notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
drc: only override default cycle_multiplier
[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 VITA
33#include <psp2/kernel/sysmem.h>
34static int sceBlock;
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
43#define noinline __attribute__((noinline,noclone))
44#ifndef ARRAY_SIZE
45#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
46#endif
47#ifndef min
48#define min(a, b) ((b) < (a) ? (b) : (a))
49#endif
50#ifndef max
51#define max(a, b) ((b) > (a) ? (b) : (a))
52#endif
53
54//#define DISASM
55//#define ASSEM_PRINT
56
57#ifdef ASSEM_PRINT
58#define assem_debug printf
59#else
60#define assem_debug(...)
61#endif
62//#define inv_debug printf
63#define inv_debug(...)
64
65#ifdef __i386__
66#include "assem_x86.h"
67#endif
68#ifdef __x86_64__
69#include "assem_x64.h"
70#endif
71#ifdef __arm__
72#include "assem_arm.h"
73#endif
74#ifdef __aarch64__
75#include "assem_arm64.h"
76#endif
77
78#define RAM_SIZE 0x200000
79#define MAXBLOCK 4096
80#define MAX_OUTPUT_BLOCK_SIZE 262144
81
82struct ndrc_mem
83{
84 u_char translation_cache[1 << TARGET_SIZE_2];
85 struct
86 {
87 struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
88 const void *f[2048 / sizeof(void *)];
89 } tramp;
90};
91
92#ifdef BASE_ADDR_DYNAMIC
93static struct ndrc_mem *ndrc;
94#else
95static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
96static struct ndrc_mem *ndrc = &ndrc_;
97#endif
98
99// stubs
100enum stub_type {
101 CC_STUB = 1,
102 FP_STUB = 2,
103 LOADB_STUB = 3,
104 LOADH_STUB = 4,
105 LOADW_STUB = 5,
106 LOADD_STUB = 6,
107 LOADBU_STUB = 7,
108 LOADHU_STUB = 8,
109 STOREB_STUB = 9,
110 STOREH_STUB = 10,
111 STOREW_STUB = 11,
112 STORED_STUB = 12,
113 STORELR_STUB = 13,
114 INVCODE_STUB = 14,
115};
116
117struct regstat
118{
119 signed char regmap_entry[HOST_REGS];
120 signed char regmap[HOST_REGS];
121 uint64_t wasdirty;
122 uint64_t dirty;
123 uint64_t u;
124 u_int wasconst;
125 u_int isconst;
126 u_int loadedconst; // host regs that have constants loaded
127 u_int waswritten; // MIPS regs that were used as store base before
128};
129
130// note: asm depends on this layout
131struct ll_entry
132{
133 u_int vaddr;
134 u_int reg_sv_flags;
135 void *addr;
136 struct ll_entry *next;
137};
138
139struct ht_entry
140{
141 u_int vaddr[2];
142 void *tcaddr[2];
143};
144
145struct code_stub
146{
147 enum stub_type type;
148 void *addr;
149 void *retaddr;
150 u_int a;
151 uintptr_t b;
152 uintptr_t c;
153 u_int d;
154 u_int e;
155};
156
157struct link_entry
158{
159 void *addr;
160 u_int target;
161 u_int ext;
162};
163
164static struct decoded_insn
165{
166 u_char itype;
167 u_char opcode;
168 u_char opcode2;
169 u_char rs1;
170 u_char rs2;
171 u_char rt1;
172 u_char rt2;
173 u_char lt1;
174 u_char bt:1;
175 u_char ooo:1;
176 u_char is_ds:1;
177 u_char is_jump:1;
178 u_char is_ujump:1;
179} dops[MAXBLOCK];
180
181 // used by asm:
182 u_char *out;
183 struct ht_entry hash_table[65536] __attribute__((aligned(16)));
184 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
185 struct ll_entry *jump_dirty[4096];
186
187 static struct ll_entry *jump_out[4096];
188 static u_int start;
189 static u_int *source;
190 static char insn[MAXBLOCK][10];
191 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
192 static uint64_t gte_rt[MAXBLOCK];
193 static uint64_t gte_unneeded[MAXBLOCK];
194 static u_int smrv[32]; // speculated MIPS register values
195 static u_int smrv_strong; // mask or regs that are likely to have correct values
196 static u_int smrv_weak; // same, but somewhat less likely
197 static u_int smrv_strong_next; // same, but after current insn executes
198 static u_int smrv_weak_next;
199 static int imm[MAXBLOCK];
200 static u_int ba[MAXBLOCK];
201 static uint64_t unneeded_reg[MAXBLOCK];
202 static uint64_t branch_unneeded_reg[MAXBLOCK];
203 static signed char regmap_pre[MAXBLOCK][HOST_REGS]; // pre-instruction i?
204 // contains 'real' consts at [i] insn, but may differ from what's actually
205 // loaded in host reg as 'final' value is always loaded, see get_final_value()
206 static uint32_t current_constmap[HOST_REGS];
207 static uint32_t constmap[MAXBLOCK][HOST_REGS];
208 static struct regstat regs[MAXBLOCK];
209 static struct regstat branch_regs[MAXBLOCK];
210 static signed char minimum_free_regs[MAXBLOCK];
211 static u_int needed_reg[MAXBLOCK];
212 static u_int wont_dirty[MAXBLOCK];
213 static u_int will_dirty[MAXBLOCK];
214 static int ccadj[MAXBLOCK];
215 static int slen;
216 static void *instr_addr[MAXBLOCK];
217 static struct link_entry link_addr[MAXBLOCK];
218 static int linkcount;
219 static struct code_stub stubs[MAXBLOCK*3];
220 static int stubcount;
221 static u_int literals[1024][2];
222 static int literalcount;
223 static int is_delayslot;
224 static char shadow[1048576] __attribute__((aligned(16)));
225 static void *copy;
226 static int expirep;
227 static u_int stop_after_jal;
228 static u_int f1_hack; // 0 - off, ~0 - capture address, else addr
229#ifndef RAM_FIXED
230 static uintptr_t ram_offset;
231#else
232 static const uintptr_t ram_offset=0;
233#endif
234
235 int new_dynarec_hacks;
236 int new_dynarec_hacks_pergame;
237 int new_dynarec_hacks_old;
238 int new_dynarec_did_compile;
239
240 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
241
242 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
243 extern int last_count; // last absolute target, often = next_interupt
244 extern int pcaddr;
245 extern int pending_exception;
246 extern int branch_target;
247 extern uintptr_t mini_ht[32][2];
248 extern u_char restore_candidate[512];
249
250 /* registers that may be allocated */
251 /* 1-31 gpr */
252#define LOREG 32 // lo
253#define HIREG 33 // hi
254//#define FSREG 34 // FPU status (FCSR)
255#define CSREG 35 // Coprocessor status
256#define CCREG 36 // Cycle count
257#define INVCP 37 // Pointer to invalid_code
258//#define MMREG 38 // Pointer to memory_map
259//#define ROREG 39 // ram offset (if rdram!=0x80000000)
260#define TEMPREG 40
261#define FTEMP 40 // FPU temporary register
262#define PTEMP 41 // Prefetch temporary register
263//#define TLREG 42 // TLB mapping offset
264#define RHASH 43 // Return address hash
265#define RHTBL 44 // Return address hash table address
266#define RTEMP 45 // JR/JALR address register
267#define MAXREG 45
268#define AGEN1 46 // Address generation temporary register
269//#define AGEN2 47 // Address generation temporary register
270//#define MGEN1 48 // Maptable address generation temporary register
271//#define MGEN2 49 // Maptable address generation temporary register
272#define BTREG 50 // Branch target temporary register
273
274 /* instruction types */
275#define NOP 0 // No operation
276#define LOAD 1 // Load
277#define STORE 2 // Store
278#define LOADLR 3 // Unaligned load
279#define STORELR 4 // Unaligned store
280#define MOV 5 // Move
281#define ALU 6 // Arithmetic/logic
282#define MULTDIV 7 // Multiply/divide
283#define SHIFT 8 // Shift by register
284#define SHIFTIMM 9// Shift by immediate
285#define IMM16 10 // 16-bit immediate
286#define RJUMP 11 // Unconditional jump to register
287#define UJUMP 12 // Unconditional jump
288#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
289#define SJUMP 14 // Conditional branch (regimm format)
290#define COP0 15 // Coprocessor 0
291#define COP1 16 // Coprocessor 1
292#define C1LS 17 // Coprocessor 1 load/store
293//#define FJUMP 18 // Conditional branch (floating point)
294//#define FLOAT 19 // Floating point unit
295//#define FCONV 20 // Convert integer to float
296//#define FCOMP 21 // Floating point compare (sets FSREG)
297#define SYSCALL 22// SYSCALL
298#define OTHER 23 // Other
299#define SPAN 24 // Branch/delay slot spans 2 pages
300#define NI 25 // Not implemented
301#define HLECALL 26// PCSX fake opcodes for HLE
302#define COP2 27 // Coprocessor 2 move
303#define C2LS 28 // Coprocessor 2 load/store
304#define C2OP 29 // Coprocessor 2 operation
305#define INTCALL 30// Call interpreter to handle rare corner cases
306
307 /* branch codes */
308#define TAKEN 1
309#define NOTTAKEN 2
310#define NULLDS 3
311
312#define DJT_1 (void *)1l // no function, just a label in assem_debug log
313#define DJT_2 (void *)2l
314
315// asm linkage
316int new_recompile_block(u_int addr);
317void *get_addr_ht(u_int vaddr);
318void invalidate_block(u_int block);
319void invalidate_addr(u_int addr);
320void remove_hash(int vaddr);
321void dyna_linker();
322void dyna_linker_ds();
323void verify_code();
324void verify_code_ds();
325void cc_interrupt();
326void fp_exception();
327void fp_exception_ds();
328void jump_to_new_pc();
329void call_gteStall();
330void new_dyna_leave();
331
332// Needed by assembler
333static void wb_register(signed char r,signed char regmap[],uint64_t dirty);
334static void wb_dirtys(signed char i_regmap[],uint64_t i_dirty);
335static void wb_needed_dirtys(signed char i_regmap[],uint64_t i_dirty,int addr);
336static void load_all_regs(signed char i_regmap[]);
337static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
338static void load_regs_entry(int t);
339static void load_all_consts(signed char regmap[],u_int dirty,int i);
340static u_int get_host_reglist(const signed char *regmap);
341
342static int verify_dirty(const u_int *ptr);
343static int get_final_value(int hr, int i, int *value);
344static void add_stub(enum stub_type type, void *addr, void *retaddr,
345 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
346static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
347 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
348static void add_to_linker(void *addr, u_int target, int ext);
349static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override);
350static void *get_direct_memhandler(void *table, u_int addr,
351 enum stub_type type, uintptr_t *addr_host);
352static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
353static void pass_args(int a0, int a1);
354static void emit_far_jump(const void *f);
355static void emit_far_call(const void *f);
356
357static void mprotect_w_x(void *start, void *end, int is_x)
358{
359#ifdef NO_WRITE_EXEC
360 #if defined(VITA)
361 // *Open* enables write on all memory that was
362 // allocated by sceKernelAllocMemBlockForVM()?
363 if (is_x)
364 sceKernelCloseVMDomain();
365 else
366 sceKernelOpenVMDomain();
367 #else
368 u_long mstart = (u_long)start & ~4095ul;
369 u_long mend = (u_long)end;
370 if (mprotect((void *)mstart, mend - mstart,
371 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
372 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
373 #endif
374#endif
375}
376
377static void start_tcache_write(void *start, void *end)
378{
379 mprotect_w_x(start, end, 0);
380}
381
382static void end_tcache_write(void *start, void *end)
383{
384#if defined(__arm__) || defined(__aarch64__)
385 size_t len = (char *)end - (char *)start;
386 #if defined(__BLACKBERRY_QNX__)
387 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
388 #elif defined(__MACH__)
389 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
390 #elif defined(VITA)
391 sceKernelSyncVMDomain(sceBlock, start, len);
392 #elif defined(_3DS)
393 ctr_flush_invalidate_cache();
394 #elif defined(__aarch64__)
395 // as of 2021, __clear_cache() is still broken on arm64
396 // so here is a custom one :(
397 clear_cache_arm64(start, end);
398 #else
399 __clear_cache(start, end);
400 #endif
401 (void)len;
402#endif
403
404 mprotect_w_x(start, end, 1);
405}
406
407static void *start_block(void)
408{
409 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
410 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
411 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
412 start_tcache_write(out, end);
413 return out;
414}
415
416static void end_block(void *start)
417{
418 end_tcache_write(start, out);
419}
420
421// also takes care of w^x mappings when patching code
422static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
423
424static void mark_clear_cache(void *target)
425{
426 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
427 u_int mask = 1u << ((offset >> 12) & 31);
428 if (!(needs_clear_cache[offset >> 17] & mask)) {
429 char *start = (char *)((uintptr_t)target & ~4095l);
430 start_tcache_write(start, start + 4095);
431 needs_clear_cache[offset >> 17] |= mask;
432 }
433}
434
435// Clearing the cache is rather slow on ARM Linux, so mark the areas
436// that need to be cleared, and then only clear these areas once.
437static void do_clear_cache(void)
438{
439 int i, j;
440 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
441 {
442 u_int bitmap = needs_clear_cache[i];
443 if (!bitmap)
444 continue;
445 for (j = 0; j < 32; j++)
446 {
447 u_char *start, *end;
448 if (!(bitmap & (1<<j)))
449 continue;
450
451 start = ndrc->translation_cache + i*131072 + j*4096;
452 end = start + 4095;
453 for (j++; j < 32; j++) {
454 if (!(bitmap & (1<<j)))
455 break;
456 end += 4096;
457 }
458 end_tcache_write(start, end);
459 }
460 needs_clear_cache[i] = 0;
461 }
462}
463
464//#define DEBUG_CYCLE_COUNT 1
465
466#define NO_CYCLE_PENALTY_THR 12
467
468int cycle_multiplier = CYCLE_MULT_DEFAULT; // 100 for 1.0
469int cycle_multiplier_override;
470int cycle_multiplier_old;
471
472static int CLOCK_ADJUST(int x)
473{
474 int m = cycle_multiplier_override && cycle_multiplier == CYCLE_MULT_DEFAULT
475 ? cycle_multiplier_override : cycle_multiplier;
476 int s=(x>>31)|1;
477 return (x * m + s * 50) / 100;
478}
479
480static int ds_writes_rjump_rs(int i)
481{
482 return dops[i].rs1 != 0 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2);
483}
484
485static u_int get_page(u_int vaddr)
486{
487 u_int page=vaddr&~0xe0000000;
488 if (page < 0x1000000)
489 page &= ~0x0e00000; // RAM mirrors
490 page>>=12;
491 if(page>2048) page=2048+(page&2047);
492 return page;
493}
494
495// no virtual mem in PCSX
496static u_int get_vpage(u_int vaddr)
497{
498 return get_page(vaddr);
499}
500
501static struct ht_entry *hash_table_get(u_int vaddr)
502{
503 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
504}
505
506static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
507{
508 ht_bin->vaddr[1] = ht_bin->vaddr[0];
509 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
510 ht_bin->vaddr[0] = vaddr;
511 ht_bin->tcaddr[0] = tcaddr;
512}
513
514// some messy ari64's code, seems to rely on unsigned 32bit overflow
515static int doesnt_expire_soon(void *tcaddr)
516{
517 u_int diff = (u_int)((u_char *)tcaddr - out) << (32-TARGET_SIZE_2);
518 return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
519}
520
521// Get address from virtual address
522// This is called from the recompiled JR/JALR instructions
523void noinline *get_addr(u_int vaddr)
524{
525 u_int page=get_page(vaddr);
526 u_int vpage=get_vpage(vaddr);
527 struct ll_entry *head;
528 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
529 head=jump_in[page];
530 while(head!=NULL) {
531 if(head->vaddr==vaddr) {
532 //printf("TRACE: count=%d next=%d (get_addr match %x: %p)\n",Count,next_interupt,vaddr,head->addr);
533 hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
534 return head->addr;
535 }
536 head=head->next;
537 }
538 head=jump_dirty[vpage];
539 while(head!=NULL) {
540 if(head->vaddr==vaddr) {
541 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %p)\n",Count,next_interupt,vaddr,head->addr);
542 // Don't restore blocks which are about to expire from the cache
543 if (doesnt_expire_soon(head->addr))
544 if (verify_dirty(head->addr)) {
545 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
546 invalid_code[vaddr>>12]=0;
547 inv_code_start=inv_code_end=~0;
548 if(vpage<2048) {
549 restore_candidate[vpage>>3]|=1<<(vpage&7);
550 }
551 else restore_candidate[page>>3]|=1<<(page&7);
552 struct ht_entry *ht_bin = hash_table_get(vaddr);
553 if (ht_bin->vaddr[0] == vaddr)
554 ht_bin->tcaddr[0] = head->addr; // Replace existing entry
555 else
556 hash_table_add(ht_bin, vaddr, head->addr);
557
558 return head->addr;
559 }
560 }
561 head=head->next;
562 }
563 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
564 int r=new_recompile_block(vaddr);
565 if(r==0) return get_addr(vaddr);
566 // Execute in unmapped page, generate pagefault execption
567 Status|=2;
568 Cause=(vaddr<<31)|0x8;
569 EPC=(vaddr&1)?vaddr-5:vaddr;
570 BadVAddr=(vaddr&~1);
571 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
572 EntryHi=BadVAddr&0xFFFFE000;
573 return get_addr_ht(0x80000000);
574}
575// Look up address in hash table first
576void *get_addr_ht(u_int vaddr)
577{
578 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
579 const struct ht_entry *ht_bin = hash_table_get(vaddr);
580 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
581 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
582 return get_addr(vaddr);
583}
584
585void clear_all_regs(signed char regmap[])
586{
587 int hr;
588 for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
589}
590
591static signed char get_reg(const signed char regmap[],int r)
592{
593 int hr;
594 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap[hr]==r) return hr;
595 return -1;
596}
597
598// Find a register that is available for two consecutive cycles
599static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
600{
601 int hr;
602 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
603 return -1;
604}
605
606int count_free_regs(signed char regmap[])
607{
608 int count=0;
609 int hr;
610 for(hr=0;hr<HOST_REGS;hr++)
611 {
612 if(hr!=EXCLUDE_REG) {
613 if(regmap[hr]<0) count++;
614 }
615 }
616 return count;
617}
618
619void dirty_reg(struct regstat *cur,signed char reg)
620{
621 int hr;
622 if(!reg) return;
623 for (hr=0;hr<HOST_REGS;hr++) {
624 if((cur->regmap[hr]&63)==reg) {
625 cur->dirty|=1<<hr;
626 }
627 }
628}
629
630static void set_const(struct regstat *cur, signed char reg, uint32_t value)
631{
632 int hr;
633 if(!reg) return;
634 for (hr=0;hr<HOST_REGS;hr++) {
635 if(cur->regmap[hr]==reg) {
636 cur->isconst|=1<<hr;
637 current_constmap[hr]=value;
638 }
639 }
640}
641
642static void clear_const(struct regstat *cur, signed char reg)
643{
644 int hr;
645 if(!reg) return;
646 for (hr=0;hr<HOST_REGS;hr++) {
647 if((cur->regmap[hr]&63)==reg) {
648 cur->isconst&=~(1<<hr);
649 }
650 }
651}
652
653static int is_const(struct regstat *cur, signed char reg)
654{
655 int hr;
656 if(reg<0) return 0;
657 if(!reg) return 1;
658 for (hr=0;hr<HOST_REGS;hr++) {
659 if((cur->regmap[hr]&63)==reg) {
660 return (cur->isconst>>hr)&1;
661 }
662 }
663 return 0;
664}
665
666static uint32_t get_const(struct regstat *cur, signed char reg)
667{
668 int hr;
669 if(!reg) return 0;
670 for (hr=0;hr<HOST_REGS;hr++) {
671 if(cur->regmap[hr]==reg) {
672 return current_constmap[hr];
673 }
674 }
675 SysPrintf("Unknown constant in r%d\n",reg);
676 abort();
677}
678
679// Least soon needed registers
680// Look at the next ten instructions and see which registers
681// will be used. Try not to reallocate these.
682void lsn(u_char hsn[], int i, int *preferred_reg)
683{
684 int j;
685 int b=-1;
686 for(j=0;j<9;j++)
687 {
688 if(i+j>=slen) {
689 j=slen-i-1;
690 break;
691 }
692 if (dops[i+j].is_ujump)
693 {
694 // Don't go past an unconditonal jump
695 j++;
696 break;
697 }
698 }
699 for(;j>=0;j--)
700 {
701 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
702 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
703 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
704 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
705 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
706 // Stores can allocate zero
707 hsn[dops[i+j].rs1]=j;
708 hsn[dops[i+j].rs2]=j;
709 }
710 // On some architectures stores need invc_ptr
711 #if defined(HOST_IMM8)
712 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR || (dops[i+j].opcode&0x3b)==0x39 || (dops[i+j].opcode&0x3b)==0x3a) {
713 hsn[INVCP]=j;
714 }
715 #endif
716 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
717 {
718 hsn[CCREG]=j;
719 b=j;
720 }
721 }
722 if(b>=0)
723 {
724 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
725 {
726 // Follow first branch
727 int t=(ba[i+b]-start)>>2;
728 j=7-b;if(t+j>=slen) j=slen-t-1;
729 for(;j>=0;j--)
730 {
731 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
732 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
733 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
734 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
735 }
736 }
737 // TODO: preferred register based on backward branch
738 }
739 // Delay slot should preferably not overwrite branch conditions or cycle count
740 if (i > 0 && dops[i-1].is_jump) {
741 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
742 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
743 hsn[CCREG]=1;
744 // ...or hash tables
745 hsn[RHASH]=1;
746 hsn[RHTBL]=1;
747 }
748 // Coprocessor load/store needs FTEMP, even if not declared
749 if(dops[i].itype==C1LS||dops[i].itype==C2LS) {
750 hsn[FTEMP]=0;
751 }
752 // Load L/R also uses FTEMP as a temporary register
753 if(dops[i].itype==LOADLR) {
754 hsn[FTEMP]=0;
755 }
756 // Also SWL/SWR/SDL/SDR
757 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
758 hsn[FTEMP]=0;
759 }
760 // Don't remove the miniht registers
761 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
762 {
763 hsn[RHASH]=0;
764 hsn[RHTBL]=0;
765 }
766}
767
768// We only want to allocate registers if we're going to use them again soon
769int needed_again(int r, int i)
770{
771 int j;
772 int b=-1;
773 int rn=10;
774
775 if (i > 0 && dops[i-1].is_ujump)
776 {
777 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
778 return 0; // Don't need any registers if exiting the block
779 }
780 for(j=0;j<9;j++)
781 {
782 if(i+j>=slen) {
783 j=slen-i-1;
784 break;
785 }
786 if (dops[i+j].is_ujump)
787 {
788 // Don't go past an unconditonal jump
789 j++;
790 break;
791 }
792 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
793 {
794 break;
795 }
796 }
797 for(;j>=1;j--)
798 {
799 if(dops[i+j].rs1==r) rn=j;
800 if(dops[i+j].rs2==r) rn=j;
801 if((unneeded_reg[i+j]>>r)&1) rn=10;
802 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
803 {
804 b=j;
805 }
806 }
807 if(rn<10) return 1;
808 (void)b;
809 return 0;
810}
811
812// Try to match register allocations at the end of a loop with those
813// at the beginning
814int loop_reg(int i, int r, int hr)
815{
816 int j,k;
817 for(j=0;j<9;j++)
818 {
819 if(i+j>=slen) {
820 j=slen-i-1;
821 break;
822 }
823 if (dops[i+j].is_ujump)
824 {
825 // Don't go past an unconditonal jump
826 j++;
827 break;
828 }
829 }
830 k=0;
831 if(i>0){
832 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
833 k--;
834 }
835 for(;k<j;k++)
836 {
837 assert(r < 64);
838 if((unneeded_reg[i+k]>>r)&1) return hr;
839 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
840 {
841 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
842 {
843 int t=(ba[i+k]-start)>>2;
844 int reg=get_reg(regs[t].regmap_entry,r);
845 if(reg>=0) return reg;
846 //reg=get_reg(regs[t+1].regmap_entry,r);
847 //if(reg>=0) return reg;
848 }
849 }
850 }
851 return hr;
852}
853
854
855// Allocate every register, preserving source/target regs
856void alloc_all(struct regstat *cur,int i)
857{
858 int hr;
859
860 for(hr=0;hr<HOST_REGS;hr++) {
861 if(hr!=EXCLUDE_REG) {
862 if(((cur->regmap[hr]&63)!=dops[i].rs1)&&((cur->regmap[hr]&63)!=dops[i].rs2)&&
863 ((cur->regmap[hr]&63)!=dops[i].rt1)&&((cur->regmap[hr]&63)!=dops[i].rt2))
864 {
865 cur->regmap[hr]=-1;
866 cur->dirty&=~(1<<hr);
867 }
868 // Don't need zeros
869 if((cur->regmap[hr]&63)==0)
870 {
871 cur->regmap[hr]=-1;
872 cur->dirty&=~(1<<hr);
873 }
874 }
875 }
876}
877
878#ifndef NDEBUG
879static int host_tempreg_in_use;
880
881static void host_tempreg_acquire(void)
882{
883 assert(!host_tempreg_in_use);
884 host_tempreg_in_use = 1;
885}
886
887static void host_tempreg_release(void)
888{
889 host_tempreg_in_use = 0;
890}
891#else
892static void host_tempreg_acquire(void) {}
893static void host_tempreg_release(void) {}
894#endif
895
896#ifdef ASSEM_PRINT
897extern void gen_interupt();
898extern void do_insn_cmp();
899#define FUNCNAME(f) { f, " " #f }
900static const struct {
901 void *addr;
902 const char *name;
903} function_names[] = {
904 FUNCNAME(cc_interrupt),
905 FUNCNAME(gen_interupt),
906 FUNCNAME(get_addr_ht),
907 FUNCNAME(get_addr),
908 FUNCNAME(jump_handler_read8),
909 FUNCNAME(jump_handler_read16),
910 FUNCNAME(jump_handler_read32),
911 FUNCNAME(jump_handler_write8),
912 FUNCNAME(jump_handler_write16),
913 FUNCNAME(jump_handler_write32),
914 FUNCNAME(invalidate_addr),
915 FUNCNAME(jump_to_new_pc),
916 FUNCNAME(call_gteStall),
917 FUNCNAME(new_dyna_leave),
918 FUNCNAME(pcsx_mtc0),
919 FUNCNAME(pcsx_mtc0_ds),
920#ifdef DRC_DBG
921 FUNCNAME(do_insn_cmp),
922#endif
923#ifdef __arm__
924 FUNCNAME(verify_code),
925#endif
926};
927
928static const char *func_name(const void *a)
929{
930 int i;
931 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
932 if (function_names[i].addr == a)
933 return function_names[i].name;
934 return "";
935}
936#else
937#define func_name(x) ""
938#endif
939
940#ifdef __i386__
941#include "assem_x86.c"
942#endif
943#ifdef __x86_64__
944#include "assem_x64.c"
945#endif
946#ifdef __arm__
947#include "assem_arm.c"
948#endif
949#ifdef __aarch64__
950#include "assem_arm64.c"
951#endif
952
953static void *get_trampoline(const void *f)
954{
955 size_t i;
956
957 for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
958 if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
959 break;
960 }
961 if (i == ARRAY_SIZE(ndrc->tramp.f)) {
962 SysPrintf("trampoline table is full, last func %p\n", f);
963 abort();
964 }
965 if (ndrc->tramp.f[i] == NULL) {
966 start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
967 ndrc->tramp.f[i] = f;
968 end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
969 }
970 return &ndrc->tramp.ops[i];
971}
972
973static void emit_far_jump(const void *f)
974{
975 if (can_jump_or_call(f)) {
976 emit_jmp(f);
977 return;
978 }
979
980 f = get_trampoline(f);
981 emit_jmp(f);
982}
983
984static void emit_far_call(const void *f)
985{
986 if (can_jump_or_call(f)) {
987 emit_call(f);
988 return;
989 }
990
991 f = get_trampoline(f);
992 emit_call(f);
993}
994
995// Add virtual address mapping to linked list
996void ll_add(struct ll_entry **head,int vaddr,void *addr)
997{
998 struct ll_entry *new_entry;
999 new_entry=malloc(sizeof(struct ll_entry));
1000 assert(new_entry!=NULL);
1001 new_entry->vaddr=vaddr;
1002 new_entry->reg_sv_flags=0;
1003 new_entry->addr=addr;
1004 new_entry->next=*head;
1005 *head=new_entry;
1006}
1007
1008void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
1009{
1010 ll_add(head,vaddr,addr);
1011 (*head)->reg_sv_flags=reg_sv_flags;
1012}
1013
1014// Check if an address is already compiled
1015// but don't return addresses which are about to expire from the cache
1016void *check_addr(u_int vaddr)
1017{
1018 struct ht_entry *ht_bin = hash_table_get(vaddr);
1019 size_t i;
1020 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1021 if (ht_bin->vaddr[i] == vaddr)
1022 if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
1023 if (isclean(ht_bin->tcaddr[i]))
1024 return ht_bin->tcaddr[i];
1025 }
1026 u_int page=get_page(vaddr);
1027 struct ll_entry *head;
1028 head=jump_in[page];
1029 while (head != NULL) {
1030 if (head->vaddr == vaddr) {
1031 if (doesnt_expire_soon(head->addr)) {
1032 // Update existing entry with current address
1033 if (ht_bin->vaddr[0] == vaddr) {
1034 ht_bin->tcaddr[0] = head->addr;
1035 return head->addr;
1036 }
1037 if (ht_bin->vaddr[1] == vaddr) {
1038 ht_bin->tcaddr[1] = head->addr;
1039 return head->addr;
1040 }
1041 // Insert into hash table with low priority.
1042 // Don't evict existing entries, as they are probably
1043 // addresses that are being accessed frequently.
1044 if (ht_bin->vaddr[0] == -1) {
1045 ht_bin->vaddr[0] = vaddr;
1046 ht_bin->tcaddr[0] = head->addr;
1047 }
1048 else if (ht_bin->vaddr[1] == -1) {
1049 ht_bin->vaddr[1] = vaddr;
1050 ht_bin->tcaddr[1] = head->addr;
1051 }
1052 return head->addr;
1053 }
1054 }
1055 head=head->next;
1056 }
1057 return 0;
1058}
1059
1060void remove_hash(int vaddr)
1061{
1062 //printf("remove hash: %x\n",vaddr);
1063 struct ht_entry *ht_bin = hash_table_get(vaddr);
1064 if (ht_bin->vaddr[1] == vaddr) {
1065 ht_bin->vaddr[1] = -1;
1066 ht_bin->tcaddr[1] = NULL;
1067 }
1068 if (ht_bin->vaddr[0] == vaddr) {
1069 ht_bin->vaddr[0] = ht_bin->vaddr[1];
1070 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
1071 ht_bin->vaddr[1] = -1;
1072 ht_bin->tcaddr[1] = NULL;
1073 }
1074}
1075
1076static void ll_remove_matching_addrs(struct ll_entry **head,
1077 uintptr_t base_offs_s, int shift)
1078{
1079 struct ll_entry *next;
1080 while(*head) {
1081 uintptr_t o1 = (u_char *)(*head)->addr - ndrc->translation_cache;
1082 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1083 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1084 {
1085 inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
1086 remove_hash((*head)->vaddr);
1087 next=(*head)->next;
1088 free(*head);
1089 *head=next;
1090 }
1091 else
1092 {
1093 head=&((*head)->next);
1094 }
1095 }
1096}
1097
1098// Remove all entries from linked list
1099void ll_clear(struct ll_entry **head)
1100{
1101 struct ll_entry *cur;
1102 struct ll_entry *next;
1103 if((cur=*head)) {
1104 *head=0;
1105 while(cur) {
1106 next=cur->next;
1107 free(cur);
1108 cur=next;
1109 }
1110 }
1111}
1112
1113// Dereference the pointers and remove if it matches
1114static void ll_kill_pointers(struct ll_entry *head,
1115 uintptr_t base_offs_s, int shift)
1116{
1117 while(head) {
1118 u_char *ptr = get_pointer(head->addr);
1119 uintptr_t o1 = ptr - ndrc->translation_cache;
1120 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1121 inv_debug("EXP: Lookup pointer to %p at %p (%x)\n",ptr,head->addr,head->vaddr);
1122 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1123 {
1124 inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
1125 void *host_addr=find_extjump_insn(head->addr);
1126 mark_clear_cache(host_addr);
1127 set_jump_target(host_addr, head->addr);
1128 }
1129 head=head->next;
1130 }
1131}
1132
1133// This is called when we write to a compiled block (see do_invstub)
1134static void invalidate_page(u_int page)
1135{
1136 struct ll_entry *head;
1137 struct ll_entry *next;
1138 head=jump_in[page];
1139 jump_in[page]=0;
1140 while(head!=NULL) {
1141 inv_debug("INVALIDATE: %x\n",head->vaddr);
1142 remove_hash(head->vaddr);
1143 next=head->next;
1144 free(head);
1145 head=next;
1146 }
1147 head=jump_out[page];
1148 jump_out[page]=0;
1149 while(head!=NULL) {
1150 inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
1151 void *host_addr=find_extjump_insn(head->addr);
1152 mark_clear_cache(host_addr);
1153 set_jump_target(host_addr, head->addr); // point back to dyna_linker
1154 next=head->next;
1155 free(head);
1156 head=next;
1157 }
1158}
1159
1160static void invalidate_block_range(u_int block, u_int first, u_int last)
1161{
1162 u_int page=get_page(block<<12);
1163 //printf("first=%d last=%d\n",first,last);
1164 invalidate_page(page);
1165 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1166 assert(last<page+5);
1167 // Invalidate the adjacent pages if a block crosses a 4K boundary
1168 while(first<page) {
1169 invalidate_page(first);
1170 first++;
1171 }
1172 for(first=page+1;first<last;first++) {
1173 invalidate_page(first);
1174 }
1175 do_clear_cache();
1176
1177 // Don't trap writes
1178 invalid_code[block]=1;
1179
1180 #ifdef USE_MINI_HT
1181 memset(mini_ht,-1,sizeof(mini_ht));
1182 #endif
1183}
1184
1185void invalidate_block(u_int block)
1186{
1187 u_int page=get_page(block<<12);
1188 u_int vpage=get_vpage(block<<12);
1189 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1190 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1191 u_int first,last;
1192 first=last=page;
1193 struct ll_entry *head;
1194 head=jump_dirty[vpage];
1195 //printf("page=%d vpage=%d\n",page,vpage);
1196 while(head!=NULL) {
1197 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1198 u_char *start, *end;
1199 get_bounds(head->addr, &start, &end);
1200 //printf("start: %p end: %p\n", start, end);
1201 if (page < 2048 && start >= rdram && end < rdram+RAM_SIZE) {
1202 if (((start-rdram)>>12) <= page && ((end-1-rdram)>>12) >= page) {
1203 if ((((start-rdram)>>12)&2047) < first) first = ((start-rdram)>>12)&2047;
1204 if ((((end-1-rdram)>>12)&2047) > last) last = ((end-1-rdram)>>12)&2047;
1205 }
1206 }
1207 }
1208 head=head->next;
1209 }
1210 invalidate_block_range(block,first,last);
1211}
1212
1213void invalidate_addr(u_int addr)
1214{
1215 //static int rhits;
1216 // this check is done by the caller
1217 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1218 u_int page=get_vpage(addr);
1219 if(page<2048) { // RAM
1220 struct ll_entry *head;
1221 u_int addr_min=~0, addr_max=0;
1222 u_int mask=RAM_SIZE-1;
1223 u_int addr_main=0x80000000|(addr&mask);
1224 int pg1;
1225 inv_code_start=addr_main&~0xfff;
1226 inv_code_end=addr_main|0xfff;
1227 pg1=page;
1228 if (pg1>0) {
1229 // must check previous page too because of spans..
1230 pg1--;
1231 inv_code_start-=0x1000;
1232 }
1233 for(;pg1<=page;pg1++) {
1234 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1235 u_char *start_h, *end_h;
1236 u_int start, end;
1237 get_bounds(head->addr, &start_h, &end_h);
1238 start = (uintptr_t)start_h - ram_offset;
1239 end = (uintptr_t)end_h - ram_offset;
1240 if(start<=addr_main&&addr_main<end) {
1241 if(start<addr_min) addr_min=start;
1242 if(end>addr_max) addr_max=end;
1243 }
1244 else if(addr_main<start) {
1245 if(start<inv_code_end)
1246 inv_code_end=start-1;
1247 }
1248 else {
1249 if(end>inv_code_start)
1250 inv_code_start=end;
1251 }
1252 }
1253 }
1254 if (addr_min!=~0) {
1255 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1256 inv_code_start=inv_code_end=~0;
1257 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1258 return;
1259 }
1260 else {
1261 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1262 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1263 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1264 return;
1265 }
1266 }
1267 invalidate_block(addr>>12);
1268}
1269
1270// This is called when loading a save state.
1271// Anything could have changed, so invalidate everything.
1272void invalidate_all_pages(void)
1273{
1274 u_int page;
1275 for(page=0;page<4096;page++)
1276 invalidate_page(page);
1277 for(page=0;page<1048576;page++)
1278 if(!invalid_code[page]) {
1279 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1280 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1281 }
1282 #ifdef USE_MINI_HT
1283 memset(mini_ht,-1,sizeof(mini_ht));
1284 #endif
1285 do_clear_cache();
1286}
1287
1288static void do_invstub(int n)
1289{
1290 literal_pool(20);
1291 u_int reglist=stubs[n].a;
1292 set_jump_target(stubs[n].addr, out);
1293 save_regs(reglist);
1294 if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
1295 emit_far_call(invalidate_addr);
1296 restore_regs(reglist);
1297 emit_jmp(stubs[n].retaddr); // return address
1298}
1299
1300// Add an entry to jump_out after making a link
1301// src should point to code by emit_extjump2()
1302void add_jump_out(u_int vaddr,void *src)
1303{
1304 u_int page=get_page(vaddr);
1305 inv_debug("add_jump_out: %p -> %x (%d)\n",src,vaddr,page);
1306 check_extjump2(src);
1307 ll_add(jump_out+page,vaddr,src);
1308 //inv_debug("add_jump_out: to %p\n",get_pointer(src));
1309}
1310
1311// If a code block was found to be unmodified (bit was set in
1312// restore_candidate) and it remains unmodified (bit is clear
1313// in invalid_code) then move the entries for that 4K page from
1314// the dirty list to the clean list.
1315void clean_blocks(u_int page)
1316{
1317 struct ll_entry *head;
1318 inv_debug("INV: clean_blocks page=%d\n",page);
1319 head=jump_dirty[page];
1320 while(head!=NULL) {
1321 if(!invalid_code[head->vaddr>>12]) {
1322 // Don't restore blocks which are about to expire from the cache
1323 if (doesnt_expire_soon(head->addr)) {
1324 if(verify_dirty(head->addr)) {
1325 u_char *start, *end;
1326 //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
1327 u_int i;
1328 u_int inv=0;
1329 get_bounds(head->addr, &start, &end);
1330 if (start - rdram < RAM_SIZE) {
1331 for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
1332 inv|=invalid_code[i];
1333 }
1334 }
1335 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1336 inv=1;
1337 }
1338 if(!inv) {
1339 void *clean_addr = get_clean_addr(head->addr);
1340 if (doesnt_expire_soon(clean_addr)) {
1341 u_int ppage=page;
1342 inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, clean_addr);
1343 //printf("page=%x, addr=%x\n",page,head->vaddr);
1344 //assert(head->vaddr>>12==(page|0x80000));
1345 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1346 struct ht_entry *ht_bin = hash_table_get(head->vaddr);
1347 if (ht_bin->vaddr[0] == head->vaddr)
1348 ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
1349 if (ht_bin->vaddr[1] == head->vaddr)
1350 ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
1351 }
1352 }
1353 }
1354 }
1355 }
1356 head=head->next;
1357 }
1358}
1359
1360/* Register allocation */
1361
1362// Note: registers are allocated clean (unmodified state)
1363// if you intend to modify the register, you must call dirty_reg().
1364static void alloc_reg(struct regstat *cur,int i,signed char reg)
1365{
1366 int r,hr;
1367 int preferred_reg = (reg&7);
1368 if(reg==CCREG) preferred_reg=HOST_CCREG;
1369 if(reg==PTEMP||reg==FTEMP) preferred_reg=12;
1370
1371 // Don't allocate unused registers
1372 if((cur->u>>reg)&1) return;
1373
1374 // see if it's already allocated
1375 for(hr=0;hr<HOST_REGS;hr++)
1376 {
1377 if(cur->regmap[hr]==reg) return;
1378 }
1379
1380 // Keep the same mapping if the register was already allocated in a loop
1381 preferred_reg = loop_reg(i,reg,preferred_reg);
1382
1383 // Try to allocate the preferred register
1384 if(cur->regmap[preferred_reg]==-1) {
1385 cur->regmap[preferred_reg]=reg;
1386 cur->dirty&=~(1<<preferred_reg);
1387 cur->isconst&=~(1<<preferred_reg);
1388 return;
1389 }
1390 r=cur->regmap[preferred_reg];
1391 assert(r < 64);
1392 if((cur->u>>r)&1) {
1393 cur->regmap[preferred_reg]=reg;
1394 cur->dirty&=~(1<<preferred_reg);
1395 cur->isconst&=~(1<<preferred_reg);
1396 return;
1397 }
1398
1399 // Clear any unneeded registers
1400 // We try to keep the mapping consistent, if possible, because it
1401 // makes branches easier (especially loops). So we try to allocate
1402 // first (see above) before removing old mappings. If this is not
1403 // possible then go ahead and clear out the registers that are no
1404 // longer needed.
1405 for(hr=0;hr<HOST_REGS;hr++)
1406 {
1407 r=cur->regmap[hr];
1408 if(r>=0) {
1409 assert(r < 64);
1410 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1411 }
1412 }
1413 // Try to allocate any available register, but prefer
1414 // registers that have not been used recently.
1415 if(i>0) {
1416 for(hr=0;hr<HOST_REGS;hr++) {
1417 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1418 if(regs[i-1].regmap[hr]!=dops[i-1].rs1&&regs[i-1].regmap[hr]!=dops[i-1].rs2&&regs[i-1].regmap[hr]!=dops[i-1].rt1&&regs[i-1].regmap[hr]!=dops[i-1].rt2) {
1419 cur->regmap[hr]=reg;
1420 cur->dirty&=~(1<<hr);
1421 cur->isconst&=~(1<<hr);
1422 return;
1423 }
1424 }
1425 }
1426 }
1427 // Try to allocate any available register
1428 for(hr=0;hr<HOST_REGS;hr++) {
1429 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1430 cur->regmap[hr]=reg;
1431 cur->dirty&=~(1<<hr);
1432 cur->isconst&=~(1<<hr);
1433 return;
1434 }
1435 }
1436
1437 // Ok, now we have to evict someone
1438 // Pick a register we hopefully won't need soon
1439 u_char hsn[MAXREG+1];
1440 memset(hsn,10,sizeof(hsn));
1441 int j;
1442 lsn(hsn,i,&preferred_reg);
1443 //printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",cur->regmap[0],cur->regmap[1],cur->regmap[2],cur->regmap[3],cur->regmap[5],cur->regmap[6],cur->regmap[7]);
1444 //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]);
1445 if(i>0) {
1446 // Don't evict the cycle count at entry points, otherwise the entry
1447 // stub will have to write it.
1448 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1449 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1450 for(j=10;j>=3;j--)
1451 {
1452 // Alloc preferred register if available
1453 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1454 for(hr=0;hr<HOST_REGS;hr++) {
1455 // Evict both parts of a 64-bit register
1456 if((cur->regmap[hr]&63)==r) {
1457 cur->regmap[hr]=-1;
1458 cur->dirty&=~(1<<hr);
1459 cur->isconst&=~(1<<hr);
1460 }
1461 }
1462 cur->regmap[preferred_reg]=reg;
1463 return;
1464 }
1465 for(r=1;r<=MAXREG;r++)
1466 {
1467 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1468 for(hr=0;hr<HOST_REGS;hr++) {
1469 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1470 if(cur->regmap[hr]==r) {
1471 cur->regmap[hr]=reg;
1472 cur->dirty&=~(1<<hr);
1473 cur->isconst&=~(1<<hr);
1474 return;
1475 }
1476 }
1477 }
1478 }
1479 }
1480 }
1481 }
1482 for(j=10;j>=0;j--)
1483 {
1484 for(r=1;r<=MAXREG;r++)
1485 {
1486 if(hsn[r]==j) {
1487 for(hr=0;hr<HOST_REGS;hr++) {
1488 if(cur->regmap[hr]==r) {
1489 cur->regmap[hr]=reg;
1490 cur->dirty&=~(1<<hr);
1491 cur->isconst&=~(1<<hr);
1492 return;
1493 }
1494 }
1495 }
1496 }
1497 }
1498 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1499}
1500
1501// Allocate a temporary register. This is done without regard to
1502// dirty status or whether the register we request is on the unneeded list
1503// Note: This will only allocate one register, even if called multiple times
1504static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1505{
1506 int r,hr;
1507 int preferred_reg = -1;
1508
1509 // see if it's already allocated
1510 for(hr=0;hr<HOST_REGS;hr++)
1511 {
1512 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1513 }
1514
1515 // Try to allocate any available register
1516 for(hr=HOST_REGS-1;hr>=0;hr--) {
1517 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1518 cur->regmap[hr]=reg;
1519 cur->dirty&=~(1<<hr);
1520 cur->isconst&=~(1<<hr);
1521 return;
1522 }
1523 }
1524
1525 // Find an unneeded register
1526 for(hr=HOST_REGS-1;hr>=0;hr--)
1527 {
1528 r=cur->regmap[hr];
1529 if(r>=0) {
1530 assert(r < 64);
1531 if((cur->u>>r)&1) {
1532 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1533 cur->regmap[hr]=reg;
1534 cur->dirty&=~(1<<hr);
1535 cur->isconst&=~(1<<hr);
1536 return;
1537 }
1538 }
1539 }
1540 }
1541
1542 // Ok, now we have to evict someone
1543 // Pick a register we hopefully won't need soon
1544 // TODO: we might want to follow unconditional jumps here
1545 // TODO: get rid of dupe code and make this into a function
1546 u_char hsn[MAXREG+1];
1547 memset(hsn,10,sizeof(hsn));
1548 int j;
1549 lsn(hsn,i,&preferred_reg);
1550 //printf("hsn: %d %d %d %d %d %d %d\n",hsn[cur->regmap[0]&63],hsn[cur->regmap[1]&63],hsn[cur->regmap[2]&63],hsn[cur->regmap[3]&63],hsn[cur->regmap[5]&63],hsn[cur->regmap[6]&63],hsn[cur->regmap[7]&63]);
1551 if(i>0) {
1552 // Don't evict the cycle count at entry points, otherwise the entry
1553 // stub will have to write it.
1554 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1555 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1556 for(j=10;j>=3;j--)
1557 {
1558 for(r=1;r<=MAXREG;r++)
1559 {
1560 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1561 for(hr=0;hr<HOST_REGS;hr++) {
1562 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1563 if(cur->regmap[hr]==r) {
1564 cur->regmap[hr]=reg;
1565 cur->dirty&=~(1<<hr);
1566 cur->isconst&=~(1<<hr);
1567 return;
1568 }
1569 }
1570 }
1571 }
1572 }
1573 }
1574 }
1575 for(j=10;j>=0;j--)
1576 {
1577 for(r=1;r<=MAXREG;r++)
1578 {
1579 if(hsn[r]==j) {
1580 for(hr=0;hr<HOST_REGS;hr++) {
1581 if(cur->regmap[hr]==r) {
1582 cur->regmap[hr]=reg;
1583 cur->dirty&=~(1<<hr);
1584 cur->isconst&=~(1<<hr);
1585 return;
1586 }
1587 }
1588 }
1589 }
1590 }
1591 SysPrintf("This shouldn't happen");abort();
1592}
1593
1594static void mov_alloc(struct regstat *current,int i)
1595{
1596 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1597 // logically this is needed but just won't work, no idea why
1598 //alloc_cc(current,i); // for stalls
1599 //dirty_reg(current,CCREG);
1600 }
1601
1602 // Note: Don't need to actually alloc the source registers
1603 //alloc_reg(current,i,dops[i].rs1);
1604 alloc_reg(current,i,dops[i].rt1);
1605
1606 clear_const(current,dops[i].rs1);
1607 clear_const(current,dops[i].rt1);
1608 dirty_reg(current,dops[i].rt1);
1609}
1610
1611static void shiftimm_alloc(struct regstat *current,int i)
1612{
1613 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1614 {
1615 if(dops[i].rt1) {
1616 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1617 else dops[i].lt1=dops[i].rs1;
1618 alloc_reg(current,i,dops[i].rt1);
1619 dirty_reg(current,dops[i].rt1);
1620 if(is_const(current,dops[i].rs1)) {
1621 int v=get_const(current,dops[i].rs1);
1622 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1623 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1624 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1625 }
1626 else clear_const(current,dops[i].rt1);
1627 }
1628 }
1629 else
1630 {
1631 clear_const(current,dops[i].rs1);
1632 clear_const(current,dops[i].rt1);
1633 }
1634
1635 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1636 {
1637 assert(0);
1638 }
1639 if(dops[i].opcode2==0x3c) // DSLL32
1640 {
1641 assert(0);
1642 }
1643 if(dops[i].opcode2==0x3e) // DSRL32
1644 {
1645 assert(0);
1646 }
1647 if(dops[i].opcode2==0x3f) // DSRA32
1648 {
1649 assert(0);
1650 }
1651}
1652
1653static void shift_alloc(struct regstat *current,int i)
1654{
1655 if(dops[i].rt1) {
1656 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1657 {
1658 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1659 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1660 alloc_reg(current,i,dops[i].rt1);
1661 if(dops[i].rt1==dops[i].rs2) {
1662 alloc_reg_temp(current,i,-1);
1663 minimum_free_regs[i]=1;
1664 }
1665 } else { // DSLLV/DSRLV/DSRAV
1666 assert(0);
1667 }
1668 clear_const(current,dops[i].rs1);
1669 clear_const(current,dops[i].rs2);
1670 clear_const(current,dops[i].rt1);
1671 dirty_reg(current,dops[i].rt1);
1672 }
1673}
1674
1675static void alu_alloc(struct regstat *current,int i)
1676{
1677 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1678 if(dops[i].rt1) {
1679 if(dops[i].rs1&&dops[i].rs2) {
1680 alloc_reg(current,i,dops[i].rs1);
1681 alloc_reg(current,i,dops[i].rs2);
1682 }
1683 else {
1684 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1685 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1686 }
1687 alloc_reg(current,i,dops[i].rt1);
1688 }
1689 }
1690 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1691 if(dops[i].rt1) {
1692 alloc_reg(current,i,dops[i].rs1);
1693 alloc_reg(current,i,dops[i].rs2);
1694 alloc_reg(current,i,dops[i].rt1);
1695 }
1696 }
1697 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1698 if(dops[i].rt1) {
1699 if(dops[i].rs1&&dops[i].rs2) {
1700 alloc_reg(current,i,dops[i].rs1);
1701 alloc_reg(current,i,dops[i].rs2);
1702 }
1703 else
1704 {
1705 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1706 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1707 }
1708 alloc_reg(current,i,dops[i].rt1);
1709 }
1710 }
1711 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1712 assert(0);
1713 }
1714 clear_const(current,dops[i].rs1);
1715 clear_const(current,dops[i].rs2);
1716 clear_const(current,dops[i].rt1);
1717 dirty_reg(current,dops[i].rt1);
1718}
1719
1720static void imm16_alloc(struct regstat *current,int i)
1721{
1722 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1723 else dops[i].lt1=dops[i].rs1;
1724 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1725 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1726 assert(0);
1727 }
1728 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1729 clear_const(current,dops[i].rs1);
1730 clear_const(current,dops[i].rt1);
1731 }
1732 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1733 if(is_const(current,dops[i].rs1)) {
1734 int v=get_const(current,dops[i].rs1);
1735 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1736 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1737 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1738 }
1739 else clear_const(current,dops[i].rt1);
1740 }
1741 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1742 if(is_const(current,dops[i].rs1)) {
1743 int v=get_const(current,dops[i].rs1);
1744 set_const(current,dops[i].rt1,v+imm[i]);
1745 }
1746 else clear_const(current,dops[i].rt1);
1747 }
1748 else {
1749 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1750 }
1751 dirty_reg(current,dops[i].rt1);
1752}
1753
1754static void load_alloc(struct regstat *current,int i)
1755{
1756 clear_const(current,dops[i].rt1);
1757 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
1758 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
1759 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1760 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
1761 alloc_reg(current,i,dops[i].rt1);
1762 assert(get_reg(current->regmap,dops[i].rt1)>=0);
1763 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
1764 {
1765 assert(0);
1766 }
1767 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1768 {
1769 assert(0);
1770 }
1771 dirty_reg(current,dops[i].rt1);
1772 // LWL/LWR need a temporary register for the old value
1773 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1774 {
1775 alloc_reg(current,i,FTEMP);
1776 alloc_reg_temp(current,i,-1);
1777 minimum_free_regs[i]=1;
1778 }
1779 }
1780 else
1781 {
1782 // Load to r0 or unneeded register (dummy load)
1783 // but we still need a register to calculate the address
1784 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1785 {
1786 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1787 }
1788 alloc_reg_temp(current,i,-1);
1789 minimum_free_regs[i]=1;
1790 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1791 {
1792 assert(0);
1793 }
1794 }
1795}
1796
1797void store_alloc(struct regstat *current,int i)
1798{
1799 clear_const(current,dops[i].rs2);
1800 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
1801 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1802 alloc_reg(current,i,dops[i].rs2);
1803 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
1804 assert(0);
1805 }
1806 #if defined(HOST_IMM8)
1807 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1808 else alloc_reg(current,i,INVCP);
1809 #endif
1810 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
1811 alloc_reg(current,i,FTEMP);
1812 }
1813 // We need a temporary register for address generation
1814 alloc_reg_temp(current,i,-1);
1815 minimum_free_regs[i]=1;
1816}
1817
1818void c1ls_alloc(struct regstat *current,int i)
1819{
1820 //clear_const(current,dops[i].rs1); // FIXME
1821 clear_const(current,dops[i].rt1);
1822 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1823 alloc_reg(current,i,CSREG); // Status
1824 alloc_reg(current,i,FTEMP);
1825 if(dops[i].opcode==0x35||dops[i].opcode==0x3d) { // 64-bit LDC1/SDC1
1826 assert(0);
1827 }
1828 #if defined(HOST_IMM8)
1829 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1830 else if((dops[i].opcode&0x3b)==0x39) // SWC1/SDC1
1831 alloc_reg(current,i,INVCP);
1832 #endif
1833 // We need a temporary register for address generation
1834 alloc_reg_temp(current,i,-1);
1835}
1836
1837void c2ls_alloc(struct regstat *current,int i)
1838{
1839 clear_const(current,dops[i].rt1);
1840 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1841 alloc_reg(current,i,FTEMP);
1842 #if defined(HOST_IMM8)
1843 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1844 if((dops[i].opcode&0x3b)==0x3a) // SWC2/SDC2
1845 alloc_reg(current,i,INVCP);
1846 #endif
1847 // We need a temporary register for address generation
1848 alloc_reg_temp(current,i,-1);
1849 minimum_free_regs[i]=1;
1850}
1851
1852#ifndef multdiv_alloc
1853void multdiv_alloc(struct regstat *current,int i)
1854{
1855 // case 0x18: MULT
1856 // case 0x19: MULTU
1857 // case 0x1A: DIV
1858 // case 0x1B: DIVU
1859 // case 0x1C: DMULT
1860 // case 0x1D: DMULTU
1861 // case 0x1E: DDIV
1862 // case 0x1F: DDIVU
1863 clear_const(current,dops[i].rs1);
1864 clear_const(current,dops[i].rs2);
1865 alloc_cc(current,i); // for stalls
1866 if(dops[i].rs1&&dops[i].rs2)
1867 {
1868 if((dops[i].opcode2&4)==0) // 32-bit
1869 {
1870 current->u&=~(1LL<<HIREG);
1871 current->u&=~(1LL<<LOREG);
1872 alloc_reg(current,i,HIREG);
1873 alloc_reg(current,i,LOREG);
1874 alloc_reg(current,i,dops[i].rs1);
1875 alloc_reg(current,i,dops[i].rs2);
1876 dirty_reg(current,HIREG);
1877 dirty_reg(current,LOREG);
1878 }
1879 else // 64-bit
1880 {
1881 assert(0);
1882 }
1883 }
1884 else
1885 {
1886 // Multiply by zero is zero.
1887 // MIPS does not have a divide by zero exception.
1888 // The result is undefined, we return zero.
1889 alloc_reg(current,i,HIREG);
1890 alloc_reg(current,i,LOREG);
1891 dirty_reg(current,HIREG);
1892 dirty_reg(current,LOREG);
1893 }
1894}
1895#endif
1896
1897void cop0_alloc(struct regstat *current,int i)
1898{
1899 if(dops[i].opcode2==0) // MFC0
1900 {
1901 if(dops[i].rt1) {
1902 clear_const(current,dops[i].rt1);
1903 alloc_all(current,i);
1904 alloc_reg(current,i,dops[i].rt1);
1905 dirty_reg(current,dops[i].rt1);
1906 }
1907 }
1908 else if(dops[i].opcode2==4) // MTC0
1909 {
1910 if(dops[i].rs1){
1911 clear_const(current,dops[i].rs1);
1912 alloc_reg(current,i,dops[i].rs1);
1913 alloc_all(current,i);
1914 }
1915 else {
1916 alloc_all(current,i); // FIXME: Keep r0
1917 current->u&=~1LL;
1918 alloc_reg(current,i,0);
1919 }
1920 }
1921 else
1922 {
1923 // TLBR/TLBWI/TLBWR/TLBP/ERET
1924 assert(dops[i].opcode2==0x10);
1925 alloc_all(current,i);
1926 }
1927 minimum_free_regs[i]=HOST_REGS;
1928}
1929
1930static void cop2_alloc(struct regstat *current,int i)
1931{
1932 if (dops[i].opcode2 < 3) // MFC2/CFC2
1933 {
1934 alloc_cc(current,i); // for stalls
1935 dirty_reg(current,CCREG);
1936 if(dops[i].rt1){
1937 clear_const(current,dops[i].rt1);
1938 alloc_reg(current,i,dops[i].rt1);
1939 dirty_reg(current,dops[i].rt1);
1940 }
1941 }
1942 else if (dops[i].opcode2 > 3) // MTC2/CTC2
1943 {
1944 if(dops[i].rs1){
1945 clear_const(current,dops[i].rs1);
1946 alloc_reg(current,i,dops[i].rs1);
1947 }
1948 else {
1949 current->u&=~1LL;
1950 alloc_reg(current,i,0);
1951 }
1952 }
1953 alloc_reg_temp(current,i,-1);
1954 minimum_free_regs[i]=1;
1955}
1956
1957void c2op_alloc(struct regstat *current,int i)
1958{
1959 alloc_cc(current,i); // for stalls
1960 dirty_reg(current,CCREG);
1961 alloc_reg_temp(current,i,-1);
1962}
1963
1964void syscall_alloc(struct regstat *current,int i)
1965{
1966 alloc_cc(current,i);
1967 dirty_reg(current,CCREG);
1968 alloc_all(current,i);
1969 minimum_free_regs[i]=HOST_REGS;
1970 current->isconst=0;
1971}
1972
1973void delayslot_alloc(struct regstat *current,int i)
1974{
1975 switch(dops[i].itype) {
1976 case UJUMP:
1977 case CJUMP:
1978 case SJUMP:
1979 case RJUMP:
1980 case SYSCALL:
1981 case HLECALL:
1982 case SPAN:
1983 assem_debug("jump in the delay slot. this shouldn't happen.\n");//abort();
1984 SysPrintf("Disabled speculative precompilation\n");
1985 stop_after_jal=1;
1986 break;
1987 case IMM16:
1988 imm16_alloc(current,i);
1989 break;
1990 case LOAD:
1991 case LOADLR:
1992 load_alloc(current,i);
1993 break;
1994 case STORE:
1995 case STORELR:
1996 store_alloc(current,i);
1997 break;
1998 case ALU:
1999 alu_alloc(current,i);
2000 break;
2001 case SHIFT:
2002 shift_alloc(current,i);
2003 break;
2004 case MULTDIV:
2005 multdiv_alloc(current,i);
2006 break;
2007 case SHIFTIMM:
2008 shiftimm_alloc(current,i);
2009 break;
2010 case MOV:
2011 mov_alloc(current,i);
2012 break;
2013 case COP0:
2014 cop0_alloc(current,i);
2015 break;
2016 case COP1:
2017 break;
2018 case COP2:
2019 cop2_alloc(current,i);
2020 break;
2021 case C1LS:
2022 c1ls_alloc(current,i);
2023 break;
2024 case C2LS:
2025 c2ls_alloc(current,i);
2026 break;
2027 case C2OP:
2028 c2op_alloc(current,i);
2029 break;
2030 }
2031}
2032
2033// Special case where a branch and delay slot span two pages in virtual memory
2034static void pagespan_alloc(struct regstat *current,int i)
2035{
2036 current->isconst=0;
2037 current->wasconst=0;
2038 regs[i].wasconst=0;
2039 minimum_free_regs[i]=HOST_REGS;
2040 alloc_all(current,i);
2041 alloc_cc(current,i);
2042 dirty_reg(current,CCREG);
2043 if(dops[i].opcode==3) // JAL
2044 {
2045 alloc_reg(current,i,31);
2046 dirty_reg(current,31);
2047 }
2048 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
2049 {
2050 alloc_reg(current,i,dops[i].rs1);
2051 if (dops[i].rt1!=0) {
2052 alloc_reg(current,i,dops[i].rt1);
2053 dirty_reg(current,dops[i].rt1);
2054 }
2055 }
2056 if((dops[i].opcode&0x2E)==4) // BEQ/BNE/BEQL/BNEL
2057 {
2058 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2059 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
2060 }
2061 else
2062 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
2063 {
2064 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2065 }
2066 //else ...
2067}
2068
2069static void add_stub(enum stub_type type, void *addr, void *retaddr,
2070 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2071{
2072 assert(stubcount < ARRAY_SIZE(stubs));
2073 stubs[stubcount].type = type;
2074 stubs[stubcount].addr = addr;
2075 stubs[stubcount].retaddr = retaddr;
2076 stubs[stubcount].a = a;
2077 stubs[stubcount].b = b;
2078 stubs[stubcount].c = c;
2079 stubs[stubcount].d = d;
2080 stubs[stubcount].e = e;
2081 stubcount++;
2082}
2083
2084static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2085 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2086{
2087 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2088}
2089
2090// Write out a single register
2091static void wb_register(signed char r,signed char regmap[],uint64_t dirty)
2092{
2093 int hr;
2094 for(hr=0;hr<HOST_REGS;hr++) {
2095 if(hr!=EXCLUDE_REG) {
2096 if((regmap[hr]&63)==r) {
2097 if((dirty>>hr)&1) {
2098 assert(regmap[hr]<64);
2099 emit_storereg(r,hr);
2100 }
2101 }
2102 }
2103 }
2104}
2105
2106static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2107{
2108 //if(dirty_pre==dirty) return;
2109 int hr,reg;
2110 for(hr=0;hr<HOST_REGS;hr++) {
2111 if(hr!=EXCLUDE_REG) {
2112 reg=pre[hr];
2113 if(((~u)>>(reg&63))&1) {
2114 if(reg>0) {
2115 if(((dirty_pre&~dirty)>>hr)&1) {
2116 if(reg>0&&reg<34) {
2117 emit_storereg(reg,hr);
2118 }
2119 else if(reg>=64) {
2120 assert(0);
2121 }
2122 }
2123 }
2124 }
2125 }
2126 }
2127}
2128
2129// trashes r2
2130static void pass_args(int a0, int a1)
2131{
2132 if(a0==1&&a1==0) {
2133 // must swap
2134 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2135 }
2136 else if(a0!=0&&a1==0) {
2137 emit_mov(a1,1);
2138 if (a0>=0) emit_mov(a0,0);
2139 }
2140 else {
2141 if(a0>=0&&a0!=0) emit_mov(a0,0);
2142 if(a1>=0&&a1!=1) emit_mov(a1,1);
2143 }
2144}
2145
2146static void alu_assemble(int i,struct regstat *i_regs)
2147{
2148 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2149 if(dops[i].rt1) {
2150 signed char s1,s2,t;
2151 t=get_reg(i_regs->regmap,dops[i].rt1);
2152 if(t>=0) {
2153 s1=get_reg(i_regs->regmap,dops[i].rs1);
2154 s2=get_reg(i_regs->regmap,dops[i].rs2);
2155 if(dops[i].rs1&&dops[i].rs2) {
2156 assert(s1>=0);
2157 assert(s2>=0);
2158 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2159 else emit_add(s1,s2,t);
2160 }
2161 else if(dops[i].rs1) {
2162 if(s1>=0) emit_mov(s1,t);
2163 else emit_loadreg(dops[i].rs1,t);
2164 }
2165 else if(dops[i].rs2) {
2166 if(s2>=0) {
2167 if(dops[i].opcode2&2) emit_neg(s2,t);
2168 else emit_mov(s2,t);
2169 }
2170 else {
2171 emit_loadreg(dops[i].rs2,t);
2172 if(dops[i].opcode2&2) emit_neg(t,t);
2173 }
2174 }
2175 else emit_zeroreg(t);
2176 }
2177 }
2178 }
2179 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2180 assert(0);
2181 }
2182 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2183 if(dops[i].rt1) {
2184 signed char s1l,s2l,t;
2185 {
2186 t=get_reg(i_regs->regmap,dops[i].rt1);
2187 //assert(t>=0);
2188 if(t>=0) {
2189 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2190 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2191 if(dops[i].rs2==0) // rx<r0
2192 {
2193 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2194 assert(s1l>=0);
2195 emit_shrimm(s1l,31,t);
2196 }
2197 else // SLTU (unsigned can not be less than zero, 0<0)
2198 emit_zeroreg(t);
2199 }
2200 else if(dops[i].rs1==0) // r0<rx
2201 {
2202 assert(s2l>=0);
2203 if(dops[i].opcode2==0x2a) // SLT
2204 emit_set_gz32(s2l,t);
2205 else // SLTU (set if not zero)
2206 emit_set_nz32(s2l,t);
2207 }
2208 else{
2209 assert(s1l>=0);assert(s2l>=0);
2210 if(dops[i].opcode2==0x2a) // SLT
2211 emit_set_if_less32(s1l,s2l,t);
2212 else // SLTU
2213 emit_set_if_carry32(s1l,s2l,t);
2214 }
2215 }
2216 }
2217 }
2218 }
2219 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2220 if(dops[i].rt1) {
2221 signed char s1l,s2l,tl;
2222 tl=get_reg(i_regs->regmap,dops[i].rt1);
2223 {
2224 if(tl>=0) {
2225 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2226 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2227 if(dops[i].rs1&&dops[i].rs2) {
2228 assert(s1l>=0);
2229 assert(s2l>=0);
2230 if(dops[i].opcode2==0x24) { // AND
2231 emit_and(s1l,s2l,tl);
2232 } else
2233 if(dops[i].opcode2==0x25) { // OR
2234 emit_or(s1l,s2l,tl);
2235 } else
2236 if(dops[i].opcode2==0x26) { // XOR
2237 emit_xor(s1l,s2l,tl);
2238 } else
2239 if(dops[i].opcode2==0x27) { // NOR
2240 emit_or(s1l,s2l,tl);
2241 emit_not(tl,tl);
2242 }
2243 }
2244 else
2245 {
2246 if(dops[i].opcode2==0x24) { // AND
2247 emit_zeroreg(tl);
2248 } else
2249 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2250 if(dops[i].rs1){
2251 if(s1l>=0) emit_mov(s1l,tl);
2252 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2253 }
2254 else
2255 if(dops[i].rs2){
2256 if(s2l>=0) emit_mov(s2l,tl);
2257 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2258 }
2259 else emit_zeroreg(tl);
2260 } else
2261 if(dops[i].opcode2==0x27) { // NOR
2262 if(dops[i].rs1){
2263 if(s1l>=0) emit_not(s1l,tl);
2264 else {
2265 emit_loadreg(dops[i].rs1,tl);
2266 emit_not(tl,tl);
2267 }
2268 }
2269 else
2270 if(dops[i].rs2){
2271 if(s2l>=0) emit_not(s2l,tl);
2272 else {
2273 emit_loadreg(dops[i].rs2,tl);
2274 emit_not(tl,tl);
2275 }
2276 }
2277 else emit_movimm(-1,tl);
2278 }
2279 }
2280 }
2281 }
2282 }
2283 }
2284}
2285
2286void imm16_assemble(int i,struct regstat *i_regs)
2287{
2288 if (dops[i].opcode==0x0f) { // LUI
2289 if(dops[i].rt1) {
2290 signed char t;
2291 t=get_reg(i_regs->regmap,dops[i].rt1);
2292 //assert(t>=0);
2293 if(t>=0) {
2294 if(!((i_regs->isconst>>t)&1))
2295 emit_movimm(imm[i]<<16,t);
2296 }
2297 }
2298 }
2299 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2300 if(dops[i].rt1) {
2301 signed char s,t;
2302 t=get_reg(i_regs->regmap,dops[i].rt1);
2303 s=get_reg(i_regs->regmap,dops[i].rs1);
2304 if(dops[i].rs1) {
2305 //assert(t>=0);
2306 //assert(s>=0);
2307 if(t>=0) {
2308 if(!((i_regs->isconst>>t)&1)) {
2309 if(s<0) {
2310 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2311 emit_addimm(t,imm[i],t);
2312 }else{
2313 if(!((i_regs->wasconst>>s)&1))
2314 emit_addimm(s,imm[i],t);
2315 else
2316 emit_movimm(constmap[i][s]+imm[i],t);
2317 }
2318 }
2319 }
2320 } else {
2321 if(t>=0) {
2322 if(!((i_regs->isconst>>t)&1))
2323 emit_movimm(imm[i],t);
2324 }
2325 }
2326 }
2327 }
2328 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2329 if(dops[i].rt1) {
2330 signed char sl,tl;
2331 tl=get_reg(i_regs->regmap,dops[i].rt1);
2332 sl=get_reg(i_regs->regmap,dops[i].rs1);
2333 if(tl>=0) {
2334 if(dops[i].rs1) {
2335 assert(sl>=0);
2336 emit_addimm(sl,imm[i],tl);
2337 } else {
2338 emit_movimm(imm[i],tl);
2339 }
2340 }
2341 }
2342 }
2343 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2344 if(dops[i].rt1) {
2345 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2346 signed char sl,t;
2347 t=get_reg(i_regs->regmap,dops[i].rt1);
2348 sl=get_reg(i_regs->regmap,dops[i].rs1);
2349 //assert(t>=0);
2350 if(t>=0) {
2351 if(dops[i].rs1>0) {
2352 if(dops[i].opcode==0x0a) { // SLTI
2353 if(sl<0) {
2354 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2355 emit_slti32(t,imm[i],t);
2356 }else{
2357 emit_slti32(sl,imm[i],t);
2358 }
2359 }
2360 else { // SLTIU
2361 if(sl<0) {
2362 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2363 emit_sltiu32(t,imm[i],t);
2364 }else{
2365 emit_sltiu32(sl,imm[i],t);
2366 }
2367 }
2368 }else{
2369 // SLTI(U) with r0 is just stupid,
2370 // nonetheless examples can be found
2371 if(dops[i].opcode==0x0a) // SLTI
2372 if(0<imm[i]) emit_movimm(1,t);
2373 else emit_zeroreg(t);
2374 else // SLTIU
2375 {
2376 if(imm[i]) emit_movimm(1,t);
2377 else emit_zeroreg(t);
2378 }
2379 }
2380 }
2381 }
2382 }
2383 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2384 if(dops[i].rt1) {
2385 signed char sl,tl;
2386 tl=get_reg(i_regs->regmap,dops[i].rt1);
2387 sl=get_reg(i_regs->regmap,dops[i].rs1);
2388 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2389 if(dops[i].opcode==0x0c) //ANDI
2390 {
2391 if(dops[i].rs1) {
2392 if(sl<0) {
2393 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2394 emit_andimm(tl,imm[i],tl);
2395 }else{
2396 if(!((i_regs->wasconst>>sl)&1))
2397 emit_andimm(sl,imm[i],tl);
2398 else
2399 emit_movimm(constmap[i][sl]&imm[i],tl);
2400 }
2401 }
2402 else
2403 emit_zeroreg(tl);
2404 }
2405 else
2406 {
2407 if(dops[i].rs1) {
2408 if(sl<0) {
2409 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2410 }
2411 if(dops[i].opcode==0x0d) { // ORI
2412 if(sl<0) {
2413 emit_orimm(tl,imm[i],tl);
2414 }else{
2415 if(!((i_regs->wasconst>>sl)&1))
2416 emit_orimm(sl,imm[i],tl);
2417 else
2418 emit_movimm(constmap[i][sl]|imm[i],tl);
2419 }
2420 }
2421 if(dops[i].opcode==0x0e) { // XORI
2422 if(sl<0) {
2423 emit_xorimm(tl,imm[i],tl);
2424 }else{
2425 if(!((i_regs->wasconst>>sl)&1))
2426 emit_xorimm(sl,imm[i],tl);
2427 else
2428 emit_movimm(constmap[i][sl]^imm[i],tl);
2429 }
2430 }
2431 }
2432 else {
2433 emit_movimm(imm[i],tl);
2434 }
2435 }
2436 }
2437 }
2438 }
2439}
2440
2441void shiftimm_assemble(int i,struct regstat *i_regs)
2442{
2443 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2444 {
2445 if(dops[i].rt1) {
2446 signed char s,t;
2447 t=get_reg(i_regs->regmap,dops[i].rt1);
2448 s=get_reg(i_regs->regmap,dops[i].rs1);
2449 //assert(t>=0);
2450 if(t>=0&&!((i_regs->isconst>>t)&1)){
2451 if(dops[i].rs1==0)
2452 {
2453 emit_zeroreg(t);
2454 }
2455 else
2456 {
2457 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2458 if(imm[i]) {
2459 if(dops[i].opcode2==0) // SLL
2460 {
2461 emit_shlimm(s<0?t:s,imm[i],t);
2462 }
2463 if(dops[i].opcode2==2) // SRL
2464 {
2465 emit_shrimm(s<0?t:s,imm[i],t);
2466 }
2467 if(dops[i].opcode2==3) // SRA
2468 {
2469 emit_sarimm(s<0?t:s,imm[i],t);
2470 }
2471 }else{
2472 // Shift by zero
2473 if(s>=0 && s!=t) emit_mov(s,t);
2474 }
2475 }
2476 }
2477 //emit_storereg(dops[i].rt1,t); //DEBUG
2478 }
2479 }
2480 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2481 {
2482 assert(0);
2483 }
2484 if(dops[i].opcode2==0x3c) // DSLL32
2485 {
2486 assert(0);
2487 }
2488 if(dops[i].opcode2==0x3e) // DSRL32
2489 {
2490 assert(0);
2491 }
2492 if(dops[i].opcode2==0x3f) // DSRA32
2493 {
2494 assert(0);
2495 }
2496}
2497
2498#ifndef shift_assemble
2499static void shift_assemble(int i,struct regstat *i_regs)
2500{
2501 signed char s,t,shift;
2502 if (dops[i].rt1 == 0)
2503 return;
2504 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2505 t = get_reg(i_regs->regmap, dops[i].rt1);
2506 s = get_reg(i_regs->regmap, dops[i].rs1);
2507 shift = get_reg(i_regs->regmap, dops[i].rs2);
2508 if (t < 0)
2509 return;
2510
2511 if(dops[i].rs1==0)
2512 emit_zeroreg(t);
2513 else if(dops[i].rs2==0) {
2514 assert(s>=0);
2515 if(s!=t) emit_mov(s,t);
2516 }
2517 else {
2518 host_tempreg_acquire();
2519 emit_andimm(shift,31,HOST_TEMPREG);
2520 switch(dops[i].opcode2) {
2521 case 4: // SLLV
2522 emit_shl(s,HOST_TEMPREG,t);
2523 break;
2524 case 6: // SRLV
2525 emit_shr(s,HOST_TEMPREG,t);
2526 break;
2527 case 7: // SRAV
2528 emit_sar(s,HOST_TEMPREG,t);
2529 break;
2530 default:
2531 assert(0);
2532 }
2533 host_tempreg_release();
2534 }
2535}
2536
2537#endif
2538
2539enum {
2540 MTYPE_8000 = 0,
2541 MTYPE_8020,
2542 MTYPE_0000,
2543 MTYPE_A000,
2544 MTYPE_1F80,
2545};
2546
2547static int get_ptr_mem_type(u_int a)
2548{
2549 if(a < 0x00200000) {
2550 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2551 // return wrong, must use memhandler for BIOS self-test to pass
2552 // 007 does similar stuff from a00 mirror, weird stuff
2553 return MTYPE_8000;
2554 return MTYPE_0000;
2555 }
2556 if(0x1f800000 <= a && a < 0x1f801000)
2557 return MTYPE_1F80;
2558 if(0x80200000 <= a && a < 0x80800000)
2559 return MTYPE_8020;
2560 if(0xa0000000 <= a && a < 0xa0200000)
2561 return MTYPE_A000;
2562 return MTYPE_8000;
2563}
2564
2565static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
2566{
2567 void *jaddr = NULL;
2568 int type=0;
2569 int mr=dops[i].rs1;
2570 if(((smrv_strong|smrv_weak)>>mr)&1) {
2571 type=get_ptr_mem_type(smrv[mr]);
2572 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2573 }
2574 else {
2575 // use the mirror we are running on
2576 type=get_ptr_mem_type(start);
2577 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2578 }
2579
2580 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2581 host_tempreg_acquire();
2582 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2583 addr=*addr_reg_override=HOST_TEMPREG;
2584 type=0;
2585 }
2586 else if(type==MTYPE_0000) { // RAM 0 mirror
2587 host_tempreg_acquire();
2588 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2589 addr=*addr_reg_override=HOST_TEMPREG;
2590 type=0;
2591 }
2592 else if(type==MTYPE_A000) { // RAM A mirror
2593 host_tempreg_acquire();
2594 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2595 addr=*addr_reg_override=HOST_TEMPREG;
2596 type=0;
2597 }
2598 else if(type==MTYPE_1F80) { // scratchpad
2599 if (psxH == (void *)0x1f800000) {
2600 host_tempreg_acquire();
2601 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2602 emit_cmpimm(HOST_TEMPREG,0x1000);
2603 host_tempreg_release();
2604 jaddr=out;
2605 emit_jc(0);
2606 }
2607 else {
2608 // do the usual RAM check, jump will go to the right handler
2609 type=0;
2610 }
2611 }
2612
2613 if(type==0)
2614 {
2615 emit_cmpimm(addr,RAM_SIZE);
2616 jaddr=out;
2617 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2618 // Hint to branch predictor that the branch is unlikely to be taken
2619 if(dops[i].rs1>=28)
2620 emit_jno_unlikely(0);
2621 else
2622 #endif
2623 emit_jno(0);
2624 if(ram_offset!=0) {
2625 host_tempreg_acquire();
2626 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2627 addr=*addr_reg_override=HOST_TEMPREG;
2628 }
2629 }
2630
2631 return jaddr;
2632}
2633
2634// return memhandler, or get directly accessable address and return 0
2635static void *get_direct_memhandler(void *table, u_int addr,
2636 enum stub_type type, uintptr_t *addr_host)
2637{
2638 uintptr_t l1, l2 = 0;
2639 l1 = ((uintptr_t *)table)[addr>>12];
2640 if ((l1 & (1ul << (sizeof(l1)*8-1))) == 0) {
2641 uintptr_t v = l1 << 1;
2642 *addr_host = v + addr;
2643 return NULL;
2644 }
2645 else {
2646 l1 <<= 1;
2647 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2648 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2649 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2650 l2=((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2651 else
2652 l2=((uintptr_t *)l1)[(addr&0xfff)/4];
2653 if ((l2 & (1<<31)) == 0) {
2654 uintptr_t v = l2 << 1;
2655 *addr_host = v + (addr&0xfff);
2656 return NULL;
2657 }
2658 return (void *)(l2 << 1);
2659 }
2660}
2661
2662static u_int get_host_reglist(const signed char *regmap)
2663{
2664 u_int reglist = 0, hr;
2665 for (hr = 0; hr < HOST_REGS; hr++) {
2666 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2667 reglist |= 1 << hr;
2668 }
2669 return reglist;
2670}
2671
2672static u_int reglist_exclude(u_int reglist, int r1, int r2)
2673{
2674 if (r1 >= 0)
2675 reglist &= ~(1u << r1);
2676 if (r2 >= 0)
2677 reglist &= ~(1u << r2);
2678 return reglist;
2679}
2680
2681// find a temp caller-saved register not in reglist (so assumed to be free)
2682static int reglist_find_free(u_int reglist)
2683{
2684 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2685 if (free_regs == 0)
2686 return -1;
2687 return __builtin_ctz(free_regs);
2688}
2689
2690static void load_assemble(int i, const struct regstat *i_regs)
2691{
2692 int s,tl,addr;
2693 int offset;
2694 void *jaddr=0;
2695 int memtarget=0,c=0;
2696 int fastio_reg_override=-1;
2697 u_int reglist=get_host_reglist(i_regs->regmap);
2698 tl=get_reg(i_regs->regmap,dops[i].rt1);
2699 s=get_reg(i_regs->regmap,dops[i].rs1);
2700 offset=imm[i];
2701 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2702 if(s>=0) {
2703 c=(i_regs->wasconst>>s)&1;
2704 if (c) {
2705 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2706 }
2707 }
2708 //printf("load_assemble: c=%d\n",c);
2709 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2710 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2711 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2712 ||dops[i].rt1==0) {
2713 // could be FIFO, must perform the read
2714 // ||dummy read
2715 assem_debug("(forced read)\n");
2716 tl=get_reg(i_regs->regmap,-1);
2717 assert(tl>=0);
2718 }
2719 if(offset||s<0||c) addr=tl;
2720 else addr=s;
2721 //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2722 if(tl>=0) {
2723 //printf("load_assemble: c=%d\n",c);
2724 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2725 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2726 reglist&=~(1<<tl);
2727 if(!c) {
2728 #ifdef R29_HACK
2729 // Strmnnrmn's speed hack
2730 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2731 #endif
2732 {
2733 jaddr=emit_fastpath_cmp_jump(i,addr,&fastio_reg_override);
2734 }
2735 }
2736 else if(ram_offset&&memtarget) {
2737 host_tempreg_acquire();
2738 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2739 fastio_reg_override=HOST_TEMPREG;
2740 }
2741 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2742 if (dops[i].opcode==0x20) { // LB
2743 if(!c||memtarget) {
2744 if(!dummy) {
2745 {
2746 int x=0,a=tl;
2747 if(!c) a=addr;
2748 if(fastio_reg_override>=0) a=fastio_reg_override;
2749
2750 emit_movsbl_indexed(x,a,tl);
2751 }
2752 }
2753 if(jaddr)
2754 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2755 }
2756 else
2757 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2758 }
2759 if (dops[i].opcode==0x21) { // LH
2760 if(!c||memtarget) {
2761 if(!dummy) {
2762 int x=0,a=tl;
2763 if(!c) a=addr;
2764 if(fastio_reg_override>=0) a=fastio_reg_override;
2765 emit_movswl_indexed(x,a,tl);
2766 }
2767 if(jaddr)
2768 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2769 }
2770 else
2771 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2772 }
2773 if (dops[i].opcode==0x23) { // LW
2774 if(!c||memtarget) {
2775 if(!dummy) {
2776 int a=addr;
2777 if(fastio_reg_override>=0) a=fastio_reg_override;
2778 emit_readword_indexed(0,a,tl);
2779 }
2780 if(jaddr)
2781 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2782 }
2783 else
2784 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2785 }
2786 if (dops[i].opcode==0x24) { // LBU
2787 if(!c||memtarget) {
2788 if(!dummy) {
2789 int x=0,a=tl;
2790 if(!c) a=addr;
2791 if(fastio_reg_override>=0) a=fastio_reg_override;
2792
2793 emit_movzbl_indexed(x,a,tl);
2794 }
2795 if(jaddr)
2796 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2797 }
2798 else
2799 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2800 }
2801 if (dops[i].opcode==0x25) { // LHU
2802 if(!c||memtarget) {
2803 if(!dummy) {
2804 int x=0,a=tl;
2805 if(!c) a=addr;
2806 if(fastio_reg_override>=0) a=fastio_reg_override;
2807 emit_movzwl_indexed(x,a,tl);
2808 }
2809 if(jaddr)
2810 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2811 }
2812 else
2813 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj[i],reglist);
2814 }
2815 if (dops[i].opcode==0x27) { // LWU
2816 assert(0);
2817 }
2818 if (dops[i].opcode==0x37) { // LD
2819 assert(0);
2820 }
2821 }
2822 if (fastio_reg_override == HOST_TEMPREG)
2823 host_tempreg_release();
2824}
2825
2826#ifndef loadlr_assemble
2827static void loadlr_assemble(int i, const struct regstat *i_regs)
2828{
2829 int s,tl,temp,temp2,addr;
2830 int offset;
2831 void *jaddr=0;
2832 int memtarget=0,c=0;
2833 int fastio_reg_override=-1;
2834 u_int reglist=get_host_reglist(i_regs->regmap);
2835 tl=get_reg(i_regs->regmap,dops[i].rt1);
2836 s=get_reg(i_regs->regmap,dops[i].rs1);
2837 temp=get_reg(i_regs->regmap,-1);
2838 temp2=get_reg(i_regs->regmap,FTEMP);
2839 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
2840 assert(addr<0);
2841 offset=imm[i];
2842 reglist|=1<<temp;
2843 if(offset||s<0||c) addr=temp2;
2844 else addr=s;
2845 if(s>=0) {
2846 c=(i_regs->wasconst>>s)&1;
2847 if(c) {
2848 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2849 }
2850 }
2851 if(!c) {
2852 emit_shlimm(addr,3,temp);
2853 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2854 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
2855 }else{
2856 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
2857 }
2858 jaddr=emit_fastpath_cmp_jump(i,temp2,&fastio_reg_override);
2859 }
2860 else {
2861 if(ram_offset&&memtarget) {
2862 host_tempreg_acquire();
2863 emit_addimm(temp2,ram_offset,HOST_TEMPREG);
2864 fastio_reg_override=HOST_TEMPREG;
2865 }
2866 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2867 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
2868 }else{
2869 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
2870 }
2871 }
2872 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
2873 if(!c||memtarget) {
2874 int a=temp2;
2875 if(fastio_reg_override>=0) a=fastio_reg_override;
2876 emit_readword_indexed(0,a,temp2);
2877 if(fastio_reg_override==HOST_TEMPREG) host_tempreg_release();
2878 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj[i],reglist);
2879 }
2880 else
2881 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj[i],reglist);
2882 if(dops[i].rt1) {
2883 assert(tl>=0);
2884 emit_andimm(temp,24,temp);
2885 if (dops[i].opcode==0x22) // LWL
2886 emit_xorimm(temp,24,temp);
2887 host_tempreg_acquire();
2888 emit_movimm(-1,HOST_TEMPREG);
2889 if (dops[i].opcode==0x26) {
2890 emit_shr(temp2,temp,temp2);
2891 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
2892 }else{
2893 emit_shl(temp2,temp,temp2);
2894 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
2895 }
2896 host_tempreg_release();
2897 emit_or(temp2,tl,tl);
2898 }
2899 //emit_storereg(dops[i].rt1,tl); // DEBUG
2900 }
2901 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
2902 assert(0);
2903 }
2904}
2905#endif
2906
2907void store_assemble(int i, const struct regstat *i_regs)
2908{
2909 int s,tl;
2910 int addr,temp;
2911 int offset;
2912 void *jaddr=0;
2913 enum stub_type type;
2914 int memtarget=0,c=0;
2915 int agr=AGEN1+(i&1);
2916 int fastio_reg_override=-1;
2917 u_int reglist=get_host_reglist(i_regs->regmap);
2918 tl=get_reg(i_regs->regmap,dops[i].rs2);
2919 s=get_reg(i_regs->regmap,dops[i].rs1);
2920 temp=get_reg(i_regs->regmap,agr);
2921 if(temp<0) temp=get_reg(i_regs->regmap,-1);
2922 offset=imm[i];
2923 if(s>=0) {
2924 c=(i_regs->wasconst>>s)&1;
2925 if(c) {
2926 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2927 }
2928 }
2929 assert(tl>=0);
2930 assert(temp>=0);
2931 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2932 if(offset||s<0||c) addr=temp;
2933 else addr=s;
2934 if(!c) {
2935 jaddr=emit_fastpath_cmp_jump(i,addr,&fastio_reg_override);
2936 }
2937 else if(ram_offset&&memtarget) {
2938 host_tempreg_acquire();
2939 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2940 fastio_reg_override=HOST_TEMPREG;
2941 }
2942
2943 if (dops[i].opcode==0x28) { // SB
2944 if(!c||memtarget) {
2945 int x=0,a=temp;
2946 if(!c) a=addr;
2947 if(fastio_reg_override>=0) a=fastio_reg_override;
2948 emit_writebyte_indexed(tl,x,a);
2949 }
2950 type=STOREB_STUB;
2951 }
2952 if (dops[i].opcode==0x29) { // SH
2953 if(!c||memtarget) {
2954 int x=0,a=temp;
2955 if(!c) a=addr;
2956 if(fastio_reg_override>=0) a=fastio_reg_override;
2957 emit_writehword_indexed(tl,x,a);
2958 }
2959 type=STOREH_STUB;
2960 }
2961 if (dops[i].opcode==0x2B) { // SW
2962 if(!c||memtarget) {
2963 int a=addr;
2964 if(fastio_reg_override>=0) a=fastio_reg_override;
2965 emit_writeword_indexed(tl,0,a);
2966 }
2967 type=STOREW_STUB;
2968 }
2969 if (dops[i].opcode==0x3F) { // SD
2970 assert(0);
2971 type=STORED_STUB;
2972 }
2973 if(fastio_reg_override==HOST_TEMPREG)
2974 host_tempreg_release();
2975 if(jaddr) {
2976 // PCSX store handlers don't check invcode again
2977 reglist|=1<<addr;
2978 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2979 jaddr=0;
2980 }
2981 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
2982 if(!c||memtarget) {
2983 #ifdef DESTRUCTIVE_SHIFT
2984 // The x86 shift operation is 'destructive'; it overwrites the
2985 // source register, so we need to make a copy first and use that.
2986 addr=temp;
2987 #endif
2988 #if defined(HOST_IMM8)
2989 int ir=get_reg(i_regs->regmap,INVCP);
2990 assert(ir>=0);
2991 emit_cmpmem_indexedsr12_reg(ir,addr,1);
2992 #else
2993 emit_cmpmem_indexedsr12_imm(invalid_code,addr,1);
2994 #endif
2995 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
2996 emit_callne(invalidate_addr_reg[addr]);
2997 #else
2998 void *jaddr2 = out;
2999 emit_jne(0);
3000 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3001 #endif
3002 }
3003 }
3004 u_int addr_val=constmap[i][s]+offset;
3005 if(jaddr) {
3006 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
3007 } else if(c&&!memtarget) {
3008 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj[i],reglist);
3009 }
3010 // basic current block modification detection..
3011 // not looking back as that should be in mips cache already
3012 // (see Spyro2 title->attract mode)
3013 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3014 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3015 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3016 if(i_regs->regmap==regs[i].regmap) {
3017 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3018 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3019 emit_movimm(start+i*4+4,0);
3020 emit_writeword(0,&pcaddr);
3021 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3022 emit_far_call(get_addr_ht);
3023 emit_jmpreg(0);
3024 }
3025 }
3026}
3027
3028static void storelr_assemble(int i, const struct regstat *i_regs)
3029{
3030 int s,tl;
3031 int temp;
3032 int offset;
3033 void *jaddr=0;
3034 void *case1, *case2, *case3;
3035 void *done0, *done1, *done2;
3036 int memtarget=0,c=0;
3037 int agr=AGEN1+(i&1);
3038 u_int reglist=get_host_reglist(i_regs->regmap);
3039 tl=get_reg(i_regs->regmap,dops[i].rs2);
3040 s=get_reg(i_regs->regmap,dops[i].rs1);
3041 temp=get_reg(i_regs->regmap,agr);
3042 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3043 offset=imm[i];
3044 if(s>=0) {
3045 c=(i_regs->isconst>>s)&1;
3046 if(c) {
3047 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3048 }
3049 }
3050 assert(tl>=0);
3051 assert(temp>=0);
3052 if(!c) {
3053 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3054 if(!offset&&s!=temp) emit_mov(s,temp);
3055 jaddr=out;
3056 emit_jno(0);
3057 }
3058 else
3059 {
3060 if(!memtarget||!dops[i].rs1) {
3061 jaddr=out;
3062 emit_jmp(0);
3063 }
3064 }
3065 if(ram_offset)
3066 emit_addimm_no_flags(ram_offset,temp);
3067
3068 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3069 assert(0);
3070 }
3071
3072 emit_xorimm(temp,3,temp);
3073 emit_testimm(temp,2);
3074 case2=out;
3075 emit_jne(0);
3076 emit_testimm(temp,1);
3077 case1=out;
3078 emit_jne(0);
3079 // 0
3080 if (dops[i].opcode==0x2A) { // SWL
3081 emit_writeword_indexed(tl,0,temp);
3082 }
3083 else if (dops[i].opcode==0x2E) { // SWR
3084 emit_writebyte_indexed(tl,3,temp);
3085 }
3086 else
3087 assert(0);
3088 done0=out;
3089 emit_jmp(0);
3090 // 1
3091 set_jump_target(case1, out);
3092 if (dops[i].opcode==0x2A) { // SWL
3093 // Write 3 msb into three least significant bytes
3094 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3095 emit_writehword_indexed(tl,-1,temp);
3096 if(dops[i].rs2) emit_rorimm(tl,16,tl);
3097 emit_writebyte_indexed(tl,1,temp);
3098 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3099 }
3100 else if (dops[i].opcode==0x2E) { // SWR
3101 // Write two lsb into two most significant bytes
3102 emit_writehword_indexed(tl,1,temp);
3103 }
3104 done1=out;
3105 emit_jmp(0);
3106 // 2
3107 set_jump_target(case2, out);
3108 emit_testimm(temp,1);
3109 case3=out;
3110 emit_jne(0);
3111 if (dops[i].opcode==0x2A) { // SWL
3112 // Write two msb into two least significant bytes
3113 if(dops[i].rs2) emit_rorimm(tl,16,tl);
3114 emit_writehword_indexed(tl,-2,temp);
3115 if(dops[i].rs2) emit_rorimm(tl,16,tl);
3116 }
3117 else if (dops[i].opcode==0x2E) { // SWR
3118 // Write 3 lsb into three most significant bytes
3119 emit_writebyte_indexed(tl,-1,temp);
3120 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3121 emit_writehword_indexed(tl,0,temp);
3122 if(dops[i].rs2) emit_rorimm(tl,24,tl);
3123 }
3124 done2=out;
3125 emit_jmp(0);
3126 // 3
3127 set_jump_target(case3, out);
3128 if (dops[i].opcode==0x2A) { // SWL
3129 // Write msb into least significant byte
3130 if(dops[i].rs2) emit_rorimm(tl,24,tl);
3131 emit_writebyte_indexed(tl,-3,temp);
3132 if(dops[i].rs2) emit_rorimm(tl,8,tl);
3133 }
3134 else if (dops[i].opcode==0x2E) { // SWR
3135 // Write entire word
3136 emit_writeword_indexed(tl,-3,temp);
3137 }
3138 set_jump_target(done0, out);
3139 set_jump_target(done1, out);
3140 set_jump_target(done2, out);
3141 if(!c||!memtarget)
3142 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj[i],reglist);
3143 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3144 emit_addimm_no_flags(-ram_offset,temp);
3145 #if defined(HOST_IMM8)
3146 int ir=get_reg(i_regs->regmap,INVCP);
3147 assert(ir>=0);
3148 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3149 #else
3150 emit_cmpmem_indexedsr12_imm(invalid_code,temp,1);
3151 #endif
3152 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3153 emit_callne(invalidate_addr_reg[temp]);
3154 #else
3155 void *jaddr2 = out;
3156 emit_jne(0);
3157 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3158 #endif
3159 }
3160}
3161
3162static void cop0_assemble(int i,struct regstat *i_regs)
3163{
3164 if(dops[i].opcode2==0) // MFC0
3165 {
3166 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3167 u_int copr=(source[i]>>11)&0x1f;
3168 //assert(t>=0); // Why does this happen? OOT is weird
3169 if(t>=0&&dops[i].rt1!=0) {
3170 emit_readword(&reg_cop0[copr],t);
3171 }
3172 }
3173 else if(dops[i].opcode2==4) // MTC0
3174 {
3175 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3176 char copr=(source[i]>>11)&0x1f;
3177 assert(s>=0);
3178 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3179 if(copr==9||copr==11||copr==12||copr==13) {
3180 emit_readword(&last_count,HOST_TEMPREG);
3181 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3182 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3183 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3184 emit_writeword(HOST_CCREG,&Count);
3185 }
3186 // What a mess. The status register (12) can enable interrupts,
3187 // so needs a special case to handle a pending interrupt.
3188 // The interrupt must be taken immediately, because a subsequent
3189 // instruction might disable interrupts again.
3190 if(copr==12||copr==13) {
3191 if (is_delayslot) {
3192 // burn cycles to cause cc_interrupt, which will
3193 // reschedule next_interupt. Relies on CCREG from above.
3194 assem_debug("MTC0 DS %d\n", copr);
3195 emit_writeword(HOST_CCREG,&last_count);
3196 emit_movimm(0,HOST_CCREG);
3197 emit_storereg(CCREG,HOST_CCREG);
3198 emit_loadreg(dops[i].rs1,1);
3199 emit_movimm(copr,0);
3200 emit_far_call(pcsx_mtc0_ds);
3201 emit_loadreg(dops[i].rs1,s);
3202 return;
3203 }
3204 emit_movimm(start+i*4+4,HOST_TEMPREG);
3205 emit_writeword(HOST_TEMPREG,&pcaddr);
3206 emit_movimm(0,HOST_TEMPREG);
3207 emit_writeword(HOST_TEMPREG,&pending_exception);
3208 }
3209 if(s==HOST_CCREG)
3210 emit_loadreg(dops[i].rs1,1);
3211 else if(s!=1)
3212 emit_mov(s,1);
3213 emit_movimm(copr,0);
3214 emit_far_call(pcsx_mtc0);
3215 if(copr==9||copr==11||copr==12||copr==13) {
3216 emit_readword(&Count,HOST_CCREG);
3217 emit_readword(&next_interupt,HOST_TEMPREG);
3218 emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3219 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3220 emit_writeword(HOST_TEMPREG,&last_count);
3221 emit_storereg(CCREG,HOST_CCREG);
3222 }
3223 if(copr==12||copr==13) {
3224 assert(!is_delayslot);
3225 emit_readword(&pending_exception,14);
3226 emit_test(14,14);
3227 void *jaddr = out;
3228 emit_jeq(0);
3229 emit_readword(&pcaddr, 0);
3230 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3231 emit_far_call(get_addr_ht);
3232 emit_jmpreg(0);
3233 set_jump_target(jaddr, out);
3234 }
3235 emit_loadreg(dops[i].rs1,s);
3236 }
3237 else
3238 {
3239 assert(dops[i].opcode2==0x10);
3240 //if((source[i]&0x3f)==0x10) // RFE
3241 {
3242 emit_readword(&Status,0);
3243 emit_andimm(0,0x3c,1);
3244 emit_andimm(0,~0xf,0);
3245 emit_orrshr_imm(1,2,0);
3246 emit_writeword(0,&Status);
3247 }
3248 }
3249}
3250
3251static void cop1_unusable(int i,struct regstat *i_regs)
3252{
3253 // XXX: should just just do the exception instead
3254 //if(!cop1_usable)
3255 {
3256 void *jaddr=out;
3257 emit_jmp(0);
3258 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3259 }
3260}
3261
3262static void cop1_assemble(int i,struct regstat *i_regs)
3263{
3264 cop1_unusable(i, i_regs);
3265}
3266
3267static void c1ls_assemble(int i,struct regstat *i_regs)
3268{
3269 cop1_unusable(i, i_regs);
3270}
3271
3272// FP_STUB
3273static void do_cop1stub(int n)
3274{
3275 literal_pool(256);
3276 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3277 set_jump_target(stubs[n].addr, out);
3278 int i=stubs[n].a;
3279// int rs=stubs[n].b;
3280 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3281 int ds=stubs[n].d;
3282 if(!ds) {
3283 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3284 //if(i_regs!=&regs[i]) printf("oops: regs[i]=%x i_regs=%x",(int)&regs[i],(int)i_regs);
3285 }
3286 //else {printf("fp exception in delay slot\n");}
3287 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3288 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3289 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3290 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3291 emit_far_jump(ds?fp_exception_ds:fp_exception);
3292}
3293
3294static int cop2_is_stalling_op(int i, int *cycles)
3295{
3296 if (dops[i].opcode == 0x3a) { // SWC2
3297 *cycles = 0;
3298 return 1;
3299 }
3300 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3301 *cycles = 0;
3302 return 1;
3303 }
3304 if (dops[i].itype == C2OP) {
3305 *cycles = gte_cycletab[source[i] & 0x3f];
3306 return 1;
3307 }
3308 // ... what about MTC2/CTC2/LWC2?
3309 return 0;
3310}
3311
3312#if 0
3313static void log_gte_stall(int stall, u_int cycle)
3314{
3315 if ((u_int)stall <= 44)
3316 printf("x stall %2d %u\n", stall, cycle + last_count);
3317}
3318
3319static void emit_log_gte_stall(int i, int stall, u_int reglist)
3320{
3321 save_regs(reglist);
3322 if (stall > 0)
3323 emit_movimm(stall, 0);
3324 else
3325 emit_mov(HOST_TEMPREG, 0);
3326 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
3327 emit_far_call(log_gte_stall);
3328 restore_regs(reglist);
3329}
3330#endif
3331
3332static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3333{
3334 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3335 int rtmp = reglist_find_free(reglist);
3336
3337 if (HACK_ENABLED(NDHACK_NO_STALLS))
3338 return;
3339 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3340 // happens occasionally... cc evicted? Don't bother then
3341 //printf("no cc %08x\n", start + i*4);
3342 return;
3343 }
3344 if (!dops[i].bt) {
3345 for (j = i - 1; j >= 0; j--) {
3346 //if (dops[j].is_ds) break;
3347 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3348 break;
3349 }
3350 j = max(j, 0);
3351 }
3352 cycles_passed = CLOCK_ADJUST(ccadj[i] - ccadj[j]);
3353 if (other_gte_op_cycles >= 0)
3354 stall = other_gte_op_cycles - cycles_passed;
3355 else if (cycles_passed >= 44)
3356 stall = 0; // can't stall
3357 if (stall == -MAXBLOCK && rtmp >= 0) {
3358 // unknown stall, do the expensive runtime check
3359 assem_debug("; cop2_do_stall_check\n");
3360#if 0 // too slow
3361 save_regs(reglist);
3362 emit_movimm(gte_cycletab[op], 0);
3363 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]), 1);
3364 emit_far_call(call_gteStall);
3365 restore_regs(reglist);
3366#else
3367 host_tempreg_acquire();
3368 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3369 emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
3370 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3371 emit_cmpimm(HOST_TEMPREG, 44);
3372 emit_cmovb_reg(rtmp, HOST_CCREG);
3373 //emit_log_gte_stall(i, 0, reglist);
3374 host_tempreg_release();
3375#endif
3376 }
3377 else if (stall > 0) {
3378 //emit_log_gte_stall(i, stall, reglist);
3379 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3380 }
3381
3382 // save gteBusyCycle, if needed
3383 if (gte_cycletab[op] == 0)
3384 return;
3385 other_gte_op_cycles = -1;
3386 for (j = i + 1; j < slen; j++) {
3387 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3388 break;
3389 if (dops[j].is_jump) {
3390 // check ds
3391 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3392 j++;
3393 break;
3394 }
3395 }
3396 if (other_gte_op_cycles >= 0)
3397 // will handle stall when assembling that op
3398 return;
3399 cycles_passed = CLOCK_ADJUST(ccadj[min(j, slen -1)] - ccadj[i]);
3400 if (cycles_passed >= 44)
3401 return;
3402 assem_debug("; save gteBusyCycle\n");
3403 host_tempreg_acquire();
3404#if 0
3405 emit_readword(&last_count, HOST_TEMPREG);
3406 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3407 emit_addimm(HOST_TEMPREG, CLOCK_ADJUST(ccadj[i]), HOST_TEMPREG);
3408 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3409 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3410#else
3411 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + gte_cycletab[op], HOST_TEMPREG);
3412 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3413#endif
3414 host_tempreg_release();
3415}
3416
3417static int is_mflohi(int i)
3418{
3419 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3420}
3421
3422static int check_multdiv(int i, int *cycles)
3423{
3424 if (dops[i].itype != MULTDIV)
3425 return 0;
3426 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3427 *cycles = 11; // approx from 7 11 14
3428 else
3429 *cycles = 37;
3430 return 1;
3431}
3432
3433static void multdiv_prepare_stall(int i, const struct regstat *i_regs)
3434{
3435 int j, found = 0, c = 0;
3436 if (HACK_ENABLED(NDHACK_NO_STALLS))
3437 return;
3438 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3439 // happens occasionally... cc evicted? Don't bother then
3440 return;
3441 }
3442 for (j = i + 1; j < slen; j++) {
3443 if (dops[j].bt)
3444 break;
3445 if ((found = is_mflohi(j)))
3446 break;
3447 if (dops[j].is_jump) {
3448 // check ds
3449 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3450 j++;
3451 break;
3452 }
3453 }
3454 if (found)
3455 // handle all in multdiv_do_stall()
3456 return;
3457 check_multdiv(i, &c);
3458 assert(c > 0);
3459 assem_debug("; muldiv prepare stall %d\n", c);
3460 host_tempreg_acquire();
3461 emit_addimm(HOST_CCREG, CLOCK_ADJUST(ccadj[i]) + c, HOST_TEMPREG);
3462 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3463 host_tempreg_release();
3464}
3465
3466static void multdiv_do_stall(int i, const struct regstat *i_regs)
3467{
3468 int j, known_cycles = 0;
3469 u_int reglist = get_host_reglist(i_regs->regmap);
3470 int rtmp = get_reg(i_regs->regmap, -1);
3471 if (rtmp < 0)
3472 rtmp = reglist_find_free(reglist);
3473 if (HACK_ENABLED(NDHACK_NO_STALLS))
3474 return;
3475 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3476 // happens occasionally... cc evicted? Don't bother then
3477 //printf("no cc/rtmp %08x\n", start + i*4);
3478 return;
3479 }
3480 if (!dops[i].bt) {
3481 for (j = i - 1; j >= 0; j--) {
3482 if (dops[j].is_ds) break;
3483 if (check_multdiv(j, &known_cycles) || dops[j].bt)
3484 break;
3485 if (is_mflohi(j))
3486 // already handled by this op
3487 return;
3488 }
3489 j = max(j, 0);
3490 }
3491 if (known_cycles > 0) {
3492 known_cycles -= CLOCK_ADJUST(ccadj[i] - ccadj[j]);
3493 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3494 if (known_cycles > 0)
3495 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3496 return;
3497 }
3498 assem_debug("; muldiv stall unresolved\n");
3499 host_tempreg_acquire();
3500 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3501 emit_addimm(rtmp, -CLOCK_ADJUST(ccadj[i]), rtmp);
3502 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3503 emit_cmpimm(HOST_TEMPREG, 37);
3504 emit_cmovb_reg(rtmp, HOST_CCREG);
3505 //emit_log_gte_stall(i, 0, reglist);
3506 host_tempreg_release();
3507}
3508
3509static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3510{
3511 switch (copr) {
3512 case 1:
3513 case 3:
3514 case 5:
3515 case 8:
3516 case 9:
3517 case 10:
3518 case 11:
3519 emit_readword(&reg_cop2d[copr],tl);
3520 emit_signextend16(tl,tl);
3521 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3522 break;
3523 case 7:
3524 case 16:
3525 case 17:
3526 case 18:
3527 case 19:
3528 emit_readword(&reg_cop2d[copr],tl);
3529 emit_andimm(tl,0xffff,tl);
3530 emit_writeword(tl,&reg_cop2d[copr]);
3531 break;
3532 case 15:
3533 emit_readword(&reg_cop2d[14],tl); // SXY2
3534 emit_writeword(tl,&reg_cop2d[copr]);
3535 break;
3536 case 28:
3537 case 29:
3538 c2op_mfc2_29_assemble(tl,temp);
3539 break;
3540 default:
3541 emit_readword(&reg_cop2d[copr],tl);
3542 break;
3543 }
3544}
3545
3546static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3547{
3548 switch (copr) {
3549 case 15:
3550 emit_readword(&reg_cop2d[13],temp); // SXY1
3551 emit_writeword(sl,&reg_cop2d[copr]);
3552 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3553 emit_readword(&reg_cop2d[14],temp); // SXY2
3554 emit_writeword(sl,&reg_cop2d[14]);
3555 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3556 break;
3557 case 28:
3558 emit_andimm(sl,0x001f,temp);
3559 emit_shlimm(temp,7,temp);
3560 emit_writeword(temp,&reg_cop2d[9]);
3561 emit_andimm(sl,0x03e0,temp);
3562 emit_shlimm(temp,2,temp);
3563 emit_writeword(temp,&reg_cop2d[10]);
3564 emit_andimm(sl,0x7c00,temp);
3565 emit_shrimm(temp,3,temp);
3566 emit_writeword(temp,&reg_cop2d[11]);
3567 emit_writeword(sl,&reg_cop2d[28]);
3568 break;
3569 case 30:
3570 emit_xorsar_imm(sl,sl,31,temp);
3571#if defined(HAVE_ARMV5) || defined(__aarch64__)
3572 emit_clz(temp,temp);
3573#else
3574 emit_movs(temp,HOST_TEMPREG);
3575 emit_movimm(0,temp);
3576 emit_jeq((int)out+4*4);
3577 emit_addpl_imm(temp,1,temp);
3578 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3579 emit_jns((int)out-2*4);
3580#endif
3581 emit_writeword(sl,&reg_cop2d[30]);
3582 emit_writeword(temp,&reg_cop2d[31]);
3583 break;
3584 case 31:
3585 break;
3586 default:
3587 emit_writeword(sl,&reg_cop2d[copr]);
3588 break;
3589 }
3590}
3591
3592static void c2ls_assemble(int i, const struct regstat *i_regs)
3593{
3594 int s,tl;
3595 int ar;
3596 int offset;
3597 int memtarget=0,c=0;
3598 void *jaddr2=NULL;
3599 enum stub_type type;
3600 int agr=AGEN1+(i&1);
3601 int fastio_reg_override=-1;
3602 u_int reglist=get_host_reglist(i_regs->regmap);
3603 u_int copr=(source[i]>>16)&0x1f;
3604 s=get_reg(i_regs->regmap,dops[i].rs1);
3605 tl=get_reg(i_regs->regmap,FTEMP);
3606 offset=imm[i];
3607 assert(dops[i].rs1>0);
3608 assert(tl>=0);
3609
3610 if(i_regs->regmap[HOST_CCREG]==CCREG)
3611 reglist&=~(1<<HOST_CCREG);
3612
3613 // get the address
3614 if (dops[i].opcode==0x3a) { // SWC2
3615 ar=get_reg(i_regs->regmap,agr);
3616 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3617 reglist|=1<<ar;
3618 } else { // LWC2
3619 ar=tl;
3620 }
3621 if(s>=0) c=(i_regs->wasconst>>s)&1;
3622 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3623 if (!offset&&!c&&s>=0) ar=s;
3624 assert(ar>=0);
3625
3626 cop2_do_stall_check(0, i, i_regs, reglist);
3627
3628 if (dops[i].opcode==0x3a) { // SWC2
3629 cop2_get_dreg(copr,tl,-1);
3630 type=STOREW_STUB;
3631 }
3632 else
3633 type=LOADW_STUB;
3634
3635 if(c&&!memtarget) {
3636 jaddr2=out;
3637 emit_jmp(0); // inline_readstub/inline_writestub?
3638 }
3639 else {
3640 if(!c) {
3641 jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
3642 }
3643 else if(ram_offset&&memtarget) {
3644 host_tempreg_acquire();
3645 emit_addimm(ar,ram_offset,HOST_TEMPREG);
3646 fastio_reg_override=HOST_TEMPREG;
3647 }
3648 if (dops[i].opcode==0x32) { // LWC2
3649 int a=ar;
3650 if(fastio_reg_override>=0) a=fastio_reg_override;
3651 emit_readword_indexed(0,a,tl);
3652 }
3653 if (dops[i].opcode==0x3a) { // SWC2
3654 #ifdef DESTRUCTIVE_SHIFT
3655 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3656 #endif
3657 int a=ar;
3658 if(fastio_reg_override>=0) a=fastio_reg_override;
3659 emit_writeword_indexed(tl,0,a);
3660 }
3661 }
3662 if(fastio_reg_override==HOST_TEMPREG)
3663 host_tempreg_release();
3664 if(jaddr2)
3665 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj[i],reglist);
3666 if(dops[i].opcode==0x3a) // SWC2
3667 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3668#if defined(HOST_IMM8)
3669 int ir=get_reg(i_regs->regmap,INVCP);
3670 assert(ir>=0);
3671 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3672#else
3673 emit_cmpmem_indexedsr12_imm(invalid_code,ar,1);
3674#endif
3675 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3676 emit_callne(invalidate_addr_reg[ar]);
3677 #else
3678 void *jaddr3 = out;
3679 emit_jne(0);
3680 add_stub(INVCODE_STUB,jaddr3,out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3681 #endif
3682 }
3683 if (dops[i].opcode==0x32) { // LWC2
3684 host_tempreg_acquire();
3685 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3686 host_tempreg_release();
3687 }
3688}
3689
3690static void cop2_assemble(int i, const struct regstat *i_regs)
3691{
3692 u_int copr = (source[i]>>11) & 0x1f;
3693 signed char temp = get_reg(i_regs->regmap, -1);
3694
3695 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3696 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3697 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3698 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3699 reglist = reglist_exclude(reglist, tl, -1);
3700 }
3701 cop2_do_stall_check(0, i, i_regs, reglist);
3702 }
3703 if (dops[i].opcode2==0) { // MFC2
3704 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3705 if(tl>=0&&dops[i].rt1!=0)
3706 cop2_get_dreg(copr,tl,temp);
3707 }
3708 else if (dops[i].opcode2==4) { // MTC2
3709 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3710 cop2_put_dreg(copr,sl,temp);
3711 }
3712 else if (dops[i].opcode2==2) // CFC2
3713 {
3714 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3715 if(tl>=0&&dops[i].rt1!=0)
3716 emit_readword(&reg_cop2c[copr],tl);
3717 }
3718 else if (dops[i].opcode2==6) // CTC2
3719 {
3720 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3721 switch(copr) {
3722 case 4:
3723 case 12:
3724 case 20:
3725 case 26:
3726 case 27:
3727 case 29:
3728 case 30:
3729 emit_signextend16(sl,temp);
3730 break;
3731 case 31:
3732 c2op_ctc2_31_assemble(sl,temp);
3733 break;
3734 default:
3735 temp=sl;
3736 break;
3737 }
3738 emit_writeword(temp,&reg_cop2c[copr]);
3739 assert(sl>=0);
3740 }
3741}
3742
3743static void do_unalignedwritestub(int n)
3744{
3745 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
3746 literal_pool(256);
3747 set_jump_target(stubs[n].addr, out);
3748
3749 int i=stubs[n].a;
3750 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3751 int addr=stubs[n].b;
3752 u_int reglist=stubs[n].e;
3753 signed char *i_regmap=i_regs->regmap;
3754 int temp2=get_reg(i_regmap,FTEMP);
3755 int rt;
3756 rt=get_reg(i_regmap,dops[i].rs2);
3757 assert(rt>=0);
3758 assert(addr>=0);
3759 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
3760 reglist|=(1<<addr);
3761 reglist&=~(1<<temp2);
3762
3763 // don't bother with it and call write handler
3764 save_regs(reglist);
3765 pass_args(addr,rt);
3766 int cc=get_reg(i_regmap,CCREG);
3767 if(cc<0)
3768 emit_loadreg(CCREG,2);
3769 emit_addimm(cc<0?2:cc,CLOCK_ADJUST((int)stubs[n].d+1),2);
3770 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
3771 emit_addimm(0,-CLOCK_ADJUST((int)stubs[n].d+1),cc<0?2:cc);
3772 if(cc<0)
3773 emit_storereg(CCREG,2);
3774 restore_regs(reglist);
3775 emit_jmp(stubs[n].retaddr); // return address
3776}
3777
3778#ifndef multdiv_assemble
3779void multdiv_assemble(int i,struct regstat *i_regs)
3780{
3781 printf("Need multdiv_assemble for this architecture.\n");
3782 abort();
3783}
3784#endif
3785
3786static void mov_assemble(int i,struct regstat *i_regs)
3787{
3788 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
3789 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
3790 if(dops[i].rt1) {
3791 signed char sl,tl;
3792 tl=get_reg(i_regs->regmap,dops[i].rt1);
3793 //assert(tl>=0);
3794 if(tl>=0) {
3795 sl=get_reg(i_regs->regmap,dops[i].rs1);
3796 if(sl>=0) emit_mov(sl,tl);
3797 else emit_loadreg(dops[i].rs1,tl);
3798 }
3799 }
3800 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
3801 multdiv_do_stall(i, i_regs);
3802}
3803
3804// call interpreter, exception handler, things that change pc/regs/cycles ...
3805static void call_c_cpu_handler(int i, const struct regstat *i_regs, u_int pc, void *func)
3806{
3807 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3808 assert(ccreg==HOST_CCREG);
3809 assert(!is_delayslot);
3810 (void)ccreg;
3811
3812 emit_movimm(pc,3); // Get PC
3813 emit_readword(&last_count,2);
3814 emit_writeword(3,&psxRegs.pc);
3815 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
3816 emit_add(2,HOST_CCREG,2);
3817 emit_writeword(2,&psxRegs.cycle);
3818 emit_far_call(func);
3819 emit_far_jump(jump_to_new_pc);
3820}
3821
3822static void syscall_assemble(int i,struct regstat *i_regs)
3823{
3824 emit_movimm(0x20,0); // cause code
3825 emit_movimm(0,1); // not in delay slot
3826 call_c_cpu_handler(i,i_regs,start+i*4,psxException);
3827}
3828
3829static void hlecall_assemble(int i,struct regstat *i_regs)
3830{
3831 void *hlefunc = psxNULL;
3832 uint32_t hleCode = source[i] & 0x03ffffff;
3833 if (hleCode < ARRAY_SIZE(psxHLEt))
3834 hlefunc = psxHLEt[hleCode];
3835
3836 call_c_cpu_handler(i,i_regs,start+i*4+4,hlefunc);
3837}
3838
3839static void intcall_assemble(int i,struct regstat *i_regs)
3840{
3841 call_c_cpu_handler(i,i_regs,start+i*4,execI);
3842}
3843
3844static void speculate_mov(int rs,int rt)
3845{
3846 if(rt!=0) {
3847 smrv_strong_next|=1<<rt;
3848 smrv[rt]=smrv[rs];
3849 }
3850}
3851
3852static void speculate_mov_weak(int rs,int rt)
3853{
3854 if(rt!=0) {
3855 smrv_weak_next|=1<<rt;
3856 smrv[rt]=smrv[rs];
3857 }
3858}
3859
3860static void speculate_register_values(int i)
3861{
3862 if(i==0) {
3863 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
3864 // gp,sp are likely to stay the same throughout the block
3865 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
3866 smrv_weak_next=~smrv_strong_next;
3867 //printf(" llr %08x\n", smrv[4]);
3868 }
3869 smrv_strong=smrv_strong_next;
3870 smrv_weak=smrv_weak_next;
3871 switch(dops[i].itype) {
3872 case ALU:
3873 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
3874 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
3875 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
3876 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
3877 else {
3878 smrv_strong_next&=~(1<<dops[i].rt1);
3879 smrv_weak_next&=~(1<<dops[i].rt1);
3880 }
3881 break;
3882 case SHIFTIMM:
3883 smrv_strong_next&=~(1<<dops[i].rt1);
3884 smrv_weak_next&=~(1<<dops[i].rt1);
3885 // fallthrough
3886 case IMM16:
3887 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
3888 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
3889 if(hr>=0) {
3890 if(get_final_value(hr,i,&value))
3891 smrv[dops[i].rt1]=value;
3892 else smrv[dops[i].rt1]=constmap[i][hr];
3893 smrv_strong_next|=1<<dops[i].rt1;
3894 }
3895 }
3896 else {
3897 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
3898 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
3899 }
3900 break;
3901 case LOAD:
3902 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
3903 // special case for BIOS
3904 smrv[dops[i].rt1]=0xa0000000;
3905 smrv_strong_next|=1<<dops[i].rt1;
3906 break;
3907 }
3908 // fallthrough
3909 case SHIFT:
3910 case LOADLR:
3911 case MOV:
3912 smrv_strong_next&=~(1<<dops[i].rt1);
3913 smrv_weak_next&=~(1<<dops[i].rt1);
3914 break;
3915 case COP0:
3916 case COP2:
3917 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
3918 smrv_strong_next&=~(1<<dops[i].rt1);
3919 smrv_weak_next&=~(1<<dops[i].rt1);
3920 }
3921 break;
3922 case C2LS:
3923 if (dops[i].opcode==0x32) { // LWC2
3924 smrv_strong_next&=~(1<<dops[i].rt1);
3925 smrv_weak_next&=~(1<<dops[i].rt1);
3926 }
3927 break;
3928 }
3929#if 0
3930 int r=4;
3931 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
3932 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
3933#endif
3934}
3935
3936static void ds_assemble(int i,struct regstat *i_regs)
3937{
3938 speculate_register_values(i);
3939 is_delayslot=1;
3940 switch(dops[i].itype) {
3941 case ALU:
3942 alu_assemble(i,i_regs);break;
3943 case IMM16:
3944 imm16_assemble(i,i_regs);break;
3945 case SHIFT:
3946 shift_assemble(i,i_regs);break;
3947 case SHIFTIMM:
3948 shiftimm_assemble(i,i_regs);break;
3949 case LOAD:
3950 load_assemble(i,i_regs);break;
3951 case LOADLR:
3952 loadlr_assemble(i,i_regs);break;
3953 case STORE:
3954 store_assemble(i,i_regs);break;
3955 case STORELR:
3956 storelr_assemble(i,i_regs);break;
3957 case COP0:
3958 cop0_assemble(i,i_regs);break;
3959 case COP1:
3960 cop1_assemble(i,i_regs);break;
3961 case C1LS:
3962 c1ls_assemble(i,i_regs);break;
3963 case COP2:
3964 cop2_assemble(i,i_regs);break;
3965 case C2LS:
3966 c2ls_assemble(i,i_regs);break;
3967 case C2OP:
3968 c2op_assemble(i,i_regs);break;
3969 case MULTDIV:
3970 multdiv_assemble(i,i_regs);
3971 multdiv_prepare_stall(i,i_regs);
3972 break;
3973 case MOV:
3974 mov_assemble(i,i_regs);break;
3975 case SYSCALL:
3976 case HLECALL:
3977 case INTCALL:
3978 case SPAN:
3979 case UJUMP:
3980 case RJUMP:
3981 case CJUMP:
3982 case SJUMP:
3983 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
3984 }
3985 is_delayslot=0;
3986}
3987
3988// Is the branch target a valid internal jump?
3989static int internal_branch(int addr)
3990{
3991 if(addr&1) return 0; // Indirect (register) jump
3992 if(addr>=start && addr<start+slen*4-4)
3993 {
3994 return 1;
3995 }
3996 return 0;
3997}
3998
3999static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4000{
4001 int hr;
4002 for(hr=0;hr<HOST_REGS;hr++) {
4003 if(hr!=EXCLUDE_REG) {
4004 if(pre[hr]!=entry[hr]) {
4005 if(pre[hr]>=0) {
4006 if((dirty>>hr)&1) {
4007 if(get_reg(entry,pre[hr])<0) {
4008 assert(pre[hr]<64);
4009 if(!((u>>pre[hr])&1))
4010 emit_storereg(pre[hr],hr);
4011 }
4012 }
4013 }
4014 }
4015 }
4016 }
4017 // Move from one register to another (no writeback)
4018 for(hr=0;hr<HOST_REGS;hr++) {
4019 if(hr!=EXCLUDE_REG) {
4020 if(pre[hr]!=entry[hr]) {
4021 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
4022 int nr;
4023 if((nr=get_reg(entry,pre[hr]))>=0) {
4024 emit_mov(hr,nr);
4025 }
4026 }
4027 }
4028 }
4029 }
4030}
4031
4032// Load the specified registers
4033// This only loads the registers given as arguments because
4034// we don't want to load things that will be overwritten
4035static void load_regs(signed char entry[],signed char regmap[],int rs1,int rs2)
4036{
4037 int hr;
4038 // Load 32-bit regs
4039 for(hr=0;hr<HOST_REGS;hr++) {
4040 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4041 if(entry[hr]!=regmap[hr]) {
4042 if(regmap[hr]==rs1||regmap[hr]==rs2)
4043 {
4044 if(regmap[hr]==0) {
4045 emit_zeroreg(hr);
4046 }
4047 else
4048 {
4049 emit_loadreg(regmap[hr],hr);
4050 }
4051 }
4052 }
4053 }
4054 }
4055}
4056
4057// Load registers prior to the start of a loop
4058// so that they are not loaded within the loop
4059static void loop_preload(signed char pre[],signed char entry[])
4060{
4061 int hr;
4062 for(hr=0;hr<HOST_REGS;hr++) {
4063 if(hr!=EXCLUDE_REG) {
4064 if(pre[hr]!=entry[hr]) {
4065 if(entry[hr]>=0) {
4066 if(get_reg(pre,entry[hr])<0) {
4067 assem_debug("loop preload:\n");
4068 //printf("loop preload: %d\n",hr);
4069 if(entry[hr]==0) {
4070 emit_zeroreg(hr);
4071 }
4072 else if(entry[hr]<TEMPREG)
4073 {
4074 emit_loadreg(entry[hr],hr);
4075 }
4076 else if(entry[hr]-64<TEMPREG)
4077 {
4078 emit_loadreg(entry[hr],hr);
4079 }
4080 }
4081 }
4082 }
4083 }
4084 }
4085}
4086
4087// Generate address for load/store instruction
4088// goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4089void address_generation(int i,struct regstat *i_regs,signed char entry[])
4090{
4091 if(dops[i].itype==LOAD||dops[i].itype==LOADLR||dops[i].itype==STORE||dops[i].itype==STORELR||dops[i].itype==C1LS||dops[i].itype==C2LS) {
4092 int ra=-1;
4093 int agr=AGEN1+(i&1);
4094 if(dops[i].itype==LOAD) {
4095 ra=get_reg(i_regs->regmap,dops[i].rt1);
4096 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4097 assert(ra>=0);
4098 }
4099 if(dops[i].itype==LOADLR) {
4100 ra=get_reg(i_regs->regmap,FTEMP);
4101 }
4102 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4103 ra=get_reg(i_regs->regmap,agr);
4104 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4105 }
4106 if(dops[i].itype==C1LS||dops[i].itype==C2LS) {
4107 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4108 ra=get_reg(i_regs->regmap,FTEMP);
4109 else { // SWC1/SDC1/SWC2/SDC2
4110 ra=get_reg(i_regs->regmap,agr);
4111 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4112 }
4113 }
4114 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4115 if(ra>=0) {
4116 int offset=imm[i];
4117 int c=(i_regs->wasconst>>rs)&1;
4118 if(dops[i].rs1==0) {
4119 // Using r0 as a base address
4120 if(!entry||entry[ra]!=agr) {
4121 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4122 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4123 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4124 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4125 }else{
4126 emit_movimm(offset,ra);
4127 }
4128 } // else did it in the previous cycle
4129 }
4130 else if(rs<0) {
4131 if(!entry||entry[ra]!=dops[i].rs1)
4132 emit_loadreg(dops[i].rs1,ra);
4133 //if(!entry||entry[ra]!=dops[i].rs1)
4134 // printf("poor load scheduling!\n");
4135 }
4136 else if(c) {
4137 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4138 if(!entry||entry[ra]!=agr) {
4139 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4140 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4141 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4142 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4143 }else{
4144 emit_movimm(constmap[i][rs]+offset,ra);
4145 regs[i].loadedconst|=1<<ra;
4146 }
4147 } // else did it in the previous cycle
4148 } // else load_consts already did it
4149 }
4150 if(offset&&!c&&dops[i].rs1) {
4151 if(rs>=0) {
4152 emit_addimm(rs,offset,ra);
4153 }else{
4154 emit_addimm(ra,offset,ra);
4155 }
4156 }
4157 }
4158 }
4159 // Preload constants for next instruction
4160 if(dops[i+1].itype==LOAD||dops[i+1].itype==LOADLR||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS||dops[i+1].itype==C2LS) {
4161 int agr,ra;
4162 // Actual address
4163 agr=AGEN1+((i+1)&1);
4164 ra=get_reg(i_regs->regmap,agr);
4165 if(ra>=0) {
4166 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4167 int offset=imm[i+1];
4168 int c=(regs[i+1].wasconst>>rs)&1;
4169 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4170 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4171 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4172 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4173 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4174 }else{
4175 emit_movimm(constmap[i+1][rs]+offset,ra);
4176 regs[i+1].loadedconst|=1<<ra;
4177 }
4178 }
4179 else if(dops[i+1].rs1==0) {
4180 // Using r0 as a base address
4181 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4182 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4183 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4184 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4185 }else{
4186 emit_movimm(offset,ra);
4187 }
4188 }
4189 }
4190 }
4191}
4192
4193static int get_final_value(int hr, int i, int *value)
4194{
4195 int reg=regs[i].regmap[hr];
4196 while(i<slen-1) {
4197 if(regs[i+1].regmap[hr]!=reg) break;
4198 if(!((regs[i+1].isconst>>hr)&1)) break;
4199 if(dops[i+1].bt) break;
4200 i++;
4201 }
4202 if(i<slen-1) {
4203 if (dops[i].is_jump) {
4204 *value=constmap[i][hr];
4205 return 1;
4206 }
4207 if(!dops[i+1].bt) {
4208 if (dops[i+1].is_jump) {
4209 // Load in delay slot, out-of-order execution
4210 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4211 {
4212 // Precompute load address
4213 *value=constmap[i][hr]+imm[i+2];
4214 return 1;
4215 }
4216 }
4217 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4218 {
4219 // Precompute load address
4220 *value=constmap[i][hr]+imm[i+1];
4221 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4222 return 1;
4223 }
4224 }
4225 }
4226 *value=constmap[i][hr];
4227 //printf("c=%lx\n",(long)constmap[i][hr]);
4228 if(i==slen-1) return 1;
4229 assert(reg < 64);
4230 return !((unneeded_reg[i+1]>>reg)&1);
4231}
4232
4233// Load registers with known constants
4234static void load_consts(signed char pre[],signed char regmap[],int i)
4235{
4236 int hr,hr2;
4237 // propagate loaded constant flags
4238 if(i==0||dops[i].bt)
4239 regs[i].loadedconst=0;
4240 else {
4241 for(hr=0;hr<HOST_REGS;hr++) {
4242 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4243 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4244 {
4245 regs[i].loadedconst|=1<<hr;
4246 }
4247 }
4248 }
4249 // Load 32-bit regs
4250 for(hr=0;hr<HOST_REGS;hr++) {
4251 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4252 //if(entry[hr]!=regmap[hr]) {
4253 if(!((regs[i].loadedconst>>hr)&1)) {
4254 assert(regmap[hr]<64);
4255 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4256 int value,similar=0;
4257 if(get_final_value(hr,i,&value)) {
4258 // see if some other register has similar value
4259 for(hr2=0;hr2<HOST_REGS;hr2++) {
4260 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4261 if(is_similar_value(value,constmap[i][hr2])) {
4262 similar=1;
4263 break;
4264 }
4265 }
4266 }
4267 if(similar) {
4268 int value2;
4269 if(get_final_value(hr2,i,&value2)) // is this needed?
4270 emit_movimm_from(value2,hr2,value,hr);
4271 else
4272 emit_movimm(value,hr);
4273 }
4274 else if(value==0) {
4275 emit_zeroreg(hr);
4276 }
4277 else {
4278 emit_movimm(value,hr);
4279 }
4280 }
4281 regs[i].loadedconst|=1<<hr;
4282 }
4283 }
4284 }
4285 }
4286}
4287
4288void load_all_consts(signed char regmap[], u_int dirty, int i)
4289{
4290 int hr;
4291 // Load 32-bit regs
4292 for(hr=0;hr<HOST_REGS;hr++) {
4293 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4294 assert(regmap[hr] < 64);
4295 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4296 int value=constmap[i][hr];
4297 if(value==0) {
4298 emit_zeroreg(hr);
4299 }
4300 else {
4301 emit_movimm(value,hr);
4302 }
4303 }
4304 }
4305 }
4306}
4307
4308// Write out all dirty registers (except cycle count)
4309static void wb_dirtys(signed char i_regmap[],uint64_t i_dirty)
4310{
4311 int hr;
4312 for(hr=0;hr<HOST_REGS;hr++) {
4313 if(hr!=EXCLUDE_REG) {
4314 if(i_regmap[hr]>0) {
4315 if(i_regmap[hr]!=CCREG) {
4316 if((i_dirty>>hr)&1) {
4317 assert(i_regmap[hr]<64);
4318 emit_storereg(i_regmap[hr],hr);
4319 }
4320 }
4321 }
4322 }
4323 }
4324}
4325
4326// Write out dirty registers that we need to reload (pair with load_needed_regs)
4327// This writes the registers not written by store_regs_bt
4328void wb_needed_dirtys(signed char i_regmap[],uint64_t i_dirty,int addr)
4329{
4330 int hr;
4331 int t=(addr-start)>>2;
4332 for(hr=0;hr<HOST_REGS;hr++) {
4333 if(hr!=EXCLUDE_REG) {
4334 if(i_regmap[hr]>0) {
4335 if(i_regmap[hr]!=CCREG) {
4336 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4337 if((i_dirty>>hr)&1) {
4338 assert(i_regmap[hr]<64);
4339 emit_storereg(i_regmap[hr],hr);
4340 }
4341 }
4342 }
4343 }
4344 }
4345 }
4346}
4347
4348// Load all registers (except cycle count)
4349void load_all_regs(signed char i_regmap[])
4350{
4351 int hr;
4352 for(hr=0;hr<HOST_REGS;hr++) {
4353 if(hr!=EXCLUDE_REG) {
4354 if(i_regmap[hr]==0) {
4355 emit_zeroreg(hr);
4356 }
4357 else
4358 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4359 {
4360 emit_loadreg(i_regmap[hr],hr);
4361 }
4362 }
4363 }
4364}
4365
4366// Load all current registers also needed by next instruction
4367void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
4368{
4369 int hr;
4370 for(hr=0;hr<HOST_REGS;hr++) {
4371 if(hr!=EXCLUDE_REG) {
4372 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4373 if(i_regmap[hr]==0) {
4374 emit_zeroreg(hr);
4375 }
4376 else
4377 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4378 {
4379 emit_loadreg(i_regmap[hr],hr);
4380 }
4381 }
4382 }
4383 }
4384}
4385
4386// Load all regs, storing cycle count if necessary
4387void load_regs_entry(int t)
4388{
4389 int hr;
4390 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4391 else if(ccadj[t]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[t]),HOST_CCREG);
4392 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4393 emit_storereg(CCREG,HOST_CCREG);
4394 }
4395 // Load 32-bit regs
4396 for(hr=0;hr<HOST_REGS;hr++) {
4397 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4398 if(regs[t].regmap_entry[hr]==0) {
4399 emit_zeroreg(hr);
4400 }
4401 else if(regs[t].regmap_entry[hr]!=CCREG)
4402 {
4403 emit_loadreg(regs[t].regmap_entry[hr],hr);
4404 }
4405 }
4406 }
4407}
4408
4409// Store dirty registers prior to branch
4410void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4411{
4412 if(internal_branch(addr))
4413 {
4414 int t=(addr-start)>>2;
4415 int hr;
4416 for(hr=0;hr<HOST_REGS;hr++) {
4417 if(hr!=EXCLUDE_REG) {
4418 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4419 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4420 if((i_dirty>>hr)&1) {
4421 assert(i_regmap[hr]<64);
4422 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4423 emit_storereg(i_regmap[hr],hr);
4424 }
4425 }
4426 }
4427 }
4428 }
4429 }
4430 else
4431 {
4432 // Branch out of this block, write out all dirty regs
4433 wb_dirtys(i_regmap,i_dirty);
4434 }
4435}
4436
4437// Load all needed registers for branch target
4438static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4439{
4440 //if(addr>=start && addr<(start+slen*4))
4441 if(internal_branch(addr))
4442 {
4443 int t=(addr-start)>>2;
4444 int hr;
4445 // Store the cycle count before loading something else
4446 if(i_regmap[HOST_CCREG]!=CCREG) {
4447 assert(i_regmap[HOST_CCREG]==-1);
4448 }
4449 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4450 emit_storereg(CCREG,HOST_CCREG);
4451 }
4452 // Load 32-bit regs
4453 for(hr=0;hr<HOST_REGS;hr++) {
4454 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4455 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4456 if(regs[t].regmap_entry[hr]==0) {
4457 emit_zeroreg(hr);
4458 }
4459 else if(regs[t].regmap_entry[hr]!=CCREG)
4460 {
4461 emit_loadreg(regs[t].regmap_entry[hr],hr);
4462 }
4463 }
4464 }
4465 }
4466 }
4467}
4468
4469static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4470{
4471 if(addr>=start && addr<start+slen*4-4)
4472 {
4473 int t=(addr-start)>>2;
4474 int hr;
4475 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4476 for(hr=0;hr<HOST_REGS;hr++)
4477 {
4478 if(hr!=EXCLUDE_REG)
4479 {
4480 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4481 {
4482 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4483 {
4484 return 0;
4485 }
4486 else
4487 if((i_dirty>>hr)&1)
4488 {
4489 if(i_regmap[hr]<TEMPREG)
4490 {
4491 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4492 return 0;
4493 }
4494 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4495 {
4496 assert(0);
4497 }
4498 }
4499 }
4500 else // Same register but is it 32-bit or dirty?
4501 if(i_regmap[hr]>=0)
4502 {
4503 if(!((regs[t].dirty>>hr)&1))
4504 {
4505 if((i_dirty>>hr)&1)
4506 {
4507 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4508 {
4509 //printf("%x: dirty no match\n",addr);
4510 return 0;
4511 }
4512 }
4513 }
4514 }
4515 }
4516 }
4517 // Delay slots are not valid branch targets
4518 //if(t>0&&(dops[t-1].is_jump) return 0;
4519 // Delay slots require additional processing, so do not match
4520 if(dops[t].is_ds) return 0;
4521 }
4522 else
4523 {
4524 int hr;
4525 for(hr=0;hr<HOST_REGS;hr++)
4526 {
4527 if(hr!=EXCLUDE_REG)
4528 {
4529 if(i_regmap[hr]>=0)
4530 {
4531 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4532 {
4533 if((i_dirty>>hr)&1)
4534 {
4535 return 0;
4536 }
4537 }
4538 }
4539 }
4540 }
4541 }
4542 return 1;
4543}
4544
4545#ifdef DRC_DBG
4546static void drc_dbg_emit_do_cmp(int i)
4547{
4548 extern void do_insn_cmp();
4549 //extern int cycle;
4550 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4551
4552 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4553 save_regs(reglist);
4554 // write out changed consts to match the interpreter
4555 if (i > 0 && !dops[i].bt) {
4556 for (hr = 0; hr < HOST_REGS; hr++) {
4557 int reg = regs[i-1].regmap[hr];
4558 if (hr == EXCLUDE_REG || reg < 0)
4559 continue;
4560 if (!((regs[i-1].isconst >> hr) & 1))
4561 continue;
4562 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4563 continue;
4564 emit_movimm(constmap[i-1][hr],0);
4565 emit_storereg(reg, 0);
4566 }
4567 }
4568 emit_movimm(start+i*4,0);
4569 emit_writeword(0,&pcaddr);
4570 emit_far_call(do_insn_cmp);
4571 //emit_readword(&cycle,0);
4572 //emit_addimm(0,2,0);
4573 //emit_writeword(0,&cycle);
4574 (void)get_reg2;
4575 restore_regs(reglist);
4576 assem_debug("\\\\do_insn_cmp\n");
4577}
4578#else
4579#define drc_dbg_emit_do_cmp(x)
4580#endif
4581
4582// Used when a branch jumps into the delay slot of another branch
4583static void ds_assemble_entry(int i)
4584{
4585 int t=(ba[i]-start)>>2;
4586 if (!instr_addr[t])
4587 instr_addr[t] = out;
4588 assem_debug("Assemble delay slot at %x\n",ba[i]);
4589 assem_debug("<->\n");
4590 drc_dbg_emit_do_cmp(t);
4591 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4592 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4593 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4594 address_generation(t,&regs[t],regs[t].regmap_entry);
4595 if(dops[t].itype==STORE||dops[t].itype==STORELR||(dops[t].opcode&0x3b)==0x39||(dops[t].opcode&0x3b)==0x3a)
4596 load_regs(regs[t].regmap_entry,regs[t].regmap,INVCP,INVCP);
4597 is_delayslot=0;
4598 switch(dops[t].itype) {
4599 case ALU:
4600 alu_assemble(t,&regs[t]);break;
4601 case IMM16:
4602 imm16_assemble(t,&regs[t]);break;
4603 case SHIFT:
4604 shift_assemble(t,&regs[t]);break;
4605 case SHIFTIMM:
4606 shiftimm_assemble(t,&regs[t]);break;
4607 case LOAD:
4608 load_assemble(t,&regs[t]);break;
4609 case LOADLR:
4610 loadlr_assemble(t,&regs[t]);break;
4611 case STORE:
4612 store_assemble(t,&regs[t]);break;
4613 case STORELR:
4614 storelr_assemble(t,&regs[t]);break;
4615 case COP0:
4616 cop0_assemble(t,&regs[t]);break;
4617 case COP1:
4618 cop1_assemble(t,&regs[t]);break;
4619 case C1LS:
4620 c1ls_assemble(t,&regs[t]);break;
4621 case COP2:
4622 cop2_assemble(t,&regs[t]);break;
4623 case C2LS:
4624 c2ls_assemble(t,&regs[t]);break;
4625 case C2OP:
4626 c2op_assemble(t,&regs[t]);break;
4627 case MULTDIV:
4628 multdiv_assemble(t,&regs[t]);
4629 multdiv_prepare_stall(i,&regs[t]);
4630 break;
4631 case MOV:
4632 mov_assemble(t,&regs[t]);break;
4633 case SYSCALL:
4634 case HLECALL:
4635 case INTCALL:
4636 case SPAN:
4637 case UJUMP:
4638 case RJUMP:
4639 case CJUMP:
4640 case SJUMP:
4641 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4642 }
4643 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4644 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4645 if(internal_branch(ba[i]+4))
4646 assem_debug("branch: internal\n");
4647 else
4648 assem_debug("branch: external\n");
4649 assert(internal_branch(ba[i]+4));
4650 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4651 emit_jmp(0);
4652}
4653
4654static void emit_extjump(void *addr, u_int target)
4655{
4656 emit_extjump2(addr, target, dyna_linker);
4657}
4658
4659static void emit_extjump_ds(void *addr, u_int target)
4660{
4661 emit_extjump2(addr, target, dyna_linker_ds);
4662}
4663
4664// Load 2 immediates optimizing for small code size
4665static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4666{
4667 emit_movimm(imm1,rt1);
4668 emit_movimm_from(imm1,rt1,imm2,rt2);
4669}
4670
4671void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
4672{
4673 int count;
4674 void *jaddr;
4675 void *idle=NULL;
4676 int t=0;
4677 if(dops[i].itype==RJUMP)
4678 {
4679 *adj=0;
4680 }
4681 //if(ba[i]>=start && ba[i]<(start+slen*4))
4682 if(internal_branch(ba[i]))
4683 {
4684 t=(ba[i]-start)>>2;
4685 if(dops[t].is_ds) *adj=-1; // Branch into delay slot adds an extra cycle
4686 else *adj=ccadj[t];
4687 }
4688 else
4689 {
4690 *adj=0;
4691 }
4692 count=ccadj[i];
4693 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4694 // Idle loop
4695 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4696 idle=out;
4697 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4698 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4699 jaddr=out;
4700 emit_jmp(0);
4701 }
4702 else if(*adj==0||invert) {
4703 int cycles=CLOCK_ADJUST(count+2);
4704 // faster loop HACK
4705#if 0
4706 if (t&&*adj) {
4707 int rel=t-i;
4708 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4709 cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
4710 }
4711#endif
4712 emit_addimm_and_set_flags(cycles,HOST_CCREG);
4713 jaddr=out;
4714 emit_jns(0);
4715 }
4716 else
4717 {
4718 emit_cmpimm(HOST_CCREG,-CLOCK_ADJUST(count+2));
4719 jaddr=out;
4720 emit_jns(0);
4721 }
4722 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:(count+2),i,addr,taken,0);
4723}
4724
4725static void do_ccstub(int n)
4726{
4727 literal_pool(256);
4728 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
4729 set_jump_target(stubs[n].addr, out);
4730 int i=stubs[n].b;
4731 if(stubs[n].d==NULLDS) {
4732 // Delay slot instruction is nullified ("likely" branch)
4733 wb_dirtys(regs[i].regmap,regs[i].dirty);
4734 }
4735 else if(stubs[n].d!=TAKEN) {
4736 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
4737 }
4738 else {
4739 if(internal_branch(ba[i]))
4740 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4741 }
4742 if(stubs[n].c!=-1)
4743 {
4744 // Save PC as return address
4745 emit_movimm(stubs[n].c,EAX);
4746 emit_writeword(EAX,&pcaddr);
4747 }
4748 else
4749 {
4750 // Return address depends on which way the branch goes
4751 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
4752 {
4753 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
4754 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
4755 if(dops[i].rs1==0)
4756 {
4757 s1l=s2l;
4758 s2l=-1;
4759 }
4760 else if(dops[i].rs2==0)
4761 {
4762 s2l=-1;
4763 }
4764 assert(s1l>=0);
4765 #ifdef DESTRUCTIVE_WRITEBACK
4766 if(dops[i].rs1) {
4767 if((branch_regs[i].dirty>>s1l)&&1)
4768 emit_loadreg(dops[i].rs1,s1l);
4769 }
4770 else {
4771 if((branch_regs[i].dirty>>s1l)&1)
4772 emit_loadreg(dops[i].rs2,s1l);
4773 }
4774 if(s2l>=0)
4775 if((branch_regs[i].dirty>>s2l)&1)
4776 emit_loadreg(dops[i].rs2,s2l);
4777 #endif
4778 int hr=0;
4779 int addr=-1,alt=-1,ntaddr=-1;
4780 while(hr<HOST_REGS)
4781 {
4782 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4783 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4784 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4785 {
4786 addr=hr++;break;
4787 }
4788 hr++;
4789 }
4790 while(hr<HOST_REGS)
4791 {
4792 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4793 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4794 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4795 {
4796 alt=hr++;break;
4797 }
4798 hr++;
4799 }
4800 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
4801 {
4802 while(hr<HOST_REGS)
4803 {
4804 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4805 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4806 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4807 {
4808 ntaddr=hr;break;
4809 }
4810 hr++;
4811 }
4812 assert(hr<HOST_REGS);
4813 }
4814 if((dops[i].opcode&0x2f)==4) // BEQ
4815 {
4816 #ifdef HAVE_CMOV_IMM
4817 if(s2l>=0) emit_cmp(s1l,s2l);
4818 else emit_test(s1l,s1l);
4819 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4820 #else
4821 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4822 if(s2l>=0) emit_cmp(s1l,s2l);
4823 else emit_test(s1l,s1l);
4824 emit_cmovne_reg(alt,addr);
4825 #endif
4826 }
4827 if((dops[i].opcode&0x2f)==5) // BNE
4828 {
4829 #ifdef HAVE_CMOV_IMM
4830 if(s2l>=0) emit_cmp(s1l,s2l);
4831 else emit_test(s1l,s1l);
4832 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4833 #else
4834 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4835 if(s2l>=0) emit_cmp(s1l,s2l);
4836 else emit_test(s1l,s1l);
4837 emit_cmovne_reg(alt,addr);
4838 #endif
4839 }
4840 if((dops[i].opcode&0x2f)==6) // BLEZ
4841 {
4842 //emit_movimm(ba[i],alt);
4843 //emit_movimm(start+i*4+8,addr);
4844 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4845 emit_cmpimm(s1l,1);
4846 emit_cmovl_reg(alt,addr);
4847 }
4848 if((dops[i].opcode&0x2f)==7) // BGTZ
4849 {
4850 //emit_movimm(ba[i],addr);
4851 //emit_movimm(start+i*4+8,ntaddr);
4852 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
4853 emit_cmpimm(s1l,1);
4854 emit_cmovl_reg(ntaddr,addr);
4855 }
4856 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
4857 {
4858 //emit_movimm(ba[i],alt);
4859 //emit_movimm(start+i*4+8,addr);
4860 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4861 emit_test(s1l,s1l);
4862 emit_cmovs_reg(alt,addr);
4863 }
4864 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
4865 {
4866 //emit_movimm(ba[i],addr);
4867 //emit_movimm(start+i*4+8,alt);
4868 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4869 emit_test(s1l,s1l);
4870 emit_cmovs_reg(alt,addr);
4871 }
4872 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
4873 if(source[i]&0x10000) // BC1T
4874 {
4875 //emit_movimm(ba[i],alt);
4876 //emit_movimm(start+i*4+8,addr);
4877 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4878 emit_testimm(s1l,0x800000);
4879 emit_cmovne_reg(alt,addr);
4880 }
4881 else // BC1F
4882 {
4883 //emit_movimm(ba[i],addr);
4884 //emit_movimm(start+i*4+8,alt);
4885 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4886 emit_testimm(s1l,0x800000);
4887 emit_cmovne_reg(alt,addr);
4888 }
4889 }
4890 emit_writeword(addr,&pcaddr);
4891 }
4892 else
4893 if(dops[i].itype==RJUMP)
4894 {
4895 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
4896 if (ds_writes_rjump_rs(i)) {
4897 r=get_reg(branch_regs[i].regmap,RTEMP);
4898 }
4899 emit_writeword(r,&pcaddr);
4900 }
4901 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
4902 }
4903 // Update cycle count
4904 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
4905 if(stubs[n].a) emit_addimm(HOST_CCREG,CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
4906 emit_far_call(cc_interrupt);
4907 if(stubs[n].a) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
4908 if(stubs[n].d==TAKEN) {
4909 if(internal_branch(ba[i]))
4910 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
4911 else if(dops[i].itype==RJUMP) {
4912 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
4913 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
4914 else
4915 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
4916 }
4917 }else if(stubs[n].d==NOTTAKEN) {
4918 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
4919 else load_all_regs(branch_regs[i].regmap);
4920 }else if(stubs[n].d==NULLDS) {
4921 // Delay slot instruction is nullified ("likely" branch)
4922 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
4923 else load_all_regs(regs[i].regmap);
4924 }else{
4925 load_all_regs(branch_regs[i].regmap);
4926 }
4927 if (stubs[n].retaddr)
4928 emit_jmp(stubs[n].retaddr);
4929 else
4930 do_jump_vaddr(stubs[n].e);
4931}
4932
4933static void add_to_linker(void *addr, u_int target, int ext)
4934{
4935 assert(linkcount < ARRAY_SIZE(link_addr));
4936 link_addr[linkcount].addr = addr;
4937 link_addr[linkcount].target = target;
4938 link_addr[linkcount].ext = ext;
4939 linkcount++;
4940}
4941
4942static void ujump_assemble_write_ra(int i)
4943{
4944 int rt;
4945 unsigned int return_address;
4946 rt=get_reg(branch_regs[i].regmap,31);
4947 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]);
4948 //assert(rt>=0);
4949 return_address=start+i*4+8;
4950 if(rt>=0) {
4951 #ifdef USE_MINI_HT
4952 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
4953 int temp=-1; // note: must be ds-safe
4954 #ifdef HOST_TEMPREG
4955 temp=HOST_TEMPREG;
4956 #endif
4957 if(temp>=0) do_miniht_insert(return_address,rt,temp);
4958 else emit_movimm(return_address,rt);
4959 }
4960 else
4961 #endif
4962 {
4963 #ifdef REG_PREFETCH
4964 if(temp>=0)
4965 {
4966 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4967 }
4968 #endif
4969 emit_movimm(return_address,rt); // PC into link register
4970 #ifdef IMM_PREFETCH
4971 emit_prefetch(hash_table_get(return_address));
4972 #endif
4973 }
4974 }
4975}
4976
4977static void ujump_assemble(int i,struct regstat *i_regs)
4978{
4979 int ra_done=0;
4980 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4981 address_generation(i+1,i_regs,regs[i].regmap_entry);
4982 #ifdef REG_PREFETCH
4983 int temp=get_reg(branch_regs[i].regmap,PTEMP);
4984 if(dops[i].rt1==31&&temp>=0)
4985 {
4986 signed char *i_regmap=i_regs->regmap;
4987 int return_address=start+i*4+8;
4988 if(get_reg(branch_regs[i].regmap,31)>0)
4989 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4990 }
4991 #endif
4992 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
4993 ujump_assemble_write_ra(i); // writeback ra for DS
4994 ra_done=1;
4995 }
4996 ds_assemble(i+1,i_regs);
4997 uint64_t bc_unneeded=branch_regs[i].u;
4998 bc_unneeded|=1|(1LL<<dops[i].rt1);
4999 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5000 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5001 if(!ra_done&&dops[i].rt1==31)
5002 ujump_assemble_write_ra(i);
5003 int cc,adj;
5004 cc=get_reg(branch_regs[i].regmap,CCREG);
5005 assert(cc==HOST_CCREG);
5006 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5007 #ifdef REG_PREFETCH
5008 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5009 #endif
5010 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5011 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5012 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5013 if(internal_branch(ba[i]))
5014 assem_debug("branch: internal\n");
5015 else
5016 assem_debug("branch: external\n");
5017 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5018 ds_assemble_entry(i);
5019 }
5020 else {
5021 add_to_linker(out,ba[i],internal_branch(ba[i]));
5022 emit_jmp(0);
5023 }
5024}
5025
5026static void rjump_assemble_write_ra(int i)
5027{
5028 int rt,return_address;
5029 assert(dops[i+1].rt1!=dops[i].rt1);
5030 assert(dops[i+1].rt2!=dops[i].rt1);
5031 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5032 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]);
5033 assert(rt>=0);
5034 return_address=start+i*4+8;
5035 #ifdef REG_PREFETCH
5036 if(temp>=0)
5037 {
5038 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5039 }
5040 #endif
5041 emit_movimm(return_address,rt); // PC into link register
5042 #ifdef IMM_PREFETCH
5043 emit_prefetch(hash_table_get(return_address));
5044 #endif
5045}
5046
5047static void rjump_assemble(int i,struct regstat *i_regs)
5048{
5049 int temp;
5050 int rs,cc;
5051 int ra_done=0;
5052 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5053 assert(rs>=0);
5054 if (ds_writes_rjump_rs(i)) {
5055 // Delay slot abuse, make a copy of the branch address register
5056 temp=get_reg(branch_regs[i].regmap,RTEMP);
5057 assert(temp>=0);
5058 assert(regs[i].regmap[temp]==RTEMP);
5059 emit_mov(rs,temp);
5060 rs=temp;
5061 }
5062 address_generation(i+1,i_regs,regs[i].regmap_entry);
5063 #ifdef REG_PREFETCH
5064 if(dops[i].rt1==31)
5065 {
5066 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5067 signed char *i_regmap=i_regs->regmap;
5068 int return_address=start+i*4+8;
5069 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5070 }
5071 }
5072 #endif
5073 #ifdef USE_MINI_HT
5074 if(dops[i].rs1==31) {
5075 int rh=get_reg(regs[i].regmap,RHASH);
5076 if(rh>=0) do_preload_rhash(rh);
5077 }
5078 #endif
5079 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5080 rjump_assemble_write_ra(i);
5081 ra_done=1;
5082 }
5083 ds_assemble(i+1,i_regs);
5084 uint64_t bc_unneeded=branch_regs[i].u;
5085 bc_unneeded|=1|(1LL<<dops[i].rt1);
5086 bc_unneeded&=~(1LL<<dops[i].rs1);
5087 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5088 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5089 if(!ra_done&&dops[i].rt1!=0)
5090 rjump_assemble_write_ra(i);
5091 cc=get_reg(branch_regs[i].regmap,CCREG);
5092 assert(cc==HOST_CCREG);
5093 (void)cc;
5094 #ifdef USE_MINI_HT
5095 int rh=get_reg(branch_regs[i].regmap,RHASH);
5096 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5097 if(dops[i].rs1==31) {
5098 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5099 do_preload_rhtbl(ht);
5100 do_rhash(rs,rh);
5101 }
5102 #endif
5103 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5104 #ifdef DESTRUCTIVE_WRITEBACK
5105 if((branch_regs[i].dirty>>rs)&1) {
5106 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5107 emit_loadreg(dops[i].rs1,rs);
5108 }
5109 }
5110 #endif
5111 #ifdef REG_PREFETCH
5112 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5113 #endif
5114 #ifdef USE_MINI_HT
5115 if(dops[i].rs1==31) {
5116 do_miniht_load(ht,rh);
5117 }
5118 #endif
5119 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5120 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5121 //assert(adj==0);
5122 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5123 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5124 if(dops[i+1].itype==COP0&&(source[i+1]&0x3f)==0x10)
5125 // special case for RFE
5126 emit_jmp(0);
5127 else
5128 emit_jns(0);
5129 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5130 #ifdef USE_MINI_HT
5131 if(dops[i].rs1==31) {
5132 do_miniht_jump(rs,rh,ht);
5133 }
5134 else
5135 #endif
5136 {
5137 do_jump_vaddr(rs);
5138 }
5139 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5140 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5141 #endif
5142}
5143
5144static void cjump_assemble(int i,struct regstat *i_regs)
5145{
5146 signed char *i_regmap=i_regs->regmap;
5147 int cc;
5148 int match;
5149 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5150 assem_debug("match=%d\n",match);
5151 int s1l,s2l;
5152 int unconditional=0,nop=0;
5153 int invert=0;
5154 int internal=internal_branch(ba[i]);
5155 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5156 if(!match) invert=1;
5157 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5158 if(i>(ba[i]-start)>>2) invert=1;
5159 #endif
5160 #ifdef __aarch64__
5161 invert=1; // because of near cond. branches
5162 #endif
5163
5164 if(dops[i].ooo) {
5165 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5166 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5167 }
5168 else {
5169 s1l=get_reg(i_regmap,dops[i].rs1);
5170 s2l=get_reg(i_regmap,dops[i].rs2);
5171 }
5172 if(dops[i].rs1==0&&dops[i].rs2==0)
5173 {
5174 if(dops[i].opcode&1) nop=1;
5175 else unconditional=1;
5176 //assert(dops[i].opcode!=5);
5177 //assert(dops[i].opcode!=7);
5178 //assert(dops[i].opcode!=0x15);
5179 //assert(dops[i].opcode!=0x17);
5180 }
5181 else if(dops[i].rs1==0)
5182 {
5183 s1l=s2l;
5184 s2l=-1;
5185 }
5186 else if(dops[i].rs2==0)
5187 {
5188 s2l=-1;
5189 }
5190
5191 if(dops[i].ooo) {
5192 // Out of order execution (delay slot first)
5193 //printf("OOOE\n");
5194 address_generation(i+1,i_regs,regs[i].regmap_entry);
5195 ds_assemble(i+1,i_regs);
5196 int adj;
5197 uint64_t bc_unneeded=branch_regs[i].u;
5198 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5199 bc_unneeded|=1;
5200 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5201 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5202 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5203 cc=get_reg(branch_regs[i].regmap,CCREG);
5204 assert(cc==HOST_CCREG);
5205 if(unconditional)
5206 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5207 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5208 //assem_debug("cycle count (adj)\n");
5209 if(unconditional) {
5210 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5211 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5212 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5213 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5214 if(internal)
5215 assem_debug("branch: internal\n");
5216 else
5217 assem_debug("branch: external\n");
5218 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5219 ds_assemble_entry(i);
5220 }
5221 else {
5222 add_to_linker(out,ba[i],internal);
5223 emit_jmp(0);
5224 }
5225 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5226 if(((u_int)out)&7) emit_addnop(0);
5227 #endif
5228 }
5229 }
5230 else if(nop) {
5231 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5232 void *jaddr=out;
5233 emit_jns(0);
5234 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5235 }
5236 else {
5237 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5238 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5239 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5240
5241 //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]);
5242 assert(s1l>=0);
5243 if(dops[i].opcode==4) // BEQ
5244 {
5245 if(s2l>=0) emit_cmp(s1l,s2l);
5246 else emit_test(s1l,s1l);
5247 if(invert){
5248 nottaken=out;
5249 emit_jne(DJT_1);
5250 }else{
5251 add_to_linker(out,ba[i],internal);
5252 emit_jeq(0);
5253 }
5254 }
5255 if(dops[i].opcode==5) // BNE
5256 {
5257 if(s2l>=0) emit_cmp(s1l,s2l);
5258 else emit_test(s1l,s1l);
5259 if(invert){
5260 nottaken=out;
5261 emit_jeq(DJT_1);
5262 }else{
5263 add_to_linker(out,ba[i],internal);
5264 emit_jne(0);
5265 }
5266 }
5267 if(dops[i].opcode==6) // BLEZ
5268 {
5269 emit_cmpimm(s1l,1);
5270 if(invert){
5271 nottaken=out;
5272 emit_jge(DJT_1);
5273 }else{
5274 add_to_linker(out,ba[i],internal);
5275 emit_jl(0);
5276 }
5277 }
5278 if(dops[i].opcode==7) // BGTZ
5279 {
5280 emit_cmpimm(s1l,1);
5281 if(invert){
5282 nottaken=out;
5283 emit_jl(DJT_1);
5284 }else{
5285 add_to_linker(out,ba[i],internal);
5286 emit_jge(0);
5287 }
5288 }
5289 if(invert) {
5290 if(taken) set_jump_target(taken, out);
5291 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5292 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5293 if(adj) {
5294 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5295 add_to_linker(out,ba[i],internal);
5296 }else{
5297 emit_addnop(13);
5298 add_to_linker(out,ba[i],internal*2);
5299 }
5300 emit_jmp(0);
5301 }else
5302 #endif
5303 {
5304 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5305 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5306 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5307 if(internal)
5308 assem_debug("branch: internal\n");
5309 else
5310 assem_debug("branch: external\n");
5311 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5312 ds_assemble_entry(i);
5313 }
5314 else {
5315 add_to_linker(out,ba[i],internal);
5316 emit_jmp(0);
5317 }
5318 }
5319 set_jump_target(nottaken, out);
5320 }
5321
5322 if(nottaken1) set_jump_target(nottaken1, out);
5323 if(adj) {
5324 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5325 }
5326 } // (!unconditional)
5327 } // if(ooo)
5328 else
5329 {
5330 // In-order execution (branch first)
5331 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5332 if(!unconditional&&!nop) {
5333 //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]);
5334 assert(s1l>=0);
5335 if((dops[i].opcode&0x2f)==4) // BEQ
5336 {
5337 if(s2l>=0) emit_cmp(s1l,s2l);
5338 else emit_test(s1l,s1l);
5339 nottaken=out;
5340 emit_jne(DJT_2);
5341 }
5342 if((dops[i].opcode&0x2f)==5) // BNE
5343 {
5344 if(s2l>=0) emit_cmp(s1l,s2l);
5345 else emit_test(s1l,s1l);
5346 nottaken=out;
5347 emit_jeq(DJT_2);
5348 }
5349 if((dops[i].opcode&0x2f)==6) // BLEZ
5350 {
5351 emit_cmpimm(s1l,1);
5352 nottaken=out;
5353 emit_jge(DJT_2);
5354 }
5355 if((dops[i].opcode&0x2f)==7) // BGTZ
5356 {
5357 emit_cmpimm(s1l,1);
5358 nottaken=out;
5359 emit_jl(DJT_2);
5360 }
5361 } // if(!unconditional)
5362 int adj;
5363 uint64_t ds_unneeded=branch_regs[i].u;
5364 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5365 ds_unneeded|=1;
5366 // branch taken
5367 if(!nop) {
5368 if(taken) set_jump_target(taken, out);
5369 assem_debug("1:\n");
5370 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5371 // load regs
5372 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5373 address_generation(i+1,&branch_regs[i],0);
5374 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5375 ds_assemble(i+1,&branch_regs[i]);
5376 cc=get_reg(branch_regs[i].regmap,CCREG);
5377 if(cc==-1) {
5378 emit_loadreg(CCREG,cc=HOST_CCREG);
5379 // CHECK: Is the following instruction (fall thru) allocated ok?
5380 }
5381 assert(cc==HOST_CCREG);
5382 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5383 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5384 assem_debug("cycle count (adj)\n");
5385 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5386 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5387 if(internal)
5388 assem_debug("branch: internal\n");
5389 else
5390 assem_debug("branch: external\n");
5391 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5392 ds_assemble_entry(i);
5393 }
5394 else {
5395 add_to_linker(out,ba[i],internal);
5396 emit_jmp(0);
5397 }
5398 }
5399 // branch not taken
5400 if(!unconditional) {
5401 if(nottaken1) set_jump_target(nottaken1, out);
5402 set_jump_target(nottaken, out);
5403 assem_debug("2:\n");
5404 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5405 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5406 address_generation(i+1,&branch_regs[i],0);
5407 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5408 ds_assemble(i+1,&branch_regs[i]);
5409 cc=get_reg(branch_regs[i].regmap,CCREG);
5410 if (cc == -1) {
5411 // Cycle count isn't in a register, temporarily load it then write it out
5412 emit_loadreg(CCREG,HOST_CCREG);
5413 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5414 void *jaddr=out;
5415 emit_jns(0);
5416 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5417 emit_storereg(CCREG,HOST_CCREG);
5418 }
5419 else{
5420 cc=get_reg(i_regmap,CCREG);
5421 assert(cc==HOST_CCREG);
5422 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5423 void *jaddr=out;
5424 emit_jns(0);
5425 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5426 }
5427 }
5428 }
5429}
5430
5431static void sjump_assemble(int i,struct regstat *i_regs)
5432{
5433 signed char *i_regmap=i_regs->regmap;
5434 int cc;
5435 int match;
5436 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5437 assem_debug("smatch=%d\n",match);
5438 int s1l;
5439 int unconditional=0,nevertaken=0;
5440 int invert=0;
5441 int internal=internal_branch(ba[i]);
5442 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5443 if(!match) invert=1;
5444 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5445 if(i>(ba[i]-start)>>2) invert=1;
5446 #endif
5447 #ifdef __aarch64__
5448 invert=1; // because of near cond. branches
5449 #endif
5450
5451 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5452 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5453
5454 if(dops[i].ooo) {
5455 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5456 }
5457 else {
5458 s1l=get_reg(i_regmap,dops[i].rs1);
5459 }
5460 if(dops[i].rs1==0)
5461 {
5462 if(dops[i].opcode2&1) unconditional=1;
5463 else nevertaken=1;
5464 // These are never taken (r0 is never less than zero)
5465 //assert(dops[i].opcode2!=0);
5466 //assert(dops[i].opcode2!=2);
5467 //assert(dops[i].opcode2!=0x10);
5468 //assert(dops[i].opcode2!=0x12);
5469 }
5470
5471 if(dops[i].ooo) {
5472 // Out of order execution (delay slot first)
5473 //printf("OOOE\n");
5474 address_generation(i+1,i_regs,regs[i].regmap_entry);
5475 ds_assemble(i+1,i_regs);
5476 int adj;
5477 uint64_t bc_unneeded=branch_regs[i].u;
5478 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5479 bc_unneeded|=1;
5480 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5481 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5482 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5483 if(dops[i].rt1==31) {
5484 int rt,return_address;
5485 rt=get_reg(branch_regs[i].regmap,31);
5486 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]);
5487 if(rt>=0) {
5488 // Save the PC even if the branch is not taken
5489 return_address=start+i*4+8;
5490 emit_movimm(return_address,rt); // PC into link register
5491 #ifdef IMM_PREFETCH
5492 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5493 #endif
5494 }
5495 }
5496 cc=get_reg(branch_regs[i].regmap,CCREG);
5497 assert(cc==HOST_CCREG);
5498 if(unconditional)
5499 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5500 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5501 assem_debug("cycle count (adj)\n");
5502 if(unconditional) {
5503 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5504 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5505 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5506 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5507 if(internal)
5508 assem_debug("branch: internal\n");
5509 else
5510 assem_debug("branch: external\n");
5511 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5512 ds_assemble_entry(i);
5513 }
5514 else {
5515 add_to_linker(out,ba[i],internal);
5516 emit_jmp(0);
5517 }
5518 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5519 if(((u_int)out)&7) emit_addnop(0);
5520 #endif
5521 }
5522 }
5523 else if(nevertaken) {
5524 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5525 void *jaddr=out;
5526 emit_jns(0);
5527 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5528 }
5529 else {
5530 void *nottaken = NULL;
5531 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5532 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5533 {
5534 assert(s1l>=0);
5535 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5536 {
5537 emit_test(s1l,s1l);
5538 if(invert){
5539 nottaken=out;
5540 emit_jns(DJT_1);
5541 }else{
5542 add_to_linker(out,ba[i],internal);
5543 emit_js(0);
5544 }
5545 }
5546 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5547 {
5548 emit_test(s1l,s1l);
5549 if(invert){
5550 nottaken=out;
5551 emit_js(DJT_1);
5552 }else{
5553 add_to_linker(out,ba[i],internal);
5554 emit_jns(0);
5555 }
5556 }
5557 }
5558
5559 if(invert) {
5560 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5561 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5562 if(adj) {
5563 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5564 add_to_linker(out,ba[i],internal);
5565 }else{
5566 emit_addnop(13);
5567 add_to_linker(out,ba[i],internal*2);
5568 }
5569 emit_jmp(0);
5570 }else
5571 #endif
5572 {
5573 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5574 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5575 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5576 if(internal)
5577 assem_debug("branch: internal\n");
5578 else
5579 assem_debug("branch: external\n");
5580 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5581 ds_assemble_entry(i);
5582 }
5583 else {
5584 add_to_linker(out,ba[i],internal);
5585 emit_jmp(0);
5586 }
5587 }
5588 set_jump_target(nottaken, out);
5589 }
5590
5591 if(adj) {
5592 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5593 }
5594 } // (!unconditional)
5595 } // if(ooo)
5596 else
5597 {
5598 // In-order execution (branch first)
5599 //printf("IOE\n");
5600 void *nottaken = NULL;
5601 if(dops[i].rt1==31) {
5602 int rt,return_address;
5603 rt=get_reg(branch_regs[i].regmap,31);
5604 if(rt>=0) {
5605 // Save the PC even if the branch is not taken
5606 return_address=start+i*4+8;
5607 emit_movimm(return_address,rt); // PC into link register
5608 #ifdef IMM_PREFETCH
5609 emit_prefetch(hash_table_get(return_address));
5610 #endif
5611 }
5612 }
5613 if(!unconditional) {
5614 //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]);
5615 assert(s1l>=0);
5616 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5617 {
5618 emit_test(s1l,s1l);
5619 nottaken=out;
5620 emit_jns(DJT_1);
5621 }
5622 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5623 {
5624 emit_test(s1l,s1l);
5625 nottaken=out;
5626 emit_js(DJT_1);
5627 }
5628 } // if(!unconditional)
5629 int adj;
5630 uint64_t ds_unneeded=branch_regs[i].u;
5631 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5632 ds_unneeded|=1;
5633 // branch taken
5634 if(!nevertaken) {
5635 //assem_debug("1:\n");
5636 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5637 // load regs
5638 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5639 address_generation(i+1,&branch_regs[i],0);
5640 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5641 ds_assemble(i+1,&branch_regs[i]);
5642 cc=get_reg(branch_regs[i].regmap,CCREG);
5643 if(cc==-1) {
5644 emit_loadreg(CCREG,cc=HOST_CCREG);
5645 // CHECK: Is the following instruction (fall thru) allocated ok?
5646 }
5647 assert(cc==HOST_CCREG);
5648 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5649 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5650 assem_debug("cycle count (adj)\n");
5651 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5652 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5653 if(internal)
5654 assem_debug("branch: internal\n");
5655 else
5656 assem_debug("branch: external\n");
5657 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5658 ds_assemble_entry(i);
5659 }
5660 else {
5661 add_to_linker(out,ba[i],internal);
5662 emit_jmp(0);
5663 }
5664 }
5665 // branch not taken
5666 if(!unconditional) {
5667 set_jump_target(nottaken, out);
5668 assem_debug("1:\n");
5669 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5670 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5671 address_generation(i+1,&branch_regs[i],0);
5672 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5673 ds_assemble(i+1,&branch_regs[i]);
5674 cc=get_reg(branch_regs[i].regmap,CCREG);
5675 if (cc == -1) {
5676 // Cycle count isn't in a register, temporarily load it then write it out
5677 emit_loadreg(CCREG,HOST_CCREG);
5678 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5679 void *jaddr=out;
5680 emit_jns(0);
5681 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5682 emit_storereg(CCREG,HOST_CCREG);
5683 }
5684 else{
5685 cc=get_reg(i_regmap,CCREG);
5686 assert(cc==HOST_CCREG);
5687 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5688 void *jaddr=out;
5689 emit_jns(0);
5690 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5691 }
5692 }
5693 }
5694}
5695
5696static void pagespan_assemble(int i,struct regstat *i_regs)
5697{
5698 int s1l=get_reg(i_regs->regmap,dops[i].rs1);
5699 int s2l=get_reg(i_regs->regmap,dops[i].rs2);
5700 void *taken = NULL;
5701 void *nottaken = NULL;
5702 int unconditional=0;
5703 if(dops[i].rs1==0)
5704 {
5705 s1l=s2l;
5706 s2l=-1;
5707 }
5708 else if(dops[i].rs2==0)
5709 {
5710 s2l=-1;
5711 }
5712 int hr=0;
5713 int addr=-1,alt=-1,ntaddr=-1;
5714 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5715 else {
5716 while(hr<HOST_REGS)
5717 {
5718 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5719 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5720 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5721 {
5722 addr=hr++;break;
5723 }
5724 hr++;
5725 }
5726 }
5727 while(hr<HOST_REGS)
5728 {
5729 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5730 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5731 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5732 {
5733 alt=hr++;break;
5734 }
5735 hr++;
5736 }
5737 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5738 {
5739 while(hr<HOST_REGS)
5740 {
5741 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5742 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5743 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5744 {
5745 ntaddr=hr;break;
5746 }
5747 hr++;
5748 }
5749 }
5750 assert(hr<HOST_REGS);
5751 if((dops[i].opcode&0x2e)==4||dops[i].opcode==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5752 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
5753 }
5754 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5755 if(dops[i].opcode==2) // J
5756 {
5757 unconditional=1;
5758 }
5759 if(dops[i].opcode==3) // JAL
5760 {
5761 // TODO: mini_ht
5762 int rt=get_reg(i_regs->regmap,31);
5763 emit_movimm(start+i*4+8,rt);
5764 unconditional=1;
5765 }
5766 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
5767 {
5768 emit_mov(s1l,addr);
5769 if(dops[i].opcode2==9) // JALR
5770 {
5771 int rt=get_reg(i_regs->regmap,dops[i].rt1);
5772 emit_movimm(start+i*4+8,rt);
5773 }
5774 }
5775 if((dops[i].opcode&0x3f)==4) // BEQ
5776 {
5777 if(dops[i].rs1==dops[i].rs2)
5778 {
5779 unconditional=1;
5780 }
5781 else
5782 #ifdef HAVE_CMOV_IMM
5783 if(1) {
5784 if(s2l>=0) emit_cmp(s1l,s2l);
5785 else emit_test(s1l,s1l);
5786 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5787 }
5788 else
5789 #endif
5790 {
5791 assert(s1l>=0);
5792 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5793 if(s2l>=0) emit_cmp(s1l,s2l);
5794 else emit_test(s1l,s1l);
5795 emit_cmovne_reg(alt,addr);
5796 }
5797 }
5798 if((dops[i].opcode&0x3f)==5) // BNE
5799 {
5800 #ifdef HAVE_CMOV_IMM
5801 if(s2l>=0) emit_cmp(s1l,s2l);
5802 else emit_test(s1l,s1l);
5803 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5804 #else
5805 assert(s1l>=0);
5806 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5807 if(s2l>=0) emit_cmp(s1l,s2l);
5808 else emit_test(s1l,s1l);
5809 emit_cmovne_reg(alt,addr);
5810 #endif
5811 }
5812 if((dops[i].opcode&0x3f)==0x14) // BEQL
5813 {
5814 if(s2l>=0) emit_cmp(s1l,s2l);
5815 else emit_test(s1l,s1l);
5816 if(nottaken) set_jump_target(nottaken, out);
5817 nottaken=out;
5818 emit_jne(0);
5819 }
5820 if((dops[i].opcode&0x3f)==0x15) // BNEL
5821 {
5822 if(s2l>=0) emit_cmp(s1l,s2l);
5823 else emit_test(s1l,s1l);
5824 nottaken=out;
5825 emit_jeq(0);
5826 if(taken) set_jump_target(taken, out);
5827 }
5828 if((dops[i].opcode&0x3f)==6) // BLEZ
5829 {
5830 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5831 emit_cmpimm(s1l,1);
5832 emit_cmovl_reg(alt,addr);
5833 }
5834 if((dops[i].opcode&0x3f)==7) // BGTZ
5835 {
5836 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5837 emit_cmpimm(s1l,1);
5838 emit_cmovl_reg(ntaddr,addr);
5839 }
5840 if((dops[i].opcode&0x3f)==0x16) // BLEZL
5841 {
5842 assert((dops[i].opcode&0x3f)!=0x16);
5843 }
5844 if((dops[i].opcode&0x3f)==0x17) // BGTZL
5845 {
5846 assert((dops[i].opcode&0x3f)!=0x17);
5847 }
5848 assert(dops[i].opcode!=1); // BLTZ/BGEZ
5849
5850 //FIXME: Check CSREG
5851 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5852 if((source[i]&0x30000)==0) // BC1F
5853 {
5854 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5855 emit_testimm(s1l,0x800000);
5856 emit_cmovne_reg(alt,addr);
5857 }
5858 if((source[i]&0x30000)==0x10000) // BC1T
5859 {
5860 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5861 emit_testimm(s1l,0x800000);
5862 emit_cmovne_reg(alt,addr);
5863 }
5864 if((source[i]&0x30000)==0x20000) // BC1FL
5865 {
5866 emit_testimm(s1l,0x800000);
5867 nottaken=out;
5868 emit_jne(0);
5869 }
5870 if((source[i]&0x30000)==0x30000) // BC1TL
5871 {
5872 emit_testimm(s1l,0x800000);
5873 nottaken=out;
5874 emit_jeq(0);
5875 }
5876 }
5877
5878 assert(i_regs->regmap[HOST_CCREG]==CCREG);
5879 wb_dirtys(regs[i].regmap,regs[i].dirty);
5880 if(unconditional)
5881 {
5882 emit_movimm(ba[i],HOST_BTREG);
5883 }
5884 else if(addr!=HOST_BTREG)
5885 {
5886 emit_mov(addr,HOST_BTREG);
5887 }
5888 void *branch_addr=out;
5889 emit_jmp(0);
5890 int target_addr=start+i*4+5;
5891 void *stub=out;
5892 void *compiled_target_addr=check_addr(target_addr);
5893 emit_extjump_ds(branch_addr, target_addr);
5894 if(compiled_target_addr) {
5895 set_jump_target(branch_addr, compiled_target_addr);
5896 add_jump_out(target_addr,stub);
5897 }
5898 else set_jump_target(branch_addr, stub);
5899}
5900
5901// Assemble the delay slot for the above
5902static void pagespan_ds()
5903{
5904 assem_debug("initial delay slot:\n");
5905 u_int vaddr=start+1;
5906 u_int page=get_page(vaddr);
5907 u_int vpage=get_vpage(vaddr);
5908 ll_add(jump_dirty+vpage,vaddr,(void *)out);
5909 do_dirty_stub_ds(slen*4);
5910 ll_add(jump_in+page,vaddr,(void *)out);
5911 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
5912 if(regs[0].regmap[HOST_CCREG]!=CCREG)
5913 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
5914 if(regs[0].regmap[HOST_BTREG]!=BTREG)
5915 emit_writeword(HOST_BTREG,&branch_target);
5916 load_regs(regs[0].regmap_entry,regs[0].regmap,dops[0].rs1,dops[0].rs2);
5917 address_generation(0,&regs[0],regs[0].regmap_entry);
5918 if(dops[0].itype==STORE||dops[0].itype==STORELR||(dops[0].opcode&0x3b)==0x39||(dops[0].opcode&0x3b)==0x3a)
5919 load_regs(regs[0].regmap_entry,regs[0].regmap,INVCP,INVCP);
5920 is_delayslot=0;
5921 switch(dops[0].itype) {
5922 case ALU:
5923 alu_assemble(0,&regs[0]);break;
5924 case IMM16:
5925 imm16_assemble(0,&regs[0]);break;
5926 case SHIFT:
5927 shift_assemble(0,&regs[0]);break;
5928 case SHIFTIMM:
5929 shiftimm_assemble(0,&regs[0]);break;
5930 case LOAD:
5931 load_assemble(0,&regs[0]);break;
5932 case LOADLR:
5933 loadlr_assemble(0,&regs[0]);break;
5934 case STORE:
5935 store_assemble(0,&regs[0]);break;
5936 case STORELR:
5937 storelr_assemble(0,&regs[0]);break;
5938 case COP0:
5939 cop0_assemble(0,&regs[0]);break;
5940 case COP1:
5941 cop1_assemble(0,&regs[0]);break;
5942 case C1LS:
5943 c1ls_assemble(0,&regs[0]);break;
5944 case COP2:
5945 cop2_assemble(0,&regs[0]);break;
5946 case C2LS:
5947 c2ls_assemble(0,&regs[0]);break;
5948 case C2OP:
5949 c2op_assemble(0,&regs[0]);break;
5950 case MULTDIV:
5951 multdiv_assemble(0,&regs[0]);
5952 multdiv_prepare_stall(0,&regs[0]);
5953 break;
5954 case MOV:
5955 mov_assemble(0,&regs[0]);break;
5956 case SYSCALL:
5957 case HLECALL:
5958 case INTCALL:
5959 case SPAN:
5960 case UJUMP:
5961 case RJUMP:
5962 case CJUMP:
5963 case SJUMP:
5964 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5965 }
5966 int btaddr=get_reg(regs[0].regmap,BTREG);
5967 if(btaddr<0) {
5968 btaddr=get_reg(regs[0].regmap,-1);
5969 emit_readword(&branch_target,btaddr);
5970 }
5971 assert(btaddr!=HOST_CCREG);
5972 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
5973#ifdef HOST_IMM8
5974 host_tempreg_acquire();
5975 emit_movimm(start+4,HOST_TEMPREG);
5976 emit_cmp(btaddr,HOST_TEMPREG);
5977 host_tempreg_release();
5978#else
5979 emit_cmpimm(btaddr,start+4);
5980#endif
5981 void *branch = out;
5982 emit_jeq(0);
5983 store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
5984 do_jump_vaddr(btaddr);
5985 set_jump_target(branch, out);
5986 store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
5987 load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
5988}
5989
5990// Basic liveness analysis for MIPS registers
5991void unneeded_registers(int istart,int iend,int r)
5992{
5993 int i;
5994 uint64_t u,gte_u,b,gte_b;
5995 uint64_t temp_u,temp_gte_u=0;
5996 uint64_t gte_u_unknown=0;
5997 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
5998 gte_u_unknown=~0ll;
5999 if(iend==slen-1) {
6000 u=1;
6001 gte_u=gte_u_unknown;
6002 }else{
6003 //u=unneeded_reg[iend+1];
6004 u=1;
6005 gte_u=gte_unneeded[iend+1];
6006 }
6007
6008 for (i=iend;i>=istart;i--)
6009 {
6010 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6011 if(dops[i].is_jump)
6012 {
6013 // If subroutine call, flag return address as a possible branch target
6014 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
6015
6016 if(ba[i]<start || ba[i]>=(start+slen*4))
6017 {
6018 // Branch out of this block, flush all regs
6019 u=1;
6020 gte_u=gte_u_unknown;
6021 branch_unneeded_reg[i]=u;
6022 // Merge in delay slot
6023 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6024 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6025 u|=1;
6026 gte_u|=gte_rt[i+1];
6027 gte_u&=~gte_rs[i+1];
6028 }
6029 else
6030 {
6031 // Internal branch, flag target
6032 dops[(ba[i]-start)>>2].bt=1;
6033 if(ba[i]<=start+i*4) {
6034 // Backward branch
6035 if(dops[i].is_ujump)
6036 {
6037 // Unconditional branch
6038 temp_u=1;
6039 temp_gte_u=0;
6040 } else {
6041 // Conditional branch (not taken case)
6042 temp_u=unneeded_reg[i+2];
6043 temp_gte_u&=gte_unneeded[i+2];
6044 }
6045 // Merge in delay slot
6046 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6047 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6048 temp_u|=1;
6049 temp_gte_u|=gte_rt[i+1];
6050 temp_gte_u&=~gte_rs[i+1];
6051 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
6052 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
6053 temp_u|=1;
6054 temp_gte_u|=gte_rt[i];
6055 temp_gte_u&=~gte_rs[i];
6056 unneeded_reg[i]=temp_u;
6057 gte_unneeded[i]=temp_gte_u;
6058 // Only go three levels deep. This recursion can take an
6059 // excessive amount of time if there are a lot of nested loops.
6060 if(r<2) {
6061 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6062 }else{
6063 unneeded_reg[(ba[i]-start)>>2]=1;
6064 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
6065 }
6066 } /*else*/ if(1) {
6067 if (dops[i].is_ujump)
6068 {
6069 // Unconditional branch
6070 u=unneeded_reg[(ba[i]-start)>>2];
6071 gte_u=gte_unneeded[(ba[i]-start)>>2];
6072 branch_unneeded_reg[i]=u;
6073 // Merge in delay slot
6074 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6075 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6076 u|=1;
6077 gte_u|=gte_rt[i+1];
6078 gte_u&=~gte_rs[i+1];
6079 } else {
6080 // Conditional branch
6081 b=unneeded_reg[(ba[i]-start)>>2];
6082 gte_b=gte_unneeded[(ba[i]-start)>>2];
6083 branch_unneeded_reg[i]=b;
6084 // Branch delay slot
6085 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6086 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6087 b|=1;
6088 gte_b|=gte_rt[i+1];
6089 gte_b&=~gte_rs[i+1];
6090 u&=b;
6091 gte_u&=gte_b;
6092 if(i<slen-1) {
6093 branch_unneeded_reg[i]&=unneeded_reg[i+2];
6094 } else {
6095 branch_unneeded_reg[i]=1;
6096 }
6097 }
6098 }
6099 }
6100 }
6101 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6102 {
6103 // SYSCALL instruction (software interrupt)
6104 u=1;
6105 }
6106 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6107 {
6108 // ERET instruction (return from interrupt)
6109 u=1;
6110 }
6111 //u=1; // DEBUG
6112 // Written registers are unneeded
6113 u|=1LL<<dops[i].rt1;
6114 u|=1LL<<dops[i].rt2;
6115 gte_u|=gte_rt[i];
6116 // Accessed registers are needed
6117 u&=~(1LL<<dops[i].rs1);
6118 u&=~(1LL<<dops[i].rs2);
6119 gte_u&=~gte_rs[i];
6120 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
6121 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
6122 // Source-target dependencies
6123 // R0 is always unneeded
6124 u|=1;
6125 // Save it
6126 unneeded_reg[i]=u;
6127 gte_unneeded[i]=gte_u;
6128 /*
6129 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6130 printf("U:");
6131 int r;
6132 for(r=1;r<=CCREG;r++) {
6133 if((unneeded_reg[i]>>r)&1) {
6134 if(r==HIREG) printf(" HI");
6135 else if(r==LOREG) printf(" LO");
6136 else printf(" r%d",r);
6137 }
6138 }
6139 printf("\n");
6140 */
6141 }
6142}
6143
6144// Write back dirty registers as soon as we will no longer modify them,
6145// so that we don't end up with lots of writes at the branches.
6146void clean_registers(int istart,int iend,int wr)
6147{
6148 int i;
6149 int r;
6150 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
6151 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
6152 if(iend==slen-1) {
6153 will_dirty_i=will_dirty_next=0;
6154 wont_dirty_i=wont_dirty_next=0;
6155 }else{
6156 will_dirty_i=will_dirty_next=will_dirty[iend+1];
6157 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
6158 }
6159 for (i=iend;i>=istart;i--)
6160 {
6161 if(dops[i].is_jump)
6162 {
6163 if(ba[i]<start || ba[i]>=(start+slen*4))
6164 {
6165 // Branch out of this block, flush all regs
6166 if (dops[i].is_ujump)
6167 {
6168 // Unconditional branch
6169 will_dirty_i=0;
6170 wont_dirty_i=0;
6171 // Merge in delay slot (will dirty)
6172 for(r=0;r<HOST_REGS;r++) {
6173 if(r!=EXCLUDE_REG) {
6174 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6175 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6176 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6177 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6178 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6179 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6180 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6181 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6182 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6183 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6184 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6185 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6186 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6187 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6188 }
6189 }
6190 }
6191 else
6192 {
6193 // Conditional branch
6194 will_dirty_i=0;
6195 wont_dirty_i=wont_dirty_next;
6196 // Merge in delay slot (will dirty)
6197 for(r=0;r<HOST_REGS;r++) {
6198 if(r!=EXCLUDE_REG) {
6199 if (1) { // !dops[i].likely) {
6200 // Might not dirty if likely branch is not taken
6201 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6202 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6203 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6204 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6205 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6206 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
6207 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6208 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6209 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6210 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6211 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6212 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6213 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6214 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6215 }
6216 }
6217 }
6218 }
6219 // Merge in delay slot (wont dirty)
6220 for(r=0;r<HOST_REGS;r++) {
6221 if(r!=EXCLUDE_REG) {
6222 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6223 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6224 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6225 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6226 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6227 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6228 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6229 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6230 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6231 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6232 }
6233 }
6234 if(wr) {
6235 #ifndef DESTRUCTIVE_WRITEBACK
6236 branch_regs[i].dirty&=wont_dirty_i;
6237 #endif
6238 branch_regs[i].dirty|=will_dirty_i;
6239 }
6240 }
6241 else
6242 {
6243 // Internal branch
6244 if(ba[i]<=start+i*4) {
6245 // Backward branch
6246 if (dops[i].is_ujump)
6247 {
6248 // Unconditional branch
6249 temp_will_dirty=0;
6250 temp_wont_dirty=0;
6251 // Merge in delay slot (will dirty)
6252 for(r=0;r<HOST_REGS;r++) {
6253 if(r!=EXCLUDE_REG) {
6254 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6255 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6256 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6257 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6258 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6259 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6260 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6261 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6262 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6263 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6264 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6265 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6266 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6267 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6268 }
6269 }
6270 } else {
6271 // Conditional branch (not taken case)
6272 temp_will_dirty=will_dirty_next;
6273 temp_wont_dirty=wont_dirty_next;
6274 // Merge in delay slot (will dirty)
6275 for(r=0;r<HOST_REGS;r++) {
6276 if(r!=EXCLUDE_REG) {
6277 if (1) { // !dops[i].likely) {
6278 // Will not dirty if likely branch is not taken
6279 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6280 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6281 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6282 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6283 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6284 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
6285 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6286 //if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6287 //if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6288 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6289 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6290 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6291 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6292 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6293 }
6294 }
6295 }
6296 }
6297 // Merge in delay slot (wont dirty)
6298 for(r=0;r<HOST_REGS;r++) {
6299 if(r!=EXCLUDE_REG) {
6300 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6301 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6302 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6303 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6304 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6305 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6306 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6307 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6308 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6309 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6310 }
6311 }
6312 // Deal with changed mappings
6313 if(i<iend) {
6314 for(r=0;r<HOST_REGS;r++) {
6315 if(r!=EXCLUDE_REG) {
6316 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
6317 temp_will_dirty&=~(1<<r);
6318 temp_wont_dirty&=~(1<<r);
6319 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6320 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6321 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6322 } else {
6323 temp_will_dirty|=1<<r;
6324 temp_wont_dirty|=1<<r;
6325 }
6326 }
6327 }
6328 }
6329 }
6330 if(wr) {
6331 will_dirty[i]=temp_will_dirty;
6332 wont_dirty[i]=temp_wont_dirty;
6333 clean_registers((ba[i]-start)>>2,i-1,0);
6334 }else{
6335 // Limit recursion. It can take an excessive amount
6336 // of time if there are a lot of nested loops.
6337 will_dirty[(ba[i]-start)>>2]=0;
6338 wont_dirty[(ba[i]-start)>>2]=-1;
6339 }
6340 }
6341 /*else*/ if(1)
6342 {
6343 if (dops[i].is_ujump)
6344 {
6345 // Unconditional branch
6346 will_dirty_i=0;
6347 wont_dirty_i=0;
6348 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6349 for(r=0;r<HOST_REGS;r++) {
6350 if(r!=EXCLUDE_REG) {
6351 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6352 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
6353 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6354 }
6355 if(branch_regs[i].regmap[r]>=0) {
6356 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6357 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6358 }
6359 }
6360 }
6361 //}
6362 // Merge in delay slot
6363 for(r=0;r<HOST_REGS;r++) {
6364 if(r!=EXCLUDE_REG) {
6365 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6366 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6367 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6368 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6369 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6370 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6371 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6372 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6373 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6374 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6375 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6376 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6377 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6378 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6379 }
6380 }
6381 } else {
6382 // Conditional branch
6383 will_dirty_i=will_dirty_next;
6384 wont_dirty_i=wont_dirty_next;
6385 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6386 for(r=0;r<HOST_REGS;r++) {
6387 if(r!=EXCLUDE_REG) {
6388 signed char target_reg=branch_regs[i].regmap[r];
6389 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6390 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
6391 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6392 }
6393 else if(target_reg>=0) {
6394 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6395 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6396 }
6397 }
6398 }
6399 //}
6400 // Merge in delay slot
6401 for(r=0;r<HOST_REGS;r++) {
6402 if(r!=EXCLUDE_REG) {
6403 if (1) { // !dops[i].likely) {
6404 // Might not dirty if likely branch is not taken
6405 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6406 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6407 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6408 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6409 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6410 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6411 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6412 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6413 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6414 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6415 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6416 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6417 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6418 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6419 }
6420 }
6421 }
6422 }
6423 // Merge in delay slot (won't dirty)
6424 for(r=0;r<HOST_REGS;r++) {
6425 if(r!=EXCLUDE_REG) {
6426 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6427 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6428 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6429 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6430 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6431 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6432 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6433 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6434 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6435 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6436 }
6437 }
6438 if(wr) {
6439 #ifndef DESTRUCTIVE_WRITEBACK
6440 branch_regs[i].dirty&=wont_dirty_i;
6441 #endif
6442 branch_regs[i].dirty|=will_dirty_i;
6443 }
6444 }
6445 }
6446 }
6447 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6448 {
6449 // SYSCALL instruction (software interrupt)
6450 will_dirty_i=0;
6451 wont_dirty_i=0;
6452 }
6453 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6454 {
6455 // ERET instruction (return from interrupt)
6456 will_dirty_i=0;
6457 wont_dirty_i=0;
6458 }
6459 will_dirty_next=will_dirty_i;
6460 wont_dirty_next=wont_dirty_i;
6461 for(r=0;r<HOST_REGS;r++) {
6462 if(r!=EXCLUDE_REG) {
6463 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6464 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6465 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6466 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6467 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6468 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6469 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6470 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6471 if(i>istart) {
6472 if (!dops[i].is_jump)
6473 {
6474 // Don't store a register immediately after writing it,
6475 // may prevent dual-issue.
6476 if((regs[i].regmap[r]&63)==dops[i-1].rt1) wont_dirty_i|=1<<r;
6477 if((regs[i].regmap[r]&63)==dops[i-1].rt2) wont_dirty_i|=1<<r;
6478 }
6479 }
6480 }
6481 }
6482 // Save it
6483 will_dirty[i]=will_dirty_i;
6484 wont_dirty[i]=wont_dirty_i;
6485 // Mark registers that won't be dirtied as not dirty
6486 if(wr) {
6487 regs[i].dirty|=will_dirty_i;
6488 #ifndef DESTRUCTIVE_WRITEBACK
6489 regs[i].dirty&=wont_dirty_i;
6490 if(dops[i].is_jump)
6491 {
6492 if (i < iend-1 && !dops[i].is_ujump) {
6493 for(r=0;r<HOST_REGS;r++) {
6494 if(r!=EXCLUDE_REG) {
6495 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
6496 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
6497 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6498 }
6499 }
6500 }
6501 }
6502 else
6503 {
6504 if(i<iend) {
6505 for(r=0;r<HOST_REGS;r++) {
6506 if(r!=EXCLUDE_REG) {
6507 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
6508 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
6509 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6510 }
6511 }
6512 }
6513 }
6514 #endif
6515 //}
6516 }
6517 // Deal with changed mappings
6518 temp_will_dirty=will_dirty_i;
6519 temp_wont_dirty=wont_dirty_i;
6520 for(r=0;r<HOST_REGS;r++) {
6521 if(r!=EXCLUDE_REG) {
6522 int nr;
6523 if(regs[i].regmap[r]==regmap_pre[i][r]) {
6524 if(wr) {
6525 #ifndef DESTRUCTIVE_WRITEBACK
6526 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6527 #endif
6528 regs[i].wasdirty|=will_dirty_i&(1<<r);
6529 }
6530 }
6531 else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
6532 // Register moved to a different register
6533 will_dirty_i&=~(1<<r);
6534 wont_dirty_i&=~(1<<r);
6535 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
6536 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
6537 if(wr) {
6538 #ifndef DESTRUCTIVE_WRITEBACK
6539 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6540 #endif
6541 regs[i].wasdirty|=will_dirty_i&(1<<r);
6542 }
6543 }
6544 else {
6545 will_dirty_i&=~(1<<r);
6546 wont_dirty_i&=~(1<<r);
6547 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6548 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6549 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6550 } else {
6551 wont_dirty_i|=1<<r;
6552 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
6553 }
6554 }
6555 }
6556 }
6557 }
6558}
6559
6560#ifdef DISASM
6561 /* disassembly */
6562void disassemble_inst(int i)
6563{
6564 if (dops[i].bt) printf("*"); else printf(" ");
6565 switch(dops[i].itype) {
6566 case UJUMP:
6567 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6568 case CJUMP:
6569 printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):*ba);break;
6570 case SJUMP:
6571 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;
6572 case RJUMP:
6573 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6574 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6575 else
6576 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6577 break;
6578 case SPAN:
6579 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,ba[i]);break;
6580 case IMM16:
6581 if(dops[i].opcode==0xf) //LUI
6582 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6583 else
6584 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6585 break;
6586 case LOAD:
6587 case LOADLR:
6588 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6589 break;
6590 case STORE:
6591 case STORELR:
6592 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6593 break;
6594 case ALU:
6595 case SHIFT:
6596 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6597 break;
6598 case MULTDIV:
6599 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6600 break;
6601 case SHIFTIMM:
6602 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6603 break;
6604 case MOV:
6605 if((dops[i].opcode2&0x1d)==0x10)
6606 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6607 else if((dops[i].opcode2&0x1d)==0x11)
6608 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6609 else
6610 printf (" %x: %s\n",start+i*4,insn[i]);
6611 break;
6612 case COP0:
6613 if(dops[i].opcode2==0)
6614 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6615 else if(dops[i].opcode2==4)
6616 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6617 else printf (" %x: %s\n",start+i*4,insn[i]);
6618 break;
6619 case COP1:
6620 if(dops[i].opcode2<3)
6621 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6622 else if(dops[i].opcode2>3)
6623 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6624 else printf (" %x: %s\n",start+i*4,insn[i]);
6625 break;
6626 case COP2:
6627 if(dops[i].opcode2<3)
6628 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6629 else if(dops[i].opcode2>3)
6630 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6631 else printf (" %x: %s\n",start+i*4,insn[i]);
6632 break;
6633 case C1LS:
6634 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6635 break;
6636 case C2LS:
6637 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6638 break;
6639 case INTCALL:
6640 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6641 break;
6642 default:
6643 //printf (" %s %8x\n",insn[i],source[i]);
6644 printf (" %x: %s\n",start+i*4,insn[i]);
6645 }
6646}
6647#else
6648static void disassemble_inst(int i) {}
6649#endif // DISASM
6650
6651#define DRC_TEST_VAL 0x74657374
6652
6653static void new_dynarec_test(void)
6654{
6655 int (*testfunc)(void);
6656 void *beginning;
6657 int ret[2];
6658 size_t i;
6659
6660 // check structure linkage
6661 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6662 {
6663 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6664 }
6665
6666 SysPrintf("testing if we can run recompiled code...\n");
6667 ((volatile u_int *)out)[0]++; // make cache dirty
6668
6669 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6670 out = ndrc->translation_cache;
6671 beginning = start_block();
6672 emit_movimm(DRC_TEST_VAL + i, 0); // test
6673 emit_ret();
6674 literal_pool(0);
6675 end_block(beginning);
6676 testfunc = beginning;
6677 ret[i] = testfunc();
6678 }
6679
6680 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6681 SysPrintf("test passed.\n");
6682 else
6683 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6684 out = ndrc->translation_cache;
6685}
6686
6687// clear the state completely, instead of just marking
6688// things invalid like invalidate_all_pages() does
6689void new_dynarec_clear_full(void)
6690{
6691 int n;
6692 out = ndrc->translation_cache;
6693 memset(invalid_code,1,sizeof(invalid_code));
6694 memset(hash_table,0xff,sizeof(hash_table));
6695 memset(mini_ht,-1,sizeof(mini_ht));
6696 memset(restore_candidate,0,sizeof(restore_candidate));
6697 memset(shadow,0,sizeof(shadow));
6698 copy=shadow;
6699 expirep=16384; // Expiry pointer, +2 blocks
6700 pending_exception=0;
6701 literalcount=0;
6702 stop_after_jal=0;
6703 inv_code_start=inv_code_end=~0;
6704 f1_hack=0;
6705 // TLB
6706 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6707 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6708 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6709
6710 cycle_multiplier_old = cycle_multiplier;
6711 new_dynarec_hacks_old = new_dynarec_hacks;
6712}
6713
6714void new_dynarec_init(void)
6715{
6716 SysPrintf("Init new dynarec\n");
6717
6718#ifdef BASE_ADDR_DYNAMIC
6719 #ifdef VITA
6720 sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
6721 if (sceBlock < 0)
6722 SysPrintf("sceKernelAllocMemBlockForVM failed\n");
6723 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6724 if (ret < 0)
6725 SysPrintf("sceKernelGetMemBlockBase failed\n");
6726 #else
6727 uintptr_t desired_addr = 0;
6728 #ifdef __ELF__
6729 extern char _end;
6730 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6731 #endif
6732 ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
6733 PROT_READ | PROT_WRITE | PROT_EXEC,
6734 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
6735 if (ndrc == MAP_FAILED) {
6736 SysPrintf("mmap() failed: %s\n", strerror(errno));
6737 abort();
6738 }
6739 #endif
6740#else
6741 #ifndef NO_WRITE_EXEC
6742 // not all systems allow execute in data segment by default
6743 if (mprotect(ndrc, sizeof(ndrc->translation_cache) + sizeof(ndrc->tramp.ops),
6744 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6745 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6746 #endif
6747#endif
6748 out = ndrc->translation_cache;
6749 cycle_multiplier=200;
6750 new_dynarec_clear_full();
6751#ifdef HOST_IMM8
6752 // Copy this into local area so we don't have to put it in every literal pool
6753 invc_ptr=invalid_code;
6754#endif
6755 arch_init();
6756 new_dynarec_test();
6757#ifndef RAM_FIXED
6758 ram_offset=(uintptr_t)rdram-0x80000000;
6759#endif
6760 if (ram_offset!=0)
6761 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6762}
6763
6764void new_dynarec_cleanup(void)
6765{
6766 int n;
6767#ifdef BASE_ADDR_DYNAMIC
6768 #ifdef VITA
6769 sceKernelFreeMemBlock(sceBlock);
6770 sceBlock = -1;
6771 #else
6772 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6773 SysPrintf("munmap() failed\n");
6774 #endif
6775#endif
6776 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6777 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6778 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6779 #ifdef ROM_COPY
6780 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
6781 #endif
6782}
6783
6784static u_int *get_source_start(u_int addr, u_int *limit)
6785{
6786 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6787 cycle_multiplier_override = 0;
6788
6789 if (addr < 0x00200000 ||
6790 (0xa0000000 <= addr && addr < 0xa0200000))
6791 {
6792 // used for BIOS calls mostly?
6793 *limit = (addr&0xa0000000)|0x00200000;
6794 return (u_int *)(rdram + (addr&0x1fffff));
6795 }
6796 else if (!Config.HLE && (
6797 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6798 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6799 {
6800 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6801 // but timings in PCSX are too tied to the interpreter's BIAS
6802 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6803 cycle_multiplier_override = 200;
6804
6805 *limit = (addr & 0xfff00000) | 0x80000;
6806 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6807 }
6808 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6809 *limit = (addr & 0x80600000) + 0x00200000;
6810 return (u_int *)(rdram + (addr&0x1fffff));
6811 }
6812 return NULL;
6813}
6814
6815static u_int scan_for_ret(u_int addr)
6816{
6817 u_int limit = 0;
6818 u_int *mem;
6819
6820 mem = get_source_start(addr, &limit);
6821 if (mem == NULL)
6822 return addr;
6823
6824 if (limit > addr + 0x1000)
6825 limit = addr + 0x1000;
6826 for (; addr < limit; addr += 4, mem++) {
6827 if (*mem == 0x03e00008) // jr $ra
6828 return addr + 8;
6829 }
6830 return addr;
6831}
6832
6833struct savestate_block {
6834 uint32_t addr;
6835 uint32_t regflags;
6836};
6837
6838static int addr_cmp(const void *p1_, const void *p2_)
6839{
6840 const struct savestate_block *p1 = p1_, *p2 = p2_;
6841 return p1->addr - p2->addr;
6842}
6843
6844int new_dynarec_save_blocks(void *save, int size)
6845{
6846 struct savestate_block *blocks = save;
6847 int maxcount = size / sizeof(blocks[0]);
6848 struct savestate_block tmp_blocks[1024];
6849 struct ll_entry *head;
6850 int p, s, d, o, bcnt;
6851 u_int addr;
6852
6853 o = 0;
6854 for (p = 0; p < ARRAY_SIZE(jump_in); p++) {
6855 bcnt = 0;
6856 for (head = jump_in[p]; head != NULL; head = head->next) {
6857 tmp_blocks[bcnt].addr = head->vaddr;
6858 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
6859 bcnt++;
6860 }
6861 if (bcnt < 1)
6862 continue;
6863 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6864
6865 addr = tmp_blocks[0].addr;
6866 for (s = d = 0; s < bcnt; s++) {
6867 if (tmp_blocks[s].addr < addr)
6868 continue;
6869 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6870 tmp_blocks[d++] = tmp_blocks[s];
6871 addr = scan_for_ret(tmp_blocks[s].addr);
6872 }
6873
6874 if (o + d > maxcount)
6875 d = maxcount - o;
6876 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
6877 o += d;
6878 }
6879
6880 return o * sizeof(blocks[0]);
6881}
6882
6883void new_dynarec_load_blocks(const void *save, int size)
6884{
6885 const struct savestate_block *blocks = save;
6886 int count = size / sizeof(blocks[0]);
6887 u_int regs_save[32];
6888 uint32_t f;
6889 int i, b;
6890
6891 get_addr(psxRegs.pc);
6892
6893 // change GPRs for speculation to at least partially work..
6894 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6895 for (i = 1; i < 32; i++)
6896 psxRegs.GPR.r[i] = 0x80000000;
6897
6898 for (b = 0; b < count; b++) {
6899 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6900 if (f & 1)
6901 psxRegs.GPR.r[i] = 0x1f800000;
6902 }
6903
6904 get_addr(blocks[b].addr);
6905
6906 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6907 if (f & 1)
6908 psxRegs.GPR.r[i] = 0x80000000;
6909 }
6910 }
6911
6912 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6913}
6914
6915int new_recompile_block(u_int addr)
6916{
6917 u_int pagelimit = 0;
6918 u_int state_rflags = 0;
6919 int i;
6920
6921 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
6922 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
6923 //if(debug)
6924 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
6925
6926 // this is just for speculation
6927 for (i = 1; i < 32; i++) {
6928 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
6929 state_rflags |= 1 << i;
6930 }
6931
6932 start = (u_int)addr&~3;
6933 //assert(((u_int)addr&1)==0); // start-in-delay-slot flag
6934 new_dynarec_did_compile=1;
6935 if (Config.HLE && start == 0x80001000) // hlecall
6936 {
6937 // XXX: is this enough? Maybe check hleSoftCall?
6938 void *beginning=start_block();
6939 u_int page=get_page(start);
6940
6941 invalid_code[start>>12]=0;
6942 emit_movimm(start,0);
6943 emit_writeword(0,&pcaddr);
6944 emit_far_jump(new_dyna_leave);
6945 literal_pool(0);
6946 end_block(beginning);
6947 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
6948 return 0;
6949 }
6950 else if (f1_hack == ~0u || (f1_hack != 0 && start == f1_hack)) {
6951 void *beginning = start_block();
6952 u_int page = get_page(start);
6953 emit_readword(&psxRegs.GPR.n.sp, 0);
6954 emit_readptr(&mem_rtab, 1);
6955 emit_shrimm(0, 12, 2);
6956 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
6957 emit_addimm(0, 0x18, 0);
6958 emit_adds_ptr(1, 1, 1);
6959 emit_ldr_dualindexed(1, 0, 0);
6960 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
6961 emit_far_call(get_addr_ht);
6962 emit_jmpreg(0); // jr k0
6963 literal_pool(0);
6964 end_block(beginning);
6965
6966 ll_add_flags(jump_in + page, start, state_rflags, beginning);
6967 SysPrintf("F1 hack to %08x\n", start);
6968 f1_hack = start;
6969 return 0;
6970 }
6971
6972 source = get_source_start(start, &pagelimit);
6973 if (source == NULL) {
6974 SysPrintf("Compile at bogus memory address: %08x\n", addr);
6975 abort();
6976 }
6977
6978 /* Pass 1: disassemble */
6979 /* Pass 2: register dependencies, branch targets */
6980 /* Pass 3: register allocation */
6981 /* Pass 4: branch dependencies */
6982 /* Pass 5: pre-alloc */
6983 /* Pass 6: optimize clean/dirty state */
6984 /* Pass 7: flag 32-bit registers */
6985 /* Pass 8: assembly */
6986 /* Pass 9: linker */
6987 /* Pass 10: garbage collection / free memory */
6988
6989 int j;
6990 int done=0;
6991 unsigned int type,op,op2;
6992
6993 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
6994
6995 /* Pass 1 disassembly */
6996
6997 for(i=0;!done;i++) {
6998 dops[i].bt=0;
6999 dops[i].ooo=0;
7000 op2=0;
7001 minimum_free_regs[i]=0;
7002 dops[i].opcode=op=source[i]>>26;
7003 switch(op)
7004 {
7005 case 0x00: strcpy(insn[i],"special"); type=NI;
7006 op2=source[i]&0x3f;
7007 switch(op2)
7008 {
7009 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
7010 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
7011 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
7012 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
7013 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
7014 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
7015 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
7016 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
7017 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
7018 case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
7019 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
7020 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
7021 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
7022 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
7023 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
7024 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
7025 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
7026 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
7027 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
7028 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
7029 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
7030 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
7031 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
7032 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
7033 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
7034 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
7035 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
7036 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
7037 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
7038 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
7039 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
7040 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
7041 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
7042 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
7043 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
7044#if 0
7045 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
7046 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
7047 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
7048 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
7049 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
7050 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
7051 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
7052 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
7053 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
7054 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
7055 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
7056 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
7057 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
7058 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
7059 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
7060 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
7061 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
7062#endif
7063 }
7064 break;
7065 case 0x01: strcpy(insn[i],"regimm"); type=NI;
7066 op2=(source[i]>>16)&0x1f;
7067 switch(op2)
7068 {
7069 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
7070 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
7071 //case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
7072 //case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
7073 //case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
7074 //case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
7075 //case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
7076 //case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
7077 //case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
7078 //case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
7079 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
7080 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
7081 //case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
7082 //case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
7083 }
7084 break;
7085 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
7086 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
7087 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
7088 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
7089 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
7090 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
7091 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
7092 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
7093 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
7094 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
7095 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
7096 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
7097 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
7098 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
7099 case 0x10: strcpy(insn[i],"cop0"); type=NI;
7100 op2=(source[i]>>21)&0x1f;
7101 switch(op2)
7102 {
7103 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
7104 case 0x02: strcpy(insn[i],"CFC0"); type=COP0; break;
7105 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
7106 case 0x06: strcpy(insn[i],"CTC0"); type=COP0; break;
7107 case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
7108 }
7109 break;
7110 case 0x11: strcpy(insn[i],"cop1"); type=COP1;
7111 op2=(source[i]>>21)&0x1f;
7112 break;
7113#if 0
7114 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
7115 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
7116 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
7117 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
7118 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
7119 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
7120 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
7121 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
7122#endif
7123 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
7124 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
7125 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
7126 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
7127 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
7128 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
7129 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
7130#if 0
7131 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
7132#endif
7133 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
7134 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
7135 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
7136 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
7137#if 0
7138 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
7139 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
7140#endif
7141 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
7142 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
7143 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
7144 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
7145#if 0
7146 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
7147 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
7148 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
7149#endif
7150 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
7151 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
7152#if 0
7153 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
7154 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
7155 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
7156#endif
7157 case 0x12: strcpy(insn[i],"COP2"); type=NI;
7158 op2=(source[i]>>21)&0x1f;
7159 //if (op2 & 0x10)
7160 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
7161 if (gte_handlers[source[i]&0x3f]!=NULL) {
7162 if (gte_regnames[source[i]&0x3f]!=NULL)
7163 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
7164 else
7165 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
7166 type=C2OP;
7167 }
7168 }
7169 else switch(op2)
7170 {
7171 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
7172 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
7173 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
7174 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
7175 }
7176 break;
7177 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
7178 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
7179 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
7180 default: strcpy(insn[i],"???"); type=NI;
7181 SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
7182 break;
7183 }
7184 dops[i].itype=type;
7185 dops[i].opcode2=op2;
7186 /* Get registers/immediates */
7187 dops[i].lt1=0;
7188 gte_rs[i]=gte_rt[i]=0;
7189 switch(type) {
7190 case LOAD:
7191 dops[i].rs1=(source[i]>>21)&0x1f;
7192 dops[i].rs2=0;
7193 dops[i].rt1=(source[i]>>16)&0x1f;
7194 dops[i].rt2=0;
7195 imm[i]=(short)source[i];
7196 break;
7197 case STORE:
7198 case STORELR:
7199 dops[i].rs1=(source[i]>>21)&0x1f;
7200 dops[i].rs2=(source[i]>>16)&0x1f;
7201 dops[i].rt1=0;
7202 dops[i].rt2=0;
7203 imm[i]=(short)source[i];
7204 break;
7205 case LOADLR:
7206 // LWL/LWR only load part of the register,
7207 // therefore the target register must be treated as a source too
7208 dops[i].rs1=(source[i]>>21)&0x1f;
7209 dops[i].rs2=(source[i]>>16)&0x1f;
7210 dops[i].rt1=(source[i]>>16)&0x1f;
7211 dops[i].rt2=0;
7212 imm[i]=(short)source[i];
7213 break;
7214 case IMM16:
7215 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
7216 else dops[i].rs1=(source[i]>>21)&0x1f;
7217 dops[i].rs2=0;
7218 dops[i].rt1=(source[i]>>16)&0x1f;
7219 dops[i].rt2=0;
7220 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
7221 imm[i]=(unsigned short)source[i];
7222 }else{
7223 imm[i]=(short)source[i];
7224 }
7225 break;
7226 case UJUMP:
7227 dops[i].rs1=0;
7228 dops[i].rs2=0;
7229 dops[i].rt1=0;
7230 dops[i].rt2=0;
7231 // The JAL instruction writes to r31.
7232 if (op&1) {
7233 dops[i].rt1=31;
7234 }
7235 dops[i].rs2=CCREG;
7236 break;
7237 case RJUMP:
7238 dops[i].rs1=(source[i]>>21)&0x1f;
7239 dops[i].rs2=0;
7240 dops[i].rt1=0;
7241 dops[i].rt2=0;
7242 // The JALR instruction writes to rd.
7243 if (op2&1) {
7244 dops[i].rt1=(source[i]>>11)&0x1f;
7245 }
7246 dops[i].rs2=CCREG;
7247 break;
7248 case CJUMP:
7249 dops[i].rs1=(source[i]>>21)&0x1f;
7250 dops[i].rs2=(source[i]>>16)&0x1f;
7251 dops[i].rt1=0;
7252 dops[i].rt2=0;
7253 if(op&2) { // BGTZ/BLEZ
7254 dops[i].rs2=0;
7255 }
7256 break;
7257 case SJUMP:
7258 dops[i].rs1=(source[i]>>21)&0x1f;
7259 dops[i].rs2=CCREG;
7260 dops[i].rt1=0;
7261 dops[i].rt2=0;
7262 if(op2&0x10) { // BxxAL
7263 dops[i].rt1=31;
7264 // NOTE: If the branch is not taken, r31 is still overwritten
7265 }
7266 break;
7267 case ALU:
7268 dops[i].rs1=(source[i]>>21)&0x1f; // source
7269 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
7270 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7271 dops[i].rt2=0;
7272 break;
7273 case MULTDIV:
7274 dops[i].rs1=(source[i]>>21)&0x1f; // source
7275 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
7276 dops[i].rt1=HIREG;
7277 dops[i].rt2=LOREG;
7278 break;
7279 case MOV:
7280 dops[i].rs1=0;
7281 dops[i].rs2=0;
7282 dops[i].rt1=0;
7283 dops[i].rt2=0;
7284 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
7285 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
7286 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
7287 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
7288 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
7289 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
7290 break;
7291 case SHIFT:
7292 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
7293 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
7294 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7295 dops[i].rt2=0;
7296 break;
7297 case SHIFTIMM:
7298 dops[i].rs1=(source[i]>>16)&0x1f;
7299 dops[i].rs2=0;
7300 dops[i].rt1=(source[i]>>11)&0x1f;
7301 dops[i].rt2=0;
7302 imm[i]=(source[i]>>6)&0x1f;
7303 // DSxx32 instructions
7304 if(op2>=0x3c) imm[i]|=0x20;
7305 break;
7306 case COP0:
7307 dops[i].rs1=0;
7308 dops[i].rs2=0;
7309 dops[i].rt1=0;
7310 dops[i].rt2=0;
7311 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
7312 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
7313 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
7314 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
7315 break;
7316 case COP1:
7317 dops[i].rs1=0;
7318 dops[i].rs2=0;
7319 dops[i].rt1=0;
7320 dops[i].rt2=0;
7321 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
7322 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
7323 dops[i].rs2=CSREG;
7324 break;
7325 case COP2:
7326 dops[i].rs1=0;
7327 dops[i].rs2=0;
7328 dops[i].rt1=0;
7329 dops[i].rt2=0;
7330 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
7331 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
7332 dops[i].rs2=CSREG;
7333 int gr=(source[i]>>11)&0x1F;
7334 switch(op2)
7335 {
7336 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
7337 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
7338 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
7339 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
7340 }
7341 break;
7342 case C1LS:
7343 dops[i].rs1=(source[i]>>21)&0x1F;
7344 dops[i].rs2=CSREG;
7345 dops[i].rt1=0;
7346 dops[i].rt2=0;
7347 imm[i]=(short)source[i];
7348 break;
7349 case C2LS:
7350 dops[i].rs1=(source[i]>>21)&0x1F;
7351 dops[i].rs2=0;
7352 dops[i].rt1=0;
7353 dops[i].rt2=0;
7354 imm[i]=(short)source[i];
7355 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
7356 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
7357 break;
7358 case C2OP:
7359 dops[i].rs1=0;
7360 dops[i].rs2=0;
7361 dops[i].rt1=0;
7362 dops[i].rt2=0;
7363 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
7364 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
7365 gte_rt[i]|=1ll<<63; // every op changes flags
7366 if((source[i]&0x3f)==GTE_MVMVA) {
7367 int v = (source[i] >> 15) & 3;
7368 gte_rs[i]&=~0xe3fll;
7369 if(v==3) gte_rs[i]|=0xe00ll;
7370 else gte_rs[i]|=3ll<<(v*2);
7371 }
7372 break;
7373 case SYSCALL:
7374 case HLECALL:
7375 case INTCALL:
7376 dops[i].rs1=CCREG;
7377 dops[i].rs2=0;
7378 dops[i].rt1=0;
7379 dops[i].rt2=0;
7380 break;
7381 default:
7382 dops[i].rs1=0;
7383 dops[i].rs2=0;
7384 dops[i].rt1=0;
7385 dops[i].rt2=0;
7386 }
7387 /* Calculate branch target addresses */
7388 if(type==UJUMP)
7389 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7390 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
7391 ba[i]=start+i*4+8; // Ignore never taken branch
7392 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
7393 ba[i]=start+i*4+8; // Ignore never taken branch
7394 else if(type==CJUMP||type==SJUMP)
7395 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7396 else ba[i]=-1;
7397
7398 /* simplify always (not)taken branches */
7399 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
7400 dops[i].rs1 = dops[i].rs2 = 0;
7401 if (!(op & 1)) {
7402 dops[i].itype = type = UJUMP;
7403 dops[i].rs2 = CCREG;
7404 }
7405 }
7406 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
7407 dops[i].itype = type = UJUMP;
7408
7409 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
7410 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
7411
7412 /* messy cases to just pass over to the interpreter */
7413 if (i > 0 && dops[i-1].is_jump) {
7414 int do_in_intrp=0;
7415 // branch in delay slot?
7416 if (dops[i].is_jump) {
7417 // don't handle first branch and call interpreter if it's hit
7418 SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
7419 do_in_intrp=1;
7420 }
7421 // basic load delay detection
7422 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
7423 int t=(ba[i-1]-start)/4;
7424 if(0 <= t && t < i &&(dops[i].rt1==dops[t].rs1||dops[i].rt1==dops[t].rs2)&&dops[t].itype!=CJUMP&&dops[t].itype!=SJUMP) {
7425 // jump target wants DS result - potential load delay effect
7426 SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
7427 do_in_intrp=1;
7428 dops[t+1].bt=1; // expected return from interpreter
7429 }
7430 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&&
7431 !(i>=3&&dops[i-3].is_jump)) {
7432 // v0 overwrite like this is a sign of trouble, bail out
7433 SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
7434 do_in_intrp=1;
7435 }
7436 }
7437 if(do_in_intrp) {
7438 dops[i-1].rs1=CCREG;
7439 dops[i-1].rs2=dops[i-1].rt1=dops[i-1].rt2=0;
7440 ba[i-1]=-1;
7441 dops[i-1].itype=INTCALL;
7442 done=2;
7443 i--; // don't compile the DS
7444 }
7445 }
7446
7447 /* Is this the end of the block? */
7448 if (i > 0 && dops[i-1].is_ujump) {
7449 if(dops[i-1].rt1==0) { // Continue past subroutine call (JAL)
7450 done=2;
7451 }
7452 else {
7453 if(stop_after_jal) done=1;
7454 // Stop on BREAK
7455 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7456 }
7457 // Don't recompile stuff that's already compiled
7458 if(check_addr(start+i*4+4)) done=1;
7459 // Don't get too close to the limit
7460 if(i>MAXBLOCK/2) done=1;
7461 }
7462 if(dops[i].itype==SYSCALL&&stop_after_jal) done=1;
7463 if(dops[i].itype==HLECALL||dops[i].itype==INTCALL) done=2;
7464 if(done==2) {
7465 // Does the block continue due to a branch?
7466 for(j=i-1;j>=0;j--)
7467 {
7468 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7469 if(ba[j]==start+i*4+4) done=j=0;
7470 if(ba[j]==start+i*4+8) done=j=0;
7471 }
7472 }
7473 //assert(i<MAXBLOCK-1);
7474 if(start+i*4==pagelimit-4) done=1;
7475 assert(start+i*4<pagelimit);
7476 if (i==MAXBLOCK-1) done=1;
7477 // Stop if we're compiling junk
7478 if(dops[i].itype==NI&&dops[i].opcode==0x11) {
7479 done=stop_after_jal=1;
7480 SysPrintf("Disabled speculative precompilation\n");
7481 }
7482 }
7483 slen=i;
7484 if (dops[i-1].is_jump) {
7485 if(start+i*4==pagelimit) {
7486 dops[i-1].itype=SPAN;
7487 }
7488 }
7489 assert(slen>0);
7490
7491 /* spacial hack(s) */
7492 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
7493 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
7494 && dops[i-7].itype == STORE)
7495 {
7496 i = i-8;
7497 if (dops[i].itype == IMM16)
7498 i--;
7499 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
7500 if (dops[i].itype == STORELR && dops[i].rs1 == 6
7501 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
7502 {
7503 SysPrintf("F1 hack from %08x\n", start);
7504 f1_hack = ~0u;
7505 }
7506 }
7507
7508 /* Pass 2 - Register dependencies and branch targets */
7509
7510 unneeded_registers(0,slen-1,0);
7511
7512 /* Pass 3 - Register allocation */
7513
7514 struct regstat current; // Current register allocations/status
7515 current.dirty=0;
7516 current.u=unneeded_reg[0];
7517 clear_all_regs(current.regmap);
7518 alloc_reg(&current,0,CCREG);
7519 dirty_reg(&current,CCREG);
7520 current.isconst=0;
7521 current.wasconst=0;
7522 current.waswritten=0;
7523 int ds=0;
7524 int cc=0;
7525 int hr=-1;
7526
7527 if((u_int)addr&1) {
7528 // First instruction is delay slot
7529 cc=-1;
7530 dops[1].bt=1;
7531 ds=1;
7532 unneeded_reg[0]=1;
7533 current.regmap[HOST_BTREG]=BTREG;
7534 }
7535
7536 for(i=0;i<slen;i++)
7537 {
7538 if(dops[i].bt)
7539 {
7540 int hr;
7541 for(hr=0;hr<HOST_REGS;hr++)
7542 {
7543 // Is this really necessary?
7544 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7545 }
7546 current.isconst=0;
7547 current.waswritten=0;
7548 }
7549
7550 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7551 regs[i].wasconst=current.isconst;
7552 regs[i].wasdirty=current.dirty;
7553 regs[i].loadedconst=0;
7554 if (!dops[i].is_jump) {
7555 if(i+1<slen) {
7556 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7557 current.u|=1;
7558 } else {
7559 current.u=1;
7560 }
7561 } else {
7562 if(i+1<slen) {
7563 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7564 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7565 current.u|=1;
7566 } else { SysPrintf("oops, branch at end of block with no delay slot\n");abort(); }
7567 }
7568 dops[i].is_ds=ds;
7569 if(ds) {
7570 ds=0; // Skip delay slot, already allocated as part of branch
7571 // ...but we need to alloc it in case something jumps here
7572 if(i+1<slen) {
7573 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7574 }else{
7575 current.u=branch_unneeded_reg[i-1];
7576 }
7577 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7578 current.u|=1;
7579 struct regstat temp;
7580 memcpy(&temp,&current,sizeof(current));
7581 temp.wasdirty=temp.dirty;
7582 // TODO: Take into account unconditional branches, as below
7583 delayslot_alloc(&temp,i);
7584 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7585 regs[i].wasdirty=temp.wasdirty;
7586 regs[i].dirty=temp.dirty;
7587 regs[i].isconst=0;
7588 regs[i].wasconst=0;
7589 current.isconst=0;
7590 // Create entry (branch target) regmap
7591 for(hr=0;hr<HOST_REGS;hr++)
7592 {
7593 int r=temp.regmap[hr];
7594 if(r>=0) {
7595 if(r!=regmap_pre[i][hr]) {
7596 regs[i].regmap_entry[hr]=-1;
7597 }
7598 else
7599 {
7600 assert(r < 64);
7601 if((current.u>>r)&1) {
7602 regs[i].regmap_entry[hr]=-1;
7603 regs[i].regmap[hr]=-1;
7604 //Don't clear regs in the delay slot as the branch might need them
7605 //current.regmap[hr]=-1;
7606 }else
7607 regs[i].regmap_entry[hr]=r;
7608 }
7609 } else {
7610 // First instruction expects CCREG to be allocated
7611 if(i==0&&hr==HOST_CCREG)
7612 regs[i].regmap_entry[hr]=CCREG;
7613 else
7614 regs[i].regmap_entry[hr]=-1;
7615 }
7616 }
7617 }
7618 else { // Not delay slot
7619 switch(dops[i].itype) {
7620 case UJUMP:
7621 //current.isconst=0; // DEBUG
7622 //current.wasconst=0; // DEBUG
7623 //regs[i].wasconst=0; // DEBUG
7624 clear_const(&current,dops[i].rt1);
7625 alloc_cc(&current,i);
7626 dirty_reg(&current,CCREG);
7627 if (dops[i].rt1==31) {
7628 alloc_reg(&current,i,31);
7629 dirty_reg(&current,31);
7630 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7631 //assert(dops[i+1].rt1!=dops[i].rt1);
7632 #ifdef REG_PREFETCH
7633 alloc_reg(&current,i,PTEMP);
7634 #endif
7635 }
7636 dops[i].ooo=1;
7637 delayslot_alloc(&current,i+1);
7638 //current.isconst=0; // DEBUG
7639 ds=1;
7640 //printf("i=%d, isconst=%x\n",i,current.isconst);
7641 break;
7642 case RJUMP:
7643 //current.isconst=0;
7644 //current.wasconst=0;
7645 //regs[i].wasconst=0;
7646 clear_const(&current,dops[i].rs1);
7647 clear_const(&current,dops[i].rt1);
7648 alloc_cc(&current,i);
7649 dirty_reg(&current,CCREG);
7650 if (!ds_writes_rjump_rs(i)) {
7651 alloc_reg(&current,i,dops[i].rs1);
7652 if (dops[i].rt1!=0) {
7653 alloc_reg(&current,i,dops[i].rt1);
7654 dirty_reg(&current,dops[i].rt1);
7655 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7656 assert(dops[i+1].rt1!=dops[i].rt1);
7657 #ifdef REG_PREFETCH
7658 alloc_reg(&current,i,PTEMP);
7659 #endif
7660 }
7661 #ifdef USE_MINI_HT
7662 if(dops[i].rs1==31) { // JALR
7663 alloc_reg(&current,i,RHASH);
7664 alloc_reg(&current,i,RHTBL);
7665 }
7666 #endif
7667 delayslot_alloc(&current,i+1);
7668 } else {
7669 // The delay slot overwrites our source register,
7670 // allocate a temporary register to hold the old value.
7671 current.isconst=0;
7672 current.wasconst=0;
7673 regs[i].wasconst=0;
7674 delayslot_alloc(&current,i+1);
7675 current.isconst=0;
7676 alloc_reg(&current,i,RTEMP);
7677 }
7678 //current.isconst=0; // DEBUG
7679 dops[i].ooo=1;
7680 ds=1;
7681 break;
7682 case CJUMP:
7683 //current.isconst=0;
7684 //current.wasconst=0;
7685 //regs[i].wasconst=0;
7686 clear_const(&current,dops[i].rs1);
7687 clear_const(&current,dops[i].rs2);
7688 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7689 {
7690 alloc_cc(&current,i);
7691 dirty_reg(&current,CCREG);
7692 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7693 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7694 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7695 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7696 // The delay slot overwrites one of our conditions.
7697 // Allocate the branch condition registers instead.
7698 current.isconst=0;
7699 current.wasconst=0;
7700 regs[i].wasconst=0;
7701 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7702 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7703 }
7704 else
7705 {
7706 dops[i].ooo=1;
7707 delayslot_alloc(&current,i+1);
7708 }
7709 }
7710 else
7711 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7712 {
7713 alloc_cc(&current,i);
7714 dirty_reg(&current,CCREG);
7715 alloc_reg(&current,i,dops[i].rs1);
7716 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7717 // The delay slot overwrites one of our conditions.
7718 // Allocate the branch condition registers instead.
7719 current.isconst=0;
7720 current.wasconst=0;
7721 regs[i].wasconst=0;
7722 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7723 }
7724 else
7725 {
7726 dops[i].ooo=1;
7727 delayslot_alloc(&current,i+1);
7728 }
7729 }
7730 else
7731 // Don't alloc the delay slot yet because we might not execute it
7732 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7733 {
7734 current.isconst=0;
7735 current.wasconst=0;
7736 regs[i].wasconst=0;
7737 alloc_cc(&current,i);
7738 dirty_reg(&current,CCREG);
7739 alloc_reg(&current,i,dops[i].rs1);
7740 alloc_reg(&current,i,dops[i].rs2);
7741 }
7742 else
7743 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7744 {
7745 current.isconst=0;
7746 current.wasconst=0;
7747 regs[i].wasconst=0;
7748 alloc_cc(&current,i);
7749 dirty_reg(&current,CCREG);
7750 alloc_reg(&current,i,dops[i].rs1);
7751 }
7752 ds=1;
7753 //current.isconst=0;
7754 break;
7755 case SJUMP:
7756 //current.isconst=0;
7757 //current.wasconst=0;
7758 //regs[i].wasconst=0;
7759 clear_const(&current,dops[i].rs1);
7760 clear_const(&current,dops[i].rt1);
7761 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7762 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7763 {
7764 alloc_cc(&current,i);
7765 dirty_reg(&current,CCREG);
7766 alloc_reg(&current,i,dops[i].rs1);
7767 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7768 alloc_reg(&current,i,31);
7769 dirty_reg(&current,31);
7770 //#ifdef REG_PREFETCH
7771 //alloc_reg(&current,i,PTEMP);
7772 //#endif
7773 }
7774 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) // The delay slot overwrites the branch condition.
7775 ||(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
7776 // Allocate the branch condition registers instead.
7777 current.isconst=0;
7778 current.wasconst=0;
7779 regs[i].wasconst=0;
7780 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7781 }
7782 else
7783 {
7784 dops[i].ooo=1;
7785 delayslot_alloc(&current,i+1);
7786 }
7787 }
7788 else
7789 // Don't alloc the delay slot yet because we might not execute it
7790 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7791 {
7792 current.isconst=0;
7793 current.wasconst=0;
7794 regs[i].wasconst=0;
7795 alloc_cc(&current,i);
7796 dirty_reg(&current,CCREG);
7797 alloc_reg(&current,i,dops[i].rs1);
7798 }
7799 ds=1;
7800 //current.isconst=0;
7801 break;
7802 case IMM16:
7803 imm16_alloc(&current,i);
7804 break;
7805 case LOAD:
7806 case LOADLR:
7807 load_alloc(&current,i);
7808 break;
7809 case STORE:
7810 case STORELR:
7811 store_alloc(&current,i);
7812 break;
7813 case ALU:
7814 alu_alloc(&current,i);
7815 break;
7816 case SHIFT:
7817 shift_alloc(&current,i);
7818 break;
7819 case MULTDIV:
7820 multdiv_alloc(&current,i);
7821 break;
7822 case SHIFTIMM:
7823 shiftimm_alloc(&current,i);
7824 break;
7825 case MOV:
7826 mov_alloc(&current,i);
7827 break;
7828 case COP0:
7829 cop0_alloc(&current,i);
7830 break;
7831 case COP1:
7832 break;
7833 case COP2:
7834 cop2_alloc(&current,i);
7835 break;
7836 case C1LS:
7837 c1ls_alloc(&current,i);
7838 break;
7839 case C2LS:
7840 c2ls_alloc(&current,i);
7841 break;
7842 case C2OP:
7843 c2op_alloc(&current,i);
7844 break;
7845 case SYSCALL:
7846 case HLECALL:
7847 case INTCALL:
7848 syscall_alloc(&current,i);
7849 break;
7850 case SPAN:
7851 pagespan_alloc(&current,i);
7852 break;
7853 }
7854
7855 // Create entry (branch target) regmap
7856 for(hr=0;hr<HOST_REGS;hr++)
7857 {
7858 int r,or;
7859 r=current.regmap[hr];
7860 if(r>=0) {
7861 if(r!=regmap_pre[i][hr]) {
7862 // TODO: delay slot (?)
7863 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7864 if(or<0||(r&63)>=TEMPREG){
7865 regs[i].regmap_entry[hr]=-1;
7866 }
7867 else
7868 {
7869 // Just move it to a different register
7870 regs[i].regmap_entry[hr]=r;
7871 // If it was dirty before, it's still dirty
7872 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r&63);
7873 }
7874 }
7875 else
7876 {
7877 // Unneeded
7878 if(r==0){
7879 regs[i].regmap_entry[hr]=0;
7880 }
7881 else
7882 {
7883 assert(r<64);
7884 if((current.u>>r)&1) {
7885 regs[i].regmap_entry[hr]=-1;
7886 //regs[i].regmap[hr]=-1;
7887 current.regmap[hr]=-1;
7888 }else
7889 regs[i].regmap_entry[hr]=r;
7890 }
7891 }
7892 } else {
7893 // Branches expect CCREG to be allocated at the target
7894 if(regmap_pre[i][hr]==CCREG)
7895 regs[i].regmap_entry[hr]=CCREG;
7896 else
7897 regs[i].regmap_entry[hr]=-1;
7898 }
7899 }
7900 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7901 }
7902
7903 if(i>0&&(dops[i-1].itype==STORE||dops[i-1].itype==STORELR||(dops[i-1].itype==C2LS&&dops[i-1].opcode==0x3a))&&(u_int)imm[i-1]<0x800)
7904 current.waswritten|=1<<dops[i-1].rs1;
7905 current.waswritten&=~(1<<dops[i].rt1);
7906 current.waswritten&=~(1<<dops[i].rt2);
7907 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
7908 current.waswritten&=~(1<<dops[i].rs1);
7909
7910 /* Branch post-alloc */
7911 if(i>0)
7912 {
7913 current.wasdirty=current.dirty;
7914 switch(dops[i-1].itype) {
7915 case UJUMP:
7916 memcpy(&branch_regs[i-1],&current,sizeof(current));
7917 branch_regs[i-1].isconst=0;
7918 branch_regs[i-1].wasconst=0;
7919 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7920 alloc_cc(&branch_regs[i-1],i-1);
7921 dirty_reg(&branch_regs[i-1],CCREG);
7922 if(dops[i-1].rt1==31) { // JAL
7923 alloc_reg(&branch_regs[i-1],i-1,31);
7924 dirty_reg(&branch_regs[i-1],31);
7925 }
7926 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7927 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7928 break;
7929 case RJUMP:
7930 memcpy(&branch_regs[i-1],&current,sizeof(current));
7931 branch_regs[i-1].isconst=0;
7932 branch_regs[i-1].wasconst=0;
7933 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7934 alloc_cc(&branch_regs[i-1],i-1);
7935 dirty_reg(&branch_regs[i-1],CCREG);
7936 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7937 if(dops[i-1].rt1!=0) { // JALR
7938 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7939 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7940 }
7941 #ifdef USE_MINI_HT
7942 if(dops[i-1].rs1==31) { // JALR
7943 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7944 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7945 }
7946 #endif
7947 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7948 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7949 break;
7950 case CJUMP:
7951 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7952 {
7953 alloc_cc(&current,i-1);
7954 dirty_reg(&current,CCREG);
7955 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7956 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7957 // The delay slot overwrote one of our conditions
7958 // Delay slot goes after the test (in order)
7959 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7960 current.u|=1;
7961 delayslot_alloc(&current,i);
7962 current.isconst=0;
7963 }
7964 else
7965 {
7966 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7967 // Alloc the branch condition registers
7968 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7969 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7970 }
7971 memcpy(&branch_regs[i-1],&current,sizeof(current));
7972 branch_regs[i-1].isconst=0;
7973 branch_regs[i-1].wasconst=0;
7974 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7975 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7976 }
7977 else
7978 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7979 {
7980 alloc_cc(&current,i-1);
7981 dirty_reg(&current,CCREG);
7982 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7983 // The delay slot overwrote the branch condition
7984 // Delay slot goes after the test (in order)
7985 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7986 current.u|=1;
7987 delayslot_alloc(&current,i);
7988 current.isconst=0;
7989 }
7990 else
7991 {
7992 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7993 // Alloc the branch condition register
7994 alloc_reg(&current,i-1,dops[i-1].rs1);
7995 }
7996 memcpy(&branch_regs[i-1],&current,sizeof(current));
7997 branch_regs[i-1].isconst=0;
7998 branch_regs[i-1].wasconst=0;
7999 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
8000 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8001 }
8002 else
8003 // Alloc the delay slot in case the branch is taken
8004 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
8005 {
8006 memcpy(&branch_regs[i-1],&current,sizeof(current));
8007 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
8008 alloc_cc(&branch_regs[i-1],i);
8009 dirty_reg(&branch_regs[i-1],CCREG);
8010 delayslot_alloc(&branch_regs[i-1],i);
8011 branch_regs[i-1].isconst=0;
8012 alloc_reg(&current,i,CCREG); // Not taken path
8013 dirty_reg(&current,CCREG);
8014 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8015 }
8016 else
8017 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
8018 {
8019 memcpy(&branch_regs[i-1],&current,sizeof(current));
8020 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
8021 alloc_cc(&branch_regs[i-1],i);
8022 dirty_reg(&branch_regs[i-1],CCREG);
8023 delayslot_alloc(&branch_regs[i-1],i);
8024 branch_regs[i-1].isconst=0;
8025 alloc_reg(&current,i,CCREG); // Not taken path
8026 dirty_reg(&current,CCREG);
8027 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8028 }
8029 break;
8030 case SJUMP:
8031 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
8032 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
8033 {
8034 alloc_cc(&current,i-1);
8035 dirty_reg(&current,CCREG);
8036 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
8037 // The delay slot overwrote the branch condition
8038 // Delay slot goes after the test (in order)
8039 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8040 current.u|=1;
8041 delayslot_alloc(&current,i);
8042 current.isconst=0;
8043 }
8044 else
8045 {
8046 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
8047 // Alloc the branch condition register
8048 alloc_reg(&current,i-1,dops[i-1].rs1);
8049 }
8050 memcpy(&branch_regs[i-1],&current,sizeof(current));
8051 branch_regs[i-1].isconst=0;
8052 branch_regs[i-1].wasconst=0;
8053 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
8054 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8055 }
8056 else
8057 // Alloc the delay slot in case the branch is taken
8058 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
8059 {
8060 memcpy(&branch_regs[i-1],&current,sizeof(current));
8061 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2)|(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2)))|1;
8062 alloc_cc(&branch_regs[i-1],i);
8063 dirty_reg(&branch_regs[i-1],CCREG);
8064 delayslot_alloc(&branch_regs[i-1],i);
8065 branch_regs[i-1].isconst=0;
8066 alloc_reg(&current,i,CCREG); // Not taken path
8067 dirty_reg(&current,CCREG);
8068 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8069 }
8070 // FIXME: BLTZAL/BGEZAL
8071 if(dops[i-1].opcode2&0x10) { // BxxZAL
8072 alloc_reg(&branch_regs[i-1],i-1,31);
8073 dirty_reg(&branch_regs[i-1],31);
8074 }
8075 break;
8076 }
8077
8078 if (dops[i-1].is_ujump)
8079 {
8080 if(dops[i-1].rt1==31) // JAL/JALR
8081 {
8082 // Subroutine call will return here, don't alloc any registers
8083 current.dirty=0;
8084 clear_all_regs(current.regmap);
8085 alloc_reg(&current,i,CCREG);
8086 dirty_reg(&current,CCREG);
8087 }
8088 else if(i+1<slen)
8089 {
8090 // Internal branch will jump here, match registers to caller
8091 current.dirty=0;
8092 clear_all_regs(current.regmap);
8093 alloc_reg(&current,i,CCREG);
8094 dirty_reg(&current,CCREG);
8095 for(j=i-1;j>=0;j--)
8096 {
8097 if(ba[j]==start+i*4+4) {
8098 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
8099 current.dirty=branch_regs[j].dirty;
8100 break;
8101 }
8102 }
8103 while(j>=0) {
8104 if(ba[j]==start+i*4+4) {
8105 for(hr=0;hr<HOST_REGS;hr++) {
8106 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
8107 current.regmap[hr]=-1;
8108 }
8109 current.dirty&=branch_regs[j].dirty;
8110 }
8111 }
8112 j--;
8113 }
8114 }
8115 }
8116 }
8117
8118 // Count cycles in between branches
8119 ccadj[i]=cc;
8120 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
8121 {
8122 cc=0;
8123 }
8124#if !defined(DRC_DBG)
8125 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
8126 {
8127 // this should really be removed since the real stalls have been implemented,
8128 // but doing so causes sizeable perf regression against the older version
8129 u_int gtec = gte_cycletab[source[i] & 0x3f];
8130 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
8131 }
8132 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
8133 {
8134 cc+=4;
8135 }
8136 else if(dops[i].itype==C2LS)
8137 {
8138 // same as with C2OP
8139 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
8140 }
8141#endif
8142 else
8143 {
8144 cc++;
8145 }
8146
8147 if(!dops[i].is_ds) {
8148 regs[i].dirty=current.dirty;
8149 regs[i].isconst=current.isconst;
8150 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
8151 }
8152 for(hr=0;hr<HOST_REGS;hr++) {
8153 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
8154 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
8155 regs[i].wasconst&=~(1<<hr);
8156 }
8157 }
8158 }
8159 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
8160 regs[i].waswritten=current.waswritten;
8161 }
8162
8163 /* Pass 4 - Cull unused host registers */
8164
8165 uint64_t nr=0;
8166
8167 for (i=slen-1;i>=0;i--)
8168 {
8169 int hr;
8170 if(dops[i].is_jump)
8171 {
8172 if(ba[i]<start || ba[i]>=(start+slen*4))
8173 {
8174 // Branch out of this block, don't need anything
8175 nr=0;
8176 }
8177 else
8178 {
8179 // Internal branch
8180 // Need whatever matches the target
8181 nr=0;
8182 int t=(ba[i]-start)>>2;
8183 for(hr=0;hr<HOST_REGS;hr++)
8184 {
8185 if(regs[i].regmap_entry[hr]>=0) {
8186 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
8187 }
8188 }
8189 }
8190 // Conditional branch may need registers for following instructions
8191 if (!dops[i].is_ujump)
8192 {
8193 if(i<slen-2) {
8194 nr|=needed_reg[i+2];
8195 for(hr=0;hr<HOST_REGS;hr++)
8196 {
8197 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
8198 //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]);
8199 }
8200 }
8201 }
8202 // Don't need stuff which is overwritten
8203 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8204 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8205 // Merge in delay slot
8206 for(hr=0;hr<HOST_REGS;hr++)
8207 {
8208 if(dops[i+1].rt1&&dops[i+1].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8209 if(dops[i+1].rt2&&dops[i+1].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8210 if(dops[i+1].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8211 if(dops[i+1].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8212 if(dops[i+1].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8213 if(dops[i+1].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8214 if(dops[i+1].itype==STORE || dops[i+1].itype==STORELR || (dops[i+1].opcode&0x3b)==0x39 || (dops[i+1].opcode&0x3b)==0x3a) {
8215 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8216 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8217 }
8218 }
8219 }
8220 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8221 {
8222 // SYSCALL instruction (software interrupt)
8223 nr=0;
8224 }
8225 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8226 {
8227 // ERET instruction (return from interrupt)
8228 nr=0;
8229 }
8230 else // Non-branch
8231 {
8232 if(i<slen-1) {
8233 for(hr=0;hr<HOST_REGS;hr++) {
8234 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8235 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8236 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8237 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8238 }
8239 }
8240 }
8241 for(hr=0;hr<HOST_REGS;hr++)
8242 {
8243 // Overwritten registers are not needed
8244 if(dops[i].rt1&&dops[i].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8245 if(dops[i].rt2&&dops[i].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8246 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8247 // Source registers are needed
8248 if(dops[i].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8249 if(dops[i].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8250 if(dops[i].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8251 if(dops[i].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8252 if(dops[i].itype==STORE || dops[i].itype==STORELR || (dops[i].opcode&0x3b)==0x39 || (dops[i].opcode&0x3b)==0x3a) {
8253 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8254 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8255 }
8256 // Don't store a register immediately after writing it,
8257 // may prevent dual-issue.
8258 // But do so if this is a branch target, otherwise we
8259 // might have to load the register before the branch.
8260 if(i>0&&!dops[i].bt&&((regs[i].wasdirty>>hr)&1)) {
8261 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
8262 if(dops[i-1].rt1==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8263 if(dops[i-1].rt2==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8264 }
8265 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
8266 if(dops[i-1].rt1==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8267 if(dops[i-1].rt2==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8268 }
8269 }
8270 }
8271 // Cycle count is needed at branches. Assume it is needed at the target too.
8272 if(i==0||dops[i].bt||dops[i].itype==CJUMP||dops[i].itype==SPAN) {
8273 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8274 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8275 }
8276 // Save it
8277 needed_reg[i]=nr;
8278
8279 // Deallocate unneeded registers
8280 for(hr=0;hr<HOST_REGS;hr++)
8281 {
8282 if(!((nr>>hr)&1)) {
8283 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8284 if(dops[i].is_jump)
8285 {
8286 int map=0,temp=0;
8287 if(dops[i+1].itype==STORE || dops[i+1].itype==STORELR ||
8288 (dops[i+1].opcode&0x3b)==0x39 || (dops[i+1].opcode&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
8289 map=INVCP;
8290 }
8291 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR ||
8292 dops[i+1].itype==C1LS || dops[i+1].itype==C2LS)
8293 temp=FTEMP;
8294 if((regs[i].regmap[hr]&63)!=dops[i].rs1 && (regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8295 (regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8296 (regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8297 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8298 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
8299 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8300 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8301 regs[i].regmap[hr]!=map )
8302 {
8303 regs[i].regmap[hr]=-1;
8304 regs[i].isconst&=~(1<<hr);
8305 if((branch_regs[i].regmap[hr]&63)!=dops[i].rs1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8306 (branch_regs[i].regmap[hr]&63)!=dops[i].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8307 (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8308 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8309 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8310 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8311 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8312 branch_regs[i].regmap[hr]!=map)
8313 {
8314 branch_regs[i].regmap[hr]=-1;
8315 branch_regs[i].regmap_entry[hr]=-1;
8316 if (!dops[i].is_ujump)
8317 {
8318 if (i < slen-2) {
8319 regmap_pre[i+2][hr]=-1;
8320 regs[i+2].wasconst&=~(1<<hr);
8321 }
8322 }
8323 }
8324 }
8325 }
8326 else
8327 {
8328 // Non-branch
8329 if(i>0)
8330 {
8331 int map=-1,temp=-1;
8332 if(dops[i].itype==STORE || dops[i].itype==STORELR ||
8333 (dops[i].opcode&0x3b)==0x39 || (dops[i].opcode&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
8334 map=INVCP;
8335 }
8336 if(dops[i].itype==LOADLR || dops[i].itype==STORELR ||
8337 dops[i].itype==C1LS || dops[i].itype==C2LS)
8338 temp=FTEMP;
8339 if((regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8340 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8341 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map &&
8342 (dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG))
8343 {
8344 if(i<slen-1&&!dops[i].is_ds) {
8345 assert(regs[i].regmap[hr]<64);
8346 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8347 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8348 {
8349 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8350 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8351 }
8352 regmap_pre[i+1][hr]=-1;
8353 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8354 regs[i+1].wasconst&=~(1<<hr);
8355 }
8356 regs[i].regmap[hr]=-1;
8357 regs[i].isconst&=~(1<<hr);
8358 }
8359 }
8360 }
8361 } // if needed
8362 } // for hr
8363 }
8364
8365 /* Pass 5 - Pre-allocate registers */
8366
8367 // If a register is allocated during a loop, try to allocate it for the
8368 // entire loop, if possible. This avoids loading/storing registers
8369 // inside of the loop.
8370
8371 signed char f_regmap[HOST_REGS];
8372 clear_all_regs(f_regmap);
8373 for(i=0;i<slen-1;i++)
8374 {
8375 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8376 {
8377 if(ba[i]>=start && ba[i]<(start+i*4))
8378 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8379 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8380 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8381 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8382 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8383 {
8384 int t=(ba[i]-start)>>2;
8385 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8386 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8387 for(hr=0;hr<HOST_REGS;hr++)
8388 {
8389 if(regs[i].regmap[hr]>=0) {
8390 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8391 // dealloc old register
8392 int n;
8393 for(n=0;n<HOST_REGS;n++)
8394 {
8395 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8396 }
8397 // and alloc new one
8398 f_regmap[hr]=regs[i].regmap[hr];
8399 }
8400 }
8401 if(branch_regs[i].regmap[hr]>=0) {
8402 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8403 // dealloc old register
8404 int n;
8405 for(n=0;n<HOST_REGS;n++)
8406 {
8407 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8408 }
8409 // and alloc new one
8410 f_regmap[hr]=branch_regs[i].regmap[hr];
8411 }
8412 }
8413 if(dops[i].ooo) {
8414 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8415 f_regmap[hr]=branch_regs[i].regmap[hr];
8416 }else{
8417 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8418 f_regmap[hr]=branch_regs[i].regmap[hr];
8419 }
8420 // Avoid dirty->clean transition
8421 #ifdef DESTRUCTIVE_WRITEBACK
8422 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;
8423 #endif
8424 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8425 // case above, however it's always a good idea. We can't hoist the
8426 // load if the register was already allocated, so there's no point
8427 // wasting time analyzing most of these cases. It only "succeeds"
8428 // when the mapping was different and the load can be replaced with
8429 // a mov, which is of negligible benefit. So such cases are
8430 // skipped below.
8431 if(f_regmap[hr]>0) {
8432 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8433 int r=f_regmap[hr];
8434 for(j=t;j<=i;j++)
8435 {
8436 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8437 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8438 assert(r < 64);
8439 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
8440 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8441 int k;
8442 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8443 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8444 if(r>63) {
8445 if(get_reg(regs[i].regmap,r&63)<0) break;
8446 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
8447 }
8448 k=i;
8449 while(k>1&&regs[k-1].regmap[hr]==-1) {
8450 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8451 //printf("no free regs for store %x\n",start+(k-1)*4);
8452 break;
8453 }
8454 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8455 //printf("no-match due to different register\n");
8456 break;
8457 }
8458 if (dops[k-2].is_jump) {
8459 //printf("no-match due to branch\n");
8460 break;
8461 }
8462 // call/ret fast path assumes no registers allocated
8463 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8464 break;
8465 }
8466 assert(r < 64);
8467 k--;
8468 }
8469 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8470 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8471 while(k<i) {
8472 regs[k].regmap_entry[hr]=f_regmap[hr];
8473 regs[k].regmap[hr]=f_regmap[hr];
8474 regmap_pre[k+1][hr]=f_regmap[hr];
8475 regs[k].wasdirty&=~(1<<hr);
8476 regs[k].dirty&=~(1<<hr);
8477 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8478 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8479 regs[k].wasconst&=~(1<<hr);
8480 regs[k].isconst&=~(1<<hr);
8481 k++;
8482 }
8483 }
8484 else {
8485 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8486 break;
8487 }
8488 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8489 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8490 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8491 regs[i].regmap_entry[hr]=f_regmap[hr];
8492 regs[i].regmap[hr]=f_regmap[hr];
8493 regs[i].wasdirty&=~(1<<hr);
8494 regs[i].dirty&=~(1<<hr);
8495 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8496 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8497 regs[i].wasconst&=~(1<<hr);
8498 regs[i].isconst&=~(1<<hr);
8499 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8500 branch_regs[i].wasdirty&=~(1<<hr);
8501 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8502 branch_regs[i].regmap[hr]=f_regmap[hr];
8503 branch_regs[i].dirty&=~(1<<hr);
8504 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8505 branch_regs[i].wasconst&=~(1<<hr);
8506 branch_regs[i].isconst&=~(1<<hr);
8507 if (!dops[i].is_ujump) {
8508 regmap_pre[i+2][hr]=f_regmap[hr];
8509 regs[i+2].wasdirty&=~(1<<hr);
8510 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8511 }
8512 }
8513 }
8514 for(k=t;k<j;k++) {
8515 // Alloc register clean at beginning of loop,
8516 // but may dirty it in pass 6
8517 regs[k].regmap_entry[hr]=f_regmap[hr];
8518 regs[k].regmap[hr]=f_regmap[hr];
8519 regs[k].dirty&=~(1<<hr);
8520 regs[k].wasconst&=~(1<<hr);
8521 regs[k].isconst&=~(1<<hr);
8522 if (dops[k].is_jump) {
8523 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8524 branch_regs[k].regmap[hr]=f_regmap[hr];
8525 branch_regs[k].dirty&=~(1<<hr);
8526 branch_regs[k].wasconst&=~(1<<hr);
8527 branch_regs[k].isconst&=~(1<<hr);
8528 if (!dops[k].is_ujump) {
8529 regmap_pre[k+2][hr]=f_regmap[hr];
8530 regs[k+2].wasdirty&=~(1<<hr);
8531 }
8532 }
8533 else
8534 {
8535 regmap_pre[k+1][hr]=f_regmap[hr];
8536 regs[k+1].wasdirty&=~(1<<hr);
8537 }
8538 }
8539 if(regs[j].regmap[hr]==f_regmap[hr])
8540 regs[j].regmap_entry[hr]=f_regmap[hr];
8541 break;
8542 }
8543 if(j==i) break;
8544 if(regs[j].regmap[hr]>=0)
8545 break;
8546 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8547 //printf("no-match due to different register\n");
8548 break;
8549 }
8550 if (dops[j].is_ujump)
8551 {
8552 // Stop on unconditional branch
8553 break;
8554 }
8555 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8556 {
8557 if(dops[j].ooo) {
8558 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8559 break;
8560 }else{
8561 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8562 break;
8563 }
8564 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8565 //printf("no-match due to different register (branch)\n");
8566 break;
8567 }
8568 }
8569 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8570 //printf("No free regs for store %x\n",start+j*4);
8571 break;
8572 }
8573 assert(f_regmap[hr]<64);
8574 }
8575 }
8576 }
8577 }
8578 }
8579 }else{
8580 // Non branch or undetermined branch target
8581 for(hr=0;hr<HOST_REGS;hr++)
8582 {
8583 if(hr!=EXCLUDE_REG) {
8584 if(regs[i].regmap[hr]>=0) {
8585 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8586 // dealloc old register
8587 int n;
8588 for(n=0;n<HOST_REGS;n++)
8589 {
8590 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8591 }
8592 // and alloc new one
8593 f_regmap[hr]=regs[i].regmap[hr];
8594 }
8595 }
8596 }
8597 }
8598 // Try to restore cycle count at branch targets
8599 if(dops[i].bt) {
8600 for(j=i;j<slen-1;j++) {
8601 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8602 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8603 //printf("no free regs for store %x\n",start+j*4);
8604 break;
8605 }
8606 }
8607 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8608 int k=i;
8609 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8610 while(k<j) {
8611 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8612 regs[k].regmap[HOST_CCREG]=CCREG;
8613 regmap_pre[k+1][HOST_CCREG]=CCREG;
8614 regs[k+1].wasdirty|=1<<HOST_CCREG;
8615 regs[k].dirty|=1<<HOST_CCREG;
8616 regs[k].wasconst&=~(1<<HOST_CCREG);
8617 regs[k].isconst&=~(1<<HOST_CCREG);
8618 k++;
8619 }
8620 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8621 }
8622 // Work backwards from the branch target
8623 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8624 {
8625 //printf("Extend backwards\n");
8626 int k;
8627 k=i;
8628 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8629 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8630 //printf("no free regs for store %x\n",start+(k-1)*4);
8631 break;
8632 }
8633 k--;
8634 }
8635 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8636 //printf("Extend CC, %x ->\n",start+k*4);
8637 while(k<=i) {
8638 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8639 regs[k].regmap[HOST_CCREG]=CCREG;
8640 regmap_pre[k+1][HOST_CCREG]=CCREG;
8641 regs[k+1].wasdirty|=1<<HOST_CCREG;
8642 regs[k].dirty|=1<<HOST_CCREG;
8643 regs[k].wasconst&=~(1<<HOST_CCREG);
8644 regs[k].isconst&=~(1<<HOST_CCREG);
8645 k++;
8646 }
8647 }
8648 else {
8649 //printf("Fail Extend CC, %x ->\n",start+k*4);
8650 }
8651 }
8652 }
8653 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8654 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8655 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8656 {
8657 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8658 }
8659 }
8660 }
8661
8662 // This allocates registers (if possible) one instruction prior
8663 // to use, which can avoid a load-use penalty on certain CPUs.
8664 for(i=0;i<slen-1;i++)
8665 {
8666 if (!i || !dops[i-1].is_jump)
8667 {
8668 if(!dops[i+1].bt)
8669 {
8670 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8671 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8672 {
8673 if(dops[i+1].rs1) {
8674 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8675 {
8676 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8677 {
8678 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8679 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8680 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8681 regs[i].isconst&=~(1<<hr);
8682 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8683 constmap[i][hr]=constmap[i+1][hr];
8684 regs[i+1].wasdirty&=~(1<<hr);
8685 regs[i].dirty&=~(1<<hr);
8686 }
8687 }
8688 }
8689 if(dops[i+1].rs2) {
8690 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8691 {
8692 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8693 {
8694 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8695 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8696 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8697 regs[i].isconst&=~(1<<hr);
8698 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8699 constmap[i][hr]=constmap[i+1][hr];
8700 regs[i+1].wasdirty&=~(1<<hr);
8701 regs[i].dirty&=~(1<<hr);
8702 }
8703 }
8704 }
8705 // Preload target address for load instruction (non-constant)
8706 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8707 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8708 {
8709 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8710 {
8711 regs[i].regmap[hr]=dops[i+1].rs1;
8712 regmap_pre[i+1][hr]=dops[i+1].rs1;
8713 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8714 regs[i].isconst&=~(1<<hr);
8715 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8716 constmap[i][hr]=constmap[i+1][hr];
8717 regs[i+1].wasdirty&=~(1<<hr);
8718 regs[i].dirty&=~(1<<hr);
8719 }
8720 }
8721 }
8722 // Load source into target register
8723 if(dops[i+1].lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8724 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8725 {
8726 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8727 {
8728 regs[i].regmap[hr]=dops[i+1].rs1;
8729 regmap_pre[i+1][hr]=dops[i+1].rs1;
8730 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8731 regs[i].isconst&=~(1<<hr);
8732 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8733 constmap[i][hr]=constmap[i+1][hr];
8734 regs[i+1].wasdirty&=~(1<<hr);
8735 regs[i].dirty&=~(1<<hr);
8736 }
8737 }
8738 }
8739 // Address for store instruction (non-constant)
8740 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8741 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8742 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8743 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8744 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8745 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
8746 assert(hr>=0);
8747 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8748 {
8749 regs[i].regmap[hr]=dops[i+1].rs1;
8750 regmap_pre[i+1][hr]=dops[i+1].rs1;
8751 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8752 regs[i].isconst&=~(1<<hr);
8753 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8754 constmap[i][hr]=constmap[i+1][hr];
8755 regs[i+1].wasdirty&=~(1<<hr);
8756 regs[i].dirty&=~(1<<hr);
8757 }
8758 }
8759 }
8760 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8761 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8762 int nr;
8763 hr=get_reg(regs[i+1].regmap,FTEMP);
8764 assert(hr>=0);
8765 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8766 {
8767 regs[i].regmap[hr]=dops[i+1].rs1;
8768 regmap_pre[i+1][hr]=dops[i+1].rs1;
8769 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8770 regs[i].isconst&=~(1<<hr);
8771 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8772 constmap[i][hr]=constmap[i+1][hr];
8773 regs[i+1].wasdirty&=~(1<<hr);
8774 regs[i].dirty&=~(1<<hr);
8775 }
8776 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8777 {
8778 // move it to another register
8779 regs[i+1].regmap[hr]=-1;
8780 regmap_pre[i+2][hr]=-1;
8781 regs[i+1].regmap[nr]=FTEMP;
8782 regmap_pre[i+2][nr]=FTEMP;
8783 regs[i].regmap[nr]=dops[i+1].rs1;
8784 regmap_pre[i+1][nr]=dops[i+1].rs1;
8785 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8786 regs[i].isconst&=~(1<<nr);
8787 regs[i+1].isconst&=~(1<<nr);
8788 regs[i].dirty&=~(1<<nr);
8789 regs[i+1].wasdirty&=~(1<<nr);
8790 regs[i+1].dirty&=~(1<<nr);
8791 regs[i+2].wasdirty&=~(1<<nr);
8792 }
8793 }
8794 }
8795 if(dops[i+1].itype==LOAD||dops[i+1].itype==LOADLR||dops[i+1].itype==STORE||dops[i+1].itype==STORELR/*||dops[i+1].itype==C1LS||||dops[i+1].itype==C2LS*/) {
8796 if(dops[i+1].itype==LOAD)
8797 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8798 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8799 hr=get_reg(regs[i+1].regmap,FTEMP);
8800 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
8801 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8802 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8803 }
8804 if(hr>=0&&regs[i].regmap[hr]<0) {
8805 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8806 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8807 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8808 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8809 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8810 regs[i].isconst&=~(1<<hr);
8811 regs[i+1].wasdirty&=~(1<<hr);
8812 regs[i].dirty&=~(1<<hr);
8813 }
8814 }
8815 }
8816 }
8817 }
8818 }
8819 }
8820
8821 /* Pass 6 - Optimize clean/dirty state */
8822 clean_registers(0,slen-1,1);
8823
8824 /* Pass 7 - Identify 32-bit registers */
8825 for (i=slen-1;i>=0;i--)
8826 {
8827 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8828 {
8829 // Conditional branch
8830 if((source[i]>>16)!=0x1000&&i<slen-2) {
8831 // Mark this address as a branch target since it may be called
8832 // upon return from interrupt
8833 dops[i+2].bt=1;
8834 }
8835 }
8836 }
8837
8838 if(dops[slen-1].itype==SPAN) {
8839 dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
8840 }
8841
8842#ifdef DISASM
8843 /* Debug/disassembly */
8844 for(i=0;i<slen;i++)
8845 {
8846 printf("U:");
8847 int r;
8848 for(r=1;r<=CCREG;r++) {
8849 if((unneeded_reg[i]>>r)&1) {
8850 if(r==HIREG) printf(" HI");
8851 else if(r==LOREG) printf(" LO");
8852 else printf(" r%d",r);
8853 }
8854 }
8855 printf("\n");
8856 #if defined(__i386__) || defined(__x86_64__)
8857 printf("pre: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7]);
8858 #endif
8859 #ifdef __arm__
8860 printf("pre: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][4],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7],regmap_pre[i][8],regmap_pre[i][9],regmap_pre[i][10],regmap_pre[i][12]);
8861 #endif
8862 #if defined(__i386__) || defined(__x86_64__)
8863 printf("needs: ");
8864 if(needed_reg[i]&1) printf("eax ");
8865 if((needed_reg[i]>>1)&1) printf("ecx ");
8866 if((needed_reg[i]>>2)&1) printf("edx ");
8867 if((needed_reg[i]>>3)&1) printf("ebx ");
8868 if((needed_reg[i]>>5)&1) printf("ebp ");
8869 if((needed_reg[i]>>6)&1) printf("esi ");
8870 if((needed_reg[i]>>7)&1) printf("edi ");
8871 printf("\n");
8872 printf("entry: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7]);
8873 printf("dirty: ");
8874 if(regs[i].wasdirty&1) printf("eax ");
8875 if((regs[i].wasdirty>>1)&1) printf("ecx ");
8876 if((regs[i].wasdirty>>2)&1) printf("edx ");
8877 if((regs[i].wasdirty>>3)&1) printf("ebx ");
8878 if((regs[i].wasdirty>>5)&1) printf("ebp ");
8879 if((regs[i].wasdirty>>6)&1) printf("esi ");
8880 if((regs[i].wasdirty>>7)&1) printf("edi ");
8881 #endif
8882 #ifdef __arm__
8883 printf("entry: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[4],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7],regs[i].regmap_entry[8],regs[i].regmap_entry[9],regs[i].regmap_entry[10],regs[i].regmap_entry[12]);
8884 printf("dirty: ");
8885 if(regs[i].wasdirty&1) printf("r0 ");
8886 if((regs[i].wasdirty>>1)&1) printf("r1 ");
8887 if((regs[i].wasdirty>>2)&1) printf("r2 ");
8888 if((regs[i].wasdirty>>3)&1) printf("r3 ");
8889 if((regs[i].wasdirty>>4)&1) printf("r4 ");
8890 if((regs[i].wasdirty>>5)&1) printf("r5 ");
8891 if((regs[i].wasdirty>>6)&1) printf("r6 ");
8892 if((regs[i].wasdirty>>7)&1) printf("r7 ");
8893 if((regs[i].wasdirty>>8)&1) printf("r8 ");
8894 if((regs[i].wasdirty>>9)&1) printf("r9 ");
8895 if((regs[i].wasdirty>>10)&1) printf("r10 ");
8896 if((regs[i].wasdirty>>12)&1) printf("r12 ");
8897 #endif
8898 printf("\n");
8899 disassemble_inst(i);
8900 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
8901 #if defined(__i386__) || defined(__x86_64__)
8902 printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7]);
8903 if(regs[i].dirty&1) printf("eax ");
8904 if((regs[i].dirty>>1)&1) printf("ecx ");
8905 if((regs[i].dirty>>2)&1) printf("edx ");
8906 if((regs[i].dirty>>3)&1) printf("ebx ");
8907 if((regs[i].dirty>>5)&1) printf("ebp ");
8908 if((regs[i].dirty>>6)&1) printf("esi ");
8909 if((regs[i].dirty>>7)&1) printf("edi ");
8910 #endif
8911 #ifdef __arm__
8912 printf("r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[4],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7],regs[i].regmap[8],regs[i].regmap[9],regs[i].regmap[10],regs[i].regmap[12]);
8913 if(regs[i].dirty&1) printf("r0 ");
8914 if((regs[i].dirty>>1)&1) printf("r1 ");
8915 if((regs[i].dirty>>2)&1) printf("r2 ");
8916 if((regs[i].dirty>>3)&1) printf("r3 ");
8917 if((regs[i].dirty>>4)&1) printf("r4 ");
8918 if((regs[i].dirty>>5)&1) printf("r5 ");
8919 if((regs[i].dirty>>6)&1) printf("r6 ");
8920 if((regs[i].dirty>>7)&1) printf("r7 ");
8921 if((regs[i].dirty>>8)&1) printf("r8 ");
8922 if((regs[i].dirty>>9)&1) printf("r9 ");
8923 if((regs[i].dirty>>10)&1) printf("r10 ");
8924 if((regs[i].dirty>>12)&1) printf("r12 ");
8925 #endif
8926 printf("\n");
8927 if(regs[i].isconst) {
8928 printf("constants: ");
8929 #if defined(__i386__) || defined(__x86_64__)
8930 if(regs[i].isconst&1) printf("eax=%x ",(u_int)constmap[i][0]);
8931 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(u_int)constmap[i][1]);
8932 if((regs[i].isconst>>2)&1) printf("edx=%x ",(u_int)constmap[i][2]);
8933 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(u_int)constmap[i][3]);
8934 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(u_int)constmap[i][5]);
8935 if((regs[i].isconst>>6)&1) printf("esi=%x ",(u_int)constmap[i][6]);
8936 if((regs[i].isconst>>7)&1) printf("edi=%x ",(u_int)constmap[i][7]);
8937 #endif
8938 #if defined(__arm__) || defined(__aarch64__)
8939 int r;
8940 for (r = 0; r < ARRAY_SIZE(constmap[i]); r++)
8941 if ((regs[i].isconst >> r) & 1)
8942 printf(" r%d=%x", r, (u_int)constmap[i][r]);
8943 #endif
8944 printf("\n");
8945 }
8946 if(dops[i].is_jump) {
8947 #if defined(__i386__) || defined(__x86_64__)
8948 printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",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]);
8949 if(branch_regs[i].dirty&1) printf("eax ");
8950 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
8951 if((branch_regs[i].dirty>>2)&1) printf("edx ");
8952 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
8953 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
8954 if((branch_regs[i].dirty>>6)&1) printf("esi ");
8955 if((branch_regs[i].dirty>>7)&1) printf("edi ");
8956 #endif
8957 #ifdef __arm__
8958 printf("branch(%d): r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",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[4],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7],branch_regs[i].regmap[8],branch_regs[i].regmap[9],branch_regs[i].regmap[10],branch_regs[i].regmap[12]);
8959 if(branch_regs[i].dirty&1) printf("r0 ");
8960 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
8961 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
8962 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
8963 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
8964 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
8965 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
8966 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
8967 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
8968 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
8969 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
8970 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
8971 #endif
8972 }
8973 }
8974#endif // DISASM
8975
8976 /* Pass 8 - Assembly */
8977 linkcount=0;stubcount=0;
8978 ds=0;is_delayslot=0;
8979 u_int dirty_pre=0;
8980 void *beginning=start_block();
8981 if((u_int)addr&1) {
8982 ds=1;
8983 pagespan_ds();
8984 }
8985 void *instr_addr0_override = NULL;
8986
8987 if (start == 0x80030000) {
8988 // nasty hack for the fastbios thing
8989 // override block entry to this code
8990 instr_addr0_override = out;
8991 emit_movimm(start,0);
8992 // abuse io address var as a flag that we
8993 // have already returned here once
8994 emit_readword(&address,1);
8995 emit_writeword(0,&pcaddr);
8996 emit_writeword(0,&address);
8997 emit_cmp(0,1);
8998 #ifdef __aarch64__
8999 emit_jeq(out + 4*2);
9000 emit_far_jump(new_dyna_leave);
9001 #else
9002 emit_jne(new_dyna_leave);
9003 #endif
9004 }
9005 for(i=0;i<slen;i++)
9006 {
9007 //if(ds) printf("ds: ");
9008 disassemble_inst(i);
9009 if(ds) {
9010 ds=0; // Skip delay slot
9011 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9012 instr_addr[i] = NULL;
9013 } else {
9014 speculate_register_values(i);
9015 #ifndef DESTRUCTIVE_WRITEBACK
9016 if (i < 2 || !dops[i-2].is_ujump)
9017 {
9018 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9019 }
9020 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9021 dirty_pre=branch_regs[i].dirty;
9022 }else{
9023 dirty_pre=regs[i].dirty;
9024 }
9025 #endif
9026 // write back
9027 if (i < 2 || !dops[i-2].is_ujump)
9028 {
9029 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9030 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9031 }
9032 // branch target entry point
9033 instr_addr[i] = out;
9034 assem_debug("<->\n");
9035 drc_dbg_emit_do_cmp(i);
9036
9037 // load regs
9038 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9039 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9040 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9041 address_generation(i,&regs[i],regs[i].regmap_entry);
9042 load_consts(regmap_pre[i],regs[i].regmap,i);
9043 if(dops[i].is_jump)
9044 {
9045 // Load the delay slot registers if necessary
9046 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))
9047 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9048 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))
9049 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9050 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a)
9051 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9052 }
9053 else if(i+1<slen)
9054 {
9055 // Preload registers for following instruction
9056 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9057 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
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)
9060 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9061 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9062 }
9063 // TODO: if(is_ooo(i)) address_generation(i+1);
9064 if(dops[i].itype==CJUMP)
9065 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
9066 if(dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].opcode&0x3b)==0x39||(dops[i].opcode&0x3b)==0x3a)
9067 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9068 // assemble
9069 switch(dops[i].itype) {
9070 case ALU:
9071 alu_assemble(i,&regs[i]);break;
9072 case IMM16:
9073 imm16_assemble(i,&regs[i]);break;
9074 case SHIFT:
9075 shift_assemble(i,&regs[i]);break;
9076 case SHIFTIMM:
9077 shiftimm_assemble(i,&regs[i]);break;
9078 case LOAD:
9079 load_assemble(i,&regs[i]);break;
9080 case LOADLR:
9081 loadlr_assemble(i,&regs[i]);break;
9082 case STORE:
9083 store_assemble(i,&regs[i]);break;
9084 case STORELR:
9085 storelr_assemble(i,&regs[i]);break;
9086 case COP0:
9087 cop0_assemble(i,&regs[i]);break;
9088 case COP1:
9089 cop1_assemble(i,&regs[i]);break;
9090 case C1LS:
9091 c1ls_assemble(i,&regs[i]);break;
9092 case COP2:
9093 cop2_assemble(i,&regs[i]);break;
9094 case C2LS:
9095 c2ls_assemble(i,&regs[i]);break;
9096 case C2OP:
9097 c2op_assemble(i,&regs[i]);break;
9098 case MULTDIV:
9099 multdiv_assemble(i,&regs[i]);
9100 multdiv_prepare_stall(i,&regs[i]);
9101 break;
9102 case MOV:
9103 mov_assemble(i,&regs[i]);break;
9104 case SYSCALL:
9105 syscall_assemble(i,&regs[i]);break;
9106 case HLECALL:
9107 hlecall_assemble(i,&regs[i]);break;
9108 case INTCALL:
9109 intcall_assemble(i,&regs[i]);break;
9110 case UJUMP:
9111 ujump_assemble(i,&regs[i]);ds=1;break;
9112 case RJUMP:
9113 rjump_assemble(i,&regs[i]);ds=1;break;
9114 case CJUMP:
9115 cjump_assemble(i,&regs[i]);ds=1;break;
9116 case SJUMP:
9117 sjump_assemble(i,&regs[i]);ds=1;break;
9118 case SPAN:
9119 pagespan_assemble(i,&regs[i]);break;
9120 }
9121 if (dops[i].is_ujump)
9122 literal_pool(1024);
9123 else
9124 literal_pool_jumpover(256);
9125 }
9126 }
9127
9128 assert(slen > 0);
9129 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9130 // no ending needed for this block since INTCALL never returns
9131 }
9132 // If the block did not end with an unconditional branch,
9133 // add a jump to the next instruction.
9134 else if (i > 1) {
9135 if (!dops[i-2].is_ujump && dops[i-1].itype != SPAN) {
9136 assert(!dops[i-1].is_jump);
9137 assert(i==slen);
9138 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9139 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9140 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9141 emit_loadreg(CCREG,HOST_CCREG);
9142 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
9143 }
9144 else
9145 {
9146 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9147 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9148 }
9149 add_to_linker(out,start+i*4,0);
9150 emit_jmp(0);
9151 }
9152 }
9153 else
9154 {
9155 assert(i>0);
9156 assert(!dops[i-1].is_jump);
9157 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9158 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9159 emit_loadreg(CCREG,HOST_CCREG);
9160 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
9161 add_to_linker(out,start+i*4,0);
9162 emit_jmp(0);
9163 }
9164
9165 // TODO: delay slot stubs?
9166 // Stubs
9167 for(i=0;i<stubcount;i++)
9168 {
9169 switch(stubs[i].type)
9170 {
9171 case LOADB_STUB:
9172 case LOADH_STUB:
9173 case LOADW_STUB:
9174 case LOADD_STUB:
9175 case LOADBU_STUB:
9176 case LOADHU_STUB:
9177 do_readstub(i);break;
9178 case STOREB_STUB:
9179 case STOREH_STUB:
9180 case STOREW_STUB:
9181 case STORED_STUB:
9182 do_writestub(i);break;
9183 case CC_STUB:
9184 do_ccstub(i);break;
9185 case INVCODE_STUB:
9186 do_invstub(i);break;
9187 case FP_STUB:
9188 do_cop1stub(i);break;
9189 case STORELR_STUB:
9190 do_unalignedwritestub(i);break;
9191 }
9192 }
9193
9194 if (instr_addr0_override)
9195 instr_addr[0] = instr_addr0_override;
9196
9197 /* Pass 9 - Linker */
9198 for(i=0;i<linkcount;i++)
9199 {
9200 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9201 literal_pool(64);
9202 if (!link_addr[i].ext)
9203 {
9204 void *stub = out;
9205 void *addr = check_addr(link_addr[i].target);
9206 emit_extjump(link_addr[i].addr, link_addr[i].target);
9207 if (addr) {
9208 set_jump_target(link_addr[i].addr, addr);
9209 add_jump_out(link_addr[i].target,stub);
9210 }
9211 else
9212 set_jump_target(link_addr[i].addr, stub);
9213 }
9214 else
9215 {
9216 // Internal branch
9217 int target=(link_addr[i].target-start)>>2;
9218 assert(target>=0&&target<slen);
9219 assert(instr_addr[target]);
9220 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9221 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9222 //#else
9223 set_jump_target(link_addr[i].addr, instr_addr[target]);
9224 //#endif
9225 }
9226 }
9227
9228 u_int source_len = slen*4;
9229 if (dops[slen-1].itype == INTCALL && source_len > 4)
9230 // no need to treat the last instruction as compiled
9231 // as interpreter fully handles it
9232 source_len -= 4;
9233
9234 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9235 copy = shadow;
9236
9237 // External Branch Targets (jump_in)
9238 for(i=0;i<slen;i++)
9239 {
9240 if(dops[i].bt||i==0)
9241 {
9242 if(instr_addr[i]) // TODO - delay slots (=null)
9243 {
9244 u_int vaddr=start+i*4;
9245 u_int page=get_page(vaddr);
9246 u_int vpage=get_vpage(vaddr);
9247 literal_pool(256);
9248 {
9249 assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
9250 assem_debug("jump_in: %x\n",start+i*4);
9251 ll_add(jump_dirty+vpage,vaddr,out);
9252 void *entry_point = do_dirty_stub(i, source_len);
9253 ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
9254 // If there was an existing entry in the hash table,
9255 // replace it with the new address.
9256 // Don't add new entries. We'll insert the
9257 // ones that actually get used in check_addr().
9258 struct ht_entry *ht_bin = hash_table_get(vaddr);
9259 if (ht_bin->vaddr[0] == vaddr)
9260 ht_bin->tcaddr[0] = entry_point;
9261 if (ht_bin->vaddr[1] == vaddr)
9262 ht_bin->tcaddr[1] = entry_point;
9263 }
9264 }
9265 }
9266 }
9267 // Write out the literal pool if necessary
9268 literal_pool(0);
9269 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9270 // Align code
9271 if(((u_int)out)&7) emit_addnop(13);
9272 #endif
9273 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9274 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9275 memcpy(copy, source, source_len);
9276 copy += source_len;
9277
9278 end_block(beginning);
9279
9280 // If we're within 256K of the end of the buffer,
9281 // start over from the beginning. (Is 256K enough?)
9282 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9283 out = ndrc->translation_cache;
9284
9285 // Trap writes to any of the pages we compiled
9286 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
9287 invalid_code[i]=0;
9288 }
9289 inv_code_start=inv_code_end=~0;
9290
9291 // for PCSX we need to mark all mirrors too
9292 if(get_page(start)<(RAM_SIZE>>12))
9293 for(i=start>>12;i<=(start+slen*4)>>12;i++)
9294 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
9295 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
9296 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
9297
9298 /* Pass 10 - Free memory by expiring oldest blocks */
9299
9300 int end=(((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
9301 while(expirep!=end)
9302 {
9303 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
9304 uintptr_t base_offs = ((uintptr_t)(expirep >> 13) << shift); // Base offset of this block
9305 uintptr_t base_offs_s = base_offs >> shift;
9306 inv_debug("EXP: Phase %d\n",expirep);
9307 switch((expirep>>11)&3)
9308 {
9309 case 0:
9310 // Clear jump_in and jump_dirty
9311 ll_remove_matching_addrs(jump_in+(expirep&2047),base_offs_s,shift);
9312 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base_offs_s,shift);
9313 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base_offs_s,shift);
9314 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base_offs_s,shift);
9315 break;
9316 case 1:
9317 // Clear pointers
9318 ll_kill_pointers(jump_out[expirep&2047],base_offs_s,shift);
9319 ll_kill_pointers(jump_out[(expirep&2047)+2048],base_offs_s,shift);
9320 break;
9321 case 2:
9322 // Clear hash table
9323 for(i=0;i<32;i++) {
9324 struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
9325 uintptr_t o1 = (u_char *)ht_bin->tcaddr[1] - ndrc->translation_cache;
9326 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9327 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9328 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
9329 ht_bin->vaddr[1] = -1;
9330 ht_bin->tcaddr[1] = NULL;
9331 }
9332 o1 = (u_char *)ht_bin->tcaddr[0] - ndrc->translation_cache;
9333 o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9334 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9335 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
9336 ht_bin->vaddr[0] = ht_bin->vaddr[1];
9337 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
9338 ht_bin->vaddr[1] = -1;
9339 ht_bin->tcaddr[1] = NULL;
9340 }
9341 }
9342 break;
9343 case 3:
9344 // Clear jump_out
9345 if((expirep&2047)==0)
9346 do_clear_cache();
9347 ll_remove_matching_addrs(jump_out+(expirep&2047),base_offs_s,shift);
9348 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base_offs_s,shift);
9349 break;
9350 }
9351 expirep=(expirep+1)&65535;
9352 }
9353 return 0;
9354}
9355
9356// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork