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 the interpreter (3)
[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 (CCREG)
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, u_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// this doesn't normally happen
752static noinline u_int generate_exception(u_int pc)
753{
754 //if (execBreakCheck(&psxRegs, pc))
755 // return psxRegs.pc;
756
757 // generate an address or bus error
758 psxRegs.CP0.n.Cause &= 0x300;
759 psxRegs.CP0.n.EPC = pc;
760 if (pc & 3) {
761 psxRegs.CP0.n.Cause |= R3000E_AdEL << 2;
762 psxRegs.CP0.n.BadVAddr = pc;
763#ifdef DRC_DBG
764 last_count -= 2;
765#endif
766 } else
767 psxRegs.CP0.n.Cause |= R3000E_IBE << 2;
768 return (psxRegs.pc = 0x80000080);
769}
770
771// Get address from virtual address
772// This is called from the recompiled JR/JALR instructions
773static void noinline *get_addr(u_int vaddr, int can_compile)
774{
775 u_int start_page = get_page_prev(vaddr);
776 u_int i, page, end_page = get_page(vaddr);
777 void *found_clean = NULL;
778
779 stat_inc(stat_jump_in_lookups);
780 for (page = start_page; page <= end_page; page++) {
781 const struct block_info *block;
782 for (block = blocks[page]; block != NULL; block = block->next) {
783 if (vaddr < block->start)
784 break;
785 if (block->is_dirty || vaddr >= block->start + block->len)
786 continue;
787 for (i = 0; i < block->jump_in_cnt; i++)
788 if (block->jump_in[i].vaddr == vaddr)
789 break;
790 if (i == block->jump_in_cnt)
791 continue;
792 found_clean = block->jump_in[i].addr;
793 hash_table_add(vaddr, found_clean);
794 return found_clean;
795 }
796 }
797 found_clean = try_restore_block(vaddr, start_page, end_page);
798 if (found_clean)
799 return found_clean;
800
801 if (!can_compile)
802 return NULL;
803
804 int r = new_recompile_block(vaddr);
805 if (likely(r == 0))
806 return ndrc_get_addr_ht(vaddr);
807
808 return ndrc_get_addr_ht(generate_exception(vaddr));
809}
810
811// Look up address in hash table first
812void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile)
813{
814 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
815 const struct ht_entry *ht_bin = hash_table_get(vaddr);
816 u_int vaddr_a = vaddr & ~3;
817 stat_inc(stat_ht_lookups);
818 if (ht_bin->vaddr[0] == vaddr_a) return ht_bin->tcaddr[0];
819 if (ht_bin->vaddr[1] == vaddr_a) return ht_bin->tcaddr[1];
820 return get_addr(vaddr, can_compile);
821}
822
823void *ndrc_get_addr_ht(u_int vaddr)
824{
825 return ndrc_get_addr_ht_param(vaddr, 1);
826}
827
828static void clear_all_regs(signed char regmap[])
829{
830 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
831}
832
833// get_reg: get allocated host reg from mips reg
834// returns -1 if no such mips reg was allocated
835#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
836
837extern signed char get_reg(const signed char regmap[], signed char r);
838
839#else
840
841static signed char get_reg(const signed char regmap[], signed char r)
842{
843 int hr;
844 for (hr = 0; hr < HOST_REGS; hr++) {
845 if (hr == EXCLUDE_REG)
846 continue;
847 if (regmap[hr] == r)
848 return hr;
849 }
850 return -1;
851}
852
853#endif
854
855// get reg suitable for writing
856static signed char get_reg_w(const signed char regmap[], signed char r)
857{
858 return r == 0 ? -1 : get_reg(regmap, r);
859}
860
861// get reg as mask bit (1 << hr)
862static u_int get_regm(const signed char regmap[], signed char r)
863{
864 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
865}
866
867static signed char get_reg_temp(const signed char regmap[])
868{
869 int hr;
870 for (hr = 0; hr < HOST_REGS; hr++) {
871 if (hr == EXCLUDE_REG)
872 continue;
873 if (regmap[hr] == (signed char)-1)
874 return hr;
875 }
876 return -1;
877}
878
879// Find a register that is available for two consecutive cycles
880static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
881{
882 int hr;
883 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
884 return -1;
885}
886
887// reverse reg map: mips -> host
888#define RRMAP_SIZE 64
889static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
890 u_int *regs_can_change)
891{
892 u_int r, hr, hr_can_change = 0;
893 memset(rrmap, -1, RRMAP_SIZE);
894 for (hr = 0; hr < HOST_REGS; )
895 {
896 r = regmap[hr];
897 rrmap[r & (RRMAP_SIZE - 1)] = hr;
898 // only add mips $1-$31+$lo, others shifted out
899 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
900 hr++;
901 if (hr == EXCLUDE_REG)
902 hr++;
903 }
904 hr_can_change |= 1u << (rrmap[33] & 31);
905 hr_can_change |= 1u << (rrmap[CCREG] & 31);
906 hr_can_change &= ~(1u << 31);
907 *regs_can_change = hr_can_change;
908}
909
910// same as get_reg, but takes rrmap
911static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
912{
913 assert(0 <= r && r < RRMAP_SIZE);
914 return rrmap[r];
915}
916
917static int count_free_regs(const signed char regmap[])
918{
919 int count=0;
920 int hr;
921 for(hr=0;hr<HOST_REGS;hr++)
922 {
923 if(hr!=EXCLUDE_REG) {
924 if(regmap[hr]<0) count++;
925 }
926 }
927 return count;
928}
929
930static void dirty_reg(struct regstat *cur, signed char reg)
931{
932 int hr;
933 if (!reg) return;
934 hr = get_reg(cur->regmap, reg);
935 if (hr >= 0)
936 cur->dirty |= 1<<hr;
937}
938
939static void set_const(struct regstat *cur, signed char reg, uint32_t value)
940{
941 int hr;
942 if (!reg) return;
943 hr = get_reg(cur->regmap, reg);
944 if (hr >= 0) {
945 cur->isconst |= 1<<hr;
946 current_constmap[hr] = value;
947 }
948}
949
950static void clear_const(struct regstat *cur, signed char reg)
951{
952 int hr;
953 if (!reg) return;
954 hr = get_reg(cur->regmap, reg);
955 if (hr >= 0)
956 cur->isconst &= ~(1<<hr);
957}
958
959static int is_const(const struct regstat *cur, signed char reg)
960{
961 int hr;
962 if (reg < 0) return 0;
963 if (!reg) return 1;
964 hr = get_reg(cur->regmap, reg);
965 if (hr >= 0)
966 return (cur->isconst>>hr)&1;
967 return 0;
968}
969
970static uint32_t get_const(const struct regstat *cur, signed char reg)
971{
972 int hr;
973 if (!reg) return 0;
974 hr = get_reg(cur->regmap, reg);
975 if (hr >= 0)
976 return current_constmap[hr];
977
978 SysPrintf("Unknown constant in r%d\n", reg);
979 abort();
980}
981
982// Least soon needed registers
983// Look at the next ten instructions and see which registers
984// will be used. Try not to reallocate these.
985static void lsn(u_char hsn[], int i, int *preferred_reg)
986{
987 int j;
988 int b=-1;
989 for(j=0;j<9;j++)
990 {
991 if(i+j>=slen) {
992 j=slen-i-1;
993 break;
994 }
995 if (dops[i+j].is_ujump)
996 {
997 // Don't go past an unconditonal jump
998 j++;
999 break;
1000 }
1001 }
1002 for(;j>=0;j--)
1003 {
1004 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
1005 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
1006 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
1007 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
1008 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
1009 // Stores can allocate zero
1010 hsn[dops[i+j].rs1]=j;
1011 hsn[dops[i+j].rs2]=j;
1012 }
1013 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
1014 hsn[ROREG] = j;
1015 // On some architectures stores need invc_ptr
1016 #if defined(HOST_IMM8)
1017 if (dops[i+j].is_store)
1018 hsn[INVCP] = j;
1019 #endif
1020 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1021 {
1022 hsn[CCREG]=j;
1023 b=j;
1024 }
1025 }
1026 if(b>=0)
1027 {
1028 if(cinfo[i+b].ba>=start && cinfo[i+b].ba<(start+slen*4))
1029 {
1030 // Follow first branch
1031 int t=(cinfo[i+b].ba-start)>>2;
1032 j=7-b;if(t+j>=slen) j=slen-t-1;
1033 for(;j>=0;j--)
1034 {
1035 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1036 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1037 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1038 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1039 }
1040 }
1041 // TODO: preferred register based on backward branch
1042 }
1043 // Delay slot should preferably not overwrite branch conditions or cycle count
1044 if (i > 0 && dops[i-1].is_jump) {
1045 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1046 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1047 hsn[CCREG]=1;
1048 // ...or hash tables
1049 hsn[RHASH]=1;
1050 hsn[RHTBL]=1;
1051 }
1052 // Coprocessor load/store needs FTEMP, even if not declared
1053 if(dops[i].itype==C2LS) {
1054 hsn[FTEMP]=0;
1055 }
1056 // Load L/R also uses FTEMP as a temporary register
1057 if(dops[i].itype==LOADLR) {
1058 hsn[FTEMP]=0;
1059 }
1060 // Also SWL/SWR/SDL/SDR
1061 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
1062 hsn[FTEMP]=0;
1063 }
1064 // Don't remove the miniht registers
1065 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1066 {
1067 hsn[RHASH]=0;
1068 hsn[RHTBL]=0;
1069 }
1070}
1071
1072// We only want to allocate registers if we're going to use them again soon
1073static int needed_again(int r, int i)
1074{
1075 int j;
1076 int b=-1;
1077 int rn=10;
1078
1079 if (i > 0 && dops[i-1].is_ujump)
1080 {
1081 if(cinfo[i-1].ba<start || cinfo[i-1].ba>start+slen*4-4)
1082 return 0; // Don't need any registers if exiting the block
1083 }
1084 for(j=0;j<9;j++)
1085 {
1086 if(i+j>=slen) {
1087 j=slen-i-1;
1088 break;
1089 }
1090 if (dops[i+j].is_ujump)
1091 {
1092 // Don't go past an unconditonal jump
1093 j++;
1094 break;
1095 }
1096 if (dops[i+j].is_exception)
1097 {
1098 break;
1099 }
1100 }
1101 for(;j>=1;j--)
1102 {
1103 if(dops[i+j].rs1==r) rn=j;
1104 if(dops[i+j].rs2==r) rn=j;
1105 if((unneeded_reg[i+j]>>r)&1) rn=10;
1106 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1107 {
1108 b=j;
1109 }
1110 }
1111 if(rn<10) return 1;
1112 (void)b;
1113 return 0;
1114}
1115
1116// Try to match register allocations at the end of a loop with those
1117// at the beginning
1118static int loop_reg(int i, int r, int hr)
1119{
1120 int j,k;
1121 for(j=0;j<9;j++)
1122 {
1123 if(i+j>=slen) {
1124 j=slen-i-1;
1125 break;
1126 }
1127 if (dops[i+j].is_ujump)
1128 {
1129 // Don't go past an unconditonal jump
1130 j++;
1131 break;
1132 }
1133 }
1134 k=0;
1135 if(i>0){
1136 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1137 k--;
1138 }
1139 for(;k<j;k++)
1140 {
1141 assert(r < 64);
1142 if((unneeded_reg[i+k]>>r)&1) return hr;
1143 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1144 {
1145 if(cinfo[i+k].ba>=start && cinfo[i+k].ba<(start+i*4))
1146 {
1147 int t=(cinfo[i+k].ba-start)>>2;
1148 int reg=get_reg(regs[t].regmap_entry,r);
1149 if(reg>=0) return reg;
1150 //reg=get_reg(regs[t+1].regmap_entry,r);
1151 //if(reg>=0) return reg;
1152 }
1153 }
1154 }
1155 return hr;
1156}
1157
1158
1159// Allocate every register, preserving source/target regs
1160static void alloc_all(struct regstat *cur,int i)
1161{
1162 int hr;
1163
1164 for(hr=0;hr<HOST_REGS;hr++) {
1165 if(hr!=EXCLUDE_REG) {
1166 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1167 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1168 {
1169 cur->regmap[hr]=-1;
1170 cur->dirty&=~(1<<hr);
1171 }
1172 // Don't need zeros
1173 if(cur->regmap[hr]==0)
1174 {
1175 cur->regmap[hr]=-1;
1176 cur->dirty&=~(1<<hr);
1177 }
1178 }
1179 }
1180}
1181
1182#ifndef NDEBUG
1183static int host_tempreg_in_use;
1184
1185static void host_tempreg_acquire(void)
1186{
1187 assert(!host_tempreg_in_use);
1188 host_tempreg_in_use = 1;
1189}
1190
1191static void host_tempreg_release(void)
1192{
1193 host_tempreg_in_use = 0;
1194}
1195#else
1196static void host_tempreg_acquire(void) {}
1197static void host_tempreg_release(void) {}
1198#endif
1199
1200#ifdef ASSEM_PRINT
1201extern void gen_interupt();
1202extern void do_insn_cmp();
1203#define FUNCNAME(f) { f, " " #f }
1204static const struct {
1205 void *addr;
1206 const char *name;
1207} function_names[] = {
1208 FUNCNAME(cc_interrupt),
1209 FUNCNAME(gen_interupt),
1210 FUNCNAME(ndrc_get_addr_ht),
1211 FUNCNAME(jump_handler_read8),
1212 FUNCNAME(jump_handler_read16),
1213 FUNCNAME(jump_handler_read32),
1214 FUNCNAME(jump_handler_write8),
1215 FUNCNAME(jump_handler_write16),
1216 FUNCNAME(jump_handler_write32),
1217 FUNCNAME(ndrc_write_invalidate_one),
1218 FUNCNAME(ndrc_write_invalidate_many),
1219 FUNCNAME(jump_to_new_pc),
1220 FUNCNAME(jump_break),
1221 FUNCNAME(jump_break_ds),
1222 FUNCNAME(jump_syscall),
1223 FUNCNAME(jump_syscall_ds),
1224 FUNCNAME(jump_overflow),
1225 FUNCNAME(jump_overflow_ds),
1226 FUNCNAME(jump_addrerror),
1227 FUNCNAME(jump_addrerror_ds),
1228 FUNCNAME(call_gteStall),
1229 FUNCNAME(new_dyna_leave),
1230 FUNCNAME(pcsx_mtc0),
1231 FUNCNAME(pcsx_mtc0_ds),
1232 FUNCNAME(execI),
1233#ifdef __aarch64__
1234 FUNCNAME(do_memhandler_pre),
1235 FUNCNAME(do_memhandler_post),
1236#endif
1237#ifdef DRC_DBG
1238 FUNCNAME(do_insn_cmp),
1239#endif
1240};
1241
1242static const char *func_name(const void *a)
1243{
1244 int i;
1245 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1246 if (function_names[i].addr == a)
1247 return function_names[i].name;
1248 return "";
1249}
1250
1251static const char *fpofs_name(u_int ofs)
1252{
1253 u_int *p = (u_int *)&dynarec_local + ofs/sizeof(u_int);
1254 static char buf[64];
1255 switch (ofs) {
1256 #define ofscase(x) case LO_##x: return " ; " #x
1257 ofscase(next_interupt);
1258 ofscase(cycle_count);
1259 ofscase(last_count);
1260 ofscase(pending_exception);
1261 ofscase(stop);
1262 ofscase(address);
1263 ofscase(lo);
1264 ofscase(hi);
1265 ofscase(PC);
1266 ofscase(cycle);
1267 ofscase(mem_rtab);
1268 ofscase(mem_wtab);
1269 ofscase(psxH_ptr);
1270 ofscase(invc_ptr);
1271 ofscase(ram_offset);
1272 #undef ofscase
1273 }
1274 buf[0] = 0;
1275 if (psxRegs.GPR.r <= p && p < &psxRegs.GPR.r[32])
1276 snprintf(buf, sizeof(buf), " ; r%d", (int)(p - psxRegs.GPR.r));
1277 else if (psxRegs.CP0.r <= p && p < &psxRegs.CP0.r[32])
1278 snprintf(buf, sizeof(buf), " ; cp0 $%d", (int)(p - psxRegs.CP0.r));
1279 else if (psxRegs.CP2D.r <= p && p < &psxRegs.CP2D.r[32])
1280 snprintf(buf, sizeof(buf), " ; cp2d $%d", (int)(p - psxRegs.CP2D.r));
1281 else if (psxRegs.CP2C.r <= p && p < &psxRegs.CP2C.r[32])
1282 snprintf(buf, sizeof(buf), " ; cp2c $%d", (int)(p - psxRegs.CP2C.r));
1283 return buf;
1284}
1285#else
1286#define func_name(x) ""
1287#define fpofs_name(x) ""
1288#endif
1289
1290#ifdef __i386__
1291#include "assem_x86.c"
1292#endif
1293#ifdef __x86_64__
1294#include "assem_x64.c"
1295#endif
1296#ifdef __arm__
1297#include "assem_arm.c"
1298#endif
1299#ifdef __aarch64__
1300#include "assem_arm64.c"
1301#endif
1302
1303static void *get_trampoline(const void *f)
1304{
1305 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1306 size_t i;
1307
1308 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1309 if (tramp->f[i] == f || tramp->f[i] == NULL)
1310 break;
1311 }
1312 if (i == ARRAY_SIZE(tramp->f)) {
1313 SysPrintf("trampoline table is full, last func %p\n", f);
1314 abort();
1315 }
1316 if (tramp->f[i] == NULL) {
1317 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1318 tramp->f[i] = f;
1319 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1320#ifdef HAVE_LIBNX
1321 // invalidate the RX mirror (unsure if necessary, but just in case...)
1322 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1323#endif
1324 }
1325 return &ndrc->tramp.ops[i];
1326}
1327
1328static void emit_far_jump(const void *f)
1329{
1330 if (can_jump_or_call(f)) {
1331 emit_jmp(f);
1332 return;
1333 }
1334
1335 f = get_trampoline(f);
1336 emit_jmp(f);
1337}
1338
1339static void emit_far_call(const void *f)
1340{
1341 if (can_jump_or_call(f)) {
1342 emit_call(f);
1343 return;
1344 }
1345
1346 f = get_trampoline(f);
1347 emit_call(f);
1348}
1349
1350// Check if an address is already compiled
1351// but don't return addresses which are about to expire from the cache
1352static void *check_addr(u_int vaddr)
1353{
1354 struct ht_entry *ht_bin = hash_table_get(vaddr);
1355 size_t i;
1356 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1357 if (ht_bin->vaddr[i] == vaddr)
1358 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1359 return ht_bin->tcaddr[i];
1360 }
1361
1362 // refactor to get_addr_nocompile?
1363 u_int start_page = get_page_prev(vaddr);
1364 u_int page, end_page = get_page(vaddr);
1365
1366 stat_inc(stat_jump_in_lookups);
1367 for (page = start_page; page <= end_page; page++) {
1368 const struct block_info *block;
1369 for (block = blocks[page]; block != NULL; block = block->next) {
1370 if (vaddr < block->start)
1371 break;
1372 if (block->is_dirty || vaddr >= block->start + block->len)
1373 continue;
1374 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1375 continue;
1376 for (i = 0; i < block->jump_in_cnt; i++)
1377 if (block->jump_in[i].vaddr == vaddr)
1378 break;
1379 if (i == block->jump_in_cnt)
1380 continue;
1381
1382 // Update existing entry with current address
1383 void *addr = block->jump_in[i].addr;
1384 if (ht_bin->vaddr[0] == vaddr) {
1385 ht_bin->tcaddr[0] = addr;
1386 return addr;
1387 }
1388 if (ht_bin->vaddr[1] == vaddr) {
1389 ht_bin->tcaddr[1] = addr;
1390 return addr;
1391 }
1392 // Insert into hash table with low priority.
1393 // Don't evict existing entries, as they are probably
1394 // addresses that are being accessed frequently.
1395 if (ht_bin->vaddr[0] == -1) {
1396 ht_bin->vaddr[0] = vaddr;
1397 ht_bin->tcaddr[0] = addr;
1398 }
1399 else if (ht_bin->vaddr[1] == -1) {
1400 ht_bin->vaddr[1] = vaddr;
1401 ht_bin->tcaddr[1] = addr;
1402 }
1403 return addr;
1404 }
1405 }
1406 return NULL;
1407}
1408
1409static void blocks_clear(struct block_info **head)
1410{
1411 struct block_info *cur, *next;
1412
1413 if ((cur = *head)) {
1414 *head = NULL;
1415 while (cur) {
1416 next = cur->next;
1417 free(cur);
1418 cur = next;
1419 }
1420 }
1421}
1422
1423static int blocks_remove_matching_addrs(struct block_info **head,
1424 u_int base_offs, int shift)
1425{
1426 struct block_info *next;
1427 int hit = 0;
1428 while (*head) {
1429 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1430 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1431 invalidate_block(*head);
1432 next = (*head)->next;
1433 free(*head);
1434 *head = next;
1435 stat_dec(stat_blocks);
1436 hit = 1;
1437 }
1438 else
1439 {
1440 head = &((*head)->next);
1441 }
1442 }
1443 return hit;
1444}
1445
1446// This is called when we write to a compiled block (see do_invstub)
1447static void unlink_jumps_vaddr_range(u_int start, u_int end)
1448{
1449 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1450 int i;
1451
1452 for (page = start_page; page <= end_page; page++) {
1453 struct jump_info *ji = jumps[page];
1454 if (ji == NULL)
1455 continue;
1456 for (i = 0; i < ji->count; ) {
1457 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1458 i++;
1459 continue;
1460 }
1461
1462 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1463 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1464 void *host_addr = find_extjump_insn(ji->e[i].stub);
1465 mark_clear_cache(host_addr);
1466 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1467
1468 stat_dec(stat_links);
1469 ji->count--;
1470 if (i < ji->count) {
1471 ji->e[i] = ji->e[ji->count];
1472 continue;
1473 }
1474 i++;
1475 }
1476 }
1477}
1478
1479static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1480{
1481 int i;
1482 if (ji == NULL)
1483 return;
1484 for (i = 0; i < ji->count; ) {
1485 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1486 if (((tc_offs ^ base_offs) >> shift) != 0) {
1487 i++;
1488 continue;
1489 }
1490
1491 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1492 stat_dec(stat_links);
1493 ji->count--;
1494 if (i < ji->count) {
1495 ji->e[i] = ji->e[ji->count];
1496 continue;
1497 }
1498 i++;
1499 }
1500}
1501
1502static void invalidate_block(struct block_info *block)
1503{
1504 u_int i;
1505
1506 block->is_dirty = 1;
1507 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1508 for (i = 0; i < block->jump_in_cnt; i++)
1509 hash_table_remove(block->jump_in[i].vaddr);
1510}
1511
1512static int invalidate_range(u_int start, u_int end,
1513 u32 *inv_start_ret, u32 *inv_end_ret)
1514{
1515 struct block_info *last_block = NULL;
1516 u_int start_page = get_page_prev(start);
1517 u_int end_page = get_page(end - 1);
1518 u_int start_m = pmmask(start);
1519 u_int end_m = pmmask(end - 1);
1520 u_int inv_start, inv_end;
1521 u_int blk_start_m, blk_end_m;
1522 u_int page;
1523 int hit = 0;
1524
1525 // additional area without code (to supplement invalid_code[]), [start, end)
1526 // avoids excessive ndrc_write_invalidate*() calls
1527 inv_start = start_m & ~0xfff;
1528 inv_end = end_m | 0xfff;
1529
1530 for (page = start_page; page <= end_page; page++) {
1531 struct block_info *block;
1532 for (block = blocks[page]; block != NULL; block = block->next) {
1533 if (block->is_dirty)
1534 continue;
1535 last_block = block;
1536 blk_end_m = pmmask(block->start + block->len);
1537 if (blk_end_m <= start_m) {
1538 inv_start = max(inv_start, blk_end_m);
1539 continue;
1540 }
1541 blk_start_m = pmmask(block->start);
1542 if (end_m <= blk_start_m) {
1543 inv_end = min(inv_end, blk_start_m - 1);
1544 continue;
1545 }
1546 if (!block->source) // "hack" block - leave it alone
1547 continue;
1548
1549 hit++;
1550 invalidate_block(block);
1551 stat_inc(stat_inv_hits);
1552 }
1553 }
1554
1555 if (!hit && last_block && last_block->source) {
1556 // could be some leftover unused block, uselessly trapping writes
1557 last_block->inv_near_misses++;
1558 if (last_block->inv_near_misses > 128) {
1559 invalidate_block(last_block);
1560 stat_inc(stat_inv_hits);
1561 hit++;
1562 }
1563 }
1564 if (hit) {
1565 do_clear_cache();
1566#ifdef USE_MINI_HT
1567 memset(mini_ht, -1, sizeof(mini_ht));
1568#endif
1569 }
1570
1571 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1572 // the whole page is empty now
1573 mark_invalid_code(start, 1, 1);
1574
1575 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1576 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1577 return hit;
1578}
1579
1580void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1581{
1582 invalidate_range(start, end, NULL, NULL);
1583}
1584
1585static void ndrc_write_invalidate_many(u_int start, u_int end)
1586{
1587 // this check is done by the caller
1588 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1589 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1590#ifdef INV_DEBUG_W
1591 int invc = invalid_code[start >> 12];
1592 u_int len = end - start;
1593 if (ret)
1594 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1595 else
1596 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1597 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1598 check_for_block_changes(start, end);
1599#endif
1600 stat_inc(stat_inv_addr_calls);
1601 (void)ret;
1602}
1603
1604void ndrc_write_invalidate_one(u_int addr)
1605{
1606 ndrc_write_invalidate_many(addr, addr + 4);
1607}
1608
1609// This is called when loading a save state.
1610// Anything could have changed, so invalidate everything.
1611void new_dynarec_invalidate_all_pages(void)
1612{
1613 struct block_info *block;
1614 u_int page;
1615 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1616 for (block = blocks[page]; block != NULL; block = block->next) {
1617 if (block->is_dirty)
1618 continue;
1619 if (!block->source) // hack block?
1620 continue;
1621 invalidate_block(block);
1622 }
1623 }
1624
1625 #ifdef USE_MINI_HT
1626 memset(mini_ht, -1, sizeof(mini_ht));
1627 #endif
1628 do_clear_cache();
1629}
1630
1631// Add an entry to jump_out after making a link
1632// src should point to code by emit_extjump()
1633void ndrc_add_jump_out(u_int vaddr, void *src)
1634{
1635 inv_debug("ndrc_add_jump_out: %p -> %x\n", src, vaddr);
1636 u_int page = get_page(vaddr);
1637 struct jump_info *ji;
1638
1639 stat_inc(stat_links);
1640 check_extjump2(src);
1641 ji = jumps[page];
1642 if (ji == NULL) {
1643 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1644 ji->alloc = 16;
1645 ji->count = 0;
1646 }
1647 else if (ji->count >= ji->alloc) {
1648 ji->alloc += 16;
1649 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1650 }
1651 jumps[page] = ji;
1652 ji->e[ji->count].target_vaddr = vaddr;
1653 ji->e[ji->count].stub = src;
1654 ji->count++;
1655}
1656
1657/* Register allocation */
1658
1659// Note: registers are allocated clean (unmodified state)
1660// if you intend to modify the register, you must call dirty_reg().
1661static void alloc_reg(struct regstat *cur,int i,signed char reg)
1662{
1663 int r,hr;
1664 int preferred_reg = PREFERRED_REG_FIRST
1665 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1666 if (reg == CCREG) preferred_reg = HOST_CCREG;
1667 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1668 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1669 assert(reg >= 0);
1670
1671 // Don't allocate unused registers
1672 if((cur->u>>reg)&1) return;
1673
1674 // see if it's already allocated
1675 if (get_reg(cur->regmap, reg) >= 0)
1676 return;
1677
1678 // Keep the same mapping if the register was already allocated in a loop
1679 preferred_reg = loop_reg(i,reg,preferred_reg);
1680
1681 // Try to allocate the preferred register
1682 if(cur->regmap[preferred_reg]==-1) {
1683 cur->regmap[preferred_reg]=reg;
1684 cur->dirty&=~(1<<preferred_reg);
1685 cur->isconst&=~(1<<preferred_reg);
1686 return;
1687 }
1688 r=cur->regmap[preferred_reg];
1689 assert(r < 64);
1690 if((cur->u>>r)&1) {
1691 cur->regmap[preferred_reg]=reg;
1692 cur->dirty&=~(1<<preferred_reg);
1693 cur->isconst&=~(1<<preferred_reg);
1694 return;
1695 }
1696
1697 // Clear any unneeded registers
1698 // We try to keep the mapping consistent, if possible, because it
1699 // makes branches easier (especially loops). So we try to allocate
1700 // first (see above) before removing old mappings. If this is not
1701 // possible then go ahead and clear out the registers that are no
1702 // longer needed.
1703 for(hr=0;hr<HOST_REGS;hr++)
1704 {
1705 r=cur->regmap[hr];
1706 if(r>=0) {
1707 assert(r < 64);
1708 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1709 }
1710 }
1711
1712 // Try to allocate any available register, but prefer
1713 // registers that have not been used recently.
1714 if (i > 0) {
1715 for (hr = PREFERRED_REG_FIRST; ; ) {
1716 if (cur->regmap[hr] < 0) {
1717 int oldreg = regs[i-1].regmap[hr];
1718 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1719 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1720 {
1721 cur->regmap[hr]=reg;
1722 cur->dirty&=~(1<<hr);
1723 cur->isconst&=~(1<<hr);
1724 return;
1725 }
1726 }
1727 hr++;
1728 if (hr == EXCLUDE_REG)
1729 hr++;
1730 if (hr == HOST_REGS)
1731 hr = 0;
1732 if (hr == PREFERRED_REG_FIRST)
1733 break;
1734 }
1735 }
1736
1737 // Try to allocate any available register
1738 for (hr = PREFERRED_REG_FIRST; ; ) {
1739 if (cur->regmap[hr] < 0) {
1740 cur->regmap[hr]=reg;
1741 cur->dirty&=~(1<<hr);
1742 cur->isconst&=~(1<<hr);
1743 return;
1744 }
1745 hr++;
1746 if (hr == EXCLUDE_REG)
1747 hr++;
1748 if (hr == HOST_REGS)
1749 hr = 0;
1750 if (hr == PREFERRED_REG_FIRST)
1751 break;
1752 }
1753
1754 // Ok, now we have to evict someone
1755 // Pick a register we hopefully won't need soon
1756 u_char hsn[MAXREG+1];
1757 memset(hsn,10,sizeof(hsn));
1758 int j;
1759 lsn(hsn,i,&preferred_reg);
1760 //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]);
1761 //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]);
1762 if(i>0) {
1763 // Don't evict the cycle count at entry points, otherwise the entry
1764 // stub will have to write it.
1765 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1766 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1767 for(j=10;j>=3;j--)
1768 {
1769 // Alloc preferred register if available
1770 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1771 for(hr=0;hr<HOST_REGS;hr++) {
1772 // Evict both parts of a 64-bit register
1773 if(cur->regmap[hr]==r) {
1774 cur->regmap[hr]=-1;
1775 cur->dirty&=~(1<<hr);
1776 cur->isconst&=~(1<<hr);
1777 }
1778 }
1779 cur->regmap[preferred_reg]=reg;
1780 return;
1781 }
1782 for(r=1;r<=MAXREG;r++)
1783 {
1784 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1785 for(hr=0;hr<HOST_REGS;hr++) {
1786 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1787 if(cur->regmap[hr]==r) {
1788 cur->regmap[hr]=reg;
1789 cur->dirty&=~(1<<hr);
1790 cur->isconst&=~(1<<hr);
1791 return;
1792 }
1793 }
1794 }
1795 }
1796 }
1797 }
1798 }
1799 for(j=10;j>=0;j--)
1800 {
1801 for(r=1;r<=MAXREG;r++)
1802 {
1803 if(hsn[r]==j) {
1804 for(hr=0;hr<HOST_REGS;hr++) {
1805 if(cur->regmap[hr]==r) {
1806 cur->regmap[hr]=reg;
1807 cur->dirty&=~(1<<hr);
1808 cur->isconst&=~(1<<hr);
1809 return;
1810 }
1811 }
1812 }
1813 }
1814 }
1815 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1816}
1817
1818// Allocate a temporary register. This is done without regard to
1819// dirty status or whether the register we request is on the unneeded list
1820// Note: This will only allocate one register, even if called multiple times
1821static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1822{
1823 int r,hr;
1824 int preferred_reg = -1;
1825
1826 // see if it's already allocated
1827 for(hr=0;hr<HOST_REGS;hr++)
1828 {
1829 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1830 }
1831
1832 // Try to allocate any available register
1833 for(hr=HOST_REGS-1;hr>=0;hr--) {
1834 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1835 cur->regmap[hr]=reg;
1836 cur->dirty&=~(1<<hr);
1837 cur->isconst&=~(1<<hr);
1838 return;
1839 }
1840 }
1841
1842 // Find an unneeded register
1843 for(hr=HOST_REGS-1;hr>=0;hr--)
1844 {
1845 r=cur->regmap[hr];
1846 if(r>=0) {
1847 assert(r < 64);
1848 if((cur->u>>r)&1) {
1849 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1850 cur->regmap[hr]=reg;
1851 cur->dirty&=~(1<<hr);
1852 cur->isconst&=~(1<<hr);
1853 return;
1854 }
1855 }
1856 }
1857 }
1858
1859 // Ok, now we have to evict someone
1860 // Pick a register we hopefully won't need soon
1861 // TODO: we might want to follow unconditional jumps here
1862 // TODO: get rid of dupe code and make this into a function
1863 u_char hsn[MAXREG+1];
1864 memset(hsn,10,sizeof(hsn));
1865 int j;
1866 lsn(hsn,i,&preferred_reg);
1867 //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]);
1868 if(i>0) {
1869 // Don't evict the cycle count at entry points, otherwise the entry
1870 // stub will have to write it.
1871 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1872 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1873 for(j=10;j>=3;j--)
1874 {
1875 for(r=1;r<=MAXREG;r++)
1876 {
1877 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1878 for(hr=0;hr<HOST_REGS;hr++) {
1879 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1880 if(cur->regmap[hr]==r) {
1881 cur->regmap[hr]=reg;
1882 cur->dirty&=~(1<<hr);
1883 cur->isconst&=~(1<<hr);
1884 return;
1885 }
1886 }
1887 }
1888 }
1889 }
1890 }
1891 }
1892 for(j=10;j>=0;j--)
1893 {
1894 for(r=1;r<=MAXREG;r++)
1895 {
1896 if(hsn[r]==j) {
1897 for(hr=0;hr<HOST_REGS;hr++) {
1898 if(cur->regmap[hr]==r) {
1899 cur->regmap[hr]=reg;
1900 cur->dirty&=~(1<<hr);
1901 cur->isconst&=~(1<<hr);
1902 return;
1903 }
1904 }
1905 }
1906 }
1907 }
1908 SysPrintf("This shouldn't happen");abort();
1909}
1910
1911static void mov_alloc(struct regstat *current,int i)
1912{
1913 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1914 alloc_cc(current,i); // for stalls
1915 dirty_reg(current,CCREG);
1916 }
1917
1918 // Note: Don't need to actually alloc the source registers
1919 //alloc_reg(current,i,dops[i].rs1);
1920 alloc_reg(current,i,dops[i].rt1);
1921
1922 clear_const(current,dops[i].rs1);
1923 clear_const(current,dops[i].rt1);
1924 dirty_reg(current,dops[i].rt1);
1925}
1926
1927static void shiftimm_alloc(struct regstat *current,int i)
1928{
1929 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1930 {
1931 if(dops[i].rt1) {
1932 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1933 else dops[i].use_lt1=!!dops[i].rs1;
1934 alloc_reg(current,i,dops[i].rt1);
1935 dirty_reg(current,dops[i].rt1);
1936 if(is_const(current,dops[i].rs1)) {
1937 int v=get_const(current,dops[i].rs1);
1938 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<cinfo[i].imm);
1939 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>cinfo[i].imm);
1940 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>cinfo[i].imm);
1941 }
1942 else clear_const(current,dops[i].rt1);
1943 }
1944 }
1945 else
1946 {
1947 clear_const(current,dops[i].rs1);
1948 clear_const(current,dops[i].rt1);
1949 }
1950
1951 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1952 {
1953 assert(0);
1954 }
1955 if(dops[i].opcode2==0x3c) // DSLL32
1956 {
1957 assert(0);
1958 }
1959 if(dops[i].opcode2==0x3e) // DSRL32
1960 {
1961 assert(0);
1962 }
1963 if(dops[i].opcode2==0x3f) // DSRA32
1964 {
1965 assert(0);
1966 }
1967}
1968
1969static void shift_alloc(struct regstat *current,int i)
1970{
1971 if(dops[i].rt1) {
1972 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1973 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1974 alloc_reg(current,i,dops[i].rt1);
1975 if(dops[i].rt1==dops[i].rs2) {
1976 alloc_reg_temp(current,i,-1);
1977 cinfo[i].min_free_regs=1;
1978 }
1979 clear_const(current,dops[i].rs1);
1980 clear_const(current,dops[i].rs2);
1981 clear_const(current,dops[i].rt1);
1982 dirty_reg(current,dops[i].rt1);
1983 }
1984}
1985
1986static void alu_alloc(struct regstat *current,int i)
1987{
1988 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1989 if(dops[i].rt1) {
1990 if(dops[i].rs1&&dops[i].rs2) {
1991 alloc_reg(current,i,dops[i].rs1);
1992 alloc_reg(current,i,dops[i].rs2);
1993 }
1994 else {
1995 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1996 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1997 }
1998 alloc_reg(current,i,dops[i].rt1);
1999 }
2000 if (dops[i].may_except) {
2001 alloc_cc(current, i); // for exceptions
2002 alloc_reg_temp(current, i, -1);
2003 cinfo[i].min_free_regs = 1;
2004 }
2005 }
2006 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2007 if(dops[i].rt1) {
2008 alloc_reg(current,i,dops[i].rs1);
2009 alloc_reg(current,i,dops[i].rs2);
2010 alloc_reg(current,i,dops[i].rt1);
2011 }
2012 }
2013 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2014 if(dops[i].rt1) {
2015 if(dops[i].rs1&&dops[i].rs2) {
2016 alloc_reg(current,i,dops[i].rs1);
2017 alloc_reg(current,i,dops[i].rs2);
2018 }
2019 else
2020 {
2021 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2022 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
2023 }
2024 alloc_reg(current,i,dops[i].rt1);
2025 }
2026 }
2027 clear_const(current,dops[i].rs1);
2028 clear_const(current,dops[i].rs2);
2029 clear_const(current,dops[i].rt1);
2030 dirty_reg(current,dops[i].rt1);
2031}
2032
2033static void imm16_alloc(struct regstat *current,int i)
2034{
2035 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2036 else dops[i].use_lt1=!!dops[i].rs1;
2037 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
2038 if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2039 clear_const(current,dops[i].rs1);
2040 clear_const(current,dops[i].rt1);
2041 }
2042 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2043 if(is_const(current,dops[i].rs1)) {
2044 int v=get_const(current,dops[i].rs1);
2045 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&cinfo[i].imm);
2046 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|cinfo[i].imm);
2047 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^cinfo[i].imm);
2048 }
2049 else clear_const(current,dops[i].rt1);
2050 }
2051 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2052 if(is_const(current,dops[i].rs1)) {
2053 int v=get_const(current,dops[i].rs1);
2054 set_const(current,dops[i].rt1,v+cinfo[i].imm);
2055 }
2056 else clear_const(current,dops[i].rt1);
2057 if (dops[i].may_except) {
2058 alloc_cc(current, i); // for exceptions
2059 alloc_reg_temp(current, i, -1);
2060 cinfo[i].min_free_regs = 1;
2061 }
2062 }
2063 else {
2064 set_const(current,dops[i].rt1,cinfo[i].imm<<16); // LUI
2065 }
2066 dirty_reg(current,dops[i].rt1);
2067}
2068
2069static void load_alloc(struct regstat *current,int i)
2070{
2071 int need_temp = 0;
2072 clear_const(current,dops[i].rt1);
2073 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2074 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2075 if (needed_again(dops[i].rs1, i))
2076 alloc_reg(current, i, dops[i].rs1);
2077 if (ram_offset)
2078 alloc_reg(current, i, ROREG);
2079 if (dops[i].may_except) {
2080 alloc_cc(current, i); // for exceptions
2081 dirty_reg(current, CCREG);
2082 need_temp = 1;
2083 }
2084 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2085 alloc_reg(current,i,dops[i].rt1);
2086 assert(get_reg_w(current->regmap, dops[i].rt1)>=0);
2087 dirty_reg(current,dops[i].rt1);
2088 // LWL/LWR need a temporary register for the old value
2089 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2090 {
2091 alloc_reg(current,i,FTEMP);
2092 need_temp = 1;
2093 }
2094 }
2095 else
2096 {
2097 // Load to r0 or unneeded register (dummy load)
2098 // but we still need a register to calculate the address
2099 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2100 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2101 need_temp = 1;
2102 }
2103 if (need_temp) {
2104 alloc_reg_temp(current, i, -1);
2105 cinfo[i].min_free_regs = 1;
2106 }
2107}
2108
2109static void store_alloc(struct regstat *current,int i)
2110{
2111 clear_const(current,dops[i].rs2);
2112 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2113 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2114 alloc_reg(current,i,dops[i].rs2);
2115 if (ram_offset)
2116 alloc_reg(current, i, ROREG);
2117 #if defined(HOST_IMM8)
2118 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2119 alloc_reg(current, i, INVCP);
2120 #endif
2121 if (dops[i].opcode == 0x2a || dops[i].opcode == 0x2e) { // SWL/SWL
2122 alloc_reg(current,i,FTEMP);
2123 }
2124 if (dops[i].may_except) {
2125 alloc_cc(current, i); // for exceptions
2126 dirty_reg(current, CCREG);
2127 }
2128 // We need a temporary register for address generation
2129 alloc_reg_temp(current,i,-1);
2130 cinfo[i].min_free_regs=1;
2131}
2132
2133static void c2ls_alloc(struct regstat *current,int i)
2134{
2135 clear_const(current,dops[i].rt1);
2136 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2137 alloc_reg(current,i,FTEMP);
2138 if (ram_offset)
2139 alloc_reg(current, i, ROREG);
2140 #if defined(HOST_IMM8)
2141 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2142 if (dops[i].opcode == 0x3a) // SWC2
2143 alloc_reg(current,i,INVCP);
2144 #endif
2145 if (dops[i].may_except) {
2146 alloc_cc(current, i); // for exceptions
2147 dirty_reg(current, CCREG);
2148 }
2149 // We need a temporary register for address generation
2150 alloc_reg_temp(current,i,-1);
2151 cinfo[i].min_free_regs=1;
2152}
2153
2154#ifndef multdiv_alloc
2155static void multdiv_alloc(struct regstat *current,int i)
2156{
2157 // case 0x18: MULT
2158 // case 0x19: MULTU
2159 // case 0x1A: DIV
2160 // case 0x1B: DIVU
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 current->u&=~(1LL<<HIREG);
2167 current->u&=~(1LL<<LOREG);
2168 alloc_reg(current,i,HIREG);
2169 alloc_reg(current,i,LOREG);
2170 alloc_reg(current,i,dops[i].rs1);
2171 alloc_reg(current,i,dops[i].rs2);
2172 dirty_reg(current,HIREG);
2173 dirty_reg(current,LOREG);
2174 }
2175 else
2176 {
2177 // Multiply by zero is zero.
2178 // MIPS does not have a divide by zero exception.
2179 alloc_reg(current,i,HIREG);
2180 alloc_reg(current,i,LOREG);
2181 dirty_reg(current,HIREG);
2182 dirty_reg(current,LOREG);
2183 if (dops[i].rs1 && ((dops[i].opcode2 & 0x3e) == 0x1a)) // div(u) 0
2184 alloc_reg(current, i, dops[i].rs1);
2185 }
2186}
2187#endif
2188
2189static void cop0_alloc(struct regstat *current,int i)
2190{
2191 if(dops[i].opcode2==0) // MFC0
2192 {
2193 if(dops[i].rt1) {
2194 clear_const(current,dops[i].rt1);
2195 alloc_reg(current,i,dops[i].rt1);
2196 dirty_reg(current,dops[i].rt1);
2197 }
2198 }
2199 else if(dops[i].opcode2==4) // MTC0
2200 {
2201 if (((source[i]>>11)&0x1e) == 12) {
2202 alloc_cc(current, i);
2203 dirty_reg(current, CCREG);
2204 }
2205 if(dops[i].rs1){
2206 clear_const(current,dops[i].rs1);
2207 alloc_reg(current,i,dops[i].rs1);
2208 alloc_all(current,i);
2209 }
2210 else {
2211 alloc_all(current,i); // FIXME: Keep r0
2212 current->u&=~1LL;
2213 alloc_reg(current,i,0);
2214 }
2215 cinfo[i].min_free_regs = HOST_REGS;
2216 }
2217}
2218
2219static void rfe_alloc(struct regstat *current, int i)
2220{
2221 alloc_all(current, i);
2222 cinfo[i].min_free_regs = HOST_REGS;
2223}
2224
2225static void cop2_alloc(struct regstat *current,int i)
2226{
2227 if (dops[i].opcode2 < 3) // MFC2/CFC2
2228 {
2229 alloc_cc(current,i); // for stalls
2230 dirty_reg(current,CCREG);
2231 if(dops[i].rt1){
2232 clear_const(current,dops[i].rt1);
2233 alloc_reg(current,i,dops[i].rt1);
2234 dirty_reg(current,dops[i].rt1);
2235 }
2236 }
2237 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2238 {
2239 if(dops[i].rs1){
2240 clear_const(current,dops[i].rs1);
2241 alloc_reg(current,i,dops[i].rs1);
2242 }
2243 else {
2244 current->u&=~1LL;
2245 alloc_reg(current,i,0);
2246 }
2247 }
2248 alloc_reg_temp(current,i,-1);
2249 cinfo[i].min_free_regs=1;
2250}
2251
2252static void c2op_alloc(struct regstat *current,int i)
2253{
2254 alloc_cc(current,i); // for stalls
2255 dirty_reg(current,CCREG);
2256 alloc_reg_temp(current,i,-1);
2257}
2258
2259static void syscall_alloc(struct regstat *current,int i)
2260{
2261 alloc_cc(current,i);
2262 dirty_reg(current,CCREG);
2263 alloc_all(current,i);
2264 cinfo[i].min_free_regs=HOST_REGS;
2265 current->isconst=0;
2266}
2267
2268static void delayslot_alloc(struct regstat *current,int i)
2269{
2270 switch(dops[i].itype) {
2271 case UJUMP:
2272 case CJUMP:
2273 case SJUMP:
2274 case RJUMP:
2275 case SYSCALL:
2276 case HLECALL:
2277 case IMM16:
2278 imm16_alloc(current,i);
2279 break;
2280 case LOAD:
2281 case LOADLR:
2282 load_alloc(current,i);
2283 break;
2284 case STORE:
2285 case STORELR:
2286 store_alloc(current,i);
2287 break;
2288 case ALU:
2289 alu_alloc(current,i);
2290 break;
2291 case SHIFT:
2292 shift_alloc(current,i);
2293 break;
2294 case MULTDIV:
2295 multdiv_alloc(current,i);
2296 break;
2297 case SHIFTIMM:
2298 shiftimm_alloc(current,i);
2299 break;
2300 case MOV:
2301 mov_alloc(current,i);
2302 break;
2303 case COP0:
2304 cop0_alloc(current,i);
2305 break;
2306 case RFE:
2307 rfe_alloc(current,i);
2308 break;
2309 case COP2:
2310 cop2_alloc(current,i);
2311 break;
2312 case C2LS:
2313 c2ls_alloc(current,i);
2314 break;
2315 case C2OP:
2316 c2op_alloc(current,i);
2317 break;
2318 }
2319}
2320
2321static void add_stub(enum stub_type type, void *addr, void *retaddr,
2322 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2323{
2324 assert(stubcount < ARRAY_SIZE(stubs));
2325 stubs[stubcount].type = type;
2326 stubs[stubcount].addr = addr;
2327 stubs[stubcount].retaddr = retaddr;
2328 stubs[stubcount].a = a;
2329 stubs[stubcount].b = b;
2330 stubs[stubcount].c = c;
2331 stubs[stubcount].d = d;
2332 stubs[stubcount].e = e;
2333 stubcount++;
2334}
2335
2336static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2337 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2338{
2339 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2340}
2341
2342// Write out a single register
2343static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2344{
2345 int hr;
2346 for(hr=0;hr<HOST_REGS;hr++) {
2347 if(hr!=EXCLUDE_REG) {
2348 if(regmap[hr]==r) {
2349 if((dirty>>hr)&1) {
2350 assert(regmap[hr]<64);
2351 emit_storereg(r,hr);
2352 }
2353 }
2354 }
2355 }
2356}
2357
2358static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2359{
2360 //if(dirty_pre==dirty) return;
2361 int hr, r;
2362 for (hr = 0; hr < HOST_REGS; hr++) {
2363 r = pre[hr];
2364 if (r < 1 || r > 33 || ((u >> r) & 1))
2365 continue;
2366 if (((dirty_pre & ~dirty) >> hr) & 1)
2367 emit_storereg(r, hr);
2368 }
2369}
2370
2371// trashes r2
2372static void pass_args(int a0, int a1)
2373{
2374 if(a0==1&&a1==0) {
2375 // must swap
2376 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2377 }
2378 else if(a0!=0&&a1==0) {
2379 emit_mov(a1,1);
2380 if (a0>=0) emit_mov(a0,0);
2381 }
2382 else {
2383 if(a0>=0&&a0!=0) emit_mov(a0,0);
2384 if(a1>=0&&a1!=1) emit_mov(a1,1);
2385 }
2386}
2387
2388static void alu_assemble(int i, const struct regstat *i_regs, int ccadj_)
2389{
2390 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2391 int do_oflow = dops[i].may_except; // ADD/SUB with exceptions enabled
2392 if (dops[i].rt1 || do_oflow) {
2393 int do_exception_check = 0;
2394 signed char s1, s2, t, tmp;
2395 t = get_reg_w(i_regs->regmap, dops[i].rt1);
2396 tmp = get_reg_temp(i_regs->regmap);
2397 if (do_oflow)
2398 assert(tmp >= 0);
2399 //if (t < 0 && do_oflow) // broken s2
2400 // t = tmp;
2401 if (t >= 0) {
2402 s1 = get_reg(i_regs->regmap, dops[i].rs1);
2403 s2 = get_reg(i_regs->regmap, dops[i].rs2);
2404 if (dops[i].rs1 && dops[i].rs2) {
2405 assert(s1>=0);
2406 assert(s2>=0);
2407 if (dops[i].opcode2 & 2) {
2408 if (do_oflow) {
2409 emit_subs(s1, s2, tmp);
2410 do_exception_check = 1;
2411 }
2412 else
2413 emit_sub(s1,s2,t);
2414 }
2415 else {
2416 if (do_oflow) {
2417 emit_adds(s1, s2, tmp);
2418 do_exception_check = 1;
2419 }
2420 else
2421 emit_add(s1,s2,t);
2422 }
2423 }
2424 else if(dops[i].rs1) {
2425 if(s1>=0) emit_mov(s1,t);
2426 else emit_loadreg(dops[i].rs1,t);
2427 }
2428 else if(dops[i].rs2) {
2429 if (s2 < 0) {
2430 emit_loadreg(dops[i].rs2, t);
2431 s2 = t;
2432 }
2433 if (dops[i].opcode2 & 2) {
2434 if (do_oflow) {
2435 emit_negs(s2, tmp);
2436 do_exception_check = 1;
2437 }
2438 else
2439 emit_neg(s2, t);
2440 }
2441 else if (s2 != t)
2442 emit_mov(s2, t);
2443 }
2444 else
2445 emit_zeroreg(t);
2446 }
2447 if (do_exception_check) {
2448 void *jaddr = out;
2449 emit_jo(0);
2450 if (t >= 0 && tmp != t)
2451 emit_mov(tmp, t);
2452 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2453 }
2454 }
2455 }
2456 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2457 if(dops[i].rt1) {
2458 signed char s1l,s2l,t;
2459 {
2460 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2461 //assert(t>=0);
2462 if(t>=0) {
2463 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2464 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2465 if(dops[i].rs2==0) // rx<r0
2466 {
2467 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2468 assert(s1l>=0);
2469 emit_shrimm(s1l,31,t);
2470 }
2471 else // SLTU (unsigned can not be less than zero, 0<0)
2472 emit_zeroreg(t);
2473 }
2474 else if(dops[i].rs1==0) // r0<rx
2475 {
2476 assert(s2l>=0);
2477 if(dops[i].opcode2==0x2a) // SLT
2478 emit_set_gz32(s2l,t);
2479 else // SLTU (set if not zero)
2480 emit_set_nz32(s2l,t);
2481 }
2482 else{
2483 assert(s1l>=0);assert(s2l>=0);
2484 if(dops[i].opcode2==0x2a) // SLT
2485 emit_set_if_less32(s1l,s2l,t);
2486 else // SLTU
2487 emit_set_if_carry32(s1l,s2l,t);
2488 }
2489 }
2490 }
2491 }
2492 }
2493 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2494 if(dops[i].rt1) {
2495 signed char s1l,s2l,tl;
2496 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2497 {
2498 if(tl>=0) {
2499 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2500 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2501 if(dops[i].rs1&&dops[i].rs2) {
2502 assert(s1l>=0);
2503 assert(s2l>=0);
2504 if(dops[i].opcode2==0x24) { // AND
2505 emit_and(s1l,s2l,tl);
2506 } else
2507 if(dops[i].opcode2==0x25) { // OR
2508 emit_or(s1l,s2l,tl);
2509 } else
2510 if(dops[i].opcode2==0x26) { // XOR
2511 emit_xor(s1l,s2l,tl);
2512 } else
2513 if(dops[i].opcode2==0x27) { // NOR
2514 emit_or(s1l,s2l,tl);
2515 emit_not(tl,tl);
2516 }
2517 }
2518 else
2519 {
2520 if(dops[i].opcode2==0x24) { // AND
2521 emit_zeroreg(tl);
2522 } else
2523 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2524 if(dops[i].rs1){
2525 if(s1l>=0) emit_mov(s1l,tl);
2526 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2527 }
2528 else
2529 if(dops[i].rs2){
2530 if(s2l>=0) emit_mov(s2l,tl);
2531 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2532 }
2533 else emit_zeroreg(tl);
2534 } else
2535 if(dops[i].opcode2==0x27) { // NOR
2536 if(dops[i].rs1){
2537 if(s1l>=0) emit_not(s1l,tl);
2538 else {
2539 emit_loadreg(dops[i].rs1,tl);
2540 emit_not(tl,tl);
2541 }
2542 }
2543 else
2544 if(dops[i].rs2){
2545 if(s2l>=0) emit_not(s2l,tl);
2546 else {
2547 emit_loadreg(dops[i].rs2,tl);
2548 emit_not(tl,tl);
2549 }
2550 }
2551 else emit_movimm(-1,tl);
2552 }
2553 }
2554 }
2555 }
2556 }
2557 }
2558}
2559
2560static void imm16_assemble(int i, const struct regstat *i_regs, int ccadj_)
2561{
2562 if (dops[i].opcode==0x0f) { // LUI
2563 if(dops[i].rt1) {
2564 signed char t;
2565 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2566 //assert(t>=0);
2567 if(t>=0) {
2568 if(!((i_regs->isconst>>t)&1))
2569 emit_movimm(cinfo[i].imm<<16,t);
2570 }
2571 }
2572 }
2573 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2574 int is_addi = dops[i].may_except;
2575 if (dops[i].rt1 || is_addi) {
2576 signed char s, t, tmp;
2577 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2578 s=get_reg(i_regs->regmap,dops[i].rs1);
2579 if(dops[i].rs1) {
2580 tmp = get_reg_temp(i_regs->regmap);
2581 if (is_addi) {
2582 assert(tmp >= 0);
2583 if (t < 0) t = tmp;
2584 }
2585 if(t>=0) {
2586 if(!((i_regs->isconst>>t)&1)) {
2587 int sum, do_exception_check = 0;
2588 if (s < 0) {
2589 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2590 if (is_addi) {
2591 emit_addimm_and_set_flags3(t, cinfo[i].imm, tmp);
2592 do_exception_check = 1;
2593 }
2594 else
2595 emit_addimm(t, cinfo[i].imm, t);
2596 } else {
2597 if (!((i_regs->wasconst >> s) & 1)) {
2598 if (is_addi) {
2599 emit_addimm_and_set_flags3(s, cinfo[i].imm, tmp);
2600 do_exception_check = 1;
2601 }
2602 else
2603 emit_addimm(s, cinfo[i].imm, t);
2604 }
2605 else {
2606 int oflow = add_overflow(constmap[i][s], cinfo[i].imm, sum);
2607 if (is_addi && oflow)
2608 do_exception_check = 2;
2609 else
2610 emit_movimm(sum, t);
2611 }
2612 }
2613 if (do_exception_check) {
2614 void *jaddr = out;
2615 if (do_exception_check == 2)
2616 emit_jmp(0);
2617 else {
2618 emit_jo(0);
2619 if (tmp != t)
2620 emit_mov(tmp, t);
2621 }
2622 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2623 }
2624 }
2625 }
2626 } else {
2627 if(t>=0) {
2628 if(!((i_regs->isconst>>t)&1))
2629 emit_movimm(cinfo[i].imm,t);
2630 }
2631 }
2632 }
2633 }
2634 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2635 if(dops[i].rt1) {
2636 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2637 signed char sl,t;
2638 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2639 sl=get_reg(i_regs->regmap,dops[i].rs1);
2640 //assert(t>=0);
2641 if(t>=0) {
2642 if(dops[i].rs1>0) {
2643 if(dops[i].opcode==0x0a) { // SLTI
2644 if(sl<0) {
2645 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2646 emit_slti32(t,cinfo[i].imm,t);
2647 }else{
2648 emit_slti32(sl,cinfo[i].imm,t);
2649 }
2650 }
2651 else { // SLTIU
2652 if(sl<0) {
2653 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2654 emit_sltiu32(t,cinfo[i].imm,t);
2655 }else{
2656 emit_sltiu32(sl,cinfo[i].imm,t);
2657 }
2658 }
2659 }else{
2660 // SLTI(U) with r0 is just stupid,
2661 // nonetheless examples can be found
2662 if(dops[i].opcode==0x0a) // SLTI
2663 if(0<cinfo[i].imm) emit_movimm(1,t);
2664 else emit_zeroreg(t);
2665 else // SLTIU
2666 {
2667 if(cinfo[i].imm) emit_movimm(1,t);
2668 else emit_zeroreg(t);
2669 }
2670 }
2671 }
2672 }
2673 }
2674 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2675 if(dops[i].rt1) {
2676 signed char sl,tl;
2677 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2678 sl=get_reg(i_regs->regmap,dops[i].rs1);
2679 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2680 if(dops[i].opcode==0x0c) //ANDI
2681 {
2682 if(dops[i].rs1) {
2683 if(sl<0) {
2684 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2685 emit_andimm(tl,cinfo[i].imm,tl);
2686 }else{
2687 if(!((i_regs->wasconst>>sl)&1))
2688 emit_andimm(sl,cinfo[i].imm,tl);
2689 else
2690 emit_movimm(constmap[i][sl]&cinfo[i].imm,tl);
2691 }
2692 }
2693 else
2694 emit_zeroreg(tl);
2695 }
2696 else
2697 {
2698 if(dops[i].rs1) {
2699 if(sl<0) {
2700 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2701 }
2702 if(dops[i].opcode==0x0d) { // ORI
2703 if(sl<0) {
2704 emit_orimm(tl,cinfo[i].imm,tl);
2705 }else{
2706 if(!((i_regs->wasconst>>sl)&1))
2707 emit_orimm(sl,cinfo[i].imm,tl);
2708 else
2709 emit_movimm(constmap[i][sl]|cinfo[i].imm,tl);
2710 }
2711 }
2712 if(dops[i].opcode==0x0e) { // XORI
2713 if(sl<0) {
2714 emit_xorimm(tl,cinfo[i].imm,tl);
2715 }else{
2716 if(!((i_regs->wasconst>>sl)&1))
2717 emit_xorimm(sl,cinfo[i].imm,tl);
2718 else
2719 emit_movimm(constmap[i][sl]^cinfo[i].imm,tl);
2720 }
2721 }
2722 }
2723 else {
2724 emit_movimm(cinfo[i].imm,tl);
2725 }
2726 }
2727 }
2728 }
2729 }
2730}
2731
2732static void shiftimm_assemble(int i, const struct regstat *i_regs)
2733{
2734 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2735 {
2736 if(dops[i].rt1) {
2737 signed char s,t;
2738 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2739 s=get_reg(i_regs->regmap,dops[i].rs1);
2740 //assert(t>=0);
2741 if(t>=0&&!((i_regs->isconst>>t)&1)){
2742 if(dops[i].rs1==0)
2743 {
2744 emit_zeroreg(t);
2745 }
2746 else
2747 {
2748 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2749 if(cinfo[i].imm) {
2750 if(dops[i].opcode2==0) // SLL
2751 {
2752 emit_shlimm(s<0?t:s,cinfo[i].imm,t);
2753 }
2754 if(dops[i].opcode2==2) // SRL
2755 {
2756 emit_shrimm(s<0?t:s,cinfo[i].imm,t);
2757 }
2758 if(dops[i].opcode2==3) // SRA
2759 {
2760 emit_sarimm(s<0?t:s,cinfo[i].imm,t);
2761 }
2762 }else{
2763 // Shift by zero
2764 if(s>=0 && s!=t) emit_mov(s,t);
2765 }
2766 }
2767 }
2768 //emit_storereg(dops[i].rt1,t); //DEBUG
2769 }
2770 }
2771 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2772 {
2773 assert(0);
2774 }
2775 if(dops[i].opcode2==0x3c) // DSLL32
2776 {
2777 assert(0);
2778 }
2779 if(dops[i].opcode2==0x3e) // DSRL32
2780 {
2781 assert(0);
2782 }
2783 if(dops[i].opcode2==0x3f) // DSRA32
2784 {
2785 assert(0);
2786 }
2787}
2788
2789#ifndef shift_assemble
2790static void shift_assemble(int i, const struct regstat *i_regs)
2791{
2792 signed char s,t,shift;
2793 if (dops[i].rt1 == 0)
2794 return;
2795 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2796 t = get_reg(i_regs->regmap, dops[i].rt1);
2797 s = get_reg(i_regs->regmap, dops[i].rs1);
2798 shift = get_reg(i_regs->regmap, dops[i].rs2);
2799 if (t < 0)
2800 return;
2801
2802 if(dops[i].rs1==0)
2803 emit_zeroreg(t);
2804 else if(dops[i].rs2==0) {
2805 assert(s>=0);
2806 if(s!=t) emit_mov(s,t);
2807 }
2808 else {
2809 host_tempreg_acquire();
2810 emit_andimm(shift,31,HOST_TEMPREG);
2811 switch(dops[i].opcode2) {
2812 case 4: // SLLV
2813 emit_shl(s,HOST_TEMPREG,t);
2814 break;
2815 case 6: // SRLV
2816 emit_shr(s,HOST_TEMPREG,t);
2817 break;
2818 case 7: // SRAV
2819 emit_sar(s,HOST_TEMPREG,t);
2820 break;
2821 default:
2822 assert(0);
2823 }
2824 host_tempreg_release();
2825 }
2826}
2827
2828#endif
2829
2830enum {
2831 MTYPE_8000 = 0,
2832 MTYPE_8020,
2833 MTYPE_0000,
2834 MTYPE_A000,
2835 MTYPE_1F80,
2836};
2837
2838static int get_ptr_mem_type(u_int a)
2839{
2840 if(a < 0x00200000) {
2841 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2842 // return wrong, must use memhandler for BIOS self-test to pass
2843 // 007 does similar stuff from a00 mirror, weird stuff
2844 return MTYPE_8000;
2845 return MTYPE_0000;
2846 }
2847 if(0x1f800000 <= a && a < 0x1f801000)
2848 return MTYPE_1F80;
2849 if(0x80200000 <= a && a < 0x80800000)
2850 return MTYPE_8020;
2851 if(0xa0000000 <= a && a < 0xa0200000)
2852 return MTYPE_A000;
2853 return MTYPE_8000;
2854}
2855
2856static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2857{
2858 int r = get_reg(i_regs->regmap, ROREG);
2859 if (r < 0 && host_tempreg_free) {
2860 host_tempreg_acquire();
2861 emit_loadreg(ROREG, r = HOST_TEMPREG);
2862 }
2863 if (r < 0)
2864 abort();
2865 return r;
2866}
2867
2868static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2869 int addr, int *offset_reg, int *addr_reg_override, int ccadj_)
2870{
2871 void *jaddr = NULL;
2872 int type = 0;
2873 int mr = dops[i].rs1;
2874 assert(addr >= 0);
2875 *offset_reg = -1;
2876 if(((smrv_strong|smrv_weak)>>mr)&1) {
2877 type=get_ptr_mem_type(smrv[mr]);
2878 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2879 }
2880 else {
2881 // use the mirror we are running on
2882 type=get_ptr_mem_type(start);
2883 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2884 }
2885
2886 if (dops[i].may_except) {
2887 // alignment check
2888 u_int op = dops[i].opcode;
2889 int mask = ((op & 0x37) == 0x21 || op == 0x25) ? 1 : 3; // LH/SH/LHU
2890 void *jaddr;
2891 emit_testimm(addr, mask);
2892 jaddr = out;
2893 emit_jne(0);
2894 add_stub_r(ALIGNMENT_STUB, jaddr, out, i, addr, i_regs, ccadj_, 0);
2895 }
2896
2897 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2898 host_tempreg_acquire();
2899 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2900 addr=*addr_reg_override=HOST_TEMPREG;
2901 type=0;
2902 }
2903 else if(type==MTYPE_0000) { // RAM 0 mirror
2904 host_tempreg_acquire();
2905 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2906 addr=*addr_reg_override=HOST_TEMPREG;
2907 type=0;
2908 }
2909 else if(type==MTYPE_A000) { // RAM A mirror
2910 host_tempreg_acquire();
2911 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2912 addr=*addr_reg_override=HOST_TEMPREG;
2913 type=0;
2914 }
2915 else if(type==MTYPE_1F80) { // scratchpad
2916 if (psxH == (void *)0x1f800000) {
2917 host_tempreg_acquire();
2918 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2919 emit_cmpimm(HOST_TEMPREG,0x1000);
2920 host_tempreg_release();
2921 jaddr=out;
2922 emit_jc(0);
2923 }
2924 else {
2925 // do the usual RAM check, jump will go to the right handler
2926 type=0;
2927 }
2928 }
2929
2930 if (type == 0) // need ram check
2931 {
2932 emit_cmpimm(addr,RAM_SIZE);
2933 jaddr = out;
2934 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2935 // Hint to branch predictor that the branch is unlikely to be taken
2936 if (dops[i].rs1 >= 28)
2937 emit_jno_unlikely(0);
2938 else
2939 #endif
2940 emit_jno(0);
2941 if (ram_offset != 0)
2942 *offset_reg = get_ro_reg(i_regs, 0);
2943 }
2944
2945 return jaddr;
2946}
2947
2948// return memhandler, or get directly accessable address and return 0
2949static void *get_direct_memhandler(void *table, u_int addr,
2950 enum stub_type type, uintptr_t *addr_host)
2951{
2952 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2953 uintptr_t l1, l2 = 0;
2954 l1 = ((uintptr_t *)table)[addr>>12];
2955 if (!(l1 & msb)) {
2956 uintptr_t v = l1 << 1;
2957 *addr_host = v + addr;
2958 return NULL;
2959 }
2960 else {
2961 l1 <<= 1;
2962 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2963 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2964 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2965 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2966 else
2967 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2968 if (!(l2 & msb)) {
2969 uintptr_t v = l2 << 1;
2970 *addr_host = v + (addr&0xfff);
2971 return NULL;
2972 }
2973 return (void *)(l2 << 1);
2974 }
2975}
2976
2977static u_int get_host_reglist(const signed char *regmap)
2978{
2979 u_int reglist = 0, hr;
2980 for (hr = 0; hr < HOST_REGS; hr++) {
2981 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2982 reglist |= 1 << hr;
2983 }
2984 return reglist;
2985}
2986
2987static u_int reglist_exclude(u_int reglist, int r1, int r2)
2988{
2989 if (r1 >= 0)
2990 reglist &= ~(1u << r1);
2991 if (r2 >= 0)
2992 reglist &= ~(1u << r2);
2993 return reglist;
2994}
2995
2996// find a temp caller-saved register not in reglist (so assumed to be free)
2997static int reglist_find_free(u_int reglist)
2998{
2999 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
3000 if (free_regs == 0)
3001 return -1;
3002 return __builtin_ctz(free_regs);
3003}
3004
3005static void do_load_word(int a, int rt, int offset_reg)
3006{
3007 if (offset_reg >= 0)
3008 emit_ldr_dualindexed(offset_reg, a, rt);
3009 else
3010 emit_readword_indexed(0, a, rt);
3011}
3012
3013static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
3014{
3015 if (offset_reg < 0) {
3016 emit_writeword_indexed(rt, ofs, a);
3017 return;
3018 }
3019 if (ofs != 0)
3020 emit_addimm(a, ofs, a);
3021 emit_str_dualindexed(offset_reg, a, rt);
3022 if (ofs != 0 && preseve_a)
3023 emit_addimm(a, -ofs, a);
3024}
3025
3026static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
3027{
3028 if (offset_reg < 0) {
3029 emit_writehword_indexed(rt, ofs, a);
3030 return;
3031 }
3032 if (ofs != 0)
3033 emit_addimm(a, ofs, a);
3034 emit_strh_dualindexed(offset_reg, a, rt);
3035 if (ofs != 0 && preseve_a)
3036 emit_addimm(a, -ofs, a);
3037}
3038
3039static void do_store_byte(int a, int rt, int offset_reg)
3040{
3041 if (offset_reg >= 0)
3042 emit_strb_dualindexed(offset_reg, a, rt);
3043 else
3044 emit_writebyte_indexed(rt, 0, a);
3045}
3046
3047static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
3048{
3049 int addr = cinfo[i].addr;
3050 int s,tl;
3051 int offset;
3052 void *jaddr=0;
3053 int memtarget=0,c=0;
3054 int offset_reg = -1;
3055 int fastio_reg_override = -1;
3056 u_int reglist=get_host_reglist(i_regs->regmap);
3057 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3058 s=get_reg(i_regs->regmap,dops[i].rs1);
3059 offset=cinfo[i].imm;
3060 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3061 if(s>=0) {
3062 c=(i_regs->wasconst>>s)&1;
3063 if (c) {
3064 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3065 }
3066 }
3067 //printf("load_assemble: c=%d\n",c);
3068 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3069 if(tl<0 && ((!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80) || dops[i].rt1==0)) {
3070 // could be FIFO, must perform the read
3071 // ||dummy read
3072 assem_debug("(forced read)\n");
3073 tl = get_reg_temp(i_regs->regmap); // may be == addr
3074 assert(tl>=0);
3075 }
3076 assert(addr >= 0);
3077 if(tl>=0) {
3078 //printf("load_assemble: c=%d\n",c);
3079 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3080 reglist&=~(1<<tl);
3081 if(!c) {
3082 #ifdef R29_HACK
3083 // Strmnnrmn's speed hack
3084 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3085 #endif
3086 {
3087 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3088 &offset_reg, &fastio_reg_override, ccadj_);
3089 }
3090 }
3091 else if (ram_offset && memtarget) {
3092 offset_reg = get_ro_reg(i_regs, 0);
3093 }
3094 int dummy=(dops[i].rt1==0)||(tl!=get_reg_w(i_regs->regmap, dops[i].rt1)); // ignore loads to r0 and unneeded reg
3095 switch (dops[i].opcode) {
3096 case 0x20: // LB
3097 if(!c||memtarget) {
3098 if(!dummy) {
3099 int a = addr;
3100 if (fastio_reg_override >= 0)
3101 a = fastio_reg_override;
3102
3103 if (offset_reg >= 0)
3104 emit_ldrsb_dualindexed(offset_reg, a, tl);
3105 else
3106 emit_movsbl_indexed(0, a, tl);
3107 }
3108 if(jaddr)
3109 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3110 }
3111 else
3112 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3113 break;
3114 case 0x21: // LH
3115 if(!c||memtarget) {
3116 if(!dummy) {
3117 int a = addr;
3118 if (fastio_reg_override >= 0)
3119 a = fastio_reg_override;
3120 if (offset_reg >= 0)
3121 emit_ldrsh_dualindexed(offset_reg, a, tl);
3122 else
3123 emit_movswl_indexed(0, a, tl);
3124 }
3125 if(jaddr)
3126 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3127 }
3128 else
3129 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3130 break;
3131 case 0x23: // LW
3132 if(!c||memtarget) {
3133 if(!dummy) {
3134 int a = addr;
3135 if (fastio_reg_override >= 0)
3136 a = fastio_reg_override;
3137 do_load_word(a, tl, offset_reg);
3138 }
3139 if(jaddr)
3140 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3141 }
3142 else
3143 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3144 break;
3145 case 0x24: // LBU
3146 if(!c||memtarget) {
3147 if(!dummy) {
3148 int a = addr;
3149 if (fastio_reg_override >= 0)
3150 a = fastio_reg_override;
3151
3152 if (offset_reg >= 0)
3153 emit_ldrb_dualindexed(offset_reg, a, tl);
3154 else
3155 emit_movzbl_indexed(0, a, tl);
3156 }
3157 if(jaddr)
3158 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3159 }
3160 else
3161 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3162 break;
3163 case 0x25: // LHU
3164 if(!c||memtarget) {
3165 if(!dummy) {
3166 int a = addr;
3167 if (fastio_reg_override >= 0)
3168 a = fastio_reg_override;
3169 if (offset_reg >= 0)
3170 emit_ldrh_dualindexed(offset_reg, a, tl);
3171 else
3172 emit_movzwl_indexed(0, a, tl);
3173 }
3174 if(jaddr)
3175 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3176 }
3177 else
3178 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3179 break;
3180 default:
3181 assert(0);
3182 }
3183 } // tl >= 0
3184 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3185 host_tempreg_release();
3186}
3187
3188#ifndef loadlr_assemble
3189static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3190{
3191 int addr = cinfo[i].addr;
3192 int s,tl,temp,temp2;
3193 int offset;
3194 void *jaddr=0;
3195 int memtarget=0,c=0;
3196 int offset_reg = -1;
3197 int fastio_reg_override = -1;
3198 u_int reglist=get_host_reglist(i_regs->regmap);
3199 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3200 s=get_reg(i_regs->regmap,dops[i].rs1);
3201 temp=get_reg_temp(i_regs->regmap);
3202 temp2=get_reg(i_regs->regmap,FTEMP);
3203 offset=cinfo[i].imm;
3204 reglist|=1<<temp;
3205 assert(addr >= 0);
3206 if(s>=0) {
3207 c=(i_regs->wasconst>>s)&1;
3208 if(c) {
3209 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3210 }
3211 }
3212 if(!c) {
3213 emit_shlimm(addr,3,temp);
3214 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3215 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3216 }else{
3217 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3218 }
3219 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3220 &offset_reg, &fastio_reg_override, ccadj_);
3221 }
3222 else {
3223 if (ram_offset && memtarget) {
3224 offset_reg = get_ro_reg(i_regs, 0);
3225 }
3226 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3227 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3228 }else{
3229 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3230 }
3231 }
3232 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3233 if(!c||memtarget) {
3234 int a = temp2;
3235 if (fastio_reg_override >= 0)
3236 a = fastio_reg_override;
3237 do_load_word(a, temp2, offset_reg);
3238 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3239 host_tempreg_release();
3240 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3241 }
3242 else
3243 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3244 if(dops[i].rt1) {
3245 assert(tl>=0);
3246 emit_andimm(temp,24,temp);
3247 if (dops[i].opcode==0x22) // LWL
3248 emit_xorimm(temp,24,temp);
3249 host_tempreg_acquire();
3250 emit_movimm(-1,HOST_TEMPREG);
3251 if (dops[i].opcode==0x26) {
3252 emit_shr(temp2,temp,temp2);
3253 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3254 }else{
3255 emit_shl(temp2,temp,temp2);
3256 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3257 }
3258 host_tempreg_release();
3259 emit_or(temp2,tl,tl);
3260 }
3261 //emit_storereg(dops[i].rt1,tl); // DEBUG
3262 }
3263 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3264 assert(0);
3265 }
3266}
3267#endif
3268
3269static void do_invstub(int n)
3270{
3271 literal_pool(20);
3272 assem_debug("do_invstub\n");
3273 u_int reglist = stubs[n].a;
3274 u_int addrr = stubs[n].b;
3275 int ofs_start = stubs[n].c;
3276 int ofs_end = stubs[n].d;
3277 int len = ofs_end - ofs_start;
3278 u_int rightr = 0;
3279
3280 set_jump_target(stubs[n].addr, out);
3281 save_regs(reglist);
3282 if (addrr != 0 || ofs_start != 0)
3283 emit_addimm(addrr, ofs_start, 0);
3284 emit_readword(&inv_code_start, 2);
3285 emit_readword(&inv_code_end, 3);
3286 if (len != 0)
3287 emit_addimm(0, len + 4, (rightr = 1));
3288 emit_cmp(0, 2);
3289 emit_cmpcs(3, rightr);
3290 void *jaddr = out;
3291 emit_jc(0);
3292 void *func = (len != 0)
3293 ? (void *)ndrc_write_invalidate_many
3294 : (void *)ndrc_write_invalidate_one;
3295 emit_far_call(func);
3296 set_jump_target(jaddr, out);
3297 restore_regs(reglist);
3298 emit_jmp(stubs[n].retaddr);
3299}
3300
3301static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3302{
3303 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3304 return;
3305 // this can't be used any more since we started to check exact
3306 // block boundaries in invalidate_range()
3307 //if (i_regs->waswritten & (1<<dops[i].rs1))
3308 // return;
3309 // (naively) assume nobody will run code from stack
3310 if (dops[i].rs1 == 29)
3311 return;
3312
3313 int j, imm_maxdiff = 32, imm_min = cinfo[i].imm, imm_max = cinfo[i].imm, count = 1;
3314 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3315 && abs(cinfo[i+1].imm - cinfo[i].imm) <= imm_maxdiff)
3316 return;
3317 for (j = i - 1; j >= 0; j--) {
3318 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3319 || abs(cinfo[j].imm - cinfo[j+1].imm) > imm_maxdiff)
3320 break;
3321 count++;
3322 if (imm_min > cinfo[j].imm)
3323 imm_min = cinfo[j].imm;
3324 if (imm_max < cinfo[j].imm)
3325 imm_max = cinfo[j].imm;
3326 }
3327#if defined(HOST_IMM8)
3328 int ir = get_reg(i_regs->regmap, INVCP);
3329 assert(ir >= 0);
3330 host_tempreg_acquire();
3331 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3332#else
3333 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3334 #error not handled
3335#endif
3336#ifdef INVALIDATE_USE_COND_CALL
3337 if (count == 1) {
3338 emit_cmpimm(HOST_TEMPREG, 1);
3339 emit_callne(invalidate_addr_reg[addr]);
3340 host_tempreg_release();
3341 return;
3342 }
3343#endif
3344 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3345 host_tempreg_release();
3346 imm_min -= cinfo[i].imm;
3347 imm_max -= cinfo[i].imm;
3348 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3349 addr, imm_min, imm_max, 0);
3350}
3351
3352static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3353{
3354 int s,tl;
3355 int addr = cinfo[i].addr;
3356 int offset;
3357 void *jaddr=0;
3358 enum stub_type type=0;
3359 int memtarget=0,c=0;
3360 int offset_reg = -1;
3361 int fastio_reg_override = -1;
3362 u_int reglist=get_host_reglist(i_regs->regmap);
3363 tl=get_reg(i_regs->regmap,dops[i].rs2);
3364 s=get_reg(i_regs->regmap,dops[i].rs1);
3365 offset=cinfo[i].imm;
3366 if(s>=0) {
3367 c=(i_regs->wasconst>>s)&1;
3368 if(c) {
3369 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3370 }
3371 }
3372 assert(tl>=0);
3373 assert(addr >= 0);
3374 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3375 if (!c) {
3376 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3377 &offset_reg, &fastio_reg_override, ccadj_);
3378 }
3379 else if (ram_offset && memtarget) {
3380 offset_reg = get_ro_reg(i_regs, 0);
3381 }
3382
3383 switch (dops[i].opcode) {
3384 case 0x28: // SB
3385 if(!c||memtarget) {
3386 int a = addr;
3387 if (fastio_reg_override >= 0)
3388 a = fastio_reg_override;
3389 do_store_byte(a, tl, offset_reg);
3390 }
3391 type = STOREB_STUB;
3392 break;
3393 case 0x29: // SH
3394 if(!c||memtarget) {
3395 int a = addr;
3396 if (fastio_reg_override >= 0)
3397 a = fastio_reg_override;
3398 do_store_hword(a, 0, tl, offset_reg, 1);
3399 }
3400 type = STOREH_STUB;
3401 break;
3402 case 0x2B: // SW
3403 if(!c||memtarget) {
3404 int a = addr;
3405 if (fastio_reg_override >= 0)
3406 a = fastio_reg_override;
3407 do_store_word(a, 0, tl, offset_reg, 1);
3408 }
3409 type = STOREW_STUB;
3410 break;
3411 default:
3412 assert(0);
3413 }
3414 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3415 host_tempreg_release();
3416 if(jaddr) {
3417 // PCSX store handlers don't check invcode again
3418 reglist|=1<<addr;
3419 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3420 jaddr=0;
3421 }
3422 {
3423 if(!c||memtarget) {
3424 do_store_smc_check(i, i_regs, reglist, addr);
3425 }
3426 }
3427 u_int addr_val=constmap[i][s]+offset;
3428 if(jaddr) {
3429 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3430 } else if(c&&!memtarget) {
3431 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3432 }
3433 // basic current block modification detection..
3434 // not looking back as that should be in mips cache already
3435 // (see Spyro2 title->attract mode)
3436 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3437 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3438 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3439 if(i_regs->regmap==regs[i].regmap) {
3440 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3441 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3442 emit_movimm(start+i*4+4,0);
3443 emit_writeword(0,&pcaddr);
3444 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3445 emit_far_call(ndrc_get_addr_ht);
3446 emit_jmpreg(0);
3447 }
3448 }
3449}
3450
3451static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3452{
3453 int addr = cinfo[i].addr;
3454 int s,tl;
3455 int offset;
3456 void *jaddr=0;
3457 void *case1, *case23, *case3;
3458 void *done0, *done1, *done2;
3459 int memtarget=0,c=0;
3460 int offset_reg = -1;
3461 u_int reglist=get_host_reglist(i_regs->regmap);
3462 tl=get_reg(i_regs->regmap,dops[i].rs2);
3463 s=get_reg(i_regs->regmap,dops[i].rs1);
3464 offset=cinfo[i].imm;
3465 if(s>=0) {
3466 c=(i_regs->isconst>>s)&1;
3467 if(c) {
3468 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3469 }
3470 }
3471 assert(tl>=0);
3472 assert(addr >= 0);
3473 if(!c) {
3474 emit_cmpimm(addr, RAM_SIZE);
3475 jaddr=out;
3476 emit_jno(0);
3477 }
3478 else
3479 {
3480 if(!memtarget||!dops[i].rs1) {
3481 jaddr=out;
3482 emit_jmp(0);
3483 }
3484 }
3485 if (ram_offset)
3486 offset_reg = get_ro_reg(i_regs, 0);
3487
3488 emit_testimm(addr,2);
3489 case23=out;
3490 emit_jne(0);
3491 emit_testimm(addr,1);
3492 case1=out;
3493 emit_jne(0);
3494 // 0
3495 if (dops[i].opcode == 0x2A) { // SWL
3496 // Write msb into least significant byte
3497 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3498 do_store_byte(addr, tl, offset_reg);
3499 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3500 }
3501 else if (dops[i].opcode == 0x2E) { // SWR
3502 // Write entire word
3503 do_store_word(addr, 0, tl, offset_reg, 1);
3504 }
3505 done0 = out;
3506 emit_jmp(0);
3507 // 1
3508 set_jump_target(case1, out);
3509 if (dops[i].opcode == 0x2A) { // SWL
3510 // Write two msb into two least significant bytes
3511 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3512 do_store_hword(addr, -1, tl, offset_reg, 0);
3513 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3514 }
3515 else if (dops[i].opcode == 0x2E) { // SWR
3516 // Write 3 lsb into three most significant bytes
3517 do_store_byte(addr, tl, offset_reg);
3518 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3519 do_store_hword(addr, 1, tl, offset_reg, 0);
3520 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3521 }
3522 done1=out;
3523 emit_jmp(0);
3524 // 2,3
3525 set_jump_target(case23, out);
3526 emit_testimm(addr,1);
3527 case3 = out;
3528 emit_jne(0);
3529 // 2
3530 if (dops[i].opcode==0x2A) { // SWL
3531 // Write 3 msb into three least significant bytes
3532 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3533 do_store_hword(addr, -2, tl, offset_reg, 1);
3534 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3535 do_store_byte(addr, tl, offset_reg);
3536 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3537 }
3538 else if (dops[i].opcode == 0x2E) { // SWR
3539 // Write two lsb into two most significant bytes
3540 do_store_hword(addr, 0, tl, offset_reg, 1);
3541 }
3542 done2 = out;
3543 emit_jmp(0);
3544 // 3
3545 set_jump_target(case3, out);
3546 if (dops[i].opcode == 0x2A) { // SWL
3547 do_store_word(addr, -3, tl, offset_reg, 0);
3548 }
3549 else if (dops[i].opcode == 0x2E) { // SWR
3550 do_store_byte(addr, tl, offset_reg);
3551 }
3552 set_jump_target(done0, out);
3553 set_jump_target(done1, out);
3554 set_jump_target(done2, out);
3555 if (offset_reg == HOST_TEMPREG)
3556 host_tempreg_release();
3557 if(!c||!memtarget)
3558 add_stub_r(STORELR_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3559 do_store_smc_check(i, i_regs, reglist, addr);
3560}
3561
3562static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3563{
3564 if(dops[i].opcode2==0) // MFC0
3565 {
3566 signed char t=get_reg_w(i_regs->regmap, dops[i].rt1);
3567 u_int copr=(source[i]>>11)&0x1f;
3568 if(t>=0&&dops[i].rt1!=0) {
3569 emit_readword(&reg_cop0[copr],t);
3570 }
3571 }
3572 else if(dops[i].opcode2==4) // MTC0
3573 {
3574 int s = get_reg(i_regs->regmap, dops[i].rs1);
3575 int cc = get_reg(i_regs->regmap, CCREG);
3576 char copr=(source[i]>>11)&0x1f;
3577 assert(s>=0);
3578 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3579 if (copr == 12 || copr == 13) {
3580 emit_readword(&last_count,HOST_TEMPREG);
3581 if (cc != HOST_CCREG)
3582 emit_loadreg(CCREG, HOST_CCREG);
3583 emit_add(HOST_CCREG, HOST_TEMPREG, HOST_CCREG);
3584 emit_addimm(HOST_CCREG, ccadj_ + 2, HOST_CCREG);
3585 emit_writeword(HOST_CCREG, &psxRegs.cycle);
3586 if (is_delayslot) {
3587 // burn cycles to cause cc_interrupt, which will
3588 // reschedule next_interupt. Relies on CCREG from above.
3589 assem_debug("MTC0 DS %d\n", copr);
3590 emit_writeword(HOST_CCREG,&last_count);
3591 emit_movimm(0,HOST_CCREG);
3592 emit_storereg(CCREG,HOST_CCREG);
3593 emit_loadreg(dops[i].rs1,1);
3594 emit_movimm(copr,0);
3595 emit_far_call(pcsx_mtc0_ds);
3596 emit_loadreg(dops[i].rs1,s);
3597 return;
3598 }
3599 emit_movimm(start+i*4+4,HOST_TEMPREG);
3600 emit_writeword(HOST_TEMPREG,&pcaddr);
3601 emit_movimm(0,HOST_TEMPREG);
3602 emit_writeword(HOST_TEMPREG,&pending_exception);
3603 }
3604 if( s != 1)
3605 emit_mov(s, 1);
3606 emit_movimm(copr, 0);
3607 emit_far_call(pcsx_mtc0);
3608 if (copr == 12 || copr == 13) {
3609 emit_readword(&psxRegs.cycle,HOST_CCREG);
3610 emit_readword(&last_count,HOST_TEMPREG);
3611 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3612 //emit_writeword(HOST_TEMPREG,&last_count);
3613 assert(!is_delayslot);
3614 emit_readword(&pending_exception,HOST_TEMPREG);
3615 emit_test(HOST_TEMPREG,HOST_TEMPREG);
3616 void *jaddr = out;
3617 emit_jeq(0);
3618 emit_readword(&pcaddr, 0);
3619 emit_far_call(ndrc_get_addr_ht);
3620 emit_jmpreg(0);
3621 set_jump_target(jaddr, out);
3622 emit_addimm(HOST_CCREG, -ccadj_ - 2, HOST_CCREG);
3623 if (cc != HOST_CCREG)
3624 emit_storereg(CCREG, HOST_CCREG);
3625 }
3626 emit_loadreg(dops[i].rs1,s);
3627 }
3628}
3629
3630static void rfe_assemble(int i, const struct regstat *i_regs)
3631{
3632 emit_readword(&psxRegs.CP0.n.SR, 0);
3633 emit_andimm(0, 0x3c, 1);
3634 emit_andimm(0, ~0xf, 0);
3635 emit_orrshr_imm(1, 2, 0);
3636 emit_writeword(0, &psxRegs.CP0.n.SR);
3637}
3638
3639static int cop2_is_stalling_op(int i, int *cycles)
3640{
3641 if (dops[i].opcode == 0x3a) { // SWC2
3642 *cycles = 0;
3643 return 1;
3644 }
3645 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3646 *cycles = 0;
3647 return 1;
3648 }
3649 if (dops[i].itype == C2OP) {
3650 *cycles = gte_cycletab[source[i] & 0x3f];
3651 return 1;
3652 }
3653 // ... what about MTC2/CTC2/LWC2?
3654 return 0;
3655}
3656
3657#if 0
3658static void log_gte_stall(int stall, u_int cycle)
3659{
3660 if ((u_int)stall <= 44)
3661 printf("x stall %2d %u\n", stall, cycle + last_count);
3662}
3663
3664static void emit_log_gte_stall(int i, int stall, u_int reglist)
3665{
3666 save_regs(reglist);
3667 if (stall > 0)
3668 emit_movimm(stall, 0);
3669 else
3670 emit_mov(HOST_TEMPREG, 0);
3671 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3672 emit_far_call(log_gte_stall);
3673 restore_regs(reglist);
3674}
3675#endif
3676
3677static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3678{
3679 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3680 int rtmp = reglist_find_free(reglist);
3681
3682 if (HACK_ENABLED(NDHACK_NO_STALLS))
3683 return;
3684 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3685 // happens occasionally... cc evicted? Don't bother then
3686 //printf("no cc %08x\n", start + i*4);
3687 return;
3688 }
3689 if (!dops[i].bt) {
3690 for (j = i - 1; j >= 0; j--) {
3691 //if (dops[j].is_ds) break;
3692 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3693 break;
3694 if (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj)
3695 break;
3696 }
3697 j = max(j, 0);
3698 }
3699 cycles_passed = cinfo[i].ccadj - cinfo[j].ccadj;
3700 if (other_gte_op_cycles >= 0)
3701 stall = other_gte_op_cycles - cycles_passed;
3702 else if (cycles_passed >= 44)
3703 stall = 0; // can't stall
3704 if (stall == -MAXBLOCK && rtmp >= 0) {
3705 // unknown stall, do the expensive runtime check
3706 assem_debug("; cop2_do_stall_check\n");
3707#if 0 // too slow
3708 save_regs(reglist);
3709 emit_movimm(gte_cycletab[op], 0);
3710 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3711 emit_far_call(call_gteStall);
3712 restore_regs(reglist);
3713#else
3714 host_tempreg_acquire();
3715 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3716 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3717 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3718 emit_cmpimm(HOST_TEMPREG, 44);
3719 emit_cmovb_reg(rtmp, HOST_CCREG);
3720 //emit_log_gte_stall(i, 0, reglist);
3721 host_tempreg_release();
3722#endif
3723 }
3724 else if (stall > 0) {
3725 //emit_log_gte_stall(i, stall, reglist);
3726 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3727 }
3728
3729 // save gteBusyCycle, if needed
3730 if (gte_cycletab[op] == 0)
3731 return;
3732 other_gte_op_cycles = -1;
3733 for (j = i + 1; j < slen; j++) {
3734 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3735 break;
3736 if (dops[j].is_jump) {
3737 // check ds
3738 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3739 j++;
3740 break;
3741 }
3742 }
3743 if (other_gte_op_cycles >= 0)
3744 // will handle stall when assembling that op
3745 return;
3746 cycles_passed = cinfo[min(j, slen -1)].ccadj - cinfo[i].ccadj;
3747 if (cycles_passed >= 44)
3748 return;
3749 assem_debug("; save gteBusyCycle\n");
3750 host_tempreg_acquire();
3751#if 0
3752 emit_readword(&last_count, HOST_TEMPREG);
3753 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3754 emit_addimm(HOST_TEMPREG, cinfo[i].ccadj, HOST_TEMPREG);
3755 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3756 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3757#else
3758 emit_addimm(HOST_CCREG, cinfo[i].ccadj + gte_cycletab[op], HOST_TEMPREG);
3759 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3760#endif
3761 host_tempreg_release();
3762}
3763
3764static int is_mflohi(int i)
3765{
3766 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3767}
3768
3769static int check_multdiv(int i, int *cycles)
3770{
3771 if (dops[i].itype != MULTDIV)
3772 return 0;
3773 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3774 *cycles = 11; // approx from 7 11 14
3775 else
3776 *cycles = 37;
3777 return 1;
3778}
3779
3780static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3781{
3782 int j, found = 0, c = 0;
3783 if (HACK_ENABLED(NDHACK_NO_STALLS))
3784 return;
3785 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3786 // happens occasionally... cc evicted? Don't bother then
3787 return;
3788 }
3789 for (j = i + 1; j < slen; j++) {
3790 if (dops[j].bt)
3791 break;
3792 if ((found = is_mflohi(j)))
3793 break;
3794 if (dops[j].is_jump) {
3795 // check ds
3796 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3797 j++;
3798 break;
3799 }
3800 }
3801 if (found)
3802 // handle all in multdiv_do_stall()
3803 return;
3804 check_multdiv(i, &c);
3805 assert(c > 0);
3806 assem_debug("; muldiv prepare stall %d\n", c);
3807 host_tempreg_acquire();
3808 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3809 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3810 host_tempreg_release();
3811}
3812
3813static void multdiv_do_stall(int i, const struct regstat *i_regs)
3814{
3815 int j, known_cycles = 0;
3816 u_int reglist = get_host_reglist(i_regs->regmap);
3817 int rtmp = get_reg_temp(i_regs->regmap);
3818 if (rtmp < 0)
3819 rtmp = reglist_find_free(reglist);
3820 if (HACK_ENABLED(NDHACK_NO_STALLS))
3821 return;
3822 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3823 // happens occasionally... cc evicted? Don't bother then
3824 //printf("no cc/rtmp %08x\n", start + i*4);
3825 return;
3826 }
3827 if (!dops[i].bt) {
3828 for (j = i - 1; j >= 0; j--) {
3829 if (dops[j].is_ds) break;
3830 if (check_multdiv(j, &known_cycles))
3831 break;
3832 if (is_mflohi(j))
3833 // already handled by this op
3834 return;
3835 if (dops[j].bt || (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj))
3836 break;
3837 }
3838 j = max(j, 0);
3839 }
3840 if (known_cycles > 0) {
3841 known_cycles -= cinfo[i].ccadj - cinfo[j].ccadj;
3842 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3843 if (known_cycles > 0)
3844 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3845 return;
3846 }
3847 assem_debug("; muldiv stall unresolved\n");
3848 host_tempreg_acquire();
3849 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3850 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3851 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3852 emit_cmpimm(HOST_TEMPREG, 37);
3853 emit_cmovb_reg(rtmp, HOST_CCREG);
3854 //emit_log_gte_stall(i, 0, reglist);
3855 host_tempreg_release();
3856}
3857
3858static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3859{
3860 switch (copr) {
3861 case 1:
3862 case 3:
3863 case 5:
3864 case 8:
3865 case 9:
3866 case 10:
3867 case 11:
3868 emit_readword(&reg_cop2d[copr],tl);
3869 emit_signextend16(tl,tl);
3870 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3871 break;
3872 case 7:
3873 case 16:
3874 case 17:
3875 case 18:
3876 case 19:
3877 emit_readword(&reg_cop2d[copr],tl);
3878 emit_andimm(tl,0xffff,tl);
3879 emit_writeword(tl,&reg_cop2d[copr]);
3880 break;
3881 case 15:
3882 emit_readword(&reg_cop2d[14],tl); // SXY2
3883 emit_writeword(tl,&reg_cop2d[copr]);
3884 break;
3885 case 28:
3886 case 29:
3887 c2op_mfc2_29_assemble(tl,temp);
3888 break;
3889 default:
3890 emit_readword(&reg_cop2d[copr],tl);
3891 break;
3892 }
3893}
3894
3895static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3896{
3897 switch (copr) {
3898 case 15:
3899 emit_readword(&reg_cop2d[13],temp); // SXY1
3900 emit_writeword(sl,&reg_cop2d[copr]);
3901 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3902 emit_readword(&reg_cop2d[14],temp); // SXY2
3903 emit_writeword(sl,&reg_cop2d[14]);
3904 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3905 break;
3906 case 28:
3907 emit_andimm(sl,0x001f,temp);
3908 emit_shlimm(temp,7,temp);
3909 emit_writeword(temp,&reg_cop2d[9]);
3910 emit_andimm(sl,0x03e0,temp);
3911 emit_shlimm(temp,2,temp);
3912 emit_writeword(temp,&reg_cop2d[10]);
3913 emit_andimm(sl,0x7c00,temp);
3914 emit_shrimm(temp,3,temp);
3915 emit_writeword(temp,&reg_cop2d[11]);
3916 emit_writeword(sl,&reg_cop2d[28]);
3917 break;
3918 case 30:
3919 emit_xorsar_imm(sl,sl,31,temp);
3920#if defined(HAVE_ARMV5) || defined(__aarch64__)
3921 emit_clz(temp,temp);
3922#else
3923 emit_movs(temp,HOST_TEMPREG);
3924 emit_movimm(0,temp);
3925 emit_jeq((int)out+4*4);
3926 emit_addpl_imm(temp,1,temp);
3927 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3928 emit_jns((int)out-2*4);
3929#endif
3930 emit_writeword(sl,&reg_cop2d[30]);
3931 emit_writeword(temp,&reg_cop2d[31]);
3932 break;
3933 case 31:
3934 break;
3935 default:
3936 emit_writeword(sl,&reg_cop2d[copr]);
3937 break;
3938 }
3939}
3940
3941static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3942{
3943 int s,tl;
3944 int ar;
3945 int offset;
3946 int memtarget=0,c=0;
3947 void *jaddr2=NULL;
3948 enum stub_type type;
3949 int offset_reg = -1;
3950 int fastio_reg_override = -1;
3951 u_int reglist=get_host_reglist(i_regs->regmap);
3952 u_int copr=(source[i]>>16)&0x1f;
3953 s=get_reg(i_regs->regmap,dops[i].rs1);
3954 tl=get_reg(i_regs->regmap,FTEMP);
3955 offset=cinfo[i].imm;
3956 assert(tl>=0);
3957
3958 if(i_regs->regmap[HOST_CCREG]==CCREG)
3959 reglist&=~(1<<HOST_CCREG);
3960
3961 // get the address
3962 ar = cinfo[i].addr;
3963 assert(ar >= 0);
3964 if (dops[i].opcode==0x3a) { // SWC2
3965 reglist |= 1<<ar;
3966 }
3967 if(s>=0) c=(i_regs->wasconst>>s)&1;
3968 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3969
3970 cop2_do_stall_check(0, i, i_regs, reglist);
3971
3972 if (dops[i].opcode==0x3a) { // SWC2
3973 cop2_get_dreg(copr,tl,-1);
3974 type=STOREW_STUB;
3975 }
3976 else
3977 type=LOADW_STUB;
3978
3979 if(c&&!memtarget) {
3980 jaddr2=out;
3981 emit_jmp(0); // inline_readstub/inline_writestub?
3982 }
3983 else {
3984 if(!c) {
3985 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3986 &offset_reg, &fastio_reg_override, ccadj_);
3987 }
3988 else if (ram_offset && memtarget) {
3989 offset_reg = get_ro_reg(i_regs, 0);
3990 }
3991 switch (dops[i].opcode) {
3992 case 0x32: { // LWC2
3993 int a = ar;
3994 if (fastio_reg_override >= 0)
3995 a = fastio_reg_override;
3996 do_load_word(a, tl, offset_reg);
3997 break;
3998 }
3999 case 0x3a: { // SWC2
4000 #ifdef DESTRUCTIVE_SHIFT
4001 if(!offset&&!c&&s>=0) emit_mov(s,ar);
4002 #endif
4003 int a = ar;
4004 if (fastio_reg_override >= 0)
4005 a = fastio_reg_override;
4006 do_store_word(a, 0, tl, offset_reg, 1);
4007 break;
4008 }
4009 default:
4010 assert(0);
4011 }
4012 }
4013 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
4014 host_tempreg_release();
4015 if(jaddr2)
4016 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
4017 if(dops[i].opcode==0x3a) // SWC2
4018 do_store_smc_check(i, i_regs, reglist, ar);
4019 if (dops[i].opcode==0x32) { // LWC2
4020 host_tempreg_acquire();
4021 cop2_put_dreg(copr,tl,HOST_TEMPREG);
4022 host_tempreg_release();
4023 }
4024}
4025
4026static void cop2_assemble(int i, const struct regstat *i_regs)
4027{
4028 u_int copr = (source[i]>>11) & 0x1f;
4029 signed char temp = get_reg_temp(i_regs->regmap);
4030
4031 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
4032 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
4033 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
4034 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
4035 reglist = reglist_exclude(reglist, tl, -1);
4036 }
4037 cop2_do_stall_check(0, i, i_regs, reglist);
4038 }
4039 if (dops[i].opcode2==0) { // MFC2
4040 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4041 if(tl>=0&&dops[i].rt1!=0)
4042 cop2_get_dreg(copr,tl,temp);
4043 }
4044 else if (dops[i].opcode2==4) { // MTC2
4045 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4046 cop2_put_dreg(copr,sl,temp);
4047 }
4048 else if (dops[i].opcode2==2) // CFC2
4049 {
4050 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4051 if(tl>=0&&dops[i].rt1!=0)
4052 emit_readword(&reg_cop2c[copr],tl);
4053 }
4054 else if (dops[i].opcode2==6) // CTC2
4055 {
4056 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4057 switch(copr) {
4058 case 4:
4059 case 12:
4060 case 20:
4061 case 26:
4062 case 27:
4063 case 29:
4064 case 30:
4065 emit_signextend16(sl,temp);
4066 break;
4067 case 31:
4068 c2op_ctc2_31_assemble(sl,temp);
4069 break;
4070 default:
4071 temp=sl;
4072 break;
4073 }
4074 emit_writeword(temp,&reg_cop2c[copr]);
4075 assert(sl>=0);
4076 }
4077}
4078
4079static void do_unalignedwritestub(int n)
4080{
4081 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4082 literal_pool(256);
4083 set_jump_target(stubs[n].addr, out);
4084
4085 int i=stubs[n].a;
4086 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4087 int addr=stubs[n].b;
4088 u_int reglist=stubs[n].e;
4089 signed char *i_regmap=i_regs->regmap;
4090 int temp2=get_reg(i_regmap,FTEMP);
4091 int rt;
4092 rt=get_reg(i_regmap,dops[i].rs2);
4093 assert(rt>=0);
4094 assert(addr>=0);
4095 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4096 reglist|=(1<<addr);
4097 reglist&=~(1<<temp2);
4098
4099 // don't bother with it and call write handler
4100 save_regs(reglist);
4101 pass_args(addr,rt);
4102 int cc=get_reg(i_regmap,CCREG);
4103 if(cc<0)
4104 emit_loadreg(CCREG,2);
4105 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4106 emit_movimm(start + i*4,3);
4107 emit_writeword(3,&psxRegs.pc);
4108 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4109 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4110 if(cc<0)
4111 emit_storereg(CCREG,2);
4112 restore_regs(reglist);
4113 emit_jmp(stubs[n].retaddr); // return address
4114}
4115
4116static void do_overflowstub(int n)
4117{
4118 assem_debug("do_overflowstub %x\n", start + (u_int)stubs[n].a * 4);
4119 literal_pool(24);
4120 int i = stubs[n].a;
4121 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4122 int ccadj = stubs[n].d;
4123 set_jump_target(stubs[n].addr, out);
4124 wb_dirtys(regs[i].regmap, regs[i].dirty);
4125 exception_assemble(i, i_regs, ccadj);
4126}
4127
4128static void do_alignmentstub(int n)
4129{
4130 assem_debug("do_alignmentstub %x\n", start + (u_int)stubs[n].a * 4);
4131 literal_pool(24);
4132 int i = stubs[n].a;
4133 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4134 int ccadj = stubs[n].d;
4135 int is_store = dops[i].itype == STORE || dops[i].opcode == 0x3A; // SWC2
4136 int cause = (dops[i].opcode & 3) << 28;
4137 cause |= is_store ? (R3000E_AdES << 2) : (R3000E_AdEL << 2);
4138 set_jump_target(stubs[n].addr, out);
4139 wb_dirtys(regs[i].regmap, regs[i].dirty);
4140 if (stubs[n].b != 1)
4141 emit_mov(stubs[n].b, 1); // faulting address
4142 emit_movimm(cause, 0);
4143 exception_assemble(i, i_regs, ccadj);
4144}
4145
4146#ifndef multdiv_assemble
4147void multdiv_assemble(int i,struct regstat *i_regs)
4148{
4149 printf("Need multdiv_assemble for this architecture.\n");
4150 abort();
4151}
4152#endif
4153
4154static void mov_assemble(int i, const struct regstat *i_regs)
4155{
4156 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
4157 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
4158 if(dops[i].rt1) {
4159 signed char sl,tl;
4160 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4161 //assert(tl>=0);
4162 if(tl>=0) {
4163 sl=get_reg(i_regs->regmap,dops[i].rs1);
4164 if(sl>=0) emit_mov(sl,tl);
4165 else emit_loadreg(dops[i].rs1,tl);
4166 }
4167 }
4168 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
4169 multdiv_do_stall(i, i_regs);
4170}
4171
4172// call interpreter, exception handler, things that change pc/regs/cycles ...
4173static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
4174{
4175 signed char ccreg=get_reg(i_regs->regmap,CCREG);
4176 assert(ccreg==HOST_CCREG);
4177 assert(!is_delayslot);
4178 (void)ccreg;
4179
4180 emit_movimm(pc,3); // Get PC
4181 emit_readword(&last_count,2);
4182 emit_writeword(3,&psxRegs.pc);
4183 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
4184 emit_add(2,HOST_CCREG,2);
4185 emit_writeword(2,&psxRegs.cycle);
4186 emit_addimm_ptr(FP,(u_char *)&psxRegs - (u_char *)&dynarec_local,0);
4187 emit_far_call(func);
4188 emit_far_jump(jump_to_new_pc);
4189}
4190
4191static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_)
4192{
4193 // 'break' tends to be littered around to catch things like
4194 // division by 0 and is almost never executed, so don't emit much code here
4195 void *func;
4196 if (dops[i].itype == ALU || dops[i].itype == IMM16)
4197 func = is_delayslot ? jump_overflow_ds : jump_overflow;
4198 else if (dops[i].itype == LOAD || dops[i].itype == STORE)
4199 func = is_delayslot ? jump_addrerror_ds : jump_addrerror;
4200 else if (dops[i].opcode2 == 0x0C)
4201 func = is_delayslot ? jump_syscall_ds : jump_syscall;
4202 else
4203 func = is_delayslot ? jump_break_ds : jump_break;
4204 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) // evicted
4205 emit_loadreg(CCREG, HOST_CCREG);
4206 emit_movimm(start + i*4, 2); // pc
4207 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4208 emit_far_jump(func);
4209}
4210
4211static void hlecall_bad()
4212{
4213 assert(0);
4214}
4215
4216static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4217{
4218 void *hlefunc = hlecall_bad;
4219 uint32_t hleCode = source[i] & 0x03ffffff;
4220 if (hleCode < ARRAY_SIZE(psxHLEt))
4221 hlefunc = psxHLEt[hleCode];
4222
4223 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4224}
4225
4226static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4227{
4228 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4229}
4230
4231static void speculate_mov(int rs,int rt)
4232{
4233 if(rt!=0) {
4234 smrv_strong_next|=1<<rt;
4235 smrv[rt]=smrv[rs];
4236 }
4237}
4238
4239static void speculate_mov_weak(int rs,int rt)
4240{
4241 if(rt!=0) {
4242 smrv_weak_next|=1<<rt;
4243 smrv[rt]=smrv[rs];
4244 }
4245}
4246
4247static void speculate_register_values(int i)
4248{
4249 if(i==0) {
4250 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4251 // gp,sp are likely to stay the same throughout the block
4252 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4253 smrv_weak_next=~smrv_strong_next;
4254 //printf(" llr %08x\n", smrv[4]);
4255 }
4256 smrv_strong=smrv_strong_next;
4257 smrv_weak=smrv_weak_next;
4258 switch(dops[i].itype) {
4259 case ALU:
4260 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4261 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4262 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4263 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4264 else {
4265 smrv_strong_next&=~(1<<dops[i].rt1);
4266 smrv_weak_next&=~(1<<dops[i].rt1);
4267 }
4268 break;
4269 case SHIFTIMM:
4270 smrv_strong_next&=~(1<<dops[i].rt1);
4271 smrv_weak_next&=~(1<<dops[i].rt1);
4272 // fallthrough
4273 case IMM16:
4274 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4275 int hr = get_reg_w(regs[i].regmap, dops[i].rt1);
4276 u_int value;
4277 if(hr>=0) {
4278 if(get_final_value(hr,i,&value))
4279 smrv[dops[i].rt1]=value;
4280 else smrv[dops[i].rt1]=constmap[i][hr];
4281 smrv_strong_next|=1<<dops[i].rt1;
4282 }
4283 }
4284 else {
4285 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4286 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4287 }
4288 break;
4289 case LOAD:
4290 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4291 // special case for BIOS
4292 smrv[dops[i].rt1]=0xa0000000;
4293 smrv_strong_next|=1<<dops[i].rt1;
4294 break;
4295 }
4296 // fallthrough
4297 case SHIFT:
4298 case LOADLR:
4299 case MOV:
4300 smrv_strong_next&=~(1<<dops[i].rt1);
4301 smrv_weak_next&=~(1<<dops[i].rt1);
4302 break;
4303 case COP0:
4304 case COP2:
4305 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4306 smrv_strong_next&=~(1<<dops[i].rt1);
4307 smrv_weak_next&=~(1<<dops[i].rt1);
4308 }
4309 break;
4310 case C2LS:
4311 if (dops[i].opcode==0x32) { // LWC2
4312 smrv_strong_next&=~(1<<dops[i].rt1);
4313 smrv_weak_next&=~(1<<dops[i].rt1);
4314 }
4315 break;
4316 }
4317#if 0
4318 int r=4;
4319 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4320 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4321#endif
4322}
4323
4324static void ujump_assemble(int i, const struct regstat *i_regs);
4325static void rjump_assemble(int i, const struct regstat *i_regs);
4326static void cjump_assemble(int i, const struct regstat *i_regs);
4327static void sjump_assemble(int i, const struct regstat *i_regs);
4328
4329static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4330{
4331 int ds = 0;
4332 switch (dops[i].itype) {
4333 case ALU:
4334 alu_assemble(i, i_regs, ccadj_);
4335 break;
4336 case IMM16:
4337 imm16_assemble(i, i_regs, ccadj_);
4338 break;
4339 case SHIFT:
4340 shift_assemble(i, i_regs);
4341 break;
4342 case SHIFTIMM:
4343 shiftimm_assemble(i, i_regs);
4344 break;
4345 case LOAD:
4346 load_assemble(i, i_regs, ccadj_);
4347 break;
4348 case LOADLR:
4349 loadlr_assemble(i, i_regs, ccadj_);
4350 break;
4351 case STORE:
4352 store_assemble(i, i_regs, ccadj_);
4353 break;
4354 case STORELR:
4355 storelr_assemble(i, i_regs, ccadj_);
4356 break;
4357 case COP0:
4358 cop0_assemble(i, i_regs, ccadj_);
4359 break;
4360 case RFE:
4361 rfe_assemble(i, i_regs);
4362 break;
4363 case COP2:
4364 cop2_assemble(i, i_regs);
4365 break;
4366 case C2LS:
4367 c2ls_assemble(i, i_regs, ccadj_);
4368 break;
4369 case C2OP:
4370 c2op_assemble(i, i_regs);
4371 break;
4372 case MULTDIV:
4373 multdiv_assemble(i, i_regs);
4374 multdiv_prepare_stall(i, i_regs, ccadj_);
4375 break;
4376 case MOV:
4377 mov_assemble(i, i_regs);
4378 break;
4379 case SYSCALL:
4380 exception_assemble(i, i_regs, ccadj_);
4381 break;
4382 case HLECALL:
4383 hlecall_assemble(i, i_regs, ccadj_);
4384 break;
4385 case INTCALL:
4386 intcall_assemble(i, i_regs, ccadj_);
4387 break;
4388 case UJUMP:
4389 ujump_assemble(i, i_regs);
4390 ds = 1;
4391 break;
4392 case RJUMP:
4393 rjump_assemble(i, i_regs);
4394 ds = 1;
4395 break;
4396 case CJUMP:
4397 cjump_assemble(i, i_regs);
4398 ds = 1;
4399 break;
4400 case SJUMP:
4401 sjump_assemble(i, i_regs);
4402 ds = 1;
4403 break;
4404 case NOP:
4405 case OTHER:
4406 // not handled, just skip
4407 break;
4408 default:
4409 assert(0);
4410 }
4411 return ds;
4412}
4413
4414static void ds_assemble(int i, const struct regstat *i_regs)
4415{
4416 speculate_register_values(i);
4417 is_delayslot = 1;
4418 switch (dops[i].itype) {
4419 case SYSCALL:
4420 case HLECALL:
4421 case INTCALL:
4422 case UJUMP:
4423 case RJUMP:
4424 case CJUMP:
4425 case SJUMP:
4426 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4427 break;
4428 default:
4429 assemble(i, i_regs, cinfo[i].ccadj);
4430 }
4431 is_delayslot = 0;
4432}
4433
4434// Is the branch target a valid internal jump?
4435static int internal_branch(int addr)
4436{
4437 if(addr&1) return 0; // Indirect (register) jump
4438 if(addr>=start && addr<start+slen*4-4)
4439 {
4440 return 1;
4441 }
4442 return 0;
4443}
4444
4445static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4446{
4447 int hr;
4448 for(hr=0;hr<HOST_REGS;hr++) {
4449 if(hr!=EXCLUDE_REG) {
4450 if(pre[hr]!=entry[hr]) {
4451 if(pre[hr]>=0) {
4452 if((dirty>>hr)&1) {
4453 if(get_reg(entry,pre[hr])<0) {
4454 assert(pre[hr]<64);
4455 if(!((u>>pre[hr])&1))
4456 emit_storereg(pre[hr],hr);
4457 }
4458 }
4459 }
4460 }
4461 }
4462 }
4463 // Move from one register to another (no writeback)
4464 for(hr=0;hr<HOST_REGS;hr++) {
4465 if(hr!=EXCLUDE_REG) {
4466 if(pre[hr]!=entry[hr]) {
4467 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4468 int nr;
4469 if((nr=get_reg(entry,pre[hr]))>=0) {
4470 emit_mov(hr,nr);
4471 }
4472 }
4473 }
4474 }
4475 }
4476}
4477
4478// Load the specified registers
4479// This only loads the registers given as arguments because
4480// we don't want to load things that will be overwritten
4481static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4482{
4483 int hr = get_reg(regmap, rs);
4484 if (hr >= 0 && entry[hr] != regmap[hr])
4485 emit_loadreg(regmap[hr], hr);
4486}
4487
4488static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4489{
4490 load_reg(entry, regmap, rs1);
4491 if (rs1 != rs2)
4492 load_reg(entry, regmap, rs2);
4493}
4494
4495// Load registers prior to the start of a loop
4496// so that they are not loaded within the loop
4497static void loop_preload(signed char pre[],signed char entry[])
4498{
4499 int hr;
4500 for (hr = 0; hr < HOST_REGS; hr++) {
4501 int r = entry[hr];
4502 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4503 assem_debug("loop preload:\n");
4504 if (r < TEMPREG)
4505 emit_loadreg(r, hr);
4506 }
4507 }
4508}
4509
4510// Generate address for load/store instruction
4511// goes to AGEN (or temp) for writes, FTEMP for LOADLR and cop1/2 loads
4512// AGEN is assigned by pass5b_preallocate2
4513static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4514{
4515 if (dops[i].is_load || dops[i].is_store) {
4516 int ra = -1;
4517 int agr = AGEN1 + (i&1);
4518 if(dops[i].itype==LOAD) {
4519 if (!dops[i].may_except)
4520 ra = get_reg_w(i_regs->regmap, dops[i].rt1); // reuse dest for agen
4521 if (ra < 0)
4522 ra = get_reg_temp(i_regs->regmap);
4523 }
4524 if(dops[i].itype==LOADLR) {
4525 ra=get_reg(i_regs->regmap,FTEMP);
4526 }
4527 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4528 ra=get_reg(i_regs->regmap,agr);
4529 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4530 }
4531 if(dops[i].itype==C2LS) {
4532 if (dops[i].opcode == 0x32) // LWC2
4533 ra=get_reg(i_regs->regmap,FTEMP);
4534 else { // SWC2
4535 ra=get_reg(i_regs->regmap,agr);
4536 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4537 }
4538 }
4539 int rs = get_reg(i_regs->regmap, dops[i].rs1);
4540 //if(ra>=0)
4541 {
4542 int offset = cinfo[i].imm;
4543 int add_offset = offset != 0;
4544 int c=(i_regs->wasconst>>rs)&1;
4545 if(dops[i].rs1==0) {
4546 // Using r0 as a base address
4547 assert(ra >= 0);
4548 if(!entry||entry[ra]!=agr) {
4549 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4550 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4551 }else{
4552 emit_movimm(offset,ra);
4553 }
4554 } // else did it in the previous cycle
4555 cinfo[i].addr = ra;
4556 add_offset = 0;
4557 }
4558 else if (rs < 0) {
4559 assert(ra >= 0);
4560 if (!entry || entry[ra] != dops[i].rs1)
4561 emit_loadreg(dops[i].rs1, ra);
4562 cinfo[i].addr = ra;
4563 //if(!entry||entry[ra]!=dops[i].rs1)
4564 // printf("poor load scheduling!\n");
4565 }
4566 else if(c) {
4567 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4568 assert(ra >= 0);
4569 if(!entry||entry[ra]!=agr) {
4570 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4571 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4572 }else{
4573 emit_movimm(constmap[i][rs]+offset,ra);
4574 regs[i].loadedconst|=1<<ra;
4575 }
4576 } // else did it in the previous cycle
4577 cinfo[i].addr = ra;
4578 }
4579 else // else load_consts already did it
4580 cinfo[i].addr = rs;
4581 add_offset = 0;
4582 }
4583 else if (dops[i].itype == STORELR) { // overwrites addr
4584 assert(ra >= 0);
4585 assert(rs != ra);
4586 emit_mov(rs, ra);
4587 cinfo[i].addr = ra;
4588 }
4589 else
4590 cinfo[i].addr = rs;
4591 if (add_offset) {
4592 assert(ra >= 0);
4593 if(rs>=0) {
4594 emit_addimm(rs,offset,ra);
4595 }else{
4596 emit_addimm(ra,offset,ra);
4597 }
4598 cinfo[i].addr = ra;
4599 }
4600 }
4601 assert(cinfo[i].addr >= 0);
4602 }
4603 // Preload constants for next instruction
4604 if (dops[i+1].is_load || dops[i+1].is_store) {
4605 int agr,ra;
4606 // Actual address
4607 agr=AGEN1+((i+1)&1);
4608 ra=get_reg(i_regs->regmap,agr);
4609 if(ra>=0) {
4610 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4611 int offset=cinfo[i+1].imm;
4612 int c=(regs[i+1].wasconst>>rs)&1;
4613 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4614 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4615 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4616 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4617 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4618 }else{
4619 emit_movimm(constmap[i+1][rs]+offset,ra);
4620 regs[i+1].loadedconst|=1<<ra;
4621 }
4622 }
4623 else if(dops[i+1].rs1==0) {
4624 // Using r0 as a base address
4625 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4626 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4627 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4628 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4629 }else{
4630 emit_movimm(offset,ra);
4631 }
4632 }
4633 }
4634 }
4635}
4636
4637static int get_final_value(int hr, int i, u_int *value)
4638{
4639 int reg=regs[i].regmap[hr];
4640 while(i<slen-1) {
4641 if(regs[i+1].regmap[hr]!=reg) break;
4642 if(!((regs[i+1].isconst>>hr)&1)) break;
4643 if(dops[i+1].bt) break;
4644 i++;
4645 }
4646 if(i<slen-1) {
4647 if (dops[i].is_jump) {
4648 *value=constmap[i][hr];
4649 return 1;
4650 }
4651 if(!dops[i+1].bt) {
4652 if (dops[i+1].is_jump) {
4653 // Load in delay slot, out-of-order execution
4654 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4655 {
4656 // Precompute load address
4657 *value=constmap[i][hr]+cinfo[i+2].imm;
4658 return 1;
4659 }
4660 }
4661 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4662 {
4663 // Precompute load address
4664 *value=constmap[i][hr]+cinfo[i+1].imm;
4665 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],cinfo[i+1].imm);
4666 return 1;
4667 }
4668 }
4669 }
4670 *value=constmap[i][hr];
4671 //printf("c=%lx\n",(long)constmap[i][hr]);
4672 if(i==slen-1) return 1;
4673 assert(reg < 64);
4674 return !((unneeded_reg[i+1]>>reg)&1);
4675}
4676
4677// Load registers with known constants
4678static void load_consts(signed char pre[],signed char regmap[],int i)
4679{
4680 int hr,hr2;
4681 // propagate loaded constant flags
4682 if(i==0||dops[i].bt)
4683 regs[i].loadedconst=0;
4684 else {
4685 for(hr=0;hr<HOST_REGS;hr++) {
4686 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4687 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4688 {
4689 regs[i].loadedconst|=1<<hr;
4690 }
4691 }
4692 }
4693 // Load 32-bit regs
4694 for(hr=0;hr<HOST_REGS;hr++) {
4695 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4696 //if(entry[hr]!=regmap[hr]) {
4697 if(!((regs[i].loadedconst>>hr)&1)) {
4698 assert(regmap[hr]<64);
4699 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4700 u_int value, similar=0;
4701 if(get_final_value(hr,i,&value)) {
4702 // see if some other register has similar value
4703 for(hr2=0;hr2<HOST_REGS;hr2++) {
4704 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4705 if(is_similar_value(value,constmap[i][hr2])) {
4706 similar=1;
4707 break;
4708 }
4709 }
4710 }
4711 if(similar) {
4712 u_int value2;
4713 if(get_final_value(hr2,i,&value2)) // is this needed?
4714 emit_movimm_from(value2,hr2,value,hr);
4715 else
4716 emit_movimm(value,hr);
4717 }
4718 else if(value==0) {
4719 emit_zeroreg(hr);
4720 }
4721 else {
4722 emit_movimm(value,hr);
4723 }
4724 }
4725 regs[i].loadedconst|=1<<hr;
4726 }
4727 }
4728 }
4729 }
4730}
4731
4732static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4733{
4734 int hr;
4735 // Load 32-bit regs
4736 for(hr=0;hr<HOST_REGS;hr++) {
4737 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4738 assert(regmap[hr] < 64);
4739 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4740 int value=constmap[i][hr];
4741 if(value==0) {
4742 emit_zeroreg(hr);
4743 }
4744 else {
4745 emit_movimm(value,hr);
4746 }
4747 }
4748 }
4749 }
4750}
4751
4752// Write out all dirty registers (except cycle count)
4753static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4754{
4755 int hr;
4756 for(hr=0;hr<HOST_REGS;hr++) {
4757 if(hr!=EXCLUDE_REG) {
4758 if(i_regmap[hr]>0) {
4759 if(i_regmap[hr]!=CCREG) {
4760 if((i_dirty>>hr)&1) {
4761 assert(i_regmap[hr]<64);
4762 emit_storereg(i_regmap[hr],hr);
4763 }
4764 }
4765 }
4766 }
4767 }
4768}
4769
4770// Write out dirty registers that we need to reload (pair with load_needed_regs)
4771// This writes the registers not written by store_regs_bt
4772static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4773{
4774 int hr;
4775 int t=(addr-start)>>2;
4776 for(hr=0;hr<HOST_REGS;hr++) {
4777 if(hr!=EXCLUDE_REG) {
4778 if(i_regmap[hr]>0) {
4779 if(i_regmap[hr]!=CCREG) {
4780 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4781 if((i_dirty>>hr)&1) {
4782 assert(i_regmap[hr]<64);
4783 emit_storereg(i_regmap[hr],hr);
4784 }
4785 }
4786 }
4787 }
4788 }
4789 }
4790}
4791
4792// Load all registers (except cycle count)
4793static void load_all_regs(const signed char i_regmap[])
4794{
4795 int hr;
4796 for(hr=0;hr<HOST_REGS;hr++) {
4797 if(hr!=EXCLUDE_REG) {
4798 if(i_regmap[hr]==0) {
4799 emit_zeroreg(hr);
4800 }
4801 else
4802 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4803 {
4804 emit_loadreg(i_regmap[hr],hr);
4805 }
4806 }
4807 }
4808}
4809
4810// Load all current registers also needed by next instruction
4811static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4812{
4813 int hr;
4814 for(hr=0;hr<HOST_REGS;hr++) {
4815 if(hr!=EXCLUDE_REG) {
4816 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4817 if(i_regmap[hr]==0) {
4818 emit_zeroreg(hr);
4819 }
4820 else
4821 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4822 {
4823 emit_loadreg(i_regmap[hr],hr);
4824 }
4825 }
4826 }
4827 }
4828}
4829
4830// Load all regs, storing cycle count if necessary
4831static void load_regs_entry(int t)
4832{
4833 int hr;
4834 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4835 else if(cinfo[t].ccadj) emit_addimm(HOST_CCREG,-cinfo[t].ccadj,HOST_CCREG);
4836 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4837 emit_storereg(CCREG,HOST_CCREG);
4838 }
4839 // Load 32-bit regs
4840 for(hr=0;hr<HOST_REGS;hr++) {
4841 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4842 if(regs[t].regmap_entry[hr]==0) {
4843 emit_zeroreg(hr);
4844 }
4845 else if(regs[t].regmap_entry[hr]!=CCREG)
4846 {
4847 emit_loadreg(regs[t].regmap_entry[hr],hr);
4848 }
4849 }
4850 }
4851}
4852
4853// Store dirty registers prior to branch
4854static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4855{
4856 if(internal_branch(addr))
4857 {
4858 int t=(addr-start)>>2;
4859 int hr;
4860 for(hr=0;hr<HOST_REGS;hr++) {
4861 if(hr!=EXCLUDE_REG) {
4862 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4863 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4864 if((i_dirty>>hr)&1) {
4865 assert(i_regmap[hr]<64);
4866 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4867 emit_storereg(i_regmap[hr],hr);
4868 }
4869 }
4870 }
4871 }
4872 }
4873 }
4874 else
4875 {
4876 // Branch out of this block, write out all dirty regs
4877 wb_dirtys(i_regmap,i_dirty);
4878 }
4879}
4880
4881// Load all needed registers for branch target
4882static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4883{
4884 //if(addr>=start && addr<(start+slen*4))
4885 if(internal_branch(addr))
4886 {
4887 int t=(addr-start)>>2;
4888 int hr;
4889 // Store the cycle count before loading something else
4890 if(i_regmap[HOST_CCREG]!=CCREG) {
4891 assert(i_regmap[HOST_CCREG]==-1);
4892 }
4893 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4894 emit_storereg(CCREG,HOST_CCREG);
4895 }
4896 // Load 32-bit regs
4897 for(hr=0;hr<HOST_REGS;hr++) {
4898 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4899 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4900 if(regs[t].regmap_entry[hr]==0) {
4901 emit_zeroreg(hr);
4902 }
4903 else if(regs[t].regmap_entry[hr]!=CCREG)
4904 {
4905 emit_loadreg(regs[t].regmap_entry[hr],hr);
4906 }
4907 }
4908 }
4909 }
4910 }
4911}
4912
4913static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4914{
4915 if(addr>=start && addr<start+slen*4-4)
4916 {
4917 int t=(addr-start)>>2;
4918 int hr;
4919 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4920 for(hr=0;hr<HOST_REGS;hr++)
4921 {
4922 if(hr!=EXCLUDE_REG)
4923 {
4924 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4925 {
4926 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4927 {
4928 return 0;
4929 }
4930 else
4931 if((i_dirty>>hr)&1)
4932 {
4933 if(i_regmap[hr]<TEMPREG)
4934 {
4935 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4936 return 0;
4937 }
4938 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4939 {
4940 assert(0);
4941 }
4942 }
4943 }
4944 else // Same register but is it 32-bit or dirty?
4945 if(i_regmap[hr]>=0)
4946 {
4947 if(!((regs[t].dirty>>hr)&1))
4948 {
4949 if((i_dirty>>hr)&1)
4950 {
4951 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4952 {
4953 //printf("%x: dirty no match\n",addr);
4954 return 0;
4955 }
4956 }
4957 }
4958 }
4959 }
4960 }
4961 // Delay slots are not valid branch targets
4962 //if(t>0&&(dops[t-1].is_jump) return 0;
4963 // Delay slots require additional processing, so do not match
4964 if(dops[t].is_ds) return 0;
4965 }
4966 else
4967 {
4968 int hr;
4969 for(hr=0;hr<HOST_REGS;hr++)
4970 {
4971 if(hr!=EXCLUDE_REG)
4972 {
4973 if(i_regmap[hr]>=0)
4974 {
4975 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4976 {
4977 if((i_dirty>>hr)&1)
4978 {
4979 return 0;
4980 }
4981 }
4982 }
4983 }
4984 }
4985 }
4986 return 1;
4987}
4988
4989#ifdef DRC_DBG
4990static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4991{
4992 extern void do_insn_cmp();
4993 //extern int cycle;
4994 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4995
4996 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4997 save_regs(reglist);
4998 // write out changed consts to match the interpreter
4999 if (i > 0 && !dops[i].bt) {
5000 for (hr = 0; hr < HOST_REGS; hr++) {
5001 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
5002 if (hr == EXCLUDE_REG || reg <= 0)
5003 continue;
5004 if (!((regs[i-1].isconst >> hr) & 1))
5005 continue;
5006 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
5007 continue;
5008 emit_movimm(constmap[i-1][hr],0);
5009 emit_storereg(reg, 0);
5010 }
5011 }
5012 emit_movimm(start+i*4,0);
5013 emit_writeword(0,&pcaddr);
5014 int cc = get_reg(regs[i].regmap_entry, CCREG);
5015 if (cc < 0)
5016 emit_loadreg(CCREG, cc = 0);
5017 emit_addimm(cc, ccadj_, 0);
5018 emit_writeword(0, &psxRegs.cycle);
5019 emit_far_call(do_insn_cmp);
5020 //emit_readword(&cycle,0);
5021 //emit_addimm(0,2,0);
5022 //emit_writeword(0,&cycle);
5023 (void)get_reg2;
5024 restore_regs(reglist);
5025 assem_debug("\\\\do_insn_cmp\n");
5026}
5027#else
5028#define drc_dbg_emit_do_cmp(x,y)
5029#endif
5030
5031// Used when a branch jumps into the delay slot of another branch
5032static void ds_assemble_entry(int i)
5033{
5034 int t = (cinfo[i].ba - start) >> 2;
5035 int ccadj_ = -CLOCK_ADJUST(1);
5036 if (!instr_addr[t])
5037 instr_addr[t] = out;
5038 assem_debug("Assemble delay slot at %x\n",cinfo[i].ba);
5039 assem_debug("<->\n");
5040 drc_dbg_emit_do_cmp(t, ccadj_);
5041 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
5042 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
5043 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
5044 address_generation(t,&regs[t],regs[t].regmap_entry);
5045 if (ram_offset && (dops[t].is_load || dops[t].is_store))
5046 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
5047 if (dops[t].is_store)
5048 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
5049 is_delayslot=0;
5050 switch (dops[t].itype) {
5051 case SYSCALL:
5052 case HLECALL:
5053 case INTCALL:
5054 case UJUMP:
5055 case RJUMP:
5056 case CJUMP:
5057 case SJUMP:
5058 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5059 break;
5060 default:
5061 assemble(t, &regs[t], ccadj_);
5062 }
5063 store_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5064 load_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5065 if(internal_branch(cinfo[i].ba+4))
5066 assem_debug("branch: internal\n");
5067 else
5068 assem_debug("branch: external\n");
5069 assert(internal_branch(cinfo[i].ba+4));
5070 add_to_linker(out,cinfo[i].ba+4,internal_branch(cinfo[i].ba+4));
5071 emit_jmp(0);
5072}
5073
5074// Load 2 immediates optimizing for small code size
5075static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
5076{
5077 emit_movimm(imm1,rt1);
5078 emit_movimm_from(imm1,rt1,imm2,rt2);
5079}
5080
5081static void do_cc(int i, const signed char i_regmap[], int *adj,
5082 int addr, int taken, int invert)
5083{
5084 int count, count_plus2;
5085 void *jaddr;
5086 void *idle=NULL;
5087 int t=0;
5088 if(dops[i].itype==RJUMP)
5089 {
5090 *adj=0;
5091 }
5092 //if(cinfo[i].ba>=start && cinfo[i].ba<(start+slen*4))
5093 if(internal_branch(cinfo[i].ba))
5094 {
5095 t=(cinfo[i].ba-start)>>2;
5096 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5097 else *adj=cinfo[t].ccadj;
5098 }
5099 else
5100 {
5101 *adj=0;
5102 }
5103 count = cinfo[i].ccadj;
5104 count_plus2 = count + CLOCK_ADJUST(2);
5105 if(taken==TAKEN && i==(cinfo[i].ba-start)>>2 && source[i+1]==0) {
5106 // Idle loop
5107 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5108 idle=out;
5109 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5110 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5111 jaddr=out;
5112 emit_jmp(0);
5113 }
5114 else if(*adj==0||invert) {
5115 int cycles = count_plus2;
5116 // faster loop HACK
5117#if 0
5118 if (t&&*adj) {
5119 int rel=t-i;
5120 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5121 cycles=*adj+count+2-*adj;
5122 }
5123#endif
5124 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5125 jaddr = out;
5126 emit_jns(0);
5127 }
5128 else
5129 {
5130 emit_cmpimm(HOST_CCREG, -count_plus2);
5131 jaddr = out;
5132 emit_jns(0);
5133 }
5134 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5135}
5136
5137static void do_ccstub(int n)
5138{
5139 literal_pool(256);
5140 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5141 set_jump_target(stubs[n].addr, out);
5142 int i=stubs[n].b;
5143 if(stubs[n].d==NULLDS) {
5144 // Delay slot instruction is nullified ("likely" branch)
5145 wb_dirtys(regs[i].regmap,regs[i].dirty);
5146 }
5147 else if(stubs[n].d!=TAKEN) {
5148 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5149 }
5150 else {
5151 if(internal_branch(cinfo[i].ba))
5152 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5153 }
5154 if(stubs[n].c!=-1)
5155 {
5156 // Save PC as return address
5157 emit_movimm(stubs[n].c,0);
5158 emit_writeword(0,&pcaddr);
5159 }
5160 else
5161 {
5162 // Return address depends on which way the branch goes
5163 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5164 {
5165 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5166 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5167 if(dops[i].rs1==0)
5168 {
5169 s1l=s2l;
5170 s2l=-1;
5171 }
5172 else if(dops[i].rs2==0)
5173 {
5174 s2l=-1;
5175 }
5176 assert(s1l>=0);
5177 #ifdef DESTRUCTIVE_WRITEBACK
5178 if(dops[i].rs1) {
5179 if((branch_regs[i].dirty>>s1l)&&1)
5180 emit_loadreg(dops[i].rs1,s1l);
5181 }
5182 else {
5183 if((branch_regs[i].dirty>>s1l)&1)
5184 emit_loadreg(dops[i].rs2,s1l);
5185 }
5186 if(s2l>=0)
5187 if((branch_regs[i].dirty>>s2l)&1)
5188 emit_loadreg(dops[i].rs2,s2l);
5189 #endif
5190 int hr=0;
5191 int addr=-1,alt=-1,ntaddr=-1;
5192 while(hr<HOST_REGS)
5193 {
5194 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5195 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5196 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5197 {
5198 addr=hr++;break;
5199 }
5200 hr++;
5201 }
5202 while(hr<HOST_REGS)
5203 {
5204 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5205 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5206 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5207 {
5208 alt=hr++;break;
5209 }
5210 hr++;
5211 }
5212 if ((dops[i].opcode & 0x3e) == 6) // BLEZ/BGTZ needs another register
5213 {
5214 while(hr<HOST_REGS)
5215 {
5216 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5217 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5218 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5219 {
5220 ntaddr=hr;break;
5221 }
5222 hr++;
5223 }
5224 assert(hr<HOST_REGS);
5225 }
5226 if (dops[i].opcode == 4) // BEQ
5227 {
5228 #ifdef HAVE_CMOV_IMM
5229 if(s2l>=0) emit_cmp(s1l,s2l);
5230 else emit_test(s1l,s1l);
5231 emit_cmov2imm_e_ne_compact(cinfo[i].ba,start+i*4+8,addr);
5232 #else
5233 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,alt);
5234 if(s2l>=0) emit_cmp(s1l,s2l);
5235 else emit_test(s1l,s1l);
5236 emit_cmovne_reg(alt,addr);
5237 #endif
5238 }
5239 else if (dops[i].opcode == 5) // BNE
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(start+i*4+8,cinfo[i].ba,addr);
5245 #else
5246 emit_mov2imm_compact(start+i*4+8,addr,cinfo[i].ba,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 == 6) // BLEZ
5253 {
5254 //emit_movimm(cinfo[i].ba,alt);
5255 //emit_movimm(start+i*4+8,addr);
5256 emit_mov2imm_compact(cinfo[i].ba,alt,start+i*4+8,addr);
5257 emit_cmpimm(s1l,1);
5258 emit_cmovl_reg(alt,addr);
5259 }
5260 else if (dops[i].opcode == 7) // BGTZ
5261 {
5262 //emit_movimm(cinfo[i].ba,addr);
5263 //emit_movimm(start+i*4+8,ntaddr);
5264 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,ntaddr);
5265 emit_cmpimm(s1l,1);
5266 emit_cmovl_reg(ntaddr,addr);
5267 }
5268 else if (dops[i].itype == SJUMP) // BLTZ/BGEZ
5269 {
5270 //emit_movimm(cinfo[i].ba,alt);
5271 //emit_movimm(start+i*4+8,addr);
5272 if (dops[i].rs1) {
5273 emit_mov2imm_compact(cinfo[i].ba,
5274 (dops[i].opcode2 & 1) ? addr : alt, start + i*4 + 8,
5275 (dops[i].opcode2 & 1) ? alt : addr);
5276 emit_test(s1l,s1l);
5277 emit_cmovs_reg(alt,addr);
5278 }
5279 else
5280 emit_movimm((dops[i].opcode2 & 1) ? cinfo[i].ba : start + i*4 + 8, addr);
5281 }
5282 emit_writeword(addr, &pcaddr);
5283 }
5284 else
5285 if(dops[i].itype==RJUMP)
5286 {
5287 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5288 if (ds_writes_rjump_rs(i)) {
5289 r=get_reg(branch_regs[i].regmap,RTEMP);
5290 }
5291 emit_writeword(r,&pcaddr);
5292 }
5293 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5294 }
5295 // Update cycle count
5296 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5297 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5298 emit_far_call(cc_interrupt);
5299 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5300 if(stubs[n].d==TAKEN) {
5301 if(internal_branch(cinfo[i].ba))
5302 load_needed_regs(branch_regs[i].regmap,regs[(cinfo[i].ba-start)>>2].regmap_entry);
5303 else if(dops[i].itype==RJUMP) {
5304 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5305 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5306 else
5307 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5308 }
5309 }else if(stubs[n].d==NOTTAKEN) {
5310 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5311 else load_all_regs(branch_regs[i].regmap);
5312 }else if(stubs[n].d==NULLDS) {
5313 // Delay slot instruction is nullified ("likely" branch)
5314 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5315 else load_all_regs(regs[i].regmap);
5316 }else{
5317 load_all_regs(branch_regs[i].regmap);
5318 }
5319 if (stubs[n].retaddr)
5320 emit_jmp(stubs[n].retaddr);
5321 else
5322 do_jump_vaddr(stubs[n].e);
5323}
5324
5325static void add_to_linker(void *addr, u_int target, int is_internal)
5326{
5327 assert(linkcount < ARRAY_SIZE(link_addr));
5328 link_addr[linkcount].addr = addr;
5329 link_addr[linkcount].target = target;
5330 link_addr[linkcount].internal = is_internal;
5331 linkcount++;
5332}
5333
5334static void ujump_assemble_write_ra(int i)
5335{
5336 int rt;
5337 unsigned int return_address;
5338 rt=get_reg(branch_regs[i].regmap,31);
5339 //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]);
5340 //assert(rt>=0);
5341 return_address=start+i*4+8;
5342 if(rt>=0) {
5343 #ifdef USE_MINI_HT
5344 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5345 int temp=-1; // note: must be ds-safe
5346 #ifdef HOST_TEMPREG
5347 temp=HOST_TEMPREG;
5348 #endif
5349 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5350 else emit_movimm(return_address,rt);
5351 }
5352 else
5353 #endif
5354 {
5355 #ifdef REG_PREFETCH
5356 if(temp>=0)
5357 {
5358 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5359 }
5360 #endif
5361 if (!((regs[i].loadedconst >> rt) & 1))
5362 emit_movimm(return_address, rt); // PC into link register
5363 #ifdef IMM_PREFETCH
5364 emit_prefetch(hash_table_get(return_address));
5365 #endif
5366 }
5367 }
5368}
5369
5370static void ujump_assemble(int i, const struct regstat *i_regs)
5371{
5372 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5373 address_generation(i+1,i_regs,regs[i].regmap_entry);
5374 #ifdef REG_PREFETCH
5375 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5376 if(dops[i].rt1==31&&temp>=0)
5377 {
5378 signed char *i_regmap=i_regs->regmap;
5379 int return_address=start+i*4+8;
5380 if(get_reg(branch_regs[i].regmap,31)>0)
5381 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5382 }
5383 #endif
5384 if (dops[i].rt1 == 31)
5385 ujump_assemble_write_ra(i); // writeback ra for DS
5386 ds_assemble(i+1,i_regs);
5387 uint64_t bc_unneeded=branch_regs[i].u;
5388 bc_unneeded|=1|(1LL<<dops[i].rt1);
5389 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5390 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5391 int cc,adj;
5392 cc=get_reg(branch_regs[i].regmap,CCREG);
5393 assert(cc==HOST_CCREG);
5394 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5395 #ifdef REG_PREFETCH
5396 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5397 #endif
5398 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5399 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5400 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5401 if(internal_branch(cinfo[i].ba))
5402 assem_debug("branch: internal\n");
5403 else
5404 assem_debug("branch: external\n");
5405 if (internal_branch(cinfo[i].ba) && dops[(cinfo[i].ba-start)>>2].is_ds) {
5406 ds_assemble_entry(i);
5407 }
5408 else {
5409 add_to_linker(out,cinfo[i].ba,internal_branch(cinfo[i].ba));
5410 emit_jmp(0);
5411 }
5412}
5413
5414static void rjump_assemble_write_ra(int i)
5415{
5416 int rt,return_address;
5417 rt=get_reg_w(branch_regs[i].regmap, dops[i].rt1);
5418 //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]);
5419 assert(rt>=0);
5420 return_address=start+i*4+8;
5421 #ifdef REG_PREFETCH
5422 if(temp>=0)
5423 {
5424 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5425 }
5426 #endif
5427 if (!((regs[i].loadedconst >> rt) & 1))
5428 emit_movimm(return_address, rt); // PC into link register
5429 #ifdef IMM_PREFETCH
5430 emit_prefetch(hash_table_get(return_address));
5431 #endif
5432}
5433
5434static void rjump_assemble(int i, const struct regstat *i_regs)
5435{
5436 int temp;
5437 int rs,cc;
5438 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5439 assert(rs>=0);
5440 if (ds_writes_rjump_rs(i)) {
5441 // Delay slot abuse, make a copy of the branch address register
5442 temp=get_reg(branch_regs[i].regmap,RTEMP);
5443 assert(temp>=0);
5444 assert(regs[i].regmap[temp]==RTEMP);
5445 emit_mov(rs,temp);
5446 rs=temp;
5447 }
5448 address_generation(i+1,i_regs,regs[i].regmap_entry);
5449 #ifdef REG_PREFETCH
5450 if(dops[i].rt1==31)
5451 {
5452 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5453 signed char *i_regmap=i_regs->regmap;
5454 int return_address=start+i*4+8;
5455 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5456 }
5457 }
5458 #endif
5459 #ifdef USE_MINI_HT
5460 if(dops[i].rs1==31) {
5461 int rh=get_reg(regs[i].regmap,RHASH);
5462 if(rh>=0) do_preload_rhash(rh);
5463 }
5464 #endif
5465 if (dops[i].rt1 != 0)
5466 rjump_assemble_write_ra(i);
5467 ds_assemble(i+1,i_regs);
5468 uint64_t bc_unneeded=branch_regs[i].u;
5469 bc_unneeded|=1|(1LL<<dops[i].rt1);
5470 bc_unneeded&=~(1LL<<dops[i].rs1);
5471 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5472 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5473 cc=get_reg(branch_regs[i].regmap,CCREG);
5474 assert(cc==HOST_CCREG);
5475 (void)cc;
5476 #ifdef USE_MINI_HT
5477 int rh=get_reg(branch_regs[i].regmap,RHASH);
5478 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5479 if(dops[i].rs1==31) {
5480 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5481 do_preload_rhtbl(ht);
5482 do_rhash(rs,rh);
5483 }
5484 #endif
5485 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5486 #ifdef DESTRUCTIVE_WRITEBACK
5487 if((branch_regs[i].dirty>>rs)&1) {
5488 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5489 emit_loadreg(dops[i].rs1,rs);
5490 }
5491 }
5492 #endif
5493 #ifdef REG_PREFETCH
5494 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5495 #endif
5496 #ifdef USE_MINI_HT
5497 if(dops[i].rs1==31) {
5498 do_miniht_load(ht,rh);
5499 }
5500 #endif
5501 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5502 //if(adj) emit_addimm(cc,2*(cinfo[i].ccadj+2-adj),cc); // ??? - Shouldn't happen
5503 //assert(adj==0);
5504 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5505 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5506 if (dops[i+1].itype == RFE)
5507 // special case for RFE
5508 emit_jmp(0);
5509 else
5510 emit_jns(0);
5511 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5512 #ifdef USE_MINI_HT
5513 if(dops[i].rs1==31) {
5514 do_miniht_jump(rs,rh,ht);
5515 }
5516 else
5517 #endif
5518 {
5519 do_jump_vaddr(rs);
5520 }
5521 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5522 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5523 #endif
5524}
5525
5526static void cjump_assemble(int i, const struct regstat *i_regs)
5527{
5528 const signed char *i_regmap = i_regs->regmap;
5529 int cc;
5530 int match;
5531 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5532 assem_debug("match=%d\n",match);
5533 int s1l,s2l;
5534 int unconditional=0,nop=0;
5535 int invert=0;
5536 int internal=internal_branch(cinfo[i].ba);
5537 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5538 if(!match) invert=1;
5539 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5540 if(i>(cinfo[i].ba-start)>>2) invert=1;
5541 #endif
5542 #ifdef __aarch64__
5543 invert=1; // because of near cond. branches
5544 #endif
5545
5546 if(dops[i].ooo) {
5547 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5548 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5549 }
5550 else {
5551 s1l=get_reg(i_regmap,dops[i].rs1);
5552 s2l=get_reg(i_regmap,dops[i].rs2);
5553 }
5554 if(dops[i].rs1==0&&dops[i].rs2==0)
5555 {
5556 if(dops[i].opcode&1) nop=1;
5557 else unconditional=1;
5558 //assert(dops[i].opcode!=5);
5559 //assert(dops[i].opcode!=7);
5560 //assert(dops[i].opcode!=0x15);
5561 //assert(dops[i].opcode!=0x17);
5562 }
5563 else if(dops[i].rs1==0)
5564 {
5565 s1l=s2l;
5566 s2l=-1;
5567 }
5568 else if(dops[i].rs2==0)
5569 {
5570 s2l=-1;
5571 }
5572
5573 if(dops[i].ooo) {
5574 // Out of order execution (delay slot first)
5575 //printf("OOOE\n");
5576 address_generation(i+1,i_regs,regs[i].regmap_entry);
5577 ds_assemble(i+1,i_regs);
5578 int adj;
5579 uint64_t bc_unneeded=branch_regs[i].u;
5580 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5581 bc_unneeded|=1;
5582 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5583 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5584 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5585 cc=get_reg(branch_regs[i].regmap,CCREG);
5586 assert(cc==HOST_CCREG);
5587 if(unconditional)
5588 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5589 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5590 //assem_debug("cycle count (adj)\n");
5591 if(unconditional) {
5592 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5593 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5594 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5595 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5596 if(internal)
5597 assem_debug("branch: internal\n");
5598 else
5599 assem_debug("branch: external\n");
5600 if (internal && dops[(cinfo[i].ba-start)>>2].is_ds) {
5601 ds_assemble_entry(i);
5602 }
5603 else {
5604 add_to_linker(out,cinfo[i].ba,internal);
5605 emit_jmp(0);
5606 }
5607 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5608 if(((u_int)out)&7) emit_addnop(0);
5609 #endif
5610 }
5611 }
5612 else if(nop) {
5613 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5614 void *jaddr=out;
5615 emit_jns(0);
5616 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5617 }
5618 else {
5619 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5620 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5621 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5622
5623 //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]);
5624 assert(s1l>=0);
5625 if(dops[i].opcode==4) // BEQ
5626 {
5627 if(s2l>=0) emit_cmp(s1l,s2l);
5628 else emit_test(s1l,s1l);
5629 if(invert){
5630 nottaken=out;
5631 emit_jne(DJT_1);
5632 }else{
5633 add_to_linker(out,cinfo[i].ba,internal);
5634 emit_jeq(0);
5635 }
5636 }
5637 if(dops[i].opcode==5) // BNE
5638 {
5639 if(s2l>=0) emit_cmp(s1l,s2l);
5640 else emit_test(s1l,s1l);
5641 if(invert){
5642 nottaken=out;
5643 emit_jeq(DJT_1);
5644 }else{
5645 add_to_linker(out,cinfo[i].ba,internal);
5646 emit_jne(0);
5647 }
5648 }
5649 if(dops[i].opcode==6) // BLEZ
5650 {
5651 emit_cmpimm(s1l,1);
5652 if(invert){
5653 nottaken=out;
5654 emit_jge(DJT_1);
5655 }else{
5656 add_to_linker(out,cinfo[i].ba,internal);
5657 emit_jl(0);
5658 }
5659 }
5660 if(dops[i].opcode==7) // BGTZ
5661 {
5662 emit_cmpimm(s1l,1);
5663 if(invert){
5664 nottaken=out;
5665 emit_jl(DJT_1);
5666 }else{
5667 add_to_linker(out,cinfo[i].ba,internal);
5668 emit_jge(0);
5669 }
5670 }
5671 if(invert) {
5672 if(taken) set_jump_target(taken, out);
5673 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5674 if (match && (!internal || !dops[(cinfo[i].ba-start)>>2].is_ds)) {
5675 if(adj) {
5676 emit_addimm(cc,-adj,cc);
5677 add_to_linker(out,cinfo[i].ba,internal);
5678 }else{
5679 emit_addnop(13);
5680 add_to_linker(out,cinfo[i].ba,internal*2);
5681 }
5682 emit_jmp(0);
5683 }else
5684 #endif
5685 {
5686 if(adj) emit_addimm(cc,-adj,cc);
5687 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5688 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5689 if(internal)
5690 assem_debug("branch: internal\n");
5691 else
5692 assem_debug("branch: external\n");
5693 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5694 ds_assemble_entry(i);
5695 }
5696 else {
5697 add_to_linker(out,cinfo[i].ba,internal);
5698 emit_jmp(0);
5699 }
5700 }
5701 set_jump_target(nottaken, out);
5702 }
5703
5704 if(nottaken1) set_jump_target(nottaken1, out);
5705 if(adj) {
5706 if(!invert) emit_addimm(cc,adj,cc);
5707 }
5708 } // (!unconditional)
5709 } // if(ooo)
5710 else
5711 {
5712 // In-order execution (branch first)
5713 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5714 if(!unconditional&&!nop) {
5715 //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]);
5716 assert(s1l>=0);
5717 if((dops[i].opcode&0x2f)==4) // BEQ
5718 {
5719 if(s2l>=0) emit_cmp(s1l,s2l);
5720 else emit_test(s1l,s1l);
5721 nottaken=out;
5722 emit_jne(DJT_2);
5723 }
5724 if((dops[i].opcode&0x2f)==5) // BNE
5725 {
5726 if(s2l>=0) emit_cmp(s1l,s2l);
5727 else emit_test(s1l,s1l);
5728 nottaken=out;
5729 emit_jeq(DJT_2);
5730 }
5731 if((dops[i].opcode&0x2f)==6) // BLEZ
5732 {
5733 emit_cmpimm(s1l,1);
5734 nottaken=out;
5735 emit_jge(DJT_2);
5736 }
5737 if((dops[i].opcode&0x2f)==7) // BGTZ
5738 {
5739 emit_cmpimm(s1l,1);
5740 nottaken=out;
5741 emit_jl(DJT_2);
5742 }
5743 } // if(!unconditional)
5744 int adj;
5745 uint64_t ds_unneeded=branch_regs[i].u;
5746 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5747 ds_unneeded|=1;
5748 // branch taken
5749 if(!nop) {
5750 if(taken) set_jump_target(taken, out);
5751 assem_debug("1:\n");
5752 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5753 // load regs
5754 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5755 address_generation(i+1,&branch_regs[i],0);
5756 if (ram_offset)
5757 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5758 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5759 ds_assemble(i+1,&branch_regs[i]);
5760 cc=get_reg(branch_regs[i].regmap,CCREG);
5761 if(cc==-1) {
5762 emit_loadreg(CCREG,cc=HOST_CCREG);
5763 // CHECK: Is the following instruction (fall thru) allocated ok?
5764 }
5765 assert(cc==HOST_CCREG);
5766 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5767 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5768 assem_debug("cycle count (adj)\n");
5769 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5770 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5771 if(internal)
5772 assem_debug("branch: internal\n");
5773 else
5774 assem_debug("branch: external\n");
5775 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5776 ds_assemble_entry(i);
5777 }
5778 else {
5779 add_to_linker(out,cinfo[i].ba,internal);
5780 emit_jmp(0);
5781 }
5782 }
5783 // branch not taken
5784 if(!unconditional) {
5785 if(nottaken1) set_jump_target(nottaken1, out);
5786 set_jump_target(nottaken, out);
5787 assem_debug("2:\n");
5788 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5789 // load regs
5790 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5791 address_generation(i+1,&branch_regs[i],0);
5792 if (ram_offset)
5793 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5794 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5795 ds_assemble(i+1,&branch_regs[i]);
5796 cc=get_reg(branch_regs[i].regmap,CCREG);
5797 if (cc == -1) {
5798 // Cycle count isn't in a register, temporarily load it then write it out
5799 emit_loadreg(CCREG,HOST_CCREG);
5800 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5801 void *jaddr=out;
5802 emit_jns(0);
5803 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5804 emit_storereg(CCREG,HOST_CCREG);
5805 }
5806 else{
5807 cc=get_reg(i_regmap,CCREG);
5808 assert(cc==HOST_CCREG);
5809 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5810 void *jaddr=out;
5811 emit_jns(0);
5812 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5813 }
5814 }
5815 }
5816}
5817
5818static void sjump_assemble(int i, const struct regstat *i_regs)
5819{
5820 const signed char *i_regmap = i_regs->regmap;
5821 int cc;
5822 int match;
5823 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5824 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5825 int s1l;
5826 int unconditional=0,nevertaken=0;
5827 int invert=0;
5828 int internal=internal_branch(cinfo[i].ba);
5829 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5830 if(!match) invert=1;
5831 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5832 if(i>(cinfo[i].ba-start)>>2) invert=1;
5833 #endif
5834 #ifdef __aarch64__
5835 invert=1; // because of near cond. branches
5836 #endif
5837
5838 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5839 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5840
5841 if(dops[i].ooo) {
5842 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5843 }
5844 else {
5845 s1l=get_reg(i_regmap,dops[i].rs1);
5846 }
5847 if(dops[i].rs1==0)
5848 {
5849 if(dops[i].opcode2&1) unconditional=1;
5850 else nevertaken=1;
5851 // These are never taken (r0 is never less than zero)
5852 //assert(dops[i].opcode2!=0);
5853 //assert(dops[i].opcode2!=2);
5854 //assert(dops[i].opcode2!=0x10);
5855 //assert(dops[i].opcode2!=0x12);
5856 }
5857
5858 if(dops[i].ooo) {
5859 // Out of order execution (delay slot first)
5860 //printf("OOOE\n");
5861 address_generation(i+1,i_regs,regs[i].regmap_entry);
5862 ds_assemble(i+1,i_regs);
5863 int adj;
5864 uint64_t bc_unneeded=branch_regs[i].u;
5865 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5866 bc_unneeded|=1;
5867 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5868 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5869 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5870 if(dops[i].rt1==31) {
5871 int rt,return_address;
5872 rt=get_reg(branch_regs[i].regmap,31);
5873 //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]);
5874 if(rt>=0) {
5875 // Save the PC even if the branch is not taken
5876 return_address=start+i*4+8;
5877 emit_movimm(return_address,rt); // PC into link register
5878 #ifdef IMM_PREFETCH
5879 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5880 #endif
5881 }
5882 }
5883 cc=get_reg(branch_regs[i].regmap,CCREG);
5884 assert(cc==HOST_CCREG);
5885 if(unconditional)
5886 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5887 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5888 assem_debug("cycle count (adj)\n");
5889 if(unconditional) {
5890 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5891 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5892 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5893 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5894 if(internal)
5895 assem_debug("branch: internal\n");
5896 else
5897 assem_debug("branch: external\n");
5898 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5899 ds_assemble_entry(i);
5900 }
5901 else {
5902 add_to_linker(out,cinfo[i].ba,internal);
5903 emit_jmp(0);
5904 }
5905 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5906 if(((u_int)out)&7) emit_addnop(0);
5907 #endif
5908 }
5909 }
5910 else if(nevertaken) {
5911 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5912 void *jaddr=out;
5913 emit_jns(0);
5914 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5915 }
5916 else {
5917 void *nottaken = NULL;
5918 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5919 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5920 {
5921 assert(s1l>=0);
5922 if ((dops[i].opcode2 & 1) == 0) // BLTZ/BLTZAL
5923 {
5924 emit_test(s1l,s1l);
5925 if(invert){
5926 nottaken=out;
5927 emit_jns(DJT_1);
5928 }else{
5929 add_to_linker(out,cinfo[i].ba,internal);
5930 emit_js(0);
5931 }
5932 }
5933 else // BGEZ/BGEZAL
5934 {
5935 emit_test(s1l,s1l);
5936 if(invert){
5937 nottaken=out;
5938 emit_js(DJT_1);
5939 }else{
5940 add_to_linker(out,cinfo[i].ba,internal);
5941 emit_jns(0);
5942 }
5943 }
5944 }
5945
5946 if(invert) {
5947 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5948 if (match && (!internal || !dops[(cinfo[i].ba - start) >> 2].is_ds)) {
5949 if(adj) {
5950 emit_addimm(cc,-adj,cc);
5951 add_to_linker(out,cinfo[i].ba,internal);
5952 }else{
5953 emit_addnop(13);
5954 add_to_linker(out,cinfo[i].ba,internal*2);
5955 }
5956 emit_jmp(0);
5957 }else
5958 #endif
5959 {
5960 if(adj) emit_addimm(cc,-adj,cc);
5961 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5962 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5963 if(internal)
5964 assem_debug("branch: internal\n");
5965 else
5966 assem_debug("branch: external\n");
5967 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5968 ds_assemble_entry(i);
5969 }
5970 else {
5971 add_to_linker(out,cinfo[i].ba,internal);
5972 emit_jmp(0);
5973 }
5974 }
5975 set_jump_target(nottaken, out);
5976 }
5977
5978 if(adj) {
5979 if(!invert) emit_addimm(cc,adj,cc);
5980 }
5981 } // (!unconditional)
5982 } // if(ooo)
5983 else
5984 {
5985 // In-order execution (branch first)
5986 //printf("IOE\n");
5987 void *nottaken = NULL;
5988 if (!unconditional && !nevertaken) {
5989 assert(s1l >= 0);
5990 emit_test(s1l, s1l);
5991 }
5992 if (dops[i].rt1 == 31) {
5993 int rt, return_address;
5994 rt = get_reg(branch_regs[i].regmap,31);
5995 if(rt >= 0) {
5996 // Save the PC even if the branch is not taken
5997 return_address = start + i*4+8;
5998 emit_movimm(return_address, rt); // PC into link register
5999 #ifdef IMM_PREFETCH
6000 emit_prefetch(hash_table_get(return_address));
6001 #endif
6002 }
6003 }
6004 if (!unconditional && !nevertaken) {
6005 nottaken = out;
6006 if (!(dops[i].opcode2 & 1)) // BLTZ/BLTZAL
6007 emit_jns(DJT_1);
6008 else // BGEZ/BGEZAL
6009 emit_js(DJT_1);
6010 }
6011 int adj;
6012 uint64_t ds_unneeded=branch_regs[i].u;
6013 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6014 ds_unneeded|=1;
6015 // branch taken
6016 if(!nevertaken) {
6017 //assem_debug("1:\n");
6018 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6019 // load regs
6020 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6021 address_generation(i+1,&branch_regs[i],0);
6022 if (ram_offset)
6023 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6024 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6025 ds_assemble(i+1,&branch_regs[i]);
6026 cc=get_reg(branch_regs[i].regmap,CCREG);
6027 if(cc==-1) {
6028 emit_loadreg(CCREG,cc=HOST_CCREG);
6029 // CHECK: Is the following instruction (fall thru) allocated ok?
6030 }
6031 assert(cc==HOST_CCREG);
6032 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6033 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
6034 assem_debug("cycle count (adj)\n");
6035 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
6036 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
6037 if(internal)
6038 assem_debug("branch: internal\n");
6039 else
6040 assem_debug("branch: external\n");
6041 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
6042 ds_assemble_entry(i);
6043 }
6044 else {
6045 add_to_linker(out,cinfo[i].ba,internal);
6046 emit_jmp(0);
6047 }
6048 }
6049 // branch not taken
6050 if(!unconditional) {
6051 if (!nevertaken) {
6052 assert(nottaken);
6053 set_jump_target(nottaken, out);
6054 }
6055 assem_debug("1:\n");
6056 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
6057 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
6058 address_generation(i+1,&branch_regs[i],0);
6059 if (ram_offset)
6060 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
6061 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
6062 ds_assemble(i+1,&branch_regs[i]);
6063 cc=get_reg(branch_regs[i].regmap,CCREG);
6064 if (cc == -1) {
6065 // Cycle count isn't in a register, temporarily load it then write it out
6066 emit_loadreg(CCREG,HOST_CCREG);
6067 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
6068 void *jaddr=out;
6069 emit_jns(0);
6070 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6071 emit_storereg(CCREG,HOST_CCREG);
6072 }
6073 else{
6074 cc=get_reg(i_regmap,CCREG);
6075 assert(cc==HOST_CCREG);
6076 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
6077 void *jaddr=out;
6078 emit_jns(0);
6079 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6080 }
6081 }
6082 }
6083}
6084
6085static void check_regmap(signed char *regmap)
6086{
6087#ifndef NDEBUG
6088 int i,j;
6089 for (i = 0; i < HOST_REGS; i++) {
6090 if (regmap[i] < 0)
6091 continue;
6092 for (j = i + 1; j < HOST_REGS; j++)
6093 assert(regmap[i] != regmap[j]);
6094 }
6095#endif
6096}
6097
6098#ifdef DISASM
6099#include <inttypes.h>
6100static char insn[MAXBLOCK][10];
6101
6102#define set_mnemonic(i_, n_) \
6103 strcpy(insn[i_], n_)
6104
6105void print_regmap(const char *name, const signed char *regmap)
6106{
6107 char buf[5];
6108 int i, l;
6109 fputs(name, stdout);
6110 for (i = 0; i < HOST_REGS; i++) {
6111 l = 0;
6112 if (regmap[i] >= 0)
6113 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6114 for (; l < 3; l++)
6115 buf[l] = ' ';
6116 buf[l] = 0;
6117 printf(" r%d=%s", i, buf);
6118 }
6119 fputs("\n", stdout);
6120}
6121
6122 /* disassembly */
6123void disassemble_inst(int i)
6124{
6125 if (dops[i].bt) printf("*"); else printf(" ");
6126 switch(dops[i].itype) {
6127 case UJUMP:
6128 printf (" %x: %s %8x\n",start+i*4,insn[i],cinfo[i].ba);break;
6129 case CJUMP:
6130 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;
6131 case SJUMP:
6132 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;
6133 case RJUMP:
6134 if (dops[i].opcode2 == 9 && dops[i].rt1 != 31)
6135 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6136 else
6137 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6138 break;
6139 case IMM16:
6140 if(dops[i].opcode==0xf) //LUI
6141 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,cinfo[i].imm&0xffff);
6142 else
6143 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6144 break;
6145 case LOAD:
6146 case LOADLR:
6147 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6148 break;
6149 case STORE:
6150 case STORELR:
6151 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,cinfo[i].imm);
6152 break;
6153 case ALU:
6154 case SHIFT:
6155 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6156 break;
6157 case MULTDIV:
6158 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6159 break;
6160 case SHIFTIMM:
6161 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6162 break;
6163 case MOV:
6164 if((dops[i].opcode2&0x1d)==0x10)
6165 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6166 else if((dops[i].opcode2&0x1d)==0x11)
6167 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6168 else
6169 printf (" %x: %s\n",start+i*4,insn[i]);
6170 break;
6171 case COP0:
6172 if(dops[i].opcode2==0)
6173 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6174 else if(dops[i].opcode2==4)
6175 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6176 else printf (" %x: %s\n",start+i*4,insn[i]);
6177 break;
6178 case COP2:
6179 if(dops[i].opcode2<3)
6180 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6181 else if(dops[i].opcode2>3)
6182 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6183 else printf (" %x: %s\n",start+i*4,insn[i]);
6184 break;
6185 case C2LS:
6186 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,cinfo[i].imm);
6187 break;
6188 case INTCALL:
6189 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6190 break;
6191 default:
6192 //printf (" %s %8x\n",insn[i],source[i]);
6193 printf (" %x: %s\n",start+i*4,insn[i]);
6194 }
6195 #ifndef REGMAP_PRINT
6196 return;
6197 #endif
6198 printf("D: %"PRIx64" WD: %"PRIx64" U: %"PRIx64" hC: %x hWC: %x hLC: %x\n",
6199 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i],
6200 regs[i].isconst, regs[i].wasconst, regs[i].loadedconst);
6201 print_regmap("pre: ", regmap_pre[i]);
6202 print_regmap("entry: ", regs[i].regmap_entry);
6203 print_regmap("map: ", regs[i].regmap);
6204 if (dops[i].is_jump) {
6205 print_regmap("bentry:", branch_regs[i].regmap_entry);
6206 print_regmap("bmap: ", branch_regs[i].regmap);
6207 }
6208}
6209#else
6210#define set_mnemonic(i_, n_)
6211static void disassemble_inst(int i) {}
6212#endif // DISASM
6213
6214#define DRC_TEST_VAL 0x74657374
6215
6216static noinline void new_dynarec_test(void)
6217{
6218 int (*testfunc)(void);
6219 void *beginning;
6220 int ret[2];
6221 size_t i;
6222
6223 // check structure linkage
6224 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6225 {
6226 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6227 }
6228
6229 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6230 new_dynarec_test, out);
6231 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6232
6233 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6234 out = ndrc->translation_cache;
6235 beginning = start_block();
6236 emit_movimm(DRC_TEST_VAL + i, 0); // test
6237 emit_ret();
6238 literal_pool(0);
6239 end_block(beginning);
6240 testfunc = beginning;
6241 ret[i] = testfunc();
6242 }
6243
6244 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6245 SysPrintf("test passed.\n");
6246 else
6247 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6248 out = ndrc->translation_cache;
6249}
6250
6251// clear the state completely, instead of just marking
6252// things invalid like invalidate_all_pages() does
6253void new_dynarec_clear_full(void)
6254{
6255 int n;
6256 out = ndrc->translation_cache;
6257 memset(invalid_code,1,sizeof(invalid_code));
6258 memset(hash_table,0xff,sizeof(hash_table));
6259 memset(mini_ht,-1,sizeof(mini_ht));
6260 memset(shadow,0,sizeof(shadow));
6261 copy=shadow;
6262 expirep = EXPIRITY_OFFSET;
6263 pending_exception=0;
6264 literalcount=0;
6265 stop_after_jal=0;
6266 inv_code_start=inv_code_end=~0;
6267 hack_addr=0;
6268 f1_hack=0;
6269 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6270 blocks_clear(&blocks[n]);
6271 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6272 free(jumps[n]);
6273 jumps[n] = NULL;
6274 }
6275 stat_clear(stat_blocks);
6276 stat_clear(stat_links);
6277
6278 cycle_multiplier_old = Config.cycle_multiplier;
6279 new_dynarec_hacks_old = new_dynarec_hacks;
6280}
6281
6282void new_dynarec_init(void)
6283{
6284 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6285
6286#ifdef _3DS
6287 check_rosalina();
6288#endif
6289#ifdef BASE_ADDR_DYNAMIC
6290 #ifdef VITA
6291 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6292 if (sceBlock <= 0)
6293 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6294 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6295 if (ret < 0)
6296 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6297 sceKernelOpenVMDomain();
6298 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6299 #elif defined(_MSC_VER)
6300 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6301 PAGE_EXECUTE_READWRITE);
6302 #elif defined(HAVE_LIBNX)
6303 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6304 if (R_FAILED(rc))
6305 SysPrintf("jitCreate failed: %08x\n", rc);
6306 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6307 jitTransitionToWritable(&g_jit);
6308 ndrc = g_jit.rx_addr;
6309 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6310 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6311 #else
6312 uintptr_t desired_addr = 0;
6313 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6314 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6315 int fd = -1;
6316 #ifdef __ELF__
6317 extern char _end;
6318 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6319 #endif
6320 #ifdef TC_WRITE_OFFSET
6321 // mostly for testing
6322 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6323 ftruncate(fd, sizeof(*ndrc));
6324 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6325 (flags = MAP_SHARED), fd, 0);
6326 assert(mw != MAP_FAILED);
6327 prot = PROT_READ | PROT_EXEC;
6328 #endif
6329 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6330 if (ndrc == MAP_FAILED) {
6331 SysPrintf("mmap() failed: %s\n", strerror(errno));
6332 abort();
6333 }
6334 #ifdef TC_WRITE_OFFSET
6335 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6336 #endif
6337 #endif
6338#else
6339 #ifndef NO_WRITE_EXEC
6340 // not all systems allow execute in data segment by default
6341 // size must be 4K aligned for 3DS?
6342 if (mprotect(ndrc, sizeof(*ndrc),
6343 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6344 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6345 #endif
6346#endif
6347 out = ndrc->translation_cache;
6348 new_dynarec_clear_full();
6349#ifdef HOST_IMM8
6350 // Copy this into local area so we don't have to put it in every literal pool
6351 invc_ptr=invalid_code;
6352#endif
6353 arch_init();
6354 new_dynarec_test();
6355 ram_offset=(uintptr_t)rdram-0x80000000;
6356 if (ram_offset!=0)
6357 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6358 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6359 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6360}
6361
6362void new_dynarec_cleanup(void)
6363{
6364 int n;
6365#ifdef BASE_ADDR_DYNAMIC
6366 #ifdef VITA
6367 // sceBlock is managed by retroarch's bootstrap code
6368 //sceKernelFreeMemBlock(sceBlock);
6369 //sceBlock = -1;
6370 #elif defined(HAVE_LIBNX)
6371 jitClose(&g_jit);
6372 ndrc = NULL;
6373 #else
6374 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6375 SysPrintf("munmap() failed\n");
6376 ndrc = NULL;
6377 #endif
6378#endif
6379 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6380 blocks_clear(&blocks[n]);
6381 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6382 free(jumps[n]);
6383 jumps[n] = NULL;
6384 }
6385 stat_clear(stat_blocks);
6386 stat_clear(stat_links);
6387 new_dynarec_print_stats();
6388}
6389
6390static u_int *get_source_start(u_int addr, u_int *limit)
6391{
6392 if (addr < 0x00200000 ||
6393 (0xa0000000 <= addr && addr < 0xa0200000))
6394 {
6395 // used for BIOS calls mostly?
6396 *limit = (addr&0xa0000000)|0x00200000;
6397 return (u_int *)(rdram + (addr&0x1fffff));
6398 }
6399 else if (!Config.HLE && (
6400 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6401 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6402 {
6403 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6404 // but timings in PCSX are too tied to the interpreter's 2-per-insn assumption
6405 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6406 cycle_multiplier_active = 200;
6407
6408 *limit = (addr & 0xfff00000) | 0x80000;
6409 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6410 }
6411 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6412 *limit = (addr & 0x80600000) + 0x00200000;
6413 return (u_int *)(rdram + (addr&0x1fffff));
6414 }
6415 return NULL;
6416}
6417
6418static u_int scan_for_ret(u_int addr)
6419{
6420 u_int limit = 0;
6421 u_int *mem;
6422
6423 mem = get_source_start(addr, &limit);
6424 if (mem == NULL)
6425 return addr;
6426
6427 if (limit > addr + 0x1000)
6428 limit = addr + 0x1000;
6429 for (; addr < limit; addr += 4, mem++) {
6430 if (*mem == 0x03e00008) // jr $ra
6431 return addr + 8;
6432 }
6433 return addr;
6434}
6435
6436struct savestate_block {
6437 uint32_t addr;
6438 uint32_t regflags;
6439};
6440
6441static int addr_cmp(const void *p1_, const void *p2_)
6442{
6443 const struct savestate_block *p1 = p1_, *p2 = p2_;
6444 return p1->addr - p2->addr;
6445}
6446
6447int new_dynarec_save_blocks(void *save, int size)
6448{
6449 struct savestate_block *sblocks = save;
6450 int maxcount = size / sizeof(sblocks[0]);
6451 struct savestate_block tmp_blocks[1024];
6452 struct block_info *block;
6453 int p, s, d, o, bcnt;
6454 u_int addr;
6455
6456 o = 0;
6457 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6458 bcnt = 0;
6459 for (block = blocks[p]; block != NULL; block = block->next) {
6460 if (block->is_dirty)
6461 continue;
6462 tmp_blocks[bcnt].addr = block->start;
6463 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6464 bcnt++;
6465 }
6466 if (bcnt < 1)
6467 continue;
6468 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6469
6470 addr = tmp_blocks[0].addr;
6471 for (s = d = 0; s < bcnt; s++) {
6472 if (tmp_blocks[s].addr < addr)
6473 continue;
6474 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6475 tmp_blocks[d++] = tmp_blocks[s];
6476 addr = scan_for_ret(tmp_blocks[s].addr);
6477 }
6478
6479 if (o + d > maxcount)
6480 d = maxcount - o;
6481 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6482 o += d;
6483 }
6484
6485 return o * sizeof(sblocks[0]);
6486}
6487
6488void new_dynarec_load_blocks(const void *save, int size)
6489{
6490 const struct savestate_block *sblocks = save;
6491 int count = size / sizeof(sblocks[0]);
6492 struct block_info *block;
6493 u_int regs_save[32];
6494 u_int page;
6495 uint32_t f;
6496 int i, b;
6497
6498 // restore clean blocks, if any
6499 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6500 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6501 if (!block->is_dirty)
6502 continue;
6503 assert(block->source && block->copy);
6504 if (memcmp(block->source, block->copy, block->len))
6505 continue;
6506
6507 // see try_restore_block
6508 block->is_dirty = 0;
6509 mark_invalid_code(block->start, block->len, 0);
6510 i++;
6511 }
6512 }
6513 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6514
6515 // change GPRs for speculation to at least partially work..
6516 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6517 for (i = 1; i < 32; i++)
6518 psxRegs.GPR.r[i] = 0x80000000;
6519
6520 for (b = 0; b < count; b++) {
6521 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6522 if (f & 1)
6523 psxRegs.GPR.r[i] = 0x1f800000;
6524 }
6525
6526 ndrc_get_addr_ht(sblocks[b].addr);
6527
6528 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6529 if (f & 1)
6530 psxRegs.GPR.r[i] = 0x80000000;
6531 }
6532 }
6533
6534 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6535}
6536
6537void new_dynarec_print_stats(void)
6538{
6539#ifdef STAT_PRINT
6540 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6541 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6542 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6543 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6544 out - ndrc->translation_cache, stat_blocks, stat_links);
6545 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6546 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6547 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6548#endif
6549}
6550
6551static int apply_hacks(void)
6552{
6553 int i;
6554 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6555 return 0;
6556 /* special hack(s) */
6557 for (i = 0; i < slen - 4; i++)
6558 {
6559 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6560 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6561 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6562 && cinfo[i+3].imm == 0x6e40 && dops[i+3].rs1 == 2)
6563 {
6564 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6565 dops[i + 3].itype = NOP;
6566 }
6567 }
6568 i = slen;
6569 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6570 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6571 && dops[i-7].itype == STORE)
6572 {
6573 i = i-8;
6574 if (dops[i].itype == IMM16)
6575 i--;
6576 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6577 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6578 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6579 {
6580 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6581 f1_hack = 1;
6582 return 1;
6583 }
6584 }
6585 return 0;
6586}
6587
6588static int is_ld_use_hazard(int ld_rt, const struct decoded_insn *op)
6589{
6590 return ld_rt != 0 && (ld_rt == op->rs1 || ld_rt == op->rs2)
6591 && op->itype != LOADLR && op->itype != CJUMP && op->itype != SJUMP;
6592}
6593
6594static void force_intcall(int i)
6595{
6596 memset(&dops[i], 0, sizeof(dops[i]));
6597 dops[i].itype = INTCALL;
6598 dops[i].rs1 = CCREG;
6599 dops[i].is_exception = 1;
6600 cinfo[i].ba = -1;
6601}
6602
6603static void disassemble_one(int i, u_int src)
6604{
6605 unsigned int type, op, op2, op3;
6606 memset(&dops[i], 0, sizeof(dops[i]));
6607 memset(&cinfo[i], 0, sizeof(cinfo[i]));
6608 cinfo[i].ba = -1;
6609 cinfo[i].addr = -1;
6610 dops[i].opcode = op = src >> 26;
6611 op2 = 0;
6612 type = INTCALL;
6613 set_mnemonic(i, "???");
6614 switch(op)
6615 {
6616 case 0x00: set_mnemonic(i, "special");
6617 op2 = src & 0x3f;
6618 switch(op2)
6619 {
6620 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6621 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6622 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6623 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6624 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6625 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6626 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6627 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6628 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6629 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6630 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6631 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6632 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6633 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6634 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6635 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6636 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6637 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6638 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6639 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6640 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6641 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6642 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6643 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6644 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6645 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6646 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6647 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6648 }
6649 break;
6650 case 0x01: set_mnemonic(i, "regimm");
6651 type = SJUMP;
6652 op2 = (src >> 16) & 0x1f;
6653 switch(op2)
6654 {
6655 case 0x10: set_mnemonic(i, "BLTZAL"); break;
6656 case 0x11: set_mnemonic(i, "BGEZAL"); break;
6657 default:
6658 if (op2 & 1)
6659 set_mnemonic(i, "BGEZ");
6660 else
6661 set_mnemonic(i, "BLTZ");
6662 }
6663 break;
6664 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6665 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6666 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6667 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6668 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6669 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6670 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6671 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6672 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6673 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6674 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6675 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6676 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6677 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6678 case 0x10: set_mnemonic(i, "COP0");
6679 op2 = (src >> 21) & 0x1f;
6680 if (op2 & 0x10) {
6681 op3 = src & 0x1f;
6682 switch (op3)
6683 {
6684 case 0x01: case 0x02: case 0x06: case 0x08: type = INTCALL; break;
6685 case 0x10: set_mnemonic(i, "RFE"); type=RFE; break;
6686 default: type = OTHER; break;
6687 }
6688 break;
6689 }
6690 switch(op2)
6691 {
6692 u32 rd;
6693 case 0x00:
6694 set_mnemonic(i, "MFC0");
6695 rd = (src >> 11) & 0x1F;
6696 if (!(0x00000417u & (1u << rd)))
6697 type = COP0;
6698 break;
6699 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6700 case 0x02:
6701 case 0x06: type = INTCALL; break;
6702 default: type = OTHER; break;
6703 }
6704 break;
6705 case 0x11: set_mnemonic(i, "COP1");
6706 op2 = (src >> 21) & 0x1f;
6707 break;
6708 case 0x12: set_mnemonic(i, "COP2");
6709 op2 = (src >> 21) & 0x1f;
6710 if (op2 & 0x10) {
6711 type = OTHER;
6712 if (gte_handlers[src & 0x3f] != NULL) {
6713#ifdef DISASM
6714 if (gte_regnames[src & 0x3f] != NULL)
6715 strcpy(insn[i], gte_regnames[src & 0x3f]);
6716 else
6717 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", src & 0x3f);
6718#endif
6719 type = C2OP;
6720 }
6721 }
6722 else switch(op2)
6723 {
6724 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6725 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6726 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6727 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6728 }
6729 break;
6730 case 0x13: set_mnemonic(i, "COP3");
6731 op2 = (src >> 21) & 0x1f;
6732 break;
6733 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6734 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6735 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6736 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6737 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6738 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6739 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6740 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6741 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6742 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6743 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6744 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6745 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6746 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6747 case 0x3B:
6748 if (Config.HLE && (src & 0x03ffffff) < ARRAY_SIZE(psxHLEt)) {
6749 set_mnemonic(i, "HLECALL");
6750 type = HLECALL;
6751 }
6752 break;
6753 default:
6754 break;
6755 }
6756 if (type == INTCALL)
6757 SysPrintf("NI %08x @%08x (%08x)\n", src, start + i*4, start);
6758 dops[i].itype=type;
6759 dops[i].opcode2=op2;
6760 /* Get registers/immediates */
6761 dops[i].use_lt1=0;
6762 gte_rs[i]=gte_rt[i]=0;
6763 dops[i].rs1 = 0;
6764 dops[i].rs2 = 0;
6765 dops[i].rt1 = 0;
6766 dops[i].rt2 = 0;
6767 switch(type) {
6768 case LOAD:
6769 dops[i].rs1 = (src >> 21) & 0x1f;
6770 dops[i].rt1 = (src >> 16) & 0x1f;
6771 cinfo[i].imm = (short)src;
6772 break;
6773 case STORE:
6774 case STORELR:
6775 dops[i].rs1 = (src >> 21) & 0x1f;
6776 dops[i].rs2 = (src >> 16) & 0x1f;
6777 cinfo[i].imm = (short)src;
6778 break;
6779 case LOADLR:
6780 // LWL/LWR only load part of the register,
6781 // therefore the target register must be treated as a source too
6782 dops[i].rs1 = (src >> 21) & 0x1f;
6783 dops[i].rs2 = (src >> 16) & 0x1f;
6784 dops[i].rt1 = (src >> 16) & 0x1f;
6785 cinfo[i].imm = (short)src;
6786 break;
6787 case IMM16:
6788 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6789 else dops[i].rs1 = (src >> 21) & 0x1f;
6790 dops[i].rs2 = 0;
6791 dops[i].rt1 = (src >> 16) & 0x1f;
6792 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6793 cinfo[i].imm = (unsigned short)src;
6794 }else{
6795 cinfo[i].imm = (short)src;
6796 }
6797 break;
6798 case UJUMP:
6799 // The JAL instruction writes to r31.
6800 if (op&1) {
6801 dops[i].rt1=31;
6802 }
6803 dops[i].rs2=CCREG;
6804 break;
6805 case RJUMP:
6806 dops[i].rs1 = (src >> 21) & 0x1f;
6807 // The JALR instruction writes to rd.
6808 if (op2&1) {
6809 dops[i].rt1 = (src >> 11) & 0x1f;
6810 }
6811 dops[i].rs2=CCREG;
6812 break;
6813 case CJUMP:
6814 dops[i].rs1 = (src >> 21) & 0x1f;
6815 dops[i].rs2 = (src >> 16) & 0x1f;
6816 if(op&2) { // BGTZ/BLEZ
6817 dops[i].rs2=0;
6818 }
6819 break;
6820 case SJUMP:
6821 dops[i].rs1 = (src >> 21) & 0x1f;
6822 dops[i].rs2 = CCREG;
6823 if (op2 == 0x10 || op2 == 0x11) { // BxxAL
6824 dops[i].rt1 = 31;
6825 // NOTE: If the branch is not taken, r31 is still overwritten
6826 }
6827 break;
6828 case ALU:
6829 dops[i].rs1=(src>>21)&0x1f; // source
6830 dops[i].rs2=(src>>16)&0x1f; // subtract amount
6831 dops[i].rt1=(src>>11)&0x1f; // destination
6832 break;
6833 case MULTDIV:
6834 dops[i].rs1=(src>>21)&0x1f; // source
6835 dops[i].rs2=(src>>16)&0x1f; // divisor
6836 dops[i].rt1=HIREG;
6837 dops[i].rt2=LOREG;
6838 break;
6839 case MOV:
6840 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6841 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6842 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6843 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6844 if((op2&0x1d)==0x10) dops[i].rt1=(src>>11)&0x1f; // MFxx
6845 if((op2&0x1d)==0x11) dops[i].rs1=(src>>21)&0x1f; // MTxx
6846 break;
6847 case SHIFT:
6848 dops[i].rs1=(src>>16)&0x1f; // target of shift
6849 dops[i].rs2=(src>>21)&0x1f; // shift amount
6850 dops[i].rt1=(src>>11)&0x1f; // destination
6851 break;
6852 case SHIFTIMM:
6853 dops[i].rs1=(src>>16)&0x1f;
6854 dops[i].rs2=0;
6855 dops[i].rt1=(src>>11)&0x1f;
6856 cinfo[i].imm=(src>>6)&0x1f;
6857 break;
6858 case COP0:
6859 if(op2==0) dops[i].rt1=(src>>16)&0x1F; // MFC0
6860 if(op2==4) dops[i].rs1=(src>>16)&0x1F; // MTC0
6861 if(op2==4&&((src>>11)&0x1e)==12) dops[i].rs2=CCREG;
6862 break;
6863 case COP2:
6864 if(op2<3) dops[i].rt1=(src>>16)&0x1F; // MFC2/CFC2
6865 if(op2>3) dops[i].rs1=(src>>16)&0x1F; // MTC2/CTC2
6866 int gr=(src>>11)&0x1F;
6867 switch(op2)
6868 {
6869 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6870 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6871 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6872 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6873 }
6874 break;
6875 case C2LS:
6876 dops[i].rs1=(src>>21)&0x1F;
6877 cinfo[i].imm=(short)src;
6878 if(op==0x32) gte_rt[i]=1ll<<((src>>16)&0x1F); // LWC2
6879 else gte_rs[i]=1ll<<((src>>16)&0x1F); // SWC2
6880 break;
6881 case C2OP:
6882 gte_rs[i]=gte_reg_reads[src&0x3f];
6883 gte_rt[i]=gte_reg_writes[src&0x3f];
6884 gte_rt[i]|=1ll<<63; // every op changes flags
6885 if((src&0x3f)==GTE_MVMVA) {
6886 int v = (src >> 15) & 3;
6887 gte_rs[i]&=~0xe3fll;
6888 if(v==3) gte_rs[i]|=0xe00ll;
6889 else gte_rs[i]|=3ll<<(v*2);
6890 }
6891 break;
6892 case SYSCALL:
6893 case HLECALL:
6894 case INTCALL:
6895 dops[i].rs1=CCREG;
6896 break;
6897 default:
6898 break;
6899 }
6900}
6901
6902static noinline void pass1_disassemble(u_int pagelimit)
6903{
6904 int i, j, done = 0, ni_count = 0;
6905
6906 for (i = 0; !done; i++)
6907 {
6908 int force_j_to_interpreter = 0;
6909 unsigned int type, op, op2;
6910
6911 disassemble_one(i, source[i]);
6912 type = dops[i].itype;
6913 op = dops[i].opcode;
6914 op2 = dops[i].opcode2;
6915
6916 /* Calculate branch target addresses */
6917 if(type==UJUMP)
6918 cinfo[i].ba=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6919 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
6920 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6921 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
6922 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6923 else if(type==CJUMP||type==SJUMP)
6924 cinfo[i].ba=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6925
6926 /* simplify always (not)taken branches */
6927 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
6928 dops[i].rs1 = dops[i].rs2 = 0;
6929 if (!(op & 1)) {
6930 dops[i].itype = type = UJUMP;
6931 dops[i].rs2 = CCREG;
6932 }
6933 }
6934 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
6935 dops[i].itype = type = UJUMP;
6936
6937 dops[i].is_jump = type == RJUMP || type == UJUMP || type == CJUMP || type == SJUMP;
6938 dops[i].is_ujump = type == RJUMP || type == UJUMP;
6939 dops[i].is_load = type == LOAD || type == LOADLR || op == 0x32; // LWC2
6940 dops[i].is_delay_load = (dops[i].is_load || (source[i] & 0xf3d00000) == 0x40000000); // MFC/CFC
6941 dops[i].is_store = type == STORE || type == STORELR || op == 0x3a; // SWC2
6942 dops[i].is_exception = type == SYSCALL || type == HLECALL || type == INTCALL;
6943 dops[i].may_except = dops[i].is_exception || (type == ALU && (op2 == 0x20 || op2 == 0x22)) || op == 8;
6944
6945 if (((op & 0x37) == 0x21 || op == 0x25) // LH/SH/LHU
6946 && ((cinfo[i].imm & 1) || Config.PreciseExceptions))
6947 dops[i].may_except = 1;
6948 if (((op & 0x37) == 0x23 || (op & 0x37) == 0x32) // LW/SW/LWC2/SWC2
6949 && ((cinfo[i].imm & 3) || Config.PreciseExceptions))
6950 dops[i].may_except = 1;
6951
6952 /* rare messy cases to just pass over to the interpreter */
6953 if (i > 0 && dops[i-1].is_jump) {
6954 j = i - 1;
6955 // branch in delay slot?
6956 if (dops[i].is_jump) {
6957 // don't handle first branch and call interpreter if it's hit
6958 SysPrintf("branch in DS @%08x (%08x)\n", start + i*4, start);
6959 force_j_to_interpreter = 1;
6960 }
6961 // load delay detection through a branch
6962 else if (dops[i].is_delay_load && dops[i].rt1 != 0) {
6963 const struct decoded_insn *dop = NULL;
6964 int t = -1;
6965 if (cinfo[i-1].ba != -1) {
6966 t = (cinfo[i-1].ba - start) / 4;
6967 if (t < 0 || t > i) {
6968 u_int limit = 0;
6969 u_int *mem = get_source_start(cinfo[i-1].ba, &limit);
6970 if (mem != NULL) {
6971 disassemble_one(MAXBLOCK - 1, mem[0]);
6972 dop = &dops[MAXBLOCK - 1];
6973 }
6974 }
6975 else
6976 dop = &dops[t];
6977 }
6978 if ((dop && is_ld_use_hazard(dops[i].rt1, dop))
6979 || (!dop && Config.PreciseExceptions)) {
6980 // jump target wants DS result - potential load delay effect
6981 SysPrintf("load delay in DS @%08x (%08x)\n", start + i*4, start);
6982 force_j_to_interpreter = 1;
6983 if (0 <= t && t < i)
6984 dops[t + 1].bt = 1; // expected return from interpreter
6985 }
6986 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&&
6987 !(i>=3&&dops[i-3].is_jump)) {
6988 // v0 overwrite like this is a sign of trouble, bail out
6989 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
6990 force_j_to_interpreter = 1;
6991 }
6992 }
6993 }
6994 else if (i > 0 && dops[i-1].is_delay_load
6995 && is_ld_use_hazard(dops[i-1].rt1, &dops[i])
6996 && (i < 2 || !dops[i-2].is_ujump)) {
6997 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
6998 for (j = i - 1; j > 0 && dops[j-1].is_delay_load; j--)
6999 if (dops[j-1].rt1 != dops[i-1].rt1)
7000 break;
7001 force_j_to_interpreter = 1;
7002 }
7003 if (force_j_to_interpreter) {
7004 force_intcall(j);
7005 done = 2;
7006 i = j; // don't compile the problematic branch/load/etc
7007 }
7008 if (dops[i].is_exception && i > 0 && dops[i-1].is_jump) {
7009 SysPrintf("exception in DS @%08x (%08x)\n", start + i*4, start);
7010 i--;
7011 force_intcall(i);
7012 done = 2;
7013 }
7014 if (i >= 2 && (source[i-2] & 0xffe0f800) == 0x40806000) // MTC0 $12
7015 dops[i].bt = 1;
7016 if (i >= 1 && (source[i-1] & 0xffe0f800) == 0x40806800) // MTC0 $13
7017 dops[i].bt = 1;
7018
7019 /* Is this the end of the block? */
7020 if (i > 0 && dops[i-1].is_ujump) {
7021 if (dops[i-1].rt1 == 0) { // not jal
7022 int found_bbranch = 0, t = (cinfo[i-1].ba - start) / 4;
7023 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
7024 // scan for a branch back to i+1
7025 for (j = t; j < t + 64; j++) {
7026 int tmpop = source[j] >> 26;
7027 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
7028 int t2 = j + 1 + (int)(signed short)source[j];
7029 if (t2 == i + 1) {
7030 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
7031 found_bbranch = 1;
7032 break;
7033 }
7034 }
7035 }
7036 }
7037 if (!found_bbranch)
7038 done = 2;
7039 }
7040 else {
7041 if(stop_after_jal) done=1;
7042 // Stop on BREAK
7043 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7044 }
7045 // Don't recompile stuff that's already compiled
7046 if(check_addr(start+i*4+4)) done=1;
7047 // Don't get too close to the limit
7048 if(i>MAXBLOCK/2) done=1;
7049 }
7050 if (dops[i].itype == HLECALL)
7051 stop = 1;
7052 else if (dops[i].itype == INTCALL)
7053 stop = 2;
7054 else if (dops[i].is_exception)
7055 done = stop_after_jal ? 1 : 2;
7056 if (done == 2) {
7057 // Does the block continue due to a branch?
7058 for(j=i-1;j>=0;j--)
7059 {
7060 if(cinfo[j].ba==start+i*4) done=j=0; // Branch into delay slot
7061 if(cinfo[j].ba==start+i*4+4) done=j=0;
7062 if(cinfo[j].ba==start+i*4+8) done=j=0;
7063 }
7064 }
7065 //assert(i<MAXBLOCK-1);
7066 if(start+i*4==pagelimit-4) done=1;
7067 assert(start+i*4<pagelimit);
7068 if (i==MAXBLOCK-1) done=1;
7069 // Stop if we're compiling junk
7070 if (dops[i].itype == INTCALL && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7071 done=stop_after_jal=1;
7072 SysPrintf("Disabled speculative precompilation\n");
7073 }
7074 }
7075 while (i > 0 && dops[i-1].is_jump)
7076 i--;
7077 assert(i > 0);
7078 assert(!dops[i-1].is_jump);
7079 slen = i;
7080}
7081
7082// Basic liveness analysis for MIPS registers
7083static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7084{
7085 int i;
7086 uint64_t u,gte_u,b,gte_b;
7087 uint64_t temp_u,temp_gte_u=0;
7088 uint64_t gte_u_unknown=0;
7089 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7090 gte_u_unknown=~0ll;
7091 if(iend==slen-1) {
7092 u=1;
7093 gte_u=gte_u_unknown;
7094 }else{
7095 //u=unneeded_reg[iend+1];
7096 u=1;
7097 gte_u=gte_unneeded[iend+1];
7098 }
7099
7100 for (i=iend;i>=istart;i--)
7101 {
7102 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7103 if(dops[i].is_jump)
7104 {
7105 // If subroutine call, flag return address as a possible branch target
7106 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7107
7108 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7109 {
7110 // Branch out of this block, flush all regs
7111 u=1;
7112 gte_u=gte_u_unknown;
7113 branch_unneeded_reg[i]=u;
7114 // Merge in delay slot
7115 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7116 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7117 u|=1;
7118 gte_u|=gte_rt[i+1];
7119 gte_u&=~gte_rs[i+1];
7120 }
7121 else
7122 {
7123 // Internal branch, flag target
7124 dops[(cinfo[i].ba-start)>>2].bt=1;
7125 if(cinfo[i].ba<=start+i*4) {
7126 // Backward branch
7127 if(dops[i].is_ujump)
7128 {
7129 // Unconditional branch
7130 temp_u=1;
7131 temp_gte_u=0;
7132 } else {
7133 // Conditional branch (not taken case)
7134 temp_u=unneeded_reg[i+2];
7135 temp_gte_u&=gte_unneeded[i+2];
7136 }
7137 // Merge in delay slot
7138 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7139 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7140 temp_u|=1;
7141 temp_gte_u|=gte_rt[i+1];
7142 temp_gte_u&=~gte_rs[i+1];
7143 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7144 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7145 temp_u|=1;
7146 temp_gte_u|=gte_rt[i];
7147 temp_gte_u&=~gte_rs[i];
7148 unneeded_reg[i]=temp_u;
7149 gte_unneeded[i]=temp_gte_u;
7150 // Only go three levels deep. This recursion can take an
7151 // excessive amount of time if there are a lot of nested loops.
7152 if(r<2) {
7153 pass2_unneeded_regs((cinfo[i].ba-start)>>2,i-1,r+1);
7154 }else{
7155 unneeded_reg[(cinfo[i].ba-start)>>2]=1;
7156 gte_unneeded[(cinfo[i].ba-start)>>2]=gte_u_unknown;
7157 }
7158 } /*else*/ if(1) {
7159 if (dops[i].is_ujump)
7160 {
7161 // Unconditional branch
7162 u=unneeded_reg[(cinfo[i].ba-start)>>2];
7163 gte_u=gte_unneeded[(cinfo[i].ba-start)>>2];
7164 branch_unneeded_reg[i]=u;
7165 // Merge in delay slot
7166 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7167 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7168 u|=1;
7169 gte_u|=gte_rt[i+1];
7170 gte_u&=~gte_rs[i+1];
7171 } else {
7172 // Conditional branch
7173 b=unneeded_reg[(cinfo[i].ba-start)>>2];
7174 gte_b=gte_unneeded[(cinfo[i].ba-start)>>2];
7175 branch_unneeded_reg[i]=b;
7176 // Branch delay slot
7177 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7178 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7179 b|=1;
7180 gte_b|=gte_rt[i+1];
7181 gte_b&=~gte_rs[i+1];
7182 u&=b;
7183 gte_u&=gte_b;
7184 if(i<slen-1) {
7185 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7186 } else {
7187 branch_unneeded_reg[i]=1;
7188 }
7189 }
7190 }
7191 }
7192 }
7193 else if(dops[i].may_except)
7194 {
7195 // SYSCALL instruction, etc or conditional exception
7196 u=1;
7197 }
7198 else if (dops[i].itype == RFE)
7199 {
7200 u=1;
7201 }
7202 //u=1; // DEBUG
7203 // Written registers are unneeded
7204 u|=1LL<<dops[i].rt1;
7205 u|=1LL<<dops[i].rt2;
7206 gte_u|=gte_rt[i];
7207 // Accessed registers are needed
7208 u&=~(1LL<<dops[i].rs1);
7209 u&=~(1LL<<dops[i].rs2);
7210 gte_u&=~gte_rs[i];
7211 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7212 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7213 // Source-target dependencies
7214 // R0 is always unneeded
7215 u|=1;
7216 // Save it
7217 unneeded_reg[i]=u;
7218 gte_unneeded[i]=gte_u;
7219 /*
7220 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7221 printf("U:");
7222 int r;
7223 for(r=1;r<=CCREG;r++) {
7224 if((unneeded_reg[i]>>r)&1) {
7225 if(r==HIREG) printf(" HI");
7226 else if(r==LOREG) printf(" LO");
7227 else printf(" r%d",r);
7228 }
7229 }
7230 printf("\n");
7231 */
7232 }
7233}
7234
7235static noinline void pass3_register_alloc(u_int addr)
7236{
7237 struct regstat current; // Current register allocations/status
7238 clear_all_regs(current.regmap_entry);
7239 clear_all_regs(current.regmap);
7240 current.wasdirty = current.dirty = 0;
7241 current.u = unneeded_reg[0];
7242 alloc_reg(&current, 0, CCREG);
7243 dirty_reg(&current, CCREG);
7244 current.wasconst = 0;
7245 current.isconst = 0;
7246 current.loadedconst = 0;
7247 //current.waswritten = 0;
7248 int ds=0;
7249 int cc=0;
7250 int hr;
7251 int i, j;
7252
7253 if (addr & 1) {
7254 // First instruction is delay slot
7255 cc=-1;
7256 dops[1].bt=1;
7257 ds=1;
7258 unneeded_reg[0]=1;
7259 current.regmap[HOST_BTREG]=BTREG;
7260 }
7261
7262 for(i=0;i<slen;i++)
7263 {
7264 if(dops[i].bt)
7265 {
7266 for(hr=0;hr<HOST_REGS;hr++)
7267 {
7268 // Is this really necessary?
7269 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7270 }
7271 current.isconst=0;
7272 //current.waswritten=0;
7273 }
7274
7275 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7276 regs[i].wasconst=current.isconst;
7277 regs[i].wasdirty=current.dirty;
7278 regs[i].dirty=0;
7279 regs[i].u=0;
7280 regs[i].isconst=0;
7281 regs[i].loadedconst=0;
7282 if (!dops[i].is_jump) {
7283 if(i+1<slen) {
7284 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7285 current.u|=1;
7286 } else {
7287 current.u=1;
7288 }
7289 } else {
7290 if(i+1<slen) {
7291 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7292 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7293 current.u|=1;
7294 } else {
7295 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7296 abort();
7297 }
7298 }
7299 dops[i].is_ds=ds;
7300 if(ds) {
7301 ds=0; // Skip delay slot, already allocated as part of branch
7302 // ...but we need to alloc it in case something jumps here
7303 if(i+1<slen) {
7304 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7305 }else{
7306 current.u=branch_unneeded_reg[i-1];
7307 }
7308 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7309 current.u|=1;
7310 struct regstat temp;
7311 memcpy(&temp,&current,sizeof(current));
7312 temp.wasdirty=temp.dirty;
7313 // TODO: Take into account unconditional branches, as below
7314 delayslot_alloc(&temp,i);
7315 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7316 regs[i].wasdirty=temp.wasdirty;
7317 regs[i].dirty=temp.dirty;
7318 regs[i].isconst=0;
7319 regs[i].wasconst=0;
7320 current.isconst=0;
7321 // Create entry (branch target) regmap
7322 for(hr=0;hr<HOST_REGS;hr++)
7323 {
7324 int r=temp.regmap[hr];
7325 if(r>=0) {
7326 if(r!=regmap_pre[i][hr]) {
7327 regs[i].regmap_entry[hr]=-1;
7328 }
7329 else
7330 {
7331 assert(r < 64);
7332 if((current.u>>r)&1) {
7333 regs[i].regmap_entry[hr]=-1;
7334 regs[i].regmap[hr]=-1;
7335 //Don't clear regs in the delay slot as the branch might need them
7336 //current.regmap[hr]=-1;
7337 }else
7338 regs[i].regmap_entry[hr]=r;
7339 }
7340 } else {
7341 // First instruction expects CCREG to be allocated
7342 if(i==0&&hr==HOST_CCREG)
7343 regs[i].regmap_entry[hr]=CCREG;
7344 else
7345 regs[i].regmap_entry[hr]=-1;
7346 }
7347 }
7348 }
7349 else { // Not delay slot
7350 switch(dops[i].itype) {
7351 case UJUMP:
7352 //current.isconst=0; // DEBUG
7353 //current.wasconst=0; // DEBUG
7354 //regs[i].wasconst=0; // DEBUG
7355 clear_const(&current,dops[i].rt1);
7356 alloc_cc(&current,i);
7357 dirty_reg(&current,CCREG);
7358 if (dops[i].rt1==31) {
7359 alloc_reg(&current,i,31);
7360 dirty_reg(&current,31);
7361 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7362 //assert(dops[i+1].rt1!=dops[i].rt1);
7363 #ifdef REG_PREFETCH
7364 alloc_reg(&current,i,PTEMP);
7365 #endif
7366 }
7367 dops[i].ooo=1;
7368 delayslot_alloc(&current,i+1);
7369 //current.isconst=0; // DEBUG
7370 ds=1;
7371 break;
7372 case RJUMP:
7373 //current.isconst=0;
7374 //current.wasconst=0;
7375 //regs[i].wasconst=0;
7376 clear_const(&current,dops[i].rs1);
7377 clear_const(&current,dops[i].rt1);
7378 alloc_cc(&current,i);
7379 dirty_reg(&current,CCREG);
7380 if (!ds_writes_rjump_rs(i)) {
7381 alloc_reg(&current,i,dops[i].rs1);
7382 if (dops[i].rt1!=0) {
7383 alloc_reg(&current,i,dops[i].rt1);
7384 dirty_reg(&current,dops[i].rt1);
7385 #ifdef REG_PREFETCH
7386 alloc_reg(&current,i,PTEMP);
7387 #endif
7388 }
7389 #ifdef USE_MINI_HT
7390 if(dops[i].rs1==31) { // JALR
7391 alloc_reg(&current,i,RHASH);
7392 alloc_reg(&current,i,RHTBL);
7393 }
7394 #endif
7395 delayslot_alloc(&current,i+1);
7396 } else {
7397 // The delay slot overwrites our source register,
7398 // allocate a temporary register to hold the old value.
7399 current.isconst=0;
7400 current.wasconst=0;
7401 regs[i].wasconst=0;
7402 delayslot_alloc(&current,i+1);
7403 current.isconst=0;
7404 alloc_reg(&current,i,RTEMP);
7405 }
7406 //current.isconst=0; // DEBUG
7407 dops[i].ooo=1;
7408 ds=1;
7409 break;
7410 case CJUMP:
7411 //current.isconst=0;
7412 //current.wasconst=0;
7413 //regs[i].wasconst=0;
7414 clear_const(&current,dops[i].rs1);
7415 clear_const(&current,dops[i].rs2);
7416 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7417 {
7418 alloc_cc(&current,i);
7419 dirty_reg(&current,CCREG);
7420 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7421 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7422 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7423 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7424 // The delay slot overwrites one of our conditions.
7425 // Allocate the branch condition registers instead.
7426 current.isconst=0;
7427 current.wasconst=0;
7428 regs[i].wasconst=0;
7429 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7430 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7431 }
7432 else
7433 {
7434 dops[i].ooo=1;
7435 delayslot_alloc(&current,i+1);
7436 }
7437 }
7438 else
7439 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7440 {
7441 alloc_cc(&current,i);
7442 dirty_reg(&current,CCREG);
7443 alloc_reg(&current,i,dops[i].rs1);
7444 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7445 // The delay slot overwrites one of our conditions.
7446 // Allocate the branch condition registers instead.
7447 current.isconst=0;
7448 current.wasconst=0;
7449 regs[i].wasconst=0;
7450 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7451 }
7452 else
7453 {
7454 dops[i].ooo=1;
7455 delayslot_alloc(&current,i+1);
7456 }
7457 }
7458 else
7459 // Don't alloc the delay slot yet because we might not execute it
7460 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7461 {
7462 current.isconst=0;
7463 current.wasconst=0;
7464 regs[i].wasconst=0;
7465 alloc_cc(&current,i);
7466 dirty_reg(&current,CCREG);
7467 alloc_reg(&current,i,dops[i].rs1);
7468 alloc_reg(&current,i,dops[i].rs2);
7469 }
7470 else
7471 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7472 {
7473 current.isconst=0;
7474 current.wasconst=0;
7475 regs[i].wasconst=0;
7476 alloc_cc(&current,i);
7477 dirty_reg(&current,CCREG);
7478 alloc_reg(&current,i,dops[i].rs1);
7479 }
7480 ds=1;
7481 //current.isconst=0;
7482 break;
7483 case SJUMP:
7484 clear_const(&current,dops[i].rs1);
7485 clear_const(&current,dops[i].rt1);
7486 {
7487 alloc_cc(&current,i);
7488 dirty_reg(&current,CCREG);
7489 alloc_reg(&current,i,dops[i].rs1);
7490 if (dops[i].rt1 == 31) { // BLTZAL/BGEZAL
7491 alloc_reg(&current,i,31);
7492 dirty_reg(&current,31);
7493 }
7494 if ((dops[i].rs1 &&
7495 (dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7496 ||(dops[i].rt1 == 31 && dops[i].rs1 == 31) // overwrites it's own condition
7497 ||(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
7498 // Allocate the branch condition registers instead.
7499 current.isconst=0;
7500 current.wasconst=0;
7501 regs[i].wasconst=0;
7502 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7503 }
7504 else
7505 {
7506 dops[i].ooo=1;
7507 delayslot_alloc(&current,i+1);
7508 }
7509 }
7510 ds=1;
7511 //current.isconst=0;
7512 break;
7513 case IMM16:
7514 imm16_alloc(&current,i);
7515 break;
7516 case LOAD:
7517 case LOADLR:
7518 load_alloc(&current,i);
7519 break;
7520 case STORE:
7521 case STORELR:
7522 store_alloc(&current,i);
7523 break;
7524 case ALU:
7525 alu_alloc(&current,i);
7526 break;
7527 case SHIFT:
7528 shift_alloc(&current,i);
7529 break;
7530 case MULTDIV:
7531 multdiv_alloc(&current,i);
7532 break;
7533 case SHIFTIMM:
7534 shiftimm_alloc(&current,i);
7535 break;
7536 case MOV:
7537 mov_alloc(&current,i);
7538 break;
7539 case COP0:
7540 cop0_alloc(&current,i);
7541 break;
7542 case RFE:
7543 rfe_alloc(&current,i);
7544 break;
7545 case COP2:
7546 cop2_alloc(&current,i);
7547 break;
7548 case C2LS:
7549 c2ls_alloc(&current,i);
7550 break;
7551 case C2OP:
7552 c2op_alloc(&current,i);
7553 break;
7554 case SYSCALL:
7555 case HLECALL:
7556 case INTCALL:
7557 syscall_alloc(&current,i);
7558 break;
7559 }
7560
7561 // Create entry (branch target) regmap
7562 for(hr=0;hr<HOST_REGS;hr++)
7563 {
7564 int r,or;
7565 r=current.regmap[hr];
7566 if(r>=0) {
7567 if(r!=regmap_pre[i][hr]) {
7568 // TODO: delay slot (?)
7569 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7570 if(or<0||r>=TEMPREG){
7571 regs[i].regmap_entry[hr]=-1;
7572 }
7573 else
7574 {
7575 // Just move it to a different register
7576 regs[i].regmap_entry[hr]=r;
7577 // If it was dirty before, it's still dirty
7578 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7579 }
7580 }
7581 else
7582 {
7583 // Unneeded
7584 if(r==0){
7585 regs[i].regmap_entry[hr]=0;
7586 }
7587 else
7588 {
7589 assert(r<64);
7590 if((current.u>>r)&1) {
7591 regs[i].regmap_entry[hr]=-1;
7592 //regs[i].regmap[hr]=-1;
7593 current.regmap[hr]=-1;
7594 }else
7595 regs[i].regmap_entry[hr]=r;
7596 }
7597 }
7598 } else {
7599 // Branches expect CCREG to be allocated at the target
7600 if(regmap_pre[i][hr]==CCREG)
7601 regs[i].regmap_entry[hr]=CCREG;
7602 else
7603 regs[i].regmap_entry[hr]=-1;
7604 }
7605 }
7606 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7607 }
7608
7609#if 0 // see do_store_smc_check()
7610 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)
7611 current.waswritten|=1<<dops[i-1].rs1;
7612 current.waswritten&=~(1<<dops[i].rt1);
7613 current.waswritten&=~(1<<dops[i].rt2);
7614 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)cinfo[i].imm>=0x800)
7615 current.waswritten&=~(1<<dops[i].rs1);
7616#endif
7617
7618 /* Branch post-alloc */
7619 if(i>0)
7620 {
7621 current.wasdirty=current.dirty;
7622 switch(dops[i-1].itype) {
7623 case UJUMP:
7624 memcpy(&branch_regs[i-1],&current,sizeof(current));
7625 branch_regs[i-1].isconst=0;
7626 branch_regs[i-1].wasconst=0;
7627 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7628 alloc_cc(&branch_regs[i-1],i-1);
7629 dirty_reg(&branch_regs[i-1],CCREG);
7630 if(dops[i-1].rt1==31) { // JAL
7631 alloc_reg(&branch_regs[i-1],i-1,31);
7632 dirty_reg(&branch_regs[i-1],31);
7633 }
7634 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7635 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7636 break;
7637 case RJUMP:
7638 memcpy(&branch_regs[i-1],&current,sizeof(current));
7639 branch_regs[i-1].isconst=0;
7640 branch_regs[i-1].wasconst=0;
7641 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7642 alloc_cc(&branch_regs[i-1],i-1);
7643 dirty_reg(&branch_regs[i-1],CCREG);
7644 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7645 if(dops[i-1].rt1!=0) { // JALR
7646 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7647 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7648 }
7649 #ifdef USE_MINI_HT
7650 if(dops[i-1].rs1==31) { // JALR
7651 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7652 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7653 }
7654 #endif
7655 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7656 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7657 break;
7658 case CJUMP:
7659 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7660 {
7661 alloc_cc(&current,i-1);
7662 dirty_reg(&current,CCREG);
7663 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7664 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7665 // The delay slot overwrote one of our conditions
7666 // Delay slot goes after the test (in order)
7667 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7668 current.u|=1;
7669 delayslot_alloc(&current,i);
7670 current.isconst=0;
7671 }
7672 else
7673 {
7674 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7675 // Alloc the branch condition registers
7676 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7677 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7678 }
7679 memcpy(&branch_regs[i-1],&current,sizeof(current));
7680 branch_regs[i-1].isconst=0;
7681 branch_regs[i-1].wasconst=0;
7682 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7683 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7684 }
7685 else
7686 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7687 {
7688 alloc_cc(&current,i-1);
7689 dirty_reg(&current,CCREG);
7690 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7691 // The delay slot overwrote the branch condition
7692 // Delay slot goes after the test (in order)
7693 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7694 current.u|=1;
7695 delayslot_alloc(&current,i);
7696 current.isconst=0;
7697 }
7698 else
7699 {
7700 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7701 // Alloc the branch condition register
7702 alloc_reg(&current,i-1,dops[i-1].rs1);
7703 }
7704 memcpy(&branch_regs[i-1],&current,sizeof(current));
7705 branch_regs[i-1].isconst=0;
7706 branch_regs[i-1].wasconst=0;
7707 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7708 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7709 }
7710 else
7711 // Alloc the delay slot in case the branch is taken
7712 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
7713 {
7714 memcpy(&branch_regs[i-1],&current,sizeof(current));
7715 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;
7716 alloc_cc(&branch_regs[i-1],i);
7717 dirty_reg(&branch_regs[i-1],CCREG);
7718 delayslot_alloc(&branch_regs[i-1],i);
7719 branch_regs[i-1].isconst=0;
7720 alloc_reg(&current,i,CCREG); // Not taken path
7721 dirty_reg(&current,CCREG);
7722 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7723 }
7724 else
7725 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
7726 {
7727 memcpy(&branch_regs[i-1],&current,sizeof(current));
7728 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;
7729 alloc_cc(&branch_regs[i-1],i);
7730 dirty_reg(&branch_regs[i-1],CCREG);
7731 delayslot_alloc(&branch_regs[i-1],i);
7732 branch_regs[i-1].isconst=0;
7733 alloc_reg(&current,i,CCREG); // Not taken path
7734 dirty_reg(&current,CCREG);
7735 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7736 }
7737 break;
7738 case SJUMP:
7739 {
7740 alloc_cc(&current,i-1);
7741 dirty_reg(&current,CCREG);
7742 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7743 // The delay slot overwrote the branch condition
7744 // Delay slot goes after the test (in order)
7745 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7746 current.u|=1;
7747 delayslot_alloc(&current,i);
7748 current.isconst=0;
7749 }
7750 else
7751 {
7752 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7753 // Alloc the branch condition register
7754 alloc_reg(&current,i-1,dops[i-1].rs1);
7755 }
7756 memcpy(&branch_regs[i-1],&current,sizeof(current));
7757 branch_regs[i-1].isconst=0;
7758 branch_regs[i-1].wasconst=0;
7759 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7760 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7761 }
7762 // FIXME: BLTZAL/BGEZAL
7763 if ((dops[i-1].opcode2 & 0x1e) == 0x10) { // BxxZAL
7764 alloc_reg(&branch_regs[i-1],i-1,31);
7765 dirty_reg(&branch_regs[i-1],31);
7766 }
7767 break;
7768 }
7769
7770 if (dops[i-1].is_ujump)
7771 {
7772 if(dops[i-1].rt1==31) // JAL/JALR
7773 {
7774 // Subroutine call will return here, don't alloc any registers
7775 current.dirty=0;
7776 clear_all_regs(current.regmap);
7777 alloc_reg(&current,i,CCREG);
7778 dirty_reg(&current,CCREG);
7779 }
7780 else if(i+1<slen)
7781 {
7782 // Internal branch will jump here, match registers to caller
7783 current.dirty=0;
7784 clear_all_regs(current.regmap);
7785 alloc_reg(&current,i,CCREG);
7786 dirty_reg(&current,CCREG);
7787 for(j=i-1;j>=0;j--)
7788 {
7789 if(cinfo[j].ba==start+i*4+4) {
7790 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7791 current.dirty=branch_regs[j].dirty;
7792 break;
7793 }
7794 }
7795 while(j>=0) {
7796 if(cinfo[j].ba==start+i*4+4) {
7797 for(hr=0;hr<HOST_REGS;hr++) {
7798 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7799 current.regmap[hr]=-1;
7800 }
7801 current.dirty&=branch_regs[j].dirty;
7802 }
7803 }
7804 j--;
7805 }
7806 }
7807 }
7808 }
7809
7810 // Count cycles in between branches
7811 cinfo[i].ccadj = CLOCK_ADJUST(cc);
7812 if (i > 0 && (dops[i-1].is_jump || dops[i].is_exception))
7813 {
7814 cc=0;
7815 }
7816#if !defined(DRC_DBG)
7817 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7818 {
7819 // this should really be removed since the real stalls have been implemented,
7820 // but doing so causes sizeable perf regression against the older version
7821 u_int gtec = gte_cycletab[source[i] & 0x3f];
7822 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7823 }
7824 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7825 {
7826 cc+=4;
7827 }
7828 else if(dops[i].itype==C2LS)
7829 {
7830 // same as with C2OP
7831 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7832 }
7833#endif
7834 else
7835 {
7836 cc++;
7837 }
7838
7839 if(!dops[i].is_ds) {
7840 regs[i].dirty=current.dirty;
7841 regs[i].isconst=current.isconst;
7842 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7843 }
7844 for(hr=0;hr<HOST_REGS;hr++) {
7845 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7846 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7847 regs[i].wasconst&=~(1<<hr);
7848 }
7849 }
7850 }
7851 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7852 //regs[i].waswritten=current.waswritten;
7853 }
7854}
7855
7856static noinline void pass4_cull_unused_regs(void)
7857{
7858 u_int last_needed_regs[4] = {0,0,0,0};
7859 u_int nr=0;
7860 int i;
7861
7862 for (i=slen-1;i>=0;i--)
7863 {
7864 int hr;
7865 __builtin_prefetch(regs[i-2].regmap);
7866 if(dops[i].is_jump)
7867 {
7868 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7869 {
7870 // Branch out of this block, don't need anything
7871 nr=0;
7872 }
7873 else
7874 {
7875 // Internal branch
7876 // Need whatever matches the target
7877 nr=0;
7878 int t=(cinfo[i].ba-start)>>2;
7879 for(hr=0;hr<HOST_REGS;hr++)
7880 {
7881 if(regs[i].regmap_entry[hr]>=0) {
7882 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7883 }
7884 }
7885 }
7886 // Conditional branch may need registers for following instructions
7887 if (!dops[i].is_ujump)
7888 {
7889 if(i<slen-2) {
7890 nr |= last_needed_regs[(i+2) & 3];
7891 for(hr=0;hr<HOST_REGS;hr++)
7892 {
7893 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7894 //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]);
7895 }
7896 }
7897 }
7898 // Don't need stuff which is overwritten
7899 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7900 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7901 // Merge in delay slot
7902 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
7903 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
7904 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
7905 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
7906 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
7907 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
7908 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
7909 nr |= get_regm(regmap_pre[i], ROREG);
7910 nr |= get_regm(regs[i].regmap_entry, ROREG);
7911 }
7912 if (dops[i+1].is_store) {
7913 nr |= get_regm(regmap_pre[i], INVCP);
7914 nr |= get_regm(regs[i].regmap_entry, INVCP);
7915 }
7916 }
7917 else if (dops[i].is_exception)
7918 {
7919 // SYSCALL instruction, etc
7920 nr=0;
7921 }
7922 else // Non-branch
7923 {
7924 if(i<slen-1) {
7925 for(hr=0;hr<HOST_REGS;hr++) {
7926 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7927 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7928 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7929 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7930 }
7931 }
7932 }
7933 // Overwritten registers are not needed
7934 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
7935 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
7936 nr &= ~get_regm(regs[i].regmap, FTEMP);
7937 // Source registers are needed
7938 nr |= get_regm(regmap_pre[i], dops[i].rs1);
7939 nr |= get_regm(regmap_pre[i], dops[i].rs2);
7940 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
7941 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
7942 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
7943 nr |= get_regm(regmap_pre[i], ROREG);
7944 nr |= get_regm(regs[i].regmap_entry, ROREG);
7945 }
7946 if (dops[i].is_store) {
7947 nr |= get_regm(regmap_pre[i], INVCP);
7948 nr |= get_regm(regs[i].regmap_entry, INVCP);
7949 }
7950
7951 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
7952 for(hr=0;hr<HOST_REGS;hr++)
7953 {
7954 // Don't store a register immediately after writing it,
7955 // may prevent dual-issue.
7956 // But do so if this is a branch target, otherwise we
7957 // might have to load the register before the branch.
7958 if((regs[i].wasdirty>>hr)&1) {
7959 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7960 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
7961 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
7962 }
7963 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7964 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
7965 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
7966 }
7967 }
7968 }
7969 // Cycle count is needed at branches. Assume it is needed at the target too.
7970 if (i == 0 || dops[i].bt || dops[i].may_except || dops[i].itype == CJUMP) {
7971 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7972 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7973 }
7974 // Save it
7975 last_needed_regs[i & 3] = nr;
7976
7977 // Deallocate unneeded registers
7978 for(hr=0;hr<HOST_REGS;hr++)
7979 {
7980 if(!((nr>>hr)&1)) {
7981 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
7982 if(dops[i].is_jump)
7983 {
7984 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
7985 if (dops[i+1].is_load || dops[i+1].is_store)
7986 map1 = ROREG;
7987 if (dops[i+1].is_store)
7988 map2 = INVCP;
7989 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
7990 temp = FTEMP;
7991 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
7992 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
7993 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
7994 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
7995 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
7996 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
7997 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
7998 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
7999 {
8000 regs[i].regmap[hr]=-1;
8001 regs[i].isconst&=~(1<<hr);
8002 regs[i].dirty&=~(1<<hr);
8003 regs[i+1].wasdirty&=~(1<<hr);
8004 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
8005 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
8006 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
8007 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8008 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8009 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8010 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8011 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8012 {
8013 branch_regs[i].regmap[hr]=-1;
8014 branch_regs[i].regmap_entry[hr]=-1;
8015 if (!dops[i].is_ujump)
8016 {
8017 if (i < slen-2) {
8018 regmap_pre[i+2][hr]=-1;
8019 regs[i+2].wasconst&=~(1<<hr);
8020 }
8021 }
8022 }
8023 }
8024 }
8025 else
8026 {
8027 // Non-branch
8028 if(i>0)
8029 {
8030 int map1 = -1, map2 = -1, temp=-1;
8031 if (dops[i].is_load || dops[i].is_store)
8032 map1 = ROREG;
8033 if (dops[i].is_store)
8034 map2 = INVCP;
8035 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8036 temp = FTEMP;
8037 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8038 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8039 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8040 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8041 regs[i].regmap[hr] != CCREG)
8042 {
8043 if(i<slen-1&&!dops[i].is_ds) {
8044 assert(regs[i].regmap[hr]<64);
8045 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8046 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8047 {
8048 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8049 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8050 }
8051 regmap_pre[i+1][hr]=-1;
8052 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8053 regs[i+1].wasconst&=~(1<<hr);
8054 }
8055 regs[i].regmap[hr]=-1;
8056 regs[i].isconst&=~(1<<hr);
8057 regs[i].dirty&=~(1<<hr);
8058 regs[i+1].wasdirty&=~(1<<hr);
8059 }
8060 }
8061 }
8062 } // if needed
8063 } // for hr
8064 }
8065}
8066
8067// If a register is allocated during a loop, try to allocate it for the
8068// entire loop, if possible. This avoids loading/storing registers
8069// inside of the loop.
8070static noinline void pass5a_preallocate1(void)
8071{
8072 int i, j, hr;
8073 signed char f_regmap[HOST_REGS];
8074 clear_all_regs(f_regmap);
8075 for(i=0;i<slen-1;i++)
8076 {
8077 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8078 {
8079 if(cinfo[i].ba>=start && cinfo[i].ba<(start+i*4))
8080 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8081 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8082 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8083 ||dops[i+1].itype==SHIFT
8084 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8085 {
8086 int t=(cinfo[i].ba-start)>>2;
8087 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8088 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8089 for(hr=0;hr<HOST_REGS;hr++)
8090 {
8091 if(regs[i].regmap[hr]>=0) {
8092 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8093 // dealloc old register
8094 int n;
8095 for(n=0;n<HOST_REGS;n++)
8096 {
8097 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8098 }
8099 // and alloc new one
8100 f_regmap[hr]=regs[i].regmap[hr];
8101 }
8102 }
8103 if(branch_regs[i].regmap[hr]>=0) {
8104 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8105 // dealloc old register
8106 int n;
8107 for(n=0;n<HOST_REGS;n++)
8108 {
8109 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8110 }
8111 // and alloc new one
8112 f_regmap[hr]=branch_regs[i].regmap[hr];
8113 }
8114 }
8115 if(dops[i].ooo) {
8116 if(count_free_regs(regs[i].regmap)<=cinfo[i+1].min_free_regs)
8117 f_regmap[hr]=branch_regs[i].regmap[hr];
8118 }else{
8119 if(count_free_regs(branch_regs[i].regmap)<=cinfo[i+1].min_free_regs)
8120 f_regmap[hr]=branch_regs[i].regmap[hr];
8121 }
8122 // Avoid dirty->clean transition
8123 #ifdef DESTRUCTIVE_WRITEBACK
8124 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;
8125 #endif
8126 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8127 // case above, however it's always a good idea. We can't hoist the
8128 // load if the register was already allocated, so there's no point
8129 // wasting time analyzing most of these cases. It only "succeeds"
8130 // when the mapping was different and the load can be replaced with
8131 // a mov, which is of negligible benefit. So such cases are
8132 // skipped below.
8133 if(f_regmap[hr]>0) {
8134 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8135 int r=f_regmap[hr];
8136 for(j=t;j<=i;j++)
8137 {
8138 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8139 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8140 assert(r < 64);
8141 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8142 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8143 int k;
8144 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8145 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8146 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8147 k=i;
8148 while(k>1&&regs[k-1].regmap[hr]==-1) {
8149 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8150 //printf("no free regs for store %x\n",start+(k-1)*4);
8151 break;
8152 }
8153 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8154 //printf("no-match due to different register\n");
8155 break;
8156 }
8157 if (dops[k-2].is_jump) {
8158 //printf("no-match due to branch\n");
8159 break;
8160 }
8161 // call/ret fast path assumes no registers allocated
8162 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8163 break;
8164 }
8165 k--;
8166 }
8167 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8168 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8169 while(k<i) {
8170 regs[k].regmap_entry[hr]=f_regmap[hr];
8171 regs[k].regmap[hr]=f_regmap[hr];
8172 regmap_pre[k+1][hr]=f_regmap[hr];
8173 regs[k].wasdirty&=~(1<<hr);
8174 regs[k].dirty&=~(1<<hr);
8175 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8176 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8177 regs[k].wasconst&=~(1<<hr);
8178 regs[k].isconst&=~(1<<hr);
8179 k++;
8180 }
8181 }
8182 else {
8183 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8184 break;
8185 }
8186 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8187 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8188 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8189 regs[i].regmap_entry[hr]=f_regmap[hr];
8190 regs[i].regmap[hr]=f_regmap[hr];
8191 regs[i].wasdirty&=~(1<<hr);
8192 regs[i].dirty&=~(1<<hr);
8193 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8194 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8195 regs[i].wasconst&=~(1<<hr);
8196 regs[i].isconst&=~(1<<hr);
8197 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8198 branch_regs[i].wasdirty&=~(1<<hr);
8199 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8200 branch_regs[i].regmap[hr]=f_regmap[hr];
8201 branch_regs[i].dirty&=~(1<<hr);
8202 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8203 branch_regs[i].wasconst&=~(1<<hr);
8204 branch_regs[i].isconst&=~(1<<hr);
8205 if (!dops[i].is_ujump) {
8206 regmap_pre[i+2][hr]=f_regmap[hr];
8207 regs[i+2].wasdirty&=~(1<<hr);
8208 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8209 }
8210 }
8211 }
8212 for(k=t;k<j;k++) {
8213 // Alloc register clean at beginning of loop,
8214 // but may dirty it in pass 6
8215 regs[k].regmap_entry[hr]=f_regmap[hr];
8216 regs[k].regmap[hr]=f_regmap[hr];
8217 regs[k].dirty&=~(1<<hr);
8218 regs[k].wasconst&=~(1<<hr);
8219 regs[k].isconst&=~(1<<hr);
8220 if (dops[k].is_jump) {
8221 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8222 branch_regs[k].regmap[hr]=f_regmap[hr];
8223 branch_regs[k].dirty&=~(1<<hr);
8224 branch_regs[k].wasconst&=~(1<<hr);
8225 branch_regs[k].isconst&=~(1<<hr);
8226 if (!dops[k].is_ujump) {
8227 regmap_pre[k+2][hr]=f_regmap[hr];
8228 regs[k+2].wasdirty&=~(1<<hr);
8229 }
8230 }
8231 else
8232 {
8233 regmap_pre[k+1][hr]=f_regmap[hr];
8234 regs[k+1].wasdirty&=~(1<<hr);
8235 }
8236 }
8237 if(regs[j].regmap[hr]==f_regmap[hr])
8238 regs[j].regmap_entry[hr]=f_regmap[hr];
8239 break;
8240 }
8241 if(j==i) break;
8242 if(regs[j].regmap[hr]>=0)
8243 break;
8244 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8245 //printf("no-match due to different register\n");
8246 break;
8247 }
8248 if (dops[j].is_ujump)
8249 {
8250 // Stop on unconditional branch
8251 break;
8252 }
8253 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8254 {
8255 if(dops[j].ooo) {
8256 if(count_free_regs(regs[j].regmap)<=cinfo[j+1].min_free_regs)
8257 break;
8258 }else{
8259 if(count_free_regs(branch_regs[j].regmap)<=cinfo[j+1].min_free_regs)
8260 break;
8261 }
8262 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8263 //printf("no-match due to different register (branch)\n");
8264 break;
8265 }
8266 }
8267 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8268 //printf("No free regs for store %x\n",start+j*4);
8269 break;
8270 }
8271 assert(f_regmap[hr]<64);
8272 }
8273 }
8274 }
8275 }
8276 }
8277 }else{
8278 // Non branch or undetermined branch target
8279 for(hr=0;hr<HOST_REGS;hr++)
8280 {
8281 if(hr!=EXCLUDE_REG) {
8282 if(regs[i].regmap[hr]>=0) {
8283 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8284 // dealloc old register
8285 int n;
8286 for(n=0;n<HOST_REGS;n++)
8287 {
8288 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8289 }
8290 // and alloc new one
8291 f_regmap[hr]=regs[i].regmap[hr];
8292 }
8293 }
8294 }
8295 }
8296 // Try to restore cycle count at branch targets
8297 if(dops[i].bt) {
8298 for(j=i;j<slen-1;j++) {
8299 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8300 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8301 //printf("no free regs for store %x\n",start+j*4);
8302 break;
8303 }
8304 }
8305 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8306 int k=i;
8307 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8308 while(k<j) {
8309 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8310 regs[k].regmap[HOST_CCREG]=CCREG;
8311 regmap_pre[k+1][HOST_CCREG]=CCREG;
8312 regs[k+1].wasdirty|=1<<HOST_CCREG;
8313 regs[k].dirty|=1<<HOST_CCREG;
8314 regs[k].wasconst&=~(1<<HOST_CCREG);
8315 regs[k].isconst&=~(1<<HOST_CCREG);
8316 k++;
8317 }
8318 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8319 }
8320 // Work backwards from the branch target
8321 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8322 {
8323 //printf("Extend backwards\n");
8324 int k;
8325 k=i;
8326 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8327 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8328 //printf("no free regs for store %x\n",start+(k-1)*4);
8329 break;
8330 }
8331 k--;
8332 }
8333 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8334 //printf("Extend CC, %x ->\n",start+k*4);
8335 while(k<=i) {
8336 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8337 regs[k].regmap[HOST_CCREG]=CCREG;
8338 regmap_pre[k+1][HOST_CCREG]=CCREG;
8339 regs[k+1].wasdirty|=1<<HOST_CCREG;
8340 regs[k].dirty|=1<<HOST_CCREG;
8341 regs[k].wasconst&=~(1<<HOST_CCREG);
8342 regs[k].isconst&=~(1<<HOST_CCREG);
8343 k++;
8344 }
8345 }
8346 else {
8347 //printf("Fail Extend CC, %x ->\n",start+k*4);
8348 }
8349 }
8350 }
8351 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=SHIFT&&
8352 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8353 dops[i].itype!=IMM16&&dops[i].itype!=LOAD)
8354 {
8355 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8356 }
8357 }
8358 }
8359}
8360
8361// This allocates registers (if possible) one instruction prior
8362// to use, which can avoid a load-use penalty on certain CPUs.
8363static noinline void pass5b_preallocate2(void)
8364{
8365 int i, hr;
8366 for(i=0;i<slen-1;i++)
8367 {
8368 if (!i || !dops[i-1].is_jump)
8369 {
8370 if(!dops[i+1].bt)
8371 {
8372 int j, can_steal = 1;
8373 for (j = i; j < i + 2; j++) {
8374 int free_regs = 0;
8375 if (cinfo[j].min_free_regs == 0)
8376 continue;
8377 for (hr = 0; hr < HOST_REGS; hr++)
8378 if (hr != EXCLUDE_REG && regs[j].regmap[hr] < 0)
8379 free_regs++;
8380 if (free_regs <= cinfo[j].min_free_regs) {
8381 can_steal = 0;
8382 break;
8383 }
8384 }
8385 if (!can_steal)
8386 continue;
8387 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8388 ||(dops[i].itype==COP2&&dops[i].opcode2<3))
8389 {
8390 if(dops[i+1].rs1) {
8391 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8392 {
8393 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8394 {
8395 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8396 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8397 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8398 regs[i].isconst&=~(1<<hr);
8399 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8400 constmap[i][hr]=constmap[i+1][hr];
8401 regs[i+1].wasdirty&=~(1<<hr);
8402 regs[i].dirty&=~(1<<hr);
8403 }
8404 }
8405 }
8406 if(dops[i+1].rs2) {
8407 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8408 {
8409 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8410 {
8411 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8412 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8413 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8414 regs[i].isconst&=~(1<<hr);
8415 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8416 constmap[i][hr]=constmap[i+1][hr];
8417 regs[i+1].wasdirty&=~(1<<hr);
8418 regs[i].dirty&=~(1<<hr);
8419 }
8420 }
8421 }
8422 // Preload target address for load instruction (non-constant)
8423 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8424 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8425 {
8426 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8427 {
8428 regs[i].regmap[hr]=dops[i+1].rs1;
8429 regmap_pre[i+1][hr]=dops[i+1].rs1;
8430 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8431 regs[i].isconst&=~(1<<hr);
8432 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8433 constmap[i][hr]=constmap[i+1][hr];
8434 regs[i+1].wasdirty&=~(1<<hr);
8435 regs[i].dirty&=~(1<<hr);
8436 }
8437 }
8438 }
8439 // Load source into target register
8440 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8441 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8442 {
8443 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8444 {
8445 regs[i].regmap[hr]=dops[i+1].rs1;
8446 regmap_pre[i+1][hr]=dops[i+1].rs1;
8447 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8448 regs[i].isconst&=~(1<<hr);
8449 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8450 constmap[i][hr]=constmap[i+1][hr];
8451 regs[i+1].wasdirty&=~(1<<hr);
8452 regs[i].dirty&=~(1<<hr);
8453 }
8454 }
8455 }
8456 // Address for store instruction (non-constant)
8457 if (dops[i+1].is_store) { // SB/SH/SW/SWC2
8458 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8459 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8460 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8461 else {
8462 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8463 regs[i+1].isconst&=~(1<<hr);
8464 regs[i+1].dirty&=~(1<<hr);
8465 regs[i+2].wasdirty&=~(1<<hr);
8466 }
8467 assert(hr>=0);
8468 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8469 {
8470 regs[i].regmap[hr]=dops[i+1].rs1;
8471 regmap_pre[i+1][hr]=dops[i+1].rs1;
8472 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8473 regs[i].isconst&=~(1<<hr);
8474 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8475 constmap[i][hr]=constmap[i+1][hr];
8476 regs[i+1].wasdirty&=~(1<<hr);
8477 regs[i].dirty&=~(1<<hr);
8478 }
8479 }
8480 }
8481 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) { // LWC2
8482 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8483 int nr;
8484 hr=get_reg(regs[i+1].regmap,FTEMP);
8485 assert(hr>=0);
8486 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8487 {
8488 regs[i].regmap[hr]=dops[i+1].rs1;
8489 regmap_pre[i+1][hr]=dops[i+1].rs1;
8490 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8491 regs[i].isconst&=~(1<<hr);
8492 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8493 constmap[i][hr]=constmap[i+1][hr];
8494 regs[i+1].wasdirty&=~(1<<hr);
8495 regs[i].dirty&=~(1<<hr);
8496 }
8497 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8498 {
8499 // move it to another register
8500 regs[i+1].regmap[hr]=-1;
8501 regmap_pre[i+2][hr]=-1;
8502 regs[i+1].regmap[nr]=FTEMP;
8503 regmap_pre[i+2][nr]=FTEMP;
8504 regs[i].regmap[nr]=dops[i+1].rs1;
8505 regmap_pre[i+1][nr]=dops[i+1].rs1;
8506 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8507 regs[i].isconst&=~(1<<nr);
8508 regs[i+1].isconst&=~(1<<nr);
8509 regs[i].dirty&=~(1<<nr);
8510 regs[i+1].wasdirty&=~(1<<nr);
8511 regs[i+1].dirty&=~(1<<nr);
8512 regs[i+2].wasdirty&=~(1<<nr);
8513 }
8514 }
8515 }
8516 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*/) {
8517 hr = -1;
8518 if(dops[i+1].itype==LOAD)
8519 hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1);
8520 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) // LWC2
8521 hr=get_reg(regs[i+1].regmap,FTEMP);
8522 if (dops[i+1].is_store) {
8523 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8524 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8525 }
8526 if(hr>=0&&regs[i].regmap[hr]<0) {
8527 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8528 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8529 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8530 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8531 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8532 regs[i].isconst&=~(1<<hr);
8533 regs[i+1].wasdirty&=~(1<<hr);
8534 regs[i].dirty&=~(1<<hr);
8535 }
8536 }
8537 }
8538 }
8539 }
8540 }
8541 }
8542}
8543
8544// Write back dirty registers as soon as we will no longer modify them,
8545// so that we don't end up with lots of writes at the branches.
8546static noinline void pass6_clean_registers(int istart, int iend, int wr)
8547{
8548 static u_int wont_dirty[MAXBLOCK];
8549 static u_int will_dirty[MAXBLOCK];
8550 int i;
8551 int r;
8552 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8553 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8554 if(iend==slen-1) {
8555 will_dirty_i=will_dirty_next=0;
8556 wont_dirty_i=wont_dirty_next=0;
8557 }else{
8558 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8559 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8560 }
8561 for (i=iend;i>=istart;i--)
8562 {
8563 signed char rregmap_i[RRMAP_SIZE];
8564 u_int hr_candirty = 0;
8565 assert(HOST_REGS < 32);
8566 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8567 __builtin_prefetch(regs[i-1].regmap);
8568 if(dops[i].is_jump)
8569 {
8570 signed char branch_rregmap_i[RRMAP_SIZE];
8571 u_int branch_hr_candirty = 0;
8572 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8573 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
8574 {
8575 // Branch out of this block, flush all regs
8576 will_dirty_i = 0;
8577 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8578 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8579 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8580 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8581 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8582 will_dirty_i &= branch_hr_candirty;
8583 if (dops[i].is_ujump)
8584 {
8585 // Unconditional branch
8586 wont_dirty_i = 0;
8587 // Merge in delay slot (will dirty)
8588 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8589 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8590 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8591 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8592 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8593 will_dirty_i &= hr_candirty;
8594 }
8595 else
8596 {
8597 // Conditional branch
8598 wont_dirty_i = wont_dirty_next;
8599 // Merge in delay slot (will dirty)
8600 // (the original code had no explanation why these 2 are commented out)
8601 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8602 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8603 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8604 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8605 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8606 will_dirty_i &= hr_candirty;
8607 }
8608 // Merge in delay slot (wont dirty)
8609 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8610 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8611 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8612 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8613 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8614 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8615 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8616 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8617 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8618 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8619 wont_dirty_i &= ~(1u << 31);
8620 if(wr) {
8621 #ifndef DESTRUCTIVE_WRITEBACK
8622 branch_regs[i].dirty&=wont_dirty_i;
8623 #endif
8624 branch_regs[i].dirty|=will_dirty_i;
8625 }
8626 }
8627 else
8628 {
8629 // Internal branch
8630 if(cinfo[i].ba<=start+i*4) {
8631 // Backward branch
8632 if (dops[i].is_ujump)
8633 {
8634 // Unconditional branch
8635 temp_will_dirty=0;
8636 temp_wont_dirty=0;
8637 // Merge in delay slot (will dirty)
8638 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8639 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8640 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8641 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8642 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8643 temp_will_dirty &= branch_hr_candirty;
8644 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8645 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8646 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8647 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8648 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8649 temp_will_dirty &= hr_candirty;
8650 } else {
8651 // Conditional branch (not taken case)
8652 temp_will_dirty=will_dirty_next;
8653 temp_wont_dirty=wont_dirty_next;
8654 // Merge in delay slot (will dirty)
8655 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8656 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8657 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8658 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8659 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8660 temp_will_dirty &= branch_hr_candirty;
8661 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8662 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8663 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8664 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8665 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8666 temp_will_dirty &= hr_candirty;
8667 }
8668 // Merge in delay slot (wont dirty)
8669 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8670 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8671 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8672 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8673 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8674 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8675 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8676 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8677 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8678 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8679 temp_wont_dirty &= ~(1u << 31);
8680 // Deal with changed mappings
8681 if(i<iend) {
8682 for(r=0;r<HOST_REGS;r++) {
8683 if(r!=EXCLUDE_REG) {
8684 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8685 temp_will_dirty&=~(1<<r);
8686 temp_wont_dirty&=~(1<<r);
8687 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8688 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8689 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8690 } else {
8691 temp_will_dirty|=1<<r;
8692 temp_wont_dirty|=1<<r;
8693 }
8694 }
8695 }
8696 }
8697 }
8698 if(wr) {
8699 will_dirty[i]=temp_will_dirty;
8700 wont_dirty[i]=temp_wont_dirty;
8701 pass6_clean_registers((cinfo[i].ba-start)>>2,i-1,0);
8702 }else{
8703 // Limit recursion. It can take an excessive amount
8704 // of time if there are a lot of nested loops.
8705 will_dirty[(cinfo[i].ba-start)>>2]=0;
8706 wont_dirty[(cinfo[i].ba-start)>>2]=-1;
8707 }
8708 }
8709 /*else*/ if(1)
8710 {
8711 if (dops[i].is_ujump)
8712 {
8713 // Unconditional branch
8714 will_dirty_i=0;
8715 wont_dirty_i=0;
8716 //if(cinfo[i].ba>start+i*4) { // Disable recursion (for debugging)
8717 for(r=0;r<HOST_REGS;r++) {
8718 if(r!=EXCLUDE_REG) {
8719 if(branch_regs[i].regmap[r]==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8720 will_dirty_i|=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8721 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8722 }
8723 if(branch_regs[i].regmap[r]>=0) {
8724 will_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8725 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8726 }
8727 }
8728 }
8729 //}
8730 // Merge in delay slot
8731 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8732 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8733 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8734 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8735 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8736 will_dirty_i &= branch_hr_candirty;
8737 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8738 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8739 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8740 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8741 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8742 will_dirty_i &= hr_candirty;
8743 } else {
8744 // Conditional branch
8745 will_dirty_i=will_dirty_next;
8746 wont_dirty_i=wont_dirty_next;
8747 //if(cinfo[i].ba>start+i*4) // Disable recursion (for debugging)
8748 for(r=0;r<HOST_REGS;r++) {
8749 if(r!=EXCLUDE_REG) {
8750 signed char target_reg=branch_regs[i].regmap[r];
8751 if(target_reg==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8752 will_dirty_i&=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8753 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8754 }
8755 else if(target_reg>=0) {
8756 will_dirty_i&=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8757 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8758 }
8759 }
8760 }
8761 // Merge in delay slot
8762 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8763 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8764 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8765 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8766 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8767 will_dirty_i &= branch_hr_candirty;
8768 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8769 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8770 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8771 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8772 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8773 will_dirty_i &= hr_candirty;
8774 }
8775 // Merge in delay slot (won't dirty)
8776 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8777 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8778 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8779 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8780 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8781 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8782 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8783 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8784 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8785 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8786 wont_dirty_i &= ~(1u << 31);
8787 if(wr) {
8788 #ifndef DESTRUCTIVE_WRITEBACK
8789 branch_regs[i].dirty&=wont_dirty_i;
8790 #endif
8791 branch_regs[i].dirty|=will_dirty_i;
8792 }
8793 }
8794 }
8795 }
8796 else if (dops[i].is_exception)
8797 {
8798 // SYSCALL instruction, etc
8799 will_dirty_i=0;
8800 wont_dirty_i=0;
8801 }
8802 will_dirty_next=will_dirty_i;
8803 wont_dirty_next=wont_dirty_i;
8804 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8805 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8806 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8807 will_dirty_i &= hr_candirty;
8808 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8809 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8810 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8811 wont_dirty_i &= ~(1u << 31);
8812 if (i > istart && !dops[i].is_jump) {
8813 // Don't store a register immediately after writing it,
8814 // may prevent dual-issue.
8815 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8816 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8817 }
8818 // Save it
8819 will_dirty[i]=will_dirty_i;
8820 wont_dirty[i]=wont_dirty_i;
8821 // Mark registers that won't be dirtied as not dirty
8822 if(wr) {
8823 regs[i].dirty|=will_dirty_i;
8824 #ifndef DESTRUCTIVE_WRITEBACK
8825 regs[i].dirty&=wont_dirty_i;
8826 if(dops[i].is_jump)
8827 {
8828 if (i < iend-1 && !dops[i].is_ujump) {
8829 for(r=0;r<HOST_REGS;r++) {
8830 if(r!=EXCLUDE_REG) {
8831 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8832 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8833 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8834 }
8835 }
8836 }
8837 }
8838 else
8839 {
8840 if(i<iend) {
8841 for(r=0;r<HOST_REGS;r++) {
8842 if(r!=EXCLUDE_REG) {
8843 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8844 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8845 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8846 }
8847 }
8848 }
8849 }
8850 #endif
8851 }
8852 // Deal with changed mappings
8853 temp_will_dirty=will_dirty_i;
8854 temp_wont_dirty=wont_dirty_i;
8855 for(r=0;r<HOST_REGS;r++) {
8856 if(r!=EXCLUDE_REG) {
8857 int nr;
8858 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8859 if(wr) {
8860 #ifndef DESTRUCTIVE_WRITEBACK
8861 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8862 #endif
8863 regs[i].wasdirty|=will_dirty_i&(1<<r);
8864 }
8865 }
8866 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8867 // Register moved to a different register
8868 will_dirty_i&=~(1<<r);
8869 wont_dirty_i&=~(1<<r);
8870 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8871 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8872 if(wr) {
8873 #ifndef DESTRUCTIVE_WRITEBACK
8874 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8875 #endif
8876 regs[i].wasdirty|=will_dirty_i&(1<<r);
8877 }
8878 }
8879 else {
8880 will_dirty_i&=~(1<<r);
8881 wont_dirty_i&=~(1<<r);
8882 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8883 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8884 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8885 } else {
8886 wont_dirty_i|=1<<r;
8887 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8888 }
8889 }
8890 }
8891 }
8892 }
8893}
8894
8895static noinline void pass10_expire_blocks(void)
8896{
8897 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
8898 // not sizeof(ndrc->translation_cache) due to vita hack
8899 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
8900 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
8901 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
8902 int hit;
8903
8904 for (; expirep != end; expirep = ((expirep + step) & step_mask))
8905 {
8906 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
8907 u_int block_i = expirep / step & (PAGE_COUNT - 1);
8908 u_int phase = (expirep >> (base_shift - 1)) & 1u;
8909 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
8910 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
8911 (long)(out - ndrc->translation_cache), phase);
8912 }
8913
8914 if (!phase) {
8915 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
8916 if (hit) {
8917 do_clear_cache();
8918 #ifdef USE_MINI_HT
8919 memset(mini_ht, -1, sizeof(mini_ht));
8920 #endif
8921 }
8922 }
8923 else
8924 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
8925 }
8926}
8927
8928static struct block_info *new_block_info(u_int start, u_int len,
8929 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
8930{
8931 struct block_info **b_pptr;
8932 struct block_info *block;
8933 u_int page = get_page(start);
8934
8935 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
8936 assert(block);
8937 assert(jump_in_count > 0);
8938 block->source = source;
8939 block->copy = copy;
8940 block->start = start;
8941 block->len = len;
8942 block->reg_sv_flags = 0;
8943 block->tc_offs = beginning - ndrc->translation_cache;
8944 //block->tc_len = out - beginning;
8945 block->is_dirty = 0;
8946 block->inv_near_misses = 0;
8947 block->jump_in_cnt = jump_in_count;
8948
8949 // insert sorted by start mirror-unmasked vaddr
8950 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
8951 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
8952 block->next = *b_pptr;
8953 *b_pptr = block;
8954 break;
8955 }
8956 }
8957 stat_inc(stat_blocks);
8958 return block;
8959}
8960
8961static int new_recompile_block(u_int addr)
8962{
8963 u_int pagelimit = 0;
8964 u_int state_rflags = 0;
8965 int i;
8966
8967 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
8968
8969 if (addr & 3) {
8970 if (addr != hack_addr) {
8971 SysPrintf("game crash @%08x, ra=%08x\n", addr, psxRegs.GPR.n.ra);
8972 hack_addr = addr;
8973 }
8974 return -1;
8975 }
8976
8977 // this is just for speculation
8978 for (i = 1; i < 32; i++) {
8979 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
8980 state_rflags |= 1 << i;
8981 }
8982
8983 start = addr;
8984 new_dynarec_did_compile=1;
8985 if (Config.HLE && start == 0x80001000) // hlecall
8986 {
8987 // XXX: is this enough? Maybe check hleSoftCall?
8988 void *beginning = start_block();
8989
8990 emit_movimm(start,0);
8991 emit_writeword(0,&pcaddr);
8992 emit_far_jump(new_dyna_leave);
8993 literal_pool(0);
8994 end_block(beginning);
8995 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8996 block->jump_in[0].vaddr = start;
8997 block->jump_in[0].addr = beginning;
8998 return 0;
8999 }
9000 else if (f1_hack && hack_addr == 0) {
9001 void *beginning = start_block();
9002 emit_movimm(start, 0);
9003 emit_writeword(0, &hack_addr);
9004 emit_readword(&psxRegs.GPR.n.sp, 0);
9005 emit_readptr(&mem_rtab, 1);
9006 emit_shrimm(0, 12, 2);
9007 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
9008 emit_addimm(0, 0x18, 0);
9009 emit_adds_ptr(1, 1, 1);
9010 emit_ldr_dualindexed(1, 0, 0);
9011 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
9012 emit_far_call(ndrc_get_addr_ht);
9013 emit_jmpreg(0); // jr k0
9014 literal_pool(0);
9015 end_block(beginning);
9016
9017 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
9018 block->jump_in[0].vaddr = start;
9019 block->jump_in[0].addr = beginning;
9020 SysPrintf("F1 hack to %08x\n", start);
9021 return 0;
9022 }
9023
9024 cycle_multiplier_active = Config.cycle_multiplier_override && Config.cycle_multiplier == CYCLE_MULT_DEFAULT
9025 ? Config.cycle_multiplier_override : Config.cycle_multiplier;
9026
9027 source = get_source_start(start, &pagelimit);
9028 if (source == NULL) {
9029 if (addr != hack_addr) {
9030 SysPrintf("Compile at bogus memory address: %08x\n", addr);
9031 hack_addr = addr;
9032 }
9033 //abort();
9034 return -1;
9035 }
9036
9037 /* Pass 1: disassemble */
9038 /* Pass 2: register dependencies, branch targets */
9039 /* Pass 3: register allocation */
9040 /* Pass 4: branch dependencies */
9041 /* Pass 5: pre-alloc */
9042 /* Pass 6: optimize clean/dirty state */
9043 /* Pass 7: flag 32-bit registers */
9044 /* Pass 8: assembly */
9045 /* Pass 9: linker */
9046 /* Pass 10: garbage collection / free memory */
9047
9048 /* Pass 1 disassembly */
9049
9050 pass1_disassemble(pagelimit);
9051
9052 int clear_hack_addr = apply_hacks();
9053
9054 /* Pass 2 - Register dependencies and branch targets */
9055
9056 pass2_unneeded_regs(0,slen-1,0);
9057
9058 /* Pass 3 - Register allocation */
9059
9060 pass3_register_alloc(addr);
9061
9062 /* Pass 4 - Cull unused host registers */
9063
9064 pass4_cull_unused_regs();
9065
9066 /* Pass 5 - Pre-allocate registers */
9067
9068 pass5a_preallocate1();
9069 pass5b_preallocate2();
9070
9071 /* Pass 6 - Optimize clean/dirty state */
9072 pass6_clean_registers(0, slen-1, 1);
9073
9074 /* Pass 7 - Identify 32-bit registers */
9075 for (i=slen-1;i>=0;i--)
9076 {
9077 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9078 {
9079 // Conditional branch
9080 if((source[i]>>16)!=0x1000&&i<slen-2) {
9081 // Mark this address as a branch target since it may be called
9082 // upon return from interrupt
9083 dops[i+2].bt=1;
9084 }
9085 }
9086 }
9087
9088 /* Pass 8 - Assembly */
9089 linkcount=0;stubcount=0;
9090 is_delayslot=0;
9091 u_int dirty_pre=0;
9092 void *beginning=start_block();
9093 void *instr_addr0_override = NULL;
9094 int ds = 0;
9095
9096 if (start == 0x80030000) {
9097 // nasty hack for the fastbios thing
9098 // override block entry to this code
9099 instr_addr0_override = out;
9100 emit_movimm(start,0);
9101 // abuse io address var as a flag that we
9102 // have already returned here once
9103 emit_readword(&address,1);
9104 emit_writeword(0,&pcaddr);
9105 emit_writeword(0,&address);
9106 emit_cmp(0,1);
9107 #ifdef __aarch64__
9108 emit_jeq(out + 4*2);
9109 emit_far_jump(new_dyna_leave);
9110 #else
9111 emit_jne(new_dyna_leave);
9112 #endif
9113 }
9114 for(i=0;i<slen;i++)
9115 {
9116 __builtin_prefetch(regs[i+1].regmap);
9117 check_regmap(regmap_pre[i]);
9118 check_regmap(regs[i].regmap_entry);
9119 check_regmap(regs[i].regmap);
9120 //if(ds) printf("ds: ");
9121 disassemble_inst(i);
9122 if(ds) {
9123 ds=0; // Skip delay slot
9124 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9125 instr_addr[i] = NULL;
9126 } else {
9127 speculate_register_values(i);
9128 #ifndef DESTRUCTIVE_WRITEBACK
9129 if (i < 2 || !dops[i-2].is_ujump)
9130 {
9131 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9132 }
9133 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9134 dirty_pre=branch_regs[i].dirty;
9135 }else{
9136 dirty_pre=regs[i].dirty;
9137 }
9138 #endif
9139 // write back
9140 if (i < 2 || !dops[i-2].is_ujump)
9141 {
9142 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9143 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9144 }
9145 // branch target entry point
9146 instr_addr[i] = out;
9147 assem_debug("<->\n");
9148 drc_dbg_emit_do_cmp(i, cinfo[i].ccadj);
9149 if (clear_hack_addr) {
9150 emit_movimm(0, 0);
9151 emit_writeword(0, &hack_addr);
9152 clear_hack_addr = 0;
9153 }
9154
9155 // load regs
9156 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9157 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9158 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9159 address_generation(i,&regs[i],regs[i].regmap_entry);
9160 load_consts(regmap_pre[i],regs[i].regmap,i);
9161 if(dops[i].is_jump)
9162 {
9163 // Load the delay slot registers if necessary
9164 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))
9165 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9166 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))
9167 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9168 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9169 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9170 if (dops[i+1].is_store)
9171 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9172 }
9173 else if(i+1<slen)
9174 {
9175 // Preload registers for following instruction
9176 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9177 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9178 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9179 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9180 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9181 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9182 }
9183 // TODO: if(is_ooo(i)) address_generation(i+1);
9184 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9185 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9186 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9187 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9188 if (dops[i].is_store)
9189 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9190
9191 ds = assemble(i, &regs[i], cinfo[i].ccadj);
9192
9193 if (dops[i].is_ujump)
9194 literal_pool(1024);
9195 else
9196 literal_pool_jumpover(256);
9197 }
9198 }
9199
9200 assert(slen > 0);
9201 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9202 // no ending needed for this block since INTCALL never returns
9203 }
9204 // If the block did not end with an unconditional branch,
9205 // add a jump to the next instruction.
9206 else if (i > 1) {
9207 if (!dops[i-2].is_ujump) {
9208 assert(!dops[i-1].is_jump);
9209 assert(i==slen);
9210 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9211 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9212 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9213 emit_loadreg(CCREG,HOST_CCREG);
9214 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9215 }
9216 else
9217 {
9218 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9219 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9220 }
9221 add_to_linker(out,start+i*4,0);
9222 emit_jmp(0);
9223 }
9224 }
9225 else
9226 {
9227 assert(i>0);
9228 assert(!dops[i-1].is_jump);
9229 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9230 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9231 emit_loadreg(CCREG,HOST_CCREG);
9232 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9233 add_to_linker(out,start+i*4,0);
9234 emit_jmp(0);
9235 }
9236
9237 // Stubs
9238 for(i = 0; i < stubcount; i++)
9239 {
9240 switch(stubs[i].type)
9241 {
9242 case LOADB_STUB:
9243 case LOADH_STUB:
9244 case LOADW_STUB:
9245 case LOADBU_STUB:
9246 case LOADHU_STUB:
9247 do_readstub(i);break;
9248 case STOREB_STUB:
9249 case STOREH_STUB:
9250 case STOREW_STUB:
9251 do_writestub(i);break;
9252 case CC_STUB:
9253 do_ccstub(i);break;
9254 case INVCODE_STUB:
9255 do_invstub(i);break;
9256 case STORELR_STUB:
9257 do_unalignedwritestub(i);break;
9258 case OVERFLOW_STUB:
9259 do_overflowstub(i); break;
9260 case ALIGNMENT_STUB:
9261 do_alignmentstub(i); break;
9262 default:
9263 assert(0);
9264 }
9265 }
9266
9267 if (instr_addr0_override)
9268 instr_addr[0] = instr_addr0_override;
9269
9270#if 0
9271 /* check for improper expiration */
9272 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9273 int j;
9274 if (!jumps[i])
9275 continue;
9276 for (j = 0; j < jumps[i]->count; j++)
9277 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9278 }
9279#endif
9280
9281 /* Pass 9 - Linker */
9282 for(i=0;i<linkcount;i++)
9283 {
9284 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9285 literal_pool(64);
9286 if (!link_addr[i].internal)
9287 {
9288 void *stub = out;
9289 void *addr = check_addr(link_addr[i].target);
9290 emit_extjump(link_addr[i].addr, link_addr[i].target);
9291 if (addr) {
9292 set_jump_target(link_addr[i].addr, addr);
9293 ndrc_add_jump_out(link_addr[i].target,stub);
9294 }
9295 else
9296 set_jump_target(link_addr[i].addr, stub);
9297 }
9298 else
9299 {
9300 // Internal branch
9301 int target=(link_addr[i].target-start)>>2;
9302 assert(target>=0&&target<slen);
9303 assert(instr_addr[target]);
9304 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9305 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9306 //#else
9307 set_jump_target(link_addr[i].addr, instr_addr[target]);
9308 //#endif
9309 }
9310 }
9311
9312 u_int source_len = slen*4;
9313 if (dops[slen-1].itype == INTCALL && source_len > 4)
9314 // no need to treat the last instruction as compiled
9315 // as interpreter fully handles it
9316 source_len -= 4;
9317
9318 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9319 copy = shadow;
9320
9321 // External Branch Targets (jump_in)
9322 int jump_in_count = 1;
9323 assert(instr_addr[0]);
9324 for (i = 1; i < slen; i++)
9325 {
9326 if (dops[i].bt && instr_addr[i])
9327 jump_in_count++;
9328 }
9329
9330 struct block_info *block =
9331 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9332 block->reg_sv_flags = state_rflags;
9333
9334 int jump_in_i = 0;
9335 for (i = 0; i < slen; i++)
9336 {
9337 if ((i == 0 || dops[i].bt) && instr_addr[i])
9338 {
9339 assem_debug("%p (%d) <- %8x\n", instr_addr[i], i, start + i*4);
9340 u_int vaddr = start + i*4;
9341
9342 literal_pool(256);
9343 void *entry = out;
9344 load_regs_entry(i);
9345 if (entry == out)
9346 entry = instr_addr[i];
9347 else
9348 emit_jmp(instr_addr[i]);
9349
9350 block->jump_in[jump_in_i].vaddr = vaddr;
9351 block->jump_in[jump_in_i].addr = entry;
9352 jump_in_i++;
9353 }
9354 }
9355 assert(jump_in_i == jump_in_count);
9356 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9357 // Write out the literal pool if necessary
9358 literal_pool(0);
9359 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9360 // Align code
9361 if(((u_int)out)&7) emit_addnop(13);
9362 #endif
9363 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9364 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9365 memcpy(copy, source, source_len);
9366 copy += source_len;
9367
9368 end_block(beginning);
9369
9370 // If we're within 256K of the end of the buffer,
9371 // start over from the beginning. (Is 256K enough?)
9372 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9373 out = ndrc->translation_cache;
9374
9375 // Trap writes to any of the pages we compiled
9376 mark_invalid_code(start, slen*4, 0);
9377
9378 /* Pass 10 - Free memory by expiring oldest blocks */
9379
9380 pass10_expire_blocks();
9381
9382#ifdef ASSEM_PRINT
9383 fflush(stdout);
9384#endif
9385 stat_inc(stat_bc_direct);
9386 return 0;
9387}
9388
9389// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork