notaz.gp2x.de / pcsx_rearmed.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
drc: put passes in their own functions
[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#include "arm_features.h"
39
40#define noinline __attribute__((noinline,noclone))
41#ifndef ARRAY_SIZE
42#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
43#endif
44#ifndef min
45#define min(a, b) ((b) < (a) ? (b) : (a))
46#endif
47#ifndef max
48#define max(a, b) ((b) > (a) ? (b) : (a))
49#endif
50
51//#define DISASM
52//#define ASSEM_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
122// regmap_pre[i] - regs before [i] insn starts; dirty things here that
123// don't match .regmap will be written back
124// [i].regmap_entry - regs that must be set up if someone jumps here
125// [i].regmap - regs [i] insn will read/(over)write
126// branch_regs[i].* - same as above but for branches, takes delay slot into account
127struct regstat
128{
129 signed char regmap_entry[HOST_REGS];
130 signed char regmap[HOST_REGS];
131 uint64_t wasdirty;
132 uint64_t dirty;
133 uint64_t u;
134 u_int wasconst; // before; for example 'lw r2, (r2)' wasconst is true
135 u_int isconst; // ... but isconst is false when r2 is known
136 u_int loadedconst; // host regs that have constants loaded
137 u_int waswritten; // MIPS regs that were used as store base before
138};
139
140// note: asm depends on this layout
141struct ll_entry
142{
143 u_int vaddr;
144 u_int reg_sv_flags;
145 void *addr;
146 struct ll_entry *next;
147};
148
149struct ht_entry
150{
151 u_int vaddr[2];
152 void *tcaddr[2];
153};
154
155struct code_stub
156{
157 enum stub_type type;
158 void *addr;
159 void *retaddr;
160 u_int a;
161 uintptr_t b;
162 uintptr_t c;
163 u_int d;
164 u_int e;
165};
166
167struct link_entry
168{
169 void *addr;
170 u_int target;
171 u_int ext;
172};
173
174static struct decoded_insn
175{
176 u_char itype;
177 u_char opcode;
178 u_char opcode2;
179 u_char rs1;
180 u_char rs2;
181 u_char rt1;
182 u_char rt2;
183 u_char use_lt1:1;
184 u_char bt:1;
185 u_char ooo:1;
186 u_char is_ds:1;
187 u_char is_jump:1;
188 u_char is_ujump:1;
189 u_char is_load:1;
190 u_char is_store:1;
191} dops[MAXBLOCK];
192
193 // used by asm:
194 u_char *out;
195 struct ht_entry hash_table[65536] __attribute__((aligned(16)));
196 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
197 struct ll_entry *jump_dirty[4096];
198
199 static struct ll_entry *jump_out[4096];
200 static u_int start;
201 static u_int *source;
202 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
203 static uint64_t gte_rt[MAXBLOCK];
204 static uint64_t gte_unneeded[MAXBLOCK];
205 static u_int smrv[32]; // speculated MIPS register values
206 static u_int smrv_strong; // mask or regs that are likely to have correct values
207 static u_int smrv_weak; // same, but somewhat less likely
208 static u_int smrv_strong_next; // same, but after current insn executes
209 static u_int smrv_weak_next;
210 static int imm[MAXBLOCK];
211 static u_int ba[MAXBLOCK];
212 static uint64_t unneeded_reg[MAXBLOCK];
213 static uint64_t branch_unneeded_reg[MAXBLOCK];
214 // see 'struct regstat' for a description
215 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
216 // contains 'real' consts at [i] insn, but may differ from what's actually
217 // loaded in host reg as 'final' value is always loaded, see get_final_value()
218 static uint32_t current_constmap[HOST_REGS];
219 static uint32_t constmap[MAXBLOCK][HOST_REGS];
220 static struct regstat regs[MAXBLOCK];
221 static struct regstat branch_regs[MAXBLOCK];
222 static signed char minimum_free_regs[MAXBLOCK];
223 static u_int needed_reg[MAXBLOCK];
224 static u_int wont_dirty[MAXBLOCK];
225 static u_int will_dirty[MAXBLOCK];
226 static int ccadj[MAXBLOCK];
227 static int slen;
228 static void *instr_addr[MAXBLOCK];
229 static struct link_entry link_addr[MAXBLOCK];
230 static int linkcount;
231 static struct code_stub stubs[MAXBLOCK*3];
232 static int stubcount;
233 static u_int literals[1024][2];
234 static int literalcount;
235 static int is_delayslot;
236 static char shadow[1048576] __attribute__((aligned(16)));
237 static void *copy;
238 static int expirep;
239 static u_int stop_after_jal;
240 static u_int f1_hack;
241
242 int new_dynarec_hacks;
243 int new_dynarec_hacks_pergame;
244 int new_dynarec_hacks_old;
245 int new_dynarec_did_compile;
246
247 #define HACK_ENABLED(x) ((new_dynarec_hacks | new_dynarec_hacks_pergame) & (x))
248
249 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
250 extern int last_count; // last absolute target, often = next_interupt
251 extern int pcaddr;
252 extern int pending_exception;
253 extern int branch_target;
254 extern uintptr_t ram_offset;
255 extern uintptr_t mini_ht[32][2];
256 extern u_char restore_candidate[512];
257
258 /* registers that may be allocated */
259 /* 1-31 gpr */
260#define LOREG 32 // lo
261#define HIREG 33 // hi
262//#define FSREG 34 // FPU status (FCSR)
263#define CSREG 35 // Coprocessor status
264#define CCREG 36 // Cycle count
265#define INVCP 37 // Pointer to invalid_code
266//#define MMREG 38 // Pointer to memory_map
267#define ROREG 39 // ram offset (if rdram!=0x80000000)
268#define TEMPREG 40
269#define FTEMP 40 // FPU temporary register
270#define PTEMP 41 // Prefetch temporary register
271//#define TLREG 42 // TLB mapping offset
272#define RHASH 43 // Return address hash
273#define RHTBL 44 // Return address hash table address
274#define RTEMP 45 // JR/JALR address register
275#define MAXREG 45
276#define AGEN1 46 // Address generation temporary register
277//#define AGEN2 47 // Address generation temporary register
278//#define MGEN1 48 // Maptable address generation temporary register
279//#define MGEN2 49 // Maptable address generation temporary register
280#define BTREG 50 // Branch target temporary register
281
282 /* instruction types */
283#define NOP 0 // No operation
284#define LOAD 1 // Load
285#define STORE 2 // Store
286#define LOADLR 3 // Unaligned load
287#define STORELR 4 // Unaligned store
288#define MOV 5 // Move
289#define ALU 6 // Arithmetic/logic
290#define MULTDIV 7 // Multiply/divide
291#define SHIFT 8 // Shift by register
292#define SHIFTIMM 9// Shift by immediate
293#define IMM16 10 // 16-bit immediate
294#define RJUMP 11 // Unconditional jump to register
295#define UJUMP 12 // Unconditional jump
296#define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
297#define SJUMP 14 // Conditional branch (regimm format)
298#define COP0 15 // Coprocessor 0
299#define COP1 16 // Coprocessor 1
300#define C1LS 17 // Coprocessor 1 load/store
301//#define FJUMP 18 // Conditional branch (floating point)
302//#define FLOAT 19 // Floating point unit
303//#define FCONV 20 // Convert integer to float
304//#define FCOMP 21 // Floating point compare (sets FSREG)
305#define SYSCALL 22// SYSCALL,BREAK
306#define OTHER 23 // Other
307#define SPAN 24 // Branch/delay slot spans 2 pages
308#define NI 25 // Not implemented
309#define HLECALL 26// PCSX fake opcodes for HLE
310#define COP2 27 // Coprocessor 2 move
311#define C2LS 28 // Coprocessor 2 load/store
312#define C2OP 29 // Coprocessor 2 operation
313#define INTCALL 30// Call interpreter to handle rare corner cases
314
315 /* branch codes */
316#define TAKEN 1
317#define NOTTAKEN 2
318#define NULLDS 3
319
320#define DJT_1 (void *)1l // no function, just a label in assem_debug log
321#define DJT_2 (void *)2l
322
323// asm linkage
324int new_recompile_block(u_int addr);
325void *get_addr_ht(u_int vaddr);
326void invalidate_block(u_int block);
327void invalidate_addr(u_int addr);
328void remove_hash(int vaddr);
329void dyna_linker();
330void dyna_linker_ds();
331void verify_code();
332void verify_code_ds();
333void cc_interrupt();
334void fp_exception();
335void fp_exception_ds();
336void jump_syscall (u_int u0, u_int u1, u_int pc);
337void jump_syscall_ds(u_int u0, u_int u1, u_int pc);
338void jump_break (u_int u0, u_int u1, u_int pc);
339void jump_break_ds(u_int u0, u_int u1, u_int pc);
340void jump_to_new_pc();
341void call_gteStall();
342void clean_blocks(u_int page);
343void add_jump_out(u_int vaddr, void *src);
344void new_dyna_leave();
345
346// Needed by assembler
347static void wb_register(signed char r, const signed char regmap[], uint64_t dirty);
348static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty);
349static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr);
350static void load_all_regs(const signed char i_regmap[]);
351static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[]);
352static void load_regs_entry(int t);
353static void load_all_consts(const signed char regmap[], u_int dirty, int i);
354static u_int get_host_reglist(const signed char *regmap);
355
356static int verify_dirty(const u_int *ptr);
357static int get_final_value(int hr, int i, int *value);
358static void add_stub(enum stub_type type, void *addr, void *retaddr,
359 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
360static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
361 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist);
362static void add_to_linker(void *addr, u_int target, int ext);
363static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
364 int addr, int *offset_reg, int *addr_reg_override);
365static void *get_direct_memhandler(void *table, u_int addr,
366 enum stub_type type, uintptr_t *addr_host);
367static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist);
368static void pass_args(int a0, int a1);
369static void emit_far_jump(const void *f);
370static void emit_far_call(const void *f);
371
372#ifdef VITA
373#include <psp2/kernel/sysmem.h>
374static int sceBlock;
375// note: this interacts with RetroArch's Vita bootstrap code: bootstrap/vita/sbrk.c
376extern int getVMBlock();
377int _newlib_vm_size_user = sizeof(*ndrc);
378#endif
379
380static void mprotect_w_x(void *start, void *end, int is_x)
381{
382#ifdef NO_WRITE_EXEC
383 #if defined(VITA)
384 // *Open* enables write on all memory that was
385 // allocated by sceKernelAllocMemBlockForVM()?
386 if (is_x)
387 sceKernelCloseVMDomain();
388 else
389 sceKernelOpenVMDomain();
390 #else
391 u_long mstart = (u_long)start & ~4095ul;
392 u_long mend = (u_long)end;
393 if (mprotect((void *)mstart, mend - mstart,
394 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
395 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
396 #endif
397#endif
398}
399
400static void start_tcache_write(void *start, void *end)
401{
402 mprotect_w_x(start, end, 0);
403}
404
405static void end_tcache_write(void *start, void *end)
406{
407#if defined(__arm__) || defined(__aarch64__)
408 size_t len = (char *)end - (char *)start;
409 #if defined(__BLACKBERRY_QNX__)
410 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
411 #elif defined(__MACH__)
412 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
413 #elif defined(VITA)
414 sceKernelSyncVMDomain(sceBlock, start, len);
415 #elif defined(_3DS)
416 ctr_flush_invalidate_cache();
417 #elif defined(__aarch64__)
418 // as of 2021, __clear_cache() is still broken on arm64
419 // so here is a custom one :(
420 clear_cache_arm64(start, end);
421 #else
422 __clear_cache(start, end);
423 #endif
424 (void)len;
425#endif
426
427 mprotect_w_x(start, end, 1);
428}
429
430static void *start_block(void)
431{
432 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
433 if (end > ndrc->translation_cache + sizeof(ndrc->translation_cache))
434 end = ndrc->translation_cache + sizeof(ndrc->translation_cache);
435 start_tcache_write(out, end);
436 return out;
437}
438
439static void end_block(void *start)
440{
441 end_tcache_write(start, out);
442}
443
444// also takes care of w^x mappings when patching code
445static u_int needs_clear_cache[1<<(TARGET_SIZE_2-17)];
446
447static void mark_clear_cache(void *target)
448{
449 uintptr_t offset = (u_char *)target - ndrc->translation_cache;
450 u_int mask = 1u << ((offset >> 12) & 31);
451 if (!(needs_clear_cache[offset >> 17] & mask)) {
452 char *start = (char *)((uintptr_t)target & ~4095l);
453 start_tcache_write(start, start + 4095);
454 needs_clear_cache[offset >> 17] |= mask;
455 }
456}
457
458// Clearing the cache is rather slow on ARM Linux, so mark the areas
459// that need to be cleared, and then only clear these areas once.
460static void do_clear_cache(void)
461{
462 int i, j;
463 for (i = 0; i < (1<<(TARGET_SIZE_2-17)); i++)
464 {
465 u_int bitmap = needs_clear_cache[i];
466 if (!bitmap)
467 continue;
468 for (j = 0; j < 32; j++)
469 {
470 u_char *start, *end;
471 if (!(bitmap & (1<<j)))
472 continue;
473
474 start = ndrc->translation_cache + i*131072 + j*4096;
475 end = start + 4095;
476 for (j++; j < 32; j++) {
477 if (!(bitmap & (1<<j)))
478 break;
479 end += 4096;
480 }
481 end_tcache_write(start, end);
482 }
483 needs_clear_cache[i] = 0;
484 }
485}
486
487//#define DEBUG_CYCLE_COUNT 1
488
489#define NO_CYCLE_PENALTY_THR 12
490
491int cycle_multiplier = CYCLE_MULT_DEFAULT; // 100 for 1.0
492int cycle_multiplier_override;
493int cycle_multiplier_old;
494static int cycle_multiplier_active;
495
496static int CLOCK_ADJUST(int x)
497{
498 int m = cycle_multiplier_active;
499 int s = (x >> 31) | 1;
500 return (x * m + s * 50) / 100;
501}
502
503static int ds_writes_rjump_rs(int i)
504{
505 return dops[i].rs1 != 0 && (dops[i].rs1 == dops[i+1].rt1 || dops[i].rs1 == dops[i+1].rt2);
506}
507
508static u_int get_page(u_int vaddr)
509{
510 u_int page=vaddr&~0xe0000000;
511 if (page < 0x1000000)
512 page &= ~0x0e00000; // RAM mirrors
513 page>>=12;
514 if(page>2048) page=2048+(page&2047);
515 return page;
516}
517
518// no virtual mem in PCSX
519static u_int get_vpage(u_int vaddr)
520{
521 return get_page(vaddr);
522}
523
524static struct ht_entry *hash_table_get(u_int vaddr)
525{
526 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
527}
528
529static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
530{
531 ht_bin->vaddr[1] = ht_bin->vaddr[0];
532 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
533 ht_bin->vaddr[0] = vaddr;
534 ht_bin->tcaddr[0] = tcaddr;
535}
536
537// some messy ari64's code, seems to rely on unsigned 32bit overflow
538static int doesnt_expire_soon(void *tcaddr)
539{
540 u_int diff = (u_int)((u_char *)tcaddr - out) << (32-TARGET_SIZE_2);
541 return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
542}
543
544// Get address from virtual address
545// This is called from the recompiled JR/JALR instructions
546void noinline *get_addr(u_int vaddr)
547{
548 u_int page=get_page(vaddr);
549 u_int vpage=get_vpage(vaddr);
550 struct ll_entry *head;
551 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
552 head=jump_in[page];
553 while(head!=NULL) {
554 if(head->vaddr==vaddr) {
555 //printf("TRACE: count=%d next=%d (get_addr match %x: %p)\n",Count,next_interupt,vaddr,head->addr);
556 hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
557 return head->addr;
558 }
559 head=head->next;
560 }
561 head=jump_dirty[vpage];
562 while(head!=NULL) {
563 if(head->vaddr==vaddr) {
564 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %p)\n",Count,next_interupt,vaddr,head->addr);
565 // Don't restore blocks which are about to expire from the cache
566 if (doesnt_expire_soon(head->addr))
567 if (verify_dirty(head->addr)) {
568 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
569 invalid_code[vaddr>>12]=0;
570 inv_code_start=inv_code_end=~0;
571 if(vpage<2048) {
572 restore_candidate[vpage>>3]|=1<<(vpage&7);
573 }
574 else restore_candidate[page>>3]|=1<<(page&7);
575 struct ht_entry *ht_bin = hash_table_get(vaddr);
576 if (ht_bin->vaddr[0] == vaddr)
577 ht_bin->tcaddr[0] = head->addr; // Replace existing entry
578 else
579 hash_table_add(ht_bin, vaddr, head->addr);
580
581 return head->addr;
582 }
583 }
584 head=head->next;
585 }
586 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
587 int r=new_recompile_block(vaddr);
588 if(r==0) return get_addr(vaddr);
589 // generate an address error
590 Status|=2;
591 Cause=(vaddr<<31)|(4<<2);
592 EPC=(vaddr&1)?vaddr-5:vaddr;
593 BadVAddr=(vaddr&~1);
594 return get_addr_ht(0x80000080);
595}
596// Look up address in hash table first
597void *get_addr_ht(u_int vaddr)
598{
599 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
600 const struct ht_entry *ht_bin = hash_table_get(vaddr);
601 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
602 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
603 return get_addr(vaddr);
604}
605
606static void clear_all_regs(signed char regmap[])
607{
608 memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
609}
610
611#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
612
613extern signed char get_reg(const signed char regmap[], signed char r);
614
615#else
616
617static signed char get_reg(const signed char regmap[], signed char r)
618{
619 int hr;
620 for (hr = 0; hr < HOST_REGS; hr++) {
621 if (hr == EXCLUDE_REG)
622 continue;
623 if (regmap[hr] == r)
624 return hr;
625 }
626 return -1;
627}
628
629#endif
630
631static signed char get_reg_temp(const signed char regmap[])
632{
633 int hr;
634 for (hr = 0; hr < HOST_REGS; hr++) {
635 if (hr == EXCLUDE_REG)
636 continue;
637 if (regmap[hr] == (signed char)-1)
638 return hr;
639 }
640 return -1;
641}
642
643// Find a register that is available for two consecutive cycles
644static signed char get_reg2(signed char regmap1[], const signed char regmap2[], int r)
645{
646 int hr;
647 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&&regmap1[hr]==r&&regmap2[hr]==r) return hr;
648 return -1;
649}
650
651// reverse reg map: mips -> host
652#define RRMAP_SIZE 64
653static void make_rregs(const signed char regmap[], signed char rrmap[RRMAP_SIZE],
654 u_int *regs_can_change)
655{
656 u_int r, hr, hr_can_change = 0;
657 memset(rrmap, -1, RRMAP_SIZE);
658 for (hr = 0; hr < HOST_REGS; )
659 {
660 r = regmap[hr];
661 rrmap[r & (RRMAP_SIZE - 1)] = hr;
662 // only add mips $1-$31+$lo, others shifted out
663 hr_can_change |= (uint64_t)1 << (hr + ((r - 1) & 32));
664 hr++;
665 if (hr == EXCLUDE_REG)
666 hr++;
667 }
668 hr_can_change |= 1u << (rrmap[33] & 31);
669 hr_can_change |= 1u << (rrmap[CCREG] & 31);
670 hr_can_change &= ~(1u << 31);
671 *regs_can_change = hr_can_change;
672}
673
674// same as get_reg, but takes rrmap
675static signed char get_rreg(signed char rrmap[RRMAP_SIZE], signed char r)
676{
677 assert(0 <= r && r < RRMAP_SIZE);
678 return rrmap[r];
679}
680
681static int count_free_regs(const signed char regmap[])
682{
683 int count=0;
684 int hr;
685 for(hr=0;hr<HOST_REGS;hr++)
686 {
687 if(hr!=EXCLUDE_REG) {
688 if(regmap[hr]<0) count++;
689 }
690 }
691 return count;
692}
693
694static void dirty_reg(struct regstat *cur, signed char reg)
695{
696 int hr;
697 if (!reg) return;
698 hr = get_reg(cur->regmap, reg);
699 if (hr >= 0)
700 cur->dirty |= 1<<hr;
701}
702
703static void set_const(struct regstat *cur, signed char reg, uint32_t value)
704{
705 int hr;
706 if (!reg) return;
707 hr = get_reg(cur->regmap, reg);
708 if (hr >= 0) {
709 cur->isconst |= 1<<hr;
710 current_constmap[hr] = value;
711 }
712}
713
714static void clear_const(struct regstat *cur, signed char reg)
715{
716 int hr;
717 if (!reg) return;
718 hr = get_reg(cur->regmap, reg);
719 if (hr >= 0)
720 cur->isconst &= ~(1<<hr);
721}
722
723static int is_const(const struct regstat *cur, signed char reg)
724{
725 int hr;
726 if (reg < 0) return 0;
727 if (!reg) return 1;
728 hr = get_reg(cur->regmap, reg);
729 if (hr >= 0)
730 return (cur->isconst>>hr)&1;
731 return 0;
732}
733
734static uint32_t get_const(const struct regstat *cur, signed char reg)
735{
736 int hr;
737 if (!reg) return 0;
738 hr = get_reg(cur->regmap, reg);
739 if (hr >= 0)
740 return current_constmap[hr];
741
742 SysPrintf("Unknown constant in r%d\n", reg);
743 abort();
744}
745
746// Least soon needed registers
747// Look at the next ten instructions and see which registers
748// will be used. Try not to reallocate these.
749static void lsn(u_char hsn[], int i, int *preferred_reg)
750{
751 int j;
752 int b=-1;
753 for(j=0;j<9;j++)
754 {
755 if(i+j>=slen) {
756 j=slen-i-1;
757 break;
758 }
759 if (dops[i+j].is_ujump)
760 {
761 // Don't go past an unconditonal jump
762 j++;
763 break;
764 }
765 }
766 for(;j>=0;j--)
767 {
768 if(dops[i+j].rs1) hsn[dops[i+j].rs1]=j;
769 if(dops[i+j].rs2) hsn[dops[i+j].rs2]=j;
770 if(dops[i+j].rt1) hsn[dops[i+j].rt1]=j;
771 if(dops[i+j].rt2) hsn[dops[i+j].rt2]=j;
772 if(dops[i+j].itype==STORE || dops[i+j].itype==STORELR) {
773 // Stores can allocate zero
774 hsn[dops[i+j].rs1]=j;
775 hsn[dops[i+j].rs2]=j;
776 }
777 if (ram_offset && (dops[i+j].is_load || dops[i+j].is_store))
778 hsn[ROREG] = j;
779 // On some architectures stores need invc_ptr
780 #if defined(HOST_IMM8)
781 if (dops[i+j].is_store)
782 hsn[INVCP] = j;
783 #endif
784 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
785 {
786 hsn[CCREG]=j;
787 b=j;
788 }
789 }
790 if(b>=0)
791 {
792 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
793 {
794 // Follow first branch
795 int t=(ba[i+b]-start)>>2;
796 j=7-b;if(t+j>=slen) j=slen-t-1;
797 for(;j>=0;j--)
798 {
799 if(dops[t+j].rs1) if(hsn[dops[t+j].rs1]>j+b+2) hsn[dops[t+j].rs1]=j+b+2;
800 if(dops[t+j].rs2) if(hsn[dops[t+j].rs2]>j+b+2) hsn[dops[t+j].rs2]=j+b+2;
801 //if(dops[t+j].rt1) if(hsn[dops[t+j].rt1]>j+b+2) hsn[dops[t+j].rt1]=j+b+2;
802 //if(dops[t+j].rt2) if(hsn[dops[t+j].rt2]>j+b+2) hsn[dops[t+j].rt2]=j+b+2;
803 }
804 }
805 // TODO: preferred register based on backward branch
806 }
807 // Delay slot should preferably not overwrite branch conditions or cycle count
808 if (i > 0 && dops[i-1].is_jump) {
809 if(dops[i-1].rs1) if(hsn[dops[i-1].rs1]>1) hsn[dops[i-1].rs1]=1;
810 if(dops[i-1].rs2) if(hsn[dops[i-1].rs2]>1) hsn[dops[i-1].rs2]=1;
811 hsn[CCREG]=1;
812 // ...or hash tables
813 hsn[RHASH]=1;
814 hsn[RHTBL]=1;
815 }
816 // Coprocessor load/store needs FTEMP, even if not declared
817 if(dops[i].itype==C2LS) {
818 hsn[FTEMP]=0;
819 }
820 // Load L/R also uses FTEMP as a temporary register
821 if(dops[i].itype==LOADLR) {
822 hsn[FTEMP]=0;
823 }
824 // Also SWL/SWR/SDL/SDR
825 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) {
826 hsn[FTEMP]=0;
827 }
828 // Don't remove the miniht registers
829 if(dops[i].itype==UJUMP||dops[i].itype==RJUMP)
830 {
831 hsn[RHASH]=0;
832 hsn[RHTBL]=0;
833 }
834}
835
836// We only want to allocate registers if we're going to use them again soon
837static int needed_again(int r, int i)
838{
839 int j;
840 int b=-1;
841 int rn=10;
842
843 if (i > 0 && dops[i-1].is_ujump)
844 {
845 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
846 return 0; // Don't need any registers if exiting the block
847 }
848 for(j=0;j<9;j++)
849 {
850 if(i+j>=slen) {
851 j=slen-i-1;
852 break;
853 }
854 if (dops[i+j].is_ujump)
855 {
856 // Don't go past an unconditonal jump
857 j++;
858 break;
859 }
860 if(dops[i+j].itype==SYSCALL||dops[i+j].itype==HLECALL||dops[i+j].itype==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
861 {
862 break;
863 }
864 }
865 for(;j>=1;j--)
866 {
867 if(dops[i+j].rs1==r) rn=j;
868 if(dops[i+j].rs2==r) rn=j;
869 if((unneeded_reg[i+j]>>r)&1) rn=10;
870 if(i+j>=0&&(dops[i+j].itype==UJUMP||dops[i+j].itype==CJUMP||dops[i+j].itype==SJUMP))
871 {
872 b=j;
873 }
874 }
875 if(rn<10) return 1;
876 (void)b;
877 return 0;
878}
879
880// Try to match register allocations at the end of a loop with those
881// at the beginning
882static int loop_reg(int i, int r, int hr)
883{
884 int j,k;
885 for(j=0;j<9;j++)
886 {
887 if(i+j>=slen) {
888 j=slen-i-1;
889 break;
890 }
891 if (dops[i+j].is_ujump)
892 {
893 // Don't go past an unconditonal jump
894 j++;
895 break;
896 }
897 }
898 k=0;
899 if(i>0){
900 if(dops[i-1].itype==UJUMP||dops[i-1].itype==CJUMP||dops[i-1].itype==SJUMP)
901 k--;
902 }
903 for(;k<j;k++)
904 {
905 assert(r < 64);
906 if((unneeded_reg[i+k]>>r)&1) return hr;
907 if(i+k>=0&&(dops[i+k].itype==UJUMP||dops[i+k].itype==CJUMP||dops[i+k].itype==SJUMP))
908 {
909 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
910 {
911 int t=(ba[i+k]-start)>>2;
912 int reg=get_reg(regs[t].regmap_entry,r);
913 if(reg>=0) return reg;
914 //reg=get_reg(regs[t+1].regmap_entry,r);
915 //if(reg>=0) return reg;
916 }
917 }
918 }
919 return hr;
920}
921
922
923// Allocate every register, preserving source/target regs
924static void alloc_all(struct regstat *cur,int i)
925{
926 int hr;
927
928 for(hr=0;hr<HOST_REGS;hr++) {
929 if(hr!=EXCLUDE_REG) {
930 if((cur->regmap[hr]!=dops[i].rs1)&&(cur->regmap[hr]!=dops[i].rs2)&&
931 (cur->regmap[hr]!=dops[i].rt1)&&(cur->regmap[hr]!=dops[i].rt2))
932 {
933 cur->regmap[hr]=-1;
934 cur->dirty&=~(1<<hr);
935 }
936 // Don't need zeros
937 if(cur->regmap[hr]==0)
938 {
939 cur->regmap[hr]=-1;
940 cur->dirty&=~(1<<hr);
941 }
942 }
943 }
944}
945
946#ifndef NDEBUG
947static int host_tempreg_in_use;
948
949static void host_tempreg_acquire(void)
950{
951 assert(!host_tempreg_in_use);
952 host_tempreg_in_use = 1;
953}
954
955static void host_tempreg_release(void)
956{
957 host_tempreg_in_use = 0;
958}
959#else
960static void host_tempreg_acquire(void) {}
961static void host_tempreg_release(void) {}
962#endif
963
964#ifdef ASSEM_PRINT
965extern void gen_interupt();
966extern void do_insn_cmp();
967#define FUNCNAME(f) { f, " " #f }
968static const struct {
969 void *addr;
970 const char *name;
971} function_names[] = {
972 FUNCNAME(cc_interrupt),
973 FUNCNAME(gen_interupt),
974 FUNCNAME(get_addr_ht),
975 FUNCNAME(get_addr),
976 FUNCNAME(jump_handler_read8),
977 FUNCNAME(jump_handler_read16),
978 FUNCNAME(jump_handler_read32),
979 FUNCNAME(jump_handler_write8),
980 FUNCNAME(jump_handler_write16),
981 FUNCNAME(jump_handler_write32),
982 FUNCNAME(invalidate_addr),
983 FUNCNAME(jump_to_new_pc),
984 FUNCNAME(jump_break),
985 FUNCNAME(jump_break_ds),
986 FUNCNAME(jump_syscall),
987 FUNCNAME(jump_syscall_ds),
988 FUNCNAME(call_gteStall),
989 FUNCNAME(clean_blocks),
990 FUNCNAME(new_dyna_leave),
991 FUNCNAME(pcsx_mtc0),
992 FUNCNAME(pcsx_mtc0_ds),
993#ifdef DRC_DBG
994 FUNCNAME(do_insn_cmp),
995#endif
996#ifdef __arm__
997 FUNCNAME(verify_code),
998#endif
999};
1000
1001static const char *func_name(const void *a)
1002{
1003 int i;
1004 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
1005 if (function_names[i].addr == a)
1006 return function_names[i].name;
1007 return "";
1008}
1009#else
1010#define func_name(x) ""
1011#endif
1012
1013#ifdef __i386__
1014#include "assem_x86.c"
1015#endif
1016#ifdef __x86_64__
1017#include "assem_x64.c"
1018#endif
1019#ifdef __arm__
1020#include "assem_arm.c"
1021#endif
1022#ifdef __aarch64__
1023#include "assem_arm64.c"
1024#endif
1025
1026static void *get_trampoline(const void *f)
1027{
1028 size_t i;
1029
1030 for (i = 0; i < ARRAY_SIZE(ndrc->tramp.f); i++) {
1031 if (ndrc->tramp.f[i] == f || ndrc->tramp.f[i] == NULL)
1032 break;
1033 }
1034 if (i == ARRAY_SIZE(ndrc->tramp.f)) {
1035 SysPrintf("trampoline table is full, last func %p\n", f);
1036 abort();
1037 }
1038 if (ndrc->tramp.f[i] == NULL) {
1039 start_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
1040 ndrc->tramp.f[i] = f;
1041 end_tcache_write(&ndrc->tramp.f[i], &ndrc->tramp.f[i + 1]);
1042 }
1043 return &ndrc->tramp.ops[i];
1044}
1045
1046static void emit_far_jump(const void *f)
1047{
1048 if (can_jump_or_call(f)) {
1049 emit_jmp(f);
1050 return;
1051 }
1052
1053 f = get_trampoline(f);
1054 emit_jmp(f);
1055}
1056
1057static void emit_far_call(const void *f)
1058{
1059 if (can_jump_or_call(f)) {
1060 emit_call(f);
1061 return;
1062 }
1063
1064 f = get_trampoline(f);
1065 emit_call(f);
1066}
1067
1068// Add virtual address mapping to linked list
1069static void ll_add(struct ll_entry **head,int vaddr,void *addr)
1070{
1071 struct ll_entry *new_entry;
1072 new_entry=malloc(sizeof(struct ll_entry));
1073 assert(new_entry!=NULL);
1074 new_entry->vaddr=vaddr;
1075 new_entry->reg_sv_flags=0;
1076 new_entry->addr=addr;
1077 new_entry->next=*head;
1078 *head=new_entry;
1079}
1080
1081static void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
1082{
1083 ll_add(head,vaddr,addr);
1084 (*head)->reg_sv_flags=reg_sv_flags;
1085}
1086
1087// Check if an address is already compiled
1088// but don't return addresses which are about to expire from the cache
1089static void *check_addr(u_int vaddr)
1090{
1091 struct ht_entry *ht_bin = hash_table_get(vaddr);
1092 size_t i;
1093 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
1094 if (ht_bin->vaddr[i] == vaddr)
1095 if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
1096 if (isclean(ht_bin->tcaddr[i]))
1097 return ht_bin->tcaddr[i];
1098 }
1099 u_int page=get_page(vaddr);
1100 struct ll_entry *head;
1101 head=jump_in[page];
1102 while (head != NULL) {
1103 if (head->vaddr == vaddr) {
1104 if (doesnt_expire_soon(head->addr)) {
1105 // Update existing entry with current address
1106 if (ht_bin->vaddr[0] == vaddr) {
1107 ht_bin->tcaddr[0] = head->addr;
1108 return head->addr;
1109 }
1110 if (ht_bin->vaddr[1] == vaddr) {
1111 ht_bin->tcaddr[1] = head->addr;
1112 return head->addr;
1113 }
1114 // Insert into hash table with low priority.
1115 // Don't evict existing entries, as they are probably
1116 // addresses that are being accessed frequently.
1117 if (ht_bin->vaddr[0] == -1) {
1118 ht_bin->vaddr[0] = vaddr;
1119 ht_bin->tcaddr[0] = head->addr;
1120 }
1121 else if (ht_bin->vaddr[1] == -1) {
1122 ht_bin->vaddr[1] = vaddr;
1123 ht_bin->tcaddr[1] = head->addr;
1124 }
1125 return head->addr;
1126 }
1127 }
1128 head=head->next;
1129 }
1130 return 0;
1131}
1132
1133void remove_hash(int vaddr)
1134{
1135 //printf("remove hash: %x\n",vaddr);
1136 struct ht_entry *ht_bin = hash_table_get(vaddr);
1137 if (ht_bin->vaddr[1] == vaddr) {
1138 ht_bin->vaddr[1] = -1;
1139 ht_bin->tcaddr[1] = NULL;
1140 }
1141 if (ht_bin->vaddr[0] == vaddr) {
1142 ht_bin->vaddr[0] = ht_bin->vaddr[1];
1143 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
1144 ht_bin->vaddr[1] = -1;
1145 ht_bin->tcaddr[1] = NULL;
1146 }
1147}
1148
1149static void ll_remove_matching_addrs(struct ll_entry **head,
1150 uintptr_t base_offs_s, int shift)
1151{
1152 struct ll_entry *next;
1153 while(*head) {
1154 uintptr_t o1 = (u_char *)(*head)->addr - ndrc->translation_cache;
1155 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1156 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1157 {
1158 inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
1159 remove_hash((*head)->vaddr);
1160 next=(*head)->next;
1161 free(*head);
1162 *head=next;
1163 }
1164 else
1165 {
1166 head=&((*head)->next);
1167 }
1168 }
1169}
1170
1171// Remove all entries from linked list
1172static void ll_clear(struct ll_entry **head)
1173{
1174 struct ll_entry *cur;
1175 struct ll_entry *next;
1176 if((cur=*head)) {
1177 *head=0;
1178 while(cur) {
1179 next=cur->next;
1180 free(cur);
1181 cur=next;
1182 }
1183 }
1184}
1185
1186// Dereference the pointers and remove if it matches
1187static void ll_kill_pointers(struct ll_entry *head,
1188 uintptr_t base_offs_s, int shift)
1189{
1190 while(head) {
1191 u_char *ptr = get_pointer(head->addr);
1192 uintptr_t o1 = ptr - ndrc->translation_cache;
1193 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
1194 inv_debug("EXP: Lookup pointer to %p at %p (%x)\n",ptr,head->addr,head->vaddr);
1195 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s)
1196 {
1197 inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
1198 void *host_addr=find_extjump_insn(head->addr);
1199 mark_clear_cache(host_addr);
1200 set_jump_target(host_addr, head->addr);
1201 }
1202 head=head->next;
1203 }
1204}
1205
1206// This is called when we write to a compiled block (see do_invstub)
1207static void invalidate_page(u_int page)
1208{
1209 struct ll_entry *head;
1210 struct ll_entry *next;
1211 head=jump_in[page];
1212 jump_in[page]=0;
1213 while(head!=NULL) {
1214 inv_debug("INVALIDATE: %x\n",head->vaddr);
1215 remove_hash(head->vaddr);
1216 next=head->next;
1217 free(head);
1218 head=next;
1219 }
1220 head=jump_out[page];
1221 jump_out[page]=0;
1222 while(head!=NULL) {
1223 inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
1224 void *host_addr=find_extjump_insn(head->addr);
1225 mark_clear_cache(host_addr);
1226 set_jump_target(host_addr, head->addr); // point back to dyna_linker
1227 next=head->next;
1228 free(head);
1229 head=next;
1230 }
1231}
1232
1233static void invalidate_block_range(u_int block, u_int first, u_int last)
1234{
1235 u_int page=get_page(block<<12);
1236 //printf("first=%d last=%d\n",first,last);
1237 invalidate_page(page);
1238 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1239 assert(last<page+5);
1240 // Invalidate the adjacent pages if a block crosses a 4K boundary
1241 while(first<page) {
1242 invalidate_page(first);
1243 first++;
1244 }
1245 for(first=page+1;first<last;first++) {
1246 invalidate_page(first);
1247 }
1248 do_clear_cache();
1249
1250 // Don't trap writes
1251 invalid_code[block]=1;
1252
1253 #ifdef USE_MINI_HT
1254 memset(mini_ht,-1,sizeof(mini_ht));
1255 #endif
1256}
1257
1258void invalidate_block(u_int block)
1259{
1260 u_int page=get_page(block<<12);
1261 u_int vpage=get_vpage(block<<12);
1262 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1263 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1264 u_int first,last;
1265 first=last=page;
1266 struct ll_entry *head;
1267 head=jump_dirty[vpage];
1268 //printf("page=%d vpage=%d\n",page,vpage);
1269 while(head!=NULL) {
1270 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1271 u_char *start, *end;
1272 get_bounds(head->addr, &start, &end);
1273 //printf("start: %p end: %p\n", start, end);
1274 if (page < 2048 && start >= rdram && end < rdram+RAM_SIZE) {
1275 if (((start-rdram)>>12) <= page && ((end-1-rdram)>>12) >= page) {
1276 if ((((start-rdram)>>12)&2047) < first) first = ((start-rdram)>>12)&2047;
1277 if ((((end-1-rdram)>>12)&2047) > last) last = ((end-1-rdram)>>12)&2047;
1278 }
1279 }
1280 }
1281 head=head->next;
1282 }
1283 invalidate_block_range(block,first,last);
1284}
1285
1286void invalidate_addr(u_int addr)
1287{
1288 //static int rhits;
1289 // this check is done by the caller
1290 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1291 u_int page=get_vpage(addr);
1292 if(page<2048) { // RAM
1293 struct ll_entry *head;
1294 u_int addr_min=~0, addr_max=0;
1295 u_int mask=RAM_SIZE-1;
1296 u_int addr_main=0x80000000|(addr&mask);
1297 int pg1;
1298 inv_code_start=addr_main&~0xfff;
1299 inv_code_end=addr_main|0xfff;
1300 pg1=page;
1301 if (pg1>0) {
1302 // must check previous page too because of spans..
1303 pg1--;
1304 inv_code_start-=0x1000;
1305 }
1306 for(;pg1<=page;pg1++) {
1307 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1308 u_char *start_h, *end_h;
1309 u_int start, end;
1310 get_bounds(head->addr, &start_h, &end_h);
1311 start = (uintptr_t)start_h - ram_offset;
1312 end = (uintptr_t)end_h - ram_offset;
1313 if(start<=addr_main&&addr_main<end) {
1314 if(start<addr_min) addr_min=start;
1315 if(end>addr_max) addr_max=end;
1316 }
1317 else if(addr_main<start) {
1318 if(start<inv_code_end)
1319 inv_code_end=start-1;
1320 }
1321 else {
1322 if(end>inv_code_start)
1323 inv_code_start=end;
1324 }
1325 }
1326 }
1327 if (addr_min!=~0) {
1328 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1329 inv_code_start=inv_code_end=~0;
1330 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1331 return;
1332 }
1333 else {
1334 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1335 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1336 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1337 return;
1338 }
1339 }
1340 invalidate_block(addr>>12);
1341}
1342
1343// This is called when loading a save state.
1344// Anything could have changed, so invalidate everything.
1345void invalidate_all_pages(void)
1346{
1347 u_int page;
1348 for(page=0;page<4096;page++)
1349 invalidate_page(page);
1350 for(page=0;page<1048576;page++)
1351 if(!invalid_code[page]) {
1352 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1353 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1354 }
1355 #ifdef USE_MINI_HT
1356 memset(mini_ht,-1,sizeof(mini_ht));
1357 #endif
1358 do_clear_cache();
1359}
1360
1361static void do_invstub(int n)
1362{
1363 literal_pool(20);
1364 u_int reglist=stubs[n].a;
1365 set_jump_target(stubs[n].addr, out);
1366 save_regs(reglist);
1367 if(stubs[n].b!=0) emit_mov(stubs[n].b,0);
1368 emit_far_call(invalidate_addr);
1369 restore_regs(reglist);
1370 emit_jmp(stubs[n].retaddr); // return address
1371}
1372
1373// Add an entry to jump_out after making a link
1374// src should point to code by emit_extjump2()
1375void add_jump_out(u_int vaddr,void *src)
1376{
1377 u_int page=get_page(vaddr);
1378 inv_debug("add_jump_out: %p -> %x (%d)\n",src,vaddr,page);
1379 check_extjump2(src);
1380 ll_add(jump_out+page,vaddr,src);
1381 //inv_debug("add_jump_out: to %p\n",get_pointer(src));
1382}
1383
1384// If a code block was found to be unmodified (bit was set in
1385// restore_candidate) and it remains unmodified (bit is clear
1386// in invalid_code) then move the entries for that 4K page from
1387// the dirty list to the clean list.
1388void clean_blocks(u_int page)
1389{
1390 struct ll_entry *head;
1391 inv_debug("INV: clean_blocks page=%d\n",page);
1392 head=jump_dirty[page];
1393 while(head!=NULL) {
1394 if(!invalid_code[head->vaddr>>12]) {
1395 // Don't restore blocks which are about to expire from the cache
1396 if (doesnt_expire_soon(head->addr)) {
1397 if(verify_dirty(head->addr)) {
1398 u_char *start, *end;
1399 //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
1400 u_int i;
1401 u_int inv=0;
1402 get_bounds(head->addr, &start, &end);
1403 if (start - rdram < RAM_SIZE) {
1404 for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
1405 inv|=invalid_code[i];
1406 }
1407 }
1408 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1409 inv=1;
1410 }
1411 if(!inv) {
1412 void *clean_addr = get_clean_addr(head->addr);
1413 if (doesnt_expire_soon(clean_addr)) {
1414 u_int ppage=page;
1415 inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, clean_addr);
1416 //printf("page=%x, addr=%x\n",page,head->vaddr);
1417 //assert(head->vaddr>>12==(page|0x80000));
1418 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1419 struct ht_entry *ht_bin = hash_table_get(head->vaddr);
1420 if (ht_bin->vaddr[0] == head->vaddr)
1421 ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
1422 if (ht_bin->vaddr[1] == head->vaddr)
1423 ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
1424 }
1425 }
1426 }
1427 }
1428 }
1429 head=head->next;
1430 }
1431}
1432
1433/* Register allocation */
1434
1435// Note: registers are allocated clean (unmodified state)
1436// if you intend to modify the register, you must call dirty_reg().
1437static void alloc_reg(struct regstat *cur,int i,signed char reg)
1438{
1439 int r,hr;
1440 int preferred_reg = PREFERRED_REG_FIRST
1441 + reg % (PREFERRED_REG_LAST - PREFERRED_REG_FIRST + 1);
1442 if (reg == CCREG) preferred_reg = HOST_CCREG;
1443 if (reg == PTEMP || reg == FTEMP) preferred_reg = 12;
1444 assert(PREFERRED_REG_FIRST != EXCLUDE_REG && EXCLUDE_REG != HOST_REGS);
1445
1446 // Don't allocate unused registers
1447 if((cur->u>>reg)&1) return;
1448
1449 // see if it's already allocated
1450 for(hr=0;hr<HOST_REGS;hr++)
1451 {
1452 if(cur->regmap[hr]==reg) return;
1453 }
1454
1455 // Keep the same mapping if the register was already allocated in a loop
1456 preferred_reg = loop_reg(i,reg,preferred_reg);
1457
1458 // Try to allocate the preferred register
1459 if(cur->regmap[preferred_reg]==-1) {
1460 cur->regmap[preferred_reg]=reg;
1461 cur->dirty&=~(1<<preferred_reg);
1462 cur->isconst&=~(1<<preferred_reg);
1463 return;
1464 }
1465 r=cur->regmap[preferred_reg];
1466 assert(r < 64);
1467 if((cur->u>>r)&1) {
1468 cur->regmap[preferred_reg]=reg;
1469 cur->dirty&=~(1<<preferred_reg);
1470 cur->isconst&=~(1<<preferred_reg);
1471 return;
1472 }
1473
1474 // Clear any unneeded registers
1475 // We try to keep the mapping consistent, if possible, because it
1476 // makes branches easier (especially loops). So we try to allocate
1477 // first (see above) before removing old mappings. If this is not
1478 // possible then go ahead and clear out the registers that are no
1479 // longer needed.
1480 for(hr=0;hr<HOST_REGS;hr++)
1481 {
1482 r=cur->regmap[hr];
1483 if(r>=0) {
1484 assert(r < 64);
1485 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1486 }
1487 }
1488
1489 // Try to allocate any available register, but prefer
1490 // registers that have not been used recently.
1491 if (i > 0) {
1492 for (hr = PREFERRED_REG_FIRST; ; ) {
1493 if (cur->regmap[hr] < 0) {
1494 int oldreg = regs[i-1].regmap[hr];
1495 if (oldreg < 0 || (oldreg != dops[i-1].rs1 && oldreg != dops[i-1].rs2
1496 && oldreg != dops[i-1].rt1 && oldreg != dops[i-1].rt2))
1497 {
1498 cur->regmap[hr]=reg;
1499 cur->dirty&=~(1<<hr);
1500 cur->isconst&=~(1<<hr);
1501 return;
1502 }
1503 }
1504 hr++;
1505 if (hr == EXCLUDE_REG)
1506 hr++;
1507 if (hr == HOST_REGS)
1508 hr = 0;
1509 if (hr == PREFERRED_REG_FIRST)
1510 break;
1511 }
1512 }
1513
1514 // Try to allocate any available register
1515 for (hr = PREFERRED_REG_FIRST; ; ) {
1516 if (cur->regmap[hr] < 0) {
1517 cur->regmap[hr]=reg;
1518 cur->dirty&=~(1<<hr);
1519 cur->isconst&=~(1<<hr);
1520 return;
1521 }
1522 hr++;
1523 if (hr == EXCLUDE_REG)
1524 hr++;
1525 if (hr == HOST_REGS)
1526 hr = 0;
1527 if (hr == PREFERRED_REG_FIRST)
1528 break;
1529 }
1530
1531 // Ok, now we have to evict someone
1532 // Pick a register we hopefully won't need soon
1533 u_char hsn[MAXREG+1];
1534 memset(hsn,10,sizeof(hsn));
1535 int j;
1536 lsn(hsn,i,&preferred_reg);
1537 //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]);
1538 //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]);
1539 if(i>0) {
1540 // Don't evict the cycle count at entry points, otherwise the entry
1541 // stub will have to write it.
1542 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1543 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1544 for(j=10;j>=3;j--)
1545 {
1546 // Alloc preferred register if available
1547 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1548 for(hr=0;hr<HOST_REGS;hr++) {
1549 // Evict both parts of a 64-bit register
1550 if(cur->regmap[hr]==r) {
1551 cur->regmap[hr]=-1;
1552 cur->dirty&=~(1<<hr);
1553 cur->isconst&=~(1<<hr);
1554 }
1555 }
1556 cur->regmap[preferred_reg]=reg;
1557 return;
1558 }
1559 for(r=1;r<=MAXREG;r++)
1560 {
1561 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1562 for(hr=0;hr<HOST_REGS;hr++) {
1563 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1564 if(cur->regmap[hr]==r) {
1565 cur->regmap[hr]=reg;
1566 cur->dirty&=~(1<<hr);
1567 cur->isconst&=~(1<<hr);
1568 return;
1569 }
1570 }
1571 }
1572 }
1573 }
1574 }
1575 }
1576 for(j=10;j>=0;j--)
1577 {
1578 for(r=1;r<=MAXREG;r++)
1579 {
1580 if(hsn[r]==j) {
1581 for(hr=0;hr<HOST_REGS;hr++) {
1582 if(cur->regmap[hr]==r) {
1583 cur->regmap[hr]=reg;
1584 cur->dirty&=~(1<<hr);
1585 cur->isconst&=~(1<<hr);
1586 return;
1587 }
1588 }
1589 }
1590 }
1591 }
1592 SysPrintf("This shouldn't happen (alloc_reg)");abort();
1593}
1594
1595// Allocate a temporary register. This is done without regard to
1596// dirty status or whether the register we request is on the unneeded list
1597// Note: This will only allocate one register, even if called multiple times
1598static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1599{
1600 int r,hr;
1601 int preferred_reg = -1;
1602
1603 // see if it's already allocated
1604 for(hr=0;hr<HOST_REGS;hr++)
1605 {
1606 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1607 }
1608
1609 // Try to allocate any available register
1610 for(hr=HOST_REGS-1;hr>=0;hr--) {
1611 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1612 cur->regmap[hr]=reg;
1613 cur->dirty&=~(1<<hr);
1614 cur->isconst&=~(1<<hr);
1615 return;
1616 }
1617 }
1618
1619 // Find an unneeded register
1620 for(hr=HOST_REGS-1;hr>=0;hr--)
1621 {
1622 r=cur->regmap[hr];
1623 if(r>=0) {
1624 assert(r < 64);
1625 if((cur->u>>r)&1) {
1626 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1627 cur->regmap[hr]=reg;
1628 cur->dirty&=~(1<<hr);
1629 cur->isconst&=~(1<<hr);
1630 return;
1631 }
1632 }
1633 }
1634 }
1635
1636 // Ok, now we have to evict someone
1637 // Pick a register we hopefully won't need soon
1638 // TODO: we might want to follow unconditional jumps here
1639 // TODO: get rid of dupe code and make this into a function
1640 u_char hsn[MAXREG+1];
1641 memset(hsn,10,sizeof(hsn));
1642 int j;
1643 lsn(hsn,i,&preferred_reg);
1644 //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]);
1645 if(i>0) {
1646 // Don't evict the cycle count at entry points, otherwise the entry
1647 // stub will have to write it.
1648 if(dops[i].bt&&hsn[CCREG]>2) hsn[CCREG]=2;
1649 if (i>1 && hsn[CCREG] > 2 && dops[i-2].is_jump) hsn[CCREG]=2;
1650 for(j=10;j>=3;j--)
1651 {
1652 for(r=1;r<=MAXREG;r++)
1653 {
1654 if(hsn[r]==j&&r!=dops[i-1].rs1&&r!=dops[i-1].rs2&&r!=dops[i-1].rt1&&r!=dops[i-1].rt2) {
1655 for(hr=0;hr<HOST_REGS;hr++) {
1656 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1657 if(cur->regmap[hr]==r) {
1658 cur->regmap[hr]=reg;
1659 cur->dirty&=~(1<<hr);
1660 cur->isconst&=~(1<<hr);
1661 return;
1662 }
1663 }
1664 }
1665 }
1666 }
1667 }
1668 }
1669 for(j=10;j>=0;j--)
1670 {
1671 for(r=1;r<=MAXREG;r++)
1672 {
1673 if(hsn[r]==j) {
1674 for(hr=0;hr<HOST_REGS;hr++) {
1675 if(cur->regmap[hr]==r) {
1676 cur->regmap[hr]=reg;
1677 cur->dirty&=~(1<<hr);
1678 cur->isconst&=~(1<<hr);
1679 return;
1680 }
1681 }
1682 }
1683 }
1684 }
1685 SysPrintf("This shouldn't happen");abort();
1686}
1687
1688static void mov_alloc(struct regstat *current,int i)
1689{
1690 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) {
1691 alloc_cc(current,i); // for stalls
1692 dirty_reg(current,CCREG);
1693 }
1694
1695 // Note: Don't need to actually alloc the source registers
1696 //alloc_reg(current,i,dops[i].rs1);
1697 alloc_reg(current,i,dops[i].rt1);
1698
1699 clear_const(current,dops[i].rs1);
1700 clear_const(current,dops[i].rt1);
1701 dirty_reg(current,dops[i].rt1);
1702}
1703
1704static void shiftimm_alloc(struct regstat *current,int i)
1705{
1706 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
1707 {
1708 if(dops[i].rt1) {
1709 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1710 else dops[i].use_lt1=!!dops[i].rs1;
1711 alloc_reg(current,i,dops[i].rt1);
1712 dirty_reg(current,dops[i].rt1);
1713 if(is_const(current,dops[i].rs1)) {
1714 int v=get_const(current,dops[i].rs1);
1715 if(dops[i].opcode2==0x00) set_const(current,dops[i].rt1,v<<imm[i]);
1716 if(dops[i].opcode2==0x02) set_const(current,dops[i].rt1,(u_int)v>>imm[i]);
1717 if(dops[i].opcode2==0x03) set_const(current,dops[i].rt1,v>>imm[i]);
1718 }
1719 else clear_const(current,dops[i].rt1);
1720 }
1721 }
1722 else
1723 {
1724 clear_const(current,dops[i].rs1);
1725 clear_const(current,dops[i].rt1);
1726 }
1727
1728 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
1729 {
1730 assert(0);
1731 }
1732 if(dops[i].opcode2==0x3c) // DSLL32
1733 {
1734 assert(0);
1735 }
1736 if(dops[i].opcode2==0x3e) // DSRL32
1737 {
1738 assert(0);
1739 }
1740 if(dops[i].opcode2==0x3f) // DSRA32
1741 {
1742 assert(0);
1743 }
1744}
1745
1746static void shift_alloc(struct regstat *current,int i)
1747{
1748 if(dops[i].rt1) {
1749 if(dops[i].opcode2<=0x07) // SLLV/SRLV/SRAV
1750 {
1751 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
1752 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
1753 alloc_reg(current,i,dops[i].rt1);
1754 if(dops[i].rt1==dops[i].rs2) {
1755 alloc_reg_temp(current,i,-1);
1756 minimum_free_regs[i]=1;
1757 }
1758 } else { // DSLLV/DSRLV/DSRAV
1759 assert(0);
1760 }
1761 clear_const(current,dops[i].rs1);
1762 clear_const(current,dops[i].rs2);
1763 clear_const(current,dops[i].rt1);
1764 dirty_reg(current,dops[i].rt1);
1765 }
1766}
1767
1768static void alu_alloc(struct regstat *current,int i)
1769{
1770 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
1771 if(dops[i].rt1) {
1772 if(dops[i].rs1&&dops[i].rs2) {
1773 alloc_reg(current,i,dops[i].rs1);
1774 alloc_reg(current,i,dops[i].rs2);
1775 }
1776 else {
1777 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1778 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1779 }
1780 alloc_reg(current,i,dops[i].rt1);
1781 }
1782 }
1783 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
1784 if(dops[i].rt1) {
1785 alloc_reg(current,i,dops[i].rs1);
1786 alloc_reg(current,i,dops[i].rs2);
1787 alloc_reg(current,i,dops[i].rt1);
1788 }
1789 }
1790 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
1791 if(dops[i].rt1) {
1792 if(dops[i].rs1&&dops[i].rs2) {
1793 alloc_reg(current,i,dops[i].rs1);
1794 alloc_reg(current,i,dops[i].rs2);
1795 }
1796 else
1797 {
1798 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1799 if(dops[i].rs2&&needed_again(dops[i].rs2,i)) alloc_reg(current,i,dops[i].rs2);
1800 }
1801 alloc_reg(current,i,dops[i].rt1);
1802 }
1803 }
1804 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1805 assert(0);
1806 }
1807 clear_const(current,dops[i].rs1);
1808 clear_const(current,dops[i].rs2);
1809 clear_const(current,dops[i].rt1);
1810 dirty_reg(current,dops[i].rt1);
1811}
1812
1813static void imm16_alloc(struct regstat *current,int i)
1814{
1815 if(dops[i].rs1&&needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1816 else dops[i].use_lt1=!!dops[i].rs1;
1817 if(dops[i].rt1) alloc_reg(current,i,dops[i].rt1);
1818 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
1819 assert(0);
1820 }
1821 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
1822 clear_const(current,dops[i].rs1);
1823 clear_const(current,dops[i].rt1);
1824 }
1825 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
1826 if(is_const(current,dops[i].rs1)) {
1827 int v=get_const(current,dops[i].rs1);
1828 if(dops[i].opcode==0x0c) set_const(current,dops[i].rt1,v&imm[i]);
1829 if(dops[i].opcode==0x0d) set_const(current,dops[i].rt1,v|imm[i]);
1830 if(dops[i].opcode==0x0e) set_const(current,dops[i].rt1,v^imm[i]);
1831 }
1832 else clear_const(current,dops[i].rt1);
1833 }
1834 else if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
1835 if(is_const(current,dops[i].rs1)) {
1836 int v=get_const(current,dops[i].rs1);
1837 set_const(current,dops[i].rt1,v+imm[i]);
1838 }
1839 else clear_const(current,dops[i].rt1);
1840 }
1841 else {
1842 set_const(current,dops[i].rt1,imm[i]<<16); // LUI
1843 }
1844 dirty_reg(current,dops[i].rt1);
1845}
1846
1847static void load_alloc(struct regstat *current,int i)
1848{
1849 clear_const(current,dops[i].rt1);
1850 //if(dops[i].rs1!=dops[i].rt1&&needed_again(dops[i].rs1,i)) clear_const(current,dops[i].rs1); // Does this help or hurt?
1851 if(!dops[i].rs1) current->u&=~1LL; // Allow allocating r0 if it's the source register
1852 if (needed_again(dops[i].rs1, i))
1853 alloc_reg(current, i, dops[i].rs1);
1854 if (ram_offset)
1855 alloc_reg(current, i, ROREG);
1856 if(dops[i].rt1&&!((current->u>>dops[i].rt1)&1)) {
1857 alloc_reg(current,i,dops[i].rt1);
1858 assert(get_reg(current->regmap,dops[i].rt1)>=0);
1859 if(dops[i].opcode==0x27||dops[i].opcode==0x37) // LWU/LD
1860 {
1861 assert(0);
1862 }
1863 else if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1864 {
1865 assert(0);
1866 }
1867 dirty_reg(current,dops[i].rt1);
1868 // LWL/LWR need a temporary register for the old value
1869 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1870 {
1871 alloc_reg(current,i,FTEMP);
1872 alloc_reg_temp(current,i,-1);
1873 minimum_free_regs[i]=1;
1874 }
1875 }
1876 else
1877 {
1878 // Load to r0 or unneeded register (dummy load)
1879 // but we still need a register to calculate the address
1880 if(dops[i].opcode==0x22||dops[i].opcode==0x26)
1881 {
1882 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1883 }
1884 alloc_reg_temp(current,i,-1);
1885 minimum_free_regs[i]=1;
1886 if(dops[i].opcode==0x1A||dops[i].opcode==0x1B) // LDL/LDR
1887 {
1888 assert(0);
1889 }
1890 }
1891}
1892
1893static void store_alloc(struct regstat *current,int i)
1894{
1895 clear_const(current,dops[i].rs2);
1896 if(!(dops[i].rs2)) current->u&=~1LL; // Allow allocating r0 if necessary
1897 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1898 alloc_reg(current,i,dops[i].rs2);
1899 if(dops[i].opcode==0x2c||dops[i].opcode==0x2d||dops[i].opcode==0x3f) { // 64-bit SDL/SDR/SD
1900 assert(0);
1901 }
1902 if (ram_offset)
1903 alloc_reg(current, i, ROREG);
1904 #if defined(HOST_IMM8)
1905 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1906 alloc_reg(current, i, INVCP);
1907 #endif
1908 if(dops[i].opcode==0x2a||dops[i].opcode==0x2e||dops[i].opcode==0x2c||dops[i].opcode==0x2d) { // SWL/SWL/SDL/SDR
1909 alloc_reg(current,i,FTEMP);
1910 }
1911 // We need a temporary register for address generation
1912 alloc_reg_temp(current,i,-1);
1913 minimum_free_regs[i]=1;
1914}
1915
1916static void c1ls_alloc(struct regstat *current,int i)
1917{
1918 clear_const(current,dops[i].rt1);
1919 alloc_reg(current,i,CSREG); // Status
1920}
1921
1922static void c2ls_alloc(struct regstat *current,int i)
1923{
1924 clear_const(current,dops[i].rt1);
1925 if(needed_again(dops[i].rs1,i)) alloc_reg(current,i,dops[i].rs1);
1926 alloc_reg(current,i,FTEMP);
1927 if (ram_offset)
1928 alloc_reg(current, i, ROREG);
1929 #if defined(HOST_IMM8)
1930 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1931 if (dops[i].opcode == 0x3a) // SWC2
1932 alloc_reg(current,i,INVCP);
1933 #endif
1934 // We need a temporary register for address generation
1935 alloc_reg_temp(current,i,-1);
1936 minimum_free_regs[i]=1;
1937}
1938
1939#ifndef multdiv_alloc
1940static void multdiv_alloc(struct regstat *current,int i)
1941{
1942 // case 0x18: MULT
1943 // case 0x19: MULTU
1944 // case 0x1A: DIV
1945 // case 0x1B: DIVU
1946 // case 0x1C: DMULT
1947 // case 0x1D: DMULTU
1948 // case 0x1E: DDIV
1949 // case 0x1F: DDIVU
1950 clear_const(current,dops[i].rs1);
1951 clear_const(current,dops[i].rs2);
1952 alloc_cc(current,i); // for stalls
1953 if(dops[i].rs1&&dops[i].rs2)
1954 {
1955 if((dops[i].opcode2&4)==0) // 32-bit
1956 {
1957 current->u&=~(1LL<<HIREG);
1958 current->u&=~(1LL<<LOREG);
1959 alloc_reg(current,i,HIREG);
1960 alloc_reg(current,i,LOREG);
1961 alloc_reg(current,i,dops[i].rs1);
1962 alloc_reg(current,i,dops[i].rs2);
1963 dirty_reg(current,HIREG);
1964 dirty_reg(current,LOREG);
1965 }
1966 else // 64-bit
1967 {
1968 assert(0);
1969 }
1970 }
1971 else
1972 {
1973 // Multiply by zero is zero.
1974 // MIPS does not have a divide by zero exception.
1975 // The result is undefined, we return zero.
1976 alloc_reg(current,i,HIREG);
1977 alloc_reg(current,i,LOREG);
1978 dirty_reg(current,HIREG);
1979 dirty_reg(current,LOREG);
1980 }
1981}
1982#endif
1983
1984static void cop0_alloc(struct regstat *current,int i)
1985{
1986 if(dops[i].opcode2==0) // MFC0
1987 {
1988 if(dops[i].rt1) {
1989 clear_const(current,dops[i].rt1);
1990 alloc_all(current,i);
1991 alloc_reg(current,i,dops[i].rt1);
1992 dirty_reg(current,dops[i].rt1);
1993 }
1994 }
1995 else if(dops[i].opcode2==4) // MTC0
1996 {
1997 if(dops[i].rs1){
1998 clear_const(current,dops[i].rs1);
1999 alloc_reg(current,i,dops[i].rs1);
2000 alloc_all(current,i);
2001 }
2002 else {
2003 alloc_all(current,i); // FIXME: Keep r0
2004 current->u&=~1LL;
2005 alloc_reg(current,i,0);
2006 }
2007 }
2008 else
2009 {
2010 // TLBR/TLBWI/TLBWR/TLBP/ERET
2011 assert(dops[i].opcode2==0x10);
2012 alloc_all(current,i);
2013 }
2014 minimum_free_regs[i]=HOST_REGS;
2015}
2016
2017static void cop2_alloc(struct regstat *current,int i)
2018{
2019 if (dops[i].opcode2 < 3) // MFC2/CFC2
2020 {
2021 alloc_cc(current,i); // for stalls
2022 dirty_reg(current,CCREG);
2023 if(dops[i].rt1){
2024 clear_const(current,dops[i].rt1);
2025 alloc_reg(current,i,dops[i].rt1);
2026 dirty_reg(current,dops[i].rt1);
2027 }
2028 }
2029 else if (dops[i].opcode2 > 3) // MTC2/CTC2
2030 {
2031 if(dops[i].rs1){
2032 clear_const(current,dops[i].rs1);
2033 alloc_reg(current,i,dops[i].rs1);
2034 }
2035 else {
2036 current->u&=~1LL;
2037 alloc_reg(current,i,0);
2038 }
2039 }
2040 alloc_reg_temp(current,i,-1);
2041 minimum_free_regs[i]=1;
2042}
2043
2044static void c2op_alloc(struct regstat *current,int i)
2045{
2046 alloc_cc(current,i); // for stalls
2047 dirty_reg(current,CCREG);
2048 alloc_reg_temp(current,i,-1);
2049}
2050
2051static void syscall_alloc(struct regstat *current,int i)
2052{
2053 alloc_cc(current,i);
2054 dirty_reg(current,CCREG);
2055 alloc_all(current,i);
2056 minimum_free_regs[i]=HOST_REGS;
2057 current->isconst=0;
2058}
2059
2060static void delayslot_alloc(struct regstat *current,int i)
2061{
2062 switch(dops[i].itype) {
2063 case UJUMP:
2064 case CJUMP:
2065 case SJUMP:
2066 case RJUMP:
2067 case SYSCALL:
2068 case HLECALL:
2069 case SPAN:
2070 assem_debug("jump in the delay slot. this shouldn't happen.\n");//abort();
2071 SysPrintf("Disabled speculative precompilation\n");
2072 stop_after_jal=1;
2073 break;
2074 case IMM16:
2075 imm16_alloc(current,i);
2076 break;
2077 case LOAD:
2078 case LOADLR:
2079 load_alloc(current,i);
2080 break;
2081 case STORE:
2082 case STORELR:
2083 store_alloc(current,i);
2084 break;
2085 case ALU:
2086 alu_alloc(current,i);
2087 break;
2088 case SHIFT:
2089 shift_alloc(current,i);
2090 break;
2091 case MULTDIV:
2092 multdiv_alloc(current,i);
2093 break;
2094 case SHIFTIMM:
2095 shiftimm_alloc(current,i);
2096 break;
2097 case MOV:
2098 mov_alloc(current,i);
2099 break;
2100 case COP0:
2101 cop0_alloc(current,i);
2102 break;
2103 case COP1:
2104 break;
2105 case COP2:
2106 cop2_alloc(current,i);
2107 break;
2108 case C1LS:
2109 c1ls_alloc(current,i);
2110 break;
2111 case C2LS:
2112 c2ls_alloc(current,i);
2113 break;
2114 case C2OP:
2115 c2op_alloc(current,i);
2116 break;
2117 }
2118}
2119
2120// Special case where a branch and delay slot span two pages in virtual memory
2121static void pagespan_alloc(struct regstat *current,int i)
2122{
2123 current->isconst=0;
2124 current->wasconst=0;
2125 regs[i].wasconst=0;
2126 minimum_free_regs[i]=HOST_REGS;
2127 alloc_all(current,i);
2128 alloc_cc(current,i);
2129 dirty_reg(current,CCREG);
2130 if(dops[i].opcode==3) // JAL
2131 {
2132 alloc_reg(current,i,31);
2133 dirty_reg(current,31);
2134 }
2135 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
2136 {
2137 alloc_reg(current,i,dops[i].rs1);
2138 if (dops[i].rt1!=0) {
2139 alloc_reg(current,i,dops[i].rt1);
2140 dirty_reg(current,dops[i].rt1);
2141 }
2142 }
2143 if((dops[i].opcode&0x2E)==4) // BEQ/BNE/BEQL/BNEL
2144 {
2145 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2146 if(dops[i].rs2) alloc_reg(current,i,dops[i].rs2);
2147 }
2148 else
2149 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
2150 {
2151 if(dops[i].rs1) alloc_reg(current,i,dops[i].rs1);
2152 }
2153 //else ...
2154}
2155
2156static void add_stub(enum stub_type type, void *addr, void *retaddr,
2157 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
2158{
2159 assert(stubcount < ARRAY_SIZE(stubs));
2160 stubs[stubcount].type = type;
2161 stubs[stubcount].addr = addr;
2162 stubs[stubcount].retaddr = retaddr;
2163 stubs[stubcount].a = a;
2164 stubs[stubcount].b = b;
2165 stubs[stubcount].c = c;
2166 stubs[stubcount].d = d;
2167 stubs[stubcount].e = e;
2168 stubcount++;
2169}
2170
2171static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
2172 int i, int addr_reg, const struct regstat *i_regs, int ccadj, u_int reglist)
2173{
2174 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
2175}
2176
2177// Write out a single register
2178static void wb_register(signed char r, const signed char regmap[], uint64_t dirty)
2179{
2180 int hr;
2181 for(hr=0;hr<HOST_REGS;hr++) {
2182 if(hr!=EXCLUDE_REG) {
2183 if(regmap[hr]==r) {
2184 if((dirty>>hr)&1) {
2185 assert(regmap[hr]<64);
2186 emit_storereg(r,hr);
2187 }
2188 }
2189 }
2190 }
2191}
2192
2193static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
2194{
2195 //if(dirty_pre==dirty) return;
2196 int hr,reg;
2197 for(hr=0;hr<HOST_REGS;hr++) {
2198 if(hr!=EXCLUDE_REG) {
2199 reg=pre[hr];
2200 if(((~u)>>reg)&1) {
2201 if(reg>0) {
2202 if(((dirty_pre&~dirty)>>hr)&1) {
2203 if(reg>0&&reg<34) {
2204 emit_storereg(reg,hr);
2205 }
2206 else if(reg>=64) {
2207 assert(0);
2208 }
2209 }
2210 }
2211 }
2212 }
2213 }
2214}
2215
2216// trashes r2
2217static void pass_args(int a0, int a1)
2218{
2219 if(a0==1&&a1==0) {
2220 // must swap
2221 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2222 }
2223 else if(a0!=0&&a1==0) {
2224 emit_mov(a1,1);
2225 if (a0>=0) emit_mov(a0,0);
2226 }
2227 else {
2228 if(a0>=0&&a0!=0) emit_mov(a0,0);
2229 if(a1>=0&&a1!=1) emit_mov(a1,1);
2230 }
2231}
2232
2233static void alu_assemble(int i, const struct regstat *i_regs)
2234{
2235 if(dops[i].opcode2>=0x20&&dops[i].opcode2<=0x23) { // ADD/ADDU/SUB/SUBU
2236 if(dops[i].rt1) {
2237 signed char s1,s2,t;
2238 t=get_reg(i_regs->regmap,dops[i].rt1);
2239 if(t>=0) {
2240 s1=get_reg(i_regs->regmap,dops[i].rs1);
2241 s2=get_reg(i_regs->regmap,dops[i].rs2);
2242 if(dops[i].rs1&&dops[i].rs2) {
2243 assert(s1>=0);
2244 assert(s2>=0);
2245 if(dops[i].opcode2&2) emit_sub(s1,s2,t);
2246 else emit_add(s1,s2,t);
2247 }
2248 else if(dops[i].rs1) {
2249 if(s1>=0) emit_mov(s1,t);
2250 else emit_loadreg(dops[i].rs1,t);
2251 }
2252 else if(dops[i].rs2) {
2253 if(s2>=0) {
2254 if(dops[i].opcode2&2) emit_neg(s2,t);
2255 else emit_mov(s2,t);
2256 }
2257 else {
2258 emit_loadreg(dops[i].rs2,t);
2259 if(dops[i].opcode2&2) emit_neg(t,t);
2260 }
2261 }
2262 else emit_zeroreg(t);
2263 }
2264 }
2265 }
2266 if(dops[i].opcode2>=0x2c&&dops[i].opcode2<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2267 assert(0);
2268 }
2269 if(dops[i].opcode2==0x2a||dops[i].opcode2==0x2b) { // SLT/SLTU
2270 if(dops[i].rt1) {
2271 signed char s1l,s2l,t;
2272 {
2273 t=get_reg(i_regs->regmap,dops[i].rt1);
2274 //assert(t>=0);
2275 if(t>=0) {
2276 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2277 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2278 if(dops[i].rs2==0) // rx<r0
2279 {
2280 if(dops[i].opcode2==0x2a&&dops[i].rs1!=0) { // SLT
2281 assert(s1l>=0);
2282 emit_shrimm(s1l,31,t);
2283 }
2284 else // SLTU (unsigned can not be less than zero, 0<0)
2285 emit_zeroreg(t);
2286 }
2287 else if(dops[i].rs1==0) // r0<rx
2288 {
2289 assert(s2l>=0);
2290 if(dops[i].opcode2==0x2a) // SLT
2291 emit_set_gz32(s2l,t);
2292 else // SLTU (set if not zero)
2293 emit_set_nz32(s2l,t);
2294 }
2295 else{
2296 assert(s1l>=0);assert(s2l>=0);
2297 if(dops[i].opcode2==0x2a) // SLT
2298 emit_set_if_less32(s1l,s2l,t);
2299 else // SLTU
2300 emit_set_if_carry32(s1l,s2l,t);
2301 }
2302 }
2303 }
2304 }
2305 }
2306 if(dops[i].opcode2>=0x24&&dops[i].opcode2<=0x27) { // AND/OR/XOR/NOR
2307 if(dops[i].rt1) {
2308 signed char s1l,s2l,tl;
2309 tl=get_reg(i_regs->regmap,dops[i].rt1);
2310 {
2311 if(tl>=0) {
2312 s1l=get_reg(i_regs->regmap,dops[i].rs1);
2313 s2l=get_reg(i_regs->regmap,dops[i].rs2);
2314 if(dops[i].rs1&&dops[i].rs2) {
2315 assert(s1l>=0);
2316 assert(s2l>=0);
2317 if(dops[i].opcode2==0x24) { // AND
2318 emit_and(s1l,s2l,tl);
2319 } else
2320 if(dops[i].opcode2==0x25) { // OR
2321 emit_or(s1l,s2l,tl);
2322 } else
2323 if(dops[i].opcode2==0x26) { // XOR
2324 emit_xor(s1l,s2l,tl);
2325 } else
2326 if(dops[i].opcode2==0x27) { // NOR
2327 emit_or(s1l,s2l,tl);
2328 emit_not(tl,tl);
2329 }
2330 }
2331 else
2332 {
2333 if(dops[i].opcode2==0x24) { // AND
2334 emit_zeroreg(tl);
2335 } else
2336 if(dops[i].opcode2==0x25||dops[i].opcode2==0x26) { // OR/XOR
2337 if(dops[i].rs1){
2338 if(s1l>=0) emit_mov(s1l,tl);
2339 else emit_loadreg(dops[i].rs1,tl); // CHECK: regmap_entry?
2340 }
2341 else
2342 if(dops[i].rs2){
2343 if(s2l>=0) emit_mov(s2l,tl);
2344 else emit_loadreg(dops[i].rs2,tl); // CHECK: regmap_entry?
2345 }
2346 else emit_zeroreg(tl);
2347 } else
2348 if(dops[i].opcode2==0x27) { // NOR
2349 if(dops[i].rs1){
2350 if(s1l>=0) emit_not(s1l,tl);
2351 else {
2352 emit_loadreg(dops[i].rs1,tl);
2353 emit_not(tl,tl);
2354 }
2355 }
2356 else
2357 if(dops[i].rs2){
2358 if(s2l>=0) emit_not(s2l,tl);
2359 else {
2360 emit_loadreg(dops[i].rs2,tl);
2361 emit_not(tl,tl);
2362 }
2363 }
2364 else emit_movimm(-1,tl);
2365 }
2366 }
2367 }
2368 }
2369 }
2370 }
2371}
2372
2373static void imm16_assemble(int i, const struct regstat *i_regs)
2374{
2375 if (dops[i].opcode==0x0f) { // LUI
2376 if(dops[i].rt1) {
2377 signed char t;
2378 t=get_reg(i_regs->regmap,dops[i].rt1);
2379 //assert(t>=0);
2380 if(t>=0) {
2381 if(!((i_regs->isconst>>t)&1))
2382 emit_movimm(imm[i]<<16,t);
2383 }
2384 }
2385 }
2386 if(dops[i].opcode==0x08||dops[i].opcode==0x09) { // ADDI/ADDIU
2387 if(dops[i].rt1) {
2388 signed char s,t;
2389 t=get_reg(i_regs->regmap,dops[i].rt1);
2390 s=get_reg(i_regs->regmap,dops[i].rs1);
2391 if(dops[i].rs1) {
2392 //assert(t>=0);
2393 //assert(s>=0);
2394 if(t>=0) {
2395 if(!((i_regs->isconst>>t)&1)) {
2396 if(s<0) {
2397 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2398 emit_addimm(t,imm[i],t);
2399 }else{
2400 if(!((i_regs->wasconst>>s)&1))
2401 emit_addimm(s,imm[i],t);
2402 else
2403 emit_movimm(constmap[i][s]+imm[i],t);
2404 }
2405 }
2406 }
2407 } else {
2408 if(t>=0) {
2409 if(!((i_regs->isconst>>t)&1))
2410 emit_movimm(imm[i],t);
2411 }
2412 }
2413 }
2414 }
2415 if(dops[i].opcode==0x18||dops[i].opcode==0x19) { // DADDI/DADDIU
2416 if(dops[i].rt1) {
2417 signed char sl,tl;
2418 tl=get_reg(i_regs->regmap,dops[i].rt1);
2419 sl=get_reg(i_regs->regmap,dops[i].rs1);
2420 if(tl>=0) {
2421 if(dops[i].rs1) {
2422 assert(sl>=0);
2423 emit_addimm(sl,imm[i],tl);
2424 } else {
2425 emit_movimm(imm[i],tl);
2426 }
2427 }
2428 }
2429 }
2430 else if(dops[i].opcode==0x0a||dops[i].opcode==0x0b) { // SLTI/SLTIU
2431 if(dops[i].rt1) {
2432 //assert(dops[i].rs1!=0); // r0 might be valid, but it's probably a bug
2433 signed char sl,t;
2434 t=get_reg(i_regs->regmap,dops[i].rt1);
2435 sl=get_reg(i_regs->regmap,dops[i].rs1);
2436 //assert(t>=0);
2437 if(t>=0) {
2438 if(dops[i].rs1>0) {
2439 if(dops[i].opcode==0x0a) { // SLTI
2440 if(sl<0) {
2441 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2442 emit_slti32(t,imm[i],t);
2443 }else{
2444 emit_slti32(sl,imm[i],t);
2445 }
2446 }
2447 else { // SLTIU
2448 if(sl<0) {
2449 if(i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2450 emit_sltiu32(t,imm[i],t);
2451 }else{
2452 emit_sltiu32(sl,imm[i],t);
2453 }
2454 }
2455 }else{
2456 // SLTI(U) with r0 is just stupid,
2457 // nonetheless examples can be found
2458 if(dops[i].opcode==0x0a) // SLTI
2459 if(0<imm[i]) emit_movimm(1,t);
2460 else emit_zeroreg(t);
2461 else // SLTIU
2462 {
2463 if(imm[i]) emit_movimm(1,t);
2464 else emit_zeroreg(t);
2465 }
2466 }
2467 }
2468 }
2469 }
2470 else if(dops[i].opcode>=0x0c&&dops[i].opcode<=0x0e) { // ANDI/ORI/XORI
2471 if(dops[i].rt1) {
2472 signed char sl,tl;
2473 tl=get_reg(i_regs->regmap,dops[i].rt1);
2474 sl=get_reg(i_regs->regmap,dops[i].rs1);
2475 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2476 if(dops[i].opcode==0x0c) //ANDI
2477 {
2478 if(dops[i].rs1) {
2479 if(sl<0) {
2480 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2481 emit_andimm(tl,imm[i],tl);
2482 }else{
2483 if(!((i_regs->wasconst>>sl)&1))
2484 emit_andimm(sl,imm[i],tl);
2485 else
2486 emit_movimm(constmap[i][sl]&imm[i],tl);
2487 }
2488 }
2489 else
2490 emit_zeroreg(tl);
2491 }
2492 else
2493 {
2494 if(dops[i].rs1) {
2495 if(sl<0) {
2496 if(i_regs->regmap_entry[tl]!=dops[i].rs1) emit_loadreg(dops[i].rs1,tl);
2497 }
2498 if(dops[i].opcode==0x0d) { // ORI
2499 if(sl<0) {
2500 emit_orimm(tl,imm[i],tl);
2501 }else{
2502 if(!((i_regs->wasconst>>sl)&1))
2503 emit_orimm(sl,imm[i],tl);
2504 else
2505 emit_movimm(constmap[i][sl]|imm[i],tl);
2506 }
2507 }
2508 if(dops[i].opcode==0x0e) { // XORI
2509 if(sl<0) {
2510 emit_xorimm(tl,imm[i],tl);
2511 }else{
2512 if(!((i_regs->wasconst>>sl)&1))
2513 emit_xorimm(sl,imm[i],tl);
2514 else
2515 emit_movimm(constmap[i][sl]^imm[i],tl);
2516 }
2517 }
2518 }
2519 else {
2520 emit_movimm(imm[i],tl);
2521 }
2522 }
2523 }
2524 }
2525 }
2526}
2527
2528static void shiftimm_assemble(int i, const struct regstat *i_regs)
2529{
2530 if(dops[i].opcode2<=0x3) // SLL/SRL/SRA
2531 {
2532 if(dops[i].rt1) {
2533 signed char s,t;
2534 t=get_reg(i_regs->regmap,dops[i].rt1);
2535 s=get_reg(i_regs->regmap,dops[i].rs1);
2536 //assert(t>=0);
2537 if(t>=0&&!((i_regs->isconst>>t)&1)){
2538 if(dops[i].rs1==0)
2539 {
2540 emit_zeroreg(t);
2541 }
2542 else
2543 {
2544 if(s<0&&i_regs->regmap_entry[t]!=dops[i].rs1) emit_loadreg(dops[i].rs1,t);
2545 if(imm[i]) {
2546 if(dops[i].opcode2==0) // SLL
2547 {
2548 emit_shlimm(s<0?t:s,imm[i],t);
2549 }
2550 if(dops[i].opcode2==2) // SRL
2551 {
2552 emit_shrimm(s<0?t:s,imm[i],t);
2553 }
2554 if(dops[i].opcode2==3) // SRA
2555 {
2556 emit_sarimm(s<0?t:s,imm[i],t);
2557 }
2558 }else{
2559 // Shift by zero
2560 if(s>=0 && s!=t) emit_mov(s,t);
2561 }
2562 }
2563 }
2564 //emit_storereg(dops[i].rt1,t); //DEBUG
2565 }
2566 }
2567 if(dops[i].opcode2>=0x38&&dops[i].opcode2<=0x3b) // DSLL/DSRL/DSRA
2568 {
2569 assert(0);
2570 }
2571 if(dops[i].opcode2==0x3c) // DSLL32
2572 {
2573 assert(0);
2574 }
2575 if(dops[i].opcode2==0x3e) // DSRL32
2576 {
2577 assert(0);
2578 }
2579 if(dops[i].opcode2==0x3f) // DSRA32
2580 {
2581 assert(0);
2582 }
2583}
2584
2585#ifndef shift_assemble
2586static void shift_assemble(int i, const struct regstat *i_regs)
2587{
2588 signed char s,t,shift;
2589 if (dops[i].rt1 == 0)
2590 return;
2591 assert(dops[i].opcode2<=0x07); // SLLV/SRLV/SRAV
2592 t = get_reg(i_regs->regmap, dops[i].rt1);
2593 s = get_reg(i_regs->regmap, dops[i].rs1);
2594 shift = get_reg(i_regs->regmap, dops[i].rs2);
2595 if (t < 0)
2596 return;
2597
2598 if(dops[i].rs1==0)
2599 emit_zeroreg(t);
2600 else if(dops[i].rs2==0) {
2601 assert(s>=0);
2602 if(s!=t) emit_mov(s,t);
2603 }
2604 else {
2605 host_tempreg_acquire();
2606 emit_andimm(shift,31,HOST_TEMPREG);
2607 switch(dops[i].opcode2) {
2608 case 4: // SLLV
2609 emit_shl(s,HOST_TEMPREG,t);
2610 break;
2611 case 6: // SRLV
2612 emit_shr(s,HOST_TEMPREG,t);
2613 break;
2614 case 7: // SRAV
2615 emit_sar(s,HOST_TEMPREG,t);
2616 break;
2617 default:
2618 assert(0);
2619 }
2620 host_tempreg_release();
2621 }
2622}
2623
2624#endif
2625
2626enum {
2627 MTYPE_8000 = 0,
2628 MTYPE_8020,
2629 MTYPE_0000,
2630 MTYPE_A000,
2631 MTYPE_1F80,
2632};
2633
2634static int get_ptr_mem_type(u_int a)
2635{
2636 if(a < 0x00200000) {
2637 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2638 // return wrong, must use memhandler for BIOS self-test to pass
2639 // 007 does similar stuff from a00 mirror, weird stuff
2640 return MTYPE_8000;
2641 return MTYPE_0000;
2642 }
2643 if(0x1f800000 <= a && a < 0x1f801000)
2644 return MTYPE_1F80;
2645 if(0x80200000 <= a && a < 0x80800000)
2646 return MTYPE_8020;
2647 if(0xa0000000 <= a && a < 0xa0200000)
2648 return MTYPE_A000;
2649 return MTYPE_8000;
2650}
2651
2652static int get_ro_reg(const struct regstat *i_regs, int host_tempreg_free)
2653{
2654 int r = get_reg(i_regs->regmap, ROREG);
2655 if (r < 0 && host_tempreg_free) {
2656 host_tempreg_acquire();
2657 emit_loadreg(ROREG, r = HOST_TEMPREG);
2658 }
2659 if (r < 0)
2660 abort();
2661 return r;
2662}
2663
2664static void *emit_fastpath_cmp_jump(int i, const struct regstat *i_regs,
2665 int addr, int *offset_reg, int *addr_reg_override)
2666{
2667 void *jaddr = NULL;
2668 int type = 0;
2669 int mr = dops[i].rs1;
2670 *offset_reg = -1;
2671 if(((smrv_strong|smrv_weak)>>mr)&1) {
2672 type=get_ptr_mem_type(smrv[mr]);
2673 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2674 }
2675 else {
2676 // use the mirror we are running on
2677 type=get_ptr_mem_type(start);
2678 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2679 }
2680
2681 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2682 host_tempreg_acquire();
2683 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2684 addr=*addr_reg_override=HOST_TEMPREG;
2685 type=0;
2686 }
2687 else if(type==MTYPE_0000) { // RAM 0 mirror
2688 host_tempreg_acquire();
2689 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2690 addr=*addr_reg_override=HOST_TEMPREG;
2691 type=0;
2692 }
2693 else if(type==MTYPE_A000) { // RAM A mirror
2694 host_tempreg_acquire();
2695 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2696 addr=*addr_reg_override=HOST_TEMPREG;
2697 type=0;
2698 }
2699 else if(type==MTYPE_1F80) { // scratchpad
2700 if (psxH == (void *)0x1f800000) {
2701 host_tempreg_acquire();
2702 emit_xorimm(addr,0x1f800000,HOST_TEMPREG);
2703 emit_cmpimm(HOST_TEMPREG,0x1000);
2704 host_tempreg_release();
2705 jaddr=out;
2706 emit_jc(0);
2707 }
2708 else {
2709 // do the usual RAM check, jump will go to the right handler
2710 type=0;
2711 }
2712 }
2713
2714 if (type == 0) // need ram check
2715 {
2716 emit_cmpimm(addr,RAM_SIZE);
2717 jaddr = out;
2718 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2719 // Hint to branch predictor that the branch is unlikely to be taken
2720 if (dops[i].rs1 >= 28)
2721 emit_jno_unlikely(0);
2722 else
2723 #endif
2724 emit_jno(0);
2725 if (ram_offset != 0)
2726 *offset_reg = get_ro_reg(i_regs, 0);
2727 }
2728
2729 return jaddr;
2730}
2731
2732// return memhandler, or get directly accessable address and return 0
2733static void *get_direct_memhandler(void *table, u_int addr,
2734 enum stub_type type, uintptr_t *addr_host)
2735{
2736 uintptr_t msb = 1ull << (sizeof(uintptr_t)*8 - 1);
2737 uintptr_t l1, l2 = 0;
2738 l1 = ((uintptr_t *)table)[addr>>12];
2739 if (!(l1 & msb)) {
2740 uintptr_t v = l1 << 1;
2741 *addr_host = v + addr;
2742 return NULL;
2743 }
2744 else {
2745 l1 <<= 1;
2746 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2747 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2748 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2749 l2 = ((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2750 else
2751 l2 = ((uintptr_t *)l1)[(addr&0xfff)/4];
2752 if (!(l2 & msb)) {
2753 uintptr_t v = l2 << 1;
2754 *addr_host = v + (addr&0xfff);
2755 return NULL;
2756 }
2757 return (void *)(l2 << 1);
2758 }
2759}
2760
2761static u_int get_host_reglist(const signed char *regmap)
2762{
2763 u_int reglist = 0, hr;
2764 for (hr = 0; hr < HOST_REGS; hr++) {
2765 if (hr != EXCLUDE_REG && regmap[hr] >= 0)
2766 reglist |= 1 << hr;
2767 }
2768 return reglist;
2769}
2770
2771static u_int reglist_exclude(u_int reglist, int r1, int r2)
2772{
2773 if (r1 >= 0)
2774 reglist &= ~(1u << r1);
2775 if (r2 >= 0)
2776 reglist &= ~(1u << r2);
2777 return reglist;
2778}
2779
2780// find a temp caller-saved register not in reglist (so assumed to be free)
2781static int reglist_find_free(u_int reglist)
2782{
2783 u_int free_regs = ~reglist & CALLER_SAVE_REGS;
2784 if (free_regs == 0)
2785 return -1;
2786 return __builtin_ctz(free_regs);
2787}
2788
2789static void do_load_word(int a, int rt, int offset_reg)
2790{
2791 if (offset_reg >= 0)
2792 emit_ldr_dualindexed(offset_reg, a, rt);
2793 else
2794 emit_readword_indexed(0, a, rt);
2795}
2796
2797static void do_store_word(int a, int ofs, int rt, int offset_reg, int preseve_a)
2798{
2799 if (offset_reg < 0) {
2800 emit_writeword_indexed(rt, ofs, a);
2801 return;
2802 }
2803 if (ofs != 0)
2804 emit_addimm(a, ofs, a);
2805 emit_str_dualindexed(offset_reg, a, rt);
2806 if (ofs != 0 && preseve_a)
2807 emit_addimm(a, -ofs, a);
2808}
2809
2810static void do_store_hword(int a, int ofs, int rt, int offset_reg, int preseve_a)
2811{
2812 if (offset_reg < 0) {
2813 emit_writehword_indexed(rt, ofs, a);
2814 return;
2815 }
2816 if (ofs != 0)
2817 emit_addimm(a, ofs, a);
2818 emit_strh_dualindexed(offset_reg, a, rt);
2819 if (ofs != 0 && preseve_a)
2820 emit_addimm(a, -ofs, a);
2821}
2822
2823static void do_store_byte(int a, int rt, int offset_reg)
2824{
2825 if (offset_reg >= 0)
2826 emit_strb_dualindexed(offset_reg, a, rt);
2827 else
2828 emit_writebyte_indexed(rt, 0, a);
2829}
2830
2831static void load_assemble(int i, const struct regstat *i_regs, int ccadj_)
2832{
2833 int s,tl,addr;
2834 int offset;
2835 void *jaddr=0;
2836 int memtarget=0,c=0;
2837 int offset_reg = -1;
2838 int fastio_reg_override = -1;
2839 u_int reglist=get_host_reglist(i_regs->regmap);
2840 tl=get_reg(i_regs->regmap,dops[i].rt1);
2841 s=get_reg(i_regs->regmap,dops[i].rs1);
2842 offset=imm[i];
2843 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2844 if(s>=0) {
2845 c=(i_regs->wasconst>>s)&1;
2846 if (c) {
2847 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2848 }
2849 }
2850 //printf("load_assemble: c=%d\n",c);
2851 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2852 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2853 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2854 ||dops[i].rt1==0) {
2855 // could be FIFO, must perform the read
2856 // ||dummy read
2857 assem_debug("(forced read)\n");
2858 tl=get_reg_temp(i_regs->regmap);
2859 assert(tl>=0);
2860 }
2861 if(offset||s<0||c) addr=tl;
2862 else addr=s;
2863 //if(tl<0) tl=get_reg_temp(i_regs->regmap);
2864 if(tl>=0) {
2865 //printf("load_assemble: c=%d\n",c);
2866 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2867 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2868 reglist&=~(1<<tl);
2869 if(!c) {
2870 #ifdef R29_HACK
2871 // Strmnnrmn's speed hack
2872 if(dops[i].rs1!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2873 #endif
2874 {
2875 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
2876 &offset_reg, &fastio_reg_override);
2877 }
2878 }
2879 else if (ram_offset && memtarget) {
2880 offset_reg = get_ro_reg(i_regs, 0);
2881 }
2882 int dummy=(dops[i].rt1==0)||(tl!=get_reg(i_regs->regmap,dops[i].rt1)); // ignore loads to r0 and unneeded reg
2883 switch (dops[i].opcode) {
2884 case 0x20: // LB
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_ldrsb_dualindexed(offset_reg, a, tl);
2894 else
2895 emit_movsbl_indexed(0, a, tl);
2896 }
2897 if(jaddr)
2898 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2899 }
2900 else
2901 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2902 break;
2903 case 0x21: // LH
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_ldrsh_dualindexed(offset_reg, a, tl);
2912 else
2913 emit_movswl_indexed(0, a, tl);
2914 }
2915 if(jaddr)
2916 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2917 }
2918 else
2919 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2920 break;
2921 case 0x23: // LW
2922 if(!c||memtarget) {
2923 if(!dummy) {
2924 int a = addr;
2925 if (fastio_reg_override >= 0)
2926 a = fastio_reg_override;
2927 do_load_word(a, tl, offset_reg);
2928 }
2929 if(jaddr)
2930 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2931 }
2932 else
2933 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2934 break;
2935 case 0x24: // LBU
2936 if(!c||memtarget) {
2937 if(!dummy) {
2938 int a = tl;
2939 if (!c) a = addr;
2940 if (fastio_reg_override >= 0)
2941 a = fastio_reg_override;
2942
2943 if (offset_reg >= 0)
2944 emit_ldrb_dualindexed(offset_reg, a, tl);
2945 else
2946 emit_movzbl_indexed(0, a, tl);
2947 }
2948 if(jaddr)
2949 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2950 }
2951 else
2952 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2953 break;
2954 case 0x25: // LHU
2955 if(!c||memtarget) {
2956 if(!dummy) {
2957 int a = tl;
2958 if(!c) a = addr;
2959 if (fastio_reg_override >= 0)
2960 a = fastio_reg_override;
2961 if (offset_reg >= 0)
2962 emit_ldrh_dualindexed(offset_reg, a, tl);
2963 else
2964 emit_movzwl_indexed(0, a, tl);
2965 }
2966 if(jaddr)
2967 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj_,reglist);
2968 }
2969 else
2970 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,dops[i].rt1,ccadj_,reglist);
2971 break;
2972 case 0x27: // LWU
2973 case 0x37: // LD
2974 default:
2975 assert(0);
2976 }
2977 }
2978 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
2979 host_tempreg_release();
2980}
2981
2982#ifndef loadlr_assemble
2983static void loadlr_assemble(int i, const struct regstat *i_regs, int ccadj_)
2984{
2985 int s,tl,temp,temp2,addr;
2986 int offset;
2987 void *jaddr=0;
2988 int memtarget=0,c=0;
2989 int offset_reg = -1;
2990 int fastio_reg_override = -1;
2991 u_int reglist=get_host_reglist(i_regs->regmap);
2992 tl=get_reg(i_regs->regmap,dops[i].rt1);
2993 s=get_reg(i_regs->regmap,dops[i].rs1);
2994 temp=get_reg_temp(i_regs->regmap);
2995 temp2=get_reg(i_regs->regmap,FTEMP);
2996 addr=get_reg(i_regs->regmap,AGEN1+(i&1));
2997 assert(addr<0);
2998 offset=imm[i];
2999 reglist|=1<<temp;
3000 if(offset||s<0||c) addr=temp2;
3001 else addr=s;
3002 if(s>=0) {
3003 c=(i_regs->wasconst>>s)&1;
3004 if(c) {
3005 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3006 }
3007 }
3008 if(!c) {
3009 emit_shlimm(addr,3,temp);
3010 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3011 emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR
3012 }else{
3013 emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR
3014 }
3015 jaddr = emit_fastpath_cmp_jump(i, i_regs, temp2,
3016 &offset_reg, &fastio_reg_override);
3017 }
3018 else {
3019 if (ram_offset && memtarget) {
3020 offset_reg = get_ro_reg(i_regs, 0);
3021 }
3022 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
3023 emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR
3024 }else{
3025 emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR
3026 }
3027 }
3028 if (dops[i].opcode==0x22||dops[i].opcode==0x26) { // LWL/LWR
3029 if(!c||memtarget) {
3030 int a = temp2;
3031 if (fastio_reg_override >= 0)
3032 a = fastio_reg_override;
3033 do_load_word(a, temp2, offset_reg);
3034 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3035 host_tempreg_release();
3036 if(jaddr) add_stub_r(LOADW_STUB,jaddr,out,i,temp2,i_regs,ccadj_,reglist);
3037 }
3038 else
3039 inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj_,reglist);
3040 if(dops[i].rt1) {
3041 assert(tl>=0);
3042 emit_andimm(temp,24,temp);
3043 if (dops[i].opcode==0x22) // LWL
3044 emit_xorimm(temp,24,temp);
3045 host_tempreg_acquire();
3046 emit_movimm(-1,HOST_TEMPREG);
3047 if (dops[i].opcode==0x26) {
3048 emit_shr(temp2,temp,temp2);
3049 emit_bic_lsr(tl,HOST_TEMPREG,temp,tl);
3050 }else{
3051 emit_shl(temp2,temp,temp2);
3052 emit_bic_lsl(tl,HOST_TEMPREG,temp,tl);
3053 }
3054 host_tempreg_release();
3055 emit_or(temp2,tl,tl);
3056 }
3057 //emit_storereg(dops[i].rt1,tl); // DEBUG
3058 }
3059 if (dops[i].opcode==0x1A||dops[i].opcode==0x1B) { // LDL/LDR
3060 assert(0);
3061 }
3062}
3063#endif
3064
3065static void store_assemble(int i, const struct regstat *i_regs, int ccadj_)
3066{
3067 int s,tl;
3068 int addr,temp;
3069 int offset;
3070 void *jaddr=0;
3071 enum stub_type type=0;
3072 int memtarget=0,c=0;
3073 int agr=AGEN1+(i&1);
3074 int offset_reg = -1;
3075 int fastio_reg_override = -1;
3076 u_int reglist=get_host_reglist(i_regs->regmap);
3077 tl=get_reg(i_regs->regmap,dops[i].rs2);
3078 s=get_reg(i_regs->regmap,dops[i].rs1);
3079 temp=get_reg(i_regs->regmap,agr);
3080 if(temp<0) temp=get_reg_temp(i_regs->regmap);
3081 offset=imm[i];
3082 if(s>=0) {
3083 c=(i_regs->wasconst>>s)&1;
3084 if(c) {
3085 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3086 }
3087 }
3088 assert(tl>=0);
3089 assert(temp>=0);
3090 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
3091 if(offset||s<0||c) addr=temp;
3092 else addr=s;
3093 if (!c) {
3094 jaddr = emit_fastpath_cmp_jump(i, i_regs, addr,
3095 &offset_reg, &fastio_reg_override);
3096 }
3097 else if (ram_offset && memtarget) {
3098 offset_reg = get_ro_reg(i_regs, 0);
3099 }
3100
3101 switch (dops[i].opcode) {
3102 case 0x28: // SB
3103 if(!c||memtarget) {
3104 int a = temp;
3105 if (!c) a = addr;
3106 if (fastio_reg_override >= 0)
3107 a = fastio_reg_override;
3108 do_store_byte(a, tl, offset_reg);
3109 }
3110 type = STOREB_STUB;
3111 break;
3112 case 0x29: // SH
3113 if(!c||memtarget) {
3114 int a = temp;
3115 if (!c) a = addr;
3116 if (fastio_reg_override >= 0)
3117 a = fastio_reg_override;
3118 do_store_hword(a, 0, tl, offset_reg, 1);
3119 }
3120 type = STOREH_STUB;
3121 break;
3122 case 0x2B: // SW
3123 if(!c||memtarget) {
3124 int a = addr;
3125 if (fastio_reg_override >= 0)
3126 a = fastio_reg_override;
3127 do_store_word(a, 0, tl, offset_reg, 1);
3128 }
3129 type = STOREW_STUB;
3130 break;
3131 case 0x3F: // SD
3132 default:
3133 assert(0);
3134 }
3135 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3136 host_tempreg_release();
3137 if(jaddr) {
3138 // PCSX store handlers don't check invcode again
3139 reglist|=1<<addr;
3140 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3141 jaddr=0;
3142 }
3143 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3144 if(!c||memtarget) {
3145 #ifdef DESTRUCTIVE_SHIFT
3146 // The x86 shift operation is 'destructive'; it overwrites the
3147 // source register, so we need to make a copy first and use that.
3148 addr=temp;
3149 #endif
3150 #if defined(HOST_IMM8)
3151 int ir=get_reg(i_regs->regmap,INVCP);
3152 assert(ir>=0);
3153 emit_cmpmem_indexedsr12_reg(ir,addr,1);
3154 #else
3155 emit_cmpmem_indexedsr12_imm(invalid_code,addr,1);
3156 #endif
3157 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3158 emit_callne(invalidate_addr_reg[addr]);
3159 #else
3160 void *jaddr2 = out;
3161 emit_jne(0);
3162 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),addr,0,0,0);
3163 #endif
3164 }
3165 }
3166 u_int addr_val=constmap[i][s]+offset;
3167 if(jaddr) {
3168 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj_,reglist);
3169 } else if(c&&!memtarget) {
3170 inline_writestub(type,i,addr_val,i_regs->regmap,dops[i].rs2,ccadj_,reglist);
3171 }
3172 // basic current block modification detection..
3173 // not looking back as that should be in mips cache already
3174 // (see Spyro2 title->attract mode)
3175 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
3176 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
3177 assert(i_regs->regmap==regs[i].regmap); // not delay slot
3178 if(i_regs->regmap==regs[i].regmap) {
3179 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3180 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
3181 emit_movimm(start+i*4+4,0);
3182 emit_writeword(0,&pcaddr);
3183 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3184 emit_far_call(get_addr_ht);
3185 emit_jmpreg(0);
3186 }
3187 }
3188}
3189
3190static void storelr_assemble(int i, const struct regstat *i_regs, int ccadj_)
3191{
3192 int s,tl;
3193 int temp;
3194 int offset;
3195 void *jaddr=0;
3196 void *case1, *case23, *case3;
3197 void *done0, *done1, *done2;
3198 int memtarget=0,c=0;
3199 int agr=AGEN1+(i&1);
3200 int offset_reg = -1;
3201 u_int reglist=get_host_reglist(i_regs->regmap);
3202 tl=get_reg(i_regs->regmap,dops[i].rs2);
3203 s=get_reg(i_regs->regmap,dops[i].rs1);
3204 temp=get_reg(i_regs->regmap,agr);
3205 if(temp<0) temp=get_reg_temp(i_regs->regmap);
3206 offset=imm[i];
3207 if(s>=0) {
3208 c=(i_regs->isconst>>s)&1;
3209 if(c) {
3210 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
3211 }
3212 }
3213 assert(tl>=0);
3214 assert(temp>=0);
3215 if(!c) {
3216 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
3217 if(!offset&&s!=temp) emit_mov(s,temp);
3218 jaddr=out;
3219 emit_jno(0);
3220 }
3221 else
3222 {
3223 if(!memtarget||!dops[i].rs1) {
3224 jaddr=out;
3225 emit_jmp(0);
3226 }
3227 }
3228 if (ram_offset)
3229 offset_reg = get_ro_reg(i_regs, 0);
3230
3231 if (dops[i].opcode==0x2C||dops[i].opcode==0x2D) { // SDL/SDR
3232 assert(0);
3233 }
3234
3235 emit_testimm(temp,2);
3236 case23=out;
3237 emit_jne(0);
3238 emit_testimm(temp,1);
3239 case1=out;
3240 emit_jne(0);
3241 // 0
3242 if (dops[i].opcode == 0x2A) { // SWL
3243 // Write msb into least significant byte
3244 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3245 do_store_byte(temp, tl, offset_reg);
3246 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3247 }
3248 else if (dops[i].opcode == 0x2E) { // SWR
3249 // Write entire word
3250 do_store_word(temp, 0, tl, offset_reg, 1);
3251 }
3252 done0 = out;
3253 emit_jmp(0);
3254 // 1
3255 set_jump_target(case1, out);
3256 if (dops[i].opcode == 0x2A) { // SWL
3257 // Write two msb into two least significant bytes
3258 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3259 do_store_hword(temp, -1, tl, offset_reg, 0);
3260 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3261 }
3262 else if (dops[i].opcode == 0x2E) { // SWR
3263 // Write 3 lsb into three most significant bytes
3264 do_store_byte(temp, tl, offset_reg);
3265 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3266 do_store_hword(temp, 1, tl, offset_reg, 0);
3267 if (dops[i].rs2) emit_rorimm(tl, 24, tl);
3268 }
3269 done1=out;
3270 emit_jmp(0);
3271 // 2,3
3272 set_jump_target(case23, out);
3273 emit_testimm(temp,1);
3274 case3 = out;
3275 emit_jne(0);
3276 // 2
3277 if (dops[i].opcode==0x2A) { // SWL
3278 // Write 3 msb into three least significant bytes
3279 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3280 do_store_hword(temp, -2, tl, offset_reg, 1);
3281 if (dops[i].rs2) emit_rorimm(tl, 16, tl);
3282 do_store_byte(temp, tl, offset_reg);
3283 if (dops[i].rs2) emit_rorimm(tl, 8, tl);
3284 }
3285 else if (dops[i].opcode == 0x2E) { // SWR
3286 // Write two lsb into two most significant bytes
3287 do_store_hword(temp, 0, tl, offset_reg, 1);
3288 }
3289 done2 = out;
3290 emit_jmp(0);
3291 // 3
3292 set_jump_target(case3, out);
3293 if (dops[i].opcode == 0x2A) { // SWL
3294 do_store_word(temp, -3, tl, offset_reg, 0);
3295 }
3296 else if (dops[i].opcode == 0x2E) { // SWR
3297 do_store_byte(temp, tl, offset_reg);
3298 }
3299 set_jump_target(done0, out);
3300 set_jump_target(done1, out);
3301 set_jump_target(done2, out);
3302 if (offset_reg == HOST_TEMPREG)
3303 host_tempreg_release();
3304 if(!c||!memtarget)
3305 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj_,reglist);
3306 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3307 #if defined(HOST_IMM8)
3308 int ir=get_reg(i_regs->regmap,INVCP);
3309 assert(ir>=0);
3310 emit_cmpmem_indexedsr12_reg(ir,temp,1);
3311 #else
3312 emit_cmpmem_indexedsr12_imm(invalid_code,temp,1);
3313 #endif
3314 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3315 emit_callne(invalidate_addr_reg[temp]);
3316 #else
3317 void *jaddr2 = out;
3318 emit_jne(0);
3319 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),temp,0,0,0);
3320 #endif
3321 }
3322}
3323
3324static void cop0_assemble(int i, const struct regstat *i_regs, int ccadj_)
3325{
3326 if(dops[i].opcode2==0) // MFC0
3327 {
3328 signed char t=get_reg(i_regs->regmap,dops[i].rt1);
3329 u_int copr=(source[i]>>11)&0x1f;
3330 //assert(t>=0); // Why does this happen? OOT is weird
3331 if(t>=0&&dops[i].rt1!=0) {
3332 emit_readword(&reg_cop0[copr],t);
3333 }
3334 }
3335 else if(dops[i].opcode2==4) // MTC0
3336 {
3337 signed char s=get_reg(i_regs->regmap,dops[i].rs1);
3338 char copr=(source[i]>>11)&0x1f;
3339 assert(s>=0);
3340 wb_register(dops[i].rs1,i_regs->regmap,i_regs->dirty);
3341 if(copr==9||copr==11||copr==12||copr==13) {
3342 emit_readword(&last_count,HOST_TEMPREG);
3343 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
3344 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3345 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3346 emit_writeword(HOST_CCREG,&Count);
3347 }
3348 // What a mess. The status register (12) can enable interrupts,
3349 // so needs a special case to handle a pending interrupt.
3350 // The interrupt must be taken immediately, because a subsequent
3351 // instruction might disable interrupts again.
3352 if(copr==12||copr==13) {
3353 if (is_delayslot) {
3354 // burn cycles to cause cc_interrupt, which will
3355 // reschedule next_interupt. Relies on CCREG from above.
3356 assem_debug("MTC0 DS %d\n", copr);
3357 emit_writeword(HOST_CCREG,&last_count);
3358 emit_movimm(0,HOST_CCREG);
3359 emit_storereg(CCREG,HOST_CCREG);
3360 emit_loadreg(dops[i].rs1,1);
3361 emit_movimm(copr,0);
3362 emit_far_call(pcsx_mtc0_ds);
3363 emit_loadreg(dops[i].rs1,s);
3364 return;
3365 }
3366 emit_movimm(start+i*4+4,HOST_TEMPREG);
3367 emit_writeword(HOST_TEMPREG,&pcaddr);
3368 emit_movimm(0,HOST_TEMPREG);
3369 emit_writeword(HOST_TEMPREG,&pending_exception);
3370 }
3371 if(s==HOST_CCREG)
3372 emit_loadreg(dops[i].rs1,1);
3373 else if(s!=1)
3374 emit_mov(s,1);
3375 emit_movimm(copr,0);
3376 emit_far_call(pcsx_mtc0);
3377 if(copr==9||copr==11||copr==12||copr==13) {
3378 emit_readword(&Count,HOST_CCREG);
3379 emit_readword(&next_interupt,HOST_TEMPREG);
3380 emit_addimm(HOST_CCREG,-ccadj_,HOST_CCREG);
3381 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3382 emit_writeword(HOST_TEMPREG,&last_count);
3383 emit_storereg(CCREG,HOST_CCREG);
3384 }
3385 if(copr==12||copr==13) {
3386 assert(!is_delayslot);
3387 emit_readword(&pending_exception,14);
3388 emit_test(14,14);
3389 void *jaddr = out;
3390 emit_jeq(0);
3391 emit_readword(&pcaddr, 0);
3392 emit_addimm(HOST_CCREG,2,HOST_CCREG);
3393 emit_far_call(get_addr_ht);
3394 emit_jmpreg(0);
3395 set_jump_target(jaddr, out);
3396 }
3397 emit_loadreg(dops[i].rs1,s);
3398 }
3399 else
3400 {
3401 assert(dops[i].opcode2==0x10);
3402 //if((source[i]&0x3f)==0x10) // RFE
3403 {
3404 emit_readword(&Status,0);
3405 emit_andimm(0,0x3c,1);
3406 emit_andimm(0,~0xf,0);
3407 emit_orrshr_imm(1,2,0);
3408 emit_writeword(0,&Status);
3409 }
3410 }
3411}
3412
3413static void cop1_unusable(int i, const struct regstat *i_regs)
3414{
3415 // XXX: should just just do the exception instead
3416 //if(!cop1_usable)
3417 {
3418 void *jaddr=out;
3419 emit_jmp(0);
3420 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3421 }
3422}
3423
3424static void cop1_assemble(int i, const struct regstat *i_regs)
3425{
3426 cop1_unusable(i, i_regs);
3427}
3428
3429static void c1ls_assemble(int i, const struct regstat *i_regs)
3430{
3431 cop1_unusable(i, i_regs);
3432}
3433
3434// FP_STUB
3435static void do_cop1stub(int n)
3436{
3437 literal_pool(256);
3438 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3439 set_jump_target(stubs[n].addr, out);
3440 int i=stubs[n].a;
3441// int rs=stubs[n].b;
3442 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3443 int ds=stubs[n].d;
3444 if(!ds) {
3445 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3446 //if(i_regs!=&regs[i]) printf("oops: regs[i]=%x i_regs=%x",(int)&regs[i],(int)i_regs);
3447 }
3448 //else {printf("fp exception in delay slot\n");}
3449 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3450 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3451 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3452 emit_addimm(HOST_CCREG,ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3453 emit_far_jump(ds?fp_exception_ds:fp_exception);
3454}
3455
3456static int cop2_is_stalling_op(int i, int *cycles)
3457{
3458 if (dops[i].opcode == 0x3a) { // SWC2
3459 *cycles = 0;
3460 return 1;
3461 }
3462 if (dops[i].itype == COP2 && (dops[i].opcode2 == 0 || dops[i].opcode2 == 2)) { // MFC2/CFC2
3463 *cycles = 0;
3464 return 1;
3465 }
3466 if (dops[i].itype == C2OP) {
3467 *cycles = gte_cycletab[source[i] & 0x3f];
3468 return 1;
3469 }
3470 // ... what about MTC2/CTC2/LWC2?
3471 return 0;
3472}
3473
3474#if 0
3475static void log_gte_stall(int stall, u_int cycle)
3476{
3477 if ((u_int)stall <= 44)
3478 printf("x stall %2d %u\n", stall, cycle + last_count);
3479}
3480
3481static void emit_log_gte_stall(int i, int stall, u_int reglist)
3482{
3483 save_regs(reglist);
3484 if (stall > 0)
3485 emit_movimm(stall, 0);
3486 else
3487 emit_mov(HOST_TEMPREG, 0);
3488 emit_addimm(HOST_CCREG, ccadj[i], 1);
3489 emit_far_call(log_gte_stall);
3490 restore_regs(reglist);
3491}
3492#endif
3493
3494static void cop2_do_stall_check(u_int op, int i, const struct regstat *i_regs, u_int reglist)
3495{
3496 int j = i, other_gte_op_cycles = -1, stall = -MAXBLOCK, cycles_passed;
3497 int rtmp = reglist_find_free(reglist);
3498
3499 if (HACK_ENABLED(NDHACK_NO_STALLS))
3500 return;
3501 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3502 // happens occasionally... cc evicted? Don't bother then
3503 //printf("no cc %08x\n", start + i*4);
3504 return;
3505 }
3506 if (!dops[i].bt) {
3507 for (j = i - 1; j >= 0; j--) {
3508 //if (dops[j].is_ds) break;
3509 if (cop2_is_stalling_op(j, &other_gte_op_cycles) || dops[j].bt)
3510 break;
3511 if (j > 0 && ccadj[j - 1] > ccadj[j])
3512 break;
3513 }
3514 j = max(j, 0);
3515 }
3516 cycles_passed = ccadj[i] - ccadj[j];
3517 if (other_gte_op_cycles >= 0)
3518 stall = other_gte_op_cycles - cycles_passed;
3519 else if (cycles_passed >= 44)
3520 stall = 0; // can't stall
3521 if (stall == -MAXBLOCK && rtmp >= 0) {
3522 // unknown stall, do the expensive runtime check
3523 assem_debug("; cop2_do_stall_check\n");
3524#if 0 // too slow
3525 save_regs(reglist);
3526 emit_movimm(gte_cycletab[op], 0);
3527 emit_addimm(HOST_CCREG, ccadj[i], 1);
3528 emit_far_call(call_gteStall);
3529 restore_regs(reglist);
3530#else
3531 host_tempreg_acquire();
3532 emit_readword(&psxRegs.gteBusyCycle, rtmp);
3533 emit_addimm(rtmp, -ccadj[i], rtmp);
3534 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3535 emit_cmpimm(HOST_TEMPREG, 44);
3536 emit_cmovb_reg(rtmp, HOST_CCREG);
3537 //emit_log_gte_stall(i, 0, reglist);
3538 host_tempreg_release();
3539#endif
3540 }
3541 else if (stall > 0) {
3542 //emit_log_gte_stall(i, stall, reglist);
3543 emit_addimm(HOST_CCREG, stall, HOST_CCREG);
3544 }
3545
3546 // save gteBusyCycle, if needed
3547 if (gte_cycletab[op] == 0)
3548 return;
3549 other_gte_op_cycles = -1;
3550 for (j = i + 1; j < slen; j++) {
3551 if (cop2_is_stalling_op(j, &other_gte_op_cycles))
3552 break;
3553 if (dops[j].is_jump) {
3554 // check ds
3555 if (j + 1 < slen && cop2_is_stalling_op(j + 1, &other_gte_op_cycles))
3556 j++;
3557 break;
3558 }
3559 }
3560 if (other_gte_op_cycles >= 0)
3561 // will handle stall when assembling that op
3562 return;
3563 cycles_passed = ccadj[min(j, slen -1)] - ccadj[i];
3564 if (cycles_passed >= 44)
3565 return;
3566 assem_debug("; save gteBusyCycle\n");
3567 host_tempreg_acquire();
3568#if 0
3569 emit_readword(&last_count, HOST_TEMPREG);
3570 emit_add(HOST_TEMPREG, HOST_CCREG, HOST_TEMPREG);
3571 emit_addimm(HOST_TEMPREG, ccadj[i], HOST_TEMPREG);
3572 emit_addimm(HOST_TEMPREG, gte_cycletab[op]), HOST_TEMPREG);
3573 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3574#else
3575 emit_addimm(HOST_CCREG, ccadj[i] + gte_cycletab[op], HOST_TEMPREG);
3576 emit_writeword(HOST_TEMPREG, &psxRegs.gteBusyCycle);
3577#endif
3578 host_tempreg_release();
3579}
3580
3581static int is_mflohi(int i)
3582{
3583 return (dops[i].itype == MOV && (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG));
3584}
3585
3586static int check_multdiv(int i, int *cycles)
3587{
3588 if (dops[i].itype != MULTDIV)
3589 return 0;
3590 if (dops[i].opcode2 == 0x18 || dops[i].opcode2 == 0x19) // MULT(U)
3591 *cycles = 11; // approx from 7 11 14
3592 else
3593 *cycles = 37;
3594 return 1;
3595}
3596
3597static void multdiv_prepare_stall(int i, const struct regstat *i_regs, int ccadj_)
3598{
3599 int j, found = 0, c = 0;
3600 if (HACK_ENABLED(NDHACK_NO_STALLS))
3601 return;
3602 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG) {
3603 // happens occasionally... cc evicted? Don't bother then
3604 return;
3605 }
3606 for (j = i + 1; j < slen; j++) {
3607 if (dops[j].bt)
3608 break;
3609 if ((found = is_mflohi(j)))
3610 break;
3611 if (dops[j].is_jump) {
3612 // check ds
3613 if (j + 1 < slen && (found = is_mflohi(j + 1)))
3614 j++;
3615 break;
3616 }
3617 }
3618 if (found)
3619 // handle all in multdiv_do_stall()
3620 return;
3621 check_multdiv(i, &c);
3622 assert(c > 0);
3623 assem_debug("; muldiv prepare stall %d\n", c);
3624 host_tempreg_acquire();
3625 emit_addimm(HOST_CCREG, ccadj_ + c, HOST_TEMPREG);
3626 emit_writeword(HOST_TEMPREG, &psxRegs.muldivBusyCycle);
3627 host_tempreg_release();
3628}
3629
3630static void multdiv_do_stall(int i, const struct regstat *i_regs)
3631{
3632 int j, known_cycles = 0;
3633 u_int reglist = get_host_reglist(i_regs->regmap);
3634 int rtmp = get_reg_temp(i_regs->regmap);
3635 if (rtmp < 0)
3636 rtmp = reglist_find_free(reglist);
3637 if (HACK_ENABLED(NDHACK_NO_STALLS))
3638 return;
3639 if (get_reg(i_regs->regmap, CCREG) != HOST_CCREG || rtmp < 0) {
3640 // happens occasionally... cc evicted? Don't bother then
3641 //printf("no cc/rtmp %08x\n", start + i*4);
3642 return;
3643 }
3644 if (!dops[i].bt) {
3645 for (j = i - 1; j >= 0; j--) {
3646 if (dops[j].is_ds) break;
3647 if (check_multdiv(j, &known_cycles))
3648 break;
3649 if (is_mflohi(j))
3650 // already handled by this op
3651 return;
3652 if (dops[j].bt || (j > 0 && ccadj[j - 1] > ccadj[j]))
3653 break;
3654 }
3655 j = max(j, 0);
3656 }
3657 if (known_cycles > 0) {
3658 known_cycles -= ccadj[i] - ccadj[j];
3659 assem_debug("; muldiv stall resolved %d\n", known_cycles);
3660 if (known_cycles > 0)
3661 emit_addimm(HOST_CCREG, known_cycles, HOST_CCREG);
3662 return;
3663 }
3664 assem_debug("; muldiv stall unresolved\n");
3665 host_tempreg_acquire();
3666 emit_readword(&psxRegs.muldivBusyCycle, rtmp);
3667 emit_addimm(rtmp, -ccadj[i], rtmp);
3668 emit_sub(rtmp, HOST_CCREG, HOST_TEMPREG);
3669 emit_cmpimm(HOST_TEMPREG, 37);
3670 emit_cmovb_reg(rtmp, HOST_CCREG);
3671 //emit_log_gte_stall(i, 0, reglist);
3672 host_tempreg_release();
3673}
3674
3675static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3676{
3677 switch (copr) {
3678 case 1:
3679 case 3:
3680 case 5:
3681 case 8:
3682 case 9:
3683 case 10:
3684 case 11:
3685 emit_readword(&reg_cop2d[copr],tl);
3686 emit_signextend16(tl,tl);
3687 emit_writeword(tl,&reg_cop2d[copr]); // hmh
3688 break;
3689 case 7:
3690 case 16:
3691 case 17:
3692 case 18:
3693 case 19:
3694 emit_readword(&reg_cop2d[copr],tl);
3695 emit_andimm(tl,0xffff,tl);
3696 emit_writeword(tl,&reg_cop2d[copr]);
3697 break;
3698 case 15:
3699 emit_readword(&reg_cop2d[14],tl); // SXY2
3700 emit_writeword(tl,&reg_cop2d[copr]);
3701 break;
3702 case 28:
3703 case 29:
3704 c2op_mfc2_29_assemble(tl,temp);
3705 break;
3706 default:
3707 emit_readword(&reg_cop2d[copr],tl);
3708 break;
3709 }
3710}
3711
3712static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3713{
3714 switch (copr) {
3715 case 15:
3716 emit_readword(&reg_cop2d[13],temp); // SXY1
3717 emit_writeword(sl,&reg_cop2d[copr]);
3718 emit_writeword(temp,&reg_cop2d[12]); // SXY0
3719 emit_readword(&reg_cop2d[14],temp); // SXY2
3720 emit_writeword(sl,&reg_cop2d[14]);
3721 emit_writeword(temp,&reg_cop2d[13]); // SXY1
3722 break;
3723 case 28:
3724 emit_andimm(sl,0x001f,temp);
3725 emit_shlimm(temp,7,temp);
3726 emit_writeword(temp,&reg_cop2d[9]);
3727 emit_andimm(sl,0x03e0,temp);
3728 emit_shlimm(temp,2,temp);
3729 emit_writeword(temp,&reg_cop2d[10]);
3730 emit_andimm(sl,0x7c00,temp);
3731 emit_shrimm(temp,3,temp);
3732 emit_writeword(temp,&reg_cop2d[11]);
3733 emit_writeword(sl,&reg_cop2d[28]);
3734 break;
3735 case 30:
3736 emit_xorsar_imm(sl,sl,31,temp);
3737#if defined(HAVE_ARMV5) || defined(__aarch64__)
3738 emit_clz(temp,temp);
3739#else
3740 emit_movs(temp,HOST_TEMPREG);
3741 emit_movimm(0,temp);
3742 emit_jeq((int)out+4*4);
3743 emit_addpl_imm(temp,1,temp);
3744 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3745 emit_jns((int)out-2*4);
3746#endif
3747 emit_writeword(sl,&reg_cop2d[30]);
3748 emit_writeword(temp,&reg_cop2d[31]);
3749 break;
3750 case 31:
3751 break;
3752 default:
3753 emit_writeword(sl,&reg_cop2d[copr]);
3754 break;
3755 }
3756}
3757
3758static void c2ls_assemble(int i, const struct regstat *i_regs, int ccadj_)
3759{
3760 int s,tl;
3761 int ar;
3762 int offset;
3763 int memtarget=0,c=0;
3764 void *jaddr2=NULL;
3765 enum stub_type type;
3766 int agr=AGEN1+(i&1);
3767 int offset_reg = -1;
3768 int fastio_reg_override = -1;
3769 u_int reglist=get_host_reglist(i_regs->regmap);
3770 u_int copr=(source[i]>>16)&0x1f;
3771 s=get_reg(i_regs->regmap,dops[i].rs1);
3772 tl=get_reg(i_regs->regmap,FTEMP);
3773 offset=imm[i];
3774 assert(dops[i].rs1>0);
3775 assert(tl>=0);
3776
3777 if(i_regs->regmap[HOST_CCREG]==CCREG)
3778 reglist&=~(1<<HOST_CCREG);
3779
3780 // get the address
3781 if (dops[i].opcode==0x3a) { // SWC2
3782 ar=get_reg(i_regs->regmap,agr);
3783 if(ar<0) ar=get_reg_temp(i_regs->regmap);
3784 reglist|=1<<ar;
3785 } else { // LWC2
3786 ar=tl;
3787 }
3788 if(s>=0) c=(i_regs->wasconst>>s)&1;
3789 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3790 if (!offset&&!c&&s>=0) ar=s;
3791 assert(ar>=0);
3792
3793 cop2_do_stall_check(0, i, i_regs, reglist);
3794
3795 if (dops[i].opcode==0x3a) { // SWC2
3796 cop2_get_dreg(copr,tl,-1);
3797 type=STOREW_STUB;
3798 }
3799 else
3800 type=LOADW_STUB;
3801
3802 if(c&&!memtarget) {
3803 jaddr2=out;
3804 emit_jmp(0); // inline_readstub/inline_writestub?
3805 }
3806 else {
3807 if(!c) {
3808 jaddr2 = emit_fastpath_cmp_jump(i, i_regs, ar,
3809 &offset_reg, &fastio_reg_override);
3810 }
3811 else if (ram_offset && memtarget) {
3812 offset_reg = get_ro_reg(i_regs, 0);
3813 }
3814 switch (dops[i].opcode) {
3815 case 0x32: { // LWC2
3816 int a = ar;
3817 if (fastio_reg_override >= 0)
3818 a = fastio_reg_override;
3819 do_load_word(a, tl, offset_reg);
3820 break;
3821 }
3822 case 0x3a: { // SWC2
3823 #ifdef DESTRUCTIVE_SHIFT
3824 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3825 #endif
3826 int a = ar;
3827 if (fastio_reg_override >= 0)
3828 a = fastio_reg_override;
3829 do_store_word(a, 0, tl, offset_reg, 1);
3830 break;
3831 }
3832 default:
3833 assert(0);
3834 }
3835 }
3836 if (fastio_reg_override == HOST_TEMPREG || offset_reg == HOST_TEMPREG)
3837 host_tempreg_release();
3838 if(jaddr2)
3839 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj_,reglist);
3840 if(dops[i].opcode==0x3a) // SWC2
3841 if(!(i_regs->waswritten&(1<<dops[i].rs1)) && !HACK_ENABLED(NDHACK_NO_SMC_CHECK)) {
3842#if defined(HOST_IMM8)
3843 int ir=get_reg(i_regs->regmap,INVCP);
3844 assert(ir>=0);
3845 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3846#else
3847 emit_cmpmem_indexedsr12_imm(invalid_code,ar,1);
3848#endif
3849 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3850 emit_callne(invalidate_addr_reg[ar]);
3851 #else
3852 void *jaddr3 = out;
3853 emit_jne(0);
3854 add_stub(INVCODE_STUB,jaddr3,out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3855 #endif
3856 }
3857 if (dops[i].opcode==0x32) { // LWC2
3858 host_tempreg_acquire();
3859 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3860 host_tempreg_release();
3861 }
3862}
3863
3864static void cop2_assemble(int i, const struct regstat *i_regs)
3865{
3866 u_int copr = (source[i]>>11) & 0x1f;
3867 signed char temp = get_reg_temp(i_regs->regmap);
3868
3869 if (!HACK_ENABLED(NDHACK_NO_STALLS)) {
3870 u_int reglist = reglist_exclude(get_host_reglist(i_regs->regmap), temp, -1);
3871 if (dops[i].opcode2 == 0 || dops[i].opcode2 == 2) { // MFC2/CFC2
3872 signed char tl = get_reg(i_regs->regmap, dops[i].rt1);
3873 reglist = reglist_exclude(reglist, tl, -1);
3874 }
3875 cop2_do_stall_check(0, i, i_regs, reglist);
3876 }
3877 if (dops[i].opcode2==0) { // MFC2
3878 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3879 if(tl>=0&&dops[i].rt1!=0)
3880 cop2_get_dreg(copr,tl,temp);
3881 }
3882 else if (dops[i].opcode2==4) { // MTC2
3883 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3884 cop2_put_dreg(copr,sl,temp);
3885 }
3886 else if (dops[i].opcode2==2) // CFC2
3887 {
3888 signed char tl=get_reg(i_regs->regmap,dops[i].rt1);
3889 if(tl>=0&&dops[i].rt1!=0)
3890 emit_readword(&reg_cop2c[copr],tl);
3891 }
3892 else if (dops[i].opcode2==6) // CTC2
3893 {
3894 signed char sl=get_reg(i_regs->regmap,dops[i].rs1);
3895 switch(copr) {
3896 case 4:
3897 case 12:
3898 case 20:
3899 case 26:
3900 case 27:
3901 case 29:
3902 case 30:
3903 emit_signextend16(sl,temp);
3904 break;
3905 case 31:
3906 c2op_ctc2_31_assemble(sl,temp);
3907 break;
3908 default:
3909 temp=sl;
3910 break;
3911 }
3912 emit_writeword(temp,&reg_cop2c[copr]);
3913 assert(sl>=0);
3914 }
3915}
3916
3917static void do_unalignedwritestub(int n)
3918{
3919 assem_debug("do_unalignedwritestub %x\n",start+stubs[n].a*4);
3920 literal_pool(256);
3921 set_jump_target(stubs[n].addr, out);
3922
3923 int i=stubs[n].a;
3924 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3925 int addr=stubs[n].b;
3926 u_int reglist=stubs[n].e;
3927 signed char *i_regmap=i_regs->regmap;
3928 int temp2=get_reg(i_regmap,FTEMP);
3929 int rt;
3930 rt=get_reg(i_regmap,dops[i].rs2);
3931 assert(rt>=0);
3932 assert(addr>=0);
3933 assert(dops[i].opcode==0x2a||dops[i].opcode==0x2e); // SWL/SWR only implemented
3934 reglist|=(1<<addr);
3935 reglist&=~(1<<temp2);
3936
3937 // don't bother with it and call write handler
3938 save_regs(reglist);
3939 pass_args(addr,rt);
3940 int cc=get_reg(i_regmap,CCREG);
3941 if(cc<0)
3942 emit_loadreg(CCREG,2);
3943 emit_addimm(cc<0?2:cc,(int)stubs[n].d+1,2);
3944 emit_far_call((dops[i].opcode==0x2a?jump_handle_swl:jump_handle_swr));
3945 emit_addimm(0,-((int)stubs[n].d+1),cc<0?2:cc);
3946 if(cc<0)
3947 emit_storereg(CCREG,2);
3948 restore_regs(reglist);
3949 emit_jmp(stubs[n].retaddr); // return address
3950}
3951
3952#ifndef multdiv_assemble
3953void multdiv_assemble(int i,struct regstat *i_regs)
3954{
3955 printf("Need multdiv_assemble for this architecture.\n");
3956 abort();
3957}
3958#endif
3959
3960static void mov_assemble(int i, const struct regstat *i_regs)
3961{
3962 //if(dops[i].opcode2==0x10||dops[i].opcode2==0x12) { // MFHI/MFLO
3963 //if(dops[i].opcode2==0x11||dops[i].opcode2==0x13) { // MTHI/MTLO
3964 if(dops[i].rt1) {
3965 signed char sl,tl;
3966 tl=get_reg(i_regs->regmap,dops[i].rt1);
3967 //assert(tl>=0);
3968 if(tl>=0) {
3969 sl=get_reg(i_regs->regmap,dops[i].rs1);
3970 if(sl>=0) emit_mov(sl,tl);
3971 else emit_loadreg(dops[i].rs1,tl);
3972 }
3973 }
3974 if (dops[i].rs1 == HIREG || dops[i].rs1 == LOREG) // MFHI/MFLO
3975 multdiv_do_stall(i, i_regs);
3976}
3977
3978// call interpreter, exception handler, things that change pc/regs/cycles ...
3979static void call_c_cpu_handler(int i, const struct regstat *i_regs, int ccadj_, u_int pc, void *func)
3980{
3981 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3982 assert(ccreg==HOST_CCREG);
3983 assert(!is_delayslot);
3984 (void)ccreg;
3985
3986 emit_movimm(pc,3); // Get PC
3987 emit_readword(&last_count,2);
3988 emit_writeword(3,&psxRegs.pc);
3989 emit_addimm(HOST_CCREG,ccadj_,HOST_CCREG);
3990 emit_add(2,HOST_CCREG,2);
3991 emit_writeword(2,&psxRegs.cycle);
3992 emit_far_call(func);
3993 emit_far_jump(jump_to_new_pc);
3994}
3995
3996static void syscall_assemble(int i, const struct regstat *i_regs, int ccadj_)
3997{
3998 // 'break' tends to be littered around to catch things like
3999 // division by 0 and is almost never executed, so don't emit much code here
4000 void *func = (dops[i].opcode2 == 0x0C)
4001 ? (is_delayslot ? jump_syscall_ds : jump_syscall)
4002 : (is_delayslot ? jump_break_ds : jump_break);
4003 assert(get_reg(i_regs->regmap, CCREG) == HOST_CCREG);
4004 emit_movimm(start + i*4, 2); // pc
4005 emit_addimm(HOST_CCREG, ccadj_ + CLOCK_ADJUST(1), HOST_CCREG);
4006 emit_far_jump(func);
4007}
4008
4009static void hlecall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4010{
4011 void *hlefunc = psxNULL;
4012 uint32_t hleCode = source[i] & 0x03ffffff;
4013 if (hleCode < ARRAY_SIZE(psxHLEt))
4014 hlefunc = psxHLEt[hleCode];
4015
4016 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4+4, hlefunc);
4017}
4018
4019static void intcall_assemble(int i, const struct regstat *i_regs, int ccadj_)
4020{
4021 call_c_cpu_handler(i, i_regs, ccadj_, start + i*4, execI);
4022}
4023
4024static void speculate_mov(int rs,int rt)
4025{
4026 if(rt!=0) {
4027 smrv_strong_next|=1<<rt;
4028 smrv[rt]=smrv[rs];
4029 }
4030}
4031
4032static void speculate_mov_weak(int rs,int rt)
4033{
4034 if(rt!=0) {
4035 smrv_weak_next|=1<<rt;
4036 smrv[rt]=smrv[rs];
4037 }
4038}
4039
4040static void speculate_register_values(int i)
4041{
4042 if(i==0) {
4043 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
4044 // gp,sp are likely to stay the same throughout the block
4045 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
4046 smrv_weak_next=~smrv_strong_next;
4047 //printf(" llr %08x\n", smrv[4]);
4048 }
4049 smrv_strong=smrv_strong_next;
4050 smrv_weak=smrv_weak_next;
4051 switch(dops[i].itype) {
4052 case ALU:
4053 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4054 else if((smrv_strong>>dops[i].rs2)&1) speculate_mov(dops[i].rs2,dops[i].rt1);
4055 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4056 else if((smrv_weak>>dops[i].rs2)&1) speculate_mov_weak(dops[i].rs2,dops[i].rt1);
4057 else {
4058 smrv_strong_next&=~(1<<dops[i].rt1);
4059 smrv_weak_next&=~(1<<dops[i].rt1);
4060 }
4061 break;
4062 case SHIFTIMM:
4063 smrv_strong_next&=~(1<<dops[i].rt1);
4064 smrv_weak_next&=~(1<<dops[i].rt1);
4065 // fallthrough
4066 case IMM16:
4067 if(dops[i].rt1&&is_const(&regs[i],dops[i].rt1)) {
4068 int value,hr=get_reg(regs[i].regmap,dops[i].rt1);
4069 if(hr>=0) {
4070 if(get_final_value(hr,i,&value))
4071 smrv[dops[i].rt1]=value;
4072 else smrv[dops[i].rt1]=constmap[i][hr];
4073 smrv_strong_next|=1<<dops[i].rt1;
4074 }
4075 }
4076 else {
4077 if ((smrv_strong>>dops[i].rs1)&1) speculate_mov(dops[i].rs1,dops[i].rt1);
4078 else if((smrv_weak>>dops[i].rs1)&1) speculate_mov_weak(dops[i].rs1,dops[i].rt1);
4079 }
4080 break;
4081 case LOAD:
4082 if(start<0x2000&&(dops[i].rt1==26||(smrv[dops[i].rt1]>>24)==0xa0)) {
4083 // special case for BIOS
4084 smrv[dops[i].rt1]=0xa0000000;
4085 smrv_strong_next|=1<<dops[i].rt1;
4086 break;
4087 }
4088 // fallthrough
4089 case SHIFT:
4090 case LOADLR:
4091 case MOV:
4092 smrv_strong_next&=~(1<<dops[i].rt1);
4093 smrv_weak_next&=~(1<<dops[i].rt1);
4094 break;
4095 case COP0:
4096 case COP2:
4097 if(dops[i].opcode2==0||dops[i].opcode2==2) { // MFC/CFC
4098 smrv_strong_next&=~(1<<dops[i].rt1);
4099 smrv_weak_next&=~(1<<dops[i].rt1);
4100 }
4101 break;
4102 case C2LS:
4103 if (dops[i].opcode==0x32) { // LWC2
4104 smrv_strong_next&=~(1<<dops[i].rt1);
4105 smrv_weak_next&=~(1<<dops[i].rt1);
4106 }
4107 break;
4108 }
4109#if 0
4110 int r=4;
4111 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
4112 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
4113#endif
4114}
4115
4116static void ujump_assemble(int i, const struct regstat *i_regs);
4117static void rjump_assemble(int i, const struct regstat *i_regs);
4118static void cjump_assemble(int i, const struct regstat *i_regs);
4119static void sjump_assemble(int i, const struct regstat *i_regs);
4120static void pagespan_assemble(int i, const struct regstat *i_regs);
4121
4122static int assemble(int i, const struct regstat *i_regs, int ccadj_)
4123{
4124 int ds = 0;
4125 switch (dops[i].itype) {
4126 case ALU:
4127 alu_assemble(i, i_regs);
4128 break;
4129 case IMM16:
4130 imm16_assemble(i, i_regs);
4131 break;
4132 case SHIFT:
4133 shift_assemble(i, i_regs);
4134 break;
4135 case SHIFTIMM:
4136 shiftimm_assemble(i, i_regs);
4137 break;
4138 case LOAD:
4139 load_assemble(i, i_regs, ccadj_);
4140 break;
4141 case LOADLR:
4142 loadlr_assemble(i, i_regs, ccadj_);
4143 break;
4144 case STORE:
4145 store_assemble(i, i_regs, ccadj_);
4146 break;
4147 case STORELR:
4148 storelr_assemble(i, i_regs, ccadj_);
4149 break;
4150 case COP0:
4151 cop0_assemble(i, i_regs, ccadj_);
4152 break;
4153 case COP1:
4154 cop1_assemble(i, i_regs);
4155 break;
4156 case C1LS:
4157 c1ls_assemble(i, i_regs);
4158 break;
4159 case COP2:
4160 cop2_assemble(i, i_regs);
4161 break;
4162 case C2LS:
4163 c2ls_assemble(i, i_regs, ccadj_);
4164 break;
4165 case C2OP:
4166 c2op_assemble(i, i_regs);
4167 break;
4168 case MULTDIV:
4169 multdiv_assemble(i, i_regs);
4170 multdiv_prepare_stall(i, i_regs, ccadj_);
4171 break;
4172 case MOV:
4173 mov_assemble(i, i_regs);
4174 break;
4175 case SYSCALL:
4176 syscall_assemble(i, i_regs, ccadj_);
4177 break;
4178 case HLECALL:
4179 hlecall_assemble(i, i_regs, ccadj_);
4180 break;
4181 case INTCALL:
4182 intcall_assemble(i, i_regs, ccadj_);
4183 break;
4184 case UJUMP:
4185 ujump_assemble(i, i_regs);
4186 ds = 1;
4187 break;
4188 case RJUMP:
4189 rjump_assemble(i, i_regs);
4190 ds = 1;
4191 break;
4192 case CJUMP:
4193 cjump_assemble(i, i_regs);
4194 ds = 1;
4195 break;
4196 case SJUMP:
4197 sjump_assemble(i, i_regs);
4198 ds = 1;
4199 break;
4200 case SPAN:
4201 pagespan_assemble(i, i_regs);
4202 break;
4203 case NOP:
4204 case OTHER:
4205 case NI:
4206 // not handled, just skip
4207 break;
4208 default:
4209 assert(0);
4210 }
4211 return ds;
4212}
4213
4214static void ds_assemble(int i, const struct regstat *i_regs)
4215{
4216 speculate_register_values(i);
4217 is_delayslot = 1;
4218 switch (dops[i].itype) {
4219 case SYSCALL:
4220 case HLECALL:
4221 case INTCALL:
4222 case SPAN:
4223 case UJUMP:
4224 case RJUMP:
4225 case CJUMP:
4226 case SJUMP:
4227 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4228 break;
4229 default:
4230 assemble(i, i_regs, ccadj[i]);
4231 }
4232 is_delayslot = 0;
4233}
4234
4235// Is the branch target a valid internal jump?
4236static int internal_branch(int addr)
4237{
4238 if(addr&1) return 0; // Indirect (register) jump
4239 if(addr>=start && addr<start+slen*4-4)
4240 {
4241 return 1;
4242 }
4243 return 0;
4244}
4245
4246static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
4247{
4248 int hr;
4249 for(hr=0;hr<HOST_REGS;hr++) {
4250 if(hr!=EXCLUDE_REG) {
4251 if(pre[hr]!=entry[hr]) {
4252 if(pre[hr]>=0) {
4253 if((dirty>>hr)&1) {
4254 if(get_reg(entry,pre[hr])<0) {
4255 assert(pre[hr]<64);
4256 if(!((u>>pre[hr])&1))
4257 emit_storereg(pre[hr],hr);
4258 }
4259 }
4260 }
4261 }
4262 }
4263 }
4264 // Move from one register to another (no writeback)
4265 for(hr=0;hr<HOST_REGS;hr++) {
4266 if(hr!=EXCLUDE_REG) {
4267 if(pre[hr]!=entry[hr]) {
4268 if(pre[hr]>=0&&pre[hr]<TEMPREG) {
4269 int nr;
4270 if((nr=get_reg(entry,pre[hr]))>=0) {
4271 emit_mov(hr,nr);
4272 }
4273 }
4274 }
4275 }
4276 }
4277}
4278
4279// Load the specified registers
4280// This only loads the registers given as arguments because
4281// we don't want to load things that will be overwritten
4282static void load_regs(signed char entry[],signed char regmap[],int rs1,int rs2)
4283{
4284 int hr;
4285 // Load 32-bit regs
4286 for(hr=0;hr<HOST_REGS;hr++) {
4287 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4288 if(entry[hr]!=regmap[hr]) {
4289 if(regmap[hr]==rs1||regmap[hr]==rs2)
4290 {
4291 if(regmap[hr]==0) {
4292 emit_zeroreg(hr);
4293 }
4294 else
4295 {
4296 emit_loadreg(regmap[hr],hr);
4297 }
4298 }
4299 }
4300 }
4301 }
4302}
4303
4304// Load registers prior to the start of a loop
4305// so that they are not loaded within the loop
4306static void loop_preload(signed char pre[],signed char entry[])
4307{
4308 int hr;
4309 for(hr=0;hr<HOST_REGS;hr++) {
4310 if(hr!=EXCLUDE_REG) {
4311 if(pre[hr]!=entry[hr]) {
4312 if(entry[hr]>=0) {
4313 if(get_reg(pre,entry[hr])<0) {
4314 assem_debug("loop preload:\n");
4315 //printf("loop preload: %d\n",hr);
4316 if(entry[hr]==0) {
4317 emit_zeroreg(hr);
4318 }
4319 else if(entry[hr]<TEMPREG)
4320 {
4321 emit_loadreg(entry[hr],hr);
4322 }
4323 else if(entry[hr]-64<TEMPREG)
4324 {
4325 emit_loadreg(entry[hr],hr);
4326 }
4327 }
4328 }
4329 }
4330 }
4331 }
4332}
4333
4334// Generate address for load/store instruction
4335// goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
4336static void address_generation(int i, const struct regstat *i_regs, signed char entry[])
4337{
4338 if (dops[i].is_load || dops[i].is_store) {
4339 int ra=-1;
4340 int agr=AGEN1+(i&1);
4341 if(dops[i].itype==LOAD) {
4342 ra=get_reg(i_regs->regmap,dops[i].rt1);
4343 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4344 assert(ra>=0);
4345 }
4346 if(dops[i].itype==LOADLR) {
4347 ra=get_reg(i_regs->regmap,FTEMP);
4348 }
4349 if(dops[i].itype==STORE||dops[i].itype==STORELR) {
4350 ra=get_reg(i_regs->regmap,agr);
4351 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4352 }
4353 if(dops[i].itype==C2LS) {
4354 if ((dops[i].opcode&0x3b)==0x31||(dops[i].opcode&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
4355 ra=get_reg(i_regs->regmap,FTEMP);
4356 else { // SWC1/SDC1/SWC2/SDC2
4357 ra=get_reg(i_regs->regmap,agr);
4358 if(ra<0) ra=get_reg_temp(i_regs->regmap);
4359 }
4360 }
4361 int rs=get_reg(i_regs->regmap,dops[i].rs1);
4362 if(ra>=0) {
4363 int offset=imm[i];
4364 int c=(i_regs->wasconst>>rs)&1;
4365 if(dops[i].rs1==0) {
4366 // Using r0 as a base address
4367 if(!entry||entry[ra]!=agr) {
4368 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4369 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4370 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4371 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4372 }else{
4373 emit_movimm(offset,ra);
4374 }
4375 } // else did it in the previous cycle
4376 }
4377 else if(rs<0) {
4378 if(!entry||entry[ra]!=dops[i].rs1)
4379 emit_loadreg(dops[i].rs1,ra);
4380 //if(!entry||entry[ra]!=dops[i].rs1)
4381 // printf("poor load scheduling!\n");
4382 }
4383 else if(c) {
4384 if(dops[i].rs1!=dops[i].rt1||dops[i].itype!=LOAD) {
4385 if(!entry||entry[ra]!=agr) {
4386 if (dops[i].opcode==0x22||dops[i].opcode==0x26) {
4387 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4388 }else if (dops[i].opcode==0x1a||dops[i].opcode==0x1b) {
4389 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4390 }else{
4391 emit_movimm(constmap[i][rs]+offset,ra);
4392 regs[i].loadedconst|=1<<ra;
4393 }
4394 } // else did it in the previous cycle
4395 } // else load_consts already did it
4396 }
4397 if(offset&&!c&&dops[i].rs1) {
4398 if(rs>=0) {
4399 emit_addimm(rs,offset,ra);
4400 }else{
4401 emit_addimm(ra,offset,ra);
4402 }
4403 }
4404 }
4405 }
4406 // Preload constants for next instruction
4407 if (dops[i+1].is_load || dops[i+1].is_store) {
4408 int agr,ra;
4409 // Actual address
4410 agr=AGEN1+((i+1)&1);
4411 ra=get_reg(i_regs->regmap,agr);
4412 if(ra>=0) {
4413 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
4414 int offset=imm[i+1];
4415 int c=(regs[i+1].wasconst>>rs)&1;
4416 if(c&&(dops[i+1].rs1!=dops[i+1].rt1||dops[i+1].itype!=LOAD)) {
4417 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4418 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
4419 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4420 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
4421 }else{
4422 emit_movimm(constmap[i+1][rs]+offset,ra);
4423 regs[i+1].loadedconst|=1<<ra;
4424 }
4425 }
4426 else if(dops[i+1].rs1==0) {
4427 // Using r0 as a base address
4428 if (dops[i+1].opcode==0x22||dops[i+1].opcode==0x26) {
4429 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
4430 }else if (dops[i+1].opcode==0x1a||dops[i+1].opcode==0x1b) {
4431 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
4432 }else{
4433 emit_movimm(offset,ra);
4434 }
4435 }
4436 }
4437 }
4438}
4439
4440static int get_final_value(int hr, int i, int *value)
4441{
4442 int reg=regs[i].regmap[hr];
4443 while(i<slen-1) {
4444 if(regs[i+1].regmap[hr]!=reg) break;
4445 if(!((regs[i+1].isconst>>hr)&1)) break;
4446 if(dops[i+1].bt) break;
4447 i++;
4448 }
4449 if(i<slen-1) {
4450 if (dops[i].is_jump) {
4451 *value=constmap[i][hr];
4452 return 1;
4453 }
4454 if(!dops[i+1].bt) {
4455 if (dops[i+1].is_jump) {
4456 // Load in delay slot, out-of-order execution
4457 if(dops[i+2].itype==LOAD&&dops[i+2].rs1==reg&&dops[i+2].rt1==reg&&((regs[i+1].wasconst>>hr)&1))
4458 {
4459 // Precompute load address
4460 *value=constmap[i][hr]+imm[i+2];
4461 return 1;
4462 }
4463 }
4464 if(dops[i+1].itype==LOAD&&dops[i+1].rs1==reg&&dops[i+1].rt1==reg)
4465 {
4466 // Precompute load address
4467 *value=constmap[i][hr]+imm[i+1];
4468 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
4469 return 1;
4470 }
4471 }
4472 }
4473 *value=constmap[i][hr];
4474 //printf("c=%lx\n",(long)constmap[i][hr]);
4475 if(i==slen-1) return 1;
4476 assert(reg < 64);
4477 return !((unneeded_reg[i+1]>>reg)&1);
4478}
4479
4480// Load registers with known constants
4481static void load_consts(signed char pre[],signed char regmap[],int i)
4482{
4483 int hr,hr2;
4484 // propagate loaded constant flags
4485 if(i==0||dops[i].bt)
4486 regs[i].loadedconst=0;
4487 else {
4488 for(hr=0;hr<HOST_REGS;hr++) {
4489 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
4490 &&regmap[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
4491 {
4492 regs[i].loadedconst|=1<<hr;
4493 }
4494 }
4495 }
4496 // Load 32-bit regs
4497 for(hr=0;hr<HOST_REGS;hr++) {
4498 if(hr!=EXCLUDE_REG&&regmap[hr]>=0) {
4499 //if(entry[hr]!=regmap[hr]) {
4500 if(!((regs[i].loadedconst>>hr)&1)) {
4501 assert(regmap[hr]<64);
4502 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4503 int value,similar=0;
4504 if(get_final_value(hr,i,&value)) {
4505 // see if some other register has similar value
4506 for(hr2=0;hr2<HOST_REGS;hr2++) {
4507 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
4508 if(is_similar_value(value,constmap[i][hr2])) {
4509 similar=1;
4510 break;
4511 }
4512 }
4513 }
4514 if(similar) {
4515 int value2;
4516 if(get_final_value(hr2,i,&value2)) // is this needed?
4517 emit_movimm_from(value2,hr2,value,hr);
4518 else
4519 emit_movimm(value,hr);
4520 }
4521 else if(value==0) {
4522 emit_zeroreg(hr);
4523 }
4524 else {
4525 emit_movimm(value,hr);
4526 }
4527 }
4528 regs[i].loadedconst|=1<<hr;
4529 }
4530 }
4531 }
4532 }
4533}
4534
4535static void load_all_consts(const signed char regmap[], u_int dirty, int i)
4536{
4537 int hr;
4538 // Load 32-bit regs
4539 for(hr=0;hr<HOST_REGS;hr++) {
4540 if(hr!=EXCLUDE_REG&&regmap[hr]>=0&&((dirty>>hr)&1)) {
4541 assert(regmap[hr] < 64);
4542 if(((regs[i].isconst>>hr)&1)&&regmap[hr]>0) {
4543 int value=constmap[i][hr];
4544 if(value==0) {
4545 emit_zeroreg(hr);
4546 }
4547 else {
4548 emit_movimm(value,hr);
4549 }
4550 }
4551 }
4552 }
4553}
4554
4555// Write out all dirty registers (except cycle count)
4556static void wb_dirtys(const signed char i_regmap[], uint64_t i_dirty)
4557{
4558 int hr;
4559 for(hr=0;hr<HOST_REGS;hr++) {
4560 if(hr!=EXCLUDE_REG) {
4561 if(i_regmap[hr]>0) {
4562 if(i_regmap[hr]!=CCREG) {
4563 if((i_dirty>>hr)&1) {
4564 assert(i_regmap[hr]<64);
4565 emit_storereg(i_regmap[hr],hr);
4566 }
4567 }
4568 }
4569 }
4570 }
4571}
4572
4573// Write out dirty registers that we need to reload (pair with load_needed_regs)
4574// This writes the registers not written by store_regs_bt
4575static void wb_needed_dirtys(const signed char i_regmap[], uint64_t i_dirty, int addr)
4576{
4577 int hr;
4578 int t=(addr-start)>>2;
4579 for(hr=0;hr<HOST_REGS;hr++) {
4580 if(hr!=EXCLUDE_REG) {
4581 if(i_regmap[hr]>0) {
4582 if(i_regmap[hr]!=CCREG) {
4583 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
4584 if((i_dirty>>hr)&1) {
4585 assert(i_regmap[hr]<64);
4586 emit_storereg(i_regmap[hr],hr);
4587 }
4588 }
4589 }
4590 }
4591 }
4592 }
4593}
4594
4595// Load all registers (except cycle count)
4596static void load_all_regs(const signed char i_regmap[])
4597{
4598 int hr;
4599 for(hr=0;hr<HOST_REGS;hr++) {
4600 if(hr!=EXCLUDE_REG) {
4601 if(i_regmap[hr]==0) {
4602 emit_zeroreg(hr);
4603 }
4604 else
4605 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4606 {
4607 emit_loadreg(i_regmap[hr],hr);
4608 }
4609 }
4610 }
4611}
4612
4613// Load all current registers also needed by next instruction
4614static void load_needed_regs(const signed char i_regmap[], const signed char next_regmap[])
4615{
4616 int hr;
4617 for(hr=0;hr<HOST_REGS;hr++) {
4618 if(hr!=EXCLUDE_REG) {
4619 if(get_reg(next_regmap,i_regmap[hr])>=0) {
4620 if(i_regmap[hr]==0) {
4621 emit_zeroreg(hr);
4622 }
4623 else
4624 if(i_regmap[hr]>0 && i_regmap[hr]<TEMPREG && i_regmap[hr]!=CCREG)
4625 {
4626 emit_loadreg(i_regmap[hr],hr);
4627 }
4628 }
4629 }
4630 }
4631}
4632
4633// Load all regs, storing cycle count if necessary
4634static void load_regs_entry(int t)
4635{
4636 int hr;
4637 if(dops[t].is_ds) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
4638 else if(ccadj[t]) emit_addimm(HOST_CCREG,-ccadj[t],HOST_CCREG);
4639 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4640 emit_storereg(CCREG,HOST_CCREG);
4641 }
4642 // Load 32-bit regs
4643 for(hr=0;hr<HOST_REGS;hr++) {
4644 if(regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4645 if(regs[t].regmap_entry[hr]==0) {
4646 emit_zeroreg(hr);
4647 }
4648 else if(regs[t].regmap_entry[hr]!=CCREG)
4649 {
4650 emit_loadreg(regs[t].regmap_entry[hr],hr);
4651 }
4652 }
4653 }
4654}
4655
4656// Store dirty registers prior to branch
4657static void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4658{
4659 if(internal_branch(addr))
4660 {
4661 int t=(addr-start)>>2;
4662 int hr;
4663 for(hr=0;hr<HOST_REGS;hr++) {
4664 if(hr!=EXCLUDE_REG) {
4665 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
4666 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
4667 if((i_dirty>>hr)&1) {
4668 assert(i_regmap[hr]<64);
4669 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4670 emit_storereg(i_regmap[hr],hr);
4671 }
4672 }
4673 }
4674 }
4675 }
4676 }
4677 else
4678 {
4679 // Branch out of this block, write out all dirty regs
4680 wb_dirtys(i_regmap,i_dirty);
4681 }
4682}
4683
4684// Load all needed registers for branch target
4685static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4686{
4687 //if(addr>=start && addr<(start+slen*4))
4688 if(internal_branch(addr))
4689 {
4690 int t=(addr-start)>>2;
4691 int hr;
4692 // Store the cycle count before loading something else
4693 if(i_regmap[HOST_CCREG]!=CCREG) {
4694 assert(i_regmap[HOST_CCREG]==-1);
4695 }
4696 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4697 emit_storereg(CCREG,HOST_CCREG);
4698 }
4699 // Load 32-bit regs
4700 for(hr=0;hr<HOST_REGS;hr++) {
4701 if(hr!=EXCLUDE_REG&&regs[t].regmap_entry[hr]>=0&&regs[t].regmap_entry[hr]<TEMPREG) {
4702 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4703 if(regs[t].regmap_entry[hr]==0) {
4704 emit_zeroreg(hr);
4705 }
4706 else if(regs[t].regmap_entry[hr]!=CCREG)
4707 {
4708 emit_loadreg(regs[t].regmap_entry[hr],hr);
4709 }
4710 }
4711 }
4712 }
4713 }
4714}
4715
4716static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4717{
4718 if(addr>=start && addr<start+slen*4-4)
4719 {
4720 int t=(addr-start)>>2;
4721 int hr;
4722 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4723 for(hr=0;hr<HOST_REGS;hr++)
4724 {
4725 if(hr!=EXCLUDE_REG)
4726 {
4727 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4728 {
4729 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4730 {
4731 return 0;
4732 }
4733 else
4734 if((i_dirty>>hr)&1)
4735 {
4736 if(i_regmap[hr]<TEMPREG)
4737 {
4738 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4739 return 0;
4740 }
4741 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4742 {
4743 assert(0);
4744 }
4745 }
4746 }
4747 else // Same register but is it 32-bit or dirty?
4748 if(i_regmap[hr]>=0)
4749 {
4750 if(!((regs[t].dirty>>hr)&1))
4751 {
4752 if((i_dirty>>hr)&1)
4753 {
4754 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4755 {
4756 //printf("%x: dirty no match\n",addr);
4757 return 0;
4758 }
4759 }
4760 }
4761 }
4762 }
4763 }
4764 // Delay slots are not valid branch targets
4765 //if(t>0&&(dops[t-1].is_jump) return 0;
4766 // Delay slots require additional processing, so do not match
4767 if(dops[t].is_ds) return 0;
4768 }
4769 else
4770 {
4771 int hr;
4772 for(hr=0;hr<HOST_REGS;hr++)
4773 {
4774 if(hr!=EXCLUDE_REG)
4775 {
4776 if(i_regmap[hr]>=0)
4777 {
4778 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4779 {
4780 if((i_dirty>>hr)&1)
4781 {
4782 return 0;
4783 }
4784 }
4785 }
4786 }
4787 }
4788 }
4789 return 1;
4790}
4791
4792#ifdef DRC_DBG
4793static void drc_dbg_emit_do_cmp(int i, int ccadj_)
4794{
4795 extern void do_insn_cmp();
4796 //extern int cycle;
4797 u_int hr, reglist = get_host_reglist(regs[i].regmap);
4798
4799 assem_debug("//do_insn_cmp %08x\n", start+i*4);
4800 save_regs(reglist);
4801 // write out changed consts to match the interpreter
4802 if (i > 0 && !dops[i].bt) {
4803 for (hr = 0; hr < HOST_REGS; hr++) {
4804 int reg = regs[i].regmap_entry[hr]; // regs[i-1].regmap[hr];
4805 if (hr == EXCLUDE_REG || reg < 0)
4806 continue;
4807 if (!((regs[i-1].isconst >> hr) & 1))
4808 continue;
4809 if (i > 1 && reg == regs[i-2].regmap[hr] && constmap[i-1][hr] == constmap[i-2][hr])
4810 continue;
4811 emit_movimm(constmap[i-1][hr],0);
4812 emit_storereg(reg, 0);
4813 }
4814 }
4815 emit_movimm(start+i*4,0);
4816 emit_writeword(0,&pcaddr);
4817 int cc = get_reg(regs[i].regmap_entry, CCREG);
4818 if (cc < 0)
4819 emit_loadreg(CCREG, cc = 0);
4820 emit_addimm(cc, ccadj_, 0);
4821 emit_writeword(0, &psxRegs.cycle);
4822 emit_far_call(do_insn_cmp);
4823 //emit_readword(&cycle,0);
4824 //emit_addimm(0,2,0);
4825 //emit_writeword(0,&cycle);
4826 (void)get_reg2;
4827 restore_regs(reglist);
4828 assem_debug("\\\\do_insn_cmp\n");
4829}
4830#else
4831#define drc_dbg_emit_do_cmp(x,y)
4832#endif
4833
4834// Used when a branch jumps into the delay slot of another branch
4835static void ds_assemble_entry(int i)
4836{
4837 int t = (ba[i] - start) >> 2;
4838 int ccadj_ = -CLOCK_ADJUST(1);
4839 if (!instr_addr[t])
4840 instr_addr[t] = out;
4841 assem_debug("Assemble delay slot at %x\n",ba[i]);
4842 assem_debug("<->\n");
4843 drc_dbg_emit_do_cmp(t, ccadj_);
4844 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&&regs[t].regmap[HOST_CCREG]!=CCREG)
4845 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4846 load_regs(regs[t].regmap_entry,regs[t].regmap,dops[t].rs1,dops[t].rs2);
4847 address_generation(t,&regs[t],regs[t].regmap_entry);
4848 if (ram_offset && (dops[t].is_load || dops[t].is_store))
4849 load_regs(regs[t].regmap_entry,regs[t].regmap,ROREG,ROREG);
4850 if (dops[t].is_store)
4851 load_regs(regs[t].regmap_entry,regs[t].regmap,INVCP,INVCP);
4852 is_delayslot=0;
4853 switch (dops[t].itype) {
4854 case SYSCALL:
4855 case HLECALL:
4856 case INTCALL:
4857 case SPAN:
4858 case UJUMP:
4859 case RJUMP:
4860 case CJUMP:
4861 case SJUMP:
4862 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4863 break;
4864 default:
4865 assemble(t, &regs[t], ccadj_);
4866 }
4867 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4868 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4869 if(internal_branch(ba[i]+4))
4870 assem_debug("branch: internal\n");
4871 else
4872 assem_debug("branch: external\n");
4873 assert(internal_branch(ba[i]+4));
4874 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4875 emit_jmp(0);
4876}
4877
4878static void emit_extjump(void *addr, u_int target)
4879{
4880 emit_extjump2(addr, target, dyna_linker);
4881}
4882
4883static void emit_extjump_ds(void *addr, u_int target)
4884{
4885 emit_extjump2(addr, target, dyna_linker_ds);
4886}
4887
4888// Load 2 immediates optimizing for small code size
4889static void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2)
4890{
4891 emit_movimm(imm1,rt1);
4892 emit_movimm_from(imm1,rt1,imm2,rt2);
4893}
4894
4895static void do_cc(int i, const signed char i_regmap[], int *adj,
4896 int addr, int taken, int invert)
4897{
4898 int count, count_plus2;
4899 void *jaddr;
4900 void *idle=NULL;
4901 int t=0;
4902 if(dops[i].itype==RJUMP)
4903 {
4904 *adj=0;
4905 }
4906 //if(ba[i]>=start && ba[i]<(start+slen*4))
4907 if(internal_branch(ba[i]))
4908 {
4909 t=(ba[i]-start)>>2;
4910 if(dops[t].is_ds) *adj=-CLOCK_ADJUST(1); // Branch into delay slot adds an extra cycle
4911 else *adj=ccadj[t];
4912 }
4913 else
4914 {
4915 *adj=0;
4916 }
4917 count = ccadj[i];
4918 count_plus2 = count + CLOCK_ADJUST(2);
4919 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4920 // Idle loop
4921 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4922 idle=out;
4923 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4924 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4925 jaddr=out;
4926 emit_jmp(0);
4927 }
4928 else if(*adj==0||invert) {
4929 int cycles = count_plus2;
4930 // faster loop HACK
4931#if 0
4932 if (t&&*adj) {
4933 int rel=t-i;
4934 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4935 cycles=*adj+count+2-*adj;
4936 }
4937#endif
4938 emit_addimm_and_set_flags(cycles, HOST_CCREG);
4939 jaddr = out;
4940 emit_jns(0);
4941 }
4942 else
4943 {
4944 emit_cmpimm(HOST_CCREG, -count_plus2);
4945 jaddr = out;
4946 emit_jns(0);
4947 }
4948 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:count_plus2,i,addr,taken,0);
4949}
4950
4951static void do_ccstub(int n)
4952{
4953 literal_pool(256);
4954 assem_debug("do_ccstub %x\n",start+(u_int)stubs[n].b*4);
4955 set_jump_target(stubs[n].addr, out);
4956 int i=stubs[n].b;
4957 if(stubs[n].d==NULLDS) {
4958 // Delay slot instruction is nullified ("likely" branch)
4959 wb_dirtys(regs[i].regmap,regs[i].dirty);
4960 }
4961 else if(stubs[n].d!=TAKEN) {
4962 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
4963 }
4964 else {
4965 if(internal_branch(ba[i]))
4966 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4967 }
4968 if(stubs[n].c!=-1)
4969 {
4970 // Save PC as return address
4971 emit_movimm(stubs[n].c,EAX);
4972 emit_writeword(EAX,&pcaddr);
4973 }
4974 else
4975 {
4976 // Return address depends on which way the branch goes
4977 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
4978 {
4979 int s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
4980 int s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
4981 if(dops[i].rs1==0)
4982 {
4983 s1l=s2l;
4984 s2l=-1;
4985 }
4986 else if(dops[i].rs2==0)
4987 {
4988 s2l=-1;
4989 }
4990 assert(s1l>=0);
4991 #ifdef DESTRUCTIVE_WRITEBACK
4992 if(dops[i].rs1) {
4993 if((branch_regs[i].dirty>>s1l)&&1)
4994 emit_loadreg(dops[i].rs1,s1l);
4995 }
4996 else {
4997 if((branch_regs[i].dirty>>s1l)&1)
4998 emit_loadreg(dops[i].rs2,s1l);
4999 }
5000 if(s2l>=0)
5001 if((branch_regs[i].dirty>>s2l)&1)
5002 emit_loadreg(dops[i].rs2,s2l);
5003 #endif
5004 int hr=0;
5005 int addr=-1,alt=-1,ntaddr=-1;
5006 while(hr<HOST_REGS)
5007 {
5008 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5009 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5010 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5011 {
5012 addr=hr++;break;
5013 }
5014 hr++;
5015 }
5016 while(hr<HOST_REGS)
5017 {
5018 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5019 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5020 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5021 {
5022 alt=hr++;break;
5023 }
5024 hr++;
5025 }
5026 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5027 {
5028 while(hr<HOST_REGS)
5029 {
5030 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5031 branch_regs[i].regmap[hr]!=dops[i].rs1 &&
5032 branch_regs[i].regmap[hr]!=dops[i].rs2 )
5033 {
5034 ntaddr=hr;break;
5035 }
5036 hr++;
5037 }
5038 assert(hr<HOST_REGS);
5039 }
5040 if((dops[i].opcode&0x2f)==4) // BEQ
5041 {
5042 #ifdef HAVE_CMOV_IMM
5043 if(s2l>=0) emit_cmp(s1l,s2l);
5044 else emit_test(s1l,s1l);
5045 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5046 #else
5047 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5048 if(s2l>=0) emit_cmp(s1l,s2l);
5049 else emit_test(s1l,s1l);
5050 emit_cmovne_reg(alt,addr);
5051 #endif
5052 }
5053 if((dops[i].opcode&0x2f)==5) // BNE
5054 {
5055 #ifdef HAVE_CMOV_IMM
5056 if(s2l>=0) emit_cmp(s1l,s2l);
5057 else emit_test(s1l,s1l);
5058 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5059 #else
5060 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5061 if(s2l>=0) emit_cmp(s1l,s2l);
5062 else emit_test(s1l,s1l);
5063 emit_cmovne_reg(alt,addr);
5064 #endif
5065 }
5066 if((dops[i].opcode&0x2f)==6) // BLEZ
5067 {
5068 //emit_movimm(ba[i],alt);
5069 //emit_movimm(start+i*4+8,addr);
5070 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5071 emit_cmpimm(s1l,1);
5072 emit_cmovl_reg(alt,addr);
5073 }
5074 if((dops[i].opcode&0x2f)==7) // BGTZ
5075 {
5076 //emit_movimm(ba[i],addr);
5077 //emit_movimm(start+i*4+8,ntaddr);
5078 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5079 emit_cmpimm(s1l,1);
5080 emit_cmovl_reg(ntaddr,addr);
5081 }
5082 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==0) // BLTZ
5083 {
5084 //emit_movimm(ba[i],alt);
5085 //emit_movimm(start+i*4+8,addr);
5086 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5087 emit_test(s1l,s1l);
5088 emit_cmovs_reg(alt,addr);
5089 }
5090 if((dops[i].opcode==1)&&(dops[i].opcode2&0x2D)==1) // BGEZ
5091 {
5092 //emit_movimm(ba[i],addr);
5093 //emit_movimm(start+i*4+8,alt);
5094 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5095 emit_test(s1l,s1l);
5096 emit_cmovs_reg(alt,addr);
5097 }
5098 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
5099 if(source[i]&0x10000) // BC1T
5100 {
5101 //emit_movimm(ba[i],alt);
5102 //emit_movimm(start+i*4+8,addr);
5103 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5104 emit_testimm(s1l,0x800000);
5105 emit_cmovne_reg(alt,addr);
5106 }
5107 else // BC1F
5108 {
5109 //emit_movimm(ba[i],addr);
5110 //emit_movimm(start+i*4+8,alt);
5111 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5112 emit_testimm(s1l,0x800000);
5113 emit_cmovne_reg(alt,addr);
5114 }
5115 }
5116 emit_writeword(addr,&pcaddr);
5117 }
5118 else
5119 if(dops[i].itype==RJUMP)
5120 {
5121 int r=get_reg(branch_regs[i].regmap,dops[i].rs1);
5122 if (ds_writes_rjump_rs(i)) {
5123 r=get_reg(branch_regs[i].regmap,RTEMP);
5124 }
5125 emit_writeword(r,&pcaddr);
5126 }
5127 else {SysPrintf("Unknown branch type in do_ccstub\n");abort();}
5128 }
5129 // Update cycle count
5130 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
5131 if(stubs[n].a) emit_addimm(HOST_CCREG,(int)stubs[n].a,HOST_CCREG);
5132 emit_far_call(cc_interrupt);
5133 if(stubs[n].a) emit_addimm(HOST_CCREG,-(int)stubs[n].a,HOST_CCREG);
5134 if(stubs[n].d==TAKEN) {
5135 if(internal_branch(ba[i]))
5136 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
5137 else if(dops[i].itype==RJUMP) {
5138 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
5139 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
5140 else
5141 emit_loadreg(dops[i].rs1,get_reg(branch_regs[i].regmap,dops[i].rs1));
5142 }
5143 }else if(stubs[n].d==NOTTAKEN) {
5144 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
5145 else load_all_regs(branch_regs[i].regmap);
5146 }else if(stubs[n].d==NULLDS) {
5147 // Delay slot instruction is nullified ("likely" branch)
5148 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
5149 else load_all_regs(regs[i].regmap);
5150 }else{
5151 load_all_regs(branch_regs[i].regmap);
5152 }
5153 if (stubs[n].retaddr)
5154 emit_jmp(stubs[n].retaddr);
5155 else
5156 do_jump_vaddr(stubs[n].e);
5157}
5158
5159static void add_to_linker(void *addr, u_int target, int ext)
5160{
5161 assert(linkcount < ARRAY_SIZE(link_addr));
5162 link_addr[linkcount].addr = addr;
5163 link_addr[linkcount].target = target;
5164 link_addr[linkcount].ext = ext;
5165 linkcount++;
5166}
5167
5168static void ujump_assemble_write_ra(int i)
5169{
5170 int rt;
5171 unsigned int return_address;
5172 rt=get_reg(branch_regs[i].regmap,31);
5173 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]);
5174 //assert(rt>=0);
5175 return_address=start+i*4+8;
5176 if(rt>=0) {
5177 #ifdef USE_MINI_HT
5178 if(internal_branch(return_address)&&dops[i+1].rt1!=31) {
5179 int temp=-1; // note: must be ds-safe
5180 #ifdef HOST_TEMPREG
5181 temp=HOST_TEMPREG;
5182 #endif
5183 if(temp>=0) do_miniht_insert(return_address,rt,temp);
5184 else emit_movimm(return_address,rt);
5185 }
5186 else
5187 #endif
5188 {
5189 #ifdef REG_PREFETCH
5190 if(temp>=0)
5191 {
5192 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5193 }
5194 #endif
5195 emit_movimm(return_address,rt); // PC into link register
5196 #ifdef IMM_PREFETCH
5197 emit_prefetch(hash_table_get(return_address));
5198 #endif
5199 }
5200 }
5201}
5202
5203static void ujump_assemble(int i, const struct regstat *i_regs)
5204{
5205 int ra_done=0;
5206 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5207 address_generation(i+1,i_regs,regs[i].regmap_entry);
5208 #ifdef REG_PREFETCH
5209 int temp=get_reg(branch_regs[i].regmap,PTEMP);
5210 if(dops[i].rt1==31&&temp>=0)
5211 {
5212 signed char *i_regmap=i_regs->regmap;
5213 int return_address=start+i*4+8;
5214 if(get_reg(branch_regs[i].regmap,31)>0)
5215 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5216 }
5217 #endif
5218 if(dops[i].rt1==31&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5219 ujump_assemble_write_ra(i); // writeback ra for DS
5220 ra_done=1;
5221 }
5222 ds_assemble(i+1,i_regs);
5223 uint64_t bc_unneeded=branch_regs[i].u;
5224 bc_unneeded|=1|(1LL<<dops[i].rt1);
5225 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5226 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5227 if(!ra_done&&dops[i].rt1==31)
5228 ujump_assemble_write_ra(i);
5229 int cc,adj;
5230 cc=get_reg(branch_regs[i].regmap,CCREG);
5231 assert(cc==HOST_CCREG);
5232 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5233 #ifdef REG_PREFETCH
5234 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5235 #endif
5236 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5237 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5238 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5239 if(internal_branch(ba[i]))
5240 assem_debug("branch: internal\n");
5241 else
5242 assem_debug("branch: external\n");
5243 if (internal_branch(ba[i]) && dops[(ba[i]-start)>>2].is_ds) {
5244 ds_assemble_entry(i);
5245 }
5246 else {
5247 add_to_linker(out,ba[i],internal_branch(ba[i]));
5248 emit_jmp(0);
5249 }
5250}
5251
5252static void rjump_assemble_write_ra(int i)
5253{
5254 int rt,return_address;
5255 assert(dops[i+1].rt1!=dops[i].rt1);
5256 assert(dops[i+1].rt2!=dops[i].rt1);
5257 rt=get_reg(branch_regs[i].regmap,dops[i].rt1);
5258 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]);
5259 assert(rt>=0);
5260 return_address=start+i*4+8;
5261 #ifdef REG_PREFETCH
5262 if(temp>=0)
5263 {
5264 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5265 }
5266 #endif
5267 emit_movimm(return_address,rt); // PC into link register
5268 #ifdef IMM_PREFETCH
5269 emit_prefetch(hash_table_get(return_address));
5270 #endif
5271}
5272
5273static void rjump_assemble(int i, const struct regstat *i_regs)
5274{
5275 int temp;
5276 int rs,cc;
5277 int ra_done=0;
5278 rs=get_reg(branch_regs[i].regmap,dops[i].rs1);
5279 assert(rs>=0);
5280 if (ds_writes_rjump_rs(i)) {
5281 // Delay slot abuse, make a copy of the branch address register
5282 temp=get_reg(branch_regs[i].regmap,RTEMP);
5283 assert(temp>=0);
5284 assert(regs[i].regmap[temp]==RTEMP);
5285 emit_mov(rs,temp);
5286 rs=temp;
5287 }
5288 address_generation(i+1,i_regs,regs[i].regmap_entry);
5289 #ifdef REG_PREFETCH
5290 if(dops[i].rt1==31)
5291 {
5292 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
5293 signed char *i_regmap=i_regs->regmap;
5294 int return_address=start+i*4+8;
5295 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
5296 }
5297 }
5298 #endif
5299 #ifdef USE_MINI_HT
5300 if(dops[i].rs1==31) {
5301 int rh=get_reg(regs[i].regmap,RHASH);
5302 if(rh>=0) do_preload_rhash(rh);
5303 }
5304 #endif
5305 if(dops[i].rt1!=0&&(dops[i].rt1==dops[i+1].rs1||dops[i].rt1==dops[i+1].rs2)) {
5306 rjump_assemble_write_ra(i);
5307 ra_done=1;
5308 }
5309 ds_assemble(i+1,i_regs);
5310 uint64_t bc_unneeded=branch_regs[i].u;
5311 bc_unneeded|=1|(1LL<<dops[i].rt1);
5312 bc_unneeded&=~(1LL<<dops[i].rs1);
5313 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5314 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,CCREG);
5315 if(!ra_done&&dops[i].rt1!=0)
5316 rjump_assemble_write_ra(i);
5317 cc=get_reg(branch_regs[i].regmap,CCREG);
5318 assert(cc==HOST_CCREG);
5319 (void)cc;
5320 #ifdef USE_MINI_HT
5321 int rh=get_reg(branch_regs[i].regmap,RHASH);
5322 int ht=get_reg(branch_regs[i].regmap,RHTBL);
5323 if(dops[i].rs1==31) {
5324 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
5325 do_preload_rhtbl(ht);
5326 do_rhash(rs,rh);
5327 }
5328 #endif
5329 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5330 #ifdef DESTRUCTIVE_WRITEBACK
5331 if((branch_regs[i].dirty>>rs)&1) {
5332 if(dops[i].rs1!=dops[i+1].rt1&&dops[i].rs1!=dops[i+1].rt2) {
5333 emit_loadreg(dops[i].rs1,rs);
5334 }
5335 }
5336 #endif
5337 #ifdef REG_PREFETCH
5338 if(dops[i].rt1==31&&temp>=0) emit_prefetchreg(temp);
5339 #endif
5340 #ifdef USE_MINI_HT
5341 if(dops[i].rs1==31) {
5342 do_miniht_load(ht,rh);
5343 }
5344 #endif
5345 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
5346 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
5347 //assert(adj==0);
5348 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5349 add_stub(CC_STUB,out,NULL,0,i,-1,TAKEN,rs);
5350 if(dops[i+1].itype==COP0&&(source[i+1]&0x3f)==0x10)
5351 // special case for RFE
5352 emit_jmp(0);
5353 else
5354 emit_jns(0);
5355 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
5356 #ifdef USE_MINI_HT
5357 if(dops[i].rs1==31) {
5358 do_miniht_jump(rs,rh,ht);
5359 }
5360 else
5361 #endif
5362 {
5363 do_jump_vaddr(rs);
5364 }
5365 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5366 if(dops[i].rt1!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
5367 #endif
5368}
5369
5370static void cjump_assemble(int i, const struct regstat *i_regs)
5371{
5372 const signed char *i_regmap = i_regs->regmap;
5373 int cc;
5374 int match;
5375 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5376 assem_debug("match=%d\n",match);
5377 int s1l,s2l;
5378 int unconditional=0,nop=0;
5379 int invert=0;
5380 int internal=internal_branch(ba[i]);
5381 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5382 if(!match) invert=1;
5383 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5384 if(i>(ba[i]-start)>>2) invert=1;
5385 #endif
5386 #ifdef __aarch64__
5387 invert=1; // because of near cond. branches
5388 #endif
5389
5390 if(dops[i].ooo) {
5391 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5392 s2l=get_reg(branch_regs[i].regmap,dops[i].rs2);
5393 }
5394 else {
5395 s1l=get_reg(i_regmap,dops[i].rs1);
5396 s2l=get_reg(i_regmap,dops[i].rs2);
5397 }
5398 if(dops[i].rs1==0&&dops[i].rs2==0)
5399 {
5400 if(dops[i].opcode&1) nop=1;
5401 else unconditional=1;
5402 //assert(dops[i].opcode!=5);
5403 //assert(dops[i].opcode!=7);
5404 //assert(dops[i].opcode!=0x15);
5405 //assert(dops[i].opcode!=0x17);
5406 }
5407 else if(dops[i].rs1==0)
5408 {
5409 s1l=s2l;
5410 s2l=-1;
5411 }
5412 else if(dops[i].rs2==0)
5413 {
5414 s2l=-1;
5415 }
5416
5417 if(dops[i].ooo) {
5418 // Out of order execution (delay slot first)
5419 //printf("OOOE\n");
5420 address_generation(i+1,i_regs,regs[i].regmap_entry);
5421 ds_assemble(i+1,i_regs);
5422 int adj;
5423 uint64_t bc_unneeded=branch_regs[i].u;
5424 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5425 bc_unneeded|=1;
5426 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5427 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs2);
5428 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5429 cc=get_reg(branch_regs[i].regmap,CCREG);
5430 assert(cc==HOST_CCREG);
5431 if(unconditional)
5432 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5433 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5434 //assem_debug("cycle count (adj)\n");
5435 if(unconditional) {
5436 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5437 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5438 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5439 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5440 if(internal)
5441 assem_debug("branch: internal\n");
5442 else
5443 assem_debug("branch: external\n");
5444 if (internal && dops[(ba[i]-start)>>2].is_ds) {
5445 ds_assemble_entry(i);
5446 }
5447 else {
5448 add_to_linker(out,ba[i],internal);
5449 emit_jmp(0);
5450 }
5451 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5452 if(((u_int)out)&7) emit_addnop(0);
5453 #endif
5454 }
5455 }
5456 else if(nop) {
5457 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5458 void *jaddr=out;
5459 emit_jns(0);
5460 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5461 }
5462 else {
5463 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5464 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5465 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5466
5467 //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]);
5468 assert(s1l>=0);
5469 if(dops[i].opcode==4) // BEQ
5470 {
5471 if(s2l>=0) emit_cmp(s1l,s2l);
5472 else emit_test(s1l,s1l);
5473 if(invert){
5474 nottaken=out;
5475 emit_jne(DJT_1);
5476 }else{
5477 add_to_linker(out,ba[i],internal);
5478 emit_jeq(0);
5479 }
5480 }
5481 if(dops[i].opcode==5) // BNE
5482 {
5483 if(s2l>=0) emit_cmp(s1l,s2l);
5484 else emit_test(s1l,s1l);
5485 if(invert){
5486 nottaken=out;
5487 emit_jeq(DJT_1);
5488 }else{
5489 add_to_linker(out,ba[i],internal);
5490 emit_jne(0);
5491 }
5492 }
5493 if(dops[i].opcode==6) // BLEZ
5494 {
5495 emit_cmpimm(s1l,1);
5496 if(invert){
5497 nottaken=out;
5498 emit_jge(DJT_1);
5499 }else{
5500 add_to_linker(out,ba[i],internal);
5501 emit_jl(0);
5502 }
5503 }
5504 if(dops[i].opcode==7) // BGTZ
5505 {
5506 emit_cmpimm(s1l,1);
5507 if(invert){
5508 nottaken=out;
5509 emit_jl(DJT_1);
5510 }else{
5511 add_to_linker(out,ba[i],internal);
5512 emit_jge(0);
5513 }
5514 }
5515 if(invert) {
5516 if(taken) set_jump_target(taken, out);
5517 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5518 if (match && (!internal || !dops[(ba[i]-start)>>2].is_ds)) {
5519 if(adj) {
5520 emit_addimm(cc,-adj,cc);
5521 add_to_linker(out,ba[i],internal);
5522 }else{
5523 emit_addnop(13);
5524 add_to_linker(out,ba[i],internal*2);
5525 }
5526 emit_jmp(0);
5527 }else
5528 #endif
5529 {
5530 if(adj) emit_addimm(cc,-adj,cc);
5531 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5532 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5533 if(internal)
5534 assem_debug("branch: internal\n");
5535 else
5536 assem_debug("branch: external\n");
5537 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5538 ds_assemble_entry(i);
5539 }
5540 else {
5541 add_to_linker(out,ba[i],internal);
5542 emit_jmp(0);
5543 }
5544 }
5545 set_jump_target(nottaken, out);
5546 }
5547
5548 if(nottaken1) set_jump_target(nottaken1, out);
5549 if(adj) {
5550 if(!invert) emit_addimm(cc,adj,cc);
5551 }
5552 } // (!unconditional)
5553 } // if(ooo)
5554 else
5555 {
5556 // In-order execution (branch first)
5557 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
5558 if(!unconditional&&!nop) {
5559 //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]);
5560 assert(s1l>=0);
5561 if((dops[i].opcode&0x2f)==4) // BEQ
5562 {
5563 if(s2l>=0) emit_cmp(s1l,s2l);
5564 else emit_test(s1l,s1l);
5565 nottaken=out;
5566 emit_jne(DJT_2);
5567 }
5568 if((dops[i].opcode&0x2f)==5) // BNE
5569 {
5570 if(s2l>=0) emit_cmp(s1l,s2l);
5571 else emit_test(s1l,s1l);
5572 nottaken=out;
5573 emit_jeq(DJT_2);
5574 }
5575 if((dops[i].opcode&0x2f)==6) // BLEZ
5576 {
5577 emit_cmpimm(s1l,1);
5578 nottaken=out;
5579 emit_jge(DJT_2);
5580 }
5581 if((dops[i].opcode&0x2f)==7) // BGTZ
5582 {
5583 emit_cmpimm(s1l,1);
5584 nottaken=out;
5585 emit_jl(DJT_2);
5586 }
5587 } // if(!unconditional)
5588 int adj;
5589 uint64_t ds_unneeded=branch_regs[i].u;
5590 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5591 ds_unneeded|=1;
5592 // branch taken
5593 if(!nop) {
5594 if(taken) set_jump_target(taken, out);
5595 assem_debug("1:\n");
5596 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5597 // load regs
5598 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5599 address_generation(i+1,&branch_regs[i],0);
5600 if (ram_offset)
5601 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5602 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5603 ds_assemble(i+1,&branch_regs[i]);
5604 cc=get_reg(branch_regs[i].regmap,CCREG);
5605 if(cc==-1) {
5606 emit_loadreg(CCREG,cc=HOST_CCREG);
5607 // CHECK: Is the following instruction (fall thru) allocated ok?
5608 }
5609 assert(cc==HOST_CCREG);
5610 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5611 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5612 assem_debug("cycle count (adj)\n");
5613 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5614 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5615 if(internal)
5616 assem_debug("branch: internal\n");
5617 else
5618 assem_debug("branch: external\n");
5619 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5620 ds_assemble_entry(i);
5621 }
5622 else {
5623 add_to_linker(out,ba[i],internal);
5624 emit_jmp(0);
5625 }
5626 }
5627 // branch not taken
5628 if(!unconditional) {
5629 if(nottaken1) set_jump_target(nottaken1, out);
5630 set_jump_target(nottaken, out);
5631 assem_debug("2:\n");
5632 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5633 // load regs
5634 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5635 address_generation(i+1,&branch_regs[i],0);
5636 if (ram_offset)
5637 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5638 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5639 ds_assemble(i+1,&branch_regs[i]);
5640 cc=get_reg(branch_regs[i].regmap,CCREG);
5641 if (cc == -1) {
5642 // Cycle count isn't in a register, temporarily load it then write it out
5643 emit_loadreg(CCREG,HOST_CCREG);
5644 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5645 void *jaddr=out;
5646 emit_jns(0);
5647 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5648 emit_storereg(CCREG,HOST_CCREG);
5649 }
5650 else{
5651 cc=get_reg(i_regmap,CCREG);
5652 assert(cc==HOST_CCREG);
5653 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5654 void *jaddr=out;
5655 emit_jns(0);
5656 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5657 }
5658 }
5659 }
5660}
5661
5662static void sjump_assemble(int i, const struct regstat *i_regs)
5663{
5664 const signed char *i_regmap = i_regs->regmap;
5665 int cc;
5666 int match;
5667 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5668 assem_debug("smatch=%d ooo=%d\n", match, dops[i].ooo);
5669 int s1l;
5670 int unconditional=0,nevertaken=0;
5671 int invert=0;
5672 int internal=internal_branch(ba[i]);
5673 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
5674 if(!match) invert=1;
5675 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5676 if(i>(ba[i]-start)>>2) invert=1;
5677 #endif
5678 #ifdef __aarch64__
5679 invert=1; // because of near cond. branches
5680 #endif
5681
5682 //if(dops[i].opcode2>=0x10) return; // FIXME (BxxZAL)
5683 //assert(dops[i].opcode2<0x10||dops[i].rs1==0); // FIXME (BxxZAL)
5684
5685 if(dops[i].ooo) {
5686 s1l=get_reg(branch_regs[i].regmap,dops[i].rs1);
5687 }
5688 else {
5689 s1l=get_reg(i_regmap,dops[i].rs1);
5690 }
5691 if(dops[i].rs1==0)
5692 {
5693 if(dops[i].opcode2&1) unconditional=1;
5694 else nevertaken=1;
5695 // These are never taken (r0 is never less than zero)
5696 //assert(dops[i].opcode2!=0);
5697 //assert(dops[i].opcode2!=2);
5698 //assert(dops[i].opcode2!=0x10);
5699 //assert(dops[i].opcode2!=0x12);
5700 }
5701
5702 if(dops[i].ooo) {
5703 // Out of order execution (delay slot first)
5704 //printf("OOOE\n");
5705 address_generation(i+1,i_regs,regs[i].regmap_entry);
5706 ds_assemble(i+1,i_regs);
5707 int adj;
5708 uint64_t bc_unneeded=branch_regs[i].u;
5709 bc_unneeded&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
5710 bc_unneeded|=1;
5711 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5712 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i].rs1,dops[i].rs1);
5713 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5714 if(dops[i].rt1==31) {
5715 int rt,return_address;
5716 rt=get_reg(branch_regs[i].regmap,31);
5717 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]);
5718 if(rt>=0) {
5719 // Save the PC even if the branch is not taken
5720 return_address=start+i*4+8;
5721 emit_movimm(return_address,rt); // PC into link register
5722 #ifdef IMM_PREFETCH
5723 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5724 #endif
5725 }
5726 }
5727 cc=get_reg(branch_regs[i].regmap,CCREG);
5728 assert(cc==HOST_CCREG);
5729 if(unconditional)
5730 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5731 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5732 assem_debug("cycle count (adj)\n");
5733 if(unconditional) {
5734 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5735 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5736 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5737 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5738 if(internal)
5739 assem_debug("branch: internal\n");
5740 else
5741 assem_debug("branch: external\n");
5742 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5743 ds_assemble_entry(i);
5744 }
5745 else {
5746 add_to_linker(out,ba[i],internal);
5747 emit_jmp(0);
5748 }
5749 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5750 if(((u_int)out)&7) emit_addnop(0);
5751 #endif
5752 }
5753 }
5754 else if(nevertaken) {
5755 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5756 void *jaddr=out;
5757 emit_jns(0);
5758 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5759 }
5760 else {
5761 void *nottaken = NULL;
5762 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5763 if(adj&&!invert) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5764 {
5765 assert(s1l>=0);
5766 if((dops[i].opcode2&0xf)==0) // BLTZ/BLTZAL
5767 {
5768 emit_test(s1l,s1l);
5769 if(invert){
5770 nottaken=out;
5771 emit_jns(DJT_1);
5772 }else{
5773 add_to_linker(out,ba[i],internal);
5774 emit_js(0);
5775 }
5776 }
5777 if((dops[i].opcode2&0xf)==1) // BGEZ/BLTZAL
5778 {
5779 emit_test(s1l,s1l);
5780 if(invert){
5781 nottaken=out;
5782 emit_js(DJT_1);
5783 }else{
5784 add_to_linker(out,ba[i],internal);
5785 emit_jns(0);
5786 }
5787 }
5788 }
5789
5790 if(invert) {
5791 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5792 if (match && (!internal || !dops[(ba[i] - start) >> 2].is_ds)) {
5793 if(adj) {
5794 emit_addimm(cc,-adj,cc);
5795 add_to_linker(out,ba[i],internal);
5796 }else{
5797 emit_addnop(13);
5798 add_to_linker(out,ba[i],internal*2);
5799 }
5800 emit_jmp(0);
5801 }else
5802 #endif
5803 {
5804 if(adj) emit_addimm(cc,-adj,cc);
5805 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5806 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5807 if(internal)
5808 assem_debug("branch: internal\n");
5809 else
5810 assem_debug("branch: external\n");
5811 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5812 ds_assemble_entry(i);
5813 }
5814 else {
5815 add_to_linker(out,ba[i],internal);
5816 emit_jmp(0);
5817 }
5818 }
5819 set_jump_target(nottaken, out);
5820 }
5821
5822 if(adj) {
5823 if(!invert) emit_addimm(cc,adj,cc);
5824 }
5825 } // (!unconditional)
5826 } // if(ooo)
5827 else
5828 {
5829 // In-order execution (branch first)
5830 //printf("IOE\n");
5831 void *nottaken = NULL;
5832 if(dops[i].rt1==31) {
5833 int rt,return_address;
5834 rt=get_reg(branch_regs[i].regmap,31);
5835 if(rt>=0) {
5836 // Save the PC even if the branch is not taken
5837 return_address=start+i*4+8;
5838 emit_movimm(return_address,rt); // PC into link register
5839 #ifdef IMM_PREFETCH
5840 emit_prefetch(hash_table_get(return_address));
5841 #endif
5842 }
5843 }
5844 if(!unconditional) {
5845 //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]);
5846 assert(s1l>=0);
5847 if((dops[i].opcode2&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5848 {
5849 emit_test(s1l,s1l);
5850 nottaken=out;
5851 emit_jns(DJT_1);
5852 }
5853 if((dops[i].opcode2&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5854 {
5855 emit_test(s1l,s1l);
5856 nottaken=out;
5857 emit_js(DJT_1);
5858 }
5859 } // if(!unconditional)
5860 int adj;
5861 uint64_t ds_unneeded=branch_regs[i].u;
5862 ds_unneeded&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
5863 ds_unneeded|=1;
5864 // branch taken
5865 if(!nevertaken) {
5866 //assem_debug("1:\n");
5867 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5868 // load regs
5869 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5870 address_generation(i+1,&branch_regs[i],0);
5871 if (ram_offset)
5872 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5873 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5874 ds_assemble(i+1,&branch_regs[i]);
5875 cc=get_reg(branch_regs[i].regmap,CCREG);
5876 if(cc==-1) {
5877 emit_loadreg(CCREG,cc=HOST_CCREG);
5878 // CHECK: Is the following instruction (fall thru) allocated ok?
5879 }
5880 assert(cc==HOST_CCREG);
5881 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5882 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5883 assem_debug("cycle count (adj)\n");
5884 if(adj) emit_addimm(cc, ccadj[i] + CLOCK_ADJUST(2) - adj, cc);
5885 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5886 if(internal)
5887 assem_debug("branch: internal\n");
5888 else
5889 assem_debug("branch: external\n");
5890 if (internal && dops[(ba[i] - start) >> 2].is_ds) {
5891 ds_assemble_entry(i);
5892 }
5893 else {
5894 add_to_linker(out,ba[i],internal);
5895 emit_jmp(0);
5896 }
5897 }
5898 // branch not taken
5899 if(!unconditional) {
5900 set_jump_target(nottaken, out);
5901 assem_debug("1:\n");
5902 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5903 load_regs(regs[i].regmap,branch_regs[i].regmap,dops[i+1].rs1,dops[i+1].rs2);
5904 address_generation(i+1,&branch_regs[i],0);
5905 if (ram_offset)
5906 load_regs(regs[i].regmap,branch_regs[i].regmap,ROREG,ROREG);
5907 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5908 ds_assemble(i+1,&branch_regs[i]);
5909 cc=get_reg(branch_regs[i].regmap,CCREG);
5910 if (cc == -1) {
5911 // Cycle count isn't in a register, temporarily load it then write it out
5912 emit_loadreg(CCREG,HOST_CCREG);
5913 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5914 void *jaddr=out;
5915 emit_jns(0);
5916 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5917 emit_storereg(CCREG,HOST_CCREG);
5918 }
5919 else{
5920 cc=get_reg(i_regmap,CCREG);
5921 assert(cc==HOST_CCREG);
5922 emit_addimm_and_set_flags(ccadj[i] + CLOCK_ADJUST(2), cc);
5923 void *jaddr=out;
5924 emit_jns(0);
5925 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5926 }
5927 }
5928 }
5929}
5930
5931static void pagespan_assemble(int i, const struct regstat *i_regs)
5932{
5933 int s1l=get_reg(i_regs->regmap,dops[i].rs1);
5934 int s2l=get_reg(i_regs->regmap,dops[i].rs2);
5935 void *taken = NULL;
5936 void *nottaken = NULL;
5937 int unconditional=0;
5938 if(dops[i].rs1==0)
5939 {
5940 s1l=s2l;
5941 s2l=-1;
5942 }
5943 else if(dops[i].rs2==0)
5944 {
5945 s2l=-1;
5946 }
5947 int hr=0;
5948 int addr=-1,alt=-1,ntaddr=-1;
5949 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5950 else {
5951 while(hr<HOST_REGS)
5952 {
5953 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5954 i_regs->regmap[hr]!=dops[i].rs1 &&
5955 i_regs->regmap[hr]!=dops[i].rs2 )
5956 {
5957 addr=hr++;break;
5958 }
5959 hr++;
5960 }
5961 }
5962 while(hr<HOST_REGS)
5963 {
5964 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5965 i_regs->regmap[hr]!=dops[i].rs1 &&
5966 i_regs->regmap[hr]!=dops[i].rs2 )
5967 {
5968 alt=hr++;break;
5969 }
5970 hr++;
5971 }
5972 if((dops[i].opcode&0x2E)==6) // BLEZ/BGTZ needs another register
5973 {
5974 while(hr<HOST_REGS)
5975 {
5976 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5977 i_regs->regmap[hr]!=dops[i].rs1 &&
5978 i_regs->regmap[hr]!=dops[i].rs2 )
5979 {
5980 ntaddr=hr;break;
5981 }
5982 hr++;
5983 }
5984 }
5985 assert(hr<HOST_REGS);
5986 if((dops[i].opcode&0x2e)==4||dops[i].opcode==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5987 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
5988 }
5989 emit_addimm(HOST_CCREG, ccadj[i] + CLOCK_ADJUST(2), HOST_CCREG);
5990 if(dops[i].opcode==2) // J
5991 {
5992 unconditional=1;
5993 }
5994 if(dops[i].opcode==3) // JAL
5995 {
5996 // TODO: mini_ht
5997 int rt=get_reg(i_regs->regmap,31);
5998 emit_movimm(start+i*4+8,rt);
5999 unconditional=1;
6000 }
6001 if(dops[i].opcode==0&&(dops[i].opcode2&0x3E)==8) // JR/JALR
6002 {
6003 emit_mov(s1l,addr);
6004 if(dops[i].opcode2==9) // JALR
6005 {
6006 int rt=get_reg(i_regs->regmap,dops[i].rt1);
6007 emit_movimm(start+i*4+8,rt);
6008 }
6009 }
6010 if((dops[i].opcode&0x3f)==4) // BEQ
6011 {
6012 if(dops[i].rs1==dops[i].rs2)
6013 {
6014 unconditional=1;
6015 }
6016 else
6017 #ifdef HAVE_CMOV_IMM
6018 if(1) {
6019 if(s2l>=0) emit_cmp(s1l,s2l);
6020 else emit_test(s1l,s1l);
6021 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
6022 }
6023 else
6024 #endif
6025 {
6026 assert(s1l>=0);
6027 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6028 if(s2l>=0) emit_cmp(s1l,s2l);
6029 else emit_test(s1l,s1l);
6030 emit_cmovne_reg(alt,addr);
6031 }
6032 }
6033 if((dops[i].opcode&0x3f)==5) // BNE
6034 {
6035 #ifdef HAVE_CMOV_IMM
6036 if(s2l>=0) emit_cmp(s1l,s2l);
6037 else emit_test(s1l,s1l);
6038 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
6039 #else
6040 assert(s1l>=0);
6041 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
6042 if(s2l>=0) emit_cmp(s1l,s2l);
6043 else emit_test(s1l,s1l);
6044 emit_cmovne_reg(alt,addr);
6045 #endif
6046 }
6047 if((dops[i].opcode&0x3f)==0x14) // BEQL
6048 {
6049 if(s2l>=0) emit_cmp(s1l,s2l);
6050 else emit_test(s1l,s1l);
6051 if(nottaken) set_jump_target(nottaken, out);
6052 nottaken=out;
6053 emit_jne(0);
6054 }
6055 if((dops[i].opcode&0x3f)==0x15) // BNEL
6056 {
6057 if(s2l>=0) emit_cmp(s1l,s2l);
6058 else emit_test(s1l,s1l);
6059 nottaken=out;
6060 emit_jeq(0);
6061 if(taken) set_jump_target(taken, out);
6062 }
6063 if((dops[i].opcode&0x3f)==6) // BLEZ
6064 {
6065 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6066 emit_cmpimm(s1l,1);
6067 emit_cmovl_reg(alt,addr);
6068 }
6069 if((dops[i].opcode&0x3f)==7) // BGTZ
6070 {
6071 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
6072 emit_cmpimm(s1l,1);
6073 emit_cmovl_reg(ntaddr,addr);
6074 }
6075 if((dops[i].opcode&0x3f)==0x16) // BLEZL
6076 {
6077 assert((dops[i].opcode&0x3f)!=0x16);
6078 }
6079 if((dops[i].opcode&0x3f)==0x17) // BGTZL
6080 {
6081 assert((dops[i].opcode&0x3f)!=0x17);
6082 }
6083 assert(dops[i].opcode!=1); // BLTZ/BGEZ
6084
6085 //FIXME: Check CSREG
6086 if(dops[i].opcode==0x11 && dops[i].opcode2==0x08 ) {
6087 if((source[i]&0x30000)==0) // BC1F
6088 {
6089 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
6090 emit_testimm(s1l,0x800000);
6091 emit_cmovne_reg(alt,addr);
6092 }
6093 if((source[i]&0x30000)==0x10000) // BC1T
6094 {
6095 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
6096 emit_testimm(s1l,0x800000);
6097 emit_cmovne_reg(alt,addr);
6098 }
6099 if((source[i]&0x30000)==0x20000) // BC1FL
6100 {
6101 emit_testimm(s1l,0x800000);
6102 nottaken=out;
6103 emit_jne(0);
6104 }
6105 if((source[i]&0x30000)==0x30000) // BC1TL
6106 {
6107 emit_testimm(s1l,0x800000);
6108 nottaken=out;
6109 emit_jeq(0);
6110 }
6111 }
6112
6113 assert(i_regs->regmap[HOST_CCREG]==CCREG);
6114 wb_dirtys(regs[i].regmap,regs[i].dirty);
6115 if(unconditional)
6116 {
6117 emit_movimm(ba[i],HOST_BTREG);
6118 }
6119 else if(addr!=HOST_BTREG)
6120 {
6121 emit_mov(addr,HOST_BTREG);
6122 }
6123 void *branch_addr=out;
6124 emit_jmp(0);
6125 int target_addr=start+i*4+5;
6126 void *stub=out;
6127 void *compiled_target_addr=check_addr(target_addr);
6128 emit_extjump_ds(branch_addr, target_addr);
6129 if(compiled_target_addr) {
6130 set_jump_target(branch_addr, compiled_target_addr);
6131 add_jump_out(target_addr,stub);
6132 }
6133 else set_jump_target(branch_addr, stub);
6134}
6135
6136// Assemble the delay slot for the above
6137static void pagespan_ds()
6138{
6139 assem_debug("initial delay slot:\n");
6140 u_int vaddr=start+1;
6141 u_int page=get_page(vaddr);
6142 u_int vpage=get_vpage(vaddr);
6143 ll_add(jump_dirty+vpage,vaddr,(void *)out);
6144 do_dirty_stub_ds(slen*4);
6145 ll_add(jump_in+page,vaddr,(void *)out);
6146 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
6147 if(regs[0].regmap[HOST_CCREG]!=CCREG)
6148 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
6149 if(regs[0].regmap[HOST_BTREG]!=BTREG)
6150 emit_writeword(HOST_BTREG,&branch_target);
6151 load_regs(regs[0].regmap_entry,regs[0].regmap,dops[0].rs1,dops[0].rs2);
6152 address_generation(0,&regs[0],regs[0].regmap_entry);
6153 if (ram_offset && (dops[0].is_load || dops[0].is_store))
6154 load_regs(regs[0].regmap_entry,regs[0].regmap,ROREG,ROREG);
6155 if (dops[0].is_store)
6156 load_regs(regs[0].regmap_entry,regs[0].regmap,INVCP,INVCP);
6157 is_delayslot=0;
6158 switch (dops[0].itype) {
6159 case SYSCALL:
6160 case HLECALL:
6161 case INTCALL:
6162 case SPAN:
6163 case UJUMP:
6164 case RJUMP:
6165 case CJUMP:
6166 case SJUMP:
6167 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
6168 break;
6169 default:
6170 assemble(0, &regs[0], 0);
6171 }
6172 int btaddr=get_reg(regs[0].regmap,BTREG);
6173 if(btaddr<0) {
6174 btaddr=get_reg_temp(regs[0].regmap);
6175 emit_readword(&branch_target,btaddr);
6176 }
6177 assert(btaddr!=HOST_CCREG);
6178 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
6179#ifdef HOST_IMM8
6180 host_tempreg_acquire();
6181 emit_movimm(start+4,HOST_TEMPREG);
6182 emit_cmp(btaddr,HOST_TEMPREG);
6183 host_tempreg_release();
6184#else
6185 emit_cmpimm(btaddr,start+4);
6186#endif
6187 void *branch = out;
6188 emit_jeq(0);
6189 store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
6190 do_jump_vaddr(btaddr);
6191 set_jump_target(branch, out);
6192 store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
6193 load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
6194}
6195
6196static void check_regmap(signed char *regmap)
6197{
6198#ifndef NDEBUG
6199 int i,j;
6200 for (i = 0; i < HOST_REGS; i++) {
6201 if (regmap[i] < 0)
6202 continue;
6203 for (j = i + 1; j < HOST_REGS; j++)
6204 assert(regmap[i] != regmap[j]);
6205 }
6206#endif
6207}
6208
6209#ifdef DISASM
6210#include <inttypes.h>
6211static char insn[MAXBLOCK][10];
6212
6213#define set_mnemonic(i_, n_) \
6214 strcpy(insn[i_], n_)
6215
6216void print_regmap(const char *name, const signed char *regmap)
6217{
6218 char buf[5];
6219 int i, l;
6220 fputs(name, stdout);
6221 for (i = 0; i < HOST_REGS; i++) {
6222 l = 0;
6223 if (regmap[i] >= 0)
6224 l = snprintf(buf, sizeof(buf), "$%d", regmap[i]);
6225 for (; l < 3; l++)
6226 buf[l] = ' ';
6227 buf[l] = 0;
6228 printf(" r%d=%s", i, buf);
6229 }
6230 fputs("\n", stdout);
6231}
6232
6233 /* disassembly */
6234void disassemble_inst(int i)
6235{
6236 if (dops[i].bt) printf("*"); else printf(" ");
6237 switch(dops[i].itype) {
6238 case UJUMP:
6239 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6240 case CJUMP:
6241 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;
6242 case SJUMP:
6243 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;
6244 case RJUMP:
6245 if (dops[i].opcode==0x9&&dops[i].rt1!=31)
6246 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1);
6247 else
6248 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6249 break;
6250 case SPAN:
6251 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2,ba[i]);break;
6252 case IMM16:
6253 if(dops[i].opcode==0xf) //LUI
6254 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],dops[i].rt1,imm[i]&0xffff);
6255 else
6256 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6257 break;
6258 case LOAD:
6259 case LOADLR:
6260 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6261 break;
6262 case STORE:
6263 case STORELR:
6264 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],dops[i].rs2,dops[i].rs1,imm[i]);
6265 break;
6266 case ALU:
6267 case SHIFT:
6268 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,dops[i].rs2);
6269 break;
6270 case MULTDIV:
6271 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],dops[i].rs1,dops[i].rs2);
6272 break;
6273 case SHIFTIMM:
6274 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],dops[i].rt1,dops[i].rs1,imm[i]);
6275 break;
6276 case MOV:
6277 if((dops[i].opcode2&0x1d)==0x10)
6278 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rt1);
6279 else if((dops[i].opcode2&0x1d)==0x11)
6280 printf (" %x: %s r%d\n",start+i*4,insn[i],dops[i].rs1);
6281 else
6282 printf (" %x: %s\n",start+i*4,insn[i]);
6283 break;
6284 case COP0:
6285 if(dops[i].opcode2==0)
6286 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC0
6287 else if(dops[i].opcode2==4)
6288 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC0
6289 else printf (" %x: %s\n",start+i*4,insn[i]);
6290 break;
6291 case COP1:
6292 if(dops[i].opcode2<3)
6293 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC1
6294 else if(dops[i].opcode2>3)
6295 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC1
6296 else printf (" %x: %s\n",start+i*4,insn[i]);
6297 break;
6298 case COP2:
6299 if(dops[i].opcode2<3)
6300 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rt1,(source[i]>>11)&0x1f); // MFC2
6301 else if(dops[i].opcode2>3)
6302 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],dops[i].rs1,(source[i]>>11)&0x1f); // MTC2
6303 else printf (" %x: %s\n",start+i*4,insn[i]);
6304 break;
6305 case C1LS:
6306 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6307 break;
6308 case C2LS:
6309 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,dops[i].rs1,imm[i]);
6310 break;
6311 case INTCALL:
6312 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6313 break;
6314 default:
6315 //printf (" %s %8x\n",insn[i],source[i]);
6316 printf (" %x: %s\n",start+i*4,insn[i]);
6317 }
6318 return;
6319 printf("D: %"PRIu64" WD: %"PRIu64" U: %"PRIu64"\n",
6320 regs[i].dirty, regs[i].wasdirty, unneeded_reg[i]);
6321 print_regmap("pre: ", regmap_pre[i]);
6322 print_regmap("entry: ", regs[i].regmap_entry);
6323 print_regmap("map: ", regs[i].regmap);
6324 if (dops[i].is_jump) {
6325 print_regmap("bentry:", branch_regs[i].regmap_entry);
6326 print_regmap("bmap: ", branch_regs[i].regmap);
6327 }
6328}
6329#else
6330#define set_mnemonic(i_, n_)
6331static void disassemble_inst(int i) {}
6332#endif // DISASM
6333
6334#define DRC_TEST_VAL 0x74657374
6335
6336static void new_dynarec_test(void)
6337{
6338 int (*testfunc)(void);
6339 void *beginning;
6340 int ret[2];
6341 size_t i;
6342
6343 // check structure linkage
6344 if ((u_char *)rcnts - (u_char *)&psxRegs != sizeof(psxRegs))
6345 {
6346 SysPrintf("linkage_arm* miscompilation/breakage detected.\n");
6347 }
6348
6349 SysPrintf("testing if we can run recompiled code @%p...\n", out);
6350 ((volatile u_int *)out)[0]++; // make cache dirty
6351
6352 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6353 out = ndrc->translation_cache;
6354 beginning = start_block();
6355 emit_movimm(DRC_TEST_VAL + i, 0); // test
6356 emit_ret();
6357 literal_pool(0);
6358 end_block(beginning);
6359 testfunc = beginning;
6360 ret[i] = testfunc();
6361 }
6362
6363 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6364 SysPrintf("test passed.\n");
6365 else
6366 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6367 out = ndrc->translation_cache;
6368}
6369
6370// clear the state completely, instead of just marking
6371// things invalid like invalidate_all_pages() does
6372void new_dynarec_clear_full(void)
6373{
6374 int n;
6375 out = ndrc->translation_cache;
6376 memset(invalid_code,1,sizeof(invalid_code));
6377 memset(hash_table,0xff,sizeof(hash_table));
6378 memset(mini_ht,-1,sizeof(mini_ht));
6379 memset(restore_candidate,0,sizeof(restore_candidate));
6380 memset(shadow,0,sizeof(shadow));
6381 copy=shadow;
6382 expirep=16384; // Expiry pointer, +2 blocks
6383 pending_exception=0;
6384 literalcount=0;
6385 stop_after_jal=0;
6386 inv_code_start=inv_code_end=~0;
6387 hack_addr=0;
6388 f1_hack=0;
6389 // TLB
6390 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6391 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6392 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6393
6394 cycle_multiplier_old = cycle_multiplier;
6395 new_dynarec_hacks_old = new_dynarec_hacks;
6396}
6397
6398void new_dynarec_init(void)
6399{
6400 SysPrintf("Init new dynarec, ndrc size %x\n", (int)sizeof(*ndrc));
6401
6402#ifdef _3DS
6403 check_rosalina();
6404#endif
6405#ifdef BASE_ADDR_DYNAMIC
6406 #ifdef VITA
6407 sceBlock = getVMBlock(); //sceKernelAllocMemBlockForVM("code", sizeof(*ndrc));
6408 if (sceBlock <= 0)
6409 SysPrintf("sceKernelAllocMemBlockForVM failed: %x\n", sceBlock);
6410 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&ndrc);
6411 if (ret < 0)
6412 SysPrintf("sceKernelGetMemBlockBase failed: %x\n", ret);
6413 sceKernelOpenVMDomain();
6414 sceClibPrintf("translation_cache = 0x%08lx\n ", (long)ndrc->translation_cache);
6415 #elif defined(_MSC_VER)
6416 ndrc = VirtualAlloc(NULL, sizeof(*ndrc), MEM_COMMIT | MEM_RESERVE,
6417 PAGE_EXECUTE_READWRITE);
6418 #else
6419 uintptr_t desired_addr = 0;
6420 #ifdef __ELF__
6421 extern char _end;
6422 desired_addr = ((uintptr_t)&_end + 0xffffff) & ~0xffffffl;
6423 #endif
6424 ndrc = mmap((void *)desired_addr, sizeof(*ndrc),
6425 PROT_READ | PROT_WRITE | PROT_EXEC,
6426 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
6427 if (ndrc == MAP_FAILED) {
6428 SysPrintf("mmap() failed: %s\n", strerror(errno));
6429 abort();
6430 }
6431 #endif
6432#else
6433 #ifndef NO_WRITE_EXEC
6434 // not all systems allow execute in data segment by default
6435 // size must be 4K aligned for 3DS?
6436 if (mprotect(ndrc, sizeof(*ndrc),
6437 PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6438 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6439 #endif
6440#endif
6441 out = ndrc->translation_cache;
6442 cycle_multiplier=200;
6443 new_dynarec_clear_full();
6444#ifdef HOST_IMM8
6445 // Copy this into local area so we don't have to put it in every literal pool
6446 invc_ptr=invalid_code;
6447#endif
6448 arch_init();
6449 new_dynarec_test();
6450 ram_offset=(uintptr_t)rdram-0x80000000;
6451 if (ram_offset!=0)
6452 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6453}
6454
6455void new_dynarec_cleanup(void)
6456{
6457 int n;
6458#ifdef BASE_ADDR_DYNAMIC
6459 #ifdef VITA
6460 // sceBlock is managed by retroarch's bootstrap code
6461 //sceKernelFreeMemBlock(sceBlock);
6462 //sceBlock = -1;
6463 #else
6464 if (munmap(ndrc, sizeof(*ndrc)) < 0)
6465 SysPrintf("munmap() failed\n");
6466 #endif
6467#endif
6468 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6469 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6470 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6471 #ifdef ROM_COPY
6472 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
6473 #endif
6474}
6475
6476static u_int *get_source_start(u_int addr, u_int *limit)
6477{
6478 if (addr < 0x00200000 ||
6479 (0xa0000000 <= addr && addr < 0xa0200000))
6480 {
6481 // used for BIOS calls mostly?
6482 *limit = (addr&0xa0000000)|0x00200000;
6483 return (u_int *)(rdram + (addr&0x1fffff));
6484 }
6485 else if (!Config.HLE && (
6486 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6487 (0xbfc00000 <= addr && addr < 0xbfc80000)))
6488 {
6489 // BIOS. The multiplier should be much higher as it's uncached 8bit mem,
6490 // but timings in PCSX are too tied to the interpreter's BIAS
6491 if (!HACK_ENABLED(NDHACK_OVERRIDE_CYCLE_M))
6492 cycle_multiplier_active = 200;
6493
6494 *limit = (addr & 0xfff00000) | 0x80000;
6495 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6496 }
6497 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6498 *limit = (addr & 0x80600000) + 0x00200000;
6499 return (u_int *)(rdram + (addr&0x1fffff));
6500 }
6501 return NULL;
6502}
6503
6504static u_int scan_for_ret(u_int addr)
6505{
6506 u_int limit = 0;
6507 u_int *mem;
6508
6509 mem = get_source_start(addr, &limit);
6510 if (mem == NULL)
6511 return addr;
6512
6513 if (limit > addr + 0x1000)
6514 limit = addr + 0x1000;
6515 for (; addr < limit; addr += 4, mem++) {
6516 if (*mem == 0x03e00008) // jr $ra
6517 return addr + 8;
6518 }
6519 return addr;
6520}
6521
6522struct savestate_block {
6523 uint32_t addr;
6524 uint32_t regflags;
6525};
6526
6527static int addr_cmp(const void *p1_, const void *p2_)
6528{
6529 const struct savestate_block *p1 = p1_, *p2 = p2_;
6530 return p1->addr - p2->addr;
6531}
6532
6533int new_dynarec_save_blocks(void *save, int size)
6534{
6535 struct savestate_block *blocks = save;
6536 int maxcount = size / sizeof(blocks[0]);
6537 struct savestate_block tmp_blocks[1024];
6538 struct ll_entry *head;
6539 int p, s, d, o, bcnt;
6540 u_int addr;
6541
6542 o = 0;
6543 for (p = 0; p < ARRAY_SIZE(jump_in); p++) {
6544 bcnt = 0;
6545 for (head = jump_in[p]; head != NULL; head = head->next) {
6546 tmp_blocks[bcnt].addr = head->vaddr;
6547 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
6548 bcnt++;
6549 }
6550 if (bcnt < 1)
6551 continue;
6552 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6553
6554 addr = tmp_blocks[0].addr;
6555 for (s = d = 0; s < bcnt; s++) {
6556 if (tmp_blocks[s].addr < addr)
6557 continue;
6558 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6559 tmp_blocks[d++] = tmp_blocks[s];
6560 addr = scan_for_ret(tmp_blocks[s].addr);
6561 }
6562
6563 if (o + d > maxcount)
6564 d = maxcount - o;
6565 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
6566 o += d;
6567 }
6568
6569 return o * sizeof(blocks[0]);
6570}
6571
6572void new_dynarec_load_blocks(const void *save, int size)
6573{
6574 const struct savestate_block *blocks = save;
6575 int count = size / sizeof(blocks[0]);
6576 u_int regs_save[32];
6577 uint32_t f;
6578 int i, b;
6579
6580 get_addr(psxRegs.pc);
6581
6582 // change GPRs for speculation to at least partially work..
6583 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6584 for (i = 1; i < 32; i++)
6585 psxRegs.GPR.r[i] = 0x80000000;
6586
6587 for (b = 0; b < count; b++) {
6588 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6589 if (f & 1)
6590 psxRegs.GPR.r[i] = 0x1f800000;
6591 }
6592
6593 get_addr(blocks[b].addr);
6594
6595 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6596 if (f & 1)
6597 psxRegs.GPR.r[i] = 0x80000000;
6598 }
6599 }
6600
6601 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6602}
6603
6604static int apply_hacks(void)
6605{
6606 int i;
6607 if (HACK_ENABLED(NDHACK_NO_COMPAT_HACKS))
6608 return 0;
6609 /* special hack(s) */
6610 for (i = 0; i < slen - 4; i++)
6611 {
6612 // lui a4, 0xf200; jal <rcnt_read>; addu a0, 2; slti v0, 28224
6613 if (source[i] == 0x3c04f200 && dops[i+1].itype == UJUMP
6614 && source[i+2] == 0x34840002 && dops[i+3].opcode == 0x0a
6615 && imm[i+3] == 0x6e40 && dops[i+3].rs1 == 2)
6616 {
6617 SysPrintf("PE2 hack @%08x\n", start + (i+3)*4);
6618 dops[i + 3].itype = NOP;
6619 }
6620 }
6621 i = slen;
6622 if (i > 10 && source[i-1] == 0 && source[i-2] == 0x03e00008
6623 && source[i-4] == 0x8fbf0018 && source[i-6] == 0x00c0f809
6624 && dops[i-7].itype == STORE)
6625 {
6626 i = i-8;
6627 if (dops[i].itype == IMM16)
6628 i--;
6629 // swl r2, 15(r6); swr r2, 12(r6); sw r6, *; jalr r6
6630 if (dops[i].itype == STORELR && dops[i].rs1 == 6
6631 && dops[i-1].itype == STORELR && dops[i-1].rs1 == 6)
6632 {
6633 SysPrintf("F1 hack from %08x, old dst %08x\n", start, hack_addr);
6634 f1_hack = 1;
6635 return 1;
6636 }
6637 }
6638 return 0;
6639}
6640
6641static noinline void pass1_disassemble(u_int pagelimit)
6642{
6643 int i, j, done = 0, ni_count = 0;
6644 unsigned int type,op,op2;
6645
6646 for (i = 0; !done; i++)
6647 {
6648 memset(&dops[i], 0, sizeof(dops[i]));
6649 op2=0;
6650 minimum_free_regs[i]=0;
6651 dops[i].opcode=op=source[i]>>26;
6652 switch(op)
6653 {
6654 case 0x00: set_mnemonic(i, "special"); type=NI;
6655 op2=source[i]&0x3f;
6656 switch(op2)
6657 {
6658 case 0x00: set_mnemonic(i, "SLL"); type=SHIFTIMM; break;
6659 case 0x02: set_mnemonic(i, "SRL"); type=SHIFTIMM; break;
6660 case 0x03: set_mnemonic(i, "SRA"); type=SHIFTIMM; break;
6661 case 0x04: set_mnemonic(i, "SLLV"); type=SHIFT; break;
6662 case 0x06: set_mnemonic(i, "SRLV"); type=SHIFT; break;
6663 case 0x07: set_mnemonic(i, "SRAV"); type=SHIFT; break;
6664 case 0x08: set_mnemonic(i, "JR"); type=RJUMP; break;
6665 case 0x09: set_mnemonic(i, "JALR"); type=RJUMP; break;
6666 case 0x0C: set_mnemonic(i, "SYSCALL"); type=SYSCALL; break;
6667 case 0x0D: set_mnemonic(i, "BREAK"); type=SYSCALL; break;
6668 case 0x0F: set_mnemonic(i, "SYNC"); type=OTHER; break;
6669 case 0x10: set_mnemonic(i, "MFHI"); type=MOV; break;
6670 case 0x11: set_mnemonic(i, "MTHI"); type=MOV; break;
6671 case 0x12: set_mnemonic(i, "MFLO"); type=MOV; break;
6672 case 0x13: set_mnemonic(i, "MTLO"); type=MOV; break;
6673 case 0x18: set_mnemonic(i, "MULT"); type=MULTDIV; break;
6674 case 0x19: set_mnemonic(i, "MULTU"); type=MULTDIV; break;
6675 case 0x1A: set_mnemonic(i, "DIV"); type=MULTDIV; break;
6676 case 0x1B: set_mnemonic(i, "DIVU"); type=MULTDIV; break;
6677 case 0x20: set_mnemonic(i, "ADD"); type=ALU; break;
6678 case 0x21: set_mnemonic(i, "ADDU"); type=ALU; break;
6679 case 0x22: set_mnemonic(i, "SUB"); type=ALU; break;
6680 case 0x23: set_mnemonic(i, "SUBU"); type=ALU; break;
6681 case 0x24: set_mnemonic(i, "AND"); type=ALU; break;
6682 case 0x25: set_mnemonic(i, "OR"); type=ALU; break;
6683 case 0x26: set_mnemonic(i, "XOR"); type=ALU; break;
6684 case 0x27: set_mnemonic(i, "NOR"); type=ALU; break;
6685 case 0x2A: set_mnemonic(i, "SLT"); type=ALU; break;
6686 case 0x2B: set_mnemonic(i, "SLTU"); type=ALU; break;
6687 case 0x30: set_mnemonic(i, "TGE"); type=NI; break;
6688 case 0x31: set_mnemonic(i, "TGEU"); type=NI; break;
6689 case 0x32: set_mnemonic(i, "TLT"); type=NI; break;
6690 case 0x33: set_mnemonic(i, "TLTU"); type=NI; break;
6691 case 0x34: set_mnemonic(i, "TEQ"); type=NI; break;
6692 case 0x36: set_mnemonic(i, "TNE"); type=NI; break;
6693#if 0
6694 case 0x14: set_mnemonic(i, "DSLLV"); type=SHIFT; break;
6695 case 0x16: set_mnemonic(i, "DSRLV"); type=SHIFT; break;
6696 case 0x17: set_mnemonic(i, "DSRAV"); type=SHIFT; break;
6697 case 0x1C: set_mnemonic(i, "DMULT"); type=MULTDIV; break;
6698 case 0x1D: set_mnemonic(i, "DMULTU"); type=MULTDIV; break;
6699 case 0x1E: set_mnemonic(i, "DDIV"); type=MULTDIV; break;
6700 case 0x1F: set_mnemonic(i, "DDIVU"); type=MULTDIV; break;
6701 case 0x2C: set_mnemonic(i, "DADD"); type=ALU; break;
6702 case 0x2D: set_mnemonic(i, "DADDU"); type=ALU; break;
6703 case 0x2E: set_mnemonic(i, "DSUB"); type=ALU; break;
6704 case 0x2F: set_mnemonic(i, "DSUBU"); type=ALU; break;
6705 case 0x38: set_mnemonic(i, "DSLL"); type=SHIFTIMM; break;
6706 case 0x3A: set_mnemonic(i, "DSRL"); type=SHIFTIMM; break;
6707 case 0x3B: set_mnemonic(i, "DSRA"); type=SHIFTIMM; break;
6708 case 0x3C: set_mnemonic(i, "DSLL32"); type=SHIFTIMM; break;
6709 case 0x3E: set_mnemonic(i, "DSRL32"); type=SHIFTIMM; break;
6710 case 0x3F: set_mnemonic(i, "DSRA32"); type=SHIFTIMM; break;
6711#endif
6712 }
6713 break;
6714 case 0x01: set_mnemonic(i, "regimm"); type=NI;
6715 op2=(source[i]>>16)&0x1f;
6716 switch(op2)
6717 {
6718 case 0x00: set_mnemonic(i, "BLTZ"); type=SJUMP; break;
6719 case 0x01: set_mnemonic(i, "BGEZ"); type=SJUMP; break;
6720 //case 0x02: set_mnemonic(i, "BLTZL"); type=SJUMP; break;
6721 //case 0x03: set_mnemonic(i, "BGEZL"); type=SJUMP; break;
6722 //case 0x08: set_mnemonic(i, "TGEI"); type=NI; break;
6723 //case 0x09: set_mnemonic(i, "TGEIU"); type=NI; break;
6724 //case 0x0A: set_mnemonic(i, "TLTI"); type=NI; break;
6725 //case 0x0B: set_mnemonic(i, "TLTIU"); type=NI; break;
6726 //case 0x0C: set_mnemonic(i, "TEQI"); type=NI; break;
6727 //case 0x0E: set_mnemonic(i, "TNEI"); type=NI; break;
6728 case 0x10: set_mnemonic(i, "BLTZAL"); type=SJUMP; break;
6729 case 0x11: set_mnemonic(i, "BGEZAL"); type=SJUMP; break;
6730 //case 0x12: set_mnemonic(i, "BLTZALL"); type=SJUMP; break;
6731 //case 0x13: set_mnemonic(i, "BGEZALL"); type=SJUMP; break;
6732 }
6733 break;
6734 case 0x02: set_mnemonic(i, "J"); type=UJUMP; break;
6735 case 0x03: set_mnemonic(i, "JAL"); type=UJUMP; break;
6736 case 0x04: set_mnemonic(i, "BEQ"); type=CJUMP; break;
6737 case 0x05: set_mnemonic(i, "BNE"); type=CJUMP; break;
6738 case 0x06: set_mnemonic(i, "BLEZ"); type=CJUMP; break;
6739 case 0x07: set_mnemonic(i, "BGTZ"); type=CJUMP; break;
6740 case 0x08: set_mnemonic(i, "ADDI"); type=IMM16; break;
6741 case 0x09: set_mnemonic(i, "ADDIU"); type=IMM16; break;
6742 case 0x0A: set_mnemonic(i, "SLTI"); type=IMM16; break;
6743 case 0x0B: set_mnemonic(i, "SLTIU"); type=IMM16; break;
6744 case 0x0C: set_mnemonic(i, "ANDI"); type=IMM16; break;
6745 case 0x0D: set_mnemonic(i, "ORI"); type=IMM16; break;
6746 case 0x0E: set_mnemonic(i, "XORI"); type=IMM16; break;
6747 case 0x0F: set_mnemonic(i, "LUI"); type=IMM16; break;
6748 case 0x10: set_mnemonic(i, "cop0"); type=NI;
6749 op2=(source[i]>>21)&0x1f;
6750 switch(op2)
6751 {
6752 case 0x00: set_mnemonic(i, "MFC0"); type=COP0; break;
6753 case 0x02: set_mnemonic(i, "CFC0"); type=COP0; break;
6754 case 0x04: set_mnemonic(i, "MTC0"); type=COP0; break;
6755 case 0x06: set_mnemonic(i, "CTC0"); type=COP0; break;
6756 case 0x10: set_mnemonic(i, "RFE"); type=COP0; break;
6757 }
6758 break;
6759 case 0x11: set_mnemonic(i, "cop1"); type=COP1;
6760 op2=(source[i]>>21)&0x1f;
6761 break;
6762#if 0
6763 case 0x14: set_mnemonic(i, "BEQL"); type=CJUMP; break;
6764 case 0x15: set_mnemonic(i, "BNEL"); type=CJUMP; break;
6765 case 0x16: set_mnemonic(i, "BLEZL"); type=CJUMP; break;
6766 case 0x17: set_mnemonic(i, "BGTZL"); type=CJUMP; break;
6767 case 0x18: set_mnemonic(i, "DADDI"); type=IMM16; break;
6768 case 0x19: set_mnemonic(i, "DADDIU"); type=IMM16; break;
6769 case 0x1A: set_mnemonic(i, "LDL"); type=LOADLR; break;
6770 case 0x1B: set_mnemonic(i, "LDR"); type=LOADLR; break;
6771#endif
6772 case 0x20: set_mnemonic(i, "LB"); type=LOAD; break;
6773 case 0x21: set_mnemonic(i, "LH"); type=LOAD; break;
6774 case 0x22: set_mnemonic(i, "LWL"); type=LOADLR; break;
6775 case 0x23: set_mnemonic(i, "LW"); type=LOAD; break;
6776 case 0x24: set_mnemonic(i, "LBU"); type=LOAD; break;
6777 case 0x25: set_mnemonic(i, "LHU"); type=LOAD; break;
6778 case 0x26: set_mnemonic(i, "LWR"); type=LOADLR; break;
6779#if 0
6780 case 0x27: set_mnemonic(i, "LWU"); type=LOAD; break;
6781#endif
6782 case 0x28: set_mnemonic(i, "SB"); type=STORE; break;
6783 case 0x29: set_mnemonic(i, "SH"); type=STORE; break;
6784 case 0x2A: set_mnemonic(i, "SWL"); type=STORELR; break;
6785 case 0x2B: set_mnemonic(i, "SW"); type=STORE; break;
6786#if 0
6787 case 0x2C: set_mnemonic(i, "SDL"); type=STORELR; break;
6788 case 0x2D: set_mnemonic(i, "SDR"); type=STORELR; break;
6789#endif
6790 case 0x2E: set_mnemonic(i, "SWR"); type=STORELR; break;
6791 case 0x2F: set_mnemonic(i, "CACHE"); type=NOP; break;
6792 case 0x30: set_mnemonic(i, "LL"); type=NI; break;
6793 case 0x31: set_mnemonic(i, "LWC1"); type=C1LS; break;
6794#if 0
6795 case 0x34: set_mnemonic(i, "LLD"); type=NI; break;
6796 case 0x35: set_mnemonic(i, "LDC1"); type=C1LS; break;
6797 case 0x37: set_mnemonic(i, "LD"); type=LOAD; break;
6798#endif
6799 case 0x38: set_mnemonic(i, "SC"); type=NI; break;
6800 case 0x39: set_mnemonic(i, "SWC1"); type=C1LS; break;
6801#if 0
6802 case 0x3C: set_mnemonic(i, "SCD"); type=NI; break;
6803 case 0x3D: set_mnemonic(i, "SDC1"); type=C1LS; break;
6804 case 0x3F: set_mnemonic(i, "SD"); type=STORE; break;
6805#endif
6806 case 0x12: set_mnemonic(i, "COP2"); type=NI;
6807 op2=(source[i]>>21)&0x1f;
6808 //if (op2 & 0x10)
6809 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
6810 if (gte_handlers[source[i]&0x3f]!=NULL) {
6811#ifdef DISASM
6812 if (gte_regnames[source[i]&0x3f]!=NULL)
6813 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
6814 else
6815 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
6816#endif
6817 type=C2OP;
6818 }
6819 }
6820 else switch(op2)
6821 {
6822 case 0x00: set_mnemonic(i, "MFC2"); type=COP2; break;
6823 case 0x02: set_mnemonic(i, "CFC2"); type=COP2; break;
6824 case 0x04: set_mnemonic(i, "MTC2"); type=COP2; break;
6825 case 0x06: set_mnemonic(i, "CTC2"); type=COP2; break;
6826 }
6827 break;
6828 case 0x32: set_mnemonic(i, "LWC2"); type=C2LS; break;
6829 case 0x3A: set_mnemonic(i, "SWC2"); type=C2LS; break;
6830 case 0x3B: set_mnemonic(i, "HLECALL"); type=HLECALL; break;
6831 default: set_mnemonic(i, "???"); type=NI;
6832 SysPrintf("NI %08x @%08x (%08x)\n", source[i], start + i*4, start);
6833 break;
6834 }
6835 dops[i].itype=type;
6836 dops[i].opcode2=op2;
6837 /* Get registers/immediates */
6838 dops[i].use_lt1=0;
6839 gte_rs[i]=gte_rt[i]=0;
6840 switch(type) {
6841 case LOAD:
6842 dops[i].rs1=(source[i]>>21)&0x1f;
6843 dops[i].rs2=0;
6844 dops[i].rt1=(source[i]>>16)&0x1f;
6845 dops[i].rt2=0;
6846 imm[i]=(short)source[i];
6847 break;
6848 case STORE:
6849 case STORELR:
6850 dops[i].rs1=(source[i]>>21)&0x1f;
6851 dops[i].rs2=(source[i]>>16)&0x1f;
6852 dops[i].rt1=0;
6853 dops[i].rt2=0;
6854 imm[i]=(short)source[i];
6855 break;
6856 case LOADLR:
6857 // LWL/LWR only load part of the register,
6858 // therefore the target register must be treated as a source too
6859 dops[i].rs1=(source[i]>>21)&0x1f;
6860 dops[i].rs2=(source[i]>>16)&0x1f;
6861 dops[i].rt1=(source[i]>>16)&0x1f;
6862 dops[i].rt2=0;
6863 imm[i]=(short)source[i];
6864 break;
6865 case IMM16:
6866 if (op==0x0f) dops[i].rs1=0; // LUI instruction has no source register
6867 else dops[i].rs1=(source[i]>>21)&0x1f;
6868 dops[i].rs2=0;
6869 dops[i].rt1=(source[i]>>16)&0x1f;
6870 dops[i].rt2=0;
6871 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6872 imm[i]=(unsigned short)source[i];
6873 }else{
6874 imm[i]=(short)source[i];
6875 }
6876 break;
6877 case UJUMP:
6878 dops[i].rs1=0;
6879 dops[i].rs2=0;
6880 dops[i].rt1=0;
6881 dops[i].rt2=0;
6882 // The JAL instruction writes to r31.
6883 if (op&1) {
6884 dops[i].rt1=31;
6885 }
6886 dops[i].rs2=CCREG;
6887 break;
6888 case RJUMP:
6889 dops[i].rs1=(source[i]>>21)&0x1f;
6890 dops[i].rs2=0;
6891 dops[i].rt1=0;
6892 dops[i].rt2=0;
6893 // The JALR instruction writes to rd.
6894 if (op2&1) {
6895 dops[i].rt1=(source[i]>>11)&0x1f;
6896 }
6897 dops[i].rs2=CCREG;
6898 break;
6899 case CJUMP:
6900 dops[i].rs1=(source[i]>>21)&0x1f;
6901 dops[i].rs2=(source[i]>>16)&0x1f;
6902 dops[i].rt1=0;
6903 dops[i].rt2=0;
6904 if(op&2) { // BGTZ/BLEZ
6905 dops[i].rs2=0;
6906 }
6907 break;
6908 case SJUMP:
6909 dops[i].rs1=(source[i]>>21)&0x1f;
6910 dops[i].rs2=CCREG;
6911 dops[i].rt1=0;
6912 dops[i].rt2=0;
6913 if(op2&0x10) { // BxxAL
6914 dops[i].rt1=31;
6915 // NOTE: If the branch is not taken, r31 is still overwritten
6916 }
6917 break;
6918 case ALU:
6919 dops[i].rs1=(source[i]>>21)&0x1f; // source
6920 dops[i].rs2=(source[i]>>16)&0x1f; // subtract amount
6921 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6922 dops[i].rt2=0;
6923 break;
6924 case MULTDIV:
6925 dops[i].rs1=(source[i]>>21)&0x1f; // source
6926 dops[i].rs2=(source[i]>>16)&0x1f; // divisor
6927 dops[i].rt1=HIREG;
6928 dops[i].rt2=LOREG;
6929 break;
6930 case MOV:
6931 dops[i].rs1=0;
6932 dops[i].rs2=0;
6933 dops[i].rt1=0;
6934 dops[i].rt2=0;
6935 if(op2==0x10) dops[i].rs1=HIREG; // MFHI
6936 if(op2==0x11) dops[i].rt1=HIREG; // MTHI
6937 if(op2==0x12) dops[i].rs1=LOREG; // MFLO
6938 if(op2==0x13) dops[i].rt1=LOREG; // MTLO
6939 if((op2&0x1d)==0x10) dops[i].rt1=(source[i]>>11)&0x1f; // MFxx
6940 if((op2&0x1d)==0x11) dops[i].rs1=(source[i]>>21)&0x1f; // MTxx
6941 break;
6942 case SHIFT:
6943 dops[i].rs1=(source[i]>>16)&0x1f; // target of shift
6944 dops[i].rs2=(source[i]>>21)&0x1f; // shift amount
6945 dops[i].rt1=(source[i]>>11)&0x1f; // destination
6946 dops[i].rt2=0;
6947 break;
6948 case SHIFTIMM:
6949 dops[i].rs1=(source[i]>>16)&0x1f;
6950 dops[i].rs2=0;
6951 dops[i].rt1=(source[i]>>11)&0x1f;
6952 dops[i].rt2=0;
6953 imm[i]=(source[i]>>6)&0x1f;
6954 // DSxx32 instructions
6955 if(op2>=0x3c) imm[i]|=0x20;
6956 break;
6957 case COP0:
6958 dops[i].rs1=0;
6959 dops[i].rs2=0;
6960 dops[i].rt1=0;
6961 dops[i].rt2=0;
6962 if(op2==0||op2==2) dops[i].rt1=(source[i]>>16)&0x1F; // MFC0/CFC0
6963 if(op2==4||op2==6) dops[i].rs1=(source[i]>>16)&0x1F; // MTC0/CTC0
6964 if(op2==4&&((source[i]>>11)&0x1f)==12) dops[i].rt2=CSREG; // Status
6965 if(op2==16) if((source[i]&0x3f)==0x18) dops[i].rs2=CCREG; // ERET
6966 break;
6967 case COP1:
6968 dops[i].rs1=0;
6969 dops[i].rs2=0;
6970 dops[i].rt1=0;
6971 dops[i].rt2=0;
6972 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
6973 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
6974 dops[i].rs2=CSREG;
6975 break;
6976 case COP2:
6977 dops[i].rs1=0;
6978 dops[i].rs2=0;
6979 dops[i].rt1=0;
6980 dops[i].rt2=0;
6981 if(op2<3) dops[i].rt1=(source[i]>>16)&0x1F; // MFC2/CFC2
6982 if(op2>3) dops[i].rs1=(source[i]>>16)&0x1F; // MTC2/CTC2
6983 dops[i].rs2=CSREG;
6984 int gr=(source[i]>>11)&0x1F;
6985 switch(op2)
6986 {
6987 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6988 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6989 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6990 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6991 }
6992 break;
6993 case C1LS:
6994 dops[i].rs1=(source[i]>>21)&0x1F;
6995 dops[i].rs2=CSREG;
6996 dops[i].rt1=0;
6997 dops[i].rt2=0;
6998 imm[i]=(short)source[i];
6999 break;
7000 case C2LS:
7001 dops[i].rs1=(source[i]>>21)&0x1F;
7002 dops[i].rs2=0;
7003 dops[i].rt1=0;
7004 dops[i].rt2=0;
7005 imm[i]=(short)source[i];
7006 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
7007 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
7008 break;
7009 case C2OP:
7010 dops[i].rs1=0;
7011 dops[i].rs2=0;
7012 dops[i].rt1=0;
7013 dops[i].rt2=0;
7014 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
7015 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
7016 gte_rt[i]|=1ll<<63; // every op changes flags
7017 if((source[i]&0x3f)==GTE_MVMVA) {
7018 int v = (source[i] >> 15) & 3;
7019 gte_rs[i]&=~0xe3fll;
7020 if(v==3) gte_rs[i]|=0xe00ll;
7021 else gte_rs[i]|=3ll<<(v*2);
7022 }
7023 break;
7024 case SYSCALL:
7025 case HLECALL:
7026 case INTCALL:
7027 dops[i].rs1=CCREG;
7028 dops[i].rs2=0;
7029 dops[i].rt1=0;
7030 dops[i].rt2=0;
7031 break;
7032 default:
7033 dops[i].rs1=0;
7034 dops[i].rs2=0;
7035 dops[i].rt1=0;
7036 dops[i].rt2=0;
7037 }
7038 /* Calculate branch target addresses */
7039 if(type==UJUMP)
7040 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
7041 else if(type==CJUMP&&dops[i].rs1==dops[i].rs2&&(op&1))
7042 ba[i]=start+i*4+8; // Ignore never taken branch
7043 else if(type==SJUMP&&dops[i].rs1==0&&!(op2&1))
7044 ba[i]=start+i*4+8; // Ignore never taken branch
7045 else if(type==CJUMP||type==SJUMP)
7046 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
7047 else ba[i]=-1;
7048
7049 /* simplify always (not)taken branches */
7050 if (type == CJUMP && dops[i].rs1 == dops[i].rs2) {
7051 dops[i].rs1 = dops[i].rs2 = 0;
7052 if (!(op & 1)) {
7053 dops[i].itype = type = UJUMP;
7054 dops[i].rs2 = CCREG;
7055 }
7056 }
7057 else if (type == SJUMP && dops[i].rs1 == 0 && (op2 & 1))
7058 dops[i].itype = type = UJUMP;
7059
7060 dops[i].is_jump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP || dops[i].itype == CJUMP || dops[i].itype == SJUMP);
7061 dops[i].is_ujump = (dops[i].itype == RJUMP || dops[i].itype == UJUMP); // || (source[i] >> 16) == 0x1000 // beq r0,r0
7062 dops[i].is_load = (dops[i].itype == LOAD || dops[i].itype == LOADLR || op == 0x32); // LWC2
7063 dops[i].is_store = (dops[i].itype == STORE || dops[i].itype == STORELR || op == 0x3a); // SWC2
7064
7065 /* messy cases to just pass over to the interpreter */
7066 if (i > 0 && dops[i-1].is_jump) {
7067 int do_in_intrp=0;
7068 // branch in delay slot?
7069 if (dops[i].is_jump) {
7070 // don't handle first branch and call interpreter if it's hit
7071 SysPrintf("branch in delay slot @%08x (%08x)\n", start + i*4, start);
7072 do_in_intrp=1;
7073 }
7074 // basic load delay detection
7075 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&dops[i].rt1!=0) {
7076 int t=(ba[i-1]-start)/4;
7077 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) {
7078 // jump target wants DS result - potential load delay effect
7079 SysPrintf("load delay @%08x (%08x)\n", start + i*4, start);
7080 do_in_intrp=1;
7081 dops[t+1].bt=1; // expected return from interpreter
7082 }
7083 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&&
7084 !(i>=3&&dops[i-3].is_jump)) {
7085 // v0 overwrite like this is a sign of trouble, bail out
7086 SysPrintf("v0 overwrite @%08x (%08x)\n", start + i*4, start);
7087 do_in_intrp=1;
7088 }
7089 }
7090 if (do_in_intrp) {
7091 memset(&dops[i-1], 0, sizeof(dops[i-1]));
7092 dops[i-1].itype = INTCALL;
7093 dops[i-1].rs1 = CCREG;
7094 ba[i-1] = -1;
7095 done = 2;
7096 i--; // don't compile the DS
7097 }
7098 }
7099
7100 /* Is this the end of the block? */
7101 if (i > 0 && dops[i-1].is_ujump) {
7102 if(dops[i-1].rt1==0) { // Continue past subroutine call (JAL)
7103 done=2;
7104 }
7105 else {
7106 if(stop_after_jal) done=1;
7107 // Stop on BREAK
7108 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7109 }
7110 // Don't recompile stuff that's already compiled
7111 if(check_addr(start+i*4+4)) done=1;
7112 // Don't get too close to the limit
7113 if(i>MAXBLOCK/2) done=1;
7114 }
7115 if (dops[i].itype == SYSCALL || dops[i].itype == HLECALL || dops[i].itype == INTCALL)
7116 done = stop_after_jal ? 1 : 2;
7117 if (done == 2) {
7118 // Does the block continue due to a branch?
7119 for(j=i-1;j>=0;j--)
7120 {
7121 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7122 if(ba[j]==start+i*4+4) done=j=0;
7123 if(ba[j]==start+i*4+8) done=j=0;
7124 }
7125 }
7126 //assert(i<MAXBLOCK-1);
7127 if(start+i*4==pagelimit-4) done=1;
7128 assert(start+i*4<pagelimit);
7129 if (i==MAXBLOCK-1) done=1;
7130 // Stop if we're compiling junk
7131 if(dops[i].itype == NI && (++ni_count > 8 || dops[i].opcode == 0x11)) {
7132 done=stop_after_jal=1;
7133 SysPrintf("Disabled speculative precompilation\n");
7134 }
7135 }
7136 slen=i;
7137 if (dops[i-1].is_jump) {
7138 if(start+i*4==pagelimit) {
7139 dops[i-1].itype=SPAN;
7140 }
7141 }
7142 assert(slen>0);
7143}
7144
7145// Basic liveness analysis for MIPS registers
7146static noinline void pass2_unneeded_regs(int istart,int iend,int r)
7147{
7148 int i;
7149 uint64_t u,gte_u,b,gte_b;
7150 uint64_t temp_u,temp_gte_u=0;
7151 uint64_t gte_u_unknown=0;
7152 if (HACK_ENABLED(NDHACK_GTE_UNNEEDED))
7153 gte_u_unknown=~0ll;
7154 if(iend==slen-1) {
7155 u=1;
7156 gte_u=gte_u_unknown;
7157 }else{
7158 //u=unneeded_reg[iend+1];
7159 u=1;
7160 gte_u=gte_unneeded[iend+1];
7161 }
7162
7163 for (i=iend;i>=istart;i--)
7164 {
7165 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
7166 if(dops[i].is_jump)
7167 {
7168 // If subroutine call, flag return address as a possible branch target
7169 if(dops[i].rt1==31 && i<slen-2) dops[i+2].bt=1;
7170
7171 if(ba[i]<start || ba[i]>=(start+slen*4))
7172 {
7173 // Branch out of this block, flush all regs
7174 u=1;
7175 gte_u=gte_u_unknown;
7176 branch_unneeded_reg[i]=u;
7177 // Merge in delay slot
7178 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7179 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7180 u|=1;
7181 gte_u|=gte_rt[i+1];
7182 gte_u&=~gte_rs[i+1];
7183 }
7184 else
7185 {
7186 // Internal branch, flag target
7187 dops[(ba[i]-start)>>2].bt=1;
7188 if(ba[i]<=start+i*4) {
7189 // Backward branch
7190 if(dops[i].is_ujump)
7191 {
7192 // Unconditional branch
7193 temp_u=1;
7194 temp_gte_u=0;
7195 } else {
7196 // Conditional branch (not taken case)
7197 temp_u=unneeded_reg[i+2];
7198 temp_gte_u&=gte_unneeded[i+2];
7199 }
7200 // Merge in delay slot
7201 temp_u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7202 temp_u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7203 temp_u|=1;
7204 temp_gte_u|=gte_rt[i+1];
7205 temp_gte_u&=~gte_rs[i+1];
7206 temp_u|=(1LL<<dops[i].rt1)|(1LL<<dops[i].rt2);
7207 temp_u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7208 temp_u|=1;
7209 temp_gte_u|=gte_rt[i];
7210 temp_gte_u&=~gte_rs[i];
7211 unneeded_reg[i]=temp_u;
7212 gte_unneeded[i]=temp_gte_u;
7213 // Only go three levels deep. This recursion can take an
7214 // excessive amount of time if there are a lot of nested loops.
7215 if(r<2) {
7216 pass2_unneeded_regs((ba[i]-start)>>2,i-1,r+1);
7217 }else{
7218 unneeded_reg[(ba[i]-start)>>2]=1;
7219 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
7220 }
7221 } /*else*/ if(1) {
7222 if (dops[i].is_ujump)
7223 {
7224 // Unconditional branch
7225 u=unneeded_reg[(ba[i]-start)>>2];
7226 gte_u=gte_unneeded[(ba[i]-start)>>2];
7227 branch_unneeded_reg[i]=u;
7228 // Merge in delay slot
7229 u|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7230 u&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7231 u|=1;
7232 gte_u|=gte_rt[i+1];
7233 gte_u&=~gte_rs[i+1];
7234 } else {
7235 // Conditional branch
7236 b=unneeded_reg[(ba[i]-start)>>2];
7237 gte_b=gte_unneeded[(ba[i]-start)>>2];
7238 branch_unneeded_reg[i]=b;
7239 // Branch delay slot
7240 b|=(1LL<<dops[i+1].rt1)|(1LL<<dops[i+1].rt2);
7241 b&=~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7242 b|=1;
7243 gte_b|=gte_rt[i+1];
7244 gte_b&=~gte_rs[i+1];
7245 u&=b;
7246 gte_u&=gte_b;
7247 if(i<slen-1) {
7248 branch_unneeded_reg[i]&=unneeded_reg[i+2];
7249 } else {
7250 branch_unneeded_reg[i]=1;
7251 }
7252 }
7253 }
7254 }
7255 }
7256 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
7257 {
7258 // SYSCALL instruction (software interrupt)
7259 u=1;
7260 }
7261 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
7262 {
7263 // ERET instruction (return from interrupt)
7264 u=1;
7265 }
7266 //u=1; // DEBUG
7267 // Written registers are unneeded
7268 u|=1LL<<dops[i].rt1;
7269 u|=1LL<<dops[i].rt2;
7270 gte_u|=gte_rt[i];
7271 // Accessed registers are needed
7272 u&=~(1LL<<dops[i].rs1);
7273 u&=~(1LL<<dops[i].rs2);
7274 gte_u&=~gte_rs[i];
7275 if(gte_rs[i]&&dops[i].rt1&&(unneeded_reg[i+1]&(1ll<<dops[i].rt1)))
7276 gte_u|=gte_rs[i]&gte_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
7277 // Source-target dependencies
7278 // R0 is always unneeded
7279 u|=1;
7280 // Save it
7281 unneeded_reg[i]=u;
7282 gte_unneeded[i]=gte_u;
7283 /*
7284 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
7285 printf("U:");
7286 int r;
7287 for(r=1;r<=CCREG;r++) {
7288 if((unneeded_reg[i]>>r)&1) {
7289 if(r==HIREG) printf(" HI");
7290 else if(r==LOREG) printf(" LO");
7291 else printf(" r%d",r);
7292 }
7293 }
7294 printf("\n");
7295 */
7296 }
7297}
7298
7299static noinline void pass3_register_alloc(u_int addr)
7300{
7301 struct regstat current; // Current register allocations/status
7302 clear_all_regs(current.regmap_entry);
7303 clear_all_regs(current.regmap);
7304 current.wasdirty = current.dirty = 0;
7305 current.u = unneeded_reg[0];
7306 alloc_reg(&current, 0, CCREG);
7307 dirty_reg(&current, CCREG);
7308 current.wasconst = 0;
7309 current.isconst = 0;
7310 current.loadedconst = 0;
7311 current.waswritten = 0;
7312 int ds=0;
7313 int cc=0;
7314 int hr;
7315 int i, j;
7316
7317 if (addr & 1) {
7318 // First instruction is delay slot
7319 cc=-1;
7320 dops[1].bt=1;
7321 ds=1;
7322 unneeded_reg[0]=1;
7323 current.regmap[HOST_BTREG]=BTREG;
7324 }
7325
7326 for(i=0;i<slen;i++)
7327 {
7328 if(dops[i].bt)
7329 {
7330 for(hr=0;hr<HOST_REGS;hr++)
7331 {
7332 // Is this really necessary?
7333 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7334 }
7335 current.isconst=0;
7336 current.waswritten=0;
7337 }
7338
7339 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7340 regs[i].wasconst=current.isconst;
7341 regs[i].wasdirty=current.dirty;
7342 regs[i].dirty=0;
7343 regs[i].u=0;
7344 regs[i].isconst=0;
7345 regs[i].loadedconst=0;
7346 if (!dops[i].is_jump) {
7347 if(i+1<slen) {
7348 current.u=unneeded_reg[i+1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7349 current.u|=1;
7350 } else {
7351 current.u=1;
7352 }
7353 } else {
7354 if(i+1<slen) {
7355 current.u=branch_unneeded_reg[i]&~((1LL<<dops[i+1].rs1)|(1LL<<dops[i+1].rs2));
7356 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7357 current.u|=1;
7358 } else {
7359 SysPrintf("oops, branch at end of block with no delay slot @%08x\n", start + i*4);
7360 abort();
7361 }
7362 }
7363 dops[i].is_ds=ds;
7364 if(ds) {
7365 ds=0; // Skip delay slot, already allocated as part of branch
7366 // ...but we need to alloc it in case something jumps here
7367 if(i+1<slen) {
7368 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7369 }else{
7370 current.u=branch_unneeded_reg[i-1];
7371 }
7372 current.u&=~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7373 current.u|=1;
7374 struct regstat temp;
7375 memcpy(&temp,&current,sizeof(current));
7376 temp.wasdirty=temp.dirty;
7377 // TODO: Take into account unconditional branches, as below
7378 delayslot_alloc(&temp,i);
7379 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7380 regs[i].wasdirty=temp.wasdirty;
7381 regs[i].dirty=temp.dirty;
7382 regs[i].isconst=0;
7383 regs[i].wasconst=0;
7384 current.isconst=0;
7385 // Create entry (branch target) regmap
7386 for(hr=0;hr<HOST_REGS;hr++)
7387 {
7388 int r=temp.regmap[hr];
7389 if(r>=0) {
7390 if(r!=regmap_pre[i][hr]) {
7391 regs[i].regmap_entry[hr]=-1;
7392 }
7393 else
7394 {
7395 assert(r < 64);
7396 if((current.u>>r)&1) {
7397 regs[i].regmap_entry[hr]=-1;
7398 regs[i].regmap[hr]=-1;
7399 //Don't clear regs in the delay slot as the branch might need them
7400 //current.regmap[hr]=-1;
7401 }else
7402 regs[i].regmap_entry[hr]=r;
7403 }
7404 } else {
7405 // First instruction expects CCREG to be allocated
7406 if(i==0&&hr==HOST_CCREG)
7407 regs[i].regmap_entry[hr]=CCREG;
7408 else
7409 regs[i].regmap_entry[hr]=-1;
7410 }
7411 }
7412 }
7413 else { // Not delay slot
7414 switch(dops[i].itype) {
7415 case UJUMP:
7416 //current.isconst=0; // DEBUG
7417 //current.wasconst=0; // DEBUG
7418 //regs[i].wasconst=0; // DEBUG
7419 clear_const(&current,dops[i].rt1);
7420 alloc_cc(&current,i);
7421 dirty_reg(&current,CCREG);
7422 if (dops[i].rt1==31) {
7423 alloc_reg(&current,i,31);
7424 dirty_reg(&current,31);
7425 //assert(dops[i+1].rs1!=31&&dops[i+1].rs2!=31);
7426 //assert(dops[i+1].rt1!=dops[i].rt1);
7427 #ifdef REG_PREFETCH
7428 alloc_reg(&current,i,PTEMP);
7429 #endif
7430 }
7431 dops[i].ooo=1;
7432 delayslot_alloc(&current,i+1);
7433 //current.isconst=0; // DEBUG
7434 ds=1;
7435 //printf("i=%d, isconst=%x\n",i,current.isconst);
7436 break;
7437 case RJUMP:
7438 //current.isconst=0;
7439 //current.wasconst=0;
7440 //regs[i].wasconst=0;
7441 clear_const(&current,dops[i].rs1);
7442 clear_const(&current,dops[i].rt1);
7443 alloc_cc(&current,i);
7444 dirty_reg(&current,CCREG);
7445 if (!ds_writes_rjump_rs(i)) {
7446 alloc_reg(&current,i,dops[i].rs1);
7447 if (dops[i].rt1!=0) {
7448 alloc_reg(&current,i,dops[i].rt1);
7449 dirty_reg(&current,dops[i].rt1);
7450 assert(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt1);
7451 assert(dops[i+1].rt1!=dops[i].rt1);
7452 #ifdef REG_PREFETCH
7453 alloc_reg(&current,i,PTEMP);
7454 #endif
7455 }
7456 #ifdef USE_MINI_HT
7457 if(dops[i].rs1==31) { // JALR
7458 alloc_reg(&current,i,RHASH);
7459 alloc_reg(&current,i,RHTBL);
7460 }
7461 #endif
7462 delayslot_alloc(&current,i+1);
7463 } else {
7464 // The delay slot overwrites our source register,
7465 // allocate a temporary register to hold the old value.
7466 current.isconst=0;
7467 current.wasconst=0;
7468 regs[i].wasconst=0;
7469 delayslot_alloc(&current,i+1);
7470 current.isconst=0;
7471 alloc_reg(&current,i,RTEMP);
7472 }
7473 //current.isconst=0; // DEBUG
7474 dops[i].ooo=1;
7475 ds=1;
7476 break;
7477 case CJUMP:
7478 //current.isconst=0;
7479 //current.wasconst=0;
7480 //regs[i].wasconst=0;
7481 clear_const(&current,dops[i].rs1);
7482 clear_const(&current,dops[i].rs2);
7483 if((dops[i].opcode&0x3E)==4) // BEQ/BNE
7484 {
7485 alloc_cc(&current,i);
7486 dirty_reg(&current,CCREG);
7487 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7488 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7489 if((dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2))||
7490 (dops[i].rs2&&(dops[i].rs2==dops[i+1].rt1||dops[i].rs2==dops[i+1].rt2))) {
7491 // The delay slot overwrites one of our conditions.
7492 // Allocate the branch condition registers instead.
7493 current.isconst=0;
7494 current.wasconst=0;
7495 regs[i].wasconst=0;
7496 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7497 if(dops[i].rs2) alloc_reg(&current,i,dops[i].rs2);
7498 }
7499 else
7500 {
7501 dops[i].ooo=1;
7502 delayslot_alloc(&current,i+1);
7503 }
7504 }
7505 else
7506 if((dops[i].opcode&0x3E)==6) // BLEZ/BGTZ
7507 {
7508 alloc_cc(&current,i);
7509 dirty_reg(&current,CCREG);
7510 alloc_reg(&current,i,dops[i].rs1);
7511 if(dops[i].rs1&&(dops[i].rs1==dops[i+1].rt1||dops[i].rs1==dops[i+1].rt2)) {
7512 // The delay slot overwrites one of our conditions.
7513 // Allocate the branch condition registers instead.
7514 current.isconst=0;
7515 current.wasconst=0;
7516 regs[i].wasconst=0;
7517 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7518 }
7519 else
7520 {
7521 dops[i].ooo=1;
7522 delayslot_alloc(&current,i+1);
7523 }
7524 }
7525 else
7526 // Don't alloc the delay slot yet because we might not execute it
7527 if((dops[i].opcode&0x3E)==0x14) // BEQL/BNEL
7528 {
7529 current.isconst=0;
7530 current.wasconst=0;
7531 regs[i].wasconst=0;
7532 alloc_cc(&current,i);
7533 dirty_reg(&current,CCREG);
7534 alloc_reg(&current,i,dops[i].rs1);
7535 alloc_reg(&current,i,dops[i].rs2);
7536 }
7537 else
7538 if((dops[i].opcode&0x3E)==0x16) // BLEZL/BGTZL
7539 {
7540 current.isconst=0;
7541 current.wasconst=0;
7542 regs[i].wasconst=0;
7543 alloc_cc(&current,i);
7544 dirty_reg(&current,CCREG);
7545 alloc_reg(&current,i,dops[i].rs1);
7546 }
7547 ds=1;
7548 //current.isconst=0;
7549 break;
7550 case SJUMP:
7551 //current.isconst=0;
7552 //current.wasconst=0;
7553 //regs[i].wasconst=0;
7554 clear_const(&current,dops[i].rs1);
7555 clear_const(&current,dops[i].rt1);
7556 //if((dops[i].opcode2&0x1E)==0x0) // BLTZ/BGEZ
7557 if((dops[i].opcode2&0x0E)==0x0) // BLTZ/BGEZ
7558 {
7559 alloc_cc(&current,i);
7560 dirty_reg(&current,CCREG);
7561 alloc_reg(&current,i,dops[i].rs1);
7562 if (dops[i].rt1==31) { // BLTZAL/BGEZAL
7563 alloc_reg(&current,i,31);
7564 dirty_reg(&current,31);
7565 //#ifdef REG_PREFETCH
7566 //alloc_reg(&current,i,PTEMP);
7567 //#endif
7568 }
7569 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.
7570 ||(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
7571 // Allocate the branch condition registers instead.
7572 current.isconst=0;
7573 current.wasconst=0;
7574 regs[i].wasconst=0;
7575 if(dops[i].rs1) alloc_reg(&current,i,dops[i].rs1);
7576 }
7577 else
7578 {
7579 dops[i].ooo=1;
7580 delayslot_alloc(&current,i+1);
7581 }
7582 }
7583 else
7584 // Don't alloc the delay slot yet because we might not execute it
7585 if((dops[i].opcode2&0x1E)==0x2) // BLTZL/BGEZL
7586 {
7587 current.isconst=0;
7588 current.wasconst=0;
7589 regs[i].wasconst=0;
7590 alloc_cc(&current,i);
7591 dirty_reg(&current,CCREG);
7592 alloc_reg(&current,i,dops[i].rs1);
7593 }
7594 ds=1;
7595 //current.isconst=0;
7596 break;
7597 case IMM16:
7598 imm16_alloc(&current,i);
7599 break;
7600 case LOAD:
7601 case LOADLR:
7602 load_alloc(&current,i);
7603 break;
7604 case STORE:
7605 case STORELR:
7606 store_alloc(&current,i);
7607 break;
7608 case ALU:
7609 alu_alloc(&current,i);
7610 break;
7611 case SHIFT:
7612 shift_alloc(&current,i);
7613 break;
7614 case MULTDIV:
7615 multdiv_alloc(&current,i);
7616 break;
7617 case SHIFTIMM:
7618 shiftimm_alloc(&current,i);
7619 break;
7620 case MOV:
7621 mov_alloc(&current,i);
7622 break;
7623 case COP0:
7624 cop0_alloc(&current,i);
7625 break;
7626 case COP1:
7627 break;
7628 case COP2:
7629 cop2_alloc(&current,i);
7630 break;
7631 case C1LS:
7632 c1ls_alloc(&current,i);
7633 break;
7634 case C2LS:
7635 c2ls_alloc(&current,i);
7636 break;
7637 case C2OP:
7638 c2op_alloc(&current,i);
7639 break;
7640 case SYSCALL:
7641 case HLECALL:
7642 case INTCALL:
7643 syscall_alloc(&current,i);
7644 break;
7645 case SPAN:
7646 pagespan_alloc(&current,i);
7647 break;
7648 }
7649
7650 // Create entry (branch target) regmap
7651 for(hr=0;hr<HOST_REGS;hr++)
7652 {
7653 int r,or;
7654 r=current.regmap[hr];
7655 if(r>=0) {
7656 if(r!=regmap_pre[i][hr]) {
7657 // TODO: delay slot (?)
7658 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7659 if(or<0||r>=TEMPREG){
7660 regs[i].regmap_entry[hr]=-1;
7661 }
7662 else
7663 {
7664 // Just move it to a different register
7665 regs[i].regmap_entry[hr]=r;
7666 // If it was dirty before, it's still dirty
7667 if((regs[i].wasdirty>>or)&1) dirty_reg(&current,r);
7668 }
7669 }
7670 else
7671 {
7672 // Unneeded
7673 if(r==0){
7674 regs[i].regmap_entry[hr]=0;
7675 }
7676 else
7677 {
7678 assert(r<64);
7679 if((current.u>>r)&1) {
7680 regs[i].regmap_entry[hr]=-1;
7681 //regs[i].regmap[hr]=-1;
7682 current.regmap[hr]=-1;
7683 }else
7684 regs[i].regmap_entry[hr]=r;
7685 }
7686 }
7687 } else {
7688 // Branches expect CCREG to be allocated at the target
7689 if(regmap_pre[i][hr]==CCREG)
7690 regs[i].regmap_entry[hr]=CCREG;
7691 else
7692 regs[i].regmap_entry[hr]=-1;
7693 }
7694 }
7695 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7696 }
7697
7698 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)
7699 current.waswritten|=1<<dops[i-1].rs1;
7700 current.waswritten&=~(1<<dops[i].rt1);
7701 current.waswritten&=~(1<<dops[i].rt2);
7702 if((dops[i].itype==STORE||dops[i].itype==STORELR||(dops[i].itype==C2LS&&dops[i].opcode==0x3a))&&(u_int)imm[i]>=0x800)
7703 current.waswritten&=~(1<<dops[i].rs1);
7704
7705 /* Branch post-alloc */
7706 if(i>0)
7707 {
7708 current.wasdirty=current.dirty;
7709 switch(dops[i-1].itype) {
7710 case UJUMP:
7711 memcpy(&branch_regs[i-1],&current,sizeof(current));
7712 branch_regs[i-1].isconst=0;
7713 branch_regs[i-1].wasconst=0;
7714 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7715 alloc_cc(&branch_regs[i-1],i-1);
7716 dirty_reg(&branch_regs[i-1],CCREG);
7717 if(dops[i-1].rt1==31) { // JAL
7718 alloc_reg(&branch_regs[i-1],i-1,31);
7719 dirty_reg(&branch_regs[i-1],31);
7720 }
7721 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7722 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7723 break;
7724 case RJUMP:
7725 memcpy(&branch_regs[i-1],&current,sizeof(current));
7726 branch_regs[i-1].isconst=0;
7727 branch_regs[i-1].wasconst=0;
7728 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7729 alloc_cc(&branch_regs[i-1],i-1);
7730 dirty_reg(&branch_regs[i-1],CCREG);
7731 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rs1);
7732 if(dops[i-1].rt1!=0) { // JALR
7733 alloc_reg(&branch_regs[i-1],i-1,dops[i-1].rt1);
7734 dirty_reg(&branch_regs[i-1],dops[i-1].rt1);
7735 }
7736 #ifdef USE_MINI_HT
7737 if(dops[i-1].rs1==31) { // JALR
7738 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7739 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7740 }
7741 #endif
7742 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7743 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7744 break;
7745 case CJUMP:
7746 if((dops[i-1].opcode&0x3E)==4) // BEQ/BNE
7747 {
7748 alloc_cc(&current,i-1);
7749 dirty_reg(&current,CCREG);
7750 if((dops[i-1].rs1&&(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2))||
7751 (dops[i-1].rs2&&(dops[i-1].rs2==dops[i].rt1||dops[i-1].rs2==dops[i].rt2))) {
7752 // The delay slot overwrote one of our conditions
7753 // Delay slot goes after the test (in order)
7754 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7755 current.u|=1;
7756 delayslot_alloc(&current,i);
7757 current.isconst=0;
7758 }
7759 else
7760 {
7761 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i-1].rs1)|(1LL<<dops[i-1].rs2));
7762 // Alloc the branch condition registers
7763 if(dops[i-1].rs1) alloc_reg(&current,i-1,dops[i-1].rs1);
7764 if(dops[i-1].rs2) alloc_reg(&current,i-1,dops[i-1].rs2);
7765 }
7766 memcpy(&branch_regs[i-1],&current,sizeof(current));
7767 branch_regs[i-1].isconst=0;
7768 branch_regs[i-1].wasconst=0;
7769 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7770 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7771 }
7772 else
7773 if((dops[i-1].opcode&0x3E)==6) // BLEZ/BGTZ
7774 {
7775 alloc_cc(&current,i-1);
7776 dirty_reg(&current,CCREG);
7777 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7778 // The delay slot overwrote the branch condition
7779 // Delay slot goes after the test (in order)
7780 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7781 current.u|=1;
7782 delayslot_alloc(&current,i);
7783 current.isconst=0;
7784 }
7785 else
7786 {
7787 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7788 // Alloc the branch condition register
7789 alloc_reg(&current,i-1,dops[i-1].rs1);
7790 }
7791 memcpy(&branch_regs[i-1],&current,sizeof(current));
7792 branch_regs[i-1].isconst=0;
7793 branch_regs[i-1].wasconst=0;
7794 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7795 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7796 }
7797 else
7798 // Alloc the delay slot in case the branch is taken
7799 if((dops[i-1].opcode&0x3E)==0x14) // BEQL/BNEL
7800 {
7801 memcpy(&branch_regs[i-1],&current,sizeof(current));
7802 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;
7803 alloc_cc(&branch_regs[i-1],i);
7804 dirty_reg(&branch_regs[i-1],CCREG);
7805 delayslot_alloc(&branch_regs[i-1],i);
7806 branch_regs[i-1].isconst=0;
7807 alloc_reg(&current,i,CCREG); // Not taken path
7808 dirty_reg(&current,CCREG);
7809 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7810 }
7811 else
7812 if((dops[i-1].opcode&0x3E)==0x16) // BLEZL/BGTZL
7813 {
7814 memcpy(&branch_regs[i-1],&current,sizeof(current));
7815 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;
7816 alloc_cc(&branch_regs[i-1],i);
7817 dirty_reg(&branch_regs[i-1],CCREG);
7818 delayslot_alloc(&branch_regs[i-1],i);
7819 branch_regs[i-1].isconst=0;
7820 alloc_reg(&current,i,CCREG); // Not taken path
7821 dirty_reg(&current,CCREG);
7822 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7823 }
7824 break;
7825 case SJUMP:
7826 //if((dops[i-1].opcode2&0x1E)==0) // BLTZ/BGEZ
7827 if((dops[i-1].opcode2&0x0E)==0) // BLTZ/BGEZ
7828 {
7829 alloc_cc(&current,i-1);
7830 dirty_reg(&current,CCREG);
7831 if(dops[i-1].rs1==dops[i].rt1||dops[i-1].rs1==dops[i].rt2) {
7832 // The delay slot overwrote the branch condition
7833 // Delay slot goes after the test (in order)
7834 current.u=branch_unneeded_reg[i-1]&~((1LL<<dops[i].rs1)|(1LL<<dops[i].rs2));
7835 current.u|=1;
7836 delayslot_alloc(&current,i);
7837 current.isconst=0;
7838 }
7839 else
7840 {
7841 current.u=branch_unneeded_reg[i-1]&~(1LL<<dops[i-1].rs1);
7842 // Alloc the branch condition register
7843 alloc_reg(&current,i-1,dops[i-1].rs1);
7844 }
7845 memcpy(&branch_regs[i-1],&current,sizeof(current));
7846 branch_regs[i-1].isconst=0;
7847 branch_regs[i-1].wasconst=0;
7848 memcpy(&branch_regs[i-1].regmap_entry,&current.regmap,sizeof(current.regmap));
7849 memcpy(constmap[i],constmap[i-1],sizeof(constmap[i]));
7850 }
7851 else
7852 // Alloc the delay slot in case the branch is taken
7853 if((dops[i-1].opcode2&0x1E)==2) // BLTZL/BGEZL
7854 {
7855 memcpy(&branch_regs[i-1],&current,sizeof(current));
7856 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;
7857 alloc_cc(&branch_regs[i-1],i);
7858 dirty_reg(&branch_regs[i-1],CCREG);
7859 delayslot_alloc(&branch_regs[i-1],i);
7860 branch_regs[i-1].isconst=0;
7861 alloc_reg(&current,i,CCREG); // Not taken path
7862 dirty_reg(&current,CCREG);
7863 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7864 }
7865 // FIXME: BLTZAL/BGEZAL
7866 if(dops[i-1].opcode2&0x10) { // BxxZAL
7867 alloc_reg(&branch_regs[i-1],i-1,31);
7868 dirty_reg(&branch_regs[i-1],31);
7869 }
7870 break;
7871 }
7872
7873 if (dops[i-1].is_ujump)
7874 {
7875 if(dops[i-1].rt1==31) // JAL/JALR
7876 {
7877 // Subroutine call will return here, don't alloc any registers
7878 current.dirty=0;
7879 clear_all_regs(current.regmap);
7880 alloc_reg(&current,i,CCREG);
7881 dirty_reg(&current,CCREG);
7882 }
7883 else if(i+1<slen)
7884 {
7885 // Internal branch will jump here, match registers to caller
7886 current.dirty=0;
7887 clear_all_regs(current.regmap);
7888 alloc_reg(&current,i,CCREG);
7889 dirty_reg(&current,CCREG);
7890 for(j=i-1;j>=0;j--)
7891 {
7892 if(ba[j]==start+i*4+4) {
7893 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7894 current.dirty=branch_regs[j].dirty;
7895 break;
7896 }
7897 }
7898 while(j>=0) {
7899 if(ba[j]==start+i*4+4) {
7900 for(hr=0;hr<HOST_REGS;hr++) {
7901 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7902 current.regmap[hr]=-1;
7903 }
7904 current.dirty&=branch_regs[j].dirty;
7905 }
7906 }
7907 j--;
7908 }
7909 }
7910 }
7911 }
7912
7913 // Count cycles in between branches
7914 ccadj[i] = CLOCK_ADJUST(cc);
7915 if (i > 0 && (dops[i-1].is_jump || dops[i].itype == SYSCALL || dops[i].itype == HLECALL))
7916 {
7917 cc=0;
7918 }
7919#if !defined(DRC_DBG)
7920 else if(dops[i].itype==C2OP&&gte_cycletab[source[i]&0x3f]>2)
7921 {
7922 // this should really be removed since the real stalls have been implemented,
7923 // but doing so causes sizeable perf regression against the older version
7924 u_int gtec = gte_cycletab[source[i] & 0x3f];
7925 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? gtec/2 : 2;
7926 }
7927 else if(i>1&&dops[i].itype==STORE&&dops[i-1].itype==STORE&&dops[i-2].itype==STORE&&!dops[i].bt)
7928 {
7929 cc+=4;
7930 }
7931 else if(dops[i].itype==C2LS)
7932 {
7933 // same as with C2OP
7934 cc += HACK_ENABLED(NDHACK_NO_STALLS) ? 4 : 2;
7935 }
7936#endif
7937 else
7938 {
7939 cc++;
7940 }
7941
7942 if(!dops[i].is_ds) {
7943 regs[i].dirty=current.dirty;
7944 regs[i].isconst=current.isconst;
7945 memcpy(constmap[i],current_constmap,sizeof(constmap[i]));
7946 }
7947 for(hr=0;hr<HOST_REGS;hr++) {
7948 if(hr!=EXCLUDE_REG&&regs[i].regmap[hr]>=0) {
7949 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7950 regs[i].wasconst&=~(1<<hr);
7951 }
7952 }
7953 }
7954 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7955 regs[i].waswritten=current.waswritten;
7956 }
7957}
7958
7959static noinline void pass4_cull_unused_regs(void)
7960{
7961 u_int nr=0;
7962 int i;
7963
7964 for (i=slen-1;i>=0;i--)
7965 {
7966 int hr;
7967 if(dops[i].is_jump)
7968 {
7969 if(ba[i]<start || ba[i]>=(start+slen*4))
7970 {
7971 // Branch out of this block, don't need anything
7972 nr=0;
7973 }
7974 else
7975 {
7976 // Internal branch
7977 // Need whatever matches the target
7978 nr=0;
7979 int t=(ba[i]-start)>>2;
7980 for(hr=0;hr<HOST_REGS;hr++)
7981 {
7982 if(regs[i].regmap_entry[hr]>=0) {
7983 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7984 }
7985 }
7986 }
7987 // Conditional branch may need registers for following instructions
7988 if (!dops[i].is_ujump)
7989 {
7990 if(i<slen-2) {
7991 nr|=needed_reg[i+2];
7992 for(hr=0;hr<HOST_REGS;hr++)
7993 {
7994 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7995 //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]);
7996 }
7997 }
7998 }
7999 // Don't need stuff which is overwritten
8000 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8001 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8002 // Merge in delay slot
8003 for(hr=0;hr<HOST_REGS;hr++)
8004 {
8005 if(dops[i+1].rt1&&dops[i+1].rt1==regs[i].regmap[hr]) nr&=~(1<<hr);
8006 if(dops[i+1].rt2&&dops[i+1].rt2==regs[i].regmap[hr]) nr&=~(1<<hr);
8007 if(dops[i+1].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8008 if(dops[i+1].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8009 if(dops[i+1].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8010 if(dops[i+1].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8011 if(ram_offset && (dops[i+1].is_load || dops[i+1].is_store)) {
8012 if(regmap_pre[i][hr]==ROREG) nr|=1<<hr;
8013 if(regs[i].regmap_entry[hr]==ROREG) nr|=1<<hr;
8014 }
8015 if(dops[i+1].is_store) {
8016 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8017 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8018 }
8019 }
8020 }
8021 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8022 {
8023 // SYSCALL instruction (software interrupt)
8024 nr=0;
8025 }
8026 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8027 {
8028 // ERET instruction (return from interrupt)
8029 nr=0;
8030 }
8031 else // Non-branch
8032 {
8033 if(i<slen-1) {
8034 for(hr=0;hr<HOST_REGS;hr++) {
8035 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
8036 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
8037 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
8038 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
8039 }
8040 }
8041 }
8042 for(hr=0;hr<HOST_REGS;hr++)
8043 {
8044 // Overwritten registers are not needed
8045 if(dops[i].rt1&&dops[i].rt1==regs[i].regmap[hr]) nr&=~(1<<hr);
8046 if(dops[i].rt2&&dops[i].rt2==regs[i].regmap[hr]) nr&=~(1<<hr);
8047 if(FTEMP==regs[i].regmap[hr]) nr&=~(1<<hr);
8048 // Source registers are needed
8049 if(dops[i].rs1==regmap_pre[i][hr]) nr|=1<<hr;
8050 if(dops[i].rs2==regmap_pre[i][hr]) nr|=1<<hr;
8051 if(dops[i].rs1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8052 if(dops[i].rs2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8053 if(ram_offset && (dops[i].is_load || dops[i].is_store)) {
8054 if(regmap_pre[i][hr]==ROREG) nr|=1<<hr;
8055 if(regs[i].regmap_entry[hr]==ROREG) nr|=1<<hr;
8056 }
8057 if(dops[i].is_store) {
8058 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
8059 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
8060 }
8061 // Don't store a register immediately after writing it,
8062 // may prevent dual-issue.
8063 // But do so if this is a branch target, otherwise we
8064 // might have to load the register before the branch.
8065 if(i>0&&!dops[i].bt&&((regs[i].wasdirty>>hr)&1)) {
8066 if((regmap_pre[i][hr]>0&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
8067 if(dops[i-1].rt1==regmap_pre[i][hr]) nr|=1<<hr;
8068 if(dops[i-1].rt2==regmap_pre[i][hr]) nr|=1<<hr;
8069 }
8070 if((regs[i].regmap_entry[hr]>0&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
8071 if(dops[i-1].rt1==regs[i].regmap_entry[hr]) nr|=1<<hr;
8072 if(dops[i-1].rt2==regs[i].regmap_entry[hr]) nr|=1<<hr;
8073 }
8074 }
8075 }
8076 // Cycle count is needed at branches. Assume it is needed at the target too.
8077 if(i==0||dops[i].bt||dops[i].itype==CJUMP||dops[i].itype==SPAN) {
8078 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8079 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
8080 }
8081 // Save it
8082 needed_reg[i]=nr;
8083
8084 // Deallocate unneeded registers
8085 for(hr=0;hr<HOST_REGS;hr++)
8086 {
8087 if(!((nr>>hr)&1)) {
8088 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
8089 if(dops[i].is_jump)
8090 {
8091 int map1 = 0, map2 = 0, temp = 0; // or -1 ??
8092 if (dops[i+1].is_load || dops[i+1].is_store)
8093 map1 = ROREG;
8094 if (dops[i+1].is_store)
8095 map2 = INVCP;
8096 if(dops[i+1].itype==LOADLR || dops[i+1].itype==STORELR || dops[i+1].itype==C2LS)
8097 temp = FTEMP;
8098 if(regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8099 regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8100 regs[i].regmap[hr]!=dops[i+1].rt1 && regs[i].regmap[hr]!=dops[i+1].rt2 &&
8101 regs[i].regmap[hr]!=dops[i+1].rs1 && regs[i].regmap[hr]!=dops[i+1].rs2 &&
8102 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=PTEMP &&
8103 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
8104 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
8105 regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2)
8106 {
8107 regs[i].regmap[hr]=-1;
8108 regs[i].isconst&=~(1<<hr);
8109 regs[i].dirty&=~(1<<hr);
8110 regs[i+1].wasdirty&=~(1<<hr);
8111 if(branch_regs[i].regmap[hr]!=dops[i].rs1 && branch_regs[i].regmap[hr]!=dops[i].rs2 &&
8112 branch_regs[i].regmap[hr]!=dops[i].rt1 && branch_regs[i].regmap[hr]!=dops[i].rt2 &&
8113 branch_regs[i].regmap[hr]!=dops[i+1].rt1 && branch_regs[i].regmap[hr]!=dops[i+1].rt2 &&
8114 branch_regs[i].regmap[hr]!=dops[i+1].rs1 && branch_regs[i].regmap[hr]!=dops[i+1].rs2 &&
8115 branch_regs[i].regmap[hr]!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
8116 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
8117 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
8118 branch_regs[i].regmap[hr]!=map1 && branch_regs[i].regmap[hr]!=map2)
8119 {
8120 branch_regs[i].regmap[hr]=-1;
8121 branch_regs[i].regmap_entry[hr]=-1;
8122 if (!dops[i].is_ujump)
8123 {
8124 if (i < slen-2) {
8125 regmap_pre[i+2][hr]=-1;
8126 regs[i+2].wasconst&=~(1<<hr);
8127 }
8128 }
8129 }
8130 }
8131 }
8132 else
8133 {
8134 // Non-branch
8135 if(i>0)
8136 {
8137 int map1 = -1, map2 = -1, temp=-1;
8138 if (dops[i].is_load || dops[i].is_store)
8139 map1 = ROREG;
8140 if (dops[i].is_store)
8141 map2 = INVCP;
8142 if (dops[i].itype==LOADLR || dops[i].itype==STORELR || dops[i].itype==C2LS)
8143 temp = FTEMP;
8144 if(regs[i].regmap[hr]!=dops[i].rt1 && regs[i].regmap[hr]!=dops[i].rt2 &&
8145 regs[i].regmap[hr]!=dops[i].rs1 && regs[i].regmap[hr]!=dops[i].rs2 &&
8146 regs[i].regmap[hr]!=temp && regs[i].regmap[hr]!=map1 && regs[i].regmap[hr]!=map2 &&
8147 //(dops[i].itype!=SPAN||regs[i].regmap[hr]!=CCREG)
8148 regs[i].regmap[hr] != CCREG)
8149 {
8150 if(i<slen-1&&!dops[i].is_ds) {
8151 assert(regs[i].regmap[hr]<64);
8152 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]>0)
8153 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
8154 {
8155 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
8156 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
8157 }
8158 regmap_pre[i+1][hr]=-1;
8159 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
8160 regs[i+1].wasconst&=~(1<<hr);
8161 }
8162 regs[i].regmap[hr]=-1;
8163 regs[i].isconst&=~(1<<hr);
8164 regs[i].dirty&=~(1<<hr);
8165 regs[i+1].wasdirty&=~(1<<hr);
8166 }
8167 }
8168 }
8169 } // if needed
8170 } // for hr
8171 }
8172}
8173
8174// If a register is allocated during a loop, try to allocate it for the
8175// entire loop, if possible. This avoids loading/storing registers
8176// inside of the loop.
8177static noinline void pass5a_preallocate1(void)
8178{
8179 int i, j, hr;
8180 signed char f_regmap[HOST_REGS];
8181 clear_all_regs(f_regmap);
8182 for(i=0;i<slen-1;i++)
8183 {
8184 if(dops[i].itype==UJUMP||dops[i].itype==CJUMP||dops[i].itype==SJUMP)
8185 {
8186 if(ba[i]>=start && ba[i]<(start+i*4))
8187 if(dops[i+1].itype==NOP||dops[i+1].itype==MOV||dops[i+1].itype==ALU
8188 ||dops[i+1].itype==SHIFTIMM||dops[i+1].itype==IMM16||dops[i+1].itype==LOAD
8189 ||dops[i+1].itype==STORE||dops[i+1].itype==STORELR||dops[i+1].itype==C1LS
8190 ||dops[i+1].itype==SHIFT||dops[i+1].itype==COP1
8191 ||dops[i+1].itype==COP2||dops[i+1].itype==C2LS||dops[i+1].itype==C2OP)
8192 {
8193 int t=(ba[i]-start)>>2;
8194 if(t > 0 && !dops[t-1].is_jump) // loop_preload can't handle jumps into delay slots
8195 if(t<2||(dops[t-2].itype!=UJUMP&&dops[t-2].itype!=RJUMP)||dops[t-2].rt1!=31) // call/ret assumes no registers allocated
8196 for(hr=0;hr<HOST_REGS;hr++)
8197 {
8198 if(regs[i].regmap[hr]>=0) {
8199 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8200 // dealloc old register
8201 int n;
8202 for(n=0;n<HOST_REGS;n++)
8203 {
8204 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8205 }
8206 // and alloc new one
8207 f_regmap[hr]=regs[i].regmap[hr];
8208 }
8209 }
8210 if(branch_regs[i].regmap[hr]>=0) {
8211 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8212 // dealloc old register
8213 int n;
8214 for(n=0;n<HOST_REGS;n++)
8215 {
8216 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8217 }
8218 // and alloc new one
8219 f_regmap[hr]=branch_regs[i].regmap[hr];
8220 }
8221 }
8222 if(dops[i].ooo) {
8223 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8224 f_regmap[hr]=branch_regs[i].regmap[hr];
8225 }else{
8226 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8227 f_regmap[hr]=branch_regs[i].regmap[hr];
8228 }
8229 // Avoid dirty->clean transition
8230 #ifdef DESTRUCTIVE_WRITEBACK
8231 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;
8232 #endif
8233 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8234 // case above, however it's always a good idea. We can't hoist the
8235 // load if the register was already allocated, so there's no point
8236 // wasting time analyzing most of these cases. It only "succeeds"
8237 // when the mapping was different and the load can be replaced with
8238 // a mov, which is of negligible benefit. So such cases are
8239 // skipped below.
8240 if(f_regmap[hr]>0) {
8241 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8242 int r=f_regmap[hr];
8243 for(j=t;j<=i;j++)
8244 {
8245 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8246 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8247 assert(r < 64);
8248 if(regs[j].regmap[hr]==f_regmap[hr]&&f_regmap[hr]<TEMPREG) {
8249 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8250 int k;
8251 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8252 if(get_reg(regs[i].regmap,f_regmap[hr])>=0) break;
8253 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8254 k=i;
8255 while(k>1&&regs[k-1].regmap[hr]==-1) {
8256 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8257 //printf("no free regs for store %x\n",start+(k-1)*4);
8258 break;
8259 }
8260 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8261 //printf("no-match due to different register\n");
8262 break;
8263 }
8264 if (dops[k-2].is_jump) {
8265 //printf("no-match due to branch\n");
8266 break;
8267 }
8268 // call/ret fast path assumes no registers allocated
8269 if(k>2&&(dops[k-3].itype==UJUMP||dops[k-3].itype==RJUMP)&&dops[k-3].rt1==31) {
8270 break;
8271 }
8272 k--;
8273 }
8274 if(regs[k-1].regmap[hr]==f_regmap[hr]&&regmap_pre[k][hr]==f_regmap[hr]) {
8275 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8276 while(k<i) {
8277 regs[k].regmap_entry[hr]=f_regmap[hr];
8278 regs[k].regmap[hr]=f_regmap[hr];
8279 regmap_pre[k+1][hr]=f_regmap[hr];
8280 regs[k].wasdirty&=~(1<<hr);
8281 regs[k].dirty&=~(1<<hr);
8282 regs[k].wasdirty|=(1<<hr)&regs[k-1].dirty;
8283 regs[k].dirty|=(1<<hr)&regs[k].wasdirty;
8284 regs[k].wasconst&=~(1<<hr);
8285 regs[k].isconst&=~(1<<hr);
8286 k++;
8287 }
8288 }
8289 else {
8290 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8291 break;
8292 }
8293 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8294 if(regs[i-1].regmap[hr]==f_regmap[hr]&&regmap_pre[i][hr]==f_regmap[hr]) {
8295 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8296 regs[i].regmap_entry[hr]=f_regmap[hr];
8297 regs[i].regmap[hr]=f_regmap[hr];
8298 regs[i].wasdirty&=~(1<<hr);
8299 regs[i].dirty&=~(1<<hr);
8300 regs[i].wasdirty|=(1<<hr)&regs[i-1].dirty;
8301 regs[i].dirty|=(1<<hr)&regs[i-1].dirty;
8302 regs[i].wasconst&=~(1<<hr);
8303 regs[i].isconst&=~(1<<hr);
8304 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8305 branch_regs[i].wasdirty&=~(1<<hr);
8306 branch_regs[i].wasdirty|=(1<<hr)&regs[i].dirty;
8307 branch_regs[i].regmap[hr]=f_regmap[hr];
8308 branch_regs[i].dirty&=~(1<<hr);
8309 branch_regs[i].dirty|=(1<<hr)&regs[i].dirty;
8310 branch_regs[i].wasconst&=~(1<<hr);
8311 branch_regs[i].isconst&=~(1<<hr);
8312 if (!dops[i].is_ujump) {
8313 regmap_pre[i+2][hr]=f_regmap[hr];
8314 regs[i+2].wasdirty&=~(1<<hr);
8315 regs[i+2].wasdirty|=(1<<hr)&regs[i].dirty;
8316 }
8317 }
8318 }
8319 for(k=t;k<j;k++) {
8320 // Alloc register clean at beginning of loop,
8321 // but may dirty it in pass 6
8322 regs[k].regmap_entry[hr]=f_regmap[hr];
8323 regs[k].regmap[hr]=f_regmap[hr];
8324 regs[k].dirty&=~(1<<hr);
8325 regs[k].wasconst&=~(1<<hr);
8326 regs[k].isconst&=~(1<<hr);
8327 if (dops[k].is_jump) {
8328 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8329 branch_regs[k].regmap[hr]=f_regmap[hr];
8330 branch_regs[k].dirty&=~(1<<hr);
8331 branch_regs[k].wasconst&=~(1<<hr);
8332 branch_regs[k].isconst&=~(1<<hr);
8333 if (!dops[k].is_ujump) {
8334 regmap_pre[k+2][hr]=f_regmap[hr];
8335 regs[k+2].wasdirty&=~(1<<hr);
8336 }
8337 }
8338 else
8339 {
8340 regmap_pre[k+1][hr]=f_regmap[hr];
8341 regs[k+1].wasdirty&=~(1<<hr);
8342 }
8343 }
8344 if(regs[j].regmap[hr]==f_regmap[hr])
8345 regs[j].regmap_entry[hr]=f_regmap[hr];
8346 break;
8347 }
8348 if(j==i) break;
8349 if(regs[j].regmap[hr]>=0)
8350 break;
8351 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8352 //printf("no-match due to different register\n");
8353 break;
8354 }
8355 if (dops[j].is_ujump)
8356 {
8357 // Stop on unconditional branch
8358 break;
8359 }
8360 if(dops[j].itype==CJUMP||dops[j].itype==SJUMP)
8361 {
8362 if(dops[j].ooo) {
8363 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8364 break;
8365 }else{
8366 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8367 break;
8368 }
8369 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8370 //printf("no-match due to different register (branch)\n");
8371 break;
8372 }
8373 }
8374 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8375 //printf("No free regs for store %x\n",start+j*4);
8376 break;
8377 }
8378 assert(f_regmap[hr]<64);
8379 }
8380 }
8381 }
8382 }
8383 }
8384 }else{
8385 // Non branch or undetermined branch target
8386 for(hr=0;hr<HOST_REGS;hr++)
8387 {
8388 if(hr!=EXCLUDE_REG) {
8389 if(regs[i].regmap[hr]>=0) {
8390 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8391 // dealloc old register
8392 int n;
8393 for(n=0;n<HOST_REGS;n++)
8394 {
8395 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8396 }
8397 // and alloc new one
8398 f_regmap[hr]=regs[i].regmap[hr];
8399 }
8400 }
8401 }
8402 }
8403 // Try to restore cycle count at branch targets
8404 if(dops[i].bt) {
8405 for(j=i;j<slen-1;j++) {
8406 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8407 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8408 //printf("no free regs for store %x\n",start+j*4);
8409 break;
8410 }
8411 }
8412 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8413 int k=i;
8414 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8415 while(k<j) {
8416 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8417 regs[k].regmap[HOST_CCREG]=CCREG;
8418 regmap_pre[k+1][HOST_CCREG]=CCREG;
8419 regs[k+1].wasdirty|=1<<HOST_CCREG;
8420 regs[k].dirty|=1<<HOST_CCREG;
8421 regs[k].wasconst&=~(1<<HOST_CCREG);
8422 regs[k].isconst&=~(1<<HOST_CCREG);
8423 k++;
8424 }
8425 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8426 }
8427 // Work backwards from the branch target
8428 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8429 {
8430 //printf("Extend backwards\n");
8431 int k;
8432 k=i;
8433 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8434 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8435 //printf("no free regs for store %x\n",start+(k-1)*4);
8436 break;
8437 }
8438 k--;
8439 }
8440 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8441 //printf("Extend CC, %x ->\n",start+k*4);
8442 while(k<=i) {
8443 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8444 regs[k].regmap[HOST_CCREG]=CCREG;
8445 regmap_pre[k+1][HOST_CCREG]=CCREG;
8446 regs[k+1].wasdirty|=1<<HOST_CCREG;
8447 regs[k].dirty|=1<<HOST_CCREG;
8448 regs[k].wasconst&=~(1<<HOST_CCREG);
8449 regs[k].isconst&=~(1<<HOST_CCREG);
8450 k++;
8451 }
8452 }
8453 else {
8454 //printf("Fail Extend CC, %x ->\n",start+k*4);
8455 }
8456 }
8457 }
8458 if(dops[i].itype!=STORE&&dops[i].itype!=STORELR&&dops[i].itype!=C1LS&&dops[i].itype!=SHIFT&&
8459 dops[i].itype!=NOP&&dops[i].itype!=MOV&&dops[i].itype!=ALU&&dops[i].itype!=SHIFTIMM&&
8460 dops[i].itype!=IMM16&&dops[i].itype!=LOAD&&dops[i].itype!=COP1)
8461 {
8462 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8463 }
8464 }
8465 }
8466}
8467
8468// This allocates registers (if possible) one instruction prior
8469// to use, which can avoid a load-use penalty on certain CPUs.
8470static noinline void pass5b_preallocate2(void)
8471{
8472 int i, hr;
8473 for(i=0;i<slen-1;i++)
8474 {
8475 if (!i || !dops[i-1].is_jump)
8476 {
8477 if(!dops[i+1].bt)
8478 {
8479 if(dops[i].itype==ALU||dops[i].itype==MOV||dops[i].itype==LOAD||dops[i].itype==SHIFTIMM||dops[i].itype==IMM16
8480 ||((dops[i].itype==COP1||dops[i].itype==COP2)&&dops[i].opcode2<3))
8481 {
8482 if(dops[i+1].rs1) {
8483 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs1))>=0)
8484 {
8485 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8486 {
8487 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8488 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8489 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8490 regs[i].isconst&=~(1<<hr);
8491 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8492 constmap[i][hr]=constmap[i+1][hr];
8493 regs[i+1].wasdirty&=~(1<<hr);
8494 regs[i].dirty&=~(1<<hr);
8495 }
8496 }
8497 }
8498 if(dops[i+1].rs2) {
8499 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rs2))>=0)
8500 {
8501 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8502 {
8503 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8504 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8505 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8506 regs[i].isconst&=~(1<<hr);
8507 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8508 constmap[i][hr]=constmap[i+1][hr];
8509 regs[i+1].wasdirty&=~(1<<hr);
8510 regs[i].dirty&=~(1<<hr);
8511 }
8512 }
8513 }
8514 // Preload target address for load instruction (non-constant)
8515 if(dops[i+1].itype==LOAD&&dops[i+1].rs1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8516 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8517 {
8518 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8519 {
8520 regs[i].regmap[hr]=dops[i+1].rs1;
8521 regmap_pre[i+1][hr]=dops[i+1].rs1;
8522 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8523 regs[i].isconst&=~(1<<hr);
8524 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8525 constmap[i][hr]=constmap[i+1][hr];
8526 regs[i+1].wasdirty&=~(1<<hr);
8527 regs[i].dirty&=~(1<<hr);
8528 }
8529 }
8530 }
8531 // Load source into target register
8532 if(dops[i+1].use_lt1&&get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8533 if((hr=get_reg(regs[i+1].regmap,dops[i+1].rt1))>=0)
8534 {
8535 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8536 {
8537 regs[i].regmap[hr]=dops[i+1].rs1;
8538 regmap_pre[i+1][hr]=dops[i+1].rs1;
8539 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8540 regs[i].isconst&=~(1<<hr);
8541 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8542 constmap[i][hr]=constmap[i+1][hr];
8543 regs[i+1].wasdirty&=~(1<<hr);
8544 regs[i].dirty&=~(1<<hr);
8545 }
8546 }
8547 }
8548 // Address for store instruction (non-constant)
8549 if(dops[i+1].itype==STORE||dops[i+1].itype==STORELR
8550 ||(dops[i+1].opcode&0x3b)==0x39||(dops[i+1].opcode&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8551 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8552 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8553 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8554 else {
8555 regs[i+1].regmap[hr]=AGEN1+((i+1)&1);
8556 regs[i+1].isconst&=~(1<<hr);
8557 }
8558 assert(hr>=0);
8559 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8560 {
8561 regs[i].regmap[hr]=dops[i+1].rs1;
8562 regmap_pre[i+1][hr]=dops[i+1].rs1;
8563 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8564 regs[i].isconst&=~(1<<hr);
8565 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8566 constmap[i][hr]=constmap[i+1][hr];
8567 regs[i+1].wasdirty&=~(1<<hr);
8568 regs[i].dirty&=~(1<<hr);
8569 }
8570 }
8571 }
8572 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8573 if(get_reg(regs[i+1].regmap,dops[i+1].rs1)<0) {
8574 int nr;
8575 hr=get_reg(regs[i+1].regmap,FTEMP);
8576 assert(hr>=0);
8577 if(regs[i].regmap[hr]<0&&regs[i+1].regmap_entry[hr]<0)
8578 {
8579 regs[i].regmap[hr]=dops[i+1].rs1;
8580 regmap_pre[i+1][hr]=dops[i+1].rs1;
8581 regs[i+1].regmap_entry[hr]=dops[i+1].rs1;
8582 regs[i].isconst&=~(1<<hr);
8583 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8584 constmap[i][hr]=constmap[i+1][hr];
8585 regs[i+1].wasdirty&=~(1<<hr);
8586 regs[i].dirty&=~(1<<hr);
8587 }
8588 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8589 {
8590 // move it to another register
8591 regs[i+1].regmap[hr]=-1;
8592 regmap_pre[i+2][hr]=-1;
8593 regs[i+1].regmap[nr]=FTEMP;
8594 regmap_pre[i+2][nr]=FTEMP;
8595 regs[i].regmap[nr]=dops[i+1].rs1;
8596 regmap_pre[i+1][nr]=dops[i+1].rs1;
8597 regs[i+1].regmap_entry[nr]=dops[i+1].rs1;
8598 regs[i].isconst&=~(1<<nr);
8599 regs[i+1].isconst&=~(1<<nr);
8600 regs[i].dirty&=~(1<<nr);
8601 regs[i+1].wasdirty&=~(1<<nr);
8602 regs[i+1].dirty&=~(1<<nr);
8603 regs[i+2].wasdirty&=~(1<<nr);
8604 }
8605 }
8606 }
8607 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*/) {
8608 hr = -1;
8609 if(dops[i+1].itype==LOAD)
8610 hr=get_reg(regs[i+1].regmap,dops[i+1].rt1);
8611 if(dops[i+1].itype==LOADLR||(dops[i+1].opcode&0x3b)==0x31||(dops[i+1].opcode&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8612 hr=get_reg(regs[i+1].regmap,FTEMP);
8613 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
8614 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8615 if(hr<0) hr=get_reg_temp(regs[i+1].regmap);
8616 }
8617 if(hr>=0&&regs[i].regmap[hr]<0) {
8618 int rs=get_reg(regs[i+1].regmap,dops[i+1].rs1);
8619 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8620 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8621 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8622 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8623 regs[i].isconst&=~(1<<hr);
8624 regs[i+1].wasdirty&=~(1<<hr);
8625 regs[i].dirty&=~(1<<hr);
8626 }
8627 }
8628 }
8629 }
8630 }
8631 }
8632 }
8633}
8634
8635// Write back dirty registers as soon as we will no longer modify them,
8636// so that we don't end up with lots of writes at the branches.
8637static noinline void pass6_clean_registers(int istart, int iend, int wr)
8638{
8639 int i;
8640 int r;
8641 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
8642 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
8643 if(iend==slen-1) {
8644 will_dirty_i=will_dirty_next=0;
8645 wont_dirty_i=wont_dirty_next=0;
8646 }else{
8647 will_dirty_i=will_dirty_next=will_dirty[iend+1];
8648 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
8649 }
8650 for (i=iend;i>=istart;i--)
8651 {
8652 signed char rregmap_i[RRMAP_SIZE];
8653 u_int hr_candirty = 0;
8654 assert(HOST_REGS < 32);
8655 make_rregs(regs[i].regmap, rregmap_i, &hr_candirty);
8656 __builtin_prefetch(regs[i-1].regmap);
8657 if(dops[i].is_jump)
8658 {
8659 signed char branch_rregmap_i[RRMAP_SIZE];
8660 u_int branch_hr_candirty = 0;
8661 make_rregs(branch_regs[i].regmap, branch_rregmap_i, &branch_hr_candirty);
8662 if(ba[i]<start || ba[i]>=(start+slen*4))
8663 {
8664 // Branch out of this block, flush all regs
8665 will_dirty_i = 0;
8666 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8667 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8668 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8669 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8670 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8671 will_dirty_i &= branch_hr_candirty;
8672 if (dops[i].is_ujump)
8673 {
8674 // Unconditional branch
8675 wont_dirty_i = 0;
8676 // Merge in delay slot (will dirty)
8677 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8678 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8679 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8680 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8681 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8682 will_dirty_i &= hr_candirty;
8683 }
8684 else
8685 {
8686 // Conditional branch
8687 wont_dirty_i = wont_dirty_next;
8688 // Merge in delay slot (will dirty)
8689 // (the original code had no explanation why these 2 are commented out)
8690 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8691 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8692 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8693 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8694 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8695 will_dirty_i &= hr_candirty;
8696 }
8697 // Merge in delay slot (wont dirty)
8698 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8699 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8700 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8701 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8702 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8703 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8704 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8705 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8706 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8707 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8708 wont_dirty_i &= ~(1u << 31);
8709 if(wr) {
8710 #ifndef DESTRUCTIVE_WRITEBACK
8711 branch_regs[i].dirty&=wont_dirty_i;
8712 #endif
8713 branch_regs[i].dirty|=will_dirty_i;
8714 }
8715 }
8716 else
8717 {
8718 // Internal branch
8719 if(ba[i]<=start+i*4) {
8720 // Backward branch
8721 if (dops[i].is_ujump)
8722 {
8723 // Unconditional branch
8724 temp_will_dirty=0;
8725 temp_wont_dirty=0;
8726 // Merge in delay slot (will dirty)
8727 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8728 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8729 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8730 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8731 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8732 temp_will_dirty &= branch_hr_candirty;
8733 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8734 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8735 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8736 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8737 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8738 temp_will_dirty &= hr_candirty;
8739 } else {
8740 // Conditional branch (not taken case)
8741 temp_will_dirty=will_dirty_next;
8742 temp_wont_dirty=wont_dirty_next;
8743 // Merge in delay slot (will dirty)
8744 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8745 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8746 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8747 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8748 temp_will_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8749 temp_will_dirty &= branch_hr_candirty;
8750 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8751 //temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8752 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8753 temp_will_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8754 temp_will_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8755 temp_will_dirty &= hr_candirty;
8756 }
8757 // Merge in delay slot (wont dirty)
8758 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8759 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8760 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8761 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8762 temp_wont_dirty |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8763 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8764 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8765 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8766 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8767 temp_wont_dirty |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8768 temp_wont_dirty &= ~(1u << 31);
8769 // Deal with changed mappings
8770 if(i<iend) {
8771 for(r=0;r<HOST_REGS;r++) {
8772 if(r!=EXCLUDE_REG) {
8773 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
8774 temp_will_dirty&=~(1<<r);
8775 temp_wont_dirty&=~(1<<r);
8776 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8777 temp_will_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8778 temp_wont_dirty|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8779 } else {
8780 temp_will_dirty|=1<<r;
8781 temp_wont_dirty|=1<<r;
8782 }
8783 }
8784 }
8785 }
8786 }
8787 if(wr) {
8788 will_dirty[i]=temp_will_dirty;
8789 wont_dirty[i]=temp_wont_dirty;
8790 pass6_clean_registers((ba[i]-start)>>2,i-1,0);
8791 }else{
8792 // Limit recursion. It can take an excessive amount
8793 // of time if there are a lot of nested loops.
8794 will_dirty[(ba[i]-start)>>2]=0;
8795 wont_dirty[(ba[i]-start)>>2]=-1;
8796 }
8797 }
8798 /*else*/ if(1)
8799 {
8800 if (dops[i].is_ujump)
8801 {
8802 // Unconditional branch
8803 will_dirty_i=0;
8804 wont_dirty_i=0;
8805 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
8806 for(r=0;r<HOST_REGS;r++) {
8807 if(r!=EXCLUDE_REG) {
8808 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
8809 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
8810 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
8811 }
8812 if(branch_regs[i].regmap[r]>=0) {
8813 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8814 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>branch_regs[i].regmap[r])&1)<<r;
8815 }
8816 }
8817 }
8818 //}
8819 // Merge in delay slot
8820 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8821 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8822 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8823 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8824 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8825 will_dirty_i &= branch_hr_candirty;
8826 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8827 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8828 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8829 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8830 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8831 will_dirty_i &= hr_candirty;
8832 } else {
8833 // Conditional branch
8834 will_dirty_i=will_dirty_next;
8835 wont_dirty_i=wont_dirty_next;
8836 //if(ba[i]>start+i*4) // Disable recursion (for debugging)
8837 for(r=0;r<HOST_REGS;r++) {
8838 if(r!=EXCLUDE_REG) {
8839 signed char target_reg=branch_regs[i].regmap[r];
8840 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
8841 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
8842 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
8843 }
8844 else if(target_reg>=0) {
8845 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>target_reg)&1)<<r;
8846 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>target_reg)&1)<<r;
8847 }
8848 }
8849 }
8850 // Merge in delay slot
8851 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8852 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8853 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8854 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8855 will_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8856 will_dirty_i &= branch_hr_candirty;
8857 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8858 //will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8859 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8860 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8861 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8862 will_dirty_i &= hr_candirty;
8863 }
8864 // Merge in delay slot (won't dirty)
8865 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8866 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8867 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt1) & 31);
8868 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i+1].rt2) & 31);
8869 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8870 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt1) & 31);
8871 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i].rt2) & 31);
8872 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt1) & 31);
8873 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, dops[i+1].rt2) & 31);
8874 wont_dirty_i |= 1u << (get_rreg(branch_rregmap_i, CCREG) & 31);
8875 wont_dirty_i &= ~(1u << 31);
8876 if(wr) {
8877 #ifndef DESTRUCTIVE_WRITEBACK
8878 branch_regs[i].dirty&=wont_dirty_i;
8879 #endif
8880 branch_regs[i].dirty|=will_dirty_i;
8881 }
8882 }
8883 }
8884 }
8885 else if(dops[i].itype==SYSCALL||dops[i].itype==HLECALL||dops[i].itype==INTCALL)
8886 {
8887 // SYSCALL instruction (software interrupt)
8888 will_dirty_i=0;
8889 wont_dirty_i=0;
8890 }
8891 else if(dops[i].itype==COP0 && (source[i]&0x3f)==0x18)
8892 {
8893 // ERET instruction (return from interrupt)
8894 will_dirty_i=0;
8895 wont_dirty_i=0;
8896 }
8897 will_dirty_next=will_dirty_i;
8898 wont_dirty_next=wont_dirty_i;
8899 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8900 will_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8901 will_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8902 will_dirty_i &= hr_candirty;
8903 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt1) & 31);
8904 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i].rt2) & 31);
8905 wont_dirty_i |= 1u << (get_rreg(rregmap_i, CCREG) & 31);
8906 wont_dirty_i &= ~(1u << 31);
8907 if (i > istart && !dops[i].is_jump) {
8908 // Don't store a register immediately after writing it,
8909 // may prevent dual-issue.
8910 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt1) & 31);
8911 wont_dirty_i |= 1u << (get_rreg(rregmap_i, dops[i-1].rt2) & 31);
8912 }
8913 // Save it
8914 will_dirty[i]=will_dirty_i;
8915 wont_dirty[i]=wont_dirty_i;
8916 // Mark registers that won't be dirtied as not dirty
8917 if(wr) {
8918 regs[i].dirty|=will_dirty_i;
8919 #ifndef DESTRUCTIVE_WRITEBACK
8920 regs[i].dirty&=wont_dirty_i;
8921 if(dops[i].is_jump)
8922 {
8923 if (i < iend-1 && !dops[i].is_ujump) {
8924 for(r=0;r<HOST_REGS;r++) {
8925 if(r!=EXCLUDE_REG) {
8926 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
8927 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
8928 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8929 }
8930 }
8931 }
8932 }
8933 else
8934 {
8935 if(i<iend) {
8936 for(r=0;r<HOST_REGS;r++) {
8937 if(r!=EXCLUDE_REG) {
8938 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
8939 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
8940 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
8941 }
8942 }
8943 }
8944 }
8945 #endif
8946 }
8947 // Deal with changed mappings
8948 temp_will_dirty=will_dirty_i;
8949 temp_wont_dirty=wont_dirty_i;
8950 for(r=0;r<HOST_REGS;r++) {
8951 if(r!=EXCLUDE_REG) {
8952 int nr;
8953 if(regs[i].regmap[r]==regmap_pre[i][r]) {
8954 if(wr) {
8955 #ifndef DESTRUCTIVE_WRITEBACK
8956 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8957 #endif
8958 regs[i].wasdirty|=will_dirty_i&(1<<r);
8959 }
8960 }
8961 else if(regmap_pre[i][r]>=0&&(nr=get_rreg(rregmap_i,regmap_pre[i][r]))>=0) {
8962 // Register moved to a different register
8963 will_dirty_i&=~(1<<r);
8964 wont_dirty_i&=~(1<<r);
8965 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
8966 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
8967 if(wr) {
8968 #ifndef DESTRUCTIVE_WRITEBACK
8969 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
8970 #endif
8971 regs[i].wasdirty|=will_dirty_i&(1<<r);
8972 }
8973 }
8974 else {
8975 will_dirty_i&=~(1<<r);
8976 wont_dirty_i&=~(1<<r);
8977 if(regmap_pre[i][r]>0 && regmap_pre[i][r]<34) {
8978 will_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8979 wont_dirty_i|=((unneeded_reg[i]>>regmap_pre[i][r])&1)<<r;
8980 } else {
8981 wont_dirty_i|=1<<r;
8982 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
8983 }
8984 }
8985 }
8986 }
8987 }
8988}
8989
8990static noinline void pass10_expire_blocks(void)
8991{
8992 int i, end;
8993 end = (((out-ndrc->translation_cache)>>(TARGET_SIZE_2-16)) + 16384) & 65535;
8994 while (expirep != end)
8995 {
8996 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
8997 uintptr_t base_offs = ((uintptr_t)(expirep >> 13) << shift); // Base offset of this block
8998 uintptr_t base_offs_s = base_offs >> shift;
8999 inv_debug("EXP: Phase %d\n",expirep);
9000 switch((expirep>>11)&3)
9001 {
9002 case 0:
9003 // Clear jump_in and jump_dirty
9004 ll_remove_matching_addrs(jump_in+(expirep&2047),base_offs_s,shift);
9005 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base_offs_s,shift);
9006 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base_offs_s,shift);
9007 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base_offs_s,shift);
9008 break;
9009 case 1:
9010 // Clear pointers
9011 ll_kill_pointers(jump_out[expirep&2047],base_offs_s,shift);
9012 ll_kill_pointers(jump_out[(expirep&2047)+2048],base_offs_s,shift);
9013 break;
9014 case 2:
9015 // Clear hash table
9016 for(i=0;i<32;i++) {
9017 struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
9018 uintptr_t o1 = (u_char *)ht_bin->tcaddr[1] - ndrc->translation_cache;
9019 uintptr_t o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9020 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9021 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
9022 ht_bin->vaddr[1] = -1;
9023 ht_bin->tcaddr[1] = NULL;
9024 }
9025 o1 = (u_char *)ht_bin->tcaddr[0] - ndrc->translation_cache;
9026 o2 = o1 - MAX_OUTPUT_BLOCK_SIZE;
9027 if ((o1 >> shift) == base_offs_s || (o2 >> shift) == base_offs_s) {
9028 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
9029 ht_bin->vaddr[0] = ht_bin->vaddr[1];
9030 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
9031 ht_bin->vaddr[1] = -1;
9032 ht_bin->tcaddr[1] = NULL;
9033 }
9034 }
9035 break;
9036 case 3:
9037 // Clear jump_out
9038 if((expirep&2047)==0)
9039 do_clear_cache();
9040 ll_remove_matching_addrs(jump_out+(expirep&2047),base_offs_s,shift);
9041 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base_offs_s,shift);
9042 break;
9043 }
9044 expirep=(expirep+1)&65535;
9045 }
9046}
9047
9048int new_recompile_block(u_int addr)
9049{
9050 u_int pagelimit = 0;
9051 u_int state_rflags = 0;
9052 int i;
9053
9054 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
9055
9056 // this is just for speculation
9057 for (i = 1; i < 32; i++) {
9058 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
9059 state_rflags |= 1 << i;
9060 }
9061
9062 start = (u_int)addr&~3;
9063 //assert(((u_int)addr&1)==0); // start-in-delay-slot flag
9064 new_dynarec_did_compile=1;
9065 if (Config.HLE && start == 0x80001000) // hlecall
9066 {
9067 // XXX: is this enough? Maybe check hleSoftCall?
9068 void *beginning=start_block();
9069 u_int page=get_page(start);
9070
9071 invalid_code[start>>12]=0;
9072 emit_movimm(start,0);
9073 emit_writeword(0,&pcaddr);
9074 emit_far_jump(new_dyna_leave);
9075 literal_pool(0);
9076 end_block(beginning);
9077 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
9078 return 0;
9079 }
9080 else if (f1_hack && hack_addr == 0) {
9081 void *beginning = start_block();
9082 u_int page = get_page(start);
9083 emit_movimm(start, 0);
9084 emit_writeword(0, &hack_addr);
9085 emit_readword(&psxRegs.GPR.n.sp, 0);
9086 emit_readptr(&mem_rtab, 1);
9087 emit_shrimm(0, 12, 2);
9088 emit_readptr_dualindexedx_ptrlen(1, 2, 1);
9089 emit_addimm(0, 0x18, 0);
9090 emit_adds_ptr(1, 1, 1);
9091 emit_ldr_dualindexed(1, 0, 0);
9092 emit_writeword(0, &psxRegs.GPR.r[26]); // lw k0, 0x18(sp)
9093 emit_far_call(get_addr_ht);
9094 emit_jmpreg(0); // jr k0
9095 literal_pool(0);
9096 end_block(beginning);
9097
9098 ll_add_flags(jump_in + page, start, state_rflags, beginning);
9099 SysPrintf("F1 hack to %08x\n", start);
9100 return 0;
9101 }
9102
9103 cycle_multiplier_active = cycle_multiplier_override && cycle_multiplier == CYCLE_MULT_DEFAULT
9104 ? cycle_multiplier_override : cycle_multiplier;
9105
9106 source = get_source_start(start, &pagelimit);
9107 if (source == NULL) {
9108 if (addr != hack_addr) {
9109 SysPrintf("Compile at bogus memory address: %08x\n", addr);
9110 hack_addr = addr;
9111 }
9112 //abort();
9113 return -1;
9114 }
9115
9116 /* Pass 1: disassemble */
9117 /* Pass 2: register dependencies, branch targets */
9118 /* Pass 3: register allocation */
9119 /* Pass 4: branch dependencies */
9120 /* Pass 5: pre-alloc */
9121 /* Pass 6: optimize clean/dirty state */
9122 /* Pass 7: flag 32-bit registers */
9123 /* Pass 8: assembly */
9124 /* Pass 9: linker */
9125 /* Pass 10: garbage collection / free memory */
9126
9127 /* Pass 1 disassembly */
9128
9129 pass1_disassemble(pagelimit);
9130
9131 int clear_hack_addr = apply_hacks();
9132
9133 /* Pass 2 - Register dependencies and branch targets */
9134
9135 pass2_unneeded_regs(0,slen-1,0);
9136
9137 /* Pass 3 - Register allocation */
9138
9139 pass3_register_alloc(addr);
9140
9141 /* Pass 4 - Cull unused host registers */
9142
9143 pass4_cull_unused_regs();
9144
9145 /* Pass 5 - Pre-allocate registers */
9146
9147 pass5a_preallocate1();
9148 pass5b_preallocate2();
9149
9150 /* Pass 6 - Optimize clean/dirty state */
9151 pass6_clean_registers(0, slen-1, 1);
9152
9153 /* Pass 7 - Identify 32-bit registers */
9154 for (i=slen-1;i>=0;i--)
9155 {
9156 if(dops[i].itype==CJUMP||dops[i].itype==SJUMP)
9157 {
9158 // Conditional branch
9159 if((source[i]>>16)!=0x1000&&i<slen-2) {
9160 // Mark this address as a branch target since it may be called
9161 // upon return from interrupt
9162 dops[i+2].bt=1;
9163 }
9164 }
9165 }
9166
9167 if(dops[slen-1].itype==SPAN) {
9168 dops[slen-1].bt=1; // Mark as a branch target so instruction can restart after exception
9169 }
9170
9171 /* Pass 8 - Assembly */
9172 linkcount=0;stubcount=0;
9173 is_delayslot=0;
9174 u_int dirty_pre=0;
9175 void *beginning=start_block();
9176 int ds = 0;
9177 if((u_int)addr&1) {
9178 ds=1;
9179 pagespan_ds();
9180 }
9181 void *instr_addr0_override = NULL;
9182
9183 if (start == 0x80030000) {
9184 // nasty hack for the fastbios thing
9185 // override block entry to this code
9186 instr_addr0_override = out;
9187 emit_movimm(start,0);
9188 // abuse io address var as a flag that we
9189 // have already returned here once
9190 emit_readword(&address,1);
9191 emit_writeword(0,&pcaddr);
9192 emit_writeword(0,&address);
9193 emit_cmp(0,1);
9194 #ifdef __aarch64__
9195 emit_jeq(out + 4*2);
9196 emit_far_jump(new_dyna_leave);
9197 #else
9198 emit_jne(new_dyna_leave);
9199 #endif
9200 }
9201 for(i=0;i<slen;i++)
9202 {
9203 __builtin_prefetch(regs[i+1].regmap);
9204 check_regmap(regmap_pre[i]);
9205 check_regmap(regs[i].regmap_entry);
9206 check_regmap(regs[i].regmap);
9207 //if(ds) printf("ds: ");
9208 disassemble_inst(i);
9209 if(ds) {
9210 ds=0; // Skip delay slot
9211 if(dops[i].bt) assem_debug("OOPS - branch into delay slot\n");
9212 instr_addr[i] = NULL;
9213 } else {
9214 speculate_register_values(i);
9215 #ifndef DESTRUCTIVE_WRITEBACK
9216 if (i < 2 || !dops[i-2].is_ujump)
9217 {
9218 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
9219 }
9220 if((dops[i].itype==CJUMP||dops[i].itype==SJUMP)) {
9221 dirty_pre=branch_regs[i].dirty;
9222 }else{
9223 dirty_pre=regs[i].dirty;
9224 }
9225 #endif
9226 // write back
9227 if (i < 2 || !dops[i-2].is_ujump)
9228 {
9229 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
9230 loop_preload(regmap_pre[i],regs[i].regmap_entry);
9231 }
9232 // branch target entry point
9233 instr_addr[i] = out;
9234 assem_debug("<->\n");
9235 drc_dbg_emit_do_cmp(i, ccadj[i]);
9236 if (clear_hack_addr) {
9237 emit_movimm(0, 0);
9238 emit_writeword(0, &hack_addr);
9239 clear_hack_addr = 0;
9240 }
9241
9242 // load regs
9243 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&&regs[i].regmap[HOST_CCREG]!=CCREG)
9244 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
9245 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i].rs1,dops[i].rs2);
9246 address_generation(i,&regs[i],regs[i].regmap_entry);
9247 load_consts(regmap_pre[i],regs[i].regmap,i);
9248 if(dops[i].is_jump)
9249 {
9250 // Load the delay slot registers if necessary
9251 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))
9252 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9253 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))
9254 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9255 if (ram_offset && (dops[i+1].is_load || dops[i+1].is_store))
9256 load_regs(regs[i].regmap_entry,regs[i].regmap,ROREG,ROREG);
9257 if (dops[i+1].is_store)
9258 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9259 }
9260 else if(i+1<slen)
9261 {
9262 // Preload registers for following instruction
9263 if(dops[i+1].rs1!=dops[i].rs1&&dops[i+1].rs1!=dops[i].rs2)
9264 if(dops[i+1].rs1!=dops[i].rt1&&dops[i+1].rs1!=dops[i].rt2)
9265 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs1,dops[i+1].rs1);
9266 if(dops[i+1].rs2!=dops[i+1].rs1&&dops[i+1].rs2!=dops[i].rs1&&dops[i+1].rs2!=dops[i].rs2)
9267 if(dops[i+1].rs2!=dops[i].rt1&&dops[i+1].rs2!=dops[i].rt2)
9268 load_regs(regs[i].regmap_entry,regs[i].regmap,dops[i+1].rs2,dops[i+1].rs2);
9269 }
9270 // TODO: if(is_ooo(i)) address_generation(i+1);
9271 if (!dops[i].is_jump || dops[i].itype == CJUMP)
9272 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
9273 if (ram_offset && (dops[i].is_load || dops[i].is_store))
9274 load_regs(regs[i].regmap_entry,regs[i].regmap,ROREG,ROREG);
9275 if (dops[i].is_store)
9276 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
9277
9278 ds = assemble(i, &regs[i], ccadj[i]);
9279
9280 if (dops[i].is_ujump)
9281 literal_pool(1024);
9282 else
9283 literal_pool_jumpover(256);
9284 }
9285 }
9286
9287 assert(slen > 0);
9288 if (slen > 0 && dops[slen-1].itype == INTCALL) {
9289 // no ending needed for this block since INTCALL never returns
9290 }
9291 // If the block did not end with an unconditional branch,
9292 // add a jump to the next instruction.
9293 else if (i > 1) {
9294 if (!dops[i-2].is_ujump && dops[i-1].itype != SPAN) {
9295 assert(!dops[i-1].is_jump);
9296 assert(i==slen);
9297 if(dops[i-2].itype!=CJUMP&&dops[i-2].itype!=SJUMP) {
9298 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9299 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9300 emit_loadreg(CCREG,HOST_CCREG);
9301 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9302 }
9303 else
9304 {
9305 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
9306 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
9307 }
9308 add_to_linker(out,start+i*4,0);
9309 emit_jmp(0);
9310 }
9311 }
9312 else
9313 {
9314 assert(i>0);
9315 assert(!dops[i-1].is_jump);
9316 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
9317 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
9318 emit_loadreg(CCREG,HOST_CCREG);
9319 emit_addimm(HOST_CCREG, ccadj[i-1] + CLOCK_ADJUST(1), HOST_CCREG);
9320 add_to_linker(out,start+i*4,0);
9321 emit_jmp(0);
9322 }
9323
9324 // TODO: delay slot stubs?
9325 // Stubs
9326 for(i=0;i<stubcount;i++)
9327 {
9328 switch(stubs[i].type)
9329 {
9330 case LOADB_STUB:
9331 case LOADH_STUB:
9332 case LOADW_STUB:
9333 case LOADD_STUB:
9334 case LOADBU_STUB:
9335 case LOADHU_STUB:
9336 do_readstub(i);break;
9337 case STOREB_STUB:
9338 case STOREH_STUB:
9339 case STOREW_STUB:
9340 case STORED_STUB:
9341 do_writestub(i);break;
9342 case CC_STUB:
9343 do_ccstub(i);break;
9344 case INVCODE_STUB:
9345 do_invstub(i);break;
9346 case FP_STUB:
9347 do_cop1stub(i);break;
9348 case STORELR_STUB:
9349 do_unalignedwritestub(i);break;
9350 }
9351 }
9352
9353 if (instr_addr0_override)
9354 instr_addr[0] = instr_addr0_override;
9355
9356 /* Pass 9 - Linker */
9357 for(i=0;i<linkcount;i++)
9358 {
9359 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
9360 literal_pool(64);
9361 if (!link_addr[i].ext)
9362 {
9363 void *stub = out;
9364 void *addr = check_addr(link_addr[i].target);
9365 emit_extjump(link_addr[i].addr, link_addr[i].target);
9366 if (addr) {
9367 set_jump_target(link_addr[i].addr, addr);
9368 add_jump_out(link_addr[i].target,stub);
9369 }
9370 else
9371 set_jump_target(link_addr[i].addr, stub);
9372 }
9373 else
9374 {
9375 // Internal branch
9376 int target=(link_addr[i].target-start)>>2;
9377 assert(target>=0&&target<slen);
9378 assert(instr_addr[target]);
9379 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9380 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
9381 //#else
9382 set_jump_target(link_addr[i].addr, instr_addr[target]);
9383 //#endif
9384 }
9385 }
9386
9387 u_int source_len = slen*4;
9388 if (dops[slen-1].itype == INTCALL && source_len > 4)
9389 // no need to treat the last instruction as compiled
9390 // as interpreter fully handles it
9391 source_len -= 4;
9392
9393 if ((u_char *)copy + source_len > (u_char *)shadow + sizeof(shadow))
9394 copy = shadow;
9395
9396 // External Branch Targets (jump_in)
9397 for(i=0;i<slen;i++)
9398 {
9399 if(dops[i].bt||i==0)
9400 {
9401 if(instr_addr[i]) // TODO - delay slots (=null)
9402 {
9403 u_int vaddr=start+i*4;
9404 u_int page=get_page(vaddr);
9405 u_int vpage=get_vpage(vaddr);
9406 literal_pool(256);
9407 {
9408 assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
9409 assem_debug("jump_in: %x\n",start+i*4);
9410 ll_add(jump_dirty+vpage,vaddr,out);
9411 void *entry_point = do_dirty_stub(i, source_len);
9412 ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
9413 // If there was an existing entry in the hash table,
9414 // replace it with the new address.
9415 // Don't add new entries. We'll insert the
9416 // ones that actually get used in check_addr().
9417 struct ht_entry *ht_bin = hash_table_get(vaddr);
9418 if (ht_bin->vaddr[0] == vaddr)
9419 ht_bin->tcaddr[0] = entry_point;
9420 if (ht_bin->vaddr[1] == vaddr)
9421 ht_bin->tcaddr[1] = entry_point;
9422 }
9423 }
9424 }
9425 }
9426 // Write out the literal pool if necessary
9427 literal_pool(0);
9428 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
9429 // Align code
9430 if(((u_int)out)&7) emit_addnop(13);
9431 #endif
9432 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
9433 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
9434 memcpy(copy, source, source_len);
9435 copy += source_len;
9436
9437 end_block(beginning);
9438
9439 // If we're within 256K of the end of the buffer,
9440 // start over from the beginning. (Is 256K enough?)
9441 if (out > ndrc->translation_cache + sizeof(ndrc->translation_cache) - MAX_OUTPUT_BLOCK_SIZE)
9442 out = ndrc->translation_cache;
9443
9444 // Trap writes to any of the pages we compiled
9445 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
9446 invalid_code[i]=0;
9447 }
9448 inv_code_start=inv_code_end=~0;
9449
9450 // for PCSX we need to mark all mirrors too
9451 if(get_page(start)<(RAM_SIZE>>12))
9452 for(i=start>>12;i<=(start+slen*4)>>12;i++)
9453 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
9454 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
9455 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
9456
9457 /* Pass 10 - Free memory by expiring oldest blocks */
9458
9459 pass10_expire_blocks();
9460
9461#ifdef ASSEM_PRINT
9462 fflush(stdout);
9463#endif
9464 return 0;
9465}
9466
9467// vim:shiftwidth=2:expandtab
ARM optimized PCSX fork