notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
drc: adjust MAXBLOCK
[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 2048
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 u_int wasdirty;
162 u_int dirty;
163 u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
164 u_int isconst; // ... but isconst is false when r2 is known (hr)
165 u_int loadedconst; // host regs that have constants loaded
166 u_int noevict; // can't evict this hr (alloced by current op)
167 //u_int waswritten; // MIPS regs that were used as store base before
168 uint64_t u;
169};
170
171struct ht_entry
172{
173 u_int vaddr[2];
174 void *tcaddr[2];
175};
176
177struct code_stub
178{
179 enum stub_type type;
180 void *addr;
181 void *retaddr;
182 u_int a;
183 uintptr_t b;
184 uintptr_t c;
185 u_int d;
186 u_int e;
187};
188
189struct link_entry
190{
191 void *addr;
192 u_int target;
193 u_int internal;
194};
195
196struct block_info
197{
198 struct block_info *next;
199 const void *source;
200 const void *copy;
201 u_int start; // vaddr of the block start
202 u_int len; // of the whole block source
203 u_int tc_offs;
204 //u_int tc_len;
205 u_int reg_sv_flags;
206 u_char is_dirty;
207 u_char inv_near_misses;
208 u_short jump_in_cnt;
209 struct {
210 u_int vaddr;
211 void *addr;
212 } jump_in[0];
213};
214
215struct jump_info
216{
217 int alloc;
218 int count;
219 struct {
220 u_int target_vaddr;
221 void *stub;
222 } e[0];
223};
224
225static struct decoded_insn
226{
227 u_char itype;
228 u_char opcode; // bits 31-26
229 u_char opcode2; // (depends on opcode)
230 u_char rs1;
231 u_char rs2;
232 u_char rt1;
233 u_char rt2;
234 u_char use_lt1:1;
235 u_char bt:1;
236 u_char ooo:1;
237 u_char is_ds:1;
238 u_char is_jump:1;
239 u_char is_ujump:1;
240 u_char is_load:1;
241 u_char is_store:1;
242 u_char is_delay_load:1; // is_load + MFC/CFC
243 u_char is_exception:1; // unconditional, also interp. fallback
244 u_char may_except:1; // might generate an exception
245} dops[MAXBLOCK];
246
247static struct compile_info
248{
249 int imm;
250 u_int ba;
251 int ccadj;
252 signed char min_free_regs;
253 signed char addr;
254 signed char reserved[2];
255} cinfo[MAXBLOCK];
256
257 static u_char *out;
258 static char invalid_code[0x100000];
259 static struct ht_entry hash_table[65536];
260 static struct block_info *blocks[PAGE_COUNT];
261 static struct jump_info *jumps[PAGE_COUNT];
262 static u_int start;
263 static u_int *source;
264 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
265 static uint64_t gte_rt[MAXBLOCK];
266 static uint64_t gte_unneeded[MAXBLOCK];
267 static u_int smrv[32]; // speculated MIPS register values
268 static u_int smrv_strong; // mask or regs that are likely to have correct values
269 static u_int smrv_weak; // same, but somewhat less likely
270 static u_int smrv_strong_next; // same, but after current insn executes
271 static u_int smrv_weak_next;
272 static uint64_t unneeded_reg[MAXBLOCK];
273 static uint64_t branch_unneeded_reg[MAXBLOCK];
274 // see 'struct regstat' for a description
275 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
276 // contains 'real' consts at [i] insn, but may differ from what's actually
277 // loaded in host reg as 'final' value is always loaded, see get_final_value()
278 static uint32_t current_constmap[HOST_REGS];
279 static uint32_t constmap[MAXBLOCK][HOST_REGS];
280 static struct regstat regs[MAXBLOCK];
281 static struct regstat branch_regs[MAXBLOCK];
282 static int slen;
283 static void *instr_addr[MAXBLOCK];
284 static struct link_entry link_addr[MAXBLOCK];
285 static int linkcount;
286 static struct code_stub stubs[MAXBLOCK*3];
287 static int stubcount;
288 static u_int literals[1024][2];
289 static int literalcount;
290 static int is_delayslot;
291 static char shadow[1048576] __attribute__((aligned(16)));
292 static void *copy;
293 static u_int expirep;
294 static u_int stop_after_jal;
295 static u_int f1_hack;
296#ifdef STAT_PRINT
297 static int stat_bc_direct;
298 static int stat_bc_pre;
299 static int stat_bc_restore;
300 static int stat_ht_lookups;
301 static int stat_jump_in_lookups;
302 static int stat_restore_tries;
303 static int stat_restore_compares;
304 static int stat_inv_addr_calls;
305 static int stat_inv_hits;
306 static int stat_blocks;
307 static int stat_links;
308 #define stat_inc(s) s++
309 #define stat_dec(s) s--
310 #define stat_clear(s) s = 0
311#else
312 #define stat_inc(s)
313 #define stat_dec(s)
314 #define stat_clear(s)
315#endif
316
317 int new_dynarec_hacks;
318 int new_dynarec_hacks_pergame;
319 int new_dynarec_hacks_old;
320 int new_dynarec_did_compile;
321
322 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
323
324 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0 (CCREG)
325 extern int last_count; // last absolute target, often = next_interupt
326 extern int pcaddr;
327 extern int pending_exception;
328 extern int branch_target;
329 extern uintptr_t ram_offset;
330 extern uintptr_t mini_ht[32][2];
331
332 /* registers that may be allocated */
333 /* 1-31 gpr */
334#define LOREG 32 // lo
335#define HIREG 33 // hi
336//#define FSREG 34 // FPU status (FCSR)
337//#define CSREG 35 // Coprocessor status
338#define CCREG 36 // Cycle count
339#define INVCP 37 // Pointer to invalid_code
340//#define MMREG 38 // Pointer to memory_map
341#define ROREG 39 // ram offset (if psxM != 0x80000000)
342#define TEMPREG 40
343#define FTEMP 40 // Load/store temporary register (was fpu)
344#define PTEMP 41 // Prefetch temporary register
345//#define TLREG 42 // TLB mapping offset
346#define RHASH 43 // Return address hash
347#define RHTBL 44 // Return address hash table address
348#define RTEMP 45 // JR/JALR address register
349#define MAXREG 45
350#define AGEN1 46 // Address generation temporary register (pass5b_preallocate2)
351//#define AGEN2 47 // Address generation temporary register
352
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
383#define DJT_1 (void *)1l // no function, just a label in assem_debug log
384#define DJT_2 (void *)2l
385
386// asm linkage
387void dyna_linker();
388void cc_interrupt();
389void jump_syscall (u_int u0, u_int u1, u_int pc);
390void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
391void jump_break (u_int u0, u_int u1, u_int pc);
392void jump_break_ds(u_int u0, u_int u1, u_int pc);
393void jump_overflow (u_int u0, u_int u1, u_int pc);
394void jump_overflow_ds(u_int u0, u_int u1, u_int pc);
395void jump_addrerror (u_int cause, u_int addr, u_int pc);
396void jump_addrerror_ds(u_int cause, u_int addr, u_int pc);
397void jump_to_new_pc();
398void call_gteStall();
399void new_dyna_leave();
400
401void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile);
402void *ndrc_get_addr_ht(u_int vaddr);
403void ndrc_add_jump_out(u_int vaddr, void *src);
404void ndrc_write_invalidate_one(u_int addr);
405static void ndrc_write_invalidate_many(u_int addr, u_int end);
406
407static int new_recompile_block(u_int addr);
408static void invalidate_block(struct block_info *block);
409static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_);
410
411// Needed by assembler
412static void wb_register(signed char r, const signed char regmap[], u_int dirty);
413static void wb_dirtys(const signed char i_regmap[], u_int i_dirty);
414static void wb_needed_dirtys(const signed char i_regmap[], u_int i_dirty, int addr);
415static void load_all_regs(const signed char i_regmap[]);
416static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
417static void load_regs_entry(int t);
418static void load_all_consts(const signed char regmap[], u_int dirty, int i);
419static u_int get_host_reglist(const signed char *regmap);
420
421static int get_final_value(int hr, int i, u_int *value);
422static void add_stub(enum stub_type type, void *addr, void *retaddr,
423 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
424static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
425 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
426static void add_to_linker(void *addr, u_int target, int ext);
427static void *get_direct_memhandler(void *table, u_int addr,
428 enum stub_type type, uintptr_t *addr_host);
429static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
430static void pass_args(int a0, int a1);
431static void emit_far_jump(const void *f);
432static void emit_far_call(const void *f);
433
434#ifdef VITA
435#include <psp2/kernel/sysmem.h>
436static int sceBlock;
437// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
438extern int getVMBlock();
439int _newlib_vm_size_user = sizeof(*ndrc);
440#endif
441
442static void mprotect_w_x(void *start, void *end, int is_x)
443{
444#ifdef NO_WRITE_EXEC
445 #if defined(VITA)
446 // *Open* enables write on all memory that was
447 // allocated by sceKernelAllocMemBlockForVM()?
448 if (is_x)
449 sceKernelCloseVMDomain();
450 else
451 sceKernelOpenVMDomain();
452 #elif defined(HAVE_LIBNX)
453 Result rc;
454 // check to avoid the full flush in jitTransitionToExecutable()
455 if (g_jit.type != JitType_CodeMemory) {
456 if (is_x)
457 rc = jitTransitionToExecutable(&g_jit);
458 else
459 rc = jitTransitionToWritable(&g_jit);
460 if (R_FAILED(rc))
461 ;//SysPrintf("jitTransition %d %08x\n", is_x, rc);
462 }
463 #elif defined(TC_WRITE_OFFSET)
464 // separated rx and rw areas are always available
465 #else
466 u_long mstart = (u_long)start & ~4095ul;
467 u_long mend = (u_long)end;
468 if (mprotect((void *)mstart, mend - mstart,
469 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
470 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
471 #endif
472#endif
473}
474
475static void start_tcache_write(void *start, void *end)
476{
477 mprotect_w_x(start, end, 0);
478}
479
480static void end_tcache_write(void *start, void *end)
481{
482#if defined(__arm__) || defined(__aarch64__)
483 size_t len = (char *)end - (char *)start;
484 #if defined(__BLACKBERRY_QNX__)
485 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
486 #elif defined(__MACH__)
487 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
488 #elif defined(VITA)
489 sceKernelSyncVMDomain(sceBlock, start, len);
490 #elif defined(_3DS)
491 ctr_flush_invalidate_cache();
492 #elif defined(HAVE_LIBNX)
493 if (g_jit.type == JitType_CodeMemory) {
494 armDCacheClean(start, len);
495 armICacheInvalidate((char *)start - ndrc_write_ofs, len);
496 // as of v4.2.1 libnx lacks isb
497 __asm__ volatile("isb" ::: "memory");
498 }
499 #elif defined(__aarch64__)
500 // as of 2021, __clear_cache() is still broken on arm64
501 // so here is a custom one :(
502 clear_cache_arm64(start, end);
503 #else
504 __clear_cache(start, end);
505 #endif
506 (void)len;
507#endif
508
509 mprotect_w_x(start, end, 1);
510}
511
512static void *start_block(void)
513{
514 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
515 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
516 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
517 start_tcache_write(NDRC_WRITE_OFFSET(out), NDRC_WRITE_OFFSET(end));
518 return out;
519}
520
521static void end_block(void *start)
522{
523 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(out));
524}
525
526#ifdef NDRC_CACHE_FLUSH_ALL
527
528static int needs_clear_cache;
529
530static void mark_clear_cache(void *target)
531{
532 if (!needs_clear_cache) {
533 start_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
534 needs_clear_cache = 1;
535 }
536}
537
538static void do_clear_cache(void)
539{
540 if (needs_clear_cache) {
541 end_tcache_write(NDRC_WRITE_OFFSET(ndrc), NDRC_WRITE_OFFSET(ndrc + 1));
542 needs_clear_cache = 0;
543 }
544}
545
546#else
547
548// also takes care of w^x mappings when patching code
549static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
550
551static void mark_clear_cache(void *target)
552{
553 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
554 u_int mask = 1u << ((offset >> 12) & 31);
555 if (!(needs_clear_cache[offset >> 17] & mask)) {
556 char *start = (char *)NDRC_WRITE_OFFSET((uintptr_t)target & ~4095l);
557 start_tcache_write(start, start + 4095);
558 needs_clear_cache[offset >> 17] |= mask;
559 }
560}
561
562// Clearing the cache is rather slow on ARM Linux, so mark the areas
563// that need to be cleared, and then only clear these areas once.
564static void do_clear_cache(void)
565{
566 int i, j;
567 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
568 {
569 u_int bitmap = needs_clear_cache[i];
570 if (!bitmap)
571 continue;
572 for (j = 0; j < 32; j++)
573 {
574 u_char *start, *end;
575 if (!(bitmap & (1u << j)))
576 continue;
577
578 start = ndrc->translation_cache + i*131072 + j*4096;
579 end = start + 4095;
580 for (j++; j < 32; j++) {
581 if (!(bitmap & (1u << j)))
582 break;
583 end += 4096;
584 }
585 end_tcache_write(NDRC_WRITE_OFFSET(start), NDRC_WRITE_OFFSET(end));
586 }
587 needs_clear_cache[i] = 0;
588 }
589}
590
591#endif // NDRC_CACHE_FLUSH_ALL
592
593#define NO_CYCLE_PENALTY_THR 12
594
595int cycle_multiplier_old;
596static int cycle_multiplier_active;
597
598static int CLOCK_ADJUST(int x)
599{
600 int m = cycle_multiplier_active;
601 int s = (x >> 31) | 1;
602 return (x * m + s * 50) / 100;
603}
604
605static int ds_writes_rjump_rs(int i)
606{
607 return dops[i].rs1 != 0
608 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2
609 || dops[i].rs1 == dops[i].rt1); // overwrites itself - same effect
610}
611
612// psx addr mirror masking (for invalidation)
613static u_int pmmask(u_int vaddr)
614{
615 vaddr &= ~0xe0000000;
616 if (vaddr < 0x01000000)
617 vaddr &= ~0x00e00000; // RAM mirrors
618 return vaddr;
619}
620
621static u_int get_page(u_int vaddr)
622{
623 u_int page = pmmask(vaddr) >> 12;
624 if (page >= PAGE_COUNT / 2)
625 page = PAGE_COUNT / 2 + (page & (PAGE_COUNT / 2 - 1));
626 return page;
627}
628
629// get a page for looking for a block that has vaddr
630// (needed because the block may start in previous page)
631static u_int get_page_prev(u_int vaddr)
632{
633 assert(MAXBLOCK <= (1 << 12));
634 u_int page = get_page(vaddr);
635 if (page & 511)
636 page--;
637 return page;
638}
639
640static struct ht_entry *hash_table_get(u_int vaddr)
641{
642 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
643}
644
645static void hash_table_add(u_int vaddr, void *tcaddr)
646{
647 struct ht_entry *ht_bin = hash_table_get(vaddr);
648 assert(tcaddr);
649 ht_bin->vaddr[1] = ht_bin->vaddr[0];
650 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
651 ht_bin->vaddr[0] = vaddr;
652 ht_bin->tcaddr[0] = tcaddr;
653}
654
655static void hash_table_remove(int vaddr)
656{
657 //printf("remove hash: %x\n",vaddr);
658 struct ht_entry *ht_bin = hash_table_get(vaddr);
659 if (ht_bin->vaddr[1] == vaddr) {
660 ht_bin->vaddr[1] = -1;
661 ht_bin->tcaddr[1] = NULL;
662 }
663 if (ht_bin->vaddr[0] == vaddr) {
664 ht_bin->vaddr[0] = ht_bin->vaddr[1];
665 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
666 ht_bin->vaddr[1] = -1;
667 ht_bin->tcaddr[1] = NULL;
668 }
669}
670
671static void mark_invalid_code(u_int vaddr, u_int len, char invalid)
672{
673 u_int vaddr_m = vaddr & 0x1fffffff;
674 u_int i, j;
675 for (i = vaddr_m & ~0xfff; i < vaddr_m + len; i += 0x1000) {
676 // ram mirrors, but should not hurt bios
677 for (j = 0; j < 0x800000; j += 0x200000) {
678 invalid_code[(i|j) >> 12] =
679 invalid_code[(i|j|0x80000000u) >> 12] =
680 invalid_code[(i|j|0xa0000000u) >> 12] = invalid;
681 }
682 }
683 if (!invalid && vaddr + len > inv_code_start && vaddr <= inv_code_end)
684 inv_code_start = inv_code_end = ~0;
685}
686
687static int doesnt_expire_soon(u_char *tcaddr)
688{
689 u_int diff = (u_int)(tcaddr - out) & ((1u << TARGET_SIZE_2) - 1u);
690 return diff > EXPIRITY_OFFSET + MAX_OUTPUT_BLOCK_SIZE;
691}
692
693static unused void check_for_block_changes(u_int start, u_int end)
694{
695 u_int start_page = get_page_prev(start);
696 u_int end_page = get_page(end - 1);
697 u_int page;
698
699 for (page = start_page; page <= end_page; page++) {
700 struct block_info *block;
701 for (block = blocks[page]; block != NULL; block = block->next) {
702 if (block->is_dirty)
703 continue;
704 if (memcmp(block->source, block->copy, block->len)) {
705 printf("bad block %08x-%08x %016llx %016llx @%08x\n",
706 block->start, block->start + block->len,
707 *(long long *)block->source, *(long long *)block->copy, psxRegs.pc);
708 fflush(stdout);
709 abort();
710 }
711 }
712 }
713}
714
715static void *try_restore_block(u_int vaddr, u_int start_page, u_int end_page)
716{
717 void *found_clean = NULL;
718 u_int i, page;
719
720 stat_inc(stat_restore_tries);
721 for (page = start_page; page <= end_page; page++) {
722 struct block_info *block;
723 for (block = blocks[page]; block != NULL; block = block->next) {
724 if (vaddr < block->start)
725 break;
726 if (!block->is_dirty || vaddr >= block->start + block->len)
727 continue;
728 for (i = 0; i < block->jump_in_cnt; i++)
729 if (block->jump_in[i].vaddr == vaddr)
730 break;
731 if (i == block->jump_in_cnt)
732 continue;
733 assert(block->source && block->copy);
734 stat_inc(stat_restore_compares);
735 if (memcmp(block->source, block->copy, block->len))
736 continue;
737
738 block->is_dirty = block->inv_near_misses = 0;
739 found_clean = block->jump_in[i].addr;
740 hash_table_add(vaddr, found_clean);
741 mark_invalid_code(block->start, block->len, 0);
742 stat_inc(stat_bc_restore);
743 inv_debug("INV: restored %08x %p (%d)\n", vaddr, found_clean, block->jump_in_cnt);
744 return found_clean;
745 }
746 }
747 return NULL;
748}
749
750// this doesn't normally happen
751static noinline u_int generate_exception(u_int pc)
752{
753 //if (execBreakCheck(&psxRegs, pc))
754 // return psxRegs.pc;
755
756 // generate an address or bus error
757 psxRegs.CP0.n.Cause &= 0x300;
758 psxRegs.CP0.n.EPC = pc;
759 if (pc & 3) {
760 psxRegs.CP0.n.Cause |= R3000E_AdEL << 2;
761 psxRegs.CP0.n.BadVAddr = pc;
762#ifdef DRC_DBG
763 last_count -= 2;
764#endif
765 } else
766 psxRegs.CP0.n.Cause |= R3000E_IBE << 2;
767 return (psxRegs.pc = 0x80000080);
768}
769
770// Get address from virtual address
771// This is called from the recompiled JR/JALR instructions
772static void noinline *get_addr(u_int vaddr, int can_compile)
773{
774 u_int start_page = get_page_prev(vaddr);
775 u_int i, page, end_page = get_page(vaddr);
776 void *found_clean = NULL;
777
778 stat_inc(stat_jump_in_lookups);
779 for (page = start_page; page <= end_page; page++) {
780 const struct block_info *block;
781 for (block = blocks[page]; block != NULL; block = block->next) {
782 if (vaddr < block->start)
783 break;
784 if (block->is_dirty || vaddr >= block->start + block->len)
785 continue;
786 for (i = 0; i < block->jump_in_cnt; i++)
787 if (block->jump_in[i].vaddr == vaddr)
788 break;
789 if (i == block->jump_in_cnt)
790 continue;
791 found_clean = block->jump_in[i].addr;
792 hash_table_add(vaddr, found_clean);
793 return found_clean;
794 }
795 }
796 found_clean = try_restore_block(vaddr, start_page, end_page);
797 if (found_clean)
798 return found_clean;
799
800 if (!can_compile)
801 return NULL;
802
803 int r = new_recompile_block(vaddr);
804 if (likely(r == 0))
805 return ndrc_get_addr_ht(vaddr);
806
807 return ndrc_get_addr_ht(generate_exception(vaddr));
808}
809
810// Look up address in hash table first
811void *ndrc_get_addr_ht_param(u_int vaddr, int can_compile)
812{
813 //check_for_block_changes(vaddr, vaddr + MAXBLOCK);
814 const struct ht_entry *ht_bin = hash_table_get(vaddr);
815 u_int vaddr_a = vaddr & ~3;
816 stat_inc(stat_ht_lookups);
817 if (ht_bin->vaddr[0] == vaddr_a) return ht_bin->tcaddr[0];
818 if (ht_bin->vaddr[1] == vaddr_a) return ht_bin->tcaddr[1];
819 return get_addr(vaddr, can_compile);
820}
821
822void *ndrc_get_addr_ht(u_int vaddr)
823{
824 return ndrc_get_addr_ht_param(vaddr, 1);
825}
826
827static void clear_all_regs(signed char regmap[])
828{
829 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
830}
831
832// get_reg: get allocated host reg from mips reg
833// returns -1 if no such mips reg was allocated
834#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
835
836extern signed char get_reg(const signed char regmap[], signed char r);
837
838#else
839
840static signed char get_reg(const signed char regmap[], signed char r)
841{
842 int hr;
843 for (hr = 0; hr < HOST_REGS; hr++) {
844 if (hr == EXCLUDE_REG)
845 continue;
846 if (regmap[hr] == r)
847 return hr;
848 }
849 return -1;
850}
851
852#endif
853
854// get reg suitable for writing
855static signed char get_reg_w(const signed char regmap[], signed char r)
856{
857 return r == 0 ? -1 : get_reg(regmap, r);
858}
859
860// get reg as mask bit (1 << hr)
861static u_int get_regm(const signed char regmap[], signed char r)
862{
863 return (1u << (get_reg(regmap, r) & 31)) & ~(1u << 31);
864}
865
866static signed char get_reg_temp(const signed char regmap[])
867{
868 int hr;
869 for (hr = 0; hr < HOST_REGS; hr++) {
870 if (hr == EXCLUDE_REG)
871 continue;
872 if (regmap[hr] == (signed char)-1)
873 return hr;
874 }
875 return -1;
876}
877
878// Find a register that is available for two consecutive cycles
879static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
880{
881 int hr;
882 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
883 return -1;
884}
885
886// reverse reg map: mips -> host
887#define RRMAP_SIZE 64
888static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
889 u_int *regs_can_change)
890{
891 u_int r, hr, hr_can_change = 0;
892 memset(rrmap, -1, RRMAP_SIZE);
893 for (hr = 0; hr < HOST_REGS; )
894 {
895 r = regmap[hr];
896 rrmap[r & (RRMAP_SIZE - 1)] = hr;
897 // only add mips $1-$31+$lo, others shifted out
898 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
899 hr++;
900 if (hr == EXCLUDE_REG)
901 hr++;
902 }
903 hr_can_change |= 1u << (rrmap[33] & 31);
904 hr_can_change |= 1u << (rrmap[CCREG] & 31);
905 hr_can_change &= ~(1u << 31);
906 *regs_can_change = hr_can_change;
907}
908
909// same as get_reg, but takes rrmap
910static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
911{
912 assert(0 <= r && r < RRMAP_SIZE);
913 return rrmap[r];
914}
915
916static int count_free_regs(const signed char regmap[])
917{
918 int count=0;
919 int hr;
920 for(hr=0;hr<HOST_REGS;hr++)
921 {
922 if(hr!=EXCLUDE_REG) {
923 if(regmap[hr]<0) count++;
924 }
925 }
926 return count;
927}
928
929static void dirty_reg(struct regstat *cur, signed char reg)
930{
931 int hr;
932 if (!reg) return;
933 hr = get_reg(cur->regmap, reg);
934 if (hr >= 0)
935 cur->dirty |= 1<<hr;
936}
937
938static void set_const(struct regstat *cur, signed char reg, uint32_t value)
939{
940 int hr;
941 if (!reg) return;
942 hr = get_reg(cur->regmap, reg);
943 if (hr >= 0) {
944 cur->isconst |= 1<<hr;
945 current_constmap[hr] = value;
946 }
947}
948
949static void clear_const(struct regstat *cur, signed char reg)
950{
951 int hr;
952 if (!reg) return;
953 hr = get_reg(cur->regmap, reg);
954 if (hr >= 0)
955 cur->isconst &= ~(1<<hr);
956}
957
958static int is_const(const struct regstat *cur, signed char reg)
959{
960 int hr;
961 if (reg < 0) return 0;
962 if (!reg) return 1;
963 hr = get_reg(cur->regmap, reg);
964 if (hr >= 0)
965 return (cur->isconst>>hr)&1;
966 return 0;
967}
968
969static uint32_t get_const(const struct regstat *cur, signed char reg)
970{
971 int hr;
972 if (!reg) return 0;
973 hr = get_reg(cur->regmap, reg);
974 if (hr >= 0)
975 return current_constmap[hr];
976
977 SysPrintf("Unknown constant in r%d\n", reg);
978 abort();
979}
980
981// Least soon needed registers
982// Look at the next ten instructions and see which registers
983// will be used. Try not to reallocate these.
984static void lsn(u_char hsn[], int i)
985{
986 int j;
987 int b=-1;
988 for(j=0;j<9;j++)
989 {
990 if(i+j>=slen) {
991 j=slen-i-1;
992 break;
993 }
994 if (dops[i+j].is_ujump)
995 {
996 // Don't go past an unconditonal jump
997 j++;
998 break;
999 }
1000 }
1001 for(;j>=0;j--)
1002 {
1003 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
1004 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
1005 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
1006 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
1007 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
1008 // Stores can allocate zero
1009 hsn[dops[i+j].rs1]=j;
1010 hsn[dops[i+j].rs2]=j;
1011 }
1012 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
1013 hsn[ROREG] = j;
1014 // On some architectures stores need invc_ptr
1015 #if defined(HOST_IMM8)
1016 if (dops[i+j].is_store)
1017 hsn[INVCP] = j;
1018 #endif
1019 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1020 {
1021 hsn[CCREG]=j;
1022 b=j;
1023 }
1024 }
1025 if(b>=0)
1026 {
1027 if(cinfo[i+b].ba>=start && cinfo[i+b].ba<(start+slen*4))
1028 {
1029 // Follow first branch
1030 int t=(cinfo[i+b].ba-start)>>2;
1031 j=7-b;if(t+j>=slen) j=slen-t-1;
1032 for(;j>=0;j--)
1033 {
1034 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
1035 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
1036 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
1037 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
1038 }
1039 }
1040 // TODO: preferred register based on backward branch
1041 }
1042 // Delay slot should preferably not overwrite branch conditions or cycle count
1043 if (i > 0 && dops[i-1].is_jump) {
1044 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
1045 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
1046 hsn[CCREG]=1;
1047 // ...or hash tables
1048 hsn[RHASH]=1;
1049 hsn[RHTBL]=1;
1050 }
1051 // Coprocessor load/store needs FTEMP, even if not declared
1052 if(dops[i].itype==C2LS) {
1053 hsn[FTEMP]=0;
1054 }
1055 // Load/store L/R also uses FTEMP as a temporary register
1056 if (dops[i].itype == LOADLR || dops[i].itype == STORELR) {
1057 hsn[FTEMP]=0;
1058 }
1059 // Don't remove the miniht registers
1060 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
1061 {
1062 hsn[RHASH]=0;
1063 hsn[RHTBL]=0;
1064 }
1065}
1066
1067// We only want to allocate registers if we're going to use them again soon
1068static int needed_again(int r, int i)
1069{
1070 int j;
1071 int b=-1;
1072 int rn=10;
1073
1074 if (i > 0 && dops[i-1].is_ujump)
1075 {
1076 if(cinfo[i-1].ba<start || cinfo[i-1].ba>start+slen*4-4)
1077 return 0; // Don't need any registers if exiting the block
1078 }
1079 for(j=0;j<9;j++)
1080 {
1081 if(i+j>=slen) {
1082 j=slen-i-1;
1083 break;
1084 }
1085 if (dops[i+j].is_ujump)
1086 {
1087 // Don't go past an unconditonal jump
1088 j++;
1089 break;
1090 }
1091 if (dops[i+j].is_exception)
1092 {
1093 break;
1094 }
1095 }
1096 for(;j>=1;j--)
1097 {
1098 if(dops[i+j].rs1==r) rn=j;
1099 if(dops[i+j].rs2==r) rn=j;
1100 if((unneeded_reg[i+j]>>r)&1) rn=10;
1101 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
1102 {
1103 b=j;
1104 }
1105 }
1106 if(rn<10) return 1;
1107 (void)b;
1108 return 0;
1109}
1110
1111// Try to match register allocations at the end of a loop with those
1112// at the beginning
1113static int loop_reg(int i, int r, int hr)
1114{
1115 int j,k;
1116 for(j=0;j<9;j++)
1117 {
1118 if(i+j>=slen) {
1119 j=slen-i-1;
1120 break;
1121 }
1122 if (dops[i+j].is_ujump)
1123 {
1124 // Don't go past an unconditonal jump
1125 j++;
1126 break;
1127 }
1128 }
1129 k=0;
1130 if(i>0){
1131 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
1132 k--;
1133 }
1134 for(;k<j;k++)
1135 {
1136 assert(r < 64);
1137 if((unneeded_reg[i+k]>>r)&1) return hr;
1138 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
1139 {
1140 if(cinfo[i+k].ba>=start && cinfo[i+k].ba<(start+i*4))
1141 {
1142 int t=(cinfo[i+k].ba-start)>>2;
1143 int reg=get_reg(regs[t].regmap_entry,r);
1144 if(reg>=0) return reg;
1145 //reg=get_reg(regs[t+1].regmap_entry,r);
1146 //if(reg>=0) return reg;
1147 }
1148 }
1149 }
1150 return hr;
1151}
1152
1153
1154// Allocate every register, preserving source/target regs
1155static void alloc_all(struct regstat *cur,int i)
1156{
1157 int hr;
1158
1159 for(hr=0;hr<HOST_REGS;hr++) {
1160 if(hr!=EXCLUDE_REG) {
1161 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
1162 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
1163 {
1164 cur->regmap[hr]=-1;
1165 cur->dirty&=~(1<<hr);
1166 }
1167 // Don't need zeros
1168 if(cur->regmap[hr]==0)
1169 {
1170 cur->regmap[hr]=-1;
1171 cur->dirty&=~(1<<hr);
1172 }
1173 }
1174 }
1175}
1176
1177#ifndef NDEBUG
1178static int host_tempreg_in_use;
1179
1180static void host_tempreg_acquire(void)
1181{
1182 assert(!host_tempreg_in_use);
1183 host_tempreg_in_use = 1;
1184}
1185
1186static void host_tempreg_release(void)
1187{
1188 host_tempreg_in_use = 0;
1189}
1190#else
1191static void host_tempreg_acquire(void) {}
1192static void host_tempreg_release(void) {}
1193#endif
1194
1195#ifdef ASSEM_PRINT
1196extern void gen_interupt();
1197extern void do_insn_cmp();
1198#define FUNCNAME(f) { f, " " #f }
1199static const struct {
1200 void *addr;
1201 const char *name;
1202} function_names[] = {
1203 FUNCNAME(cc_interrupt),
1204 FUNCNAME(gen_interupt),
1205 FUNCNAME(ndrc_get_addr_ht),
1206 FUNCNAME(jump_handler_read8),
1207 FUNCNAME(jump_handler_read16),
1208 FUNCNAME(jump_handler_read32),
1209 FUNCNAME(jump_handler_write8),
1210 FUNCNAME(jump_handler_write16),
1211 FUNCNAME(jump_handler_write32),
1212 FUNCNAME(ndrc_write_invalidate_one),
1213 FUNCNAME(ndrc_write_invalidate_many),
1214 FUNCNAME(jump_to_new_pc),
1215 FUNCNAME(jump_break),
1216 FUNCNAME(jump_break_ds),
1217 FUNCNAME(jump_syscall),
1218 FUNCNAME(jump_syscall_ds),
1219 FUNCNAME(jump_overflow),
1220 FUNCNAME(jump_overflow_ds),
1221 FUNCNAME(jump_addrerror),
1222 FUNCNAME(jump_addrerror_ds),
1223 FUNCNAME(call_gteStall),
1224 FUNCNAME(new_dyna_leave),
1225 FUNCNAME(pcsx_mtc0),
1226 FUNCNAME(pcsx_mtc0_ds),
1227 FUNCNAME(execI),
1228#ifdef __aarch64__
1229 FUNCNAME(do_memhandler_pre),
1230 FUNCNAME(do_memhandler_post),
1231#endif
1232#ifdef DRC_DBG
1233# ifdef __aarch64__
1234 FUNCNAME(do_insn_cmp_arm64),
1235# else
1236 FUNCNAME(do_insn_cmp),
1237# endif
1238#endif
1239};
1240
1241static const char *func_name(const void *a)
1242{
1243 int i;
1244 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1245 if (function_names[i].addr == a)
1246 return function_names[i].name;
1247 return "";
1248}
1249
1250static const char *fpofs_name(u_int ofs)
1251{
1252 u_int *p = (u_int *)&dynarec_local + ofs/sizeof(u_int);
1253 static char buf[64];
1254 switch (ofs) {
1255 #define ofscase(x) case LO_##x: return " ; " #x
1256 ofscase(next_interupt);
1257 ofscase(cycle_count);
1258 ofscase(last_count);
1259 ofscase(pending_exception);
1260 ofscase(stop);
1261 ofscase(address);
1262 ofscase(lo);
1263 ofscase(hi);
1264 ofscase(PC);
1265 ofscase(cycle);
1266 ofscase(mem_rtab);
1267 ofscase(mem_wtab);
1268 ofscase(psxH_ptr);
1269 ofscase(invc_ptr);
1270 ofscase(ram_offset);
1271 #undef ofscase
1272 }
1273 buf[0] = 0;
1274 if (psxRegs.GPR.r <= p && p < &psxRegs.GPR.r[32])
1275 snprintf(buf, sizeof(buf), " ; r%d", (int)(p - psxRegs.GPR.r));
1276 else if (psxRegs.CP0.r <= p && p < &psxRegs.CP0.r[32])
1277 snprintf(buf, sizeof(buf), " ; cp0 $%d", (int)(p - psxRegs.CP0.r));
1278 else if (psxRegs.CP2D.r <= p && p < &psxRegs.CP2D.r[32])
1279 snprintf(buf, sizeof(buf), " ; cp2d $%d", (int)(p - psxRegs.CP2D.r));
1280 else if (psxRegs.CP2C.r <= p && p < &psxRegs.CP2C.r[32])
1281 snprintf(buf, sizeof(buf), " ; cp2c $%d", (int)(p - psxRegs.CP2C.r));
1282 return buf;
1283}
1284#else
1285#define func_name(x) ""
1286#define fpofs_name(x) ""
1287#endif
1288
1289#ifdef __i386__
1290#include "assem_x86.c"
1291#endif
1292#ifdef __x86_64__
1293#include "assem_x64.c"
1294#endif
1295#ifdef __arm__
1296#include "assem_arm.c"
1297#endif
1298#ifdef __aarch64__
1299#include "assem_arm64.c"
1300#endif
1301
1302static void *get_trampoline(const void *f)
1303{
1304 struct ndrc_tramp *tramp = NDRC_WRITE_OFFSET(&ndrc->tramp);
1305 size_t i;
1306
1307 for (i = 0; i < ARRAY_SIZE(tramp->f); i++) {
1308 if (tramp->f[i] == f || tramp->f[i] == NULL)
1309 break;
1310 }
1311 if (i == ARRAY_SIZE(tramp->f)) {
1312 SysPrintf("trampoline table is full, last func %p\n", f);
1313 abort();
1314 }
1315 if (tramp->f[i] == NULL) {
1316 start_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1317 tramp->f[i] = f;
1318 end_tcache_write(&tramp->f[i], &tramp->f[i + 1]);
1319#ifdef HAVE_LIBNX
1320 // invalidate the RX mirror (unsure if necessary, but just in case...)
1321 armDCacheFlush(&ndrc->tramp.f[i], sizeof(ndrc->tramp.f[i]));
1322#endif
1323 }
1324 return &ndrc->tramp.ops[i];
1325}
1326
1327static void emit_far_jump(const void *f)
1328{
1329 if (can_jump_or_call(f)) {
1330 emit_jmp(f);
1331 return;
1332 }
1333
1334 f = get_trampoline(f);
1335 emit_jmp(f);
1336}
1337
1338static void emit_far_call(const void *f)
1339{
1340 if (can_jump_or_call(f)) {
1341 emit_call(f);
1342 return;
1343 }
1344
1345 f = get_trampoline(f);
1346 emit_call(f);
1347}
1348
1349// Check if an address is already compiled
1350// but don't return addresses which are about to expire from the cache
1351static void *check_addr(u_int vaddr)
1352{
1353 struct ht_entry *ht_bin = hash_table_get(vaddr);
1354 size_t i;
1355 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1356 if (ht_bin->vaddr[i] == vaddr)
1357 if (doesnt_expire_soon(ht_bin->tcaddr[i]))
1358 return ht_bin->tcaddr[i];
1359 }
1360
1361 // refactor to get_addr_nocompile?
1362 u_int start_page = get_page_prev(vaddr);
1363 u_int page, end_page = get_page(vaddr);
1364
1365 stat_inc(stat_jump_in_lookups);
1366 for (page = start_page; page <= end_page; page++) {
1367 const struct block_info *block;
1368 for (block = blocks[page]; block != NULL; block = block->next) {
1369 if (vaddr < block->start)
1370 break;
1371 if (block->is_dirty || vaddr >= block->start + block->len)
1372 continue;
1373 if (!doesnt_expire_soon(ndrc->translation_cache + block->tc_offs))
1374 continue;
1375 for (i = 0; i < block->jump_in_cnt; i++)
1376 if (block->jump_in[i].vaddr == vaddr)
1377 break;
1378 if (i == block->jump_in_cnt)
1379 continue;
1380
1381 // Update existing entry with current address
1382 void *addr = block->jump_in[i].addr;
1383 if (ht_bin->vaddr[0] == vaddr) {
1384 ht_bin->tcaddr[0] = addr;
1385 return addr;
1386 }
1387 if (ht_bin->vaddr[1] == vaddr) {
1388 ht_bin->tcaddr[1] = addr;
1389 return addr;
1390 }
1391 // Insert into hash table with low priority.
1392 // Don't evict existing entries, as they are probably
1393 // addresses that are being accessed frequently.
1394 if (ht_bin->vaddr[0] == -1) {
1395 ht_bin->vaddr[0] = vaddr;
1396 ht_bin->tcaddr[0] = addr;
1397 }
1398 else if (ht_bin->vaddr[1] == -1) {
1399 ht_bin->vaddr[1] = vaddr;
1400 ht_bin->tcaddr[1] = addr;
1401 }
1402 return addr;
1403 }
1404 }
1405 return NULL;
1406}
1407
1408static void blocks_clear(struct block_info **head)
1409{
1410 struct block_info *cur, *next;
1411
1412 if ((cur = *head)) {
1413 *head = NULL;
1414 while (cur) {
1415 next = cur->next;
1416 free(cur);
1417 cur = next;
1418 }
1419 }
1420}
1421
1422static int blocks_remove_matching_addrs(struct block_info **head,
1423 u_int base_offs, int shift)
1424{
1425 struct block_info *next;
1426 int hit = 0;
1427 while (*head) {
1428 if ((((*head)->tc_offs ^ base_offs) >> shift) == 0) {
1429 inv_debug("EXP: rm block %08x (tc_offs %x)\n", (*head)->start, (*head)->tc_offs);
1430 invalidate_block(*head);
1431 next = (*head)->next;
1432 free(*head);
1433 *head = next;
1434 stat_dec(stat_blocks);
1435 hit = 1;
1436 }
1437 else
1438 {
1439 head = &((*head)->next);
1440 }
1441 }
1442 return hit;
1443}
1444
1445// This is called when we write to a compiled block (see do_invstub)
1446static void unlink_jumps_vaddr_range(u_int start, u_int end)
1447{
1448 u_int page, start_page = get_page(start), end_page = get_page(end - 1);
1449 int i;
1450
1451 for (page = start_page; page <= end_page; page++) {
1452 struct jump_info *ji = jumps[page];
1453 if (ji == NULL)
1454 continue;
1455 for (i = 0; i < ji->count; ) {
1456 if (ji->e[i].target_vaddr < start || ji->e[i].target_vaddr >= end) {
1457 i++;
1458 continue;
1459 }
1460
1461 inv_debug("INV: rm link to %08x (tc_offs %zx)\n", ji->e[i].target_vaddr,
1462 (u_char *)ji->e[i].stub - ndrc->translation_cache);
1463 void *host_addr = find_extjump_insn(ji->e[i].stub);
1464 mark_clear_cache(host_addr);
1465 set_jump_target(host_addr, ji->e[i].stub); // point back to dyna_linker stub
1466
1467 stat_dec(stat_links);
1468 ji->count--;
1469 if (i < ji->count) {
1470 ji->e[i] = ji->e[ji->count];
1471 continue;
1472 }
1473 i++;
1474 }
1475 }
1476}
1477
1478static void unlink_jumps_tc_range(struct jump_info *ji, u_int base_offs, int shift)
1479{
1480 int i;
1481 if (ji == NULL)
1482 return;
1483 for (i = 0; i < ji->count; ) {
1484 u_int tc_offs = (u_char *)ji->e[i].stub - ndrc->translation_cache;
1485 if (((tc_offs ^ base_offs) >> shift) != 0) {
1486 i++;
1487 continue;
1488 }
1489
1490 inv_debug("EXP: rm link to %08x (tc_offs %x)\n", ji->e[i].target_vaddr, tc_offs);
1491 stat_dec(stat_links);
1492 ji->count--;
1493 if (i < ji->count) {
1494 ji->e[i] = ji->e[ji->count];
1495 continue;
1496 }
1497 i++;
1498 }
1499}
1500
1501static void invalidate_block(struct block_info *block)
1502{
1503 u_int i;
1504
1505 block->is_dirty = 1;
1506 unlink_jumps_vaddr_range(block->start, block->start + block->len);
1507 for (i = 0; i < block->jump_in_cnt; i++)
1508 hash_table_remove(block->jump_in[i].vaddr);
1509}
1510
1511static int invalidate_range(u_int start, u_int end,
1512 u32 *inv_start_ret, u32 *inv_end_ret)
1513{
1514 struct block_info *last_block = NULL;
1515 u_int start_page = get_page_prev(start);
1516 u_int end_page = get_page(end - 1);
1517 u_int start_m = pmmask(start);
1518 u_int end_m = pmmask(end - 1);
1519 u_int inv_start, inv_end;
1520 u_int blk_start_m, blk_end_m;
1521 u_int page;
1522 int hit = 0;
1523
1524 // additional area without code (to supplement invalid_code[]), [start, end)
1525 // avoids excessive ndrc_write_invalidate*() calls
1526 inv_start = start_m & ~0xfff;
1527 inv_end = end_m | 0xfff;
1528
1529 for (page = start_page; page <= end_page; page++) {
1530 struct block_info *block;
1531 for (block = blocks[page]; block != NULL; block = block->next) {
1532 if (block->is_dirty)
1533 continue;
1534 last_block = block;
1535 blk_end_m = pmmask(block->start + block->len);
1536 if (blk_end_m <= start_m) {
1537 inv_start = max(inv_start, blk_end_m);
1538 continue;
1539 }
1540 blk_start_m = pmmask(block->start);
1541 if (end_m <= blk_start_m) {
1542 inv_end = min(inv_end, blk_start_m - 1);
1543 continue;
1544 }
1545 if (!block->source) // "hack" block - leave it alone
1546 continue;
1547
1548 hit++;
1549 invalidate_block(block);
1550 stat_inc(stat_inv_hits);
1551 }
1552 }
1553
1554 if (!hit && last_block && last_block->source) {
1555 // could be some leftover unused block, uselessly trapping writes
1556 last_block->inv_near_misses++;
1557 if (last_block->inv_near_misses > 128) {
1558 invalidate_block(last_block);
1559 stat_inc(stat_inv_hits);
1560 hit++;
1561 }
1562 }
1563 if (hit) {
1564 do_clear_cache();
1565#ifdef USE_MINI_HT
1566 memset(mini_ht, -1, sizeof(mini_ht));
1567#endif
1568 }
1569
1570 if (inv_start <= (start_m & ~0xfff) && inv_end >= (start_m | 0xfff))
1571 // the whole page is empty now
1572 mark_invalid_code(start, 1, 1);
1573
1574 if (inv_start_ret) *inv_start_ret = inv_start | (start & 0xe0000000);
1575 if (inv_end_ret) *inv_end_ret = inv_end | (end & 0xe0000000);
1576 return hit;
1577}
1578
1579void new_dynarec_invalidate_range(unsigned int start, unsigned int end)
1580{
1581 invalidate_range(start, end, NULL, NULL);
1582}
1583
1584static void ndrc_write_invalidate_many(u_int start, u_int end)
1585{
1586 // this check is done by the caller
1587 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1588 int ret = invalidate_range(start, end, &inv_code_start, &inv_code_end);
1589#ifdef INV_DEBUG_W
1590 int invc = invalid_code[start >> 12];
1591 u_int len = end - start;
1592 if (ret)
1593 printf("INV ADDR: %08x/%02x hit %d blocks\n", start, len, ret);
1594 else
1595 printf("INV ADDR: %08x/%02x miss, inv %08x-%08x invc %d->%d\n", start, len,
1596 inv_code_start, inv_code_end, invc, invalid_code[start >> 12]);
1597 check_for_block_changes(start, end);
1598#endif
1599 stat_inc(stat_inv_addr_calls);
1600 (void)ret;
1601}
1602
1603void ndrc_write_invalidate_one(u_int addr)
1604{
1605 ndrc_write_invalidate_many(addr, addr + 4);
1606}
1607
1608// This is called when loading a save state.
1609// Anything could have changed, so invalidate everything.
1610void new_dynarec_invalidate_all_pages(void)
1611{
1612 struct block_info *block;
1613 u_int page;
1614 for (page = 0; page < ARRAY_SIZE(blocks); page++) {
1615 for (block = blocks[page]; block != NULL; block = block->next) {
1616 if (block->is_dirty)
1617 continue;
1618 if (!block->source) // hack block?
1619 continue;
1620 invalidate_block(block);
1621 }
1622 }
1623
1624 #ifdef USE_MINI_HT
1625 memset(mini_ht, -1, sizeof(mini_ht));
1626 #endif
1627 do_clear_cache();
1628}
1629
1630// Add an entry to jump_out after making a link
1631// src should point to code by emit_extjump()
1632void ndrc_add_jump_out(u_int vaddr, void *src)
1633{
1634 inv_debug("ndrc_add_jump_out: %p -> %x\n", src, vaddr);
1635 u_int page = get_page(vaddr);
1636 struct jump_info *ji;
1637
1638 stat_inc(stat_links);
1639 check_extjump2(src);
1640 ji = jumps[page];
1641 if (ji == NULL) {
1642 ji = malloc(sizeof(*ji) + sizeof(ji->e[0]) * 16);
1643 ji->alloc = 16;
1644 ji->count = 0;
1645 }
1646 else if (ji->count >= ji->alloc) {
1647 ji->alloc += 16;
1648 ji = realloc(ji, sizeof(*ji) + sizeof(ji->e[0]) * ji->alloc);
1649 }
1650 jumps[page] = ji;
1651 ji->e[ji->count].target_vaddr = vaddr;
1652 ji->e[ji->count].stub = src;
1653 ji->count++;
1654}
1655
1656/* Register allocation */
1657
1658static void alloc_set(struct regstat *cur, int reg, int hr)
1659{
1660 cur->regmap[hr] = reg;
1661 cur->dirty &= ~(1u << hr);
1662 cur->isconst &= ~(1u << hr);
1663 cur->noevict |= 1u << hr;
1664}
1665
1666static void evict_alloc_reg(struct regstat *cur, int i, int reg, int preferred_hr)
1667{
1668 u_char hsn[MAXREG+1];
1669 int j, r, hr;
1670 memset(hsn, 10, sizeof(hsn));
1671 lsn(hsn, i);
1672 //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]);
1673 if(i>0) {
1674 // Don't evict the cycle count at entry points, otherwise the entry
1675 // stub will have to write it.
1676 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1677 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1678 for(j=10;j>=3;j--)
1679 {
1680 // Alloc preferred register if available
1681 if (!((cur->noevict >> preferred_hr) & 1)
1682 && hsn[cur->regmap[preferred_hr]] == j)
1683 {
1684 alloc_set(cur, reg, preferred_hr);
1685 return;
1686 }
1687 for(r=1;r<=MAXREG;r++)
1688 {
1689 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1690 for(hr=0;hr<HOST_REGS;hr++) {
1691 if (hr == EXCLUDE_REG || ((cur->noevict >> hr) & 1))
1692 continue;
1693 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1694 if(cur->regmap[hr]==r) {
1695 alloc_set(cur, reg, hr);
1696 return;
1697 }
1698 }
1699 }
1700 }
1701 }
1702 }
1703 }
1704 for(j=10;j>=0;j--)
1705 {
1706 for(r=1;r<=MAXREG;r++)
1707 {
1708 if(hsn[r]==j) {
1709 for(hr=0;hr<HOST_REGS;hr++) {
1710 if (hr == EXCLUDE_REG || ((cur->noevict >> hr) & 1))
1711 continue;
1712 if(cur->regmap[hr]==r) {
1713 alloc_set(cur, reg, hr);
1714 return;
1715 }
1716 }
1717 }
1718 }
1719 }
1720 SysPrintf("This shouldn't happen (evict_alloc_reg)\n");
1721 abort();
1722}
1723
1724// Note: registers are allocated clean (unmodified state)
1725// if you intend to modify the register, you must call dirty_reg().
1726static void alloc_reg(struct regstat *cur,int i,signed char reg)
1727{
1728 int r,hr;
1729 int preferred_reg = PREFERRED_REG_FIRST
1730 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1731 if (reg == CCREG) preferred_reg = HOST_CCREG;
1732 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1733 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1734 assert(reg >= 0);
1735
1736 // Don't allocate unused registers
1737 if((cur->u>>reg)&1) return;
1738
1739 // see if it's already allocated
1740 if ((hr = get_reg(cur->regmap, reg)) >= 0) {
1741 cur->noevict |= 1u << hr;
1742 return;
1743 }
1744
1745 // Keep the same mapping if the register was already allocated in a loop
1746 preferred_reg = loop_reg(i,reg,preferred_reg);
1747
1748 // Try to allocate the preferred register
1749 if (cur->regmap[preferred_reg] == -1) {
1750 alloc_set(cur, reg, preferred_reg);
1751 return;
1752 }
1753 r=cur->regmap[preferred_reg];
1754 assert(r < 64);
1755 if((cur->u>>r)&1) {
1756 alloc_set(cur, reg, preferred_reg);
1757 return;
1758 }
1759
1760 // Clear any unneeded registers
1761 // We try to keep the mapping consistent, if possible, because it
1762 // makes branches easier (especially loops). So we try to allocate
1763 // first (see above) before removing old mappings. If this is not
1764 // possible then go ahead and clear out the registers that are no
1765 // longer needed.
1766 for(hr=0;hr<HOST_REGS;hr++)
1767 {
1768 r=cur->regmap[hr];
1769 if(r>=0) {
1770 assert(r < 64);
1771 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1772 }
1773 }
1774
1775 // Try to allocate any available register, but prefer
1776 // registers that have not been used recently.
1777 if (i > 0) {
1778 for (hr = PREFERRED_REG_FIRST; ; ) {
1779 if (cur->regmap[hr] < 0) {
1780 int oldreg = regs[i-1].regmap[hr];
1781 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1782 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1783 {
1784 alloc_set(cur, reg, hr);
1785 return;
1786 }
1787 }
1788 hr++;
1789 if (hr == EXCLUDE_REG)
1790 hr++;
1791 if (hr == HOST_REGS)
1792 hr = 0;
1793 if (hr == PREFERRED_REG_FIRST)
1794 break;
1795 }
1796 }
1797
1798 // Try to allocate any available register
1799 for (hr = PREFERRED_REG_FIRST; ; ) {
1800 if (cur->regmap[hr] < 0) {
1801 alloc_set(cur, reg, hr);
1802 return;
1803 }
1804 hr++;
1805 if (hr == EXCLUDE_REG)
1806 hr++;
1807 if (hr == HOST_REGS)
1808 hr = 0;
1809 if (hr == PREFERRED_REG_FIRST)
1810 break;
1811 }
1812
1813 // Ok, now we have to evict someone
1814 // Pick a register we hopefully won't need soon
1815 evict_alloc_reg(cur, i, reg, preferred_reg);
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
1825 // see if it's already allocated
1826 for (hr = 0; hr < HOST_REGS; hr++)
1827 {
1828 if (hr != EXCLUDE_REG && cur->regmap[hr] == reg) {
1829 cur->noevict |= 1u << hr;
1830 return;
1831 }
1832 }
1833
1834 // Try to allocate any available register
1835 for(hr=HOST_REGS-1;hr>=0;hr--) {
1836 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1837 alloc_set(cur, reg, 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 alloc_set(cur, reg, hr);
1851 return;
1852 }
1853 }
1854 }
1855 }
1856
1857 // Ok, now we have to evict someone
1858 // Pick a register we hopefully won't need soon
1859 evict_alloc_reg(cur, i, reg, 0);
1860}
1861
1862static void mov_alloc(struct regstat *current,int i)
1863{
1864 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1865 alloc_cc(current,i); // for stalls
1866 dirty_reg(current,CCREG);
1867 }
1868
1869 // Note: Don't need to actually alloc the source registers
1870 //alloc_reg(current,i,dops[i].rs1);
1871 alloc_reg(current,i,dops[i].rt1);
1872
1873 clear_const(current,dops[i].rs1);
1874 clear_const(current,dops[i].rt1);
1875 dirty_reg(current,dops[i].rt1);
1876}
1877
1878static void shiftimm_alloc(struct regstat *current,int i)
1879{
1880 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1881 {
1882 if(dops[i].rt1) {
1883 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1884 else dops[i].use_lt1=!!dops[i].rs1;
1885 alloc_reg(current,i,dops[i].rt1);
1886 dirty_reg(current,dops[i].rt1);
1887 if(is_const(current,dops[i].rs1)) {
1888 int v=get_const(current,dops[i].rs1);
1889 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<cinfo[i].imm);
1890 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>cinfo[i].imm);
1891 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>cinfo[i].imm);
1892 }
1893 else clear_const(current,dops[i].rt1);
1894 }
1895 }
1896 else
1897 {
1898 clear_const(current,dops[i].rs1);
1899 clear_const(current,dops[i].rt1);
1900 }
1901
1902 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1903 {
1904 assert(0);
1905 }
1906 if(dops[i].opcode2==0x3c) // DSLL32
1907 {
1908 assert(0);
1909 }
1910 if(dops[i].opcode2==0x3e) // DSRL32
1911 {
1912 assert(0);
1913 }
1914 if(dops[i].opcode2==0x3f) // DSRA32
1915 {
1916 assert(0);
1917 }
1918}
1919
1920static void shift_alloc(struct regstat *current,int i)
1921{
1922 if(dops[i].rt1) {
1923 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1924 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1925 alloc_reg(current,i,dops[i].rt1);
1926 if(dops[i].rt1==dops[i].rs2) {
1927 alloc_reg_temp(current,i,-1);
1928 cinfo[i].min_free_regs=1;
1929 }
1930 clear_const(current,dops[i].rs1);
1931 clear_const(current,dops[i].rs2);
1932 clear_const(current,dops[i].rt1);
1933 dirty_reg(current,dops[i].rt1);
1934 }
1935}
1936
1937static void alu_alloc(struct regstat *current,int i)
1938{
1939 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1940 if(dops[i].rt1) {
1941 if(dops[i].rs1&&dops[i].rs2) {
1942 alloc_reg(current,i,dops[i].rs1);
1943 alloc_reg(current,i,dops[i].rs2);
1944 }
1945 else {
1946 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1947 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1948 }
1949 alloc_reg(current,i,dops[i].rt1);
1950 }
1951 if (dops[i].may_except) {
1952 alloc_cc_optional(current, i); // for exceptions
1953 alloc_reg_temp(current, i, -1);
1954 cinfo[i].min_free_regs = 1;
1955 }
1956 }
1957 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1958 if(dops[i].rt1) {
1959 alloc_reg(current,i,dops[i].rs1);
1960 alloc_reg(current,i,dops[i].rs2);
1961 alloc_reg(current,i,dops[i].rt1);
1962 }
1963 }
1964 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1965 if(dops[i].rt1) {
1966 if(dops[i].rs1&&dops[i].rs2) {
1967 alloc_reg(current,i,dops[i].rs1);
1968 alloc_reg(current,i,dops[i].rs2);
1969 }
1970 else
1971 {
1972 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1973 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1974 }
1975 alloc_reg(current,i,dops[i].rt1);
1976 }
1977 }
1978 clear_const(current,dops[i].rs1);
1979 clear_const(current,dops[i].rs2);
1980 clear_const(current,dops[i].rt1);
1981 dirty_reg(current,dops[i].rt1);
1982}
1983
1984static void imm16_alloc(struct regstat *current,int i)
1985{
1986 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1987 else dops[i].use_lt1=!!dops[i].rs1;
1988 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1989 if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1990 clear_const(current,dops[i].rs1);
1991 clear_const(current,dops[i].rt1);
1992 }
1993 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1994 if(is_const(current,dops[i].rs1)) {
1995 int v=get_const(current,dops[i].rs1);
1996 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&cinfo[i].imm);
1997 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|cinfo[i].imm);
1998 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^cinfo[i].imm);
1999 }
2000 else clear_const(current,dops[i].rt1);
2001 }
2002 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2003 if(is_const(current,dops[i].rs1)) {
2004 int v=get_const(current,dops[i].rs1);
2005 set_const(current,dops[i].rt1,v+cinfo[i].imm);
2006 }
2007 else clear_const(current,dops[i].rt1);
2008 if (dops[i].may_except) {
2009 alloc_cc_optional(current, i); // for exceptions
2010 alloc_reg_temp(current, i, -1);
2011 cinfo[i].min_free_regs = 1;
2012 }
2013 }
2014 else {
2015 set_const(current,dops[i].rt1,cinfo[i].imm<<16); // LUI
2016 }
2017 dirty_reg(current,dops[i].rt1);
2018}
2019
2020static void load_alloc(struct regstat *current,int i)
2021{
2022 int need_temp = 0;
2023 clear_const(current,dops[i].rt1);
2024 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
2025 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
2026 if (needed_again(dops[i].rs1, i))
2027 alloc_reg(current, i, dops[i].rs1);
2028 if (ram_offset)
2029 alloc_reg(current, i, ROREG);
2030 if (dops[i].may_except) {
2031 alloc_cc_optional(current, i); // for exceptions
2032 need_temp = 1;
2033 }
2034 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
2035 alloc_reg(current,i,dops[i].rt1);
2036 assert(get_reg_w(current->regmap, dops[i].rt1)>=0);
2037 dirty_reg(current,dops[i].rt1);
2038 // LWL/LWR need a temporary register for the old value
2039 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2040 {
2041 alloc_reg(current,i,FTEMP);
2042 need_temp = 1;
2043 }
2044 }
2045 else
2046 {
2047 // Load to r0 or unneeded register (dummy load)
2048 // but we still need a register to calculate the address
2049 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
2050 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
2051 need_temp = 1;
2052 }
2053 if (need_temp) {
2054 alloc_reg_temp(current, i, -1);
2055 cinfo[i].min_free_regs = 1;
2056 }
2057}
2058
2059// this may eat up to 7 registers
2060static void store_alloc(struct regstat *current, int i)
2061{
2062 clear_const(current,dops[i].rs2);
2063 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
2064 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2065 alloc_reg(current,i,dops[i].rs2);
2066 if (ram_offset)
2067 alloc_reg(current, i, ROREG);
2068 #if defined(HOST_IMM8)
2069 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2070 alloc_reg(current, i, INVCP);
2071 #endif
2072 if (dops[i].opcode == 0x2a || dops[i].opcode == 0x2e) { // SWL/SWL
2073 alloc_reg(current,i,FTEMP);
2074 }
2075 if (dops[i].may_except)
2076 alloc_cc_optional(current, i); // for exceptions
2077 // We need a temporary register for address generation
2078 alloc_reg_temp(current,i,-1);
2079 cinfo[i].min_free_regs=1;
2080}
2081
2082static void c2ls_alloc(struct regstat *current, int i)
2083{
2084 clear_const(current,dops[i].rt1);
2085 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
2086 alloc_reg(current,i,FTEMP);
2087 if (ram_offset)
2088 alloc_reg(current, i, ROREG);
2089 #if defined(HOST_IMM8)
2090 // On CPUs without 32-bit immediates we need a pointer to invalid_code
2091 if (dops[i].opcode == 0x3a) // SWC2
2092 alloc_reg(current,i,INVCP);
2093 #endif
2094 if (dops[i].may_except)
2095 alloc_cc_optional(current, i); // for exceptions
2096 // We need a temporary register for address generation
2097 alloc_reg_temp(current,i,-1);
2098 cinfo[i].min_free_regs=1;
2099}
2100
2101#ifndef multdiv_alloc
2102static void multdiv_alloc(struct regstat *current,int i)
2103{
2104 // case 0x18: MULT
2105 // case 0x19: MULTU
2106 // case 0x1A: DIV
2107 // case 0x1B: DIVU
2108 clear_const(current,dops[i].rs1);
2109 clear_const(current,dops[i].rs2);
2110 alloc_cc(current,i); // for stalls
2111 dirty_reg(current,CCREG);
2112 if(dops[i].rs1&&dops[i].rs2)
2113 {
2114 current->u&=~(1LL<<HIREG);
2115 current->u&=~(1LL<<LOREG);
2116 alloc_reg(current,i,HIREG);
2117 alloc_reg(current,i,LOREG);
2118 alloc_reg(current,i,dops[i].rs1);
2119 alloc_reg(current,i,dops[i].rs2);
2120 dirty_reg(current,HIREG);
2121 dirty_reg(current,LOREG);
2122 }
2123 else
2124 {
2125 // Multiply by zero is zero.
2126 // MIPS does not have a divide by zero exception.
2127 alloc_reg(current,i,HIREG);
2128 alloc_reg(current,i,LOREG);
2129 dirty_reg(current,HIREG);
2130 dirty_reg(current,LOREG);
2131 if (dops[i].rs1 && ((dops[i].opcode2 & 0x3e) == 0x1a)) // div(u) 0
2132 alloc_reg(current, i, dops[i].rs1);
2133 }
2134}
2135#endif
2136
2137static void cop0_alloc(struct regstat *current,int i)
2138{
2139 if(dops[i].opcode2==0) // MFC0
2140 {
2141 if(dops[i].rt1) {
2142 clear_const(current,dops[i].rt1);
2143 alloc_reg(current,i,dops[i].rt1);
2144 dirty_reg(current,dops[i].rt1);
2145 }
2146 }
2147 else if(dops[i].opcode2==4) // MTC0
2148 {
2149 if (((source[i]>>11)&0x1e) == 12) {
2150 alloc_cc(current, i);
2151 dirty_reg(current, CCREG);
2152 }
2153 if(dops[i].rs1){
2154 clear_const(current,dops[i].rs1);
2155 alloc_reg(current,i,dops[i].rs1);
2156 alloc_all(current,i);
2157 }
2158 else {
2159 alloc_all(current,i); // FIXME: Keep r0
2160 current->u&=~1LL;
2161 alloc_reg(current,i,0);
2162 }
2163 cinfo[i].min_free_regs = HOST_REGS;
2164 }
2165}
2166
2167static void rfe_alloc(struct regstat *current, int i)
2168{
2169 alloc_all(current, i);
2170 cinfo[i].min_free_regs = HOST_REGS;
2171}
2172
2173static void cop2_alloc(struct regstat *current,int i)
2174{
2175 if (dops[i].opcode2 < 3) // MFC2/CFC2
2176 {
2177 alloc_cc(current,i); // for stalls
2178 dirty_reg(current,CCREG);
2179 if(dops[i].rt1){
2180 clear_const(current,dops[i].rt1);
2181 alloc_reg(current,i,dops[i].rt1);
2182 dirty_reg(current,dops[i].rt1);
2183 }
2184 }
2185 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2186 {
2187 if(dops[i].rs1){
2188 clear_const(current,dops[i].rs1);
2189 alloc_reg(current,i,dops[i].rs1);
2190 }
2191 else {
2192 current->u&=~1LL;
2193 alloc_reg(current,i,0);
2194 }
2195 }
2196 alloc_reg_temp(current,i,-1);
2197 cinfo[i].min_free_regs=1;
2198}
2199
2200static void c2op_alloc(struct regstat *current,int i)
2201{
2202 alloc_cc(current,i); // for stalls
2203 dirty_reg(current,CCREG);
2204 alloc_reg_temp(current,i,-1);
2205}
2206
2207static void syscall_alloc(struct regstat *current,int i)
2208{
2209 alloc_cc(current,i);
2210 dirty_reg(current,CCREG);
2211 alloc_all(current,i);
2212 cinfo[i].min_free_regs=HOST_REGS;
2213 current->isconst=0;
2214}
2215
2216static void delayslot_alloc(struct regstat *current,int i)
2217{
2218 switch(dops[i].itype) {
2219 case UJUMP:
2220 case CJUMP:
2221 case SJUMP:
2222 case RJUMP:
2223 case SYSCALL:
2224 case HLECALL:
2225 case IMM16:
2226 imm16_alloc(current,i);
2227 break;
2228 case LOAD:
2229 case LOADLR:
2230 load_alloc(current,i);
2231 break;
2232 case STORE:
2233 case STORELR:
2234 store_alloc(current,i);
2235 break;
2236 case ALU:
2237 alu_alloc(current,i);
2238 break;
2239 case SHIFT:
2240 shift_alloc(current,i);
2241 break;
2242 case MULTDIV:
2243 multdiv_alloc(current,i);
2244 break;
2245 case SHIFTIMM:
2246 shiftimm_alloc(current,i);
2247 break;
2248 case MOV:
2249 mov_alloc(current,i);
2250 break;
2251 case COP0:
2252 cop0_alloc(current,i);
2253 break;
2254 case RFE:
2255 rfe_alloc(current,i);
2256 break;
2257 case COP2:
2258 cop2_alloc(current,i);
2259 break;
2260 case C2LS:
2261 c2ls_alloc(current,i);
2262 break;
2263 case C2OP:
2264 c2op_alloc(current,i);
2265 break;
2266 }
2267}
2268
2269static void add_stub(enum stub_type type, void *addr, void *retaddr,
2270 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2271{
2272 assert(stubcount < ARRAY_SIZE(stubs));
2273 stubs[stubcount].type = type;
2274 stubs[stubcount].addr = addr;
2275 stubs[stubcount].retaddr = retaddr;
2276 stubs[stubcount].a = a;
2277 stubs[stubcount].b = b;
2278 stubs[stubcount].c = c;
2279 stubs[stubcount].d = d;
2280 stubs[stubcount].e = e;
2281 stubcount++;
2282}
2283
2284static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2285 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2286{
2287 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2288}
2289
2290// Write out a single register
2291static void wb_register(signed char r, const signed char regmap[], u_int dirty)
2292{
2293 int hr;
2294 for(hr=0;hr<HOST_REGS;hr++) {
2295 if(hr!=EXCLUDE_REG) {
2296 if(regmap[hr]==r) {
2297 if((dirty>>hr)&1) {
2298 assert(regmap[hr]<64);
2299 emit_storereg(r,hr);
2300 }
2301 break;
2302 }
2303 }
2304 }
2305}
2306
2307static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2308{
2309 //if(dirty_pre==dirty) return;
2310 int hr, r;
2311 for (hr = 0; hr < HOST_REGS; hr++) {
2312 r = pre[hr];
2313 if (r < 1 || r > 33 || ((u >> r) & 1))
2314 continue;
2315 if (((dirty_pre & ~dirty) >> hr) & 1)
2316 emit_storereg(r, hr);
2317 }
2318}
2319
2320// trashes r2
2321static void pass_args(int a0, int a1)
2322{
2323 if(a0==1&&a1==0) {
2324 // must swap
2325 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2326 }
2327 else if(a0!=0&&a1==0) {
2328 emit_mov(a1,1);
2329 if (a0>=0) emit_mov(a0,0);
2330 }
2331 else {
2332 if(a0>=0&&a0!=0) emit_mov(a0,0);
2333 if(a1>=0&&a1!=1) emit_mov(a1,1);
2334 }
2335}
2336
2337static void alu_assemble(int i, const struct regstat *i_regs, int ccadj_)
2338{
2339 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2340 int do_oflow = dops[i].may_except; // ADD/SUB with exceptions enabled
2341 if (dops[i].rt1 || do_oflow) {
2342 int do_exception_check = 0;
2343 signed char s1, s2, t, tmp;
2344 t = get_reg_w(i_regs->regmap, dops[i].rt1);
2345 tmp = get_reg_temp(i_regs->regmap);
2346 if (do_oflow)
2347 assert(tmp >= 0);
2348 if (t < 0 && do_oflow)
2349 t = tmp;
2350 if (t >= 0) {
2351 s1 = get_reg(i_regs->regmap, dops[i].rs1);
2352 s2 = get_reg(i_regs->regmap, dops[i].rs2);
2353 if (dops[i].rs1 && dops[i].rs2) {
2354 assert(s1>=0);
2355 assert(s2>=0);
2356 if (dops[i].opcode2 & 2) {
2357 if (do_oflow) {
2358 emit_subs(s1, s2, tmp);
2359 do_exception_check = 1;
2360 }
2361 else
2362 emit_sub(s1,s2,t);
2363 }
2364 else {
2365 if (do_oflow) {
2366 emit_adds(s1, s2, tmp);
2367 do_exception_check = 1;
2368 }
2369 else
2370 emit_add(s1,s2,t);
2371 }
2372 }
2373 else if(dops[i].rs1) {
2374 if(s1>=0) emit_mov(s1,t);
2375 else emit_loadreg(dops[i].rs1,t);
2376 }
2377 else if(dops[i].rs2) {
2378 if (s2 < 0) {
2379 emit_loadreg(dops[i].rs2, t);
2380 s2 = t;
2381 }
2382 if (dops[i].opcode2 & 2) {
2383 if (do_oflow) {
2384 emit_negs(s2, tmp);
2385 do_exception_check = 1;
2386 }
2387 else
2388 emit_neg(s2, t);
2389 }
2390 else if (s2 != t)
2391 emit_mov(s2, t);
2392 }
2393 else
2394 emit_zeroreg(t);
2395 }
2396 if (do_exception_check) {
2397 void *jaddr = out;
2398 emit_jo(0);
2399 if (t >= 0 && tmp != t)
2400 emit_mov(tmp, t);
2401 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2402 }
2403 }
2404 }
2405 else if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2406 if(dops[i].rt1) {
2407 signed char s1l,s2l,t;
2408 {
2409 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2410 //assert(t>=0);
2411 if(t>=0) {
2412 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2413 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2414 if(dops[i].rs2==0) // rx<r0
2415 {
2416 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2417 assert(s1l>=0);
2418 emit_shrimm(s1l,31,t);
2419 }
2420 else // SLTU (unsigned can not be less than zero, 0<0)
2421 emit_zeroreg(t);
2422 }
2423 else if(dops[i].rs1==0) // r0<rx
2424 {
2425 assert(s2l>=0);
2426 if(dops[i].opcode2==0x2a) // SLT
2427 emit_set_gz32(s2l,t);
2428 else // SLTU (set if not zero)
2429 emit_set_nz32(s2l,t);
2430 }
2431 else{
2432 assert(s1l>=0);assert(s2l>=0);
2433 if(dops[i].opcode2==0x2a) // SLT
2434 emit_set_if_less32(s1l,s2l,t);
2435 else // SLTU
2436 emit_set_if_carry32(s1l,s2l,t);
2437 }
2438 }
2439 }
2440 }
2441 }
2442 else if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2443 if(dops[i].rt1) {
2444 signed char s1l,s2l,tl;
2445 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2446 {
2447 if(tl>=0) {
2448 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2449 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2450 if(dops[i].rs1&&dops[i].rs2) {
2451 assert(s1l>=0);
2452 assert(s2l>=0);
2453 if(dops[i].opcode2==0x24) { // AND
2454 emit_and(s1l,s2l,tl);
2455 } else
2456 if(dops[i].opcode2==0x25) { // OR
2457 emit_or(s1l,s2l,tl);
2458 } else
2459 if(dops[i].opcode2==0x26) { // XOR
2460 emit_xor(s1l,s2l,tl);
2461 } else
2462 if(dops[i].opcode2==0x27) { // NOR
2463 emit_or(s1l,s2l,tl);
2464 emit_not(tl,tl);
2465 }
2466 }
2467 else
2468 {
2469 if(dops[i].opcode2==0x24) { // AND
2470 emit_zeroreg(tl);
2471 } else
2472 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2473 if(dops[i].rs1){
2474 if(s1l>=0) emit_mov(s1l,tl);
2475 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2476 }
2477 else
2478 if(dops[i].rs2){
2479 if(s2l>=0) emit_mov(s2l,tl);
2480 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2481 }
2482 else emit_zeroreg(tl);
2483 } else
2484 if(dops[i].opcode2==0x27) { // NOR
2485 if(dops[i].rs1){
2486 if(s1l>=0) emit_not(s1l,tl);
2487 else {
2488 emit_loadreg(dops[i].rs1,tl);
2489 emit_not(tl,tl);
2490 }
2491 }
2492 else
2493 if(dops[i].rs2){
2494 if(s2l>=0) emit_not(s2l,tl);
2495 else {
2496 emit_loadreg(dops[i].rs2,tl);
2497 emit_not(tl,tl);
2498 }
2499 }
2500 else emit_movimm(-1,tl);
2501 }
2502 }
2503 }
2504 }
2505 }
2506 }
2507}
2508
2509static void imm16_assemble(int i, const struct regstat *i_regs, int ccadj_)
2510{
2511 if (dops[i].opcode==0x0f) { // LUI
2512 if(dops[i].rt1) {
2513 signed char t;
2514 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2515 //assert(t>=0);
2516 if(t>=0) {
2517 if(!((i_regs->isconst>>t)&1))
2518 emit_movimm(cinfo[i].imm<<16,t);
2519 }
2520 }
2521 }
2522 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2523 int is_addi = dops[i].may_except;
2524 if (dops[i].rt1 || is_addi) {
2525 signed char s, t, tmp;
2526 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2527 s=get_reg(i_regs->regmap,dops[i].rs1);
2528 if(dops[i].rs1) {
2529 tmp = get_reg_temp(i_regs->regmap);
2530 if (is_addi) {
2531 assert(tmp >= 0);
2532 if (t < 0) t = tmp;
2533 }
2534 if(t>=0) {
2535 if(!((i_regs->isconst>>t)&1)) {
2536 int sum, do_exception_check = 0;
2537 if (s < 0) {
2538 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2539 if (is_addi) {
2540 emit_addimm_and_set_flags3(t, cinfo[i].imm, tmp);
2541 do_exception_check = 1;
2542 }
2543 else
2544 emit_addimm(t, cinfo[i].imm, t);
2545 } else {
2546 if (!((i_regs->wasconst >> s) & 1)) {
2547 if (is_addi) {
2548 emit_addimm_and_set_flags3(s, cinfo[i].imm, tmp);
2549 do_exception_check = 1;
2550 }
2551 else
2552 emit_addimm(s, cinfo[i].imm, t);
2553 }
2554 else {
2555 int oflow = add_overflow(constmap[i][s], cinfo[i].imm, sum);
2556 if (is_addi && oflow)
2557 do_exception_check = 2;
2558 else
2559 emit_movimm(sum, t);
2560 }
2561 }
2562 if (do_exception_check) {
2563 void *jaddr = out;
2564 if (do_exception_check == 2)
2565 emit_jmp(0);
2566 else {
2567 emit_jo(0);
2568 if (tmp != t)
2569 emit_mov(tmp, t);
2570 }
2571 add_stub_r(OVERFLOW_STUB, jaddr, out, i, 0, i_regs, ccadj_, 0);
2572 }
2573 }
2574 }
2575 } else {
2576 if(t>=0) {
2577 if(!((i_regs->isconst>>t)&1))
2578 emit_movimm(cinfo[i].imm,t);
2579 }
2580 }
2581 }
2582 }
2583 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2584 if(dops[i].rt1) {
2585 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2586 signed char sl,t;
2587 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2588 sl=get_reg(i_regs->regmap,dops[i].rs1);
2589 //assert(t>=0);
2590 if(t>=0) {
2591 if(dops[i].rs1>0) {
2592 if(dops[i].opcode==0x0a) { // SLTI
2593 if(sl<0) {
2594 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2595 emit_slti32(t,cinfo[i].imm,t);
2596 }else{
2597 emit_slti32(sl,cinfo[i].imm,t);
2598 }
2599 }
2600 else { // SLTIU
2601 if(sl<0) {
2602 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2603 emit_sltiu32(t,cinfo[i].imm,t);
2604 }else{
2605 emit_sltiu32(sl,cinfo[i].imm,t);
2606 }
2607 }
2608 }else{
2609 // SLTI(U) with r0 is just stupid,
2610 // nonetheless examples can be found
2611 if(dops[i].opcode==0x0a) // SLTI
2612 if(0<cinfo[i].imm) emit_movimm(1,t);
2613 else emit_zeroreg(t);
2614 else // SLTIU
2615 {
2616 if(cinfo[i].imm) emit_movimm(1,t);
2617 else emit_zeroreg(t);
2618 }
2619 }
2620 }
2621 }
2622 }
2623 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2624 if(dops[i].rt1) {
2625 signed char sl,tl;
2626 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
2627 sl=get_reg(i_regs->regmap,dops[i].rs1);
2628 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2629 if(dops[i].opcode==0x0c) //ANDI
2630 {
2631 if(dops[i].rs1) {
2632 if(sl<0) {
2633 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2634 emit_andimm(tl,cinfo[i].imm,tl);
2635 }else{
2636 if(!((i_regs->wasconst>>sl)&1))
2637 emit_andimm(sl,cinfo[i].imm,tl);
2638 else
2639 emit_movimm(constmap[i][sl]&cinfo[i].imm,tl);
2640 }
2641 }
2642 else
2643 emit_zeroreg(tl);
2644 }
2645 else
2646 {
2647 if(dops[i].rs1) {
2648 if(sl<0) {
2649 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2650 }
2651 if(dops[i].opcode==0x0d) { // ORI
2652 if(sl<0) {
2653 emit_orimm(tl,cinfo[i].imm,tl);
2654 }else{
2655 if(!((i_regs->wasconst>>sl)&1))
2656 emit_orimm(sl,cinfo[i].imm,tl);
2657 else
2658 emit_movimm(constmap[i][sl]|cinfo[i].imm,tl);
2659 }
2660 }
2661 if(dops[i].opcode==0x0e) { // XORI
2662 if(sl<0) {
2663 emit_xorimm(tl,cinfo[i].imm,tl);
2664 }else{
2665 if(!((i_regs->wasconst>>sl)&1))
2666 emit_xorimm(sl,cinfo[i].imm,tl);
2667 else
2668 emit_movimm(constmap[i][sl]^cinfo[i].imm,tl);
2669 }
2670 }
2671 }
2672 else {
2673 emit_movimm(cinfo[i].imm,tl);
2674 }
2675 }
2676 }
2677 }
2678 }
2679}
2680
2681static void shiftimm_assemble(int i, const struct regstat *i_regs)
2682{
2683 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2684 {
2685 if(dops[i].rt1) {
2686 signed char s,t;
2687 t=get_reg_w(i_regs->regmap, dops[i].rt1);
2688 s=get_reg(i_regs->regmap,dops[i].rs1);
2689 //assert(t>=0);
2690 if(t>=0&&!((i_regs->isconst>>t)&1)){
2691 if(dops[i].rs1==0)
2692 {
2693 emit_zeroreg(t);
2694 }
2695 else
2696 {
2697 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2698 if(cinfo[i].imm) {
2699 if(dops[i].opcode2==0) // SLL
2700 {
2701 emit_shlimm(s<0?t:s,cinfo[i].imm,t);
2702 }
2703 if(dops[i].opcode2==2) // SRL
2704 {
2705 emit_shrimm(s<0?t:s,cinfo[i].imm,t);
2706 }
2707 if(dops[i].opcode2==3) // SRA
2708 {
2709 emit_sarimm(s<0?t:s,cinfo[i].imm,t);
2710 }
2711 }else{
2712 // Shift by zero
2713 if(s>=0 && s!=t) emit_mov(s,t);
2714 }
2715 }
2716 }
2717 //emit_storereg(dops[i].rt1,t); //DEBUG
2718 }
2719 }
2720 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2721 {
2722 assert(0);
2723 }
2724 if(dops[i].opcode2==0x3c) // DSLL32
2725 {
2726 assert(0);
2727 }
2728 if(dops[i].opcode2==0x3e) // DSRL32
2729 {
2730 assert(0);
2731 }
2732 if(dops[i].opcode2==0x3f) // DSRA32
2733 {
2734 assert(0);
2735 }
2736}
2737
2738#ifndef shift_assemble
2739static void shift_assemble(int i, const struct regstat *i_regs)
2740{
2741 signed char s,t,shift;
2742 if (dops[i].rt1 == 0)
2743 return;
2744 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2745 t = get_reg(i_regs->regmap, dops[i].rt1);
2746 s = get_reg(i_regs->regmap, dops[i].rs1);
2747 shift = get_reg(i_regs->regmap, dops[i].rs2);
2748 if (t < 0)
2749 return;
2750
2751 if(dops[i].rs1==0)
2752 emit_zeroreg(t);
2753 else if(dops[i].rs2==0) {
2754 assert(s>=0);
2755 if(s!=t) emit_mov(s,t);
2756 }
2757 else {
2758 host_tempreg_acquire();
2759 emit_andimm(shift,31,HOST_TEMPREG);
2760 switch(dops[i].opcode2) {
2761 case 4: // SLLV
2762 emit_shl(s,HOST_TEMPREG,t);
2763 break;
2764 case 6: // SRLV
2765 emit_shr(s,HOST_TEMPREG,t);
2766 break;
2767 case 7: // SRAV
2768 emit_sar(s,HOST_TEMPREG,t);
2769 break;
2770 default:
2771 assert(0);
2772 }
2773 host_tempreg_release();
2774 }
2775}
2776
2777#endif
2778
2779enum {
2780 MTYPE_8000 = 0,
2781 MTYPE_8020,
2782 MTYPE_0000,
2783 MTYPE_A000,
2784 MTYPE_1F80,
2785};
2786
2787static int get_ptr_mem_type(u_int a)
2788{
2789 if(a < 0x00200000) {
2790 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2791 // return wrong, must use memhandler for BIOS self-test to pass
2792 // 007 does similar stuff from a00 mirror, weird stuff
2793 return MTYPE_8000;
2794 return MTYPE_0000;
2795 }
2796 if(0x1f800000 <= a && a < 0x1f801000)
2797 return MTYPE_1F80;
2798 if(0x80200000 <= a && a < 0x80800000)
2799 return MTYPE_8020;
2800 if(0xa0000000 <= a && a < 0xa0200000)
2801 return MTYPE_A000;
2802 return MTYPE_8000;
2803}
2804
2805static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2806{
2807 int r = get_reg(i_regs->regmap, ROREG);
2808 if (r < 0 && host_tempreg_free) {
2809 host_tempreg_acquire();
2810 emit_loadreg(ROREG, r = HOST_TEMPREG);
2811 }
2812 if (r < 0)
2813 abort();
2814 return r;
2815}
2816
2817static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2818 int addr, int *offset_reg, int *addr_reg_override, int ccadj_)
2819{
2820 void *jaddr = NULL;
2821 int type = 0;
2822 int mr = dops[i].rs1;
2823 assert(addr >= 0);
2824 *offset_reg = -1;
2825 if(((smrv_strong|smrv_weak)>>mr)&1) {
2826 type=get_ptr_mem_type(smrv[mr]);
2827 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2828 }
2829 else {
2830 // use the mirror we are running on
2831 type=get_ptr_mem_type(start);
2832 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2833 }
2834
2835 if (dops[i].may_except) {
2836 // alignment check
2837 u_int op = dops[i].opcode;
2838 int mask = ((op & 0x37) == 0x21 || op == 0x25) ? 1 : 3; // LH/SH/LHU
2839 void *jaddr2;
2840 emit_testimm(addr, mask);
2841 jaddr2 = out;
2842 emit_jne(0);
2843 add_stub_r(ALIGNMENT_STUB, jaddr2, out, i, addr, i_regs, ccadj_, 0);
2844 }
2845
2846 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2847 host_tempreg_acquire();
2848 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2849 addr=*addr_reg_override=HOST_TEMPREG;
2850 type=0;
2851 }
2852 else if(type==MTYPE_0000) { // RAM 0 mirror
2853 host_tempreg_acquire();
2854 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2855 addr=*addr_reg_override=HOST_TEMPREG;
2856 type=0;
2857 }
2858 else if(type==MTYPE_A000) { // RAM A mirror
2859 host_tempreg_acquire();
2860 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2861 addr=*addr_reg_override=HOST_TEMPREG;
2862 type=0;
2863 }
2864 else if(type==MTYPE_1F80) { // scratchpad
2865 if (psxH == (void *)0x1f800000) {
2866 host_tempreg_acquire();
2867 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2868 emit_cmpimm(HOST_TEMPREG,0x1000);
2869 host_tempreg_release();
2870 jaddr=out;
2871 emit_jc(0);
2872 }
2873 else {
2874 // do the usual RAM check, jump will go to the right handler
2875 type=0;
2876 }
2877 }
2878
2879 if (type == 0) // need ram check
2880 {
2881 emit_cmpimm(addr,RAM_SIZE);
2882 jaddr = out;
2883 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2884 // Hint to branch predictor that the branch is unlikely to be taken
2885 if (dops[i].rs1 >= 28)
2886 emit_jno_unlikely(0);
2887 else
2888 #endif
2889 emit_jno(0);
2890 if (ram_offset != 0)
2891 *offset_reg = get_ro_reg(i_regs, 0);
2892 }
2893
2894 return jaddr;
2895}
2896
2897// return memhandler, or get directly accessable address and return 0
2898static void *get_direct_memhandler(void *table, u_int addr,
2899 enum stub_type type, uintptr_t *addr_host)
2900{
2901 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2902 uintptr_t l1, l2 = 0;
2903 l1 = ((uintptr_t *)table)[addr>>12];
2904 if (!(l1 & msb)) {
2905 uintptr_t v = l1 << 1;
2906 *addr_host = v + addr;
2907 return NULL;
2908 }
2909 else {
2910 l1 <<= 1;
2911 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2912 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2913 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2914 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2915 else
2916 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2917 if (!(l2 & msb)) {
2918 uintptr_t v = l2 << 1;
2919 *addr_host = v + (addr&0xfff);
2920 return NULL;
2921 }
2922 return (void *)(l2 << 1);
2923 }
2924}
2925
2926static u_int get_host_reglist(const signed char *regmap)
2927{
2928 u_int reglist = 0, hr;
2929 for (hr = 0; hr < HOST_REGS; hr++) {
2930 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2931 reglist |= 1 << hr;
2932 }
2933 return reglist;
2934}
2935
2936static u_int reglist_exclude(u_int reglist, int r1, int r2)
2937{
2938 if (r1 >= 0)
2939 reglist &= ~(1u << r1);
2940 if (r2 >= 0)
2941 reglist &= ~(1u << r2);
2942 return reglist;
2943}
2944
2945// find a temp caller-saved register not in reglist (so assumed to be free)
2946static int reglist_find_free(u_int reglist)
2947{
2948 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2949 if (free_regs == 0)
2950 return -1;
2951 return __builtin_ctz(free_regs);
2952}
2953
2954static void do_load_word(int a, int rt, int offset_reg)
2955{
2956 if (offset_reg >= 0)
2957 emit_ldr_dualindexed(offset_reg, a, rt);
2958 else
2959 emit_readword_indexed(0, a, rt);
2960}
2961
2962static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2963{
2964 if (offset_reg < 0) {
2965 emit_writeword_indexed(rt, ofs, a);
2966 return;
2967 }
2968 if (ofs != 0)
2969 emit_addimm(a, ofs, a);
2970 emit_str_dualindexed(offset_reg, a, rt);
2971 if (ofs != 0 && preseve_a)
2972 emit_addimm(a, -ofs, a);
2973}
2974
2975static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2976{
2977 if (offset_reg < 0) {
2978 emit_writehword_indexed(rt, ofs, a);
2979 return;
2980 }
2981 if (ofs != 0)
2982 emit_addimm(a, ofs, a);
2983 emit_strh_dualindexed(offset_reg, a, rt);
2984 if (ofs != 0 && preseve_a)
2985 emit_addimm(a, -ofs, a);
2986}
2987
2988static void do_store_byte(int a, int rt, int offset_reg)
2989{
2990 if (offset_reg >= 0)
2991 emit_strb_dualindexed(offset_reg, a, rt);
2992 else
2993 emit_writebyte_indexed(rt, 0, a);
2994}
2995
2996static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2997{
2998 int addr = cinfo[i].addr;
2999 int s,tl;
3000 int offset;
3001 void *jaddr=0;
3002 int memtarget=0,c=0;
3003 int offset_reg = -1;
3004 int fastio_reg_override = -1;
3005 u_int reglist=get_host_reglist(i_regs->regmap);
3006 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3007 s=get_reg(i_regs->regmap,dops[i].rs1);
3008 offset=cinfo[i].imm;
3009 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3010 if(s>=0) {
3011 c=(i_regs->wasconst>>s)&1;
3012 if (c) {
3013 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3014 }
3015 }
3016 //printf("load_assemble: c=%d\n",c);
3017 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3018 if(tl<0 && ((!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80) || dops[i].rt1==0)) {
3019 // could be FIFO, must perform the read
3020 // ||dummy read
3021 assem_debug("(forced read)\n");
3022 tl = get_reg_temp(i_regs->regmap); // may be == addr
3023 assert(tl>=0);
3024 }
3025 assert(addr >= 0);
3026 if(tl>=0) {
3027 //printf("load_assemble: c=%d\n",c);
3028 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
3029 reglist&=~(1<<tl);
3030 if(!c) {
3031 #ifdef R29_HACK
3032 // Strmnnrmn's speed hack
3033 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
3034 #endif
3035 {
3036 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3037 &offset_reg, &fastio_reg_override, ccadj_);
3038 }
3039 }
3040 else if (ram_offset && memtarget) {
3041 offset_reg = get_ro_reg(i_regs, 0);
3042 }
3043 int dummy=(dops[i].rt1==0)||(tl!=get_reg_w(i_regs->regmap, dops[i].rt1)); // ignore loads to r0 and unneeded reg
3044 switch (dops[i].opcode) {
3045 case 0x20: // LB
3046 if(!c||memtarget) {
3047 if(!dummy) {
3048 int a = addr;
3049 if (fastio_reg_override >= 0)
3050 a = fastio_reg_override;
3051
3052 if (offset_reg >= 0)
3053 emit_ldrsb_dualindexed(offset_reg, a, tl);
3054 else
3055 emit_movsbl_indexed(0, a, tl);
3056 }
3057 if(jaddr)
3058 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3059 }
3060 else
3061 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3062 break;
3063 case 0x21: // LH
3064 if(!c||memtarget) {
3065 if(!dummy) {
3066 int a = addr;
3067 if (fastio_reg_override >= 0)
3068 a = fastio_reg_override;
3069 if (offset_reg >= 0)
3070 emit_ldrsh_dualindexed(offset_reg, a, tl);
3071 else
3072 emit_movswl_indexed(0, a, tl);
3073 }
3074 if(jaddr)
3075 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3076 }
3077 else
3078 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3079 break;
3080 case 0x23: // LW
3081 if(!c||memtarget) {
3082 if(!dummy) {
3083 int a = addr;
3084 if (fastio_reg_override >= 0)
3085 a = fastio_reg_override;
3086 do_load_word(a, tl, offset_reg);
3087 }
3088 if(jaddr)
3089 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3090 }
3091 else
3092 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3093 break;
3094 case 0x24: // LBU
3095 if(!c||memtarget) {
3096 if(!dummy) {
3097 int a = addr;
3098 if (fastio_reg_override >= 0)
3099 a = fastio_reg_override;
3100
3101 if (offset_reg >= 0)
3102 emit_ldrb_dualindexed(offset_reg, a, tl);
3103 else
3104 emit_movzbl_indexed(0, a, tl);
3105 }
3106 if(jaddr)
3107 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3108 }
3109 else
3110 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3111 break;
3112 case 0x25: // LHU
3113 if(!c||memtarget) {
3114 if(!dummy) {
3115 int a = addr;
3116 if (fastio_reg_override >= 0)
3117 a = fastio_reg_override;
3118 if (offset_reg >= 0)
3119 emit_ldrh_dualindexed(offset_reg, a, tl);
3120 else
3121 emit_movzwl_indexed(0, a, tl);
3122 }
3123 if(jaddr)
3124 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3125 }
3126 else
3127 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
3128 break;
3129 default:
3130 assert(0);
3131 }
3132 } // tl >= 0
3133 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3134 host_tempreg_release();
3135}
3136
3137#ifndef loadlr_assemble
3138static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3139{
3140 int addr = cinfo[i].addr;
3141 int s,tl,temp,temp2;
3142 int offset;
3143 void *jaddr=0;
3144 int memtarget=0,c=0;
3145 int offset_reg = -1;
3146 int fastio_reg_override = -1;
3147 u_int reglist=get_host_reglist(i_regs->regmap);
3148 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3149 s=get_reg(i_regs->regmap,dops[i].rs1);
3150 temp=get_reg_temp(i_regs->regmap);
3151 temp2=get_reg(i_regs->regmap,FTEMP);
3152 offset=cinfo[i].imm;
3153 reglist|=1<<temp;
3154 assert(addr >= 0);
3155 if(s>=0) {
3156 c=(i_regs->wasconst>>s)&1;
3157 if(c) {
3158 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3159 }
3160 }
3161 if(!c) {
3162 emit_shlimm(addr,3,temp);
3163 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3164 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3165 }else{
3166 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3167 }
3168 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3169 &offset_reg, &fastio_reg_override, ccadj_);
3170 }
3171 else {
3172 if (ram_offset && memtarget) {
3173 offset_reg = get_ro_reg(i_regs, 0);
3174 }
3175 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3176 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3177 }else{
3178 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3179 }
3180 }
3181 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3182 if(!c||memtarget) {
3183 int a = temp2;
3184 if (fastio_reg_override >= 0)
3185 a = fastio_reg_override;
3186 do_load_word(a, temp2, offset_reg);
3187 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3188 host_tempreg_release();
3189 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3190 }
3191 else
3192 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3193 if(dops[i].rt1) {
3194 assert(tl>=0);
3195 emit_andimm(temp,24,temp);
3196 if (dops[i].opcode==0x22) // LWL
3197 emit_xorimm(temp,24,temp);
3198 host_tempreg_acquire();
3199 emit_movimm(-1,HOST_TEMPREG);
3200 if (dops[i].opcode==0x26) {
3201 emit_shr(temp2,temp,temp2);
3202 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3203 }else{
3204 emit_shl(temp2,temp,temp2);
3205 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3206 }
3207 host_tempreg_release();
3208 emit_or(temp2,tl,tl);
3209 }
3210 //emit_storereg(dops[i].rt1,tl); // DEBUG
3211 }
3212 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3213 assert(0);
3214 }
3215}
3216#endif
3217
3218static void do_invstub(int n)
3219{
3220 literal_pool(20);
3221 assem_debug("do_invstub\n");
3222 u_int reglist = stubs[n].a;
3223 u_int addrr = stubs[n].b;
3224 int ofs_start = stubs[n].c;
3225 int ofs_end = stubs[n].d;
3226 int len = ofs_end - ofs_start;
3227 u_int rightr = 0;
3228
3229 set_jump_target(stubs[n].addr, out);
3230 save_regs(reglist);
3231 if (addrr != 0 || ofs_start != 0)
3232 emit_addimm(addrr, ofs_start, 0);
3233 emit_readword(&inv_code_start, 2);
3234 emit_readword(&inv_code_end, 3);
3235 if (len != 0)
3236 emit_addimm(0, len + 4, (rightr = 1));
3237 emit_cmp(0, 2);
3238 emit_cmpcs(3, rightr);
3239 void *jaddr = out;
3240 emit_jc(0);
3241 void *func = (len != 0)
3242 ? (void *)ndrc_write_invalidate_many
3243 : (void *)ndrc_write_invalidate_one;
3244 emit_far_call(func);
3245 set_jump_target(jaddr, out);
3246 restore_regs(reglist);
3247 emit_jmp(stubs[n].retaddr);
3248}
3249
3250static void do_store_smc_check(int i, const struct regstat *i_regs, u_int reglist, int addr)
3251{
3252 if (HACK_ENABLED(NDHACK_NO_SMC_CHECK))
3253 return;
3254 // this can't be used any more since we started to check exact
3255 // block boundaries in invalidate_range()
3256 //if (i_regs->waswritten & (1<<dops[i].rs1))
3257 // return;
3258 // (naively) assume nobody will run code from stack
3259 if (dops[i].rs1 == 29)
3260 return;
3261
3262 int j, imm_maxdiff = 32, imm_min = cinfo[i].imm, imm_max = cinfo[i].imm, count = 1;
3263 if (i < slen - 1 && dops[i+1].is_store && dops[i+1].rs1 == dops[i].rs1
3264 && abs(cinfo[i+1].imm - cinfo[i].imm) <= imm_maxdiff)
3265 return;
3266 for (j = i - 1; j >= 0; j--) {
3267 if (!dops[j].is_store || dops[j].rs1 != dops[i].rs1
3268 || abs(cinfo[j].imm - cinfo[j+1].imm) > imm_maxdiff)
3269 break;
3270 count++;
3271 if (imm_min > cinfo[j].imm)
3272 imm_min = cinfo[j].imm;
3273 if (imm_max < cinfo[j].imm)
3274 imm_max = cinfo[j].imm;
3275 }
3276#if defined(HOST_IMM8)
3277 int ir = get_reg(i_regs->regmap, INVCP);
3278 assert(ir >= 0);
3279 host_tempreg_acquire();
3280 emit_ldrb_indexedsr12_reg(ir, addr, HOST_TEMPREG);
3281#else
3282 emit_cmpmem_indexedsr12_imm(invalid_code, addr, 1);
3283 #error not handled
3284#endif
3285#ifdef INVALIDATE_USE_COND_CALL
3286 if (count == 1) {
3287 emit_cmpimm(HOST_TEMPREG, 1);
3288 emit_callne(invalidate_addr_reg[addr]);
3289 host_tempreg_release();
3290 return;
3291 }
3292#endif
3293 void *jaddr = emit_cbz(HOST_TEMPREG, 0);
3294 host_tempreg_release();
3295 imm_min -= cinfo[i].imm;
3296 imm_max -= cinfo[i].imm;
3297 add_stub(INVCODE_STUB, jaddr, out, reglist|(1<<HOST_CCREG),
3298 addr, imm_min, imm_max, 0);
3299}
3300
3301static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3302{
3303 int s,tl;
3304 int addr = cinfo[i].addr;
3305 int offset;
3306 void *jaddr=0;
3307 enum stub_type type=0;
3308 int memtarget=0,c=0;
3309 int offset_reg = -1;
3310 int fastio_reg_override = -1;
3311 u_int reglist=get_host_reglist(i_regs->regmap);
3312 tl=get_reg(i_regs->regmap,dops[i].rs2);
3313 s=get_reg(i_regs->regmap,dops[i].rs1);
3314 offset=cinfo[i].imm;
3315 if(s>=0) {
3316 c=(i_regs->wasconst>>s)&1;
3317 if(c) {
3318 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3319 }
3320 }
3321 assert(tl>=0);
3322 assert(addr >= 0);
3323 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3324 if (!c) {
3325 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3326 &offset_reg, &fastio_reg_override, ccadj_);
3327 }
3328 else if (ram_offset && memtarget) {
3329 offset_reg = get_ro_reg(i_regs, 0);
3330 }
3331
3332 switch (dops[i].opcode) {
3333 case 0x28: // SB
3334 if(!c||memtarget) {
3335 int a = addr;
3336 if (fastio_reg_override >= 0)
3337 a = fastio_reg_override;
3338 do_store_byte(a, tl, offset_reg);
3339 }
3340 type = STOREB_STUB;
3341 break;
3342 case 0x29: // SH
3343 if(!c||memtarget) {
3344 int a = addr;
3345 if (fastio_reg_override >= 0)
3346 a = fastio_reg_override;
3347 do_store_hword(a, 0, tl, offset_reg, 1);
3348 }
3349 type = STOREH_STUB;
3350 break;
3351 case 0x2B: // SW
3352 if(!c||memtarget) {
3353 int a = addr;
3354 if (fastio_reg_override >= 0)
3355 a = fastio_reg_override;
3356 do_store_word(a, 0, tl, offset_reg, 1);
3357 }
3358 type = STOREW_STUB;
3359 break;
3360 default:
3361 assert(0);
3362 }
3363 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3364 host_tempreg_release();
3365 if(jaddr) {
3366 // PCSX store handlers don't check invcode again
3367 reglist|=1<<addr;
3368 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3369 jaddr=0;
3370 }
3371 {
3372 if(!c||memtarget) {
3373 do_store_smc_check(i, i_regs, reglist, addr);
3374 }
3375 }
3376 u_int addr_val=constmap[i][s]+offset;
3377 if(jaddr) {
3378 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3379 } else if(c&&!memtarget) {
3380 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3381 }
3382 // basic current block modification detection..
3383 // not looking back as that should be in mips cache already
3384 // (see Spyro2 title->attract mode)
3385 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3386 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3387 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3388 if(i_regs->regmap==regs[i].regmap) {
3389 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3390 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3391 emit_movimm(start+i*4+4,0);
3392 emit_writeword(0,&pcaddr);
3393 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3394 emit_far_call(ndrc_get_addr_ht);
3395 emit_jmpreg(0);
3396 }
3397 }
3398}
3399
3400static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3401{
3402 int addr = cinfo[i].addr;
3403 int s,tl;
3404 int offset;
3405 void *jaddr=0;
3406 void *case1, *case23, *case3;
3407 void *done0, *done1, *done2;
3408 int memtarget=0,c=0;
3409 int offset_reg = -1;
3410 u_int reglist=get_host_reglist(i_regs->regmap);
3411 tl=get_reg(i_regs->regmap,dops[i].rs2);
3412 s=get_reg(i_regs->regmap,dops[i].rs1);
3413 offset=cinfo[i].imm;
3414 if(s>=0) {
3415 c=(i_regs->isconst>>s)&1;
3416 if(c) {
3417 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3418 }
3419 }
3420 assert(tl>=0);
3421 assert(addr >= 0);
3422 if(!c) {
3423 emit_cmpimm(addr, RAM_SIZE);
3424 jaddr=out;
3425 emit_jno(0);
3426 }
3427 else
3428 {
3429 if(!memtarget||!dops[i].rs1) {
3430 jaddr=out;
3431 emit_jmp(0);
3432 }
3433 }
3434 if (ram_offset)
3435 offset_reg = get_ro_reg(i_regs, 0);
3436
3437 emit_testimm(addr,2);
3438 case23=out;
3439 emit_jne(0);
3440 emit_testimm(addr,1);
3441 case1=out;
3442 emit_jne(0);
3443 // 0
3444 if (dops[i].opcode == 0x2A) { // SWL
3445 // Write msb into least significant byte
3446 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3447 do_store_byte(addr, tl, offset_reg);
3448 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3449 }
3450 else if (dops[i].opcode == 0x2E) { // SWR
3451 // Write entire word
3452 do_store_word(addr, 0, tl, offset_reg, 1);
3453 }
3454 done0 = out;
3455 emit_jmp(0);
3456 // 1
3457 set_jump_target(case1, out);
3458 if (dops[i].opcode == 0x2A) { // SWL
3459 // Write two msb into two least significant bytes
3460 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3461 do_store_hword(addr, -1, tl, offset_reg, 0);
3462 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3463 }
3464 else if (dops[i].opcode == 0x2E) { // SWR
3465 // Write 3 lsb into three most significant bytes
3466 do_store_byte(addr, tl, offset_reg);
3467 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3468 do_store_hword(addr, 1, tl, offset_reg, 0);
3469 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3470 }
3471 done1=out;
3472 emit_jmp(0);
3473 // 2,3
3474 set_jump_target(case23, out);
3475 emit_testimm(addr,1);
3476 case3 = out;
3477 emit_jne(0);
3478 // 2
3479 if (dops[i].opcode==0x2A) { // SWL
3480 // Write 3 msb into three least significant bytes
3481 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3482 do_store_hword(addr, -2, tl, offset_reg, 1);
3483 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3484 do_store_byte(addr, tl, offset_reg);
3485 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3486 }
3487 else if (dops[i].opcode == 0x2E) { // SWR
3488 // Write two lsb into two most significant bytes
3489 do_store_hword(addr, 0, tl, offset_reg, 1);
3490 }
3491 done2 = out;
3492 emit_jmp(0);
3493 // 3
3494 set_jump_target(case3, out);
3495 if (dops[i].opcode == 0x2A) { // SWL
3496 do_store_word(addr, -3, tl, offset_reg, 0);
3497 }
3498 else if (dops[i].opcode == 0x2E) { // SWR
3499 do_store_byte(addr, tl, offset_reg);
3500 }
3501 set_jump_target(done0, out);
3502 set_jump_target(done1, out);
3503 set_jump_target(done2, out);
3504 if (offset_reg == HOST_TEMPREG)
3505 host_tempreg_release();
3506 if(!c||!memtarget)
3507 add_stub_r(STORELR_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3508 do_store_smc_check(i, i_regs, reglist, addr);
3509}
3510
3511static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3512{
3513 if(dops[i].opcode2==0) // MFC0
3514 {
3515 signed char t=get_reg_w(i_regs->regmap, dops[i].rt1);
3516 u_int copr=(source[i]>>11)&0x1f;
3517 if(t>=0&&dops[i].rt1!=0) {
3518 emit_readword(&reg_cop0[copr],t);
3519 }
3520 }
3521 else if(dops[i].opcode2==4) // MTC0
3522 {
3523 int s = get_reg(i_regs->regmap, dops[i].rs1);
3524 int cc = get_reg(i_regs->regmap, CCREG);
3525 char copr=(source[i]>>11)&0x1f;
3526 assert(s>=0);
3527 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3528 if (copr == 12 || copr == 13) {
3529 emit_readword(&last_count,HOST_TEMPREG);
3530 if (cc != HOST_CCREG)
3531 emit_loadreg(CCREG, HOST_CCREG);
3532 emit_add(HOST_CCREG, HOST_TEMPREG, HOST_CCREG);
3533 emit_addimm(HOST_CCREG, ccadj_ + 2, HOST_CCREG);
3534 emit_writeword(HOST_CCREG, &psxRegs.cycle);
3535 if (is_delayslot) {
3536 // burn cycles to cause cc_interrupt, which will
3537 // reschedule next_interupt. Relies on CCREG from above.
3538 assem_debug("MTC0 DS %d\n", copr);
3539 emit_writeword(HOST_CCREG,&last_count);
3540 emit_movimm(0,HOST_CCREG);
3541 emit_storereg(CCREG,HOST_CCREG);
3542 emit_loadreg(dops[i].rs1,1);
3543 emit_movimm(copr,0);
3544 emit_far_call(pcsx_mtc0_ds);
3545 emit_loadreg(dops[i].rs1,s);
3546 return;
3547 }
3548 emit_movimm(start+i*4+4,HOST_TEMPREG);
3549 emit_writeword(HOST_TEMPREG,&pcaddr);
3550 emit_movimm(0,HOST_TEMPREG);
3551 emit_writeword(HOST_TEMPREG,&pending_exception);
3552 }
3553 if( s != 1)
3554 emit_mov(s, 1);
3555 emit_movimm(copr, 0);
3556 emit_far_call(pcsx_mtc0);
3557 if (copr == 12 || copr == 13) {
3558 emit_readword(&psxRegs.cycle,HOST_CCREG);
3559 emit_readword(&last_count,HOST_TEMPREG);
3560 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3561 //emit_writeword(HOST_TEMPREG,&last_count);
3562 assert(!is_delayslot);
3563 emit_readword(&pending_exception,HOST_TEMPREG);
3564 emit_test(HOST_TEMPREG,HOST_TEMPREG);
3565 void *jaddr = out;
3566 emit_jeq(0);
3567 emit_readword(&pcaddr, 0);
3568 emit_far_call(ndrc_get_addr_ht);
3569 emit_jmpreg(0);
3570 set_jump_target(jaddr, out);
3571 emit_addimm(HOST_CCREG, -ccadj_ - 2, HOST_CCREG);
3572 if (cc != HOST_CCREG)
3573 emit_storereg(CCREG, HOST_CCREG);
3574 }
3575 emit_loadreg(dops[i].rs1,s);
3576 }
3577}
3578
3579static void rfe_assemble(int i, const struct regstat *i_regs)
3580{
3581 emit_readword(&psxRegs.CP0.n.SR, 0);
3582 emit_andimm(0, 0x3c, 1);
3583 emit_andimm(0, ~0xf, 0);
3584 emit_orrshr_imm(1, 2, 0);
3585 emit_writeword(0, &psxRegs.CP0.n.SR);
3586}
3587
3588static int cop2_is_stalling_op(int i, int *cycles)
3589{
3590 if (dops[i].opcode == 0x3a) { // SWC2
3591 *cycles = 0;
3592 return 1;
3593 }
3594 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3595 *cycles = 0;
3596 return 1;
3597 }
3598 if (dops[i].itype == C2OP) {
3599 *cycles = gte_cycletab[source[i] & 0x3f];
3600 return 1;
3601 }
3602 // ... what about MTC2/CTC2/LWC2?
3603 return 0;
3604}
3605
3606#if 0
3607static void log_gte_stall(int stall, u_int cycle)
3608{
3609 if ((u_int)stall <= 44)
3610 printf("x stall %2d %u\n", stall, cycle + last_count);
3611}
3612
3613static void emit_log_gte_stall(int i, int stall, u_int reglist)
3614{
3615 save_regs(reglist);
3616 if (stall > 0)
3617 emit_movimm(stall, 0);
3618 else
3619 emit_mov(HOST_TEMPREG, 0);
3620 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3621 emit_far_call(log_gte_stall);
3622 restore_regs(reglist);
3623}
3624#endif
3625
3626static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3627{
3628 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3629 int rtmp = reglist_find_free(reglist);
3630
3631 if (HACK_ENABLED(NDHACK_NO_STALLS))
3632 return;
3633 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3634 // happens occasionally... cc evicted? Don't bother then
3635 //printf("no cc %08x\n", start + i*4);
3636 return;
3637 }
3638 if (!dops[i].bt) {
3639 for (j = i - 1; j >= 0; j--) {
3640 //if (dops[j].is_ds) break;
3641 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3642 break;
3643 if (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj)
3644 break;
3645 }
3646 j = max(j, 0);
3647 }
3648 cycles_passed = cinfo[i].ccadj - cinfo[j].ccadj;
3649 if (other_gte_op_cycles >= 0)
3650 stall = other_gte_op_cycles - cycles_passed;
3651 else if (cycles_passed >= 44)
3652 stall = 0; // can't stall
3653 if (stall == -MAXBLOCK && rtmp >= 0) {
3654 // unknown stall, do the expensive runtime check
3655 assem_debug("; cop2_do_stall_check\n");
3656#if 0 // too slow
3657 save_regs(reglist);
3658 emit_movimm(gte_cycletab[op], 0);
3659 emit_addimm(HOST_CCREG, cinfo[i].ccadj, 1);
3660 emit_far_call(call_gteStall);
3661 restore_regs(reglist);
3662#else
3663 host_tempreg_acquire();
3664 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3665 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3666 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3667 emit_cmpimm(HOST_TEMPREG, 44);
3668 emit_cmovb_reg(rtmp, HOST_CCREG);
3669 //emit_log_gte_stall(i, 0, reglist);
3670 host_tempreg_release();
3671#endif
3672 }
3673 else if (stall > 0) {
3674 //emit_log_gte_stall(i, stall, reglist);
3675 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3676 }
3677
3678 // save gteBusyCycle, if needed
3679 if (gte_cycletab[op] == 0)
3680 return;
3681 other_gte_op_cycles = -1;
3682 for (j = i + 1; j < slen; j++) {
3683 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3684 break;
3685 if (dops[j].is_jump) {
3686 // check ds
3687 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3688 j++;
3689 break;
3690 }
3691 }
3692 if (other_gte_op_cycles >= 0)
3693 // will handle stall when assembling that op
3694 return;
3695 cycles_passed = cinfo[min(j, slen -1)].ccadj - cinfo[i].ccadj;
3696 if (cycles_passed >= 44)
3697 return;
3698 assem_debug("; save gteBusyCycle\n");
3699 host_tempreg_acquire();
3700#if 0
3701 emit_readword(&last_count, HOST_TEMPREG);
3702 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3703 emit_addimm(HOST_TEMPREG, cinfo[i].ccadj, HOST_TEMPREG);
3704 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3705 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3706#else
3707 emit_addimm(HOST_CCREG, cinfo[i].ccadj + gte_cycletab[op], HOST_TEMPREG);
3708 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3709#endif
3710 host_tempreg_release();
3711}
3712
3713static int is_mflohi(int i)
3714{
3715 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3716}
3717
3718static int check_multdiv(int i, int *cycles)
3719{
3720 if (dops[i].itype != MULTDIV)
3721 return 0;
3722 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3723 *cycles = 11; // approx from 7 11 14
3724 else
3725 *cycles = 37;
3726 return 1;
3727}
3728
3729static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3730{
3731 int j, found = 0, c = 0;
3732 if (HACK_ENABLED(NDHACK_NO_STALLS))
3733 return;
3734 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3735 // happens occasionally... cc evicted? Don't bother then
3736 return;
3737 }
3738 for (j = i + 1; j < slen; j++) {
3739 if (dops[j].bt)
3740 break;
3741 if ((found = is_mflohi(j)))
3742 break;
3743 if (dops[j].is_jump) {
3744 // check ds
3745 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3746 j++;
3747 break;
3748 }
3749 }
3750 if (found)
3751 // handle all in multdiv_do_stall()
3752 return;
3753 check_multdiv(i, &c);
3754 assert(c > 0);
3755 assem_debug("; muldiv prepare stall %d\n", c);
3756 host_tempreg_acquire();
3757 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3758 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3759 host_tempreg_release();
3760}
3761
3762static void multdiv_do_stall(int i, const struct regstat *i_regs)
3763{
3764 int j, known_cycles = 0;
3765 u_int reglist = get_host_reglist(i_regs->regmap);
3766 int rtmp = get_reg_temp(i_regs->regmap);
3767 if (rtmp < 0)
3768 rtmp = reglist_find_free(reglist);
3769 if (HACK_ENABLED(NDHACK_NO_STALLS))
3770 return;
3771 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3772 // happens occasionally... cc evicted? Don't bother then
3773 //printf("no cc/rtmp %08x\n", start + i*4);
3774 return;
3775 }
3776 if (!dops[i].bt) {
3777 for (j = i - 1; j >= 0; j--) {
3778 if (dops[j].is_ds) break;
3779 if (check_multdiv(j, &known_cycles))
3780 break;
3781 if (is_mflohi(j))
3782 // already handled by this op
3783 return;
3784 if (dops[j].bt || (j > 0 && cinfo[j - 1].ccadj > cinfo[j].ccadj))
3785 break;
3786 }
3787 j = max(j, 0);
3788 }
3789 if (known_cycles > 0) {
3790 known_cycles -= cinfo[i].ccadj - cinfo[j].ccadj;
3791 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3792 if (known_cycles > 0)
3793 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3794 return;
3795 }
3796 assem_debug("; muldiv stall unresolved\n");
3797 host_tempreg_acquire();
3798 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3799 emit_addimm(rtmp, -cinfo[i].ccadj, rtmp);
3800 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3801 emit_cmpimm(HOST_TEMPREG, 37);
3802 emit_cmovb_reg(rtmp, HOST_CCREG);
3803 //emit_log_gte_stall(i, 0, reglist);
3804 host_tempreg_release();
3805}
3806
3807static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3808{
3809 switch (copr) {
3810 case 1:
3811 case 3:
3812 case 5:
3813 case 8:
3814 case 9:
3815 case 10:
3816 case 11:
3817 emit_readword(&reg_cop2d[copr],tl);
3818 emit_signextend16(tl,tl);
3819 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3820 break;
3821 case 7:
3822 case 16:
3823 case 17:
3824 case 18:
3825 case 19:
3826 emit_readword(&reg_cop2d[copr],tl);
3827 emit_andimm(tl,0xffff,tl);
3828 emit_writeword(tl,&reg_cop2d[copr]);
3829 break;
3830 case 15:
3831 emit_readword(&reg_cop2d[14],tl); // SXY2
3832 emit_writeword(tl,&reg_cop2d[copr]);
3833 break;
3834 case 28:
3835 case 29:
3836 c2op_mfc2_29_assemble(tl,temp);
3837 break;
3838 default:
3839 emit_readword(&reg_cop2d[copr],tl);
3840 break;
3841 }
3842}
3843
3844static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3845{
3846 switch (copr) {
3847 case 15:
3848 emit_readword(&reg_cop2d[13],temp); // SXY1
3849 emit_writeword(sl,&reg_cop2d[copr]);
3850 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3851 emit_readword(&reg_cop2d[14],temp); // SXY2
3852 emit_writeword(sl,&reg_cop2d[14]);
3853 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3854 break;
3855 case 28:
3856 emit_andimm(sl,0x001f,temp);
3857 emit_shlimm(temp,7,temp);
3858 emit_writeword(temp,&reg_cop2d[9]);
3859 emit_andimm(sl,0x03e0,temp);
3860 emit_shlimm(temp,2,temp);
3861 emit_writeword(temp,&reg_cop2d[10]);
3862 emit_andimm(sl,0x7c00,temp);
3863 emit_shrimm(temp,3,temp);
3864 emit_writeword(temp,&reg_cop2d[11]);
3865 emit_writeword(sl,&reg_cop2d[28]);
3866 break;
3867 case 30:
3868 emit_xorsar_imm(sl,sl,31,temp);
3869#if defined(HAVE_ARMV5) || defined(__aarch64__)
3870 emit_clz(temp,temp);
3871#else
3872 emit_movs(temp,HOST_TEMPREG);
3873 emit_movimm(0,temp);
3874 emit_jeq((int)out+4*4);
3875 emit_addpl_imm(temp,1,temp);
3876 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3877 emit_jns((int)out-2*4);
3878#endif
3879 emit_writeword(sl,&reg_cop2d[30]);
3880 emit_writeword(temp,&reg_cop2d[31]);
3881 break;
3882 case 31:
3883 break;
3884 default:
3885 emit_writeword(sl,&reg_cop2d[copr]);
3886 break;
3887 }
3888}
3889
3890static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3891{
3892 int s,tl;
3893 int ar;
3894 int offset;
3895 int memtarget=0,c=0;
3896 void *jaddr2=NULL;
3897 enum stub_type type;
3898 int offset_reg = -1;
3899 int fastio_reg_override = -1;
3900 u_int reglist=get_host_reglist(i_regs->regmap);
3901 u_int copr=(source[i]>>16)&0x1f;
3902 s=get_reg(i_regs->regmap,dops[i].rs1);
3903 tl=get_reg(i_regs->regmap,FTEMP);
3904 offset=cinfo[i].imm;
3905 assert(tl>=0);
3906
3907 if(i_regs->regmap[HOST_CCREG]==CCREG)
3908 reglist&=~(1<<HOST_CCREG);
3909
3910 // get the address
3911 ar = cinfo[i].addr;
3912 assert(ar >= 0);
3913 if (dops[i].opcode==0x3a) { // SWC2
3914 reglist |= 1<<ar;
3915 }
3916 if(s>=0) c=(i_regs->wasconst>>s)&1;
3917 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3918
3919 cop2_do_stall_check(0, i, i_regs, reglist);
3920
3921 if (dops[i].opcode==0x3a) { // SWC2
3922 cop2_get_dreg(copr,tl,-1);
3923 type=STOREW_STUB;
3924 }
3925 else
3926 type=LOADW_STUB;
3927
3928 if(c&&!memtarget) {
3929 jaddr2=out;
3930 emit_jmp(0); // inline_readstub/inline_writestub?
3931 }
3932 else {
3933 if(!c) {
3934 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3935 &offset_reg, &fastio_reg_override, ccadj_);
3936 }
3937 else if (ram_offset && memtarget) {
3938 offset_reg = get_ro_reg(i_regs, 0);
3939 }
3940 switch (dops[i].opcode) {
3941 case 0x32: { // LWC2
3942 int a = ar;
3943 if (fastio_reg_override >= 0)
3944 a = fastio_reg_override;
3945 do_load_word(a, tl, offset_reg);
3946 break;
3947 }
3948 case 0x3a: { // SWC2
3949 #ifdef DESTRUCTIVE_SHIFT
3950 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3951 #endif
3952 int a = ar;
3953 if (fastio_reg_override >= 0)
3954 a = fastio_reg_override;
3955 do_store_word(a, 0, tl, offset_reg, 1);
3956 break;
3957 }
3958 default:
3959 assert(0);
3960 }
3961 }
3962 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3963 host_tempreg_release();
3964 if(jaddr2)
3965 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3966 if(dops[i].opcode==0x3a) // SWC2
3967 do_store_smc_check(i, i_regs, reglist, ar);
3968 if (dops[i].opcode==0x32) { // LWC2
3969 host_tempreg_acquire();
3970 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3971 host_tempreg_release();
3972 }
3973}
3974
3975static void cop2_assemble(int i, const struct regstat *i_regs)
3976{
3977 u_int copr = (source[i]>>11) & 0x1f;
3978 signed char temp = get_reg_temp(i_regs->regmap);
3979
3980 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3981 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3982 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3983 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3984 reglist = reglist_exclude(reglist, tl, -1);
3985 }
3986 cop2_do_stall_check(0, i, i_regs, reglist);
3987 }
3988 if (dops[i].opcode2==0) { // MFC2
3989 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
3990 if(tl>=0&&dops[i].rt1!=0)
3991 cop2_get_dreg(copr,tl,temp);
3992 }
3993 else if (dops[i].opcode2==4) { // MTC2
3994 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3995 cop2_put_dreg(copr,sl,temp);
3996 }
3997 else if (dops[i].opcode2==2) // CFC2
3998 {
3999 signed char tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4000 if(tl>=0&&dops[i].rt1!=0)
4001 emit_readword(&reg_cop2c[copr],tl);
4002 }
4003 else if (dops[i].opcode2==6) // CTC2
4004 {
4005 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
4006 switch(copr) {
4007 case 4:
4008 case 12:
4009 case 20:
4010 case 26:
4011 case 27:
4012 case 29:
4013 case 30:
4014 emit_signextend16(sl,temp);
4015 break;
4016 case 31:
4017 c2op_ctc2_31_assemble(sl,temp);
4018 break;
4019 default:
4020 temp=sl;
4021 break;
4022 }
4023 emit_writeword(temp,&reg_cop2c[copr]);
4024 assert(sl>=0);
4025 }
4026}
4027
4028static void do_unalignedwritestub(int n)
4029{
4030 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
4031 literal_pool(256);
4032 set_jump_target(stubs[n].addr, out);
4033
4034 int i=stubs[n].a;
4035 struct regstat *i_regs=(struct regstat *)stubs[n].c;
4036 int addr=stubs[n].b;
4037 u_int reglist=stubs[n].e;
4038 signed char *i_regmap=i_regs->regmap;
4039 int temp2=get_reg(i_regmap,FTEMP);
4040 int rt;
4041 rt=get_reg(i_regmap,dops[i].rs2);
4042 assert(rt>=0);
4043 assert(addr>=0);
4044 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
4045 reglist|=(1<<addr);
4046 reglist&=~(1<<temp2);
4047
4048 // don't bother with it and call write handler
4049 save_regs(reglist);
4050 pass_args(addr,rt);
4051 int cc=get_reg(i_regmap,CCREG);
4052 if(cc<0)
4053 emit_loadreg(CCREG,2);
4054 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
4055 emit_movimm(start + i*4,3);
4056 emit_writeword(3,&psxRegs.pc);
4057 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
4058 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
4059 if(cc<0)
4060 emit_storereg(CCREG,2);
4061 restore_regs(reglist);
4062 emit_jmp(stubs[n].retaddr); // return address
4063}
4064
4065static void do_overflowstub(int n)
4066{
4067 assem_debug("do_overflowstub %x\n", start + (u_int)stubs[n].a * 4);
4068 literal_pool(24);
4069 int i = stubs[n].a;
4070 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4071 int ccadj = stubs[n].d;
4072 set_jump_target(stubs[n].addr, out);
4073 wb_dirtys(regs[i].regmap, regs[i].dirty);
4074 exception_assemble(i, i_regs, ccadj);
4075}
4076
4077static void do_alignmentstub(int n)
4078{
4079 assem_debug("do_alignmentstub %x\n", start + (u_int)stubs[n].a * 4);
4080 literal_pool(24);
4081 int i = stubs[n].a;
4082 struct regstat *i_regs = (struct regstat *)stubs[n].c;
4083 int ccadj = stubs[n].d;
4084 int is_store = dops[i].itype == STORE || dops[i].opcode == 0x3A; // SWC2
4085 int cause = (dops[i].opcode & 3) << 28;
4086 cause |= is_store ? (R3000E_AdES << 2) : (R3000E_AdEL << 2);
4087 set_jump_target(stubs[n].addr, out);
4088 wb_dirtys(regs[i].regmap, regs[i].dirty);
4089 if (stubs[n].b != 1)
4090 emit_mov(stubs[n].b, 1); // faulting address
4091 emit_movimm(cause, 0);
4092 exception_assemble(i, i_regs, ccadj);
4093}
4094
4095#ifndef multdiv_assemble
4096void multdiv_assemble(int i,struct regstat *i_regs)
4097{
4098 printf("Need multdiv_assemble for this architecture.\n");
4099 abort();
4100}
4101#endif
4102
4103static void mov_assemble(int i, const struct regstat *i_regs)
4104{
4105 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
4106 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
4107 if(dops[i].rt1) {
4108 signed char sl,tl;
4109 tl=get_reg_w(i_regs->regmap, dops[i].rt1);
4110 //assert(tl>=0);
4111 if(tl>=0) {
4112 sl=get_reg(i_regs->regmap,dops[i].rs1);
4113 if(sl>=0) emit_mov(sl,tl);
4114 else emit_loadreg(dops[i].rs1,tl);
4115 }
4116 }
4117 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
4118 multdiv_do_stall(i, i_regs);
4119}
4120
4121// call interpreter, exception handler, things that change pc/regs/cycles ...
4122static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
4123{
4124 signed char ccreg=get_reg(i_regs->regmap,CCREG);
4125 assert(ccreg==HOST_CCREG);
4126 assert(!is_delayslot);
4127 (void)ccreg;
4128
4129 emit_movimm(pc,3); // Get PC
4130 emit_readword(&last_count,2);
4131 emit_writeword(3,&psxRegs.pc);
4132 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
4133 emit_add(2,HOST_CCREG,2);
4134 emit_writeword(2,&psxRegs.cycle);
4135 emit_addimm_ptr(FP,(u_char *)&psxRegs - (u_char *)&dynarec_local,0);
4136 emit_far_call(func);
4137 emit_far_jump(jump_to_new_pc);
4138}
4139
4140static void exception_assemble(int i, const struct regstat *i_regs, int ccadj_)
4141{
4142 // 'break' tends to be littered around to catch things like
4143 // division by 0 and is almost never executed, so don't emit much code here
4144 void *func;
4145 if (dops[i].itype == ALU || dops[i].itype == IMM16)
4146 func = is_delayslot ? jump_overflow_ds : jump_overflow;
4147 else if (dops[i].itype == LOAD || dops[i].itype == STORE)
4148 func = is_delayslot ? jump_addrerror_ds : jump_addrerror;
4149 else if (dops[i].opcode2 == 0x0C)
4150 func = is_delayslot ? jump_syscall_ds : jump_syscall;
4151 else
4152 func = is_delayslot ? jump_break_ds : jump_break;
4153 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) // evicted
4154 emit_loadreg(CCREG, HOST_CCREG);
4155 emit_movimm(start + i*4, 2); // pc
4156 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4157 emit_far_jump(func);
4158}
4159
4160static void hlecall_bad()
4161{
4162 assert(0);
4163}
4164
4165static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4166{
4167 void *hlefunc = hlecall_bad;
4168 uint32_t hleCode = source[i] & 0x03ffffff;
4169 if (hleCode < ARRAY_SIZE(psxHLEt))
4170 hlefunc = psxHLEt[hleCode];
4171
4172 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4173}
4174
4175static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4176{
4177 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4178}
4179
4180static void speculate_mov(int rs,int rt)
4181{
4182 if(rt!=0) {
4183 smrv_strong_next|=1<<rt;
4184 smrv[rt]=smrv[rs];
4185 }
4186}
4187
4188static void speculate_mov_weak(int rs,int rt)
4189{
4190 if(rt!=0) {
4191 smrv_weak_next|=1<<rt;
4192 smrv[rt]=smrv[rs];
4193 }
4194}
4195
4196static void speculate_register_values(int i)
4197{
4198 if(i==0) {
4199 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4200 // gp,sp are likely to stay the same throughout the block
4201 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4202 smrv_weak_next=~smrv_strong_next;
4203 //printf(" llr %08x\n", smrv[4]);
4204 }
4205 smrv_strong=smrv_strong_next;
4206 smrv_weak=smrv_weak_next;
4207 switch(dops[i].itype) {
4208 case ALU:
4209 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4210 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4211 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4212 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4213 else {
4214 smrv_strong_next&=~(1<<dops[i].rt1);
4215 smrv_weak_next&=~(1<<dops[i].rt1);
4216 }
4217 break;
4218 case SHIFTIMM:
4219 smrv_strong_next&=~(1<<dops[i].rt1);
4220 smrv_weak_next&=~(1<<dops[i].rt1);
4221 // fallthrough
4222 case IMM16:
4223 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4224 int hr = get_reg_w(regs[i].regmap, dops[i].rt1);
4225 u_int value;
4226 if(hr>=0) {
4227 if(get_final_value(hr,i,&value))
4228 smrv[dops[i].rt1]=value;
4229 else smrv[dops[i].rt1]=constmap[i][hr];
4230 smrv_strong_next|=1<<dops[i].rt1;
4231 }
4232 }
4233 else {
4234 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4235 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4236 }
4237 break;
4238 case LOAD:
4239 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4240 // special case for BIOS
4241 smrv[dops[i].rt1]=0xa0000000;
4242 smrv_strong_next|=1<<dops[i].rt1;
4243 break;
4244 }
4245 // fallthrough
4246 case SHIFT:
4247 case LOADLR:
4248 case MOV:
4249 smrv_strong_next&=~(1<<dops[i].rt1);
4250 smrv_weak_next&=~(1<<dops[i].rt1);
4251 break;
4252 case COP0:
4253 case COP2:
4254 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4255 smrv_strong_next&=~(1<<dops[i].rt1);
4256 smrv_weak_next&=~(1<<dops[i].rt1);
4257 }
4258 break;
4259 case C2LS:
4260 if (dops[i].opcode==0x32) { // LWC2
4261 smrv_strong_next&=~(1<<dops[i].rt1);
4262 smrv_weak_next&=~(1<<dops[i].rt1);
4263 }
4264 break;
4265 }
4266#if 0
4267 int r=4;
4268 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4269 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4270#endif
4271}
4272
4273static void ujump_assemble(int i, const struct regstat *i_regs);
4274static void rjump_assemble(int i, const struct regstat *i_regs);
4275static void cjump_assemble(int i, const struct regstat *i_regs);
4276static void sjump_assemble(int i, const struct regstat *i_regs);
4277
4278static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4279{
4280 int ds = 0;
4281 switch (dops[i].itype) {
4282 case ALU:
4283 alu_assemble(i, i_regs, ccadj_);
4284 break;
4285 case IMM16:
4286 imm16_assemble(i, i_regs, ccadj_);
4287 break;
4288 case SHIFT:
4289 shift_assemble(i, i_regs);
4290 break;
4291 case SHIFTIMM:
4292 shiftimm_assemble(i, i_regs);
4293 break;
4294 case LOAD:
4295 load_assemble(i, i_regs, ccadj_);
4296 break;
4297 case LOADLR:
4298 loadlr_assemble(i, i_regs, ccadj_);
4299 break;
4300 case STORE:
4301 store_assemble(i, i_regs, ccadj_);
4302 break;
4303 case STORELR:
4304 storelr_assemble(i, i_regs, ccadj_);
4305 break;
4306 case COP0:
4307 cop0_assemble(i, i_regs, ccadj_);
4308 break;
4309 case RFE:
4310 rfe_assemble(i, i_regs);
4311 break;
4312 case COP2:
4313 cop2_assemble(i, i_regs);
4314 break;
4315 case C2LS:
4316 c2ls_assemble(i, i_regs, ccadj_);
4317 break;
4318 case C2OP:
4319 c2op_assemble(i, i_regs);
4320 break;
4321 case MULTDIV:
4322 multdiv_assemble(i, i_regs);
4323 multdiv_prepare_stall(i, i_regs, ccadj_);
4324 break;
4325 case MOV:
4326 mov_assemble(i, i_regs);
4327 break;
4328 case SYSCALL:
4329 exception_assemble(i, i_regs, ccadj_);
4330 break;
4331 case HLECALL:
4332 hlecall_assemble(i, i_regs, ccadj_);
4333 break;
4334 case INTCALL:
4335 intcall_assemble(i, i_regs, ccadj_);
4336 break;
4337 case UJUMP:
4338 ujump_assemble(i, i_regs);
4339 ds = 1;
4340 break;
4341 case RJUMP:
4342 rjump_assemble(i, i_regs);
4343 ds = 1;
4344 break;
4345 case CJUMP:
4346 cjump_assemble(i, i_regs);
4347 ds = 1;
4348 break;
4349 case SJUMP:
4350 sjump_assemble(i, i_regs);
4351 ds = 1;
4352 break;
4353 case NOP:
4354 case OTHER:
4355 // not handled, just skip
4356 break;
4357 default:
4358 assert(0);
4359 }
4360 return ds;
4361}
4362
4363static void ds_assemble(int i, const struct regstat *i_regs)
4364{
4365 speculate_register_values(i);
4366 is_delayslot = 1;
4367 switch (dops[i].itype) {
4368 case SYSCALL:
4369 case HLECALL:
4370 case INTCALL:
4371 case UJUMP:
4372 case RJUMP:
4373 case CJUMP:
4374 case SJUMP:
4375 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4376 break;
4377 default:
4378 assemble(i, i_regs, cinfo[i].ccadj);
4379 }
4380 is_delayslot = 0;
4381}
4382
4383// Is the branch target a valid internal jump?
4384static int internal_branch(int addr)
4385{
4386 if(addr&1) return 0; // Indirect (register) jump
4387 if(addr>=start && addr<start+slen*4-4)
4388 {
4389 return 1;
4390 }
4391 return 0;
4392}
4393
4394static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4395{
4396 int hr;
4397 for(hr=0;hr<HOST_REGS;hr++) {
4398 if(hr!=EXCLUDE_REG) {
4399 if(pre[hr]!=entry[hr]) {
4400 if(pre[hr]>=0) {
4401 if((dirty>>hr)&1) {
4402 if(get_reg(entry,pre[hr])<0) {
4403 assert(pre[hr]<64);
4404 if(!((u>>pre[hr])&1))
4405 emit_storereg(pre[hr],hr);
4406 }
4407 }
4408 }
4409 }
4410 }
4411 }
4412 // Move from one register to another (no writeback)
4413 for(hr=0;hr<HOST_REGS;hr++) {
4414 if(hr!=EXCLUDE_REG) {
4415 if(pre[hr]!=entry[hr]) {
4416 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4417 int nr;
4418 if((nr=get_reg(entry,pre[hr]))>=0) {
4419 emit_mov(hr,nr);
4420 }
4421 }
4422 }
4423 }
4424 }
4425}
4426
4427// Load the specified registers
4428// This only loads the registers given as arguments because
4429// we don't want to load things that will be overwritten
4430static inline void load_reg(signed char entry[], signed char regmap[], int rs)
4431{
4432 int hr = get_reg(regmap, rs);
4433 if (hr >= 0 && entry[hr] != regmap[hr])
4434 emit_loadreg(regmap[hr], hr);
4435}
4436
4437static void load_regs(signed char entry[], signed char regmap[], int rs1, int rs2)
4438{
4439 load_reg(entry, regmap, rs1);
4440 if (rs1 != rs2)
4441 load_reg(entry, regmap, rs2);
4442}
4443
4444// Load registers prior to the start of a loop
4445// so that they are not loaded within the loop
4446static void loop_preload(signed char pre[],signed char entry[])
4447{
4448 int hr;
4449 for (hr = 0; hr < HOST_REGS; hr++) {
4450 int r = entry[hr];
4451 if (r >= 0 && pre[hr] != r && get_reg(pre, r) < 0) {
4452 assem_debug("loop preload:\n");
4453 if (r < TEMPREG)
4454 emit_loadreg(r, hr);
4455 }
4456 }
4457}
4458
4459// Generate address for load/store instruction
4460// goes to AGEN (or temp) for writes, FTEMP for LOADLR and cop1/2 loads
4461// AGEN is assigned by pass5b_preallocate2
4462static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4463{
4464 if (dops[i].is_load || dops[i].is_store) {
4465 int ra = -1;
4466 int agr = AGEN1 + (i&1);
4467 if(dops[i].itype==LOAD) {
4468 if (!dops[i].may_except)
4469 ra = get_reg_w(i_regs->regmap, dops[i].rt1); // reuse dest for agen
4470 if (ra < 0)
4471 ra = get_reg_temp(i_regs->regmap);
4472 }
4473 if(dops[i].itype==LOADLR) {
4474 ra=get_reg(i_regs->regmap,FTEMP);
4475 }
4476 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4477 ra=get_reg(i_regs->regmap,agr);
4478 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4479 }
4480 if(dops[i].itype==C2LS) {
4481 if (dops[i].opcode == 0x32) // LWC2
4482 ra=get_reg(i_regs->regmap,FTEMP);
4483 else { // SWC2
4484 ra=get_reg(i_regs->regmap,agr);
4485 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4486 }
4487 }
4488 int rs = get_reg(i_regs->regmap, dops[i].rs1);
4489 //if(ra>=0)
4490 {
4491 int offset = cinfo[i].imm;
4492 int add_offset = offset != 0;
4493 int c = rs >= 0 && ((i_regs->wasconst >> rs) & 1);
4494 if(dops[i].rs1==0) {
4495 // Using r0 as a base address
4496 assert(ra >= 0);
4497 if(!entry||entry[ra]!=agr) {
4498 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4499 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4500 }else{
4501 emit_movimm(offset,ra);
4502 }
4503 } // else did it in the previous cycle
4504 cinfo[i].addr = ra;
4505 add_offset = 0;
4506 }
4507 else if (rs < 0) {
4508 assert(ra >= 0);
4509 if (!entry || entry[ra] != dops[i].rs1)
4510 emit_loadreg(dops[i].rs1, ra);
4511 cinfo[i].addr = ra;
4512 //if(!entry||entry[ra]!=dops[i].rs1)
4513 // printf("poor load scheduling!\n");
4514 }
4515 else if(c) {
4516 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4517 assert(ra >= 0);
4518 if(!entry||entry[ra]!=agr) {
4519 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4520 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4521 }else{
4522 emit_movimm(constmap[i][rs]+offset,ra);
4523 regs[i].loadedconst|=1<<ra;
4524 }
4525 } // else did it in the previous cycle
4526 cinfo[i].addr = ra;
4527 }
4528 else // else load_consts already did it
4529 cinfo[i].addr = rs;
4530 add_offset = 0;
4531 }
4532 else if (dops[i].itype == STORELR) { // overwrites addr
4533 assert(ra >= 0);
4534 assert(rs != ra);
4535 emit_mov(rs, ra);
4536 cinfo[i].addr = ra;
4537 }
4538 else
4539 cinfo[i].addr = rs;
4540 if (add_offset) {
4541 assert(ra >= 0);
4542 if(rs>=0) {
4543 emit_addimm(rs,offset,ra);
4544 }else{
4545 emit_addimm(ra,offset,ra);
4546 }
4547 cinfo[i].addr = ra;
4548 }
4549 }
4550 assert(cinfo[i].addr >= 0);
4551 }
4552 // Preload constants for next instruction
4553 if (dops[i+1].is_load || dops[i+1].is_store) {
4554 int agr,ra;
4555 // Actual address
4556 agr=AGEN1+((i+1)&1);
4557 ra=get_reg(i_regs->regmap,agr);
4558 if(ra>=0) {
4559 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4560 int offset=cinfo[i+1].imm;
4561 int c=(regs[i+1].wasconst>>rs)&1;
4562 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4563 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4564 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4565 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4566 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4567 }else{
4568 emit_movimm(constmap[i+1][rs]+offset,ra);
4569 regs[i+1].loadedconst|=1<<ra;
4570 }
4571 }
4572 else if(dops[i+1].rs1==0) {
4573 // Using r0 as a base address
4574 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4575 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4576 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4577 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4578 }else{
4579 emit_movimm(offset,ra);
4580 }
4581 }
4582 }
4583 }
4584}
4585
4586static int get_final_value(int hr, int i, u_int *value)
4587{
4588 int reg=regs[i].regmap[hr];
4589 while(i<slen-1) {
4590 if(regs[i+1].regmap[hr]!=reg) break;
4591 if(!((regs[i+1].isconst>>hr)&1)) break;
4592 if(dops[i+1].bt) break;
4593 i++;
4594 }
4595 if(i<slen-1) {
4596 if (dops[i].is_jump) {
4597 *value=constmap[i][hr];
4598 return 1;
4599 }
4600 if(!dops[i+1].bt) {
4601 if (dops[i+1].is_jump) {
4602 // Load in delay slot, out-of-order execution
4603 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4604 {
4605 // Precompute load address
4606 *value=constmap[i][hr]+cinfo[i+2].imm;
4607 return 1;
4608 }
4609 }
4610 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4611 {
4612 // Precompute load address
4613 *value=constmap[i][hr]+cinfo[i+1].imm;
4614 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],cinfo[i+1].imm);
4615 return 1;
4616 }
4617 }
4618 }
4619 *value=constmap[i][hr];
4620 //printf("c=%lx\n",(long)constmap[i][hr]);
4621 if(i==slen-1) return 1;
4622 assert(reg < 64);
4623 return !((unneeded_reg[i+1]>>reg)&1);
4624}
4625
4626// Load registers with known constants
4627static void load_consts(signed char pre[],signed char regmap[],int i)
4628{
4629 int hr,hr2;
4630 // propagate loaded constant flags
4631 if(i==0||dops[i].bt)
4632 regs[i].loadedconst=0;
4633 else {
4634 for(hr=0;hr<HOST_REGS;hr++) {
4635 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4636 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4637 {
4638 regs[i].loadedconst|=1<<hr;
4639 }
4640 }
4641 }
4642 // Load 32-bit regs
4643 for(hr=0;hr<HOST_REGS;hr++) {
4644 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4645 //if(entry[hr]!=regmap[hr]) {
4646 if(!((regs[i].loadedconst>>hr)&1)) {
4647 assert(regmap[hr]<64);
4648 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4649 u_int value, similar=0;
4650 if(get_final_value(hr,i,&value)) {
4651 // see if some other register has similar value
4652 for(hr2=0;hr2<HOST_REGS;hr2++) {
4653 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4654 if(is_similar_value(value,constmap[i][hr2])) {
4655 similar=1;
4656 break;
4657 }
4658 }
4659 }
4660 if(similar) {
4661 u_int value2;
4662 if(get_final_value(hr2,i,&value2)) // is this needed?
4663 emit_movimm_from(value2,hr2,value,hr);
4664 else
4665 emit_movimm(value,hr);
4666 }
4667 else if(value==0) {
4668 emit_zeroreg(hr);
4669 }
4670 else {
4671 emit_movimm(value,hr);
4672 }
4673 }
4674 regs[i].loadedconst|=1<<hr;
4675 }
4676 }
4677 }
4678 }
4679}
4680
4681static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4682{
4683 int hr;
4684 // Load 32-bit regs
4685 for(hr=0;hr<HOST_REGS;hr++) {
4686 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4687 assert(regmap[hr] < 64);
4688 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4689 int value=constmap[i][hr];
4690 if(value==0) {
4691 emit_zeroreg(hr);
4692 }
4693 else {
4694 emit_movimm(value,hr);
4695 }
4696 }
4697 }
4698 }
4699}
4700
4701// Write out all dirty registers (except cycle count)
4702#ifndef wb_dirtys
4703static void wb_dirtys(const signed char i_regmap[], u_int i_dirty)
4704{
4705 int hr;
4706 for(hr=0;hr<HOST_REGS;hr++) {
4707 if(hr!=EXCLUDE_REG) {
4708 if(i_regmap[hr]>0) {
4709 if(i_regmap[hr]!=CCREG) {
4710 if((i_dirty>>hr)&1) {
4711 assert(i_regmap[hr]<64);
4712 emit_storereg(i_regmap[hr],hr);
4713 }
4714 }
4715 }
4716 }
4717 }
4718}
4719#endif
4720
4721// Write out dirty registers that we need to reload (pair with load_needed_regs)
4722// This writes the registers not written by store_regs_bt
4723static void wb_needed_dirtys(const signed char i_regmap[], u_int i_dirty, int addr)
4724{
4725 int hr;
4726 int t=(addr-start)>>2;
4727 for(hr=0;hr<HOST_REGS;hr++) {
4728 if(hr!=EXCLUDE_REG) {
4729 if(i_regmap[hr]>0) {
4730 if(i_regmap[hr]!=CCREG) {
4731 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4732 if((i_dirty>>hr)&1) {
4733 assert(i_regmap[hr]<64);
4734 emit_storereg(i_regmap[hr],hr);
4735 }
4736 }
4737 }
4738 }
4739 }
4740 }
4741}
4742
4743// Load all registers (except cycle count)
4744#ifndef load_all_regs
4745static void load_all_regs(const signed char i_regmap[])
4746{
4747 int hr;
4748 for(hr=0;hr<HOST_REGS;hr++) {
4749 if(hr!=EXCLUDE_REG) {
4750 if(i_regmap[hr]==0) {
4751 emit_zeroreg(hr);
4752 }
4753 else
4754 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4755 {
4756 emit_loadreg(i_regmap[hr],hr);
4757 }
4758 }
4759 }
4760}
4761#endif
4762
4763// Load all current registers also needed by next instruction
4764static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4765{
4766 signed char regmap_sel[HOST_REGS];
4767 int hr;
4768 for (hr = 0; hr < HOST_REGS; hr++) {
4769 regmap_sel[hr] = -1;
4770 if (hr != EXCLUDE_REG)
4771 if (next_regmap[hr] == i_regmap[hr] || get_reg(next_regmap, i_regmap[hr]) >= 0)
4772 regmap_sel[hr] = i_regmap[hr];
4773 }
4774 load_all_regs(regmap_sel);
4775}
4776
4777// Load all regs, storing cycle count if necessary
4778static void load_regs_entry(int t)
4779{
4780 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4781 else if(cinfo[t].ccadj) emit_addimm(HOST_CCREG,-cinfo[t].ccadj,HOST_CCREG);
4782 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4783 emit_storereg(CCREG,HOST_CCREG);
4784 }
4785 load_all_regs(regs[t].regmap_entry);
4786}
4787
4788// Store dirty registers prior to branch
4789static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4790{
4791 if(internal_branch(addr))
4792 {
4793 int t=(addr-start)>>2;
4794 int hr;
4795 for(hr=0;hr<HOST_REGS;hr++) {
4796 if(hr!=EXCLUDE_REG) {
4797 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4798 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4799 if((i_dirty>>hr)&1) {
4800 assert(i_regmap[hr]<64);
4801 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4802 emit_storereg(i_regmap[hr],hr);
4803 }
4804 }
4805 }
4806 }
4807 }
4808 }
4809 else
4810 {
4811 // Branch out of this block, write out all dirty regs
4812 wb_dirtys(i_regmap,i_dirty);
4813 }
4814}
4815
4816// Load all needed registers for branch target
4817static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4818{
4819 //if(addr>=start && addr<(start+slen*4))
4820 if(internal_branch(addr))
4821 {
4822 int t=(addr-start)>>2;
4823 int hr;
4824 // Store the cycle count before loading something else
4825 if(i_regmap[HOST_CCREG]!=CCREG) {
4826 assert(i_regmap[HOST_CCREG]==-1);
4827 }
4828 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4829 emit_storereg(CCREG,HOST_CCREG);
4830 }
4831 // Load 32-bit regs
4832 for(hr=0;hr<HOST_REGS;hr++) {
4833 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4834 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4835 if(regs[t].regmap_entry[hr]==0) {
4836 emit_zeroreg(hr);
4837 }
4838 else if(regs[t].regmap_entry[hr]!=CCREG)
4839 {
4840 emit_loadreg(regs[t].regmap_entry[hr],hr);
4841 }
4842 }
4843 }
4844 }
4845 }
4846}
4847
4848static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4849{
4850 if(addr>=start && addr<start+slen*4-4)
4851 {
4852 int t=(addr-start)>>2;
4853 int hr;
4854 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4855 for(hr=0;hr<HOST_REGS;hr++)
4856 {
4857 if(hr!=EXCLUDE_REG)
4858 {
4859 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4860 {
4861 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4862 {
4863 return 0;
4864 }
4865 else
4866 if((i_dirty>>hr)&1)
4867 {
4868 if(i_regmap[hr]<TEMPREG)
4869 {
4870 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4871 return 0;
4872 }
4873 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4874 {
4875 assert(0);
4876 }
4877 }
4878 }
4879 else // Same register but is it 32-bit or dirty?
4880 if(i_regmap[hr]>=0)
4881 {
4882 if(!((regs[t].dirty>>hr)&1))
4883 {
4884 if((i_dirty>>hr)&1)
4885 {
4886 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4887 {
4888 //printf("%x: dirty no match\n",addr);
4889 return 0;
4890 }
4891 }
4892 }
4893 }
4894 }
4895 }
4896 // Delay slots are not valid branch targets
4897 //if(t>0&&(dops[t-1].is_jump) return 0;
4898 // Delay slots require additional processing, so do not match
4899 if(dops[t].is_ds) return 0;
4900 }
4901 else
4902 {
4903 int hr;
4904 for(hr=0;hr<HOST_REGS;hr++)
4905 {
4906 if(hr!=EXCLUDE_REG)
4907 {
4908 if(i_regmap[hr]>=0)
4909 {
4910 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4911 {
4912 if((i_dirty>>hr)&1)
4913 {
4914 return 0;
4915 }
4916 }
4917 }
4918 }
4919 }
4920 }
4921 return 1;
4922}
4923
4924#ifdef DRC_DBG
4925static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4926{
4927 extern void do_insn_cmp();
4928 //extern int cycle;
4929 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4930 reglist |= get_host_reglist(regs[i].regmap_entry);
4931 reglist &= DRC_DBG_REGMASK;
4932
4933 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4934 save_regs(reglist);
4935 // write out changed consts to match the interpreter
4936 if (i > 0 && !dops[i].bt) {
4937 for (hr = 0; hr < HOST_REGS; hr++) {
4938 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4939 if (hr == EXCLUDE_REG || reg <= 0)
4940 continue;
4941 if (!((regs[i-1].isconst >> hr) & 1))
4942 continue;
4943 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4944 continue;
4945 emit_movimm(constmap[i-1][hr],0);
4946 emit_storereg(reg, 0);
4947 }
4948 }
4949 emit_movimm(start+i*4,0);
4950 emit_writeword(0,&pcaddr);
4951 int cc = get_reg(regs[i].regmap_entry, CCREG);
4952 if (cc < 0)
4953 emit_loadreg(CCREG, cc = 0);
4954 emit_addimm(cc, ccadj_, 0);
4955 emit_writeword(0, &psxRegs.cycle);
4956 emit_far_call(do_insn_cmp);
4957 //emit_readword(&cycle,0);
4958 //emit_addimm(0,2,0);
4959 //emit_writeword(0,&cycle);
4960 (void)get_reg2;
4961 restore_regs(reglist);
4962 assem_debug("\\\\do_insn_cmp\n");
4963}
4964#else
4965#define drc_dbg_emit_do_cmp(x,y)
4966#endif
4967
4968// Used when a branch jumps into the delay slot of another branch
4969static void ds_assemble_entry(int i)
4970{
4971 int t = (cinfo[i].ba - start) >> 2;
4972 int ccadj_ = -CLOCK_ADJUST(1);
4973 if (!instr_addr[t])
4974 instr_addr[t] = out;
4975 assem_debug("Assemble delay slot at %x\n",cinfo[i].ba);
4976 assem_debug("<->\n");
4977 drc_dbg_emit_do_cmp(t, ccadj_);
4978 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4979 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4980 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4981 address_generation(t,&regs[t],regs[t].regmap_entry);
4982 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4983 load_reg(regs[t].regmap_entry,regs[t].regmap,ROREG);
4984 if (dops[t].is_store)
4985 load_reg(regs[t].regmap_entry,regs[t].regmap,INVCP);
4986 is_delayslot=0;
4987 switch (dops[t].itype) {
4988 case SYSCALL:
4989 case HLECALL:
4990 case INTCALL:
4991 case UJUMP:
4992 case RJUMP:
4993 case CJUMP:
4994 case SJUMP:
4995 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4996 break;
4997 default:
4998 assemble(t, &regs[t], ccadj_);
4999 }
5000 store_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5001 load_regs_bt(regs[t].regmap,regs[t].dirty,cinfo[i].ba+4);
5002 if(internal_branch(cinfo[i].ba+4))
5003 assem_debug("branch: internal\n");
5004 else
5005 assem_debug("branch: external\n");
5006 assert(internal_branch(cinfo[i].ba+4));
5007 add_to_linker(out,cinfo[i].ba+4,internal_branch(cinfo[i].ba+4));
5008 emit_jmp(0);
5009}
5010
5011// Load 2 immediates optimizing for small code size
5012static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
5013{
5014 emit_movimm(imm1,rt1);
5015 emit_movimm_from(imm1,rt1,imm2,rt2);
5016}
5017
5018static void do_cc(int i, const signed char i_regmap[], int *adj,
5019 int addr, int taken, int invert)
5020{
5021 int count, count_plus2;
5022 void *jaddr;
5023 void *idle=NULL;
5024 int t=0;
5025 if(dops[i].itype==RJUMP)
5026 {
5027 *adj=0;
5028 }
5029 //if(cinfo[i].ba>=start && cinfo[i].ba<(start+slen*4))
5030 if(internal_branch(cinfo[i].ba))
5031 {
5032 t=(cinfo[i].ba-start)>>2;
5033 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
5034 else *adj=cinfo[t].ccadj;
5035 }
5036 else
5037 {
5038 *adj=0;
5039 }
5040 count = cinfo[i].ccadj;
5041 count_plus2 = count + CLOCK_ADJUST(2);
5042 if(taken==TAKEN && i==(cinfo[i].ba-start)>>2 && source[i+1]==0) {
5043 // Idle loop
5044 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
5045 idle=out;
5046 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
5047 emit_andimm(HOST_CCREG,3,HOST_CCREG);
5048 jaddr=out;
5049 emit_jmp(0);
5050 }
5051 else if(*adj==0||invert) {
5052 int cycles = count_plus2;
5053 // faster loop HACK
5054#if 0
5055 if (t&&*adj) {
5056 int rel=t-i;
5057 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
5058 cycles=*adj+count+2-*adj;
5059 }
5060#endif
5061 emit_addimm_and_set_flags(cycles, HOST_CCREG);
5062 jaddr = out;
5063 emit_jns(0);
5064 }
5065 else
5066 {
5067 emit_cmpimm(HOST_CCREG, -count_plus2);
5068 jaddr = out;
5069 emit_jns(0);
5070 }
5071 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
5072}
5073
5074static void do_ccstub(int n)
5075{
5076 literal_pool(256);
5077 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
5078 set_jump_target(stubs[n].addr, out);
5079 int i=stubs[n].b;
5080 if (stubs[n].d != TAKEN) {
5081 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
5082 }
5083 else {
5084 if(internal_branch(cinfo[i].ba))
5085 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5086 }
5087 if(stubs[n].c!=-1)
5088 {
5089 // Save PC as return address
5090 emit_movimm(stubs[n].c,0);
5091 emit_writeword(0,&pcaddr);
5092 }
5093 else
5094 {
5095 // Return address depends on which way the branch goes
5096 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
5097 {
5098 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5099 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5100 if(dops[i].rs1==0)
5101 {
5102 s1l=s2l;
5103 s2l=-1;
5104 }
5105 else if(dops[i].rs2==0)
5106 {
5107 s2l=-1;
5108 }
5109 assert(s1l>=0);
5110 #ifdef DESTRUCTIVE_WRITEBACK
5111 if(dops[i].rs1) {
5112 if((branch_regs[i].dirty>>s1l)&&1)
5113 emit_loadreg(dops[i].rs1,s1l);
5114 }
5115 else {
5116 if((branch_regs[i].dirty>>s1l)&1)
5117 emit_loadreg(dops[i].rs2,s1l);
5118 }
5119 if(s2l>=0)
5120 if((branch_regs[i].dirty>>s2l)&1)
5121 emit_loadreg(dops[i].rs2,s2l);
5122 #endif
5123 int hr=0;
5124 int addr=-1,alt=-1,ntaddr=-1;
5125 while(hr<HOST_REGS)
5126 {
5127 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5128 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5129 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5130 {
5131 addr=hr++;break;
5132 }
5133 hr++;
5134 }
5135 while(hr<HOST_REGS)
5136 {
5137 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5138 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5139 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5140 {
5141 alt=hr++;break;
5142 }
5143 hr++;
5144 }
5145 if ((dops[i].opcode & 0x3e) == 6) // BLEZ/BGTZ needs another register
5146 {
5147 while(hr<HOST_REGS)
5148 {
5149 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5150 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5151 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5152 {
5153 ntaddr=hr;break;
5154 }
5155 hr++;
5156 }
5157 assert(hr<HOST_REGS);
5158 }
5159 if (dops[i].opcode == 4) // BEQ
5160 {
5161 #ifdef HAVE_CMOV_IMM
5162 if(s2l>=0) emit_cmp(s1l,s2l);
5163 else emit_test(s1l,s1l);
5164 emit_cmov2imm_e_ne_compact(cinfo[i].ba,start+i*4+8,addr);
5165 #else
5166 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,alt);
5167 if(s2l>=0) emit_cmp(s1l,s2l);
5168 else emit_test(s1l,s1l);
5169 emit_cmovne_reg(alt,addr);
5170 #endif
5171 }
5172 else if (dops[i].opcode == 5) // BNE
5173 {
5174 #ifdef HAVE_CMOV_IMM
5175 if(s2l>=0) emit_cmp(s1l,s2l);
5176 else emit_test(s1l,s1l);
5177 emit_cmov2imm_e_ne_compact(start+i*4+8,cinfo[i].ba,addr);
5178 #else
5179 emit_mov2imm_compact(start+i*4+8,addr,cinfo[i].ba,alt);
5180 if(s2l>=0) emit_cmp(s1l,s2l);
5181 else emit_test(s1l,s1l);
5182 emit_cmovne_reg(alt,addr);
5183 #endif
5184 }
5185 else if (dops[i].opcode == 6) // BLEZ
5186 {
5187 //emit_movimm(cinfo[i].ba,alt);
5188 //emit_movimm(start+i*4+8,addr);
5189 emit_mov2imm_compact(cinfo[i].ba,alt,start+i*4+8,addr);
5190 emit_cmpimm(s1l,1);
5191 emit_cmovl_reg(alt,addr);
5192 }
5193 else if (dops[i].opcode == 7) // BGTZ
5194 {
5195 //emit_movimm(cinfo[i].ba,addr);
5196 //emit_movimm(start+i*4+8,ntaddr);
5197 emit_mov2imm_compact(cinfo[i].ba,addr,start+i*4+8,ntaddr);
5198 emit_cmpimm(s1l,1);
5199 emit_cmovl_reg(ntaddr,addr);
5200 }
5201 else if (dops[i].itype == SJUMP) // BLTZ/BGEZ
5202 {
5203 //emit_movimm(cinfo[i].ba,alt);
5204 //emit_movimm(start+i*4+8,addr);
5205 if (dops[i].rs1) {
5206 emit_mov2imm_compact(cinfo[i].ba,
5207 (dops[i].opcode2 & 1) ? addr : alt, start + i*4 + 8,
5208 (dops[i].opcode2 & 1) ? alt : addr);
5209 emit_test(s1l,s1l);
5210 emit_cmovs_reg(alt,addr);
5211 }
5212 else
5213 emit_movimm((dops[i].opcode2 & 1) ? cinfo[i].ba : start + i*4 + 8, addr);
5214 }
5215 emit_writeword(addr, &pcaddr);
5216 }
5217 else
5218 if(dops[i].itype==RJUMP)
5219 {
5220 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5221 if (ds_writes_rjump_rs(i)) {
5222 r=get_reg(branch_regs[i].regmap,RTEMP);
5223 }
5224 emit_writeword(r,&pcaddr);
5225 }
5226 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5227 }
5228 // Update cycle count
5229 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5230 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5231 emit_far_call(cc_interrupt);
5232 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5233 if(stubs[n].d==TAKEN) {
5234 if(internal_branch(cinfo[i].ba))
5235 load_needed_regs(branch_regs[i].regmap,regs[(cinfo[i].ba-start)>>2].regmap_entry);
5236 else if(dops[i].itype==RJUMP) {
5237 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5238 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5239 else
5240 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5241 }
5242 }else if(stubs[n].d==NOTTAKEN) {
5243 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5244 else load_all_regs(branch_regs[i].regmap);
5245 }else{
5246 load_all_regs(branch_regs[i].regmap);
5247 }
5248 if (stubs[n].retaddr)
5249 emit_jmp(stubs[n].retaddr);
5250 else
5251 do_jump_vaddr(stubs[n].e);
5252}
5253
5254static void add_to_linker(void *addr, u_int target, int is_internal)
5255{
5256 assert(linkcount < ARRAY_SIZE(link_addr));
5257 link_addr[linkcount].addr = addr;
5258 link_addr[linkcount].target = target;
5259 link_addr[linkcount].internal = is_internal;
5260 linkcount++;
5261}
5262
5263static void ujump_assemble_write_ra(int i)
5264{
5265 int rt;
5266 unsigned int return_address;
5267 rt=get_reg(branch_regs[i].regmap,31);
5268 //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]);
5269 //assert(rt>=0);
5270 return_address=start+i*4+8;
5271 if(rt>=0) {
5272 #ifdef USE_MINI_HT
5273 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5274 int temp=-1; // note: must be ds-safe
5275 #ifdef HOST_TEMPREG
5276 temp=HOST_TEMPREG;
5277 #endif
5278 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5279 else emit_movimm(return_address,rt);
5280 }
5281 else
5282 #endif
5283 {
5284 #ifdef REG_PREFETCH
5285 if(temp>=0)
5286 {
5287 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5288 }
5289 #endif
5290 if (!((regs[i].loadedconst >> rt) & 1))
5291 emit_movimm(return_address, rt); // PC into link register
5292 #ifdef IMM_PREFETCH
5293 emit_prefetch(hash_table_get(return_address));
5294 #endif
5295 }
5296 }
5297}
5298
5299static void ujump_assemble(int i, const struct regstat *i_regs)
5300{
5301 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5302 address_generation(i+1,i_regs,regs[i].regmap_entry);
5303 #ifdef REG_PREFETCH
5304 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5305 if(dops[i].rt1==31&&temp>=0)
5306 {
5307 signed char *i_regmap=i_regs->regmap;
5308 int return_address=start+i*4+8;
5309 if(get_reg(branch_regs[i].regmap,31)>0)
5310 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5311 }
5312 #endif
5313 if (dops[i].rt1 == 31)
5314 ujump_assemble_write_ra(i); // writeback ra for DS
5315 ds_assemble(i+1,i_regs);
5316 uint64_t bc_unneeded=branch_regs[i].u;
5317 bc_unneeded|=1|(1LL<<dops[i].rt1);
5318 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5319 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5320 int cc,adj;
5321 cc=get_reg(branch_regs[i].regmap,CCREG);
5322 assert(cc==HOST_CCREG);
5323 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5324 #ifdef REG_PREFETCH
5325 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5326 #endif
5327 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5328 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5329 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5330 if(internal_branch(cinfo[i].ba))
5331 assem_debug("branch: internal\n");
5332 else
5333 assem_debug("branch: external\n");
5334 if (internal_branch(cinfo[i].ba) && dops[(cinfo[i].ba-start)>>2].is_ds) {
5335 ds_assemble_entry(i);
5336 }
5337 else {
5338 add_to_linker(out,cinfo[i].ba,internal_branch(cinfo[i].ba));
5339 emit_jmp(0);
5340 }
5341}
5342
5343static void rjump_assemble_write_ra(int i)
5344{
5345 int rt,return_address;
5346 rt=get_reg_w(branch_regs[i].regmap, dops[i].rt1);
5347 //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]);
5348 assert(rt>=0);
5349 return_address=start+i*4+8;
5350 #ifdef REG_PREFETCH
5351 if(temp>=0)
5352 {
5353 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5354 }
5355 #endif
5356 if (!((regs[i].loadedconst >> rt) & 1))
5357 emit_movimm(return_address, rt); // PC into link register
5358 #ifdef IMM_PREFETCH
5359 emit_prefetch(hash_table_get(return_address));
5360 #endif
5361}
5362
5363static void rjump_assemble(int i, const struct regstat *i_regs)
5364{
5365 int temp;
5366 int rs,cc;
5367 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5368 assert(rs>=0);
5369 if (ds_writes_rjump_rs(i)) {
5370 // Delay slot abuse, make a copy of the branch address register
5371 temp=get_reg(branch_regs[i].regmap,RTEMP);
5372 assert(temp>=0);
5373 assert(regs[i].regmap[temp]==RTEMP);
5374 emit_mov(rs,temp);
5375 rs=temp;
5376 }
5377 address_generation(i+1,i_regs,regs[i].regmap_entry);
5378 #ifdef REG_PREFETCH
5379 if(dops[i].rt1==31)
5380 {
5381 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5382 signed char *i_regmap=i_regs->regmap;
5383 int return_address=start+i*4+8;
5384 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5385 }
5386 }
5387 #endif
5388 #ifdef USE_MINI_HT
5389 if(dops[i].rs1==31) {
5390 int rh=get_reg(regs[i].regmap,RHASH);
5391 if(rh>=0) do_preload_rhash(rh);
5392 }
5393 #endif
5394 if (dops[i].rt1 != 0)
5395 rjump_assemble_write_ra(i);
5396 ds_assemble(i+1,i_regs);
5397 uint64_t bc_unneeded=branch_regs[i].u;
5398 bc_unneeded|=1|(1LL<<dops[i].rt1);
5399 bc_unneeded&=~(1LL<<dops[i].rs1);
5400 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5401 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5402 cc=get_reg(branch_regs[i].regmap,CCREG);
5403 assert(cc==HOST_CCREG);
5404 (void)cc;
5405 #ifdef USE_MINI_HT
5406 int rh=get_reg(branch_regs[i].regmap,RHASH);
5407 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5408 if(dops[i].rs1==31) {
5409 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5410 do_preload_rhtbl(ht);
5411 do_rhash(rs,rh);
5412 }
5413 #endif
5414 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5415 #ifdef DESTRUCTIVE_WRITEBACK
5416 if((branch_regs[i].dirty>>rs)&1) {
5417 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5418 emit_loadreg(dops[i].rs1,rs);
5419 }
5420 }
5421 #endif
5422 #ifdef REG_PREFETCH
5423 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5424 #endif
5425 #ifdef USE_MINI_HT
5426 if(dops[i].rs1==31) {
5427 do_miniht_load(ht,rh);
5428 }
5429 #endif
5430 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5431 //if(adj) emit_addimm(cc,2*(cinfo[i].ccadj+2-adj),cc); // ??? - Shouldn't happen
5432 //assert(adj==0);
5433 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5434 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5435 if (dops[i+1].itype == RFE)
5436 // special case for RFE
5437 emit_jmp(0);
5438 else
5439 emit_jns(0);
5440 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5441 #ifdef USE_MINI_HT
5442 if(dops[i].rs1==31) {
5443 do_miniht_jump(rs,rh,ht);
5444 }
5445 else
5446 #endif
5447 {
5448 do_jump_vaddr(rs);
5449 }
5450 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5451 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5452 #endif
5453}
5454
5455static void cjump_assemble(int i, const struct regstat *i_regs)
5456{
5457 const signed char *i_regmap = i_regs->regmap;
5458 int cc;
5459 int match;
5460 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5461 assem_debug("match=%d\n",match);
5462 int s1l,s2l;
5463 int unconditional=0,nop=0;
5464 int invert=0;
5465 int internal=internal_branch(cinfo[i].ba);
5466 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5467 if(!match) invert=1;
5468 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5469 if(i>(cinfo[i].ba-start)>>2) invert=1;
5470 #endif
5471 #ifdef __aarch64__
5472 invert=1; // because of near cond. branches
5473 #endif
5474
5475 if(dops[i].ooo) {
5476 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5477 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5478 }
5479 else {
5480 s1l=get_reg(i_regmap,dops[i].rs1);
5481 s2l=get_reg(i_regmap,dops[i].rs2);
5482 }
5483 if(dops[i].rs1==0&&dops[i].rs2==0)
5484 {
5485 if(dops[i].opcode&1) nop=1;
5486 else unconditional=1;
5487 //assert(dops[i].opcode!=5);
5488 //assert(dops[i].opcode!=7);
5489 //assert(dops[i].opcode!=0x15);
5490 //assert(dops[i].opcode!=0x17);
5491 }
5492 else if(dops[i].rs1==0)
5493 {
5494 s1l=s2l;
5495 s2l=-1;
5496 }
5497 else if(dops[i].rs2==0)
5498 {
5499 s2l=-1;
5500 }
5501
5502 if(dops[i].ooo) {
5503 // Out of order execution (delay slot first)
5504 //printf("OOOE\n");
5505 address_generation(i+1,i_regs,regs[i].regmap_entry);
5506 ds_assemble(i+1,i_regs);
5507 int adj;
5508 uint64_t bc_unneeded=branch_regs[i].u;
5509 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5510 bc_unneeded|=1;
5511 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5512 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5513 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5514 cc=get_reg(branch_regs[i].regmap,CCREG);
5515 assert(cc==HOST_CCREG);
5516 if(unconditional)
5517 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5518 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5519 //assem_debug("cycle count (adj)\n");
5520 if(unconditional) {
5521 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5522 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5523 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5524 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5525 if(internal)
5526 assem_debug("branch: internal\n");
5527 else
5528 assem_debug("branch: external\n");
5529 if (internal && dops[(cinfo[i].ba-start)>>2].is_ds) {
5530 ds_assemble_entry(i);
5531 }
5532 else {
5533 add_to_linker(out,cinfo[i].ba,internal);
5534 emit_jmp(0);
5535 }
5536 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5537 if(((u_int)out)&7) emit_addnop(0);
5538 #endif
5539 }
5540 }
5541 else if(nop) {
5542 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5543 void *jaddr=out;
5544 emit_jns(0);
5545 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5546 }
5547 else {
5548 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5549 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5550 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5551
5552 //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]);
5553 assert(s1l>=0);
5554 if(dops[i].opcode==4) // BEQ
5555 {
5556 if(s2l>=0) emit_cmp(s1l,s2l);
5557 else emit_test(s1l,s1l);
5558 if(invert){
5559 nottaken=out;
5560 emit_jne(DJT_1);
5561 }else{
5562 add_to_linker(out,cinfo[i].ba,internal);
5563 emit_jeq(0);
5564 }
5565 }
5566 if(dops[i].opcode==5) // BNE
5567 {
5568 if(s2l>=0) emit_cmp(s1l,s2l);
5569 else emit_test(s1l,s1l);
5570 if(invert){
5571 nottaken=out;
5572 emit_jeq(DJT_1);
5573 }else{
5574 add_to_linker(out,cinfo[i].ba,internal);
5575 emit_jne(0);
5576 }
5577 }
5578 if(dops[i].opcode==6) // BLEZ
5579 {
5580 emit_cmpimm(s1l,1);
5581 if(invert){
5582 nottaken=out;
5583 emit_jge(DJT_1);
5584 }else{
5585 add_to_linker(out,cinfo[i].ba,internal);
5586 emit_jl(0);
5587 }
5588 }
5589 if(dops[i].opcode==7) // BGTZ
5590 {
5591 emit_cmpimm(s1l,1);
5592 if(invert){
5593 nottaken=out;
5594 emit_jl(DJT_1);
5595 }else{
5596 add_to_linker(out,cinfo[i].ba,internal);
5597 emit_jge(0);
5598 }
5599 }
5600 if(invert) {
5601 if(taken) set_jump_target(taken, out);
5602 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5603 if (match && (!internal || !dops[(cinfo[i].ba-start)>>2].is_ds)) {
5604 if(adj) {
5605 emit_addimm(cc,-adj,cc);
5606 add_to_linker(out,cinfo[i].ba,internal);
5607 }else{
5608 emit_addnop(13);
5609 add_to_linker(out,cinfo[i].ba,internal*2);
5610 }
5611 emit_jmp(0);
5612 }else
5613 #endif
5614 {
5615 if(adj) emit_addimm(cc,-adj,cc);
5616 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5617 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5618 if(internal)
5619 assem_debug("branch: internal\n");
5620 else
5621 assem_debug("branch: external\n");
5622 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5623 ds_assemble_entry(i);
5624 }
5625 else {
5626 add_to_linker(out,cinfo[i].ba,internal);
5627 emit_jmp(0);
5628 }
5629 }
5630 set_jump_target(nottaken, out);
5631 }
5632
5633 if(nottaken1) set_jump_target(nottaken1, out);
5634 if(adj) {
5635 if(!invert) emit_addimm(cc,adj,cc);
5636 }
5637 } // (!unconditional)
5638 } // if(ooo)
5639 else
5640 {
5641 // In-order execution (branch first)
5642 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5643 if(!unconditional&&!nop) {
5644 //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]);
5645 assert(s1l>=0);
5646 if((dops[i].opcode&0x2f)==4) // BEQ
5647 {
5648 if(s2l>=0) emit_cmp(s1l,s2l);
5649 else emit_test(s1l,s1l);
5650 nottaken=out;
5651 emit_jne(DJT_2);
5652 }
5653 if((dops[i].opcode&0x2f)==5) // BNE
5654 {
5655 if(s2l>=0) emit_cmp(s1l,s2l);
5656 else emit_test(s1l,s1l);
5657 nottaken=out;
5658 emit_jeq(DJT_2);
5659 }
5660 if((dops[i].opcode&0x2f)==6) // BLEZ
5661 {
5662 emit_cmpimm(s1l,1);
5663 nottaken=out;
5664 emit_jge(DJT_2);
5665 }
5666 if((dops[i].opcode&0x2f)==7) // BGTZ
5667 {
5668 emit_cmpimm(s1l,1);
5669 nottaken=out;
5670 emit_jl(DJT_2);
5671 }
5672 } // if(!unconditional)
5673 int adj;
5674 uint64_t ds_unneeded=branch_regs[i].u;
5675 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5676 ds_unneeded|=1;
5677 // branch taken
5678 if(!nop) {
5679 if(taken) set_jump_target(taken, out);
5680 assem_debug("1:\n");
5681 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5682 // load regs
5683 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5684 address_generation(i+1,&branch_regs[i],0);
5685 if (ram_offset)
5686 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5687 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5688 ds_assemble(i+1,&branch_regs[i]);
5689 cc=get_reg(branch_regs[i].regmap,CCREG);
5690 if(cc==-1) {
5691 emit_loadreg(CCREG,cc=HOST_CCREG);
5692 // CHECK: Is the following instruction (fall thru) allocated ok?
5693 }
5694 assert(cc==HOST_CCREG);
5695 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5696 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5697 assem_debug("cycle count (adj)\n");
5698 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5699 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5700 if(internal)
5701 assem_debug("branch: internal\n");
5702 else
5703 assem_debug("branch: external\n");
5704 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5705 ds_assemble_entry(i);
5706 }
5707 else {
5708 add_to_linker(out,cinfo[i].ba,internal);
5709 emit_jmp(0);
5710 }
5711 }
5712 // branch not taken
5713 if(!unconditional) {
5714 if(nottaken1) set_jump_target(nottaken1, out);
5715 set_jump_target(nottaken, out);
5716 assem_debug("2:\n");
5717 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5718 // load regs
5719 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5720 address_generation(i+1,&branch_regs[i],0);
5721 if (ram_offset)
5722 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5723 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5724 ds_assemble(i+1,&branch_regs[i]);
5725 cc=get_reg(branch_regs[i].regmap,CCREG);
5726 if (cc == -1) {
5727 // Cycle count isn't in a register, temporarily load it then write it out
5728 emit_loadreg(CCREG,HOST_CCREG);
5729 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5730 void *jaddr=out;
5731 emit_jns(0);
5732 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5733 emit_storereg(CCREG,HOST_CCREG);
5734 }
5735 else{
5736 cc=get_reg(i_regmap,CCREG);
5737 assert(cc==HOST_CCREG);
5738 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5739 void *jaddr=out;
5740 emit_jns(0);
5741 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5742 }
5743 }
5744 }
5745}
5746
5747static void sjump_assemble(int i, const struct regstat *i_regs)
5748{
5749 const signed char *i_regmap = i_regs->regmap;
5750 int cc;
5751 int match;
5752 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5753 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5754 int s1l;
5755 int unconditional=0,nevertaken=0;
5756 int invert=0;
5757 int internal=internal_branch(cinfo[i].ba);
5758 if(i==(cinfo[i].ba-start)>>2) assem_debug("idle loop\n");
5759 if(!match) invert=1;
5760 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5761 if(i>(cinfo[i].ba-start)>>2) invert=1;
5762 #endif
5763 #ifdef __aarch64__
5764 invert=1; // because of near cond. branches
5765 #endif
5766
5767 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5768 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5769
5770 if(dops[i].ooo) {
5771 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5772 }
5773 else {
5774 s1l=get_reg(i_regmap,dops[i].rs1);
5775 }
5776 if(dops[i].rs1==0)
5777 {
5778 if(dops[i].opcode2&1) unconditional=1;
5779 else nevertaken=1;
5780 // These are never taken (r0 is never less than zero)
5781 //assert(dops[i].opcode2!=0);
5782 //assert(dops[i].opcode2!=2);
5783 //assert(dops[i].opcode2!=0x10);
5784 //assert(dops[i].opcode2!=0x12);
5785 }
5786
5787 if(dops[i].ooo) {
5788 // Out of order execution (delay slot first)
5789 //printf("OOOE\n");
5790 address_generation(i+1,i_regs,regs[i].regmap_entry);
5791 ds_assemble(i+1,i_regs);
5792 int adj;
5793 uint64_t bc_unneeded=branch_regs[i].u;
5794 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5795 bc_unneeded|=1;
5796 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5797 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5798 load_reg(regs[i].regmap,branch_regs[i].regmap,CCREG);
5799 if(dops[i].rt1==31) {
5800 int rt,return_address;
5801 rt=get_reg(branch_regs[i].regmap,31);
5802 //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]);
5803 if(rt>=0) {
5804 // Save the PC even if the branch is not taken
5805 return_address=start+i*4+8;
5806 emit_movimm(return_address,rt); // PC into link register
5807 #ifdef IMM_PREFETCH
5808 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5809 #endif
5810 }
5811 }
5812 cc=get_reg(branch_regs[i].regmap,CCREG);
5813 assert(cc==HOST_CCREG);
5814 if(unconditional)
5815 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5816 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?cinfo[i].ba:-1,unconditional);
5817 assem_debug("cycle count (adj)\n");
5818 if(unconditional) {
5819 do_cc(i,branch_regs[i].regmap,&adj,cinfo[i].ba,TAKEN,0);
5820 if(i!=(cinfo[i].ba-start)>>2 || source[i+1]!=0) {
5821 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5822 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5823 if(internal)
5824 assem_debug("branch: internal\n");
5825 else
5826 assem_debug("branch: external\n");
5827 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5828 ds_assemble_entry(i);
5829 }
5830 else {
5831 add_to_linker(out,cinfo[i].ba,internal);
5832 emit_jmp(0);
5833 }
5834 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5835 if(((u_int)out)&7) emit_addnop(0);
5836 #endif
5837 }
5838 }
5839 else if(nevertaken) {
5840 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
5841 void *jaddr=out;
5842 emit_jns(0);
5843 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5844 }
5845 else {
5846 void *nottaken = NULL;
5847 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5848 if(adj&&!invert) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5849 {
5850 assert(s1l>=0);
5851 if ((dops[i].opcode2 & 1) == 0) // BLTZ/BLTZAL
5852 {
5853 emit_test(s1l,s1l);
5854 if(invert){
5855 nottaken=out;
5856 emit_jns(DJT_1);
5857 }else{
5858 add_to_linker(out,cinfo[i].ba,internal);
5859 emit_js(0);
5860 }
5861 }
5862 else // BGEZ/BGEZAL
5863 {
5864 emit_test(s1l,s1l);
5865 if(invert){
5866 nottaken=out;
5867 emit_js(DJT_1);
5868 }else{
5869 add_to_linker(out,cinfo[i].ba,internal);
5870 emit_jns(0);
5871 }
5872 }
5873 }
5874
5875 if(invert) {
5876 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5877 if (match && (!internal || !dops[(cinfo[i].ba - start) >> 2].is_ds)) {
5878 if(adj) {
5879 emit_addimm(cc,-adj,cc);
5880 add_to_linker(out,cinfo[i].ba,internal);
5881 }else{
5882 emit_addnop(13);
5883 add_to_linker(out,cinfo[i].ba,internal*2);
5884 }
5885 emit_jmp(0);
5886 }else
5887 #endif
5888 {
5889 if(adj) emit_addimm(cc,-adj,cc);
5890 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5891 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5892 if(internal)
5893 assem_debug("branch: internal\n");
5894 else
5895 assem_debug("branch: external\n");
5896 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5897 ds_assemble_entry(i);
5898 }
5899 else {
5900 add_to_linker(out,cinfo[i].ba,internal);
5901 emit_jmp(0);
5902 }
5903 }
5904 set_jump_target(nottaken, out);
5905 }
5906
5907 if(adj) {
5908 if(!invert) emit_addimm(cc,adj,cc);
5909 }
5910 } // (!unconditional)
5911 } // if(ooo)
5912 else
5913 {
5914 // In-order execution (branch first)
5915 //printf("IOE\n");
5916 void *nottaken = NULL;
5917 if (!unconditional && !nevertaken) {
5918 assert(s1l >= 0);
5919 emit_test(s1l, s1l);
5920 }
5921 if (dops[i].rt1 == 31) {
5922 int rt, return_address;
5923 rt = get_reg(branch_regs[i].regmap,31);
5924 if(rt >= 0) {
5925 // Save the PC even if the branch is not taken
5926 return_address = start + i*4+8;
5927 emit_movimm(return_address, rt); // PC into link register
5928 #ifdef IMM_PREFETCH
5929 emit_prefetch(hash_table_get(return_address));
5930 #endif
5931 }
5932 }
5933 if (!unconditional && !nevertaken) {
5934 nottaken = out;
5935 if (!(dops[i].opcode2 & 1)) // BLTZ/BLTZAL
5936 emit_jns(DJT_1);
5937 else // BGEZ/BGEZAL
5938 emit_js(DJT_1);
5939 }
5940 int adj;
5941 uint64_t ds_unneeded=branch_regs[i].u;
5942 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5943 ds_unneeded|=1;
5944 // branch taken
5945 if(!nevertaken) {
5946 //assem_debug("1:\n");
5947 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5948 // load regs
5949 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5950 address_generation(i+1,&branch_regs[i],0);
5951 if (ram_offset)
5952 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5953 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5954 ds_assemble(i+1,&branch_regs[i]);
5955 cc=get_reg(branch_regs[i].regmap,CCREG);
5956 if(cc==-1) {
5957 emit_loadreg(CCREG,cc=HOST_CCREG);
5958 // CHECK: Is the following instruction (fall thru) allocated ok?
5959 }
5960 assert(cc==HOST_CCREG);
5961 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5962 do_cc(i,i_regmap,&adj,cinfo[i].ba,TAKEN,0);
5963 assem_debug("cycle count (adj)\n");
5964 if(adj) emit_addimm(cc, cinfo[i].ccadj + CLOCK_ADJUST(2) - adj, cc);
5965 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,cinfo[i].ba);
5966 if(internal)
5967 assem_debug("branch: internal\n");
5968 else
5969 assem_debug("branch: external\n");
5970 if (internal && dops[(cinfo[i].ba - start) >> 2].is_ds) {
5971 ds_assemble_entry(i);
5972 }
5973 else {
5974 add_to_linker(out,cinfo[i].ba,internal);
5975 emit_jmp(0);
5976 }
5977 }
5978 // branch not taken
5979 if(!unconditional) {
5980 if (!nevertaken) {
5981 assert(nottaken);
5982 set_jump_target(nottaken, out);
5983 }
5984 assem_debug("1:\n");
5985 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5986 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5987 address_generation(i+1,&branch_regs[i],0);
5988 if (ram_offset)
5989 load_reg(regs[i].regmap,branch_regs[i].regmap,ROREG);
5990 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5991 ds_assemble(i+1,&branch_regs[i]);
5992 cc=get_reg(branch_regs[i].regmap,CCREG);
5993 if (cc == -1) {
5994 // Cycle count isn't in a register, temporarily load it then write it out
5995 emit_loadreg(CCREG,HOST_CCREG);
5996 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), HOST_CCREG);
5997 void *jaddr=out;
5998 emit_jns(0);
5999 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6000 emit_storereg(CCREG,HOST_CCREG);
6001 }
6002 else{
6003 cc=get_reg(i_regmap,CCREG);
6004 assert(cc==HOST_CCREG);
6005 emit_addimm_and_set_flags(cinfo[i].ccadj + CLOCK_ADJUST(2), cc);
6006 void *jaddr=out;
6007 emit_jns(0);
6008 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
6009 }
6010 }
6011 }
6012}
6013
6014static void check_regmap(signed char *regmap)
6015{
6016#ifndef NDEBUG
6017 int i,j;
6018 for (i = 0; i < HOST_REGS; i++) {
6019 if (regmap[i] < 0)
6020 continue;
6021 for (j = i + 1; j < HOST_REGS; j++)
6022 assert(regmap[i] != regmap[j]);
6023 }
6024#endif
6025}
6026
6027#ifdef DISASM
6028#include <inttypes.h>
6029static char insn[MAXBLOCK][10];
6030
6031#define set_mnemonic(i_, n_) \
6032 strcpy(insn[i_], n_)
6033
6034void print_regmap(const char *name, const signed char *regmap)
6035{
6036 char buf[5];
6037 int i, l;
6038 fputs(name, stdout);
6039 for (i = 0; i < HOST_REGS; i++) {
6040 l = 0;
6041 if (regmap[i] >= 0)
6042 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6043 for (; l < 3; l++)
6044 buf[l] = ' ';
6045 buf[l] = 0;
6046 printf(" r%d=%s", i, buf);
6047 }
6048 fputs("\n", stdout);
6049}
6050
6051 /* disassembly */
6052void disassemble_inst(int i)
6053{
6054 if (dops[i].bt) printf("*"); else printf(" ");
6055 switch(dops[i].itype) {
6056 case UJUMP:
6057 printf (" %x: %s %8x\n",start+i*4,insn[i],cinfo[i].ba);break;
6058 case CJUMP:
6059 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;
6060 case SJUMP:
6061 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;
6062 case RJUMP:
6063 if (dops[i].opcode2 == 9 && dops[i].rt1 != 31)
6064 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6065 else
6066 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6067 break;
6068 case IMM16:
6069 if(dops[i].opcode==0xf) //LUI
6070 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,cinfo[i].imm&0xffff);
6071 else
6072 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6073 break;
6074 case LOAD:
6075 case LOADLR:
6076 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6077 break;
6078 case STORE:
6079 case STORELR:
6080 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,cinfo[i].imm);
6081 break;
6082 case ALU:
6083 case SHIFT:
6084 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6085 break;
6086 case MULTDIV:
6087 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6088 break;
6089 case SHIFTIMM:
6090 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,cinfo[i].imm);
6091 break;
6092 case MOV:
6093 if((dops[i].opcode2&0x1d)==0x10)
6094 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6095 else if((dops[i].opcode2&0x1d)==0x11)
6096 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6097 else
6098 printf (" %x: %s\n",start+i*4,insn[i]);
6099 break;
6100 case COP0:
6101 if(dops[i].opcode2==0)
6102 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6103 else if(dops[i].opcode2==4)
6104 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6105 else printf (" %x: %s\n",start+i*4,insn[i]);
6106 break;
6107 case COP2:
6108 if(dops[i].opcode2<3)
6109 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6110 else if(dops[i].opcode2>3)
6111 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6112 else printf (" %x: %s\n",start+i*4,insn[i]);
6113 break;
6114 case C2LS:
6115 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,cinfo[i].imm);
6116 break;
6117 case INTCALL:
6118 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6119 break;
6120 default:
6121 //printf (" %s %8x\n",insn[i],source[i]);
6122 printf (" %x: %s\n",start+i*4,insn[i]);
6123 }
6124 #ifndef REGMAP_PRINT
6125 return;
6126 #endif
6127 printf("D: %x WD: %x U: %"PRIx64" hC: %x hWC: %x hLC: %x\n",
6128 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i],
6129 regs[i].isconst, regs[i].wasconst, regs[i].loadedconst);
6130 print_regmap("pre: ", regmap_pre[i]);
6131 print_regmap("entry: ", regs[i].regmap_entry);
6132 print_regmap("map: ", regs[i].regmap);
6133 if (dops[i].is_jump) {
6134 print_regmap("bentry:", branch_regs[i].regmap_entry);
6135 print_regmap("bmap: ", branch_regs[i].regmap);
6136 }
6137}
6138#else
6139#define set_mnemonic(i_, n_)
6140static void disassemble_inst(int i) {}
6141#endif // DISASM
6142
6143#define DRC_TEST_VAL 0x74657374
6144
6145static noinline void new_dynarec_test(void)
6146{
6147 int (*testfunc)(void);
6148 void *beginning;
6149 int ret[2];
6150 size_t i;
6151
6152 // check structure linkage
6153 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6154 {
6155 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6156 }
6157
6158 SysPrintf("(%p) testing if we can run recompiled code @%p...\n",
6159 new_dynarec_test, out);
6160 ((volatile u_int *)NDRC_WRITE_OFFSET(out))[0]++; // make the cache dirty
6161
6162 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6163 out = ndrc->translation_cache;
6164 beginning = start_block();
6165 emit_movimm(DRC_TEST_VAL + i, 0); // test
6166 emit_ret();
6167 literal_pool(0);
6168 end_block(beginning);
6169 testfunc = beginning;
6170 ret[i] = testfunc();
6171 }
6172
6173 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6174 SysPrintf("test passed.\n");
6175 else
6176 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6177 out = ndrc->translation_cache;
6178}
6179
6180// clear the state completely, instead of just marking
6181// things invalid like invalidate_all_pages() does
6182void new_dynarec_clear_full(void)
6183{
6184 int n;
6185 out = ndrc->translation_cache;
6186 memset(invalid_code,1,sizeof(invalid_code));
6187 memset(hash_table,0xff,sizeof(hash_table));
6188 memset(mini_ht,-1,sizeof(mini_ht));
6189 memset(shadow,0,sizeof(shadow));
6190 copy=shadow;
6191 expirep = EXPIRITY_OFFSET;
6192 pending_exception=0;
6193 literalcount=0;
6194 stop_after_jal=0;
6195 inv_code_start=inv_code_end=~0;
6196 hack_addr=0;
6197 f1_hack=0;
6198 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6199 blocks_clear(&blocks[n]);
6200 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6201 free(jumps[n]);
6202 jumps[n] = NULL;
6203 }
6204 stat_clear(stat_blocks);
6205 stat_clear(stat_links);
6206
6207 cycle_multiplier_old = Config.cycle_multiplier;
6208 new_dynarec_hacks_old = new_dynarec_hacks;
6209}
6210
6211void new_dynarec_init(void)
6212{
6213 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6214
6215#ifdef _3DS
6216 check_rosalina();
6217#endif
6218#ifdef BASE_ADDR_DYNAMIC
6219 #ifdef VITA
6220 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6221 if (sceBlock <= 0)
6222 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6223 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6224 if (ret < 0)
6225 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6226 sceKernelOpenVMDomain();
6227 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6228 #elif defined(_MSC_VER)
6229 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6230 PAGE_EXECUTE_READWRITE);
6231 #elif defined(HAVE_LIBNX)
6232 Result rc = jitCreate(&g_jit, sizeof(*ndrc));
6233 if (R_FAILED(rc))
6234 SysPrintf("jitCreate failed: %08x\n", rc);
6235 SysPrintf("jitCreate: RX: %p RW: %p type: %d\n", g_jit.rx_addr, g_jit.rw_addr, g_jit.type);
6236 jitTransitionToWritable(&g_jit);
6237 ndrc = g_jit.rx_addr;
6238 ndrc_write_ofs = (char *)g_jit.rw_addr - (char *)ndrc;
6239 memset(NDRC_WRITE_OFFSET(&ndrc->tramp), 0, sizeof(ndrc->tramp));
6240 #else
6241 uintptr_t desired_addr = 0;
6242 int prot = PROT_READ | PROT_WRITE | PROT_EXEC;
6243 int flags = MAP_PRIVATE | MAP_ANONYMOUS;
6244 int fd = -1;
6245 #ifdef __ELF__
6246 extern char _end;
6247 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6248 #endif
6249 #ifdef TC_WRITE_OFFSET
6250 // mostly for testing
6251 fd = open("/dev/shm/pcsxr", O_CREAT | O_RDWR, 0600);
6252 ftruncate(fd, sizeof(*ndrc));
6253 void *mw = mmap(NULL, sizeof(*ndrc), PROT_READ | PROT_WRITE,
6254 (flags = MAP_SHARED), fd, 0);
6255 assert(mw != MAP_FAILED);
6256 prot = PROT_READ | PROT_EXEC;
6257 #endif
6258 ndrc = mmap((void *)desired_addr, sizeof(*ndrc), prot, flags, fd, 0);
6259 if (ndrc == MAP_FAILED) {
6260 SysPrintf("mmap() failed: %s\n", strerror(errno));
6261 abort();
6262 }
6263 #ifdef TC_WRITE_OFFSET
6264 ndrc_write_ofs = (char *)mw - (char *)ndrc;
6265 #endif
6266 #endif
6267#else
6268 #ifndef NO_WRITE_EXEC
6269 // not all systems allow execute in data segment by default
6270 // size must be 4K aligned for 3DS?
6271 if (mprotect(ndrc, sizeof(*ndrc),
6272 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6273 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6274 #endif
6275#endif
6276 out = ndrc->translation_cache;
6277 new_dynarec_clear_full();
6278#ifdef HOST_IMM8
6279 // Copy this into local area so we don't have to put it in every literal pool
6280 invc_ptr=invalid_code;
6281#endif
6282 arch_init();
6283 new_dynarec_test();
6284 ram_offset = (uintptr_t)psxM - 0x80000000;
6285 if (ram_offset!=0)
6286 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6287 SysPrintf("Mapped (RAM/scrp/ROM/LUTs/TC):\n");
6288 SysPrintf("%p/%p/%p/%p/%p\n", psxM, psxH, psxR, mem_rtab, out);
6289}
6290
6291void new_dynarec_cleanup(void)
6292{
6293 int n;
6294#ifdef BASE_ADDR_DYNAMIC
6295 #ifdef VITA
6296 // sceBlock is managed by retroarch's bootstrap code
6297 //sceKernelFreeMemBlock(sceBlock);
6298 //sceBlock = -1;
6299 #elif defined(HAVE_LIBNX)
6300 jitClose(&g_jit);
6301 ndrc = NULL;
6302 #else
6303 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6304 SysPrintf("munmap() failed\n");
6305 ndrc = NULL;
6306 #endif
6307#endif
6308 for (n = 0; n < ARRAY_SIZE(blocks); n++)
6309 blocks_clear(&blocks[n]);
6310 for (n = 0; n < ARRAY_SIZE(jumps); n++) {
6311 free(jumps[n]);
6312 jumps[n] = NULL;
6313 }
6314 stat_clear(stat_blocks);
6315 stat_clear(stat_links);
6316 new_dynarec_print_stats();
6317}
6318
6319static u_int *get_source_start(u_int addr, u_int *limit)
6320{
6321 if (addr < 0x00800000
6322 || (0x80000000 <= addr && addr < 0x80800000)
6323 || (0xa0000000 <= addr && addr < 0xa0800000))
6324 {
6325 // used for BIOS calls mostly?
6326 *limit = (addr & 0xa0600000) + 0x00200000;
6327 return (u_int *)(psxM + (addr & 0x1fffff));
6328 }
6329 else if (!Config.HLE && (
6330 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6331 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6332 {
6333 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6334 // but timings in PCSX are too tied to the interpreter's 2-per-insn assumption
6335 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6336 cycle_multiplier_active = 200;
6337
6338 *limit = (addr & 0xfff00000) | 0x80000;
6339 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6340 }
6341 return NULL;
6342}
6343
6344static u_int scan_for_ret(u_int addr)
6345{
6346 u_int limit = 0;
6347 u_int *mem;
6348
6349 mem = get_source_start(addr, &limit);
6350 if (mem == NULL)
6351 return addr;
6352
6353 if (limit > addr + 0x1000)
6354 limit = addr + 0x1000;
6355 for (; addr < limit; addr += 4, mem++) {
6356 if (*mem == 0x03e00008) // jr $ra
6357 return addr + 8;
6358 }
6359 return addr;
6360}
6361
6362struct savestate_block {
6363 uint32_t addr;
6364 uint32_t regflags;
6365};
6366
6367static int addr_cmp(const void *p1_, const void *p2_)
6368{
6369 const struct savestate_block *p1 = p1_, *p2 = p2_;
6370 return p1->addr - p2->addr;
6371}
6372
6373int new_dynarec_save_blocks(void *save, int size)
6374{
6375 struct savestate_block *sblocks = save;
6376 int maxcount = size / sizeof(sblocks[0]);
6377 struct savestate_block tmp_blocks[1024];
6378 struct block_info *block;
6379 int p, s, d, o, bcnt;
6380 u_int addr;
6381
6382 o = 0;
6383 for (p = 0; p < ARRAY_SIZE(blocks); p++) {
6384 bcnt = 0;
6385 for (block = blocks[p]; block != NULL; block = block->next) {
6386 if (block->is_dirty)
6387 continue;
6388 tmp_blocks[bcnt].addr = block->start;
6389 tmp_blocks[bcnt].regflags = block->reg_sv_flags;
6390 bcnt++;
6391 }
6392 if (bcnt < 1)
6393 continue;
6394 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6395
6396 addr = tmp_blocks[0].addr;
6397 for (s = d = 0; s < bcnt; s++) {
6398 if (tmp_blocks[s].addr < addr)
6399 continue;
6400 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6401 tmp_blocks[d++] = tmp_blocks[s];
6402 addr = scan_for_ret(tmp_blocks[s].addr);
6403 }
6404
6405 if (o + d > maxcount)
6406 d = maxcount - o;
6407 memcpy(&sblocks[o], tmp_blocks, d * sizeof(sblocks[0]));
6408 o += d;
6409 }
6410
6411 return o * sizeof(sblocks[0]);
6412}
6413
6414void new_dynarec_load_blocks(const void *save, int size)
6415{
6416 const struct savestate_block *sblocks = save;
6417 int count = size / sizeof(sblocks[0]);
6418 struct block_info *block;
6419 u_int regs_save[32];
6420 u_int page;
6421 uint32_t f;
6422 int i, b;
6423
6424 // restore clean blocks, if any
6425 for (page = 0, b = i = 0; page < ARRAY_SIZE(blocks); page++) {
6426 for (block = blocks[page]; block != NULL; block = block->next, b++) {
6427 if (!block->is_dirty)
6428 continue;
6429 assert(block->source && block->copy);
6430 if (memcmp(block->source, block->copy, block->len))
6431 continue;
6432
6433 // see try_restore_block
6434 block->is_dirty = 0;
6435 mark_invalid_code(block->start, block->len, 0);
6436 i++;
6437 }
6438 }
6439 inv_debug("load_blocks: %d/%d clean blocks\n", i, b);
6440
6441 // change GPRs for speculation to at least partially work..
6442 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6443 for (i = 1; i < 32; i++)
6444 psxRegs.GPR.r[i] = 0x80000000;
6445
6446 for (b = 0; b < count; b++) {
6447 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6448 if (f & 1)
6449 psxRegs.GPR.r[i] = 0x1f800000;
6450 }
6451
6452 ndrc_get_addr_ht(sblocks[b].addr);
6453
6454 for (f = sblocks[b].regflags, i = 0; f; f >>= 1, i++) {
6455 if (f & 1)
6456 psxRegs.GPR.r[i] = 0x80000000;
6457 }
6458 }
6459
6460 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6461}
6462
6463void new_dynarec_print_stats(void)
6464{
6465#ifdef STAT_PRINT
6466 printf("cc %3d,%3d,%3d lu%6d,%3d,%3d c%3d inv%3d,%3d tc_offs %zu b %u,%u\n",
6467 stat_bc_pre, stat_bc_direct, stat_bc_restore,
6468 stat_ht_lookups, stat_jump_in_lookups, stat_restore_tries,
6469 stat_restore_compares, stat_inv_addr_calls, stat_inv_hits,
6470 out - ndrc->translation_cache, stat_blocks, stat_links);
6471 stat_bc_direct = stat_bc_pre = stat_bc_restore =
6472 stat_ht_lookups = stat_jump_in_lookups = stat_restore_tries =
6473 stat_restore_compares = stat_inv_addr_calls = stat_inv_hits = 0;
6474#endif
6475}
6476
6477static int apply_hacks(void)
6478{
6479 int i;
6480 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6481 return 0;
6482 /* special hack(s) */
6483 for (i = 0; i < slen - 4; i++)
6484 {
6485 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6486 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6487 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6488 && cinfo[i+3].imm == 0x6e40 && dops[i+3].rs1 == 2)
6489 {
6490 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6491 dops[i + 3].itype = NOP;
6492 }
6493 }
6494 i = slen;
6495 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6496 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6497 && dops[i-7].itype == STORE)
6498 {
6499 i = i-8;
6500 if (dops[i].itype == IMM16)
6501 i--;
6502 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6503 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6504 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6505 {
6506 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6507 f1_hack = 1;
6508 return 1;
6509 }
6510 }
6511 return 0;
6512}
6513
6514static int is_ld_use_hazard(int ld_rt, const struct decoded_insn *op)
6515{
6516 return ld_rt != 0 && (ld_rt == op->rs1 || ld_rt == op->rs2)
6517 && op->itype != LOADLR && op->itype != CJUMP && op->itype != SJUMP;
6518}
6519
6520static void force_intcall(int i)
6521{
6522 memset(&dops[i], 0, sizeof(dops[i]));
6523 dops[i].itype = INTCALL;
6524 dops[i].rs1 = CCREG;
6525 dops[i].is_exception = 1;
6526 cinfo[i].ba = -1;
6527}
6528
6529static void disassemble_one(int i, u_int src)
6530{
6531 unsigned int type, op, op2, op3;
6532 enum ls_width_type ls_type = LS_32;
6533 memset(&dops[i], 0, sizeof(dops[i]));
6534 memset(&cinfo[i], 0, sizeof(cinfo[i]));
6535 cinfo[i].ba = -1;
6536 cinfo[i].addr = -1;
6537 dops[i].opcode = op = src >> 26;
6538 op2 = 0;
6539 type = INTCALL;
6540 set_mnemonic(i, "???");
6541 switch(op)
6542 {
6543 case 0x00: set_mnemonic(i, "special");
6544 op2 = src & 0x3f;
6545 switch(op2)
6546 {
6547 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6548 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6549 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6550 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6551 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6552 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6553 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6554 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6555 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6556 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6557 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6558 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6559 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6560 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6561 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6562 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6563 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6564 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6565 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6566 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6567 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6568 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6569 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6570 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6571 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6572 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6573 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6574 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6575 }
6576 break;
6577 case 0x01: set_mnemonic(i, "regimm");
6578 type = SJUMP;
6579 op2 = (src >> 16) & 0x1f;
6580 switch(op2)
6581 {
6582 case 0x10: set_mnemonic(i, "BLTZAL"); break;
6583 case 0x11: set_mnemonic(i, "BGEZAL"); break;
6584 default:
6585 if (op2 & 1)
6586 set_mnemonic(i, "BGEZ");
6587 else
6588 set_mnemonic(i, "BLTZ");
6589 }
6590 break;
6591 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6592 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6593 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6594 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6595 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6596 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6597 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6598 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6599 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6600 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6601 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6602 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6603 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6604 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6605 case 0x10: set_mnemonic(i, "COP0");
6606 op2 = (src >> 21) & 0x1f;
6607 if (op2 & 0x10) {
6608 op3 = src & 0x1f;
6609 switch (op3)
6610 {
6611 case 0x01: case 0x02: case 0x06: case 0x08: type = INTCALL; break;
6612 case 0x10: set_mnemonic(i, "RFE"); type=RFE; break;
6613 default: type = OTHER; break;
6614 }
6615 break;
6616 }
6617 switch(op2)
6618 {
6619 u32 rd;
6620 case 0x00:
6621 set_mnemonic(i, "MFC0");
6622 rd = (src >> 11) & 0x1F;
6623 if (!(0x00000417u & (1u << rd)))
6624 type = COP0;
6625 break;
6626 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6627 case 0x02:
6628 case 0x06: type = INTCALL; break;
6629 default: type = OTHER; break;
6630 }
6631 break;
6632 case 0x11: set_mnemonic(i, "COP1");
6633 op2 = (src >> 21) & 0x1f;
6634 break;
6635 case 0x12: set_mnemonic(i, "COP2");
6636 op2 = (src >> 21) & 0x1f;
6637 if (op2 & 0x10) {
6638 type = OTHER;
6639 if (gte_handlers[src & 0x3f] != NULL) {
6640#ifdef DISASM
6641 if (gte_regnames[src & 0x3f] != NULL)
6642 strcpy(insn[i], gte_regnames[src & 0x3f]);
6643 else
6644 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", src & 0x3f);
6645#endif
6646 type = C2OP;
6647 }
6648 }
6649 else switch(op2)
6650 {
6651 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6652 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6653 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6654 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6655 }
6656 break;
6657 case 0x13: set_mnemonic(i, "COP3");
6658 op2 = (src >> 21) & 0x1f;
6659 break;
6660 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6661 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6662 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6663 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6664 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6665 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6666 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6667 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6668 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6669 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6670 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6671 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6672 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6673 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6674 case 0x3B:
6675 if (Config.HLE && (src & 0x03ffffff) < ARRAY_SIZE(psxHLEt)) {
6676 set_mnemonic(i, "HLECALL");
6677 type = HLECALL;
6678 }
6679 break;
6680 default:
6681 break;
6682 }
6683 if (type == INTCALL)
6684 SysPrintf("NI %08x @%08x (%08x)\n", src, start + i*4, start);
6685 dops[i].itype=type;
6686 dops[i].opcode2=op2;
6687 /* Get registers/immediates */
6688 dops[i].use_lt1=0;
6689 gte_rs[i]=gte_rt[i]=0;
6690 dops[i].rs1 = 0;
6691 dops[i].rs2 = 0;
6692 dops[i].rt1 = 0;
6693 dops[i].rt2 = 0;
6694 switch(type) {
6695 case LOAD:
6696 dops[i].rs1 = (src >> 21) & 0x1f;
6697 dops[i].rt1 = (src >> 16) & 0x1f;
6698 cinfo[i].imm = (short)src;
6699 break;
6700 case STORE:
6701 case STORELR:
6702 dops[i].rs1 = (src >> 21) & 0x1f;
6703 dops[i].rs2 = (src >> 16) & 0x1f;
6704 cinfo[i].imm = (short)src;
6705 break;
6706 case LOADLR:
6707 // LWL/LWR only load part of the register,
6708 // therefore the target register must be treated as a source too
6709 dops[i].rs1 = (src >> 21) & 0x1f;
6710 dops[i].rs2 = (src >> 16) & 0x1f;
6711 dops[i].rt1 = (src >> 16) & 0x1f;
6712 cinfo[i].imm = (short)src;
6713 break;
6714 case IMM16:
6715 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6716 else dops[i].rs1 = (src >> 21) & 0x1f;
6717 dops[i].rs2 = 0;
6718 dops[i].rt1 = (src >> 16) & 0x1f;
6719 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6720 cinfo[i].imm = (unsigned short)src;
6721 }else{
6722 cinfo[i].imm = (short)src;
6723 }
6724 break;
6725 case UJUMP:
6726 // The JAL instruction writes to r31.
6727 if (op&1) {
6728 dops[i].rt1=31;
6729 }
6730 dops[i].rs2=CCREG;
6731 break;
6732 case RJUMP:
6733 dops[i].rs1 = (src >> 21) & 0x1f;
6734 // The JALR instruction writes to rd.
6735 if (op2&1) {
6736 dops[i].rt1 = (src >> 11) & 0x1f;
6737 }
6738 dops[i].rs2=CCREG;
6739 break;
6740 case CJUMP:
6741 dops[i].rs1 = (src >> 21) & 0x1f;
6742 dops[i].rs2 = (src >> 16) & 0x1f;
6743 if(op&2) { // BGTZ/BLEZ
6744 dops[i].rs2=0;
6745 }
6746 break;
6747 case SJUMP:
6748 dops[i].rs1 = (src >> 21) & 0x1f;
6749 dops[i].rs2 = CCREG;
6750 if (op2 == 0x10 || op2 == 0x11) { // BxxAL
6751 dops[i].rt1 = 31;
6752 // NOTE: If the branch is not taken, r31 is still overwritten
6753 }
6754 break;
6755 case ALU:
6756 dops[i].rs1=(src>>21)&0x1f; // source
6757 dops[i].rs2=(src>>16)&0x1f; // subtract amount
6758 dops[i].rt1=(src>>11)&0x1f; // destination
6759 break;
6760 case MULTDIV:
6761 dops[i].rs1=(src>>21)&0x1f; // source
6762 dops[i].rs2=(src>>16)&0x1f; // divisor
6763 dops[i].rt1=HIREG;
6764 dops[i].rt2=LOREG;
6765 break;
6766 case MOV:
6767 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6768 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6769 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6770 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6771 if((op2&0x1d)==0x10) dops[i].rt1=(src>>11)&0x1f; // MFxx
6772 if((op2&0x1d)==0x11) dops[i].rs1=(src>>21)&0x1f; // MTxx
6773 break;
6774 case SHIFT:
6775 dops[i].rs1=(src>>16)&0x1f; // target of shift
6776 dops[i].rs2=(src>>21)&0x1f; // shift amount
6777 dops[i].rt1=(src>>11)&0x1f; // destination
6778 break;
6779 case SHIFTIMM:
6780 dops[i].rs1=(src>>16)&0x1f;
6781 dops[i].rs2=0;
6782 dops[i].rt1=(src>>11)&0x1f;
6783 cinfo[i].imm=(src>>6)&0x1f;
6784 break;
6785 case COP0:
6786 if(op2==0) dops[i].rt1=(src>>16)&0x1F; // MFC0
6787 if(op2==4) dops[i].rs1=(src>>16)&0x1F; // MTC0
6788 if(op2==4&&((src>>11)&0x1e)==12) dops[i].rs2=CCREG;
6789 break;
6790 case COP2:
6791 if(op2<3) dops[i].rt1=(src>>16)&0x1F; // MFC2/CFC2
6792 if(op2>3) dops[i].rs1=(src>>16)&0x1F; // MTC2/CTC2
6793 int gr=(src>>11)&0x1F;
6794 switch(op2)
6795 {
6796 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6797 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6798 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6799 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6800 }
6801 break;
6802 case C2LS:
6803 dops[i].rs1=(src>>21)&0x1F;
6804 cinfo[i].imm=(short)src;
6805 if(op==0x32) gte_rt[i]=1ll<<((src>>16)&0x1F); // LWC2
6806 else gte_rs[i]=1ll<<((src>>16)&0x1F); // SWC2
6807 break;
6808 case C2OP:
6809 gte_rs[i]=gte_reg_reads[src&0x3f];
6810 gte_rt[i]=gte_reg_writes[src&0x3f];
6811 gte_rt[i]|=1ll<<63; // every op changes flags
6812 if((src&0x3f)==GTE_MVMVA) {
6813 int v = (src >> 15) & 3;
6814 gte_rs[i]&=~0xe3fll;
6815 if(v==3) gte_rs[i]|=0xe00ll;
6816 else gte_rs[i]|=3ll<<(v*2);
6817 }
6818 break;
6819 case SYSCALL:
6820 case HLECALL:
6821 case INTCALL:
6822 dops[i].rs1=CCREG;
6823 break;
6824 default:
6825 break;
6826 }
6827}
6828
6829static noinline void pass1_disassemble(u_int pagelimit)
6830{
6831 int i, j, done = 0, ni_count = 0;
6832
6833 for (i = 0; !done; i++)
6834 {
6835 int force_j_to_interpreter = 0;
6836 unsigned int type, op, op2;
6837
6838 disassemble_one(i, source[i]);
6839 type = dops[i].itype;
6840 op = dops[i].opcode;
6841 op2 = dops[i].opcode2;
6842
6843 /* Calculate branch target addresses */
6844 if(type==UJUMP)
6845 cinfo[i].ba=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6846 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
6847 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6848 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
6849 cinfo[i].ba=start+i*4+8; // Ignore never taken branch
6850 else if(type==CJUMP||type==SJUMP)
6851 cinfo[i].ba=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6852
6853 /* simplify always (not)taken branches */
6854 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
6855 dops[i].rs1 = dops[i].rs2 = 0;
6856 if (!(op & 1)) {
6857 dops[i].itype = type = UJUMP;
6858 dops[i].rs2 = CCREG;
6859 }
6860 }
6861 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
6862 dops[i].itype = type = UJUMP;
6863
6864 dops[i].is_jump = type == RJUMP || type == UJUMP || type == CJUMP || type == SJUMP;
6865 dops[i].is_ujump = type == RJUMP || type == UJUMP;
6866 dops[i].is_load = type == LOAD || type == LOADLR || op == 0x32; // LWC2
6867 dops[i].is_delay_load = (dops[i].is_load || (source[i] & 0xf3d00000) == 0x40000000); // MFC/CFC
6868 dops[i].is_store = type == STORE || type == STORELR || op == 0x3a; // SWC2
6869 dops[i].is_exception = type == SYSCALL || type == HLECALL || type == INTCALL;
6870 dops[i].may_except = dops[i].is_exception || (type == ALU && (op2 == 0x20 || op2 == 0x22)) || op == 8;
6871
6872 if (((op & 0x37) == 0x21 || op == 0x25) // LH/SH/LHU
6873 && ((cinfo[i].imm & 1) || Config.PreciseExceptions))
6874 dops[i].may_except = 1;
6875 if (((op & 0x37) == 0x23 || (op & 0x37) == 0x32) // LW/SW/LWC2/SWC2
6876 && ((cinfo[i].imm & 3) || Config.PreciseExceptions))
6877 dops[i].may_except = 1;
6878
6879 /* rare messy cases to just pass over to the interpreter */
6880 if (i > 0 && dops[i-1].is_jump) {
6881 j = i - 1;
6882 // branch in delay slot?
6883 if (dops[i].is_jump) {
6884 // don't handle first branch and call interpreter if it's hit
6885 SysPrintf("branch in DS @%08x (%08x)\n", start + i*4, start);
6886 force_j_to_interpreter = 1;
6887 }
6888 // load delay detection through a branch
6889 else if (dops[i].is_delay_load && dops[i].rt1 != 0) {
6890 const struct decoded_insn *dop = NULL;
6891 int t = -1;
6892 if (cinfo[i-1].ba != -1) {
6893 t = (cinfo[i-1].ba - start) / 4;
6894 if (t < 0 || t > i) {
6895 u_int limit = 0;
6896 u_int *mem = get_source_start(cinfo[i-1].ba, &limit);
6897 if (mem != NULL) {
6898 disassemble_one(MAXBLOCK - 1, mem[0]);
6899 dop = &dops[MAXBLOCK - 1];
6900 }
6901 }
6902 else
6903 dop = &dops[t];
6904 }
6905 if ((dop && is_ld_use_hazard(dops[i].rt1, dop))
6906 || (!dop && Config.PreciseExceptions)) {
6907 // jump target wants DS result - potential load delay effect
6908 SysPrintf("load delay in DS @%08x (%08x)\n", start + i*4, start);
6909 force_j_to_interpreter = 1;
6910 if (0 <= t && t < i)
6911 dops[t + 1].bt = 1; // expected return from interpreter
6912 }
6913 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&&
6914 !(i>=3&&dops[i-3].is_jump)) {
6915 // v0 overwrite like this is a sign of trouble, bail out
6916 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
6917 force_j_to_interpreter = 1;
6918 }
6919 }
6920 }
6921 else if (i > 0 && dops[i-1].is_delay_load
6922 && is_ld_use_hazard(dops[i-1].rt1, &dops[i])
6923 && (i < 2 || !dops[i-2].is_ujump)) {
6924 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
6925 for (j = i - 1; j > 0 && dops[j-1].is_delay_load; j--)
6926 if (dops[j-1].rt1 != dops[i-1].rt1)
6927 break;
6928 force_j_to_interpreter = 1;
6929 }
6930 if (force_j_to_interpreter) {
6931 force_intcall(j);
6932 done = 2;
6933 i = j; // don't compile the problematic branch/load/etc
6934 }
6935 if (dops[i].is_exception && i > 0 && dops[i-1].is_jump) {
6936 SysPrintf("exception in DS @%08x (%08x)\n", start + i*4, start);
6937 i--;
6938 force_intcall(i);
6939 done = 2;
6940 }
6941 if (i >= 2 && (source[i-2] & 0xffe0f800) == 0x40806000) // MTC0 $12
6942 dops[i].bt = 1;
6943 if (i >= 1 && (source[i-1] & 0xffe0f800) == 0x40806800) // MTC0 $13
6944 dops[i].bt = 1;
6945
6946 /* Is this the end of the block? */
6947 if (i > 0 && dops[i-1].is_ujump) {
6948 if (dops[i-1].rt1 == 0) { // not jal
6949 int found_bbranch = 0, t = (cinfo[i-1].ba - start) / 4;
6950 if ((u_int)(t - i) < 64 && start + (t+64)*4 < pagelimit) {
6951 // scan for a branch back to i+1
6952 for (j = t; j < t + 64; j++) {
6953 int tmpop = source[j] >> 26;
6954 if (tmpop == 1 || ((tmpop & ~3) == 4)) {
6955 int t2 = j + 1 + (int)(signed short)source[j];
6956 if (t2 == i + 1) {
6957 //printf("blk expand %08x<-%08x\n", start + (i+1)*4, start + j*4);
6958 found_bbranch = 1;
6959 break;
6960 }
6961 }
6962 }
6963 }
6964 if (!found_bbranch)
6965 done = 2;
6966 }
6967 else {
6968 if(stop_after_jal) done=1;
6969 // Stop on BREAK
6970 if((source[i+1]&0xfc00003f)==0x0d) done=1;
6971 }
6972 // Don't recompile stuff that's already compiled
6973 if(check_addr(start+i*4+4)) done=1;
6974 // Don't get too close to the limit
6975 if (i > MAXBLOCK - 64)
6976 done = 1;
6977 }
6978 if (dops[i].itype == HLECALL)
6979 stop = 1;
6980 else if (dops[i].itype == INTCALL)
6981 stop = 2;
6982 else if (dops[i].is_exception)
6983 done = stop_after_jal ? 1 : 2;
6984 if (done == 2) {
6985 // Does the block continue due to a branch?
6986 for(j=i-1;j>=0;j--)
6987 {
6988 if(cinfo[j].ba==start+i*4) done=j=0; // Branch into delay slot
6989 if(cinfo[j].ba==start+i*4+4) done=j=0;
6990 if(cinfo[j].ba==start+i*4+8) done=j=0;
6991 }
6992 }
6993 //assert(i<MAXBLOCK-1);
6994 if(start+i*4==pagelimit-4) done=1;
6995 assert(start+i*4<pagelimit);
6996 if (i == MAXBLOCK - 2)
6997 done = 1;
6998 // Stop if we're compiling junk
6999 if (dops[i].itype == INTCALL && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7000 done=stop_after_jal=1;
7001 SysPrintf("Disabled speculative precompilation\n");
7002 }
7003 }
7004 while (i > 0 && dops[i-1].is_jump)
7005 i--;
7006 assert(i > 0);
7007 assert(!dops[i-1].is_jump);
7008 slen = i;
7009}
7010
7011// Basic liveness analysis for MIPS registers
7012static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7013{
7014 int i;
7015 uint64_t u,gte_u,b,gte_b;
7016 uint64_t temp_u,temp_gte_u=0;
7017 uint64_t gte_u_unknown=0;
7018 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7019 gte_u_unknown=~0ll;
7020 if(iend==slen-1) {
7021 u=1;
7022 gte_u=gte_u_unknown;
7023 }else{
7024 //u=unneeded_reg[iend+1];
7025 u=1;
7026 gte_u=gte_unneeded[iend+1];
7027 }
7028
7029 for (i=iend;i>=istart;i--)
7030 {
7031 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7032 if(dops[i].is_jump)
7033 {
7034 // If subroutine call, flag return address as a possible branch target
7035 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7036
7037 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7038 {
7039 // Branch out of this block, flush all regs
7040 u=1;
7041 gte_u=gte_u_unknown;
7042 branch_unneeded_reg[i]=u;
7043 // Merge in delay slot
7044 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7045 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7046 u|=1;
7047 gte_u|=gte_rt[i+1];
7048 gte_u&=~gte_rs[i+1];
7049 }
7050 else
7051 {
7052 // Internal branch, flag target
7053 dops[(cinfo[i].ba-start)>>2].bt=1;
7054 if(cinfo[i].ba<=start+i*4) {
7055 // Backward branch
7056 if(dops[i].is_ujump)
7057 {
7058 // Unconditional branch
7059 temp_u=1;
7060 temp_gte_u=0;
7061 } else {
7062 // Conditional branch (not taken case)
7063 temp_u=unneeded_reg[i+2];
7064 temp_gte_u&=gte_unneeded[i+2];
7065 }
7066 // Merge in delay slot
7067 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7068 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7069 temp_u|=1;
7070 temp_gte_u|=gte_rt[i+1];
7071 temp_gte_u&=~gte_rs[i+1];
7072 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7073 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7074 temp_u|=1;
7075 temp_gte_u|=gte_rt[i];
7076 temp_gte_u&=~gte_rs[i];
7077 unneeded_reg[i]=temp_u;
7078 gte_unneeded[i]=temp_gte_u;
7079 // Only go three levels deep. This recursion can take an
7080 // excessive amount of time if there are a lot of nested loops.
7081 if(r<2) {
7082 pass2_unneeded_regs((cinfo[i].ba-start)>>2,i-1,r+1);
7083 }else{
7084 unneeded_reg[(cinfo[i].ba-start)>>2]=1;
7085 gte_unneeded[(cinfo[i].ba-start)>>2]=gte_u_unknown;
7086 }
7087 } /*else*/ if(1) {
7088 if (dops[i].is_ujump)
7089 {
7090 // Unconditional branch
7091 u=unneeded_reg[(cinfo[i].ba-start)>>2];
7092 gte_u=gte_unneeded[(cinfo[i].ba-start)>>2];
7093 branch_unneeded_reg[i]=u;
7094 // Merge in delay slot
7095 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7096 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7097 u|=1;
7098 gte_u|=gte_rt[i+1];
7099 gte_u&=~gte_rs[i+1];
7100 } else {
7101 // Conditional branch
7102 b=unneeded_reg[(cinfo[i].ba-start)>>2];
7103 gte_b=gte_unneeded[(cinfo[i].ba-start)>>2];
7104 branch_unneeded_reg[i]=b;
7105 // Branch delay slot
7106 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7107 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7108 b|=1;
7109 gte_b|=gte_rt[i+1];
7110 gte_b&=~gte_rs[i+1];
7111 u&=b;
7112 gte_u&=gte_b;
7113 if(i<slen-1) {
7114 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7115 } else {
7116 branch_unneeded_reg[i]=1;
7117 }
7118 }
7119 }
7120 }
7121 }
7122 //u=1; // DEBUG
7123 // Written registers are unneeded
7124 u|=1LL<<dops[i].rt1;
7125 u|=1LL<<dops[i].rt2;
7126 gte_u|=gte_rt[i];
7127 // Accessed registers are needed
7128 u&=~(1LL<<dops[i].rs1);
7129 u&=~(1LL<<dops[i].rs2);
7130 gte_u&=~gte_rs[i];
7131 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7132 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7133 if (dops[i].may_except || dops[i].itype == RFE)
7134 {
7135 // SYSCALL instruction, etc or conditional exception
7136 u=1;
7137 }
7138 // Source-target dependencies
7139 // R0 is always unneeded
7140 u|=1;
7141 // Save it
7142 unneeded_reg[i]=u;
7143 gte_unneeded[i]=gte_u;
7144 /*
7145 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7146 printf("U:");
7147 int r;
7148 for(r=1;r<=CCREG;r++) {
7149 if((unneeded_reg[i]>>r)&1) {
7150 if(r==HIREG) printf(" HI");
7151 else if(r==LOREG) printf(" LO");
7152 else printf(" r%d",r);
7153 }
7154 }
7155 printf("\n");
7156 */
7157 }
7158}
7159
7160static noinline void pass3_register_alloc(u_int addr)
7161{
7162 struct regstat current; // Current register allocations/status
7163 clear_all_regs(current.regmap_entry);
7164 clear_all_regs(current.regmap);
7165 current.wasdirty = current.dirty = 0;
7166 current.u = unneeded_reg[0];
7167 alloc_reg(&current, 0, CCREG);
7168 dirty_reg(&current, CCREG);
7169 current.wasconst = 0;
7170 current.isconst = 0;
7171 current.loadedconst = 0;
7172 current.noevict = 0;
7173 //current.waswritten = 0;
7174 int ds=0;
7175 int cc=0;
7176 int hr;
7177 int i, j;
7178
7179 if (addr & 1) {
7180 // First instruction is delay slot
7181 cc=-1;
7182 dops[1].bt=1;
7183 ds=1;
7184 unneeded_reg[0]=1;
7185 }
7186
7187 for(i=0;i<slen;i++)
7188 {
7189 if(dops[i].bt)
7190 {
7191 for(hr=0;hr<HOST_REGS;hr++)
7192 {
7193 // Is this really necessary?
7194 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7195 }
7196 current.isconst=0;
7197 //current.waswritten=0;
7198 }
7199
7200 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7201 regs[i].wasconst=current.isconst;
7202 regs[i].wasdirty=current.dirty;
7203 regs[i].dirty=0;
7204 regs[i].u=0;
7205 regs[i].isconst=0;
7206 regs[i].loadedconst=0;
7207 if (!dops[i].is_jump) {
7208 if(i+1<slen) {
7209 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7210 current.u|=1;
7211 } else {
7212 current.u=1;
7213 }
7214 } else {
7215 if(i+1<slen) {
7216 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7217 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7218 current.u|=1;
7219 } else {
7220 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7221 abort();
7222 }
7223 }
7224 dops[i].is_ds=ds;
7225 if(ds) {
7226 ds=0; // Skip delay slot, already allocated as part of branch
7227 // ...but we need to alloc it in case something jumps here
7228 if(i+1<slen) {
7229 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7230 }else{
7231 current.u=branch_unneeded_reg[i-1];
7232 }
7233 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7234 current.u|=1;
7235 struct regstat temp;
7236 memcpy(&temp,&current,sizeof(current));
7237 temp.wasdirty=temp.dirty;
7238 // TODO: Take into account unconditional branches, as below
7239 delayslot_alloc(&temp,i);
7240 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7241 regs[i].wasdirty=temp.wasdirty;
7242 regs[i].dirty=temp.dirty;
7243 regs[i].isconst=0;
7244 regs[i].wasconst=0;
7245 current.isconst=0;
7246 // Create entry (branch target) regmap
7247 for(hr=0;hr<HOST_REGS;hr++)
7248 {
7249 int r=temp.regmap[hr];
7250 if(r>=0) {
7251 if(r!=regmap_pre[i][hr]) {
7252 regs[i].regmap_entry[hr]=-1;
7253 }
7254 else
7255 {
7256 assert(r < 64);
7257 if((current.u>>r)&1) {
7258 regs[i].regmap_entry[hr]=-1;
7259 regs[i].regmap[hr]=-1;
7260 //Don't clear regs in the delay slot as the branch might need them
7261 //current.regmap[hr]=-1;
7262 }else
7263 regs[i].regmap_entry[hr]=r;
7264 }
7265 } else {
7266 // First instruction expects CCREG to be allocated
7267 if(i==0&&hr==HOST_CCREG)
7268 regs[i].regmap_entry[hr]=CCREG;
7269 else
7270 regs[i].regmap_entry[hr]=-1;
7271 }
7272 }
7273 }
7274 else { // Not delay slot
7275 current.noevict = 0;
7276 switch(dops[i].itype) {
7277 case UJUMP:
7278 //current.isconst=0; // DEBUG
7279 //current.wasconst=0; // DEBUG
7280 //regs[i].wasconst=0; // DEBUG
7281 clear_const(&current,dops[i].rt1);
7282 alloc_cc(&current,i);
7283 dirty_reg(&current,CCREG);
7284 if (dops[i].rt1==31) {
7285 alloc_reg(&current,i,31);
7286 dirty_reg(&current,31);
7287 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7288 //assert(dops[i+1].rt1!=dops[i].rt1);
7289 #ifdef REG_PREFETCH
7290 alloc_reg(&current,i,PTEMP);
7291 #endif
7292 }
7293 dops[i].ooo=1;
7294 delayslot_alloc(&current,i+1);
7295 //current.isconst=0; // DEBUG
7296 ds=1;
7297 break;
7298 case RJUMP:
7299 //current.isconst=0;
7300 //current.wasconst=0;
7301 //regs[i].wasconst=0;
7302 clear_const(&current,dops[i].rs1);
7303 clear_const(&current,dops[i].rt1);
7304 alloc_cc(&current,i);
7305 dirty_reg(&current,CCREG);
7306 if (!ds_writes_rjump_rs(i)) {
7307 alloc_reg(&current,i,dops[i].rs1);
7308 if (dops[i].rt1!=0) {
7309 alloc_reg(&current,i,dops[i].rt1);
7310 dirty_reg(&current,dops[i].rt1);
7311 #ifdef REG_PREFETCH
7312 alloc_reg(&current,i,PTEMP);
7313 #endif
7314 }
7315 #ifdef USE_MINI_HT
7316 if(dops[i].rs1==31) { // JALR
7317 alloc_reg(&current,i,RHASH);
7318 alloc_reg(&current,i,RHTBL);
7319 }
7320 #endif
7321 delayslot_alloc(&current,i+1);
7322 } else {
7323 // The delay slot overwrites our source register,
7324 // allocate a temporary register to hold the old value.
7325 current.isconst=0;
7326 current.wasconst=0;
7327 regs[i].wasconst=0;
7328 delayslot_alloc(&current,i+1);
7329 current.isconst=0;
7330 alloc_reg(&current,i,RTEMP);
7331 }
7332 //current.isconst=0; // DEBUG
7333 dops[i].ooo=1;
7334 ds=1;
7335 break;
7336 case CJUMP:
7337 //current.isconst=0;
7338 //current.wasconst=0;
7339 //regs[i].wasconst=0;
7340 clear_const(&current,dops[i].rs1);
7341 clear_const(&current,dops[i].rs2);
7342 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7343 {
7344 alloc_cc(&current,i);
7345 dirty_reg(&current,CCREG);
7346 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7347 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7348 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7349 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7350 // The delay slot overwrites one of our conditions.
7351 // Allocate the branch condition registers instead.
7352 current.isconst=0;
7353 current.wasconst=0;
7354 regs[i].wasconst=0;
7355 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7356 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7357 }
7358 else
7359 {
7360 dops[i].ooo=1;
7361 delayslot_alloc(&current,i+1);
7362 }
7363 }
7364 else
7365 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7366 {
7367 alloc_cc(&current,i);
7368 dirty_reg(&current,CCREG);
7369 alloc_reg(&current,i,dops[i].rs1);
7370 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7371 // The delay slot overwrites one of our conditions.
7372 // Allocate the branch condition registers instead.
7373 current.isconst=0;
7374 current.wasconst=0;
7375 regs[i].wasconst=0;
7376 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7377 }
7378 else
7379 {
7380 dops[i].ooo=1;
7381 delayslot_alloc(&current,i+1);
7382 }
7383 }
7384 else
7385 // Don't alloc the delay slot yet because we might not execute it
7386 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7387 {
7388 current.isconst=0;
7389 current.wasconst=0;
7390 regs[i].wasconst=0;
7391 alloc_cc(&current,i);
7392 dirty_reg(&current,CCREG);
7393 alloc_reg(&current,i,dops[i].rs1);
7394 alloc_reg(&current,i,dops[i].rs2);
7395 }
7396 else
7397 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7398 {
7399 current.isconst=0;
7400 current.wasconst=0;
7401 regs[i].wasconst=0;
7402 alloc_cc(&current,i);
7403 dirty_reg(&current,CCREG);
7404 alloc_reg(&current,i,dops[i].rs1);
7405 }
7406 ds=1;
7407 //current.isconst=0;
7408 break;
7409 case SJUMP:
7410 clear_const(&current,dops[i].rs1);
7411 clear_const(&current,dops[i].rt1);
7412 {
7413 alloc_cc(&current,i);
7414 dirty_reg(&current,CCREG);
7415 alloc_reg(&current,i,dops[i].rs1);
7416 if (dops[i].rt1 == 31) { // BLTZAL/BGEZAL
7417 alloc_reg(&current,i,31);
7418 dirty_reg(&current,31);
7419 }
7420 if ((dops[i].rs1 &&
7421 (dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7422 ||(dops[i].rt1 == 31 && dops[i].rs1 == 31) // overwrites it's own condition
7423 ||(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
7424 // Allocate the branch condition registers instead.
7425 current.isconst=0;
7426 current.wasconst=0;
7427 regs[i].wasconst=0;
7428 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7429 }
7430 else
7431 {
7432 dops[i].ooo=1;
7433 delayslot_alloc(&current,i+1);
7434 }
7435 }
7436 ds=1;
7437 //current.isconst=0;
7438 break;
7439 case IMM16:
7440 imm16_alloc(&current,i);
7441 break;
7442 case LOAD:
7443 case LOADLR:
7444 load_alloc(&current,i);
7445 break;
7446 case STORE:
7447 case STORELR:
7448 store_alloc(&current,i);
7449 break;
7450 case ALU:
7451 alu_alloc(&current,i);
7452 break;
7453 case SHIFT:
7454 shift_alloc(&current,i);
7455 break;
7456 case MULTDIV:
7457 multdiv_alloc(&current,i);
7458 break;
7459 case SHIFTIMM:
7460 shiftimm_alloc(&current,i);
7461 break;
7462 case MOV:
7463 mov_alloc(&current,i);
7464 break;
7465 case COP0:
7466 cop0_alloc(&current,i);
7467 break;
7468 case RFE:
7469 rfe_alloc(&current,i);
7470 break;
7471 case COP2:
7472 cop2_alloc(&current,i);
7473 break;
7474 case C2LS:
7475 c2ls_alloc(&current,i);
7476 break;
7477 case C2OP:
7478 c2op_alloc(&current,i);
7479 break;
7480 case SYSCALL:
7481 case HLECALL:
7482 case INTCALL:
7483 syscall_alloc(&current,i);
7484 break;
7485 }
7486
7487 // Create entry (branch target) regmap
7488 for(hr=0;hr<HOST_REGS;hr++)
7489 {
7490 int r,or;
7491 r=current.regmap[hr];
7492 if(r>=0) {
7493 if(r!=regmap_pre[i][hr]) {
7494 // TODO: delay slot (?)
7495 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7496 if(or<0||r>=TEMPREG){
7497 regs[i].regmap_entry[hr]=-1;
7498 }
7499 else
7500 {
7501 // Just move it to a different register
7502 regs[i].regmap_entry[hr]=r;
7503 // If it was dirty before, it's still dirty
7504 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7505 }
7506 }
7507 else
7508 {
7509 // Unneeded
7510 if(r==0){
7511 regs[i].regmap_entry[hr]=0;
7512 }
7513 else
7514 {
7515 assert(r<64);
7516 if((current.u>>r)&1) {
7517 regs[i].regmap_entry[hr]=-1;
7518 //regs[i].regmap[hr]=-1;
7519 current.regmap[hr]=-1;
7520 }else
7521 regs[i].regmap_entry[hr]=r;
7522 }
7523 }
7524 } else {
7525 // Branches expect CCREG to be allocated at the target
7526 if(regmap_pre[i][hr]==CCREG)
7527 regs[i].regmap_entry[hr]=CCREG;
7528 else
7529 regs[i].regmap_entry[hr]=-1;
7530 }
7531 }
7532 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7533 }
7534
7535#if 0 // see do_store_smc_check()
7536 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)
7537 current.waswritten|=1<<dops[i-1].rs1;
7538 current.waswritten&=~(1<<dops[i].rt1);
7539 current.waswritten&=~(1<<dops[i].rt2);
7540 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)cinfo[i].imm>=0x800)
7541 current.waswritten&=~(1<<dops[i].rs1);
7542#endif
7543
7544 /* Branch post-alloc */
7545 if(i>0)
7546 {
7547 current.wasdirty=current.dirty;
7548 switch(dops[i-1].itype) {
7549 case UJUMP:
7550 memcpy(&branch_regs[i-1],&current,sizeof(current));
7551 branch_regs[i-1].isconst=0;
7552 branch_regs[i-1].wasconst=0;
7553 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7554 alloc_cc(&branch_regs[i-1],i-1);
7555 dirty_reg(&branch_regs[i-1],CCREG);
7556 if(dops[i-1].rt1==31) { // JAL
7557 alloc_reg(&branch_regs[i-1],i-1,31);
7558 dirty_reg(&branch_regs[i-1],31);
7559 }
7560 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7561 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7562 break;
7563 case RJUMP:
7564 memcpy(&branch_regs[i-1],&current,sizeof(current));
7565 branch_regs[i-1].isconst=0;
7566 branch_regs[i-1].wasconst=0;
7567 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7568 alloc_cc(&branch_regs[i-1],i-1);
7569 dirty_reg(&branch_regs[i-1],CCREG);
7570 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7571 if(dops[i-1].rt1!=0) { // JALR
7572 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7573 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7574 }
7575 #ifdef USE_MINI_HT
7576 if(dops[i-1].rs1==31) { // JALR
7577 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7578 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7579 }
7580 #endif
7581 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7582 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7583 break;
7584 case CJUMP:
7585 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7586 {
7587 alloc_cc(&current,i-1);
7588 dirty_reg(&current,CCREG);
7589 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7590 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7591 // The delay slot overwrote one of our conditions
7592 // Delay slot goes after the test (in order)
7593 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7594 current.u|=1;
7595 delayslot_alloc(&current,i);
7596 current.isconst=0;
7597 }
7598 else
7599 {
7600 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7601 // Alloc the branch condition registers
7602 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7603 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7604 }
7605 memcpy(&branch_regs[i-1],&current,sizeof(current));
7606 branch_regs[i-1].isconst=0;
7607 branch_regs[i-1].wasconst=0;
7608 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7609 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7610 }
7611 else
7612 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7613 {
7614 alloc_cc(&current,i-1);
7615 dirty_reg(&current,CCREG);
7616 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7617 // The delay slot overwrote the branch condition
7618 // Delay slot goes after the test (in order)
7619 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7620 current.u|=1;
7621 delayslot_alloc(&current,i);
7622 current.isconst=0;
7623 }
7624 else
7625 {
7626 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7627 // Alloc the branch condition register
7628 alloc_reg(&current,i-1,dops[i-1].rs1);
7629 }
7630 memcpy(&branch_regs[i-1],&current,sizeof(current));
7631 branch_regs[i-1].isconst=0;
7632 branch_regs[i-1].wasconst=0;
7633 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7634 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7635 }
7636 break;
7637 case SJUMP:
7638 {
7639 alloc_cc(&current,i-1);
7640 dirty_reg(&current,CCREG);
7641 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7642 // The delay slot overwrote the branch condition
7643 // Delay slot goes after the test (in order)
7644 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7645 current.u|=1;
7646 delayslot_alloc(&current,i);
7647 current.isconst=0;
7648 }
7649 else
7650 {
7651 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7652 // Alloc the branch condition register
7653 alloc_reg(&current,i-1,dops[i-1].rs1);
7654 }
7655 memcpy(&branch_regs[i-1],&current,sizeof(current));
7656 branch_regs[i-1].isconst=0;
7657 branch_regs[i-1].wasconst=0;
7658 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7659 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7660 }
7661 break;
7662 }
7663
7664 if (dops[i-1].is_ujump)
7665 {
7666 if(dops[i-1].rt1==31) // JAL/JALR
7667 {
7668 // Subroutine call will return here, don't alloc any registers
7669 current.dirty=0;
7670 clear_all_regs(current.regmap);
7671 alloc_reg(&current,i,CCREG);
7672 dirty_reg(&current,CCREG);
7673 }
7674 else if(i+1<slen)
7675 {
7676 // Internal branch will jump here, match registers to caller
7677 current.dirty=0;
7678 clear_all_regs(current.regmap);
7679 alloc_reg(&current,i,CCREG);
7680 dirty_reg(&current,CCREG);
7681 for(j=i-1;j>=0;j--)
7682 {
7683 if(cinfo[j].ba==start+i*4+4) {
7684 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7685 current.dirty=branch_regs[j].dirty;
7686 break;
7687 }
7688 }
7689 while(j>=0) {
7690 if(cinfo[j].ba==start+i*4+4) {
7691 for(hr=0;hr<HOST_REGS;hr++) {
7692 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7693 current.regmap[hr]=-1;
7694 }
7695 current.dirty&=branch_regs[j].dirty;
7696 }
7697 }
7698 j--;
7699 }
7700 }
7701 }
7702 }
7703
7704 // Count cycles in between branches
7705 cinfo[i].ccadj = CLOCK_ADJUST(cc);
7706 if (i > 0 && (dops[i-1].is_jump || dops[i].is_exception))
7707 {
7708 cc=0;
7709 }
7710#if !defined(DRC_DBG)
7711 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7712 {
7713 // this should really be removed since the real stalls have been implemented,
7714 // but doing so causes sizeable perf regression against the older version
7715 u_int gtec = gte_cycletab[source[i] & 0x3f];
7716 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7717 }
7718 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7719 {
7720 cc+=4;
7721 }
7722 else if(dops[i].itype==C2LS)
7723 {
7724 // same as with C2OP
7725 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7726 }
7727#endif
7728 else
7729 {
7730 cc++;
7731 }
7732
7733 if(!dops[i].is_ds) {
7734 regs[i].dirty=current.dirty;
7735 regs[i].isconst=current.isconst;
7736 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7737 }
7738 for(hr=0;hr<HOST_REGS;hr++) {
7739 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7740 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7741 regs[i].wasconst&=~(1<<hr);
7742 }
7743 }
7744 }
7745 //regs[i].waswritten=current.waswritten;
7746 }
7747}
7748
7749static noinline void pass4_cull_unused_regs(void)
7750{
7751 u_int last_needed_regs[4] = {0,0,0,0};
7752 u_int nr=0;
7753 int i;
7754
7755 for (i=slen-1;i>=0;i--)
7756 {
7757 int hr;
7758 __builtin_prefetch(regs[i-2].regmap);
7759 if(dops[i].is_jump)
7760 {
7761 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
7762 {
7763 // Branch out of this block, don't need anything
7764 nr=0;
7765 }
7766 else
7767 {
7768 // Internal branch
7769 // Need whatever matches the target
7770 nr=0;
7771 int t=(cinfo[i].ba-start)>>2;
7772 for(hr=0;hr<HOST_REGS;hr++)
7773 {
7774 if(regs[i].regmap_entry[hr]>=0) {
7775 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7776 }
7777 }
7778 }
7779 // Conditional branch may need registers for following instructions
7780 if (!dops[i].is_ujump)
7781 {
7782 if(i<slen-2) {
7783 nr |= last_needed_regs[(i+2) & 3];
7784 for(hr=0;hr<HOST_REGS;hr++)
7785 {
7786 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7787 //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]);
7788 }
7789 }
7790 }
7791 // Don't need stuff which is overwritten
7792 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7793 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7794 // Merge in delay slot
7795 if (dops[i+1].rt1) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt1);
7796 if (dops[i+1].rt2) nr &= ~get_regm(regs[i].regmap, dops[i+1].rt2);
7797 nr |= get_regm(regmap_pre[i], dops[i+1].rs1);
7798 nr |= get_regm(regmap_pre[i], dops[i+1].rs2);
7799 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs1);
7800 nr |= get_regm(regs[i].regmap_entry, dops[i+1].rs2);
7801 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
7802 nr |= get_regm(regmap_pre[i], ROREG);
7803 nr |= get_regm(regs[i].regmap_entry, ROREG);
7804 }
7805 if (dops[i+1].is_store) {
7806 nr |= get_regm(regmap_pre[i], INVCP);
7807 nr |= get_regm(regs[i].regmap_entry, INVCP);
7808 }
7809 }
7810 else if (dops[i].is_exception)
7811 {
7812 // SYSCALL instruction, etc
7813 nr=0;
7814 }
7815 else // Non-branch
7816 {
7817 if(i<slen-1) {
7818 for(hr=0;hr<HOST_REGS;hr++) {
7819 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7820 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7821 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7822 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7823 }
7824 }
7825 }
7826 // Overwritten registers are not needed
7827 if (dops[i].rt1) nr &= ~get_regm(regs[i].regmap, dops[i].rt1);
7828 if (dops[i].rt2) nr &= ~get_regm(regs[i].regmap, dops[i].rt2);
7829 nr &= ~get_regm(regs[i].regmap, FTEMP);
7830 // Source registers are needed
7831 nr |= get_regm(regmap_pre[i], dops[i].rs1);
7832 nr |= get_regm(regmap_pre[i], dops[i].rs2);
7833 nr |= get_regm(regs[i].regmap_entry, dops[i].rs1);
7834 nr |= get_regm(regs[i].regmap_entry, dops[i].rs2);
7835 if (ram_offset && (dops[i].is_load || dops[i].is_store)) {
7836 nr |= get_regm(regmap_pre[i], ROREG);
7837 nr |= get_regm(regs[i].regmap_entry, ROREG);
7838 }
7839 if (dops[i].is_store) {
7840 nr |= get_regm(regmap_pre[i], INVCP);
7841 nr |= get_regm(regs[i].regmap_entry, INVCP);
7842 }
7843
7844 if (i > 0 && !dops[i].bt && regs[i].wasdirty)
7845 for(hr=0;hr<HOST_REGS;hr++)
7846 {
7847 // Don't store a register immediately after writing it,
7848 // may prevent dual-issue.
7849 // But do so if this is a branch target, otherwise we
7850 // might have to load the register before the branch.
7851 if((regs[i].wasdirty>>hr)&1) {
7852 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7853 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
7854 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
7855 }
7856 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7857 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
7858 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
7859 }
7860 }
7861 }
7862 // Cycle count is needed at branches. Assume it is needed at the target too.
7863 if (i == 0 || dops[i].bt || dops[i].may_except || dops[i].itype == CJUMP) {
7864 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7865 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7866 }
7867 // Save it
7868 last_needed_regs[i & 3] = nr;
7869
7870 // Deallocate unneeded registers
7871 for(hr=0;hr<HOST_REGS;hr++)
7872 {
7873 if(!((nr>>hr)&1)) {
7874 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
7875 if(dops[i].is_jump)
7876 {
7877 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
7878 if (dops[i+1].is_load || dops[i+1].is_store)
7879 map1 = ROREG;
7880 if (dops[i+1].is_store)
7881 map2 = INVCP;
7882 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
7883 temp = FTEMP;
7884 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
7885 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
7886 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
7887 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
7888 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
7889 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
7890 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
7891 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
7892 {
7893 regs[i].regmap[hr]=-1;
7894 regs[i].isconst&=~(1<<hr);
7895 regs[i].dirty&=~(1<<hr);
7896 regs[i+1].wasdirty&=~(1<<hr);
7897 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
7898 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
7899 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
7900 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
7901 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
7902 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
7903 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
7904 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
7905 {
7906 branch_regs[i].regmap[hr]=-1;
7907 branch_regs[i].regmap_entry[hr]=-1;
7908 if (!dops[i].is_ujump)
7909 {
7910 if (i < slen-2) {
7911 regmap_pre[i+2][hr]=-1;
7912 regs[i+2].wasconst&=~(1<<hr);
7913 }
7914 }
7915 }
7916 }
7917 }
7918 else
7919 {
7920 // Non-branch
7921 if(i>0)
7922 {
7923 int map1 = -1, map2 = -1, temp=-1;
7924 if (dops[i].is_load || dops[i].is_store)
7925 map1 = ROREG;
7926 if (dops[i].is_store)
7927 map2 = INVCP;
7928 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
7929 temp = FTEMP;
7930 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
7931 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
7932 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
7933 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
7934 regs[i].regmap[hr] != CCREG)
7935 {
7936 if(i<slen-1&&!dops[i].is_ds) {
7937 assert(regs[i].regmap[hr]<64);
7938 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
7939 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
7940 {
7941 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
7942 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
7943 }
7944 regmap_pre[i+1][hr]=-1;
7945 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
7946 regs[i+1].wasconst&=~(1<<hr);
7947 }
7948 regs[i].regmap[hr]=-1;
7949 regs[i].isconst&=~(1<<hr);
7950 regs[i].dirty&=~(1<<hr);
7951 regs[i+1].wasdirty&=~(1<<hr);
7952 }
7953 }
7954 }
7955 } // if needed
7956 } // for hr
7957 }
7958}
7959
7960// If a register is allocated during a loop, try to allocate it for the
7961// entire loop, if possible. This avoids loading/storing registers
7962// inside of the loop.
7963static noinline void pass5a_preallocate1(void)
7964{
7965 int i, j, hr;
7966 signed char f_regmap[HOST_REGS];
7967 clear_all_regs(f_regmap);
7968 for(i=0;i<slen-1;i++)
7969 {
7970 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
7971 {
7972 if(cinfo[i].ba>=start && cinfo[i].ba<(start+i*4))
7973 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
7974 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
7975 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR
7976 ||dops[i+1].itype==SHIFT
7977 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
7978 {
7979 int t=(cinfo[i].ba-start)>>2;
7980 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
7981 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
7982 for(hr=0;hr<HOST_REGS;hr++)
7983 {
7984 if(regs[i].regmap[hr]>=0) {
7985 if(f_regmap[hr]!=regs[i].regmap[hr]) {
7986 // dealloc old register
7987 int n;
7988 for(n=0;n<HOST_REGS;n++)
7989 {
7990 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
7991 }
7992 // and alloc new one
7993 f_regmap[hr]=regs[i].regmap[hr];
7994 }
7995 }
7996 if(branch_regs[i].regmap[hr]>=0) {
7997 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
7998 // dealloc old register
7999 int n;
8000 for(n=0;n<HOST_REGS;n++)
8001 {
8002 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8003 }
8004 // and alloc new one
8005 f_regmap[hr]=branch_regs[i].regmap[hr];
8006 }
8007 }
8008 if(dops[i].ooo) {
8009 if(count_free_regs(regs[i].regmap)<=cinfo[i+1].min_free_regs)
8010 f_regmap[hr]=branch_regs[i].regmap[hr];
8011 }else{
8012 if(count_free_regs(branch_regs[i].regmap)<=cinfo[i+1].min_free_regs)
8013 f_regmap[hr]=branch_regs[i].regmap[hr];
8014 }
8015 // Avoid dirty->clean transition
8016 #ifdef DESTRUCTIVE_WRITEBACK
8017 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;
8018 #endif
8019 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8020 // case above, however it's always a good idea. We can't hoist the
8021 // load if the register was already allocated, so there's no point
8022 // wasting time analyzing most of these cases. It only "succeeds"
8023 // when the mapping was different and the load can be replaced with
8024 // a mov, which is of negligible benefit. So such cases are
8025 // skipped below.
8026 if(f_regmap[hr]>0) {
8027 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8028 int r=f_regmap[hr];
8029 for(j=t;j<=i;j++)
8030 {
8031 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8032 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8033 assert(r < 64);
8034 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8035 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,cinfo[i].ba,start+j*4,hr,r);
8036 int k;
8037 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8038 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8039 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8040 k=i;
8041 while(k>1&&regs[k-1].regmap[hr]==-1) {
8042 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8043 //printf("no free regs for store %x\n",start+(k-1)*4);
8044 break;
8045 }
8046 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8047 //printf("no-match due to different register\n");
8048 break;
8049 }
8050 if (dops[k-2].is_jump) {
8051 //printf("no-match due to branch\n");
8052 break;
8053 }
8054 // call/ret fast path assumes no registers allocated
8055 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8056 break;
8057 }
8058 k--;
8059 }
8060 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8061 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8062 while(k<i) {
8063 regs[k].regmap_entry[hr]=f_regmap[hr];
8064 regs[k].regmap[hr]=f_regmap[hr];
8065 regmap_pre[k+1][hr]=f_regmap[hr];
8066 regs[k].wasdirty&=~(1<<hr);
8067 regs[k].dirty&=~(1<<hr);
8068 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8069 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8070 regs[k].wasconst&=~(1<<hr);
8071 regs[k].isconst&=~(1<<hr);
8072 k++;
8073 }
8074 }
8075 else {
8076 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8077 break;
8078 }
8079 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8080 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8081 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8082 regs[i].regmap_entry[hr]=f_regmap[hr];
8083 regs[i].regmap[hr]=f_regmap[hr];
8084 regs[i].wasdirty&=~(1<<hr);
8085 regs[i].dirty&=~(1<<hr);
8086 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8087 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8088 regs[i].wasconst&=~(1<<hr);
8089 regs[i].isconst&=~(1<<hr);
8090 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8091 branch_regs[i].wasdirty&=~(1<<hr);
8092 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8093 branch_regs[i].regmap[hr]=f_regmap[hr];
8094 branch_regs[i].dirty&=~(1<<hr);
8095 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8096 branch_regs[i].wasconst&=~(1<<hr);
8097 branch_regs[i].isconst&=~(1<<hr);
8098 if (!dops[i].is_ujump) {
8099 regmap_pre[i+2][hr]=f_regmap[hr];
8100 regs[i+2].wasdirty&=~(1<<hr);
8101 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8102 }
8103 }
8104 }
8105 for(k=t;k<j;k++) {
8106 // Alloc register clean at beginning of loop,
8107 // but may dirty it in pass 6
8108 regs[k].regmap_entry[hr]=f_regmap[hr];
8109 regs[k].regmap[hr]=f_regmap[hr];
8110 regs[k].dirty&=~(1<<hr);
8111 regs[k].wasconst&=~(1<<hr);
8112 regs[k].isconst&=~(1<<hr);
8113 if (dops[k].is_jump) {
8114 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8115 branch_regs[k].regmap[hr]=f_regmap[hr];
8116 branch_regs[k].dirty&=~(1<<hr);
8117 branch_regs[k].wasconst&=~(1<<hr);
8118 branch_regs[k].isconst&=~(1<<hr);
8119 if (!dops[k].is_ujump) {
8120 regmap_pre[k+2][hr]=f_regmap[hr];
8121 regs[k+2].wasdirty&=~(1<<hr);
8122 }
8123 }
8124 else
8125 {
8126 regmap_pre[k+1][hr]=f_regmap[hr];
8127 regs[k+1].wasdirty&=~(1<<hr);
8128 }
8129 }
8130 if(regs[j].regmap[hr]==f_regmap[hr])
8131 regs[j].regmap_entry[hr]=f_regmap[hr];
8132 break;
8133 }
8134 if(j==i) break;
8135 if(regs[j].regmap[hr]>=0)
8136 break;
8137 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8138 //printf("no-match due to different register\n");
8139 break;
8140 }
8141 if (dops[j].is_ujump)
8142 {
8143 // Stop on unconditional branch
8144 break;
8145 }
8146 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8147 {
8148 if(dops[j].ooo) {
8149 if(count_free_regs(regs[j].regmap)<=cinfo[j+1].min_free_regs)
8150 break;
8151 }else{
8152 if(count_free_regs(branch_regs[j].regmap)<=cinfo[j+1].min_free_regs)
8153 break;
8154 }
8155 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8156 //printf("no-match due to different register (branch)\n");
8157 break;
8158 }
8159 }
8160 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8161 //printf("No free regs for store %x\n",start+j*4);
8162 break;
8163 }
8164 assert(f_regmap[hr]<64);
8165 }
8166 }
8167 }
8168 }
8169 }
8170 }else{
8171 // Non branch or undetermined branch target
8172 for(hr=0;hr<HOST_REGS;hr++)
8173 {
8174 if(hr!=EXCLUDE_REG) {
8175 if(regs[i].regmap[hr]>=0) {
8176 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8177 // dealloc old register
8178 int n;
8179 for(n=0;n<HOST_REGS;n++)
8180 {
8181 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8182 }
8183 // and alloc new one
8184 f_regmap[hr]=regs[i].regmap[hr];
8185 }
8186 }
8187 }
8188 }
8189 // Try to restore cycle count at branch targets
8190 if(dops[i].bt) {
8191 for(j=i;j<slen-1;j++) {
8192 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8193 if(count_free_regs(regs[j].regmap)<=cinfo[j].min_free_regs) {
8194 //printf("no free regs for store %x\n",start+j*4);
8195 break;
8196 }
8197 }
8198 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8199 int k=i;
8200 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8201 while(k<j) {
8202 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8203 regs[k].regmap[HOST_CCREG]=CCREG;
8204 regmap_pre[k+1][HOST_CCREG]=CCREG;
8205 regs[k+1].wasdirty|=1<<HOST_CCREG;
8206 regs[k].dirty|=1<<HOST_CCREG;
8207 regs[k].wasconst&=~(1<<HOST_CCREG);
8208 regs[k].isconst&=~(1<<HOST_CCREG);
8209 k++;
8210 }
8211 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8212 }
8213 // Work backwards from the branch target
8214 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8215 {
8216 //printf("Extend backwards\n");
8217 int k;
8218 k=i;
8219 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8220 if(count_free_regs(regs[k-1].regmap)<=cinfo[k-1].min_free_regs) {
8221 //printf("no free regs for store %x\n",start+(k-1)*4);
8222 break;
8223 }
8224 k--;
8225 }
8226 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8227 //printf("Extend CC, %x ->\n",start+k*4);
8228 while(k<=i) {
8229 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8230 regs[k].regmap[HOST_CCREG]=CCREG;
8231 regmap_pre[k+1][HOST_CCREG]=CCREG;
8232 regs[k+1].wasdirty|=1<<HOST_CCREG;
8233 regs[k].dirty|=1<<HOST_CCREG;
8234 regs[k].wasconst&=~(1<<HOST_CCREG);
8235 regs[k].isconst&=~(1<<HOST_CCREG);
8236 k++;
8237 }
8238 }
8239 else {
8240 //printf("Fail Extend CC, %x ->\n",start+k*4);
8241 }
8242 }
8243 }
8244 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=SHIFT&&
8245 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8246 dops[i].itype!=IMM16&&dops[i].itype!=LOAD)
8247 {
8248 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8249 }
8250 }
8251 }
8252}
8253
8254// This allocates registers (if possible) one instruction prior
8255// to use, which can avoid a load-use penalty on certain CPUs.
8256static noinline void pass5b_preallocate2(void)
8257{
8258 int i, hr;
8259 for(i=0;i<slen-1;i++)
8260 {
8261 if (!i || !dops[i-1].is_jump)
8262 {
8263 if(!dops[i+1].bt)
8264 {
8265 int j, can_steal = 1;
8266 for (j = i; j < i + 2; j++) {
8267 int free_regs = 0;
8268 if (cinfo[j].min_free_regs == 0)
8269 continue;
8270 for (hr = 0; hr < HOST_REGS; hr++)
8271 if (hr != EXCLUDE_REG && regs[j].regmap[hr] < 0)
8272 free_regs++;
8273 if (free_regs <= cinfo[j].min_free_regs) {
8274 can_steal = 0;
8275 break;
8276 }
8277 }
8278 if (!can_steal)
8279 continue;
8280 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8281 ||(dops[i].itype==COP2&&dops[i].opcode2<3))
8282 {
8283 if(dops[i+1].rs1) {
8284 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8285 {
8286 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8287 {
8288 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8289 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8290 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8291 regs[i].isconst&=~(1<<hr);
8292 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8293 constmap[i][hr]=constmap[i+1][hr];
8294 regs[i+1].wasdirty&=~(1<<hr);
8295 regs[i].dirty&=~(1<<hr);
8296 }
8297 }
8298 }
8299 if(dops[i+1].rs2) {
8300 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8301 {
8302 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8303 {
8304 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8305 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8306 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8307 regs[i].isconst&=~(1<<hr);
8308 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8309 constmap[i][hr]=constmap[i+1][hr];
8310 regs[i+1].wasdirty&=~(1<<hr);
8311 regs[i].dirty&=~(1<<hr);
8312 }
8313 }
8314 }
8315 // Preload target address for load instruction (non-constant)
8316 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8317 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8318 {
8319 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8320 {
8321 regs[i].regmap[hr]=dops[i+1].rs1;
8322 regmap_pre[i+1][hr]=dops[i+1].rs1;
8323 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8324 regs[i].isconst&=~(1<<hr);
8325 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8326 constmap[i][hr]=constmap[i+1][hr];
8327 regs[i+1].wasdirty&=~(1<<hr);
8328 regs[i].dirty&=~(1<<hr);
8329 }
8330 }
8331 }
8332 // Load source into target register
8333 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8334 if((hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1))>=0)
8335 {
8336 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8337 {
8338 regs[i].regmap[hr]=dops[i+1].rs1;
8339 regmap_pre[i+1][hr]=dops[i+1].rs1;
8340 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8341 regs[i].isconst&=~(1<<hr);
8342 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8343 constmap[i][hr]=constmap[i+1][hr];
8344 regs[i+1].wasdirty&=~(1<<hr);
8345 regs[i].dirty&=~(1<<hr);
8346 }
8347 }
8348 }
8349 // Address for store instruction (non-constant)
8350 if (dops[i+1].is_store) { // SB/SH/SW/SWC2
8351 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8352 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8353 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8354 else {
8355 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8356 regs[i+1].isconst&=~(1<<hr);
8357 regs[i+1].dirty&=~(1<<hr);
8358 regs[i+2].wasdirty&=~(1<<hr);
8359 }
8360 assert(hr>=0);
8361 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8362 {
8363 regs[i].regmap[hr]=dops[i+1].rs1;
8364 regmap_pre[i+1][hr]=dops[i+1].rs1;
8365 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8366 regs[i].isconst&=~(1<<hr);
8367 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8368 constmap[i][hr]=constmap[i+1][hr];
8369 regs[i+1].wasdirty&=~(1<<hr);
8370 regs[i].dirty&=~(1<<hr);
8371 }
8372 }
8373 }
8374 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) { // LWC2
8375 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8376 int nr;
8377 hr=get_reg(regs[i+1].regmap,FTEMP);
8378 assert(hr>=0);
8379 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8380 {
8381 regs[i].regmap[hr]=dops[i+1].rs1;
8382 regmap_pre[i+1][hr]=dops[i+1].rs1;
8383 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8384 regs[i].isconst&=~(1<<hr);
8385 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8386 constmap[i][hr]=constmap[i+1][hr];
8387 regs[i+1].wasdirty&=~(1<<hr);
8388 regs[i].dirty&=~(1<<hr);
8389 }
8390 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8391 {
8392 // move it to another register
8393 regs[i+1].regmap[hr]=-1;
8394 regmap_pre[i+2][hr]=-1;
8395 regs[i+1].regmap[nr]=FTEMP;
8396 regmap_pre[i+2][nr]=FTEMP;
8397 regs[i].regmap[nr]=dops[i+1].rs1;
8398 regmap_pre[i+1][nr]=dops[i+1].rs1;
8399 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8400 regs[i].isconst&=~(1<<nr);
8401 regs[i+1].isconst&=~(1<<nr);
8402 regs[i].dirty&=~(1<<nr);
8403 regs[i+1].wasdirty&=~(1<<nr);
8404 regs[i+1].dirty&=~(1<<nr);
8405 regs[i+2].wasdirty&=~(1<<nr);
8406 }
8407 }
8408 }
8409 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*/) {
8410 hr = -1;
8411 if(dops[i+1].itype==LOAD)
8412 hr=get_reg_w(regs[i+1].regmap, dops[i+1].rt1);
8413 if (dops[i+1].itype == LOADLR || dops[i+1].opcode == 0x32) // LWC2
8414 hr=get_reg(regs[i+1].regmap,FTEMP);
8415 if (dops[i+1].is_store) {
8416 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8417 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8418 }
8419 if(hr>=0&&regs[i].regmap[hr]<0) {
8420 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8421 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8422 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8423 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8424 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8425 regs[i].isconst&=~(1<<hr);
8426 regs[i+1].wasdirty&=~(1<<hr);
8427 regs[i].dirty&=~(1<<hr);
8428 }
8429 }
8430 }
8431 }
8432 }
8433 }
8434 }
8435}
8436
8437// Write back dirty registers as soon as we will no longer modify them,
8438// so that we don't end up with lots of writes at the branches.
8439static noinline void pass6_clean_registers(int istart, int iend, int wr)
8440{
8441 static u_int wont_dirty[MAXBLOCK];
8442 static u_int will_dirty[MAXBLOCK];
8443 int i;
8444 int r;
8445 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8446 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8447 if(iend==slen-1) {
8448 will_dirty_i=will_dirty_next=0;
8449 wont_dirty_i=wont_dirty_next=0;
8450 }else{
8451 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8452 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8453 }
8454 for (i=iend;i>=istart;i--)
8455 {
8456 signed char rregmap_i[RRMAP_SIZE];
8457 u_int hr_candirty = 0;
8458 assert(HOST_REGS < 32);
8459 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8460 __builtin_prefetch(regs[i-1].regmap);
8461 if(dops[i].is_jump)
8462 {
8463 signed char branch_rregmap_i[RRMAP_SIZE];
8464 u_int branch_hr_candirty = 0;
8465 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8466 if(cinfo[i].ba<start || cinfo[i].ba>=(start+slen*4))
8467 {
8468 // Branch out of this block, flush all regs
8469 will_dirty_i = 0;
8470 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8471 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8472 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8473 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8474 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8475 will_dirty_i &= branch_hr_candirty;
8476 if (dops[i].is_ujump)
8477 {
8478 // Unconditional branch
8479 wont_dirty_i = 0;
8480 // Merge in delay slot (will dirty)
8481 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8482 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8483 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8484 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8485 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8486 will_dirty_i &= hr_candirty;
8487 }
8488 else
8489 {
8490 // Conditional branch
8491 wont_dirty_i = wont_dirty_next;
8492 // Merge in delay slot (will dirty)
8493 // (the original code had no explanation why these 2 are commented out)
8494 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8495 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8496 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8497 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8498 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8499 will_dirty_i &= hr_candirty;
8500 }
8501 // Merge in delay slot (wont dirty)
8502 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8503 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8504 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8505 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8506 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8507 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8508 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8509 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8510 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8511 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8512 wont_dirty_i &= ~(1u << 31);
8513 if(wr) {
8514 #ifndef DESTRUCTIVE_WRITEBACK
8515 branch_regs[i].dirty&=wont_dirty_i;
8516 #endif
8517 branch_regs[i].dirty|=will_dirty_i;
8518 }
8519 }
8520 else
8521 {
8522 // Internal branch
8523 if(cinfo[i].ba<=start+i*4) {
8524 // Backward branch
8525 if (dops[i].is_ujump)
8526 {
8527 // Unconditional branch
8528 temp_will_dirty=0;
8529 temp_wont_dirty=0;
8530 // Merge in delay slot (will dirty)
8531 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8532 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8533 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8534 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8535 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8536 temp_will_dirty &= branch_hr_candirty;
8537 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8538 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8539 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8540 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8541 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8542 temp_will_dirty &= hr_candirty;
8543 } else {
8544 // Conditional branch (not taken case)
8545 temp_will_dirty=will_dirty_next;
8546 temp_wont_dirty=wont_dirty_next;
8547 // Merge in delay slot (will dirty)
8548 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8549 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8550 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8551 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8552 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8553 temp_will_dirty &= branch_hr_candirty;
8554 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8555 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8556 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8557 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8558 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8559 temp_will_dirty &= hr_candirty;
8560 }
8561 // Merge in delay slot (wont dirty)
8562 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8563 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8564 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8565 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8566 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8567 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8568 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8569 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8570 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8571 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8572 temp_wont_dirty &= ~(1u << 31);
8573 // Deal with changed mappings
8574 if(i<iend) {
8575 for(r=0;r<HOST_REGS;r++) {
8576 if(r!=EXCLUDE_REG) {
8577 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8578 temp_will_dirty&=~(1<<r);
8579 temp_wont_dirty&=~(1<<r);
8580 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8581 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8582 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8583 } else {
8584 temp_will_dirty|=1<<r;
8585 temp_wont_dirty|=1<<r;
8586 }
8587 }
8588 }
8589 }
8590 }
8591 if(wr) {
8592 will_dirty[i]=temp_will_dirty;
8593 wont_dirty[i]=temp_wont_dirty;
8594 pass6_clean_registers((cinfo[i].ba-start)>>2,i-1,0);
8595 }else{
8596 // Limit recursion. It can take an excessive amount
8597 // of time if there are a lot of nested loops.
8598 will_dirty[(cinfo[i].ba-start)>>2]=0;
8599 wont_dirty[(cinfo[i].ba-start)>>2]=-1;
8600 }
8601 }
8602 /*else*/ if(1)
8603 {
8604 if (dops[i].is_ujump)
8605 {
8606 // Unconditional branch
8607 will_dirty_i=0;
8608 wont_dirty_i=0;
8609 //if(cinfo[i].ba>start+i*4) { // Disable recursion (for debugging)
8610 for(r=0;r<HOST_REGS;r++) {
8611 if(r!=EXCLUDE_REG) {
8612 if(branch_regs[i].regmap[r]==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8613 will_dirty_i|=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8614 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8615 }
8616 if(branch_regs[i].regmap[r]>=0) {
8617 will_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8618 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8619 }
8620 }
8621 }
8622 //}
8623 // Merge in delay slot
8624 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8625 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8626 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8627 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8628 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8629 will_dirty_i &= branch_hr_candirty;
8630 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8631 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8632 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8633 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8634 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8635 will_dirty_i &= hr_candirty;
8636 } else {
8637 // Conditional branch
8638 will_dirty_i=will_dirty_next;
8639 wont_dirty_i=wont_dirty_next;
8640 //if(cinfo[i].ba>start+i*4) // Disable recursion (for debugging)
8641 for(r=0;r<HOST_REGS;r++) {
8642 if(r!=EXCLUDE_REG) {
8643 signed char target_reg=branch_regs[i].regmap[r];
8644 if(target_reg==regs[(cinfo[i].ba-start)>>2].regmap_entry[r]) {
8645 will_dirty_i&=will_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8646 wont_dirty_i|=wont_dirty[(cinfo[i].ba-start)>>2]&(1<<r);
8647 }
8648 else if(target_reg>=0) {
8649 will_dirty_i&=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8650 wont_dirty_i|=((unneeded_reg[(cinfo[i].ba-start)>>2]>>target_reg)&1)<<r;
8651 }
8652 }
8653 }
8654 // Merge in delay slot
8655 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8656 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8657 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8658 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8659 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8660 will_dirty_i &= branch_hr_candirty;
8661 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8662 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8663 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8664 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8665 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8666 will_dirty_i &= hr_candirty;
8667 }
8668 // Merge in delay slot (won't dirty)
8669 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8670 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8671 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8672 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8673 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8674 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8675 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8676 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8677 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8678 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8679 wont_dirty_i &= ~(1u << 31);
8680 if(wr) {
8681 #ifndef DESTRUCTIVE_WRITEBACK
8682 branch_regs[i].dirty&=wont_dirty_i;
8683 #endif
8684 branch_regs[i].dirty|=will_dirty_i;
8685 }
8686 }
8687 }
8688 }
8689 else if (dops[i].is_exception)
8690 {
8691 // SYSCALL instruction, etc
8692 will_dirty_i=0;
8693 wont_dirty_i=0;
8694 }
8695 will_dirty_next=will_dirty_i;
8696 wont_dirty_next=wont_dirty_i;
8697 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8698 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8699 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8700 will_dirty_i &= hr_candirty;
8701 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8702 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8703 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8704 wont_dirty_i &= ~(1u << 31);
8705 if (i > istart && !dops[i].is_jump) {
8706 // Don't store a register immediately after writing it,
8707 // may prevent dual-issue.
8708 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8709 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8710 }
8711 // Save it
8712 will_dirty[i]=will_dirty_i;
8713 wont_dirty[i]=wont_dirty_i;
8714 // Mark registers that won't be dirtied as not dirty
8715 if(wr) {
8716 regs[i].dirty|=will_dirty_i;
8717 #ifndef DESTRUCTIVE_WRITEBACK
8718 regs[i].dirty&=wont_dirty_i;
8719 if(dops[i].is_jump)
8720 {
8721 if (i < iend-1 && !dops[i].is_ujump) {
8722 for(r=0;r<HOST_REGS;r++) {
8723 if(r!=EXCLUDE_REG) {
8724 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8725 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8726 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8727 }
8728 }
8729 }
8730 }
8731 else
8732 {
8733 if(i<iend) {
8734 for(r=0;r<HOST_REGS;r++) {
8735 if(r!=EXCLUDE_REG) {
8736 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8737 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8738 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8739 }
8740 }
8741 }
8742 }
8743 #endif
8744 }
8745 // Deal with changed mappings
8746 temp_will_dirty=will_dirty_i;
8747 temp_wont_dirty=wont_dirty_i;
8748 for(r=0;r<HOST_REGS;r++) {
8749 if(r!=EXCLUDE_REG) {
8750 int nr;
8751 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8752 if(wr) {
8753 #ifndef DESTRUCTIVE_WRITEBACK
8754 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8755 #endif
8756 regs[i].wasdirty|=will_dirty_i&(1<<r);
8757 }
8758 }
8759 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8760 // Register moved to a different register
8761 will_dirty_i&=~(1<<r);
8762 wont_dirty_i&=~(1<<r);
8763 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8764 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8765 if(wr) {
8766 #ifndef DESTRUCTIVE_WRITEBACK
8767 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8768 #endif
8769 regs[i].wasdirty|=will_dirty_i&(1<<r);
8770 }
8771 }
8772 else {
8773 will_dirty_i&=~(1<<r);
8774 wont_dirty_i&=~(1<<r);
8775 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8776 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8777 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8778 } else {
8779 wont_dirty_i|=1<<r;
8780 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8781 }
8782 }
8783 }
8784 }
8785 }
8786}
8787
8788static noinline void pass10_expire_blocks(void)
8789{
8790 u_int step = MAX_OUTPUT_BLOCK_SIZE / PAGE_COUNT / 2;
8791 // not sizeof(ndrc->translation_cache) due to vita hack
8792 u_int step_mask = ((1u << TARGET_SIZE_2) - 1u) & ~(step - 1u);
8793 u_int end = (out - ndrc->translation_cache + EXPIRITY_OFFSET) & step_mask;
8794 u_int base_shift = __builtin_ctz(MAX_OUTPUT_BLOCK_SIZE);
8795 int hit;
8796
8797 for (; expirep != end; expirep = ((expirep + step) & step_mask))
8798 {
8799 u_int base_offs = expirep & ~(MAX_OUTPUT_BLOCK_SIZE - 1);
8800 u_int block_i = expirep / step & (PAGE_COUNT - 1);
8801 u_int phase = (expirep >> (base_shift - 1)) & 1u;
8802 if (!(expirep & (MAX_OUTPUT_BLOCK_SIZE / 2 - 1))) {
8803 inv_debug("EXP: base_offs %x/%lx phase %u\n", base_offs,
8804 (long)(out - ndrc->translation_cache), phase);
8805 }
8806
8807 if (!phase) {
8808 hit = blocks_remove_matching_addrs(&blocks[block_i], base_offs, base_shift);
8809 if (hit) {
8810 do_clear_cache();
8811 #ifdef USE_MINI_HT
8812 memset(mini_ht, -1, sizeof(mini_ht));
8813 #endif
8814 }
8815 }
8816 else
8817 unlink_jumps_tc_range(jumps[block_i], base_offs, base_shift);
8818 }
8819}
8820
8821static struct block_info *new_block_info(u_int start, u_int len,
8822 const void *source, const void *copy, u_char *beginning, u_short jump_in_count)
8823{
8824 struct block_info **b_pptr;
8825 struct block_info *block;
8826 u_int page = get_page(start);
8827
8828 block = malloc(sizeof(*block) + jump_in_count * sizeof(block->jump_in[0]));
8829 assert(block);
8830 assert(jump_in_count > 0);
8831 block->source = source;
8832 block->copy = copy;
8833 block->start = start;
8834 block->len = len;
8835 block->reg_sv_flags = 0;
8836 block->tc_offs = beginning - ndrc->translation_cache;
8837 //block->tc_len = out - beginning;
8838 block->is_dirty = 0;
8839 block->inv_near_misses = 0;
8840 block->jump_in_cnt = jump_in_count;
8841
8842 // insert sorted by start mirror-unmasked vaddr
8843 for (b_pptr = &blocks[page]; ; b_pptr = &((*b_pptr)->next)) {
8844 if (*b_pptr == NULL || (*b_pptr)->start >= start) {
8845 block->next = *b_pptr;
8846 *b_pptr = block;
8847 break;
8848 }
8849 }
8850 stat_inc(stat_blocks);
8851 return block;
8852}
8853
8854static int new_recompile_block(u_int addr)
8855{
8856 u_int pagelimit = 0;
8857 u_int state_rflags = 0;
8858 int i;
8859
8860 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
8861
8862 if (addr & 3) {
8863 if (addr != hack_addr) {
8864 SysPrintf("game crash @%08x, ra=%08x\n", addr, psxRegs.GPR.n.ra);
8865 hack_addr = addr;
8866 }
8867 return -1;
8868 }
8869
8870 // this is just for speculation
8871 for (i = 1; i < 32; i++) {
8872 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
8873 state_rflags |= 1 << i;
8874 }
8875
8876 start = addr;
8877 new_dynarec_did_compile=1;
8878 if (Config.HLE && start == 0x80001000) // hlecall
8879 {
8880 // XXX: is this enough? Maybe check hleSoftCall?
8881 void *beginning = start_block();
8882
8883 emit_movimm(start,0);
8884 emit_writeword(0,&pcaddr);
8885 emit_far_jump(new_dyna_leave);
8886 literal_pool(0);
8887 end_block(beginning);
8888 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8889 block->jump_in[0].vaddr = start;
8890 block->jump_in[0].addr = beginning;
8891 return 0;
8892 }
8893 else if (f1_hack && hack_addr == 0) {
8894 void *beginning = start_block();
8895 emit_movimm(start, 0);
8896 emit_writeword(0, &hack_addr);
8897 emit_readword(&psxRegs.GPR.n.sp, 0);
8898 emit_readptr(&mem_rtab, 1);
8899 emit_shrimm(0, 12, 2);
8900 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
8901 emit_addimm(0, 0x18, 0);
8902 emit_adds_ptr(1, 1, 1);
8903 emit_ldr_dualindexed(1, 0, 0);
8904 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
8905 emit_far_call(ndrc_get_addr_ht);
8906 emit_jmpreg(0); // jr k0
8907 literal_pool(0);
8908 end_block(beginning);
8909
8910 struct block_info *block = new_block_info(start, 4, NULL, NULL, beginning, 1);
8911 block->jump_in[0].vaddr = start;
8912 block->jump_in[0].addr = beginning;
8913 SysPrintf("F1 hack to %08x\n", start);
8914 return 0;
8915 }
8916
8917 cycle_multiplier_active = Config.cycle_multiplier_override && Config.cycle_multiplier == CYCLE_MULT_DEFAULT
8918 ? Config.cycle_multiplier_override : Config.cycle_multiplier;
8919
8920 source = get_source_start(start, &pagelimit);
8921 if (source == NULL) {
8922 if (addr != hack_addr) {
8923 SysPrintf("Compile at bogus memory address: %08x\n", addr);
8924 hack_addr = addr;
8925 }
8926 //abort();
8927 return -1;
8928 }
8929
8930 /* Pass 1: disassemble */
8931 /* Pass 2: register dependencies, branch targets */
8932 /* Pass 3: register allocation */
8933 /* Pass 4: branch dependencies */
8934 /* Pass 5: pre-alloc */
8935 /* Pass 6: optimize clean/dirty state */
8936 /* Pass 7: flag 32-bit registers */
8937 /* Pass 8: assembly */
8938 /* Pass 9: linker */
8939 /* Pass 10: garbage collection / free memory */
8940
8941 /* Pass 1 disassembly */
8942
8943 pass1_disassemble(pagelimit);
8944
8945 int clear_hack_addr = apply_hacks();
8946
8947 /* Pass 2 - Register dependencies and branch targets */
8948
8949 pass2_unneeded_regs(0,slen-1,0);
8950
8951 /* Pass 3 - Register allocation */
8952
8953 pass3_register_alloc(addr);
8954
8955 /* Pass 4 - Cull unused host registers */
8956
8957 pass4_cull_unused_regs();
8958
8959 /* Pass 5 - Pre-allocate registers */
8960
8961 pass5a_preallocate1();
8962 pass5b_preallocate2();
8963
8964 /* Pass 6 - Optimize clean/dirty state */
8965 pass6_clean_registers(0, slen-1, 1);
8966
8967 /* Pass 7 */
8968 for (i=slen-1;i>=0;i--)
8969 {
8970 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8971 {
8972 // Conditional branch
8973 if((source[i]>>16)!=0x1000&&i<slen-2) {
8974 // Mark this address as a branch target since it may be called
8975 // upon return from interrupt
8976 dops[i+2].bt=1;
8977 }
8978 }
8979 }
8980
8981 /* Pass 8 - Assembly */
8982 linkcount=0;stubcount=0;
8983 is_delayslot=0;
8984 u_int dirty_pre=0;
8985 void *beginning=start_block();
8986 void *instr_addr0_override = NULL;
8987 int ds = 0;
8988
8989 if (start == 0x80030000) {
8990 // nasty hack for the fastbios thing
8991 // override block entry to this code
8992 instr_addr0_override = out;
8993 emit_movimm(start,0);
8994 // abuse io address var as a flag that we
8995 // have already returned here once
8996 emit_readword(&address,1);
8997 emit_writeword(0,&pcaddr);
8998 emit_writeword(0,&address);
8999 emit_cmp(0,1);
9000 #ifdef __aarch64__
9001 emit_jeq(out + 4*2);
9002 emit_far_jump(new_dyna_leave);
9003 #else
9004 emit_jne(new_dyna_leave);
9005 #endif
9006 }
9007 for(i=0;i<slen;i++)
9008 {
9009 __builtin_prefetch(regs[i+1].regmap);
9010 check_regmap(regmap_pre[i]);
9011 check_regmap(regs[i].regmap_entry);
9012 check_regmap(regs[i].regmap);
9013 //if(ds) printf("ds: ");
9014 disassemble_inst(i);
9015 if(ds) {
9016 ds=0; // Skip delay slot
9017 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9018 instr_addr[i] = NULL;
9019 } else {
9020 speculate_register_values(i);
9021 #ifndef DESTRUCTIVE_WRITEBACK
9022 if (i < 2 || !dops[i-2].is_ujump)
9023 {
9024 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9025 }
9026 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9027 dirty_pre=branch_regs[i].dirty;
9028 }else{
9029 dirty_pre=regs[i].dirty;
9030 }
9031 #endif
9032 // write back
9033 if (i < 2 || !dops[i-2].is_ujump)
9034 {
9035 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9036 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9037 }
9038 // branch target entry point
9039 instr_addr[i] = out;
9040 assem_debug("<->\n");
9041 drc_dbg_emit_do_cmp(i, cinfo[i].ccadj);
9042 if (clear_hack_addr) {
9043 emit_movimm(0, 0);
9044 emit_writeword(0, &hack_addr);
9045 clear_hack_addr = 0;
9046 }
9047
9048 // load regs
9049 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9050 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9051 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9052 address_generation(i,&regs[i],regs[i].regmap_entry);
9053 load_consts(regmap_pre[i],regs[i].regmap,i);
9054 if(dops[i].is_jump)
9055 {
9056 // Load the delay slot registers if necessary
9057 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))
9058 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9059 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))
9060 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9061 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9062 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9063 if (dops[i+1].is_store)
9064 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9065 }
9066 else if(i+1<slen)
9067 {
9068 // Preload registers for following instruction
9069 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9070 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9071 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9072 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9073 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9074 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9075 }
9076 // TODO: if(is_ooo(i)) address_generation(i+1);
9077 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9078 load_reg(regs[i].regmap_entry,regs[i].regmap,CCREG);
9079 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9080 load_reg(regs[i].regmap_entry,regs[i].regmap,ROREG);
9081 if (dops[i].is_store)
9082 load_reg(regs[i].regmap_entry,regs[i].regmap,INVCP);
9083
9084 ds = assemble(i, &regs[i], cinfo[i].ccadj);
9085
9086 if (dops[i].is_ujump)
9087 literal_pool(1024);
9088 else
9089 literal_pool_jumpover(256);
9090 }
9091 }
9092
9093 assert(slen > 0);
9094 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9095 // no ending needed for this block since INTCALL never returns
9096 }
9097 // If the block did not end with an unconditional branch,
9098 // add a jump to the next instruction.
9099 else if (i > 1) {
9100 if (!dops[i-2].is_ujump) {
9101 assert(!dops[i-1].is_jump);
9102 assert(i==slen);
9103 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9104 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9105 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9106 emit_loadreg(CCREG,HOST_CCREG);
9107 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9108 }
9109 else
9110 {
9111 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9112 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9113 }
9114 add_to_linker(out,start+i*4,0);
9115 emit_jmp(0);
9116 }
9117 }
9118 else
9119 {
9120 assert(i>0);
9121 assert(!dops[i-1].is_jump);
9122 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9123 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9124 emit_loadreg(CCREG,HOST_CCREG);
9125 emit_addimm(HOST_CCREG, cinfo[i-1].ccadj + CLOCK_ADJUST(1), HOST_CCREG);
9126 add_to_linker(out,start+i*4,0);
9127 emit_jmp(0);
9128 }
9129
9130 // Stubs
9131 for(i = 0; i < stubcount; i++)
9132 {
9133 switch(stubs[i].type)
9134 {
9135 case LOADB_STUB:
9136 case LOADH_STUB:
9137 case LOADW_STUB:
9138 case LOADBU_STUB:
9139 case LOADHU_STUB:
9140 do_readstub(i);break;
9141 case STOREB_STUB:
9142 case STOREH_STUB:
9143 case STOREW_STUB:
9144 do_writestub(i);break;
9145 case CC_STUB:
9146 do_ccstub(i);break;
9147 case INVCODE_STUB:
9148 do_invstub(i);break;
9149 case STORELR_STUB:
9150 do_unalignedwritestub(i);break;
9151 case OVERFLOW_STUB:
9152 do_overflowstub(i); break;
9153 case ALIGNMENT_STUB:
9154 do_alignmentstub(i); break;
9155 default:
9156 assert(0);
9157 }
9158 }
9159
9160 if (instr_addr0_override)
9161 instr_addr[0] = instr_addr0_override;
9162
9163#if 0
9164 /* check for improper expiration */
9165 for (i = 0; i < ARRAY_SIZE(jumps); i++) {
9166 int j;
9167 if (!jumps[i])
9168 continue;
9169 for (j = 0; j < jumps[i]->count; j++)
9170 assert(jumps[i]->e[j].stub < beginning || (u_char *)jumps[i]->e[j].stub > out);
9171 }
9172#endif
9173
9174 /* Pass 9 - Linker */
9175 for(i=0;i<linkcount;i++)
9176 {
9177 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9178 literal_pool(64);
9179 if (!link_addr[i].internal)
9180 {
9181 void *stub = out;
9182 void *addr = check_addr(link_addr[i].target);
9183 emit_extjump(link_addr[i].addr, link_addr[i].target);
9184 if (addr) {
9185 set_jump_target(link_addr[i].addr, addr);
9186 ndrc_add_jump_out(link_addr[i].target,stub);
9187 }
9188 else
9189 set_jump_target(link_addr[i].addr, stub);
9190 }
9191 else
9192 {
9193 // Internal branch
9194 int target=(link_addr[i].target-start)>>2;
9195 assert(target>=0&&target<slen);
9196 assert(instr_addr[target]);
9197 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9198 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9199 //#else
9200 set_jump_target(link_addr[i].addr, instr_addr[target]);
9201 //#endif
9202 }
9203 }
9204
9205 u_int source_len = slen*4;
9206 if (dops[slen-1].itype == INTCALL && source_len > 4)
9207 // no need to treat the last instruction as compiled
9208 // as interpreter fully handles it
9209 source_len -= 4;
9210
9211 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9212 copy = shadow;
9213
9214 // External Branch Targets (jump_in)
9215 int jump_in_count = 1;
9216 assert(instr_addr[0]);
9217 for (i = 1; i < slen; i++)
9218 {
9219 if (dops[i].bt && instr_addr[i])
9220 jump_in_count++;
9221 }
9222
9223 struct block_info *block =
9224 new_block_info(start, slen * 4, source, copy, beginning, jump_in_count);
9225 block->reg_sv_flags = state_rflags;
9226
9227 int jump_in_i = 0;
9228 for (i = 0; i < slen; i++)
9229 {
9230 if ((i == 0 || dops[i].bt) && instr_addr[i])
9231 {
9232 assem_debug("%p (%d) <- %8x\n", instr_addr[i], i, start + i*4);
9233 u_int vaddr = start + i*4;
9234
9235 literal_pool(256);
9236 void *entry = out;
9237 load_regs_entry(i);
9238 if (entry == out)
9239 entry = instr_addr[i];
9240 else
9241 emit_jmp(instr_addr[i]);
9242
9243 block->jump_in[jump_in_i].vaddr = vaddr;
9244 block->jump_in[jump_in_i].addr = entry;
9245 jump_in_i++;
9246 }
9247 }
9248 assert(jump_in_i == jump_in_count);
9249 hash_table_add(block->jump_in[0].vaddr, block->jump_in[0].addr);
9250 // Write out the literal pool if necessary
9251 literal_pool(0);
9252 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9253 // Align code
9254 if(((u_int)out)&7) emit_addnop(13);
9255 #endif
9256 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9257 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9258 memcpy(copy, source, source_len);
9259 copy += source_len;
9260
9261 end_block(beginning);
9262
9263 // If we're within 256K of the end of the buffer,
9264 // start over from the beginning. (Is 256K enough?)
9265 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9266 out = ndrc->translation_cache;
9267
9268 // Trap writes to any of the pages we compiled
9269 mark_invalid_code(start, slen*4, 0);
9270
9271 /* Pass 10 - Free memory by expiring oldest blocks */
9272
9273 pass10_expire_blocks();
9274
9275#ifdef ASSEM_PRINT
9276 fflush(stdout);
9277#endif
9278 stat_inc(stat_bc_direct);
9279 return 0;
9280}
9281
9282// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork