notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
drc: update according to interpreter (2)
[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 waswritten; // MIPS regs that were used as store base before
168};
169
170struct ht_entry
171{
172 u_int vaddr[2];
173 void *tcaddr[2];
174};
175
176struct code_stub
177{
178 enum stub_type type;
179 void *addr;
180 void *retaddr;
181 u_int a;
182 uintptr_t b;
183 uintptr_t c;
184 u_int d;
185 u_int e;
186};
187
188struct link_entry
189{
190 void *addr;
191 u_int target;
192 u_int internal;
193};
194
195struct block_info
196{
197 struct block_info *next;
198 const void *source;
199 const void *copy;
200 u_int start; // vaddr of the block start
201 u_int len; // of the whole block source
202 u_int tc_offs;
203 //u_int tc_len;
204 u_int reg_sv_flags;
205 u_char is_dirty;
206 u_char inv_near_misses;
207 u_short jump_in_cnt;
208 struct {
209 u_int vaddr;
210 void *addr;
211 } jump_in[0];
212};
213
214struct jump_info
215{
216 int alloc;
217 int count;
218 struct {
219 u_int target_vaddr;
220 void *stub;
221 } e[0];
222};
223
224static struct decoded_insn
225{
226 u_char itype;
227 u_char opcode; // bits 31-26
228 u_char opcode2; // (depends on opcode)
229 u_char rs1;
230 u_char rs2;
231 u_char rt1;
232 u_char rt2;
233 u_char use_lt1:1;
234 u_char bt:1;
235 u_char ooo:1;
236 u_char is_ds:1;
237 u_char is_jump:1;
238 u_char is_ujump:1;
239 u_char is_load:1;
240 u_char is_store:1;
241 u_char is_delay_load:1; // is_load + MFC/CFC
242 u_char is_exception:1; // unconditional, also interp. fallback
243 u_char may_except:1; // might generate an exception
244} dops[MAXBLOCK];
245
246static struct compile_info
247{
248 int imm;
249 u_int ba;
250 int ccadj;
251 signed char min_free_regs;
252 signed char addr;
253 signed char reserved[2];
254} cinfo[MAXBLOCK];
255
256 static u_char *out;
257 static char invalid_code[0x100000];
258 static struct ht_entry hash_table[65536];
259 static struct block_info *blocks[PAGE_COUNT];
260 static struct jump_info *jumps[PAGE_COUNT];
261 static u_int start;
262 static u_int *source;
263 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
264 static uint64_t gte_rt[MAXBLOCK];
265 static uint64_t gte_unneeded[MAXBLOCK];
266 static u_int smrv[32]; // speculated MIPS register values
267 static u_int smrv_strong; // mask or regs that are likely to have correct values
268 static u_int smrv_weak; // same, but somewhat less likely
269 static u_int smrv_strong_next; // same, but after current insn executes
270 static u_int smrv_weak_next;
271 static uint64_t unneeded_reg[MAXBLOCK];
272 static uint64_t branch_unneeded_reg[MAXBLOCK];
273 // see 'struct regstat' for a description
274 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
275 // contains 'real' consts at [i] insn, but may differ from what's actually
276 // loaded in host reg as 'final' value is always loaded, see get_final_value()
277 static uint32_t current_constmap[HOST_REGS];
278 static uint32_t constmap[MAXBLOCK][HOST_REGS];
279 static struct regstat regs[MAXBLOCK];
280 static struct regstat branch_regs[MAXBLOCK];
281 static int slen;
282 static void *instr_addr[MAXBLOCK];
283 static struct link_entry link_addr[MAXBLOCK];
284 static int linkcount;
285 static struct code_stub stubs[MAXBLOCK*3];
286 static int stubcount;
287 static u_int literals[1024][2];
288 static int literalcount;
289 static int is_delayslot;
290 static char shadow[1048576] __attribute__((aligned(16)));
291 static void *copy;
292 static u_int expirep;
293 static u_int stop_after_jal;
294 static u_int f1_hack;
295#ifdef STAT_PRINT
296 static int stat_bc_direct;
297 static int stat_bc_pre;
298 static int stat_bc_restore;
299 static int stat_ht_lookups;
300 static int stat_jump_in_lookups;
301 static int stat_restore_tries;
302 static int stat_restore_compares;
303 static int stat_inv_addr_calls;
304 static int stat_inv_hits;
305 static int stat_blocks;
306 static int stat_links;
307 #define stat_inc(s) s++
308 #define stat_dec(s) s--
309 #define stat_clear(s) s = 0
310#else
311 #define stat_inc(s)
312 #define stat_dec(s)
313 #define stat_clear(s)
314#endif
315
316 int new_dynarec_hacks;
317 int new_dynarec_hacks_pergame;
318 int new_dynarec_hacks_old;
319 int new_dynarec_did_compile;
320
321 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
322
323 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
324 extern int last_count; // last absolute target, often = next_interupt
325 extern int pcaddr;
326 extern int pending_exception;
327 extern int branch_target;
328 extern uintptr_t ram_offset;
329 extern uintptr_t mini_ht[32][2];
330
331 /* registers that may be allocated */
332 /* 1-31 gpr */
333#define LOREG 32 // lo
334#define HIREG 33 // hi
335//#define FSREG 34 // FPU status (FCSR)
336#define CSREG 35 // Coprocessor status
337#define CCREG 36 // Cycle count
338#define INVCP 37 // Pointer to invalid_code
339//#define MMREG 38 // Pointer to memory_map
340#define ROREG 39 // ram offset (if rdram!=0x80000000)
341#define TEMPREG 40
342#define FTEMP 40 // FPU temporary register
343#define PTEMP 41 // Prefetch temporary register
344//#define TLREG 42 // TLB mapping offset
345#define RHASH 43 // Return address hash
346#define RHTBL 44 // Return address hash table address
347#define RTEMP 45 // JR/JALR address register
348#define MAXREG 45
349#define AGEN1 46 // Address generation temporary register (pass5b_preallocate2)
350//#define AGEN2 47 // Address generation temporary register
351#define BTREG 50 // Branch target temporary register
352
353 /* instruction types */
354#define NOP 0 // No operation
355#define LOAD 1 // Load
356#define STORE 2 // Store
357#define LOADLR 3 // Unaligned load
358#define STORELR 4 // Unaligned store
359#define MOV 5 // Move (hi/lo only)
360#define ALU 6 // Arithmetic/logic
361#define MULTDIV 7 // Multiply/divide
362#define SHIFT 8 // Shift by register
363#define SHIFTIMM 9// Shift by immediate
364#define IMM16 10 // 16-bit immediate
365#define RJUMP 11 // Unconditional jump to register
366#define UJUMP 12 // Unconditional jump
367#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
368#define SJUMP 14 // Conditional branch (regimm format)
369#define COP0 15 // Coprocessor 0
370#define RFE 16
371#define SYSCALL 22// SYSCALL,BREAK
372#define OTHER 23 // Other/unknown - do nothing
373#define HLECALL 26// PCSX fake opcodes for HLE
374#define COP2 27 // Coprocessor 2 move
375#define C2LS 28 // Coprocessor 2 load/store
376#define C2OP 29 // Coprocessor 2 operation
377#define INTCALL 30// Call interpreter to handle rare corner cases
378
379 /* branch codes */
380#define TAKEN 1
381#define NOTTAKEN 2
382#define NULLDS 3
383
384#define DJT_1 (void *)1l // no function, just a label in assem_debug log
385#define DJT_2 (void *)2l
386
387// asm linkage
388void dyna_linker();
389void cc_interrupt();
390void jump_syscall (u_int u0, u_int u1, u_int pc);
391void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
392void jump_break (u_int u0, u_int u1, u_int pc);
393void jump_break_ds(u_int u0, u_int u1, u_int pc);
394void jump_overflow (u_int u0, u_int u1, u_int pc);
395void jump_overflow_ds(u_int u0, u_int u1, u_int pc);
396void jump_addrerror (u_int cause, u_int addr, u_int pc);
397void jump_addrerror_ds(u_int cause, u_int addr, u_int pc);
398void jump_to_new_pc();
399void call_gteStall();
400void new_dyna_leave();
401
402void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile);
403void *ndrc_get_addr_ht(u_int vaddr);
404void ndrc_add_jump_out(u_int vaddr, void *src);
405void ndrc_write_invalidate_one(u_int addr);
406static void ndrc_write_invalidate_many(u_int addr, u_int end);
407
408static int new_recompile_block(u_int addr);
409static void invalidate_block(struct block_info *block);
410static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_);
411
412// Needed by assembler
413static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
414static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
415static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
416static void load_all_regs(const signed char i_regmap[]);
417static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
418static void load_regs_entry(int t);
419static void load_all_consts(const signed char regmap[], u_int dirty, int i);
420static u_int get_host_reglist(const signed char *regmap);
421
422static int get_final_value(int hr, int i, int *value);
423static void add_stub(enum stub_type type, void *addr, void *retaddr,
424 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
425static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
426 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
427static void add_to_linker(void *addr, u_int target, int ext);
428static void *get_direct_memhandler(void *table, u_int addr,
429 enum stub_type type, uintptr_t *addr_host);
430static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
431static void pass_args(int a0, int a1);
432static void emit_far_jump(const void *f);
433static void emit_far_call(const void *f);
434
435#ifdef VITA
436#include <psp2/kernel/sysmem.h>
437static int sceBlock;
438// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
439extern int getVMBlock();
440int _newlib_vm_size_user = sizeof(*ndrc);
441#endif
442
443static void mprotect_w_x(void *start, void *end, int is_x)
444{
445#ifdef NO_WRITE_EXEC
446 #if defined(VITA)
447 // *Open* enables write on all memory that was
448 // allocated by sceKernelAllocMemBlockForVM()?
449 if (is_x)
450 sceKernelCloseVMDomain();
451 else
452 sceKernelOpenVMDomain();
453 #elif defined(HAVE_LIBNX)
454 Result rc;
455 // check to avoid the full flush in jitTransitionToExecutable()
456 if (g_jit.type != JitType_CodeMemory) {
457 if (is_x)
458 rc = jitTransitionToExecutable(&g_jit);
459 else
460 rc = jitTransitionToWritable(&g_jit);
461 if (R_FAILED(rc))
462 ;//SysPrintf("jitTransition %d %08x\n", is_x, rc);
463 }
464 #elif defined(TC_WRITE_OFFSET)
465 // separated rx and rw areas are always available
466 #else
467 u_long mstart = (u_long)start & ~4095ul;
468 u_long mend = (u_long)end;
469 if (mprotect((void *)mstart, mend - mstart,
470 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
471 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
472 #endif
473#endif
474}
475
476static void start_tcache_write(void *start, void *end)
477{
478 mprotect_w_x(start, end, 0);
479}
480
481static void end_tcache_write(void *start, void *end)
482{
483#if defined(__arm__) || defined(__aarch64__)
484 size_t len = (char *)end - (char *)start;
485 #if defined(__BLACKBERRY_QNX__)
486 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
487 #elif defined(__MACH__)
488 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
489 #elif defined(VITA)
490 sceKernelSyncVMDomain(sceBlock, start, len);
491 #elif defined(_3DS)
492 ctr_flush_invalidate_cache();
493 #elif defined(HAVE_LIBNX)
494 if (g_jit.type == JitType_CodeMemory) {
495 armDCacheClean(start, len);
496 armICacheInvalidate((char *)start - ndrc_write_ofs, len);
497 // as of v4.2.1 libnx lacks isb
498 __asm__ volatile("isb" ::: "memory");
499 }
500 #elif defined(__aarch64__)
501 // as of 2021, __clear_cache() is still broken on arm64
502 // so here is a custom one :(
503 clear_cache_arm64(start, end);
504 #else
505 __clear_cache(start, end);
506 #endif
507 (void)len;
508#endif
509
510 mprotect_w_x(start, end, 1);
511}
512
513static void *start_block(void)
514{
515 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
516 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
517 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
518 start_tcache_write(NDRC_WRITE_OFFSET(out), NDRC_WRITE_OFFSET(end));
519 return out;
520}
521
522static void end_block(void *start)
523{
524 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(out));
525}
526
527#ifdef NDRC_CACHE_FLUSH_ALL
528
529static int needs_clear_cache;
530
531static void mark_clear_cache(void *target)
532{
533 if (!needs_clear_cache) {
534 start_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
535 needs_clear_cache = 1;
536 }
537}
538
539static void do_clear_cache(void)
540{
541 if (needs_clear_cache) {
542 end_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
543 needs_clear_cache = 0;
544 }
545}
546
547#else
548
549// also takes care of w^x mappings when patching code
550static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
551
552static void mark_clear_cache(void *target)
553{
554 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
555 u_int mask = 1u << ((offset >> 12) & 31);
556 if (!(needs_clear_cache[offset >> 17] & mask)) {
557 char *start = (char *)NDRC_WRITE_OFFSET((uintptr_t)target & ~4095l);
558 start_tcache_write(start, start + 4095);
559 needs_clear_cache[offset >> 17] |= mask;
560 }
561}
562
563// Clearing the cache is rather slow on ARM Linux, so mark the areas
564// that need to be cleared, and then only clear these areas once.
565static void do_clear_cache(void)
566{
567 int i, j;
568 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
569 {
570 u_int bitmap = needs_clear_cache[i];
571 if (!bitmap)
572 continue;
573 for (j = 0; j < 32; j++)
574 {
575 u_char *start, *end;
576 if (!(bitmap & (1u << j)))
577 continue;
578
579 start = ndrc->translation_cache + i*131072 + j*4096;
580 end = start + 4095;
581 for (j++; j < 32; j++) {
582 if (!(bitmap & (1u << j)))
583 break;
584 end += 4096;
585 }
586 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(end));
587 }
588 needs_clear_cache[i] = 0;
589 }
590}
591
592#endif // NDRC_CACHE_FLUSH_ALL
593
594#define NO_CYCLE_PENALTY_THR 12
595
596int cycle_multiplier_old;
597static int cycle_multiplier_active;
598
599static int CLOCK_ADJUST(int x)
600{
601 int m = cycle_multiplier_active;
602 int s = (x >> 31) | 1;
603 return (x * m + s * 50) / 100;
604}
605
606static int ds_writes_rjump_rs(int i)
607{
608 return dops[i].rs1 != 0
609 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2
610 || dops[i].rs1 == dops[i].rt1); // overwrites itself - same effect
611}
612
613// psx addr mirror masking (for invalidation)
614static u_int pmmask(u_int vaddr)
615{
616 vaddr &= ~0xe0000000;
617 if (vaddr < 0x01000000)
618 vaddr &= ~0x00e00000; // RAM mirrors
619 return vaddr;
620}
621
622static u_int get_page(u_int vaddr)
623{
624 u_int page = pmmask(vaddr) >> 12;
625 if (page >= PAGE_COUNT / 2)
626 page = PAGE_COUNT / 2 + (page & (PAGE_COUNT / 2 - 1));
627 return page;
628}
629
630// get a page for looking for a block that has vaddr
631// (needed because the block may start in previous page)
632static u_int get_page_prev(u_int vaddr)
633{
634 assert(MAXBLOCK <= (1 << 12));
635 u_int page = get_page(vaddr);
636 if (page & 511)
637 page--;
638 return page;
639}
640
641static struct ht_entry *hash_table_get(u_int vaddr)
642{
643 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
644}
645
646static void hash_table_add(u_int vaddr, void *tcaddr)
647{
648 struct ht_entry *ht_bin = hash_table_get(vaddr);
649 assert(tcaddr);
650 ht_bin->vaddr[1] = ht_bin->vaddr[0];
651 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
652 ht_bin->vaddr[0] = vaddr;
653 ht_bin->tcaddr[0] = tcaddr;
654}
655
656static void hash_table_remove(int vaddr)
657{
658 //printf("remove hash: %x\n",vaddr);
659 struct ht_entry *ht_bin = hash_table_get(vaddr);
660 if (ht_bin->vaddr[1] == vaddr) {
661 ht_bin->vaddr[1] = -1;
662 ht_bin->tcaddr[1] = NULL;
663 }
664 if (ht_bin->vaddr[0] == vaddr) {
665 ht_bin->vaddr[0] = ht_bin->vaddr[1];
666 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
667 ht_bin->vaddr[1] = -1;
668 ht_bin->tcaddr[1] = NULL;
669 }
670}
671
672static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
673{
674 u_int vaddr_m = vaddr & 0x1fffffff;
675 u_int i, j;
676 for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
677 // ram mirrors, but should not hurt bios
678 for (j = 0; j < 0x800000; j += 0x200000) {
679 invalid_code[(i|j) >> 12] =
680 invalid_code[(i|j|0x80000000u) >> 12] =
681 invalid_code[(i|j|0xa0000000u) >> 12] = invalid;
682 }
683 }
684 if (!invalid && vaddr + len > inv_code_start && vaddr <= inv_code_end)
685 inv_code_start = inv_code_end = ~0;
686}
687
688static int doesnt_expire_soon(u_char *tcaddr)
689{
690 u_int diff = (u_int)(tcaddr - out) & ((1u << TARGET_SIZE_2) - 1u);
691 return diff > EXPIRITY_OFFSET + MAX_OUTPUT_BLOCK_SIZE;
692}
693
694static unused void check_for_block_changes(u_int start, u_int end)
695{
696 u_int start_page = get_page_prev(start);
697 u_int end_page = get_page(end - 1);
698 u_int page;
699
700 for (page = start_page; page <= end_page; page++) {
701 struct block_info *block;
702 for (block = blocks[page]; block != NULL; block = block->next) {
703 if (block->is_dirty)
704 continue;
705 if (memcmp(block->source, block->copy, block->len)) {
706 printf("bad block %08x-%08x %016llx %016llx @%08x\n",
707 block->start, block->start + block->len,
708 *(long long *)block->source, *(long long *)block->copy, psxRegs.pc);
709 fflush(stdout);
710 abort();
711 }
712 }
713 }
714}
715
716static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
717{
718 void *found_clean = NULL;
719 u_int i, page;
720
721 stat_inc(stat_restore_tries);
722 for (page = start_page; page <= end_page; page++) {
723 struct block_info *block;
724 for (block = blocks[page]; block != NULL; block = block->next) {
725 if (vaddr < block->start)
726 break;
727 if (!block->is_dirty || vaddr >= block->start + block->len)
728 continue;
729 for (i = 0; i < block->jump_in_cnt; i++)
730 if (block->jump_in[i].vaddr == vaddr)
731 break;
732 if (i == block->jump_in_cnt)
733 continue;
734 assert(block->source && block->copy);
735 stat_inc(stat_restore_compares);
736 if (memcmp(block->source, block->copy, block->len))
737 continue;
738
739 block->is_dirty = block->inv_near_misses = 0;
740 found_clean = block->jump_in[i].addr;
741 hash_table_add(vaddr, found_clean);
742 mark_invalid_code(block->start, block->len, 0);
743 stat_inc(stat_bc_restore);
744 inv_debug("INV: restored %08x %p (%d)\n", vaddr, found_clean, block->jump_in_cnt);
745 return found_clean;
746 }
747 }
748 return NULL;
749}
750
751// Get address from virtual address
752// This is called from the recompiled JR/JALR instructions
753static void noinline *get_addr(u_int vaddr, int can_compile)
754{
755 u_int start_page = get_page_prev(vaddr);
756 u_int i, page, end_page = get_page(vaddr);
757 void *found_clean = NULL;
758
759 stat_inc(stat_jump_in_lookups);
760 for (page = start_page; page <= end_page; page++) {
761 const struct block_info *block;
762 for (block = blocks[page]; block != NULL; block = block->next) {
763 if (vaddr < block->start)
764 break;
765 if (block->is_dirty || vaddr >= block->start + block->len)
766 continue;
767 for (i = 0; i < block->jump_in_cnt; i++)
768 if (block->jump_in[i].vaddr == vaddr)
769 break;
770 if (i == block->jump_in_cnt)
771 continue;
772 found_clean = block->jump_in[i].addr;
773 hash_table_add(vaddr, found_clean);
774 return found_clean;
775 }
776 }
777 found_clean = try_restore_block(vaddr, start_page, end_page);
778 if (found_clean)
779 return found_clean;
780
781 if (!can_compile)
782 return NULL;
783
784 int r = new_recompile_block(vaddr);
785 if (r == 0)
786 return ndrc_get_addr_ht(vaddr);
787
788 // generate an address error
789#ifdef DRC_DBG
790 last_count -= 2;
791#endif
792 psxRegs.CP0.n.Cause &= 0x300;
793 psxRegs.CP0.n.EPC = vaddr;
794 if (vaddr & 3) {
795 psxRegs.CP0.n.Cause |= R3000E_AdEL << 2;
796 psxRegs.CP0.n.BadVAddr = vaddr;
797 } else
798 psxRegs.CP0.n.Cause |= R3000E_IBE << 2;
799 psxRegs.pc = 0x80000080;
800 return ndrc_get_addr_ht(0x80000080);
801}
802
803// Look up address in hash table first
804void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile)
805{
806 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
807 const struct ht_entry *ht_bin = hash_table_get(vaddr);
808 u_int vaddr_a = vaddr & ~3;
809 stat_inc(stat_ht_lookups);
810 if (ht_bin->vaddr[0] == vaddr_a) return ht_bin->tcaddr[0];
811 if (ht_bin->vaddr[1] == vaddr_a) return ht_bin->tcaddr[1];
812 return get_addr(vaddr, can_compile);
813}
814
815void *ndrc_get_addr_ht(u_int vaddr)
816{
817 return ndrc_get_addr_ht_param(vaddr, 1);
818}
819
820static void clear_all_regs(signed char regmap[])
821{
822 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
823}
824
825// get_reg: get allocated host reg from mips reg
826// returns -1 if no such mips reg was allocated
827#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
828
829extern signed char get_reg(const signed char regmap[], signed char r);
830
831#else
832
833static signed char get_reg(const signed char regmap[], signed char r)
834{
835 int hr;
836 for (hr = 0; hr < HOST_REGS; hr++) {
837 if (hr == EXCLUDE_REG)
838 continue;
839 if (regmap[hr] == r)
840 return hr;
841 }
842 return -1;
843}
844
845#endif
846
847// get reg suitable for writing
848static signed char get_reg_w(const signed char regmap[], signed char r)
849{
850 return r == 0 ? -1 : get_reg(regmap, r);
851}
852
853// get reg as mask bit (1 << hr)
854static u_int get_regm(const signed char regmap[], signed char r)
855{
856 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
857}
858
859static signed char get_reg_temp(const signed char regmap[])
860{
861 int hr;
862 for (hr = 0; hr < HOST_REGS; hr++) {
863 if (hr == EXCLUDE_REG)
864 continue;
865 if (regmap[hr] == (signed char)-1)
866 return hr;
867 }
868 return -1;
869}
870
871// Find a register that is available for two consecutive cycles
872static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
873{
874 int hr;
875 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
876 return -1;
877}
878
879// reverse reg map: mips -> host
880#define RRMAP_SIZE 64
881static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
882 u_int *regs_can_change)
883{
884 u_int r, hr, hr_can_change = 0;
885 memset(rrmap, -1, RRMAP_SIZE);
886 for (hr = 0; hr < HOST_REGS; )
887 {
888 r = regmap[hr];
889 rrmap[r & (RRMAP_SIZE - 1)] = hr;
890 // only add mips $1-$31+$lo, others shifted out
891 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
892 hr++;
893 if (hr == EXCLUDE_REG)
894 hr++;
895 }
896 hr_can_change |= 1u << (rrmap[33] & 31);
897 hr_can_change |= 1u << (rrmap[CCREG] & 31);
898 hr_can_change &= ~(1u << 31);
899 *regs_can_change = hr_can_change;
900}
901
902// same as get_reg, but takes rrmap
903static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
904{
905 assert(0 <= r && r < RRMAP_SIZE);
906 return rrmap[r];
907}
908
909static int count_free_regs(const signed char regmap[])
910{
911 int count=0;
912 int hr;
913 for(hr=0;hr<HOST_REGS;hr++)
914 {
915 if(hr!=EXCLUDE_REG) {
916 if(regmap[hr]<0) count++;
917 }
918 }
919 return count;
920}
921
922static void dirty_reg(struct regstat *cur, signed char reg)
923{
924 int hr;
925 if (!reg) return;
926 hr = get_reg(cur->regmap, reg);
927 if (hr >= 0)
928 cur->dirty |= 1<<hr;
929}
930
931static void set_const(struct regstat *cur, signed char reg, uint32_t value)
932{
933 int hr;
934 if (!reg) return;
935 hr = get_reg(cur->regmap, reg);
936 if (hr >= 0) {
937 cur->isconst |= 1<<hr;
938 current_constmap[hr] = value;
939 }
940}
941
942static void clear_const(struct regstat *cur, signed char reg)
943{
944 int hr;
945 if (!reg) return;
946 hr = get_reg(cur->regmap, reg);
947 if (hr >= 0)
948 cur->isconst &= ~(1<<hr);
949}
950
951static int is_const(const struct regstat *cur, signed char reg)
952{
953 int hr;
954 if (reg < 0) return 0;
955 if (!reg) return 1;
956 hr = get_reg(cur->regmap, reg);
957 if (hr >= 0)
958 return (cur->isconst>>hr)&1;
959 return 0;
960}
961
962static uint32_t get_const(const struct regstat *cur, signed char reg)
963{
964 int hr;
965 if (!reg) return 0;
966 hr = get_reg(cur->regmap, reg);
967 if (hr >= 0)
968 return current_constmap[hr];
969
970 SysPrintf("Unknown constant in r%d\n", reg);
971 abort();
972}
973
974// Least soon needed registers
975// Look at the next ten instructions and see which registers
976// will be used. Try not to reallocate these.
977static void lsn(u_char hsn[], int i, int *preferred_reg)
978{
979 int j;
980 int b=-1;
981 for(j=0;j<9;j++)
982 {
983 if(i+j>=slen) {
984 j=slen-i-1;
985 break;
986 }
987 if (dops[i+j].is_ujump)
988 {
989 // Don't go past an unconditonal jump
990 j++;
991 break;
992 }
993 }
994 for(;j>=0;j--)
995 {
996 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
997 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
998 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
999 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
1000 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
1001 // Stores can allocate zero
1002 hsn[dops[i+j].rs1]=j;
1003 hsn[dops[i+j].rs2]=j;
1004 }
1005 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
1006 hsn[ROREG] = j;
1007 // On some architectures stores need invc_ptr
1008 #if defined(HOST_IMM8)
1009 if (dops[i+j].is_store)
1010 hsn[INVCP] = j;
1011 #endif
1012 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1013 {
1014 hsn[CCREG]=j;
1015 b=j;
1016 }
1017 }
1018 if(b>=0)
1019 {
1020 if(cinfo[i+b].ba>=start && cinfo[i+b].ba<(start+slen*4))
1021 {
1022 // Follow first branch
1023 int t=(cinfo[i+b].ba-start)>>2;
1024 j=7-b;if(t+j>=slen) j=slen-t-1;
1025 for(;j>=0;j--)
1026 {
1027 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1028 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1029 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1030 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1031 }
1032 }
1033 // TODO: preferred register based on backward branch
1034 }
1035 // Delay slot should preferably not overwrite branch conditions or cycle count
1036 if (i > 0 && dops[i-1].is_jump) {
1037 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1038 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1039 hsn[CCREG]=1;
1040 // ...or hash tables
1041 hsn[RHASH]=1;
1042 hsn[RHTBL]=1;
1043 }
1044 // Coprocessor load/store needs FTEMP, even if not declared
1045 if(dops[i].itype==C2LS) {
1046 hsn[FTEMP]=0;
1047 }
1048 // Load L/R also uses FTEMP as a temporary register
1049 if(dops[i].itype==LOADLR) {
1050 hsn[FTEMP]=0;
1051 }
1052 // Also SWL/SWR/SDL/SDR
1053 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
1054 hsn[FTEMP]=0;
1055 }
1056 // Don't remove the miniht registers
1057 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1058 {
1059 hsn[RHASH]=0;
1060 hsn[RHTBL]=0;
1061 }
1062}
1063
1064// We only want to allocate registers if we're going to use them again soon
1065static int needed_again(int r, int i)
1066{
1067 int j;
1068 int b=-1;
1069 int rn=10;
1070
1071 if (i > 0 && dops[i-1].is_ujump)
1072 {
1073 if(cinfo[i-1].ba<start || cinfo[i-1].ba>start+slen*4-4)
1074 return 0; // Don't need any registers if exiting the block
1075 }
1076 for(j=0;j<9;j++)
1077 {
1078 if(i+j>=slen) {
1079 j=slen-i-1;
1080 break;
1081 }
1082 if (dops[i+j].is_ujump)
1083 {
1084 // Don't go past an unconditonal jump
1085 j++;
1086 break;
1087 }
1088 if (dops[i+j].is_exception)
1089 {
1090 break;
1091 }
1092 }
1093 for(;j>=1;j--)
1094 {
1095 if(dops[i+j].rs1==r) rn=j;
1096 if(dops[i+j].rs2==r) rn=j;
1097 if((unneeded_reg[i+j]>>r)&1) rn=10;
1098 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1099 {
1100 b=j;
1101 }
1102 }
1103 if(rn<10) return 1;
1104 (void)b;
1105 return 0;
1106}
1107
1108// Try to match register allocations at the end of a loop with those
1109// at the beginning
1110static int loop_reg(int i, int r, int hr)
1111{
1112 int j,k;
1113 for(j=0;j<9;j++)
1114 {
1115 if(i+j>=slen) {
1116 j=slen-i-1;
1117 break;
1118 }
1119 if (dops[i+j].is_ujump)
1120 {
1121 // Don't go past an unconditonal jump
1122 j++;
1123 break;
1124 }
1125 }
1126 k=0;
1127 if(i>0){
1128 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1129 k--;
1130 }
1131 for(;k<j;k++)
1132 {
1133 assert(r < 64);
1134 if((unneeded_reg[i+k]>>r)&1) return hr;
1135 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1136 {
1137 if(cinfo[i+k].ba>=start && cinfo[i+k].ba<(start+i*4))
1138 {
1139 int t=(cinfo[i+k].ba-start)>>2;
1140 int reg=get_reg(regs[t].regmap_entry,r);
1141 if(reg>=0) return reg;
1142 //reg=get_reg(regs[t+1].regmap_entry,r);
1143 //if(reg>=0) return reg;
1144 }
1145 }
1146 }
1147 return hr;
1148}
1149
1150
1151// Allocate every register, preserving source/target regs
1152static void alloc_all(struct regstat *cur,int i)
1153{
1154 int hr;
1155
1156 for(hr=0;hr<HOST_REGS;hr++) {
1157 if(hr!=EXCLUDE_REG) {
1158 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1159 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1160 {
1161 cur->regmap[hr]=-1;
1162 cur->dirty&=~(1<<hr);
1163 }
1164 // Don't need zeros
1165 if(cur->regmap[hr]==0)
1166 {
1167 cur->regmap[hr]=-1;
1168 cur->dirty&=~(1<<hr);
1169 }
1170 }
1171 }
1172}
1173
1174#ifndef NDEBUG
1175static int host_tempreg_in_use;
1176
1177static void host_tempreg_acquire(void)
1178{
1179 assert(!host_tempreg_in_use);
1180 host_tempreg_in_use = 1;
1181}
1182
1183static void host_tempreg_release(void)
1184{
1185 host_tempreg_in_use = 0;
1186}
1187#else
1188static void host_tempreg_acquire(void) {}
1189static void host_tempreg_release(void) {}
1190#endif
1191
1192#ifdef ASSEM_PRINT
1193extern void gen_interupt();
1194extern void do_insn_cmp();
1195#define FUNCNAME(f) { f, " " #f }
1196static const struct {
1197 void *addr;
1198 const char *name;
1199} function_names[] = {
1200 FUNCNAME(cc_interrupt),
1201 FUNCNAME(gen_interupt),
1202 FUNCNAME(ndrc_get_addr_ht),
1203 FUNCNAME(jump_handler_read8),
1204 FUNCNAME(jump_handler_read16),
1205 FUNCNAME(jump_handler_read32),
1206 FUNCNAME(jump_handler_write8),
1207 FUNCNAME(jump_handler_write16),
1208 FUNCNAME(jump_handler_write32),
1209 FUNCNAME(ndrc_write_invalidate_one),
1210 FUNCNAME(ndrc_write_invalidate_many),
1211 FUNCNAME(jump_to_new_pc),
1212 FUNCNAME(jump_break),
1213 FUNCNAME(jump_break_ds),
1214 FUNCNAME(jump_syscall),
1215 FUNCNAME(jump_syscall_ds),
1216 FUNCNAME(jump_overflow),
1217 FUNCNAME(jump_overflow_ds),
1218 FUNCNAME(jump_addrerror),
1219 FUNCNAME(jump_addrerror_ds),
1220 FUNCNAME(call_gteStall),
1221 FUNCNAME(new_dyna_leave),
1222 FUNCNAME(pcsx_mtc0),
1223 FUNCNAME(pcsx_mtc0_ds),
1224 FUNCNAME(execI),
1225#ifdef __aarch64__
1226 FUNCNAME(do_memhandler_pre),
1227 FUNCNAME(do_memhandler_post),
1228#endif
1229#ifdef DRC_DBG
1230 FUNCNAME(do_insn_cmp),
1231#endif
1232};
1233
1234static const char *func_name(const void *a)
1235{
1236 int i;
1237 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1238 if (function_names[i].addr == a)
1239 return function_names[i].name;
1240 return "";
1241}
1242
1243static const char *fpofs_name(u_int ofs)
1244{
1245 u_int *p = (u_int *)&dynarec_local + ofs/sizeof(u_int);
1246 static char buf[64];
1247 switch (ofs) {
1248 #define ofscase(x) case LO_##x: return " ; " #x
1249 ofscase(next_interupt);
1250 ofscase(last_count);
1251 ofscase(pending_exception);
1252 ofscase(stop);
1253 ofscase(address);
1254 ofscase(lo);
1255 ofscase(hi);
1256 ofscase(PC);
1257 ofscase(cycle);
1258 ofscase(mem_rtab);
1259 ofscase(mem_wtab);
1260 ofscase(psxH_ptr);
1261 ofscase(invc_ptr);
1262 ofscase(ram_offset);
1263 #undef ofscase
1264 }
1265 buf[0] = 0;
1266 if (psxRegs.GPR.r <= p && p < &psxRegs.GPR.r[32])
1267 snprintf(buf, sizeof(buf), " ; r%d", (int)(p - psxRegs.GPR.r));
1268 else if (psxRegs.CP0.r <= p && p < &psxRegs.CP0.r[32])
1269 snprintf(buf, sizeof(buf), " ; cp0 $%d", (int)(p - psxRegs.CP0.r));
1270 else if (psxRegs.CP2D.r <= p && p < &psxRegs.CP2D.r[32])
1271 snprintf(buf, sizeof(buf), " ; cp2d $%d", (int)(p - psxRegs.CP2D.r));
1272 else if (psxRegs.CP2C.r <= p && p < &psxRegs.CP2C.r[32])
1273 snprintf(buf, sizeof(buf), " ; cp2c $%d", (int)(p - psxRegs.CP2C.r));
1274 return buf;
1275}
1276#else
1277#define func_name(x) ""
1278#define fpofs_name(x) ""
1279#endif
1280
1281#ifdef __i386__
1282#include "assem_x86.c"
1283#endif
1284#ifdef __x86_64__
1285#include "assem_x64.c"
1286#endif
1287#ifdef __arm__
1288#include "assem_arm.c"
1289#endif
1290#ifdef __aarch64__
1291#include "assem_arm64.c"
1292#endif
1293
1294static void *get_trampoline(const void *f)
1295{
1296 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1297 size_t i;
1298
1299 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1300 if (tramp->f[i] == f || tramp->f[i] == NULL)
1301 break;
1302 }
1303 if (i == ARRAY_SIZE(tramp->f)) {
1304 SysPrintf("trampoline table is full, last func %p\n", f);
1305 abort();
1306 }
1307 if (tramp->f[i] == NULL) {
1308 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1309 tramp->f[i] = f;
1310 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1311#ifdef HAVE_LIBNX
1312 // invalidate the RX mirror (unsure if necessary, but just in case...)
