can_steal = 0;
8298 break;
8299 }
8300 }
8301 if (!can_steal)
8302 continue;
8303 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8304 ||(dops[i].itype==COP2&&dops[i].opcode2<3))
8305 {
8306 if(dops[i+1].rs1) {
8307 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8308 {
8309 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8310 {
8311 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8312 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8313 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8314 regs[i].isconst&=~(1<<hr);
8315 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8316 constmap[i][hr]=constmap[i+1][hr];
8317 regs[i+1].wasdirty&=~(1<<hr);
8318 regs[i].dirty&=~(1<<hr);
8319 }
8320 }
8321 }
8322 if(dops[i+1].rs2) {
8323 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8324 {
8325 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8326 {
8327 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8328 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8329 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8330 regs[i].isconst&=~(1<<hr);
8331 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8332 constmap[i][hr]=constmap[i+1][hr];
8333 regs[i+1].wasdirty&=~(1<<hr);
8334 regs[i].dirty&=~(1<<hr);
8335 }
8336 }
8337 }
8338 // Preload target address for load instruction (non-constant)
8339 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8340 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8341 {
8342 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8343 {
8344 regs[i].regmap[hr]=dops[i+1].rs1;
8345 regmap_pre[i+1][hr]=dops[i+1].rs1;
8346 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8347 regs[i].isconst&=~(1<<hr);
8348 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8349 constmap[i][hr]=constmap[i+1][hr];
8350 regs[i+1].wasdirty&=~(1<<hr);
8351 regs[i].dirty&=~(1<<hr);
8352 }
8353 }
8354 }
8355 // Load source into target register
8356 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8357 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8358 {
8359 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8360 {
8361 regs[i].regmap[hr]=dops[i+1].rs1;
8362 regmap_pre[i+1][hr]=dops[i+1].rs1;
8363 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8364 regs[i].isconst&=~(1<<hr);
8365 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8366 constmap[i][hr]=constmap[i+1][hr];
8367 regs[i+1].wasdirty&=~(1<<hr);
8368 regs[i].dirty&=~(1<<hr);
8369 }
8370 }
8371 }
8372 // Address for store instruction (non-constant)
8373 if (dops[i+1].is_store) { // SB/SH/SW/SWC2
8374 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8375 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8376 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8377 else {
8378 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8379 regs[i+1].isconst&=~(1<<hr);
8380 regs[i+1].dirty&=~(1<<hr);
8381 regs[i+2].wasdirty&=~(1<<hr);
8382 }
8383 assert(hr>=0);
8384 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8385 {
8386 regs[i].regmap[hr]=dops[i+1].rs1;
8387 regmap_pre[i+1][hr]=dops[i+1].rs1;
8388 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8389 regs[i].isconst&=~(1<<hr);
8390 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8391 constmap[i][hr]=constmap[i+1][hr];
8392 regs[i+1].wasdirty&=~(1<<hr);
8393 regs[i].dirty&=~(1<<hr);
8394 }
8395 }
8396 }
8397 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) { // LWC2
8398 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8399 int nr;
8400 hr=get_reg(regs[i+1].regmap,FTEMP);
8401 assert(hr>=0);
8402 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8403 {
8404 regs[i].regmap[hr]=dops[i+1].rs1;
8405 regmap_pre[i+1][hr]=dops[i+1].rs1;
8406 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8407 regs[i].isconst&=~(1<<hr);
8408 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8409 constmap[i][hr]=constmap[i+1][hr];
8410 regs[i+1].wasdirty&=~(1<<hr);
8411 regs[i].dirty&=~(1<<hr);
8412 }
8413 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8414 {
8415 // move it to another register
8416 regs[i+1].regmap[hr]=-1;
8417 regmap_pre[i+2][hr]=-1;
8418 regs[i+1].regmap[nr]=FTEMP;
8419 regmap_pre[i+2][nr]=FTEMP;
8420 regs[i].regmap[nr]=dops[i+1].rs1;
8421 regmap_pre[i+1][nr]=dops[i+1].rs1;
8422 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8423 regs[i].isconst&=~(1<<nr);
8424 regs[i+1].isconst&=~(1<<nr);
8425 regs[i].dirty&=~(1<<nr);
8426 regs[i+1].wasdirty&=~(1<<nr);
8427 regs[i+1].dirty&=~(1<<nr);
8428 regs[i+2].wasdirty&=~(1<<nr);
8429 }
8430 }
8431 }
8432 if(dops[i+1].itype==LOAD||dops[i+1].itype==LOADLR||dops[i+1].itype==STORE||dops[i+1].itype==STORELR/*||dops[i+1].itype==C2LS*/) {
8433 hr = -1;
8434 if(dops[i+1].itype==LOAD)
8435 hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1);
8436 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) // LWC2
8437 hr=get_reg(regs[i+1].regmap,FTEMP);
8438 if (dops[i+1].is_store) {
8439 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8440 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8441 }
8442 if(hr>=0&&regs[i].regmap[hr]<0) {
8443 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8444 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8445 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8446 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8447 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8448 regs[i].isconst&=~(1<<hr);
8449 regs[i+1].wasdirty&=~(1<<hr);
8450 regs[i].dirty&=~(1<<hr);
8451 }
8452 }
8453 }
8454 }
8455 }
8456 }
8457 }
8458}
8459
8460// Write back dirty registers as soon as we will no longer modify them,
8461// so that we don't end up with lots of writes at the branches.
8462static noinline void pass6_clean_registers(int istart, int iend, int wr)
8463{
8464 static u_int wont_dirty[MAXBLOCK];
8465 static u_int will_dirty[MAXBLOCK];
8466 int i;
8467 int r;
8468 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8469 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8470 if(iend==slen-1) {
8471 will_dirty_i=will_dirty_next=0;
8472 wont_dirty_i=wont_dirty_next=0;
8473 }else{
8474 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8475 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8476 }
8477 for (i=iend;i>=istart;i--)
8478 {
8479 signed char rregmap_i[RRMAP_SIZE];
8480 u_int hr_candirty = 0;
8481 assert(HOST_REGS < 32);
8482 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8483 __builtin_prefetch(regs[i-1].regmap);
8484 if(dops[i].is_jump)
8485 {
8486 signed char branch_rregmap_i[RRMAP_SIZE];
8487 u_int branch_hr_candirty = 0;
8488 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8489 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
8490 {
8491 // Branch out of this block, flush all regs
8492 will_dirty_i = 0;
8493 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8494 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8495 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8496 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8497 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8498 will_dirty_i &= branch_hr_candirty;
8499 if (dops[i].is_ujump)
8500 {
8501 // Unconditional branch
8502 wont_dirty_i = 0;
8503 // Merge in delay slot (will dirty)
8504 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8505 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8506 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8507 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8508 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8509 will_dirty_i &= hr_candirty;
8510 }
8511 else
8512 {
8513 // Conditional branch
8514 wont_dirty_i = wont_dirty_next;
8515 // Merge in delay slot (will dirty)
8516 // (the original code had no explanation why these 2 are commented out)
8517 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8518 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8519 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8520 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8521 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8522 will_dirty_i &= hr_candirty;
8523 }
8524 // Merge in delay slot (wont dirty)
8525 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8526 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8527 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8528 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8529 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8530 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8531 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8532 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8533 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8534 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8535 wont_dirty_i &= ~(1u << 31);
8536 if(wr) {
8537 #ifndef DESTRUCTIVE_WRITEBACK
8538 branch_regs[i].dirty&=wont_dirty_i;
8539 #endif
8540 branch_regs[i].dirty|=will_dirty_i;
8541 }
8542 }
8543 else
8544 {
8545 // Internal branch
8546 if(cinfo[i].ba<=start+i*4) {
8547 // Backward branch
8548 if (dops[i].is_ujump)
8549 {
8550 // Unconditional branch
8551 temp_will_dirty=0;
8552 temp_wont_dirty=0;
8553 // Merge in delay slot (will dirty)
8554 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8555 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8556 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8557 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8558 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8559 temp_will_dirty &= branch_hr_candirty;
8560 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8561 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8562 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8563 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8564 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8565 temp_will_dirty &= hr_candirty;
8566 } else {
8567 // Conditional branch (not taken case)
8568 temp_will_dirty=will_dirty_next;
8569 temp_wont_dirty=wont_dirty_next;
8570 // Merge in delay slot (will dirty)
8571 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8572 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8573 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8574 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8575 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8576 temp_will_dirty &= branch_hr_candirty;
8577 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8578 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8579 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8580 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8581 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8582 temp_will_dirty &= hr_candirty;
8583 }
8584 // Merge in delay slot (wont dirty)
8585 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8586 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8587 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8588 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8589 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8590 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8591 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8592 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8593 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8594 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8595 temp_wont_dirty &= ~(1u << 31);
8596 // Deal with changed mappings
8597 if(i<iend) {
8598 for(r=0;r<HOST_REGS;r++) {
8599 if(r!=EXCLUDE_REG) {
8600 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8601 temp_will_dirty&=~(1<<r);
8602 temp_wont_dirty&=~(1<<r);
8603 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8604 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8605 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8606 } else {
8607 temp_will_dirty|=1<<r;
8608 temp_wont_dirty|=1<<r;
8609 }
8610 }
8611 }
8612 }
8613 }
8614 if(wr) {
8615 will_dirty[i]=temp_will_dirty;
8616 wont_dirty[i]=temp_wont_dirty;
8617 pass6_clean_registers((cinfo[i].ba-start)>>2,i-1,0);
8618 }else{
8619 // Limit recursion. It can take an excessive amount
8620 // of time if there are a lot of nested loops.
8621 will_dirty[(cinfo[i].ba-start)>>2]=0;
8622 wont_dirty[(cinfo[i].ba-start)>>2]=-1;
8623 }
8624 }
8625 /*else*/ if(1)
8626 {
8627 if (dops[i].is_ujump)
8628 {
8629 // Unconditional branch
8630 will_dirty_i=0;
8631 wont_dirty_i=0;
8632 //if(cinfo[i].ba>start+i*4) { // Disable recursion (for debugging)
8633 for(r=0;r<HOST_REGS;r++) {
8634 if(r!=EXCLUDE_REG) {
8635 if(branch_regs[i].regmap[r]==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8636 will_dirty_i|=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8637 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8638 }
8639 if(branch_regs[i].regmap[r]>=0) {
8640 will_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8641 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8642 }
8643 }
8644 }
8645 //}
8646 // Merge in delay slot
8647 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8648 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8649 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8650 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8651 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8652 will_dirty_i &= branch_hr_candirty;
8653 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8654 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8655 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8656 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8657 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8658 will_dirty_i &= hr_candirty;
8659 } else {
8660 // Conditional branch
8661 will_dirty_i=will_dirty_next;
8662 wont_dirty_i=wont_dirty_next;
8663 //if(cinfo[i].ba>start+i*4) // Disable recursion (for debugging)
8664 for(r=0;r<HOST_REGS;r++) {
8665 if(r!=EXCLUDE_REG) {
8666 signed char target_reg=branch_regs[i].regmap[r];
8667 if(target_reg==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8668 will_dirty_i&=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8669 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8670 }
8671 else if(target_reg>=0) {
8672 will_dirty_i&=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8673 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8674 }
8675 }
8676 }
8677 // Merge in delay slot
8678 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8679 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8680 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8681 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8682 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8683 will_dirty_i &= branch_hr_candirty;
8684 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8685 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8686 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8687 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8688 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8689 will_dirty_i &= hr_candirty;
8690 }
8691 // Merge in delay slot (won't dirty)
8692 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8693 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8694 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8695 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8696 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8697 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8698 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8699 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8700 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8701 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8702 wont_dirty_i &= ~(1u << 31);
8703 if(wr) {
8704 #ifndef DESTRUCTIVE_WRITEBACK
8705 branch_regs[i].dirty&=wont_dirty_i;
8706 #endif
8707 branch_regs[i].dirty|=will_dirty_i;
8708 }
8709 }
8710 }
8711 }
8712 else if (dops[i].is_exception)
8713 {
8714 // SYSCALL instruction, etc
8715 will_dirty_i=0;
8716 wont_dirty_i=0;
8717 }
8718 will_dirty_next=will_dirty_i;
8719 wont_dirty_next=wont_dirty_i;
8720 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8721 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8722 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8723 will_dirty_i &= hr_candirty;
8724 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8725 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8726 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8727 wont_dirty_i &= ~(1u << 31);
8728 if (i > istart && !dops[i].is_jump) {
8729 // Don't store a register immediately after writing it,
8730 // may prevent dual-issue.
8731 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8732 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8733 }
8734 // Save it
8735 will_dirty[i]=will_dirty_i;
8736 wont_dirty[i]=wont_dirty_i;
8737 // Mark registers that won't be dirtied as not dirty
8738 if(wr) {
8739 regs[i].dirty|=will_dirty_i;
8740 #ifndef DESTRUCTIVE_WRITEBACK
8741 regs[i].dirty&=wont_dirty_i;
8742 if(dops[i].is_jump)
8743 {
8744 if (i < iend-1 && !dops[i].is_ujump) {
8745 for(r=0;r<HOST_REGS;r++) {
8746 if(r!=EXCLUDE_REG) {
8747 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8748 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8749 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8750 }
8751 }
8752 }
8753 }
8754 else
8755 {
8756 if(i<iend) {
8757 for(r=0;r<HOST_REGS;r++) {
8758 if(r!=EXCLUDE_REG) {
8759 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8760 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8761 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8762 }
8763 }
8764 }
8765 }
8766 #endif
8767 }
8768 // Deal with changed mappings
8769 temp_will_dirty=will_dirty_i;
8770 temp_wont_dirty=wont_dirty_i;
8771 for(r=0;r<HOST_REGS;r++) {
8772 if(r!=EXCLUDE_REG) {
8773 int nr;
8774 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8775 if(wr) {
8776 #ifndef DESTRUCTIVE_WRITEBACK
8777 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8778 #endif
8779 regs[i].wasdirty|=will_dirty_i&(1<<r);
8780 }
8781 }
8782 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8783 // Register moved to a different register
8784 will_dirty_i&=~(1<<r);
8785 wont_dirty_i&=~(1<<r);
8786 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8787 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8788 if(wr) {
8789 #ifndef DESTRUCTIVE_WRITEBACK
8790 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8791 #endif
8792 regs[i].wasdirty|=will_dirty_i&(1<<r);
8793 }
8794 }
8795 else {
8796 will_dirty_i&=~(1<<r);
8797 wont_dirty_i&=~(1<<r);
8798 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8799 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8800 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8801 } else {
8802 wont_dirty_i|=1<<r;
8803 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8804 }
8805 }
8806 }
8807 }
8808 }
8809}
8810
8811static noinline void pass10_expire_blocks(void)
8812{
8813 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
8814 // not sizeof(ndrc->translation_cache) due to vita hack
8815 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
8816 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
8817 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
8818 int hit;
8819
8820 for (; expirep != end; expirep = ((expirep + step) & step_mask))
8821 {
8822 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
8823 u_int block_i = expirep / step & (PAGE_COUNT - 1);
8824 u_int phase = (expirep >> (base_shift - 1)) & 1u;
8825 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
8826 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
8827 (long)(out - ndrc->translation_cache), phase);
8828 }
8829
8830 if (!phase) {
8831 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
8832 if (hit) {
8833 do_clear_cache();
8834 #ifdef USE_MINI_HT
8835 memset(mini_ht, -1, sizeof(mini_ht));
8836 #endif
8837 }
8838 }
8839 else
8840 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
8841 }
8842}
8843
8844static struct block_info *new_block_info(u_int start, u_int len,
8845 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
8846{
8847 struct block_info **b_pptr;
8848 struct block_info *block;
8849 u_int page = get_page(start);
8850
8851 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
8852 assert(block);
8853 assert(jump_in_count > 0);
8854 block->source = source;
8855 block->copy = copy;
8856 block->start = start;
8857 block->len = len;
8858 block->reg_sv_flags = 0;
8859 block->tc_offs = beginning - ndrc->translation_cache;
8860 //block->tc_len = out - beginning;
8861 block->is_dirty = 0;
8862 block->inv_near_misses = 0;
8863 block->jump_in_cnt = jump_in_count;
8864
8865 // insert sorted by start mirror-unmasked vaddr
8866 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
8867 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
8868 block->next = *b_pptr;
8869 *b_pptr = block;
8870 break;
8871 }
8872 }
8873 stat_inc(stat_blocks);
8874 return block;
8875}
8876
8877static int new_recompile_block(u_int addr)
8878{
8879 u_int pagelimit = 0;
8880 u_int state_rflags = 0;
8881 int i;
8882
8883 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
8884
8885 if (addr & 3) {
8886 if (addr != hack_addr) {
8887 SysPrintf("game crash @%08x, ra=%08x\n", addr, psxRegs.GPR.n.ra);
8888 hack_addr = addr;
8889 }
8890 return -1;
8891 }
8892
8893 // this is just for speculation
8894 for (i = 1; i < 32; i++) {
8895 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
8896 state_rflags |= 1 << i;
8897 }
8898
8899 start = addr;
8900 new_dynarec_did_compile=1;
8901 if (Config.HLE && start == 0x80001000) // hlecall
8902 {
8903 // XXX: is this enough? Maybe check hleSoftCall?
8904 void *beginning = start_block();
8905
8906 emit_movimm(start,0);
8907 emit_writeword(0,&pcaddr);
8908 emit_far_jump(new_dyna_leave);
8909 literal_pool(0);
8910 end_block(beginning);
8911 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8912 block->jump_in[0].vaddr = start;
8913 block->jump_in[0].addr = beginning;
8914 return 0;
8915 }
8916 else if (f1_hack && hack_addr == 0) {
8917 void *beginning = start_block();
8918 emit_movimm(start, 0);
8919 emit_writeword(0, &hack_addr);
8920 emit_readword(&psxRegs.GPR.n.sp, 0);
8921 emit_readptr(&mem_rtab, 1);
8922 emit_shrimm(0, 12, 2);
8923 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
8924 emit_addimm(0, 0x18, 0);
8925 emit_adds_ptr(1, 1, 1);
8926 emit_ldr_dualindexed(1, 0, 0);
8927 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
8928 emit_far_call(ndrc_get_addr_ht);
8929 emit_jmpreg(0); // jr k0
8930 literal_pool(0);
8931 end_block(beginning);
8932
8933 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8934 block->jump_in[0].vaddr = start;
8935 block->jump_in[0].addr = beginning;
8936 SysPrintf("F1 hack to %08x\n", start);
8937 return 0;
8938 }
8939
8940 cycle_multiplier_active = Config.cycle_multiplier_override && Config.cycle_multiplier == CYCLE_MULT_DEFAULT
8941 ? Config.cycle_multiplier_override : Config.cycle_multiplier;
8942
8943 source = get_source_start(start, &pagelimit);
8944 if (source == NULL) {
8945 if (addr != hack_addr) {
8946 SysPrintf("Compile at bogus memory address: %08x\n", addr);
8947 hack_addr = addr;
8948 }
8949 //abort();
8950 return -1;
8951 }
8952
8953 /* Pass 1: disassemble */
8954 /* Pass 2: register dependencies, branch targets */
8955 /* Pass 3: register allocation */
8956 /* Pass 4: branch dependencies */
8957 /* Pass 5: pre-alloc */
8958 /* Pass 6: optimize clean/dirty state */
8959 /* Pass 7: flag 32-bit registers */
8960 /* Pass 8: assembly */
8961 /* Pass 9: linker */
8962 /* Pass 10: garbage collection / free memory */
8963
8964 /* Pass 1 disassembly */
8965
8966 pass1_disassemble(pagelimit);
8967
8968 int clear_hack_addr = apply_hacks();
8969
8970 /* Pass 2 - Register dependencies and branch targets */
8971
8972 pass2_unneeded_regs(0,slen-1,0);
8973
8974 /* Pass 3 - Register allocation */
8975
8976 pass3_register_alloc(addr);
8977
8978 /* Pass 4 - Cull unused host registers */
8979
8980 pass4_cull_unused_regs();
8981
8982 /* Pass 5 - Pre-allocate registers */
8983
8984 pass5a_preallocate1();
8985 pass5b_preallocate2();
8986
8987 /* Pass 6 - Optimize clean/dirty state */
8988 pass6_clean_registers(0, slen-1, 1);
8989
8990 /* Pass 7 - Identify 32-bit registers */
8991 for (i=slen-1;i>=0;i--)
8992 {
8993 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8994 {
8995 // Conditional branch
8996 if((source[i]>>16)!=0x1000&&i<slen-2) {
8997 // Mark this address as a branch target since it may be called
8998 // upon return from interrupt
8999 dops[i+2].bt=1;
9000 }
9001 }
9002 }
9003
9004 /* Pass 8 - Assembly */
9005 linkcount=0;stubcount=0;
9006 is_delayslot=0;
9007 u_int dirty_pre=0;
9008 void *beginning=start_block();
9009 void *instr_addr0_override = NULL;
9010 int ds = 0;
9011
9012 if (start == 0x80030000) {
9013 // nasty hack for the fastbios thing
9014 // override block entry to this code
9015 instr_addr0_override = out;
9016 emit_movimm(start,0);
9017 // abuse io address var as a flag that we
9018 // have already returned here once
9019 emit_readword(&address,1);
9020 emit_writeword(0,&pcaddr);
9021 emit_writeword(0,&address);
9022 emit_cmp(0,1);
9023 #ifdef __aarch64__
9024 emit_jeq(out + 4*2);
9025 emit_far_jump(new_dyna_leave);
9026 #else
9027 emit_jne(new_dyna_leave);
9028 #endif
9029 }
9030 for(i=0;i<slen;i++)
9031 {
9032 __builtin_prefetch(regs[i+1].regmap);
9033 check_regmap(regmap_pre[i]);
9034 check_regmap(regs[i].regmap_entry);
9035 check_regmap(regs[i].regmap);
9036 //if(ds) printf("ds: ");
9037 disassemble_inst(i);
9038 if(ds) {
9039 ds=0; // Skip delay slot
9040 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9041 instr_addr[i] = NULL;
9042 } else {
9043 speculate_register_values(i);
9044 #ifndef DESTRUCTIVE_WRITEBACK
9045 if (i < 2 || !dops[i-2].is_ujump)
9046 {
9047 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9048 }
9049 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9050 dirty_pre=branch_regs[i].dirty;
9051 }else{
9052 dirty_pre=regs[i].dirty;
9053 }
9054 #endif
9055 // write back
9056 if (i < 2 || !dops[i-2].is_ujump)
9057 {
9058 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9059 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9060 }
9061 // branch target entry point
9062 instr_addr[i] = out;
9063 assem_debug("<->\n");
9064 drc_dbg_emit_do_cmp(i, cinfo[i].ccadj);
9065 if (clear_hack_addr) {
9066 emit_movimm(0, 0);
9067 emit_writeword(0, &hack_addr);
9068 clear_hack_addr = 0;
9069 }
9070
9071 // load regs
9072 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9073 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9074 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9075 address_generation(i,&regs[i],regs[i].regmap_entry);
9076 load_consts(regmap_pre[i],regs[i].regmap,i);
9077 if(dops[i].is_jump)
9078 {
9079 // Load the delay slot registers if necessary
9080 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))
9081 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9082 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))
9083 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9084 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9085 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9086 if (dops[i+1].is_store)
9087 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9088 }
9089 else if(i+1<slen)
9090 {
9091 // Preload registers for following instruction
9092 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9093 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9094 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9095 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9096 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9097 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9098 }
9099 // TODO: if(is_ooo(i)) address_generation(i+1);
9100 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9101 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9102 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9103 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9104 if (dops[i].is_store)
9105 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9106
9107 ds = assemble(i, &regs[i], cinfo[i].ccadj);
9108
9109 if (dops[i].is_ujump)
9110 literal_pool(1024);
9111 else
9112 literal_pool_jumpover(256);
9113 }
9114 }
9115
9116 assert(slen > 0);
9117 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9118 // no ending needed for this block since INTCALL never returns
9119 }
9120 // If the block did not end with an unconditional branch,
9121 // add a jump to the next instruction.
9122 else if (i > 1) {
9123 if (!dops[i-2].is_ujump) {
9124 assert(!dops[i-1].is_jump);
9125 assert(i==slen);
9126 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9127 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9128 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9129 emit_loadreg(CCREG,HOST_CCREG);
9130 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9131 }
9132 else
9133 {
9134 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9135 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9136 }
9137 add_to_linker(out,start+i*4,0);
9138 emit_jmp(0);
9139 }
9140 }
9141 else
9142 {
9143 assert(i>0);
9144 assert(!dops[i-1].is_jump);
9145 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9146 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9147 emit_loadreg(CCREG,HOST_CCREG);
9148 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9149 add_to_linker(out,start+i*4,0);
9150 emit_jmp(0);
9151 }
9152
9153 // Stubs
9154 for(i = 0; i < stubcount; i++)
9155 {
9156 switch(stubs[i].type)
9157 {
9158 case LOADB_STUB:
9159 case LOADH_STUB:
9160 case LOADW_STUB:
9161 case LOADBU_STUB:
9162 case LOADHU_STUB:
9163 do_readstub(i);break;
9164 case STOREB_STUB:
9165 case STOREH_STUB:
9166 case STOREW_STUB:
9167 do_writestub(i);break;
9168 case CC_STUB:
9169 do_ccstub(i);break;
9170 case INVCODE_STUB:
9171 do_invstub(i);break;
9172 case STORELR_STUB:
9173 do_unalignedwritestub(i);break;
9174 case OVERFLOW_STUB:
9175 do_overflowstub(i); break;
9176 case ALIGNMENT_STUB:
9177 do_alignmentstub(i); break;
9178 default:
9179 assert(0);
9180 }
9181 }
9182
9183 if (instr_addr0_override)
9184 instr_addr[0] = instr_addr0_override;
9185
9186#if 0
9187 /* check for improper expiration */
9188 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9189 int j;
9190 if (!jumps[i])
9191 continue;
9192 for (j = 0; j < jumps[i]->count; j++)
9193 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9194 }
9195#endif
9196
9197 /* Pass 9 - Linker */
9198 for(i=0;i<linkcount;i++)
9199 {
9200 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9201 literal_pool(64);
9202 if (!link_addr[i].internal)
9203 {
9204 void *stub = out;
9205 void *addr = check_addr(link_addr[i].target);
9206 emit_extjump(link_addr[i].addr, link_addr[i].target);
9207 if (addr) {
9208 set_jump_target(link_addr[i].addr, addr);
9209 ndrc_add_jump_out(link_addr[i].target,stub);
9210 }
9211 else
9212 set_jump_target(link_addr[i].addr, stub);
9213 }
9214 else
9215 {
9216 // Internal branch
9217 int target=(link_addr[i].target-start)>>2;
9218 assert(target>=0&&target<slen);
9219 assert(instr_addr[target]);
9220 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9221 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9222 //#else
9223 set_jump_target(link_addr[i].addr, instr_addr[target]);
9224 //#endif
9225 }
9226 }
9227
9228 u_int source_len = slen*4;
9229 if (dops[slen-1].itype == INTCALL && source_len > 4)
9230 // no need to treat the last instruction as compiled
9231 // as interpreter fully handles it
9232 source_len -= 4;
9233
9234 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9235 copy = shadow;
9236
9237 // External Branch Targets (jump_in)
9238 int jump_in_count = 1;
9239 assert(instr_addr[0]);
9240 for (i = 1; i < slen; i++)
9241 {
9242 if (dops[i].bt && instr_addr[i])
9243 jump_in_count++;
9244 }
9245
9246 struct block_info *block =
9247 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9248 block->reg_sv_flags = state_rflags;
9249
9250 int jump_in_i = 0;
9251 for (i = 0; i < slen; i++)
9252 {
9253 if ((i == 0 || dops[i].bt) && instr_addr[i])
9254 {
9255 assem_debug("%p (%d) <- %8x\n", instr_addr[i], i, start + i*4);
9256 u_int vaddr = start + i*4;
9257
9258 literal_pool(256);
9259 void *entry = out;
9260 load_regs_entry(i);
9261 if (entry == out)
9262 entry = instr_addr[i];
9263 else
9264 emit_jmp(instr_addr[i]);
9265
9266 block->jump_in[jump_in_i].vaddr = vaddr;
9267 block->jump_in[jump_in_i].addr = entry;
9268 jump_in_i++;
9269 }
9270 }
9271 assert(jump_in_i == jump_in_count);
9272 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9273 // Write out the literal pool if necessary
9274 literal_pool(0);
9275 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9276 // Align code
9277 if(((u_int)out)&7) emit_addnop(13);
9278 #endif
9279 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9280 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9281 memcpy(copy, source, source_len);
9282 copy += source_len;
9283
9284 end_block(beginning);
9285
9286 // If we're within 256K of the end of the buffer,
9287 // start over from the beginning. (Is 256K enough?)
9288 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9289 out = ndrc->translation_cache;
9290
9291 // Trap writes to any of the pages we compiled
9292 mark_invalid_code(start, slen*4, 0);
9293
9294 /* Pass 10 - Free memory by expiring oldest blocks */
9295
9296 pass10_expire_blocks();
9297
9298#ifdef ASSEM_PRINT
9299 fflush(stdout);
9300#endif
9301 stat_inc(stat_bc_direct);
9302 return 0;
9303}
9304
9305// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork