drc: emulate break opcode
[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
33#include "new_dynarec_config.h"
34#include "../psxhle.h"
35#include "../psxinterpreter.h"
36#include "../gte.h"
37#include "emu_if.h" // emulator interface
38
39#define noinline __attribute__((noinline,noclone))
40#ifndef ARRAY_SIZE
41#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
42#endif
43#ifndef min
44#define min(a, b) ((b) < (a) ? (b) : (a))
45#endif
46#ifndef max
47#define max(a, b) ((b) > (a) ? (b) : (a))
48#endif
49
50//#define DISASM
51//#define ASSEM_PRINT
52//#define REG_ALLOC_PRINT
53
54#ifdef ASSEM_PRINT
55#define assem_debug printf
56#else
57#define assem_debug(...)
58#endif
59//#define inv_debug printf
60#define inv_debug(...)
61
62#ifdef __i386__
63#include "assem_x86.h"
64#endif
65#ifdef __x86_64__
66#include "assem_x64.h"
67#endif
68#ifdef __arm__
69#include "assem_arm.h"
70#endif
71#ifdef __aarch64__
72#include "assem_arm64.h"
73#endif
74
75#define RAM_SIZE 0x200000
76#define MAXBLOCK 4096
77#define MAX_OUTPUT_BLOCK_SIZE 262144
78
79#ifdef VITA
80// apparently Vita has a 16MB limit, so either we cut tc in half,
81// or use this hack (it's a hack because tc size was designed to be power-of-2)
82#define TC_REDUCE_BYTES 4096
83#else
84#define TC_REDUCE_BYTES 0
85#endif
86
87struct ndrc_mem
88{
89 u_char translation_cache[(1 << TARGET_SIZE_2) - TC_REDUCE_BYTES];
90 struct
91 {
92 struct tramp_insns ops[2048 / sizeof(struct tramp_insns)];
93 const void *f[2048 / sizeof(void *)];
94 } tramp;
95};
96
97#ifdef BASE_ADDR_DYNAMIC
98static struct ndrc_mem *ndrc;
99#else
100static struct ndrc_mem ndrc_ __attribute__((aligned(4096)));
101static struct ndrc_mem *ndrc = &ndrc_;
102#endif
103
104// stubs
105enum stub_type {
106 CC_STUB = 1,
107 FP_STUB = 2,
108 LOADB_STUB = 3,
109 LOADH_STUB = 4,
110 LOADW_STUB = 5,
111 LOADD_STUB = 6,
112 LOADBU_STUB = 7,
113 LOADHU_STUB = 8,
114 STOREB_STUB = 9,
115 STOREH_STUB = 10,
116 STOREW_STUB = 11,
117 STORED_STUB = 12,
118 STORELR_STUB = 13,
119 INVCODE_STUB = 14,
120};
121
122struct regstat
123{
124 signed char regmap_entry[HOST_REGS]; // pre-insn + loop preloaded regs?
125 signed char regmap[HOST_REGS];
126 uint64_t wasdirty;
127 uint64_t dirty;
128 uint64_t u;
129 u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
130 u_int isconst; // ... but isconst is false when r2 is known
131 u_int loadedconst; // host regs that have constants loaded
132 u_int waswritten; // MIPS regs that were used as store base before
133};
134
135// note: asm depends on this layout
136struct ll_entry
137{
138 u_int vaddr;
139 u_int reg_sv_flags;
140 void *addr;
141 struct ll_entry *next;
142};
143
144struct ht_entry
145{
146 u_int vaddr[2];
147 void *tcaddr[2];
148};
149
150struct code_stub
151{
152 enum stub_type type;
153 void *addr;
154 void *retaddr;
155 u_int a;
156 uintptr_t b;
157 uintptr_t c;
158 u_int d;
159 u_int e;
160};
161
162struct link_entry
163{
164 void *addr;
165 u_int target;
166 u_int ext;
167};
168
169static struct decoded_insn
170{
171 u_char itype;
172 u_char opcode;
173 u_char opcode2;
174 u_char rs1;
175 u_char rs2;
176 u_char rt1;
177 u_char rt2;
178 u_char lt1;
179 u_char bt:1;
180 u_char ooo:1;
181 u_char is_ds:1;
182 u_char is_jump:1;
183 u_char is_ujump:1;
184 u_char is_load:1;
185 u_char is_store:1;
186} dops[MAXBLOCK];
187
188 // used by asm:
189 u_char *out;
190 struct ht_entry hash_table[65536] __attribute__((aligned(16)));
191 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
192 struct ll_entry *jump_dirty[4096];
193
194 static struct ll_entry *jump_out[4096];
195 static u_int start;
196 static u_int *source;
197 static char insn[MAXBLOCK][10];
198 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
199 static uint64_t gte_rt[MAXBLOCK];
200 static uint64_t gte_unneeded[MAXBLOCK];
201 static u_int smrv[32]; // speculated MIPS register values
202 static u_int smrv_strong; // mask or regs that are likely to have correct values
203 static u_int smrv_weak; // same, but somewhat less likely
204 static u_int smrv_strong_next; // same, but after current insn executes
205 static u_int smrv_weak_next;
206 static int imm[MAXBLOCK];
207 static u_int ba[MAXBLOCK];
208 static uint64_t unneeded_reg[MAXBLOCK];
209 static uint64_t branch_unneeded_reg[MAXBLOCK];
210 // pre-instruction [i], excluding loop-preload regs?
211 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
212 // contains 'real' consts at [i] insn, but may differ from what's actually
213 // loaded in host reg as 'final' value is always loaded, see get_final_value()
214 static uint32_t current_constmap[HOST_REGS];
215 static uint32_t constmap[MAXBLOCK][HOST_REGS];
216 static struct regstat regs[MAXBLOCK];
217 static struct regstat branch_regs[MAXBLOCK];
218 static signed char minimum_free_regs[MAXBLOCK];
219 static u_int needed_reg[MAXBLOCK];
220 static u_int wont_dirty[MAXBLOCK];
221 static u_int will_dirty[MAXBLOCK];
222 static int ccadj[MAXBLOCK];
223 static int slen;
224 static void *instr_addr[MAXBLOCK];
225 static struct link_entry link_addr[MAXBLOCK];
226 static int linkcount;
227 static struct code_stub stubs[MAXBLOCK*3];
228 static int stubcount;
229 static u_int literals[1024][2];
230 static int literalcount;
231 static int is_delayslot;
232 static char shadow[1048576] __attribute__((aligned(16)));
233 static void *copy;
234 static int expirep;
235 static u_int stop_after_jal;
236 static u_int f1_hack;
237
238 int new_dynarec_hacks;
239 int new_dynarec_hacks_pergame;
240 int new_dynarec_hacks_old;
241 int new_dynarec_did_compile;
242
243 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
244
245 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
246 extern int last_count; // last absolute target, often = next_interupt
247 extern int pcaddr;
248 extern int pending_exception;
249 extern int branch_target;
250 extern uintptr_t ram_offset;
251 extern uintptr_t mini_ht[32][2];
252 extern u_char restore_candidate[512];
253
254 /* registers that may be allocated */
255 /* 1-31 gpr */
256#define LOREG 32 // lo
257#define HIREG 33 // hi
258//#define FSREG 34 // FPU status (FCSR)
259#define CSREG 35 // Coprocessor status
260#define CCREG 36 // Cycle count
261#define INVCP 37 // Pointer to invalid_code
262//#define MMREG 38 // Pointer to memory_map
263#define ROREG 39 // ram offset (if rdram!=0x80000000)
264#define TEMPREG 40
265#define FTEMP 40 // FPU temporary register
266#define PTEMP 41 // Prefetch temporary register
267//#define TLREG 42 // TLB mapping offset
268#define RHASH 43 // Return address hash
269#define RHTBL 44 // Return address hash table address
270#define RTEMP 45 // JR/JALR address register
271#define MAXREG 45
272#define AGEN1 46 // Address generation temporary register
273//#define AGEN2 47 // Address generation temporary register
274//#define MGEN1 48 // Maptable address generation temporary register
275//#define MGEN2 49 // Maptable address generation temporary register
276#define BTREG 50 // Branch target temporary register
277
278 /* instruction types */
279#define NOP 0 // No operation
280#define LOAD 1 // Load
281#define STORE 2 // Store
282#define LOADLR 3 // Unaligned load
283#define STORELR 4 // Unaligned store
284#define MOV 5 // Move
285#define ALU 6 // Arithmetic/logic
286#define MULTDIV 7 // Multiply/divide
287#define SHIFT 8 // Shift by register
288#define SHIFTIMM 9// Shift by immediate
289#define IMM16 10 // 16-bit immediate
290#define RJUMP 11 // Unconditional jump to register
291#define UJUMP 12 // Unconditional jump
292#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
293#define SJUMP 14 // Conditional branch (regimm format)
294#define COP0 15 // Coprocessor 0
295#define COP1 16 // Coprocessor 1
296#define C1LS 17 // Coprocessor 1 load/store
297//#define FJUMP 18 // Conditional branch (floating point)
298//#define FLOAT 19 // Floating point unit
299//#define FCONV 20 // Convert integer to float
300//#define FCOMP 21 // Floating point compare (sets FSREG)
301#define SYSCALL 22// SYSCALL,BREAK
302#define OTHER 23 // Other
303#define SPAN 24 // Branch/delay slot spans 2 pages
304#define NI 25 // Not implemented
305#define HLECALL 26// PCSX fake opcodes for HLE
306#define COP2 27 // Coprocessor 2 move
307#define C2LS 28 // Coprocessor 2 load/store
308#define C2OP 29 // Coprocessor 2 operation
309#define INTCALL 30// Call interpreter to handle rare corner cases
310
311 /* branch codes */
312#define TAKEN 1
313#define NOTTAKEN 2
314#define NULLDS 3
315
316#define DJT_1 (void *)1l // no function, just a label in assem_debug log
317#define DJT_2 (void *)2l
318
319// asm linkage
320int new_recompile_block(u_int addr);
321void *get_addr_ht(u_int vaddr);
322void invalidate_block(u_int block);
323void invalidate_addr(u_int addr);
324void remove_hash(int vaddr);
325void dyna_linker();
326void dyna_linker_ds();
327void verify_code();
328void verify_code_ds();
329void cc_interrupt();
330void fp_exception();
331void fp_exception_ds();
332void jump_syscall (u_int u0, u_int u1, u_int pc);
333void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
334void jump_break (u_int u0, u_int u1, u_int pc);
335void jump_break_ds(u_int u0, u_int u1, u_int pc);
336void jump_to_new_pc();
337void call_gteStall();
338void new_dyna_leave();
339
340// Needed by assembler
341static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
342static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
343static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
344static void load_all_regs(const signed char i_regmap[]);
345static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
346static void load_regs_entry(int t);
347static void load_all_consts(const signed char regmap[], u_int dirty, int i);
348static u_int get_host_reglist(const signed char *regmap);
349
350static int verify_dirty(const u_int *ptr);
351static int get_final_value(int hr, int i, int *value);
352static void add_stub(enum stub_type type, void *addr, void *retaddr,
353 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
354static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
355 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
356static void add_to_linker(void *addr, u_int target, int ext);
357static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
358 int addr, int *offset_reg, int *addr_reg_override);
359static void *get_direct_memhandler(void *table, u_int addr,
360 enum stub_type type, uintptr_t *addr_host);
361static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
362static void pass_args(int a0, int a1);
363static void emit_far_jump(const void *f);
364static void emit_far_call(const void *f);
365
366#ifdef VITA
367#include <psp2/kernel/sysmem.h>
368static int sceBlock;
369// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
370extern int getVMBlock();
371int _newlib_vm_size_user = sizeof(*ndrc);
372#endif
373
374static void mprotect_w_x(void *start, void *end, int is_x)
375{
376#ifdef NO_WRITE_EXEC
377 #if defined(VITA)
378 // *Open* enables write on all memory that was
379 // allocated by sceKernelAllocMemBlockForVM()?
380 if (is_x)
381 sceKernelCloseVMDomain();
382 else
383 sceKernelOpenVMDomain();
384 #else
385 u_long mstart = (u_long)start & ~4095ul;
386 u_long mend = (u_long)end;
387 if (mprotect((void *)mstart, mend - mstart,
388 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
389 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
390 #endif
391#endif
392}
393
394static void start_tcache_write(void *start, void *end)
395{
396 mprotect_w_x(start, end, 0);
397}
398
399static void end_tcache_write(void *start, void *end)
400{
401#if defined(__arm__) || defined(__aarch64__)
402 size_t len = (char *)end - (char *)start;
403 #if defined(__BLACKBERRY_QNX__)
404 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
405 #elif defined(__MACH__)
406 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
407 #elif defined(VITA)
408 sceKernelSyncVMDomain(sceBlock, start, len);
409 #elif defined(_3DS)
410 ctr_flush_invalidate_cache();
411 #elif defined(__aarch64__)
412 // as of 2021, __clear_cache() is still broken on arm64
413 // so here is a custom one :(
414 clear_cache_arm64(start, end);
415 #else
416 __clear_cache(start, end);
417 #endif
418 (void)len;
419#endif
420
421 mprotect_w_x(start, end, 1);
422}
423
424static void *start_block(void)
425{
426 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
427 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
428 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
429 start_tcache_write(out, end);
430 return out;
431}
432
433static void end_block(void *start)
434{
435 end_tcache_write(start, out);
436}
437
438// also takes care of w^x mappings when patching code
439static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
440
441static void mark_clear_cache(void *target)
442{
443 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
444 u_int mask = 1u << ((offset >> 12) & 31);
445 if (!(needs_clear_cache[offset >> 17] & mask)) {
446 char *start = (char *)((uintptr_t)target & ~4095l);
447 start_tcache_write(start, start + 4095);
448 needs_clear_cache[offset >> 17] |= mask;
449 }
450}
451
452// Clearing the cache is rather slow on ARM Linux, so mark the areas
453// that need to be cleared, and then only clear these areas once.
454static void do_clear_cache(void)
455{
456 int i, j;
457 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
458 {
459 u_int bitmap = needs_clear_cache[i];
460 if (!bitmap)
461 continue;
462 for (j = 0; j < 32; j++)
463 {
464 u_char *start, *end;
465 if (!(bitmap & (1<<j)))
466 continue;
467
468 start = ndrc->translation_cache + i*131072 + j*4096;
469 end = start + 4095;
470 for (j++; j < 32; j++) {
471 if (!(bitmap & (1<<j)))
472 break;
473 end += 4096;
474 }
475 end_tcache_write(start, end);
476 }
477 needs_clear_cache[i] = 0;
478 }
479}
480
481//#define DEBUG_CYCLE_COUNT 1
482
483#define NO_CYCLE_PENALTY_THR 12
484
485int cycle_multiplier = CYCLE_MULT_DEFAULT; // 100 for 1.0
486int cycle_multiplier_override;
487int cycle_multiplier_old;
488static int cycle_multiplier_active;
489
490static int CLOCK_ADJUST(int x)
491{
492 int m = cycle_multiplier_active;
493 int s = (x >> 31) | 1;
494 return (x * m + s * 50) / 100;
495}
496
497static int ds_writes_rjump_rs(int i)
498{
499 return dops[i].rs1 != 0 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2);
500}
501
502static u_int get_page(u_int vaddr)
503{
504 u_int page=vaddr&~0xe0000000;
505 if (page < 0x1000000)
506 page &= ~0x0e00000; // RAM mirrors
507 page>>=12;
508 if(page>2048) page=2048+(page&2047);
509 return page;
510}
511
512// no virtual mem in PCSX
513static u_int get_vpage(u_int vaddr)
514{
515 return get_page(vaddr);
516}
517
518static struct ht_entry *hash_table_get(u_int vaddr)
519{
520 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
521}
522
523static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
524{
525 ht_bin->vaddr[1] = ht_bin->vaddr[0];
526 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
527 ht_bin->vaddr[0] = vaddr;
528 ht_bin->tcaddr[0] = tcaddr;
529}
530
531// some messy ari64's code, seems to rely on unsigned 32bit overflow
532static int doesnt_expire_soon(void *tcaddr)
533{
534 u_int diff = (u_int)((u_char *)tcaddr - out) << (32-TARGET_SIZE_2);
535 return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
536}
537
538// Get address from virtual address
539// This is called from the recompiled JR/JALR instructions
540void noinline *get_addr(u_int vaddr)
541{
542 u_int page=get_page(vaddr);
543 u_int vpage=get_vpage(vaddr);
544 struct ll_entry *head;
545 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
546 head=jump_in[page];
547 while(head!=NULL) {
548 if(head->vaddr==vaddr) {
549 //printf("TRACE: count=%d next=%d (get_addr match %x: %p)\n",Count,next_interupt,vaddr,head->addr);
550 hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
551 return head->addr;
552 }
553 head=head->next;
554 }
555 head=jump_dirty[vpage];
556 while(head!=NULL) {
557 if(head->vaddr==vaddr) {
558 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %p)\n",Count,next_interupt,vaddr,head->addr);
559 // Don't restore blocks which are about to expire from the cache
560 if (doesnt_expire_soon(head->addr))
561 if (verify_dirty(head->addr)) {
562 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
563 invalid_code[vaddr>>12]=0;
564 inv_code_start=inv_code_end=~0;
565 if(vpage<2048) {
566 restore_candidate[vpage>>3]|=1<<(vpage&7);
567 }
568 else restore_candidate[page>>3]|=1<<(page&7);
569 struct ht_entry *ht_bin = hash_table_get(vaddr);
570 if (ht_bin->vaddr[0] == vaddr)
571 ht_bin->tcaddr[0] = head->addr; // Replace existing entry
572 else
573 hash_table_add(ht_bin, vaddr, head->addr);
574
575 return head->addr;
576 }
577 }
578 head=head->next;
579 }
580 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
581 int r=new_recompile_block(vaddr);
582 if(r==0) return get_addr(vaddr);
583 // Execute in unmapped page, generate pagefault execption
584 Status|=2;
585 Cause=(vaddr<<31)|0x8;
586 EPC=(vaddr&1)?vaddr-5:vaddr;
587 BadVAddr=(vaddr&~1);
588 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
589 EntryHi=BadVAddr&0xFFFFE000;
590 return get_addr_ht(0x80000000);
591}
592// Look up address in hash table first
593void *get_addr_ht(u_int vaddr)
594{
595 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
596 const struct ht_entry *ht_bin = hash_table_get(vaddr);
597 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
598 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
599 return get_addr(vaddr);
600}
601
602void clear_all_regs(signed char regmap[])
603{
604 int hr;
605 for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
606}
607
608static signed char get_reg(const signed char regmap[],int r)
609{
610 int hr;
611 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap[hr]==r) return hr;
612 return -1;
613}
614
615// Find a register that is available for two consecutive cycles
616static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
617{
618 int hr;
619 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
620 return -1;
621}
622
623int count_free_regs(signed char regmap[])
624{
625 int count=0;
626 int hr;
627 for(hr=0;hr<HOST_REGS;hr++)
628 {
629 if(hr!=EXCLUDE_REG) {
630 if(regmap[hr]<0) count++;
631 }
632 }
633 return count;
634}
635
636void dirty_reg(struct regstat *cur,signed char reg)
637{
638 int hr;
639 if(!reg) return;
640 for (hr=0;hr<HOST_REGS;hr++) {
641 if((cur->regmap[hr]&63)==reg) {
642 cur->dirty|=1<<hr;
643 }
644 }
645}
646
647static void set_const(struct regstat *cur, signed char reg, uint32_t value)
648{
649 int hr;
650 if(!reg) return;
651 for (hr=0;hr<HOST_REGS;hr++) {
652 if(cur->regmap[hr]==reg) {
653 cur->isconst|=1<<hr;
654 current_constmap[hr]=value;
655 }
656 }
657}
658
659static void clear_const(struct regstat *cur, signed char reg)
660{
661 int hr;
662 if(!reg) return;
663 for (hr=0;hr<HOST_REGS;hr++) {
664 if((cur->regmap[hr]&63)==reg) {
665 cur->isconst&=~(1<<hr);
666 }
667 }
668}
669
670static int is_const(struct regstat *cur, signed char reg)
671{
672 int hr;
673 if(reg<0) return 0;
674 if(!reg) return 1;
675 for (hr=0;hr<HOST_REGS;hr++) {
676 if((cur->regmap[hr]&63)==reg) {
677 return (cur->isconst>>hr)&1;
678 }
679 }
680 return 0;
681}
682
683static uint32_t get_const(struct regstat *cur, signed char reg)
684{
685 int hr;
686 if(!reg) return 0;
687 for (hr=0;hr<HOST_REGS;hr++) {
688 if(cur->regmap[hr]==reg) {
689 return current_constmap[hr];
690 }
691 }
692 SysPrintf("Unknown constant in r%d\n",reg);
693 abort();
694}
695
696// Least soon needed registers
697// Look at the next ten instructions and see which registers
698// will be used. Try not to reallocate these.
699void lsn(u_char hsn[], int i, int *preferred_reg)
700{
701 int j;
702 int b=-1;
703 for(j=0;j<9;j++)
704 {
705 if(i+j>=slen) {
706 j=slen-i-1;
707 break;
708 }
709 if (dops[i+j].is_ujump)
710 {
711 // Don't go past an unconditonal jump
712 j++;
713 break;
714 }
715 }
716 for(;j>=0;j--)
717 {
718 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
719 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
720 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
721 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
722 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
723 // Stores can allocate zero
724 hsn[dops[i+j].rs1]=j;
725 hsn[dops[i+j].rs2]=j;
726 }
727 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
728 hsn[ROREG] = j;
729 // On some architectures stores need invc_ptr
730 #if defined(HOST_IMM8)
731 if (dops[i+j].is_store)
732 hsn[INVCP] = j;
733 #endif
734 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
735 {
736 hsn[CCREG]=j;
737 b=j;
738 }
739 }
740 if(b>=0)
741 {
742 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
743 {
744 // Follow first branch
745 int t=(ba[i+b]-start)>>2;
746 j=7-b;if(t+j>=slen) j=slen-t-1;
747 for(;j>=0;j--)
748 {
749 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
750 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
751 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
752 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
753 }
754 }
755 // TODO: preferred register based on backward branch
756 }
757 // Delay slot should preferably not overwrite branch conditions or cycle count
758 if (i > 0 && dops[i-1].is_jump) {
759 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
760 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
761 hsn[CCREG]=1;
762 // ...or hash tables
763 hsn[RHASH]=1;
764 hsn[RHTBL]=1;
765 }
766 // Coprocessor load/store needs FTEMP, even if not declared
767 if(dops[i].itype==C2LS) {
768 hsn[FTEMP]=0;
769 }
770 // Load L/R also uses FTEMP as a temporary register
771 if(dops[i].itype==LOADLR) {
772 hsn[FTEMP]=0;
773 }
774 // Also SWL/SWR/SDL/SDR
775 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
776 hsn[FTEMP]=0;
777 }
778 // Don't remove the miniht registers
779 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
780 {
781 hsn[RHASH]=0;
782 hsn[RHTBL]=0;
783 }
784}
785
786// We only want to allocate registers if we're going to use them again soon
787int needed_again(int r, int i)
788{
789 int j;
790 int b=-1;
791 int rn=10;
792
793 if (i > 0 && dops[i-1].is_ujump)
794 {
795 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
796 return 0; // Don't need any registers if exiting the block
797 }
798 for(j=0;j<9;j++)
799 {
800 if(i+j>=slen) {
801 j=slen-i-1;
802 break;
803 }
804 if (dops[i+j].is_ujump)
805 {
806 // Don't go past an unconditonal jump
807 j++;
808 break;
809 }
810 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
811 {
812 break;
813 }
814 }
815 for(;j>=1;j--)
816 {
817 if(dops[i+j].rs1==r) rn=j;
818 if(dops[i+j].rs2==r) rn=j;
819 if((unneeded_reg[i+j]>>r)&1) rn=10;
820 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
821 {
822 b=j;
823 }
824 }
825 if(rn<10) return 1;
826 (void)b;
827 return 0;
828}
829
830// Try to match register allocations at the end of a loop with those
831// at the beginning
832int loop_reg(int i, int r, int hr)
833{
834 int j,k;
835 for(j=0;j<9;j++)
836 {
837 if(i+j>=slen) {
838 j=slen-i-1;
839 break;
840 }
841 if (dops[i+j].is_ujump)
842 {
843 // Don't go past an unconditonal jump
844 j++;
845 break;
846 }
847 }
848 k=0;
849 if(i>0){
850 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
851 k--;
852 }
853 for(;k<j;k++)
854 {
855 assert(r < 64);
856 if((unneeded_reg[i+k]>>r)&1) return hr;
857 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
858 {
859 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
860 {
861 int t=(ba[i+k]-start)>>2;
862 int reg=get_reg(regs[t].regmap_entry,r);
863 if(reg>=0) return reg;
864 //reg=get_reg(regs[t+1].regmap_entry,r);
865 //if(reg>=0) return reg;
866 }
867 }
868 }
869 return hr;
870}
871
872
873// Allocate every register, preserving source/target regs
874void alloc_all(struct regstat *cur,int i)
875{
876 int hr;
877
878 for(hr=0;hr<HOST_REGS;hr++) {
879 if(hr!=EXCLUDE_REG) {
880 if(((cur->regmap[hr]&63)!=dops[i].rs1)&&((cur->regmap[hr]&63)!=dops[i].rs2)&&
881 ((cur->regmap[hr]&63)!=dops[i].rt1)&&((cur->regmap[hr]&63)!=dops[i].rt2))
882 {
883 cur->regmap[hr]=-1;
884 cur->dirty&=~(1<<hr);
885 }
886 // Don't need zeros
887 if((cur->regmap[hr]&63)==0)
888 {
889 cur->regmap[hr]=-1;
890 cur->dirty&=~(1<<hr);
891 }
892 }
893 }
894}
895
896#ifndef NDEBUG
897static int host_tempreg_in_use;
898
899static void host_tempreg_acquire(void)
900{
901 assert(!host_tempreg_in_use);
902 host_tempreg_in_use = 1;
903}
904
905static void host_tempreg_release(void)
906{
907 host_tempreg_in_use = 0;
908}
909#else
910static void host_tempreg_acquire(void) {}
911static void host_tempreg_release(void) {}
912#endif
913
914#ifdef ASSEM_PRINT
915extern void gen_interupt();
916extern void do_insn_cmp();
917#define FUNCNAME(f) { f, " " #f }
918static const struct {
919 void *addr;
920 const char *name;
921} function_names[] = {
922 FUNCNAME(cc_interrupt),
923 FUNCNAME(gen_interupt),
924 FUNCNAME(get_addr_ht),
925 FUNCNAME(get_addr),
926 FUNCNAME(jump_handler_read8),
927 FUNCNAME(jump_handler_read16),
928 FUNCNAME(jump_handler_read32),
929 FUNCNAME(jump_handler_write8),
930 FUNCNAME(jump_handler_write16),
931 FUNCNAME(jump_handler_write32),
932 FUNCNAME(invalidate_addr),
933 FUNCNAME(jump_to_new_pc),
934 FUNCNAME(jump_break),
935 FUNCNAME(jump_break_ds),
936 FUNCNAME(jump_syscall),
937 FUNCNAME(jump_syscall_ds),
938 FUNCNAME(call_gteStall),
939 FUNCNAME(new_dyna_leave),
940 FUNCNAME(pcsx_mtc0),
941 FUNCNAME(pcsx_mtc0_ds),
942#ifdef DRC_DBG
943 FUNCNAME(do_insn_cmp),
944#endif
945#ifdef __arm__
946 FUNCNAME(verify_code),
947#endif
948};
949
950static const char *func_name(const void *a)
951{
952 int i;
953 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
954 if (function_names[i].addr == a)
955 return function_names[i].name;
956 return "";
957}
958#else
959#define func_name(x) ""
960#endif
961
962#ifdef __i386__
963#include "assem_x86.c"
964#endif
965#ifdef __x86_64__
966#include "assem_x64.c"
967#endif
968#ifdef __arm__
969#include "assem_arm.c"
970#endif
971#ifdef __aarch64__
972#include "assem_arm64.c"
973#endif
974
975static void *get_trampoline(const void *f)
976{
977 size_t i;
978
979 for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
980 if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
981 break;
982 }
983 if (i == ARRAY_SIZE(ndrc->tramp.f)) {
984 SysPrintf("trampoline table is full, last func %p\n", f);
985 abort();
986 }
987 if (ndrc->tramp.f[i] == NULL) {
988 start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
989 ndrc->tramp.f[i] = f;
990 end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
991 }
992 return &ndrc->tramp.ops[i];
993}
994
995static void emit_far_jump(const void *f)
996{
997 if (can_jump_or_call(f)) {
998 emit_jmp(f);
999 return;
1000 }
1001
1002 f = get_trampoline(f);
1003 emit_jmp(f);
1004}
1005
1006static void emit_far_call(const void *f)
1007{
1008 if (can_jump_or_call(f)) {
1009 emit_call(f);
1010 return;
1011 }
1012
1013 f = get_trampoline(f);
1014 emit_call(f);
1015}
1016
1017// Add virtual address mapping to linked list
1018void ll_add(struct ll_entry **head,int vaddr,void *addr)
1019{
1020 struct ll_entry *new_entry;
1021 new_entry=malloc(sizeof(struct ll_entry));
1022 assert(new_entry!=NULL);
1023 new_entry->vaddr=vaddr;
1024 new_entry->reg_sv_flags=0;
1025 new_entry->addr=addr;
1026 new_entry->next=*head;
1027 *head=new_entry;
1028}
1029
1030void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
1031{
1032 ll_add(head,vaddr,addr);
1033 (*head)->reg_sv_flags=reg_sv_flags;
1034}
1035
1036// Check if an address is already compiled
1037// but don't return addresses which are about to expire from the cache
1038void *check_addr(u_int vaddr)
1039{
1040 struct ht_entry *ht_bin = hash_table_get(vaddr);
1041 size_t i;
1042 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1043 if (ht_bin->vaddr[i] == vaddr)
1044 if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
1045 if (isclean(ht_bin->tcaddr[i]))
1046 return ht_bin->tcaddr[i];
1047 }
1048 u_int page=get_page(vaddr);
1049 struct ll_entry *head;
1050 head=jump_in[page];
1051 while (head != NULL) {
1052 if (head->vaddr == vaddr) {
1053 if (doesnt_expire_soon(head->addr)) {
1054 // Update existing entry with current address
1055 if (ht_bin->vaddr[0] == vaddr) {
1056 ht_bin->tcaddr[0] = head->addr;
1057 return head->addr;
1058 }
1059 if (ht_bin->vaddr[1] == vaddr) {
1060 ht_bin->tcaddr[1] = head->addr;
1061 return head->addr;
1062 }
1063 // Insert into hash table with low priority.
1064 // Don't evict existing entries, as they are probably
1065 // addresses that are being accessed frequently.
1066 if (ht_bin->vaddr[0] == -1) {
1067 ht_bin->vaddr[0] = vaddr;
1068 ht_bin->tcaddr[0] = head->addr;
1069 }
1070 else if (ht_bin->vaddr[1] == -1) {
1071 ht_bin->vaddr[1] = vaddr;
1072 ht_bin->tcaddr[1] = head->addr;
1073 }
1074 return head->addr;
1075 }
1076 }
1077 head=head->next;
1078 }
1079 return 0;
1080}
1081
1082void remove_hash(int vaddr)
1083{
1084 //printf("remove hash: %x\n",vaddr);
1085 struct ht_entry *ht_bin = hash_table_get(vaddr);
1086 if (ht_bin->vaddr[1] == vaddr) {
1087 ht_bin->vaddr[1] = -1;
1088 ht_bin->tcaddr[1] = NULL;
1089 }
1090 if (ht_bin->vaddr[0] == vaddr) {
1091 ht_bin->vaddr[0] = ht_bin->vaddr[1];
1092 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
1093 ht_bin->vaddr[1] = -1;
1094 ht_bin->tcaddr[1] = NULL;
1095 }
1096}
1097
1098static void ll_remove_matching_addrs(struct ll_entry **head,
1099 uintptr_t base_offs_s, int shift)
1100{
1101 struct ll_entry *next;
1102 while(*head) {
1103 uintptr_t o1 = (u_char *)(*head)->addr - ndrc->translation_cache;
1104 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1105 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1106 {
1107 inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
1108 remove_hash((*head)->vaddr);
1109 next=(*head)->next;
1110 free(*head);
1111 *head=next;
1112 }
1113 else
1114 {
1115 head=&((*head)->next);
1116 }
1117 }
1118}
1119
1120// Remove all entries from linked list
1121void ll_clear(struct ll_entry **head)
1122{
1123 struct ll_entry *cur;
1124 struct ll_entry *next;
1125 if((cur=*head)) {
1126 *head=0;
1127 while(cur) {
1128 next=cur->next;
1129 free(cur);
1130 cur=next;
1131 }
1132 }
1133}
1134
1135// Dereference the pointers and remove if it matches
1136static void ll_kill_pointers(struct ll_entry *head,
1137 uintptr_t base_offs_s, int shift)
1138{
1139 while(head) {
1140 u_char *ptr = get_pointer(head->addr);
1141 uintptr_t o1 = ptr - ndrc->translation_cache;
1142 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1143 inv_debug("EXP: Lookup pointer to %p at %p (%x)\n",ptr,head->addr,head->vaddr);
1144 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1145 {
1146 inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
1147 void *host_addr=find_extjump_insn(head->addr);
1148 mark_clear_cache(host_addr);
1149 set_jump_target(host_addr, head->addr);
1150 }
1151 head=head->next;
1152 }
1153}
1154
1155// This is called when we write to a compiled block (see do_invstub)
1156static void invalidate_page(u_int page)
1157{
1158 struct ll_entry *head;
1159 struct ll_entry *next;
1160 head=jump_in[page];
1161 jump_in[page]=0;
1162 while(head!=NULL) {
1163 inv_debug("INVALIDATE: %x\n",head->vaddr);
1164 remove_hash(head->vaddr);
1165 next=head->next;
1166 free(head);
1167 head=next;
1168 }
1169 head=jump_out[page];
1170 jump_out[page]=0;
1171 while(head!=NULL) {
1172 inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
1173 void *host_addr=find_extjump_insn(head->addr);
1174 mark_clear_cache(host_addr);
1175 set_jump_target(host_addr, head->addr); // point back to dyna_linker
1176 next=head->next;
1177 free(head);
1178 head=next;
1179 }
1180}
1181
1182static void invalidate_block_range(u_int block, u_int first, u_int last)
1183{
1184 u_int page=get_page(block<<12);
1185 //printf("first=%d last=%d\n",first,last);
1186 invalidate_page(page);
1187 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1188 assert(last<page+5);
1189 // Invalidate the adjacent pages if a block crosses a 4K boundary
1190 while(first<page) {
1191 invalidate_page(first);
1192 first++;
1193 }
1194 for(first=page+1;first<last;first++) {
1195 invalidate_page(first);
1196 }
1197 do_clear_cache();
1198
1199 // Don't trap writes
1200 invalid_code[block]=1;
1201
1202 #ifdef USE_MINI_HT
1203 memset(mini_ht,-1,sizeof(mini_ht));
1204 #endif
1205}
1206
1207void invalidate_block(u_int block)
1208{
1209 u_int page=get_page(block<<12);
1210 u_int vpage=get_vpage(block<<12);
1211 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1212 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1213 u_int first,last;
1214 first=last=page;
1215 struct ll_entry *head;
1216 head=jump_dirty[vpage];
1217 //printf("page=%d vpage=%d\n",page,vpage);
1218 while(head!=NULL) {
1219 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1220 u_char *start, *end;
1221 get_bounds(head->addr, &start, &end);
1222 //printf("start: %p end: %p\n", start, end);
1223 if (page < 2048 && start >= rdram && end < rdram+RAM_SIZE) {
1224 if (((start-rdram)>>12) <= page && ((end-1-rdram)>>12) >= page) {
1225 if ((((start-rdram)>>12)&2047) < first) first = ((start-rdram)>>12)&2047;
1226 if ((((end-1-rdram)>>12)&2047) > last) last = ((end-1-rdram)>>12)&2047;
1227 }
1228 }
1229 }
1230 head=head->next;
1231 }
1232 invalidate_block_range(block,first,last);
1233}
1234
1235void invalidate_addr(u_int addr)
1236{
1237 //static int rhits;
1238 // this check is done by the caller
1239 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1240 u_int page=get_vpage(addr);
1241 if(page<2048) { // RAM
1242 struct ll_entry *head;
1243 u_int addr_min=~0, addr_max=0;
1244 u_int mask=RAM_SIZE-1;
1245 u_int addr_main=0x80000000|(addr&mask);
1246 int pg1;
1247 inv_code_start=addr_main&~0xfff;
1248 inv_code_end=addr_main|0xfff;
1249 pg1=page;
1250 if (pg1>0) {
1251 // must check previous page too because of spans..
1252 pg1--;
1253 inv_code_start-=0x1000;
1254 }
1255 for(;pg1<=page;pg1++) {
1256 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1257 u_char *start_h, *end_h;
1258 u_int start, end;
1259 get_bounds(head->addr, &start_h, &end_h);
1260 start = (uintptr_t)start_h - ram_offset;
1261 end = (uintptr_t)end_h - ram_offset;
1262 if(start<=addr_main&&addr_main<end) {
1263 if(start<addr_min) addr_min=start;
1264 if(end>addr_max) addr_max=end;
1265 }
1266 else if(addr_main<start) {
1267 if(start<inv_code_end)
1268 inv_code_end=start-1;
1269 }
1270 else {
1271 if(end>inv_code_start)
1272 inv_code_start=end;
1273 }
1274 }
1275 }
1276 if (addr_min!=~0) {
1277 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1278 inv_code_start=inv_code_end=~0;
1279 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1280 return;
1281 }
1282 else {
1283 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1284 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1285 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1286 return;
1287 }
1288 }
1289 invalidate_block(addr>>12);
1290}
1291
1292// This is called when loading a save state.
1293// Anything could have changed, so invalidate everything.
1294void invalidate_all_pages(void)
1295{
1296 u_int page;
1297 for(page=0;page<4096;page++)
1298 invalidate_page(page);
1299 for(page=0;page<1048576;page++)
1300 if(!invalid_code[page]) {
1301 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1302 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1303 }
1304 #ifdef USE_MINI_HT
1305 memset(mini_ht,-1,sizeof(mini_ht));
1306 #endif
1307 do_clear_cache();
1308}
1309
1310static void do_invstub(int n)
1311{
1312 literal_pool(20);
1313 u_int reglist=stubs[n].a;
1314 set_jump_target(stubs[n].addr, out);
1315 save_regs(reglist);
1316 if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
1317 emit_far_call(invalidate_addr);
1318 restore_regs(reglist);
1319 emit_jmp(stubs[n].retaddr); // return address
1320}
1321
1322// Add an entry to jump_out after making a link
1323// src should point to code by emit_extjump2()
1324void add_jump_out(u_int vaddr,void *src)
1325{
1326 u_int page=get_page(vaddr);
1327 inv_debug("add_jump_out: %p -> %x (%d)\n",src,vaddr,page);
1328 check_extjump2(src);
1329 ll_add(jump_out+page,vaddr,src);
1330 //inv_debug("add_jump_out: to %p\n",get_pointer(src));
1331}
1332
1333// If a code block was found to be unmodified (bit was set in
1334// restore_candidate) and it remains unmodified (bit is clear
1335// in invalid_code) then move the entries for that 4K page from
1336// the dirty list to the clean list.
1337void clean_blocks(u_int page)
1338{
1339 struct ll_entry *head;
1340 inv_debug("INV: clean_blocks page=%d\n",page);
1341 head=jump_dirty[page];
1342 while(head!=NULL) {
1343 if(!invalid_code[head->vaddr>>12]) {
1344 // Don't restore blocks which are about to expire from the cache
1345 if (doesnt_expire_soon(head->addr)) {
1346 if(verify_dirty(head->addr)) {
1347 u_char *start, *end;
1348 //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
1349 u_int i;
1350 u_int inv=0;
1351 get_bounds(head->addr, &start, &end);
1352 if (start - rdram < RAM_SIZE) {
1353 for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
1354 inv|=invalid_code[i];
1355 }
1356 }
1357 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1358 inv=1;
1359 }
1360 if(!inv) {
1361 void *clean_addr = get_clean_addr(head->addr);
1362 if (doesnt_expire_soon(clean_addr)) {
1363 u_int ppage=page;
1364 inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, clean_addr);
1365 //printf("page=%x, addr=%x\n",page,head->vaddr);
1366 //assert(head->vaddr>>12==(page|0x80000));
1367 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1368 struct ht_entry *ht_bin = hash_table_get(head->vaddr);
1369 if (ht_bin->vaddr[0] == head->vaddr)
1370 ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
1371 if (ht_bin->vaddr[1] == head->vaddr)
1372 ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
1373 }
1374 }
1375 }
1376 }
1377 }
1378 head=head->next;
1379 }
1380}
1381
1382/* Register allocation */
1383
1384// Note: registers are allocated clean (unmodified state)
1385// if you intend to modify the register, you must call dirty_reg().
1386static void alloc_reg(struct regstat *cur,int i,signed char reg)
1387{
1388 int r,hr;
1389 int preferred_reg = PREFERRED_REG_FIRST
1390 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1391 if (reg == CCREG) preferred_reg = HOST_CCREG;
1392 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1393 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1394
1395 // Don't allocate unused registers
1396 if((cur->u>>reg)&1) return;
1397
1398 // see if it's already allocated
1399 for(hr=0;hr<HOST_REGS;hr++)
1400 {
1401 if(cur->regmap[hr]==reg) return;
1402 }
1403
1404 // Keep the same mapping if the register was already allocated in a loop
1405 preferred_reg = loop_reg(i,reg,preferred_reg);
1406
1407 // Try to allocate the preferred register
1408 if(cur->regmap[preferred_reg]==-1) {
1409 cur->regmap[preferred_reg]=reg;
1410 cur->dirty&=~(1<<preferred_reg);
1411 cur->isconst&=~(1<<preferred_reg);
1412 return;
1413 }
1414 r=cur->regmap[preferred_reg];
1415 assert(r < 64);
1416 if((cur->u>>r)&1) {
1417 cur->regmap[preferred_reg]=reg;
1418 cur->dirty&=~(1<<preferred_reg);
1419 cur->isconst&=~(1<<preferred_reg);
1420 return;
1421 }
1422
1423 // Clear any unneeded registers
1424 // We try to keep the mapping consistent, if possible, because it
1425 // makes branches easier (especially loops). So we try to allocate
1426 // first (see above) before removing old mappings. If this is not
1427 // possible then go ahead and clear out the registers that are no
1428 // longer needed.
1429 for(hr=0;hr<HOST_REGS;hr++)
1430 {
1431 r=cur->regmap[hr];
1432 if(r>=0) {
1433 assert(r < 64);
1434 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1435 }
1436 }
1437
1438 // Try to allocate any available register, but prefer
1439 // registers that have not been used recently.
1440 if (i > 0) {
1441 for (hr = PREFERRED_REG_FIRST; ; ) {
1442 if (cur->regmap[hr] < 0) {
1443 int oldreg = regs[i-1].regmap[hr];
1444 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1445 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1446 {
1447 cur->regmap[hr]=reg;
1448 cur->dirty&=~(1<<hr);
1449 cur->isconst&=~(1<<hr);
1450 return;
1451 }
1452 }
1453 hr++;
1454 if (hr == EXCLUDE_REG)
1455 hr++;
1456 if (hr == HOST_REGS)
1457 hr = 0;
1458 if (hr == PREFERRED_REG_FIRST)
1459 break;
1460 }
1461 }
1462
1463 // Try to allocate any available register
1464 for (hr = PREFERRED_REG_FIRST; ; ) {
1465 if (cur->regmap[hr] < 0) {
1466 cur->regmap[hr]=reg;
1467 cur->dirty&=~(1<<hr);
1468 cur->isconst&=~(1<<hr);
1469 return;
1470 }
1471 hr++;
1472 if (hr == EXCLUDE_REG)
1473 hr++;
1474 if (hr == HOST_REGS)
1475 hr = 0;
1476 if (hr == PREFERRED_REG_FIRST)
1477 break;
1478 }
1479
1480 // Ok, now we have to evict someone
1481 // Pick a register we hopefully won't need soon
1482 u_char hsn[MAXREG+1];
1483 memset(hsn,10,sizeof(hsn));
1484 int j;
1485 lsn(hsn,i,&preferred_reg);
1486 //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]);
1487 //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]);
1488 if(i>0) {
1489 // Don't evict the cycle count at entry points, otherwise the entry
1490 // stub will have to write it.
1491 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1492 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1493 for(j=10;j>=3;j--)
1494 {
1495 // Alloc preferred register if available
1496 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1497 for(hr=0;hr<HOST_REGS;hr++) {
1498 // Evict both parts of a 64-bit register
1499 if((cur->regmap[hr]&63)==r) {
1500 cur->regmap[hr]=-1;
1501 cur->dirty&=~(1<<hr);
1502 cur->isconst&=~(1<<hr);
1503 }
1504 }
1505 cur->regmap[preferred_reg]=reg;
1506 return;
1507 }
1508 for(r=1;r<=MAXREG;r++)
1509 {
1510 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1511 for(hr=0;hr<HOST_REGS;hr++) {
1512 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1513 if(cur->regmap[hr]==r) {
1514 cur->regmap[hr]=reg;
1515 cur->dirty&=~(1<<hr);
1516 cur->isconst&=~(1<<hr);
1517 return;
1518 }
1519 }
1520 }
1521 }
1522 }
1523 }
1524 }
1525 for(j=10;j>=0;j--)
1526 {
1527 for(r=1;r<=MAXREG;r++)
1528 {
1529 if(hsn[r]==j) {
1530 for(hr=0;hr<HOST_REGS;hr++) {
1531 if(cur->regmap[hr]==r) {
1532 cur->regmap[hr]=reg;
1533 cur->dirty&=~(1<<hr);
1534 cur->isconst&=~(1<<hr);
1535 return;
1536 }
1537 }
1538 }
1539 }
1540 }
1541 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1542}
1543
1544// Allocate a temporary register. This is done without regard to
1545// dirty status or whether the register we request is on the unneeded list
1546// Note: This will only allocate one register, even if called multiple times
1547static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1548{
1549 int r,hr;
1550 int preferred_reg = -1;
1551
1552 // see if it's already allocated
1553 for(hr=0;hr<HOST_REGS;hr++)
1554 {
1555 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1556 }
1557
1558 // Try to allocate any available register
1559 for(hr=HOST_REGS-1;hr>=0;hr--) {
1560 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1561 cur->regmap[hr]=reg;
1562 cur->dirty&=~(1<<hr);
1563 cur->isconst&=~(1<<hr);
1564 return;
1565 }
1566 }
1567
1568 // Find an unneeded register
1569 for(hr=HOST_REGS-1;hr>=0;hr--)
1570 {
1571 r=cur->regmap[hr];
1572 if(r>=0) {
1573 assert(r < 64);
1574 if((cur->u>>r)&1) {
1575 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1576 cur->regmap[hr]=reg;
1577 cur->dirty&=~(1<<hr);
1578 cur->isconst&=~(1<<hr);
1579 return;
1580 }
1581 }
1582 }
1583 }
1584
1585 // Ok, now we have to evict someone
1586 // Pick a register we hopefully won't need soon
1587 // TODO: we might want to follow unconditional jumps here
1588 // TODO: get rid of dupe code and make this into a function
1589 u_char hsn[MAXREG+1];
1590 memset(hsn,10,sizeof(hsn));
1591 int j;
1592 lsn(hsn,i,&preferred_reg);
1593 //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]);
1594 if(i>0) {
1595 // Don't evict the cycle count at entry points, otherwise the entry
1596 // stub will have to write it.
1597 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1598 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1599 for(j=10;j>=3;j--)
1600 {
1601 for(r=1;r<=MAXREG;r++)
1602 {
1603 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1604 for(hr=0;hr<HOST_REGS;hr++) {
1605 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1606 if(cur->regmap[hr]==r) {
1607 cur->regmap[hr]=reg;
1608 cur->dirty&=~(1<<hr);
1609 cur->isconst&=~(1<<hr);
1610 return;
1611 }
1612 }
1613 }
1614 }
1615 }
1616 }
1617 }
1618 for(j=10;j>=0;j--)
1619 {
1620 for(r=1;r<=MAXREG;r++)
1621 {
1622 if(hsn[r]==j) {
1623 for(hr=0;hr<HOST_REGS;hr++) {
1624 if(cur->regmap[hr]==r) {
1625 cur->regmap[hr]=reg;
1626 cur->dirty&=~(1<<hr);
1627 cur->isconst&=~(1<<hr);
1628 return;
1629 }
1630 }
1631 }
1632 }
1633 }
1634 SysPrintf("This shouldn't happen");abort();
1635}
1636
1637static void mov_alloc(struct regstat *current,int i)
1638{
1639 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1640 alloc_cc(current,i); // for stalls
1641 dirty_reg(current,CCREG);
1642 }
1643
1644 // Note: Don't need to actually alloc the source registers
1645 //alloc_reg(current,i,dops[i].rs1);
1646 alloc_reg(current,i,dops[i].rt1);
1647
1648 clear_const(current,dops[i].rs1);
1649 clear_const(current,dops[i].rt1);
1650 dirty_reg(current,dops[i].rt1);
1651}
1652
1653static void shiftimm_alloc(struct regstat *current,int i)
1654{
1655 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1656 {
1657 if(dops[i].rt1) {
1658 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1659 else dops[i].lt1=dops[i].rs1;
1660 alloc_reg(current,i,dops[i].rt1);
1661 dirty_reg(current,dops[i].rt1);
1662 if(is_const(current,dops[i].rs1)) {
1663 int v=get_const(current,dops[i].rs1);
1664 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1665 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1666 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1667 }
1668 else clear_const(current,dops[i].rt1);
1669 }
1670 }
1671 else
1672 {
1673 clear_const(current,dops[i].rs1);
1674 clear_const(current,dops[i].rt1);
1675 }
1676
1677 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1678 {
1679 assert(0);
1680 }
1681 if(dops[i].opcode2==0x3c) // DSLL32
1682 {
1683 assert(0);
1684 }
1685 if(dops[i].opcode2==0x3e) // DSRL32
1686 {
1687 assert(0);
1688 }
1689 if(dops[i].opcode2==0x3f) // DSRA32
1690 {
1691 assert(0);
1692 }
1693}
1694
1695static void shift_alloc(struct regstat *current,int i)
1696{
1697 if(dops[i].rt1) {
1698 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1699 {
1700 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1701 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1702 alloc_reg(current,i,dops[i].rt1);
1703 if(dops[i].rt1==dops[i].rs2) {
1704 alloc_reg_temp(current,i,-1);
1705 minimum_free_regs[i]=1;
1706 }
1707 } else { // DSLLV/DSRLV/DSRAV
1708 assert(0);
1709 }
1710 clear_const(current,dops[i].rs1);
1711 clear_const(current,dops[i].rs2);
1712 clear_const(current,dops[i].rt1);
1713 dirty_reg(current,dops[i].rt1);
1714 }
1715}
1716
1717static void alu_alloc(struct regstat *current,int i)
1718{
1719 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1720 if(dops[i].rt1) {
1721 if(dops[i].rs1&&dops[i].rs2) {
1722 alloc_reg(current,i,dops[i].rs1);
1723 alloc_reg(current,i,dops[i].rs2);
1724 }
1725 else {
1726 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1727 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1728 }
1729 alloc_reg(current,i,dops[i].rt1);
1730 }
1731 }
1732 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1733 if(dops[i].rt1) {
1734 alloc_reg(current,i,dops[i].rs1);
1735 alloc_reg(current,i,dops[i].rs2);
1736 alloc_reg(current,i,dops[i].rt1);
1737 }
1738 }
1739 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1740 if(dops[i].rt1) {
1741 if(dops[i].rs1&&dops[i].rs2) {
1742 alloc_reg(current,i,dops[i].rs1);
1743 alloc_reg(current,i,dops[i].rs2);
1744 }
1745 else
1746 {
1747 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1748 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1749 }
1750 alloc_reg(current,i,dops[i].rt1);
1751 }
1752 }
1753 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1754 assert(0);
1755 }
1756 clear_const(current,dops[i].rs1);
1757 clear_const(current,dops[i].rs2);
1758 clear_const(current,dops[i].rt1);
1759 dirty_reg(current,dops[i].rt1);
1760}
1761
1762static void imm16_alloc(struct regstat *current,int i)
1763{
1764 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1765 else dops[i].lt1=dops[i].rs1;
1766 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1767 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1768 assert(0);
1769 }
1770 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1771 clear_const(current,dops[i].rs1);
1772 clear_const(current,dops[i].rt1);
1773 }
1774 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1775 if(is_const(current,dops[i].rs1)) {
1776 int v=get_const(current,dops[i].rs1);
1777 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1778 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1779 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1780 }
1781 else clear_const(current,dops[i].rt1);
1782 }
1783 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1784 if(is_const(current,dops[i].rs1)) {
1785 int v=get_const(current,dops[i].rs1);
1786 set_const(current,dops[i].rt1,v+imm[i]);
1787 }
1788 else clear_const(current,dops[i].rt1);
1789 }
1790 else {
1791 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1792 }
1793 dirty_reg(current,dops[i].rt1);
1794}
1795
1796static void load_alloc(struct regstat *current,int i)
1797{
1798 clear_const(current,dops[i].rt1);
1799 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
1800 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
1801 if (needed_again(dops[i].rs1, i))
1802 alloc_reg(current, i, dops[i].rs1);
1803 if (ram_offset)
1804 alloc_reg(current, i, ROREG);
1805 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
1806 alloc_reg(current,i,dops[i].rt1);
1807 assert(get_reg(current->regmap,dops[i].rt1)>=0);
1808 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
1809 {
1810 assert(0);
1811 }
1812 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1813 {
1814 assert(0);
1815 }
1816 dirty_reg(current,dops[i].rt1);
1817 // LWL/LWR need a temporary register for the old value
1818 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1819 {
1820 alloc_reg(current,i,FTEMP);
1821 alloc_reg_temp(current,i,-1);
1822 minimum_free_regs[i]=1;
1823 }
1824 }
1825 else
1826 {
1827 // Load to r0 or unneeded register (dummy load)
1828 // but we still need a register to calculate the address
1829 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1830 {
1831 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1832 }
1833 alloc_reg_temp(current,i,-1);
1834 minimum_free_regs[i]=1;
1835 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1836 {
1837 assert(0);
1838 }
1839 }
1840}
1841
1842void store_alloc(struct regstat *current,int i)
1843{
1844 clear_const(current,dops[i].rs2);
1845 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
1846 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1847 alloc_reg(current,i,dops[i].rs2);
1848 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
1849 assert(0);
1850 }
1851 if (ram_offset)
1852 alloc_reg(current, i, ROREG);
1853 #if defined(HOST_IMM8)
1854 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1855 alloc_reg(current, i, INVCP);
1856 #endif
1857 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
1858 alloc_reg(current,i,FTEMP);
1859 }
1860 // We need a temporary register for address generation
1861 alloc_reg_temp(current,i,-1);
1862 minimum_free_regs[i]=1;
1863}
1864
1865void c1ls_alloc(struct regstat *current,int i)
1866{
1867 clear_const(current,dops[i].rt1);
1868 alloc_reg(current,i,CSREG); // Status
1869}
1870
1871void c2ls_alloc(struct regstat *current,int i)
1872{
1873 clear_const(current,dops[i].rt1);
1874 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1875 alloc_reg(current,i,FTEMP);
1876 if (ram_offset)
1877 alloc_reg(current, i, ROREG);
1878 #if defined(HOST_IMM8)
1879 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1880 if (dops[i].opcode == 0x3a) // SWC2
1881 alloc_reg(current,i,INVCP);
1882 #endif
1883 // We need a temporary register for address generation
1884 alloc_reg_temp(current,i,-1);
1885 minimum_free_regs[i]=1;
1886}
1887
1888#ifndef multdiv_alloc
1889void multdiv_alloc(struct regstat *current,int i)
1890{
1891 // case 0x18: MULT
1892 // case 0x19: MULTU
1893 // case 0x1A: DIV
1894 // case 0x1B: DIVU
1895 // case 0x1C: DMULT
1896 // case 0x1D: DMULTU
1897 // case 0x1E: DDIV
1898 // case 0x1F: DDIVU
1899 clear_const(current,dops[i].rs1);
1900 clear_const(current,dops[i].rs2);
1901 alloc_cc(current,i); // for stalls
1902 if(dops[i].rs1&&dops[i].rs2)
1903 {
1904 if((dops[i].opcode2&4)==0) // 32-bit
1905 {
1906 current->u&=~(1LL<<HIREG);
1907 current->u&=~(1LL<<LOREG);
1908 alloc_reg(current,i,HIREG);
1909 alloc_reg(current,i,LOREG);
1910 alloc_reg(current,i,dops[i].rs1);
1911 alloc_reg(current,i,dops[i].rs2);
1912 dirty_reg(current,HIREG);
1913 dirty_reg(current,LOREG);
1914 }
1915 else // 64-bit
1916 {
1917 assert(0);
1918 }
1919 }
1920 else
1921 {
1922 // Multiply by zero is zero.
1923 // MIPS does not have a divide by zero exception.
1924 // The result is undefined, we return zero.
1925 alloc_reg(current,i,HIREG);
1926 alloc_reg(current,i,LOREG);
1927 dirty_reg(current,HIREG);
1928 dirty_reg(current,LOREG);
1929 }
1930}
1931#endif
1932
1933void cop0_alloc(struct regstat *current,int i)
1934{
1935 if(dops[i].opcode2==0) // MFC0
1936 {
1937 if(dops[i].rt1) {
1938 clear_const(current,dops[i].rt1);
1939 alloc_all(current,i);
1940 alloc_reg(current,i,dops[i].rt1);
1941 dirty_reg(current,dops[i].rt1);
1942 }
1943 }
1944 else if(dops[i].opcode2==4) // MTC0
1945 {
1946 if(dops[i].rs1){
1947 clear_const(current,dops[i].rs1);
1948 alloc_reg(current,i,dops[i].rs1);
1949 alloc_all(current,i);
1950 }
1951 else {
1952 alloc_all(current,i); // FIXME: Keep r0
1953 current->u&=~1LL;
1954 alloc_reg(current,i,0);
1955 }
1956 }
1957 else
1958 {
1959 // TLBR/TLBWI/TLBWR/TLBP/ERET
1960 assert(dops[i].opcode2==0x10);
1961 alloc_all(current,i);
1962 }
1963 minimum_free_regs[i]=HOST_REGS;
1964}
1965
1966static void cop2_alloc(struct regstat *current,int i)
1967{
1968 if (dops[i].opcode2 < 3) // MFC2/CFC2
1969 {
1970 alloc_cc(current,i); // for stalls
1971 dirty_reg(current,CCREG);
1972 if(dops[i].rt1){
1973 clear_const(current,dops[i].rt1);
1974 alloc_reg(current,i,dops[i].rt1);
1975 dirty_reg(current,dops[i].rt1);
1976 }
1977 }
1978 else if (dops[i].opcode2 > 3) // MTC2/CTC2
1979 {
1980 if(dops[i].rs1){
1981 clear_const(current,dops[i].rs1);
1982 alloc_reg(current,i,dops[i].rs1);
1983 }
1984 else {
1985 current->u&=~1LL;
1986 alloc_reg(current,i,0);
1987 }
1988 }
1989 alloc_reg_temp(current,i,-1);
1990 minimum_free_regs[i]=1;
1991}
1992
1993void c2op_alloc(struct regstat *current,int i)
1994{
1995 alloc_cc(current,i); // for stalls
1996 dirty_reg(current,CCREG);
1997 alloc_reg_temp(current,i,-1);
1998}
1999
2000void syscall_alloc(struct regstat *current,int i)
2001{
2002 alloc_cc(current,i);
2003 dirty_reg(current,CCREG);
2004 alloc_all(current,i);
2005 minimum_free_regs[i]=HOST_REGS;
2006 current->isconst=0;
2007}
2008
2009void delayslot_alloc(struct regstat *current,int i)
2010{
2011 switch(dops[i].itype) {
2012 case UJUMP:
2013 case CJUMP:
2014 case SJUMP:
2015 case RJUMP:
2016 case SYSCALL:
2017 case HLECALL:
2018 case SPAN:
2019 assem_debug("jump in the delay slot. this shouldn't happen.\n");//abort();
2020 SysPrintf("Disabled speculative precompilation\n");
2021 stop_after_jal=1;
2022 break;
2023 case IMM16:
2024 imm16_alloc(current,i);
2025 break;
2026 case LOAD:
2027 case LOADLR:
2028 load_alloc(current,i);
2029 break;
2030 case STORE:
2031 case STORELR:
2032 store_alloc(current,i);
2033 break;
2034 case ALU:
2035 alu_alloc(current,i);
2036 break;
2037 case SHIFT:
2038 shift_alloc(current,i);
2039 break;
2040 case MULTDIV:
2041 multdiv_alloc(current,i);
2042 break;
2043 case SHIFTIMM:
2044 shiftimm_alloc(current,i);
2045 break;
2046 case MOV:
2047 mov_alloc(current,i);
2048 break;
2049 case COP0:
2050 cop0_alloc(current,i);
2051 break;
2052 case COP1:
2053 break;
2054 case COP2:
2055 cop2_alloc(current,i);
2056 break;
2057 case C1LS:
2058 c1ls_alloc(current,i);
2059 break;
2060 case C2LS:
2061 c2ls_alloc(current,i);
2062 break;
2063 case C2OP:
2064 c2op_alloc(current,i);
2065 break;
2066 }
2067}
2068
2069// Special case where a branch and delay slot span two pages in virtual memory
2070static void pagespan_alloc(struct regstat *current,int i)
2071{
2072 current->isconst=0;
2073 current->wasconst=0;
2074 regs[i].wasconst=0;
2075 minimum_free_regs[i]=HOST_REGS;
2076 alloc_all(current,i);
2077 alloc_cc(current,i);
2078 dirty_reg(current,CCREG);
2079 if(dops[i].opcode==3) // JAL
2080 {
2081 alloc_reg(current,i,31);
2082 dirty_reg(current,31);
2083 }
2084 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
2085 {
2086 alloc_reg(current,i,dops[i].rs1);
2087 if (dops[i].rt1!=0) {
2088 alloc_reg(current,i,dops[i].rt1);
2089 dirty_reg(current,dops[i].rt1);
2090 }
2091 }
2092 if((dops[i].opcode&0x2E)==4) // BEQ/BNE/BEQL/BNEL
2093 {
2094 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2095 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
2096 }
2097 else
2098 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
2099 {
2100 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2101 }
2102 //else ...
2103}
2104
2105static void add_stub(enum stub_type type, void *addr, void *retaddr,
2106 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2107{
2108 assert(stubcount < ARRAY_SIZE(stubs));
2109 stubs[stubcount].type = type;
2110 stubs[stubcount].addr = addr;
2111 stubs[stubcount].retaddr = retaddr;
2112 stubs[stubcount].a = a;
2113 stubs[stubcount].b = b;
2114 stubs[stubcount].c = c;
2115 stubs[stubcount].d = d;
2116 stubs[stubcount].e = e;
2117 stubcount++;
2118}
2119
2120static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2121 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2122{
2123 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2124}
2125
2126// Write out a single register
2127static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2128{
2129 int hr;
2130 for(hr=0;hr<HOST_REGS;hr++) {
2131 if(hr!=EXCLUDE_REG) {
2132 if((regmap[hr]&63)==r) {
2133 if((dirty>>hr)&1) {
2134 assert(regmap[hr]<64);
2135 emit_storereg(r,hr);
2136 }
2137 }
2138 }
2139 }
2140}
2141
2142static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2143{
2144 //if(dirty_pre==dirty) return;
2145 int hr,reg;
2146 for(hr=0;hr<HOST_REGS;hr++) {
2147 if(hr!=EXCLUDE_REG) {
2148 reg=pre[hr];
2149 if(((~u)>>(reg&63))&1) {
2150 if(reg>0) {
2151 if(((dirty_pre&~dirty)>>hr)&1) {
2152 if(reg>0&&reg<34) {
2153 emit_storereg(reg,hr);
2154 }
2155 else if(reg>=64) {
2156 assert(0);
2157 }
2158 }
2159 }
2160 }
2161 }
2162 }
2163}
2164
2165// trashes r2
2166static void pass_args(int a0, int a1)
2167{
2168 if(a0==1&&a1==0) {
2169 // must swap
2170 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2171 }
2172 else if(a0!=0&&a1==0) {
2173 emit_mov(a1,1);
2174 if (a0>=0) emit_mov(a0,0);
2175 }
2176 else {
2177 if(a0>=0&&a0!=0) emit_mov(a0,0);
2178 if(a1>=0&&a1!=1) emit_mov(a1,1);
2179 }
2180}
2181
2182static void alu_assemble(int i, const struct regstat *i_regs)
2183{
2184 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2185 if(dops[i].rt1) {
2186 signed char s1,s2,t;
2187 t=get_reg(i_regs->regmap,dops[i].rt1);
2188 if(t>=0) {
2189 s1=get_reg(i_regs->regmap,dops[i].rs1);
2190 s2=get_reg(i_regs->regmap,dops[i].rs2);
2191 if(dops[i].rs1&&dops[i].rs2) {
2192 assert(s1>=0);
2193 assert(s2>=0);
2194 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2195 else emit_add(s1,s2,t);
2196 }
2197 else if(dops[i].rs1) {
2198 if(s1>=0) emit_mov(s1,t);
2199 else emit_loadreg(dops[i].rs1,t);
2200 }
2201 else if(dops[i].rs2) {
2202 if(s2>=0) {
2203 if(dops[i].opcode2&2) emit_neg(s2,t);
2204 else emit_mov(s2,t);
2205 }
2206 else {
2207 emit_loadreg(dops[i].rs2,t);
2208 if(dops[i].opcode2&2) emit_neg(t,t);
2209 }
2210 }
2211 else emit_zeroreg(t);
2212 }
2213 }
2214 }
2215 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2216 assert(0);
2217 }
2218 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2219 if(dops[i].rt1) {
2220 signed char s1l,s2l,t;
2221 {
2222 t=get_reg(i_regs->regmap,dops[i].rt1);
2223 //assert(t>=0);
2224 if(t>=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].rs2==0) // rx<r0
2228 {
2229 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2230 assert(s1l>=0);
2231 emit_shrimm(s1l,31,t);
2232 }
2233 else // SLTU (unsigned can not be less than zero, 0<0)
2234 emit_zeroreg(t);
2235 }
2236 else if(dops[i].rs1==0) // r0<rx
2237 {
2238 assert(s2l>=0);
2239 if(dops[i].opcode2==0x2a) // SLT
2240 emit_set_gz32(s2l,t);
2241 else // SLTU (set if not zero)
2242 emit_set_nz32(s2l,t);
2243 }
2244 else{
2245 assert(s1l>=0);assert(s2l>=0);
2246 if(dops[i].opcode2==0x2a) // SLT
2247 emit_set_if_less32(s1l,s2l,t);
2248 else // SLTU
2249 emit_set_if_carry32(s1l,s2l,t);
2250 }
2251 }
2252 }
2253 }
2254 }
2255 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2256 if(dops[i].rt1) {
2257 signed char s1l,s2l,tl;
2258 tl=get_reg(i_regs->regmap,dops[i].rt1);
2259 {
2260 if(tl>=0) {
2261 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2262 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2263 if(dops[i].rs1&&dops[i].rs2) {
2264 assert(s1l>=0);
2265 assert(s2l>=0);
2266 if(dops[i].opcode2==0x24) { // AND
2267 emit_and(s1l,s2l,tl);
2268 } else
2269 if(dops[i].opcode2==0x25) { // OR
2270 emit_or(s1l,s2l,tl);
2271 } else
2272 if(dops[i].opcode2==0x26) { // XOR
2273 emit_xor(s1l,s2l,tl);
2274 } else
2275 if(dops[i].opcode2==0x27) { // NOR
2276 emit_or(s1l,s2l,tl);
2277 emit_not(tl,tl);
2278 }
2279 }
2280 else
2281 {
2282 if(dops[i].opcode2==0x24) { // AND
2283 emit_zeroreg(tl);
2284 } else
2285 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2286 if(dops[i].rs1){
2287 if(s1l>=0) emit_mov(s1l,tl);
2288 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2289 }
2290 else
2291 if(dops[i].rs2){
2292 if(s2l>=0) emit_mov(s2l,tl);
2293 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2294 }
2295 else emit_zeroreg(tl);
2296 } else
2297 if(dops[i].opcode2==0x27) { // NOR
2298 if(dops[i].rs1){
2299 if(s1l>=0) emit_not(s1l,tl);
2300 else {
2301 emit_loadreg(dops[i].rs1,tl);
2302 emit_not(tl,tl);
2303 }
2304 }
2305 else
2306 if(dops[i].rs2){
2307 if(s2l>=0) emit_not(s2l,tl);
2308 else {
2309 emit_loadreg(dops[i].rs2,tl);
2310 emit_not(tl,tl);
2311 }
2312 }
2313 else emit_movimm(-1,tl);
2314 }
2315 }
2316 }
2317 }
2318 }
2319 }
2320}
2321
2322static void imm16_assemble(int i, const struct regstat *i_regs)
2323{
2324 if (dops[i].opcode==0x0f) { // LUI
2325 if(dops[i].rt1) {
2326 signed char t;
2327 t=get_reg(i_regs->regmap,dops[i].rt1);
2328 //assert(t>=0);
2329 if(t>=0) {
2330 if(!((i_regs->isconst>>t)&1))
2331 emit_movimm(imm[i]<<16,t);
2332 }
2333 }
2334 }
2335 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2336 if(dops[i].rt1) {
2337 signed char s,t;
2338 t=get_reg(i_regs->regmap,dops[i].rt1);
2339 s=get_reg(i_regs->regmap,dops[i].rs1);
2340 if(dops[i].rs1) {
2341 //assert(t>=0);
2342 //assert(s>=0);
2343 if(t>=0) {
2344 if(!((i_regs->isconst>>t)&1)) {
2345 if(s<0) {
2346 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2347 emit_addimm(t,imm[i],t);
2348 }else{
2349 if(!((i_regs->wasconst>>s)&1))
2350 emit_addimm(s,imm[i],t);
2351 else
2352 emit_movimm(constmap[i][s]+imm[i],t);
2353 }
2354 }
2355 }
2356 } else {
2357 if(t>=0) {
2358 if(!((i_regs->isconst>>t)&1))
2359 emit_movimm(imm[i],t);
2360 }
2361 }
2362 }
2363 }
2364 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2365 if(dops[i].rt1) {
2366 signed char sl,tl;
2367 tl=get_reg(i_regs->regmap,dops[i].rt1);
2368 sl=get_reg(i_regs->regmap,dops[i].rs1);
2369 if(tl>=0) {
2370 if(dops[i].rs1) {
2371 assert(sl>=0);
2372 emit_addimm(sl,imm[i],tl);
2373 } else {
2374 emit_movimm(imm[i],tl);
2375 }
2376 }
2377 }
2378 }
2379 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2380 if(dops[i].rt1) {
2381 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2382 signed char sl,t;
2383 t=get_reg(i_regs->regmap,dops[i].rt1);
2384 sl=get_reg(i_regs->regmap,dops[i].rs1);
2385 //assert(t>=0);
2386 if(t>=0) {
2387 if(dops[i].rs1>0) {
2388 if(dops[i].opcode==0x0a) { // SLTI
2389 if(sl<0) {
2390 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2391 emit_slti32(t,imm[i],t);
2392 }else{
2393 emit_slti32(sl,imm[i],t);
2394 }
2395 }
2396 else { // SLTIU
2397 if(sl<0) {
2398 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2399 emit_sltiu32(t,imm[i],t);
2400 }else{
2401 emit_sltiu32(sl,imm[i],t);
2402 }
2403 }
2404 }else{
2405 // SLTI(U) with r0 is just stupid,
2406 // nonetheless examples can be found
2407 if(dops[i].opcode==0x0a) // SLTI
2408 if(0<imm[i]) emit_movimm(1,t);
2409 else emit_zeroreg(t);
2410 else // SLTIU
2411 {
2412 if(imm[i]) emit_movimm(1,t);
2413 else emit_zeroreg(t);
2414 }
2415 }
2416 }
2417 }
2418 }
2419 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2420 if(dops[i].rt1) {
2421 signed char sl,tl;
2422 tl=get_reg(i_regs->regmap,dops[i].rt1);
2423 sl=get_reg(i_regs->regmap,dops[i].rs1);
2424 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2425 if(dops[i].opcode==0x0c) //ANDI
2426 {
2427 if(dops[i].rs1) {
2428 if(sl<0) {
2429 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2430 emit_andimm(tl,imm[i],tl);
2431 }else{
2432 if(!((i_regs->wasconst>>sl)&1))
2433 emit_andimm(sl,imm[i],tl);
2434 else
2435 emit_movimm(constmap[i][sl]&imm[i],tl);
2436 }
2437 }
2438 else
2439 emit_zeroreg(tl);
2440 }
2441 else
2442 {
2443 if(dops[i].rs1) {
2444 if(sl<0) {
2445 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2446 }
2447 if(dops[i].opcode==0x0d) { // ORI
2448 if(sl<0) {
2449 emit_orimm(tl,imm[i],tl);
2450 }else{
2451 if(!((i_regs->wasconst>>sl)&1))
2452 emit_orimm(sl,imm[i],tl);
2453 else
2454 emit_movimm(constmap[i][sl]|imm[i],tl);
2455 }
2456 }
2457 if(dops[i].opcode==0x0e) { // XORI
2458 if(sl<0) {
2459 emit_xorimm(tl,imm[i],tl);
2460 }else{
2461 if(!((i_regs->wasconst>>sl)&1))
2462 emit_xorimm(sl,imm[i],tl);
2463 else
2464 emit_movimm(constmap[i][sl]^imm[i],tl);
2465 }
2466 }
2467 }
2468 else {
2469 emit_movimm(imm[i],tl);
2470 }
2471 }
2472 }
2473 }
2474 }
2475}
2476
2477static void shiftimm_assemble(int i, const struct regstat *i_regs)
2478{
2479 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2480 {
2481 if(dops[i].rt1) {
2482 signed char s,t;
2483 t=get_reg(i_regs->regmap,dops[i].rt1);
2484 s=get_reg(i_regs->regmap,dops[i].rs1);
2485 //assert(t>=0);
2486 if(t>=0&&!((i_regs->isconst>>t)&1)){
2487 if(dops[i].rs1==0)
2488 {
2489 emit_zeroreg(t);
2490 }
2491 else
2492 {
2493 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2494 if(imm[i]) {
2495 if(dops[i].opcode2==0) // SLL
2496 {
2497 emit_shlimm(s<0?t:s,imm[i],t);
2498 }
2499 if(dops[i].opcode2==2) // SRL
2500 {
2501 emit_shrimm(s<0?t:s,imm[i],t);
2502 }
2503 if(dops[i].opcode2==3) // SRA
2504 {
2505 emit_sarimm(s<0?t:s,imm[i],t);
2506 }
2507 }else{
2508 // Shift by zero
2509 if(s>=0 && s!=t) emit_mov(s,t);
2510 }
2511 }
2512 }
2513 //emit_storereg(dops[i].rt1,t); //DEBUG
2514 }
2515 }
2516 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2517 {
2518 assert(0);
2519 }
2520 if(dops[i].opcode2==0x3c) // DSLL32
2521 {
2522 assert(0);
2523 }
2524 if(dops[i].opcode2==0x3e) // DSRL32
2525 {
2526 assert(0);
2527 }
2528 if(dops[i].opcode2==0x3f) // DSRA32
2529 {
2530 assert(0);
2531 }
2532}
2533
2534#ifndef shift_assemble
2535static void shift_assemble(int i, const struct regstat *i_regs)
2536{
2537 signed char s,t,shift;
2538 if (dops[i].rt1 == 0)
2539 return;
2540 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2541 t = get_reg(i_regs->regmap, dops[i].rt1);
2542 s = get_reg(i_regs->regmap, dops[i].rs1);
2543 shift = get_reg(i_regs->regmap, dops[i].rs2);
2544 if (t < 0)
2545 return;
2546
2547 if(dops[i].rs1==0)
2548 emit_zeroreg(t);
2549 else if(dops[i].rs2==0) {
2550 assert(s>=0);
2551 if(s!=t) emit_mov(s,t);
2552 }
2553 else {
2554 host_tempreg_acquire();
2555 emit_andimm(shift,31,HOST_TEMPREG);
2556 switch(dops[i].opcode2) {
2557 case 4: // SLLV
2558 emit_shl(s,HOST_TEMPREG,t);
2559 break;
2560 case 6: // SRLV
2561 emit_shr(s,HOST_TEMPREG,t);
2562 break;
2563 case 7: // SRAV
2564 emit_sar(s,HOST_TEMPREG,t);
2565 break;
2566 default:
2567 assert(0);
2568 }
2569 host_tempreg_release();
2570 }
2571}
2572
2573#endif
2574
2575enum {
2576 MTYPE_8000 = 0,
2577 MTYPE_8020,
2578 MTYPE_0000,
2579 MTYPE_A000,
2580 MTYPE_1F80,
2581};
2582
2583static int get_ptr_mem_type(u_int a)
2584{
2585 if(a < 0x00200000) {
2586 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2587 // return wrong, must use memhandler for BIOS self-test to pass
2588 // 007 does similar stuff from a00 mirror, weird stuff
2589 return MTYPE_8000;
2590 return MTYPE_0000;
2591 }
2592 if(0x1f800000 <= a && a < 0x1f801000)
2593 return MTYPE_1F80;
2594 if(0x80200000 <= a && a < 0x80800000)
2595 return MTYPE_8020;
2596 if(0xa0000000 <= a && a < 0xa0200000)
2597 return MTYPE_A000;
2598 return MTYPE_8000;
2599}
2600
2601static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2602{
2603 int r = get_reg(i_regs->regmap, ROREG);
2604 if (r < 0 && host_tempreg_free) {
2605 host_tempreg_acquire();
2606 emit_loadreg(ROREG, r = HOST_TEMPREG);
2607 }
2608 if (r < 0)
2609 abort();
2610 return r;
2611}
2612
2613static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2614 int addr, int *offset_reg, int *addr_reg_override)
2615{
2616 void *jaddr = NULL;
2617 int type = 0;
2618 int mr = dops[i].rs1;
2619 *offset_reg = -1;
2620 if(((smrv_strong|smrv_weak)>>mr)&1) {
2621 type=get_ptr_mem_type(smrv[mr]);
2622 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2623 }
2624 else {
2625 // use the mirror we are running on
2626 type=get_ptr_mem_type(start);
2627 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2628 }
2629
2630 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2631 host_tempreg_acquire();
2632 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2633 addr=*addr_reg_override=HOST_TEMPREG;
2634 type=0;
2635 }
2636 else if(type==MTYPE_0000) { // RAM 0 mirror
2637 host_tempreg_acquire();
2638 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2639 addr=*addr_reg_override=HOST_TEMPREG;
2640 type=0;
2641 }
2642 else if(type==MTYPE_A000) { // RAM A mirror
2643 host_tempreg_acquire();
2644 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2645 addr=*addr_reg_override=HOST_TEMPREG;
2646 type=0;
2647 }
2648 else if(type==MTYPE_1F80) { // scratchpad
2649 if (psxH == (void *)0x1f800000) {
2650 host_tempreg_acquire();
2651 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2652 emit_cmpimm(HOST_TEMPREG,0x1000);
2653 host_tempreg_release();
2654 jaddr=out;
2655 emit_jc(0);
2656 }
2657 else {
2658 // do the usual RAM check, jump will go to the right handler
2659 type=0;
2660 }
2661 }
2662
2663 if (type == 0) // need ram check
2664 {
2665 emit_cmpimm(addr,RAM_SIZE);
2666 jaddr = out;
2667 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2668 // Hint to branch predictor that the branch is unlikely to be taken
2669 if (dops[i].rs1 >= 28)
2670 emit_jno_unlikely(0);
2671 else
2672 #endif
2673 emit_jno(0);
2674 if (ram_offset != 0)
2675 *offset_reg = get_ro_reg(i_regs, 0);
2676 }
2677
2678 return jaddr;
2679}
2680
2681// return memhandler, or get directly accessable address and return 0
2682static void *get_direct_memhandler(void *table, u_int addr,
2683 enum stub_type type, uintptr_t *addr_host)
2684{
2685 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2686 uintptr_t l1, l2 = 0;
2687 l1 = ((uintptr_t *)table)[addr>>12];
2688 if (!(l1 & msb)) {
2689 uintptr_t v = l1 << 1;
2690 *addr_host = v + addr;
2691 return NULL;
2692 }
2693 else {
2694 l1 <<= 1;
2695 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2696 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2697 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2698 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2699 else
2700 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2701 if (!(l2 & msb)) {
2702 uintptr_t v = l2 << 1;
2703 *addr_host = v + (addr&0xfff);
2704 return NULL;
2705 }
2706 return (void *)(l2 << 1);
2707 }
2708}
2709
2710static u_int get_host_reglist(const signed char *regmap)
2711{
2712 u_int reglist = 0, hr;
2713 for (hr = 0; hr < HOST_REGS; hr++) {
2714 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2715 reglist |= 1 << hr;
2716 }
2717 return reglist;
2718}
2719
2720static u_int reglist_exclude(u_int reglist, int r1, int r2)
2721{
2722 if (r1 >= 0)
2723 reglist &= ~(1u << r1);
2724 if (r2 >= 0)
2725 reglist &= ~(1u << r2);
2726 return reglist;
2727}
2728
2729// find a temp caller-saved register not in reglist (so assumed to be free)
2730static int reglist_find_free(u_int reglist)
2731{
2732 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2733 if (free_regs == 0)
2734 return -1;
2735 return __builtin_ctz(free_regs);
2736}
2737
2738static void do_load_word(int a, int rt, int offset_reg)
2739{
2740 if (offset_reg >= 0)
2741 emit_ldr_dualindexed(offset_reg, a, rt);
2742 else
2743 emit_readword_indexed(0, a, rt);
2744}
2745
2746static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2747{
2748 if (offset_reg < 0) {
2749 emit_writeword_indexed(rt, ofs, a);
2750 return;
2751 }
2752 if (ofs != 0)
2753 emit_addimm(a, ofs, a);
2754 emit_str_dualindexed(offset_reg, a, rt);
2755 if (ofs != 0 && preseve_a)
2756 emit_addimm(a, -ofs, a);
2757}
2758
2759static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2760{
2761 if (offset_reg < 0) {
2762 emit_writehword_indexed(rt, ofs, a);
2763 return;
2764 }
2765 if (ofs != 0)
2766 emit_addimm(a, ofs, a);
2767 emit_strh_dualindexed(offset_reg, a, rt);
2768 if (ofs != 0 && preseve_a)
2769 emit_addimm(a, -ofs, a);
2770}
2771
2772static void do_store_byte(int a, int rt, int offset_reg)
2773{
2774 if (offset_reg >= 0)
2775 emit_strb_dualindexed(offset_reg, a, rt);
2776 else
2777 emit_writebyte_indexed(rt, 0, a);
2778}
2779
2780static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2781{
2782 int s,tl,addr;
2783 int offset;
2784 void *jaddr=0;
2785 int memtarget=0,c=0;
2786 int offset_reg = -1;
2787 int fastio_reg_override = -1;
2788 u_int reglist=get_host_reglist(i_regs->regmap);
2789 tl=get_reg(i_regs->regmap,dops[i].rt1);
2790 s=get_reg(i_regs->regmap,dops[i].rs1);
2791 offset=imm[i];
2792 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2793 if(s>=0) {
2794 c=(i_regs->wasconst>>s)&1;
2795 if (c) {
2796 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2797 }
2798 }
2799 //printf("load_assemble: c=%d\n",c);
2800 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2801 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2802 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2803 ||dops[i].rt1==0) {
2804 // could be FIFO, must perform the read
2805 // ||dummy read
2806 assem_debug("(forced read)\n");
2807 tl=get_reg(i_regs->regmap,-1);
2808 assert(tl>=0);
2809 }
2810 if(offset||s<0||c) addr=tl;
2811 else addr=s;
2812 //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2813 if(tl>=0) {
2814 //printf("load_assemble: c=%d\n",c);
2815 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2816 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2817 reglist&=~(1<<tl);
2818 if(!c) {
2819 #ifdef R29_HACK
2820 // Strmnnrmn's speed hack
2821 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2822 #endif
2823 {
2824 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
2825 &offset_reg, &fastio_reg_override);
2826 }
2827 }
2828 else if (ram_offset && memtarget) {
2829 offset_reg = get_ro_reg(i_regs, 0);
2830 }
2831 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2832 switch (dops[i].opcode) {
2833 case 0x20: // LB
2834 if(!c||memtarget) {
2835 if(!dummy) {
2836 int a = tl;
2837 if (!c) a = addr;
2838 if (fastio_reg_override >= 0)
2839 a = fastio_reg_override;
2840
2841 if (offset_reg >= 0)
2842 emit_ldrsb_dualindexed(offset_reg, a, tl);
2843 else
2844 emit_movsbl_indexed(0, a, tl);
2845 }
2846 if(jaddr)
2847 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2848 }
2849 else
2850 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2851 break;
2852 case 0x21: // LH
2853 if(!c||memtarget) {
2854 if(!dummy) {
2855 int a = tl;
2856 if (!c) a = addr;
2857 if (fastio_reg_override >= 0)
2858 a = fastio_reg_override;
2859 if (offset_reg >= 0)
2860 emit_ldrsh_dualindexed(offset_reg, a, tl);
2861 else
2862 emit_movswl_indexed(0, a, tl);
2863 }
2864 if(jaddr)
2865 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2866 }
2867 else
2868 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2869 break;
2870 case 0x23: // LW
2871 if(!c||memtarget) {
2872 if(!dummy) {
2873 int a = addr;
2874 if (fastio_reg_override >= 0)
2875 a = fastio_reg_override;
2876 do_load_word(a, tl, offset_reg);
2877 }
2878 if(jaddr)
2879 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2880 }
2881 else
2882 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2883 break;
2884 case 0x24: // LBU
2885 if(!c||memtarget) {
2886 if(!dummy) {
2887 int a = tl;
2888 if (!c) a = addr;
2889 if (fastio_reg_override >= 0)
2890 a = fastio_reg_override;
2891
2892 if (offset_reg >= 0)
2893 emit_ldrb_dualindexed(offset_reg, a, tl);
2894 else
2895 emit_movzbl_indexed(0, a, tl);
2896 }
2897 if(jaddr)
2898 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2899 }
2900 else
2901 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2902 break;
2903 case 0x25: // LHU
2904 if(!c||memtarget) {
2905 if(!dummy) {
2906 int a = tl;
2907 if(!c) a = addr;
2908 if (fastio_reg_override >= 0)
2909 a = fastio_reg_override;
2910 if (offset_reg >= 0)
2911 emit_ldrh_dualindexed(offset_reg, a, tl);
2912 else
2913 emit_movzwl_indexed(0, a, tl);
2914 }
2915 if(jaddr)
2916 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2917 }
2918 else
2919 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2920 break;
2921 case 0x27: // LWU
2922 case 0x37: // LD
2923 default:
2924 assert(0);
2925 }
2926 }
2927 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
2928 host_tempreg_release();
2929}
2930
2931#ifndef loadlr_assemble
2932static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
2933{
2934 int s,tl,temp,temp2,addr;
2935 int offset;
2936 void *jaddr=0;
2937 int memtarget=0,c=0;
2938 int offset_reg = -1;
2939 int fastio_reg_override = -1;
2940 u_int reglist=get_host_reglist(i_regs->regmap);
2941 tl=get_reg(i_regs->regmap,dops[i].rt1);
2942 s=get_reg(i_regs->regmap,dops[i].rs1);
2943 temp=get_reg(i_regs->regmap,-1);
2944 temp2=get_reg(i_regs->regmap,FTEMP);
2945 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
2946 assert(addr<0);
2947 offset=imm[i];
2948 reglist|=1<<temp;
2949 if(offset||s<0||c) addr=temp2;
2950 else addr=s;
2951 if(s>=0) {
2952 c=(i_regs->wasconst>>s)&1;
2953 if(c) {
2954 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2955 }
2956 }
2957 if(!c) {
2958 emit_shlimm(addr,3,temp);
2959 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2960 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
2961 }else{
2962 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
2963 }
2964 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
2965 &offset_reg, &fastio_reg_override);
2966 }
2967 else {
2968 if (ram_offset && memtarget) {
2969 offset_reg = get_ro_reg(i_regs, 0);
2970 }
2971 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
2972 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
2973 }else{
2974 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
2975 }
2976 }
2977 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
2978 if(!c||memtarget) {
2979 int a = temp2;
2980 if (fastio_reg_override >= 0)
2981 a = fastio_reg_override;
2982 do_load_word(a, temp2, offset_reg);
2983 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
2984 host_tempreg_release();
2985 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
2986 }
2987 else
2988 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
2989 if(dops[i].rt1) {
2990 assert(tl>=0);
2991 emit_andimm(temp,24,temp);
2992 if (dops[i].opcode==0x22) // LWL
2993 emit_xorimm(temp,24,temp);
2994 host_tempreg_acquire();
2995 emit_movimm(-1,HOST_TEMPREG);
2996 if (dops[i].opcode==0x26) {
2997 emit_shr(temp2,temp,temp2);
2998 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
2999 }else{
3000 emit_shl(temp2,temp,temp2);
3001 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3002 }
3003 host_tempreg_release();
3004 emit_or(temp2,tl,tl);
3005 }
3006 //emit_storereg(dops[i].rt1,tl); // DEBUG
3007 }
3008 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3009 assert(0);
3010 }
3011}
3012#endif
3013
3014static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3015{
3016 int s,tl;
3017 int addr,temp;
3018 int offset;
3019 void *jaddr=0;
3020 enum stub_type type=0;
3021 int memtarget=0,c=0;
3022 int agr=AGEN1+(i&1);
3023 int offset_reg = -1;
3024 int fastio_reg_override = -1;
3025 u_int reglist=get_host_reglist(i_regs->regmap);
3026 tl=get_reg(i_regs->regmap,dops[i].rs2);
3027 s=get_reg(i_regs->regmap,dops[i].rs1);
3028 temp=get_reg(i_regs->regmap,agr);
3029 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3030 offset=imm[i];
3031 if(s>=0) {
3032 c=(i_regs->wasconst>>s)&1;
3033 if(c) {
3034 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3035 }
3036 }
3037 assert(tl>=0);
3038 assert(temp>=0);
3039 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3040 if(offset||s<0||c) addr=temp;
3041 else addr=s;
3042 if (!c) {
3043 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3044 &offset_reg, &fastio_reg_override);
3045 }
3046 else if (ram_offset && memtarget) {
3047 offset_reg = get_ro_reg(i_regs, 0);
3048 }
3049
3050 switch (dops[i].opcode) {
3051 case 0x28: // SB
3052 if(!c||memtarget) {
3053 int a = temp;
3054 if (!c) a = addr;
3055 if (fastio_reg_override >= 0)
3056 a = fastio_reg_override;
3057 do_store_byte(a, tl, offset_reg);
3058 }
3059 type = STOREB_STUB;
3060 break;
3061 case 0x29: // SH
3062 if(!c||memtarget) {
3063 int a = temp;
3064 if (!c) a = addr;
3065 if (fastio_reg_override >= 0)
3066 a = fastio_reg_override;
3067 do_store_hword(a, 0, tl, offset_reg, 1);
3068 }
3069 type = STOREH_STUB;
3070 break;
3071 case 0x2B: // SW
3072 if(!c||memtarget) {
3073 int a = addr;
3074 if (fastio_reg_override >= 0)
3075 a = fastio_reg_override;
3076 do_store_word(a, 0, tl, offset_reg, 1);
3077 }
3078 type = STOREW_STUB;
3079 break;
3080 case 0x3F: // SD
3081 default:
3082 assert(0);
3083 }
3084 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3085 host_tempreg_release();
3086 if(jaddr) {
3087 // PCSX store handlers don't check invcode again
3088 reglist|=1<<addr;
3089 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3090 jaddr=0;
3091 }
3092 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3093 if(!c||memtarget) {
3094 #ifdef DESTRUCTIVE_SHIFT
3095 // The x86 shift operation is 'destructive'; it overwrites the
3096 // source register, so we need to make a copy first and use that.
3097 addr=temp;
3098 #endif
3099 #if defined(HOST_IMM8)
3100 int ir=get_reg(i_regs->regmap,INVCP);
3101 assert(ir>=0);
3102 emit_cmpmem_indexedsr12_reg(ir,addr,1);
3103 #else
3104 emit_cmpmem_indexedsr12_imm(invalid_code,addr,1);
3105 #endif
3106 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3107 emit_callne(invalidate_addr_reg[addr]);
3108 #else
3109 void *jaddr2 = out;
3110 emit_jne(0);
3111 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3112 #endif
3113 }
3114 }
3115 u_int addr_val=constmap[i][s]+offset;
3116 if(jaddr) {
3117 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3118 } else if(c&&!memtarget) {
3119 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3120 }
3121 // basic current block modification detection..
3122 // not looking back as that should be in mips cache already
3123 // (see Spyro2 title->attract mode)
3124 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3125 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3126 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3127 if(i_regs->regmap==regs[i].regmap) {
3128 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3129 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3130 emit_movimm(start+i*4+4,0);
3131 emit_writeword(0,&pcaddr);
3132 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3133 emit_far_call(get_addr_ht);
3134 emit_jmpreg(0);
3135 }
3136 }
3137}
3138
3139static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3140{
3141 int s,tl;
3142 int temp;
3143 int offset;
3144 void *jaddr=0;
3145 void *case1, *case23, *case3;
3146 void *done0, *done1, *done2;
3147 int memtarget=0,c=0;
3148 int agr=AGEN1+(i&1);
3149 int offset_reg = -1;
3150 u_int reglist=get_host_reglist(i_regs->regmap);
3151 tl=get_reg(i_regs->regmap,dops[i].rs2);
3152 s=get_reg(i_regs->regmap,dops[i].rs1);
3153 temp=get_reg(i_regs->regmap,agr);
3154 if(temp<0) temp=get_reg(i_regs->regmap,-1);
3155 offset=imm[i];
3156 if(s>=0) {
3157 c=(i_regs->isconst>>s)&1;
3158 if(c) {
3159 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3160 }
3161 }
3162 assert(tl>=0);
3163 assert(temp>=0);
3164 if(!c) {
3165 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3166 if(!offset&&s!=temp) emit_mov(s,temp);
3167 jaddr=out;
3168 emit_jno(0);
3169 }
3170 else
3171 {
3172 if(!memtarget||!dops[i].rs1) {
3173 jaddr=out;
3174 emit_jmp(0);
3175 }
3176 }
3177 if (ram_offset)
3178 offset_reg = get_ro_reg(i_regs, 0);
3179
3180 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3181 assert(0);
3182 }
3183
3184 emit_testimm(temp,2);
3185 case23=out;
3186 emit_jne(0);
3187 emit_testimm(temp,1);
3188 case1=out;
3189 emit_jne(0);
3190 // 0
3191 if (dops[i].opcode == 0x2A) { // SWL
3192 // Write msb into least significant byte
3193 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3194 do_store_byte(temp, tl, offset_reg);
3195 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3196 }
3197 else if (dops[i].opcode == 0x2E) { // SWR
3198 // Write entire word
3199 do_store_word(temp, 0, tl, offset_reg, 1);
3200 }
3201 done0 = out;
3202 emit_jmp(0);
3203 // 1
3204 set_jump_target(case1, out);
3205 if (dops[i].opcode == 0x2A) { // SWL
3206 // Write two msb into two least significant bytes
3207 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3208 do_store_hword(temp, -1, tl, offset_reg, 0);
3209 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3210 }
3211 else if (dops[i].opcode == 0x2E) { // SWR
3212 // Write 3 lsb into three most significant bytes
3213 do_store_byte(temp, tl, offset_reg);
3214 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3215 do_store_hword(temp, 1, tl, offset_reg, 0);
3216 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3217 }
3218 done1=out;
3219 emit_jmp(0);
3220 // 2,3
3221 set_jump_target(case23, out);
3222 emit_testimm(temp,1);
3223 case3 = out;
3224 emit_jne(0);
3225 // 2
3226 if (dops[i].opcode==0x2A) { // SWL
3227 // Write 3 msb into three least significant bytes
3228 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3229 do_store_hword(temp, -2, tl, offset_reg, 1);
3230 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3231 do_store_byte(temp, tl, offset_reg);
3232 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3233 }
3234 else if (dops[i].opcode == 0x2E) { // SWR
3235 // Write two lsb into two most significant bytes
3236 do_store_hword(temp, 0, tl, offset_reg, 1);
3237 }
3238 done2 = out;
3239 emit_jmp(0);
3240 // 3
3241 set_jump_target(case3, out);
3242 if (dops[i].opcode == 0x2A) { // SWL
3243 do_store_word(temp, -3, tl, offset_reg, 0);
3244 }
3245 else if (dops[i].opcode == 0x2E) { // SWR
3246 do_store_byte(temp, tl, offset_reg);
3247 }
3248 set_jump_target(done0, out);
3249 set_jump_target(done1, out);
3250 set_jump_target(done2, out);
3251 if (offset_reg == HOST_TEMPREG)
3252 host_tempreg_release();
3253 if(!c||!memtarget)
3254 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj_,reglist);
3255 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3256 #if defined(HOST_IMM8)
3257 int ir=get_reg(i_regs->regmap,INVCP);
3258 assert(ir>=0);
3259 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3260 #else
3261 emit_cmpmem_indexedsr12_imm(invalid_code,temp,1);
3262 #endif
3263 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3264 emit_callne(invalidate_addr_reg[temp]);
3265 #else
3266 void *jaddr2 = out;
3267 emit_jne(0);
3268 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3269 #endif
3270 }
3271}
3272
3273static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3274{
3275 if(dops[i].opcode2==0) // MFC0
3276 {
3277 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3278 u_int copr=(source[i]>>11)&0x1f;
3279 //assert(t>=0); // Why does this happen? OOT is weird
3280 if(t>=0&&dops[i].rt1!=0) {
3281 emit_readword(&reg_cop0[copr],t);
3282 }
3283 }
3284 else if(dops[i].opcode2==4) // MTC0
3285 {
3286 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3287 char copr=(source[i]>>11)&0x1f;
3288 assert(s>=0);
3289 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3290 if(copr==9||copr==11||copr==12||copr==13) {
3291 emit_readword(&last_count,HOST_TEMPREG);
3292 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3293 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3294 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3295 emit_writeword(HOST_CCREG,&Count);
3296 }
3297 // What a mess. The status register (12) can enable interrupts,
3298 // so needs a special case to handle a pending interrupt.
3299 // The interrupt must be taken immediately, because a subsequent
3300 // instruction might disable interrupts again.
3301 if(copr==12||copr==13) {
3302 if (is_delayslot) {
3303 // burn cycles to cause cc_interrupt, which will
3304 // reschedule next_interupt. Relies on CCREG from above.
3305 assem_debug("MTC0 DS %d\n", copr);
3306 emit_writeword(HOST_CCREG,&last_count);
3307 emit_movimm(0,HOST_CCREG);
3308 emit_storereg(CCREG,HOST_CCREG);
3309 emit_loadreg(dops[i].rs1,1);
3310 emit_movimm(copr,0);
3311 emit_far_call(pcsx_mtc0_ds);
3312 emit_loadreg(dops[i].rs1,s);
3313 return;
3314 }
3315 emit_movimm(start+i*4+4,HOST_TEMPREG);
3316 emit_writeword(HOST_TEMPREG,&pcaddr);
3317 emit_movimm(0,HOST_TEMPREG);
3318 emit_writeword(HOST_TEMPREG,&pending_exception);
3319 }
3320 if(s==HOST_CCREG)
3321 emit_loadreg(dops[i].rs1,1);
3322 else if(s!=1)
3323 emit_mov(s,1);
3324 emit_movimm(copr,0);
3325 emit_far_call(pcsx_mtc0);
3326 if(copr==9||copr==11||copr==12||copr==13) {
3327 emit_readword(&Count,HOST_CCREG);
3328 emit_readword(&next_interupt,HOST_TEMPREG);
3329 emit_addimm(HOST_CCREG,-ccadj_,HOST_CCREG);
3330 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3331 emit_writeword(HOST_TEMPREG,&last_count);
3332 emit_storereg(CCREG,HOST_CCREG);
3333 }
3334 if(copr==12||copr==13) {
3335 assert(!is_delayslot);
3336 emit_readword(&pending_exception,14);
3337 emit_test(14,14);
3338 void *jaddr = out;
3339 emit_jeq(0);
3340 emit_readword(&pcaddr, 0);
3341 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3342 emit_far_call(get_addr_ht);
3343 emit_jmpreg(0);
3344 set_jump_target(jaddr, out);
3345 }
3346 emit_loadreg(dops[i].rs1,s);
3347 }
3348 else
3349 {
3350 assert(dops[i].opcode2==0x10);
3351 //if((source[i]&0x3f)==0x10) // RFE
3352 {
3353 emit_readword(&Status,0);
3354 emit_andimm(0,0x3c,1);
3355 emit_andimm(0,~0xf,0);
3356 emit_orrshr_imm(1,2,0);
3357 emit_writeword(0,&Status);
3358 }
3359 }
3360}
3361
3362static void cop1_unusable(int i, const struct regstat *i_regs)
3363{
3364 // XXX: should just just do the exception instead
3365 //if(!cop1_usable)
3366 {
3367 void *jaddr=out;
3368 emit_jmp(0);
3369 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3370 }
3371}
3372
3373static void cop1_assemble(int i, const struct regstat *i_regs)
3374{
3375 cop1_unusable(i, i_regs);
3376}
3377
3378static void c1ls_assemble(int i, const struct regstat *i_regs)
3379{
3380 cop1_unusable(i, i_regs);
3381}
3382
3383// FP_STUB
3384static void do_cop1stub(int n)
3385{
3386 literal_pool(256);
3387 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3388 set_jump_target(stubs[n].addr, out);
3389 int i=stubs[n].a;
3390// int rs=stubs[n].b;
3391 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3392 int ds=stubs[n].d;
3393 if(!ds) {
3394 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3395 //if(i_regs!=&regs[i]) printf("oops: regs[i]=%x i_regs=%x",(int)&regs[i],(int)i_regs);
3396 }
3397 //else {printf("fp exception in delay slot\n");}
3398 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3399 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3400 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3401 emit_addimm(HOST_CCREG,ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3402 emit_far_jump(ds?fp_exception_ds:fp_exception);
3403}
3404
3405static int cop2_is_stalling_op(int i, int *cycles)
3406{
3407 if (dops[i].opcode == 0x3a) { // SWC2
3408 *cycles = 0;
3409 return 1;
3410 }
3411 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3412 *cycles = 0;
3413 return 1;
3414 }
3415 if (dops[i].itype == C2OP) {
3416 *cycles = gte_cycletab[source[i] & 0x3f];
3417 return 1;
3418 }
3419 // ... what about MTC2/CTC2/LWC2?
3420 return 0;
3421}
3422
3423#if 0
3424static void log_gte_stall(int stall, u_int cycle)
3425{
3426 if ((u_int)stall <= 44)
3427 printf("x stall %2d %u\n", stall, cycle + last_count);
3428}
3429
3430static void emit_log_gte_stall(int i, int stall, u_int reglist)
3431{
3432 save_regs(reglist);
3433 if (stall > 0)
3434 emit_movimm(stall, 0);
3435 else
3436 emit_mov(HOST_TEMPREG, 0);
3437 emit_addimm(HOST_CCREG, ccadj[i], 1);
3438 emit_far_call(log_gte_stall);
3439 restore_regs(reglist);
3440}
3441#endif
3442
3443static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3444{
3445 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3446 int rtmp = reglist_find_free(reglist);
3447
3448 if (HACK_ENABLED(NDHACK_NO_STALLS))
3449 return;
3450 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3451 // happens occasionally... cc evicted? Don't bother then
3452 //printf("no cc %08x\n", start + i*4);
3453 return;
3454 }
3455 if (!dops[i].bt) {
3456 for (j = i - 1; j >= 0; j--) {
3457 //if (dops[j].is_ds) break;
3458 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3459 break;
3460 if (j > 0 && ccadj[j - 1] > ccadj[j])
3461 break;
3462 }
3463 j = max(j, 0);
3464 }
3465 cycles_passed = ccadj[i] - ccadj[j];
3466 if (other_gte_op_cycles >= 0)
3467 stall = other_gte_op_cycles - cycles_passed;
3468 else if (cycles_passed >= 44)
3469 stall = 0; // can't stall
3470 if (stall == -MAXBLOCK && rtmp >= 0) {
3471 // unknown stall, do the expensive runtime check
3472 assem_debug("; cop2_do_stall_check\n");
3473#if 0 // too slow
3474 save_regs(reglist);
3475 emit_movimm(gte_cycletab[op], 0);
3476 emit_addimm(HOST_CCREG, ccadj[i], 1);
3477 emit_far_call(call_gteStall);
3478 restore_regs(reglist);
3479#else
3480 host_tempreg_acquire();
3481 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3482 emit_addimm(rtmp, -ccadj[i], rtmp);
3483 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3484 emit_cmpimm(HOST_TEMPREG, 44);
3485 emit_cmovb_reg(rtmp, HOST_CCREG);
3486 //emit_log_gte_stall(i, 0, reglist);
3487 host_tempreg_release();
3488#endif
3489 }
3490 else if (stall > 0) {
3491 //emit_log_gte_stall(i, stall, reglist);
3492 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3493 }
3494
3495 // save gteBusyCycle, if needed
3496 if (gte_cycletab[op] == 0)
3497 return;
3498 other_gte_op_cycles = -1;
3499 for (j = i + 1; j < slen; j++) {
3500 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3501 break;
3502 if (dops[j].is_jump) {
3503 // check ds
3504 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3505 j++;
3506 break;
3507 }
3508 }
3509 if (other_gte_op_cycles >= 0)
3510 // will handle stall when assembling that op
3511 return;
3512 cycles_passed = ccadj[min(j, slen -1)] - ccadj[i];
3513 if (cycles_passed >= 44)
3514 return;
3515 assem_debug("; save gteBusyCycle\n");
3516 host_tempreg_acquire();
3517#if 0
3518 emit_readword(&last_count, HOST_TEMPREG);
3519 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3520 emit_addimm(HOST_TEMPREG, ccadj[i], HOST_TEMPREG);
3521 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3522 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3523#else
3524 emit_addimm(HOST_CCREG, ccadj[i] + gte_cycletab[op], HOST_TEMPREG);
3525 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3526#endif
3527 host_tempreg_release();
3528}
3529
3530static int is_mflohi(int i)
3531{
3532 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3533}
3534
3535static int check_multdiv(int i, int *cycles)
3536{
3537 if (dops[i].itype != MULTDIV)
3538 return 0;
3539 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3540 *cycles = 11; // approx from 7 11 14
3541 else
3542 *cycles = 37;
3543 return 1;
3544}
3545
3546static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3547{
3548 int j, found = 0, c = 0;
3549 if (HACK_ENABLED(NDHACK_NO_STALLS))
3550 return;
3551 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3552 // happens occasionally... cc evicted? Don't bother then
3553 return;
3554 }
3555 for (j = i + 1; j < slen; j++) {
3556 if (dops[j].bt)
3557 break;
3558 if ((found = is_mflohi(j)))
3559 break;
3560 if (dops[j].is_jump) {
3561 // check ds
3562 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3563 j++;
3564 break;
3565 }
3566 }
3567 if (found)
3568 // handle all in multdiv_do_stall()
3569 return;
3570 check_multdiv(i, &c);
3571 assert(c > 0);
3572 assem_debug("; muldiv prepare stall %d\n", c);
3573 host_tempreg_acquire();
3574 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3575 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3576 host_tempreg_release();
3577}
3578
3579static void multdiv_do_stall(int i, const struct regstat *i_regs)
3580{
3581 int j, known_cycles = 0;
3582 u_int reglist = get_host_reglist(i_regs->regmap);
3583 int rtmp = get_reg(i_regs->regmap, -1);
3584 if (rtmp < 0)
3585 rtmp = reglist_find_free(reglist);
3586 if (HACK_ENABLED(NDHACK_NO_STALLS))
3587 return;
3588 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3589 // happens occasionally... cc evicted? Don't bother then
3590 //printf("no cc/rtmp %08x\n", start + i*4);
3591 return;
3592 }
3593 if (!dops[i].bt) {
3594 for (j = i - 1; j >= 0; j--) {
3595 if (dops[j].is_ds) break;
3596 if (check_multdiv(j, &known_cycles))
3597 break;
3598 if (is_mflohi(j))
3599 // already handled by this op
3600 return;
3601 if (dops[j].bt || (j > 0 && ccadj[j - 1] > ccadj[j]))
3602 break;
3603 }
3604 j = max(j, 0);
3605 }
3606 if (known_cycles > 0) {
3607 known_cycles -= ccadj[i] - ccadj[j];
3608 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3609 if (known_cycles > 0)
3610 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3611 return;
3612 }
3613 assem_debug("; muldiv stall unresolved\n");
3614 host_tempreg_acquire();
3615 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3616 emit_addimm(rtmp, -ccadj[i], rtmp);
3617 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3618 emit_cmpimm(HOST_TEMPREG, 37);
3619 emit_cmovb_reg(rtmp, HOST_CCREG);
3620 //emit_log_gte_stall(i, 0, reglist);
3621 host_tempreg_release();
3622}
3623
3624static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3625{
3626 switch (copr) {
3627 case 1:
3628 case 3:
3629 case 5:
3630 case 8:
3631 case 9:
3632 case 10:
3633 case 11:
3634 emit_readword(&reg_cop2d[copr],tl);
3635 emit_signextend16(tl,tl);
3636 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3637 break;
3638 case 7:
3639 case 16:
3640 case 17:
3641 case 18:
3642 case 19:
3643 emit_readword(&reg_cop2d[copr],tl);
3644 emit_andimm(tl,0xffff,tl);
3645 emit_writeword(tl,&reg_cop2d[copr]);
3646 break;
3647 case 15:
3648 emit_readword(&reg_cop2d[14],tl); // SXY2
3649 emit_writeword(tl,&reg_cop2d[copr]);
3650 break;
3651 case 28:
3652 case 29:
3653 c2op_mfc2_29_assemble(tl,temp);
3654 break;
3655 default:
3656 emit_readword(&reg_cop2d[copr],tl);
3657 break;
3658 }
3659}
3660
3661static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3662{
3663 switch (copr) {
3664 case 15:
3665 emit_readword(&reg_cop2d[13],temp); // SXY1
3666 emit_writeword(sl,&reg_cop2d[copr]);
3667 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3668 emit_readword(&reg_cop2d[14],temp); // SXY2
3669 emit_writeword(sl,&reg_cop2d[14]);
3670 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3671 break;
3672 case 28:
3673 emit_andimm(sl,0x001f,temp);
3674 emit_shlimm(temp,7,temp);
3675 emit_writeword(temp,&reg_cop2d[9]);
3676 emit_andimm(sl,0x03e0,temp);
3677 emit_shlimm(temp,2,temp);
3678 emit_writeword(temp,&reg_cop2d[10]);
3679 emit_andimm(sl,0x7c00,temp);
3680 emit_shrimm(temp,3,temp);
3681 emit_writeword(temp,&reg_cop2d[11]);
3682 emit_writeword(sl,&reg_cop2d[28]);
3683 break;
3684 case 30:
3685 emit_xorsar_imm(sl,sl,31,temp);
3686#if defined(HAVE_ARMV5) || defined(__aarch64__)
3687 emit_clz(temp,temp);
3688#else
3689 emit_movs(temp,HOST_TEMPREG);
3690 emit_movimm(0,temp);
3691 emit_jeq((int)out+4*4);
3692 emit_addpl_imm(temp,1,temp);
3693 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3694 emit_jns((int)out-2*4);
3695#endif
3696 emit_writeword(sl,&reg_cop2d[30]);
3697 emit_writeword(temp,&reg_cop2d[31]);
3698 break;
3699 case 31:
3700 break;
3701 default:
3702 emit_writeword(sl,&reg_cop2d[copr]);
3703 break;
3704 }
3705}
3706
3707static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3708{
3709 int s,tl;
3710 int ar;
3711 int offset;
3712 int memtarget=0,c=0;
3713 void *jaddr2=NULL;
3714 enum stub_type type;
3715 int agr=AGEN1+(i&1);
3716 int offset_reg = -1;
3717 int fastio_reg_override = -1;
3718 u_int reglist=get_host_reglist(i_regs->regmap);
3719 u_int copr=(source[i]>>16)&0x1f;
3720 s=get_reg(i_regs->regmap,dops[i].rs1);
3721 tl=get_reg(i_regs->regmap,FTEMP);
3722 offset=imm[i];
3723 assert(dops[i].rs1>0);
3724 assert(tl>=0);
3725
3726 if(i_regs->regmap[HOST_CCREG]==CCREG)
3727 reglist&=~(1<<HOST_CCREG);
3728
3729 // get the address
3730 if (dops[i].opcode==0x3a) { // SWC2
3731 ar=get_reg(i_regs->regmap,agr);
3732 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3733 reglist|=1<<ar;
3734 } else { // LWC2
3735 ar=tl;
3736 }
3737 if(s>=0) c=(i_regs->wasconst>>s)&1;
3738 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3739 if (!offset&&!c&&s>=0) ar=s;
3740 assert(ar>=0);
3741
3742 cop2_do_stall_check(0, i, i_regs, reglist);
3743
3744 if (dops[i].opcode==0x3a) { // SWC2
3745 cop2_get_dreg(copr,tl,-1);
3746 type=STOREW_STUB;
3747 }
3748 else
3749 type=LOADW_STUB;
3750
3751 if(c&&!memtarget) {
3752 jaddr2=out;
3753 emit_jmp(0); // inline_readstub/inline_writestub?
3754 }
3755 else {
3756 if(!c) {
3757 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3758 &offset_reg, &fastio_reg_override);
3759 }
3760 else if (ram_offset && memtarget) {
3761 offset_reg = get_ro_reg(i_regs, 0);
3762 }
3763 switch (dops[i].opcode) {
3764 case 0x32: { // LWC2
3765 int a = ar;
3766 if (fastio_reg_override >= 0)
3767 a = fastio_reg_override;
3768 do_load_word(a, tl, offset_reg);
3769 break;
3770 }
3771 case 0x3a: { // SWC2
3772 #ifdef DESTRUCTIVE_SHIFT
3773 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3774 #endif
3775 int a = ar;
3776 if (fastio_reg_override >= 0)
3777 a = fastio_reg_override;
3778 do_store_word(a, 0, tl, offset_reg, 1);
3779 break;
3780 }
3781 default:
3782 assert(0);
3783 }
3784 }
3785 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3786 host_tempreg_release();
3787 if(jaddr2)
3788 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3789 if(dops[i].opcode==0x3a) // SWC2
3790 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3791#if defined(HOST_IMM8)
3792 int ir=get_reg(i_regs->regmap,INVCP);
3793 assert(ir>=0);
3794 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3795#else
3796 emit_cmpmem_indexedsr12_imm(invalid_code,ar,1);
3797#endif
3798 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3799 emit_callne(invalidate_addr_reg[ar]);
3800 #else
3801 void *jaddr3 = out;
3802 emit_jne(0);
3803 add_stub(INVCODE_STUB,jaddr3,out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3804 #endif
3805 }
3806 if (dops[i].opcode==0x32) { // LWC2
3807 host_tempreg_acquire();
3808 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3809 host_tempreg_release();
3810 }
3811}
3812
3813static void cop2_assemble(int i, const struct regstat *i_regs)
3814{
3815 u_int copr = (source[i]>>11) & 0x1f;
3816 signed char temp = get_reg(i_regs->regmap, -1);
3817
3818 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3819 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3820 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3821 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3822 reglist = reglist_exclude(reglist, tl, -1);
3823 }
3824 cop2_do_stall_check(0, i, i_regs, reglist);
3825 }
3826 if (dops[i].opcode2==0) { // MFC2
3827 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3828 if(tl>=0&&dops[i].rt1!=0)
3829 cop2_get_dreg(copr,tl,temp);
3830 }
3831 else if (dops[i].opcode2==4) { // MTC2
3832 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3833 cop2_put_dreg(copr,sl,temp);
3834 }
3835 else if (dops[i].opcode2==2) // CFC2
3836 {
3837 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3838 if(tl>=0&&dops[i].rt1!=0)
3839 emit_readword(&reg_cop2c[copr],tl);
3840 }
3841 else if (dops[i].opcode2==6) // CTC2
3842 {
3843 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3844 switch(copr) {
3845 case 4:
3846 case 12:
3847 case 20:
3848 case 26:
3849 case 27:
3850 case 29:
3851 case 30:
3852 emit_signextend16(sl,temp);
3853 break;
3854 case 31:
3855 c2op_ctc2_31_assemble(sl,temp);
3856 break;
3857 default:
3858 temp=sl;
3859 break;
3860 }
3861 emit_writeword(temp,&reg_cop2c[copr]);
3862 assert(sl>=0);
3863 }
3864}
3865
3866static void do_unalignedwritestub(int n)
3867{
3868 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
3869 literal_pool(256);
3870 set_jump_target(stubs[n].addr, out);
3871
3872 int i=stubs[n].a;
3873 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3874 int addr=stubs[n].b;
3875 u_int reglist=stubs[n].e;
3876 signed char *i_regmap=i_regs->regmap;
3877 int temp2=get_reg(i_regmap,FTEMP);
3878 int rt;
3879 rt=get_reg(i_regmap,dops[i].rs2);
3880 assert(rt>=0);
3881 assert(addr>=0);
3882 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
3883 reglist|=(1<<addr);
3884 reglist&=~(1<<temp2);
3885
3886 // don't bother with it and call write handler
3887 save_regs(reglist);
3888 pass_args(addr,rt);
3889 int cc=get_reg(i_regmap,CCREG);
3890 if(cc<0)
3891 emit_loadreg(CCREG,2);
3892 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
3893 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
3894 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
3895 if(cc<0)
3896 emit_storereg(CCREG,2);
3897 restore_regs(reglist);
3898 emit_jmp(stubs[n].retaddr); // return address
3899}
3900
3901#ifndef multdiv_assemble
3902void multdiv_assemble(int i,struct regstat *i_regs)
3903{
3904 printf("Need multdiv_assemble for this architecture.\n");
3905 abort();
3906}
3907#endif
3908
3909static void mov_assemble(int i, const struct regstat *i_regs)
3910{
3911 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
3912 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
3913 if(dops[i].rt1) {
3914 signed char sl,tl;
3915 tl=get_reg(i_regs->regmap,dops[i].rt1);
3916 //assert(tl>=0);
3917 if(tl>=0) {
3918 sl=get_reg(i_regs->regmap,dops[i].rs1);
3919 if(sl>=0) emit_mov(sl,tl);
3920 else emit_loadreg(dops[i].rs1,tl);
3921 }
3922 }
3923 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
3924 multdiv_do_stall(i, i_regs);
3925}
3926
3927// call interpreter, exception handler, things that change pc/regs/cycles ...
3928static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
3929{
3930 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3931 assert(ccreg==HOST_CCREG);
3932 assert(!is_delayslot);
3933 (void)ccreg;
3934
3935 emit_movimm(pc,3); // Get PC
3936 emit_readword(&last_count,2);
3937 emit_writeword(3,&psxRegs.pc);
3938 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3939 emit_add(2,HOST_CCREG,2);
3940 emit_writeword(2,&psxRegs.cycle);
3941 emit_far_call(func);
3942 emit_far_jump(jump_to_new_pc);
3943}
3944
3945static void syscall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3946{
3947 // 'break' tends to be littered around to catch things like
3948 // division by 0 and is almost never executed, so don't emit much code here
3949 void *func = (dops[i].opcode2 == 0x0C)
3950 ? (is_delayslot ? jump_syscall_ds : jump_syscall)
3951 : (is_delayslot ? jump_break_ds : jump_break);
3952 signed char ccreg = get_reg(i_regs->regmap, CCREG);
3953 assert(ccreg == HOST_CCREG);
3954 emit_movimm(start + i*4, 2); // pc
3955 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
3956 emit_far_jump(func);
3957}
3958
3959static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3960{
3961 void *hlefunc = psxNULL;
3962 uint32_t hleCode = source[i] & 0x03ffffff;
3963 if (hleCode < ARRAY_SIZE(psxHLEt))
3964 hlefunc = psxHLEt[hleCode];
3965
3966 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
3967}
3968
3969static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3970{
3971 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
3972}
3973
3974static void speculate_mov(int rs,int rt)
3975{
3976 if(rt!=0) {
3977 smrv_strong_next|=1<<rt;
3978 smrv[rt]=smrv[rs];
3979 }
3980}
3981
3982static void speculate_mov_weak(int rs,int rt)
3983{
3984 if(rt!=0) {
3985 smrv_weak_next|=1<<rt;
3986 smrv[rt]=smrv[rs];
3987 }
3988}
3989
3990static void speculate_register_values(int i)
3991{
3992 if(i==0) {
3993 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
3994 // gp,sp are likely to stay the same throughout the block
3995 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
3996 smrv_weak_next=~smrv_strong_next;
3997 //printf(" llr %08x\n", smrv[4]);
3998 }
3999 smrv_strong=smrv_strong_next;
4000 smrv_weak=smrv_weak_next;
4001 switch(dops[i].itype) {
4002 case ALU:
4003 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4004 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4005 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4006 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4007 else {
4008 smrv_strong_next&=~(1<<dops[i].rt1);
4009 smrv_weak_next&=~(1<<dops[i].rt1);
4010 }
4011 break;
4012 case SHIFTIMM:
4013 smrv_strong_next&=~(1<<dops[i].rt1);
4014 smrv_weak_next&=~(1<<dops[i].rt1);
4015 // fallthrough
4016 case IMM16:
4017 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4018 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
4019 if(hr>=0) {
4020 if(get_final_value(hr,i,&value))
4021 smrv[dops[i].rt1]=value;
4022 else smrv[dops[i].rt1]=constmap[i][hr];
4023 smrv_strong_next|=1<<dops[i].rt1;
4024 }
4025 }
4026 else {
4027 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4028 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4029 }
4030 break;
4031 case LOAD:
4032 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4033 // special case for BIOS
4034 smrv[dops[i].rt1]=0xa0000000;
4035 smrv_strong_next|=1<<dops[i].rt1;
4036 break;
4037 }
4038 // fallthrough
4039 case SHIFT:
4040 case LOADLR:
4041 case MOV:
4042 smrv_strong_next&=~(1<<dops[i].rt1);
4043 smrv_weak_next&=~(1<<dops[i].rt1);
4044 break;
4045 case COP0:
4046 case COP2:
4047 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4048 smrv_strong_next&=~(1<<dops[i].rt1);
4049 smrv_weak_next&=~(1<<dops[i].rt1);
4050 }
4051 break;
4052 case C2LS:
4053 if (dops[i].opcode==0x32) { // LWC2
4054 smrv_strong_next&=~(1<<dops[i].rt1);
4055 smrv_weak_next&=~(1<<dops[i].rt1);
4056 }
4057 break;
4058 }
4059#if 0
4060 int r=4;
4061 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4062 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4063#endif
4064}
4065
4066static void ujump_assemble(int i, const struct regstat *i_regs);
4067static void rjump_assemble(int i, const struct regstat *i_regs);
4068static void cjump_assemble(int i, const struct regstat *i_regs);
4069static void sjump_assemble(int i, const struct regstat *i_regs);
4070static void pagespan_assemble(int i, const struct regstat *i_regs);
4071
4072static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4073{
4074 int ds = 0;
4075 switch (dops[i].itype) {
4076 case ALU:
4077 alu_assemble(i, i_regs);
4078 break;
4079 case IMM16:
4080 imm16_assemble(i, i_regs);
4081 break;
4082 case SHIFT:
4083 shift_assemble(i, i_regs);
4084 break;
4085 case SHIFTIMM:
4086 shiftimm_assemble(i, i_regs);
4087 break;
4088 case LOAD:
4089 load_assemble(i, i_regs, ccadj_);
4090 break;
4091 case LOADLR:
4092 loadlr_assemble(i, i_regs, ccadj_);
4093 break;
4094 case STORE:
4095 store_assemble(i, i_regs, ccadj_);
4096 break;
4097 case STORELR:
4098 storelr_assemble(i, i_regs, ccadj_);
4099 break;
4100 case COP0:
4101 cop0_assemble(i, i_regs, ccadj_);
4102 break;
4103 case COP1:
4104 cop1_assemble(i, i_regs);
4105 break;
4106 case C1LS:
4107 c1ls_assemble(i, i_regs);
4108 break;
4109 case COP2:
4110 cop2_assemble(i, i_regs);
4111 break;
4112 case C2LS:
4113 c2ls_assemble(i, i_regs, ccadj_);
4114 break;
4115 case C2OP:
4116 c2op_assemble(i, i_regs);
4117 break;
4118 case MULTDIV:
4119 multdiv_assemble(i, i_regs);
4120 multdiv_prepare_stall(i, i_regs, ccadj_);
4121 break;
4122 case MOV:
4123 mov_assemble(i, i_regs);
4124 break;
4125 case SYSCALL:
4126 syscall_assemble(i, i_regs, ccadj_);
4127 break;
4128 case HLECALL:
4129 hlecall_assemble(i, i_regs, ccadj_);
4130 break;
4131 case INTCALL:
4132 intcall_assemble(i, i_regs, ccadj_);
4133 break;
4134 case UJUMP:
4135 ujump_assemble(i, i_regs);
4136 ds = 1;
4137 break;
4138 case RJUMP:
4139 rjump_assemble(i, i_regs);
4140 ds = 1;
4141 break;
4142 case CJUMP:
4143 cjump_assemble(i, i_regs);
4144 ds = 1;
4145 break;
4146 case SJUMP:
4147 sjump_assemble(i, i_regs);
4148 ds = 1;
4149 break;
4150 case SPAN:
4151 pagespan_assemble(i, i_regs);
4152 break;
4153 case NOP:
4154 case OTHER:
4155 case NI:
4156 // not handled, just skip
4157 break;
4158 default:
4159 assert(0);
4160 }
4161 return ds;
4162}
4163
4164static void ds_assemble(int i, const struct regstat *i_regs)
4165{
4166 speculate_register_values(i);
4167 is_delayslot = 1;
4168 switch (dops[i].itype) {
4169 case SYSCALL:
4170 case HLECALL:
4171 case INTCALL:
4172 case SPAN:
4173 case UJUMP:
4174 case RJUMP:
4175 case CJUMP:
4176 case SJUMP:
4177 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4178 break;
4179 default:
4180 assemble(i, i_regs, ccadj[i]);
4181 }
4182 is_delayslot = 0;
4183}
4184
4185// Is the branch target a valid internal jump?
4186static int internal_branch(int addr)
4187{
4188 if(addr&1) return 0; // Indirect (register) jump
4189 if(addr>=start && addr<start+slen*4-4)
4190 {
4191 return 1;
4192 }
4193 return 0;
4194}
4195
4196static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4197{
4198 int hr;
4199 for(hr=0;hr<HOST_REGS;hr++) {
4200 if(hr!=EXCLUDE_REG) {
4201 if(pre[hr]!=entry[hr]) {
4202 if(pre[hr]>=0) {
4203 if((dirty>>hr)&1) {
4204 if(get_reg(entry,pre[hr])<0) {
4205 assert(pre[hr]<64);
4206 if(!((u>>pre[hr])&1))
4207 emit_storereg(pre[hr],hr);
4208 }
4209 }
4210 }
4211 }
4212 }
4213 }
4214 // Move from one register to another (no writeback)
4215 for(hr=0;hr<HOST_REGS;hr++) {
4216 if(hr!=EXCLUDE_REG) {
4217 if(pre[hr]!=entry[hr]) {
4218 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
4219 int nr;
4220 if((nr=get_reg(entry,pre[hr]))>=0) {
4221 emit_mov(hr,nr);
4222 }
4223 }
4224 }
4225 }
4226 }
4227}
4228
4229// Load the specified registers
4230// This only loads the registers given as arguments because
4231// we don't want to load things that will be overwritten
4232static void load_regs(signed char entry[],signed char regmap[],int rs1,int rs2)
4233{
4234 int hr;
4235 // Load 32-bit regs
4236 for(hr=0;hr<HOST_REGS;hr++) {
4237 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4238 if(entry[hr]!=regmap[hr]) {
4239 if(regmap[hr]==rs1||regmap[hr]==rs2)
4240 {
4241 if(regmap[hr]==0) {
4242 emit_zeroreg(hr);
4243 }
4244 else
4245 {
4246 emit_loadreg(regmap[hr],hr);
4247 }
4248 }
4249 }
4250 }
4251 }
4252}
4253
4254// Load registers prior to the start of a loop
4255// so that they are not loaded within the loop
4256static void loop_preload(signed char pre[],signed char entry[])
4257{
4258 int hr;
4259 for(hr=0;hr<HOST_REGS;hr++) {
4260 if(hr!=EXCLUDE_REG) {
4261 if(pre[hr]!=entry[hr]) {
4262 if(entry[hr]>=0) {
4263 if(get_reg(pre,entry[hr])<0) {
4264 assem_debug("loop preload:\n");
4265 //printf("loop preload: %d\n",hr);
4266 if(entry[hr]==0) {
4267 emit_zeroreg(hr);
4268 }
4269 else if(entry[hr]<TEMPREG)
4270 {
4271 emit_loadreg(entry[hr],hr);
4272 }
4273 else if(entry[hr]-64<TEMPREG)
4274 {
4275 emit_loadreg(entry[hr],hr);
4276 }
4277 }
4278 }
4279 }
4280 }
4281 }
4282}
4283
4284// Generate address for load/store instruction
4285// goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4286void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4287{
4288 if (dops[i].is_load || dops[i].is_store) {
4289 int ra=-1;
4290 int agr=AGEN1+(i&1);
4291 if(dops[i].itype==LOAD) {
4292 ra=get_reg(i_regs->regmap,dops[i].rt1);
4293 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4294 assert(ra>=0);
4295 }
4296 if(dops[i].itype==LOADLR) {
4297 ra=get_reg(i_regs->regmap,FTEMP);
4298 }
4299 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4300 ra=get_reg(i_regs->regmap,agr);
4301 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4302 }
4303 if(dops[i].itype==C2LS) {
4304 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4305 ra=get_reg(i_regs->regmap,FTEMP);
4306 else { // SWC1/SDC1/SWC2/SDC2
4307 ra=get_reg(i_regs->regmap,agr);
4308 if(ra<0) ra=get_reg(i_regs->regmap,-1);
4309 }
4310 }
4311 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4312 if(ra>=0) {
4313 int offset=imm[i];
4314 int c=(i_regs->wasconst>>rs)&1;
4315 if(dops[i].rs1==0) {
4316 // Using r0 as a base address
4317 if(!entry||entry[ra]!=agr) {
4318 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4319 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4320 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4321 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4322 }else{
4323 emit_movimm(offset,ra);
4324 }
4325 } // else did it in the previous cycle
4326 }
4327 else if(rs<0) {
4328 if(!entry||entry[ra]!=dops[i].rs1)
4329 emit_loadreg(dops[i].rs1,ra);
4330 //if(!entry||entry[ra]!=dops[i].rs1)
4331 // printf("poor load scheduling!\n");
4332 }
4333 else if(c) {
4334 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4335 if(!entry||entry[ra]!=agr) {
4336 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4337 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4338 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4339 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4340 }else{
4341 emit_movimm(constmap[i][rs]+offset,ra);
4342 regs[i].loadedconst|=1<<ra;
4343 }
4344 } // else did it in the previous cycle
4345 } // else load_consts already did it
4346 }
4347 if(offset&&!c&&dops[i].rs1) {
4348 if(rs>=0) {
4349 emit_addimm(rs,offset,ra);
4350 }else{
4351 emit_addimm(ra,offset,ra);
4352 }
4353 }
4354 }
4355 }
4356 // Preload constants for next instruction
4357 if (dops[i+1].is_load || dops[i+1].is_store) {
4358 int agr,ra;
4359 // Actual address
4360 agr=AGEN1+((i+1)&1);
4361 ra=get_reg(i_regs->regmap,agr);
4362 if(ra>=0) {
4363 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4364 int offset=imm[i+1];
4365 int c=(regs[i+1].wasconst>>rs)&1;
4366 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4367 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4368 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4369 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4370 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4371 }else{
4372 emit_movimm(constmap[i+1][rs]+offset,ra);
4373 regs[i+1].loadedconst|=1<<ra;
4374 }
4375 }
4376 else if(dops[i+1].rs1==0) {
4377 // Using r0 as a base address
4378 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4379 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4380 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4381 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4382 }else{
4383 emit_movimm(offset,ra);
4384 }
4385 }
4386 }
4387 }
4388}
4389
4390static int get_final_value(int hr, int i, int *value)
4391{
4392 int reg=regs[i].regmap[hr];
4393 while(i<slen-1) {
4394 if(regs[i+1].regmap[hr]!=reg) break;
4395 if(!((regs[i+1].isconst>>hr)&1)) break;
4396 if(dops[i+1].bt) break;
4397 i++;
4398 }
4399 if(i<slen-1) {
4400 if (dops[i].is_jump) {
4401 *value=constmap[i][hr];
4402 return 1;
4403 }
4404 if(!dops[i+1].bt) {
4405 if (dops[i+1].is_jump) {
4406 // Load in delay slot, out-of-order execution
4407 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4408 {
4409 // Precompute load address
4410 *value=constmap[i][hr]+imm[i+2];
4411 return 1;
4412 }
4413 }
4414 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4415 {
4416 // Precompute load address
4417 *value=constmap[i][hr]+imm[i+1];
4418 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4419 return 1;
4420 }
4421 }
4422 }
4423 *value=constmap[i][hr];
4424 //printf("c=%lx\n",(long)constmap[i][hr]);
4425 if(i==slen-1) return 1;
4426 assert(reg < 64);
4427 return !((unneeded_reg[i+1]>>reg)&1);
4428}
4429
4430// Load registers with known constants
4431static void load_consts(signed char pre[],signed char regmap[],int i)
4432{
4433 int hr,hr2;
4434 // propagate loaded constant flags
4435 if(i==0||dops[i].bt)
4436 regs[i].loadedconst=0;
4437 else {
4438 for(hr=0;hr<HOST_REGS;hr++) {
4439 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4440 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4441 {
4442 regs[i].loadedconst|=1<<hr;
4443 }
4444 }
4445 }
4446 // Load 32-bit regs
4447 for(hr=0;hr<HOST_REGS;hr++) {
4448 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4449 //if(entry[hr]!=regmap[hr]) {
4450 if(!((regs[i].loadedconst>>hr)&1)) {
4451 assert(regmap[hr]<64);
4452 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4453 int value,similar=0;
4454 if(get_final_value(hr,i,&value)) {
4455 // see if some other register has similar value
4456 for(hr2=0;hr2<HOST_REGS;hr2++) {
4457 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4458 if(is_similar_value(value,constmap[i][hr2])) {
4459 similar=1;
4460 break;
4461 }
4462 }
4463 }
4464 if(similar) {
4465 int value2;
4466 if(get_final_value(hr2,i,&value2)) // is this needed?
4467 emit_movimm_from(value2,hr2,value,hr);
4468 else
4469 emit_movimm(value,hr);
4470 }
4471 else if(value==0) {
4472 emit_zeroreg(hr);
4473 }
4474 else {
4475 emit_movimm(value,hr);
4476 }
4477 }
4478 regs[i].loadedconst|=1<<hr;
4479 }
4480 }
4481 }
4482 }
4483}
4484
4485static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4486{
4487 int hr;
4488 // Load 32-bit regs
4489 for(hr=0;hr<HOST_REGS;hr++) {
4490 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4491 assert(regmap[hr] < 64);
4492 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4493 int value=constmap[i][hr];
4494 if(value==0) {
4495 emit_zeroreg(hr);
4496 }
4497 else {
4498 emit_movimm(value,hr);
4499 }
4500 }
4501 }
4502 }
4503}
4504
4505// Write out all dirty registers (except cycle count)
4506static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4507{
4508 int hr;
4509 for(hr=0;hr<HOST_REGS;hr++) {
4510 if(hr!=EXCLUDE_REG) {
4511 if(i_regmap[hr]>0) {
4512 if(i_regmap[hr]!=CCREG) {
4513 if((i_dirty>>hr)&1) {
4514 assert(i_regmap[hr]<64);
4515 emit_storereg(i_regmap[hr],hr);
4516 }
4517 }
4518 }
4519 }
4520 }
4521}
4522
4523// Write out dirty registers that we need to reload (pair with load_needed_regs)
4524// This writes the registers not written by store_regs_bt
4525static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4526{
4527 int hr;
4528 int t=(addr-start)>>2;
4529 for(hr=0;hr<HOST_REGS;hr++) {
4530 if(hr!=EXCLUDE_REG) {
4531 if(i_regmap[hr]>0) {
4532 if(i_regmap[hr]!=CCREG) {
4533 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4534 if((i_dirty>>hr)&1) {
4535 assert(i_regmap[hr]<64);
4536 emit_storereg(i_regmap[hr],hr);
4537 }
4538 }
4539 }
4540 }
4541 }
4542 }
4543}
4544
4545// Load all registers (except cycle count)
4546static void load_all_regs(const signed char i_regmap[])
4547{
4548 int hr;
4549 for(hr=0;hr<HOST_REGS;hr++) {
4550 if(hr!=EXCLUDE_REG) {
4551 if(i_regmap[hr]==0) {
4552 emit_zeroreg(hr);
4553 }
4554 else
4555 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4556 {
4557 emit_loadreg(i_regmap[hr],hr);
4558 }
4559 }
4560 }
4561}
4562
4563// Load all current registers also needed by next instruction
4564static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4565{
4566 int hr;
4567 for(hr=0;hr<HOST_REGS;hr++) {
4568 if(hr!=EXCLUDE_REG) {
4569 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4570 if(i_regmap[hr]==0) {
4571 emit_zeroreg(hr);
4572 }
4573 else
4574 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
4575 {
4576 emit_loadreg(i_regmap[hr],hr);
4577 }
4578 }
4579 }
4580 }
4581}
4582
4583// Load all regs, storing cycle count if necessary
4584static void load_regs_entry(int t)
4585{
4586 int hr;
4587 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4588 else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t],HOST_CCREG);
4589 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4590 emit_storereg(CCREG,HOST_CCREG);
4591 }
4592 // Load 32-bit regs
4593 for(hr=0;hr<HOST_REGS;hr++) {
4594 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4595 if(regs[t].regmap_entry[hr]==0) {
4596 emit_zeroreg(hr);
4597 }
4598 else if(regs[t].regmap_entry[hr]!=CCREG)
4599 {
4600 emit_loadreg(regs[t].regmap_entry[hr],hr);
4601 }
4602 }
4603 }
4604}
4605
4606// Store dirty registers prior to branch
4607void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4608{
4609 if(internal_branch(addr))
4610 {
4611 int t=(addr-start)>>2;
4612 int hr;
4613 for(hr=0;hr<HOST_REGS;hr++) {
4614 if(hr!=EXCLUDE_REG) {
4615 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4616 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4617 if((i_dirty>>hr)&1) {
4618 assert(i_regmap[hr]<64);
4619 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4620 emit_storereg(i_regmap[hr],hr);
4621 }
4622 }
4623 }
4624 }
4625 }
4626 }
4627 else
4628 {
4629 // Branch out of this block, write out all dirty regs
4630 wb_dirtys(i_regmap,i_dirty);
4631 }
4632}
4633
4634// Load all needed registers for branch target
4635static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4636{
4637 //if(addr>=start && addr<(start+slen*4))
4638 if(internal_branch(addr))
4639 {
4640 int t=(addr-start)>>2;
4641 int hr;
4642 // Store the cycle count before loading something else
4643 if(i_regmap[HOST_CCREG]!=CCREG) {
4644 assert(i_regmap[HOST_CCREG]==-1);
4645 }
4646 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4647 emit_storereg(CCREG,HOST_CCREG);
4648 }
4649 // Load 32-bit regs
4650 for(hr=0;hr<HOST_REGS;hr++) {
4651 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4652 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4653 if(regs[t].regmap_entry[hr]==0) {
4654 emit_zeroreg(hr);
4655 }
4656 else if(regs[t].regmap_entry[hr]!=CCREG)
4657 {
4658 emit_loadreg(regs[t].regmap_entry[hr],hr);
4659 }
4660 }
4661 }
4662 }
4663 }
4664}
4665
4666static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4667{
4668 if(addr>=start && addr<start+slen*4-4)
4669 {
4670 int t=(addr-start)>>2;
4671 int hr;
4672 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4673 for(hr=0;hr<HOST_REGS;hr++)
4674 {
4675 if(hr!=EXCLUDE_REG)
4676 {
4677 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4678 {
4679 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4680 {
4681 return 0;
4682 }
4683 else
4684 if((i_dirty>>hr)&1)
4685 {
4686 if(i_regmap[hr]<TEMPREG)
4687 {
4688 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4689 return 0;
4690 }
4691 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4692 {
4693 assert(0);
4694 }
4695 }
4696 }
4697 else // Same register but is it 32-bit or dirty?
4698 if(i_regmap[hr]>=0)
4699 {
4700 if(!((regs[t].dirty>>hr)&1))
4701 {
4702 if((i_dirty>>hr)&1)
4703 {
4704 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4705 {
4706 //printf("%x: dirty no match\n",addr);
4707 return 0;
4708 }
4709 }
4710 }
4711 }
4712 }
4713 }
4714 // Delay slots are not valid branch targets
4715 //if(t>0&&(dops[t-1].is_jump) return 0;
4716 // Delay slots require additional processing, so do not match
4717 if(dops[t].is_ds) return 0;
4718 }
4719 else
4720 {
4721 int hr;
4722 for(hr=0;hr<HOST_REGS;hr++)
4723 {
4724 if(hr!=EXCLUDE_REG)
4725 {
4726 if(i_regmap[hr]>=0)
4727 {
4728 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4729 {
4730 if((i_dirty>>hr)&1)
4731 {
4732 return 0;
4733 }
4734 }
4735 }
4736 }
4737 }
4738 }
4739 return 1;
4740}
4741
4742#ifdef DRC_DBG
4743static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4744{
4745 extern void do_insn_cmp();
4746 //extern int cycle;
4747 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4748
4749 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4750 save_regs(reglist);
4751 // write out changed consts to match the interpreter
4752 if (i > 0 && !dops[i].bt) {
4753 for (hr = 0; hr < HOST_REGS; hr++) {
4754 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4755 if (hr == EXCLUDE_REG || reg < 0)
4756 continue;
4757 if (!((regs[i-1].isconst >> hr) & 1))
4758 continue;
4759 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4760 continue;
4761 emit_movimm(constmap[i-1][hr],0);
4762 emit_storereg(reg, 0);
4763 }
4764 }
4765 emit_movimm(start+i*4,0);
4766 emit_writeword(0,&pcaddr);
4767 int cc = get_reg(regs[i].regmap_entry, CCREG);
4768 if (cc < 0)
4769 emit_loadreg(CCREG, cc = 0);
4770 emit_addimm(cc, ccadj_, 0);
4771 emit_writeword(0, &psxRegs.cycle);
4772 emit_far_call(do_insn_cmp);
4773 //emit_readword(&cycle,0);
4774 //emit_addimm(0,2,0);
4775 //emit_writeword(0,&cycle);
4776 (void)get_reg2;
4777 restore_regs(reglist);
4778 assem_debug("\\\\do_insn_cmp\n");
4779}
4780#else
4781#define drc_dbg_emit_do_cmp(x,y)
4782#endif
4783
4784// Used when a branch jumps into the delay slot of another branch
4785static void ds_assemble_entry(int i)
4786{
4787 int t = (ba[i] - start) >> 2;
4788 int ccadj_ = -CLOCK_ADJUST(1);
4789 if (!instr_addr[t])
4790 instr_addr[t] = out;
4791 assem_debug("Assemble delay slot at %x\n",ba[i]);
4792 assem_debug("<->\n");
4793 drc_dbg_emit_do_cmp(t, ccadj_);
4794 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4795 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4796 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4797 address_generation(t,&regs[t],regs[t].regmap_entry);
4798 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4799 load_regs(regs[t].regmap_entry,regs[t].regmap,ROREG,ROREG);
4800 if (dops[t].is_store)
4801 load_regs(regs[t].regmap_entry,regs[t].regmap,INVCP,INVCP);
4802 is_delayslot=0;
4803 switch (dops[t].itype) {
4804 case SYSCALL:
4805 case HLECALL:
4806 case INTCALL:
4807 case SPAN:
4808 case UJUMP:
4809 case RJUMP:
4810 case CJUMP:
4811 case SJUMP:
4812 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4813 break;
4814 default:
4815 assemble(t, &regs[t], ccadj_);
4816 }
4817 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4818 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4819 if(internal_branch(ba[i]+4))
4820 assem_debug("branch: internal\n");
4821 else
4822 assem_debug("branch: external\n");
4823 assert(internal_branch(ba[i]+4));
4824 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4825 emit_jmp(0);
4826}
4827
4828static void emit_extjump(void *addr, u_int target)
4829{
4830 emit_extjump2(addr, target, dyna_linker);
4831}
4832
4833static void emit_extjump_ds(void *addr, u_int target)
4834{
4835 emit_extjump2(addr, target, dyna_linker_ds);
4836}
4837
4838// Load 2 immediates optimizing for small code size
4839static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4840{
4841 emit_movimm(imm1,rt1);
4842 emit_movimm_from(imm1,rt1,imm2,rt2);
4843}
4844
4845static void do_cc(int i, const signed char i_regmap[], int *adj,
4846 int addr, int taken, int invert)
4847{
4848 int count, count_plus2;
4849 void *jaddr;
4850 void *idle=NULL;
4851 int t=0;
4852 if(dops[i].itype==RJUMP)
4853 {
4854 *adj=0;
4855 }
4856 //if(ba[i]>=start && ba[i]<(start+slen*4))
4857 if(internal_branch(ba[i]))
4858 {
4859 t=(ba[i]-start)>>2;
4860 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
4861 else *adj=ccadj[t];
4862 }
4863 else
4864 {
4865 *adj=0;
4866 }
4867 count = ccadj[i];
4868 count_plus2 = count + CLOCK_ADJUST(2);
4869 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4870 // Idle loop
4871 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4872 idle=out;
4873 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4874 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4875 jaddr=out;
4876 emit_jmp(0);
4877 }
4878 else if(*adj==0||invert) {
4879 int cycles = count_plus2;
4880 // faster loop HACK
4881#if 0
4882 if (t&&*adj) {
4883 int rel=t-i;
4884 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4885 cycles=*adj+count+2-*adj;
4886 }
4887#endif
4888 emit_addimm_and_set_flags(cycles, HOST_CCREG);
4889 jaddr = out;
4890 emit_jns(0);
4891 }
4892 else
4893 {
4894 emit_cmpimm(HOST_CCREG, -count_plus2);
4895 jaddr = out;
4896 emit_jns(0);
4897 }
4898 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
4899}
4900
4901static void do_ccstub(int n)
4902{
4903 literal_pool(256);
4904 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
4905 set_jump_target(stubs[n].addr, out);
4906 int i=stubs[n].b;
4907 if(stubs[n].d==NULLDS) {
4908 // Delay slot instruction is nullified ("likely" branch)
4909 wb_dirtys(regs[i].regmap,regs[i].dirty);
4910 }
4911 else if(stubs[n].d!=TAKEN) {
4912 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
4913 }
4914 else {
4915 if(internal_branch(ba[i]))
4916 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4917 }
4918 if(stubs[n].c!=-1)
4919 {
4920 // Save PC as return address
4921 emit_movimm(stubs[n].c,EAX);
4922 emit_writeword(EAX,&pcaddr);
4923 }
4924 else
4925 {
4926 // Return address depends on which way the branch goes
4927 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
4928 {
4929 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
4930 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
4931 if(dops[i].rs1==0)
4932 {
4933 s1l=s2l;
4934 s2l=-1;
4935 }
4936 else if(dops[i].rs2==0)
4937 {
4938 s2l=-1;
4939 }
4940 assert(s1l>=0);
4941 #ifdef DESTRUCTIVE_WRITEBACK
4942 if(dops[i].rs1) {
4943 if((branch_regs[i].dirty>>s1l)&&1)
4944 emit_loadreg(dops[i].rs1,s1l);
4945 }
4946 else {
4947 if((branch_regs[i].dirty>>s1l)&1)
4948 emit_loadreg(dops[i].rs2,s1l);
4949 }
4950 if(s2l>=0)
4951 if((branch_regs[i].dirty>>s2l)&1)
4952 emit_loadreg(dops[i].rs2,s2l);
4953 #endif
4954 int hr=0;
4955 int addr=-1,alt=-1,ntaddr=-1;
4956 while(hr<HOST_REGS)
4957 {
4958 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4959 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4960 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4961 {
4962 addr=hr++;break;
4963 }
4964 hr++;
4965 }
4966 while(hr<HOST_REGS)
4967 {
4968 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4969 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4970 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4971 {
4972 alt=hr++;break;
4973 }
4974 hr++;
4975 }
4976 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
4977 {
4978 while(hr<HOST_REGS)
4979 {
4980 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4981 (branch_regs[i].regmap[hr]&63)!=dops[i].rs1 &&
4982 (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 )
4983 {
4984 ntaddr=hr;break;
4985 }
4986 hr++;
4987 }
4988 assert(hr<HOST_REGS);
4989 }
4990 if((dops[i].opcode&0x2f)==4) // BEQ
4991 {
4992 #ifdef HAVE_CMOV_IMM
4993 if(s2l>=0) emit_cmp(s1l,s2l);
4994 else emit_test(s1l,s1l);
4995 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4996 #else
4997 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4998 if(s2l>=0) emit_cmp(s1l,s2l);
4999 else emit_test(s1l,s1l);
5000 emit_cmovne_reg(alt,addr);
5001 #endif
5002 }
5003 if((dops[i].opcode&0x2f)==5) // BNE
5004 {
5005 #ifdef HAVE_CMOV_IMM
5006 if(s2l>=0) emit_cmp(s1l,s2l);
5007 else emit_test(s1l,s1l);
5008 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5009 #else
5010 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5011 if(s2l>=0) emit_cmp(s1l,s2l);
5012 else emit_test(s1l,s1l);
5013 emit_cmovne_reg(alt,addr);
5014 #endif
5015 }
5016 if((dops[i].opcode&0x2f)==6) // BLEZ
5017 {
5018 //emit_movimm(ba[i],alt);
5019 //emit_movimm(start+i*4+8,addr);
5020 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5021 emit_cmpimm(s1l,1);
5022 emit_cmovl_reg(alt,addr);
5023 }
5024 if((dops[i].opcode&0x2f)==7) // BGTZ
5025 {
5026 //emit_movimm(ba[i],addr);
5027 //emit_movimm(start+i*4+8,ntaddr);
5028 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5029 emit_cmpimm(s1l,1);
5030 emit_cmovl_reg(ntaddr,addr);
5031 }
5032 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
5033 {
5034 //emit_movimm(ba[i],alt);
5035 //emit_movimm(start+i*4+8,addr);
5036 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5037 emit_test(s1l,s1l);
5038 emit_cmovs_reg(alt,addr);
5039 }
5040 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
5041 {
5042 //emit_movimm(ba[i],addr);
5043 //emit_movimm(start+i*4+8,alt);
5044 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5045 emit_test(s1l,s1l);
5046 emit_cmovs_reg(alt,addr);
5047 }
5048 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5049 if(source[i]&0x10000) // BC1T
5050 {
5051 //emit_movimm(ba[i],alt);
5052 //emit_movimm(start+i*4+8,addr);
5053 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5054 emit_testimm(s1l,0x800000);
5055 emit_cmovne_reg(alt,addr);
5056 }
5057 else // BC1F
5058 {
5059 //emit_movimm(ba[i],addr);
5060 //emit_movimm(start+i*4+8,alt);
5061 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5062 emit_testimm(s1l,0x800000);
5063 emit_cmovne_reg(alt,addr);
5064 }
5065 }
5066 emit_writeword(addr,&pcaddr);
5067 }
5068 else
5069 if(dops[i].itype==RJUMP)
5070 {
5071 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5072 if (ds_writes_rjump_rs(i)) {
5073 r=get_reg(branch_regs[i].regmap,RTEMP);
5074 }
5075 emit_writeword(r,&pcaddr);
5076 }
5077 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5078 }
5079 // Update cycle count
5080 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5081 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5082 emit_far_call(cc_interrupt);
5083 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5084 if(stubs[n].d==TAKEN) {
5085 if(internal_branch(ba[i]))
5086 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
5087 else if(dops[i].itype==RJUMP) {
5088 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5089 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5090 else
5091 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5092 }
5093 }else if(stubs[n].d==NOTTAKEN) {
5094 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5095 else load_all_regs(branch_regs[i].regmap);
5096 }else if(stubs[n].d==NULLDS) {
5097 // Delay slot instruction is nullified ("likely" branch)
5098 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5099 else load_all_regs(regs[i].regmap);
5100 }else{
5101 load_all_regs(branch_regs[i].regmap);
5102 }
5103 if (stubs[n].retaddr)
5104 emit_jmp(stubs[n].retaddr);
5105 else
5106 do_jump_vaddr(stubs[n].e);
5107}
5108
5109static void add_to_linker(void *addr, u_int target, int ext)
5110{
5111 assert(linkcount < ARRAY_SIZE(link_addr));
5112 link_addr[linkcount].addr = addr;
5113 link_addr[linkcount].target = target;
5114 link_addr[linkcount].ext = ext;
5115 linkcount++;
5116}
5117
5118static void ujump_assemble_write_ra(int i)
5119{
5120 int rt;
5121 unsigned int return_address;
5122 rt=get_reg(branch_regs[i].regmap,31);
5123 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]);
5124 //assert(rt>=0);
5125 return_address=start+i*4+8;
5126 if(rt>=0) {
5127 #ifdef USE_MINI_HT
5128 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5129 int temp=-1; // note: must be ds-safe
5130 #ifdef HOST_TEMPREG
5131 temp=HOST_TEMPREG;
5132 #endif
5133 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5134 else emit_movimm(return_address,rt);
5135 }
5136 else
5137 #endif
5138 {
5139 #ifdef REG_PREFETCH
5140 if(temp>=0)
5141 {
5142 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5143 }
5144 #endif
5145 emit_movimm(return_address,rt); // PC into link register
5146 #ifdef IMM_PREFETCH
5147 emit_prefetch(hash_table_get(return_address));
5148 #endif
5149 }
5150 }
5151}
5152
5153static void ujump_assemble(int i, const struct regstat *i_regs)
5154{
5155 int ra_done=0;
5156 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5157 address_generation(i+1,i_regs,regs[i].regmap_entry);
5158 #ifdef REG_PREFETCH
5159 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5160 if(dops[i].rt1==31&&temp>=0)
5161 {
5162 signed char *i_regmap=i_regs->regmap;
5163 int return_address=start+i*4+8;
5164 if(get_reg(branch_regs[i].regmap,31)>0)
5165 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5166 }
5167 #endif
5168 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5169 ujump_assemble_write_ra(i); // writeback ra for DS
5170 ra_done=1;
5171 }
5172 ds_assemble(i+1,i_regs);
5173 uint64_t bc_unneeded=branch_regs[i].u;
5174 bc_unneeded|=1|(1LL<<dops[i].rt1);
5175 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5176 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5177 if(!ra_done&&dops[i].rt1==31)
5178 ujump_assemble_write_ra(i);
5179 int cc,adj;
5180 cc=get_reg(branch_regs[i].regmap,CCREG);
5181 assert(cc==HOST_CCREG);
5182 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5183 #ifdef REG_PREFETCH
5184 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5185 #endif
5186 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5187 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5188 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5189 if(internal_branch(ba[i]))
5190 assem_debug("branch: internal\n");
5191 else
5192 assem_debug("branch: external\n");
5193 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5194 ds_assemble_entry(i);
5195 }
5196 else {
5197 add_to_linker(out,ba[i],internal_branch(ba[i]));
5198 emit_jmp(0);
5199 }
5200}
5201
5202static void rjump_assemble_write_ra(int i)
5203{
5204 int rt,return_address;
5205 assert(dops[i+1].rt1!=dops[i].rt1);
5206 assert(dops[i+1].rt2!=dops[i].rt1);
5207 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5208 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]);
5209 assert(rt>=0);
5210 return_address=start+i*4+8;
5211 #ifdef REG_PREFETCH
5212 if(temp>=0)
5213 {
5214 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5215 }
5216 #endif
5217 emit_movimm(return_address,rt); // PC into link register
5218 #ifdef IMM_PREFETCH
5219 emit_prefetch(hash_table_get(return_address));
5220 #endif
5221}
5222
5223static void rjump_assemble(int i, const struct regstat *i_regs)
5224{
5225 int temp;
5226 int rs,cc;
5227 int ra_done=0;
5228 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5229 assert(rs>=0);
5230 if (ds_writes_rjump_rs(i)) {
5231 // Delay slot abuse, make a copy of the branch address register
5232 temp=get_reg(branch_regs[i].regmap,RTEMP);
5233 assert(temp>=0);
5234 assert(regs[i].regmap[temp]==RTEMP);
5235 emit_mov(rs,temp);
5236 rs=temp;
5237 }
5238 address_generation(i+1,i_regs,regs[i].regmap_entry);
5239 #ifdef REG_PREFETCH
5240 if(dops[i].rt1==31)
5241 {
5242 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5243 signed char *i_regmap=i_regs->regmap;
5244 int return_address=start+i*4+8;
5245 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5246 }
5247 }
5248 #endif
5249 #ifdef USE_MINI_HT
5250 if(dops[i].rs1==31) {
5251 int rh=get_reg(regs[i].regmap,RHASH);
5252 if(rh>=0) do_preload_rhash(rh);
5253 }
5254 #endif
5255 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5256 rjump_assemble_write_ra(i);
5257 ra_done=1;
5258 }
5259 ds_assemble(i+1,i_regs);
5260 uint64_t bc_unneeded=branch_regs[i].u;
5261 bc_unneeded|=1|(1LL<<dops[i].rt1);
5262 bc_unneeded&=~(1LL<<dops[i].rs1);
5263 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5264 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5265 if(!ra_done&&dops[i].rt1!=0)
5266 rjump_assemble_write_ra(i);
5267 cc=get_reg(branch_regs[i].regmap,CCREG);
5268 assert(cc==HOST_CCREG);
5269 (void)cc;
5270 #ifdef USE_MINI_HT
5271 int rh=get_reg(branch_regs[i].regmap,RHASH);
5272 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5273 if(dops[i].rs1==31) {
5274 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5275 do_preload_rhtbl(ht);
5276 do_rhash(rs,rh);
5277 }
5278 #endif
5279 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5280 #ifdef DESTRUCTIVE_WRITEBACK
5281 if((branch_regs[i].dirty>>rs)&1) {
5282 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5283 emit_loadreg(dops[i].rs1,rs);
5284 }
5285 }
5286 #endif
5287 #ifdef REG_PREFETCH
5288 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5289 #endif
5290 #ifdef USE_MINI_HT
5291 if(dops[i].rs1==31) {
5292 do_miniht_load(ht,rh);
5293 }
5294 #endif
5295 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5296 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5297 //assert(adj==0);
5298 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5299 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5300 if(dops[i+1].itype==COP0&&(source[i+1]&0x3f)==0x10)
5301 // special case for RFE
5302 emit_jmp(0);
5303 else
5304 emit_jns(0);
5305 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5306 #ifdef USE_MINI_HT
5307 if(dops[i].rs1==31) {
5308 do_miniht_jump(rs,rh,ht);
5309 }
5310 else
5311 #endif
5312 {
5313 do_jump_vaddr(rs);
5314 }
5315 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5316 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5317 #endif
5318}
5319
5320static void cjump_assemble(int i, const struct regstat *i_regs)
5321{
5322 const signed char *i_regmap = i_regs->regmap;
5323 int cc;
5324 int match;
5325 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5326 assem_debug("match=%d\n",match);
5327 int s1l,s2l;
5328 int unconditional=0,nop=0;
5329 int invert=0;
5330 int internal=internal_branch(ba[i]);
5331 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5332 if(!match) invert=1;
5333 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5334 if(i>(ba[i]-start)>>2) invert=1;
5335 #endif
5336 #ifdef __aarch64__
5337 invert=1; // because of near cond. branches
5338 #endif
5339
5340 if(dops[i].ooo) {
5341 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5342 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5343 }
5344 else {
5345 s1l=get_reg(i_regmap,dops[i].rs1);
5346 s2l=get_reg(i_regmap,dops[i].rs2);
5347 }
5348 if(dops[i].rs1==0&&dops[i].rs2==0)
5349 {
5350 if(dops[i].opcode&1) nop=1;
5351 else unconditional=1;
5352 //assert(dops[i].opcode!=5);
5353 //assert(dops[i].opcode!=7);
5354 //assert(dops[i].opcode!=0x15);
5355 //assert(dops[i].opcode!=0x17);
5356 }
5357 else if(dops[i].rs1==0)
5358 {
5359 s1l=s2l;
5360 s2l=-1;
5361 }
5362 else if(dops[i].rs2==0)
5363 {
5364 s2l=-1;
5365 }
5366
5367 if(dops[i].ooo) {
5368 // Out of order execution (delay slot first)
5369 //printf("OOOE\n");
5370 address_generation(i+1,i_regs,regs[i].regmap_entry);
5371 ds_assemble(i+1,i_regs);
5372 int adj;
5373 uint64_t bc_unneeded=branch_regs[i].u;
5374 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5375 bc_unneeded|=1;
5376 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5377 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5378 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5379 cc=get_reg(branch_regs[i].regmap,CCREG);
5380 assert(cc==HOST_CCREG);
5381 if(unconditional)
5382 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5383 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5384 //assem_debug("cycle count (adj)\n");
5385 if(unconditional) {
5386 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5387 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5388 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5389 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5390 if(internal)
5391 assem_debug("branch: internal\n");
5392 else
5393 assem_debug("branch: external\n");
5394 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5395 ds_assemble_entry(i);
5396 }
5397 else {
5398 add_to_linker(out,ba[i],internal);
5399 emit_jmp(0);
5400 }
5401 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5402 if(((u_int)out)&7) emit_addnop(0);
5403 #endif
5404 }
5405 }
5406 else if(nop) {
5407 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5408 void *jaddr=out;
5409 emit_jns(0);
5410 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5411 }
5412 else {
5413 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5414 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5415 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5416
5417 //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]);
5418 assert(s1l>=0);
5419 if(dops[i].opcode==4) // BEQ
5420 {
5421 if(s2l>=0) emit_cmp(s1l,s2l);
5422 else emit_test(s1l,s1l);
5423 if(invert){
5424 nottaken=out;
5425 emit_jne(DJT_1);
5426 }else{
5427 add_to_linker(out,ba[i],internal);
5428 emit_jeq(0);
5429 }
5430 }
5431 if(dops[i].opcode==5) // BNE
5432 {
5433 if(s2l>=0) emit_cmp(s1l,s2l);
5434 else emit_test(s1l,s1l);
5435 if(invert){
5436 nottaken=out;
5437 emit_jeq(DJT_1);
5438 }else{
5439 add_to_linker(out,ba[i],internal);
5440 emit_jne(0);
5441 }
5442 }
5443 if(dops[i].opcode==6) // BLEZ
5444 {
5445 emit_cmpimm(s1l,1);
5446 if(invert){
5447 nottaken=out;
5448 emit_jge(DJT_1);
5449 }else{
5450 add_to_linker(out,ba[i],internal);
5451 emit_jl(0);
5452 }
5453 }
5454 if(dops[i].opcode==7) // BGTZ
5455 {
5456 emit_cmpimm(s1l,1);
5457 if(invert){
5458 nottaken=out;
5459 emit_jl(DJT_1);
5460 }else{
5461 add_to_linker(out,ba[i],internal);
5462 emit_jge(0);
5463 }
5464 }
5465 if(invert) {
5466 if(taken) set_jump_target(taken, out);
5467 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5468 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5469 if(adj) {
5470 emit_addimm(cc,-adj,cc);
5471 add_to_linker(out,ba[i],internal);
5472 }else{
5473 emit_addnop(13);
5474 add_to_linker(out,ba[i],internal*2);
5475 }
5476 emit_jmp(0);
5477 }else
5478 #endif
5479 {
5480 if(adj) emit_addimm(cc,-adj,cc);
5481 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5482 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5483 if(internal)
5484 assem_debug("branch: internal\n");
5485 else
5486 assem_debug("branch: external\n");
5487 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5488 ds_assemble_entry(i);
5489 }
5490 else {
5491 add_to_linker(out,ba[i],internal);
5492 emit_jmp(0);
5493 }
5494 }
5495 set_jump_target(nottaken, out);
5496 }
5497
5498 if(nottaken1) set_jump_target(nottaken1, out);
5499 if(adj) {
5500 if(!invert) emit_addimm(cc,adj,cc);
5501 }
5502 } // (!unconditional)
5503 } // if(ooo)
5504 else
5505 {
5506 // In-order execution (branch first)
5507 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5508 if(!unconditional&&!nop) {
5509 //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]);
5510 assert(s1l>=0);
5511 if((dops[i].opcode&0x2f)==4) // BEQ
5512 {
5513 if(s2l>=0) emit_cmp(s1l,s2l);
5514 else emit_test(s1l,s1l);
5515 nottaken=out;
5516 emit_jne(DJT_2);
5517 }
5518 if((dops[i].opcode&0x2f)==5) // BNE
5519 {
5520 if(s2l>=0) emit_cmp(s1l,s2l);
5521 else emit_test(s1l,s1l);
5522 nottaken=out;
5523 emit_jeq(DJT_2);
5524 }
5525 if((dops[i].opcode&0x2f)==6) // BLEZ
5526 {
5527 emit_cmpimm(s1l,1);
5528 nottaken=out;
5529 emit_jge(DJT_2);
5530 }
5531 if((dops[i].opcode&0x2f)==7) // BGTZ
5532 {
5533 emit_cmpimm(s1l,1);
5534 nottaken=out;
5535 emit_jl(DJT_2);
5536 }
5537 } // if(!unconditional)
5538 int adj;
5539 uint64_t ds_unneeded=branch_regs[i].u;
5540 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5541 ds_unneeded|=1;
5542 // branch taken
5543 if(!nop) {
5544 if(taken) set_jump_target(taken, out);
5545 assem_debug("1:\n");
5546 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5547 // load regs
5548 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5549 address_generation(i+1,&branch_regs[i],0);
5550 if (ram_offset)
5551 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5552 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5553 ds_assemble(i+1,&branch_regs[i]);
5554 cc=get_reg(branch_regs[i].regmap,CCREG);
5555 if(cc==-1) {
5556 emit_loadreg(CCREG,cc=HOST_CCREG);
5557 // CHECK: Is the following instruction (fall thru) allocated ok?
5558 }
5559 assert(cc==HOST_CCREG);
5560 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5561 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5562 assem_debug("cycle count (adj)\n");
5563 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5564 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5565 if(internal)
5566 assem_debug("branch: internal\n");
5567 else
5568 assem_debug("branch: external\n");
5569 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5570 ds_assemble_entry(i);
5571 }
5572 else {
5573 add_to_linker(out,ba[i],internal);
5574 emit_jmp(0);
5575 }
5576 }
5577 // branch not taken
5578 if(!unconditional) {
5579 if(nottaken1) set_jump_target(nottaken1, out);
5580 set_jump_target(nottaken, out);
5581 assem_debug("2:\n");
5582 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5583 // load regs
5584 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5585 address_generation(i+1,&branch_regs[i],0);
5586 if (ram_offset)
5587 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5588 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5589 ds_assemble(i+1,&branch_regs[i]);
5590 cc=get_reg(branch_regs[i].regmap,CCREG);
5591 if (cc == -1) {
5592 // Cycle count isn't in a register, temporarily load it then write it out
5593 emit_loadreg(CCREG,HOST_CCREG);
5594 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5595 void *jaddr=out;
5596 emit_jns(0);
5597 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5598 emit_storereg(CCREG,HOST_CCREG);
5599 }
5600 else{
5601 cc=get_reg(i_regmap,CCREG);
5602 assert(cc==HOST_CCREG);
5603 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5604 void *jaddr=out;
5605 emit_jns(0);
5606 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5607 }
5608 }
5609 }
5610}
5611
5612static void sjump_assemble(int i, const struct regstat *i_regs)
5613{
5614 const signed char *i_regmap = i_regs->regmap;
5615 int cc;
5616 int match;
5617 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5618 assem_debug("smatch=%d\n",match);
5619 int s1l;
5620 int unconditional=0,nevertaken=0;
5621 int invert=0;
5622 int internal=internal_branch(ba[i]);
5623 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5624 if(!match) invert=1;
5625 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5626 if(i>(ba[i]-start)>>2) invert=1;
5627 #endif
5628 #ifdef __aarch64__
5629 invert=1; // because of near cond. branches
5630 #endif
5631
5632 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5633 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5634
5635 if(dops[i].ooo) {
5636 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5637 }
5638 else {
5639 s1l=get_reg(i_regmap,dops[i].rs1);
5640 }
5641 if(dops[i].rs1==0)
5642 {
5643 if(dops[i].opcode2&1) unconditional=1;
5644 else nevertaken=1;
5645 // These are never taken (r0 is never less than zero)
5646 //assert(dops[i].opcode2!=0);
5647 //assert(dops[i].opcode2!=2);
5648 //assert(dops[i].opcode2!=0x10);
5649 //assert(dops[i].opcode2!=0x12);
5650 }
5651
5652 if(dops[i].ooo) {
5653 // Out of order execution (delay slot first)
5654 //printf("OOOE\n");
5655 address_generation(i+1,i_regs,regs[i].regmap_entry);
5656 ds_assemble(i+1,i_regs);
5657 int adj;
5658 uint64_t bc_unneeded=branch_regs[i].u;
5659 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5660 bc_unneeded|=1;
5661 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5662 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5663 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5664 if(dops[i].rt1==31) {
5665 int rt,return_address;
5666 rt=get_reg(branch_regs[i].regmap,31);
5667 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]);
5668 if(rt>=0) {
5669 // Save the PC even if the branch is not taken
5670 return_address=start+i*4+8;
5671 emit_movimm(return_address,rt); // PC into link register
5672 #ifdef IMM_PREFETCH
5673 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5674 #endif
5675 }
5676 }
5677 cc=get_reg(branch_regs[i].regmap,CCREG);
5678 assert(cc==HOST_CCREG);
5679 if(unconditional)
5680 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5681 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5682 assem_debug("cycle count (adj)\n");
5683 if(unconditional) {
5684 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5685 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5686 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5687 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5688 if(internal)
5689 assem_debug("branch: internal\n");
5690 else
5691 assem_debug("branch: external\n");
5692 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5693 ds_assemble_entry(i);
5694 }
5695 else {
5696 add_to_linker(out,ba[i],internal);
5697 emit_jmp(0);
5698 }
5699 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5700 if(((u_int)out)&7) emit_addnop(0);
5701 #endif
5702 }
5703 }
5704 else if(nevertaken) {
5705 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5706 void *jaddr=out;
5707 emit_jns(0);
5708 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5709 }
5710 else {
5711 void *nottaken = NULL;
5712 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5713 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5714 {
5715 assert(s1l>=0);
5716 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5717 {
5718 emit_test(s1l,s1l);
5719 if(invert){
5720 nottaken=out;
5721 emit_jns(DJT_1);
5722 }else{
5723 add_to_linker(out,ba[i],internal);
5724 emit_js(0);
5725 }
5726 }
5727 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5728 {
5729 emit_test(s1l,s1l);
5730 if(invert){
5731 nottaken=out;
5732 emit_js(DJT_1);
5733 }else{
5734 add_to_linker(out,ba[i],internal);
5735 emit_jns(0);
5736 }
5737 }
5738 }
5739
5740 if(invert) {
5741 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5742 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5743 if(adj) {
5744 emit_addimm(cc,-adj,cc);
5745 add_to_linker(out,ba[i],internal);
5746 }else{
5747 emit_addnop(13);
5748 add_to_linker(out,ba[i],internal*2);
5749 }
5750 emit_jmp(0);
5751 }else
5752 #endif
5753 {
5754 if(adj) emit_addimm(cc,-adj,cc);
5755 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5756 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5757 if(internal)
5758 assem_debug("branch: internal\n");
5759 else
5760 assem_debug("branch: external\n");
5761 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5762 ds_assemble_entry(i);
5763 }
5764 else {
5765 add_to_linker(out,ba[i],internal);
5766 emit_jmp(0);
5767 }
5768 }
5769 set_jump_target(nottaken, out);
5770 }
5771
5772 if(adj) {
5773 if(!invert) emit_addimm(cc,adj,cc);
5774 }
5775 } // (!unconditional)
5776 } // if(ooo)
5777 else
5778 {
5779 // In-order execution (branch first)
5780 //printf("IOE\n");
5781 void *nottaken = NULL;
5782 if(dops[i].rt1==31) {
5783 int rt,return_address;
5784 rt=get_reg(branch_regs[i].regmap,31);
5785 if(rt>=0) {
5786 // Save the PC even if the branch is not taken
5787 return_address=start+i*4+8;
5788 emit_movimm(return_address,rt); // PC into link register
5789 #ifdef IMM_PREFETCH
5790 emit_prefetch(hash_table_get(return_address));
5791 #endif
5792 }
5793 }
5794 if(!unconditional) {
5795 //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]);
5796 assert(s1l>=0);
5797 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5798 {
5799 emit_test(s1l,s1l);
5800 nottaken=out;
5801 emit_jns(DJT_1);
5802 }
5803 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5804 {
5805 emit_test(s1l,s1l);
5806 nottaken=out;
5807 emit_js(DJT_1);
5808 }
5809 } // if(!unconditional)
5810 int adj;
5811 uint64_t ds_unneeded=branch_regs[i].u;
5812 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5813 ds_unneeded|=1;
5814 // branch taken
5815 if(!nevertaken) {
5816 //assem_debug("1:\n");
5817 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5818 // load regs
5819 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5820 address_generation(i+1,&branch_regs[i],0);
5821 if (ram_offset)
5822 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5823 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5824 ds_assemble(i+1,&branch_regs[i]);
5825 cc=get_reg(branch_regs[i].regmap,CCREG);
5826 if(cc==-1) {
5827 emit_loadreg(CCREG,cc=HOST_CCREG);
5828 // CHECK: Is the following instruction (fall thru) allocated ok?
5829 }
5830 assert(cc==HOST_CCREG);
5831 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5832 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5833 assem_debug("cycle count (adj)\n");
5834 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5835 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5836 if(internal)
5837 assem_debug("branch: internal\n");
5838 else
5839 assem_debug("branch: external\n");
5840 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5841 ds_assemble_entry(i);
5842 }
5843 else {
5844 add_to_linker(out,ba[i],internal);
5845 emit_jmp(0);
5846 }
5847 }
5848 // branch not taken
5849 if(!unconditional) {
5850 set_jump_target(nottaken, out);
5851 assem_debug("1:\n");
5852 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5853 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5854 address_generation(i+1,&branch_regs[i],0);
5855 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5856 ds_assemble(i+1,&branch_regs[i]);
5857 cc=get_reg(branch_regs[i].regmap,CCREG);
5858 if (cc == -1) {
5859 // Cycle count isn't in a register, temporarily load it then write it out
5860 emit_loadreg(CCREG,HOST_CCREG);
5861 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5862 void *jaddr=out;
5863 emit_jns(0);
5864 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5865 emit_storereg(CCREG,HOST_CCREG);
5866 }
5867 else{
5868 cc=get_reg(i_regmap,CCREG);
5869 assert(cc==HOST_CCREG);
5870 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5871 void *jaddr=out;
5872 emit_jns(0);
5873 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5874 }
5875 }
5876 }
5877}
5878
5879static void pagespan_assemble(int i, const struct regstat *i_regs)
5880{
5881 int s1l=get_reg(i_regs->regmap,dops[i].rs1);
5882 int s2l=get_reg(i_regs->regmap,dops[i].rs2);
5883 void *taken = NULL;
5884 void *nottaken = NULL;
5885 int unconditional=0;
5886 if(dops[i].rs1==0)
5887 {
5888 s1l=s2l;
5889 s2l=-1;
5890 }
5891 else if(dops[i].rs2==0)
5892 {
5893 s2l=-1;
5894 }
5895 int hr=0;
5896 int addr=-1,alt=-1,ntaddr=-1;
5897 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5898 else {
5899 while(hr<HOST_REGS)
5900 {
5901 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5902 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5903 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5904 {
5905 addr=hr++;break;
5906 }
5907 hr++;
5908 }
5909 }
5910 while(hr<HOST_REGS)
5911 {
5912 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5913 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5914 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5915 {
5916 alt=hr++;break;
5917 }
5918 hr++;
5919 }
5920 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5921 {
5922 while(hr<HOST_REGS)
5923 {
5924 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5925 (i_regs->regmap[hr]&63)!=dops[i].rs1 &&
5926 (i_regs->regmap[hr]&63)!=dops[i].rs2 )
5927 {
5928 ntaddr=hr;break;
5929 }
5930 hr++;
5931 }
5932 }
5933 assert(hr<HOST_REGS);
5934 if((dops[i].opcode&0x2e)==4||dops[i].opcode==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5935 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
5936 }
5937 emit_addimm(HOST_CCREG, ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5938 if(dops[i].opcode==2) // J
5939 {
5940 unconditional=1;
5941 }
5942 if(dops[i].opcode==3) // JAL
5943 {
5944 // TODO: mini_ht
5945 int rt=get_reg(i_regs->regmap,31);
5946 emit_movimm(start+i*4+8,rt);
5947 unconditional=1;
5948 }
5949 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
5950 {
5951 emit_mov(s1l,addr);
5952 if(dops[i].opcode2==9) // JALR
5953 {
5954 int rt=get_reg(i_regs->regmap,dops[i].rt1);
5955 emit_movimm(start+i*4+8,rt);
5956 }
5957 }
5958 if((dops[i].opcode&0x3f)==4) // BEQ
5959 {
5960 if(dops[i].rs1==dops[i].rs2)
5961 {
5962 unconditional=1;
5963 }
5964 else
5965 #ifdef HAVE_CMOV_IMM
5966 if(1) {
5967 if(s2l>=0) emit_cmp(s1l,s2l);
5968 else emit_test(s1l,s1l);
5969 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5970 }
5971 else
5972 #endif
5973 {
5974 assert(s1l>=0);
5975 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5976 if(s2l>=0) emit_cmp(s1l,s2l);
5977 else emit_test(s1l,s1l);
5978 emit_cmovne_reg(alt,addr);
5979 }
5980 }
5981 if((dops[i].opcode&0x3f)==5) // BNE
5982 {
5983 #ifdef HAVE_CMOV_IMM
5984 if(s2l>=0) emit_cmp(s1l,s2l);
5985 else emit_test(s1l,s1l);
5986 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5987 #else
5988 assert(s1l>=0);
5989 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5990 if(s2l>=0) emit_cmp(s1l,s2l);
5991 else emit_test(s1l,s1l);
5992 emit_cmovne_reg(alt,addr);
5993 #endif
5994 }
5995 if((dops[i].opcode&0x3f)==0x14) // BEQL
5996 {
5997 if(s2l>=0) emit_cmp(s1l,s2l);
5998 else emit_test(s1l,s1l);
5999 if(nottaken) set_jump_target(nottaken, out);
6000 nottaken=out;
6001 emit_jne(0);
6002 }
6003 if((dops[i].opcode&0x3f)==0x15) // BNEL
6004 {
6005 if(s2l>=0) emit_cmp(s1l,s2l);
6006 else emit_test(s1l,s1l);
6007 nottaken=out;
6008 emit_jeq(0);
6009 if(taken) set_jump_target(taken, out);
6010 }
6011 if((dops[i].opcode&0x3f)==6) // BLEZ
6012 {
6013 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6014 emit_cmpimm(s1l,1);
6015 emit_cmovl_reg(alt,addr);
6016 }
6017 if((dops[i].opcode&0x3f)==7) // BGTZ
6018 {
6019 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
6020 emit_cmpimm(s1l,1);
6021 emit_cmovl_reg(ntaddr,addr);
6022 }
6023 if((dops[i].opcode&0x3f)==0x16) // BLEZL
6024 {
6025 assert((dops[i].opcode&0x3f)!=0x16);
6026 }
6027 if((dops[i].opcode&0x3f)==0x17) // BGTZL
6028 {
6029 assert((dops[i].opcode&0x3f)!=0x17);
6030 }
6031 assert(dops[i].opcode!=1); // BLTZ/BGEZ
6032
6033 //FIXME: Check CSREG
6034 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
6035 if((source[i]&0x30000)==0) // BC1F
6036 {
6037 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6038 emit_testimm(s1l,0x800000);
6039 emit_cmovne_reg(alt,addr);
6040 }
6041 if((source[i]&0x30000)==0x10000) // BC1T
6042 {
6043 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6044 emit_testimm(s1l,0x800000);
6045 emit_cmovne_reg(alt,addr);
6046 }
6047 if((source[i]&0x30000)==0x20000) // BC1FL
6048 {
6049 emit_testimm(s1l,0x800000);
6050 nottaken=out;
6051 emit_jne(0);
6052 }
6053 if((source[i]&0x30000)==0x30000) // BC1TL
6054 {
6055 emit_testimm(s1l,0x800000);
6056 nottaken=out;
6057 emit_jeq(0);
6058 }
6059 }
6060
6061 assert(i_regs->regmap[HOST_CCREG]==CCREG);
6062 wb_dirtys(regs[i].regmap,regs[i].dirty);
6063 if(unconditional)
6064 {
6065 emit_movimm(ba[i],HOST_BTREG);
6066 }
6067 else if(addr!=HOST_BTREG)
6068 {
6069 emit_mov(addr,HOST_BTREG);
6070 }
6071 void *branch_addr=out;
6072 emit_jmp(0);
6073 int target_addr=start+i*4+5;
6074 void *stub=out;
6075 void *compiled_target_addr=check_addr(target_addr);
6076 emit_extjump_ds(branch_addr, target_addr);
6077 if(compiled_target_addr) {
6078 set_jump_target(branch_addr, compiled_target_addr);
6079 add_jump_out(target_addr,stub);
6080 }
6081 else set_jump_target(branch_addr, stub);
6082}
6083
6084// Assemble the delay slot for the above
6085static void pagespan_ds()
6086{
6087 assem_debug("initial delay slot:\n");
6088 u_int vaddr=start+1;
6089 u_int page=get_page(vaddr);
6090 u_int vpage=get_vpage(vaddr);
6091 ll_add(jump_dirty+vpage,vaddr,(void *)out);
6092 do_dirty_stub_ds(slen*4);
6093 ll_add(jump_in+page,vaddr,(void *)out);
6094 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
6095 if(regs[0].regmap[HOST_CCREG]!=CCREG)
6096 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
6097 if(regs[0].regmap[HOST_BTREG]!=BTREG)
6098 emit_writeword(HOST_BTREG,&branch_target);
6099 load_regs(regs[0].regmap_entry,regs[0].regmap,dops[0].rs1,dops[0].rs2);
6100 address_generation(0,&regs[0],regs[0].regmap_entry);
6101 if (ram_offset && (dops[0].is_load || dops[0].is_store))
6102 load_regs(regs[0].regmap_entry,regs[0].regmap,ROREG,ROREG);
6103 if (dops[0].is_store)
6104 load_regs(regs[0].regmap_entry,regs[0].regmap,INVCP,INVCP);
6105 is_delayslot=0;
6106 switch (dops[0].itype) {
6107 case SYSCALL:
6108 case HLECALL:
6109 case INTCALL:
6110 case SPAN:
6111 case UJUMP:
6112 case RJUMP:
6113 case CJUMP:
6114 case SJUMP:
6115 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
6116 break;
6117 default:
6118 assemble(0, &regs[0], 0);
6119 }
6120 int btaddr=get_reg(regs[0].regmap,BTREG);
6121 if(btaddr<0) {
6122 btaddr=get_reg(regs[0].regmap,-1);
6123 emit_readword(&branch_target,btaddr);
6124 }
6125 assert(btaddr!=HOST_CCREG);
6126 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
6127#ifdef HOST_IMM8
6128 host_tempreg_acquire();
6129 emit_movimm(start+4,HOST_TEMPREG);
6130 emit_cmp(btaddr,HOST_TEMPREG);
6131 host_tempreg_release();
6132#else
6133 emit_cmpimm(btaddr,start+4);
6134#endif
6135 void *branch = out;
6136 emit_jeq(0);
6137 store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
6138 do_jump_vaddr(btaddr);
6139 set_jump_target(branch, out);
6140 store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
6141 load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
6142}
6143
6144// Basic liveness analysis for MIPS registers
6145void unneeded_registers(int istart,int iend,int r)
6146{
6147 int i;
6148 uint64_t u,gte_u,b,gte_b;
6149 uint64_t temp_u,temp_gte_u=0;
6150 uint64_t gte_u_unknown=0;
6151 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
6152 gte_u_unknown=~0ll;
6153 if(iend==slen-1) {
6154 u=1;
6155 gte_u=gte_u_unknown;
6156 }else{
6157 //u=unneeded_reg[iend+1];
6158 u=1;
6159 gte_u=gte_unneeded[iend+1];
6160 }
6161
6162 for (i=iend;i>=istart;i--)
6163 {
6164 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
6165 if(dops[i].is_jump)
6166 {
6167 // If subroutine call, flag return address as a possible branch target
6168 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
6169
6170 if(ba[i]<start || ba[i]>=(start+slen*4))
6171 {
6172 // Branch out of this block, flush all regs
6173 u=1;
6174 gte_u=gte_u_unknown;
6175 branch_unneeded_reg[i]=u;
6176 // Merge in delay slot
6177 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6178 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6179 u|=1;
6180 gte_u|=gte_rt[i+1];
6181 gte_u&=~gte_rs[i+1];
6182 }
6183 else
6184 {
6185 // Internal branch, flag target
6186 dops[(ba[i]-start)>>2].bt=1;
6187 if(ba[i]<=start+i*4) {
6188 // Backward branch
6189 if(dops[i].is_ujump)
6190 {
6191 // Unconditional branch
6192 temp_u=1;
6193 temp_gte_u=0;
6194 } else {
6195 // Conditional branch (not taken case)
6196 temp_u=unneeded_reg[i+2];
6197 temp_gte_u&=gte_unneeded[i+2];
6198 }
6199 // Merge in delay slot
6200 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6201 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6202 temp_u|=1;
6203 temp_gte_u|=gte_rt[i+1];
6204 temp_gte_u&=~gte_rs[i+1];
6205 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
6206 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
6207 temp_u|=1;
6208 temp_gte_u|=gte_rt[i];
6209 temp_gte_u&=~gte_rs[i];
6210 unneeded_reg[i]=temp_u;
6211 gte_unneeded[i]=temp_gte_u;
6212 // Only go three levels deep. This recursion can take an
6213 // excessive amount of time if there are a lot of nested loops.
6214 if(r<2) {
6215 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
6216 }else{
6217 unneeded_reg[(ba[i]-start)>>2]=1;
6218 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
6219 }
6220 } /*else*/ if(1) {
6221 if (dops[i].is_ujump)
6222 {
6223 // Unconditional branch
6224 u=unneeded_reg[(ba[i]-start)>>2];
6225 gte_u=gte_unneeded[(ba[i]-start)>>2];
6226 branch_unneeded_reg[i]=u;
6227 // Merge in delay slot
6228 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6229 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6230 u|=1;
6231 gte_u|=gte_rt[i+1];
6232 gte_u&=~gte_rs[i+1];
6233 } else {
6234 // Conditional branch
6235 b=unneeded_reg[(ba[i]-start)>>2];
6236 gte_b=gte_unneeded[(ba[i]-start)>>2];
6237 branch_unneeded_reg[i]=b;
6238 // Branch delay slot
6239 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
6240 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
6241 b|=1;
6242 gte_b|=gte_rt[i+1];
6243 gte_b&=~gte_rs[i+1];
6244 u&=b;
6245 gte_u&=gte_b;
6246 if(i<slen-1) {
6247 branch_unneeded_reg[i]&=unneeded_reg[i+2];
6248 } else {
6249 branch_unneeded_reg[i]=1;
6250 }
6251 }
6252 }
6253 }
6254 }
6255 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6256 {
6257 // SYSCALL instruction (software interrupt)
6258 u=1;
6259 }
6260 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6261 {
6262 // ERET instruction (return from interrupt)
6263 u=1;
6264 }
6265 //u=1; // DEBUG
6266 // Written registers are unneeded
6267 u|=1LL<<dops[i].rt1;
6268 u|=1LL<<dops[i].rt2;
6269 gte_u|=gte_rt[i];
6270 // Accessed registers are needed
6271 u&=~(1LL<<dops[i].rs1);
6272 u&=~(1LL<<dops[i].rs2);
6273 gte_u&=~gte_rs[i];
6274 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
6275 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
6276 // Source-target dependencies
6277 // R0 is always unneeded
6278 u|=1;
6279 // Save it
6280 unneeded_reg[i]=u;
6281 gte_unneeded[i]=gte_u;
6282 /*
6283 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
6284 printf("U:");
6285 int r;
6286 for(r=1;r<=CCREG;r++) {
6287 if((unneeded_reg[i]>>r)&1) {
6288 if(r==HIREG) printf(" HI");
6289 else if(r==LOREG) printf(" LO");
6290 else printf(" r%d",r);
6291 }
6292 }
6293 printf("\n");
6294 */
6295 }
6296}
6297
6298// Write back dirty registers as soon as we will no longer modify them,
6299// so that we don't end up with lots of writes at the branches.
6300void clean_registers(int istart,int iend,int wr)
6301{
6302 int i;
6303 int r;
6304 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
6305 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
6306 if(iend==slen-1) {
6307 will_dirty_i=will_dirty_next=0;
6308 wont_dirty_i=wont_dirty_next=0;
6309 }else{
6310 will_dirty_i=will_dirty_next=will_dirty[iend+1];
6311 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
6312 }
6313 for (i=iend;i>=istart;i--)
6314 {
6315 if(dops[i].is_jump)
6316 {
6317 if(ba[i]<start || ba[i]>=(start+slen*4))
6318 {
6319 // Branch out of this block, flush all regs
6320 if (dops[i].is_ujump)
6321 {
6322 // Unconditional branch
6323 will_dirty_i=0;
6324 wont_dirty_i=0;
6325 // Merge in delay slot (will dirty)
6326 for(r=0;r<HOST_REGS;r++) {
6327 if(r!=EXCLUDE_REG) {
6328 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6329 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6330 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6331 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6332 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6333 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6334 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6335 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6336 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6337 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6338 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6339 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6340 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6341 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6342 }
6343 }
6344 }
6345 else
6346 {
6347 // Conditional branch
6348 will_dirty_i=0;
6349 wont_dirty_i=wont_dirty_next;
6350 // Merge in delay slot (will dirty)
6351 for(r=0;r<HOST_REGS;r++) {
6352 if(r!=EXCLUDE_REG) {
6353 if (1) { // !dops[i].likely) {
6354 // Might not dirty if likely branch is not taken
6355 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6356 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6357 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6358 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6359 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6360 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
6361 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6362 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6363 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6364 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6365 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6366 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6367 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6368 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6369 }
6370 }
6371 }
6372 }
6373 // Merge in delay slot (wont dirty)
6374 for(r=0;r<HOST_REGS;r++) {
6375 if(r!=EXCLUDE_REG) {
6376 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6377 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6378 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6379 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6380 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6381 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6382 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6383 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6384 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6385 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6386 }
6387 }
6388 if(wr) {
6389 #ifndef DESTRUCTIVE_WRITEBACK
6390 branch_regs[i].dirty&=wont_dirty_i;
6391 #endif
6392 branch_regs[i].dirty|=will_dirty_i;
6393 }
6394 }
6395 else
6396 {
6397 // Internal branch
6398 if(ba[i]<=start+i*4) {
6399 // Backward branch
6400 if (dops[i].is_ujump)
6401 {
6402 // Unconditional branch
6403 temp_will_dirty=0;
6404 temp_wont_dirty=0;
6405 // Merge in delay slot (will dirty)
6406 for(r=0;r<HOST_REGS;r++) {
6407 if(r!=EXCLUDE_REG) {
6408 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6409 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6410 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6411 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6412 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6413 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6414 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6415 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6416 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6417 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6418 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6419 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6420 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6421 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6422 }
6423 }
6424 } else {
6425 // Conditional branch (not taken case)
6426 temp_will_dirty=will_dirty_next;
6427 temp_wont_dirty=wont_dirty_next;
6428 // Merge in delay slot (will dirty)
6429 for(r=0;r<HOST_REGS;r++) {
6430 if(r!=EXCLUDE_REG) {
6431 if (1) { // !dops[i].likely) {
6432 // Will not dirty if likely branch is not taken
6433 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6434 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6435 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6436 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6437 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6438 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
6439 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6440 //if((regs[i].regmap[r]&63)==dops[i].rt1) temp_will_dirty|=1<<r;
6441 //if((regs[i].regmap[r]&63)==dops[i].rt2) temp_will_dirty|=1<<r;
6442 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_will_dirty|=1<<r;
6443 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_will_dirty|=1<<r;
6444 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
6445 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
6446 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
6447 }
6448 }
6449 }
6450 }
6451 // Merge in delay slot (wont dirty)
6452 for(r=0;r<HOST_REGS;r++) {
6453 if(r!=EXCLUDE_REG) {
6454 if((regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6455 if((regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6456 if((regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6457 if((regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6458 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6459 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) temp_wont_dirty|=1<<r;
6460 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) temp_wont_dirty|=1<<r;
6461 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) temp_wont_dirty|=1<<r;
6462 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) temp_wont_dirty|=1<<r;
6463 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
6464 }
6465 }
6466 // Deal with changed mappings
6467 if(i<iend) {
6468 for(r=0;r<HOST_REGS;r++) {
6469 if(r!=EXCLUDE_REG) {
6470 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
6471 temp_will_dirty&=~(1<<r);
6472 temp_wont_dirty&=~(1<<r);
6473 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6474 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6475 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6476 } else {
6477 temp_will_dirty|=1<<r;
6478 temp_wont_dirty|=1<<r;
6479 }
6480 }
6481 }
6482 }
6483 }
6484 if(wr) {
6485 will_dirty[i]=temp_will_dirty;
6486 wont_dirty[i]=temp_wont_dirty;
6487 clean_registers((ba[i]-start)>>2,i-1,0);
6488 }else{
6489 // Limit recursion. It can take an excessive amount
6490 // of time if there are a lot of nested loops.
6491 will_dirty[(ba[i]-start)>>2]=0;
6492 wont_dirty[(ba[i]-start)>>2]=-1;
6493 }
6494 }
6495 /*else*/ if(1)
6496 {
6497 if (dops[i].is_ujump)
6498 {
6499 // Unconditional branch
6500 will_dirty_i=0;
6501 wont_dirty_i=0;
6502 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6503 for(r=0;r<HOST_REGS;r++) {
6504 if(r!=EXCLUDE_REG) {
6505 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6506 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
6507 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6508 }
6509 if(branch_regs[i].regmap[r]>=0) {
6510 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6511 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
6512 }
6513 }
6514 }
6515 //}
6516 // Merge in delay slot
6517 for(r=0;r<HOST_REGS;r++) {
6518 if(r!=EXCLUDE_REG) {
6519 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6520 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6521 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6522 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6523 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6524 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6525 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6526 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6527 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6528 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6529 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6530 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6531 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6532 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6533 }
6534 }
6535 } else {
6536 // Conditional branch
6537 will_dirty_i=will_dirty_next;
6538 wont_dirty_i=wont_dirty_next;
6539 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
6540 for(r=0;r<HOST_REGS;r++) {
6541 if(r!=EXCLUDE_REG) {
6542 signed char target_reg=branch_regs[i].regmap[r];
6543 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
6544 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
6545 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
6546 }
6547 else if(target_reg>=0) {
6548 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6549 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
6550 }
6551 }
6552 }
6553 //}
6554 // Merge in delay slot
6555 for(r=0;r<HOST_REGS;r++) {
6556 if(r!=EXCLUDE_REG) {
6557 if (1) { // !dops[i].likely) {
6558 // Might not dirty if likely branch is not taken
6559 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6560 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6561 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6562 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6563 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6564 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6565 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6566 //if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6567 //if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6568 if((regs[i].regmap[r]&63)==dops[i+1].rt1) will_dirty_i|=1<<r;
6569 if((regs[i].regmap[r]&63)==dops[i+1].rt2) will_dirty_i|=1<<r;
6570 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6571 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6572 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6573 }
6574 }
6575 }
6576 }
6577 // Merge in delay slot (won't dirty)
6578 for(r=0;r<HOST_REGS;r++) {
6579 if(r!=EXCLUDE_REG) {
6580 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6581 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6582 if((regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6583 if((regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6584 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6585 if((branch_regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6586 if((branch_regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6587 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt1) wont_dirty_i|=1<<r;
6588 if((branch_regs[i].regmap[r]&63)==dops[i+1].rt2) wont_dirty_i|=1<<r;
6589 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6590 }
6591 }
6592 if(wr) {
6593 #ifndef DESTRUCTIVE_WRITEBACK
6594 branch_regs[i].dirty&=wont_dirty_i;
6595 #endif
6596 branch_regs[i].dirty|=will_dirty_i;
6597 }
6598 }
6599 }
6600 }
6601 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
6602 {
6603 // SYSCALL instruction (software interrupt)
6604 will_dirty_i=0;
6605 wont_dirty_i=0;
6606 }
6607 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
6608 {
6609 // ERET instruction (return from interrupt)
6610 will_dirty_i=0;
6611 wont_dirty_i=0;
6612 }
6613 will_dirty_next=will_dirty_i;
6614 wont_dirty_next=wont_dirty_i;
6615 for(r=0;r<HOST_REGS;r++) {
6616 if(r!=EXCLUDE_REG) {
6617 if((regs[i].regmap[r]&63)==dops[i].rt1) will_dirty_i|=1<<r;
6618 if((regs[i].regmap[r]&63)==dops[i].rt2) will_dirty_i|=1<<r;
6619 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6620 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6621 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6622 if((regs[i].regmap[r]&63)==dops[i].rt1) wont_dirty_i|=1<<r;
6623 if((regs[i].regmap[r]&63)==dops[i].rt2) wont_dirty_i|=1<<r;
6624 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6625 if(i>istart) {
6626 if (!dops[i].is_jump)
6627 {
6628 // Don't store a register immediately after writing it,
6629 // may prevent dual-issue.
6630 if((regs[i].regmap[r]&63)==dops[i-1].rt1) wont_dirty_i|=1<<r;
6631 if((regs[i].regmap[r]&63)==dops[i-1].rt2) wont_dirty_i|=1<<r;
6632 }
6633 }
6634 }
6635 }
6636 // Save it
6637 will_dirty[i]=will_dirty_i;
6638 wont_dirty[i]=wont_dirty_i;
6639 // Mark registers that won't be dirtied as not dirty
6640 if(wr) {
6641 regs[i].dirty|=will_dirty_i;
6642 #ifndef DESTRUCTIVE_WRITEBACK
6643 regs[i].dirty&=wont_dirty_i;
6644 if(dops[i].is_jump)
6645 {
6646 if (i < iend-1 && !dops[i].is_ujump) {
6647 for(r=0;r<HOST_REGS;r++) {
6648 if(r!=EXCLUDE_REG) {
6649 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
6650 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
6651 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6652 }
6653 }
6654 }
6655 }
6656 else
6657 {
6658 if(i<iend) {
6659 for(r=0;r<HOST_REGS;r++) {
6660 if(r!=EXCLUDE_REG) {
6661 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
6662 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
6663 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6664 }
6665 }
6666 }
6667 }
6668 #endif
6669 //}
6670 }
6671 // Deal with changed mappings
6672 temp_will_dirty=will_dirty_i;
6673 temp_wont_dirty=wont_dirty_i;
6674 for(r=0;r<HOST_REGS;r++) {
6675 if(r!=EXCLUDE_REG) {
6676 int nr;
6677 if(regs[i].regmap[r]==regmap_pre[i][r]) {
6678 if(wr) {
6679 #ifndef DESTRUCTIVE_WRITEBACK
6680 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6681 #endif
6682 regs[i].wasdirty|=will_dirty_i&(1<<r);
6683 }
6684 }
6685 else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
6686 // Register moved to a different register
6687 will_dirty_i&=~(1<<r);
6688 wont_dirty_i&=~(1<<r);
6689 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
6690 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
6691 if(wr) {
6692 #ifndef DESTRUCTIVE_WRITEBACK
6693 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6694 #endif
6695 regs[i].wasdirty|=will_dirty_i&(1<<r);
6696 }
6697 }
6698 else {
6699 will_dirty_i&=~(1<<r);
6700 wont_dirty_i&=~(1<<r);
6701 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6702 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6703 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6704 } else {
6705 wont_dirty_i|=1<<r;
6706 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
6707 }
6708 }
6709 }
6710 }
6711 }
6712}
6713
6714#ifdef DISASM
6715 /* disassembly */
6716void disassemble_inst(int i)
6717{
6718 if (dops[i].bt) printf("*"); else printf(" ");
6719 switch(dops[i].itype) {
6720 case UJUMP:
6721 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6722 case CJUMP:
6723 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;
6724 case SJUMP:
6725 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;
6726 case RJUMP:
6727 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6728 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6729 else
6730 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6731 break;
6732 case SPAN:
6733 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,ba[i]);break;
6734 case IMM16:
6735 if(dops[i].opcode==0xf) //LUI
6736 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6737 else
6738 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6739 break;
6740 case LOAD:
6741 case LOADLR:
6742 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6743 break;
6744 case STORE:
6745 case STORELR:
6746 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6747 break;
6748 case ALU:
6749 case SHIFT:
6750 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6751 break;
6752 case MULTDIV:
6753 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6754 break;
6755 case SHIFTIMM:
6756 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6757 break;
6758 case MOV:
6759 if((dops[i].opcode2&0x1d)==0x10)
6760 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6761 else if((dops[i].opcode2&0x1d)==0x11)
6762 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6763 else
6764 printf (" %x: %s\n",start+i*4,insn[i]);
6765 break;
6766 case COP0:
6767 if(dops[i].opcode2==0)
6768 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6769 else if(dops[i].opcode2==4)
6770 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6771 else printf (" %x: %s\n",start+i*4,insn[i]);
6772 break;
6773 case COP1:
6774 if(dops[i].opcode2<3)
6775 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6776 else if(dops[i].opcode2>3)
6777 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6778 else printf (" %x: %s\n",start+i*4,insn[i]);
6779 break;
6780 case COP2:
6781 if(dops[i].opcode2<3)
6782 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6783 else if(dops[i].opcode2>3)
6784 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6785 else printf (" %x: %s\n",start+i*4,insn[i]);
6786 break;
6787 case C1LS:
6788 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6789 break;
6790 case C2LS:
6791 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6792 break;
6793 case INTCALL:
6794 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6795 break;
6796 default:
6797 //printf (" %s %8x\n",insn[i],source[i]);
6798 printf (" %x: %s\n",start+i*4,insn[i]);
6799 }
6800}
6801#else
6802static void disassemble_inst(int i) {}
6803#endif // DISASM
6804
6805#define DRC_TEST_VAL 0x74657374
6806
6807static void new_dynarec_test(void)
6808{
6809 int (*testfunc)(void);
6810 void *beginning;
6811 int ret[2];
6812 size_t i;
6813
6814 // check structure linkage
6815 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6816 {
6817 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6818 }
6819
6820 SysPrintf("testing if we can run recompiled code @%p...\n", out);
6821 ((volatile u_int *)out)[0]++; // make cache dirty
6822
6823 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6824 out = ndrc->translation_cache;
6825 beginning = start_block();
6826 emit_movimm(DRC_TEST_VAL + i, 0); // test
6827 emit_ret();
6828 literal_pool(0);
6829 end_block(beginning);
6830 testfunc = beginning;
6831 ret[i] = testfunc();
6832 }
6833
6834 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6835 SysPrintf("test passed.\n");
6836 else
6837 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6838 out = ndrc->translation_cache;
6839}
6840
6841// clear the state completely, instead of just marking
6842// things invalid like invalidate_all_pages() does
6843void new_dynarec_clear_full(void)
6844{
6845 int n;
6846 out = ndrc->translation_cache;
6847 memset(invalid_code,1,sizeof(invalid_code));
6848 memset(hash_table,0xff,sizeof(hash_table));
6849 memset(mini_ht,-1,sizeof(mini_ht));
6850 memset(restore_candidate,0,sizeof(restore_candidate));
6851 memset(shadow,0,sizeof(shadow));
6852 copy=shadow;
6853 expirep=16384; // Expiry pointer, +2 blocks
6854 pending_exception=0;
6855 literalcount=0;
6856 stop_after_jal=0;
6857 inv_code_start=inv_code_end=~0;
6858 hack_addr=0;
6859 f1_hack=0;
6860 // TLB
6861 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6862 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6863 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6864
6865 cycle_multiplier_old = cycle_multiplier;
6866 new_dynarec_hacks_old = new_dynarec_hacks;
6867}
6868
6869void new_dynarec_init(void)
6870{
6871 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6872
6873#ifdef _3DS
6874 check_rosalina();
6875#endif
6876#ifdef BASE_ADDR_DYNAMIC
6877 #ifdef VITA
6878 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6879 if (sceBlock <= 0)
6880 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6881 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6882 if (ret < 0)
6883 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6884 sceKernelOpenVMDomain();
6885 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6886 #elif defined(_MSC_VER)
6887 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6888 PAGE_EXECUTE_READWRITE);
6889 #else
6890 uintptr_t desired_addr = 0;
6891 #ifdef __ELF__
6892 extern char _end;
6893 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6894 #endif
6895 ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
6896 PROT_READ | PROT_WRITE | PROT_EXEC,
6897 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
6898 if (ndrc == MAP_FAILED) {
6899 SysPrintf("mmap() failed: %s\n", strerror(errno));
6900 abort();
6901 }
6902 #endif
6903#else
6904 #ifndef NO_WRITE_EXEC
6905 // not all systems allow execute in data segment by default
6906 // size must be 4K aligned for 3DS?
6907 if (mprotect(ndrc, sizeof(*ndrc),
6908 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6909 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6910 #endif
6911#endif
6912 out = ndrc->translation_cache;
6913 cycle_multiplier=200;
6914 new_dynarec_clear_full();
6915#ifdef HOST_IMM8
6916 // Copy this into local area so we don't have to put it in every literal pool
6917 invc_ptr=invalid_code;
6918#endif
6919 arch_init();
6920 new_dynarec_test();
6921 ram_offset=(uintptr_t)rdram-0x80000000;
6922 if (ram_offset!=0)
6923 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6924}
6925
6926void new_dynarec_cleanup(void)
6927{
6928 int n;
6929#ifdef BASE_ADDR_DYNAMIC
6930 #ifdef VITA
6931 // sceBlock is managed by retroarch's bootstrap code
6932 //sceKernelFreeMemBlock(sceBlock);
6933 //sceBlock = -1;
6934 #else
6935 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6936 SysPrintf("munmap() failed\n");
6937 #endif
6938#endif
6939 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6940 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6941 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6942 #ifdef ROM_COPY
6943 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
6944 #endif
6945}
6946
6947static u_int *get_source_start(u_int addr, u_int *limit)
6948{
6949 if (addr < 0x00200000 ||
6950 (0xa0000000 <= addr && addr < 0xa0200000))
6951 {
6952 // used for BIOS calls mostly?
6953 *limit = (addr&0xa0000000)|0x00200000;
6954 return (u_int *)(rdram + (addr&0x1fffff));
6955 }
6956 else if (!Config.HLE && (
6957 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6958 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6959 {
6960 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6961 // but timings in PCSX are too tied to the interpreter's BIAS
6962 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6963 cycle_multiplier_active = 200;
6964
6965 *limit = (addr & 0xfff00000) | 0x80000;
6966 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6967 }
6968 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6969 *limit = (addr & 0x80600000) + 0x00200000;
6970 return (u_int *)(rdram + (addr&0x1fffff));
6971 }
6972 return NULL;
6973}
6974
6975static u_int scan_for_ret(u_int addr)
6976{
6977 u_int limit = 0;
6978 u_int *mem;
6979
6980 mem = get_source_start(addr, &limit);
6981 if (mem == NULL)
6982 return addr;
6983
6984 if (limit > addr + 0x1000)
6985 limit = addr + 0x1000;
6986 for (; addr < limit; addr += 4, mem++) {
6987 if (*mem == 0x03e00008) // jr $ra
6988 return addr + 8;
6989 }
6990 return addr;
6991}
6992
6993struct savestate_block {
6994 uint32_t addr;
6995 uint32_t regflags;
6996};
6997
6998static int addr_cmp(const void *p1_, const void *p2_)
6999{
7000 const struct savestate_block *p1 = p1_, *p2 = p2_;
7001 return p1->addr - p2->addr;
7002}
7003
7004int new_dynarec_save_blocks(void *save, int size)
7005{
7006 struct savestate_block *blocks = save;
7007 int maxcount = size / sizeof(blocks[0]);
7008 struct savestate_block tmp_blocks[1024];
7009 struct ll_entry *head;
7010 int p, s, d, o, bcnt;
7011 u_int addr;
7012
7013 o = 0;
7014 for (p = 0; p < ARRAY_SIZE(jump_in); p++) {
7015 bcnt = 0;
7016 for (head = jump_in[p]; head != NULL; head = head->next) {
7017 tmp_blocks[bcnt].addr = head->vaddr;
7018 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
7019 bcnt++;
7020 }
7021 if (bcnt < 1)
7022 continue;
7023 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
7024
7025 addr = tmp_blocks[0].addr;
7026 for (s = d = 0; s < bcnt; s++) {
7027 if (tmp_blocks[s].addr < addr)
7028 continue;
7029 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
7030 tmp_blocks[d++] = tmp_blocks[s];
7031 addr = scan_for_ret(tmp_blocks[s].addr);
7032 }
7033
7034 if (o + d > maxcount)
7035 d = maxcount - o;
7036 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
7037 o += d;
7038 }
7039
7040 return o * sizeof(blocks[0]);
7041}
7042
7043void new_dynarec_load_blocks(const void *save, int size)
7044{
7045 const struct savestate_block *blocks = save;
7046 int count = size / sizeof(blocks[0]);
7047 u_int regs_save[32];
7048 uint32_t f;
7049 int i, b;
7050
7051 get_addr(psxRegs.pc);
7052
7053 // change GPRs for speculation to at least partially work..
7054 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
7055 for (i = 1; i < 32; i++)
7056 psxRegs.GPR.r[i] = 0x80000000;
7057
7058 for (b = 0; b < count; b++) {
7059 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
7060 if (f & 1)
7061 psxRegs.GPR.r[i] = 0x1f800000;
7062 }
7063
7064 get_addr(blocks[b].addr);
7065
7066 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
7067 if (f & 1)
7068 psxRegs.GPR.r[i] = 0x80000000;
7069 }
7070 }
7071
7072 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
7073}
7074
7075static int apply_hacks(void)
7076{
7077 int i;
7078 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
7079 return 0;
7080 /* special hack(s) */
7081 for (i = 0; i < slen - 4; i++)
7082 {
7083 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
7084 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
7085 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
7086 && imm[i+3] == 0x6e40 && dops[i+3].rs1 == 2)
7087 {
7088 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
7089 dops[i + 3].itype = NOP;
7090 }
7091 }
7092 i = slen;
7093 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
7094 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
7095 && dops[i-7].itype == STORE)
7096 {
7097 i = i-8;
7098 if (dops[i].itype == IMM16)
7099 i--;
7100 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
7101 if (dops[i].itype == STORELR && dops[i].rs1 == 6
7102 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
7103 {
7104 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
7105 f1_hack = 1;
7106 return 1;
7107 }
7108 }
7109 return 0;
7110}
7111
7112int new_recompile_block(u_int addr)
7113{
7114 u_int pagelimit = 0;
7115 u_int state_rflags = 0;
7116 int i;
7117
7118 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
7119 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
7120 //if(debug)
7121 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
7122
7123 // this is just for speculation
7124 for (i = 1; i < 32; i++) {
7125 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
7126 state_rflags |= 1 << i;
7127 }
7128
7129 start = (u_int)addr&~3;
7130 //assert(((u_int)addr&1)==0); // start-in-delay-slot flag
7131 new_dynarec_did_compile=1;
7132 if (Config.HLE && start == 0x80001000) // hlecall
7133 {
7134 // XXX: is this enough? Maybe check hleSoftCall?
7135 void *beginning=start_block();
7136 u_int page=get_page(start);
7137
7138 invalid_code[start>>12]=0;
7139 emit_movimm(start,0);
7140 emit_writeword(0,&pcaddr);
7141 emit_far_jump(new_dyna_leave);
7142 literal_pool(0);
7143 end_block(beginning);
7144 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
7145 return 0;
7146 }
7147 else if (f1_hack && hack_addr == 0) {
7148 void *beginning = start_block();
7149 u_int page = get_page(start);
7150 emit_movimm(start, 0);
7151 emit_writeword(0, &hack_addr);
7152 emit_readword(&psxRegs.GPR.n.sp, 0);
7153 emit_readptr(&mem_rtab, 1);
7154 emit_shrimm(0, 12, 2);
7155 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
7156 emit_addimm(0, 0x18, 0);
7157 emit_adds_ptr(1, 1, 1);
7158 emit_ldr_dualindexed(1, 0, 0);
7159 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
7160 emit_far_call(get_addr_ht);
7161 emit_jmpreg(0); // jr k0
7162 literal_pool(0);
7163 end_block(beginning);
7164
7165 ll_add_flags(jump_in + page, start, state_rflags, beginning);
7166 SysPrintf("F1 hack to %08x\n", start);
7167 return 0;
7168 }
7169
7170 cycle_multiplier_active = cycle_multiplier_override && cycle_multiplier == CYCLE_MULT_DEFAULT
7171 ? cycle_multiplier_override : cycle_multiplier;
7172
7173 source = get_source_start(start, &pagelimit);
7174 if (source == NULL) {
7175 SysPrintf("Compile at bogus memory address: %08x\n", addr);
7176 abort();
7177 }
7178
7179 /* Pass 1: disassemble */
7180 /* Pass 2: register dependencies, branch targets */
7181 /* Pass 3: register allocation */
7182 /* Pass 4: branch dependencies */
7183 /* Pass 5: pre-alloc */
7184 /* Pass 6: optimize clean/dirty state */
7185 /* Pass 7: flag 32-bit registers */
7186 /* Pass 8: assembly */
7187 /* Pass 9: linker */
7188 /* Pass 10: garbage collection / free memory */
7189
7190 int j;
7191 int done=0;
7192 unsigned int type,op,op2;
7193
7194 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
7195
7196 /* Pass 1 disassembly */
7197
7198 for (i = 0; !done; i++)
7199 {
7200 memset(&dops[i], 0, sizeof(dops[i]));
7201 op2=0;
7202 minimum_free_regs[i]=0;
7203 dops[i].opcode=op=source[i]>>26;
7204 switch(op)
7205 {
7206 case 0x00: strcpy(insn[i],"special"); type=NI;
7207 op2=source[i]&0x3f;
7208 switch(op2)
7209 {
7210 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
7211 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
7212 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
7213 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
7214 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
7215 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
7216 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
7217 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
7218 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
7219 case 0x0D: strcpy(insn[i],"BREAK"); type=SYSCALL; break;
7220 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
7221 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
7222 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
7223 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
7224 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
7225 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
7226 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
7227 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
7228 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
7229 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
7230 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
7231 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
7232 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
7233 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
7234 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
7235 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
7236 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
7237 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
7238 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
7239 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
7240 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
7241 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
7242 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
7243 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
7244 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
7245#if 0
7246 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
7247 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
7248 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
7249 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
7250 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
7251 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
7252 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
7253 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
7254 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
7255 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
7256 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
7257 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
7258 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
7259 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
7260 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
7261 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
7262 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
7263#endif
7264 }
7265 break;
7266 case 0x01: strcpy(insn[i],"regimm"); type=NI;
7267 op2=(source[i]>>16)&0x1f;
7268 switch(op2)
7269 {
7270 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
7271 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
7272 //case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
7273 //case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
7274 //case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
7275 //case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
7276 //case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
7277 //case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
7278 //case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
7279 //case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
7280 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
7281 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
7282 //case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
7283 //case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
7284 }
7285 break;
7286 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
7287 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
7288 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
7289 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
7290 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
7291 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
7292 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
7293 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
7294 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
7295 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
7296 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
7297 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
7298 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
7299 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
7300 case 0x10: strcpy(insn[i],"cop0"); type=NI;
7301 op2=(source[i]>>21)&0x1f;
7302 switch(op2)
7303 {
7304 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
7305 case 0x02: strcpy(insn[i],"CFC0"); type=COP0; break;
7306 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
7307 case 0x06: strcpy(insn[i],"CTC0"); type=COP0; break;
7308 case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
7309 }
7310 break;
7311 case 0x11: strcpy(insn[i],"cop1"); type=COP1;
7312 op2=(source[i]>>21)&0x1f;
7313 break;
7314#if 0
7315 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
7316 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
7317 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
7318 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
7319 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
7320 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
7321 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
7322 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
7323#endif
7324 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
7325 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
7326 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
7327 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
7328 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
7329 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
7330 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
7331#if 0
7332 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
7333#endif
7334 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
7335 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
7336 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
7337 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
7338#if 0
7339 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
7340 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
7341#endif
7342 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
7343 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
7344 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
7345 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
7346#if 0
7347 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
7348 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
7349 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
7350#endif
7351 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
7352 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
7353#if 0
7354 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
7355 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
7356 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
7357#endif
7358 case 0x12: strcpy(insn[i],"COP2"); type=NI;
7359 op2=(source[i]>>21)&0x1f;
7360 //if (op2 & 0x10)
7361 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
7362 if (gte_handlers[source[i]&0x3f]!=NULL) {
7363 if (gte_regnames[source[i]&0x3f]!=NULL)
7364 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
7365 else
7366 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
7367 type=C2OP;
7368 }
7369 }
7370 else switch(op2)
7371 {
7372 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
7373 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
7374 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
7375 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
7376 }
7377 break;
7378 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
7379 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
7380 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
7381 default: strcpy(insn[i],"???"); type=NI;
7382 SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
7383 break;
7384 }
7385 dops[i].itype=type;
7386 dops[i].opcode2=op2;
7387 /* Get registers/immediates */
7388 dops[i].lt1=0;
7389 gte_rs[i]=gte_rt[i]=0;
7390 switch(type) {
7391 case LOAD:
7392 dops[i].rs1=(source[i]>>21)&0x1f;
7393 dops[i].rs2=0;
7394 dops[i].rt1=(source[i]>>16)&0x1f;
7395 dops[i].rt2=0;
7396 imm[i]=(short)source[i];
7397 break;
7398 case STORE:
7399 case STORELR:
7400 dops[i].rs1=(source[i]>>21)&0x1f;
7401 dops[i].rs2=(source[i]>>16)&0x1f;
7402 dops[i].rt1=0;
7403 dops[i].rt2=0;
7404 imm[i]=(short)source[i];
7405 break;
7406 case LOADLR:
7407 // LWL/LWR only load part of the register,
7408 // therefore the target register must be treated as a source too
7409 dops[i].rs1=(source[i]>>21)&0x1f;
7410 dops[i].rs2=(source[i]>>16)&0x1f;
7411 dops[i].rt1=(source[i]>>16)&0x1f;
7412 dops[i].rt2=0;
7413 imm[i]=(short)source[i];
7414 break;
7415 case IMM16:
7416 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
7417 else dops[i].rs1=(source[i]>>21)&0x1f;
7418 dops[i].rs2=0;
7419 dops[i].rt1=(source[i]>>16)&0x1f;
7420 dops[i].rt2=0;
7421 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
7422 imm[i]=(unsigned short)source[i];
7423 }else{
7424 imm[i]=(short)source[i];
7425 }
7426 break;
7427 case UJUMP:
7428 dops[i].rs1=0;
7429 dops[i].rs2=0;
7430 dops[i].rt1=0;
7431 dops[i].rt2=0;
7432 // The JAL instruction writes to r31.
7433 if (op&1) {
7434 dops[i].rt1=31;
7435 }
7436 dops[i].rs2=CCREG;
7437 break;
7438 case RJUMP:
7439 dops[i].rs1=(source[i]>>21)&0x1f;
7440 dops[i].rs2=0;
7441 dops[i].rt1=0;
7442 dops[i].rt2=0;
7443 // The JALR instruction writes to rd.
7444 if (op2&1) {
7445 dops[i].rt1=(source[i]>>11)&0x1f;
7446 }
7447 dops[i].rs2=CCREG;
7448 break;
7449 case CJUMP:
7450 dops[i].rs1=(source[i]>>21)&0x1f;
7451 dops[i].rs2=(source[i]>>16)&0x1f;
7452 dops[i].rt1=0;
7453 dops[i].rt2=0;
7454 if(op&2) { // BGTZ/BLEZ
7455 dops[i].rs2=0;
7456 }
7457 break;
7458 case SJUMP:
7459 dops[i].rs1=(source[i]>>21)&0x1f;
7460 dops[i].rs2=CCREG;
7461 dops[i].rt1=0;
7462 dops[i].rt2=0;
7463 if(op2&0x10) { // BxxAL
7464 dops[i].rt1=31;
7465 // NOTE: If the branch is not taken, r31 is still overwritten
7466 }
7467 break;
7468 case ALU:
7469 dops[i].rs1=(source[i]>>21)&0x1f; // source
7470 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
7471 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7472 dops[i].rt2=0;
7473 break;
7474 case MULTDIV:
7475 dops[i].rs1=(source[i]>>21)&0x1f; // source
7476 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
7477 dops[i].rt1=HIREG;
7478 dops[i].rt2=LOREG;
7479 break;
7480 case MOV:
7481 dops[i].rs1=0;
7482 dops[i].rs2=0;
7483 dops[i].rt1=0;
7484 dops[i].rt2=0;
7485 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
7486 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
7487 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
7488 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
7489 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
7490 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
7491 break;
7492 case SHIFT:
7493 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
7494 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
7495 dops[i].rt1=(source[i]>>11)&0x1f; // destination
7496 dops[i].rt2=0;
7497 break;
7498 case SHIFTIMM:
7499 dops[i].rs1=(source[i]>>16)&0x1f;
7500 dops[i].rs2=0;
7501 dops[i].rt1=(source[i]>>11)&0x1f;
7502 dops[i].rt2=0;
7503 imm[i]=(source[i]>>6)&0x1f;
7504 // DSxx32 instructions
7505 if(op2>=0x3c) imm[i]|=0x20;
7506 break;
7507 case COP0:
7508 dops[i].rs1=0;
7509 dops[i].rs2=0;
7510 dops[i].rt1=0;
7511 dops[i].rt2=0;
7512 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
7513 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
7514 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
7515 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
7516 break;
7517 case COP1:
7518 dops[i].rs1=0;
7519 dops[i].rs2=0;
7520 dops[i].rt1=0;
7521 dops[i].rt2=0;
7522 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
7523 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
7524 dops[i].rs2=CSREG;
7525 break;
7526 case COP2:
7527 dops[i].rs1=0;
7528 dops[i].rs2=0;
7529 dops[i].rt1=0;
7530 dops[i].rt2=0;
7531 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
7532 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
7533 dops[i].rs2=CSREG;
7534 int gr=(source[i]>>11)&0x1F;
7535 switch(op2)
7536 {
7537 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
7538 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
7539 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
7540 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
7541 }
7542 break;
7543 case C1LS:
7544 dops[i].rs1=(source[i]>>21)&0x1F;
7545 dops[i].rs2=CSREG;
7546 dops[i].rt1=0;
7547 dops[i].rt2=0;
7548 imm[i]=(short)source[i];
7549 break;
7550 case C2LS:
7551 dops[i].rs1=(source[i]>>21)&0x1F;
7552 dops[i].rs2=0;
7553 dops[i].rt1=0;
7554 dops[i].rt2=0;
7555 imm[i]=(short)source[i];
7556 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
7557 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
7558 break;
7559 case C2OP:
7560 dops[i].rs1=0;
7561 dops[i].rs2=0;
7562 dops[i].rt1=0;
7563 dops[i].rt2=0;
7564 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
7565 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
7566 gte_rt[i]|=1ll<<63; // every op changes flags
7567 if((source[i]&0x3f)==GTE_MVMVA) {
7568 int v = (source[i] >> 15) & 3;
7569 gte_rs[i]&=~0xe3fll;
7570 if(v==3) gte_rs[i]|=0xe00ll;
7571 else gte_rs[i]|=3ll<<(v*2);
7572 }
7573 break;
7574 case SYSCALL:
7575 case HLECALL:
7576 case INTCALL:
7577 dops[i].rs1=CCREG;
7578 dops[i].rs2=0;
7579 dops[i].rt1=0;
7580 dops[i].rt2=0;
7581 break;
7582 default:
7583 dops[i].rs1=0;
7584 dops[i].rs2=0;
7585 dops[i].rt1=0;
7586 dops[i].rt2=0;
7587 }
7588 /* Calculate branch target addresses */
7589 if(type==UJUMP)
7590 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7591 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
7592 ba[i]=start+i*4+8; // Ignore never taken branch
7593 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
7594 ba[i]=start+i*4+8; // Ignore never taken branch
7595 else if(type==CJUMP||type==SJUMP)
7596 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7597 else ba[i]=-1;
7598
7599 /* simplify always (not)taken branches */
7600 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
7601 dops[i].rs1 = dops[i].rs2 = 0;
7602 if (!(op & 1)) {
7603 dops[i].itype = type = UJUMP;
7604 dops[i].rs2 = CCREG;
7605 }
7606 }
7607 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
7608 dops[i].itype = type = UJUMP;
7609
7610 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
7611 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
7612 dops[i].is_load = (dops[i].itype == LOAD || dops[i].itype == LOADLR || op == 0x32); // LWC2
7613 dops[i].is_store = (dops[i].itype == STORE || dops[i].itype == STORELR || op == 0x3a); // SWC2
7614
7615 /* messy cases to just pass over to the interpreter */
7616 if (i > 0 && dops[i-1].is_jump) {
7617 int do_in_intrp=0;
7618 // branch in delay slot?
7619 if (dops[i].is_jump) {
7620 // don't handle first branch and call interpreter if it's hit
7621 SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
7622 do_in_intrp=1;
7623 }
7624 // basic load delay detection
7625 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
7626 int t=(ba[i-1]-start)/4;
7627 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) {
7628 // jump target wants DS result - potential load delay effect
7629 SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
7630 do_in_intrp=1;
7631 dops[t+1].bt=1; // expected return from interpreter
7632 }
7633 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&&
7634 !(i>=3&&dops[i-3].is_jump)) {
7635 // v0 overwrite like this is a sign of trouble, bail out
7636 SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
7637 do_in_intrp=1;
7638 }
7639 }
7640 if (do_in_intrp) {
7641 memset(&dops[i-1], 0, sizeof(dops[i-1]));
7642 dops[i-1].itype = INTCALL;
7643 dops[i-1].rs1 = CCREG;
7644 ba[i-1] = -1;
7645 done = 2;
7646 i--; // don't compile the DS
7647 }
7648 }
7649
7650 /* Is this the end of the block? */
7651 if (i > 0 && dops[i-1].is_ujump) {
7652 if(dops[i-1].rt1==0) { // Continue past subroutine call (JAL)
7653 done=2;
7654 }
7655 else {
7656 if(stop_after_jal) done=1;
7657 // Stop on BREAK
7658 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7659 }
7660 // Don't recompile stuff that's already compiled
7661 if(check_addr(start+i*4+4)) done=1;
7662 // Don't get too close to the limit
7663 if(i>MAXBLOCK/2) done=1;
7664 }
7665 if (dops[i].itype == SYSCALL || dops[i].itype == HLECALL || dops[i].itype == INTCALL)
7666 done = stop_after_jal ? 1 : 2;
7667 if (done == 2) {
7668 // Does the block continue due to a branch?
7669 for(j=i-1;j>=0;j--)
7670 {
7671 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7672 if(ba[j]==start+i*4+4) done=j=0;
7673 if(ba[j]==start+i*4+8) done=j=0;
7674 }
7675 }
7676 //assert(i<MAXBLOCK-1);
7677 if(start+i*4==pagelimit-4) done=1;
7678 assert(start+i*4<pagelimit);
7679 if (i==MAXBLOCK-1) done=1;
7680 // Stop if we're compiling junk
7681 if(dops[i].itype==NI&&dops[i].opcode==0x11) {
7682 done=stop_after_jal=1;
7683 SysPrintf("Disabled speculative precompilation\n");
7684 }
7685 }
7686 slen=i;
7687 if (dops[i-1].is_jump) {
7688 if(start+i*4==pagelimit) {
7689 dops[i-1].itype=SPAN;
7690 }
7691 }
7692 assert(slen>0);
7693
7694 int clear_hack_addr = apply_hacks();
7695
7696 /* Pass 2 - Register dependencies and branch targets */
7697
7698 unneeded_registers(0,slen-1,0);
7699
7700 /* Pass 3 - Register allocation */
7701
7702 struct regstat current; // Current register allocations/status
7703 current.dirty=0;
7704 current.u=unneeded_reg[0];
7705 clear_all_regs(current.regmap);
7706 alloc_reg(&current,0,CCREG);
7707 dirty_reg(&current,CCREG);
7708 current.isconst=0;
7709 current.wasconst=0;
7710 current.waswritten=0;
7711 int ds=0;
7712 int cc=0;
7713 int hr=-1;
7714
7715 if((u_int)addr&1) {
7716 // First instruction is delay slot
7717 cc=-1;
7718 dops[1].bt=1;
7719 ds=1;
7720 unneeded_reg[0]=1;
7721 current.regmap[HOST_BTREG]=BTREG;
7722 }
7723
7724 for(i=0;i<slen;i++)
7725 {
7726 if(dops[i].bt)
7727 {
7728 int hr;
7729 for(hr=0;hr<HOST_REGS;hr++)
7730 {
7731 // Is this really necessary?
7732 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7733 }
7734 current.isconst=0;
7735 current.waswritten=0;
7736 }
7737
7738 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7739 regs[i].wasconst=current.isconst;
7740 regs[i].wasdirty=current.dirty;
7741 regs[i].loadedconst=0;
7742 if (!dops[i].is_jump) {
7743 if(i+1<slen) {
7744 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7745 current.u|=1;
7746 } else {
7747 current.u=1;
7748 }
7749 } else {
7750 if(i+1<slen) {
7751 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7752 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7753 current.u|=1;
7754 } else {
7755 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7756 abort();
7757 }
7758 }
7759 dops[i].is_ds=ds;
7760 if(ds) {
7761 ds=0; // Skip delay slot, already allocated as part of branch
7762 // ...but we need to alloc it in case something jumps here
7763 if(i+1<slen) {
7764 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7765 }else{
7766 current.u=branch_unneeded_reg[i-1];
7767 }
7768 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7769 current.u|=1;
7770 struct regstat temp;
7771 memcpy(&temp,&current,sizeof(current));
7772 temp.wasdirty=temp.dirty;
7773 // TODO: Take into account unconditional branches, as below
7774 delayslot_alloc(&temp,i);
7775 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7776 regs[i].wasdirty=temp.wasdirty;
7777 regs[i].dirty=temp.dirty;
7778 regs[i].isconst=0;
7779 regs[i].wasconst=0;
7780 current.isconst=0;
7781 // Create entry (branch target) regmap
7782 for(hr=0;hr<HOST_REGS;hr++)
7783 {
7784 int r=temp.regmap[hr];
7785 if(r>=0) {
7786 if(r!=regmap_pre[i][hr]) {
7787 regs[i].regmap_entry[hr]=-1;
7788 }
7789 else
7790 {
7791 assert(r < 64);
7792 if((current.u>>r)&1) {
7793 regs[i].regmap_entry[hr]=-1;
7794 regs[i].regmap[hr]=-1;
7795 //Don't clear regs in the delay slot as the branch might need them
7796 //current.regmap[hr]=-1;
7797 }else
7798 regs[i].regmap_entry[hr]=r;
7799 }
7800 } else {
7801 // First instruction expects CCREG to be allocated
7802 if(i==0&&hr==HOST_CCREG)
7803 regs[i].regmap_entry[hr]=CCREG;
7804 else
7805 regs[i].regmap_entry[hr]=-1;
7806 }
7807 }
7808 }
7809 else { // Not delay slot
7810 switch(dops[i].itype) {
7811 case UJUMP:
7812 //current.isconst=0; // DEBUG
7813 //current.wasconst=0; // DEBUG
7814 //regs[i].wasconst=0; // DEBUG
7815 clear_const(&current,dops[i].rt1);
7816 alloc_cc(&current,i);
7817 dirty_reg(&current,CCREG);
7818 if (dops[i].rt1==31) {
7819 alloc_reg(&current,i,31);
7820 dirty_reg(&current,31);
7821 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7822 //assert(dops[i+1].rt1!=dops[i].rt1);
7823 #ifdef REG_PREFETCH
7824 alloc_reg(&current,i,PTEMP);
7825 #endif
7826 }
7827 dops[i].ooo=1;
7828 delayslot_alloc(&current,i+1);
7829 //current.isconst=0; // DEBUG
7830 ds=1;
7831 //printf("i=%d, isconst=%x\n",i,current.isconst);
7832 break;
7833 case RJUMP:
7834 //current.isconst=0;
7835 //current.wasconst=0;
7836 //regs[i].wasconst=0;
7837 clear_const(&current,dops[i].rs1);
7838 clear_const(&current,dops[i].rt1);
7839 alloc_cc(&current,i);
7840 dirty_reg(&current,CCREG);
7841 if (!ds_writes_rjump_rs(i)) {
7842 alloc_reg(&current,i,dops[i].rs1);
7843 if (dops[i].rt1!=0) {
7844 alloc_reg(&current,i,dops[i].rt1);
7845 dirty_reg(&current,dops[i].rt1);
7846 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7847 assert(dops[i+1].rt1!=dops[i].rt1);
7848 #ifdef REG_PREFETCH
7849 alloc_reg(&current,i,PTEMP);
7850 #endif
7851 }
7852 #ifdef USE_MINI_HT
7853 if(dops[i].rs1==31) { // JALR
7854 alloc_reg(&current,i,RHASH);
7855 alloc_reg(&current,i,RHTBL);
7856 }
7857 #endif
7858 delayslot_alloc(&current,i+1);
7859 } else {
7860 // The delay slot overwrites our source register,
7861 // allocate a temporary register to hold the old value.
7862 current.isconst=0;
7863 current.wasconst=0;
7864 regs[i].wasconst=0;
7865 delayslot_alloc(&current,i+1);
7866 current.isconst=0;
7867 alloc_reg(&current,i,RTEMP);
7868 }
7869 //current.isconst=0; // DEBUG
7870 dops[i].ooo=1;
7871 ds=1;
7872 break;
7873 case CJUMP:
7874 //current.isconst=0;
7875 //current.wasconst=0;
7876 //regs[i].wasconst=0;
7877 clear_const(&current,dops[i].rs1);
7878 clear_const(&current,dops[i].rs2);
7879 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7880 {
7881 alloc_cc(&current,i);
7882 dirty_reg(&current,CCREG);
7883 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7884 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7885 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7886 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7887 // The delay slot overwrites one of our conditions.
7888 // Allocate the branch condition registers instead.
7889 current.isconst=0;
7890 current.wasconst=0;
7891 regs[i].wasconst=0;
7892 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7893 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7894 }
7895 else
7896 {
7897 dops[i].ooo=1;
7898 delayslot_alloc(&current,i+1);
7899 }
7900 }
7901 else
7902 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7903 {
7904 alloc_cc(&current,i);
7905 dirty_reg(&current,CCREG);
7906 alloc_reg(&current,i,dops[i].rs1);
7907 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7908 // The delay slot overwrites one of our conditions.
7909 // Allocate the branch condition registers instead.
7910 current.isconst=0;
7911 current.wasconst=0;
7912 regs[i].wasconst=0;
7913 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7914 }
7915 else
7916 {
7917 dops[i].ooo=1;
7918 delayslot_alloc(&current,i+1);
7919 }
7920 }
7921 else
7922 // Don't alloc the delay slot yet because we might not execute it
7923 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7924 {
7925 current.isconst=0;
7926 current.wasconst=0;
7927 regs[i].wasconst=0;
7928 alloc_cc(&current,i);
7929 dirty_reg(&current,CCREG);
7930 alloc_reg(&current,i,dops[i].rs1);
7931 alloc_reg(&current,i,dops[i].rs2);
7932 }
7933 else
7934 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7935 {
7936 current.isconst=0;
7937 current.wasconst=0;
7938 regs[i].wasconst=0;
7939 alloc_cc(&current,i);
7940 dirty_reg(&current,CCREG);
7941 alloc_reg(&current,i,dops[i].rs1);
7942 }
7943 ds=1;
7944 //current.isconst=0;
7945 break;
7946 case SJUMP:
7947 //current.isconst=0;
7948 //current.wasconst=0;
7949 //regs[i].wasconst=0;
7950 clear_const(&current,dops[i].rs1);
7951 clear_const(&current,dops[i].rt1);
7952 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7953 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7954 {
7955 alloc_cc(&current,i);
7956 dirty_reg(&current,CCREG);
7957 alloc_reg(&current,i,dops[i].rs1);
7958 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7959 alloc_reg(&current,i,31);
7960 dirty_reg(&current,31);
7961 //#ifdef REG_PREFETCH
7962 //alloc_reg(&current,i,PTEMP);
7963 //#endif
7964 }
7965 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.
7966 ||(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
7967 // Allocate the branch condition registers instead.
7968 current.isconst=0;
7969 current.wasconst=0;
7970 regs[i].wasconst=0;
7971 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7972 }
7973 else
7974 {
7975 dops[i].ooo=1;
7976 delayslot_alloc(&current,i+1);
7977 }
7978 }
7979 else
7980 // Don't alloc the delay slot yet because we might not execute it
7981 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7982 {
7983 current.isconst=0;
7984 current.wasconst=0;
7985 regs[i].wasconst=0;
7986 alloc_cc(&current,i);
7987 dirty_reg(&current,CCREG);
7988 alloc_reg(&current,i,dops[i].rs1);
7989 }
7990 ds=1;
7991 //current.isconst=0;
7992 break;
7993 case IMM16:
7994 imm16_alloc(&current,i);
7995 break;
7996 case LOAD:
7997 case LOADLR:
7998 load_alloc(&current,i);
7999 break;
8000 case STORE:
8001 case STORELR:
8002 store_alloc(&current,i);
8003 break;
8004 case ALU:
8005 alu_alloc(&current,i);
8006 break;
8007 case SHIFT:
8008 shift_alloc(&current,i);
8009 break;
8010 case MULTDIV:
8011 multdiv_alloc(&current,i);
8012 break;
8013 case SHIFTIMM:
8014 shiftimm_alloc(&current,i);
8015 break;
8016 case MOV:
8017 mov_alloc(&current,i);
8018 break;
8019 case COP0:
8020 cop0_alloc(&current,i);
8021 break;
8022 case COP1:
8023 break;
8024 case COP2:
8025 cop2_alloc(&current,i);
8026 break;
8027 case C1LS:
8028 c1ls_alloc(&current,i);
8029 break;
8030 case C2LS:
8031 c2ls_alloc(&current,i);
8032 break;
8033 case C2OP:
8034 c2op_alloc(&current,i);
8035 break;
8036 case SYSCALL:
8037 case HLECALL:
8038 case INTCALL:
8039 syscall_alloc(&current,i);
8040 break;
8041 case SPAN:
8042 pagespan_alloc(&current,i);
8043 break;
8044 }
8045
8046 // Create entry (branch target) regmap
8047 for(hr=0;hr<HOST_REGS;hr++)
8048 {
8049 int r,or;
8050 r=current.regmap[hr];
8051 if(r>=0) {
8052 if(r!=regmap_pre[i][hr]) {
8053 // TODO: delay slot (?)
8054 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
8055 if(or<0||(r&63)>=TEMPREG){
8056 regs[i].regmap_entry[hr]=-1;
8057 }
8058 else
8059 {
8060 // Just move it to a different register
8061 regs[i].regmap_entry[hr]=r;
8062 // If it was dirty before, it's still dirty
8063 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r&63);
8064 }
8065 }
8066 else
8067 {
8068 // Unneeded
8069 if(r==0){
8070 regs[i].regmap_entry[hr]=0;
8071 }
8072 else
8073 {
8074 assert(r<64);
8075 if((current.u>>r)&1) {
8076 regs[i].regmap_entry[hr]=-1;
8077 //regs[i].regmap[hr]=-1;
8078 current.regmap[hr]=-1;
8079 }else
8080 regs[i].regmap_entry[hr]=r;
8081 }
8082 }
8083 } else {
8084 // Branches expect CCREG to be allocated at the target
8085 if(regmap_pre[i][hr]==CCREG)
8086 regs[i].regmap_entry[hr]=CCREG;
8087 else
8088 regs[i].regmap_entry[hr]=-1;
8089 }
8090 }
8091 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
8092 }
8093
8094 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)
8095 current.waswritten|=1<<dops[i-1].rs1;
8096 current.waswritten&=~(1<<dops[i].rt1);
8097 current.waswritten&=~(1<<dops[i].rt2);
8098 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
8099 current.waswritten&=~(1<<dops[i].rs1);
8100
8101 /* Branch post-alloc */
8102 if(i>0)
8103 {
8104 current.wasdirty=current.dirty;
8105 switch(dops[i-1].itype) {
8106 case UJUMP:
8107 memcpy(&branch_regs[i-1],&current,sizeof(current));
8108 branch_regs[i-1].isconst=0;
8109 branch_regs[i-1].wasconst=0;
8110 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
8111 alloc_cc(&branch_regs[i-1],i-1);
8112 dirty_reg(&branch_regs[i-1],CCREG);
8113 if(dops[i-1].rt1==31) { // JAL
8114 alloc_reg(&branch_regs[i-1],i-1,31);
8115 dirty_reg(&branch_regs[i-1],31);
8116 }
8117 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8118 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8119 break;
8120 case RJUMP:
8121 memcpy(&branch_regs[i-1],&current,sizeof(current));
8122 branch_regs[i-1].isconst=0;
8123 branch_regs[i-1].wasconst=0;
8124 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
8125 alloc_cc(&branch_regs[i-1],i-1);
8126 dirty_reg(&branch_regs[i-1],CCREG);
8127 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
8128 if(dops[i-1].rt1!=0) { // JALR
8129 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
8130 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
8131 }
8132 #ifdef USE_MINI_HT
8133 if(dops[i-1].rs1==31) { // JALR
8134 alloc_reg(&branch_regs[i-1],i-1,RHASH);
8135 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
8136 }
8137 #endif
8138 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8139 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8140 break;
8141 case CJUMP:
8142 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
8143 {
8144 alloc_cc(&current,i-1);
8145 dirty_reg(&current,CCREG);
8146 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
8147 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
8148 // The delay slot overwrote one of our conditions
8149 // Delay slot goes after the test (in order)
8150 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8151 current.u|=1;
8152 delayslot_alloc(&current,i);
8153 current.isconst=0;
8154 }
8155 else
8156 {
8157 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
8158 // Alloc the branch condition registers
8159 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
8160 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
8161 }
8162 memcpy(&branch_regs[i-1],&current,sizeof(current));
8163 branch_regs[i-1].isconst=0;
8164 branch_regs[i-1].wasconst=0;
8165 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
8166 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8167 }
8168 else
8169 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
8170 {
8171 alloc_cc(&current,i-1);
8172 dirty_reg(&current,CCREG);
8173 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
8174 // The delay slot overwrote the branch condition
8175 // Delay slot goes after the test (in order)
8176 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8177 current.u|=1;
8178 delayslot_alloc(&current,i);
8179 current.isconst=0;
8180 }
8181 else
8182 {
8183 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
8184 // Alloc the branch condition register
8185 alloc_reg(&current,i-1,dops[i-1].rs1);
8186 }
8187 memcpy(&branch_regs[i-1],&current,sizeof(current));
8188 branch_regs[i-1].isconst=0;
8189 branch_regs[i-1].wasconst=0;
8190 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
8191 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8192 }
8193 else
8194 // Alloc the delay slot in case the branch is taken
8195 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
8196 {
8197 memcpy(&branch_regs[i-1],&current,sizeof(current));
8198 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;
8199 alloc_cc(&branch_regs[i-1],i);
8200 dirty_reg(&branch_regs[i-1],CCREG);
8201 delayslot_alloc(&branch_regs[i-1],i);
8202 branch_regs[i-1].isconst=0;
8203 alloc_reg(&current,i,CCREG); // Not taken path
8204 dirty_reg(&current,CCREG);
8205 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8206 }
8207 else
8208 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
8209 {
8210 memcpy(&branch_regs[i-1],&current,sizeof(current));
8211 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;
8212 alloc_cc(&branch_regs[i-1],i);
8213 dirty_reg(&branch_regs[i-1],CCREG);
8214 delayslot_alloc(&branch_regs[i-1],i);
8215 branch_regs[i-1].isconst=0;
8216 alloc_reg(&current,i,CCREG); // Not taken path
8217 dirty_reg(&current,CCREG);
8218 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8219 }
8220 break;
8221 case SJUMP:
8222 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
8223 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
8224 {
8225 alloc_cc(&current,i-1);
8226 dirty_reg(&current,CCREG);
8227 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
8228 // The delay slot overwrote the branch condition
8229 // Delay slot goes after the test (in order)
8230 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
8231 current.u|=1;
8232 delayslot_alloc(&current,i);
8233 current.isconst=0;
8234 }
8235 else
8236 {
8237 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
8238 // Alloc the branch condition register
8239 alloc_reg(&current,i-1,dops[i-1].rs1);
8240 }
8241 memcpy(&branch_regs[i-1],&current,sizeof(current));
8242 branch_regs[i-1].isconst=0;
8243 branch_regs[i-1].wasconst=0;
8244 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
8245 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
8246 }
8247 else
8248 // Alloc the delay slot in case the branch is taken
8249 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
8250 {
8251 memcpy(&branch_regs[i-1],&current,sizeof(current));
8252 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;
8253 alloc_cc(&branch_regs[i-1],i);
8254 dirty_reg(&branch_regs[i-1],CCREG);
8255 delayslot_alloc(&branch_regs[i-1],i);
8256 branch_regs[i-1].isconst=0;
8257 alloc_reg(&current,i,CCREG); // Not taken path
8258 dirty_reg(&current,CCREG);
8259 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
8260 }
8261 // FIXME: BLTZAL/BGEZAL
8262 if(dops[i-1].opcode2&0x10) { // BxxZAL
8263 alloc_reg(&branch_regs[i-1],i-1,31);
8264 dirty_reg(&branch_regs[i-1],31);
8265 }
8266 break;
8267 }
8268
8269 if (dops[i-1].is_ujump)
8270 {
8271 if(dops[i-1].rt1==31) // JAL/JALR
8272 {
8273 // Subroutine call will return here, don't alloc any registers
8274 current.dirty=0;
8275 clear_all_regs(current.regmap);
8276 alloc_reg(&current,i,CCREG);
8277 dirty_reg(&current,CCREG);
8278 }
8279 else if(i+1<slen)
8280 {
8281 // Internal branch will jump here, match registers to caller
8282 current.dirty=0;
8283 clear_all_regs(current.regmap);
8284 alloc_reg(&current,i,CCREG);
8285 dirty_reg(&current,CCREG);
8286 for(j=i-1;j>=0;j--)
8287 {
8288 if(ba[j]==start+i*4+4) {
8289 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
8290 current.dirty=branch_regs[j].dirty;
8291 break;
8292 }
8293 }
8294 while(j>=0) {
8295 if(ba[j]==start+i*4+4) {
8296 for(hr=0;hr<HOST_REGS;hr++) {
8297 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
8298 current.regmap[hr]=-1;
8299 }
8300 current.dirty&=branch_regs[j].dirty;
8301 }
8302 }
8303 j--;
8304 }
8305 }
8306 }
8307 }
8308
8309 // Count cycles in between branches
8310 ccadj[i] = CLOCK_ADJUST(cc);
8311 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
8312 {
8313 cc=0;
8314 }
8315#if !defined(DRC_DBG)
8316 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
8317 {
8318 // this should really be removed since the real stalls have been implemented,
8319 // but doing so causes sizeable perf regression against the older version
8320 u_int gtec = gte_cycletab[source[i] & 0x3f];
8321 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
8322 }
8323 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
8324 {
8325 cc+=4;
8326 }
8327 else if(dops[i].itype==C2LS)
8328 {
8329 // same as with C2OP
8330 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
8331 }
8332#endif
8333 else
8334 {
8335 cc++;
8336 }
8337
8338 if(!dops[i].is_ds) {
8339 regs[i].dirty=current.dirty;
8340 regs[i].isconst=current.isconst;
8341 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
8342 }
8343 for(hr=0;hr<HOST_REGS;hr++) {
8344 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
8345 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
8346 regs[i].wasconst&=~(1<<hr);
8347 }
8348 }
8349 }
8350 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
8351 regs[i].waswritten=current.waswritten;
8352 }
8353
8354 /* Pass 4 - Cull unused host registers */
8355
8356 uint64_t nr=0;
8357
8358 for (i=slen-1;i>=0;i--)
8359 {
8360 int hr;
8361 if(dops[i].is_jump)
8362 {
8363 if(ba[i]<start || ba[i]>=(start+slen*4))
8364 {
8365 // Branch out of this block, don't need anything
8366 nr=0;
8367 }
8368 else
8369 {
8370 // Internal branch
8371 // Need whatever matches the target
8372 nr=0;
8373 int t=(ba[i]-start)>>2;
8374 for(hr=0;hr<HOST_REGS;hr++)
8375 {
8376 if(regs[i].regmap_entry[hr]>=0) {
8377 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
8378 }
8379 }
8380 }
8381 // Conditional branch may need registers for following instructions
8382 if (!dops[i].is_ujump)
8383 {
8384 if(i<slen-2) {
8385 nr|=needed_reg[i+2];
8386 for(hr=0;hr<HOST_REGS;hr++)
8387 {
8388 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
8389 //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]);
8390 }
8391 }
8392 }
8393 // Don't need stuff which is overwritten
8394 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8395 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8396 // Merge in delay slot
8397 for(hr=0;hr<HOST_REGS;hr++)
8398 {
8399 if(dops[i+1].rt1&&dops[i+1].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8400 if(dops[i+1].rt2&&dops[i+1].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8401 if(dops[i+1].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8402 if(dops[i+1].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8403 if(dops[i+1].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8404 if(dops[i+1].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8405 if(ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
8406 if(regmap_pre[i][hr]==ROREG) nr|=1<<hr;
8407 if(regs[i].regmap_entry[hr]==ROREG) nr|=1<<hr;
8408 }
8409 if(dops[i+1].is_store) {
8410 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8411 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8412 }
8413 }
8414 }
8415 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8416 {
8417 // SYSCALL instruction (software interrupt)
8418 nr=0;
8419 }
8420 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8421 {
8422 // ERET instruction (return from interrupt)
8423 nr=0;
8424 }
8425 else // Non-branch
8426 {
8427 if(i<slen-1) {
8428 for(hr=0;hr<HOST_REGS;hr++) {
8429 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8430 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8431 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8432 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8433 }
8434 }
8435 }
8436 for(hr=0;hr<HOST_REGS;hr++)
8437 {
8438 // Overwritten registers are not needed
8439 if(dops[i].rt1&&dops[i].rt1==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8440 if(dops[i].rt2&&dops[i].rt2==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8441 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
8442 // Source registers are needed
8443 if(dops[i].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8444 if(dops[i].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8445 if(dops[i].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8446 if(dops[i].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8447 if(ram_offset && (dops[i].is_load || dops[i].is_store)) {
8448 if(regmap_pre[i][hr]==ROREG) nr|=1<<hr;
8449 if(regs[i].regmap_entry[hr]==ROREG) nr|=1<<hr;
8450 }
8451 if(dops[i].is_store) {
8452 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8453 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8454 }
8455 // Don't store a register immediately after writing it,
8456 // may prevent dual-issue.
8457 // But do so if this is a branch target, otherwise we
8458 // might have to load the register before the branch.
8459 if(i>0&&!dops[i].bt&&((regs[i].wasdirty>>hr)&1)) {
8460 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
8461 if(dops[i-1].rt1==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8462 if(dops[i-1].rt2==(regmap_pre[i][hr]&63)) nr|=1<<hr;
8463 }
8464 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
8465 if(dops[i-1].rt1==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8466 if(dops[i-1].rt2==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
8467 }
8468 }
8469 }
8470 // Cycle count is needed at branches. Assume it is needed at the target too.
8471 if(i==0||dops[i].bt||dops[i].itype==CJUMP||dops[i].itype==SPAN) {
8472 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8473 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8474 }
8475 // Save it
8476 needed_reg[i]=nr;
8477
8478 // Deallocate unneeded registers
8479 for(hr=0;hr<HOST_REGS;hr++)
8480 {
8481 if(!((nr>>hr)&1)) {
8482 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8483 if(dops[i].is_jump)
8484 {
8485 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
8486 if (dops[i+1].is_load || dops[i+1].is_store)
8487 map1 = ROREG;
8488 if (dops[i+1].is_store)
8489 map2 = INVCP;
8490 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
8491 temp = FTEMP;
8492 if((regs[i].regmap[hr]&63)!=dops[i].rs1 && (regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8493 (regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8494 (regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8495 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8496 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
8497 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8498 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8499 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
8500 {
8501 regs[i].regmap[hr]=-1;
8502 regs[i].isconst&=~(1<<hr);
8503 if((branch_regs[i].regmap[hr]&63)!=dops[i].rs1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rs2 &&
8504 (branch_regs[i].regmap[hr]&63)!=dops[i].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8505 (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt1 && (branch_regs[i].regmap[hr]&63)!=dops[i+1].rt2 &&
8506 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8507 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8508 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8509 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8510 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8511 {
8512 branch_regs[i].regmap[hr]=-1;
8513 branch_regs[i].regmap_entry[hr]=-1;
8514 if (!dops[i].is_ujump)
8515 {
8516 if (i < slen-2) {
8517 regmap_pre[i+2][hr]=-1;
8518 regs[i+2].wasconst&=~(1<<hr);
8519 }
8520 }
8521 }
8522 }
8523 }
8524 else
8525 {
8526 // Non-branch
8527 if(i>0)
8528 {
8529 int map1 = -1, map2 = -1, temp=-1;
8530 if (dops[i].is_load || dops[i].is_store)
8531 map1 = ROREG;
8532 if (dops[i].is_store)
8533 map2 = INVCP;
8534 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8535 temp = FTEMP;
8536 if((regs[i].regmap[hr]&63)!=dops[i].rt1 && (regs[i].regmap[hr]&63)!=dops[i].rt2 &&
8537 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8538 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8539 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8540 regs[i].regmap[hr] != CCREG)
8541 {
8542 if(i<slen-1&&!dops[i].is_ds) {
8543 assert(regs[i].regmap[hr]<64);
8544 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8545 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8546 {
8547 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8548 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8549 }
8550 regmap_pre[i+1][hr]=-1;
8551 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8552 regs[i+1].wasconst&=~(1<<hr);
8553 }
8554 regs[i].regmap[hr]=-1;
8555 regs[i].isconst&=~(1<<hr);
8556 }
8557 }
8558 }
8559 } // if needed
8560 } // for hr
8561 }
8562
8563 /* Pass 5 - Pre-allocate registers */
8564
8565 // If a register is allocated during a loop, try to allocate it for the
8566 // entire loop, if possible. This avoids loading/storing registers
8567 // inside of the loop.
8568
8569 signed char f_regmap[HOST_REGS];
8570 clear_all_regs(f_regmap);
8571 for(i=0;i<slen-1;i++)
8572 {
8573 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8574 {
8575 if(ba[i]>=start && ba[i]<(start+i*4))
8576 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8577 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8578 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8579 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8580 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8581 {
8582 int t=(ba[i]-start)>>2;
8583 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8584 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8585 for(hr=0;hr<HOST_REGS;hr++)
8586 {
8587 if(regs[i].regmap[hr]>=0) {
8588 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8589 // dealloc old register
8590 int n;
8591 for(n=0;n<HOST_REGS;n++)
8592 {
8593 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8594 }
8595 // and alloc new one
8596 f_regmap[hr]=regs[i].regmap[hr];
8597 }
8598 }
8599 if(branch_regs[i].regmap[hr]>=0) {
8600 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8601 // dealloc old register
8602 int n;
8603 for(n=0;n<HOST_REGS;n++)
8604 {
8605 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8606 }
8607 // and alloc new one
8608 f_regmap[hr]=branch_regs[i].regmap[hr];
8609 }
8610 }
8611 if(dops[i].ooo) {
8612 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8613 f_regmap[hr]=branch_regs[i].regmap[hr];
8614 }else{
8615 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8616 f_regmap[hr]=branch_regs[i].regmap[hr];
8617 }
8618 // Avoid dirty->clean transition
8619 #ifdef DESTRUCTIVE_WRITEBACK
8620 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;
8621 #endif
8622 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8623 // case above, however it's always a good idea. We can't hoist the
8624 // load if the register was already allocated, so there's no point
8625 // wasting time analyzing most of these cases. It only "succeeds"
8626 // when the mapping was different and the load can be replaced with
8627 // a mov, which is of negligible benefit. So such cases are
8628 // skipped below.
8629 if(f_regmap[hr]>0) {
8630 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8631 int r=f_regmap[hr];
8632 for(j=t;j<=i;j++)
8633 {
8634 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8635 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8636 assert(r < 64);
8637 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
8638 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8639 int k;
8640 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8641 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8642 if(r>63) {
8643 if(get_reg(regs[i].regmap,r&63)<0) break;
8644 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
8645 }
8646 k=i;
8647 while(k>1&&regs[k-1].regmap[hr]==-1) {
8648 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8649 //printf("no free regs for store %x\n",start+(k-1)*4);
8650 break;
8651 }
8652 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8653 //printf("no-match due to different register\n");
8654 break;
8655 }
8656 if (dops[k-2].is_jump) {
8657 //printf("no-match due to branch\n");
8658 break;
8659 }
8660 // call/ret fast path assumes no registers allocated
8661 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8662 break;
8663 }
8664 assert(r < 64);
8665 k--;
8666 }
8667 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8668 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8669 while(k<i) {
8670 regs[k].regmap_entry[hr]=f_regmap[hr];
8671 regs[k].regmap[hr]=f_regmap[hr];
8672 regmap_pre[k+1][hr]=f_regmap[hr];
8673 regs[k].wasdirty&=~(1<<hr);
8674 regs[k].dirty&=~(1<<hr);
8675 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8676 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8677 regs[k].wasconst&=~(1<<hr);
8678 regs[k].isconst&=~(1<<hr);
8679 k++;
8680 }
8681 }
8682 else {
8683 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8684 break;
8685 }
8686 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8687 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8688 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8689 regs[i].regmap_entry[hr]=f_regmap[hr];
8690 regs[i].regmap[hr]=f_regmap[hr];
8691 regs[i].wasdirty&=~(1<<hr);
8692 regs[i].dirty&=~(1<<hr);
8693 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8694 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8695 regs[i].wasconst&=~(1<<hr);
8696 regs[i].isconst&=~(1<<hr);
8697 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8698 branch_regs[i].wasdirty&=~(1<<hr);
8699 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8700 branch_regs[i].regmap[hr]=f_regmap[hr];
8701 branch_regs[i].dirty&=~(1<<hr);
8702 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8703 branch_regs[i].wasconst&=~(1<<hr);
8704 branch_regs[i].isconst&=~(1<<hr);
8705 if (!dops[i].is_ujump) {
8706 regmap_pre[i+2][hr]=f_regmap[hr];
8707 regs[i+2].wasdirty&=~(1<<hr);
8708 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8709 }
8710 }
8711 }
8712 for(k=t;k<j;k++) {
8713 // Alloc register clean at beginning of loop,
8714 // but may dirty it in pass 6
8715 regs[k].regmap_entry[hr]=f_regmap[hr];
8716 regs[k].regmap[hr]=f_regmap[hr];
8717 regs[k].dirty&=~(1<<hr);
8718 regs[k].wasconst&=~(1<<hr);
8719 regs[k].isconst&=~(1<<hr);
8720 if (dops[k].is_jump) {
8721 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8722 branch_regs[k].regmap[hr]=f_regmap[hr];
8723 branch_regs[k].dirty&=~(1<<hr);
8724 branch_regs[k].wasconst&=~(1<<hr);
8725 branch_regs[k].isconst&=~(1<<hr);
8726 if (!dops[k].is_ujump) {
8727 regmap_pre[k+2][hr]=f_regmap[hr];
8728 regs[k+2].wasdirty&=~(1<<hr);
8729 }
8730 }
8731 else
8732 {
8733 regmap_pre[k+1][hr]=f_regmap[hr];
8734 regs[k+1].wasdirty&=~(1<<hr);
8735 }
8736 }
8737 if(regs[j].regmap[hr]==f_regmap[hr])
8738 regs[j].regmap_entry[hr]=f_regmap[hr];
8739 break;
8740 }
8741 if(j==i) break;
8742 if(regs[j].regmap[hr]>=0)
8743 break;
8744 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8745 //printf("no-match due to different register\n");
8746 break;
8747 }
8748 if (dops[j].is_ujump)
8749 {
8750 // Stop on unconditional branch
8751 break;
8752 }
8753 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8754 {
8755 if(dops[j].ooo) {
8756 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8757 break;
8758 }else{
8759 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8760 break;
8761 }
8762 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8763 //printf("no-match due to different register (branch)\n");
8764 break;
8765 }
8766 }
8767 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8768 //printf("No free regs for store %x\n",start+j*4);
8769 break;
8770 }
8771 assert(f_regmap[hr]<64);
8772 }
8773 }
8774 }
8775 }
8776 }
8777 }else{
8778 // Non branch or undetermined branch target
8779 for(hr=0;hr<HOST_REGS;hr++)
8780 {
8781 if(hr!=EXCLUDE_REG) {
8782 if(regs[i].regmap[hr]>=0) {
8783 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8784 // dealloc old register
8785 int n;
8786 for(n=0;n<HOST_REGS;n++)
8787 {
8788 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8789 }
8790 // and alloc new one
8791 f_regmap[hr]=regs[i].regmap[hr];
8792 }
8793 }
8794 }
8795 }
8796 // Try to restore cycle count at branch targets
8797 if(dops[i].bt) {
8798 for(j=i;j<slen-1;j++) {
8799 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8800 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8801 //printf("no free regs for store %x\n",start+j*4);
8802 break;
8803 }
8804 }
8805 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8806 int k=i;
8807 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8808 while(k<j) {
8809 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8810 regs[k].regmap[HOST_CCREG]=CCREG;
8811 regmap_pre[k+1][HOST_CCREG]=CCREG;
8812 regs[k+1].wasdirty|=1<<HOST_CCREG;
8813 regs[k].dirty|=1<<HOST_CCREG;
8814 regs[k].wasconst&=~(1<<HOST_CCREG);
8815 regs[k].isconst&=~(1<<HOST_CCREG);
8816 k++;
8817 }
8818 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8819 }
8820 // Work backwards from the branch target
8821 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8822 {
8823 //printf("Extend backwards\n");
8824 int k;
8825 k=i;
8826 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8827 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8828 //printf("no free regs for store %x\n",start+(k-1)*4);
8829 break;
8830 }
8831 k--;
8832 }
8833 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8834 //printf("Extend CC, %x ->\n",start+k*4);
8835 while(k<=i) {
8836 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8837 regs[k].regmap[HOST_CCREG]=CCREG;
8838 regmap_pre[k+1][HOST_CCREG]=CCREG;
8839 regs[k+1].wasdirty|=1<<HOST_CCREG;
8840 regs[k].dirty|=1<<HOST_CCREG;
8841 regs[k].wasconst&=~(1<<HOST_CCREG);
8842 regs[k].isconst&=~(1<<HOST_CCREG);
8843 k++;
8844 }
8845 }
8846 else {
8847 //printf("Fail Extend CC, %x ->\n",start+k*4);
8848 }
8849 }
8850 }
8851 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8852 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8853 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8854 {
8855 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8856 }
8857 }
8858 }
8859
8860 // This allocates registers (if possible) one instruction prior
8861 // to use, which can avoid a load-use penalty on certain CPUs.
8862 for(i=0;i<slen-1;i++)
8863 {
8864 if (!i || !dops[i-1].is_jump)
8865 {
8866 if(!dops[i+1].bt)
8867 {
8868 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8869 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8870 {
8871 if(dops[i+1].rs1) {
8872 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8873 {
8874 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8875 {
8876 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8877 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8878 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8879 regs[i].isconst&=~(1<<hr);
8880 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8881 constmap[i][hr]=constmap[i+1][hr];
8882 regs[i+1].wasdirty&=~(1<<hr);
8883 regs[i].dirty&=~(1<<hr);
8884 }
8885 }
8886 }
8887 if(dops[i+1].rs2) {
8888 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8889 {
8890 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8891 {
8892 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8893 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8894 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8895 regs[i].isconst&=~(1<<hr);
8896 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8897 constmap[i][hr]=constmap[i+1][hr];
8898 regs[i+1].wasdirty&=~(1<<hr);
8899 regs[i].dirty&=~(1<<hr);
8900 }
8901 }
8902 }
8903 // Preload target address for load instruction (non-constant)
8904 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8905 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8906 {
8907 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8908 {
8909 regs[i].regmap[hr]=dops[i+1].rs1;
8910 regmap_pre[i+1][hr]=dops[i+1].rs1;
8911 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8912 regs[i].isconst&=~(1<<hr);
8913 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8914 constmap[i][hr]=constmap[i+1][hr];
8915 regs[i+1].wasdirty&=~(1<<hr);
8916 regs[i].dirty&=~(1<<hr);
8917 }
8918 }
8919 }
8920 // Load source into target register
8921 if(dops[i+1].lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8922 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8923 {
8924 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8925 {
8926 regs[i].regmap[hr]=dops[i+1].rs1;
8927 regmap_pre[i+1][hr]=dops[i+1].rs1;
8928 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8929 regs[i].isconst&=~(1<<hr);
8930 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8931 constmap[i][hr]=constmap[i+1][hr];
8932 regs[i+1].wasdirty&=~(1<<hr);
8933 regs[i].dirty&=~(1<<hr);
8934 }
8935 }
8936 }
8937 // Address for store instruction (non-constant)
8938 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8939 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8940 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8941 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8942 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8943 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
8944 assert(hr>=0);
8945 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8946 {
8947 regs[i].regmap[hr]=dops[i+1].rs1;
8948 regmap_pre[i+1][hr]=dops[i+1].rs1;
8949 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8950 regs[i].isconst&=~(1<<hr);
8951 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8952 constmap[i][hr]=constmap[i+1][hr];
8953 regs[i+1].wasdirty&=~(1<<hr);
8954 regs[i].dirty&=~(1<<hr);
8955 }
8956 }
8957 }
8958 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8959 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8960 int nr;
8961 hr=get_reg(regs[i+1].regmap,FTEMP);
8962 assert(hr>=0);
8963 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8964 {
8965 regs[i].regmap[hr]=dops[i+1].rs1;
8966 regmap_pre[i+1][hr]=dops[i+1].rs1;
8967 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8968 regs[i].isconst&=~(1<<hr);
8969 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8970 constmap[i][hr]=constmap[i+1][hr];
8971 regs[i+1].wasdirty&=~(1<<hr);
8972 regs[i].dirty&=~(1<<hr);
8973 }
8974 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8975 {
8976 // move it to another register
8977 regs[i+1].regmap[hr]=-1;
8978 regmap_pre[i+2][hr]=-1;
8979 regs[i+1].regmap[nr]=FTEMP;
8980 regmap_pre[i+2][nr]=FTEMP;
8981 regs[i].regmap[nr]=dops[i+1].rs1;
8982 regmap_pre[i+1][nr]=dops[i+1].rs1;
8983 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8984 regs[i].isconst&=~(1<<nr);
8985 regs[i+1].isconst&=~(1<<nr);
8986 regs[i].dirty&=~(1<<nr);
8987 regs[i+1].wasdirty&=~(1<<nr);
8988 regs[i+1].dirty&=~(1<<nr);
8989 regs[i+2].wasdirty&=~(1<<nr);
8990 }
8991 }
8992 }
8993 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*/) {
8994 if(dops[i+1].itype==LOAD)
8995 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8996 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8997 hr=get_reg(regs[i+1].regmap,FTEMP);
8998 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
8999 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
9000 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
9001 }
9002 if(hr>=0&&regs[i].regmap[hr]<0) {
9003 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
9004 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
9005 regs[i].regmap[hr]=AGEN1+((i+1)&1);
9006 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
9007 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
9008 regs[i].isconst&=~(1<<hr);
9009 regs[i+1].wasdirty&=~(1<<hr);
9010 regs[i].dirty&=~(1<<hr);
9011 }
9012 }
9013 }
9014 }
9015 }
9016 }
9017 }
9018
9019 /* Pass 6 - Optimize clean/dirty state */
9020 clean_registers(0,slen-1,1);
9021
9022 /* Pass 7 - Identify 32-bit registers */
9023 for (i=slen-1;i>=0;i--)
9024 {
9025 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9026 {
9027 // Conditional branch
9028 if((source[i]>>16)!=0x1000&&i<slen-2) {
9029 // Mark this address as a branch target since it may be called
9030 // upon return from interrupt
9031 dops[i+2].bt=1;
9032 }
9033 }
9034 }
9035
9036 if(dops[slen-1].itype==SPAN) {
9037 dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
9038 }
9039
9040#ifdef REG_ALLOC_PRINT
9041 /* Debug/disassembly */
9042 for(i=0;i<slen;i++)
9043 {
9044 printf("U:");
9045 int r;
9046 for(r=1;r<=CCREG;r++) {
9047 if((unneeded_reg[i]>>r)&1) {
9048 if(r==HIREG) printf(" HI");
9049 else if(r==LOREG) printf(" LO");
9050 else printf(" r%d",r);
9051 }
9052 }
9053 printf("\n");
9054 #if defined(__i386__) || defined(__x86_64__)
9055 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]);
9056 #endif
9057 #ifdef __arm__
9058 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]);
9059 #endif
9060 #if defined(__i386__) || defined(__x86_64__)
9061 printf("needs: ");
9062 if(needed_reg[i]&1) printf("eax ");
9063 if((needed_reg[i]>>1)&1) printf("ecx ");
9064 if((needed_reg[i]>>2)&1) printf("edx ");
9065 if((needed_reg[i]>>3)&1) printf("ebx ");
9066 if((needed_reg[i]>>5)&1) printf("ebp ");
9067 if((needed_reg[i]>>6)&1) printf("esi ");
9068 if((needed_reg[i]>>7)&1) printf("edi ");
9069 printf("\n");
9070 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]);
9071 printf("dirty: ");
9072 if(regs[i].wasdirty&1) printf("eax ");
9073 if((regs[i].wasdirty>>1)&1) printf("ecx ");
9074 if((regs[i].wasdirty>>2)&1) printf("edx ");
9075 if((regs[i].wasdirty>>3)&1) printf("ebx ");
9076 if((regs[i].wasdirty>>5)&1) printf("ebp ");
9077 if((regs[i].wasdirty>>6)&1) printf("esi ");
9078 if((regs[i].wasdirty>>7)&1) printf("edi ");
9079 #endif
9080 #ifdef __arm__
9081 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]);
9082 printf("dirty: ");
9083 if(regs[i].wasdirty&1) printf("r0 ");
9084 if((regs[i].wasdirty>>1)&1) printf("r1 ");
9085 if((regs[i].wasdirty>>2)&1) printf("r2 ");
9086 if((regs[i].wasdirty>>3)&1) printf("r3 ");
9087 if((regs[i].wasdirty>>4)&1) printf("r4 ");
9088 if((regs[i].wasdirty>>5)&1) printf("r5 ");
9089 if((regs[i].wasdirty>>6)&1) printf("r6 ");
9090 if((regs[i].wasdirty>>7)&1) printf("r7 ");
9091 if((regs[i].wasdirty>>8)&1) printf("r8 ");
9092 if((regs[i].wasdirty>>9)&1) printf("r9 ");
9093 if((regs[i].wasdirty>>10)&1) printf("r10 ");
9094 if((regs[i].wasdirty>>12)&1) printf("r12 ");
9095 #endif
9096 printf("\n");
9097 disassemble_inst(i);
9098 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
9099 #if defined(__i386__) || defined(__x86_64__)
9100 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]);
9101 if(regs[i].dirty&1) printf("eax ");
9102 if((regs[i].dirty>>1)&1) printf("ecx ");
9103 if((regs[i].dirty>>2)&1) printf("edx ");
9104 if((regs[i].dirty>>3)&1) printf("ebx ");
9105 if((regs[i].dirty>>5)&1) printf("ebp ");
9106 if((regs[i].dirty>>6)&1) printf("esi ");
9107 if((regs[i].dirty>>7)&1) printf("edi ");
9108 #endif
9109 #ifdef __arm__
9110 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]);
9111 if(regs[i].dirty&1) printf("r0 ");
9112 if((regs[i].dirty>>1)&1) printf("r1 ");
9113 if((regs[i].dirty>>2)&1) printf("r2 ");
9114 if((regs[i].dirty>>3)&1) printf("r3 ");
9115 if((regs[i].dirty>>4)&1) printf("r4 ");
9116 if((regs[i].dirty>>5)&1) printf("r5 ");
9117 if((regs[i].dirty>>6)&1) printf("r6 ");
9118 if((regs[i].dirty>>7)&1) printf("r7 ");
9119 if((regs[i].dirty>>8)&1) printf("r8 ");
9120 if((regs[i].dirty>>9)&1) printf("r9 ");
9121 if((regs[i].dirty>>10)&1) printf("r10 ");
9122 if((regs[i].dirty>>12)&1) printf("r12 ");
9123 #endif
9124 printf("\n");
9125 if(regs[i].isconst) {
9126 printf("constants: ");
9127 #if defined(__i386__) || defined(__x86_64__)
9128 if(regs[i].isconst&1) printf("eax=%x ",(u_int)constmap[i][0]);
9129 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(u_int)constmap[i][1]);
9130 if((regs[i].isconst>>2)&1) printf("edx=%x ",(u_int)constmap[i][2]);
9131 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(u_int)constmap[i][3]);
9132 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(u_int)constmap[i][5]);
9133 if((regs[i].isconst>>6)&1) printf("esi=%x ",(u_int)constmap[i][6]);
9134 if((regs[i].isconst>>7)&1) printf("edi=%x ",(u_int)constmap[i][7]);
9135 #endif
9136 #if defined(__arm__) || defined(__aarch64__)
9137 int r;
9138 for (r = 0; r < ARRAY_SIZE(constmap[i]); r++)
9139 if ((regs[i].isconst >> r) & 1)
9140 printf(" r%d=%x", r, (u_int)constmap[i][r]);
9141 #endif
9142 printf("\n");
9143 }
9144 if(dops[i].is_jump) {
9145 #if defined(__i386__) || defined(__x86_64__)
9146 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]);
9147 if(branch_regs[i].dirty&1) printf("eax ");
9148 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
9149 if((branch_regs[i].dirty>>2)&1) printf("edx ");
9150 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
9151 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
9152 if((branch_regs[i].dirty>>6)&1) printf("esi ");
9153 if((branch_regs[i].dirty>>7)&1) printf("edi ");
9154 #endif
9155 #ifdef __arm__
9156 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]);
9157 if(branch_regs[i].dirty&1) printf("r0 ");
9158 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
9159 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
9160 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
9161 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
9162 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
9163 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
9164 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
9165 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
9166 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
9167 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
9168 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
9169 #endif
9170 }
9171 }
9172#endif // REG_ALLOC_PRINT
9173
9174 /* Pass 8 - Assembly */
9175 linkcount=0;stubcount=0;
9176 ds=0;is_delayslot=0;
9177 u_int dirty_pre=0;
9178 void *beginning=start_block();
9179 if((u_int)addr&1) {
9180 ds=1;
9181 pagespan_ds();
9182 }
9183 void *instr_addr0_override = NULL;
9184
9185 if (start == 0x80030000) {
9186 // nasty hack for the fastbios thing
9187 // override block entry to this code
9188 instr_addr0_override = out;
9189 emit_movimm(start,0);
9190 // abuse io address var as a flag that we
9191 // have already returned here once
9192 emit_readword(&address,1);
9193 emit_writeword(0,&pcaddr);
9194 emit_writeword(0,&address);
9195 emit_cmp(0,1);
9196 #ifdef __aarch64__
9197 emit_jeq(out + 4*2);
9198 emit_far_jump(new_dyna_leave);
9199 #else
9200 emit_jne(new_dyna_leave);
9201 #endif
9202 }
9203 for(i=0;i<slen;i++)
9204 {
9205 //if(ds) printf("ds: ");
9206 disassemble_inst(i);
9207 if(ds) {
9208 ds=0; // Skip delay slot
9209 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9210 instr_addr[i] = NULL;
9211 } else {
9212 speculate_register_values(i);
9213 #ifndef DESTRUCTIVE_WRITEBACK
9214 if (i < 2 || !dops[i-2].is_ujump)
9215 {
9216 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9217 }
9218 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9219 dirty_pre=branch_regs[i].dirty;
9220 }else{
9221 dirty_pre=regs[i].dirty;
9222 }
9223 #endif
9224 // write back
9225 if (i < 2 || !dops[i-2].is_ujump)
9226 {
9227 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9228 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9229 }
9230 // branch target entry point
9231 instr_addr[i] = out;
9232 assem_debug("<->\n");
9233 drc_dbg_emit_do_cmp(i, ccadj[i]);
9234 if (clear_hack_addr) {
9235 emit_movimm(0, 0);
9236 emit_writeword(0, &hack_addr);
9237 clear_hack_addr = 0;
9238 }
9239
9240 // load regs
9241 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9242 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9243 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9244 address_generation(i,&regs[i],regs[i].regmap_entry);
9245 load_consts(regmap_pre[i],regs[i].regmap,i);
9246 if(dops[i].is_jump)
9247 {
9248 // Load the delay slot registers if necessary
9249 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))
9250 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9251 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))
9252 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9253 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9254 load_regs(regs[i].regmap_entry,regs[i].regmap,ROREG,ROREG);
9255 if (dops[i+1].is_store)
9256 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9257 }
9258 else if(i+1<slen)
9259 {
9260 // Preload registers for following instruction
9261 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9262 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9263 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9264 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9265 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9266 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9267 }
9268 // TODO: if(is_ooo(i)) address_generation(i+1);
9269 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9270 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
9271 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9272 load_regs(regs[i].regmap_entry,regs[i].regmap,ROREG,ROREG);
9273 if (dops[i].is_store)
9274 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9275
9276 ds = assemble(i, &regs[i], ccadj[i]);
9277
9278 if (dops[i].is_ujump)
9279 literal_pool(1024);
9280 else
9281 literal_pool_jumpover(256);
9282 }
9283 }
9284
9285 assert(slen > 0);
9286 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9287 // no ending needed for this block since INTCALL never returns
9288 }
9289 // If the block did not end with an unconditional branch,
9290 // add a jump to the next instruction.
9291 else if (i > 1) {
9292 if (!dops[i-2].is_ujump && dops[i-1].itype != SPAN) {
9293 assert(!dops[i-1].is_jump);
9294 assert(i==slen);
9295 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9296 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9297 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9298 emit_loadreg(CCREG,HOST_CCREG);
9299 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9300 }
9301 else
9302 {
9303 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9304 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9305 }
9306 add_to_linker(out,start+i*4,0);
9307 emit_jmp(0);
9308 }
9309 }
9310 else
9311 {
9312 assert(i>0);
9313 assert(!dops[i-1].is_jump);
9314 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9315 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9316 emit_loadreg(CCREG,HOST_CCREG);
9317 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9318 add_to_linker(out,start+i*4,0);
9319 emit_jmp(0);
9320 }
9321
9322 // TODO: delay slot stubs?
9323 // Stubs
9324 for(i=0;i<stubcount;i++)
9325 {
9326 switch(stubs[i].type)
9327 {
9328 case LOADB_STUB:
9329 case LOADH_STUB:
9330 case LOADW_STUB:
9331 case LOADD_STUB:
9332 case LOADBU_STUB:
9333 case LOADHU_STUB:
9334 do_readstub(i);break;
9335 case STOREB_STUB:
9336 case STOREH_STUB:
9337 case STOREW_STUB:
9338 case STORED_STUB:
9339 do_writestub(i);break;
9340 case CC_STUB:
9341 do_ccstub(i);break;
9342 case INVCODE_STUB:
9343 do_invstub(i);break;
9344 case FP_STUB:
9345 do_cop1stub(i);break;
9346 case STORELR_STUB:
9347 do_unalignedwritestub(i);break;
9348 }
9349 }
9350
9351 if (instr_addr0_override)
9352 instr_addr[0] = instr_addr0_override;
9353
9354 /* Pass 9 - Linker */
9355 for(i=0;i<linkcount;i++)
9356 {
9357 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9358 literal_pool(64);
9359 if (!link_addr[i].ext)
9360 {
9361 void *stub = out;
9362 void *addr = check_addr(link_addr[i].target);
9363 emit_extjump(link_addr[i].addr, link_addr[i].target);
9364 if (addr) {
9365 set_jump_target(link_addr[i].addr, addr);
9366 add_jump_out(link_addr[i].target,stub);
9367 }
9368 else
9369 set_jump_target(link_addr[i].addr, stub);
9370 }
9371 else
9372 {
9373 // Internal branch
9374 int target=(link_addr[i].target-start)>>2;
9375 assert(target>=0&&target<slen);
9376 assert(instr_addr[target]);
9377 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9378 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9379 //#else
9380 set_jump_target(link_addr[i].addr, instr_addr[target]);
9381 //#endif
9382 }
9383 }
9384
9385 u_int source_len = slen*4;
9386 if (dops[slen-1].itype == INTCALL && source_len > 4)
9387 // no need to treat the last instruction as compiled
9388 // as interpreter fully handles it
9389 source_len -= 4;
9390
9391 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9392 copy = shadow;
9393
9394 // External Branch Targets (jump_in)
9395 for(i=0;i<slen;i++)
9396 {
9397 if(dops[i].bt||i==0)
9398 {
9399 if(instr_addr[i]) // TODO - delay slots (=null)
9400 {
9401 u_int vaddr=start+i*4;
9402 u_int page=get_page(vaddr);
9403 u_int vpage=get_vpage(vaddr);
9404 literal_pool(256);
9405 {
9406 assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
9407 assem_debug("jump_in: %x\n",start+i*4);
9408 ll_add(jump_dirty+vpage,vaddr,out);
9409 void *entry_point = do_dirty_stub(i, source_len);
9410 ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
9411 // If there was an existing entry in the hash table,
9412 // replace it with the new address.
9413 // Don't add new entries. We'll insert the
9414 // ones that actually get used in check_addr().
9415 struct ht_entry *ht_bin = hash_table_get(vaddr);
9416 if (ht_bin->vaddr[0] == vaddr)
9417 ht_bin->tcaddr[0] = entry_point;
9418 if (ht_bin->vaddr[1] == vaddr)
9419 ht_bin->tcaddr[1] = entry_point;
9420 }
9421 }
9422 }
9423 }
9424 // Write out the literal pool if necessary
9425 literal_pool(0);
9426 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9427 // Align code
9428 if(((u_int)out)&7) emit_addnop(13);
9429 #endif
9430 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9431 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9432 memcpy(copy, source, source_len);
9433 copy += source_len;
9434
9435 end_block(beginning);
9436
9437 // If we're within 256K of the end of the buffer,
9438 // start over from the beginning. (Is 256K enough?)
9439 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9440 out = ndrc->translation_cache;
9441
9442 // Trap writes to any of the pages we compiled
9443 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
9444 invalid_code[i]=0;
9445 }
9446 inv_code_start=inv_code_end=~0;
9447
9448 // for PCSX we need to mark all mirrors too
9449 if(get_page(start)<(RAM_SIZE>>12))
9450 for(i=start>>12;i<=(start+slen*4)>>12;i++)
9451 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
9452 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
9453 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
9454
9455 /* Pass 10 - Free memory by expiring oldest blocks */
9456
9457 int end=(((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
9458 while(expirep!=end)
9459 {
9460 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
9461 uintptr_t base_offs = ((uintptr_t)(expirep >> 13) << shift); // Base offset of this block
9462 uintptr_t base_offs_s = base_offs >> shift;
9463 inv_debug("EXP: Phase %d\n",expirep);
9464 switch((expirep>>11)&3)
9465 {
9466 case 0:
9467 // Clear jump_in and jump_dirty
9468 ll_remove_matching_addrs(jump_in+(expirep&2047),base_offs_s,shift);
9469 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base_offs_s,shift);
9470 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base_offs_s,shift);
9471 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base_offs_s,shift);
9472 break;
9473 case 1:
9474 // Clear pointers
9475 ll_kill_pointers(jump_out[expirep&2047],base_offs_s,shift);
9476 ll_kill_pointers(jump_out[(expirep&2047)+2048],base_offs_s,shift);
9477 break;
9478 case 2:
9479 // Clear hash table
9480 for(i=0;i<32;i++) {
9481 struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
9482 uintptr_t o1 = (u_char *)ht_bin->tcaddr[1] - ndrc->translation_cache;
9483 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9484 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9485 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
9486 ht_bin->vaddr[1] = -1;
9487 ht_bin->tcaddr[1] = NULL;
9488 }
9489 o1 = (u_char *)ht_bin->tcaddr[0] - ndrc->translation_cache;
9490 o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9491 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9492 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
9493 ht_bin->vaddr[0] = ht_bin->vaddr[1];
9494 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
9495 ht_bin->vaddr[1] = -1;
9496 ht_bin->tcaddr[1] = NULL;
9497 }
9498 }
9499 break;
9500 case 3:
9501 // Clear jump_out
9502 if((expirep&2047)==0)
9503 do_clear_cache();
9504 ll_remove_matching_addrs(jump_out+(expirep&2047),base_offs_s,shift);
9505 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base_offs_s,shift);
9506 break;
9507 }
9508 expirep=(expirep+1)&65535;
9509 }
9510#ifdef ASSEM_PRINT
9511 fflush(stdout);
9512#endif
9513 return 0;
9514}
9515
9516// vim:shiftwidth=2:expandtab