1313 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1314#endif
1315 }
1316 return &ndrc->tramp.ops[i];
1317}
1318
1319static void emit_far_jump(const void *f)
1320{
1321 if (can_jump_or_call(f)) {
1322 emit_jmp(f);
1323 return;
1324 }
1325
1326 f = get_trampoline(f);
1327 emit_jmp(f);
1328}
1329
1330static void emit_far_call(const void *f)
1331{
1332 if (can_jump_or_call(f)) {
1333 emit_call(f);
1334 return;
1335 }
1336
1337 f = get_trampoline(f);
1338 emit_call(f);
1339}
1340
1341// Check if an address is already compiled
1342// but don't return addresses which are about to expire from the cache
1343static void *check_addr(u_int vaddr)
1344{
1345 struct ht_entry *ht_bin = hash_table_get(vaddr);
1346 size_t i;
1347 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1348 if (ht_bin->vaddr[i] == vaddr)
1349 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1350 return ht_bin->tcaddr[i];
1351 }
1352
1353 // refactor to get_addr_nocompile?
1354 u_int start_page = get_page_prev(vaddr);
1355 u_int page, end_page = get_page(vaddr);
1356
1357 stat_inc(stat_jump_in_lookups);
1358 for (page = start_page; page <= end_page; page++) {
1359 const struct block_info *block;
1360 for (block = blocks[page]; block != NULL; block = block->next) {
1361 if (vaddr < block->start)
1362 break;
1363 if (block->is_dirty || vaddr >= block->start + block->len)
1364 continue;
1365 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1366 continue;
1367 for (i = 0; i < block->jump_in_cnt; i++)
1368 if (block->jump_in[i].vaddr == vaddr)
1369 break;
1370 if (i == block->jump_in_cnt)
1371 continue;
1372
1373 // Update existing entry with current address
1374 void *addr = block->jump_in[i].addr;
1375 if (ht_bin->vaddr[0] == vaddr) {
1376 ht_bin->tcaddr[0] = addr;
1377 return addr;
1378 }
1379 if (ht_bin->vaddr[1] == vaddr) {
1380 ht_bin->tcaddr[1] = addr;
1381 return addr;
1382 }
1383 // Insert into hash table with low priority.
1384 // Don't evict existing entries, as they are probably
1385 // addresses that are being accessed frequently.
1386 if (ht_bin->vaddr[0] == -1) {
1387 ht_bin->vaddr[0] = vaddr;
1388 ht_bin->tcaddr[0] = addr;
1389 }
1390 else if (ht_bin->vaddr[1] == -1) {
1391 ht_bin->vaddr[1] = vaddr;
1392 ht_bin->tcaddr[1] = addr;
1393 }
1394 return addr;
1395 }
1396 }
1397 return NULL;
1398}
1399
1400static void blocks_clear(struct block_info **head)
1401{
1402 struct block_info *cur, *next;
1403
1404 if ((cur = *head)) {
1405 *head = NULL;
1406 while (cur) {
1407 next = cur->next;
1408 free(cur);
1409 cur = next;
1410 }
1411 }
1412}
1413
1414static int blocks_remove_matching_addrs(struct block_info **head,
1415 u_int base_offs, int shift)
1416{
1417 struct block_info *next;
1418 int hit = 0;
1419 while (*head) {
1420 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1421 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1422 invalidate_block(*head);
1423 next = (*head)->next;
1424 free(*head);
1425 *head = next;
1426 stat_dec(stat_blocks);
1427 hit = 1;
1428 }
1429 else
1430 {
1431 head = &((*head)->next);
1432 }
1433 }
1434 return hit;
1435}
1436
1437// This is called when we write to a compiled block (see do_invstub)
1438static void unlink_jumps_vaddr_range(u_int start, u_int end)
1439{
1440 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1441 int i;
1442
1443 for (page = start_page; page <= end_page; page++) {
1444 struct jump_info *ji = jumps[page];
1445 if (ji == NULL)
1446 continue;
1447 for (i = 0; i < ji->count; ) {
1448 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1449 i++;
1450 continue;
1451 }
1452
1453 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1454 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1455 void *host_addr = find_extjump_insn(ji->e[i].stub);
1456 mark_clear_cache(host_addr);
1457 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1458
1459 stat_dec(stat_links);
1460 ji->count--;
1461 if (i < ji->count) {
1462 ji->e[i] = ji->e[ji->count];
1463 continue;
1464 }
1465 i++;
1466 }
1467 }
1468}
1469
1470static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1471{
1472 int i;
1473 if (ji == NULL)
1474 return;
1475 for (i = 0; i < ji->count; ) {
1476 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1477 if (((tc_offs ^ base_offs) >> shift) != 0) {
1478 i++;
1479 continue;
1480 }
1481
1482 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1483 stat_dec(stat_links);
1484 ji->count--;
1485 if (i < ji->count) {
1486 ji->e[i] = ji->e[ji->count];
1487 continue;
1488 }
1489 i++;
1490 }
1491}
1492
1493static void invalidate_block(struct block_info *block)
1494{
1495 u_int i;
1496
1497 block->is_dirty = 1;
1498 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1499 for (i = 0; i < block->jump_in_cnt; i++)
1500 hash_table_remove(block->jump_in[i].vaddr);
1501}
1502
1503static int invalidate_range(u_int start, u_int end,
1504 u32 *inv_start_ret, u32 *inv_end_ret)
1505{
1506 struct block_info *last_block = NULL;
1507 u_int start_page = get_page_prev(start);
1508 u_int end_page = get_page(end - 1);
1509 u_int start_m = pmmask(start);
1510 u_int end_m = pmmask(end - 1);
1511 u_int inv_start, inv_end;
1512 u_int blk_start_m, blk_end_m;
1513 u_int page;
1514 int hit = 0;
1515
1516 // additional area without code (to supplement invalid_code[]), [start, end)
1517 // avoids excessive ndrc_write_invalidate*() calls
1518 inv_start = start_m & ~0xfff;
1519 inv_end = end_m | 0xfff;
1520
1521 for (page = start_page; page <= end_page; page++) {
1522 struct block_info *block;
1523 for (block = blocks[page]; block != NULL; block = block->next) {
1524 if (block->is_dirty)
1525 continue;
1526 last_block = block;
1527 blk_end_m = pmmask(block->start + block->len);
1528 if (blk_end_m <= start_m) {
1529 inv_start = max(inv_start, blk_end_m);
1530 continue;
1531 }
1532 blk_start_m = pmmask(block->start);
1533 if (end_m <= blk_start_m) {
1534 inv_end = min(inv_end, blk_start_m - 1);
1535 continue;
1536 }
1537 if (!block->source) // "hack" block - leave it alone
1538 continue;
1539
1540 hit++;
1541 invalidate_block(block);
1542 stat_inc(stat_inv_hits);
1543 }
1544 }
1545
1546 if (!hit && last_block && last_block->source) {
1547 // could be some leftover unused block, uselessly trapping writes
1548 last_block->inv_near_misses++;
1549 if (last_block->inv_near_misses > 128) {
1550 invalidate_block(last_block);
1551 stat_inc(stat_inv_hits);
1552 hit++;
1553 }
1554 }
1555 if (hit) {
1556 do_clear_cache();
1557#ifdef USE_MINI_HT
1558 memset(mini_ht, -1, sizeof(mini_ht));
1559#endif
1560 }
1561
1562 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1563 // the whole page is empty now
1564 mark_invalid_code(start, 1, 1);
1565
1566 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1567 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1568 return hit;
1569}
1570
1571void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1572{
1573 invalidate_range(start, end, NULL, NULL);
1574}
1575
1576static void ndrc_write_invalidate_many(u_int start, u_int end)
1577{
1578 // this check is done by the caller
1579 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1580 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1581#ifdef INV_DEBUG_W
1582 int invc = invalid_code[start >> 12];
1583 u_int len = end - start;
1584 if (ret)
1585 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1586 else
1587 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1588 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1589 check_for_block_changes(start, end);
1590#endif
1591 stat_inc(stat_inv_addr_calls);
1592 (void)ret;
1593}
1594
1595void ndrc_write_invalidate_one(u_int addr)
1596{
1597 ndrc_write_invalidate_many(addr, addr + 4);
1598}
1599
1600// This is called when loading a save state.
1601// Anything could have changed, so invalidate everything.
1602void new_dynarec_invalidate_all_pages(void)
1603{
1604 struct block_info *block;
1605 u_int page;
1606 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1607 for (block = blocks[page]; block != NULL; block = block->next) {
1608 if (block->is_dirty)
1609 continue;
1610 if (!block->source) // hack block?
1611 continue;
1612 invalidate_block(block);
1613 }
1614 }
1615
1616 #ifdef USE_MINI_HT
1617 memset(mini_ht, -1, sizeof(mini_ht));
1618 #endif
1619 do_clear_cache();
1620}
1621
1622// Add an entry to jump_out after making a link
1623// src should point to code by emit_extjump()
1624void ndrc_add_jump_out(u_int vaddr, void *src)
1625{
1626 inv_debug("ndrc_add_jump_out: %p -> %x\n", src, vaddr);
1627 u_int page = get_page(vaddr);
1628 struct jump_info *ji;
1629
1630 stat_inc(stat_links);
1631 check_extjump2(src);
1632 ji = jumps[page];
1633 if (ji == NULL) {
1634 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1635 ji->alloc = 16;
1636 ji->count = 0;
1637 }
1638 else if (ji->count >= ji->alloc) {
1639 ji->alloc += 16;
1640 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1641 }
1642 jumps[page] = ji;
1643 ji->e[ji->count].target_vaddr = vaddr;
1644 ji->e[ji->count].stub = src;
1645 ji->count++;
1646}
1647
1648/* Register allocation */
1649
1650// Note: registers are allocated clean (unmodified state)
1651// if you intend to modify the register, you must call dirty_reg().
1652static void alloc_reg(struct regstat *cur,int i,signed char reg)
1653{
1654 int r,hr;
1655 int preferred_reg = PREFERRED_REG_FIRST
1656 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1657 if (reg == CCREG) preferred_reg = HOST_CCREG;
1658 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1659 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1660 assert(reg >= 0);
1661
1662 // Don't allocate unused registers
1663 if((cur->u>>reg)&1) return;
1664
1665 // see if it's already allocated
1666 if (get_reg(cur->regmap, reg) >= 0)
1667 return;
1668
1669 // Keep the same mapping if the register was already allocated in a loop
1670 preferred_reg = loop_reg(i,reg,preferred_reg);
1671
1672 // Try to allocate the preferred register
1673 if(cur->regmap[preferred_reg]==-1) {
1674 cur->regmap[preferred_reg]=reg;
1675 cur->dirty&=~(1<<preferred_reg);
1676 cur->isconst&=~(1<<preferred_reg);
1677 return;
1678 }
1679 r=cur->regmap[preferred_reg];
1680 assert(r < 64);
1681 if((cur->u>>r)&1) {
1682 cur->regmap[preferred_reg]=reg;
1683 cur->dirty&=~(1<<preferred_reg);
1684 cur->isconst&=~(1<<preferred_reg);
1685 return;
1686 }
1687
1688 // Clear any unneeded registers
1689 // We try to keep the mapping consistent, if possible, because it
1690 // makes branches easier (especially loops). So we try to allocate
1691 // first (see above) before removing old mappings. If this is not
1692 // possible then go ahead and clear out the registers that are no
1693 // longer needed.
1694 for(hr=0;hr<HOST_REGS;hr++)
1695 {
1696 r=cur->regmap[hr];
1697 if(r>=0) {
1698 assert(r < 64);
1699 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1700 }
1701 }
1702
1703 // Try to allocate any available register, but prefer
1704 // registers that have not been used recently.
1705 if (i > 0) {
1706 for (hr = PREFERRED_REG_FIRST; ; ) {
1707 if (cur->regmap[hr] < 0) {
1708 int oldreg = regs[i-1].regmap[hr];
1709 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1710 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1711 {
1712 cur->regmap[hr]=reg;
1713 cur->dirty&=~(1<<hr);
1714 cur->isconst&=~(1<<hr);
1715 return;
1716 }
1717 }
1718 hr++;
1719 if (hr == EXCLUDE_REG)
1720 hr++;
1721 if (hr == HOST_REGS)
1722 hr = 0;
1723 if (hr == PREFERRED_REG_FIRST)
1724 break;
1725 }
1726 }
1727
1728 // Try to allocate any available register
1729 for (hr = PREFERRED_REG_FIRST; ; ) {
1730 if (cur->regmap[hr] < 0) {
1731 cur->regmap[hr]=reg;
1732 cur->dirty&=~(1<<hr);
1733 cur->isconst&=~(1<<hr);
1734 return;
1735 }
1736 hr++;
1737 if (hr == EXCLUDE_REG)
1738 hr++;
1739 if (hr == HOST_REGS)
1740 hr = 0;
1741 if (hr == PREFERRED_REG_FIRST)
1742 break;
1743 }
1744
1745 // Ok, now we have to evict someone
1746 // Pick a register we hopefully won't need soon
1747 u_char hsn[MAXREG+1];
1748 memset(hsn,10,sizeof(hsn));
1749 int j;
1750 lsn(hsn,i,&preferred_reg);
1751 //printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",cur->regmap[0],cur->regmap[1],cur->regmap[2],cur->regmap[3],cur->regmap[5],cur->regmap[6],cur->regmap[7]);
1752 //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]);
1753 if(i>0) {
1754 // Don't evict the cycle count at entry points, otherwise the entry
1755 // stub will have to write it.
1756 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1757 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1758 for(j=10;j>=3;j--)
1759 {
1760 // Alloc preferred register if available
1761 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1762 for(hr=0;hr<HOST_REGS;hr++) {
1763 // Evict both parts of a 64-bit register
1764 if(cur->regmap[hr]==r) {
1765 cur->regmap[hr]=-1;
1766 cur->dirty&=~(1<<hr);
1767 cur->isconst&=~(1<<hr);
1768 }
1769 }
1770 cur->regmap[preferred_reg]=reg;
1771 return;
1772 }
1773 for(r=1;r<=MAXREG;r++)
1774 {
1775 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1776 for(hr=0;hr<HOST_REGS;hr++) {
1777 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1778 if(cur->regmap[hr]==r) {
1779 cur->regmap[hr]=reg;
1780 cur->dirty&=~(1<<hr);
1781 cur->isconst&=~(1<<hr);
1782 return;
1783 }
1784 }
1785 }
1786 }
1787 }
1788 }
1789 }
1790 for(j=10;j>=0;j--)
1791 {
1792 for(r=1;r<=MAXREG;r++)
1793 {
1794 if(hsn[r]==j) {
1795 for(hr=0;hr<HOST_REGS;hr++) {
1796 if(cur->regmap[hr]==r) {
1797 cur->regmap[hr]=reg;
1798 cur->dirty&=~(1<<hr);
1799 cur->isconst&=~(1<<hr);
1800 return;
1801 }
1802 }
1803 }
1804 }
1805 }
1806 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1807}
1808
1809// Allocate a temporary register. This is done without regard to
1810// dirty status or whether the register we request is on the unneeded list
1811// Note: This will only allocate one register, even if called multiple times
1812static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1813{
1814 int r,hr;
1815 int preferred_reg = -1;
1816
1817 // see if it's already allocated
1818 for(hr=0;hr<HOST_REGS;hr++)
1819 {
1820 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1821 }
1822
1823 // Try to allocate any available register
1824 for(hr=HOST_REGS-1;hr>=0;hr--) {
1825 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1826 cur->regmap[hr]=reg;
1827 cur->dirty&=~(1<<hr);
1828 cur->isconst&=~(1<<hr);
1829 return;
1830 }
1831 }
1832
1833 // Find an unneeded register
1834 for(hr=HOST_REGS-1;hr>=0;hr--)
1835 {
1836 r=cur->regmap[hr];
1837 if(r>=0) {
1838 assert(r < 64);
1839 if((cur->u>>r)&1) {
1840 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1841 cur->regmap[hr]=reg;
1842 cur->dirty&=~(1<<hr);
1843 cur->isconst&=~(1<<hr);
1844 return;
1845 }
1846 }
1847 }
1848 }
1849
1850 // Ok, now we have to evict someone
1851 // Pick a register we hopefully won't need soon
1852 // TODO: we might want to follow unconditional jumps here
1853 // TODO: get rid of dupe code and make this into a function
1854 u_char hsn[MAXREG+1];
1855 memset(hsn,10,sizeof(hsn));
1856 int j;
1857 lsn(hsn,i,&preferred_reg);
1858 //printf("hsn: %d %d %d %d %d %d %d\n",hsn[cur->regmap[0]&63],hsn[cur->regmap[1]&63],hsn[cur->regmap[2]&63],hsn[cur->regmap[3]&63],hsn[cur->regmap[5]&63],hsn[cur->regmap[6]&63],hsn[cur->regmap[7]&63]);
1859 if(i>0) {
1860 // Don't evict the cycle count at entry points, otherwise the entry
1861 // stub will have to write it.
1862 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1863 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1864 for(j=10;j>=3;j--)
1865 {
1866 for(r=1;r<=MAXREG;r++)
1867 {
1868 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1869 for(hr=0;hr<HOST_REGS;hr++) {
1870 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1871 if(cur->regmap[hr]==r) {
1872 cur->regmap[hr]=reg;
1873 cur->dirty&=~(1<<hr);
1874 cur->isconst&=~(1<<hr);
1875 return;
1876 }
1877 }
1878 }
1879 }
1880 }
1881 }
1882 }
1883 for(j=10;j>=0;j--)
1884 {
1885 for(r=1;r<=MAXREG;r++)
1886 {
1887 if(hsn[r]==j) {
1888 for(hr=0;hr<HOST_REGS;hr++) {
1889 if(cur->regmap[hr]==r) {
1890 cur->regmap[hr]=reg;
1891 cur->dirty&=~(1<<hr);
1892 cur->isconst&=~(1<<hr);
1893 return;
1894 }
1895 }
1896 }
1897 }
1898 }
1899 SysPrintf("This shouldn't happen");abort();
1900}
1901
1902static void mov_alloc(struct regstat *current,int i)
1903{
1904 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1905 alloc_cc(current,i); // for stalls
1906 dirty_reg(current,CCREG);
1907 }
1908
1909 // Note: Don't need to actually alloc the source registers
1910 //alloc_reg(current,i,dops[i].rs1);
1911 alloc_reg(current,i,dops[i].rt1);
1912
1913 clear_const(current,dops[i].rs1);
1914 clear_const(current,dops[i].rt1);
1915 dirty_reg(current,dops[i].rt1);
1916}
1917
1918static void shiftimm_alloc(struct regstat *current,int i)
1919{
1920 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1921 {
1922 if(dops[i].rt1) {
1923 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1924 else dops[i].use_lt1=!!dops[i].rs1;
1925 alloc_reg(current,i,dops[i].rt1);
1926 dirty_reg(current,dops[i].rt1);
1927 if(is_const(current,dops[i].rs1)) {
1928 int v=get_const(current,dops[i].rs1);
1929 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<cinfo[i].imm);
1930 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>cinfo[i].imm);
1931 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>cinfo[i].imm);
1932 }
1933 else clear_const(current,dops[i].rt1);
1934 }
1935 }
1936 else
1937 {
1938 clear_const(current,dops[i].rs1);
1939 clear_const(current,dops[i].rt1);
1940 }
1941
1942 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1943 {
1944 assert(0);
1945 }
1946 if(dops[i].opcode2==0x3c) // DSLL32
1947 {
1948 assert(0);
1949 }
1950 if(dops[i].opcode2==0x3e) // DSRL32
1951 {
1952 assert(0);
1953 }
1954 if(dops[i].opcode2==0x3f) // DSRA32
1955 {
1956 assert(0);
1957 }
1958}
1959
1960static void shift_alloc(struct regstat *current,int i)
1961{
1962 if(dops[i].rt1) {
1963 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1964 {
1965 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1966 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1967 alloc_reg(current,i,dops[i].rt1);
1968 if(dops[i].rt1==dops[i].rs2) {
1969 alloc_reg_temp(current,i,-1);
1970 cinfo[i].min_free_regs=1;
1971 }
1972 } else { // DSLLV/DSRLV/DSRAV
1973 assert(0);
1974 }
1975 clear_const(current,dops[i].rs1);
1976 clear_const(current,dops[i].rs2);
1977 clear_const(current,dops[i].rt1);
1978 dirty_reg(current,dops[i].rt1);
1979 }
1980}
1981
1982static void alu_alloc(struct regstat *current,int i)
1983{
1984 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1985 if(dops[i].rt1) {
1986 if(dops[i].rs1&&dops[i].rs2) {
1987 alloc_reg(current,i,dops[i].rs1);
1988 alloc_reg(current,i,dops[i].rs2);
1989 }
1990 else {
1991 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1992 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1993 }
1994 alloc_reg(current,i,dops[i].rt1);
1995 }
1996 if (dops[i].may_except) {
1997 alloc_cc(current, i); // for exceptions
1998 alloc_reg_temp(current, i, -1);
1999 cinfo[i].min_free_regs = 1;
2000 }
2001 }
2002 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2003 if(dops[i].rt1) {
2004 alloc_reg(current,i,dops[i].rs1);
2005 alloc_reg(current,i,dops[i].rs2);
2006 alloc_reg(current,i,dops[i].rt1);
2007 }
2008 }
2009 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2010 if(dops[i].rt1) {
2011 if(dops[i].rs1&&dops[i].rs2) {
2012 alloc_reg(current,i,dops[i].rs1);
2013 alloc_reg(current,i,dops[i].rs2);
2014 }
2015 else
2016 {
2017 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2018 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
2019 }
2020 alloc_reg(current,i,dops[i].rt1);
2021 }
2022 }
2023 clear_const(current,dops[i].rs1);
2024 clear_const(current,dops[i].rs2);
2025 clear_const(current,dops[i].rt1);
2026 dirty_reg(current,dops[i].rt1);
2027}
2028
2029static void imm16_alloc(struct regstat *current,int i)
2030{
2031 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2032 else dops[i].use_lt1=!!dops[i].rs1;
2033 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
2034 if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2035 clear_const(current,dops[i].rs1);
2036 clear_const(current,dops[i].rt1);
2037 }
2038 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2039 if(is_const(current,dops[i].rs1)) {
2040 int v=get_const(current,dops[i].rs1);
2041 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&cinfo[i].imm);
2042 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|cinfo[i].imm);
2043 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^cinfo[i].imm);
2044 }
2045 else clear_const(current,dops[i].rt1);
2046 }
2047 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2048 if(is_const(current,dops[i].rs1)) {
2049 int v=get_const(current,dops[i].rs1);
2050 set_const(current,dops[i].rt1,v+cinfo[i].imm);
2051 }
2052 else clear_const(current,dops[i].rt1);
2053 if (dops[i].may_except) {
2054 alloc_cc(current, i); // for exceptions
2055 alloc_reg_temp(current, i, -1);
2056 cinfo[i].min_free_regs = 1;
2057 }
2058 }
2059 else {
2060 set_const(current,dops[i].rt1,cinfo[i].imm<<16); // LUI
2061 }
2062 dirty_reg(current,dops[i].rt1);
2063}
2064
2065static void load_alloc(struct regstat *current,int i)
2066{
2067 int need_temp = 0;
2068 clear_const(current,dops[i].rt1);
2069 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2070 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2071 if (needed_again(dops[i].rs1, i))
2072 alloc_reg(current, i, dops[i].rs1);
2073 if (ram_offset)
2074 alloc_reg(current, i, ROREG);
2075 if (dops[i].may_except) {
2076 alloc_cc(current, i); // for exceptions
2077 dirty_reg(current, CCREG);
2078 need_temp = 1;
2079 }
2080 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2081 alloc_reg(current,i,dops[i].rt1);
2082 assert(get_reg_w(current->regmap, dops[i].rt1)>=0);
2083 dirty_reg(current,dops[i].rt1);
2084 // LWL/LWR need a temporary register for the old value
2085 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2086 {
2087 alloc_reg(current,i,FTEMP);
2088 need_temp = 1;
2089 }
2090 }
2091 else
2092 {
2093 // Load to r0 or unneeded register (dummy load)
2094 // but we still need a register to calculate the address
2095 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2096 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2097 need_temp = 1;
2098 }
2099 if (need_temp) {
2100 alloc_reg_temp(current, i, -1);
2101 cinfo[i].min_free_regs = 1;
2102 }
2103}
2104
2105static void store_alloc(struct regstat *current,int i)
2106{
2107 clear_const(current,dops[i].rs2);
2108 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2109 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2110 alloc_reg(current,i,dops[i].rs2);
2111 if (ram_offset)
2112 alloc_reg(current, i, ROREG);
2113 #if defined(HOST_IMM8)
2114 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2115 alloc_reg(current, i, INVCP);
2116 #endif
2117 if (dops[i].opcode == 0x2a || dops[i].opcode == 0x2e) { // SWL/SWL
2118 alloc_reg(current,i,FTEMP);
2119 }
2120 if (dops[i].may_except) {
2121 alloc_cc(current, i); // for exceptions
2122 dirty_reg(current, CCREG);
2123 }
2124 // We need a temporary register for address generation
2125 alloc_reg_temp(current,i,-1);
2126 cinfo[i].min_free_regs=1;
2127}
2128
2129static void c2ls_alloc(struct regstat *current,int i)
2130{
2131 clear_const(current,dops[i].rt1);
2132 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2133 alloc_reg(current,i,FTEMP);
2134 if (ram_offset)
2135 alloc_reg(current, i, ROREG);
2136 #if defined(HOST_IMM8)
2137 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2138 if (dops[i].opcode == 0x3a) // SWC2
2139 alloc_reg(current,i,INVCP);
2140 #endif
2141 if (dops[i].may_except) {
2142 alloc_cc(current, i); // for exceptions
2143 dirty_reg(current, CCREG);
2144 }
2145 // We need a temporary register for address generation
2146 alloc_reg_temp(current,i,-1);
2147 cinfo[i].min_free_regs=1;
2148}
2149
2150#ifndef multdiv_alloc
2151static void multdiv_alloc(struct regstat *current,int i)
2152{
2153 // case 0x18: MULT
2154 // case 0x19: MULTU
2155 // case 0x1A: DIV
2156 // case 0x1B: DIVU
2157 // case 0x1C: DMULT
2158 // case 0x1D: DMULTU
2159 // case 0x1E: DDIV
2160 // case 0x1F: DDIVU
2161 clear_const(current,dops[i].rs1);
2162 clear_const(current,dops[i].rs2);
2163 alloc_cc(current,i); // for stalls
2164 if(dops[i].rs1&&dops[i].rs2)
2165 {
2166 if((dops[i].opcode2&4)==0) // 32-bit
2167 {
2168 current->u&=~(1LL<<HIREG);
2169 current->u&=~(1LL<<LOREG);
2170 alloc_reg(current,i,HIREG);
2171 alloc_reg(current,i,LOREG);
2172 alloc_reg(current,i,dops[i].rs1);
2173 alloc_reg(current,i,dops[i].rs2);
2174 dirty_reg(current,HIREG);
2175 dirty_reg(current,LOREG);
2176 }
2177 else // 64-bit
2178 {
2179 assert(0);
2180 }
2181 }
2182 else
2183 {
2184 // Multiply by zero is zero.
2185 // MIPS does not have a divide by zero exception.
2186 // The result is undefined, we return zero.
2187 alloc_reg(current,i,HIREG);
2188 alloc_reg(current,i,LOREG);
2189 dirty_reg(current,HIREG);
2190 dirty_reg(current,LOREG);
2191 }
2192}
2193#endif
2194
2195static void cop0_alloc(struct regstat *current,int i)
2196{
2197 if(dops[i].opcode2==0) // MFC0
2198 {
2199 if(dops[i].rt1) {
2200 clear_const(current,dops[i].rt1);
2201 alloc_reg(current,i,dops[i].rt1);
2202 dirty_reg(current,dops[i].rt1);
2203 }
2204 }
2205 else if(dops[i].opcode2==4) // MTC0
2206 {
2207 if(dops[i].rs1){
2208 clear_const(current,dops[i].rs1);
2209 alloc_reg(current,i,dops[i].rs1);
2210 alloc_all(current,i);
2211 }
2212 else {
2213 alloc_all(current,i); // FIXME: Keep r0
2214 current->u&=~1LL;
2215 alloc_reg(current,i,0);
2216 }
2217 cinfo[i].min_free_regs = HOST_REGS;
2218 }
2219}
2220
2221static void rfe_alloc(struct regstat *current, int i)
2222{
2223 alloc_all(current, i);
2224 cinfo[i].min_free_regs = HOST_REGS;
2225}
2226
2227static void cop2_alloc(struct regstat *current,int i)
2228{
2229 if (dops[i].opcode2 < 3) // MFC2/CFC2
2230 {
2231 alloc_cc(current,i); // for stalls
2232 dirty_reg(current,CCREG);
2233 if(dops[i].rt1){
2234 clear_const(current,dops[i].rt1);
2235 alloc_reg(current,i,dops[i].rt1);
2236 dirty_reg(current,dops[i].rt1);
2237 }
2238 }
2239 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2240 {
2241 if(dops[i].rs1){
2242 clear_const(current,dops[i].rs1);
2243 alloc_reg(current,i,dops[i].rs1);
2244 }
2245 else {
2246 current->u&=~1LL;
2247 alloc_reg(current,i,0);
2248 }
2249 }
2250 alloc_reg_temp(current,i,-1);
2251 cinfo[i].min_free_regs=1;
2252}
2253
2254static void c2op_alloc(struct regstat *current,int i)
2255{
2256 alloc_cc(current,i); // for stalls
2257 dirty_reg(current,CCREG);
2258 alloc_reg_temp(current,i,-1);
2259}
2260
2261static void syscall_alloc(struct regstat *current,int i)
2262{
2263 alloc_cc(current,i);
2264 dirty_reg(current,CCREG);
2265 alloc_all(current,i);
2266 cinfo[i].min_free_regs=HOST_REGS;
2267 current->isconst=0;
2268}
2269
2270static void delayslot_alloc(struct regstat *current,int i)
2271{
2272 switch(dops[i].itype) {
2273 case UJUMP:
2274 case CJUMP:
2275 case SJUMP:
2276 case RJUMP:
2277 case SYSCALL:
2278 case HLECALL:
2279 case IMM16:
2280 imm16_alloc(current,i);
2281 break;
2282 case LOAD:
2283 case LOADLR:
2284 load_alloc(current,i);
2285 break;
2286 case STORE:
2287 case STORELR:
2288 store_alloc(current,i);
2289 break;
2290 case ALU:
2291 alu_alloc(current,i);
2292 break;
2293 case SHIFT:
2294 shift_alloc(current,i);
2295 break;
2296 case MULTDIV:
2297 multdiv_alloc(current,i);
2298 break;
2299 case SHIFTIMM:
2300 shiftimm_alloc(current,i);
2301 break;
2302 case MOV:
2303 mov_alloc(current,i);
2304 break;
2305 case COP0:
2306 cop0_alloc(current,i);
2307 break;
2308 case RFE:
2309 rfe_alloc(current,i);
2310 break;
2311 case COP2:
2312 cop2_alloc(current,i);
2313 break;
2314 case C2LS:
2315 c2ls_alloc(current,i);
2316 break;
2317 case C2OP:
2318 c2op_alloc(current,i);
2319 break;
2320 }
2321}
2322
2323static void add_stub(enum stub_type type, void *addr, void *retaddr,
2324 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2325{
2326 assert(stubcount < ARRAY_SIZE(stubs));
2327 stubs[stubcount].type = type;
2328 stubs[stubcount].addr = addr;
2329 stubs[stubcount].retaddr = retaddr;
2330 stubs[stubcount].a = a;
2331 stubs[stubcount].b = b;
2332 stubs[stubcount].c = c;
2333 stubs[stubcount].d = d;
2334 stubs[stubcount].e = e;
2335 stubcount++;
2336}
2337
2338static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2339 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2340{
2341 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2342}
2343
2344// Write out a single register
2345static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2346{
2347 int hr;
2348 for(hr=0;hr<HOST_REGS;hr++) {
2349 if(hr!=EXCLUDE_REG) {
2350 if(regmap[hr]==r) {
2351 if((dirty>>hr)&1) {
2352 assert(regmap[hr]<64);
2353 emit_storereg(r,hr);
2354 }
2355 }
2356 }
2357 }
2358}
2359
2360static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2361{
2362 //if(dirty_pre==dirty) return;
2363 int hr, r;
2364 for (hr = 0; hr < HOST_REGS; hr++) {
2365 r = pre[hr];
2366 if (r < 1 || r > 33 || ((u >> r) & 1))
2367 continue;
2368 if (((dirty_pre & ~dirty) >> hr) & 1)
2369 emit_storereg(r, hr);
2370 }
2371}
2372
2373// trashes r2
2374static void pass_args(int a0, int a1)
2375{
2376 if(a0==1&&a1==0) {
2377 // must swap
2378 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2379 }
2380 else if(a0!=0&&a1==0) {
2381 emit_mov(a1,1);
2382 if (a0>=0) emit_mov(a0,0);
2383 }
2384 else {
2385 if(a0>=0&&a0!=0) emit_mov(a0,0);
2386 if(a1>=0&&a1!=1) emit_mov(a1,1);
2387 }
2388}
2389
2390static void alu_assemble(int i, const struct regstat *i_regs, int ccadj_)
2391{
2392 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2393 int do_oflow = dops[i].may_except; // ADD/SUB with exceptions enabled
2394 if (dops[i].rt1 || do_oflow) {
2395 int do_exception_check = 0;
2396 signed char s1, s2, t, tmp;
2397 t = get_reg_w(i_regs->regmap, dops[i].rt1);
2398 tmp = get_reg_temp(i_regs->regmap);
2399 if (do_oflow)
2400 assert(tmp >= 0);
2401 //if (t < 0 && do_oflow) // broken s2
2402 // t = tmp;
2403 if (t >= 0) {
2404 s1 = get_reg(i_regs->regmap, dops[i].rs1);
2405 s2 = get_reg(i_regs->regmap, dops[i].rs2);
2406 if (dops[i].rs1 && dops[i].rs2) {
2407 assert(s1>=0);
2408 assert(s2>=0);
2409 if (dops[i].opcode2 & 2) {
2410 if (do_oflow) {
2411 emit_subs(s1, s2, tmp);
2412 do_exception_check = 1;
2413 }
2414 else
2415 emit_sub(s1,s2,t);
2416 }
2417 else {
2418 if (do_oflow) {
2419 emit_adds(s1, s2, tmp);
2420 do_exception_check = 1;
2421 }
2422 else
2423 emit_add(s1,s2,t);
2424 }
2425 }
2426 else if(dops[i].rs1) {
2427 if(s1>=0) emit_mov(s1,t);
2428 else emit_loadreg(dops[i].rs1,t);
2429 }
2430 else if(dops[i].rs2) {
2431 if (s2 < 0) {
2432 emit_loadreg(dops[i].rs2, t);
2433 s2 = t;
2434 }
2435 if (dops[i].opcode2 & 2) {
2436 if (do_oflow) {
2437 emit_negs(s2, tmp);
2438 do_exception_check = 1;
2439 }
2440 else
2441 emit_neg(s2, t);
2442 }
2443 else if (s2 != t)
2444 emit_mov(s2, t);
2445 }
2446 else
2447 emit_zeroreg(t);
2448 }
2449 if (do_exception_check) {
2450 void *jaddr = out;
2451 emit_jo(0);
2452 if (t >= 0 && tmp != t)
2453 emit_mov(tmp, t);
2454 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2455 }
2456 }
2457 }
2458 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2459 if(dops[i].rt1) {
2460 signed char s1l,s2l,t;
2461 {
2462 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2463 //assert(t>=0);
2464 if(t>=0) {
2465 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2466 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2467 if(dops[i].rs2==0) // rx<r0
2468 {
2469 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2470 assert(s1l>=0);
2471 emit_shrimm(s1l,31,t);
2472 }
2473 else // SLTU (unsigned can not be less than zero, 0<0)
2474 emit_zeroreg(t);
2475 }
2476 else if(dops[i].rs1==0) // r0<rx
2477 {
2478 assert(s2l>=0);
2479 if(dops[i].opcode2==0x2a) // SLT
2480 emit_set_gz32(s2l,t);
2481 else // SLTU (set if not zero)
2482 emit_set_nz32(s2l,t);
2483 }
2484 else{
2485 assert(s1l>=0);assert(s2l>=0);
2486 if(dops[i].opcode2==0x2a) // SLT
2487 emit_set_if_less32(s1l,s2l,t);
2488 else // SLTU
2489 emit_set_if_carry32(s1l,s2l,t);
2490 }
2491 }
2492 }
2493 }
2494 }
2495 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2496 if(dops[i].rt1) {
2497 signed char s1l,s2l,tl;
2498 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2499 {
2500 if(tl>=0) {
2501 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2502 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2503 if(dops[i].rs1&&dops[i].rs2) {
2504 assert(s1l>=0);
2505 assert(s2l>=0);
2506 if(dops[i].opcode2==0x24) { // AND
2507 emit_and(s1l,s2l,tl);
2508 } else
2509 if(dops[i].opcode2==0x25) { // OR
2510 emit_or(s1l,s2l,tl);
2511 } else
2512 if(dops[i].opcode2==0x26) { // XOR
2513 emit_xor(s1l,s2l,tl);
2514 } else
2515 if(dops[i].opcode2==0x27) { // NOR
2516 emit_or(s1l,s2l,tl);
2517 emit_not(tl,tl);
2518 }
2519 }
2520 else
2521 {
2522 if(dops[i].opcode2==0x24) { // AND
2523 emit_zeroreg(tl);
2524 } else
2525 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2526 if(dops[i].rs1){
2527 if(s1l>=0) emit_mov(s1l,tl);
2528 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2529 }
2530 else
2531 if(dops[i].rs2){
2532 if(s2l>=0) emit_mov(s2l,tl);
2533 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2534 }
2535 else emit_zeroreg(tl);
2536 } else
2537 if(dops[i].opcode2==0x27) { // NOR
2538 if(dops[i].rs1){
2539 if(s1l>=0) emit_not(s1l,tl);
2540 else {
2541 emit_loadreg(dops[i].rs1,tl);
2542 emit_not(tl,tl);
2543 }
2544 }
2545 else
2546 if(dops[i].rs2){
2547 if(s2l>=0) emit_not(s2l,tl);
2548 else {
2549 emit_loadreg(dops[i].rs2,tl);
2550 emit_not(tl,tl);
2551 }
2552 }
2553 else emit_movimm(-1,tl);
2554 }
2555 }
2556 }
2557 }
2558 }
2559 }
2560}
2561
2562static void imm16_assemble(int i, const struct regstat *i_regs, int ccadj_)
2563{
2564 if (dops[i].opcode==0x0f) { // LUI
2565 if(dops[i].rt1) {
2566 signed char t;
2567 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2568 //assert(t>=0);
2569 if(t>=0) {
2570 if(!((i_regs->isconst>>t)&1))
2571 emit_movimm(cinfo[i].imm<<16,t);
2572 }
2573 }
2574 }
2575 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2576 int is_addi = dops[i].may_except;
2577 if (dops[i].rt1 || is_addi) {
2578 signed char s, t, tmp;
2579 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2580 s=get_reg(i_regs->regmap,dops[i].rs1);
2581 if(dops[i].rs1) {
2582 tmp = get_reg_temp(i_regs->regmap);
2583 if (is_addi) {
2584 assert(tmp >= 0);
2585 if (t < 0) t = tmp;
2586 }
2587 if(t>=0) {
2588 if(!((i_regs->isconst>>t)&1)) {
2589 int sum, do_exception_check = 0;
2590 if (s < 0) {
2591 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2592 if (is_addi) {
2593 emit_addimm_and_set_flags3(t, cinfo[i].imm, tmp);
2594 do_exception_check = 1;
2595 }
2596 else
2597 emit_addimm(t, cinfo[i].imm, t);
2598 } else {
2599 if (!((i_regs->wasconst >> s) & 1)) {
2600 if (is_addi) {
2601 emit_addimm_and_set_flags3(s, cinfo[i].imm, tmp);
2602 do_exception_check = 1;
2603 }
2604 else
2605 emit_addimm(s, cinfo[i].imm, t);
2606 }
2607 else {
2608 int oflow = add_overflow(constmap[i][s], cinfo[i].imm, sum);
2609 if (is_addi && oflow)
2610 do_exception_check = 2;
2611 else
2612 emit_movimm(sum, t);
2613 }
2614 }
2615 if (do_exception_check) {
2616 void *jaddr = out;
2617 if (do_exception_check == 2)
2618 emit_jmp(0);
2619 else {
2620 emit_jo(0);
2621 if (tmp != t)
2622 emit_mov(tmp, t);
2623 }
2624 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2625 }
2626 }
2627 }
2628 } else {
2629 if(t>=0) {
2630 if(!((i_regs->isconst>>t)&1))
2631 emit_movimm(cinfo[i].imm,t);
2632 }
2633 }
2634 }
2635 }
2636 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2637 if(dops[i].rt1) {
2638 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2639 signed char sl,t;
2640 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2641 sl=get_reg(i_regs->regmap,dops[i].rs1);
2642 //assert(t>=0);
2643 if(t>=0) {
2644 if(dops[i].rs1>0) {
2645 if(dops[i].opcode==0x0a) { // SLTI
2646 if(sl<0) {
2647 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2648 emit_slti32(t,cinfo[i].imm,t);
2649 }else{
2650 emit_slti32(sl,cinfo[i].imm,t);
2651 }
2652 }
2653 else { // SLTIU
2654 if(sl<0) {
2655 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2656 emit_sltiu32(t,cinfo[i].imm,t);
2657 }else{
2658 emit_sltiu32(sl,cinfo[i].imm,t);
2659 }
2660 }
2661 }else{
2662 // SLTI(U) with r0 is just stupid,
2663 // nonetheless examples can be found
2664 if(dops[i].opcode==0x0a) // SLTI
2665 if(0<cinfo[i].imm) emit_movimm(1,t);
2666 else emit_zeroreg(t);
2667 else // SLTIU
2668 {
2669 if(cinfo[i].imm) emit_movimm(1,t);
2670 else emit_zeroreg(t);
2671 }
2672 }
2673 }
2674 }
2675 }
2676 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2677 if(dops[i].rt1) {
2678 signed char sl,tl;
2679 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2680 sl=get_reg(i_regs->regmap,dops[i].rs1);
2681 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2682 if(dops[i].opcode==0x0c) //ANDI
2683 {
2684 if(dops[i].rs1) {
2685 if(sl<0) {
2686 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2687 emit_andimm(tl,cinfo[i].imm,tl);
2688 }else{
2689 if(!((i_regs->wasconst>>sl)&1))
2690 emit_andimm(sl,cinfo[i].imm,tl);
2691 else
2692 emit_movimm(constmap[i][sl]&cinfo[i].imm,tl);
2693 }
2694 }
2695 else
2696 emit_zeroreg(tl);
2697 }
2698 else
2699 {
2700 if(dops[i].rs1) {
2701 if(sl<0) {
2702 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2703 }
2704 if(dops[i].opcode==0x0d) { // ORI
2705 if(sl<0) {
2706 emit_orimm(tl,cinfo[i].imm,tl);
2707 }else{
2708 if(!((i_regs->wasconst>>sl)&1))
2709 emit_orimm(sl,cinfo[i].imm,tl);
2710 else
2711 emit_movimm(constmap[i][sl]|cinfo[i].imm,tl);
2712 }
2713 }
2714 if(dops[i].opcode==0x0e) { // XORI
2715 if(sl<0) {
2716 emit_xorimm(tl,cinfo[i].imm,tl);
2717 }else{
2718 if(!((i_regs->wasconst>>sl)&1))
2719 emit_xorimm(sl,cinfo[i].imm,tl);
2720 else
2721 emit_movimm(constmap[i][sl]^cinfo[i].imm,tl);
2722 }
2723 }
2724 }
2725 else {
2726 emit_movimm(cinfo[i].imm,tl);
2727 }
2728 }
2729 }
2730 }
2731 }
2732}
2733
2734static void shiftimm_assemble(int i, const struct regstat *i_regs)
2735{
2736 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2737 {
2738 if(dops[i].rt1) {
2739 signed char s,t;
2740 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2741 s=get_reg(i_regs->regmap,dops[i].rs1);
2742 //assert(t>=0);
2743 if(t>=0&&!((i_regs->isconst>>t)&1)){
2744 if(dops[i].rs1==0)
2745 {
2746 emit_zeroreg(t);
2747 }
2748 else
2749 {
2750 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2751 if(cinfo[i].imm) {
2752 if(dops[i].opcode2==0) // SLL
2753 {
2754 emit_shlimm(s<0?t:s,cinfo[i].imm,t);
2755 }
2756 if(dops[i].opcode2==2) // SRL
2757 {
2758 emit_shrimm(s<0?t:s,cinfo[i].imm,t);
2759 }
2760 if(dops[i].opcode2==3) // SRA
2761 {
2762 emit_sarimm(s<0?t:s,cinfo[i].imm,t);
2763 }
2764 }else{
2765 // Shift by zero
2766 if(s>=0 && s!=t) emit_mov(s,t);
2767 }
2768 }
2769 }
2770 //emit_storereg(dops[i].rt1,t); //DEBUG
2771 }
2772 }
2773 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2774 {
2775 assert(0);
2776 }
2777 if(dops[i].opcode2==0x3c) // DSLL32
2778 {
2779 assert(0);
2780 }
2781 if(dops[i].opcode2==0x3e) // DSRL32
2782 {
2783 assert(0);
2784 }
2785 if(dops[i].opcode2==0x3f) // DSRA32
2786 {
2787 assert(0);
2788 }
2789}
2790
2791#ifndef shift_assemble
2792static void shift_assemble(int i, const struct regstat *i_regs)
2793{
2794 signed char s,t,shift;
2795 if (dops[i].rt1 == 0)
2796 return;
2797 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2798 t = get_reg(i_regs->regmap, dops[i].rt1);
2799 s = get_reg(i_regs->regmap, dops[i].rs1);
2800 shift = get_reg(i_regs->regmap, dops[i].rs2);
2801 if (t < 0)
2802 return;
2803
2804 if(dops[i].rs1==0)
2805 emit_zeroreg(t);
2806 else if(dops[i].rs2==0) {
2807 assert(s>=0);
2808 if(s!=t) emit_mov(s,t);
2809 }
2810 else {
2811 host_tempreg_acquire();
2812 emit_andimm(shift,31,HOST_TEMPREG);
2813 switch(dops[i].opcode2) {
2814 case 4: // SLLV
2815 emit_shl(s,HOST_TEMPREG,t);
2816 break;
2817 case 6: // SRLV
2818 emit_shr(s,HOST_TEMPREG,t);
2819 break;
2820 case 7: // SRAV
2821 emit_sar(s,HOST_TEMPREG,t);
2822 break;
2823 default:
2824 assert(0);
2825 }
2826 host_tempreg_release();
2827 }
2828}
2829
2830#endif
2831
2832enum {
2833 MTYPE_8000 = 0,
2834 MTYPE_8020,
2835 MTYPE_0000,
2836 MTYPE_A000,
2837 MTYPE_1F80,
2838};
2839
2840static int get_ptr_mem_type(u_int a)
2841{
2842 if(a < 0x00200000) {
2843 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2844 // return wrong, must use memhandler for BIOS self-test to pass
2845 // 007 does similar stuff from a00 mirror, weird stuff
2846 return MTYPE_8000;
2847 return MTYPE_0000;
2848 }
2849 if(0x1f800000 <= a && a < 0x1f801000)
2850 return MTYPE_1F80;
2851 if(0x80200000 <= a && a < 0x80800000)
2852 return MTYPE_8020;
2853 if(0xa0000000 <= a && a < 0xa0200000)
2854 return MTYPE_A000;
2855 return MTYPE_8000;
2856}
2857
2858static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2859{
2860 int r = get_reg(i_regs->regmap, ROREG);
2861 if (r < 0 && host_tempreg_free) {
2862 host_tempreg_acquire();
2863 emit_loadreg(ROREG, r = HOST_TEMPREG);
2864 }
2865 if (r < 0)
2866 abort();
2867 return r;
2868}
2869
2870static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2871 int addr, int *offset_reg, int *addr_reg_override, int ccadj_)
2872{
2873 void *jaddr = NULL;
2874 int type = 0;
2875 int mr = dops[i].rs1;
2876 assert(addr >= 0);
2877 *offset_reg = -1;
2878 if(((smrv_strong|smrv_weak)>>mr)&1) {
2879 type=get_ptr_mem_type(smrv[mr]);
2880 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2881 }
2882 else {
2883 // use the mirror we are running on
2884 type=get_ptr_mem_type(start);
2885 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2886 }
2887
2888 if (dops[i].may_except) {
2889 // alignment check
2890 u_int op = dops[i].opcode;
2891 int mask = ((op & 0x37) == 0x21 || op == 0x25) ? 1 : 3; // LH/SH/LHU
2892 void *jaddr;
2893 emit_testimm(addr, mask);
2894 jaddr = out;
2895 emit_jne(0);
2896 add_stub_r(ALIGNMENT_STUB, jaddr, out, i, addr, i_regs, ccadj_, 0);
2897 }
2898
2899 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2900 host_tempreg_acquire();
2901 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2902 addr=*addr_reg_override=HOST_TEMPREG;
2903 type=0;
2904 }
2905 else if(type==MTYPE_0000) { // RAM 0 mirror
2906 host_tempreg_acquire();
2907 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2908 addr=*addr_reg_override=HOST_TEMPREG;
2909 type=0;
2910 }
2911 else if(type==MTYPE_A000) { // RAM A mirror
2912 host_tempreg_acquire();
2913 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2914 addr=*addr_reg_override=HOST_TEMPREG;
2915 type=0;
2916 }
2917 else if(type==MTYPE_1F80) { // scratchpad
2918 if (psxH == (void *)0x1f800000) {
2919 host_tempreg_acquire();
2920 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2921 emit_cmpimm(HOST_TEMPREG,0x1000);
2922 host_tempreg_release();
2923 jaddr=out;
2924 emit_jc(0);
2925 }
2926 else {
2927 // do the usual RAM check, jump will go to the right handler
2928 type=0;
2929 }
2930 }
2931
2932 if (type == 0) // need ram check
2933 {
2934 emit_cmpimm(addr,RAM_SIZE);
2935 jaddr = out;
2936 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2937 // Hint to branch predictor that the branch is unlikely to be taken
2938 if (dops[i].rs1 >= 28)
2939 emit_jno_unlikely(0);
2940 else
2941 #endif
2942 emit_jno(0);
2943 if (ram_offset != 0)
2944 *offset_reg = get_ro_reg(i_regs, 0);
2945 }
2946
2947 return jaddr;
2948}
2949
2950// return memhandler, or get directly accessable address and return 0
2951static void *get_direct_memhandler(void *table, u_int addr,
2952 enum stub_type type, uintptr_t *addr_host)
2953{
2954 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2955 uintptr_t l1, l2 = 0;
2956 l1 = ((uintptr_t *)table)[addr>>12];
2957 if (!(l1 & msb)) {
2958 uintptr_t v = l1 << 1;
2959 *addr_host = v + addr;
2960 return NULL;
2961 }
2962 else {
2963 l1 <<= 1;
2964 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2965 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2966 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2967 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2968 else
2969 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2970 if (!(l2 & msb)) {
2971 uintptr_t v = l2 << 1;
2972 *addr_host = v + (addr&0xfff);
2973 return NULL;
2974 }
2975 return (void *)(l2 << 1);
2976 }
2977}
2978
2979static u_int get_host_reglist(const signed char *regmap)
2980{
2981 u_int reglist = 0, hr;
2982 for (hr = 0; hr < HOST_REGS; hr++) {
2983 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2984 reglist |= 1 << hr;
2985 }
2986 return reglist;
2987}
2988
2989static u_int reglist_exclude(u_int reglist, int r1, int r2)
2990{
2991 if (r1 >= 0)
2992 reglist &= ~(1u << r1);
2993 if (r2 >= 0)
2994 reglist &= ~(1u << r2);
2995 return reglist;
2996}
2997
2998// find a temp caller-saved register not in reglist (so assumed to be free)
2999static int reglist_find_free(u_int reglist)
3000{
3001 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
3002 if (free_regs == 0)
3003 return -1;
3004 return __builtin_ctz(free_regs);
3005}
3006
3007static void do_load_word(int a, int rt, int offset_reg)
3008{
3009 if (offset_reg >= 0)
3010 emit_ldr_dualindexed(offset_reg, a, rt);
3011 else
3012 emit_readword_indexed(0, a, rt);
3013}
3014
3015static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
3016{
3017 if (offset_reg < 0) {
3018 emit_writeword_indexed(rt, ofs, a);
3019 return;
3020 }
3021 if (ofs != 0)
3022 emit_addimm(a, ofs, a);
3023 emit_str_dualindexed(offset_reg, a, rt);
3024 if (ofs != 0 && preseve_a)
3025 emit_addimm(a, -ofs, a);
3026}
3027
3028static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
3029{
3030 if (offset_reg < 0) {
3031 emit_writehword_indexed(rt, ofs, a);
3032 return;
3033 }
3034 if (ofs != 0)
3035 emit_addimm(a, ofs, a);
3036 emit_strh_dualindexed(offset_reg, a, rt);
3037 if (ofs != 0 && preseve_a)
3038 emit_addimm(a, -ofs, a);
3039}
3040
3041static void do_store_byte(int a, int rt, int offset_reg)
3042{
3043 if (offset_reg >= 0)
3044 emit_strb_dualindexed(offset_reg, a, rt);
3045 else
3046 emit_writebyte_indexed(rt, 0, a);
3047}
3048
3049static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
3050{
3051 int addr = cinfo[i].addr;
3052 int s,tl;
3053 int offset;
3054 void *jaddr=0;
3055 int memtarget=0,c=0;
3056 int offset_reg = -1;
3057 int fastio_reg_override = -1;
3058 u_int reglist=get_host_reglist(i_regs->regmap);
3059 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3060 s=get_reg(i_regs->regmap,dops[i].rs1);
3061 offset=cinfo[i].imm;
3062 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3063 if(s>=0) {
3064 c=(i_regs->wasconst>>s)&1;
3065 if (c) {
3066 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3067 }
3068 }
3069 //printf("load_assemble: c=%d\n",c);
3070 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3071 if(tl<0 && ((!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80) || dops[i].rt1==0)) {
3072 // could be FIFO, must perform the read
3073 // ||dummy read
3074 assem_debug("(forced read)\n");
3075 tl = get_reg_temp(i_regs->regmap); // may be == addr
3076 assert(tl>=0);
3077 }
3078 assert(addr >= 0);
3079 if(tl>=0) {
3080 //printf("load_assemble: c=%d\n",c);
3081 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3082 reglist&=~(1<<tl);
3083 if(!c) {
3084 #ifdef R29_HACK
3085 // Strmnnrmn's speed hack
3086 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3087 #endif
3088 {
3089 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3090 &offset_reg, &fastio_reg_override, ccadj_);
3091 }
3092 }
3093 else if (ram_offset && memtarget) {
3094 offset_reg = get_ro_reg(i_regs, 0);
3095 }
3096 int dummy=(dops[i].rt1==0)||(tl!=get_reg_w(i_regs->regmap, dops[i].rt1)); // ignore loads to r0 and unneeded reg
3097 switch (dops[i].opcode) {
3098 case 0x20: // LB
3099 if(!c||memtarget) {
3100 if(!dummy) {
3101 int a = addr;
3102 if (fastio_reg_override >= 0)
3103 a = fastio_reg_override;
3104
3105 if (offset_reg >= 0)
3106 emit_ldrsb_dualindexed(offset_reg, a, tl);
3107 else
3108 emit_movsbl_indexed(0, a, tl);
3109 }
3110 if(jaddr)
3111 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3112 }
3113 else
3114 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3115 break;
3116 case 0x21: // LH
3117 if(!c||memtarget) {
3118 if(!dummy) {
3119 int a = addr;
3120 if (fastio_reg_override >= 0)
3121 a = fastio_reg_override;
3122 if (offset_reg >= 0)
3123 emit_ldrsh_dualindexed(offset_reg, a, tl);
3124 else
3125 emit_movswl_indexed(0, a, tl);
3126 }
3127 if(jaddr)
3128 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3129 }
3130 else
3131 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3132 break;
3133 case 0x23: // LW
3134 if(!c||memtarget) {
3135 if(!dummy) {
3136 int a = addr;
3137 if (fastio_reg_override >= 0)
3138 a = fastio_reg_override;
3139 do_load_word(a, tl, offset_reg);
3140 }
3141 if(jaddr)
3142 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3143 }
3144 else
3145 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3146 break;
3147 case 0x24: // LBU
3148 if(!c||memtarget) {
3149 if(!dummy) {
3150 int a = addr;
3151 if (fastio_reg_override >= 0)
3152 a = fastio_reg_override;
3153
3154 if (offset_reg >= 0)
3155 emit_ldrb_dualindexed(offset_reg, a, tl);
3156 else
3157 emit_movzbl_indexed(0, a, tl);
3158 }
3159 if(jaddr)
3160 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3161 }
3162 else
3163 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3164 break;
3165 case 0x25: // LHU
3166 if(!c||memtarget) {
3167 if(!dummy) {
3168 int a = addr;
3169 if (fastio_reg_override >= 0)
3170 a = fastio_reg_override;
3171 if (offset_reg >= 0)
3172 emit_ldrh_dualindexed(offset_reg, a, tl);
3173 else
3174 emit_movzwl_indexed(0, a, tl);
3175 }
3176 if(jaddr)
3177 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3178 }
3179 else
3180 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3181 break;
3182 default:
3183 assert(0);
3184 }
3185 } // tl >= 0
3186 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3187 host_tempreg_release();
3188}
3189
3190#ifndef loadlr_assemble
3191static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3192{
3193 int addr = cinfo[i].addr;
3194 int s,tl,temp,temp2;
3195 int offset;
3196 void *jaddr=0;
3197 int memtarget=0,c=0;
3198 int offset_reg = -1;
3199 int fastio_reg_override = -1;
3200 u_int reglist=get_host_reglist(i_regs->regmap);
3201 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3202 s=get_reg(i_regs->regmap,dops[i].rs1);
3203 temp=get_reg_temp(i_regs->regmap);
3204 temp2=get_reg(i_regs->regmap,FTEMP);
3205 offset=cinfo[i].imm;
3206 reglist|=1<<temp;
3207 assert(addr >= 0);
3208 if(s>=0) {
3209 c=(i_regs->wasconst>>s)&1;
3210 if(c) {
3211 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3212 }
3213 }
3214 if(!c) {
3215 emit_shlimm(addr,3,temp);
3216 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3217 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3218 }else{
3219 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3220 }
3221 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3222 &offset_reg, &fastio_reg_override, ccadj_);
3223 }
3224 else {
3225 if (ram_offset && memtarget) {
3226 offset_reg = get_ro_reg(i_regs, 0);
3227 }
3228 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3229 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3230 }else{
3231 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3232 }
3233 }
3234 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3235 if(!c||memtarget) {
3236 int a = temp2;
3237 if (fastio_reg_override >= 0)
3238 a = fastio_reg_override;
3239 do_load_word(a, temp2, offset_reg);
3240 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3241 host_tempreg_release();
3242 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3243 }
3244 else
3245 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3246 if(dops[i].rt1) {
3247 assert(tl>=0);
3248 emit_andimm(temp,24,temp);
3249 if (dops[i].opcode==0x22) // LWL
3250 emit_xorimm(temp,24,temp);
3251 host_tempreg_acquire();
3252 emit_movimm(-1,HOST_TEMPREG);
3253 if (dops[i].opcode==0x26) {
3254 emit_shr(temp2,temp,temp2);
3255 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3256 }else{
3257 emit_shl(temp2,temp,temp2);
3258 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3259 }
3260 host_tempreg_release();
3261 emit_or(temp2,tl,tl);
3262 }
3263 //emit_storereg(dops[i].rt1,tl); // DEBUG
3264 }
3265 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3266 assert(0);
3267 }
3268}
3269#endif
3270
3271static void do_invstub(int n)
3272{
3273 literal_pool(20);
3274 assem_debug("do_invstub\n");
3275 u_int reglist = stubs[n].a;
3276 u_int addrr = stubs[n].b;
3277 int ofs_start = stubs[n].c;
3278 int ofs_end = stubs[n].d;
3279 int len = ofs_end - ofs_start;
3280 u_int rightr = 0;
3281
3282 set_jump_target(stubs[n].addr, out);
3283 save_regs(reglist);
3284 if (addrr != 0 || ofs_start != 0)
3285 emit_addimm(addrr, ofs_start, 0);
3286 emit_readword(&inv_code_start, 2);
3287 emit_readword(&inv_code_end, 3);
3288 if (len != 0)
3289 emit_addimm(0, len + 4, (rightr = 1));
3290 emit_cmp(0, 2);
3291 emit_cmpcs(3, rightr);
3292 void *jaddr = out;
3293 emit_jc(0);
3294 void *func = (len != 0)
3295 ? (void *)ndrc_write_invalidate_many
3296 : (void *)ndrc_write_invalidate_one;
3297 emit_far_call(func);
3298 set_jump_target(jaddr, out);
3299 restore_regs(reglist);
3300 emit_jmp(stubs[n].retaddr);
3301}
3302
3303static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3304{
3305 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3306 return;
3307 // this can't be used any more since we started to check exact
3308 // block boundaries in invalidate_range()
3309 //if (i_regs->waswritten & (1<<dops[i].rs1))
3310 // return;
3311 // (naively) assume nobody will run code from stack
3312 if (dops[i].rs1 == 29)
3313 return;
3314
3315 int j, imm_maxdiff = 32, imm_min = cinfo[i].imm, imm_max = cinfo[i].imm, count = 1;
3316 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3317 && abs(cinfo[i+1].imm - cinfo[i].imm) <= imm_maxdiff)
3318 return;
3319 for (j = i - 1; j >= 0; j--) {
3320 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3321 || abs(cinfo[j].imm - cinfo[j+1].imm) > imm_maxdiff)
3322 break;
3323 count++;
3324 if (imm_min > cinfo[j].imm)
3325 imm_min = cinfo[j].imm;
3326 if (imm_max < cinfo[j].imm)
3327 imm_max = cinfo[j].imm;
3328 }
3329#if defined(HOST_IMM8)
3330 int ir = get_reg(i_regs->regmap, INVCP);
3331 assert(ir >= 0);
3332 host_tempreg_acquire();
3333 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3334#else
3335 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3336 #error not handled
3337#endif
3338#ifdef INVALIDATE_USE_COND_CALL
3339 if (count == 1) {
3340 emit_cmpimm(HOST_TEMPREG, 1);
3341 emit_callne(invalidate_addr_reg[addr]);
3342 host_tempreg_release();
3343 return;
3344 }
3345#endif
3346 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3347 host_tempreg_release();
3348 imm_min -= cinfo[i].imm;
3349 imm_max -= cinfo[i].imm;
3350 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3351 addr, imm_min, imm_max, 0);
3352}
3353
3354static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3355{
3356 int s,tl;
3357 int addr = cinfo[i].addr;
3358 int offset;
3359 void *jaddr=0;
3360 enum stub_type type=0;
3361 int memtarget=0,c=0;
3362 int offset_reg = -1;
3363 int fastio_reg_override = -1;
3364 u_int reglist=get_host_reglist(i_regs->regmap);
3365 tl=get_reg(i_regs->regmap,dops[i].rs2);
3366 s=get_reg(i_regs->regmap,dops[i].rs1);
3367 offset=cinfo[i].imm;
3368 if(s>=0) {
3369 c=(i_regs->wasconst>>s)&1;
3370 if(c) {
3371 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3372 }
3373 }
3374 assert(tl>=0);
3375 assert(addr >= 0);
3376 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3377 if (!c) {
3378 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3379 &offset_reg, &fastio_reg_override, ccadj_);
3380 }
3381 else if (ram_offset && memtarget) {
3382 offset_reg = get_ro_reg(i_regs, 0);
3383 }
3384
3385 switch (dops[i].opcode) {
3386 case 0x28: // SB
3387 if(!c||memtarget) {
3388 int a = addr;
3389 if (fastio_reg_override >= 0)
3390 a = fastio_reg_override;
3391 do_store_byte(a, tl, offset_reg);
3392 }
3393 type = STOREB_STUB;
3394 break;
3395 case 0x29: // SH
3396 if(!c||memtarget) {
3397 int a = addr;
3398 if (fastio_reg_override >= 0)
3399 a = fastio_reg_override;
3400 do_store_hword(a, 0, tl, offset_reg, 1);
3401 }
3402 type = STOREH_STUB;
3403 break;
3404 case 0x2B: // SW
3405 if(!c||memtarget) {
3406 int a = addr;
3407 if (fastio_reg_override >= 0)
3408 a = fastio_reg_override;
3409 do_store_word(a, 0, tl, offset_reg, 1);
3410 }
3411 type = STOREW_STUB;
3412 break;
3413 default:
3414 assert(0);
3415 }
3416 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3417 host_tempreg_release();
3418 if(jaddr) {
3419 // PCSX store handlers don't check invcode again
3420 reglist|=1<<addr;
3421 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3422 jaddr=0;
3423 }
3424 {
3425 if(!c||memtarget) {
3426 do_store_smc_check(i, i_regs, reglist, addr);
3427 }
3428 }
3429 u_int addr_val=constmap[i][s]+offset;
3430 if(jaddr) {
3431 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3432 } else if(c&&!memtarget) {
3433 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3434 }
3435 // basic current block modification detection..
3436 // not looking back as that should be in mips cache already
3437 // (see Spyro2 title->attract mode)
3438 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3439 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3440 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3441 if(i_regs->regmap==regs[i].regmap) {
3442 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3443 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3444 emit_movimm(start+i*4+4,0);
3445 emit_writeword(0,&pcaddr);
3446 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3447 emit_far_call(ndrc_get_addr_ht);
3448 emit_jmpreg(0);
3449 }
3450 }
3451}
3452
3453static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3454{
3455 int addr = cinfo[i].addr;
3456 int s,tl;
3457 int offset;
3458 void *jaddr=0;
3459 void *case1, *case23, *case3;
3460 void *done0, *done1, *done2;
3461 int memtarget=0,c=0;
3462 int offset_reg = -1;
3463 u_int reglist=get_host_reglist(i_regs->regmap);
3464 tl=get_reg(i_regs->regmap,dops[i].rs2);
3465 s=get_reg(i_regs->regmap,dops[i].rs1);
3466 offset=cinfo[i].imm;
3467 if(s>=0) {
3468 c=(i_regs->isconst>>s)&1;
3469 if(c) {
3470 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3471 }
3472 }
3473 assert(tl>=0);
3474 assert(addr >= 0);
3475 if(!c) {
3476 emit_cmpimm(addr, RAM_SIZE);
3477 if (!offset && s != addr) emit_mov(s, addr);
3478 jaddr=out;
3479 emit_jno(0);
3480 }
3481 else
3482 {
3483 if(!memtarget||!dops[i].rs1) {
3484 jaddr=out;
3485 emit_jmp(0);
3486 }
3487 }
3488 if (ram_offset)
3489 offset_reg = get_ro_reg(i_regs, 0);
3490
3491 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3492 assert(0);
3493 }
3494
3495 emit_testimm(addr,2);
3496 case23=out;
3497 emit_jne(0);
3498 emit_testimm(addr,1);
3499 case1=out;
3500 emit_jne(0);
3501 // 0
3502 if (dops[i].opcode == 0x2A) { // SWL
3503 // Write msb into least significant byte
3504 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3505 do_store_byte(addr, tl, offset_reg);
3506 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3507 }
3508 else if (dops[i].opcode == 0x2E) { // SWR
3509 // Write entire word
3510 do_store_word(addr, 0, tl, offset_reg, 1);
3511 }
3512 done0 = out;
3513 emit_jmp(0);
3514 // 1
3515 set_jump_target(case1, out);
3516 if (dops[i].opcode == 0x2A) { // SWL
3517 // Write two msb into two least significant bytes
3518 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3519 do_store_hword(addr, -1, tl, offset_reg, 0);
3520 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3521 }
3522 else if (dops[i].opcode == 0x2E) { // SWR
3523 // Write 3 lsb into three most significant bytes
3524 do_store_byte(addr, tl, offset_reg);
3525 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3526 do_store_hword(addr, 1, tl, offset_reg, 0);
3527 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3528 }
3529 done1=out;
3530 emit_jmp(0);
3531 // 2,3
3532 set_jump_target(case23, out);
3533 emit_testimm(addr,1);
3534 case3 = out;
3535 emit_jne(0);
3536 // 2
3537 if (dops[i].opcode==0x2A) { // SWL
3538 // Write 3 msb into three least significant bytes
3539 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3540 do_store_hword(addr, -2, tl, offset_reg, 1);
3541 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3542 do_store_byte(addr, tl, offset_reg);
3543 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3544 }
3545 else if (dops[i].opcode == 0x2E) { // SWR
3546 // Write two lsb into two most significant bytes
3547 do_store_hword(addr, 0, tl, offset_reg, 1);
3548 }
3549 done2 = out;
3550 emit_jmp(0);
3551 // 3
3552 set_jump_target(case3, out);
3553 if (dops[i].opcode == 0x2A) { // SWL
3554 do_store_word(addr, -3, tl, offset_reg, 0);
3555 }
3556 else if (dops[i].opcode == 0x2E) { // SWR
3557 do_store_byte(addr, tl, offset_reg);
3558 }
3559 set_jump_target(done0, out);
3560 set_jump_target(done1, out);
3561 set_jump_target(done2, out);
3562 if (offset_reg == HOST_TEMPREG)
3563 host_tempreg_release();
3564 if(!c||!memtarget)
3565 add_stub_r(STORELR_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3566 do_store_smc_check(i, i_regs, reglist, addr);
3567}
3568
3569static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3570{
3571 if(dops[i].opcode2==0) // MFC0
3572 {
3573 signed char t=get_reg_w(i_regs->regmap, dops[i].rt1);
3574 u_int copr=(source[i]>>11)&0x1f;
3575 if(t>=0&&dops[i].rt1!=0) {
3576 emit_readword(&reg_cop0[copr],t);
3577 }
3578 }
3579 else if(dops[i].opcode2==4) // MTC0
3580 {
3581 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3582 char copr=(source[i]>>11)&0x1f;
3583 assert(s>=0);
3584 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3585 if(copr==9||copr==11||copr==12||copr==13) {
3586 emit_readword(&last_count,HOST_TEMPREG);
3587 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3588 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3589 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3590 emit_writeword(HOST_CCREG,&psxRegs.cycle);
3591 }
3592 // What a mess. The status register (12) can enable interrupts,
3593 // so needs a special case to handle a pending interrupt.
3594 // The interrupt must be taken immediately, because a subsequent
3595 // instruction might disable interrupts again.
3596 if(copr==12||copr==13) {
3597 if (is_delayslot) {
3598 // burn cycles to cause cc_interrupt, which will
3599 // reschedule next_interupt. Relies on CCREG from above.
3600 assem_debug("MTC0 DS %d\n", copr);
3601 emit_writeword(HOST_CCREG,&last_count);
3602 emit_movimm(0,HOST_CCREG);
3603 emit_storereg(CCREG,HOST_CCREG);
3604 emit_loadreg(dops[i].rs1,1);
3605 emit_movimm(copr,0);
3606 emit_far_call(pcsx_mtc0_ds);
3607 emit_loadreg(dops[i].rs1,s);
3608 return;
3609 }
3610 emit_movimm(start+i*4+4,HOST_TEMPREG);
3611 emit_writeword(HOST_TEMPREG,&pcaddr);
3612 emit_movimm(0,HOST_TEMPREG);
3613 emit_writeword(HOST_TEMPREG,&pending_exception);
3614 }
3615 if(s==HOST_CCREG)
3616 emit_loadreg(dops[i].rs1,1);
3617 else if(s!=1)
3618 emit_mov(s,1);
3619 emit_movimm(copr,0);
3620 emit_far_call(pcsx_mtc0);
3621 if(copr==9||copr==11||copr==12||copr==13) {
3622 emit_readword(&psxRegs.cycle,HOST_CCREG);
3623 emit_readword(&next_interupt,HOST_TEMPREG);
3624 emit_addimm(HOST_CCREG,-ccadj_,HOST_CCREG);
3625 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3626 emit_writeword(HOST_TEMPREG,&last_count);
3627 emit_storereg(CCREG,HOST_CCREG);
3628 }
3629 if(copr==12||copr==13) {
3630 assert(!is_delayslot);
3631 emit_readword(&pending_exception,HOST_TEMPREG);
3632 emit_test(HOST_TEMPREG,HOST_TEMPREG);
3633 void *jaddr = out;
3634 emit_jeq(0);
3635 emit_readword(&pcaddr, 0);
3636 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3637 emit_far_call(ndrc_get_addr_ht);
3638 emit_jmpreg(0);
3639 set_jump_target(jaddr, out);
3640 }
3641 emit_loadreg(dops[i].rs1,s);
3642 }
3643}
3644
3645static void rfe_assemble(int i, const struct regstat *i_regs)
3646{
3647 emit_readword(&psxRegs.CP0.n.SR, 0);
3648 emit_andimm(0, 0x3c, 1);
3649 emit_andimm(0, ~0xf, 0);
3650 emit_orrshr_imm(1, 2, 0);
3651 emit_writeword(0, &psxRegs.CP0.n.SR);
3652}
3653
3654static int cop2_is_stalling_op(int i, int *cycles)
3655{
3656 if (dops[i].opcode == 0x3a) { // SWC2
3657 *cycles = 0;
3658 return 1;
3659 }
3660 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3661 *cycles = 0;
3662 return 1;
3663 }
3664 if (dops[i].itype == C2OP) {
3665 *cycles = gte_cycletab[source[i] & 0x3f];
3666 return 1;
3667 }
3668 // ... what about MTC2/CTC2/LWC2?
3669 return 0;
3670}
3671
3672#if 0
3673static void log_gte_stall(int stall, u_int cycle)
3674{
3675 if ((u_int)stall <= 44)
3676 printf("x stall %2d %u\n", stall, cycle + last_count);
3677}
3678
3679static void emit_log_gte_stall(int i, int stall, u_int reglist)
3680{
3681 save_regs(reglist);
3682 if (stall > 0)
3683 emit_movimm(stall, 0);
3684 else
3685 emit_mov(HOST_TEMPREG, 0);
3686 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3687 emit_far_call(log_gte_stall);
3688 restore_regs(reglist);
3689}
3690#endif
3691
3692static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3693{
3694 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3695 int rtmp = reglist_find_free(reglist);
3696
3697 if (HACK_ENABLED(NDHACK_NO_STALLS))
3698 return;
3699 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3700 // happens occasionally... cc evicted? Don't bother then
3701 //printf("no cc %08x\n", start + i*4);
3702 return;
3703 }
3704 if (!dops[i].bt) {
3705 for (j = i - 1; j >= 0; j--) {
3706 //if (dops[j].is_ds) break;
3707 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3708 break;
3709 if (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj)
3710 break;
3711 }
3712 j = max(j, 0);
3713 }
3714 cycles_passed = cinfo[i].ccadj - cinfo[j].ccadj;
3715 if (other_gte_op_cycles >= 0)
3716 stall = other_gte_op_cycles - cycles_passed;
3717 else if (cycles_passed >= 44)
3718 stall = 0; // can't stall
3719 if (stall == -MAXBLOCK && rtmp >= 0) {
3720 // unknown stall, do the expensive runtime check
3721 assem_debug("; cop2_do_stall_check\n");
3722#if 0 // too slow
3723 save_regs(reglist);
3724 emit_movimm(gte_cycletab[op], 0);
3725 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3726 emit_far_call(call_gteStall);
3727 restore_regs(reglist);
3728#else
3729 host_tempreg_acquire();
3730 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3731 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3732 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3733 emit_cmpimm(HOST_TEMPREG, 44);
3734 emit_cmovb_reg(rtmp, HOST_CCREG);
3735 //emit_log_gte_stall(i, 0, reglist);
3736 host_tempreg_release();
3737#endif
3738 }
3739 else if (stall > 0) {
3740 //emit_log_gte_stall(i, stall, reglist);
3741 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3742 }
3743
3744 // save gteBusyCycle, if needed
3745 if (gte_cycletab[op] == 0)
3746 return;
3747 other_gte_op_cycles = -1;
3748 for (j = i + 1; j < slen; j++) {
3749 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3750 break;
3751 if (dops[j].is_jump) {
3752 // check ds
3753 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3754 j++;
3755 break;
3756 }
3757 }
3758 if (other_gte_op_cycles >= 0)
3759 // will handle stall when assembling that op
3760 return;
3761 cycles_passed = cinfo[min(j, slen -1)].ccadj - cinfo[i].ccadj;
3762 if (cycles_passed >= 44)
3763 return;
3764 assem_debug("; save gteBusyCycle\n");
3765 host_tempreg_acquire();
3766#if 0
3767 emit_readword(&last_count, HOST_TEMPREG);
3768 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3769 emit_addimm(HOST_TEMPREG, cinfo[i].ccadj, HOST_TEMPREG);
3770 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3771 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3772#else
3773 emit_addimm(HOST_CCREG, cinfo[i].ccadj + gte_cycletab[op], HOST_TEMPREG);
3774 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3775#endif
3776 host_tempreg_release();
3777}
3778
3779static int is_mflohi(int i)
3780{
3781 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3782}
3783
3784static int check_multdiv(int i, int *cycles)
3785{
3786 if (dops[i].itype != MULTDIV)
3787 return 0;
3788 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3789 *cycles = 11; // approx from 7 11 14
3790 else
3791 *cycles = 37;
3792 return 1;
3793}
3794
3795static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3796{
3797 int j, found = 0, c = 0;
3798 if (HACK_ENABLED(NDHACK_NO_STALLS))
3799 return;
3800 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3801 // happens occasionally... cc evicted? Don't bother then
3802 return;
3803 }
3804 for (j = i + 1; j < slen; j++) {
3805 if (dops[j].bt)
3806 break;
3807 if ((found = is_mflohi(j)))
3808 break;
3809 if (dops[j].is_jump) {
3810 // check ds
3811 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3812 j++;
3813 break;
3814 }
3815 }
3816 if (found)
3817 // handle all in multdiv_do_stall()
3818 return;
3819 check_multdiv(i, &c);
3820 assert(c > 0);
3821 assem_debug("; muldiv prepare stall %d\n", c);
3822 host_tempreg_acquire();
3823 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3824 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3825 host_tempreg_release();
3826}
3827
3828static void multdiv_do_stall(int i, const struct regstat *i_regs)
3829{
3830 int j, known_cycles = 0;
3831 u_int reglist = get_host_reglist(i_regs->regmap);
3832 int rtmp = get_reg_temp(i_regs->regmap);
3833 if (rtmp < 0)
3834 rtmp = reglist_find_free(reglist);
3835 if (HACK_ENABLED(NDHACK_NO_STALLS))
3836 return;
3837 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3838 // happens occasionally... cc evicted? Don't bother then
3839 //printf("no cc/rtmp %08x\n", start + i*4);
3840 return;
3841 }
3842 if (!dops[i].bt) {
3843 for (j = i - 1; j >= 0; j--) {
3844 if (dops[j].is_ds) break;
3845 if (check_multdiv(j, &known_cycles))
3846 break;
3847 if (is_mflohi(j))
3848 // already handled by this op
3849 return;
3850 if (dops[j].bt || (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj))
3851 break;
3852 }
3853 j = max(j, 0);
3854 }
3855 if (known_cycles > 0) {
3856 known_cycles -= cinfo[i].ccadj - cinfo[j].ccadj;
3857 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3858 if (known_cycles > 0)
3859 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3860 return;
3861 }
3862 assem_debug("; muldiv stall unresolved\n");
3863 host_tempreg_acquire();
3864 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3865 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3866 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3867 emit_cmpimm(HOST_TEMPREG, 37);
3868 emit_cmovb_reg(rtmp, HOST_CCREG);
3869 //emit_log_gte_stall(i, 0, reglist);
3870 host_tempreg_release();
3871}
3872
3873static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3874{
3875 switch (copr) {
3876 case 1:
3877 case 3:
3878 case 5:
3879 case 8:
3880 case 9:
3881 case 10:
3882 case 11:
3883 emit_readword(&reg_cop2d[copr],tl);
3884 emit_signextend16(tl,tl);
3885 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3886 break;
3887 case 7:
3888 case 16:
3889 case 17:
3890 case 18:
3891 case 19:
3892 emit_readword(&reg_cop2d[copr],tl);
3893 emit_andimm(tl,0xffff,tl);
3894 emit_writeword(tl,&reg_cop2d[copr]);
3895 break;
3896 case 15:
3897 emit_readword(&reg_cop2d[14],tl); // SXY2
3898 emit_writeword(tl,&reg_cop2d[copr]);
3899 break;
3900 case 28:
3901 case 29:
3902 c2op_mfc2_29_assemble(tl,temp);
3903 break;
3904 default:
3905 emit_readword(&reg_cop2d[copr],tl);
3906 break;
3907 }
3908}
3909
3910static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3911{
3912 switch (copr) {
3913 case 15:
3914 emit_readword(&reg_cop2d[13],temp); // SXY1
3915 emit_writeword(sl,&reg_cop2d[copr]);
3916 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3917 emit_readword(&reg_cop2d[14],temp); // SXY2
3918 emit_writeword(sl,&reg_cop2d[14]);
3919 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3920 break;
3921 case 28:
3922 emit_andimm(sl,0x001f,temp);
3923 emit_shlimm(temp,7,temp);
3924 emit_writeword(temp,&reg_cop2d[9]);
3925 emit_andimm(sl,0x03e0,temp);
3926 emit_shlimm(temp,2,temp);
3927 emit_writeword(temp,&reg_cop2d[10]);
3928 emit_andimm(sl,0x7c00,temp);
3929 emit_shrimm(temp,3,temp);
3930 emit_writeword(temp,&reg_cop2d[11]);
3931 emit_writeword(sl,&reg_cop2d[28]);
3932 break;
3933 case 30:
3934 emit_xorsar_imm(sl,sl,31,temp);
3935#if defined(HAVE_ARMV5) || defined(__aarch64__)
3936 emit_clz(temp,temp);
3937#else
3938 emit_movs(temp,HOST_TEMPREG);
3939 emit_movimm(0,temp);
3940 emit_jeq((int)out+4*4);
3941 emit_addpl_imm(temp,1,temp);
3942 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3943 emit_jns((int)out-2*4);
3944#endif
3945 emit_writeword(sl,&reg_cop2d[30]);
3946 emit_writeword(temp,&reg_cop2d[31]);
3947 break;
3948 case 31:
3949 break;
3950 default:
3951 emit_writeword(sl,&reg_cop2d[copr]);
3952 break;
3953 }
3954}
3955
3956static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3957{
3958 int s,tl;
3959 int ar;
3960 int offset;
3961 int memtarget=0,c=0;
3962 void *jaddr2=NULL;
3963 enum stub_type type;
3964 int offset_reg = -1;
3965 int fastio_reg_override = -1;
3966 u_int reglist=get_host_reglist(i_regs->regmap);
3967 u_int copr=(source[i]>>16)&0x1f;
3968 s=get_reg(i_regs->regmap,dops[i].rs1);
3969 tl=get_reg(i_regs->regmap,FTEMP);
3970 offset=cinfo[i].imm;
3971 assert(dops[i].rs1>0);
3972 assert(tl>=0);
3973
3974 if(i_regs->regmap[HOST_CCREG]==CCREG)
3975 reglist&=~(1<<HOST_CCREG);
3976
3977 // get the address
3978 ar = cinfo[i].addr;
3979 assert(ar >= 0);
3980 if (dops[i].opcode==0x3a) { // SWC2
3981 reglist |= 1<<ar;
3982 }
3983 if(s>=0) c=(i_regs->wasconst>>s)&1;
3984 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3985
3986 cop2_do_stall_check(0, i, i_regs, reglist);
3987
3988 if (dops[i].opcode==0x3a) { // SWC2
3989 cop2_get_dreg(copr,tl,-1);
3990 type=STOREW_STUB;
3991 }
3992 else
3993 type=LOADW_STUB;
3994
3995 if(c&&!memtarget) {
3996 jaddr2=out;
3997 emit_jmp(0); // inline_readstub/inline_writestub?
3998 }
3999 else {
4000 if(!c) {
4001 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
4002 &offset_reg, &fastio_reg_override, ccadj_);
4003 }
4004 else if (ram_offset && memtarget) {
4005 offset_reg = get_ro_reg(i_regs, 0);
4006 }
4007 switch (dops[i].opcode) {
4008 case 0x32: { // LWC2
4009 int a = ar;
4010 if (fastio_reg_override >= 0)
4011 a = fastio_reg_override;
4012 do_load_word(a, tl, offset_reg);
4013 break;
4014 }
4015 case 0x3a: { // SWC2
4016 #ifdef DESTRUCTIVE_SHIFT
4017 if(!offset&&!c&&s>=0) emit_mov(s,ar);
4018 #endif
4019 int a = ar;
4020 if (fastio_reg_override >= 0)
4021 a = fastio_reg_override;
4022 do_store_word(a, 0, tl, offset_reg, 1);
4023 break;
4024 }
4025 default:
4026 assert(0);
4027 }
4028 }
4029 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
4030 host_tempreg_release();
4031 if(jaddr2)
4032 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
4033 if(dops[i].opcode==0x3a) // SWC2
4034 do_store_smc_check(i, i_regs, reglist, ar);
4035 if (dops[i].opcode==0x32) { // LWC2
4036 host_tempreg_acquire();
4037 cop2_put_dreg(copr,tl,HOST_TEMPREG);
4038 host_tempreg_release();
4039 }
4040}
4041
4042static void cop2_assemble(int i, const struct regstat *i_regs)
4043{
4044 u_int copr = (source[i]>>11) & 0x1f;
4045 signed char temp = get_reg_temp(i_regs->regmap);
4046
4047 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
4048 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
4049 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
4050 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
4051 reglist = reglist_exclude(reglist, tl, -1);
4052 }
4053 cop2_do_stall_check(0, i, i_regs, reglist);
4054 }
4055 if (dops[i].opcode2==0) { // MFC2
4056 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4057 if(tl>=0&&dops[i].rt1!=0)
4058 cop2_get_dreg(copr,tl,temp);
4059 }
4060 else if (dops[i].opcode2==4) { // MTC2
4061 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4062 cop2_put_dreg(copr,sl,temp);
4063 }
4064 else if (dops[i].opcode2==2) // CFC2
4065 {
4066 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4067 if(tl>=0&&dops[i].rt1!=0)
4068 emit_readword(&reg_cop2c[copr],tl);
4069 }
4070 else if (dops[i].opcode2==6) // CTC2
4071 {
4072 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4073 switch(copr) {
4074 case 4:
4075 case 12:
4076 case 20:
4077 case 26:
4078 case 27:
4079 case 29:
4080 case 30:
4081 emit_signextend16(sl,temp);
4082 break;
4083 case 31:
4084 c2op_ctc2_31_assemble(sl,temp);
4085 break;
4086 default:
4087 temp=sl;
4088 break;
4089 }
4090 emit_writeword(temp,&reg_cop2c[copr]);
4091 assert(sl>=0);
4092 }
4093}
4094
4095static void do_unalignedwritestub(int n)
4096{
4097 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4098 literal_pool(256);
4099 set_jump_target(stubs[n].addr, out);
4100
4101 int i=stubs[n].a;
4102 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4103 int addr=stubs[n].b;
4104 u_int reglist=stubs[n].e;
4105 signed char *i_regmap=i_regs->regmap;
4106 int temp2=get_reg(i_regmap,FTEMP);
4107 int rt;
4108 rt=get_reg(i_regmap,dops[i].rs2);
4109 assert(rt>=0);
4110 assert(addr>=0);
4111 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4112 reglist|=(1<<addr);
4113 reglist&=~(1<<temp2);
4114
4115 // don't bother with it and call write handler
4116 save_regs(reglist);
4117 pass_args(addr,rt);
4118 int cc=get_reg(i_regmap,CCREG);
4119 if(cc<0)
4120 emit_loadreg(CCREG,2);
4121 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4122 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4123 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4124 if(cc<0)
4125 emit_storereg(CCREG,2);
4126 restore_regs(reglist);
4127 emit_jmp(stubs[n].retaddr); // return address
4128}
4129
4130static void do_overflowstub(int n)
4131{
4132 assem_debug("do_overflowstub %x\n", start + (u_int)stubs[n].a * 4);
4133 literal_pool(24);
4134 int i = stubs[n].a;
4135 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4136 int ccadj = stubs[n].d;
4137 set_jump_target(stubs[n].addr, out);
4138 wb_dirtys(regs[i].regmap, regs[i].dirty);
4139 exception_assemble(i, i_regs, ccadj);
4140}
4141
4142static void do_alignmentstub(int n)
4143{
4144 assem_debug("do_alignmentstub %x\n", start + (u_int)stubs[n].a * 4);
4145 literal_pool(24);
4146 int i = stubs[n].a;
4147 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4148 int ccadj = stubs[n].d;
4149 int is_store = dops[i].itype == STORE || dops[i].opcode == 0x3A; // SWC2
4150 int cause = (dops[i].opcode & 3) << 28;
4151 cause |= is_store ? (R3000E_AdES << 2) : (R3000E_AdEL << 2);
4152 set_jump_target(stubs[n].addr, out);
4153 wb_dirtys(regs[i].regmap, regs[i].dirty);
4154 if (stubs[n].b != 1)
4155 emit_mov(stubs[n].b, 1); // faulting address
4156 emit_movimm(cause, 0);
4157 exception_assemble(i, i_regs, ccadj);
4158}
4159
4160#ifndef multdiv_assemble
4161void multdiv_assemble(int i,struct regstat *i_regs)
4162{
4163 printf("Need multdiv_assemble for this architecture.\n");
4164 abort();
4165}
4166#endif
4167
4168static void mov_assemble(int i, const struct regstat *i_regs)
4169{
4170 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
4171 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
4172 if(dops[i].rt1) {
4173 signed char sl,tl;
4174 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4175 //assert(tl>=0);
4176 if(tl>=0) {
4177 sl=get_reg(i_regs->regmap,dops[i].rs1);
4178 if(sl>=0) emit_mov(sl,tl);
4179 else emit_loadreg(dops[i].rs1,tl);
4180 }
4181 }
4182 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
4183 multdiv_do_stall(i, i_regs);
4184}
4185
4186// call interpreter, exception handler, things that change pc/regs/cycles ...
4187static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
4188{
4189 signed char ccreg=get_reg(i_regs->regmap,CCREG);
4190 assert(ccreg==HOST_CCREG);
4191 assert(!is_delayslot);
4192 (void)ccreg;
4193
4194 emit_movimm(pc,3); // Get PC
4195 emit_readword(&last_count,2);
4196 emit_writeword(3,&psxRegs.pc);
4197 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
4198 emit_add(2,HOST_CCREG,2);
4199 emit_writeword(2,&psxRegs.cycle);
4200 emit_addimm_ptr(FP,(u_char *)&psxRegs - (u_char *)&dynarec_local,0);
4201 emit_far_call(func);
4202 emit_far_jump(jump_to_new_pc);
4203}
4204
4205static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_)
4206{
4207 // 'break' tends to be littered around to catch things like
4208 // division by 0 and is almost never executed, so don't emit much code here
4209 void *func;
4210 if (dops[i].itype == ALU || dops[i].itype == IMM16)
4211 func = is_delayslot ? jump_overflow_ds : jump_overflow;
4212 else if (dops[i].itype == LOAD || dops[i].itype == STORE)
4213 func = is_delayslot ? jump_addrerror_ds : jump_addrerror;
4214 else if (dops[i].opcode2 == 0x0C)
4215 func = is_delayslot ? jump_syscall_ds : jump_syscall;
4216 else
4217 func = is_delayslot ? jump_break_ds : jump_break;
4218 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) // evicted
4219 emit_loadreg(CCREG, HOST_CCREG);
4220 emit_movimm(start + i*4, 2); // pc
4221 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4222 emit_far_jump(func);
4223}
4224
4225static void hlecall_bad()
4226{
4227 assert(0);
4228}
4229
4230static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4231{
4232 void *hlefunc = hlecall_bad;
4233 uint32_t hleCode = source[i] & 0x03ffffff;
4234 if (hleCode < ARRAY_SIZE(psxHLEt))
4235 hlefunc = psxHLEt[hleCode];
4236
4237 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4238}
4239
4240static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4241{
4242 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4243}
4244
4245static void speculate_mov(int rs,int rt)
4246{
4247 if(rt!=0) {
4248 smrv_strong_next|=1<<rt;
4249 smrv[rt]=smrv[rs];
4250 }
4251}
4252
4253static void speculate_mov_weak(int rs,int rt)
4254{
4255 if(rt!=0) {
4256 smrv_weak_next|=1<<rt;
4257 smrv[rt]=smrv[rs];
4258 }
4259}
4260
4261static void speculate_register_values(int i)
4262{
4263 if(i==0) {
4264 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4265 // gp,sp are likely to stay the same throughout the block
4266 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4267 smrv_weak_next=~smrv_strong_next;
4268 //printf(" llr %08x\n", smrv[4]);
4269 }
4270 smrv_strong=smrv_strong_next;
4271 smrv_weak=smrv_weak_next;
4272 switch(dops[i].itype) {
4273 case ALU:
4274 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4275 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4276 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4277 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4278 else {
4279 smrv_strong_next&=~(1<<dops[i].rt1);
4280 smrv_weak_next&=~(1<<dops[i].rt1);
4281 }
4282 break;
4283 case SHIFTIMM:
4284 smrv_strong_next&=~(1<<dops[i].rt1);
4285 smrv_weak_next&=~(1<<dops[i].rt1);
4286 // fallthrough
4287 case IMM16:
4288 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4289 int value,hr=get_reg_w(regs[i].regmap, dops[i].rt1);
4290 if(hr>=0) {
4291 if(get_final_value(hr,i,&value))
4292 smrv[dops[i].rt1]=value;
4293 else smrv[dops[i].rt1]=constmap[i][hr];
4294 smrv_strong_next|=1<<dops[i].rt1;
4295 }
4296 }
4297 else {
4298 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4299 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4300 }
4301 break;
4302 case LOAD:
4303 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4304 // special case for BIOS
4305 smrv[dops[i].rt1]=0xa0000000;
4306 smrv_strong_next|=1<<dops[i].rt1;
4307 break;
4308 }
4309 // fallthrough
4310 case SHIFT:
4311 case LOADLR:
4312 case MOV:
4313 smrv_strong_next&=~(1<<dops[i].rt1);
4314 smrv_weak_next&=~(1<<dops[i].rt1);
4315 break;
4316 case COP0:
4317 case COP2:
4318 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4319 smrv_strong_next&=~(1<<dops[i].rt1);
4320 smrv_weak_next&=~(1<<dops[i].rt1);
4321 }
4322 break;
4323 case C2LS:
4324 if (dops[i].opcode==0x32) { // LWC2
4325 smrv_strong_next&=~(1<<dops[i].rt1);
4326 smrv_weak_next&=~(1<<dops[i].rt1);
4327 }
4328 break;
4329 }
4330#if 0
4331 int r=4;
4332 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4333 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4334#endif
4335}
4336
4337static void ujump_assemble(int i, const struct regstat *i_regs);
4338static void rjump_assemble(int i, const struct regstat *i_regs);
4339static void cjump_assemble(int i, const struct regstat *i_regs);
4340static void sjump_assemble(int i, const struct regstat *i_regs);
4341
4342static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4343{
4344 int ds = 0;
4345 switch (dops[i].itype) {
4346 case ALU:
4347 alu_assemble(i, i_regs, ccadj_);
4348 break;
4349 case IMM16:
4350 imm16_assemble(i, i_regs, ccadj_);
4351 break;
4352 case SHIFT:
4353 shift_assemble(i, i_regs);
4354 break;
4355 case SHIFTIMM:
4356 shiftimm_assemble(i, i_regs);
4357 break;
4358 case LOAD:
4359 load_assemble(i, i_regs, ccadj_);
4360 break;
4361 case LOADLR:
4362 loadlr_assemble(i, i_regs, ccadj_);
4363 break;
4364 case STORE:
4365 store_assemble(i, i_regs, ccadj_);
4366 break;
4367 case STORELR:
4368 storelr_assemble(i, i_regs, ccadj_);
4369 break;
4370 case COP0:
4371 cop0_assemble(i, i_regs, ccadj_);
4372 break;
4373 case RFE:
4374 rfe_assemble(i, i_regs);
4375 break;
4376 case COP2:
4377 cop2_assemble(i, i_regs);
4378 break;
4379 case C2LS:
4380 c2ls_assemble(i, i_regs, ccadj_);
4381 break;
4382 case C2OP:
4383 c2op_assemble(i, i_regs);
4384 break;
4385 case MULTDIV:
4386 multdiv_assemble(i, i_regs);
4387 multdiv_prepare_stall(i, i_regs, ccadj_);
4388 break;
4389 case MOV:
4390 mov_assemble(i, i_regs);
4391 break;
4392 case SYSCALL:
4393 exception_assemble(i, i_regs, ccadj_);
4394 break;
4395 case HLECALL:
4396 hlecall_assemble(i, i_regs, ccadj_);
4397 break;
4398 case INTCALL:
4399 intcall_assemble(i, i_regs, ccadj_);
4400 break;
4401 case UJUMP:
4402 ujump_assemble(i, i_regs);
4403 ds = 1;
4404 break;
4405 case RJUMP:
4406 rjump_assemble(i, i_regs);
4407 ds = 1;
4408 break;
4409 case CJUMP:
4410 cjump_assemble(i, i_regs);
4411 ds = 1;
4412 break;
4413 case SJUMP:
4414 sjump_assemble(i, i_regs);
4415 ds = 1;
4416 break;
4417 case NOP:
4418 case OTHER:
4419 // not handled, just skip
4420 break;
4421 default:
4422 assert(0);
4423 }
4424 return ds;
4425}
4426
4427static void ds_assemble(int i, const struct regstat *i_regs)
4428{
4429 speculate_register_values(i);
4430 is_delayslot = 1;
4431 switch (dops[i].itype) {
4432 case SYSCALL:
4433 case HLECALL:
4434 case INTCALL:
4435 case UJUMP:
4436 case RJUMP:
4437 case CJUMP:
4438 case SJUMP:
4439 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4440 break;
4441 default:
4442 assemble(i, i_regs, cinfo[i].ccadj);
4443 }
4444 is_delayslot = 0;
4445}
4446
4447// Is the branch target a valid internal jump?
4448static int internal_branch(int addr)
4449{
4450 if(addr&1) return 0; // Indirect (register) jump
4451 if(addr>=start && addr<start+slen*4-4)
4452 {
4453 return 1;
4454 }
4455 return 0;
4456}
4457
4458static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4459{
4460 int hr;
4461 for(hr=0;hr<HOST_REGS;hr++) {
4462 if(hr!=EXCLUDE_REG) {
4463 if(pre[hr]!=entry[hr]) {
4464 if(pre[hr]>=0) {
4465 if((dirty>>hr)&1) {
4466 if(get_reg(entry,pre[hr])<0) {
4467 assert(pre[hr]<64);
4468 if(!((u>>pre[hr])&1))
4469 emit_storereg(pre[hr],hr);
4470 }
4471 }
4472 }
4473 }
4474 }
4475 }
4476 // Move from one register to another (no writeback)
4477 for(hr=0;hr<HOST_REGS;hr++) {
4478 if(hr!=EXCLUDE_REG) {
4479 if(pre[hr]!=entry[hr]) {
4480 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4481 int nr;
4482 if((nr=get_reg(entry,pre[hr]))>=0) {
4483 emit_mov(hr,nr);
4484 }
4485 }
4486 }
4487 }
4488 }
4489}
4490
4491// Load the specified registers
4492// This only loads the registers given as arguments because
4493// we don't want to load things that will be overwritten
4494static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4495{
4496 int hr = get_reg(regmap, rs);
4497 if (hr >= 0 && entry[hr] != regmap[hr])
4498 emit_loadreg(regmap[hr], hr);
4499}
4500
4501static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4502{
4503 load_reg(entry, regmap, rs1);
4504 if (rs1 != rs2)
4505 load_reg(entry, regmap, rs2);
4506}
4507
4508// Load registers prior to the start of a loop
4509// so that they are not loaded within the loop
4510static void loop_preload(signed char pre[],signed char entry[])
4511{
4512 int hr;
4513 for (hr = 0; hr < HOST_REGS; hr++) {
4514 int r = entry[hr];
4515 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4516 assem_debug("loop preload:\n");
4517 if (r < TEMPREG)
4518 emit_loadreg(r, hr);
4519 }
4520 }
4521}
4522
4523// Generate address for load/store instruction
4524// goes to AGEN (or temp) for writes, FTEMP for LOADLR and cop1/2 loads
4525// AGEN is assigned by pass5b_preallocate2
4526static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4527{
4528 if (dops[i].is_load || dops[i].is_store) {
4529 int ra = -1;
4530 int agr = AGEN1 + (i&1);
4531 if(dops[i].itype==LOAD) {
4532 if (!dops[i].may_except)
4533 ra = get_reg_w(i_regs->regmap, dops[i].rt1); // reuse dest for agen
4534 if (ra < 0)
4535 ra = get_reg_temp(i_regs->regmap);
4536 }
4537 if(dops[i].itype==LOADLR) {
4538 ra=get_reg(i_regs->regmap,FTEMP);
4539 }
4540 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4541 ra=get_reg(i_regs->regmap,agr);
4542 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4543 }
4544 if(dops[i].itype==C2LS) {
4545 if (dops[i].opcode == 0x32) // LWC2
4546 ra=get_reg(i_regs->regmap,FTEMP);
4547 else { // SWC2
4548 ra=get_reg(i_regs->regmap,agr);
4549 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4550 }
4551 }
4552 int rs = get_reg(i_regs->regmap, dops[i].rs1);
4553 //if(ra>=0)
4554 {
4555 int offset = cinfo[i].imm;
4556 int add_offset = offset != 0;
4557 int c=(i_regs->wasconst>>rs)&1;
4558 if(dops[i].rs1==0) {
4559 // Using r0 as a base address
4560 assert(ra >= 0);
4561 if(!entry||entry[ra]!=agr) {
4562 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4563 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4564 }else{
4565 emit_movimm(offset,ra);
4566 }
4567 } // else did it in the previous cycle
4568 cinfo[i].addr = ra;
4569 add_offset = 0;
4570 }
4571 else if (rs < 0) {
4572 assert(ra >= 0);
4573 if (!entry || entry[ra] != dops[i].rs1)
4574 emit_loadreg(dops[i].rs1, ra);
4575 cinfo[i].addr = ra;
4576 //if(!entry||entry[ra]!=dops[i].rs1)
4577 // printf("poor load scheduling!\n");
4578 }
4579 else if(c) {
4580 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4581 assert(ra >= 0);
4582 if(!entry||entry[ra]!=agr) {
4583 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4584 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4585 }else{
4586 emit_movimm(constmap[i][rs]+offset,ra);
4587 regs[i].loadedconst|=1<<ra;
4588 }
4589 } // else did it in the previous cycle
4590 cinfo[i].addr = ra;
4591 }
4592 else // else load_consts already did it
4593 cinfo[i].addr = rs;
4594 add_offset = 0;
4595 }
4596 else if (dops[i].itype == STORELR) { // overwrites addr
4597 assert(ra >= 0);
4598 assert(rs != ra);
4599 emit_mov(rs, ra);
4600 cinfo[i].addr = ra;
4601 }
4602 else
4603 cinfo[i].addr = rs;
4604 if (add_offset) {
4605 assert(ra >= 0);
4606 if(rs>=0) {
4607 emit_addimm(rs,offset,ra);
4608 }else{
4609 emit_addimm(ra,offset,ra);
4610 }
4611 cinfo[i].addr = ra;
4612 }
4613 }
4614 assert(cinfo[i].addr >= 0);
4615 }
4616 // Preload constants for next instruction
4617 if (dops[i+1].is_load || dops[i+1].is_store) {
4618 int agr,ra;
4619 // Actual address
4620 agr=AGEN1+((i+1)&1);
4621 ra=get_reg(i_regs->regmap,agr);
4622 if(ra>=0) {
4623 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4624 int offset=cinfo[i+1].imm;
4625 int c=(regs[i+1].wasconst>>rs)&1;
4626 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4627 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4628 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4629 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4630 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4631 }else{
4632 emit_movimm(constmap[i+1][rs]+offset,ra);
4633 regs[i+1].loadedconst|=1<<ra;
4634 }
4635 }
4636 else if(dops[i+1].rs1==0) {
4637 // Using r0 as a base address
4638 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4639 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4640 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4641 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4642 }else{
4643 emit_movimm(offset,ra);
4644 }
4645 }
4646 }
4647 }
4648}
4649
4650static int get_final_value(int hr, int i, int *value)
4651{
4652 int reg=regs[i].regmap[hr];
4653 while(i<slen-1) {
4654 if(regs[i+1].regmap[hr]!=reg) break;
4655 if(!((regs[i+1].isconst>>hr)&1)) break;
4656 if(dops[i+1].bt) break;
4657 i++;
4658 }
4659 if(i<slen-1) {
4660 if (dops[i].is_jump) {
4661 *value=constmap[i][hr];
4662 return 1;
4663 }
4664 if(!dops[i+1].bt) {
4665 if (dops[i+1].is_jump) {
4666 // Load in delay slot, out-of-order execution
4667 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4668 {
4669 // Precompute load address
4670 *value=constmap[i][hr]+cinfo[i+2].imm;
4671 return 1;
4672 }
4673 }
4674 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4675 {
4676 // Precompute load address
4677 *value=constmap[i][hr]+cinfo[i+1].imm;
4678 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],cinfo[i+1].imm);
4679 return 1;
4680 }
4681 }
4682 }
4683 *value=constmap[i][hr];
4684 //printf("c=%lx\n",(long)constmap[i][hr]);
4685 if(i==slen-1) return 1;
4686 assert(reg < 64);
4687 return !((unneeded_reg[i+1]>>reg)&1);
4688}
4689
4690// Load registers with known constants
4691static void load_consts(signed char pre[],signed char regmap[],int i)
4692{
4693 int hr,hr2;
4694 // propagate loaded constant flags
4695 if(i==0||dops[i].bt)
4696 regs[i].loadedconst=0;
4697 else {
4698 for(hr=0;hr<HOST_REGS;hr++) {
4699 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4700 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4701 {
4702 regs[i].loadedconst|=1<<hr;
4703 }
4704 }
4705 }
4706 // Load 32-bit regs
4707 for(hr=0;hr<HOST_REGS;hr++) {
4708 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4709 //if(entry[hr]!=regmap[hr]) {
4710 if(!((regs[i].loadedconst>>hr)&1)) {
4711 assert(regmap[hr]<64);
4712 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4713 int value,similar=0;
4714 if(get_final_value(hr,i,&value)) {
4715 // see if some other register has similar value
4716 for(hr2=0;hr2<HOST_REGS;hr2++) {
4717 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4718 if(is_similar_value(value,constmap[i][hr2])) {
4719 similar=1;
4720 break;
4721 }
4722 }
4723 }
4724 if(similar) {
4725 int value2;
4726 if(get_final_value(hr2,i,&value2)) // is this needed?
4727 emit_movimm_from(value2,hr2,value,hr);
4728 else
4729 emit_movimm(value,hr);
4730 }
4731 else if(value==0) {
4732 emit_zeroreg(hr);
4733 }
4734 else {
4735 emit_movimm(value,hr);
4736 }
4737 }
4738 regs[i].loadedconst|=1<<hr;
4739 }
4740 }
4741 }
4742 }
4743}
4744
4745static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4746{
4747 int hr;
4748 // Load 32-bit regs
4749 for(hr=0;hr<HOST_REGS;hr++) {
4750 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4751 assert(regmap[hr] < 64);
4752 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4753 int value=constmap[i][hr];
4754 if(value==0) {
4755 emit_zeroreg(hr);
4756 }
4757 else {
4758 emit_movimm(value,hr);
4759 }
4760 }
4761 }
4762 }
4763}
4764
4765// Write out all dirty registers (except cycle count)
4766static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4767{
4768 int hr;
4769 for(hr=0;hr<HOST_REGS;hr++) {
4770 if(hr!=EXCLUDE_REG) {
4771 if(i_regmap[hr]>0) {
4772 if(i_regmap[hr]!=CCREG) {
4773 if((i_dirty>>hr)&1) {
4774 assert(i_regmap[hr]<64);
4775 emit_storereg(i_regmap[hr],hr);
4776 }
4777 }
4778 }
4779 }
4780 }
4781}
4782
4783// Write out dirty registers that we need to reload (pair with load_needed_regs)
4784// This writes the registers not written by store_regs_bt
4785static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4786{
4787 int hr;
4788 int t=(addr-start)>>2;
4789 for(hr=0;hr<HOST_REGS;hr++) {
4790 if(hr!=EXCLUDE_REG) {
4791 if(i_regmap[hr]>0) {
4792 if(i_regmap[hr]!=CCREG) {
4793 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4794 if((i_dirty>>hr)&1) {
4795 assert(i_regmap[hr]<64);
4796 emit_storereg(i_regmap[hr],hr);
4797 }
4798 }
4799 }
4800 }
4801 }
4802 }
4803}
4804
4805// Load all registers (except cycle count)
4806static void load_all_regs(const signed char i_regmap[])
4807{
4808 int hr;
4809 for(hr=0;hr<HOST_REGS;hr++) {
4810 if(hr!=EXCLUDE_REG) {
4811 if(i_regmap[hr]==0) {
4812 emit_zeroreg(hr);
4813 }
4814 else
4815 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4816 {
4817 emit_loadreg(i_regmap[hr],hr);
4818 }
4819 }
4820 }
4821}
4822
4823// Load all current registers also needed by next instruction
4824static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4825{
4826 int hr;
4827 for(hr=0;hr<HOST_REGS;hr++) {
4828 if(hr!=EXCLUDE_REG) {
4829 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4830 if(i_regmap[hr]==0) {
4831 emit_zeroreg(hr);
4832 }
4833 else
4834 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4835 {
4836 emit_loadreg(i_regmap[hr],hr);
4837 }
4838 }
4839 }
4840 }
4841}
4842
4843// Load all regs, storing cycle count if necessary
4844static void load_regs_entry(int t)
4845{
4846 int hr;
4847 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4848 else if(cinfo[t].ccadj) emit_addimm(HOST_CCREG,-cinfo[t].ccadj,HOST_CCREG);
4849 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4850 emit_storereg(CCREG,HOST_CCREG);
4851 }
4852 // Load 32-bit regs
4853 for(hr=0;hr<HOST_REGS;hr++) {
4854 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4855 if(regs[t].regmap_entry[hr]==0) {
4856 emit_zeroreg(hr);
4857 }
4858 else if(regs[t].regmap_entry[hr]!=CCREG)
4859 {
4860 emit_loadreg(regs[t].regmap_entry[hr],hr);
4861 }
4862 }
4863 }
4864}
4865
4866// Store dirty registers prior to branch
4867static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4868{
4869 if(internal_branch(addr))
4870 {
4871 int t=(addr-start)>>2;
4872 int hr;
4873 for(hr=0;hr<HOST_REGS;hr++) {
4874 if(hr!=EXCLUDE_REG) {
4875 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4876 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4877 if((i_dirty>>hr)&1) {
4878 assert(i_regmap[hr]<64);
4879 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4880 emit_storereg(i_regmap[hr],hr);
4881 }
4882 }
4883 }
4884 }
4885 }
4886 }
4887 else
4888 {
4889 // Branch out of this block, write out all dirty regs
4890 wb_dirtys(i_regmap,i_dirty);
4891 }
4892}
4893
4894// Load all needed registers for branch target
4895static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4896{
4897 //if(addr>=start && addr<(start+slen*4))
4898 if(internal_branch(addr))
4899 {
4900 int t=(addr-start)>>2;
4901 int hr;
4902 // Store the cycle count before loading something else
4903 if(i_regmap[HOST_CCREG]!=CCREG) {
4904 assert(i_regmap[HOST_CCREG]==-1);
4905 }
4906 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4907 emit_storereg(CCREG,HOST_CCREG);
4908 }
4909 // Load 32-bit regs
4910 for(hr=0;hr<HOST_REGS;hr++) {
4911 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4912 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4913 if(regs[t].regmap_entry[hr]==0) {
4914 emit_zeroreg(hr);
4915 }
4916 else if(regs[t].regmap_entry[hr]!=CCREG)
4917 {
4918 emit_loadreg(regs[t].regmap_entry[hr],hr);
4919 }
4920 }
4921 }
4922 }
4923 }
4924}
4925
4926static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4927{
4928 if(addr>=start && addr<start+slen*4-4)
4929 {
4930 int t=(addr-start)>>2;
4931 int hr;
4932 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4933 for(hr=0;hr<HOST_REGS;hr++)
4934 {
4935 if(hr!=EXCLUDE_REG)
4936 {
4937 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4938 {
4939 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4940 {
4941 return 0;
4942 }
4943 else
4944 if((i_dirty>>hr)&1)
4945 {
4946 if(i_regmap[hr]<TEMPREG)
4947 {
4948 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4949 return 0;
4950 }
4951 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4952 {
4953 assert(0);
4954 }
4955 }
4956 }
4957 else // Same register but is it 32-bit or dirty?
4958 if(i_regmap[hr]>=0)
4959 {
4960 if(!((regs[t].dirty>>hr)&1))
4961 {
4962 if((i_dirty>>hr)&1)
4963 {
4964 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4965 {
4966 //printf("%x: dirty no match\n",addr);
4967 return 0;
4968 }
4969 }
4970 }
4971 }
4972 }
4973 }
4974 // Delay slots are not valid branch targets
4975 //if(t>0&&(dops[t-1].is_jump) return 0;
4976 // Delay slots require additional processing, so do not match
4977 if(dops[t].is_ds) return 0;
4978 }
4979 else
4980 {
4981 int hr;
4982 for(hr=0;hr<HOST_REGS;hr++)
4983 {
4984 if(hr!=EXCLUDE_REG)
4985 {
4986 if(i_regmap[hr]>=0)
4987 {
4988 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4989 {
4990 if((i_dirty>>hr)&1)
4991 {
4992 return 0;
4993 }
4994 }
4995 }
4996 }
4997 }
4998 }
4999 return 1;
5000}
5001
5002#ifdef DRC_DBG
5003static void drc_dbg_emit_do_cmp(int i, int ccadj_)
5004{
5005 extern void do_insn_cmp();
5006 //extern int cycle;
5007 u_int hr, reglist = get_host_reglist(regs[i].regmap);
5008
5009 assem_debug("//do_insn_cmp %08x\n", start+i*4);
5010 save_regs(reglist);
5011 // write out changed consts to match the interpreter
5012 if (i > 0 && !dops[i].bt) {
5013 for (hr = 0; hr < HOST_REGS; hr++) {
5014 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
5015 if (hr == EXCLUDE_REG || reg <= 0)
5016 continue;
5017 if (!((regs[i-1].isconst >> hr) & 1))
5018 continue;
5019 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
5020 continue;
5021 emit_movimm(constmap[i-1][hr],0);
5022 emit_storereg(reg, 0);
5023 }
5024 }
5025 emit_movimm(start+i*4,0);
5026 emit_writeword(0,&pcaddr);
5027 int cc = get_reg(regs[i].regmap_entry, CCREG);
5028 if (cc < 0)
5029 emit_loadreg(CCREG, cc = 0);
5030 emit_addimm(cc, ccadj_, 0);
5031 emit_writeword(0, &psxRegs.cycle);
5032 emit_far_call(do_insn_cmp);
5033 //emit_readword(&cycle,0);
5034 //emit_addimm(0,2,0);
5035 //emit_writeword(0,&cycle);
5036 (void)get_reg2;
5037 restore_regs(reglist);
5038 assem_debug("\\\\do_insn_cmp\n");
5039}
5040#else
5041#define drc_dbg_emit_do_cmp(x,y)
5042#endif
5043
5044// Used when a branch jumps into the delay slot of another branch
5045static void ds_assemble_entry(int i)
5046{
5047 int t = (cinfo[i].ba - start) >> 2;
5048 int ccadj_ = -CLOCK_ADJUST(1);
5049 if (!instr_addr[t])
5050 instr_addr[t] = out;
5051 assem_debug("Assemble delay slot at %x\n",cinfo[i].ba);
5052 assem_debug("<->\n");
5053 drc_dbg_emit_do_cmp(t, ccadj_);
5054 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
5055 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
5056 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
5057 address_generation(t,&regs[t],regs[t].regmap_entry);
5058 if (ram_offset && (dops[t].is_load || dops[t].is_store))
5059 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
5060 if (dops[t].is_store)
5061 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
5062 is_delayslot=0;
5063 switch (dops[t].itype) {
5064 case SYSCALL:
5065 case HLECALL:
5066 case INTCALL:
5067 case UJUMP:
5068 case RJUMP:
5069 case CJUMP:
5070 case SJUMP:
5071 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5072 break;
5073 default:
5074 assemble(t, &regs[t], ccadj_);
5075 }
5076 store_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5077 load_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5078 if(internal_branch(cinfo[i].ba+4))
5079 assem_debug("branch: internal\n");
5080 else
5081 assem_debug("branch: external\n");
5082 assert(internal_branch(cinfo[i].ba+4));
5083 add_to_linker(out,cinfo[i].ba+4,internal_branch(cinfo[i].ba+4));
5084 emit_jmp(0);
5085}
5086
5087// Load 2 immediates optimizing for small code size
5088static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
5089{
5090 emit_movimm(imm1,rt1);
5091 emit_movimm_from(imm1,rt1,imm2,rt2);
5092}
5093
5094static void do_cc(int i, const signed char i_regmap[], int *adj,
5095 int addr, int taken, int invert)
5096{
5097 int count, count_plus2;
5098 void *jaddr;
5099 void *idle=NULL;
5100 int t=0;
5101 if(dops[i].itype==RJUMP)
5102 {
5103 *adj=0;
5104 }
5105 //if(cinfo[i].ba>=start && cinfo[i].ba<(start+slen*4))
5106 if(internal_branch(cinfo[i].ba))
5107 {
5108 t=(cinfo[i].ba-start)>>2;
5109 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5110 else *adj=cinfo[t].ccadj;
5111 }
5112 else
5113 {
5114 *adj=0;
5115 }
5116 count = cinfo[i].ccadj;
5117 count_plus2 = count + CLOCK_ADJUST(2);
5118 if(taken==TAKEN && i==(cinfo[i].ba-start)>>2 && source[i+1]==0) {
5119 // Idle loop
5120 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5121 idle=out;
5122 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5123 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5124 jaddr=out;
5125 emit_jmp(0);
5126 }
5127 else if(*adj==0||invert) {
5128 int cycles = count_plus2;
5129 // faster loop HACK
5130#if 0
5131 if (t&&*adj) {
5132 int rel=t-i;
5133 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5134 cycles=*adj+count+2-*adj;
5135 }
5136#endif
5137 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5138 jaddr = out;
5139 emit_jns(0);
5140 }
5141 else
5142 {
5143 emit_cmpimm(HOST_CCREG, -count_plus2);
5144 jaddr = out;
5145 emit_jns(0);
5146 }
5147 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5148}
5149
5150static void do_ccstub(int n)
5151{
5152 literal_pool(256);
5153 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5154 set_jump_target(stubs[n].addr, out);
5155 int i=stubs[n].b;
5156 if(stubs[n].d==NULLDS) {
5157 // Delay slot instruction is nullified ("likely" branch)
5158 wb_dirtys(regs[i].regmap,regs[i].dirty);
5159 }
5160 else if(stubs[n].d!=TAKEN) {
5161 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5162 }
5163 else {
5164 if(internal_branch(cinfo[i].ba))
5165 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5166 }
5167 if(stubs[n].c!=-1)
5168 {
5169 // Save PC as return address
5170 emit_movimm(stubs[n].c,0);
5171 emit_writeword(0,&pcaddr);
5172 }
5173 else
5174 {
5175 // Return address depends on which way the branch goes
5176 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5177 {
5178 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5179 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5180 if(dops[i].rs1==0)
5181 {
5182 s1l=s2l;
5183 s2l=-1;
5184 }
5185 else if(dops[i].rs2==0)
5186 {
5187 s2l=-1;
5188 }
5189 assert(s1l>=0);
5190 #ifdef DESTRUCTIVE_WRITEBACK
5191 if(dops[i].rs1) {
5192 if((branch_regs[i].dirty>>s1l)&&1)
5193 emit_loadreg(dops[i].rs1,s1l);
5194 }
5195 else {
5196 if((branch_regs[i].dirty>>s1l)&1)
5197 emit_loadreg(dops[i].rs2,s1l);
5198 }
5199 if(s2l>=0)
5200 if((branch_regs[i].dirty>>s2l)&1)
5201 emit_loadreg(dops[i].rs2,s2l);
5202 #endif
5203 int hr=0;
5204 int addr=-1,alt=-1,ntaddr=-1;
5205 while(hr<HOST_REGS)
5206 {
5207 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5208 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5209 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5210 {
5211 addr=hr++;break;
5212 }
5213 hr++;
5214 }
5215 while(hr<HOST_REGS)
5216 {
5217 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5218 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5219 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5220 {
5221 alt=hr++;break;
5222 }
5223 hr++;
5224 }
5225 if ((dops[i].opcode & 0x3e) == 6) // BLEZ/BGTZ needs another register
5226 {
5227 while(hr<HOST_REGS)
5228 {
5229 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5230 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5231 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5232 {
5233 ntaddr=hr;break;
5234 }
5235 hr++;
5236 }
5237 assert(hr<HOST_REGS);
5238 }
5239 if (dops[i].opcode == 4) // BEQ
5240 {
5241 #ifdef HAVE_CMOV_IMM
5242 if(s2l>=0) emit_cmp(s1l,s2l);
5243 else emit_test(s1l,s1l);
5244 emit_cmov2imm_e_ne_compact(cinfo[i].ba,start+i*4+8,addr);
5245 #else
5246 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,alt);
5247 if(s2l>=0) emit_cmp(s1l,s2l);
5248 else emit_test(s1l,s1l);
5249 emit_cmovne_reg(alt,addr);
5250 #endif
5251 }
5252 else if (dops[i].opcode == 5) // BNE
5253 {
5254 #ifdef HAVE_CMOV_IMM
5255 if(s2l>=0) emit_cmp(s1l,s2l);
5256 else emit_test(s1l,s1l);
5257 emit_cmov2imm_e_ne_compact(start+i*4+8,cinfo[i].ba,addr);
5258 #else
5259 emit_mov2imm_compact(start+i*4+8,addr,cinfo[i].ba,alt);
5260 if(s2l>=0) emit_cmp(s1l,s2l);
5261 else emit_test(s1l,s1l);
5262 emit_cmovne_reg(alt,addr);
5263 #endif
5264 }
5265 else if (dops[i].opcode == 6) // BLEZ
5266 {
5267 //emit_movimm(cinfo[i].ba,alt);
5268 //emit_movimm(start+i*4+8,addr);
5269 emit_mov2imm_compact(cinfo[i].ba,alt,start+i*4+8,addr);
5270 emit_cmpimm(s1l,1);
5271 emit_cmovl_reg(alt,addr);
5272 }
5273 else if (dops[i].opcode == 7) // BGTZ
5274 {
5275 //emit_movimm(cinfo[i].ba,addr);
5276 //emit_movimm(start+i*4+8,ntaddr);
5277 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,ntaddr);
5278 emit_cmpimm(s1l,1);
5279 emit_cmovl_reg(ntaddr,addr);
5280 }
5281 else if (dops[i].itype == SJUMP) // BLTZ/BGEZ
5282 {
5283 //emit_movimm(cinfo[i].ba,alt);
5284 //emit_movimm(start+i*4+8,addr);
5285 emit_mov2imm_compact(cinfo[i].ba,
5286 (dops[i].opcode2 & 1) ? addr : alt, start + i*4 + 8,
5287 (dops[i].opcode2 & 1) ? alt : addr);
5288 emit_test(s1l,s1l);
5289 emit_cmovs_reg(alt,addr);
5290 }
5291 emit_writeword(addr, &pcaddr);
5292 }
5293 else
5294 if(dops[i].itype==RJUMP)
5295 {
5296 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5297 if (ds_writes_rjump_rs(i)) {
5298 r=get_reg(branch_regs[i].regmap,RTEMP);
5299 }
5300 emit_writeword(r,&pcaddr);
5301 }
5302 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5303 }
5304 // Update cycle count
5305 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5306 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5307 emit_far_call(cc_interrupt);
5308 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5309 if(stubs[n].d==TAKEN) {
5310 if(internal_branch(cinfo[i].ba))
5311 load_needed_regs(branch_regs[i].regmap,regs[(cinfo[i].ba-start)>>2].regmap_entry);
5312 else if(dops[i].itype==RJUMP) {
5313 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5314 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5315 else
5316 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5317 }
5318 }else if(stubs[n].d==NOTTAKEN) {
5319 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5320 else load_all_regs(branch_regs[i].regmap);
5321 }else if(stubs[n].d==NULLDS) {
5322 // Delay slot instruction is nullified ("likely" branch)
5323 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5324 else load_all_regs(regs[i].regmap);
5325 }else{
5326 load_all_regs(branch_regs[i].regmap);
5327 }
5328 if (stubs[n].retaddr)
5329 emit_jmp(stubs[n].retaddr);
5330 else
5331 do_jump_vaddr(stubs[n].e);
5332}
5333
5334static void add_to_linker(void *addr, u_int target, int is_internal)
5335{
5336 assert(linkcount < ARRAY_SIZE(link_addr));
5337 link_addr[linkcount].addr = addr;
5338 link_addr[linkcount].target = target;
5339 link_addr[linkcount].internal = is_internal;
5340 linkcount++;
5341}
5342
5343static void ujump_assemble_write_ra(int i)
5344{
5345 int rt;
5346 unsigned int return_address;
5347 rt=get_reg(branch_regs[i].regmap,31);
5348 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]);
5349 //assert(rt>=0);
5350 return_address=start+i*4+8;
5351 if(rt>=0) {
5352 #ifdef USE_MINI_HT
5353 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5354 int temp=-1; // note: must be ds-safe
5355 #ifdef HOST_TEMPREG
5356 temp=HOST_TEMPREG;
5357 #endif
5358 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5359 else emit_movimm(return_address,rt);
5360 }
5361 else
5362 #endif
5363 {
5364 #ifdef REG_PREFETCH
5365 if(temp>=0)
5366 {
5367 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5368 }
5369 #endif
5370 emit_movimm(return_address,rt); // PC into link register
5371 #ifdef IMM_PREFETCH
5372 emit_prefetch(hash_table_get(return_address));
5373 #endif
5374 }
5375 }
5376}
5377
5378static void ujump_assemble(int i, const struct regstat *i_regs)
5379{
5380 int ra_done=0;
5381 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5382 address_generation(i+1,i_regs,regs[i].regmap_entry);
5383 #ifdef REG_PREFETCH
5384 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5385 if(dops[i].rt1==31&&temp>=0)
5386 {
5387 signed char *i_regmap=i_regs->regmap;
5388 int return_address=start+i*4+8;
5389 if(get_reg(branch_regs[i].regmap,31)>0)
5390 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5391 }
5392 #endif
5393 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5394 ujump_assemble_write_ra(i); // writeback ra for DS
5395 ra_done=1;
5396 }
5397 ds_assemble(i+1,i_regs);
5398 uint64_t bc_unneeded=branch_regs[i].u;
5399 bc_unneeded|=1|(1LL<<dops[i].rt1);
5400 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5401 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5402 if(!ra_done&&dops[i].rt1==31)
5403 ujump_assemble_write_ra(i);
5404 int cc,adj;
5405 cc=get_reg(branch_regs[i].regmap,CCREG);
5406 assert(cc==HOST_CCREG);
5407 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5408 #ifdef REG_PREFETCH
5409 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5410 #endif
5411 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5412 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5413 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5414 if(internal_branch(cinfo[i].ba))
5415 assem_debug("branch: internal\n");
5416 else
5417 assem_debug("branch: external\n");
5418 if (internal_branch(cinfo[i].ba) && dops[(cinfo[i].ba-start)>>2].is_ds) {
5419 ds_assemble_entry(i);
5420 }
5421 else {
5422 add_to_linker(out,cinfo[i].ba,internal_branch(cinfo[i].ba));
5423 emit_jmp(0);
5424 }
5425}
5426
5427static void rjump_assemble_write_ra(int i)
5428{
5429 int rt,return_address;
5430 assert(dops[i+1].rt1!=dops[i].rt1);
5431 assert(dops[i+1].rt2!=dops[i].rt1);
5432 rt=get_reg_w(branch_regs[i].regmap, dops[i].rt1);
5433 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]);
5434 assert(rt>=0);
5435 return_address=start+i*4+8;
5436 #ifdef REG_PREFETCH
5437 if(temp>=0)
5438 {
5439 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5440 }
5441 #endif
5442 emit_movimm(return_address,rt); // PC into link register
5443 #ifdef IMM_PREFETCH
5444 emit_prefetch(hash_table_get(return_address));
5445 #endif
5446}
5447
5448static void rjump_assemble(int i, const struct regstat *i_regs)
5449{
5450 int temp;
5451 int rs,cc;
5452 int ra_done=0;
5453 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5454 assert(rs>=0);
5455 if (ds_writes_rjump_rs(i)) {
5456 // Delay slot abuse, make a copy of the branch address register
5457 temp=get_reg(branch_regs[i].regmap,RTEMP);
5458 assert(temp>=0);
5459 assert(regs[i].regmap[temp]==RTEMP);
5460 emit_mov(rs,temp);
5461 rs=temp;
5462 }
5463 address_generation(i+1,i_regs,regs[i].regmap_entry);
5464 #ifdef REG_PREFETCH
5465 if(dops[i].rt1==31)
5466 {
5467 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5468 signed char *i_regmap=i_regs->regmap;
5469 int return_address=start+i*4+8;
5470 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5471 }
5472 }
5473 #endif
5474 #ifdef USE_MINI_HT
5475 if(dops[i].rs1==31) {
5476 int rh=get_reg(regs[i].regmap,RHASH);
5477 if(rh>=0) do_preload_rhash(rh);
5478 }
5479 #endif
5480 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5481 rjump_assemble_write_ra(i);
5482 ra_done=1;
5483 }
5484 ds_assemble(i+1,i_regs);
5485 uint64_t bc_unneeded=branch_regs[i].u;
5486 bc_unneeded|=1|(1LL<<dops[i].rt1);
5487 bc_unneeded&=~(1LL<<dops[i].rs1);
5488 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5489 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5490 if(!ra_done&&dops[i].rt1!=0)
5491 rjump_assemble_write_ra(i);
5492 cc=get_reg(branch_regs[i].regmap,CCREG);
5493 assert(cc==HOST_CCREG);
5494 (void)cc;
5495 #ifdef USE_MINI_HT
5496 int rh=get_reg(branch_regs[i].regmap,RHASH);
5497 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5498 if(dops[i].rs1==31) {
5499 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5500 do_preload_rhtbl(ht);
5501 do_rhash(rs,rh);
5502 }
5503 #endif
5504 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5505 #ifdef DESTRUCTIVE_WRITEBACK
5506 if((branch_regs[i].dirty>>rs)&1) {
5507 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5508 emit_loadreg(dops[i].rs1,rs);
5509 }
5510 }
5511 #endif
5512 #ifdef REG_PREFETCH
5513 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5514 #endif
5515 #ifdef USE_MINI_HT
5516 if(dops[i].rs1==31) {
5517 do_miniht_load(ht,rh);
5518 }
5519 #endif
5520 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5521 //if(adj) emit_addimm(cc,2*(cinfo[i].ccadj+2-adj),cc); // ??? - Shouldn't happen
5522 //assert(adj==0);
5523 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5524 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5525 if (dops[i+1].itype == RFE)
5526 // special case for RFE
5527 emit_jmp(0);
5528 else
5529 emit_jns(0);
5530 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5531 #ifdef USE_MINI_HT
5532 if(dops[i].rs1==31) {
5533 do_miniht_jump(rs,rh,ht);
5534 }
5535 else
5536 #endif
5537 {
5538 do_jump_vaddr(rs);
5539 }
5540 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5541 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5542 #endif
5543}
5544
5545static void cjump_assemble(int i, const struct regstat *i_regs)
5546{
5547 const signed char *i_regmap = i_regs->regmap;
5548 int cc;
5549 int match;
5550 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5551 assem_debug("match=%d\n",match);
5552 int s1l,s2l;
5553 int unconditional=0,nop=0;
5554 int invert=0;
5555 int internal=internal_branch(cinfo[i].ba);
5556 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5557 if(!match) invert=1;
5558 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5559 if(i>(cinfo[i].ba-start)>>2) invert=1;
5560 #endif
5561 #ifdef __aarch64__
5562 invert=1; // because of near cond. branches
5563 #endif
5564
5565 if(dops[i].ooo) {
5566 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5567 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5568 }
5569 else {
5570 s1l=get_reg(i_regmap,dops[i].rs1);
5571 s2l=get_reg(i_regmap,dops[i].rs2);
5572 }
5573 if(dops[i].rs1==0&&dops[i].rs2==0)
5574 {
5575 if(dops[i].opcode&1) nop=1;
5576 else unconditional=1;
5577 //assert(dops[i].opcode!=5);
5578 //assert(dops[i].opcode!=7);
5579 //assert(dops[i].opcode!=0x15);
5580 //assert(dops[i].opcode!=0x17);
5581 }
5582 else if(dops[i].rs1==0)
5583 {
5584 s1l=s2l;
5585 s2l=-1;
5586 }
5587 else if(dops[i].rs2==0)
5588 {
5589 s2l=-1;
5590 }
5591
5592 if(dops[i].ooo) {
5593 // Out of order execution (delay slot first)
5594 //printf("OOOE\n");
5595 address_generation(i+1,i_regs,regs[i].regmap_entry);
5596 ds_assemble(i+1,i_regs);
5597 int adj;
5598 uint64_t bc_unneeded=branch_regs[i].u;
5599 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5600 bc_unneeded|=1;
5601 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5602 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5603 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5604 cc=get_reg(branch_regs[i].regmap,CCREG);
5605 assert(cc==HOST_CCREG);
5606 if(unconditional)
5607 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5608 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5609 //assem_debug("cycle count (adj)\n");
5610 if(unconditional) {
5611 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5612 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5613 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5614 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5615 if(internal)
5616 assem_debug("branch: internal\n");
5617 else
5618 assem_debug("branch: external\n");
5619 if (internal && dops[(cinfo[i].ba-start)>>2].is_ds) {
5620 ds_assemble_entry(i);
5621 }
5622 else {
5623 add_to_linker(out,cinfo[i].ba,internal);
5624 emit_jmp(0);
5625 }
5626 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5627 if(((u_int)out)&7) emit_addnop(0);
5628 #endif
5629 }
5630 }
5631 else if(nop) {
5632 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5633 void *jaddr=out;
5634 emit_jns(0);
5635 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5636 }
5637 else {
5638 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5639 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5640 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5641
5642 //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]);
5643 assert(s1l>=0);
5644 if(dops[i].opcode==4) // BEQ
5645 {
5646 if(s2l>=0) emit_cmp(s1l,s2l);
5647 else emit_test(s1l,s1l);
5648 if(invert){
5649 nottaken=out;
5650 emit_jne(DJT_1);
5651 }else{
5652 add_to_linker(out,cinfo[i].ba,internal);
5653 emit_jeq(0);
5654 }
5655 }
5656 if(dops[i].opcode==5) // BNE
5657 {
5658 if(s2l>=0) emit_cmp(s1l,s2l);
5659 else emit_test(s1l,s1l);
5660 if(invert){
5661 nottaken=out;
5662 emit_jeq(DJT_1);
5663 }else{
5664 add_to_linker(out,cinfo[i].ba,internal);
5665 emit_jne(0);
5666 }
5667 }
5668 if(dops[i].opcode==6) // BLEZ
5669 {
5670 emit_cmpimm(s1l,1);
5671 if(invert){
5672 nottaken=out;
5673 emit_jge(DJT_1);
5674 }else{
5675 add_to_linker(out,cinfo[i].ba,internal);
5676 emit_jl(0);
5677 }
5678 }
5679 if(dops[i].opcode==7) // BGTZ
5680 {
5681 emit_cmpimm(s1l,1);
5682 if(invert){
5683 nottaken=out;
5684 emit_jl(DJT_1);
5685 }else{
5686 add_to_linker(out,cinfo[i].ba,internal);
5687 emit_jge(0);
5688 }
5689 }
5690 if(invert) {
5691 if(taken) set_jump_target(taken, out);
5692 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5693 if (match && (!internal || !dops[(cinfo[i].ba-start)>>2].is_ds)) {
5694 if(adj) {
5695 emit_addimm(cc,-adj,cc);
5696 add_to_linker(out,cinfo[i].ba,internal);
5697 }else{
5698 emit_addnop(13);
5699 add_to_linker(out,cinfo[i].ba,internal*2);
5700 }
5701 emit_jmp(0);
5702 }else
5703 #endif
5704 {
5705 if(adj) emit_addimm(cc,-adj,cc);
5706 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5707 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5708 if(internal)
5709 assem_debug("branch: internal\n");
5710 else
5711 assem_debug("branch: external\n");
5712 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5713 ds_assemble_entry(i);
5714 }
5715 else {
5716 add_to_linker(out,cinfo[i].ba,internal);
5717 emit_jmp(0);
5718 }
5719 }
5720 set_jump_target(nottaken, out);
5721 }
5722
5723 if(nottaken1) set_jump_target(nottaken1, out);
5724 if(adj) {
5725 if(!invert) emit_addimm(cc,adj,cc);
5726 }
5727 } // (!unconditional)
5728 } // if(ooo)
5729 else
5730 {
5731 // In-order execution (branch first)
5732 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5733 if(!unconditional&&!nop) {
5734 //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]);
5735 assert(s1l>=0);
5736 if((dops[i].opcode&0x2f)==4) // BEQ
5737 {
5738 if(s2l>=0) emit_cmp(s1l,s2l);
5739 else emit_test(s1l,s1l);
5740 nottaken=out;
5741 emit_jne(DJT_2);
5742 }
5743 if((dops[i].opcode&0x2f)==5) // BNE
5744 {
5745 if(s2l>=0) emit_cmp(s1l,s2l);
5746 else emit_test(s1l,s1l);
5747 nottaken=out;
5748 emit_jeq(DJT_2);
5749 }
5750 if((dops[i].opcode&0x2f)==6) // BLEZ
5751 {
5752 emit_cmpimm(s1l,1);
5753 nottaken=out;
5754 emit_jge(DJT_2);
5755 }
5756 if((dops[i].opcode&0x2f)==7) // BGTZ
5757 {
5758 emit_cmpimm(s1l,1);
5759 nottaken=out;
5760 emit_jl(DJT_2);
5761 }
5762 } // if(!unconditional)
5763 int adj;
5764 uint64_t ds_unneeded=branch_regs[i].u;
5765 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5766 ds_unneeded|=1;
5767 // branch taken
5768 if(!nop) {
5769 if(taken) set_jump_target(taken, out);
5770 assem_debug("1:\n");
5771 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5772 // load regs
5773 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5774 address_generation(i+1,&branch_regs[i],0);
5775 if (ram_offset)
5776 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5777 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5778 ds_assemble(i+1,&branch_regs[i]);
5779 cc=get_reg(branch_regs[i].regmap,CCREG);
5780 if(cc==-1) {
5781 emit_loadreg(CCREG,cc=HOST_CCREG);
5782 // CHECK: Is the following instruction (fall thru) allocated ok?
5783 }
5784 assert(cc==HOST_CCREG);
5785 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5786 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5787 assem_debug("cycle count (adj)\n");
5788 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5789 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5790 if(internal)
5791 assem_debug("branch: internal\n");
5792 else
5793 assem_debug("branch: external\n");
5794 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5795 ds_assemble_entry(i);
5796 }
5797 else {
5798 add_to_linker(out,cinfo[i].ba,internal);
5799 emit_jmp(0);
5800 }
5801 }
5802 // branch not taken
5803 if(!unconditional) {
5804 if(nottaken1) set_jump_target(nottaken1, out);
5805 set_jump_target(nottaken, out);
5806 assem_debug("2:\n");
5807 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5808 // load regs
5809 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5810 address_generation(i+1,&branch_regs[i],0);
5811 if (ram_offset)
5812 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5813 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5814 ds_assemble(i+1,&branch_regs[i]);
5815 cc=get_reg(branch_regs[i].regmap,CCREG);
5816 if (cc == -1) {
5817 // Cycle count isn't in a register, temporarily load it then write it out
5818 emit_loadreg(CCREG,HOST_CCREG);
5819 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5820 void *jaddr=out;
5821 emit_jns(0);
5822 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5823 emit_storereg(CCREG,HOST_CCREG);
5824 }
5825 else{
5826 cc=get_reg(i_regmap,CCREG);
5827 assert(cc==HOST_CCREG);
5828 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5829 void *jaddr=out;
5830 emit_jns(0);
5831 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5832 }
5833 }
5834 }
5835}
5836
5837static void sjump_assemble(int i, const struct regstat *i_regs)
5838{
5839 const signed char *i_regmap = i_regs->regmap;
5840 int cc;
5841 int match;
5842 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5843 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5844 int s1l;
5845 int unconditional=0,nevertaken=0;
5846 int invert=0;
5847 int internal=internal_branch(cinfo[i].ba);
5848 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5849 if(!match) invert=1;
5850 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5851 if(i>(cinfo[i].ba-start)>>2) invert=1;
5852 #endif
5853 #ifdef __aarch64__
5854 invert=1; // because of near cond. branches
5855 #endif
5856
5857 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5858 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5859
5860 if(dops[i].ooo) {
5861 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5862 }
5863 else {
5864 s1l=get_reg(i_regmap,dops[i].rs1);
5865 }
5866 if(dops[i].rs1==0)
5867 {
5868 if(dops[i].opcode2&1) unconditional=1;
5869 else nevertaken=1;
5870 // These are never taken (r0 is never less than zero)
5871 //assert(dops[i].opcode2!=0);
5872 //assert(dops[i].opcode2!=2);
5873 //assert(dops[i].opcode2!=0x10);
5874 //assert(dops[i].opcode2!=0x12);
5875 }
5876
5877 if(dops[i].ooo) {
5878 // Out of order execution (delay slot first)
5879 //printf("OOOE\n");
5880 address_generation(i+1,i_regs,regs[i].regmap_entry);
5881 ds_assemble(i+1,i_regs);
5882 int adj;
5883 uint64_t bc_unneeded=branch_regs[i].u;
5884 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5885 bc_unneeded|=1;
5886 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5887 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5888 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5889 if(dops[i].rt1==31) {
5890 int rt,return_address;
5891 rt=get_reg(branch_regs[i].regmap,31);
5892 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]);
5893 if(rt>=0) {
5894 // Save the PC even if the branch is not taken
5895 return_address=start+i*4+8;
5896 emit_movimm(return_address,rt); // PC into link register
5897 #ifdef IMM_PREFETCH
5898 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5899 #endif
5900 }
5901 }
5902 cc=get_reg(branch_regs[i].regmap,CCREG);
5903 assert(cc==HOST_CCREG);
5904 if(unconditional)
5905 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5906 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5907 assem_debug("cycle count (adj)\n");
5908 if(unconditional) {
5909 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5910 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5911 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
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 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5925 if(((u_int)out)&7) emit_addnop(0);
5926 #endif
5927 }
5928 }
5929 else if(nevertaken) {
5930 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5931 void *jaddr=out;
5932 emit_jns(0);
5933 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5934 }
5935 else {
5936 void *nottaken = NULL;
5937 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5938 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5939 {
5940 assert(s1l>=0);
5941 if ((dops[i].opcode2 & 1) == 0) // BLTZ/BLTZAL
5942 {
5943 emit_test(s1l,s1l);
5944 if(invert){
5945 nottaken=out;
5946 emit_jns(DJT_1);
5947 }else{
5948 add_to_linker(out,cinfo[i].ba,internal);
5949 emit_js(0);
5950 }
5951 }
5952 else // BGEZ/BGEZAL
5953 {
5954 emit_test(s1l,s1l);
5955 if(invert){
5956 nottaken=out;
5957 emit_js(DJT_1);
5958 }else{
5959 add_to_linker(out,cinfo[i].ba,internal);
5960 emit_jns(0);
5961 }
5962 }
5963 }
5964
5965 if(invert) {
5966 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5967 if (match && (!internal || !dops[(cinfo[i].ba - start) >> 2].is_ds)) {
5968 if(adj) {
5969 emit_addimm(cc,-adj,cc);
5970 add_to_linker(out,cinfo[i].ba,internal);
5971 }else{
5972 emit_addnop(13);
5973 add_to_linker(out,cinfo[i].ba,internal*2);
5974 }
5975 emit_jmp(0);
5976 }else
5977 #endif
5978 {
5979 if(adj) emit_addimm(cc,-adj,cc);
5980 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5981 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5982 if(internal)
5983 assem_debug("branch: internal\n");
5984 else
5985 assem_debug("branch: external\n");
5986 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5987 ds_assemble_entry(i);
5988 }
5989 else {
5990 add_to_linker(out,cinfo[i].ba,internal);
5991 emit_jmp(0);
5992 }
5993 }
5994 set_jump_target(nottaken, out);
5995 }
5996
5997 if(adj) {
5998 if(!invert) emit_addimm(cc,adj,cc);
5999 }
6000 } // (!unconditional)
6001 } // if(ooo)
6002 else
6003 {
6004 // In-order execution (branch first)
6005 //printf("IOE\n");
6006 void *nottaken = NULL;
6007 if (!unconditional) {
6008 assert(s1l >= 0);
6009 emit_test(s1l, s1l);
6010 }
6011 if (dops[i].rt1 == 31) {
6012 int rt, return_address;
6013 rt = get_reg(branch_regs[i].regmap,31);
6014 if(rt >= 0) {
6015 // Save the PC even if the branch is not taken
6016 return_address = start + i*4+8;
6017 emit_movimm(return_address, rt); // PC into link register
6018 #ifdef IMM_PREFETCH
6019 emit_prefetch(hash_table_get(return_address));
6020 #endif
6021 }
6022 }
6023 if (!unconditional) {
6024 nottaken = out;
6025 if (!(dops[i].opcode2 & 1)) // BLTZ/BLTZAL
6026 emit_jns(DJT_1);
6027 else // BGEZ/BGEZAL
6028 emit_js(DJT_1);
6029 }
6030 int adj;
6031 uint64_t ds_unneeded=branch_regs[i].u;
6032 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6033 ds_unneeded|=1;
6034 // branch taken
6035 if(!nevertaken) {
6036 //assem_debug("1:\n");
6037 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6038 // load regs
6039 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6040 address_generation(i+1,&branch_regs[i],0);
6041 if (ram_offset)
6042 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6043 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6044 ds_assemble(i+1,&branch_regs[i]);
6045 cc=get_reg(branch_regs[i].regmap,CCREG);
6046 if(cc==-1) {
6047 emit_loadreg(CCREG,cc=HOST_CCREG);
6048 // CHECK: Is the following instruction (fall thru) allocated ok?
6049 }
6050 assert(cc==HOST_CCREG);
6051 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6052 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
6053 assem_debug("cycle count (adj)\n");
6054 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
6055 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6056 if(internal)
6057 assem_debug("branch: internal\n");
6058 else
6059 assem_debug("branch: external\n");
6060 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
6061 ds_assemble_entry(i);
6062 }
6063 else {
6064 add_to_linker(out,cinfo[i].ba,internal);
6065 emit_jmp(0);
6066 }
6067 }
6068 // branch not taken
6069 if(!unconditional) {
6070 set_jump_target(nottaken, out);
6071 assem_debug("1:\n");
6072 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6073 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6074 address_generation(i+1,&branch_regs[i],0);
6075 if (ram_offset)
6076 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6077 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6078 ds_assemble(i+1,&branch_regs[i]);
6079 cc=get_reg(branch_regs[i].regmap,CCREG);
6080 if (cc == -1) {
6081 // Cycle count isn't in a register, temporarily load it then write it out
6082 emit_loadreg(CCREG,HOST_CCREG);
6083 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
6084 void *jaddr=out;
6085 emit_jns(0);
6086 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6087 emit_storereg(CCREG,HOST_CCREG);
6088 }
6089 else{
6090 cc=get_reg(i_regmap,CCREG);
6091 assert(cc==HOST_CCREG);
6092 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
6093 void *jaddr=out;
6094 emit_jns(0);
6095 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6096 }
6097 }
6098 }
6099}
6100
6101static void check_regmap(signed char *regmap)
6102{
6103#ifndef NDEBUG
6104 int i,j;
6105 for (i = 0; i < HOST_REGS; i++) {
6106 if (regmap[i] < 0)
6107 continue;
6108 for (j = i + 1; j < HOST_REGS; j++)
6109 assert(regmap[i] != regmap[j]);
6110 }
6111#endif
6112}
6113
6114#ifdef DISASM
6115#include <inttypes.h>
6116static char insn[MAXBLOCK][10];
6117
6118#define set_mnemonic(i_, n_) \
6119 strcpy(insn[i_], n_)
6120
6121void print_regmap(const char *name, const signed char *regmap)
6122{
6123 char buf[5];
6124 int i, l;
6125 fputs(name, stdout);
6126 for (i = 0; i < HOST_REGS; i++) {
6127 l = 0;
6128 if (regmap[i] >= 0)
6129 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6130 for (; l < 3; l++)
6131 buf[l] = ' ';
6132 buf[l] = 0;
6133 printf(" r%d=%s", i, buf);
6134 }
6135 fputs("\n", stdout);
6136}
6137
6138 /* disassembly */
6139void disassemble_inst(int i)
6140{
6141 if (dops[i].bt) printf("*"); else printf(" ");
6142 switch(dops[i].itype) {
6143 case UJUMP:
6144 printf (" %x: %s %8x\n",start+i*4,insn[i],cinfo[i].ba);break;
6145 case CJUMP:
6146 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;
6147 case SJUMP:
6148 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;
6149 case RJUMP:
6150 if (dops[i].opcode2 == 9 && dops[i].rt1 != 31)
6151 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6152 else
6153 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6154 break;
6155 case IMM16:
6156 if(dops[i].opcode==0xf) //LUI
6157 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,cinfo[i].imm&0xffff);
6158 else
6159 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6160 break;
6161 case LOAD:
6162 case LOADLR:
6163 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6164 break;
6165 case STORE:
6166 case STORELR:
6167 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,cinfo[i].imm);
6168 break;
6169 case ALU:
6170 case SHIFT:
6171 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6172 break;
6173 case MULTDIV:
6174 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6175 break;
6176 case SHIFTIMM:
6177 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6178 break;
6179 case MOV:
6180 if((dops[i].opcode2&0x1d)==0x10)
6181 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6182 else if((dops[i].opcode2&0x1d)==0x11)
6183 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6184 else
6185 printf (" %x: %s\n",start+i*4,insn[i]);
6186 break;
6187 case COP0:
6188 if(dops[i].opcode2==0)
6189 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6190 else if(dops[i].opcode2==4)
6191 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6192 else printf (" %x: %s\n",start+i*4,insn[i]);
6193 break;
6194 case COP2:
6195 if(dops[i].opcode2<3)
6196 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6197 else if(dops[i].opcode2>3)
6198 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6199 else printf (" %x: %s\n",start+i*4,insn[i]);
6200 break;
6201 case C2LS:
6202 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,cinfo[i].imm);
6203 break;
6204 case INTCALL:
6205 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6206 break;
6207 default:
6208 //printf (" %s %8x\n",insn[i],source[i]);
6209 printf (" %x: %s\n",start+i*4,insn[i]);
6210 }
6211 #ifndef REGMAP_PRINT
6212 return;
6213 #endif
6214 printf("D: %"PRIx64" WD: %"PRIx64" U: %"PRIx64" hC: %x hWC: %x hLC: %x\n",
6215 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i],
6216 regs[i].isconst, regs[i].wasconst, regs[i].loadedconst);
6217 print_regmap("pre: ", regmap_pre[i]);
6218 print_regmap("entry: ", regs[i].regmap_entry);
6219 print_regmap("map: ", regs[i].regmap);
6220 if (dops[i].is_jump) {
6221 print_regmap("bentry:", branch_regs[i].regmap_entry);
6222 print_regmap("bmap: ", branch_regs[i].regmap);
6223 }
6224}
6225#else
6226#define set_mnemonic(i_, n_)
6227static void disassemble_inst(int i) {}
6228#endif // DISASM
6229
6230#define DRC_TEST_VAL 0x74657374
6231
6232static noinline void new_dynarec_test(void)
6233{
6234 int (*testfunc)(void);
6235 void *beginning;
6236 int ret[2];
6237 size_t i;
6238
6239 // check structure linkage
6240 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6241 {
6242 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6243 }
6244
6245 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6246 new_dynarec_test, out);
6247 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6248
6249 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6250 out = ndrc->translation_cache;
6251 beginning = start_block();
6252 emit_movimm(DRC_TEST_VAL + i, 0); // test
6253 emit_ret();
6254 literal_pool(0);
6255 end_block(beginning);
6256 testfunc = beginning;
6257 ret[i] = testfunc();
6258 }
6259
6260 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6261 SysPrintf("test passed.\n");
6262 else
6263 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6264 out = ndrc->translation_cache;
6265}
6266
6267// clear the state completely, instead of just marking
6268// things invalid like invalidate_all_pages() does
6269void new_dynarec_clear_full(void)
6270{
6271 int n;
6272 out = ndrc->translation_cache;
6273 memset(invalid_code,1,sizeof(invalid_code));
6274 memset(hash_table,0xff,sizeof(hash_table));
6275 memset(mini_ht,-1,sizeof(mini_ht));
6276 memset(shadow,0,sizeof(shadow));
6277 copy=shadow;
6278 expirep = EXPIRITY_OFFSET;
6279 pending_exception=0;
6280 literalcount=0;
6281 stop_after_jal=0;
6282 inv_code_start=inv_code_end=~0;
6283 hack_addr=0;
6284 f1_hack=0;
6285 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6286 blocks_clear(&blocks[n]);
6287 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6288 free(jumps[n]);
6289 jumps[n] = NULL;
6290 }
6291 stat_clear(stat_blocks);
6292 stat_clear(stat_links);
6293
6294 cycle_multiplier_old = Config.cycle_multiplier;
6295 new_dynarec_hacks_old = new_dynarec_hacks;
6296}
6297
6298void new_dynarec_init(void)
6299{
6300 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6301
6302#ifdef _3DS
6303 check_rosalina();
6304#endif
6305#ifdef BASE_ADDR_DYNAMIC
6306 #ifdef VITA
6307 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6308 if (sceBlock <= 0)
6309 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6310 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6311 if (ret < 0)
6312 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6313 sceKernelOpenVMDomain();
6314 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6315 #elif defined(_MSC_VER)
6316 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6317 PAGE_EXECUTE_READWRITE);
6318 #elif defined(HAVE_LIBNX)
6319 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6320 if (R_FAILED(rc))
6321 SysPrintf("jitCreate failed: %08x\n", rc);
6322 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6323 jitTransitionToWritable(&g_jit);
6324 ndrc = g_jit.rx_addr;
6325 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6326 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6327 #else
6328 uintptr_t desired_addr = 0;
6329 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6330 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6331 int fd = -1;
6332 #ifdef __ELF__
6333 extern char _end;
6334 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6335 #endif
6336 #ifdef TC_WRITE_OFFSET
6337 // mostly for testing
6338 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6339 ftruncate(fd, sizeof(*ndrc));
6340 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6341 (flags = MAP_SHARED), fd, 0);
6342 assert(mw != MAP_FAILED);
6343 prot = PROT_READ | PROT_EXEC;
6344 #endif
6345 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6346 if (ndrc == MAP_FAILED) {
6347 SysPrintf("mmap() failed: %s\n", strerror(errno));
6348 abort();
6349 }
6350 #ifdef TC_WRITE_OFFSET
6351 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6352 #endif
6353 #endif
6354#else
6355 #ifndef NO_WRITE_EXEC
6356 // not all systems allow execute in data segment by default
6357 // size must be 4K aligned for 3DS?
6358 if (mprotect(ndrc, sizeof(*ndrc),
6359 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6360 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6361 #endif
6362#endif
6363 out = ndrc->translation_cache;
6364 new_dynarec_clear_full();
6365#ifdef HOST_IMM8
6366 // Copy this into local area so we don't have to put it in every literal pool
6367 invc_ptr=invalid_code;
6368#endif
6369 arch_init();
6370 new_dynarec_test();
6371 ram_offset=(uintptr_t)rdram-0x80000000;
6372 if (ram_offset!=0)
6373 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6374 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6375 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6376}
6377
6378void new_dynarec_cleanup(void)
6379{
6380 int n;
6381#ifdef BASE_ADDR_DYNAMIC
6382 #ifdef VITA
6383 // sceBlock is managed by retroarch's bootstrap code
6384 //sceKernelFreeMemBlock(sceBlock);
6385 //sceBlock = -1;
6386 #elif defined(HAVE_LIBNX)
6387 jitClose(&g_jit);
6388 ndrc = NULL;
6389 #else
6390 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6391 SysPrintf("munmap() failed\n");
6392 ndrc = NULL;
6393 #endif
6394#endif
6395 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6396 blocks_clear(&blocks[n]);
6397 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6398 free(jumps[n]);
6399 jumps[n] = NULL;
6400 }
6401 stat_clear(stat_blocks);
6402 stat_clear(stat_links);
6403 new_dynarec_print_stats();
6404}
6405
6406static u_int *get_source_start(u_int addr, u_int *limit)
6407{
6408 if (addr < 0x00200000 ||
6409 (0xa0000000 <= addr && addr < 0xa0200000))
6410 {
6411 // used for BIOS calls mostly?
6412 *limit = (addr&0xa0000000)|0x00200000;
6413 return (u_int *)(rdram + (addr&0x1fffff));
6414 }
6415 else if (!Config.HLE && (
6416 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6417 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6418 {
6419 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6420 // but timings in PCSX are too tied to the interpreter's 2-per-insn assumption
6421 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6422 cycle_multiplier_active = 200;
6423
6424 *limit = (addr & 0xfff00000) | 0x80000;
6425 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6426 }
6427 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6428 *limit = (addr & 0x80600000) + 0x00200000;
6429 return (u_int *)(rdram + (addr&0x1fffff));
6430 }
6431 return NULL;
6432}
6433
6434static u_int scan_for_ret(u_int addr)
6435{
6436 u_int limit = 0;
6437 u_int *mem;
6438
6439 mem = get_source_start(addr, &limit);
6440 if (mem == NULL)
6441 return addr;
6442
6443 if (limit > addr + 0x1000)
6444 limit = addr + 0x1000;
6445 for (; addr < limit; addr += 4, mem++) {
6446 if (*mem == 0x03e00008) // jr $ra
6447 return addr + 8;
6448 }
6449 return addr;
6450}
6451
6452struct savestate_block {
6453 uint32_t addr;
6454 uint32_t regflags;
6455};
6456
6457static int addr_cmp(const void *p1_, const void *p2_)
6458{
6459 const struct savestate_block *p1 = p1_, *p2 = p2_;
6460 return p1->addr - p2->addr;
6461}
6462
6463int new_dynarec_save_blocks(void *save, int size)
6464{
6465 struct savestate_block *sblocks = save;
6466 int maxcount = size / sizeof(sblocks[0]);
6467 struct savestate_block tmp_blocks[1024];
6468 struct block_info *block;
6469 int p, s, d, o, bcnt;
6470 u_int addr;
6471
6472 o = 0;
6473 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6474 bcnt = 0;
6475 for (block = blocks[p]; block != NULL; block = block->next) {
6476 if (block->is_dirty)
6477 continue;
6478 tmp_blocks[bcnt].addr = block->start;
6479 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6480 bcnt++;
6481 }
6482 if (bcnt < 1)
6483 continue;
6484 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6485
6486 addr = tmp_blocks[0].addr;
6487 for (s = d = 0; s < bcnt; s++) {
6488 if (tmp_blocks[s].addr < addr)
6489 continue;
6490 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6491 tmp_blocks[d++] = tmp_blocks[s];
6492 addr = scan_for_ret(tmp_blocks[s].addr);
6493 }
6494
6495 if (o + d > maxcount)
6496 d = maxcount - o;
6497 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6498 o += d;
6499 }
6500
6501 return o * sizeof(sblocks[0]);
6502}
6503
6504void new_dynarec_load_blocks(const void *save, int size)
6505{
6506 const struct savestate_block *sblocks = save;
6507 int count = size / sizeof(sblocks[0]);
6508 struct block_info *block;
6509 u_int regs_save[32];
6510 u_int page;
6511 uint32_t f;
6512 int i, b;
6513
6514 // restore clean blocks, if any
6515 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6516 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6517 if (!block->is_dirty)
6518 continue;
6519 assert(block->source && block->copy);
6520 if (memcmp(block->source, block->copy, block->len))
6521 continue;
6522
6523 // see try_restore_block
6524 block->is_dirty = 0;
6525 mark_invalid_code(block->start, block->len, 0);
6526 i++;
6527 }
6528 }
6529 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6530
6531 // change GPRs for speculation to at least partially work..
6532 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6533 for (i = 1; i < 32; i++)
6534 psxRegs.GPR.r[i] = 0x80000000;
6535
6536 for (b = 0; b < count; b++) {
6537 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6538 if (f & 1)
6539 psxRegs.GPR.r[i] = 0x1f800000;
6540 }
6541
6542 ndrc_get_addr_ht(sblocks[b].addr);
6543
6544 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6545 if (f & 1)
6546 psxRegs.GPR.r[i] = 0x80000000;
6547 }
6548 }
6549
6550 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6551}
6552
6553void new_dynarec_print_stats(void)
6554{
6555#ifdef STAT_PRINT
6556 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6557 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6558 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6559 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6560 out - ndrc->translation_cache, stat_blocks, stat_links);
6561 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6562 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6563 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6564#endif
6565}
6566
6567static int apply_hacks(void)
6568{
6569 int i;
6570 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6571 return 0;
6572 /* special hack(s) */
6573 for (i = 0; i < slen - 4; i++)
6574 {
6575 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6576 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6577 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6578 && cinfo[i+3].imm == 0x6e40 && dops[i+3].rs1 == 2)
6579 {
6580 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6581 dops[i + 3].itype = NOP;
6582 }
6583 }
6584 i = slen;
6585 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6586 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6587 && dops[i-7].itype == STORE)
6588 {
6589 i = i-8;
6590 if (dops[i].itype == IMM16)
6591 i--;
6592 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6593 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6594 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6595 {
6596 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6597 f1_hack = 1;
6598 return 1;
6599 }
6600 }
6601 return 0;
6602}
6603
6604static noinline void pass1_disassemble(u_int pagelimit)
6605{
6606 int i, j, done = 0, ni_count = 0;
6607 unsigned int type,op,op2,op3;
6608
6609 for (i = 0; !done; i++)
6610 {
6611 int force_j_to_interpreter = 0;
6612 memset(&dops[i], 0, sizeof(dops[i]));
6613 memset(&cinfo[i], 0, sizeof(cinfo[i]));
6614 cinfo[i].ba = -1;
6615 cinfo[i].addr = -1;
6616 dops[i].opcode = op = source[i] >> 26;
6617 op2 = 0;
6618 type = INTCALL;
6619 set_mnemonic(i, "???");
6620 switch(op)
6621 {
6622 case 0x00: set_mnemonic(i, "special");
6623 op2=source[i]&0x3f;
6624 switch(op2)
6625 {
6626 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6627 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6628 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6629 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6630 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6631 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6632 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6633 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6634 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6635 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6636 case 0x0F: set_mnemonic(i, "SYNC"); type=OTHER; break;
6637 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6638 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6639 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6640 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6641 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6642 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6643 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6644 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6645 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6646 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6647 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6648 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6649 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6650 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6651 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6652 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6653 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6654 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6655 }
6656 break;
6657 case 0x01: set_mnemonic(i, "regimm");
6658 type = SJUMP;
6659 op2 = (source[i] >> 16) & 0x1f;
6660 switch(op2)
6661 {
6662 case 0x10: set_mnemonic(i, "BLTZAL"); break;
6663 case 0x11: set_mnemonic(i, "BGEZAL"); break;
6664 default:
6665 if (op2 & 1)
6666 set_mnemonic(i, "BGEZ");
6667 else
6668 set_mnemonic(i, "BLTZ");
6669 }
6670 break;
6671 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6672 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6673 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6674 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6675 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6676 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6677 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6678 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6679 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6680 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6681 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6682 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6683 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6684 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6685 case 0x10: set_mnemonic(i, "COP0");
6686 op2 = (source[i]>>21) & 0x1f;
6687 if (op2 & 0x10) {
6688 op3 = source[i] & 0x1f;
6689 switch (op3)
6690 {
6691 case 0x01: case 0x02: case 0x06: case 0x08: type = INTCALL; break;
6692 case 0x10: set_mnemonic(i, "RFE"); type=RFE; break;
6693 default: type = OTHER; break;
6694 }
6695 break;
6696 }
6697 switch(op2)
6698 {
6699 u32 rd;
6700 case 0x00:
6701 set_mnemonic(i, "MFC0");
6702 rd = (source[i] >> 11) & 0x1F;
6703 if (!(0x00000417u & (1u << rd)))
6704 type = COP0;
6705 break;
6706 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6707 case 0x02:
6708 case 0x06: type = INTCALL; break;
6709 default: type = OTHER; break;
6710 }
6711 break;
6712 case 0x11: set_mnemonic(i, "COP1");
6713 op2=(source[i]>>21)&0x1f;
6714 break;
6715 case 0x12: set_mnemonic(i, "COP2");
6716 op2=(source[i]>>21)&0x1f;
6717 if (op2 & 0x10) {
6718 type = OTHER;
6719 if (gte_handlers[source[i]&0x3f]!=NULL) {
6720#ifdef DISASM
6721 if (gte_regnames[source[i]&0x3f]!=NULL)
6722 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
6723 else
6724 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
6725#endif
6726 type = C2OP;
6727 }
6728 }
6729 else switch(op2)
6730 {
6731 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6732 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6733 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6734 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6735 }
6736 break;
6737 case 0x13: set_mnemonic(i, "COP3");
6738 op2=(source[i]>>21)&0x1f;
6739 break;
6740 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6741 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6742 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6743 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6744 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6745 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6746 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6747 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6748 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6749 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6750 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6751 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6752 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6753 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6754 case 0x3B:
6755 if (Config.HLE && (source[i] & 0x03ffffff) < ARRAY_SIZE(psxHLEt)) {
6756 set_mnemonic(i, "HLECALL");
6757 type = HLECALL;
6758 }
6759 break;
6760 default:
6761 break;
6762 }
6763 if (type == INTCALL)
6764 SysPrintf("NI %08x @%08x (%08x)\n", source[i], start + i*4, start);
6765 dops[i].itype=type;
6766 dops[i].opcode2=op2;
6767 /* Get registers/immediates */
6768 dops[i].use_lt1=0;
6769 gte_rs[i]=gte_rt[i]=0;
6770 dops[i].rs1 = 0;
6771 dops[i].rs2 = 0;
6772 dops[i].rt1 = 0;
6773 dops[i].rt2 = 0;
6774 switch(type) {
6775 case LOAD:
6776 dops[i].rs1=(source[i]>>21)&0x1f;
6777 dops[i].rt1=(source[i]>>16)&0x1f;
6778 cinfo[i].imm=(short)source[i];
6779 break;
6780 case STORE:
6781 case STORELR:
6782 dops[i].rs1=(source[i]>>21)&0x1f;
6783 dops[i].rs2=(source[i]>>16)&0x1f;
6784 cinfo[i].imm=(short)source[i];
6785 break;
6786 case LOADLR:
6787 // LWL/LWR only load part of the register,
6788 // therefore the target register must be treated as a source too
6789 dops[i].rs1=(source[i]>>21)&0x1f;
6790 dops[i].rs2=(source[i]>>16)&0x1f;
6791 dops[i].rt1=(source[i]>>16)&0x1f;
6792 cinfo[i].imm=(short)source[i];
6793 break;
6794 case IMM16:
6795 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6796 else dops[i].rs1=(source[i]>>21)&0x1f;
6797 dops[i].rs2=0;
6798 dops[i].rt1=(source[i]>>16)&0x1f;
6799 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6800 cinfo[i].imm=(unsigned short)source[i];
6801 }else{
6802 cinfo[i].imm=(short)source[i];
6803 }
6804 break;
6805 case UJUMP:
6806 // The JAL instruction writes to r31.
6807 if (op&1) {
6808 dops[i].rt1=31;
6809 }
6810 dops[i].rs2=CCREG;
6811 break;
6812 case RJUMP:
6813 dops[i].rs1=(source[i]>>21)&0x1f;
6814 // The JALR instruction writes to rd.
6815 if (op2&1) {
6816 dops[i].rt1=(source[i]>>11)&0x1f;
6817 }
6818 dops[i].rs2=CCREG;
6819 break;
6820 case CJUMP:
6821 dops[i].rs1=(source[i]>>21)&0x1f;
6822 dops[i].rs2=(source[i]>>16)&0x1f;
6823 if(op&2) { // BGTZ/BLEZ
6824 dops[i].rs2=0;
6825 }
6826 break;
6827 case SJUMP:
6828 dops[i].rs1=(source[i]>>21)&0x1f;
6829 dops[i].rs2=CCREG;
6830 if (op2 == 0x10 || op2 == 0x11) { // BxxAL
6831 dops[i].rt1 = 31;
6832 // NOTE: If the branch is not taken, r31 is still overwritten
6833 }
6834 break;
6835 case ALU:
6836 dops[i].rs1=(source[i]>>21)&0x1f; // source
6837 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
6838 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6839 break;
6840 case MULTDIV:
6841 dops[i].rs1=(source[i]>>21)&0x1f; // source
6842 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
6843 dops[i].rt1=HIREG;
6844 dops[i].rt2=LOREG;
6845 break;
6846 case MOV:
6847 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6848 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6849 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6850 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6851 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
6852 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
6853 break;
6854 case SHIFT:
6855 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
6856 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
6857 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6858 break;
6859 case SHIFTIMM:
6860 dops[i].rs1=(source[i]>>16)&0x1f;
6861 dops[i].rs2=0;
6862 dops[i].rt1=(source[i]>>11)&0x1f;
6863 cinfo[i].imm=(source[i]>>6)&0x1f;
6864 break;
6865 case COP0:
6866 if(op2==0) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0
6867 if(op2==4) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0
6868 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
6869 break;
6870 case COP2:
6871 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
6872 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
6873 dops[i].rs2=CSREG;
6874 int gr=(source[i]>>11)&0x1F;
6875 switch(op2)
6876 {
6877 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6878 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6879 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6880 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6881 }
6882 break;
6883 case C2LS:
6884 dops[i].rs1=(source[i]>>21)&0x1F;
6885 cinfo[i].imm=(short)source[i];
6886 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
6887 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
6888 break;
6889 case C2OP:
6890 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
6891 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
6892 gte_rt[i]|=1ll<<63; // every op changes flags
6893 if((source[i]&0x3f)==GTE_MVMVA) {
6894 int v = (source[i] >> 15) & 3;
6895 gte_rs[i]&=~0xe3fll;
6896 if(v==3) gte_rs[i]|=0xe00ll;
6897 else gte_rs[i]|=3ll<<(v*2);
6898 }
6899 break;
6900 case SYSCALL:
6901 case HLECALL:
6902 case INTCALL:
6903 dops[i].rs1=CCREG;
6904 break;
6905 default:
6906 break;
6907 }
6908 /* Calculate branch target addresses */
6909 if(type==UJUMP)
6910 cinfo[i].ba=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6911 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
6912 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6913 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
6914 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6915 else if(type==CJUMP||type==SJUMP)
6916 cinfo[i].ba=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6917
6918 /* simplify always (not)taken branches */
6919 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
6920 dops[i].rs1 = dops[i].rs2 = 0;
6921 if (!(op & 1)) {
6922 dops[i].itype = type = UJUMP;
6923 dops[i].rs2 = CCREG;
6924 }
6925 }
6926 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
6927 dops[i].itype = type = UJUMP;
6928
6929 dops[i].is_jump = type == RJUMP || type == UJUMP || type == CJUMP || type == SJUMP;
6930 dops[i].is_ujump = type == RJUMP || type == UJUMP;
6931 dops[i].is_load = type == LOAD || type == LOADLR || op == 0x32; // LWC2
6932 dops[i].is_delay_load = (dops[i].is_load || (source[i] & 0xf3d00000) == 0x40000000); // MFC/CFC
6933 dops[i].is_store = type == STORE || type == STORELR || op == 0x3a; // SWC2
6934 dops[i].is_exception = type == SYSCALL || type == HLECALL || type == INTCALL;
6935 dops[i].may_except = dops[i].is_exception || (type == ALU && (op2 == 0x20 || op2 == 0x22)) || op == 8;
6936
6937 if (((op & 0x37) == 0x21 || op == 0x25) // LH/SH/LHU
6938 && ((cinfo[i].imm & 1) || Config.PreciseExceptions))
6939 dops[i].may_except = 1;
6940 if (((op & 0x37) == 0x23 || (op & 0x37) == 0x32) // LW/SW/LWC2/SWC2
6941 && ((cinfo[i].imm & 3) || Config.PreciseExceptions))
6942 dops[i].may_except = 1;
6943
6944 /* rare messy cases to just pass over to the interpreter */
6945 if (i > 0 && dops[i-1].is_jump) {
6946 j = i - 1;
6947 // branch in delay slot?
6948 if (dops[i].is_jump) {
6949 // don't handle first branch and call interpreter if it's hit
6950 SysPrintf("branch in DS @%08x (%08x)\n", start + i*4, start);
6951 force_j_to_interpreter = 1;
6952 }
6953 // basic load delay detection through a branch
6954 else if (dops[i].is_delay_load && dops[i].rt1 != 0) {
6955 int t=(cinfo[i-1].ba-start)/4;
6956 if(0 <= t && t < i &&(dops[i].rt1==dops[t].rs1||dops[i].rt1==dops[t].rs2)&&dops[t].itype!=CJUMP&&dops[t].itype!=SJUMP) {
6957 // jump target wants DS result - potential load delay effect
6958 SysPrintf("load delay in DS @%08x (%08x)\n", start + i*4, start);
6959 force_j_to_interpreter = 1;
6960 dops[t+1].bt=1; // expected return from interpreter
6961 }
6962 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&&
6963 !(i>=3&&dops[i-3].is_jump)) {
6964 // v0 overwrite like this is a sign of trouble, bail out
6965 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
6966 force_j_to_interpreter = 1;
6967 }
6968 }
6969 }
6970 else if (i > 0 && dops[i-1].is_delay_load && dops[i-1].rt1 != 0
6971 && (dops[i].rs1 == dops[i-1].rt1 || dops[i].rs2 == dops[i-1].rt1)) {
6972 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
6973 for (j = i - 1; j > 0 && dops[j-1].is_delay_load; j--)
6974 if (dops[j-1].rt1 != dops[i-1].rt1)
6975 break;
6976 force_j_to_interpreter = 1;
6977 }
6978 if (force_j_to_interpreter) {
6979 memset(&dops[j], 0, sizeof(dops[j]));
6980 dops[j].itype = INTCALL;
6981 dops[j].rs1 = CCREG;
6982 cinfo[j].ba = -1;
6983 done = 2;
6984 i = j; // don't compile the problematic branch/load/etc
6985 }
6986
6987 /* Is this the end of the block? */
6988 if (i > 0 && dops[i-1].is_ujump) {
6989 if (dops[i-1].rt1 == 0) { // not jal
6990 int found_bbranch = 0, t = (cinfo[i-1].ba - start) / 4;
6991 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
6992 // scan for a branch back to i+1
6993 for (j = t; j < t + 64; j++) {
6994 int tmpop = source[j] >> 26;
6995 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
6996 int t2 = j + 1 + (int)(signed short)source[j];
6997 if (t2 == i + 1) {
6998 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
6999 found_bbranch = 1;
7000 break;
7001 }
7002 }
7003 }
7004 }
7005 if (!found_bbranch)
7006 done = 2;
7007 }
7008 else {
7009 if(stop_after_jal) done=1;
7010 // Stop on BREAK
7011 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7012 }
7013 // Don't recompile stuff that's already compiled
7014 if(check_addr(start+i*4+4)) done=1;
7015 // Don't get too close to the limit
7016 if(i>MAXBLOCK/2) done=1;
7017 }
7018 if (dops[i].itype == HLECALL)
7019 stop = 1;
7020 else if (dops[i].itype == INTCALL)
7021 stop = 2;
7022 else if (dops[i].is_exception)
7023 done = stop_after_jal ? 1 : 2;
7024 if (done == 2) {
7025 // Does the block continue due to a branch?
7026 for(j=i-1;j>=0;j--)
7027 {
7028 if(cinfo[j].ba==start+i*4) done=j=0; // Branch into delay slot
7029 if(cinfo[j].ba==start+i*4+4) done=j=0;
7030 if(cinfo[j].ba==start+i*4+8) done=j=0;
7031 }
7032 }
7033 //assert(i<MAXBLOCK-1);
7034 if(start+i*4==pagelimit-4) done=1;
7035 assert(start+i*4<pagelimit);
7036 if (i==MAXBLOCK-1) done=1;
7037 // Stop if we're compiling junk
7038 if (dops[i].itype == INTCALL && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7039 done=stop_after_jal=1;
7040 SysPrintf("Disabled speculative precompilation\n");
7041 }
7042 }
7043 while (i > 0 && dops[i-1].is_jump)
7044 i--;
7045 assert(i > 0);
7046 assert(!dops[i-1].is_jump);
7047 slen = i;
7048}
7049
7050// Basic liveness analysis for MIPS registers
7051static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7052{
7053 int i;
7054 uint64_t u,gte_u,b,gte_b;
7055 uint64_t temp_u,temp_gte_u=0;
7056 uint64_t gte_u_unknown=0;
7057 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7058 gte_u_unknown=~0ll;
7059 if(iend==slen-1) {
7060 u=1;
7061 gte_u=gte_u_unknown;
7062 }else{
7063 //u=unneeded_reg[iend+1];
7064 u=1;
7065 gte_u=gte_unneeded[iend+1];
7066 }
7067
7068 for (i=iend;i>=istart;i--)
7069 {
7070 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7071 if(dops[i].is_jump)
7072 {
7073 // If subroutine call, flag return address as a possible branch target
7074 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7075
7076 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7077 {
7078 // Branch out of this block, flush all regs
7079 u=1;
7080 gte_u=gte_u_unknown;
7081 branch_unneeded_reg[i]=u;
7082 // Merge in delay slot
7083 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7084 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7085 u|=1;
7086 gte_u|=gte_rt[i+1];
7087 gte_u&=~gte_rs[i+1];
7088 }
7089 else
7090 {
7091 // Internal branch, flag target
7092 dops[(cinfo[i].ba-start)>>2].bt=1;
7093 if(cinfo[i].ba<=start+i*4) {
7094 // Backward branch
7095 if(dops[i].is_ujump)
7096 {
7097 // Unconditional branch
7098 temp_u=1;
7099 temp_gte_u=0;
7100 } else {
7101 // Conditional branch (not taken case)
7102 temp_u=unneeded_reg[i+2];
7103 temp_gte_u&=gte_unneeded[i+2];
7104 }
7105 // Merge in delay slot
7106 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7107 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7108 temp_u|=1;
7109 temp_gte_u|=gte_rt[i+1];
7110 temp_gte_u&=~gte_rs[i+1];
7111 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7112 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7113 temp_u|=1;
7114 temp_gte_u|=gte_rt[i];
7115 temp_gte_u&=~gte_rs[i];
7116 unneeded_reg[i]=temp_u;
7117 gte_unneeded[i]=temp_gte_u;
7118 // Only go three levels deep. This recursion can take an
7119 // excessive amount of time if there are a lot of nested loops.
7120 if(r<2) {
7121 pass2_unneeded_regs((cinfo[i].ba-start)>>2,i-1,r+1);
7122 }else{
7123 unneeded_reg[(cinfo[i].ba-start)>>2]=1;
7124 gte_unneeded[(cinfo[i].ba-start)>>2]=gte_u_unknown;
7125 }
7126 } /*else*/ if(1) {
7127 if (dops[i].is_ujump)
7128 {
7129 // Unconditional branch
7130 u=unneeded_reg[(cinfo[i].ba-start)>>2];
7131 gte_u=gte_unneeded[(cinfo[i].ba-start)>>2];
7132 branch_unneeded_reg[i]=u;
7133 // Merge in delay slot
7134 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7135 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7136 u|=1;
7137 gte_u|=gte_rt[i+1];
7138 gte_u&=~gte_rs[i+1];
7139 } else {
7140 // Conditional branch
7141 b=unneeded_reg[(cinfo[i].ba-start)>>2];
7142 gte_b=gte_unneeded[(cinfo[i].ba-start)>>2];
7143 branch_unneeded_reg[i]=b;
7144 // Branch delay slot
7145 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7146 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7147 b|=1;
7148 gte_b|=gte_rt[i+1];
7149 gte_b&=~gte_rs[i+1];
7150 u&=b;
7151 gte_u&=gte_b;
7152 if(i<slen-1) {
7153 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7154 } else {
7155 branch_unneeded_reg[i]=1;
7156 }
7157 }
7158 }
7159 }
7160 }
7161 else if(dops[i].may_except)
7162 {
7163 // SYSCALL instruction, etc or conditional exception
7164 u=1;
7165 }
7166 else if (dops[i].itype == RFE)
7167 {
7168 u=1;
7169 }
7170 //u=1; // DEBUG
7171 // Written registers are unneeded
7172 u|=1LL<<dops[i].rt1;
7173 u|=1LL<<dops[i].rt2;
7174 gte_u|=gte_rt[i];
7175 // Accessed registers are needed
7176 u&=~(1LL<<dops[i].rs1);
7177 u&=~(1LL<<dops[i].rs2);
7178 gte_u&=~gte_rs[i];
7179 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7180 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7181 // Source-target dependencies
7182 // R0 is always unneeded
7183 u|=1;
7184 // Save it
7185 unneeded_reg[i]=u;
7186 gte_unneeded[i]=gte_u;
7187 /*
7188 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7189 printf("U:");
7190 int r;
7191 for(r=1;r<=CCREG;r++) {
7192 if((unneeded_reg[i]>>r)&1) {
7193 if(r==HIREG) printf(" HI");
7194 else if(r==LOREG) printf(" LO");
7195 else printf(" r%d",r);
7196 }
7197 }
7198 printf("\n");
7199 */
7200 }
7201}
7202
7203static noinline void pass3_register_alloc(u_int addr)
7204{
7205 struct regstat current; // Current register allocations/status
7206 clear_all_regs(current.regmap_entry);
7207 clear_all_regs(current.regmap);
7208 current.wasdirty = current.dirty = 0;
7209 current.u = unneeded_reg[0];
7210 alloc_reg(&current, 0, CCREG);
7211 dirty_reg(&current, CCREG);
7212 current.wasconst = 0;
7213 current.isconst = 0;
7214 current.loadedconst = 0;
7215 //current.waswritten = 0;
7216 int ds=0;
7217 int cc=0;
7218 int hr;
7219 int i, j;
7220
7221 if (addr & 1) {
7222 // First instruction is delay slot
7223 cc=-1;
7224 dops[1].bt=1;
7225 ds=1;
7226 unneeded_reg[0]=1;
7227 current.regmap[HOST_BTREG]=BTREG;
7228 }
7229
7230 for(i=0;i<slen;i++)
7231 {
7232 if(dops[i].bt)
7233 {
7234 for(hr=0;hr<HOST_REGS;hr++)
7235 {
7236 // Is this really necessary?
7237 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7238 }
7239 current.isconst=0;
7240 //current.waswritten=0;
7241 }
7242
7243 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7244 regs[i].wasconst=current.isconst;
7245 regs[i].wasdirty=current.dirty;
7246 regs[i].dirty=0;
7247 regs[i].u=0;
7248 regs[i].isconst=0;
7249 regs[i].loadedconst=0;
7250 if (!dops[i].is_jump) {
7251 if(i+1<slen) {
7252 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7253 current.u|=1;
7254 } else {
7255 current.u=1;
7256 }
7257 } else {
7258 if(i+1<slen) {
7259 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7260 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7261 current.u|=1;
7262 } else {
7263 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7264 abort();
7265 }
7266 }
7267 dops[i].is_ds=ds;
7268 if(ds) {
7269 ds=0; // Skip delay slot, already allocated as part of branch
7270 // ...but we need to alloc it in case something jumps here
7271 if(i+1<slen) {
7272 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7273 }else{
7274 current.u=branch_unneeded_reg[i-1];
7275 }
7276 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7277 current.u|=1;
7278 struct regstat temp;
7279 memcpy(&temp,&current,sizeof(current));
7280 temp.wasdirty=temp.dirty;
7281 // TODO: Take into account unconditional branches, as below
7282 delayslot_alloc(&temp,i);
7283 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7284 regs[i].wasdirty=temp.wasdirty;
7285 regs[i].dirty=temp.dirty;
7286 regs[i].isconst=0;
7287 regs[i].wasconst=0;
7288 current.isconst=0;
7289 // Create entry (branch target) regmap
7290 for(hr=0;hr<HOST_REGS;hr++)
7291 {
7292 int r=temp.regmap[hr];
7293 if(r>=0) {
7294 if(r!=regmap_pre[i][hr]) {
7295 regs[i].regmap_entry[hr]=-1;
7296 }
7297 else
7298 {
7299 assert(r < 64);
7300 if((current.u>>r)&1) {
7301 regs[i].regmap_entry[hr]=-1;
7302 regs[i].regmap[hr]=-1;
7303 //Don't clear regs in the delay slot as the branch might need them
7304 //current.regmap[hr]=-1;
7305 }else
7306 regs[i].regmap_entry[hr]=r;
7307 }
7308 } else {
7309 // First instruction expects CCREG to be allocated
7310 if(i==0&&hr==HOST_CCREG)
7311 regs[i].regmap_entry[hr]=CCREG;
7312 else
7313 regs[i].regmap_entry[hr]=-1;
7314 }
7315 }
7316 }
7317 else { // Not delay slot
7318 switch(dops[i].itype) {
7319 case UJUMP:
7320 //current.isconst=0; // DEBUG
7321 //current.wasconst=0; // DEBUG
7322 //regs[i].wasconst=0; // DEBUG
7323 clear_const(&current,dops[i].rt1);
7324 alloc_cc(&current,i);
7325 dirty_reg(&current,CCREG);
7326 if (dops[i].rt1==31) {
7327 alloc_reg(&current,i,31);
7328 dirty_reg(&current,31);
7329 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7330 //assert(dops[i+1].rt1!=dops[i].rt1);
7331 #ifdef REG_PREFETCH
7332 alloc_reg(&current,i,PTEMP);
7333 #endif
7334 }
7335 dops[i].ooo=1;
7336 delayslot_alloc(&current,i+1);
7337 //current.isconst=0; // DEBUG
7338 ds=1;
7339 break;
7340 case RJUMP:
7341 //current.isconst=0;
7342 //current.wasconst=0;
7343 //regs[i].wasconst=0;
7344 clear_const(&current,dops[i].rs1);
7345 clear_const(&current,dops[i].rt1);
7346 alloc_cc(&current,i);
7347 dirty_reg(&current,CCREG);
7348 if (!ds_writes_rjump_rs(i)) {
7349 alloc_reg(&current,i,dops[i].rs1);
7350 if (dops[i].rt1!=0) {
7351 alloc_reg(&current,i,dops[i].rt1);
7352 dirty_reg(&current,dops[i].rt1);
7353 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7354 assert(dops[i+1].rt1!=dops[i].rt1);
7355 #ifdef REG_PREFETCH
7356 alloc_reg(&current,i,PTEMP);
7357 #endif
7358 }
7359 #ifdef USE_MINI_HT
7360 if(dops[i].rs1==31) { // JALR
7361 alloc_reg(&current,i,RHASH);
7362 alloc_reg(&current,i,RHTBL);
7363 }
7364 #endif
7365 delayslot_alloc(&current,i+1);
7366 } else {
7367 // The delay slot overwrites our source register,
7368 // allocate a temporary register to hold the old value.
7369 current.isconst=0;
7370 current.wasconst=0;
7371 regs[i].wasconst=0;
7372 delayslot_alloc(&current,i+1);
7373 current.isconst=0;
7374 alloc_reg(&current,i,RTEMP);
7375 }
7376 //current.isconst=0; // DEBUG
7377 dops[i].ooo=1;
7378 ds=1;
7379 break;
7380 case CJUMP:
7381 //current.isconst=0;
7382 //current.wasconst=0;
7383 //regs[i].wasconst=0;
7384 clear_const(&current,dops[i].rs1);
7385 clear_const(&current,dops[i].rs2);
7386 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7387 {
7388 alloc_cc(&current,i);
7389 dirty_reg(&current,CCREG);
7390 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7391 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7392 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7393 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7394 // The delay slot overwrites one of our conditions.
7395 // Allocate the branch condition registers instead.
7396 current.isconst=0;
7397 current.wasconst=0;
7398 regs[i].wasconst=0;
7399 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7400 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7401 }
7402 else
7403 {
7404 dops[i].ooo=1;
7405 delayslot_alloc(&current,i+1);
7406 }
7407 }
7408 else
7409 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7410 {
7411 alloc_cc(&current,i);
7412 dirty_reg(&current,CCREG);
7413 alloc_reg(&current,i,dops[i].rs1);
7414 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7415 // The delay slot overwrites one of our conditions.
7416 // Allocate the branch condition registers instead.
7417 current.isconst=0;
7418 current.wasconst=0;
7419 regs[i].wasconst=0;
7420 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7421 }
7422 else
7423 {
7424 dops[i].ooo=1;
7425 delayslot_alloc(&current,i+1);
7426 }
7427 }
7428 else
7429 // Don't alloc the delay slot yet because we might not execute it
7430 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7431 {
7432 current.isconst=0;
7433 current.wasconst=0;
7434 regs[i].wasconst=0;
7435 alloc_cc(&current,i);
7436 dirty_reg(&current,CCREG);
7437 alloc_reg(&current,i,dops[i].rs1);
7438 alloc_reg(&current,i,dops[i].rs2);
7439 }
7440 else
7441 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7442 {
7443 current.isconst=0;
7444 current.wasconst=0;
7445 regs[i].wasconst=0;
7446 alloc_cc(&current,i);
7447 dirty_reg(&current,CCREG);
7448 alloc_reg(&current,i,dops[i].rs1);
7449 }
7450 ds=1;
7451 //current.isconst=0;
7452 break;
7453 case SJUMP:
7454 clear_const(&current,dops[i].rs1);
7455 clear_const(&current,dops[i].rt1);
7456 {
7457 alloc_cc(&current,i);
7458 dirty_reg(&current,CCREG);
7459 alloc_reg(&current,i,dops[i].rs1);
7460 if (dops[i].rt1 == 31) { // BLTZAL/BGEZAL
7461 alloc_reg(&current,i,31);
7462 dirty_reg(&current,31);
7463 }
7464 if ((dops[i].rs1 &&
7465 (dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7466 ||(dops[i].rt1 == 31 && dops[i].rs1 == 31) // overwrites it's own condition
7467 ||(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
7468 // Allocate the branch condition registers instead.
7469 current.isconst=0;
7470 current.wasconst=0;
7471 regs[i].wasconst=0;
7472 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7473 }
7474 else
7475 {
7476 dops[i].ooo=1;
7477 delayslot_alloc(&current,i+1);
7478 }
7479 }
7480 ds=1;
7481 //current.isconst=0;
7482 break;
7483 case IMM16:
7484 imm16_alloc(&current,i);
7485 break;
7486 case LOAD:
7487 case LOADLR:
7488 load_alloc(&current,i);
7489 break;
7490 case STORE:
7491 case STORELR:
7492 store_alloc(&current,i);
7493 break;
7494 case ALU:
7495 alu_alloc(&current,i);
7496 break;
7497 case SHIFT:
7498 shift_alloc(&current,i);
7499 break;
7500 case MULTDIV:
7501 multdiv_alloc(&current,i);
7502 break;
7503 case SHIFTIMM:
7504 shiftimm_alloc(&current,i);
7505 break;
7506 case MOV:
7507 mov_alloc(&current,i);
7508 break;
7509 case COP0:
7510 cop0_alloc(&current,i);
7511 break;
7512 case RFE:
7513 rfe_alloc(&current,i);
7514 break;
7515 case COP2:
7516 cop2_alloc(&current,i);
7517 break;
7518 case C2LS:
7519 c2ls_alloc(&current,i);
7520 break;
7521 case C2OP:
7522 c2op_alloc(&current,i);
7523 break;
7524 case SYSCALL:
7525 case HLECALL:
7526 case INTCALL:
7527 syscall_alloc(&current,i);
7528 break;
7529 }
7530
7531 // Create entry (branch target) regmap
7532 for(hr=0;hr<HOST_REGS;hr++)
7533 {
7534 int r,or;
7535 r=current.regmap[hr];
7536 if(r>=0) {
7537 if(r!=regmap_pre[i][hr]) {
7538 // TODO: delay slot (?)
7539 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7540 if(or<0||r>=TEMPREG){
7541 regs[i].regmap_entry[hr]=-1;
7542 }
7543 else
7544 {
7545 // Just move it to a different register
7546 regs[i].regmap_entry[hr]=r;
7547 // If it was dirty before, it's still dirty
7548 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7549 }
7550 }
7551 else
7552 {
7553 // Unneeded
7554 if(r==0){
7555 regs[i].regmap_entry[hr]=0;
7556 }
7557 else
7558 {
7559 assert(r<64);
7560 if((current.u>>r)&1) {
7561 regs[i].regmap_entry[hr]=-1;
7562 //regs[i].regmap[hr]=-1;
7563 current.regmap[hr]=-1;
7564 }else
7565 regs[i].regmap_entry[hr]=r;
7566 }
7567 }
7568 } else {
7569 // Branches expect CCREG to be allocated at the target
7570 if(regmap_pre[i][hr]==CCREG)
7571 regs[i].regmap_entry[hr]=CCREG;
7572 else
7573 regs[i].regmap_entry[hr]=-1;
7574 }
7575 }
7576 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7577 }
7578
7579#if 0 // see do_store_smc_check()
7580 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)
7581 current.waswritten|=1<<dops[i-1].rs1;
7582 current.waswritten&=~(1<<dops[i].rt1);
7583 current.waswritten&=~(1<<dops[i].rt2);
7584 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)cinfo[i].imm>=0x800)
7585 current.waswritten&=~(1<<dops[i].rs1);
7586#endif
7587
7588 /* Branch post-alloc */
7589 if(i>0)
7590 {
7591 current.wasdirty=current.dirty;
7592 switch(dops[i-1].itype) {
7593 case UJUMP:
7594 memcpy(&branch_regs[i-1],&current,sizeof(current));
7595 branch_regs[i-1].isconst=0;
7596 branch_regs[i-1].wasconst=0;
7597 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7598 alloc_cc(&branch_regs[i-1],i-1);
7599 dirty_reg(&branch_regs[i-1],CCREG);
7600 if(dops[i-1].rt1==31) { // JAL
7601 alloc_reg(&branch_regs[i-1],i-1,31);
7602 dirty_reg(&branch_regs[i-1],31);
7603 }
7604 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7605 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7606 break;
7607 case RJUMP:
7608 memcpy(&branch_regs[i-1],&current,sizeof(current));
7609 branch_regs[i-1].isconst=0;
7610 branch_regs[i-1].wasconst=0;
7611 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7612 alloc_cc(&branch_regs[i-1],i-1);
7613 dirty_reg(&branch_regs[i-1],CCREG);
7614 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7615 if(dops[i-1].rt1!=0) { // JALR
7616 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7617 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7618 }
7619 #ifdef USE_MINI_HT
7620 if(dops[i-1].rs1==31) { // JALR
7621 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7622 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7623 }
7624 #endif
7625 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7626 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7627 break;
7628 case CJUMP:
7629 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7630 {
7631 alloc_cc(&current,i-1);
7632 dirty_reg(&current,CCREG);
7633 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7634 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7635 // The delay slot overwrote one of our conditions
7636 // Delay slot goes after the test (in order)
7637 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7638 current.u|=1;
7639 delayslot_alloc(&current,i);
7640 current.isconst=0;
7641 }
7642 else
7643 {
7644 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7645 // Alloc the branch condition registers
7646 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7647 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7648 }
7649 memcpy(&branch_regs[i-1],&current,sizeof(current));
7650 branch_regs[i-1].isconst=0;
7651 branch_regs[i-1].wasconst=0;
7652 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7653 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7654 }
7655 else
7656 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7657 {
7658 alloc_cc(&current,i-1);
7659 dirty_reg(&current,CCREG);
7660 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7661 // The delay slot overwrote the branch condition
7662 // Delay slot goes after the test (in order)
7663 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7664 current.u|=1;
7665 delayslot_alloc(&current,i);
7666 current.isconst=0;
7667 }
7668 else
7669 {
7670 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7671 // Alloc the branch condition register
7672 alloc_reg(&current,i-1,dops[i-1].rs1);
7673 }
7674 memcpy(&branch_regs[i-1],&current,sizeof(current));
7675 branch_regs[i-1].isconst=0;
7676 branch_regs[i-1].wasconst=0;
7677 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7678 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7679 }
7680 else
7681 // Alloc the delay slot in case the branch is taken
7682 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
7683 {
7684 memcpy(&branch_regs[i-1],&current,sizeof(current));
7685 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
7686 alloc_cc(&branch_regs[i-1],i);
7687 dirty_reg(&branch_regs[i-1],CCREG);
7688 delayslot_alloc(&branch_regs[i-1],i);
7689 branch_regs[i-1].isconst=0;
7690 alloc_reg(&current,i,CCREG); // Not taken path
7691 dirty_reg(&current,CCREG);
7692 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7693 }
7694 else
7695 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
7696 {
7697 memcpy(&branch_regs[i-1],&current,sizeof(current));
7698 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
7699 alloc_cc(&branch_regs[i-1],i);
7700 dirty_reg(&branch_regs[i-1],CCREG);
7701 delayslot_alloc(&branch_regs[i-1],i);
7702 branch_regs[i-1].isconst=0;
7703 alloc_reg(&current,i,CCREG); // Not taken path
7704 dirty_reg(&current,CCREG);
7705 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7706 }
7707 break;
7708 case SJUMP:
7709 {
7710 alloc_cc(&current,i-1);
7711 dirty_reg(&current,CCREG);
7712 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7713 // The delay slot overwrote the branch condition
7714 // Delay slot goes after the test (in order)
7715 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7716 current.u|=1;
7717 delayslot_alloc(&current,i);
7718 current.isconst=0;
7719 }
7720 else
7721 {
7722 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7723 // Alloc the branch condition register
7724 alloc_reg(&current,i-1,dops[i-1].rs1);
7725 }
7726 memcpy(&branch_regs[i-1],&current,sizeof(current));
7727 branch_regs[i-1].isconst=0;
7728 branch_regs[i-1].wasconst=0;
7729 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7730 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7731 }
7732 // FIXME: BLTZAL/BGEZAL
7733 if ((dops[i-1].opcode2 & 0x1e) == 0x10) { // BxxZAL
7734 alloc_reg(&branch_regs[i-1],i-1,31);
7735 dirty_reg(&branch_regs[i-1],31);
7736 }
7737 break;
7738 }
7739
7740 if (dops[i-1].is_ujump)
7741 {
7742 if(dops[i-1].rt1==31) // JAL/JALR
7743 {
7744 // Subroutine call will return here, don't alloc any registers
7745 current.dirty=0;
7746 clear_all_regs(current.regmap);
7747 alloc_reg(&current,i,CCREG);
7748 dirty_reg(&current,CCREG);
7749 }
7750 else if(i+1<slen)
7751 {
7752 // Internal branch will jump here, match registers to caller
7753 current.dirty=0;
7754 clear_all_regs(current.regmap);
7755 alloc_reg(&current,i,CCREG);
7756 dirty_reg(&current,CCREG);
7757 for(j=i-1;j>=0;j--)
7758 {
7759 if(cinfo[j].ba==start+i*4+4) {
7760 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7761 current.dirty=branch_regs[j].dirty;
7762 break;
7763 }
7764 }
7765 while(j>=0) {
7766 if(cinfo[j].ba==start+i*4+4) {
7767 for(hr=0;hr<HOST_REGS;hr++) {
7768 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7769 current.regmap[hr]=-1;
7770 }
7771 current.dirty&=branch_regs[j].dirty;
7772 }
7773 }
7774 j--;
7775 }
7776 }
7777 }
7778 }
7779
7780 // Count cycles in between branches
7781 cinfo[i].ccadj = CLOCK_ADJUST(cc);
7782 if (i > 0 && (dops[i-1].is_jump || dops[i].is_exception))
7783 {
7784 cc=0;
7785 }
7786#if !defined(DRC_DBG)
7787 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7788 {
7789 // this should really be removed since the real stalls have been implemented,
7790 // but doing so causes sizeable perf regression against the older version
7791 u_int gtec = gte_cycletab[source[i] & 0x3f];
7792 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7793 }
7794 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7795 {
7796 cc+=4;
7797 }
7798 else if(dops[i].itype==C2LS)
7799 {
7800 // same as with C2OP
7801 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7802 }
7803#endif
7804 else
7805 {
7806 cc++;
7807 }
7808
7809 if(!dops[i].is_ds) {
7810 regs[i].dirty=current.dirty;
7811 regs[i].isconst=current.isconst;
7812 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7813 }
7814 for(hr=0;hr<HOST_REGS;hr++) {
7815 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7816 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7817 regs[i].wasconst&=~(1<<hr);
7818 }
7819 }
7820 }
7821 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7822 //regs[i].waswritten=current.waswritten;
7823 }
7824}
7825
7826static noinline void pass4_cull_unused_regs(void)
7827{
7828 u_int last_needed_regs[4] = {0,0,0,0};
7829 u_int nr=0;
7830 int i;
7831
7832 for (i=slen-1;i>=0;i--)
7833 {
7834 int hr;
7835 __builtin_prefetch(regs[i-2].regmap);
7836 if(dops[i].is_jump)
7837 {
7838 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7839 {
7840 // Branch out of this block, don't need anything
7841 nr=0;
7842 }
7843 else
7844 {
7845 // Internal branch
7846 // Need whatever matches the target
7847 nr=0;
7848 int t=(cinfo[i].ba-start)>>2;
7849 for(hr=0;hr<HOST_REGS;hr++)
7850 {
7851 if(regs[i].regmap_entry[hr]>=0) {
7852 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7853 }
7854 }
7855 }
7856 // Conditional branch may need registers for following instructions
7857 if (!dops[i].is_ujump)
7858 {
7859 if(i<slen-2) {
7860 nr |= last_needed_regs[(i+2) & 3];
7861 for(hr=0;hr<HOST_REGS;hr++)
7862 {
7863 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7864 //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]);
7865 }
7866 }
7867 }
7868 // Don't need stuff which is overwritten
7869 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7870 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7871 // Merge in delay slot
7872 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
7873 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
7874 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
7875 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
7876 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
7877 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
7878 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
7879 nr |= get_regm(regmap_pre[i], ROREG);
7880 nr |= get_regm(regs[i].regmap_entry, ROREG);
7881 }
7882 if (dops[i+1].is_store) {
7883 nr |= get_regm(regmap_pre[i], INVCP);
7884 nr |= get_regm(regs[i].regmap_entry, INVCP);
7885 }
7886 }
7887 else if (dops[i].is_exception)
7888 {
7889 // SYSCALL instruction, etc
7890 nr=0;
7891 }
7892 else // Non-branch
7893 {
7894 if(i<slen-1) {
7895 for(hr=0;hr<HOST_REGS;hr++) {
7896 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7897 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7898 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7899 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7900 }
7901 }
7902 }
7903 // Overwritten registers are not needed
7904 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
7905 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
7906 nr &= ~get_regm(regs[i].regmap, FTEMP);
7907 // Source registers are needed
7908 nr |= get_regm(regmap_pre[i], dops[i].rs1);
7909 nr |= get_regm(regmap_pre[i], dops[i].rs2);
7910 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
7911 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
7912 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
7913 nr |= get_regm(regmap_pre[i], ROREG);
7914 nr |= get_regm(regs[i].regmap_entry, ROREG);
7915 }
7916 if (dops[i].is_store) {
7917 nr |= get_regm(regmap_pre[i], INVCP);
7918 nr |= get_regm(regs[i].regmap_entry, INVCP);
7919 }
7920
7921 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
7922 for(hr=0;hr<HOST_REGS;hr++)
7923 {
7924 // Don't store a register immediately after writing it,
7925 // may prevent dual-issue.
7926 // But do so if this is a branch target, otherwise we
7927 // might have to load the register before the branch.
7928 if((regs[i].wasdirty>>hr)&1) {
7929 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7930 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
7931 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
7932 }
7933 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7934 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
7935 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
7936 }
7937 }
7938 }
7939 // Cycle count is needed at branches. Assume it is needed at the target too.
7940 if(i==0||dops[i].bt||dops[i].itype==CJUMP) {
7941 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7942 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7943 }
7944 // Save it
7945 last_needed_regs[i & 3] = nr;
7946
7947 // Deallocate unneeded registers
7948 for(hr=0;hr<HOST_REGS;hr++)
7949 {
7950 if(!((nr>>hr)&1)) {
7951 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
7952 if(dops[i].is_jump)
7953 {
7954 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
7955 if (dops[i+1].is_load || dops[i+1].is_store)
7956 map1 = ROREG;
7957 if (dops[i+1].is_store)
7958 map2 = INVCP;
7959 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
7960 temp = FTEMP;
7961 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
7962 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
7963 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
7964 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
7965 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
7966 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
7967 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
7968 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
7969 {
7970 regs[i].regmap[hr]=-1;
7971 regs[i].isconst&=~(1<<hr);
7972 regs[i].dirty&=~(1<<hr);
7973 regs[i+1].wasdirty&=~(1<<hr);
7974 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
7975 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
7976 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
7977 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
7978 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
7979 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
7980 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
7981 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
7982 {
7983 branch_regs[i].regmap[hr]=-1;
7984 branch_regs[i].regmap_entry[hr]=-1;
7985 if (!dops[i].is_ujump)
7986 {
7987 if (i < slen-2) {
7988 regmap_pre[i+2][hr]=-1;
7989 regs[i+2].wasconst&=~(1<<hr);
7990 }
7991 }
7992 }
7993 }
7994 }
7995 else
7996 {
7997 // Non-branch
7998 if(i>0)
7999 {
8000 int map1 = -1, map2 = -1, temp=-1;
8001 if (dops[i].is_load || dops[i].is_store)
8002 map1 = ROREG;
8003 if (dops[i].is_store)
8004 map2 = INVCP;
8005 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8006 temp = FTEMP;
8007 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8008 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8009 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8010 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8011 regs[i].regmap[hr] != CCREG)
8012 {
8013 if(i<slen-1&&!dops[i].is_ds) {
8014 assert(regs[i].regmap[hr]<64);
8015 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8016 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8017 {
8018 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8019 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8020 }
8021 regmap_pre[i+1][hr]=-1;
8022 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8023 regs[i+1].wasconst&=~(1<<hr);
8024 }
8025 regs[i].regmap[hr]=-1;
8026 regs[i].isconst&=~(1<<hr);
8027 regs[i].dirty&=~(1<<hr);
8028 regs[i+1].wasdirty&=~(1<<hr);
8029 }
8030 }
8031 }
8032 } // if needed
8033 } // for hr
8034 }
8035}
8036
8037// If a register is allocated during a loop, try to allocate it for the
8038// entire loop, if possible. This avoids loading/storing registers
8039// inside of the loop.
8040static noinline void pass5a_preallocate1(void)
8041{
8042 int i, j, hr;
8043 signed char f_regmap[HOST_REGS];
8044 clear_all_regs(f_regmap);
8045 for(i=0;i<slen-1;i++)
8046 {
8047 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8048 {
8049 if(cinfo[i].ba>=start && cinfo[i].ba<(start+i*4))
8050 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8051 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8052 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8053 ||dops[i+1].itype==SHIFT
8054 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8055 {
8056 int t=(cinfo[i].ba-start)>>2;
8057 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8058 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8059 for(hr=0;hr<HOST_REGS;hr++)
8060 {
8061 if(regs[i].regmap[hr]>=0) {
8062 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8063 // dealloc old register
8064 int n;
8065 for(n=0;n<HOST_REGS;n++)
8066 {
8067 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8068 }
8069 // and alloc new one
8070 f_regmap[hr]=regs[i].regmap[hr];
8071 }
8072 }
8073 if(branch_regs[i].regmap[hr]>=0) {
8074 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8075 // dealloc old register
8076 int n;
8077 for(n=0;n<HOST_REGS;n++)
8078 {
8079 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8080 }
8081 // and alloc new one
8082 f_regmap[hr]=branch_regs[i].regmap[hr];
8083 }
8084 }
8085 if(dops[i].ooo) {
8086 if(count_free_regs(regs[i].regmap)<=cinfo[i+1].min_free_regs)
8087 f_regmap[hr]=branch_regs[i].regmap[hr];
8088 }else{
8089 if(count_free_regs(branch_regs[i].regmap)<=cinfo[i+1].min_free_regs)
8090 f_regmap[hr]=branch_regs[i].regmap[hr];
8091 }
8092 // Avoid dirty->clean transition
8093 #ifdef DESTRUCTIVE_WRITEBACK
8094 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;
8095 #endif
8096 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8097 // case above, however it's always a good idea. We can't hoist the
8098 // load if the register was already allocated, so there's no point
8099 // wasting time analyzing most of these cases. It only "succeeds"
8100 // when the mapping was different and the load can be replaced with
8101 // a mov, which is of negligible benefit. So such cases are
8102 // skipped below.
8103 if(f_regmap[hr]>0) {
8104 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8105 int r=f_regmap[hr];
8106 for(j=t;j<=i;j++)
8107 {
8108 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8109 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8110 assert(r < 64);
8111 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8112 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8113 int k;
8114 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8115 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8116 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8117 k=i;
8118 while(k>1&&regs[k-1].regmap[hr]==-1) {
8119 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8120 //printf("no free regs for store %x\n",start+(k-1)*4);
8121 break;
8122 }
8123 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8124 //printf("no-match due to different register\n");
8125 break;
8126 }
8127 if (dops[k-2].is_jump) {
8128 //printf("no-match due to branch\n");
8129 break;
8130 }
8131 // call/ret fast path assumes no registers allocated
8132 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8133 break;
8134 }
8135 k--;
8136 }
8137 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8138 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8139 while(k<i) {
8140 regs[k].regmap_entry[hr]=f_regmap[hr];
8141 regs[k].regmap[hr]=f_regmap[hr];
8142 regmap_pre[k+1][hr]=f_regmap[hr];
8143 regs[k].wasdirty&=~(1<<hr);
8144 regs[k].dirty&=~(1<<hr);
8145 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8146 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8147 regs[k].wasconst&=~(1<<hr);
8148 regs[k].isconst&=~(1<<hr);
8149 k++;
8150 }
8151 }
8152 else {
8153 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8154 break;
8155 }
8156 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8157 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8158 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8159 regs[i].regmap_entry[hr]=f_regmap[hr];
8160 regs[i].regmap[hr]=f_regmap[hr];
8161 regs[i].wasdirty&=~(1<<hr);
8162 regs[i].dirty&=~(1<<hr);
8163 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8164 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8165 regs[i].wasconst&=~(1<<hr);
8166 regs[i].isconst&=~(1<<hr);
8167 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8168 branch_regs[i].wasdirty&=~(1<<hr);
8169 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8170 branch_regs[i].regmap[hr]=f_regmap[hr];
8171 branch_regs[i].dirty&=~(1<<hr);
8172 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8173 branch_regs[i].wasconst&=~(1<<hr);
8174 branch_regs[i].isconst&=~(1<<hr);
8175 if (!dops[i].is_ujump) {
8176 regmap_pre[i+2][hr]=f_regmap[hr];
8177 regs[i+2].wasdirty&=~(1<<hr);
8178 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8179 }
8180 }
8181 }
8182 for(k=t;k<j;k++) {
8183 // Alloc register clean at beginning of loop,
8184 // but may dirty it in pass 6
8185 regs[k].regmap_entry[hr]=f_regmap[hr];
8186 regs[k].regmap[hr]=f_regmap[hr];
8187 regs[k].dirty&=~(1<<hr);
8188 regs[k].wasconst&=~(1<<hr);
8189 regs[k].isconst&=~(1<<hr);
8190 if (dops[k].is_jump) {
8191 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8192 branch_regs[k].regmap[hr]=f_regmap[hr];
8193 branch_regs[k].dirty&=~(1<<hr);
8194 branch_regs[k].wasconst&=~(1<<hr);
8195 branch_regs[k].isconst&=~(1<<hr);
8196 if (!dops[k].is_ujump) {
8197 regmap_pre[k+2][hr]=f_regmap[hr];
8198 regs[k+2].wasdirty&=~(1<<hr);
8199 }
8200 }
8201 else
8202 {
8203 regmap_pre[k+1][hr]=f_regmap[hr];
8204 regs[k+1].wasdirty&=~(1<<hr);
8205 }
8206 }
8207 if(regs[j].regmap[hr]==f_regmap[hr])
8208 regs[j].regmap_entry[hr]=f_regmap[hr];
8209 break;
8210 }
8211 if(j==i) break;
8212 if(regs[j].regmap[hr]>=0)
8213 break;
8214 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8215 //printf("no-match due to different register\n");
8216 break;
8217 }
8218 if (dops[j].is_ujump)
8219 {
8220 // Stop on unconditional branch
8221 break;
8222 }
8223 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8224 {
8225 if(dops[j].ooo) {
8226 if(count_free_regs(regs[j].regmap)<=cinfo[j+1].min_free_regs)
8227 break;
8228 }else{
8229 if(count_free_regs(branch_regs[j].regmap)<=cinfo[j+1].min_free_regs)
8230 break;
8231 }
8232 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8233 //printf("no-match due to different register (branch)\n");
8234 break;
8235 }
8236 }
8237 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8238 //printf("No free regs for store %x\n",start+j*4);
8239 break;
8240 }
8241 assert(f_regmap[hr]<64);
8242 }
8243 }
8244 }
8245 }
8246 }
8247 }else{
8248 // Non branch or undetermined branch target
8249 for(hr=0;hr<HOST_REGS;hr++)
8250 {
8251 if(hr!=EXCLUDE_REG) {
8252 if(regs[i].regmap[hr]>=0) {
8253 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8254 // dealloc old register
8255 int n;
8256 for(n=0;n<HOST_REGS;n++)
8257 {
8258 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8259 }
8260 // and alloc new one
8261 f_regmap[hr]=regs[i].regmap[hr];
8262 }
8263 }
8264 }
8265 }
8266 // Try to restore cycle count at branch targets
8267 if(dops[i].bt) {
8268 for(j=i;j<slen-1;j++) {
8269 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8270 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8271 //printf("no free regs for store %x\n",start+j*4);
8272 break;
8273 }
8274 }
8275 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8276 int k=i;
8277 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8278 while(k<j) {
8279 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8280 regs[k].regmap[HOST_CCREG]=CCREG;
8281 regmap_pre[k+1][HOST_CCREG]=CCREG;
8282 regs[k+1].wasdirty|=1<<HOST_CCREG;
8283 regs[k].dirty|=1<<HOST_CCREG;
8284 regs[k].wasconst&=~(1<<HOST_CCREG);
8285 regs[k].isconst&=~(1<<HOST_CCREG);
8286 k++;
8287 }
8288 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8289 }
8290 // Work backwards from the branch target
8291 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8292 {
8293 //printf("Extend backwards\n");
8294 int k;
8295 k=i;
8296 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8297 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8298 //printf("no free regs for store %x\n",start+(k-1)*4);
8299 break;
8300 }
8301 k--;
8302 }
8303 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8304 //printf("Extend CC, %x ->\n",start+k*4);
8305 while(k<=i) {
8306 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8307 regs[k].regmap[HOST_CCREG]=CCREG;
8308 regmap_pre[k+1][HOST_CCREG]=CCREG;
8309 regs[k+1].wasdirty|=1<<HOST_CCREG;
8310 regs[k].dirty|=1<<HOST_CCREG;
8311 regs[k].wasconst&=~(1<<HOST_CCREG);
8312 regs[k].isconst&=~(1<<HOST_CCREG);
8313 k++;
8314 }
8315 }
8316 else {
8317 //printf("Fail Extend CC, %x ->\n",start+k*4);
8318 }
8319 }
8320 }
8321 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=SHIFT&&
8322 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8323 dops[i].itype!=IMM16&&dops[i].itype!=LOAD)
8324 {
8325 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8326 }
8327 }
8328 }
8329}
8330
8331// This allocates registers (if possible) one instruction prior
8332// to use, which can avoid a load-use penalty on certain CPUs.
8333static noinline void pass5b_preallocate2(void)
8334{
8335 int i, hr;
8336 for(i=0;i<slen-1;i++)
8337 {
8338 if (!i || !dops[i-1].is_jump)
8339 {
8340 if(!dops[i+1].bt)
8341 {
8342 int j, can_steal = 1;
8343 for (j = i; j < i + 2; j++) {
8344 int free_regs = 0;
8345 if (cinfo[j].min_free_regs == 0)
8346 continue;
8347 for (hr = 0; hr < HOST_REGS; hr++)
8348 if (hr != EXCLUDE_REG && regs[j].regmap[hr] < 0)
8349 free_regs++;
8350 if (free_regs <= cinfo[j].min_free_regs) {
8351 can_steal = 0;
8352 break;
8353 }
8354 }
8355 if (!can_steal)
8356 continue;
8357 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8358 ||(dops[i].itype==COP2&&dops[i].opcode2<3))
8359 {
8360 if(dops[i+1].rs1) {
8361 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8362 {
8363 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8364 {
8365 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8366 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8367 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8368 regs[i].isconst&=~(1<<hr);
8369 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8370 constmap[i][hr]=constmap[i+1][hr];
8371 regs[i+1].wasdirty&=~(1<<hr);
8372 regs[i].dirty&=~(1<<hr);
8373 }
8374 }
8375 }
8376 if(dops[i+1].rs2) {
8377 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8378 {
8379 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8380 {
8381 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8382 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8383 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8384 regs[i].isconst&=~(1<<hr);
8385 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8386 constmap[i][hr]=constmap[i+1][hr];
8387 regs[i+1].wasdirty&=~(1<<hr);
8388 regs[i].dirty&=~(1<<hr);
8389 }
8390 }
8391 }
8392 // Preload target address for load instruction (non-constant)
8393 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8394 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8395 {
8396 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8397 {
8398 regs[i].regmap[hr]=dops[i+1].rs1;
8399 regmap_pre[i+1][hr]=dops[i+1].rs1;
8400 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8401 regs[i].isconst&=~(1<<hr);
8402 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8403 constmap[i][hr]=constmap[i+1][hr];
8404 regs[i+1].wasdirty&=~(1<<hr);
8405 regs[i].dirty&=~(1<<hr);
8406 }
8407 }
8408 }
8409 // Load source into target register
8410 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8411 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8412 {
8413 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8414 {
8415 regs[i].regmap[hr]=dops[i+1].rs1;
8416 regmap_pre[i+1][hr]=dops[i+1].rs1;
8417 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8418 regs[i].isconst&=~(1<<hr);
8419 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8420 constmap[i][hr]=constmap[i+1][hr];
8421 regs[i+1].wasdirty&=~(1<<hr);
8422 regs[i].dirty&=~(1<<hr);
8423 }
8424 }
8425 }
8426 // Address for store instruction (non-constant)
8427 if (dops[i+1].is_store) { // SB/SH/SW/SWC2
8428 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8429 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8430 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8431 else {
8432 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8433 regs[i+1].isconst&=~(1<<hr);
8434 regs[i+1].dirty&=~(1<<hr);
8435 regs[i+2].wasdirty&=~(1<<hr);
8436 }
8437 assert(hr>=0);
8438 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8439 {
8440 regs[i].regmap[hr]=dops[i+1].rs1;
8441 regmap_pre[i+1][hr]=dops[i+1].rs1;
8442 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8443 regs[i].isconst&=~(1<<hr);
8444 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8445 constmap[i][hr]=constmap[i+1][hr];
8446 regs[i+1].wasdirty&=~(1<<hr);
8447 regs[i].dirty&=~(1<<hr);
8448 }
8449 }
8450 }
8451 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) { // LWC2
8452 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8453 int nr;
8454 hr=get_reg(regs[i+1].regmap,FTEMP);
8455 assert(hr>=0);
8456 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8457 {
8458 regs[i].regmap[hr]=dops[i+1].rs1;
8459 regmap_pre[i+1][hr]=dops[i+1].rs1;
8460 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8461 regs[i].isconst&=~(1<<hr);
8462 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8463 constmap[i][hr]=constmap[i+1][hr];
8464 regs[i+1].wasdirty&=~(1<<hr);
8465 regs[i].dirty&=~(1<<hr);
8466 }
8467 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8468 {
8469 // move it to another register
8470 regs[i+1].regmap[hr]=-1;
8471 regmap_pre[i+2][hr]=-1;
8472 regs[i+1].regmap[nr]=FTEMP;
8473 regmap_pre[i+2][nr]=FTEMP;
8474 regs[i].regmap[nr]=dops[i+1].rs1;
8475 regmap_pre[i+1][nr]=dops[i+1].rs1;
8476 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8477 regs[i].isconst&=~(1<<nr);
8478 regs[i+1].isconst&=~(1<<nr);
8479 regs[i].dirty&=~(1<<nr);
8480 regs[i+1].wasdirty&=~(1<<nr);
8481 regs[i+1].dirty&=~(1<<nr);
8482 regs[i+2].wasdirty&=~(1<<nr);
8483 }
8484 }
8485 }
8486 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*/) {
8487 hr = -1;
8488 if(dops[i+1].itype==LOAD)
8489 hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1);
8490 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) // LWC2
8491 hr=get_reg(regs[i+1].regmap,FTEMP);
8492 if (dops[i+1].is_store) {
8493 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8494 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8495 }
8496 if(hr>=0&&regs[i].regmap[hr]<0) {
8497 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8498 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8499 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8500 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8501 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8502 regs[i].isconst&=~(1<<hr);
8503 regs[i+1].wasdirty&=~(1<<hr);
8504 regs[i].dirty&=~(1<<hr);
8505 }
8506 }
8507 }
8508 }
8509 }
8510 }
8511 }
8512}
8513
8514// Write back dirty registers as soon as we will no longer modify them,
8515// so that we don't end up with lots of writes at the branches.
8516static noinline void pass6_clean_registers(int istart, int iend, int wr)
8517{
8518 static u_int wont_dirty[MAXBLOCK];
8519 static u_int will_dirty[MAXBLOCK];
8520 int i;
8521 int r;
8522 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8523 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8524 if(iend==slen-1) {
8525 will_dirty_i=will_dirty_next=0;
8526 wont_dirty_i=wont_dirty_next=0;
8527 }else{
8528 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8529 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8530 }
8531 for (i=iend;i>=istart;i--)
8532 {
8533 signed char rregmap_i[RRMAP_SIZE];
8534 u_int hr_candirty = 0;
8535 assert(HOST_REGS < 32);
8536 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8537 __builtin_prefetch(regs[i-1].regmap);
8538 if(dops[i].is_jump)
8539 {
8540 signed char branch_rregmap_i[RRMAP_SIZE];
8541 u_int branch_hr_candirty = 0;
8542 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8543 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
8544 {
8545 // Branch out of this block, flush all regs
8546 will_dirty_i = 0;
8547 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8548 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8549 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8550 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8551 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8552 will_dirty_i &= branch_hr_candirty;
8553 if (dops[i].is_ujump)
8554 {
8555 // Unconditional branch
8556 wont_dirty_i = 0;
8557 // Merge in delay slot (will dirty)
8558 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8559 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8560 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8561 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8562 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8563 will_dirty_i &= hr_candirty;
8564 }
8565 else
8566 {
8567 // Conditional branch
8568 wont_dirty_i = wont_dirty_next;
8569 // Merge in delay slot (will dirty)
8570 // (the original code had no explanation why these 2 are commented out)
8571 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8572 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8573 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8574 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8575 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8576 will_dirty_i &= hr_candirty;
8577 }
8578 // Merge in delay slot (wont dirty)
8579 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8580 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8581 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8582 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8583 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8584 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8585 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8586 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8587 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8588 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8589 wont_dirty_i &= ~(1u << 31);
8590 if(wr) {
8591 #ifndef DESTRUCTIVE_WRITEBACK
8592 branch_regs[i].dirty&=wont_dirty_i;
8593 #endif
8594 branch_regs[i].dirty|=will_dirty_i;
8595 }
8596 }
8597 else
8598 {
8599 // Internal branch
8600 if(cinfo[i].ba<=start+i*4) {
8601 // Backward branch
8602 if (dops[i].is_ujump)
8603 {
8604 // Unconditional branch
8605 temp_will_dirty=0;
8606 temp_wont_dirty=0;
8607 // Merge in delay slot (will dirty)
8608 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8609 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8610 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8611 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8612 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8613 temp_will_dirty &= branch_hr_candirty;
8614 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8615 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8616 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8617 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8618 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8619 temp_will_dirty &= hr_candirty;
8620 } else {
8621 // Conditional branch (not taken case)
8622 temp_will_dirty=will_dirty_next;
8623 temp_wont_dirty=wont_dirty_next;
8624 // Merge in delay slot (will dirty)
8625 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8626 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8627 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8628 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8629 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8630 temp_will_dirty &= branch_hr_candirty;
8631 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8632 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8633 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8634 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8635 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8636 temp_will_dirty &= hr_candirty;
8637 }
8638 // Merge in delay slot (wont dirty)
8639 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8640 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8641 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8642 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8643 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8644 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8645 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8646 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8647 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8648 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8649 temp_wont_dirty &= ~(1u << 31);
8650 // Deal with changed mappings
8651 if(i<iend) {
8652 for(r=0;r<HOST_REGS;r++) {
8653 if(r!=EXCLUDE_REG) {
8654 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8655 temp_will_dirty&=~(1<<r);
8656 temp_wont_dirty&=~(1<<r);
8657 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8658 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8659 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8660 } else {
8661 temp_will_dirty|=1<<r;
8662 temp_wont_dirty|=1<<r;
8663 }
8664 }
8665 }
8666 }
8667 }
8668 if(wr) {
8669 will_dirty[i]=temp_will_dirty;
8670 wont_dirty[i]=temp_wont_dirty;
8671 pass6_clean_registers((cinfo[i].ba-start)>>2,i-1,0);
8672 }else{
8673 // Limit recursion. It can take an excessive amount
8674 // of time if there are a lot of nested loops.
8675 will_dirty[(cinfo[i].ba-start)>>2]=0;
8676 wont_dirty[(cinfo[i].ba-start)>>2]=-1;
8677 }
8678 }
8679 /*else*/ if(1)
8680 {
8681 if (dops[i].is_ujump)
8682 {
8683 // Unconditional branch
8684 will_dirty_i=0;
8685 wont_dirty_i=0;
8686 //if(cinfo[i].ba>start+i*4) { // Disable recursion (for debugging)
8687 for(r=0;r<HOST_REGS;r++) {
8688 if(r!=EXCLUDE_REG) {
8689 if(branch_regs[i].regmap[r]==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8690 will_dirty_i|=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8691 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8692 }
8693 if(branch_regs[i].regmap[r]>=0) {
8694 will_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8695 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8696 }
8697 }
8698 }
8699 //}
8700 // Merge in delay slot
8701 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8702 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8703 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8704 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8705 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8706 will_dirty_i &= branch_hr_candirty;
8707 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8708 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8709 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8710 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8711 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8712 will_dirty_i &= hr_candirty;
8713 } else {
8714 // Conditional branch
8715 will_dirty_i=will_dirty_next;
8716 wont_dirty_i=wont_dirty_next;
8717 //if(cinfo[i].ba>start+i*4) // Disable recursion (for debugging)
8718 for(r=0;r<HOST_REGS;r++) {
8719 if(r!=EXCLUDE_REG) {
8720 signed char target_reg=branch_regs[i].regmap[r];
8721 if(target_reg==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8722 will_dirty_i&=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8723 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8724 }
8725 else if(target_reg>=0) {
8726 will_dirty_i&=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8727 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8728 }
8729 }
8730 }
8731 // Merge in delay slot
8732 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8733 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8734 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8735 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8736 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8737 will_dirty_i &= branch_hr_candirty;
8738 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8739 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8740 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8741 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8742 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8743 will_dirty_i &= hr_candirty;
8744 }
8745 // Merge in delay slot (won't dirty)
8746 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8747 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8748 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8749 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8750 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8751 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8752 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8753 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8754 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8755 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8756 wont_dirty_i &= ~(1u << 31);
8757 if(wr) {
8758 #ifndef DESTRUCTIVE_WRITEBACK
8759 branch_regs[i].dirty&=wont_dirty_i;
8760 #endif
8761 branch_regs[i].dirty|=will_dirty_i;
8762 }
8763 }
8764 }
8765 }
8766 else if (dops[i].is_exception)
8767 {
8768 // SYSCALL instruction, etc
8769 will_dirty_i=0;
8770 wont_dirty_i=0;
8771 }
8772 will_dirty_next=will_dirty_i;
8773 wont_dirty_next=wont_dirty_i;
8774 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8775 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8776 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8777 will_dirty_i &= hr_candirty;
8778 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8779 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8780 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8781 wont_dirty_i &= ~(1u << 31);
8782 if (i > istart && !dops[i].is_jump) {
8783 // Don't store a register immediately after writing it,
8784 // may prevent dual-issue.
8785 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8786 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8787 }
8788 // Save it
8789 will_dirty[i]=will_dirty_i;
8790 wont_dirty[i]=wont_dirty_i;
8791 // Mark registers that won't be dirtied as not dirty
8792 if(wr) {
8793 regs[i].dirty|=will_dirty_i;
8794 #ifndef DESTRUCTIVE_WRITEBACK
8795 regs[i].dirty&=wont_dirty_i;
8796 if(dops[i].is_jump)
8797 {
8798 if (i < iend-1 && !dops[i].is_ujump) {
8799 for(r=0;r<HOST_REGS;r++) {
8800 if(r!=EXCLUDE_REG) {
8801 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8802 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8803 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8804 }
8805 }
8806 }
8807 }
8808 else
8809 {
8810 if(i<iend) {
8811 for(r=0;r<HOST_REGS;r++) {
8812 if(r!=EXCLUDE_REG) {
8813 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8814 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8815 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8816 }
8817 }
8818 }
8819 }
8820 #endif
8821 }
8822 // Deal with changed mappings
8823 temp_will_dirty=will_dirty_i;
8824 temp_wont_dirty=wont_dirty_i;
8825 for(r=0;r<HOST_REGS;r++) {
8826 if(r!=EXCLUDE_REG) {
8827 int nr;
8828 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8829 if(wr) {
8830 #ifndef DESTRUCTIVE_WRITEBACK
8831 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8832 #endif
8833 regs[i].wasdirty|=will_dirty_i&(1<<r);
8834 }
8835 }
8836 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8837 // Register moved to a different register
8838 will_dirty_i&=~(1<<r);
8839 wont_dirty_i&=~(1<<r);
8840 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8841 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8842 if(wr) {
8843 #ifndef DESTRUCTIVE_WRITEBACK
8844 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8845 #endif
8846 regs[i].wasdirty|=will_dirty_i&(1<<r);
8847 }
8848 }
8849 else {
8850 will_dirty_i&=~(1<<r);
8851 wont_dirty_i&=~(1<<r);
8852 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8853 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8854 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8855 } else {
8856 wont_dirty_i|=1<<r;
8857 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8858 }
8859 }
8860 }
8861 }
8862 }
8863}
8864
8865static noinline void pass10_expire_blocks(void)
8866{
8867 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
8868 // not sizeof(ndrc->translation_cache) due to vita hack
8869 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
8870 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
8871 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
8872 int hit;
8873
8874 for (; expirep != end; expirep = ((expirep + step) & step_mask))
8875 {
8876 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
8877 u_int block_i = expirep / step & (PAGE_COUNT - 1);
8878 u_int phase = (expirep >> (base_shift - 1)) & 1u;
8879 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
8880 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
8881 (long)(out - ndrc->translation_cache), phase);
8882 }
8883
8884 if (!phase) {
8885 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
8886 if (hit) {
8887 do_clear_cache();
8888 #ifdef USE_MINI_HT
8889 memset(mini_ht, -1, sizeof(mini_ht));
8890 #endif
8891 }
8892 }
8893 else
8894 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
8895 }
8896}
8897
8898static struct block_info *new_block_info(u_int start, u_int len,
8899 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
8900{
8901 struct block_info **b_pptr;
8902 struct block_info *block;
8903 u_int page = get_page(start);
8904
8905 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
8906 assert(block);
8907 assert(jump_in_count > 0);
8908 block->source = source;
8909 block->copy = copy;
8910 block->start = start;
8911 block->len = len;
8912 block->reg_sv_flags = 0;
8913 block->tc_offs = beginning - ndrc->translation_cache;
8914 //block->tc_len = out - beginning;
8915 block->is_dirty = 0;
8916 block->inv_near_misses = 0;
8917 block->jump_in_cnt = jump_in_count;
8918
8919 // insert sorted by start mirror-unmasked vaddr
8920 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
8921 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
8922 block->next = *b_pptr;
8923 *b_pptr = block;
8924 break;
8925 }
8926 }
8927 stat_inc(stat_blocks);
8928 return block;
8929}
8930
8931static int new_recompile_block(u_int addr)
8932{
8933 u_int pagelimit = 0;
8934 u_int state_rflags = 0;
8935 int i;
8936
8937 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
8938
8939 if (addr & 3) {
8940 if (addr != hack_addr) {
8941 SysPrintf("game crash @%08x, ra=%08x\n", addr, psxRegs.GPR.n.ra);
8942 hack_addr = addr;
8943 }
8944 return -1;
8945 }
8946
8947 // this is just for speculation
8948 for (i = 1; i < 32; i++) {
8949 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
8950 state_rflags |= 1 << i;
8951 }
8952
8953 start = addr;
8954 new_dynarec_did_compile=1;
8955 if (Config.HLE && start == 0x80001000) // hlecall
8956 {
8957 // XXX: is this enough? Maybe check hleSoftCall?
8958 void *beginning = start_block();
8959
8960 emit_movimm(start,0);
8961 emit_writeword(0,&pcaddr);
8962 emit_far_jump(new_dyna_leave);
8963 literal_pool(0);
8964 end_block(beginning);
8965 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8966 block->jump_in[0].vaddr = start;
8967 block->jump_in[0].addr = beginning;
8968 return 0;
8969 }
8970 else if (f1_hack && hack_addr == 0) {
8971 void *beginning = start_block();
8972 emit_movimm(start, 0);
8973 emit_writeword(0, &hack_addr);
8974 emit_readword(&psxRegs.GPR.n.sp, 0);
8975 emit_readptr(&mem_rtab, 1);
8976 emit_shrimm(0, 12, 2);
8977 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
8978 emit_addimm(0, 0x18, 0);
8979 emit_adds_ptr(1, 1, 1);
8980 emit_ldr_dualindexed(1, 0, 0);
8981 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
8982 emit_far_call(ndrc_get_addr_ht);
8983 emit_jmpreg(0); // jr k0
8984 literal_pool(0);
8985 end_block(beginning);
8986
8987 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8988 block->jump_in[0].vaddr = start;
8989 block->jump_in[0].addr = beginning;
8990 SysPrintf("F1 hack to %08x\n", start);
8991 return 0;
8992 }
8993
8994 cycle_multiplier_active = Config.cycle_multiplier_override && Config.cycle_multiplier == CYCLE_MULT_DEFAULT
8995 ? Config.cycle_multiplier_override : Config.cycle_multiplier;
8996
8997 source = get_source_start(start, &pagelimit);
8998 if (source == NULL) {
8999 if (addr != hack_addr) {
9000 SysPrintf("Compile at bogus memory address: %08x\n", addr);
9001 hack_addr = addr;
9002 }
9003 //abort();
9004 return -1;
9005 }
9006
9007 /* Pass 1: disassemble */
9008 /* Pass 2: register dependencies, branch targets */
9009 /* Pass 3: register allocation */
9010 /* Pass 4: branch dependencies */
9011 /* Pass 5: pre-alloc */
9012 /* Pass 6: optimize clean/dirty state */
9013 /* Pass 7: flag 32-bit registers */
9014 /* Pass 8: assembly */
9015 /* Pass 9: linker */
9016 /* Pass 10: garbage collection / free memory */
9017
9018 /* Pass 1 disassembly */
9019
9020 pass1_disassemble(pagelimit);
9021
9022 int clear_hack_addr = apply_hacks();
9023
9024 /* Pass 2 - Register dependencies and branch targets */
9025
9026 pass2_unneeded_regs(0,slen-1,0);
9027
9028 /* Pass 3 - Register allocation */
9029
9030 pass3_register_alloc(addr);
9031
9032 /* Pass 4 - Cull unused host registers */
9033
9034 pass4_cull_unused_regs();
9035
9036 /* Pass 5 - Pre-allocate registers */
9037
9038 pass5a_preallocate1();
9039 pass5b_preallocate2();
9040
9041 /* Pass 6 - Optimize clean/dirty state */
9042 pass6_clean_registers(0, slen-1, 1);
9043
9044 /* Pass 7 - Identify 32-bit registers */
9045 for (i=slen-1;i>=0;i--)
9046 {
9047 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9048 {
9049 // Conditional branch
9050 if((source[i]>>16)!=0x1000&&i<slen-2) {
9051 // Mark this address as a branch target since it may be called
9052 // upon return from interrupt
9053 dops[i+2].bt=1;
9054 }
9055 }
9056 }
9057
9058 /* Pass 8 - Assembly */
9059 linkcount=0;stubcount=0;
9060 is_delayslot=0;
9061 u_int dirty_pre=0;
9062 void *beginning=start_block();
9063 void *instr_addr0_override = NULL;
9064 int ds = 0;
9065
9066 if (start == 0x80030000) {
9067 // nasty hack for the fastbios thing
9068 // override block entry to this code
9069 instr_addr0_override = out;
9070 emit_movimm(start,0);
9071 // abuse io address var as a flag that we
9072 // have already returned here once
9073 emit_readword(&address,1);
9074 emit_writeword(0,&pcaddr);
9075 emit_writeword(0,&address);
9076 emit_cmp(0,1);
9077 #ifdef __aarch64__
9078 emit_jeq(out + 4*2);
9079 emit_far_jump(new_dyna_leave);
9080 #else
9081 emit_jne(new_dyna_leave);
9082 #endif
9083 }
9084 for(i=0;i<slen;i++)
9085 {
9086 __builtin_prefetch(regs[i+1].regmap);
9087 check_regmap(regmap_pre[i]);
9088 check_regmap(regs[i].regmap_entry);
9089 check_regmap(regs[i].regmap);
9090 //if(ds) printf("ds: ");
9091 disassemble_inst(i);
9092 if(ds) {
9093 ds=0; // Skip delay slot
9094 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9095 instr_addr[i] = NULL;
9096 } else {
9097 speculate_register_values(i);
9098 #ifndef DESTRUCTIVE_WRITEBACK
9099 if (i < 2 || !dops[i-2].is_ujump)
9100 {
9101 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9102 }
9103 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9104 dirty_pre=branch_regs[i].dirty;
9105 }else{
9106 dirty_pre=regs[i].dirty;
9107 }
9108 #endif
9109 // write back
9110 if (i < 2 || !dops[i-2].is_ujump)
9111 {
9112 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9113 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9114 }
9115 // branch target entry point
9116 instr_addr[i] = out;
9117 assem_debug("<->\n");
9118 drc_dbg_emit_do_cmp(i, cinfo[i].ccadj);
9119 if (clear_hack_addr) {
9120 emit_movimm(0, 0);
9121 emit_writeword(0, &hack_addr);
9122 clear_hack_addr = 0;
9123 }
9124
9125 // load regs
9126 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9127 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9128 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9129 address_generation(i,&regs[i],regs[i].regmap_entry);
9130 load_consts(regmap_pre[i],regs[i].regmap,i);
9131 if(dops[i].is_jump)
9132 {
9133 // Load the delay slot registers if necessary
9134 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))
9135 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9136 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))
9137 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9138 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9139 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9140 if (dops[i+1].is_store)
9141 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9142 }
9143 else if(i+1<slen)
9144 {
9145 // Preload registers for following instruction
9146 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9147 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9148 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9149 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9150 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9151 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9152 }
9153 // TODO: if(is_ooo(i)) address_generation(i+1);
9154 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9155 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9156 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9157 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9158 if (dops[i].is_store)
9159 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9160
9161 ds = assemble(i, &regs[i], cinfo[i].ccadj);
9162
9163 if (dops[i].is_ujump)
9164 literal_pool(1024);
9165 else
9166 literal_pool_jumpover(256);
9167 }
9168 }
9169
9170 assert(slen > 0);
9171 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9172 // no ending needed for this block since INTCALL never returns
9173 }
9174 // If the block did not end with an unconditional branch,
9175 // add a jump to the next instruction.
9176 else if (i > 1) {
9177 if (!dops[i-2].is_ujump) {
9178 assert(!dops[i-1].is_jump);
9179 assert(i==slen);
9180 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9181 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9182 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9183 emit_loadreg(CCREG,HOST_CCREG);
9184 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9185 }
9186 else
9187 {
9188 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9189 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9190 }
9191 add_to_linker(out,start+i*4,0);
9192 emit_jmp(0);
9193 }
9194 }
9195 else
9196 {
9197 assert(i>0);
9198 assert(!dops[i-1].is_jump);
9199 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9200 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9201 emit_loadreg(CCREG,HOST_CCREG);
9202 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9203 add_to_linker(out,start+i*4,0);
9204 emit_jmp(0);
9205 }
9206
9207 // Stubs
9208 for(i = 0; i < stubcount; i++)
9209 {
9210 switch(stubs[i].type)
9211 {
9212 case LOADB_STUB:
9213 case LOADH_STUB:
9214 case LOADW_STUB:
9215 case LOADBU_STUB:
9216 case LOADHU_STUB:
9217 do_readstub(i);break;
9218 case STOREB_STUB:
9219 case STOREH_STUB:
9220 case STOREW_STUB:
9221 do_writestub(i);break;
9222 case CC_STUB:
9223 do_ccstub(i);break;
9224 case INVCODE_STUB:
9225 do_invstub(i);break;
9226 case STORELR_STUB:
9227 do_unalignedwritestub(i);break;
9228 case OVERFLOW_STUB:
9229 do_overflowstub(i); break;
9230 case ALIGNMENT_STUB:
9231 do_alignmentstub(i); break;
9232 default:
9233 assert(0);
9234 }
9235 }
9236
9237 if (instr_addr0_override)
9238 instr_addr[0] = instr_addr0_override;
9239
9240#if 0
9241 /* check for improper expiration */
9242 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9243 int j;
9244 if (!jumps[i])
9245 continue;
9246 for (j = 0; j < jumps[i]->count; j++)
9247 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9248 }
9249#endif
9250
9251 /* Pass 9 - Linker */
9252 for(i=0;i<linkcount;i++)
9253 {
9254 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9255 literal_pool(64);
9256 if (!link_addr[i].internal)
9257 {
9258 void *stub = out;
9259 void *addr = check_addr(link_addr[i].target);
9260 emit_extjump(link_addr[i].addr, link_addr[i].target);
9261 if (addr) {
9262 set_jump_target(link_addr[i].addr, addr);
9263 ndrc_add_jump_out(link_addr[i].target,stub);
9264 }
9265 else
9266 set_jump_target(link_addr[i].addr, stub);
9267 }
9268 else
9269 {
9270 // Internal branch
9271 int target=(link_addr[i].target-start)>>2;
9272 assert(target>=0&&target<slen);
9273 assert(instr_addr[target]);
9274 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9275 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9276 //#else
9277 set_jump_target(link_addr[i].addr, instr_addr[target]);
9278 //#endif
9279 }
9280 }
9281
9282 u_int source_len = slen*4;
9283 if (dops[slen-1].itype == INTCALL && source_len > 4)
9284 // no need to treat the last instruction as compiled
9285 // as interpreter fully handles it
9286 source_len -= 4;
9287
9288 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9289 copy = shadow;
9290
9291 // External Branch Targets (jump_in)
9292 int jump_in_count = 1;
9293 assert(instr_addr[0]);
9294 for (i = 1; i < slen; i++)
9295 {
9296 if (dops[i].bt && instr_addr[i])
9297 jump_in_count++;
9298 }
9299
9300 struct block_info *block =
9301 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9302 block->reg_sv_flags = state_rflags;
9303
9304 int jump_in_i = 0;
9305 for (i = 0; i < slen; i++)
9306 {
9307 if ((i == 0 || dops[i].bt) && instr_addr[i])
9308 {
9309 assem_debug("%p (%d) <- %8x\n", instr_addr[i], i, start + i*4);
9310 u_int vaddr = start + i*4;
9311
9312 literal_pool(256);
9313 void *entry = out;
9314 load_regs_entry(i);
9315 if (entry == out)
9316 entry = instr_addr[i];
9317 else
9318 emit_jmp(instr_addr[i]);
9319
9320 block->jump_in[jump_in_i].vaddr = vaddr;
9321 block->jump_in[jump_in_i].addr = entry;
9322 jump_in_i++;
9323 }
9324 }
9325 assert(jump_in_i == jump_in_count);
9326 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9327 // Write out the literal pool if necessary
9328 literal_pool(0);
9329 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9330 // Align code
9331 if(((u_int)out)&7) emit_addnop(13);
9332 #endif
9333 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9334 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9335 memcpy(copy, source, source_len);
9336 copy += source_len;
9337
9338 end_block(beginning);
9339
9340 // If we're within 256K of the end of the buffer,
9341 // start over from the beginning. (Is 256K enough?)
9342 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9343 out = ndrc->translation_cache;
9344
9345 // Trap writes to any of the pages we compiled
9346 mark_invalid_code(start, slen*4, 0);
9347
9348 /* Pass 10 - Free memory by expiring oldest blocks */
9349
9350 pass10_expire_blocks();
9351
9352#ifdef ASSEM_PRINT
9353 fflush(stdout);
9354#endif
9355 stat_inc(stat_bc_direct);
9356 return 0;
9357}
9358
9359// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork