1 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
2 * Mupen64plus - new_dynarec.c *
3 * Copyright (C) 2009-2011 Ari64 *
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. *
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. *
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 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
22 #include <stdint.h> //include for uint64_t
27 #include <libkern/OSCacheControl.h>
30 #include <3ds_utils.h>
33 #include <psp2/kernel/sysmem.h>
37 #include "new_dynarec_config.h"
38 #include "../psxhle.h" //emulator interface
39 #include "emu_if.h" //emulator interface
42 #define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
46 //#define assem_debug printf
47 //#define inv_debug printf
48 #define assem_debug(...)
49 #define inv_debug(...)
52 #include "assem_x86.h"
55 #include "assem_x64.h"
58 #include "assem_arm.h"
61 #include "assem_arm64.h"
65 #define MAX_OUTPUT_BLOCK_SIZE 262144
87 signed char regmap_entry[HOST_REGS];
88 signed char regmap[HOST_REGS];
94 u_int loadedconst; // host regs that have constants loaded
95 u_int waswritten; // MIPS regs that were used as store base before
98 // note: asm depends on this layout
104 struct ll_entry *next;
134 struct ht_entry hash_table[65536] __attribute__((aligned(16)));
135 struct ll_entry *jump_in[4096] __attribute__((aligned(16)));
136 struct ll_entry *jump_dirty[4096];
138 static struct ll_entry *jump_out[4096];
140 static u_int *source;
141 static char insn[MAXBLOCK][10];
142 static u_char itype[MAXBLOCK];
143 static u_char opcode[MAXBLOCK];
144 static u_char opcode2[MAXBLOCK];
145 static u_char bt[MAXBLOCK];
146 static u_char rs1[MAXBLOCK];
147 static u_char rs2[MAXBLOCK];
148 static u_char rt1[MAXBLOCK];
149 static u_char rt2[MAXBLOCK];
150 static u_char us1[MAXBLOCK];
151 static u_char us2[MAXBLOCK];
152 static u_char dep1[MAXBLOCK];
153 static u_char dep2[MAXBLOCK];
154 static u_char lt1[MAXBLOCK];
155 static uint64_t gte_rs[MAXBLOCK]; // gte: 32 data and 32 ctl regs
156 static uint64_t gte_rt[MAXBLOCK];
157 static uint64_t gte_unneeded[MAXBLOCK];
158 static u_int smrv[32]; // speculated MIPS register values
159 static u_int smrv_strong; // mask or regs that are likely to have correct values
160 static u_int smrv_weak; // same, but somewhat less likely
161 static u_int smrv_strong_next; // same, but after current insn executes
162 static u_int smrv_weak_next;
163 static int imm[MAXBLOCK];
164 static u_int ba[MAXBLOCK];
165 static char likely[MAXBLOCK];
166 static char is_ds[MAXBLOCK];
167 static char ooo[MAXBLOCK];
168 static uint64_t unneeded_reg[MAXBLOCK];
169 static uint64_t branch_unneeded_reg[MAXBLOCK];
170 static signed char regmap_pre[MAXBLOCK][HOST_REGS];
171 static uint64_t current_constmap[HOST_REGS];
172 static uint64_t constmap[MAXBLOCK][HOST_REGS];
173 static struct regstat regs[MAXBLOCK];
174 static struct regstat branch_regs[MAXBLOCK];
175 static signed char minimum_free_regs[MAXBLOCK];
176 static u_int needed_reg[MAXBLOCK];
177 static u_int wont_dirty[MAXBLOCK];
178 static u_int will_dirty[MAXBLOCK];
179 static int ccadj[MAXBLOCK];
181 static void *instr_addr[MAXBLOCK];
182 static struct link_entry link_addr[MAXBLOCK];
183 static int linkcount;
184 static struct code_stub stubs[MAXBLOCK*3];
185 static int stubcount;
186 static u_int literals[1024][2];
187 static int literalcount;
188 static int is_delayslot;
189 static char shadow[1048576] __attribute__((aligned(16)));
192 static u_int stop_after_jal;
194 static uintptr_t ram_offset;
196 static const uintptr_t ram_offset=0;
199 int new_dynarec_hacks;
200 int new_dynarec_did_compile;
202 extern int cycle_count; // ... until end of the timeslice, counts -N -> 0
203 extern int last_count; // last absolute target, often = next_interupt
205 extern int pending_exception;
206 extern int branch_target;
207 extern u_int mini_ht[32][2];
208 extern u_char restore_candidate[512];
210 /* registers that may be allocated */
212 #define HIREG 32 // hi
213 #define LOREG 33 // lo
214 //#define FSREG 34 // FPU status (FCSR)
215 #define CSREG 35 // Coprocessor status
216 #define CCREG 36 // Cycle count
217 #define INVCP 37 // Pointer to invalid_code
218 //#define MMREG 38 // Pointer to memory_map
219 //#define ROREG 39 // ram offset (if rdram!=0x80000000)
221 #define FTEMP 40 // FPU temporary register
222 #define PTEMP 41 // Prefetch temporary register
223 //#define TLREG 42 // TLB mapping offset
224 #define RHASH 43 // Return address hash
225 #define RHTBL 44 // Return address hash table address
226 #define RTEMP 45 // JR/JALR address register
228 #define AGEN1 46 // Address generation temporary register
229 //#define AGEN2 47 // Address generation temporary register
230 //#define MGEN1 48 // Maptable address generation temporary register
231 //#define MGEN2 49 // Maptable address generation temporary register
232 #define BTREG 50 // Branch target temporary register
234 /* instruction types */
235 #define NOP 0 // No operation
236 #define LOAD 1 // Load
237 #define STORE 2 // Store
238 #define LOADLR 3 // Unaligned load
239 #define STORELR 4 // Unaligned store
240 #define MOV 5 // Move
241 #define ALU 6 // Arithmetic/logic
242 #define MULTDIV 7 // Multiply/divide
243 #define SHIFT 8 // Shift by register
244 #define SHIFTIMM 9// Shift by immediate
245 #define IMM16 10 // 16-bit immediate
246 #define RJUMP 11 // Unconditional jump to register
247 #define UJUMP 12 // Unconditional jump
248 #define CJUMP 13 // Conditional branch (BEQ/BNE/BGTZ/BLEZ)
249 #define SJUMP 14 // Conditional branch (regimm format)
250 #define COP0 15 // Coprocessor 0
251 #define COP1 16 // Coprocessor 1
252 #define C1LS 17 // Coprocessor 1 load/store
253 //#define FJUMP 18 // Conditional branch (floating point)
254 //#define FLOAT 19 // Floating point unit
255 //#define FCONV 20 // Convert integer to float
256 //#define FCOMP 21 // Floating point compare (sets FSREG)
257 #define SYSCALL 22// SYSCALL
258 #define OTHER 23 // Other
259 #define SPAN 24 // Branch/delay slot spans 2 pages
260 #define NI 25 // Not implemented
261 #define HLECALL 26// PCSX fake opcodes for HLE
262 #define COP2 27 // Coprocessor 2 move
263 #define C2LS 28 // Coprocessor 2 load/store
264 #define C2OP 29 // Coprocessor 2 operation
265 #define INTCALL 30// Call interpreter to handle rare corner cases
273 int new_recompile_block(int addr);
274 void *get_addr_ht(u_int vaddr);
275 void invalidate_block(u_int block);
276 void invalidate_addr(u_int addr);
277 void remove_hash(int vaddr);
279 void dyna_linker_ds();
281 void verify_code_vm();
282 void verify_code_ds();
285 void fp_exception_ds();
286 void jump_syscall_hle();
289 void new_dyna_leave();
291 // Needed by assembler
292 static void wb_register(signed char r,signed char regmap[],uint64_t dirty);
293 static void wb_dirtys(signed char i_regmap[],uint64_t i_dirty);
294 static void wb_needed_dirtys(signed char i_regmap[],uint64_t i_dirty,int addr);
295 static void load_all_regs(signed char i_regmap[]);
296 static void load_needed_regs(signed char i_regmap[],signed char next_regmap[]);
297 static void load_regs_entry(int t);
298 static void load_all_consts(signed char regmap[],u_int dirty,int i);
300 static int verify_dirty(u_int *ptr);
301 static int get_final_value(int hr, int i, int *value);
302 static void add_stub(enum stub_type type, void *addr, void *retaddr,
303 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e);
304 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
305 int i, int addr_reg, struct regstat *i_regs, int ccadj, u_int reglist);
306 static void add_to_linker(void *addr, u_int target, int ext);
307 static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override);
308 static void *get_direct_memhandler(void *table, u_int addr,
309 enum stub_type type, uintptr_t *addr_host);
310 static void pass_args(int a0, int a1);
312 static void mprotect_w_x(void *start, void *end, int is_x)
316 // *Open* enables write on all memory that was
317 // allocated by sceKernelAllocMemBlockForVM()?
319 sceKernelCloseVMDomain();
321 sceKernelOpenVMDomain();
323 u_long mstart = (u_long)start & ~4095ul;
324 u_long mend = (u_long)end;
325 if (mprotect((void *)mstart, mend - mstart,
326 PROT_READ | (is_x ? PROT_EXEC : PROT_WRITE)) != 0)
327 SysPrintf("mprotect(%c) failed: %s\n", is_x ? 'x' : 'w', strerror(errno));
332 static void start_tcache_write(void *start, void *end)
334 mprotect_w_x(start, end, 0);
337 static void end_tcache_write(void *start, void *end)
340 size_t len = (char *)end - (char *)start;
341 #if defined(__BLACKBERRY_QNX__)
342 msync(start, len, MS_SYNC | MS_CACHE_ONLY | MS_INVALIDATE_ICACHE);
343 #elif defined(__MACH__)
344 sys_cache_control(kCacheFunctionPrepareForExecution, start, len);
346 sceKernelSyncVMDomain(sceBlock, start, len);
348 ctr_flush_invalidate_cache();
350 __clear_cache(start, end);
354 __clear_cache(start, end);
357 mprotect_w_x(start, end, 1);
360 static void *start_block(void)
362 u_char *end = out + MAX_OUTPUT_BLOCK_SIZE;
363 if (end > translation_cache + (1<<TARGET_SIZE_2))
364 end = translation_cache + (1<<TARGET_SIZE_2);
365 start_tcache_write(out, end);
369 static void end_block(void *start)
371 end_tcache_write(start, out);
374 //#define DEBUG_CYCLE_COUNT 1
376 #define NO_CYCLE_PENALTY_THR 12
378 int cycle_multiplier; // 100 for 1.0
380 static int CLOCK_ADJUST(int x)
383 return (x * cycle_multiplier + s * 50) / 100;
386 static u_int get_page(u_int vaddr)
388 u_int page=vaddr&~0xe0000000;
389 if (page < 0x1000000)
390 page &= ~0x0e00000; // RAM mirrors
392 if(page>2048) page=2048+(page&2047);
396 // no virtual mem in PCSX
397 static u_int get_vpage(u_int vaddr)
399 return get_page(vaddr);
402 static struct ht_entry *hash_table_get(u_int vaddr)
404 return &hash_table[((vaddr>>16)^vaddr)&0xFFFF];
407 static void hash_table_add(struct ht_entry *ht_bin, u_int vaddr, void *tcaddr)
409 ht_bin->vaddr[1] = ht_bin->vaddr[0];
410 ht_bin->tcaddr[1] = ht_bin->tcaddr[0];
411 ht_bin->vaddr[0] = vaddr;
412 ht_bin->tcaddr[0] = tcaddr;
415 // some messy ari64's code, seems to rely on unsigned 32bit overflow
416 static int doesnt_expire_soon(void *tcaddr)
418 u_int diff = (u_int)((u_char *)tcaddr - out) << (32-TARGET_SIZE_2);
419 return diff > (u_int)(0x60000000 + (MAX_OUTPUT_BLOCK_SIZE << (32-TARGET_SIZE_2)));
422 // Get address from virtual address
423 // This is called from the recompiled JR/JALR instructions
424 void *get_addr(u_int vaddr)
426 u_int page=get_page(vaddr);
427 u_int vpage=get_vpage(vaddr);
428 struct ll_entry *head;
429 //printf("TRACE: count=%d next=%d (get_addr %x,page %d)\n",Count,next_interupt,vaddr,page);
432 if(head->vaddr==vaddr) {
433 //printf("TRACE: count=%d next=%d (get_addr match %x: %p)\n",Count,next_interupt,vaddr,head->addr);
434 hash_table_add(hash_table_get(vaddr), vaddr, head->addr);
439 head=jump_dirty[vpage];
441 if(head->vaddr==vaddr) {
442 //printf("TRACE: count=%d next=%d (get_addr match dirty %x: %p)\n",Count,next_interupt,vaddr,head->addr);
443 // Don't restore blocks which are about to expire from the cache
444 if (doesnt_expire_soon(head->addr))
445 if (verify_dirty(head->addr)) {
446 //printf("restore candidate: %x (%d) d=%d\n",vaddr,page,invalid_code[vaddr>>12]);
447 invalid_code[vaddr>>12]=0;
448 inv_code_start=inv_code_end=~0;
450 restore_candidate[vpage>>3]|=1<<(vpage&7);
452 else restore_candidate[page>>3]|=1<<(page&7);
453 struct ht_entry *ht_bin = hash_table_get(vaddr);
454 if (ht_bin->vaddr[0] == vaddr)
455 ht_bin->tcaddr[0] = head->addr; // Replace existing entry
457 hash_table_add(ht_bin, vaddr, head->addr);
464 //printf("TRACE: count=%d next=%d (get_addr no-match %x)\n",Count,next_interupt,vaddr);
465 int r=new_recompile_block(vaddr);
466 if(r==0) return get_addr(vaddr);
467 // Execute in unmapped page, generate pagefault execption
469 Cause=(vaddr<<31)|0x8;
470 EPC=(vaddr&1)?vaddr-5:vaddr;
472 Context=(Context&0xFF80000F)|((BadVAddr>>9)&0x007FFFF0);
473 EntryHi=BadVAddr&0xFFFFE000;
474 return get_addr_ht(0x80000000);
476 // Look up address in hash table first
477 void *get_addr_ht(u_int vaddr)
479 //printf("TRACE: count=%d next=%d (get_addr_ht %x)\n",Count,next_interupt,vaddr);
480 const struct ht_entry *ht_bin = hash_table_get(vaddr);
481 if (ht_bin->vaddr[0] == vaddr) return ht_bin->tcaddr[0];
482 if (ht_bin->vaddr[1] == vaddr) return ht_bin->tcaddr[1];
483 return get_addr(vaddr);
486 void clear_all_regs(signed char regmap[])
489 for (hr=0;hr<HOST_REGS;hr++) regmap[hr]=-1;
492 signed char get_reg(signed char regmap[],int r)
495 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map[hr]==r) return hr;
499 // Find a register that is available for two consecutive cycles
500 signed char get_reg2(signed char regmap1[],signed char regmap2[],int r)
503 for (hr=0;hr<HOST_REGS;hr++) if(hr!=EXCLUDE_REG&®map1[hr]==r&®map2[hr]==r) return hr;
507 int count_free_regs(signed char regmap[])
511 for(hr=0;hr<HOST_REGS;hr++)
513 if(hr!=EXCLUDE_REG) {
514 if(regmap[hr]<0) count++;
520 void dirty_reg(struct regstat *cur,signed char reg)
524 for (hr=0;hr<HOST_REGS;hr++) {
525 if((cur->regmap[hr]&63)==reg) {
531 void set_const(struct regstat *cur,signed char reg,uint64_t value)
535 for (hr=0;hr<HOST_REGS;hr++) {
536 if(cur->regmap[hr]==reg) {
538 current_constmap[hr]=value;
540 else if((cur->regmap[hr]^64)==reg) {
542 current_constmap[hr]=value>>32;
547 void clear_const(struct regstat *cur,signed char reg)
551 for (hr=0;hr<HOST_REGS;hr++) {
552 if((cur->regmap[hr]&63)==reg) {
553 cur->isconst&=~(1<<hr);
558 int is_const(struct regstat *cur,signed char reg)
563 for (hr=0;hr<HOST_REGS;hr++) {
564 if((cur->regmap[hr]&63)==reg) {
565 return (cur->isconst>>hr)&1;
570 uint64_t get_const(struct regstat *cur,signed char reg)
574 for (hr=0;hr<HOST_REGS;hr++) {
575 if(cur->regmap[hr]==reg) {
576 return current_constmap[hr];
579 SysPrintf("Unknown constant in r%d\n",reg);
583 // Least soon needed registers
584 // Look at the next ten instructions and see which registers
585 // will be used. Try not to reallocate these.
586 void lsn(u_char hsn[], int i, int *preferred_reg)
596 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
598 // Don't go past an unconditonal jump
605 if(rs1[i+j]) hsn[rs1[i+j]]=j;
606 if(rs2[i+j]) hsn[rs2[i+j]]=j;
607 if(rt1[i+j]) hsn[rt1[i+j]]=j;
608 if(rt2[i+j]) hsn[rt2[i+j]]=j;
609 if(itype[i+j]==STORE || itype[i+j]==STORELR) {
610 // Stores can allocate zero
614 // On some architectures stores need invc_ptr
615 #if defined(HOST_IMM8)
616 if(itype[i+j]==STORE || itype[i+j]==STORELR || (opcode[i+j]&0x3b)==0x39 || (opcode[i+j]&0x3b)==0x3a) {
620 if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP))
628 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
630 // Follow first branch
631 int t=(ba[i+b]-start)>>2;
632 j=7-b;if(t+j>=slen) j=slen-t-1;
635 if(rs1[t+j]) if(hsn[rs1[t+j]]>j+b+2) hsn[rs1[t+j]]=j+b+2;
636 if(rs2[t+j]) if(hsn[rs2[t+j]]>j+b+2) hsn[rs2[t+j]]=j+b+2;
637 //if(rt1[t+j]) if(hsn[rt1[t+j]]>j+b+2) hsn[rt1[t+j]]=j+b+2;
638 //if(rt2[t+j]) if(hsn[rt2[t+j]]>j+b+2) hsn[rt2[t+j]]=j+b+2;
641 // TODO: preferred register based on backward branch
643 // Delay slot should preferably not overwrite branch conditions or cycle count
644 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP)) {
645 if(rs1[i-1]) if(hsn[rs1[i-1]]>1) hsn[rs1[i-1]]=1;
646 if(rs2[i-1]) if(hsn[rs2[i-1]]>1) hsn[rs2[i-1]]=1;
652 // Coprocessor load/store needs FTEMP, even if not declared
653 if(itype[i]==C1LS||itype[i]==C2LS) {
656 // Load L/R also uses FTEMP as a temporary register
657 if(itype[i]==LOADLR) {
660 // Also SWL/SWR/SDL/SDR
661 if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) {
664 // Don't remove the miniht registers
665 if(itype[i]==UJUMP||itype[i]==RJUMP)
672 // We only want to allocate registers if we're going to use them again soon
673 int needed_again(int r, int i)
679 if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000))
681 if(ba[i-1]<start || ba[i-1]>start+slen*4-4)
682 return 0; // Don't need any registers if exiting the block
690 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
692 // Don't go past an unconditonal jump
696 if(itype[i+j]==SYSCALL||itype[i+j]==HLECALL||itype[i+j]==INTCALL||((source[i+j]&0xfc00003f)==0x0d))
703 if(rs1[i+j]==r) rn=j;
704 if(rs2[i+j]==r) rn=j;
705 if((unneeded_reg[i+j]>>r)&1) rn=10;
706 if(i+j>=0&&(itype[i+j]==UJUMP||itype[i+j]==CJUMP||itype[i+j]==SJUMP))
714 if(ba[i+b]>=start && ba[i+b]<(start+slen*4))
716 // Follow first branch
718 int t=(ba[i+b]-start)>>2;
719 j=7-b;if(t+j>=slen) j=slen-t-1;
722 if(!((unneeded_reg[t+j]>>r)&1)) {
723 if(rs1[t+j]==r) if(rn>j+b+2) rn=j+b+2;
724 if(rs2[t+j]==r) if(rn>j+b+2) rn=j+b+2;
735 // Try to match register allocations at the end of a loop with those
737 int loop_reg(int i, int r, int hr)
746 if(itype[i+j]==UJUMP||itype[i+j]==RJUMP||(source[i+j]>>16)==0x1000)
748 // Don't go past an unconditonal jump
755 if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP)
761 if((unneeded_reg[i+k]>>r)&1) return hr;
762 if(i+k>=0&&(itype[i+k]==UJUMP||itype[i+k]==CJUMP||itype[i+k]==SJUMP))
764 if(ba[i+k]>=start && ba[i+k]<(start+i*4))
766 int t=(ba[i+k]-start)>>2;
767 int reg=get_reg(regs[t].regmap_entry,r);
768 if(reg>=0) return reg;
769 //reg=get_reg(regs[t+1].regmap_entry,r);
770 //if(reg>=0) return reg;
778 // Allocate every register, preserving source/target regs
779 void alloc_all(struct regstat *cur,int i)
783 for(hr=0;hr<HOST_REGS;hr++) {
784 if(hr!=EXCLUDE_REG) {
785 if(((cur->regmap[hr]&63)!=rs1[i])&&((cur->regmap[hr]&63)!=rs2[i])&&
786 ((cur->regmap[hr]&63)!=rt1[i])&&((cur->regmap[hr]&63)!=rt2[i]))
789 cur->dirty&=~(1<<hr);
792 if((cur->regmap[hr]&63)==0)
795 cur->dirty&=~(1<<hr);
802 extern void gen_interupt();
803 extern void do_insn_cmp();
804 #define FUNCNAME(f) { (intptr_t)f, " " #f }
805 static const struct {
808 } function_names[] = {
809 FUNCNAME(cc_interrupt),
810 FUNCNAME(gen_interupt),
811 FUNCNAME(get_addr_ht),
813 FUNCNAME(jump_handler_read8),
814 FUNCNAME(jump_handler_read16),
815 FUNCNAME(jump_handler_read32),
816 FUNCNAME(jump_handler_write8),
817 FUNCNAME(jump_handler_write16),
818 FUNCNAME(jump_handler_write32),
819 FUNCNAME(invalidate_addr),
820 FUNCNAME(verify_code_vm),
821 FUNCNAME(verify_code),
822 FUNCNAME(jump_hlecall),
823 FUNCNAME(jump_syscall_hle),
824 FUNCNAME(new_dyna_leave),
826 FUNCNAME(pcsx_mtc0_ds),
827 FUNCNAME(do_insn_cmp),
830 static const char *func_name(intptr_t a)
833 for (i = 0; i < sizeof(function_names)/sizeof(function_names[0]); i++)
834 if (function_names[i].addr == a)
835 return function_names[i].name;
839 #define func_name(x) ""
843 #include "assem_x86.c"
846 #include "assem_x64.c"
849 #include "assem_arm.c"
852 #include "assem_arm64.c"
855 // Add virtual address mapping to linked list
856 void ll_add(struct ll_entry **head,int vaddr,void *addr)
858 struct ll_entry *new_entry;
859 new_entry=malloc(sizeof(struct ll_entry));
860 assert(new_entry!=NULL);
861 new_entry->vaddr=vaddr;
862 new_entry->reg_sv_flags=0;
863 new_entry->addr=addr;
864 new_entry->next=*head;
868 void ll_add_flags(struct ll_entry **head,int vaddr,u_int reg_sv_flags,void *addr)
870 ll_add(head,vaddr,addr);
871 (*head)->reg_sv_flags=reg_sv_flags;
874 // Check if an address is already compiled
875 // but don't return addresses which are about to expire from the cache
876 void *check_addr(u_int vaddr)
878 struct ht_entry *ht_bin = hash_table_get(vaddr);
880 for (i = 0; i < ARRAY_SIZE(ht_bin->vaddr); i++) {
881 if (ht_bin->vaddr[i] == vaddr)
882 if (doesnt_expire_soon((u_char *)ht_bin->tcaddr[i] - MAX_OUTPUT_BLOCK_SIZE))
883 if (isclean(ht_bin->tcaddr[i]))
884 return ht_bin->tcaddr[i];
886 u_int page=get_page(vaddr);
887 struct ll_entry *head;
889 while (head != NULL) {
890 if (head->vaddr == vaddr) {
891 if (doesnt_expire_soon(head->addr)) {
892 // Update existing entry with current address
893 if (ht_bin->vaddr[0] == vaddr) {
894 ht_bin->tcaddr[0] = head->addr;
897 if (ht_bin->vaddr[1] == vaddr) {
898 ht_bin->tcaddr[1] = head->addr;
901 // Insert into hash table with low priority.
902 // Don't evict existing entries, as they are probably
903 // addresses that are being accessed frequently.
904 if (ht_bin->vaddr[0] == -1) {
905 ht_bin->vaddr[0] = vaddr;
906 ht_bin->tcaddr[0] = head->addr;
908 else if (ht_bin->vaddr[1] == -1) {
909 ht_bin->vaddr[1] = vaddr;
910 ht_bin->tcaddr[1] = head->addr;
920 void remove_hash(int vaddr)
922 //printf("remove hash: %x\n",vaddr);
923 struct ht_entry *ht_bin = hash_table_get(vaddr);
924 if (ht_bin->vaddr[1] == vaddr) {
925 ht_bin->vaddr[1] = -1;
926 ht_bin->tcaddr[1] = NULL;
928 if (ht_bin->vaddr[0] == vaddr) {
929 ht_bin->vaddr[0] = ht_bin->vaddr[1];
930 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
931 ht_bin->vaddr[1] = -1;
932 ht_bin->tcaddr[1] = NULL;
936 void ll_remove_matching_addrs(struct ll_entry **head,uintptr_t addr,int shift)
938 struct ll_entry *next;
940 if(((uintptr_t)((*head)->addr)>>shift)==(addr>>shift) ||
941 ((uintptr_t)((*head)->addr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift))
943 inv_debug("EXP: Remove pointer to %p (%x)\n",(*head)->addr,(*head)->vaddr);
944 remove_hash((*head)->vaddr);
951 head=&((*head)->next);
956 // Remove all entries from linked list
957 void ll_clear(struct ll_entry **head)
959 struct ll_entry *cur;
960 struct ll_entry *next;
971 // Dereference the pointers and remove if it matches
972 static void ll_kill_pointers(struct ll_entry *head,uintptr_t addr,int shift)
975 uintptr_t ptr = (uintptr_t)get_pointer(head->addr);
976 inv_debug("EXP: Lookup pointer to %lx at %p (%x)\n",(long)ptr,head->addr,head->vaddr);
977 if(((ptr>>shift)==(addr>>shift)) ||
978 (((ptr-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(addr>>shift)))
980 inv_debug("EXP: Kill pointer at %p (%x)\n",head->addr,head->vaddr);
981 void *host_addr=find_extjump_insn(head->addr);
982 #if defined(__arm__) || defined(__aarch64__)
983 mark_clear_cache(host_addr);
985 set_jump_target(host_addr, head->addr);
991 // This is called when we write to a compiled block (see do_invstub)
992 void invalidate_page(u_int page)
994 struct ll_entry *head;
995 struct ll_entry *next;
999 inv_debug("INVALIDATE: %x\n",head->vaddr);
1000 remove_hash(head->vaddr);
1005 head=jump_out[page];
1008 inv_debug("INVALIDATE: kill pointer to %x (%p)\n",head->vaddr,head->addr);
1009 void *host_addr=find_extjump_insn(head->addr);
1010 #if defined(__arm__) || defined(__aarch64__)
1011 mark_clear_cache(host_addr);
1013 set_jump_target(host_addr, head->addr);
1020 static void invalidate_block_range(u_int block, u_int first, u_int last)
1022 u_int page=get_page(block<<12);
1023 //printf("first=%d last=%d\n",first,last);
1024 invalidate_page(page);
1025 assert(first+5>page); // NB: this assumes MAXBLOCK<=4096 (4 pages)
1026 assert(last<page+5);
1027 // Invalidate the adjacent pages if a block crosses a 4K boundary
1029 invalidate_page(first);
1032 for(first=page+1;first<last;first++) {
1033 invalidate_page(first);
1035 #if defined(__arm__) || defined(__aarch64__)
1039 // Don't trap writes
1040 invalid_code[block]=1;
1043 memset(mini_ht,-1,sizeof(mini_ht));
1047 void invalidate_block(u_int block)
1049 u_int page=get_page(block<<12);
1050 u_int vpage=get_vpage(block<<12);
1051 inv_debug("INVALIDATE: %x (%d)\n",block<<12,page);
1052 //inv_debug("invalid_code[block]=%d\n",invalid_code[block]);
1055 struct ll_entry *head;
1056 head=jump_dirty[vpage];
1057 //printf("page=%d vpage=%d\n",page,vpage);
1059 if(vpage>2047||(head->vaddr>>12)==block) { // Ignore vaddr hash collision
1060 u_char *start, *end;
1061 get_bounds(head->addr, &start, &end);
1062 //printf("start: %p end: %p\n", start, end);
1063 if (page < 2048 && start >= rdram && end < rdram+RAM_SIZE) {
1064 if (((start-rdram)>>12) <= page && ((end-1-rdram)>>12) >= page) {
1065 if ((((start-rdram)>>12)&2047) < first) first = ((start-rdram)>>12)&2047;
1066 if ((((end-1-rdram)>>12)&2047) > last) last = ((end-1-rdram)>>12)&2047;
1072 invalidate_block_range(block,first,last);
1075 void invalidate_addr(u_int addr)
1078 // this check is done by the caller
1079 //if (inv_code_start<=addr&&addr<=inv_code_end) { rhits++; return; }
1080 u_int page=get_vpage(addr);
1081 if(page<2048) { // RAM
1082 struct ll_entry *head;
1083 u_int addr_min=~0, addr_max=0;
1084 u_int mask=RAM_SIZE-1;
1085 u_int addr_main=0x80000000|(addr&mask);
1087 inv_code_start=addr_main&~0xfff;
1088 inv_code_end=addr_main|0xfff;
1091 // must check previous page too because of spans..
1093 inv_code_start-=0x1000;
1095 for(;pg1<=page;pg1++) {
1096 for(head=jump_dirty[pg1];head!=NULL;head=head->next) {
1097 u_char *start_h, *end_h;
1099 get_bounds(head->addr, &start_h, &end_h);
1100 start = (uintptr_t)start_h - ram_offset;
1101 end = (uintptr_t)end_h - ram_offset;
1102 if(start<=addr_main&&addr_main<end) {
1103 if(start<addr_min) addr_min=start;
1104 if(end>addr_max) addr_max=end;
1106 else if(addr_main<start) {
1107 if(start<inv_code_end)
1108 inv_code_end=start-1;
1111 if(end>inv_code_start)
1117 inv_debug("INV ADDR: %08x hit %08x-%08x\n", addr, addr_min, addr_max);
1118 inv_code_start=inv_code_end=~0;
1119 invalidate_block_range(addr>>12,(addr_min&mask)>>12,(addr_max&mask)>>12);
1123 inv_code_start=(addr&~mask)|(inv_code_start&mask);
1124 inv_code_end=(addr&~mask)|(inv_code_end&mask);
1125 inv_debug("INV ADDR: %08x miss, inv %08x-%08x, sk %d\n", addr, inv_code_start, inv_code_end, 0);
1129 invalidate_block(addr>>12);
1132 // This is called when loading a save state.
1133 // Anything could have changed, so invalidate everything.
1134 void invalidate_all_pages()
1137 for(page=0;page<4096;page++)
1138 invalidate_page(page);
1139 for(page=0;page<1048576;page++)
1140 if(!invalid_code[page]) {
1141 restore_candidate[(page&2047)>>3]|=1<<(page&7);
1142 restore_candidate[((page&2047)>>3)+256]|=1<<(page&7);
1145 memset(mini_ht,-1,sizeof(mini_ht));
1149 // Add an entry to jump_out after making a link
1150 void add_link(u_int vaddr,void *src)
1152 u_int page=get_page(vaddr);
1153 inv_debug("add_link: %p -> %x (%d)\n",src,vaddr,page);
1154 int *ptr=(int *)(src+4);
1155 assert((*ptr&0x0fff0000)==0x059f0000);
1157 ll_add(jump_out+page,vaddr,src);
1158 //void *ptr=get_pointer(src);
1159 //inv_debug("add_link: Pointer is to %p\n",ptr);
1162 // If a code block was found to be unmodified (bit was set in
1163 // restore_candidate) and it remains unmodified (bit is clear
1164 // in invalid_code) then move the entries for that 4K page from
1165 // the dirty list to the clean list.
1166 void clean_blocks(u_int page)
1168 struct ll_entry *head;
1169 inv_debug("INV: clean_blocks page=%d\n",page);
1170 head=jump_dirty[page];
1172 if(!invalid_code[head->vaddr>>12]) {
1173 // Don't restore blocks which are about to expire from the cache
1174 if (doesnt_expire_soon(head->addr)) {
1175 if(verify_dirty(head->addr)) {
1176 u_char *start, *end;
1177 //printf("Possibly Restore %x (%p)\n",head->vaddr, head->addr);
1180 get_bounds(head->addr, &start, &end);
1181 if (start - rdram < RAM_SIZE) {
1182 for (i = (start-rdram+0x80000000)>>12; i <= (end-1-rdram+0x80000000)>>12; i++) {
1183 inv|=invalid_code[i];
1186 else if((signed int)head->vaddr>=(signed int)0x80000000+RAM_SIZE) {
1190 void *clean_addr = get_clean_addr(head->addr);
1191 if (doesnt_expire_soon(clean_addr)) {
1193 inv_debug("INV: Restored %x (%p/%p)\n",head->vaddr, head->addr, clean_addr);
1194 //printf("page=%x, addr=%x\n",page,head->vaddr);
1195 //assert(head->vaddr>>12==(page|0x80000));
1196 ll_add_flags(jump_in+ppage,head->vaddr,head->reg_sv_flags,clean_addr);
1197 struct ht_entry *ht_bin = hash_table_get(head->vaddr);
1198 if (ht_bin->vaddr[0] == head->vaddr)
1199 ht_bin->tcaddr[0] = clean_addr; // Replace existing entry
1200 if (ht_bin->vaddr[1] == head->vaddr)
1201 ht_bin->tcaddr[1] = clean_addr; // Replace existing entry
1211 /* Register allocation */
1213 // Note: registers are allocated clean (unmodified state)
1214 // if you intend to modify the register, you must call dirty_reg().
1215 static void alloc_reg(struct regstat *cur,int i,signed char reg)
1218 int preferred_reg = (reg&7);
1219 if(reg==CCREG) preferred_reg=HOST_CCREG;
1220 if(reg==PTEMP||reg==FTEMP) preferred_reg=12;
1222 // Don't allocate unused registers
1223 if((cur->u>>reg)&1) return;
1225 // see if it's already allocated
1226 for(hr=0;hr<HOST_REGS;hr++)
1228 if(cur->regmap[hr]==reg) return;
1231 // Keep the same mapping if the register was already allocated in a loop
1232 preferred_reg = loop_reg(i,reg,preferred_reg);
1234 // Try to allocate the preferred register
1235 if(cur->regmap[preferred_reg]==-1) {
1236 cur->regmap[preferred_reg]=reg;
1237 cur->dirty&=~(1<<preferred_reg);
1238 cur->isconst&=~(1<<preferred_reg);
1241 r=cur->regmap[preferred_reg];
1244 cur->regmap[preferred_reg]=reg;
1245 cur->dirty&=~(1<<preferred_reg);
1246 cur->isconst&=~(1<<preferred_reg);
1250 // Clear any unneeded registers
1251 // We try to keep the mapping consistent, if possible, because it
1252 // makes branches easier (especially loops). So we try to allocate
1253 // first (see above) before removing old mappings. If this is not
1254 // possible then go ahead and clear out the registers that are no
1256 for(hr=0;hr<HOST_REGS;hr++)
1261 if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;}
1264 // Try to allocate any available register, but prefer
1265 // registers that have not been used recently.
1267 for(hr=0;hr<HOST_REGS;hr++) {
1268 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1269 if(regs[i-1].regmap[hr]!=rs1[i-1]&®s[i-1].regmap[hr]!=rs2[i-1]&®s[i-1].regmap[hr]!=rt1[i-1]&®s[i-1].regmap[hr]!=rt2[i-1]) {
1270 cur->regmap[hr]=reg;
1271 cur->dirty&=~(1<<hr);
1272 cur->isconst&=~(1<<hr);
1278 // Try to allocate any available register
1279 for(hr=0;hr<HOST_REGS;hr++) {
1280 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1281 cur->regmap[hr]=reg;
1282 cur->dirty&=~(1<<hr);
1283 cur->isconst&=~(1<<hr);
1288 // Ok, now we have to evict someone
1289 // Pick a register we hopefully won't need soon
1290 u_char hsn[MAXREG+1];
1291 memset(hsn,10,sizeof(hsn));
1293 lsn(hsn,i,&preferred_reg);
1294 //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]);
1295 //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]);
1297 // Don't evict the cycle count at entry points, otherwise the entry
1298 // stub will have to write it.
1299 if(bt[i]&&hsn[CCREG]>2) hsn[CCREG]=2;
1300 if(i>1&&hsn[CCREG]>2&&(itype[i-2]==RJUMP||itype[i-2]==UJUMP||itype[i-2]==CJUMP||itype[i-2]==SJUMP)) hsn[CCREG]=2;
1303 // Alloc preferred register if available
1304 if(hsn[r=cur->regmap[preferred_reg]&63]==j) {
1305 for(hr=0;hr<HOST_REGS;hr++) {
1306 // Evict both parts of a 64-bit register
1307 if((cur->regmap[hr]&63)==r) {
1309 cur->dirty&=~(1<<hr);
1310 cur->isconst&=~(1<<hr);
1313 cur->regmap[preferred_reg]=reg;
1316 for(r=1;r<=MAXREG;r++)
1318 if(hsn[r]==j&&r!=rs1[i-1]&&r!=rs2[i-1]&&r!=rt1[i-1]&&r!=rt2[i-1]) {
1319 for(hr=0;hr<HOST_REGS;hr++) {
1320 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1321 if(cur->regmap[hr]==r+64) {
1322 cur->regmap[hr]=reg;
1323 cur->dirty&=~(1<<hr);
1324 cur->isconst&=~(1<<hr);
1329 for(hr=0;hr<HOST_REGS;hr++) {
1330 if(hr!=HOST_CCREG||j<hsn[CCREG]) {
1331 if(cur->regmap[hr]==r) {
1332 cur->regmap[hr]=reg;
1333 cur->dirty&=~(1<<hr);
1334 cur->isconst&=~(1<<hr);
1345 for(r=1;r<=MAXREG;r++)
1348 for(hr=0;hr<HOST_REGS;hr++) {
1349 if(cur->regmap[hr]==r+64) {
1350 cur->regmap[hr]=reg;
1351 cur->dirty&=~(1<<hr);
1352 cur->isconst&=~(1<<hr);
1356 for(hr=0;hr<HOST_REGS;hr++) {
1357 if(cur->regmap[hr]==r) {
1358 cur->regmap[hr]=reg;
1359 cur->dirty&=~(1<<hr);
1360 cur->isconst&=~(1<<hr);
1367 SysPrintf("This shouldn't happen (alloc_reg)");exit(1);
1370 // Allocate a temporary register. This is done without regard to
1371 // dirty status or whether the register we request is on the unneeded list
1372 // Note: This will only allocate one register, even if called multiple times
1373 static void alloc_reg_temp(struct regstat *cur,int i,signed char reg)
1376 int preferred_reg = -1;
1378 // see if it's already allocated
1379 for(hr=0;hr<HOST_REGS;hr++)
1381 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return;
1384 // Try to allocate any available register
1385 for(hr=HOST_REGS-1;hr>=0;hr--) {
1386 if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) {
1387 cur->regmap[hr]=reg;
1388 cur->dirty&=~(1<<hr);
1389 cur->isconst&=~(1<<hr);
1394 // Find an unneeded register
1395 for(hr=HOST_REGS-1;hr>=0;hr--)
1401 if(i==0||((unneeded_reg[i-1]>>r)&1)) {
1402 cur->regmap[hr]=reg;
1403 cur->dirty&=~(1<<hr);
1404 cur->isconst&=~(1<<hr);
1411 // Ok, now we have to evict someone
1412 // Pick a register we hopefully won't need soon
1413 // TODO: we might want to follow unconditional jumps here
1414 // TODO: get rid of dupe code and make this into a function
1415 u_char hsn[MAXREG+1];
1416 memset(hsn,10,sizeof(hsn));
1418 lsn(hsn,i,&preferred_reg);
1419 //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]);
1421 // Don't evict the cycle count at entry points, otherwise the entry
1422 // stub will have to write it.
1423 if(bt[i]&&hsn[CCREG]>2) hsn[CCREG]=2;
1424 if(i>1&&hsn[CCREG]>2&&(itype[i-2]==RJUMP||itype[i-2]==UJUMP||itype[i-2]==CJUMP||itype[i-2]==SJUMP)) hsn[CCREG]=2;
1427 for(r=1;r<=MAXREG;r++)
1429 if(hsn[r]==j&&r!=rs1[i-1]&&r!=rs2[i-1]&&r!=rt1[i-1]&&r!=rt2[i-1]) {
1430 for(hr=0;hr<HOST_REGS;hr++) {
1431 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1432 if(cur->regmap[hr]==r+64) {
1433 cur->regmap[hr]=reg;
1434 cur->dirty&=~(1<<hr);
1435 cur->isconst&=~(1<<hr);
1440 for(hr=0;hr<HOST_REGS;hr++) {
1441 if(hr!=HOST_CCREG||hsn[CCREG]>2) {
1442 if(cur->regmap[hr]==r) {
1443 cur->regmap[hr]=reg;
1444 cur->dirty&=~(1<<hr);
1445 cur->isconst&=~(1<<hr);
1456 for(r=1;r<=MAXREG;r++)
1459 for(hr=0;hr<HOST_REGS;hr++) {
1460 if(cur->regmap[hr]==r+64) {
1461 cur->regmap[hr]=reg;
1462 cur->dirty&=~(1<<hr);
1463 cur->isconst&=~(1<<hr);
1467 for(hr=0;hr<HOST_REGS;hr++) {
1468 if(cur->regmap[hr]==r) {
1469 cur->regmap[hr]=reg;
1470 cur->dirty&=~(1<<hr);
1471 cur->isconst&=~(1<<hr);
1478 SysPrintf("This shouldn't happen");exit(1);
1481 static void mov_alloc(struct regstat *current,int i)
1483 // Note: Don't need to actually alloc the source registers
1484 //alloc_reg(current,i,rs1[i]);
1485 alloc_reg(current,i,rt1[i]);
1487 clear_const(current,rs1[i]);
1488 clear_const(current,rt1[i]);
1489 dirty_reg(current,rt1[i]);
1492 static void shiftimm_alloc(struct regstat *current,int i)
1494 if(opcode2[i]<=0x3) // SLL/SRL/SRA
1497 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1499 alloc_reg(current,i,rt1[i]);
1500 dirty_reg(current,rt1[i]);
1501 if(is_const(current,rs1[i])) {
1502 int v=get_const(current,rs1[i]);
1503 if(opcode2[i]==0x00) set_const(current,rt1[i],v<<imm[i]);
1504 if(opcode2[i]==0x02) set_const(current,rt1[i],(u_int)v>>imm[i]);
1505 if(opcode2[i]==0x03) set_const(current,rt1[i],v>>imm[i]);
1507 else clear_const(current,rt1[i]);
1512 clear_const(current,rs1[i]);
1513 clear_const(current,rt1[i]);
1516 if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
1520 if(opcode2[i]==0x3c) // DSLL32
1524 if(opcode2[i]==0x3e) // DSRL32
1528 if(opcode2[i]==0x3f) // DSRA32
1534 static void shift_alloc(struct regstat *current,int i)
1537 if(opcode2[i]<=0x07) // SLLV/SRLV/SRAV
1539 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1540 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1541 alloc_reg(current,i,rt1[i]);
1542 if(rt1[i]==rs2[i]) {
1543 alloc_reg_temp(current,i,-1);
1544 minimum_free_regs[i]=1;
1546 } else { // DSLLV/DSRLV/DSRAV
1549 clear_const(current,rs1[i]);
1550 clear_const(current,rs2[i]);
1551 clear_const(current,rt1[i]);
1552 dirty_reg(current,rt1[i]);
1556 static void alu_alloc(struct regstat *current,int i)
1558 if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
1560 if(rs1[i]&&rs2[i]) {
1561 alloc_reg(current,i,rs1[i]);
1562 alloc_reg(current,i,rs2[i]);
1565 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1566 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1568 alloc_reg(current,i,rt1[i]);
1571 if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
1573 alloc_reg(current,i,rs1[i]);
1574 alloc_reg(current,i,rs2[i]);
1575 alloc_reg(current,i,rt1[i]);
1578 if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
1580 if(rs1[i]&&rs2[i]) {
1581 alloc_reg(current,i,rs1[i]);
1582 alloc_reg(current,i,rs2[i]);
1586 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1587 if(rs2[i]&&needed_again(rs2[i],i)) alloc_reg(current,i,rs2[i]);
1589 alloc_reg(current,i,rt1[i]);
1592 if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
1595 clear_const(current,rs1[i]);
1596 clear_const(current,rs2[i]);
1597 clear_const(current,rt1[i]);
1598 dirty_reg(current,rt1[i]);
1601 static void imm16_alloc(struct regstat *current,int i)
1603 if(rs1[i]&&needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1605 if(rt1[i]) alloc_reg(current,i,rt1[i]);
1606 if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
1609 else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
1610 clear_const(current,rs1[i]);
1611 clear_const(current,rt1[i]);
1613 else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
1614 if(is_const(current,rs1[i])) {
1615 int v=get_const(current,rs1[i]);
1616 if(opcode[i]==0x0c) set_const(current,rt1[i],v&imm[i]);
1617 if(opcode[i]==0x0d) set_const(current,rt1[i],v|imm[i]);
1618 if(opcode[i]==0x0e) set_const(current,rt1[i],v^imm[i]);
1620 else clear_const(current,rt1[i]);
1622 else if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
1623 if(is_const(current,rs1[i])) {
1624 int v=get_const(current,rs1[i]);
1625 set_const(current,rt1[i],v+imm[i]);
1627 else clear_const(current,rt1[i]);
1630 set_const(current,rt1[i],((long long)((short)imm[i]))<<16); // LUI
1632 dirty_reg(current,rt1[i]);
1635 static void load_alloc(struct regstat *current,int i)
1637 clear_const(current,rt1[i]);
1638 //if(rs1[i]!=rt1[i]&&needed_again(rs1[i],i)) clear_const(current,rs1[i]); // Does this help or hurt?
1639 if(!rs1[i]) current->u&=~1LL; // Allow allocating r0 if it's the source register
1640 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1641 if(rt1[i]&&!((current->u>>rt1[i])&1)) {
1642 alloc_reg(current,i,rt1[i]);
1643 assert(get_reg(current->regmap,rt1[i])>=0);
1644 if(opcode[i]==0x27||opcode[i]==0x37) // LWU/LD
1648 else if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1652 dirty_reg(current,rt1[i]);
1653 // LWL/LWR need a temporary register for the old value
1654 if(opcode[i]==0x22||opcode[i]==0x26)
1656 alloc_reg(current,i,FTEMP);
1657 alloc_reg_temp(current,i,-1);
1658 minimum_free_regs[i]=1;
1663 // Load to r0 or unneeded register (dummy load)
1664 // but we still need a register to calculate the address
1665 if(opcode[i]==0x22||opcode[i]==0x26)
1667 alloc_reg(current,i,FTEMP); // LWL/LWR need another temporary
1669 alloc_reg_temp(current,i,-1);
1670 minimum_free_regs[i]=1;
1671 if(opcode[i]==0x1A||opcode[i]==0x1B) // LDL/LDR
1678 void store_alloc(struct regstat *current,int i)
1680 clear_const(current,rs2[i]);
1681 if(!(rs2[i])) current->u&=~1LL; // Allow allocating r0 if necessary
1682 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1683 alloc_reg(current,i,rs2[i]);
1684 if(opcode[i]==0x2c||opcode[i]==0x2d||opcode[i]==0x3f) { // 64-bit SDL/SDR/SD
1687 #if defined(HOST_IMM8)
1688 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1689 else alloc_reg(current,i,INVCP);
1691 if(opcode[i]==0x2a||opcode[i]==0x2e||opcode[i]==0x2c||opcode[i]==0x2d) { // SWL/SWL/SDL/SDR
1692 alloc_reg(current,i,FTEMP);
1694 // We need a temporary register for address generation
1695 alloc_reg_temp(current,i,-1);
1696 minimum_free_regs[i]=1;
1699 void c1ls_alloc(struct regstat *current,int i)
1701 //clear_const(current,rs1[i]); // FIXME
1702 clear_const(current,rt1[i]);
1703 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1704 alloc_reg(current,i,CSREG); // Status
1705 alloc_reg(current,i,FTEMP);
1706 if(opcode[i]==0x35||opcode[i]==0x3d) { // 64-bit LDC1/SDC1
1709 #if defined(HOST_IMM8)
1710 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1711 else if((opcode[i]&0x3b)==0x39) // SWC1/SDC1
1712 alloc_reg(current,i,INVCP);
1714 // We need a temporary register for address generation
1715 alloc_reg_temp(current,i,-1);
1718 void c2ls_alloc(struct regstat *current,int i)
1720 clear_const(current,rt1[i]);
1721 if(needed_again(rs1[i],i)) alloc_reg(current,i,rs1[i]);
1722 alloc_reg(current,i,FTEMP);
1723 #if defined(HOST_IMM8)
1724 // On CPUs without 32-bit immediates we need a pointer to invalid_code
1725 if((opcode[i]&0x3b)==0x3a) // SWC2/SDC2
1726 alloc_reg(current,i,INVCP);
1728 // We need a temporary register for address generation
1729 alloc_reg_temp(current,i,-1);
1730 minimum_free_regs[i]=1;
1733 #ifndef multdiv_alloc
1734 void multdiv_alloc(struct regstat *current,int i)
1741 // case 0x1D: DMULTU
1744 clear_const(current,rs1[i]);
1745 clear_const(current,rs2[i]);
1748 if((opcode2[i]&4)==0) // 32-bit
1750 current->u&=~(1LL<<HIREG);
1751 current->u&=~(1LL<<LOREG);
1752 alloc_reg(current,i,HIREG);
1753 alloc_reg(current,i,LOREG);
1754 alloc_reg(current,i,rs1[i]);
1755 alloc_reg(current,i,rs2[i]);
1756 dirty_reg(current,HIREG);
1757 dirty_reg(current,LOREG);
1766 // Multiply by zero is zero.
1767 // MIPS does not have a divide by zero exception.
1768 // The result is undefined, we return zero.
1769 alloc_reg(current,i,HIREG);
1770 alloc_reg(current,i,LOREG);
1771 dirty_reg(current,HIREG);
1772 dirty_reg(current,LOREG);
1777 void cop0_alloc(struct regstat *current,int i)
1779 if(opcode2[i]==0) // MFC0
1782 clear_const(current,rt1[i]);
1783 alloc_all(current,i);
1784 alloc_reg(current,i,rt1[i]);
1785 dirty_reg(current,rt1[i]);
1788 else if(opcode2[i]==4) // MTC0
1791 clear_const(current,rs1[i]);
1792 alloc_reg(current,i,rs1[i]);
1793 alloc_all(current,i);
1796 alloc_all(current,i); // FIXME: Keep r0
1798 alloc_reg(current,i,0);
1803 // TLBR/TLBWI/TLBWR/TLBP/ERET
1804 assert(opcode2[i]==0x10);
1805 alloc_all(current,i);
1807 minimum_free_regs[i]=HOST_REGS;
1810 static void cop12_alloc(struct regstat *current,int i)
1812 alloc_reg(current,i,CSREG); // Load status
1813 if(opcode2[i]<3) // MFC1/CFC1
1816 clear_const(current,rt1[i]);
1817 alloc_reg(current,i,rt1[i]);
1818 dirty_reg(current,rt1[i]);
1820 alloc_reg_temp(current,i,-1);
1822 else if(opcode2[i]>3) // MTC1/CTC1
1825 clear_const(current,rs1[i]);
1826 alloc_reg(current,i,rs1[i]);
1830 alloc_reg(current,i,0);
1832 alloc_reg_temp(current,i,-1);
1834 minimum_free_regs[i]=1;
1837 void c2op_alloc(struct regstat *current,int i)
1839 alloc_reg_temp(current,i,-1);
1842 void syscall_alloc(struct regstat *current,int i)
1844 alloc_cc(current,i);
1845 dirty_reg(current,CCREG);
1846 alloc_all(current,i);
1847 minimum_free_regs[i]=HOST_REGS;
1851 void delayslot_alloc(struct regstat *current,int i)
1861 assem_debug("jump in the delay slot. this shouldn't happen.\n");//exit(1);
1862 SysPrintf("Disabled speculative precompilation\n");
1866 imm16_alloc(current,i);
1870 load_alloc(current,i);
1874 store_alloc(current,i);
1877 alu_alloc(current,i);
1880 shift_alloc(current,i);
1883 multdiv_alloc(current,i);
1886 shiftimm_alloc(current,i);
1889 mov_alloc(current,i);
1892 cop0_alloc(current,i);
1896 cop12_alloc(current,i);
1899 c1ls_alloc(current,i);
1902 c2ls_alloc(current,i);
1905 c2op_alloc(current,i);
1910 // Special case where a branch and delay slot span two pages in virtual memory
1911 static void pagespan_alloc(struct regstat *current,int i)
1914 current->wasconst=0;
1916 minimum_free_regs[i]=HOST_REGS;
1917 alloc_all(current,i);
1918 alloc_cc(current,i);
1919 dirty_reg(current,CCREG);
1920 if(opcode[i]==3) // JAL
1922 alloc_reg(current,i,31);
1923 dirty_reg(current,31);
1925 if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
1927 alloc_reg(current,i,rs1[i]);
1929 alloc_reg(current,i,rt1[i]);
1930 dirty_reg(current,rt1[i]);
1933 if((opcode[i]&0x2E)==4) // BEQ/BNE/BEQL/BNEL
1935 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1936 if(rs2[i]) alloc_reg(current,i,rs2[i]);
1939 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ/BLEZL/BGTZL
1941 if(rs1[i]) alloc_reg(current,i,rs1[i]);
1946 static void add_stub(enum stub_type type, void *addr, void *retaddr,
1947 u_int a, uintptr_t b, uintptr_t c, u_int d, u_int e)
1949 assert(a < ARRAY_SIZE(stubs));
1950 stubs[stubcount].type = type;
1951 stubs[stubcount].addr = addr;
1952 stubs[stubcount].retaddr = retaddr;
1953 stubs[stubcount].a = a;
1954 stubs[stubcount].b = b;
1955 stubs[stubcount].c = c;
1956 stubs[stubcount].d = d;
1957 stubs[stubcount].e = e;
1961 static void add_stub_r(enum stub_type type, void *addr, void *retaddr,
1962 int i, int addr_reg, struct regstat *i_regs, int ccadj, u_int reglist)
1964 add_stub(type, addr, retaddr, i, addr_reg, (uintptr_t)i_regs, ccadj, reglist);
1967 // Write out a single register
1968 static void wb_register(signed char r,signed char regmap[],uint64_t dirty)
1971 for(hr=0;hr<HOST_REGS;hr++) {
1972 if(hr!=EXCLUDE_REG) {
1973 if((regmap[hr]&63)==r) {
1975 assert(regmap[hr]<64);
1976 emit_storereg(r,hr);
1983 static void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t u)
1985 //if(dirty_pre==dirty) return;
1987 for(hr=0;hr<HOST_REGS;hr++) {
1988 if(hr!=EXCLUDE_REG) {
1990 if(((~u)>>(reg&63))&1) {
1992 if(((dirty_pre&~dirty)>>hr)&1) {
1994 emit_storereg(reg,hr);
2011 printf("r%d:%8x%8x ",i,((int *)(reg+i))[1],((int *)(reg+i))[0]);
2016 static void pass_args(int a0, int a1)
2020 emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0);
2022 else if(a0!=0&&a1==0) {
2024 if (a0>=0) emit_mov(a0,0);
2027 if(a0>=0&&a0!=0) emit_mov(a0,0);
2028 if(a1>=0&&a1!=1) emit_mov(a1,1);
2032 static void alu_assemble(int i,struct regstat *i_regs)
2034 if(opcode2[i]>=0x20&&opcode2[i]<=0x23) { // ADD/ADDU/SUB/SUBU
2036 signed char s1,s2,t;
2037 t=get_reg(i_regs->regmap,rt1[i]);
2039 s1=get_reg(i_regs->regmap,rs1[i]);
2040 s2=get_reg(i_regs->regmap,rs2[i]);
2041 if(rs1[i]&&rs2[i]) {
2044 if(opcode2[i]&2) emit_sub(s1,s2,t);
2045 else emit_add(s1,s2,t);
2048 if(s1>=0) emit_mov(s1,t);
2049 else emit_loadreg(rs1[i],t);
2053 if(opcode2[i]&2) emit_neg(s2,t);
2054 else emit_mov(s2,t);
2057 emit_loadreg(rs2[i],t);
2058 if(opcode2[i]&2) emit_neg(t,t);
2061 else emit_zeroreg(t);
2065 if(opcode2[i]>=0x2c&&opcode2[i]<=0x2f) { // DADD/DADDU/DSUB/DSUBU
2068 if(opcode2[i]==0x2a||opcode2[i]==0x2b) { // SLT/SLTU
2070 signed char s1l,s2l,t;
2072 t=get_reg(i_regs->regmap,rt1[i]);
2075 s1l=get_reg(i_regs->regmap,rs1[i]);
2076 s2l=get_reg(i_regs->regmap,rs2[i]);
2077 if(rs2[i]==0) // rx<r0
2080 if(opcode2[i]==0x2a) // SLT
2081 emit_shrimm(s1l,31,t);
2082 else // SLTU (unsigned can not be less than zero)
2085 else if(rs1[i]==0) // r0<rx
2088 if(opcode2[i]==0x2a) // SLT
2089 emit_set_gz32(s2l,t);
2090 else // SLTU (set if not zero)
2091 emit_set_nz32(s2l,t);
2094 assert(s1l>=0);assert(s2l>=0);
2095 if(opcode2[i]==0x2a) // SLT
2096 emit_set_if_less32(s1l,s2l,t);
2098 emit_set_if_carry32(s1l,s2l,t);
2104 if(opcode2[i]>=0x24&&opcode2[i]<=0x27) { // AND/OR/XOR/NOR
2106 signed char s1l,s2l,tl;
2107 tl=get_reg(i_regs->regmap,rt1[i]);
2110 s1l=get_reg(i_regs->regmap,rs1[i]);
2111 s2l=get_reg(i_regs->regmap,rs2[i]);
2112 if(rs1[i]&&rs2[i]) {
2115 if(opcode2[i]==0x24) { // AND
2116 emit_and(s1l,s2l,tl);
2118 if(opcode2[i]==0x25) { // OR
2119 emit_or(s1l,s2l,tl);
2121 if(opcode2[i]==0x26) { // XOR
2122 emit_xor(s1l,s2l,tl);
2124 if(opcode2[i]==0x27) { // NOR
2125 emit_or(s1l,s2l,tl);
2131 if(opcode2[i]==0x24) { // AND
2134 if(opcode2[i]==0x25||opcode2[i]==0x26) { // OR/XOR
2136 if(s1l>=0) emit_mov(s1l,tl);
2137 else emit_loadreg(rs1[i],tl); // CHECK: regmap_entry?
2141 if(s2l>=0) emit_mov(s2l,tl);
2142 else emit_loadreg(rs2[i],tl); // CHECK: regmap_entry?
2144 else emit_zeroreg(tl);
2146 if(opcode2[i]==0x27) { // NOR
2148 if(s1l>=0) emit_not(s1l,tl);
2150 emit_loadreg(rs1[i],tl);
2156 if(s2l>=0) emit_not(s2l,tl);
2158 emit_loadreg(rs2[i],tl);
2162 else emit_movimm(-1,tl);
2171 void imm16_assemble(int i,struct regstat *i_regs)
2173 if (opcode[i]==0x0f) { // LUI
2176 t=get_reg(i_regs->regmap,rt1[i]);
2179 if(!((i_regs->isconst>>t)&1))
2180 emit_movimm(imm[i]<<16,t);
2184 if(opcode[i]==0x08||opcode[i]==0x09) { // ADDI/ADDIU
2187 t=get_reg(i_regs->regmap,rt1[i]);
2188 s=get_reg(i_regs->regmap,rs1[i]);
2193 if(!((i_regs->isconst>>t)&1)) {
2195 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2196 emit_addimm(t,imm[i],t);
2198 if(!((i_regs->wasconst>>s)&1))
2199 emit_addimm(s,imm[i],t);
2201 emit_movimm(constmap[i][s]+imm[i],t);
2207 if(!((i_regs->isconst>>t)&1))
2208 emit_movimm(imm[i],t);
2213 if(opcode[i]==0x18||opcode[i]==0x19) { // DADDI/DADDIU
2215 signed char sh,sl,th,tl;
2216 th=get_reg(i_regs->regmap,rt1[i]|64);
2217 tl=get_reg(i_regs->regmap,rt1[i]);
2218 sh=get_reg(i_regs->regmap,rs1[i]|64);
2219 sl=get_reg(i_regs->regmap,rs1[i]);
2225 emit_addimm64_32(sh,sl,imm[i],th,tl);
2228 emit_addimm(sl,imm[i],tl);
2231 emit_movimm(imm[i],tl);
2232 if(th>=0) emit_movimm(((signed int)imm[i])>>31,th);
2237 else if(opcode[i]==0x0a||opcode[i]==0x0b) { // SLTI/SLTIU
2239 //assert(rs1[i]!=0); // r0 might be valid, but it's probably a bug
2241 t=get_reg(i_regs->regmap,rt1[i]);
2242 sl=get_reg(i_regs->regmap,rs1[i]);
2246 if(opcode[i]==0x0a) { // SLTI
2248 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2249 emit_slti32(t,imm[i],t);
2251 emit_slti32(sl,imm[i],t);
2256 if(i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2257 emit_sltiu32(t,imm[i],t);
2259 emit_sltiu32(sl,imm[i],t);
2263 // SLTI(U) with r0 is just stupid,
2264 // nonetheless examples can be found
2265 if(opcode[i]==0x0a) // SLTI
2266 if(0<imm[i]) emit_movimm(1,t);
2267 else emit_zeroreg(t);
2270 if(imm[i]) emit_movimm(1,t);
2271 else emit_zeroreg(t);
2277 else if(opcode[i]>=0x0c&&opcode[i]<=0x0e) { // ANDI/ORI/XORI
2279 signed char sh,sl,th,tl;
2280 th=get_reg(i_regs->regmap,rt1[i]|64);
2281 tl=get_reg(i_regs->regmap,rt1[i]);
2282 sh=get_reg(i_regs->regmap,rs1[i]|64);
2283 sl=get_reg(i_regs->regmap,rs1[i]);
2284 if(tl>=0 && !((i_regs->isconst>>tl)&1)) {
2285 if(opcode[i]==0x0c) //ANDI
2289 if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2290 emit_andimm(tl,imm[i],tl);
2292 if(!((i_regs->wasconst>>sl)&1))
2293 emit_andimm(sl,imm[i],tl);
2295 emit_movimm(constmap[i][sl]&imm[i],tl);
2300 if(th>=0) emit_zeroreg(th);
2306 if(i_regs->regmap_entry[tl]!=rs1[i]) emit_loadreg(rs1[i],tl);
2310 emit_loadreg(rs1[i]|64,th);
2315 if(opcode[i]==0x0d) { // ORI
2317 emit_orimm(tl,imm[i],tl);
2319 if(!((i_regs->wasconst>>sl)&1))
2320 emit_orimm(sl,imm[i],tl);
2322 emit_movimm(constmap[i][sl]|imm[i],tl);
2325 if(opcode[i]==0x0e) { // XORI
2327 emit_xorimm(tl,imm[i],tl);
2329 if(!((i_regs->wasconst>>sl)&1))
2330 emit_xorimm(sl,imm[i],tl);
2332 emit_movimm(constmap[i][sl]^imm[i],tl);
2337 emit_movimm(imm[i],tl);
2338 if(th>=0) emit_zeroreg(th);
2346 void shiftimm_assemble(int i,struct regstat *i_regs)
2348 if(opcode2[i]<=0x3) // SLL/SRL/SRA
2352 t=get_reg(i_regs->regmap,rt1[i]);
2353 s=get_reg(i_regs->regmap,rs1[i]);
2355 if(t>=0&&!((i_regs->isconst>>t)&1)){
2362 if(s<0&&i_regs->regmap_entry[t]!=rs1[i]) emit_loadreg(rs1[i],t);
2364 if(opcode2[i]==0) // SLL
2366 emit_shlimm(s<0?t:s,imm[i],t);
2368 if(opcode2[i]==2) // SRL
2370 emit_shrimm(s<0?t:s,imm[i],t);
2372 if(opcode2[i]==3) // SRA
2374 emit_sarimm(s<0?t:s,imm[i],t);
2378 if(s>=0 && s!=t) emit_mov(s,t);
2382 //emit_storereg(rt1[i],t); //DEBUG
2385 if(opcode2[i]>=0x38&&opcode2[i]<=0x3b) // DSLL/DSRL/DSRA
2389 if(opcode2[i]==0x3c) // DSLL32
2393 if(opcode2[i]==0x3e) // DSRL32
2397 if(opcode2[i]==0x3f) // DSRA32
2403 #ifndef shift_assemble
2404 void shift_assemble(int i,struct regstat *i_regs)
2406 printf("Need shift_assemble for this architecture.\n");
2419 static int get_ptr_mem_type(u_int a)
2421 if(a < 0x00200000) {
2422 if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0))
2423 // return wrong, must use memhandler for BIOS self-test to pass
2424 // 007 does similar stuff from a00 mirror, weird stuff
2428 if(0x1f800000 <= a && a < 0x1f801000)
2430 if(0x80200000 <= a && a < 0x80800000)
2432 if(0xa0000000 <= a && a < 0xa0200000)
2437 static void *emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override)
2442 if(((smrv_strong|smrv_weak)>>mr)&1) {
2443 type=get_ptr_mem_type(smrv[mr]);
2444 //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type);
2447 // use the mirror we are running on
2448 type=get_ptr_mem_type(start);
2449 //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type);
2452 if(type==MTYPE_8020) { // RAM 80200000+ mirror
2453 emit_andimm(addr,~0x00e00000,HOST_TEMPREG);
2454 addr=*addr_reg_override=HOST_TEMPREG;
2457 else if(type==MTYPE_0000) { // RAM 0 mirror
2458 emit_orimm(addr,0x80000000,HOST_TEMPREG);
2459 addr=*addr_reg_override=HOST_TEMPREG;
2462 else if(type==MTYPE_A000) { // RAM A mirror
2463 emit_andimm(addr,~0x20000000,HOST_TEMPREG);
2464 addr=*addr_reg_override=HOST_TEMPREG;
2467 else if(type==MTYPE_1F80) { // scratchpad
2468 if (psxH == (void *)0x1f800000) {
2469 emit_addimm(addr,-0x1f800000,HOST_TEMPREG);
2470 emit_cmpimm(HOST_TEMPREG,0x1000);
2475 // do the usual RAM check, jump will go to the right handler
2482 emit_cmpimm(addr,RAM_SIZE);
2484 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
2485 // Hint to branch predictor that the branch is unlikely to be taken
2487 emit_jno_unlikely(0);
2492 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2493 addr=*addr_reg_override=HOST_TEMPREG;
2500 // return memhandler, or get directly accessable address and return 0
2501 static void *get_direct_memhandler(void *table, u_int addr,
2502 enum stub_type type, uintptr_t *addr_host)
2504 uintptr_t l1, l2 = 0;
2505 l1 = ((uintptr_t *)table)[addr>>12];
2506 if ((l1 & (1ul << (sizeof(l1)*8-1))) == 0) {
2507 uintptr_t v = l1 << 1;
2508 *addr_host = v + addr;
2513 if (type == LOADB_STUB || type == LOADBU_STUB || type == STOREB_STUB)
2514 l2 = ((uintptr_t *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)];
2515 else if (type == LOADH_STUB || type == LOADHU_STUB || type == STOREH_STUB)
2516 l2=((uintptr_t *)l1)[0x1000/4 + (addr&0xfff)/2];
2518 l2=((uintptr_t *)l1)[(addr&0xfff)/4];
2519 if ((l2 & (1<<31)) == 0) {
2520 uintptr_t v = l2 << 1;
2521 *addr_host = v + (addr&0xfff);
2524 return (void *)(l2 << 1);
2528 static void load_assemble(int i,struct regstat *i_regs)
2533 int memtarget=0,c=0;
2534 int fastload_reg_override=0;
2536 th=get_reg(i_regs->regmap,rt1[i]|64);
2537 tl=get_reg(i_regs->regmap,rt1[i]);
2538 s=get_reg(i_regs->regmap,rs1[i]);
2540 for(hr=0;hr<HOST_REGS;hr++) {
2541 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2543 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2545 c=(i_regs->wasconst>>s)&1;
2547 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2550 //printf("load_assemble: c=%d\n",c);
2551 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2552 // FIXME: Even if the load is a NOP, we should check for pagefaults...
2553 if((tl<0&&(!c||(((u_int)constmap[i][s]+offset)>>16)==0x1f80))
2555 // could be FIFO, must perform the read
2557 assem_debug("(forced read)\n");
2558 tl=get_reg(i_regs->regmap,-1);
2561 if(offset||s<0||c) addr=tl;
2563 //if(tl<0) tl=get_reg(i_regs->regmap,-1);
2565 //printf("load_assemble: c=%d\n",c);
2566 //if(c) printf("load_assemble: const=%lx\n",(long)constmap[i][s]+offset);
2567 assert(tl>=0); // Even if the load is a NOP, we must check for pagefaults and I/O
2569 if(th>=0) reglist&=~(1<<th);
2572 // Strmnnrmn's speed hack
2573 if(rs1[i]!=29||start<0x80001000||start>=0x80000000+RAM_SIZE)
2576 jaddr=emit_fastpath_cmp_jump(i,addr,&fastload_reg_override);
2579 else if(ram_offset&&memtarget) {
2580 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2581 fastload_reg_override=HOST_TEMPREG;
2583 int dummy=(rt1[i]==0)||(tl!=get_reg(i_regs->regmap,rt1[i])); // ignore loads to r0 and unneeded reg
2584 if (opcode[i]==0x20) { // LB
2590 if(fastload_reg_override) a=fastload_reg_override;
2592 emit_movsbl_indexed(x,a,tl);
2596 add_stub_r(LOADB_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2599 inline_readstub(LOADB_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2601 if (opcode[i]==0x21) { // LH
2606 if(fastload_reg_override) a=fastload_reg_override;
2607 emit_movswl_indexed(x,a,tl);
2610 add_stub_r(LOADH_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2613 inline_readstub(LOADH_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2615 if (opcode[i]==0x23) { // LW
2619 if(fastload_reg_override) a=fastload_reg_override;
2620 emit_readword_indexed(0,a,tl);
2623 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2626 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2628 if (opcode[i]==0x24) { // LBU
2633 if(fastload_reg_override) a=fastload_reg_override;
2635 emit_movzbl_indexed(x,a,tl);
2638 add_stub_r(LOADBU_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2641 inline_readstub(LOADBU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2643 if (opcode[i]==0x25) { // LHU
2648 if(fastload_reg_override) a=fastload_reg_override;
2649 emit_movzwl_indexed(x,a,tl);
2652 add_stub_r(LOADHU_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2655 inline_readstub(LOADHU_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2657 if (opcode[i]==0x27) { // LWU
2662 if(fastload_reg_override) a=fastload_reg_override;
2663 emit_readword_indexed(0,a,tl);
2666 add_stub_r(LOADW_STUB,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2669 inline_readstub(LOADW_STUB,i,constmap[i][s]+offset,i_regs->regmap,rt1[i],ccadj[i],reglist);
2673 if (opcode[i]==0x37) { // LD
2679 #ifndef loadlr_assemble
2680 void loadlr_assemble(int i,struct regstat *i_regs)
2682 printf("Need loadlr_assemble for this architecture.\n");
2687 void store_assemble(int i,struct regstat *i_regs)
2693 enum stub_type type;
2694 int memtarget=0,c=0;
2695 int agr=AGEN1+(i&1);
2696 int faststore_reg_override=0;
2698 tl=get_reg(i_regs->regmap,rs2[i]);
2699 s=get_reg(i_regs->regmap,rs1[i]);
2700 temp=get_reg(i_regs->regmap,agr);
2701 if(temp<0) temp=get_reg(i_regs->regmap,-1);
2704 c=(i_regs->wasconst>>s)&1;
2706 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2711 for(hr=0;hr<HOST_REGS;hr++) {
2712 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2714 if(i_regs->regmap[HOST_CCREG]==CCREG) reglist&=~(1<<HOST_CCREG);
2715 if(offset||s<0||c) addr=temp;
2718 jaddr=emit_fastpath_cmp_jump(i,addr,&faststore_reg_override);
2720 else if(ram_offset&&memtarget) {
2721 emit_addimm(addr,ram_offset,HOST_TEMPREG);
2722 faststore_reg_override=HOST_TEMPREG;
2725 if (opcode[i]==0x28) { // SB
2729 if(faststore_reg_override) a=faststore_reg_override;
2730 emit_writebyte_indexed(tl,x,a);
2734 if (opcode[i]==0x29) { // SH
2738 if(faststore_reg_override) a=faststore_reg_override;
2739 emit_writehword_indexed(tl,x,a);
2743 if (opcode[i]==0x2B) { // SW
2746 if(faststore_reg_override) a=faststore_reg_override;
2747 emit_writeword_indexed(tl,0,a);
2751 if (opcode[i]==0x3F) { // SD
2756 // PCSX store handlers don't check invcode again
2758 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2761 if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
2763 #ifdef DESTRUCTIVE_SHIFT
2764 // The x86 shift operation is 'destructive'; it overwrites the
2765 // source register, so we need to make a copy first and use that.
2768 #if defined(HOST_IMM8)
2769 int ir=get_reg(i_regs->regmap,INVCP);
2771 emit_cmpmem_indexedsr12_reg(ir,addr,1);
2773 emit_cmpmem_indexedsr12_imm(invalid_code,addr,1);
2775 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
2776 emit_callne(invalidate_addr_reg[addr]);
2780 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),addr,0,0,0);
2784 u_int addr_val=constmap[i][s]+offset;
2786 add_stub_r(type,jaddr,out,i,addr,i_regs,ccadj[i],reglist);
2787 } else if(c&&!memtarget) {
2788 inline_writestub(type,i,addr_val,i_regs->regmap,rs2[i],ccadj[i],reglist);
2790 // basic current block modification detection..
2791 // not looking back as that should be in mips cache already
2792 if(c&&start+i*4<addr_val&&addr_val<start+slen*4) {
2793 SysPrintf("write to %08x hits block %08x, pc=%08x\n",addr_val,start,start+i*4);
2794 assert(i_regs->regmap==regs[i].regmap); // not delay slot
2795 if(i_regs->regmap==regs[i].regmap) {
2796 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
2797 wb_dirtys(regs[i].regmap_entry,regs[i].wasdirty);
2798 emit_movimm(start+i*4+4,0);
2799 emit_writeword(0,&pcaddr);
2800 emit_jmp(do_interrupt);
2805 void storelr_assemble(int i,struct regstat *i_regs)
2811 void *case1, *case2, *case3;
2812 void *done0, *done1, *done2;
2813 int memtarget=0,c=0;
2814 int agr=AGEN1+(i&1);
2816 tl=get_reg(i_regs->regmap,rs2[i]);
2817 s=get_reg(i_regs->regmap,rs1[i]);
2818 temp=get_reg(i_regs->regmap,agr);
2819 if(temp<0) temp=get_reg(i_regs->regmap,-1);
2822 c=(i_regs->isconst>>s)&1;
2824 memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE;
2828 for(hr=0;hr<HOST_REGS;hr++) {
2829 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
2833 emit_cmpimm(s<0||offset?temp:s,RAM_SIZE);
2834 if(!offset&&s!=temp) emit_mov(s,temp);
2840 if(!memtarget||!rs1[i]) {
2845 emit_addimm_no_flags(ram_offset,temp);
2847 if (opcode[i]==0x2C||opcode[i]==0x2D) { // SDL/SDR
2851 emit_xorimm(temp,3,temp);
2852 emit_testimm(temp,2);
2855 emit_testimm(temp,1);
2859 if (opcode[i]==0x2A) { // SWL
2860 emit_writeword_indexed(tl,0,temp);
2862 if (opcode[i]==0x2E) { // SWR
2863 emit_writebyte_indexed(tl,3,temp);
2865 if (opcode[i]==0x2C) { // SDL
2868 if (opcode[i]==0x2D) { // SDR
2874 set_jump_target(case1, out);
2875 if (opcode[i]==0x2A) { // SWL
2876 // Write 3 msb into three least significant bytes
2877 if(rs2[i]) emit_rorimm(tl,8,tl);
2878 emit_writehword_indexed(tl,-1,temp);
2879 if(rs2[i]) emit_rorimm(tl,16,tl);
2880 emit_writebyte_indexed(tl,1,temp);
2881 if(rs2[i]) emit_rorimm(tl,8,tl);
2883 if (opcode[i]==0x2E) { // SWR
2884 // Write two lsb into two most significant bytes
2885 emit_writehword_indexed(tl,1,temp);
2887 if (opcode[i]==0x2C) { // SDL
2890 if (opcode[i]==0x2D) { // SDR
2896 set_jump_target(case2, out);
2897 emit_testimm(temp,1);
2900 if (opcode[i]==0x2A) { // SWL
2901 // Write two msb into two least significant bytes
2902 if(rs2[i]) emit_rorimm(tl,16,tl);
2903 emit_writehword_indexed(tl,-2,temp);
2904 if(rs2[i]) emit_rorimm(tl,16,tl);
2906 if (opcode[i]==0x2E) { // SWR
2907 // Write 3 lsb into three most significant bytes
2908 emit_writebyte_indexed(tl,-1,temp);
2909 if(rs2[i]) emit_rorimm(tl,8,tl);
2910 emit_writehword_indexed(tl,0,temp);
2911 if(rs2[i]) emit_rorimm(tl,24,tl);
2913 if (opcode[i]==0x2C) { // SDL
2916 if (opcode[i]==0x2D) { // SDR
2922 set_jump_target(case3, out);
2923 if (opcode[i]==0x2A) { // SWL
2924 // Write msb into least significant byte
2925 if(rs2[i]) emit_rorimm(tl,24,tl);
2926 emit_writebyte_indexed(tl,-3,temp);
2927 if(rs2[i]) emit_rorimm(tl,8,tl);
2929 if (opcode[i]==0x2E) { // SWR
2930 // Write entire word
2931 emit_writeword_indexed(tl,-3,temp);
2933 if (opcode[i]==0x2C) { // SDL
2936 if (opcode[i]==0x2D) { // SDR
2939 set_jump_target(done0, out);
2940 set_jump_target(done1, out);
2941 set_jump_target(done2, out);
2942 if (opcode[i]==0x2C) { // SDL
2945 if (opcode[i]==0x2D) { // SDR
2949 add_stub_r(STORELR_STUB,jaddr,out,i,temp,i_regs,ccadj[i],reglist);
2950 if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
2951 emit_addimm_no_flags(-ram_offset,temp);
2952 #if defined(HOST_IMM8)
2953 int ir=get_reg(i_regs->regmap,INVCP);
2955 emit_cmpmem_indexedsr12_reg(ir,temp,1);
2957 emit_cmpmem_indexedsr12_imm(invalid_code,temp,1);
2959 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
2960 emit_callne(invalidate_addr_reg[temp]);
2964 add_stub(INVCODE_STUB,jaddr2,out,reglist|(1<<HOST_CCREG),temp,0,0,0);
2969 static void cop0_assemble(int i,struct regstat *i_regs)
2971 if(opcode2[i]==0) // MFC0
2973 signed char t=get_reg(i_regs->regmap,rt1[i]);
2974 u_int copr=(source[i]>>11)&0x1f;
2975 //assert(t>=0); // Why does this happen? OOT is weird
2976 if(t>=0&&rt1[i]!=0) {
2977 emit_readword(®_cop0[copr],t);
2980 else if(opcode2[i]==4) // MTC0
2982 signed char s=get_reg(i_regs->regmap,rs1[i]);
2983 char copr=(source[i]>>11)&0x1f;
2985 wb_register(rs1[i],i_regs->regmap,i_regs->dirty);
2986 if(copr==9||copr==11||copr==12||copr==13) {
2987 emit_readword(&last_count,HOST_TEMPREG);
2988 emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc
2989 emit_add(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
2990 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
2991 emit_writeword(HOST_CCREG,&Count);
2993 // What a mess. The status register (12) can enable interrupts,
2994 // so needs a special case to handle a pending interrupt.
2995 // The interrupt must be taken immediately, because a subsequent
2996 // instruction might disable interrupts again.
2997 if(copr==12||copr==13) {
2999 // burn cycles to cause cc_interrupt, which will
3000 // reschedule next_interupt. Relies on CCREG from above.
3001 assem_debug("MTC0 DS %d\n", copr);
3002 emit_writeword(HOST_CCREG,&last_count);
3003 emit_movimm(0,HOST_CCREG);
3004 emit_storereg(CCREG,HOST_CCREG);
3005 emit_loadreg(rs1[i],1);
3006 emit_movimm(copr,0);
3007 emit_call(pcsx_mtc0_ds);
3008 emit_loadreg(rs1[i],s);
3011 emit_movimm(start+i*4+4,HOST_TEMPREG);
3012 emit_writeword(HOST_TEMPREG,&pcaddr);
3013 emit_movimm(0,HOST_TEMPREG);
3014 emit_writeword(HOST_TEMPREG,&pending_exception);
3016 //else if(copr==12&&is_delayslot) emit_call((int)MTC0_R12);
3019 emit_loadreg(rs1[i],1);
3022 emit_movimm(copr,0);
3023 emit_call(pcsx_mtc0);
3024 if(copr==9||copr==11||copr==12||copr==13) {
3025 emit_readword(&Count,HOST_CCREG);
3026 emit_readword(&next_interupt,HOST_TEMPREG);
3027 emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3028 emit_sub(HOST_CCREG,HOST_TEMPREG,HOST_CCREG);
3029 emit_writeword(HOST_TEMPREG,&last_count);
3030 emit_storereg(CCREG,HOST_CCREG);
3032 if(copr==12||copr==13) {
3033 assert(!is_delayslot);
3034 emit_readword(&pending_exception,14);
3036 emit_jne(&do_interrupt);
3038 emit_loadreg(rs1[i],s);
3039 if(get_reg(i_regs->regmap,rs1[i]|64)>=0)
3040 emit_loadreg(rs1[i]|64,get_reg(i_regs->regmap,rs1[i]|64));
3044 assert(opcode2[i]==0x10);
3045 //if((source[i]&0x3f)==0x10) // RFE
3047 emit_readword(&Status,0);
3048 emit_andimm(0,0x3c,1);
3049 emit_andimm(0,~0xf,0);
3050 emit_orrshr_imm(1,2,0);
3051 emit_writeword(0,&Status);
3056 static void cop1_unusable(int i,struct regstat *i_regs)
3058 // XXX: should just just do the exception instead
3063 add_stub_r(FP_STUB,jaddr,out,i,0,i_regs,is_delayslot,0);
3067 static void cop1_assemble(int i,struct regstat *i_regs)
3069 cop1_unusable(i, i_regs);
3072 static void c1ls_assemble(int i,struct regstat *i_regs)
3074 cop1_unusable(i, i_regs);
3078 static void do_cop1stub(int n)
3081 assem_debug("do_cop1stub %x\n",start+stubs[n].a*4);
3082 set_jump_target(stubs[n].addr, out);
3084 // int rs=stubs[n].b;
3085 struct regstat *i_regs=(struct regstat *)stubs[n].c;
3088 load_all_consts(regs[i].regmap_entry,regs[i].wasdirty,i);
3089 //if(i_regs!=®s[i]) printf("oops: regs[i]=%x i_regs=%x",(int)®s[i],(int)i_regs);
3091 //else {printf("fp exception in delay slot\n");}
3092 wb_dirtys(i_regs->regmap_entry,i_regs->wasdirty);
3093 if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
3094 emit_movimm(start+(i-ds)*4,EAX); // Get PC
3095 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3096 emit_jmp(ds?fp_exception_ds:fp_exception);
3099 static void cop2_get_dreg(u_int copr,signed char tl,signed char temp)
3109 emit_readword(®_cop2d[copr],tl);
3110 emit_signextend16(tl,tl);
3111 emit_writeword(tl,®_cop2d[copr]); // hmh
3118 emit_readword(®_cop2d[copr],tl);
3119 emit_andimm(tl,0xffff,tl);
3120 emit_writeword(tl,®_cop2d[copr]);
3123 emit_readword(®_cop2d[14],tl); // SXY2
3124 emit_writeword(tl,®_cop2d[copr]);
3128 emit_readword(®_cop2d[9],temp);
3129 emit_testimm(temp,0x8000); // do we need this?
3130 emit_andimm(temp,0xf80,temp);
3131 emit_andne_imm(temp,0,temp);
3132 emit_shrimm(temp,7,tl);
3133 emit_readword(®_cop2d[10],temp);
3134 emit_testimm(temp,0x8000);
3135 emit_andimm(temp,0xf80,temp);
3136 emit_andne_imm(temp,0,temp);
3137 emit_orrshr_imm(temp,2,tl);
3138 emit_readword(®_cop2d[11],temp);
3139 emit_testimm(temp,0x8000);
3140 emit_andimm(temp,0xf80,temp);
3141 emit_andne_imm(temp,0,temp);
3142 emit_orrshl_imm(temp,3,tl);
3143 emit_writeword(tl,®_cop2d[copr]);
3146 emit_readword(®_cop2d[copr],tl);
3151 static void cop2_put_dreg(u_int copr,signed char sl,signed char temp)
3155 emit_readword(®_cop2d[13],temp); // SXY1
3156 emit_writeword(sl,®_cop2d[copr]);
3157 emit_writeword(temp,®_cop2d[12]); // SXY0
3158 emit_readword(®_cop2d[14],temp); // SXY2
3159 emit_writeword(sl,®_cop2d[14]);
3160 emit_writeword(temp,®_cop2d[13]); // SXY1
3163 emit_andimm(sl,0x001f,temp);
3164 emit_shlimm(temp,7,temp);
3165 emit_writeword(temp,®_cop2d[9]);
3166 emit_andimm(sl,0x03e0,temp);
3167 emit_shlimm(temp,2,temp);
3168 emit_writeword(temp,®_cop2d[10]);
3169 emit_andimm(sl,0x7c00,temp);
3170 emit_shrimm(temp,3,temp);
3171 emit_writeword(temp,®_cop2d[11]);
3172 emit_writeword(sl,®_cop2d[28]);
3176 emit_mvnmi(temp,temp);
3177 #if defined(HAVE_ARMV5) || defined(__aarch64__)
3178 emit_clz(temp,temp);
3180 emit_movs(temp,HOST_TEMPREG);
3181 emit_movimm(0,temp);
3182 emit_jeq((int)out+4*4);
3183 emit_addpl_imm(temp,1,temp);
3184 emit_lslpls_imm(HOST_TEMPREG,1,HOST_TEMPREG);
3185 emit_jns((int)out-2*4);
3187 emit_writeword(sl,®_cop2d[30]);
3188 emit_writeword(temp,®_cop2d[31]);
3193 emit_writeword(sl,®_cop2d[copr]);
3198 static void c2ls_assemble(int i,struct regstat *i_regs)
3203 int memtarget=0,c=0;
3205 enum stub_type type;
3206 int agr=AGEN1+(i&1);
3207 int fastio_reg_override=0;
3209 u_int copr=(source[i]>>16)&0x1f;
3210 s=get_reg(i_regs->regmap,rs1[i]);
3211 tl=get_reg(i_regs->regmap,FTEMP);
3216 for(hr=0;hr<HOST_REGS;hr++) {
3217 if(i_regs->regmap[hr]>=0) reglist|=1<<hr;
3219 if(i_regs->regmap[HOST_CCREG]==CCREG)
3220 reglist&=~(1<<HOST_CCREG);
3223 if (opcode[i]==0x3a) { // SWC2
3224 ar=get_reg(i_regs->regmap,agr);
3225 if(ar<0) ar=get_reg(i_regs->regmap,-1);
3230 if(s>=0) c=(i_regs->wasconst>>s)&1;
3231 memtarget=c&&(((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE);
3232 if (!offset&&!c&&s>=0) ar=s;
3235 if (opcode[i]==0x3a) { // SWC2
3236 cop2_get_dreg(copr,tl,HOST_TEMPREG);
3244 emit_jmp(0); // inline_readstub/inline_writestub?
3248 jaddr2=emit_fastpath_cmp_jump(i,ar,&fastio_reg_override);
3250 else if(ram_offset&&memtarget) {
3251 emit_addimm(ar,ram_offset,HOST_TEMPREG);
3252 fastio_reg_override=HOST_TEMPREG;
3254 if (opcode[i]==0x32) { // LWC2
3256 if(fastio_reg_override) a=fastio_reg_override;
3257 emit_readword_indexed(0,a,tl);
3259 if (opcode[i]==0x3a) { // SWC2
3260 #ifdef DESTRUCTIVE_SHIFT
3261 if(!offset&&!c&&s>=0) emit_mov(s,ar);
3264 if(fastio_reg_override) a=fastio_reg_override;
3265 emit_writeword_indexed(tl,0,a);
3269 add_stub_r(type,jaddr2,out,i,ar,i_regs,ccadj[i],reglist);
3270 if(opcode[i]==0x3a) // SWC2
3271 if(!(i_regs->waswritten&(1<<rs1[i]))&&!(new_dynarec_hacks&NDHACK_NO_SMC_CHECK)) {
3272 #if defined(HOST_IMM8)
3273 int ir=get_reg(i_regs->regmap,INVCP);
3275 emit_cmpmem_indexedsr12_reg(ir,ar,1);
3277 emit_cmpmem_indexedsr12_imm(invalid_code,ar,1);
3279 #if defined(HAVE_CONDITIONAL_CALL) && !defined(DESTRUCTIVE_SHIFT)
3280 emit_callne(invalidate_addr_reg[ar]);
3284 add_stub(INVCODE_STUB,jaddr3,out,reglist|(1<<HOST_CCREG),ar,0,0,0);
3287 if (opcode[i]==0x32) { // LWC2
3288 cop2_put_dreg(copr,tl,HOST_TEMPREG);
3292 static void cop2_assemble(int i,struct regstat *i_regs)
3294 u_int copr=(source[i]>>11)&0x1f;
3295 signed char temp=get_reg(i_regs->regmap,-1);
3296 if (opcode2[i]==0) { // MFC2
3297 signed char tl=get_reg(i_regs->regmap,rt1[i]);
3298 if(tl>=0&&rt1[i]!=0)
3299 cop2_get_dreg(copr,tl,temp);
3301 else if (opcode2[i]==4) { // MTC2
3302 signed char sl=get_reg(i_regs->regmap,rs1[i]);
3303 cop2_put_dreg(copr,sl,temp);
3305 else if (opcode2[i]==2) // CFC2
3307 signed char tl=get_reg(i_regs->regmap,rt1[i]);
3308 if(tl>=0&&rt1[i]!=0)
3309 emit_readword(®_cop2c[copr],tl);
3311 else if (opcode2[i]==6) // CTC2
3313 signed char sl=get_reg(i_regs->regmap,rs1[i]);
3322 emit_signextend16(sl,temp);
3325 //value = value & 0x7ffff000;
3326 //if (value & 0x7f87e000) value |= 0x80000000;
3327 emit_shrimm(sl,12,temp);
3328 emit_shlimm(temp,12,temp);
3329 emit_testimm(temp,0x7f000000);
3330 emit_testeqimm(temp,0x00870000);
3331 emit_testeqimm(temp,0x0000e000);
3332 emit_orrne_imm(temp,0x80000000,temp);
3338 emit_writeword(temp,®_cop2c[copr]);
3343 #ifndef multdiv_assemble
3344 void multdiv_assemble(int i,struct regstat *i_regs)
3346 printf("Need multdiv_assemble for this architecture.\n");
3351 void mov_assemble(int i,struct regstat *i_regs)
3353 //if(opcode2[i]==0x10||opcode2[i]==0x12) { // MFHI/MFLO
3354 //if(opcode2[i]==0x11||opcode2[i]==0x13) { // MTHI/MTLO
3356 signed char sh,sl,th,tl;
3357 th=get_reg(i_regs->regmap,rt1[i]|64);
3358 tl=get_reg(i_regs->regmap,rt1[i]);
3361 sh=get_reg(i_regs->regmap,rs1[i]|64);
3362 sl=get_reg(i_regs->regmap,rs1[i]);
3363 if(sl>=0) emit_mov(sl,tl);
3364 else emit_loadreg(rs1[i],tl);
3366 if(sh>=0) emit_mov(sh,th);
3367 else emit_loadreg(rs1[i]|64,th);
3373 void syscall_assemble(int i,struct regstat *i_regs)
3375 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3376 assert(ccreg==HOST_CCREG);
3377 assert(!is_delayslot);
3379 emit_movimm(start+i*4,EAX); // Get PC
3380 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle...
3381 emit_jmp(jump_syscall_hle); // XXX
3384 void hlecall_assemble(int i,struct regstat *i_regs)
3386 extern void psxNULL();
3387 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3388 assert(ccreg==HOST_CCREG);
3389 assert(!is_delayslot);
3391 emit_movimm(start+i*4+4,0); // Get PC
3392 uint32_t hleCode = source[i] & 0x03ffffff;
3393 if (hleCode >= ARRAY_SIZE(psxHLEt))
3394 emit_movimm((uintptr_t)psxNULL,1);
3396 emit_movimm((uintptr_t)psxHLEt[hleCode],1);
3397 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG); // XXX
3398 emit_jmp(jump_hlecall);
3401 void intcall_assemble(int i,struct regstat *i_regs)
3403 signed char ccreg=get_reg(i_regs->regmap,CCREG);
3404 assert(ccreg==HOST_CCREG);
3405 assert(!is_delayslot);
3407 emit_movimm(start+i*4,0); // Get PC
3408 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]),HOST_CCREG);
3409 emit_jmp(jump_intcall);
3412 static void speculate_mov(int rs,int rt)
3415 smrv_strong_next|=1<<rt;
3420 static void speculate_mov_weak(int rs,int rt)
3423 smrv_weak_next|=1<<rt;
3428 static void speculate_register_values(int i)
3431 memcpy(smrv,psxRegs.GPR.r,sizeof(smrv));
3432 // gp,sp are likely to stay the same throughout the block
3433 smrv_strong_next=(1<<28)|(1<<29)|(1<<30);
3434 smrv_weak_next=~smrv_strong_next;
3435 //printf(" llr %08x\n", smrv[4]);
3437 smrv_strong=smrv_strong_next;
3438 smrv_weak=smrv_weak_next;
3441 if ((smrv_strong>>rs1[i])&1) speculate_mov(rs1[i],rt1[i]);
3442 else if((smrv_strong>>rs2[i])&1) speculate_mov(rs2[i],rt1[i]);
3443 else if((smrv_weak>>rs1[i])&1) speculate_mov_weak(rs1[i],rt1[i]);
3444 else if((smrv_weak>>rs2[i])&1) speculate_mov_weak(rs2[i],rt1[i]);
3446 smrv_strong_next&=~(1<<rt1[i]);
3447 smrv_weak_next&=~(1<<rt1[i]);
3451 smrv_strong_next&=~(1<<rt1[i]);
3452 smrv_weak_next&=~(1<<rt1[i]);
3455 if(rt1[i]&&is_const(®s[i],rt1[i])) {
3456 int value,hr=get_reg(regs[i].regmap,rt1[i]);
3458 if(get_final_value(hr,i,&value))
3460 else smrv[rt1[i]]=constmap[i][hr];
3461 smrv_strong_next|=1<<rt1[i];
3465 if ((smrv_strong>>rs1[i])&1) speculate_mov(rs1[i],rt1[i]);
3466 else if((smrv_weak>>rs1[i])&1) speculate_mov_weak(rs1[i],rt1[i]);
3470 if(start<0x2000&&(rt1[i]==26||(smrv[rt1[i]]>>24)==0xa0)) {
3471 // special case for BIOS
3472 smrv[rt1[i]]=0xa0000000;
3473 smrv_strong_next|=1<<rt1[i];
3480 smrv_strong_next&=~(1<<rt1[i]);
3481 smrv_weak_next&=~(1<<rt1[i]);
3485 if(opcode2[i]==0||opcode2[i]==2) { // MFC/CFC
3486 smrv_strong_next&=~(1<<rt1[i]);
3487 smrv_weak_next&=~(1<<rt1[i]);
3491 if (opcode[i]==0x32) { // LWC2
3492 smrv_strong_next&=~(1<<rt1[i]);
3493 smrv_weak_next&=~(1<<rt1[i]);
3499 printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4,
3500 ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst);
3504 void ds_assemble(int i,struct regstat *i_regs)
3506 speculate_register_values(i);
3510 alu_assemble(i,i_regs);break;
3512 imm16_assemble(i,i_regs);break;
3514 shift_assemble(i,i_regs);break;
3516 shiftimm_assemble(i,i_regs);break;
3518 load_assemble(i,i_regs);break;
3520 loadlr_assemble(i,i_regs);break;
3522 store_assemble(i,i_regs);break;
3524 storelr_assemble(i,i_regs);break;
3526 cop0_assemble(i,i_regs);break;
3528 cop1_assemble(i,i_regs);break;
3530 c1ls_assemble(i,i_regs);break;
3532 cop2_assemble(i,i_regs);break;
3534 c2ls_assemble(i,i_regs);break;
3536 c2op_assemble(i,i_regs);break;
3538 multdiv_assemble(i,i_regs);break;
3540 mov_assemble(i,i_regs);break;
3549 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
3554 // Is the branch target a valid internal jump?
3555 static int internal_branch(int addr)
3557 if(addr&1) return 0; // Indirect (register) jump
3558 if(addr>=start && addr<start+slen*4-4)
3565 static void wb_invalidate(signed char pre[],signed char entry[],uint64_t dirty,uint64_t u)
3568 for(hr=0;hr<HOST_REGS;hr++) {
3569 if(hr!=EXCLUDE_REG) {
3570 if(pre[hr]!=entry[hr]) {
3573 if(get_reg(entry,pre[hr])<0) {
3575 if(!((u>>pre[hr])&1))
3576 emit_storereg(pre[hr],hr);
3583 // Move from one register to another (no writeback)
3584 for(hr=0;hr<HOST_REGS;hr++) {
3585 if(hr!=EXCLUDE_REG) {
3586 if(pre[hr]!=entry[hr]) {
3587 if(pre[hr]>=0&&(pre[hr]&63)<TEMPREG) {
3589 if((nr=get_reg(entry,pre[hr]))>=0) {
3598 // Load the specified registers
3599 // This only loads the registers given as arguments because
3600 // we don't want to load things that will be overwritten
3601 static void load_regs(signed char entry[],signed char regmap[],int rs1,int rs2)
3605 for(hr=0;hr<HOST_REGS;hr++) {
3606 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3607 if(entry[hr]!=regmap[hr]) {
3608 if(regmap[hr]==rs1||regmap[hr]==rs2)
3615 emit_loadreg(regmap[hr],hr);
3623 // Load registers prior to the start of a loop
3624 // so that they are not loaded within the loop
3625 static void loop_preload(signed char pre[],signed char entry[])
3628 for(hr=0;hr<HOST_REGS;hr++) {
3629 if(hr!=EXCLUDE_REG) {
3630 if(pre[hr]!=entry[hr]) {
3632 if(get_reg(pre,entry[hr])<0) {
3633 assem_debug("loop preload:\n");
3634 //printf("loop preload: %d\n",hr);
3638 else if(entry[hr]<TEMPREG)
3640 emit_loadreg(entry[hr],hr);
3642 else if(entry[hr]-64<TEMPREG)
3644 emit_loadreg(entry[hr],hr);
3653 // Generate address for load/store instruction
3654 // goes to AGEN for writes, FTEMP for LOADLR and cop1/2 loads
3655 void address_generation(int i,struct regstat *i_regs,signed char entry[])
3657 if(itype[i]==LOAD||itype[i]==LOADLR||itype[i]==STORE||itype[i]==STORELR||itype[i]==C1LS||itype[i]==C2LS) {
3659 int agr=AGEN1+(i&1);
3660 if(itype[i]==LOAD) {
3661 ra=get_reg(i_regs->regmap,rt1[i]);
3662 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3665 if(itype[i]==LOADLR) {
3666 ra=get_reg(i_regs->regmap,FTEMP);
3668 if(itype[i]==STORE||itype[i]==STORELR) {
3669 ra=get_reg(i_regs->regmap,agr);
3670 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3672 if(itype[i]==C1LS||itype[i]==C2LS) {
3673 if ((opcode[i]&0x3b)==0x31||(opcode[i]&0x3b)==0x32) // LWC1/LDC1/LWC2/LDC2
3674 ra=get_reg(i_regs->regmap,FTEMP);
3675 else { // SWC1/SDC1/SWC2/SDC2
3676 ra=get_reg(i_regs->regmap,agr);
3677 if(ra<0) ra=get_reg(i_regs->regmap,-1);
3680 int rs=get_reg(i_regs->regmap,rs1[i]);
3683 int c=(i_regs->wasconst>>rs)&1;
3685 // Using r0 as a base address
3686 if(!entry||entry[ra]!=agr) {
3687 if (opcode[i]==0x22||opcode[i]==0x26) {
3688 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
3689 }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
3690 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
3692 emit_movimm(offset,ra);
3694 } // else did it in the previous cycle
3697 if(!entry||entry[ra]!=rs1[i])
3698 emit_loadreg(rs1[i],ra);
3699 //if(!entry||entry[ra]!=rs1[i])
3700 // printf("poor load scheduling!\n");
3703 if(rs1[i]!=rt1[i]||itype[i]!=LOAD) {
3704 if(!entry||entry[ra]!=agr) {
3705 if (opcode[i]==0x22||opcode[i]==0x26) {
3706 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
3707 }else if (opcode[i]==0x1a||opcode[i]==0x1b) {
3708 emit_movimm((constmap[i][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
3710 emit_movimm(constmap[i][rs]+offset,ra);
3711 regs[i].loadedconst|=1<<ra;
3713 } // else did it in the previous cycle
3714 } // else load_consts already did it
3716 if(offset&&!c&&rs1[i]) {
3718 emit_addimm(rs,offset,ra);
3720 emit_addimm(ra,offset,ra);
3725 // Preload constants for next instruction
3726 if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS||itype[i+1]==C2LS) {
3729 agr=AGEN1+((i+1)&1);
3730 ra=get_reg(i_regs->regmap,agr);
3732 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
3733 int offset=imm[i+1];
3734 int c=(regs[i+1].wasconst>>rs)&1;
3735 if(c&&(rs1[i+1]!=rt1[i+1]||itype[i+1]!=LOAD)) {
3736 if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
3737 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFFC,ra); // LWL/LWR
3738 }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
3739 emit_movimm((constmap[i+1][rs]+offset)&0xFFFFFFF8,ra); // LDL/LDR
3741 emit_movimm(constmap[i+1][rs]+offset,ra);
3742 regs[i+1].loadedconst|=1<<ra;
3745 else if(rs1[i+1]==0) {
3746 // Using r0 as a base address
3747 if (opcode[i+1]==0x22||opcode[i+1]==0x26) {
3748 emit_movimm(offset&0xFFFFFFFC,ra); // LWL/LWR
3749 }else if (opcode[i+1]==0x1a||opcode[i+1]==0x1b) {
3750 emit_movimm(offset&0xFFFFFFF8,ra); // LDL/LDR
3752 emit_movimm(offset,ra);
3759 static int get_final_value(int hr, int i, int *value)
3761 int reg=regs[i].regmap[hr];
3763 if(regs[i+1].regmap[hr]!=reg) break;
3764 if(!((regs[i+1].isconst>>hr)&1)) break;
3769 if(itype[i]==UJUMP||itype[i]==RJUMP||itype[i]==CJUMP||itype[i]==SJUMP) {
3770 *value=constmap[i][hr];
3774 if(itype[i+1]==UJUMP||itype[i+1]==RJUMP||itype[i+1]==CJUMP||itype[i+1]==SJUMP) {
3775 // Load in delay slot, out-of-order execution
3776 if(itype[i+2]==LOAD&&rs1[i+2]==reg&&rt1[i+2]==reg&&((regs[i+1].wasconst>>hr)&1))
3778 // Precompute load address
3779 *value=constmap[i][hr]+imm[i+2];
3783 if(itype[i+1]==LOAD&&rs1[i+1]==reg&&rt1[i+1]==reg)
3785 // Precompute load address
3786 *value=constmap[i][hr]+imm[i+1];
3787 //printf("c=%x imm=%lx\n",(long)constmap[i][hr],imm[i+1]);
3792 *value=constmap[i][hr];
3793 //printf("c=%lx\n",(long)constmap[i][hr]);
3794 if(i==slen-1) return 1;
3796 return !((unneeded_reg[i+1]>>reg)&1);
3799 // Load registers with known constants
3800 static void load_consts(signed char pre[],signed char regmap[],int i)
3803 // propagate loaded constant flags
3805 regs[i].loadedconst=0;
3807 for(hr=0;hr<HOST_REGS;hr++) {
3808 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((regs[i-1].isconst>>hr)&1)&&pre[hr]==regmap[hr]
3809 &®map[hr]==regs[i-1].regmap[hr]&&((regs[i-1].loadedconst>>hr)&1))
3811 regs[i].loadedconst|=1<<hr;
3816 for(hr=0;hr<HOST_REGS;hr++) {
3817 if(hr!=EXCLUDE_REG&®map[hr]>=0) {
3818 //if(entry[hr]!=regmap[hr]) {
3819 if(!((regs[i].loadedconst>>hr)&1)) {
3820 assert(regmap[hr]<64);
3821 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
3822 int value,similar=0;
3823 if(get_final_value(hr,i,&value)) {
3824 // see if some other register has similar value
3825 for(hr2=0;hr2<HOST_REGS;hr2++) {
3826 if(hr2!=EXCLUDE_REG&&((regs[i].loadedconst>>hr2)&1)) {
3827 if(is_similar_value(value,constmap[i][hr2])) {
3835 if(get_final_value(hr2,i,&value2)) // is this needed?
3836 emit_movimm_from(value2,hr2,value,hr);
3838 emit_movimm(value,hr);
3844 emit_movimm(value,hr);
3847 regs[i].loadedconst|=1<<hr;
3854 void load_all_consts(signed char regmap[], u_int dirty, int i)
3858 for(hr=0;hr<HOST_REGS;hr++) {
3859 if(hr!=EXCLUDE_REG&®map[hr]>=0&&((dirty>>hr)&1)) {
3860 assert(regmap[hr] < 64);
3861 if(((regs[i].isconst>>hr)&1)&®map[hr]>0) {
3862 int value=constmap[i][hr];
3867 emit_movimm(value,hr);
3874 // Write out all dirty registers (except cycle count)
3875 static void wb_dirtys(signed char i_regmap[],uint64_t i_dirty)
3878 for(hr=0;hr<HOST_REGS;hr++) {
3879 if(hr!=EXCLUDE_REG) {
3880 if(i_regmap[hr]>0) {
3881 if(i_regmap[hr]!=CCREG) {
3882 if((i_dirty>>hr)&1) {
3883 assert(i_regmap[hr]<64);
3884 emit_storereg(i_regmap[hr],hr);
3892 // Write out dirty registers that we need to reload (pair with load_needed_regs)
3893 // This writes the registers not written by store_regs_bt
3894 void wb_needed_dirtys(signed char i_regmap[],uint64_t i_dirty,int addr)
3897 int t=(addr-start)>>2;
3898 for(hr=0;hr<HOST_REGS;hr++) {
3899 if(hr!=EXCLUDE_REG) {
3900 if(i_regmap[hr]>0) {
3901 if(i_regmap[hr]!=CCREG) {
3902 if(i_regmap[hr]==regs[t].regmap_entry[hr] && ((regs[t].dirty>>hr)&1)) {
3903 if((i_dirty>>hr)&1) {
3904 assert(i_regmap[hr]<64);
3905 emit_storereg(i_regmap[hr],hr);
3914 // Load all registers (except cycle count)
3915 void load_all_regs(signed char i_regmap[])
3918 for(hr=0;hr<HOST_REGS;hr++) {
3919 if(hr!=EXCLUDE_REG) {
3920 if(i_regmap[hr]==0) {
3924 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
3926 emit_loadreg(i_regmap[hr],hr);
3932 // Load all current registers also needed by next instruction
3933 void load_needed_regs(signed char i_regmap[],signed char next_regmap[])
3936 for(hr=0;hr<HOST_REGS;hr++) {
3937 if(hr!=EXCLUDE_REG) {
3938 if(get_reg(next_regmap,i_regmap[hr])>=0) {
3939 if(i_regmap[hr]==0) {
3943 if(i_regmap[hr]>0 && (i_regmap[hr]&63)<TEMPREG && i_regmap[hr]!=CCREG)
3945 emit_loadreg(i_regmap[hr],hr);
3952 // Load all regs, storing cycle count if necessary
3953 void load_regs_entry(int t)
3956 if(is_ds[t]) emit_addimm(HOST_CCREG,CLOCK_ADJUST(1),HOST_CCREG);
3957 else if(ccadj[t]) emit_addimm(HOST_CCREG,-CLOCK_ADJUST(ccadj[t]),HOST_CCREG);
3958 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
3959 emit_storereg(CCREG,HOST_CCREG);
3962 for(hr=0;hr<HOST_REGS;hr++) {
3963 if(regs[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
3964 if(regs[t].regmap_entry[hr]==0) {
3967 else if(regs[t].regmap_entry[hr]!=CCREG)
3969 emit_loadreg(regs[t].regmap_entry[hr],hr);
3975 // Store dirty registers prior to branch
3976 void store_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
3978 if(internal_branch(addr))
3980 int t=(addr-start)>>2;
3982 for(hr=0;hr<HOST_REGS;hr++) {
3983 if(hr!=EXCLUDE_REG) {
3984 if(i_regmap[hr]>0 && i_regmap[hr]!=CCREG) {
3985 if(i_regmap[hr]!=regs[t].regmap_entry[hr] || !((regs[t].dirty>>hr)&1)) {
3986 if((i_dirty>>hr)&1) {
3987 assert(i_regmap[hr]<64);
3988 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
3989 emit_storereg(i_regmap[hr],hr);
3998 // Branch out of this block, write out all dirty regs
3999 wb_dirtys(i_regmap,i_dirty);
4003 // Load all needed registers for branch target
4004 static void load_regs_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4006 //if(addr>=start && addr<(start+slen*4))
4007 if(internal_branch(addr))
4009 int t=(addr-start)>>2;
4011 // Store the cycle count before loading something else
4012 if(i_regmap[HOST_CCREG]!=CCREG) {
4013 assert(i_regmap[HOST_CCREG]==-1);
4015 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) {
4016 emit_storereg(CCREG,HOST_CCREG);
4019 for(hr=0;hr<HOST_REGS;hr++) {
4020 if(hr!=EXCLUDE_REG&®s[t].regmap_entry[hr]>=0&®s[t].regmap_entry[hr]<TEMPREG) {
4021 if(i_regmap[hr]!=regs[t].regmap_entry[hr]) {
4022 if(regs[t].regmap_entry[hr]==0) {
4025 else if(regs[t].regmap_entry[hr]!=CCREG)
4027 emit_loadreg(regs[t].regmap_entry[hr],hr);
4035 static int match_bt(signed char i_regmap[],uint64_t i_dirty,int addr)
4037 if(addr>=start && addr<start+slen*4-4)
4039 int t=(addr-start)>>2;
4041 if(regs[t].regmap_entry[HOST_CCREG]!=CCREG) return 0;
4042 for(hr=0;hr<HOST_REGS;hr++)
4046 if(i_regmap[hr]!=regs[t].regmap_entry[hr])
4048 if(regs[t].regmap_entry[hr]>=0&&(regs[t].regmap_entry[hr]|64)<TEMPREG+64)
4055 if(i_regmap[hr]<TEMPREG)
4057 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4060 else if(i_regmap[hr]>=64&&i_regmap[hr]<TEMPREG+64)
4066 else // Same register but is it 32-bit or dirty?
4069 if(!((regs[t].dirty>>hr)&1))
4073 if(!((unneeded_reg[t]>>i_regmap[hr])&1))
4075 //printf("%x: dirty no match\n",addr);
4083 // Delay slots are not valid branch targets
4084 //if(t>0&&(itype[t-1]==RJUMP||itype[t-1]==UJUMP||itype[t-1]==CJUMP||itype[t-1]==SJUMP)) return 0;
4085 // Delay slots require additional processing, so do not match
4086 if(is_ds[t]) return 0;
4091 for(hr=0;hr<HOST_REGS;hr++)
4097 if(hr!=HOST_CCREG||i_regmap[hr]!=CCREG)
4112 static void drc_dbg_emit_do_cmp(int i)
4114 extern void do_insn_cmp();
4118 for(hr=0;hr<HOST_REGS;hr++)
4119 if(regs[i].regmap[hr]>=0) reglist|=1<<hr;
4121 emit_movimm(start+i*4,0);
4122 emit_writeword(0,&pcaddr);
4123 emit_call(do_insn_cmp);
4124 //emit_readword(&cycle,0);
4125 //emit_addimm(0,2,0);
4126 //emit_writeword(0,&cycle);
4127 restore_regs(reglist);
4130 #define drc_dbg_emit_do_cmp(x)
4133 // Used when a branch jumps into the delay slot of another branch
4134 void ds_assemble_entry(int i)
4136 int t=(ba[i]-start)>>2;
4138 instr_addr[t] = out;
4139 assem_debug("Assemble delay slot at %x\n",ba[i]);
4140 assem_debug("<->\n");
4141 drc_dbg_emit_do_cmp(t);
4142 if(regs[t].regmap_entry[HOST_CCREG]==CCREG&®s[t].regmap[HOST_CCREG]!=CCREG)
4143 wb_register(CCREG,regs[t].regmap_entry,regs[t].wasdirty);
4144 load_regs(regs[t].regmap_entry,regs[t].regmap,rs1[t],rs2[t]);
4145 address_generation(t,®s[t],regs[t].regmap_entry);
4146 if(itype[t]==STORE||itype[t]==STORELR||(opcode[t]&0x3b)==0x39||(opcode[t]&0x3b)==0x3a)
4147 load_regs(regs[t].regmap_entry,regs[t].regmap,INVCP,INVCP);
4151 alu_assemble(t,®s[t]);break;
4153 imm16_assemble(t,®s[t]);break;
4155 shift_assemble(t,®s[t]);break;
4157 shiftimm_assemble(t,®s[t]);break;
4159 load_assemble(t,®s[t]);break;
4161 loadlr_assemble(t,®s[t]);break;
4163 store_assemble(t,®s[t]);break;
4165 storelr_assemble(t,®s[t]);break;
4167 cop0_assemble(t,®s[t]);break;
4169 cop1_assemble(t,®s[t]);break;
4171 c1ls_assemble(t,®s[t]);break;
4173 cop2_assemble(t,®s[t]);break;
4175 c2ls_assemble(t,®s[t]);break;
4177 c2op_assemble(t,®s[t]);break;
4179 multdiv_assemble(t,®s[t]);break;
4181 mov_assemble(t,®s[t]);break;
4190 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
4192 store_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4193 load_regs_bt(regs[t].regmap,regs[t].dirty,ba[i]+4);
4194 if(internal_branch(ba[i]+4))
4195 assem_debug("branch: internal\n");
4197 assem_debug("branch: external\n");
4198 assert(internal_branch(ba[i]+4));
4199 add_to_linker(out,ba[i]+4,internal_branch(ba[i]+4));
4203 void do_cc(int i,signed char i_regmap[],int *adj,int addr,int taken,int invert)
4213 //if(ba[i]>=start && ba[i]<(start+slen*4))
4214 if(internal_branch(ba[i]))
4217 if(is_ds[t]) *adj=-1; // Branch into delay slot adds an extra cycle
4225 if(taken==TAKEN && i==(ba[i]-start)>>2 && source[i+1]==0) {
4227 if(count&1) emit_addimm_and_set_flags(2*(count+2),HOST_CCREG);
4229 //emit_subfrommem(&idlecount,HOST_CCREG); // Count idle cycles
4230 emit_andimm(HOST_CCREG,3,HOST_CCREG);
4234 else if(*adj==0||invert) {
4235 int cycles=CLOCK_ADJUST(count+2);
4239 if(-NO_CYCLE_PENALTY_THR<rel&&rel<0)
4240 cycles=CLOCK_ADJUST(*adj)+count+2-*adj;
4242 emit_addimm_and_set_flags(cycles,HOST_CCREG);
4248 emit_cmpimm(HOST_CCREG,-CLOCK_ADJUST(count+2));
4252 add_stub(CC_STUB,jaddr,idle?idle:out,(*adj==0||invert||idle)?0:(count+2),i,addr,taken,0);
4255 static void do_ccstub(int n)
4258 assem_debug("do_ccstub %lx\n",start+stubs[n].b*4);
4259 set_jump_target(stubs[n].addr, out);
4261 if(stubs[n].d==NULLDS) {
4262 // Delay slot instruction is nullified ("likely" branch)
4263 wb_dirtys(regs[i].regmap,regs[i].dirty);
4265 else if(stubs[n].d!=TAKEN) {
4266 wb_dirtys(branch_regs[i].regmap,branch_regs[i].dirty);
4269 if(internal_branch(ba[i]))
4270 wb_needed_dirtys(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4274 // Save PC as return address
4275 emit_movimm(stubs[n].c,EAX);
4276 emit_writeword(EAX,&pcaddr);
4280 // Return address depends on which way the branch goes
4281 if(itype[i]==CJUMP||itype[i]==SJUMP)
4283 int s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4284 int s2l=get_reg(branch_regs[i].regmap,rs2[i]);
4295 #ifdef DESTRUCTIVE_WRITEBACK
4297 if((branch_regs[i].dirty>>s1l)&&1)
4298 emit_loadreg(rs1[i],s1l);
4301 if((branch_regs[i].dirty>>s1l)&1)
4302 emit_loadreg(rs2[i],s1l);
4305 if((branch_regs[i].dirty>>s2l)&1)
4306 emit_loadreg(rs2[i],s2l);
4309 int addr=-1,alt=-1,ntaddr=-1;
4312 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4313 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4314 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4322 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4323 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4324 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4330 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
4334 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
4335 (branch_regs[i].regmap[hr]&63)!=rs1[i] &&
4336 (branch_regs[i].regmap[hr]&63)!=rs2[i] )
4342 assert(hr<HOST_REGS);
4344 if((opcode[i]&0x2f)==4) // BEQ
4346 #ifdef HAVE_CMOV_IMM
4347 if(s2l>=0) emit_cmp(s1l,s2l);
4348 else emit_test(s1l,s1l);
4349 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
4351 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4352 if(s2l>=0) emit_cmp(s1l,s2l);
4353 else emit_test(s1l,s1l);
4354 emit_cmovne_reg(alt,addr);
4357 if((opcode[i]&0x2f)==5) // BNE
4359 #ifdef HAVE_CMOV_IMM
4360 if(s2l>=0) emit_cmp(s1l,s2l);
4361 else emit_test(s1l,s1l);
4362 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
4364 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
4365 if(s2l>=0) emit_cmp(s1l,s2l);
4366 else emit_test(s1l,s1l);
4367 emit_cmovne_reg(alt,addr);
4370 if((opcode[i]&0x2f)==6) // BLEZ
4372 //emit_movimm(ba[i],alt);
4373 //emit_movimm(start+i*4+8,addr);
4374 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4376 emit_cmovl_reg(alt,addr);
4378 if((opcode[i]&0x2f)==7) // BGTZ
4380 //emit_movimm(ba[i],addr);
4381 //emit_movimm(start+i*4+8,ntaddr);
4382 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
4384 emit_cmovl_reg(ntaddr,addr);
4386 if((opcode[i]==1)&&(opcode2[i]&0x2D)==0) // BLTZ
4388 //emit_movimm(ba[i],alt);
4389 //emit_movimm(start+i*4+8,addr);
4390 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4392 emit_cmovs_reg(alt,addr);
4394 if((opcode[i]==1)&&(opcode2[i]&0x2D)==1) // BGEZ
4396 //emit_movimm(ba[i],addr);
4397 //emit_movimm(start+i*4+8,alt);
4398 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4400 emit_cmovs_reg(alt,addr);
4402 if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
4403 if(source[i]&0x10000) // BC1T
4405 //emit_movimm(ba[i],alt);
4406 //emit_movimm(start+i*4+8,addr);
4407 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
4408 emit_testimm(s1l,0x800000);
4409 emit_cmovne_reg(alt,addr);
4413 //emit_movimm(ba[i],addr);
4414 //emit_movimm(start+i*4+8,alt);
4415 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
4416 emit_testimm(s1l,0x800000);
4417 emit_cmovne_reg(alt,addr);
4420 emit_writeword(addr,&pcaddr);
4425 int r=get_reg(branch_regs[i].regmap,rs1[i]);
4426 if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
4427 r=get_reg(branch_regs[i].regmap,RTEMP);
4429 emit_writeword(r,&pcaddr);
4431 else {SysPrintf("Unknown branch type in do_ccstub\n");exit(1);}
4433 // Update cycle count
4434 assert(branch_regs[i].regmap[HOST_CCREG]==CCREG||branch_regs[i].regmap[HOST_CCREG]==-1);
4435 if(stubs[n].a) emit_addimm(HOST_CCREG,CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
4436 emit_call(cc_interrupt);
4437 if(stubs[n].a) emit_addimm(HOST_CCREG,-CLOCK_ADJUST((signed int)stubs[n].a),HOST_CCREG);
4438 if(stubs[n].d==TAKEN) {
4439 if(internal_branch(ba[i]))
4440 load_needed_regs(branch_regs[i].regmap,regs[(ba[i]-start)>>2].regmap_entry);
4441 else if(itype[i]==RJUMP) {
4442 if(get_reg(branch_regs[i].regmap,RTEMP)>=0)
4443 emit_readword(&pcaddr,get_reg(branch_regs[i].regmap,RTEMP));
4445 emit_loadreg(rs1[i],get_reg(branch_regs[i].regmap,rs1[i]));
4447 }else if(stubs[n].d==NOTTAKEN) {
4448 if(i<slen-2) load_needed_regs(branch_regs[i].regmap,regmap_pre[i+2]);
4449 else load_all_regs(branch_regs[i].regmap);
4450 }else if(stubs[n].d==NULLDS) {
4451 // Delay slot instruction is nullified ("likely" branch)
4452 if(i<slen-2) load_needed_regs(regs[i].regmap,regmap_pre[i+2]);
4453 else load_all_regs(regs[i].regmap);
4455 load_all_regs(branch_regs[i].regmap);
4457 emit_jmp(stubs[n].retaddr);
4460 static void add_to_linker(void *addr, u_int target, int ext)
4462 assert(linkcount < ARRAY_SIZE(link_addr));
4463 link_addr[linkcount].addr = addr;
4464 link_addr[linkcount].target = target;
4465 link_addr[linkcount].ext = ext;
4469 static void ujump_assemble_write_ra(int i)
4472 unsigned int return_address;
4473 rt=get_reg(branch_regs[i].regmap,31);
4474 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]);
4476 return_address=start+i*4+8;
4479 if(internal_branch(return_address)&&rt1[i+1]!=31) {
4480 int temp=-1; // note: must be ds-safe
4484 if(temp>=0) do_miniht_insert(return_address,rt,temp);
4485 else emit_movimm(return_address,rt);
4493 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4496 emit_movimm(return_address,rt); // PC into link register
4498 emit_prefetch(hash_table_get(return_address));
4504 void ujump_assemble(int i,struct regstat *i_regs)
4507 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4508 address_generation(i+1,i_regs,regs[i].regmap_entry);
4510 int temp=get_reg(branch_regs[i].regmap,PTEMP);
4511 if(rt1[i]==31&&temp>=0)
4513 signed char *i_regmap=i_regs->regmap;
4514 int return_address=start+i*4+8;
4515 if(get_reg(branch_regs[i].regmap,31)>0)
4516 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4519 if(rt1[i]==31&&(rt1[i]==rs1[i+1]||rt1[i]==rs2[i+1])) {
4520 ujump_assemble_write_ra(i); // writeback ra for DS
4523 ds_assemble(i+1,i_regs);
4524 uint64_t bc_unneeded=branch_regs[i].u;
4525 bc_unneeded|=1|(1LL<<rt1[i]);
4526 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
4527 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
4528 if(!ra_done&&rt1[i]==31)
4529 ujump_assemble_write_ra(i);
4531 cc=get_reg(branch_regs[i].regmap,CCREG);
4532 assert(cc==HOST_CCREG);
4533 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4535 if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
4537 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
4538 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
4539 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4540 if(internal_branch(ba[i]))
4541 assem_debug("branch: internal\n");
4543 assem_debug("branch: external\n");
4544 if(internal_branch(ba[i])&&is_ds[(ba[i]-start)>>2]) {
4545 ds_assemble_entry(i);
4548 add_to_linker(out,ba[i],internal_branch(ba[i]));
4553 static void rjump_assemble_write_ra(int i)
4555 int rt,return_address;
4556 assert(rt1[i+1]!=rt1[i]);
4557 assert(rt2[i+1]!=rt1[i]);
4558 rt=get_reg(branch_regs[i].regmap,rt1[i]);
4559 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]);
4561 return_address=start+i*4+8;
4565 if(i_regmap[temp]!=PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4568 emit_movimm(return_address,rt); // PC into link register
4570 emit_prefetch(hash_table_get(return_address));
4574 void rjump_assemble(int i,struct regstat *i_regs)
4579 rs=get_reg(branch_regs[i].regmap,rs1[i]);
4581 if(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]) {
4582 // Delay slot abuse, make a copy of the branch address register
4583 temp=get_reg(branch_regs[i].regmap,RTEMP);
4585 assert(regs[i].regmap[temp]==RTEMP);
4589 address_generation(i+1,i_regs,regs[i].regmap_entry);
4593 if((temp=get_reg(branch_regs[i].regmap,PTEMP))>=0) {
4594 signed char *i_regmap=i_regs->regmap;
4595 int return_address=start+i*4+8;
4596 if(i_regmap[temp]==PTEMP) emit_movimm((uintptr_t)hash_table_get(return_address),temp);
4602 int rh=get_reg(regs[i].regmap,RHASH);
4603 if(rh>=0) do_preload_rhash(rh);
4606 if(rt1[i]!=0&&(rt1[i]==rs1[i+1]||rt1[i]==rs2[i+1])) {
4607 rjump_assemble_write_ra(i);
4610 ds_assemble(i+1,i_regs);
4611 uint64_t bc_unneeded=branch_regs[i].u;
4612 bc_unneeded|=1|(1LL<<rt1[i]);
4613 bc_unneeded&=~(1LL<<rs1[i]);
4614 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
4615 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i],CCREG);
4616 if(!ra_done&&rt1[i]!=0)
4617 rjump_assemble_write_ra(i);
4618 cc=get_reg(branch_regs[i].regmap,CCREG);
4619 assert(cc==HOST_CCREG);
4622 int rh=get_reg(branch_regs[i].regmap,RHASH);
4623 int ht=get_reg(branch_regs[i].regmap,RHTBL);
4625 if(regs[i].regmap[rh]!=RHASH) do_preload_rhash(rh);
4626 do_preload_rhtbl(ht);
4630 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
4631 #ifdef DESTRUCTIVE_WRITEBACK
4632 if((branch_regs[i].dirty>>rs)&1) {
4633 if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
4634 emit_loadreg(rs1[i],rs);
4639 if(rt1[i]==31&&temp>=0) emit_prefetchreg(temp);
4643 do_miniht_load(ht,rh);
4646 //do_cc(i,branch_regs[i].regmap,&adj,-1,TAKEN);
4647 //if(adj) emit_addimm(cc,2*(ccadj[i]+2-adj),cc); // ??? - Shouldn't happen
4649 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
4650 add_stub(CC_STUB,out,jump_vaddr_reg[rs],0,i,-1,TAKEN,0);
4651 if(itype[i+1]==COP0&&(source[i+1]&0x3f)==0x10)
4652 // special case for RFE
4656 //load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,-1);
4659 do_miniht_jump(rs,rh,ht);
4664 emit_jmp(jump_vaddr_reg[rs]);
4666 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4667 if(rt1[i]!=31&&i<slen-2&&(((u_int)out)&7)) emit_mov(13,13);
4671 void cjump_assemble(int i,struct regstat *i_regs)
4673 signed char *i_regmap=i_regs->regmap;
4676 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4677 assem_debug("match=%d\n",match);
4679 int unconditional=0,nop=0;
4681 int internal=internal_branch(ba[i]);
4682 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4683 if(!match) invert=1;
4684 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4685 if(i>(ba[i]-start)>>2) invert=1;
4689 s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4690 s2l=get_reg(branch_regs[i].regmap,rs2[i]);
4693 s1l=get_reg(i_regmap,rs1[i]);
4694 s2l=get_reg(i_regmap,rs2[i]);
4696 if(rs1[i]==0&&rs2[i]==0)
4698 if(opcode[i]&1) nop=1;
4699 else unconditional=1;
4700 //assert(opcode[i]!=5);
4701 //assert(opcode[i]!=7);
4702 //assert(opcode[i]!=0x15);
4703 //assert(opcode[i]!=0x17);
4716 // Out of order execution (delay slot first)
4718 address_generation(i+1,i_regs,regs[i].regmap_entry);
4719 ds_assemble(i+1,i_regs);
4721 uint64_t bc_unneeded=branch_regs[i].u;
4722 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
4724 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
4725 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i],rs2[i]);
4726 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
4727 cc=get_reg(branch_regs[i].regmap,CCREG);
4728 assert(cc==HOST_CCREG);
4730 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4731 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
4732 //assem_debug("cycle count (adj)\n");
4734 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
4735 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
4736 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
4737 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4739 assem_debug("branch: internal\n");
4741 assem_debug("branch: external\n");
4742 if(internal&&is_ds[(ba[i]-start)>>2]) {
4743 ds_assemble_entry(i);
4746 add_to_linker(out,ba[i],internal);
4749 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4750 if(((u_int)out)&7) emit_addnop(0);
4755 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
4758 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
4761 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
4762 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
4763 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
4765 //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]);
4767 if(opcode[i]==4) // BEQ
4769 if(s2l>=0) emit_cmp(s1l,s2l);
4770 else emit_test(s1l,s1l);
4773 emit_jne((void *)1l);
4775 add_to_linker(out,ba[i],internal);
4779 if(opcode[i]==5) // BNE
4781 if(s2l>=0) emit_cmp(s1l,s2l);
4782 else emit_test(s1l,s1l);
4787 add_to_linker(out,ba[i],internal);
4791 if(opcode[i]==6) // BLEZ
4798 add_to_linker(out,ba[i],internal);
4802 if(opcode[i]==7) // BGTZ
4809 add_to_linker(out,ba[i],internal);
4814 if(taken) set_jump_target(taken, out);
4815 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4816 if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
4818 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
4819 add_to_linker(out,ba[i],internal);
4822 add_to_linker(out,ba[i],internal*2);
4828 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
4829 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4830 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4832 assem_debug("branch: internal\n");
4834 assem_debug("branch: external\n");
4835 if(internal&&is_ds[(ba[i]-start)>>2]) {
4836 ds_assemble_entry(i);
4839 add_to_linker(out,ba[i],internal);
4843 set_jump_target(nottaken, out);
4846 if(nottaken1) set_jump_target(nottaken1, out);
4848 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
4850 } // (!unconditional)
4854 // In-order execution (branch first)
4855 //if(likely[i]) printf("IOL\n");
4858 void *taken = NULL, *nottaken = NULL, *nottaken1 = NULL;
4859 if(!unconditional&&!nop) {
4860 //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]);
4862 if((opcode[i]&0x2f)==4) // BEQ
4864 if(s2l>=0) emit_cmp(s1l,s2l);
4865 else emit_test(s1l,s1l);
4867 emit_jne((void *)2l);
4869 if((opcode[i]&0x2f)==5) // BNE
4871 if(s2l>=0) emit_cmp(s1l,s2l);
4872 else emit_test(s1l,s1l);
4876 if((opcode[i]&0x2f)==6) // BLEZ
4882 if((opcode[i]&0x2f)==7) // BGTZ
4888 } // if(!unconditional)
4890 uint64_t ds_unneeded=branch_regs[i].u;
4891 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
4895 if(taken) set_jump_target(taken, out);
4896 assem_debug("1:\n");
4897 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
4899 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i+1],rs2[i+1]);
4900 address_generation(i+1,&branch_regs[i],0);
4901 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
4902 ds_assemble(i+1,&branch_regs[i]);
4903 cc=get_reg(branch_regs[i].regmap,CCREG);
4905 emit_loadreg(CCREG,cc=HOST_CCREG);
4906 // CHECK: Is the following instruction (fall thru) allocated ok?
4908 assert(cc==HOST_CCREG);
4909 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4910 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
4911 assem_debug("cycle count (adj)\n");
4912 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
4913 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4915 assem_debug("branch: internal\n");
4917 assem_debug("branch: external\n");
4918 if(internal&&is_ds[(ba[i]-start)>>2]) {
4919 ds_assemble_entry(i);
4922 add_to_linker(out,ba[i],internal);
4927 if(!unconditional) {
4928 if(nottaken1) set_jump_target(nottaken1, out);
4929 set_jump_target(nottaken, out);
4930 assem_debug("2:\n");
4932 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
4933 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i+1],rs2[i+1]);
4934 address_generation(i+1,&branch_regs[i],0);
4935 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
4936 ds_assemble(i+1,&branch_regs[i]);
4938 cc=get_reg(branch_regs[i].regmap,CCREG);
4939 if(cc==-1&&!likely[i]) {
4940 // Cycle count isn't in a register, temporarily load it then write it out
4941 emit_loadreg(CCREG,HOST_CCREG);
4942 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
4945 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
4946 emit_storereg(CCREG,HOST_CCREG);
4949 cc=get_reg(i_regmap,CCREG);
4950 assert(cc==HOST_CCREG);
4951 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
4954 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
4960 void sjump_assemble(int i,struct regstat *i_regs)
4962 signed char *i_regmap=i_regs->regmap;
4965 match=match_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
4966 assem_debug("smatch=%d\n",match);
4968 int unconditional=0,nevertaken=0;
4970 int internal=internal_branch(ba[i]);
4971 if(i==(ba[i]-start)>>2) assem_debug("idle loop\n");
4972 if(!match) invert=1;
4973 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
4974 if(i>(ba[i]-start)>>2) invert=1;
4977 //if(opcode2[i]>=0x10) return; // FIXME (BxxZAL)
4978 //assert(opcode2[i]<0x10||rs1[i]==0); // FIXME (BxxZAL)
4981 s1l=get_reg(branch_regs[i].regmap,rs1[i]);
4984 s1l=get_reg(i_regmap,rs1[i]);
4988 if(opcode2[i]&1) unconditional=1;
4990 // These are never taken (r0 is never less than zero)
4991 //assert(opcode2[i]!=0);
4992 //assert(opcode2[i]!=2);
4993 //assert(opcode2[i]!=0x10);
4994 //assert(opcode2[i]!=0x12);
4998 // Out of order execution (delay slot first)
5000 address_generation(i+1,i_regs,regs[i].regmap_entry);
5001 ds_assemble(i+1,i_regs);
5003 uint64_t bc_unneeded=branch_regs[i].u;
5004 bc_unneeded&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5006 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,bc_unneeded);
5007 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i],rs1[i]);
5008 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5010 int rt,return_address;
5011 rt=get_reg(branch_regs[i].regmap,31);
5012 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]);
5014 // Save the PC even if the branch is not taken
5015 return_address=start+i*4+8;
5016 emit_movimm(return_address,rt); // PC into link register
5018 if(!nevertaken) emit_prefetch(hash_table_get(return_address));
5022 cc=get_reg(branch_regs[i].regmap,CCREG);
5023 assert(cc==HOST_CCREG);
5025 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5026 //do_cc(i,branch_regs[i].regmap,&adj,unconditional?ba[i]:-1,unconditional);
5027 assem_debug("cycle count (adj)\n");
5029 do_cc(i,branch_regs[i].regmap,&adj,ba[i],TAKEN,0);
5030 if(i!=(ba[i]-start)>>2 || source[i+1]!=0) {
5031 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5032 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5034 assem_debug("branch: internal\n");
5036 assem_debug("branch: external\n");
5037 if(internal&&is_ds[(ba[i]-start)>>2]) {
5038 ds_assemble_entry(i);
5041 add_to_linker(out,ba[i],internal);
5044 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5045 if(((u_int)out)&7) emit_addnop(0);
5049 else if(nevertaken) {
5050 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5053 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5056 void *nottaken = NULL;
5057 do_cc(i,branch_regs[i].regmap,&adj,-1,0,invert);
5058 if(adj&&!invert) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5061 if((opcode2[i]&0xf)==0) // BLTZ/BLTZAL
5068 add_to_linker(out,ba[i],internal);
5072 if((opcode2[i]&0xf)==1) // BGEZ/BLTZAL
5079 add_to_linker(out,ba[i],internal);
5086 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
5087 if(match&&(!internal||!is_ds[(ba[i]-start)>>2])) {
5089 emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5090 add_to_linker(out,ba[i],internal);
5093 add_to_linker(out,ba[i],internal*2);
5099 if(adj) emit_addimm(cc,-CLOCK_ADJUST(adj),cc);
5100 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5101 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5103 assem_debug("branch: internal\n");
5105 assem_debug("branch: external\n");
5106 if(internal&&is_ds[(ba[i]-start)>>2]) {
5107 ds_assemble_entry(i);
5110 add_to_linker(out,ba[i],internal);
5114 set_jump_target(nottaken, out);
5118 if(!invert) emit_addimm(cc,CLOCK_ADJUST(adj),cc);
5120 } // (!unconditional)
5124 // In-order execution (branch first)
5126 void *nottaken = NULL;
5128 int rt,return_address;
5129 rt=get_reg(branch_regs[i].regmap,31);
5131 // Save the PC even if the branch is not taken
5132 return_address=start+i*4+8;
5133 emit_movimm(return_address,rt); // PC into link register
5135 emit_prefetch(hash_table_get(return_address));
5139 if(!unconditional) {
5140 //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]);
5142 if((opcode2[i]&0x0d)==0) // BLTZ/BLTZL/BLTZAL/BLTZALL
5148 if((opcode2[i]&0x0d)==1) // BGEZ/BGEZL/BGEZAL/BGEZALL
5154 } // if(!unconditional)
5156 uint64_t ds_unneeded=branch_regs[i].u;
5157 ds_unneeded&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5161 //assem_debug("1:\n");
5162 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5164 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i+1],rs2[i+1]);
5165 address_generation(i+1,&branch_regs[i],0);
5166 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,INVCP);
5167 ds_assemble(i+1,&branch_regs[i]);
5168 cc=get_reg(branch_regs[i].regmap,CCREG);
5170 emit_loadreg(CCREG,cc=HOST_CCREG);
5171 // CHECK: Is the following instruction (fall thru) allocated ok?
5173 assert(cc==HOST_CCREG);
5174 store_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5175 do_cc(i,i_regmap,&adj,ba[i],TAKEN,0);
5176 assem_debug("cycle count (adj)\n");
5177 if(adj) emit_addimm(cc,CLOCK_ADJUST(ccadj[i]+2-adj),cc);
5178 load_regs_bt(branch_regs[i].regmap,branch_regs[i].dirty,ba[i]);
5180 assem_debug("branch: internal\n");
5182 assem_debug("branch: external\n");
5183 if(internal&&is_ds[(ba[i]-start)>>2]) {
5184 ds_assemble_entry(i);
5187 add_to_linker(out,ba[i],internal);
5192 if(!unconditional) {
5193 set_jump_target(nottaken, out);
5194 assem_debug("1:\n");
5196 wb_invalidate(regs[i].regmap,branch_regs[i].regmap,regs[i].dirty,ds_unneeded);
5197 load_regs(regs[i].regmap,branch_regs[i].regmap,rs1[i+1],rs2[i+1]);
5198 address_generation(i+1,&branch_regs[i],0);
5199 load_regs(regs[i].regmap,branch_regs[i].regmap,CCREG,CCREG);
5200 ds_assemble(i+1,&branch_regs[i]);
5202 cc=get_reg(branch_regs[i].regmap,CCREG);
5203 if(cc==-1&&!likely[i]) {
5204 // Cycle count isn't in a register, temporarily load it then write it out
5205 emit_loadreg(CCREG,HOST_CCREG);
5206 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5209 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,NOTTAKEN,0);
5210 emit_storereg(CCREG,HOST_CCREG);
5213 cc=get_reg(i_regmap,CCREG);
5214 assert(cc==HOST_CCREG);
5215 emit_addimm_and_set_flags(CLOCK_ADJUST(ccadj[i]+2),cc);
5218 add_stub(CC_STUB,jaddr,out,0,i,start+i*4+8,likely[i]?NULLDS:NOTTAKEN,0);
5224 static void pagespan_assemble(int i,struct regstat *i_regs)
5226 int s1l=get_reg(i_regs->regmap,rs1[i]);
5227 int s2l=get_reg(i_regs->regmap,rs2[i]);
5229 void *nottaken = NULL;
5230 int unconditional=0;
5241 int addr=-1,alt=-1,ntaddr=-1;
5242 if(i_regs->regmap[HOST_BTREG]<0) {addr=HOST_BTREG;}
5246 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG &&
5247 (i_regs->regmap[hr]&63)!=rs1[i] &&
5248 (i_regs->regmap[hr]&63)!=rs2[i] )
5257 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5258 (i_regs->regmap[hr]&63)!=rs1[i] &&
5259 (i_regs->regmap[hr]&63)!=rs2[i] )
5265 if((opcode[i]&0x2E)==6) // BLEZ/BGTZ needs another register
5269 if(hr!=EXCLUDE_REG && hr!=HOST_CCREG && hr!=HOST_BTREG &&
5270 (i_regs->regmap[hr]&63)!=rs1[i] &&
5271 (i_regs->regmap[hr]&63)!=rs2[i] )
5278 assert(hr<HOST_REGS);
5279 if((opcode[i]&0x2e)==4||opcode[i]==0x11) { // BEQ/BNE/BEQL/BNEL/BC1
5280 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
5282 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i]+2),HOST_CCREG);
5283 if(opcode[i]==2) // J
5287 if(opcode[i]==3) // JAL
5290 int rt=get_reg(i_regs->regmap,31);
5291 emit_movimm(start+i*4+8,rt);
5294 if(opcode[i]==0&&(opcode2[i]&0x3E)==8) // JR/JALR
5297 if(opcode2[i]==9) // JALR
5299 int rt=get_reg(i_regs->regmap,rt1[i]);
5300 emit_movimm(start+i*4+8,rt);
5303 if((opcode[i]&0x3f)==4) // BEQ
5310 #ifdef HAVE_CMOV_IMM
5312 if(s2l>=0) emit_cmp(s1l,s2l);
5313 else emit_test(s1l,s1l);
5314 emit_cmov2imm_e_ne_compact(ba[i],start+i*4+8,addr);
5320 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5321 if(s2l>=0) emit_cmp(s1l,s2l);
5322 else emit_test(s1l,s1l);
5323 emit_cmovne_reg(alt,addr);
5326 if((opcode[i]&0x3f)==5) // BNE
5328 #ifdef HAVE_CMOV_IMM
5329 if(s2l>=0) emit_cmp(s1l,s2l);
5330 else emit_test(s1l,s1l);
5331 emit_cmov2imm_e_ne_compact(start+i*4+8,ba[i],addr);
5334 emit_mov2imm_compact(start+i*4+8,addr,ba[i],alt);
5335 if(s2l>=0) emit_cmp(s1l,s2l);
5336 else emit_test(s1l,s1l);
5337 emit_cmovne_reg(alt,addr);
5340 if((opcode[i]&0x3f)==0x14) // BEQL
5342 if(s2l>=0) emit_cmp(s1l,s2l);
5343 else emit_test(s1l,s1l);
5344 if(nottaken) set_jump_target(nottaken, out);
5348 if((opcode[i]&0x3f)==0x15) // BNEL
5350 if(s2l>=0) emit_cmp(s1l,s2l);
5351 else emit_test(s1l,s1l);
5354 if(taken) set_jump_target(taken, out);
5356 if((opcode[i]&0x3f)==6) // BLEZ
5358 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5360 emit_cmovl_reg(alt,addr);
5362 if((opcode[i]&0x3f)==7) // BGTZ
5364 emit_mov2imm_compact(ba[i],addr,start+i*4+8,ntaddr);
5366 emit_cmovl_reg(ntaddr,addr);
5368 if((opcode[i]&0x3f)==0x16) // BLEZL
5370 assert((opcode[i]&0x3f)!=0x16);
5372 if((opcode[i]&0x3f)==0x17) // BGTZL
5374 assert((opcode[i]&0x3f)!=0x17);
5376 assert(opcode[i]!=1); // BLTZ/BGEZ
5378 //FIXME: Check CSREG
5379 if(opcode[i]==0x11 && opcode2[i]==0x08 ) {
5380 if((source[i]&0x30000)==0) // BC1F
5382 emit_mov2imm_compact(ba[i],addr,start+i*4+8,alt);
5383 emit_testimm(s1l,0x800000);
5384 emit_cmovne_reg(alt,addr);
5386 if((source[i]&0x30000)==0x10000) // BC1T
5388 emit_mov2imm_compact(ba[i],alt,start+i*4+8,addr);
5389 emit_testimm(s1l,0x800000);
5390 emit_cmovne_reg(alt,addr);
5392 if((source[i]&0x30000)==0x20000) // BC1FL
5394 emit_testimm(s1l,0x800000);
5398 if((source[i]&0x30000)==0x30000) // BC1TL
5400 emit_testimm(s1l,0x800000);
5406 assert(i_regs->regmap[HOST_CCREG]==CCREG);
5407 wb_dirtys(regs[i].regmap,regs[i].dirty);
5408 if(likely[i]||unconditional)
5410 emit_movimm(ba[i],HOST_BTREG);
5412 else if(addr!=HOST_BTREG)
5414 emit_mov(addr,HOST_BTREG);
5416 void *branch_addr=out;
5418 int target_addr=start+i*4+5;
5420 void *compiled_target_addr=check_addr(target_addr);
5421 emit_extjump_ds(branch_addr, target_addr);
5422 if(compiled_target_addr) {
5423 set_jump_target(branch_addr, compiled_target_addr);
5424 add_link(target_addr,stub);
5426 else set_jump_target(branch_addr, stub);
5429 set_jump_target(nottaken, out);
5430 wb_dirtys(regs[i].regmap,regs[i].dirty);
5431 void *branch_addr=out;
5433 int target_addr=start+i*4+8;
5435 void *compiled_target_addr=check_addr(target_addr);
5436 emit_extjump_ds(branch_addr, target_addr);
5437 if(compiled_target_addr) {
5438 set_jump_target(branch_addr, compiled_target_addr);
5439 add_link(target_addr,stub);
5441 else set_jump_target(branch_addr, stub);
5445 // Assemble the delay slot for the above
5446 static void pagespan_ds()
5448 assem_debug("initial delay slot:\n");
5449 u_int vaddr=start+1;
5450 u_int page=get_page(vaddr);
5451 u_int vpage=get_vpage(vaddr);
5452 ll_add(jump_dirty+vpage,vaddr,(void *)out);
5454 ll_add(jump_in+page,vaddr,(void *)out);
5455 assert(regs[0].regmap_entry[HOST_CCREG]==CCREG);
5456 if(regs[0].regmap[HOST_CCREG]!=CCREG)
5457 wb_register(CCREG,regs[0].regmap_entry,regs[0].wasdirty);
5458 if(regs[0].regmap[HOST_BTREG]!=BTREG)
5459 emit_writeword(HOST_BTREG,&branch_target);
5460 load_regs(regs[0].regmap_entry,regs[0].regmap,rs1[0],rs2[0]);
5461 address_generation(0,®s[0],regs[0].regmap_entry);
5462 if(itype[0]==STORE||itype[0]==STORELR||(opcode[0]&0x3b)==0x39||(opcode[0]&0x3b)==0x3a)
5463 load_regs(regs[0].regmap_entry,regs[0].regmap,INVCP,INVCP);
5467 alu_assemble(0,®s[0]);break;
5469 imm16_assemble(0,®s[0]);break;
5471 shift_assemble(0,®s[0]);break;
5473 shiftimm_assemble(0,®s[0]);break;
5475 load_assemble(0,®s[0]);break;
5477 loadlr_assemble(0,®s[0]);break;
5479 store_assemble(0,®s[0]);break;
5481 storelr_assemble(0,®s[0]);break;
5483 cop0_assemble(0,®s[0]);break;
5485 cop1_assemble(0,®s[0]);break;
5487 c1ls_assemble(0,®s[0]);break;
5489 cop2_assemble(0,®s[0]);break;
5491 c2ls_assemble(0,®s[0]);break;
5493 c2op_assemble(0,®s[0]);break;
5495 multdiv_assemble(0,®s[0]);break;
5497 mov_assemble(0,®s[0]);break;
5506 SysPrintf("Jump in the delay slot. This is probably a bug.\n");
5508 int btaddr=get_reg(regs[0].regmap,BTREG);
5510 btaddr=get_reg(regs[0].regmap,-1);
5511 emit_readword(&branch_target,btaddr);
5513 assert(btaddr!=HOST_CCREG);
5514 if(regs[0].regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG);
5516 emit_movimm(start+4,HOST_TEMPREG);
5517 emit_cmp(btaddr,HOST_TEMPREG);
5519 emit_cmpimm(btaddr,start+4);
5523 store_regs_bt(regs[0].regmap,regs[0].dirty,-1);
5524 emit_jmp(jump_vaddr_reg[btaddr]);
5525 set_jump_target(branch, out);
5526 store_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
5527 load_regs_bt(regs[0].regmap,regs[0].dirty,start+4);
5530 // Basic liveness analysis for MIPS registers
5531 void unneeded_registers(int istart,int iend,int r)
5534 uint64_t u,gte_u,b,gte_b;
5535 uint64_t temp_u,temp_gte_u=0;
5536 uint64_t gte_u_unknown=0;
5537 if(new_dynarec_hacks&NDHACK_GTE_UNNEEDED)
5541 gte_u=gte_u_unknown;
5543 //u=unneeded_reg[iend+1];
5545 gte_u=gte_unneeded[iend+1];
5548 for (i=iend;i>=istart;i--)
5550 //printf("unneeded registers i=%d (%d,%d) r=%d\n",i,istart,iend,r);
5551 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
5553 // If subroutine call, flag return address as a possible branch target
5554 if(rt1[i]==31 && i<slen-2) bt[i+2]=1;
5556 if(ba[i]<start || ba[i]>=(start+slen*4))
5558 // Branch out of this block, flush all regs
5560 gte_u=gte_u_unknown;
5561 branch_unneeded_reg[i]=u;
5562 // Merge in delay slot
5563 u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
5564 u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5567 gte_u&=~gte_rs[i+1];
5568 // If branch is "likely" (and conditional)
5569 // then we skip the delay slot on the fall-thru path
5572 u&=unneeded_reg[i+2];
5573 gte_u&=gte_unneeded[i+2];
5578 gte_u=gte_u_unknown;
5584 // Internal branch, flag target
5585 bt[(ba[i]-start)>>2]=1;
5586 if(ba[i]<=start+i*4) {
5588 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
5590 // Unconditional branch
5594 // Conditional branch (not taken case)
5595 temp_u=unneeded_reg[i+2];
5596 temp_gte_u&=gte_unneeded[i+2];
5598 // Merge in delay slot
5599 temp_u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
5600 temp_u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5602 temp_gte_u|=gte_rt[i+1];
5603 temp_gte_u&=~gte_rs[i+1];
5604 // If branch is "likely" (and conditional)
5605 // then we skip the delay slot on the fall-thru path
5608 temp_u&=unneeded_reg[i+2];
5609 temp_gte_u&=gte_unneeded[i+2];
5614 temp_gte_u=gte_u_unknown;
5617 temp_u|=(1LL<<rt1[i])|(1LL<<rt2[i]);
5618 temp_u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
5620 temp_gte_u|=gte_rt[i];
5621 temp_gte_u&=~gte_rs[i];
5622 unneeded_reg[i]=temp_u;
5623 gte_unneeded[i]=temp_gte_u;
5624 // Only go three levels deep. This recursion can take an
5625 // excessive amount of time if there are a lot of nested loops.
5627 unneeded_registers((ba[i]-start)>>2,i-1,r+1);
5629 unneeded_reg[(ba[i]-start)>>2]=1;
5630 gte_unneeded[(ba[i]-start)>>2]=gte_u_unknown;
5633 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
5635 // Unconditional branch
5636 u=unneeded_reg[(ba[i]-start)>>2];
5637 gte_u=gte_unneeded[(ba[i]-start)>>2];
5638 branch_unneeded_reg[i]=u;
5639 // Merge in delay slot
5640 u|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
5641 u&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5644 gte_u&=~gte_rs[i+1];
5646 // Conditional branch
5647 b=unneeded_reg[(ba[i]-start)>>2];
5648 gte_b=gte_unneeded[(ba[i]-start)>>2];
5649 branch_unneeded_reg[i]=b;
5650 // Branch delay slot
5651 b|=(1LL<<rt1[i+1])|(1LL<<rt2[i+1]);
5652 b&=~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
5655 gte_b&=~gte_rs[i+1];
5656 // If branch is "likely" then we skip the
5657 // delay slot on the fall-thru path
5662 u&=unneeded_reg[i+2];
5663 gte_u&=gte_unneeded[i+2];
5670 branch_unneeded_reg[i]&=unneeded_reg[i+2];
5672 branch_unneeded_reg[i]=1;
5678 else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
5680 // SYSCALL instruction (software interrupt)
5683 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
5685 // ERET instruction (return from interrupt)
5689 // Written registers are unneeded
5693 // Accessed registers are needed
5697 if(gte_rs[i]&&rt1[i]&&(unneeded_reg[i+1]&(1ll<<rt1[i])))
5698 gte_u|=gte_rs[i]>e_unneeded[i+1]; // MFC2/CFC2 to dead register, unneeded
5699 // Source-target dependencies
5700 // R0 is always unneeded
5704 gte_unneeded[i]=gte_u;
5706 printf("ur (%d,%d) %x: ",istart,iend,start+i*4);
5709 for(r=1;r<=CCREG;r++) {
5710 if((unneeded_reg[i]>>r)&1) {
5711 if(r==HIREG) printf(" HI");
5712 else if(r==LOREG) printf(" LO");
5713 else printf(" r%d",r);
5721 // Write back dirty registers as soon as we will no longer modify them,
5722 // so that we don't end up with lots of writes at the branches.
5723 void clean_registers(int istart,int iend,int wr)
5727 u_int will_dirty_i,will_dirty_next,temp_will_dirty;
5728 u_int wont_dirty_i,wont_dirty_next,temp_wont_dirty;
5730 will_dirty_i=will_dirty_next=0;
5731 wont_dirty_i=wont_dirty_next=0;
5733 will_dirty_i=will_dirty_next=will_dirty[iend+1];
5734 wont_dirty_i=wont_dirty_next=wont_dirty[iend+1];
5736 for (i=iend;i>=istart;i--)
5738 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
5740 if(ba[i]<start || ba[i]>=(start+slen*4))
5742 // Branch out of this block, flush all regs
5743 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
5745 // Unconditional branch
5748 // Merge in delay slot (will dirty)
5749 for(r=0;r<HOST_REGS;r++) {
5750 if(r!=EXCLUDE_REG) {
5751 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5752 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5753 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5754 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5755 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5756 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
5757 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5758 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5759 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5760 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5761 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5762 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5763 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
5764 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5770 // Conditional branch
5772 wont_dirty_i=wont_dirty_next;
5773 // Merge in delay slot (will dirty)
5774 for(r=0;r<HOST_REGS;r++) {
5775 if(r!=EXCLUDE_REG) {
5777 // Might not dirty if likely branch is not taken
5778 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5779 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5780 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5781 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5782 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5783 if(branch_regs[i].regmap[r]==0) will_dirty_i&=~(1<<r);
5784 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5785 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5786 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5787 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5788 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5789 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5790 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
5791 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5796 // Merge in delay slot (wont dirty)
5797 for(r=0;r<HOST_REGS;r++) {
5798 if(r!=EXCLUDE_REG) {
5799 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
5800 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
5801 if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
5802 if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
5803 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
5804 if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
5805 if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
5806 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
5807 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
5808 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
5812 #ifndef DESTRUCTIVE_WRITEBACK
5813 branch_regs[i].dirty&=wont_dirty_i;
5815 branch_regs[i].dirty|=will_dirty_i;
5821 if(ba[i]<=start+i*4) {
5823 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
5825 // Unconditional branch
5828 // Merge in delay slot (will dirty)
5829 for(r=0;r<HOST_REGS;r++) {
5830 if(r!=EXCLUDE_REG) {
5831 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
5832 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
5833 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
5834 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
5835 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
5836 if(branch_regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
5837 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
5838 if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
5839 if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
5840 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
5841 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
5842 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
5843 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
5844 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
5848 // Conditional branch (not taken case)
5849 temp_will_dirty=will_dirty_next;
5850 temp_wont_dirty=wont_dirty_next;
5851 // Merge in delay slot (will dirty)
5852 for(r=0;r<HOST_REGS;r++) {
5853 if(r!=EXCLUDE_REG) {
5855 // Will not dirty if likely branch is not taken
5856 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
5857 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
5858 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
5859 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
5860 if((branch_regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
5861 if(branch_regs[i].regmap[r]==0) temp_will_dirty&=~(1<<r);
5862 if(branch_regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
5863 //if((regs[i].regmap[r]&63)==rt1[i]) temp_will_dirty|=1<<r;
5864 //if((regs[i].regmap[r]&63)==rt2[i]) temp_will_dirty|=1<<r;
5865 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_will_dirty|=1<<r;
5866 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_will_dirty|=1<<r;
5867 if((regs[i].regmap[r]&63)>33) temp_will_dirty&=~(1<<r);
5868 if(regs[i].regmap[r]<=0) temp_will_dirty&=~(1<<r);
5869 if(regs[i].regmap[r]==CCREG) temp_will_dirty|=1<<r;
5874 // Merge in delay slot (wont dirty)
5875 for(r=0;r<HOST_REGS;r++) {
5876 if(r!=EXCLUDE_REG) {
5877 if((regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
5878 if((regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
5879 if((regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
5880 if((regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
5881 if(regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
5882 if((branch_regs[i].regmap[r]&63)==rt1[i]) temp_wont_dirty|=1<<r;
5883 if((branch_regs[i].regmap[r]&63)==rt2[i]) temp_wont_dirty|=1<<r;
5884 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) temp_wont_dirty|=1<<r;
5885 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) temp_wont_dirty|=1<<r;
5886 if(branch_regs[i].regmap[r]==CCREG) temp_wont_dirty|=1<<r;
5889 // Deal with changed mappings
5891 for(r=0;r<HOST_REGS;r++) {
5892 if(r!=EXCLUDE_REG) {
5893 if(regs[i].regmap[r]!=regmap_pre[i][r]) {
5894 temp_will_dirty&=~(1<<r);
5895 temp_wont_dirty&=~(1<<r);
5896 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
5897 temp_will_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
5898 temp_wont_dirty|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
5900 temp_will_dirty|=1<<r;
5901 temp_wont_dirty|=1<<r;
5908 will_dirty[i]=temp_will_dirty;
5909 wont_dirty[i]=temp_wont_dirty;
5910 clean_registers((ba[i]-start)>>2,i-1,0);
5912 // Limit recursion. It can take an excessive amount
5913 // of time if there are a lot of nested loops.
5914 will_dirty[(ba[i]-start)>>2]=0;
5915 wont_dirty[(ba[i]-start)>>2]=-1;
5920 if(itype[i]==RJUMP||itype[i]==UJUMP||(source[i]>>16)==0x1000)
5922 // Unconditional branch
5925 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
5926 for(r=0;r<HOST_REGS;r++) {
5927 if(r!=EXCLUDE_REG) {
5928 if(branch_regs[i].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
5929 will_dirty_i|=will_dirty[(ba[i]-start)>>2]&(1<<r);
5930 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
5932 if(branch_regs[i].regmap[r]>=0) {
5933 will_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
5934 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(branch_regs[i].regmap[r]&63))&1)<<r;
5939 // Merge in delay slot
5940 for(r=0;r<HOST_REGS;r++) {
5941 if(r!=EXCLUDE_REG) {
5942 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5943 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5944 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5945 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5946 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5947 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
5948 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5949 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5950 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5951 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5952 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5953 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5954 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
5955 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5959 // Conditional branch
5960 will_dirty_i=will_dirty_next;
5961 wont_dirty_i=wont_dirty_next;
5962 //if(ba[i]>start+i*4) { // Disable recursion (for debugging)
5963 for(r=0;r<HOST_REGS;r++) {
5964 if(r!=EXCLUDE_REG) {
5965 signed char target_reg=branch_regs[i].regmap[r];
5966 if(target_reg==regs[(ba[i]-start)>>2].regmap_entry[r]) {
5967 will_dirty_i&=will_dirty[(ba[i]-start)>>2]&(1<<r);
5968 wont_dirty_i|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
5970 else if(target_reg>=0) {
5971 will_dirty_i&=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
5972 wont_dirty_i|=((unneeded_reg[(ba[i]-start)>>2]>>(target_reg&63))&1)<<r;
5974 // Treat delay slot as part of branch too
5975 /*if(regs[i+1].regmap[r]==regs[(ba[i]-start)>>2].regmap_entry[r]) {
5976 will_dirty[i+1]&=will_dirty[(ba[i]-start)>>2]&(1<<r);
5977 wont_dirty[i+1]|=wont_dirty[(ba[i]-start)>>2]&(1<<r);
5981 will_dirty[i+1]&=~(1<<r);
5986 // Merge in delay slot
5987 for(r=0;r<HOST_REGS;r++) {
5988 if(r!=EXCLUDE_REG) {
5990 // Might not dirty if likely branch is not taken
5991 if((branch_regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5992 if((branch_regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
5993 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
5994 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
5995 if((branch_regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
5996 if(branch_regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
5997 if(branch_regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
5998 //if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
5999 //if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6000 if((regs[i].regmap[r]&63)==rt1[i+1]) will_dirty_i|=1<<r;
6001 if((regs[i].regmap[r]&63)==rt2[i+1]) will_dirty_i|=1<<r;
6002 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6003 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6004 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6009 // Merge in delay slot (won't dirty)
6010 for(r=0;r<HOST_REGS;r++) {
6011 if(r!=EXCLUDE_REG) {
6012 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6013 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6014 if((regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
6015 if((regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
6016 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6017 if((branch_regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6018 if((branch_regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6019 if((branch_regs[i].regmap[r]&63)==rt1[i+1]) wont_dirty_i|=1<<r;
6020 if((branch_regs[i].regmap[r]&63)==rt2[i+1]) wont_dirty_i|=1<<r;
6021 if(branch_regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6025 #ifndef DESTRUCTIVE_WRITEBACK
6026 branch_regs[i].dirty&=wont_dirty_i;
6028 branch_regs[i].dirty|=will_dirty_i;
6033 else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
6035 // SYSCALL instruction (software interrupt)
6039 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
6041 // ERET instruction (return from interrupt)
6045 will_dirty_next=will_dirty_i;
6046 wont_dirty_next=wont_dirty_i;
6047 for(r=0;r<HOST_REGS;r++) {
6048 if(r!=EXCLUDE_REG) {
6049 if((regs[i].regmap[r]&63)==rt1[i]) will_dirty_i|=1<<r;
6050 if((regs[i].regmap[r]&63)==rt2[i]) will_dirty_i|=1<<r;
6051 if((regs[i].regmap[r]&63)>33) will_dirty_i&=~(1<<r);
6052 if(regs[i].regmap[r]<=0) will_dirty_i&=~(1<<r);
6053 if(regs[i].regmap[r]==CCREG) will_dirty_i|=1<<r;
6054 if((regs[i].regmap[r]&63)==rt1[i]) wont_dirty_i|=1<<r;
6055 if((regs[i].regmap[r]&63)==rt2[i]) wont_dirty_i|=1<<r;
6056 if(regs[i].regmap[r]==CCREG) wont_dirty_i|=1<<r;
6058 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP)
6060 // Don't store a register immediately after writing it,
6061 // may prevent dual-issue.
6062 if((regs[i].regmap[r]&63)==rt1[i-1]) wont_dirty_i|=1<<r;
6063 if((regs[i].regmap[r]&63)==rt2[i-1]) wont_dirty_i|=1<<r;
6069 will_dirty[i]=will_dirty_i;
6070 wont_dirty[i]=wont_dirty_i;
6071 // Mark registers that won't be dirtied as not dirty
6073 /*printf("wr (%d,%d) %x will:",istart,iend,start+i*4);
6074 for(r=0;r<HOST_REGS;r++) {
6075 if((will_dirty_i>>r)&1) {
6081 //if(i==istart||(itype[i-1]!=RJUMP&&itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP)) {
6082 regs[i].dirty|=will_dirty_i;
6083 #ifndef DESTRUCTIVE_WRITEBACK
6084 regs[i].dirty&=wont_dirty_i;
6085 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
6087 if(i<iend-1&&itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
6088 for(r=0;r<HOST_REGS;r++) {
6089 if(r!=EXCLUDE_REG) {
6090 if(regs[i].regmap[r]==regmap_pre[i+2][r]) {
6091 regs[i+2].wasdirty&=wont_dirty_i|~(1<<r);
6092 }else {/*printf("i: %x (%d) mismatch(+2): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6100 for(r=0;r<HOST_REGS;r++) {
6101 if(r!=EXCLUDE_REG) {
6102 if(regs[i].regmap[r]==regmap_pre[i+1][r]) {
6103 regs[i+1].wasdirty&=wont_dirty_i|~(1<<r);
6104 }else {/*printf("i: %x (%d) mismatch(+1): %d\n",start+i*4,i,r);assert(!((wont_dirty_i>>r)&1));*/}
6112 // Deal with changed mappings
6113 temp_will_dirty=will_dirty_i;
6114 temp_wont_dirty=wont_dirty_i;
6115 for(r=0;r<HOST_REGS;r++) {
6116 if(r!=EXCLUDE_REG) {
6118 if(regs[i].regmap[r]==regmap_pre[i][r]) {
6120 #ifndef DESTRUCTIVE_WRITEBACK
6121 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6123 regs[i].wasdirty|=will_dirty_i&(1<<r);
6126 else if(regmap_pre[i][r]>=0&&(nr=get_reg(regs[i].regmap,regmap_pre[i][r]))>=0) {
6127 // Register moved to a different register
6128 will_dirty_i&=~(1<<r);
6129 wont_dirty_i&=~(1<<r);
6130 will_dirty_i|=((temp_will_dirty>>nr)&1)<<r;
6131 wont_dirty_i|=((temp_wont_dirty>>nr)&1)<<r;
6133 #ifndef DESTRUCTIVE_WRITEBACK
6134 regs[i].wasdirty&=wont_dirty_i|~(1<<r);
6136 regs[i].wasdirty|=will_dirty_i&(1<<r);
6140 will_dirty_i&=~(1<<r);
6141 wont_dirty_i&=~(1<<r);
6142 if((regmap_pre[i][r]&63)>0 && (regmap_pre[i][r]&63)<34) {
6143 will_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6144 wont_dirty_i|=((unneeded_reg[i]>>(regmap_pre[i][r]&63))&1)<<r;
6147 /*printf("i: %x (%d) mismatch: %d\n",start+i*4,i,r);assert(!((will_dirty>>r)&1));*/
6157 void disassemble_inst(int i)
6159 if (bt[i]) printf("*"); else printf(" ");
6162 printf (" %x: %s %8x\n",start+i*4,insn[i],ba[i]);break;
6164 printf (" %x: %s r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],i?start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14):*ba);break;
6166 printf (" %x: %s r%d,%8x\n",start+i*4,insn[i],rs1[i],start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14));break;
6168 if (opcode[i]==0x9&&rt1[i]!=31)
6169 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i]);
6171 printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
6174 printf (" %x: %s (pagespan) r%d,r%d,%8x\n",start+i*4,insn[i],rs1[i],rs2[i],ba[i]);break;
6176 if(opcode[i]==0xf) //LUI
6177 printf (" %x: %s r%d,%4x0000\n",start+i*4,insn[i],rt1[i],imm[i]&0xffff);
6179 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
6183 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
6187 printf (" %x: %s r%d,r%d+%x\n",start+i*4,insn[i],rs2[i],rs1[i],imm[i]);
6191 printf (" %x: %s r%d,r%d,r%d\n",start+i*4,insn[i],rt1[i],rs1[i],rs2[i]);
6194 printf (" %x: %s r%d,r%d\n",start+i*4,insn[i],rs1[i],rs2[i]);
6197 printf (" %x: %s r%d,r%d,%d\n",start+i*4,insn[i],rt1[i],rs1[i],imm[i]);
6200 if((opcode2[i]&0x1d)==0x10)
6201 printf (" %x: %s r%d\n",start+i*4,insn[i],rt1[i]);
6202 else if((opcode2[i]&0x1d)==0x11)
6203 printf (" %x: %s r%d\n",start+i*4,insn[i],rs1[i]);
6205 printf (" %x: %s\n",start+i*4,insn[i]);
6209 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC0
6210 else if(opcode2[i]==4)
6211 printf (" %x: %s r%d,cpr0[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC0
6212 else printf (" %x: %s\n",start+i*4,insn[i]);
6216 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC1
6217 else if(opcode2[i]>3)
6218 printf (" %x: %s r%d,cpr1[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC1
6219 else printf (" %x: %s\n",start+i*4,insn[i]);
6223 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rt1[i],(source[i]>>11)&0x1f); // MFC2
6224 else if(opcode2[i]>3)
6225 printf (" %x: %s r%d,cpr2[%d]\n",start+i*4,insn[i],rs1[i],(source[i]>>11)&0x1f); // MTC2
6226 else printf (" %x: %s\n",start+i*4,insn[i]);
6229 printf (" %x: %s cpr1[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
6232 printf (" %x: %s cpr2[%d],r%d+%x\n",start+i*4,insn[i],(source[i]>>16)&0x1f,rs1[i],imm[i]);
6235 printf (" %x: %s (INTCALL)\n",start+i*4,insn[i]);
6238 //printf (" %s %8x\n",insn[i],source[i]);
6239 printf (" %x: %s\n",start+i*4,insn[i]);
6243 static void disassemble_inst(int i) {}
6246 #define DRC_TEST_VAL 0x74657374
6248 static void new_dynarec_test(void)
6250 int (*testfunc)(void);
6255 // check structure linkage
6256 if ((void *)reg != (void *)&psxRegs
6257 || (u_char *)rcnts - (u_char *)reg != sizeof(psxRegs))
6259 SysPrintf("linkage_arm miscompilation/breakage detected.\n");
6262 SysPrintf("testing if we can run recompiled code...\n");
6263 ((volatile u_int *)out)[0]++; // make cache dirty
6265 for (i = 0; i < ARRAY_SIZE(ret); i++) {
6266 out = translation_cache;
6267 beginning = start_block();
6268 emit_movimm(DRC_TEST_VAL + i, 0); // test
6271 end_block(beginning);
6272 testfunc = beginning;
6273 ret[i] = testfunc();
6276 if (ret[0] == DRC_TEST_VAL && ret[1] == DRC_TEST_VAL + 1)
6277 SysPrintf("test passed.\n");
6279 SysPrintf("test failed, will likely crash soon (r=%08x %08x)\n", ret[0], ret[1]);
6280 out = translation_cache;
6283 // clear the state completely, instead of just marking
6284 // things invalid like invalidate_all_pages() does
6285 void new_dynarec_clear_full()
6288 out = translation_cache;
6289 memset(invalid_code,1,sizeof(invalid_code));
6290 memset(hash_table,0xff,sizeof(hash_table));
6291 memset(mini_ht,-1,sizeof(mini_ht));
6292 memset(restore_candidate,0,sizeof(restore_candidate));
6293 memset(shadow,0,sizeof(shadow));
6295 expirep=16384; // Expiry pointer, +2 blocks
6296 pending_exception=0;
6299 inv_code_start=inv_code_end=~0;
6301 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6302 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6303 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6306 void new_dynarec_init()
6308 SysPrintf("Init new dynarec\n");
6310 // allocate/prepare a buffer for translation cache
6311 // see assem_arm.h for some explanation
6312 #if defined(BASE_ADDR_FIXED)
6313 if (mmap(translation_cache, 1 << TARGET_SIZE_2,
6314 PROT_READ | PROT_WRITE | PROT_EXEC,
6315 MAP_PRIVATE | MAP_ANONYMOUS,
6316 -1, 0) != translation_cache) {
6317 SysPrintf("mmap() failed: %s\n", strerror(errno));
6318 SysPrintf("disable BASE_ADDR_FIXED and recompile\n");
6321 #elif defined(BASE_ADDR_DYNAMIC)
6323 sceBlock = sceKernelAllocMemBlockForVM("code", 1 << TARGET_SIZE_2);
6325 SysPrintf("sceKernelAllocMemBlockForVM failed\n");
6326 int ret = sceKernelGetMemBlockBase(sceBlock, (void **)&translation_cache);
6328 SysPrintf("sceKernelGetMemBlockBase failed\n");
6330 translation_cache = mmap (NULL, 1 << TARGET_SIZE_2,
6331 PROT_READ | PROT_WRITE | PROT_EXEC,
6332 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
6333 if (translation_cache == MAP_FAILED) {
6334 SysPrintf("mmap() failed: %s\n", strerror(errno));
6339 #ifndef NO_WRITE_EXEC
6340 // not all systems allow execute in data segment by default
6341 if (mprotect(translation_cache, 1<<TARGET_SIZE_2, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
6342 SysPrintf("mprotect() failed: %s\n", strerror(errno));
6345 out = translation_cache;
6346 cycle_multiplier=200;
6347 new_dynarec_clear_full();
6349 // Copy this into local area so we don't have to put it in every literal pool
6350 invc_ptr=invalid_code;
6355 ram_offset=(uintptr_t)rdram-0x80000000;
6358 SysPrintf("warning: RAM is not directly mapped, performance will suffer\n");
6361 void new_dynarec_cleanup()
6364 #if defined(BASE_ADDR_FIXED) || defined(BASE_ADDR_DYNAMIC)
6366 sceKernelFreeMemBlock(sceBlock);
6369 if (munmap(translation_cache, 1<<TARGET_SIZE_2) < 0)
6370 SysPrintf("munmap() failed\n");
6373 for(n=0;n<4096;n++) ll_clear(jump_in+n);
6374 for(n=0;n<4096;n++) ll_clear(jump_out+n);
6375 for(n=0;n<4096;n++) ll_clear(jump_dirty+n);
6377 if (munmap (ROM_COPY, 67108864) < 0) {SysPrintf("munmap() failed\n");}
6381 static u_int *get_source_start(u_int addr, u_int *limit)
6383 if (addr < 0x00200000 ||
6384 (0xa0000000 <= addr && addr < 0xa0200000)) {
6385 // used for BIOS calls mostly?
6386 *limit = (addr&0xa0000000)|0x00200000;
6387 return (u_int *)(rdram + (addr&0x1fffff));
6389 else if (!Config.HLE && (
6390 /* (0x9fc00000 <= addr && addr < 0x9fc80000) ||*/
6391 (0xbfc00000 <= addr && addr < 0xbfc80000))) {
6393 *limit = (addr & 0xfff00000) | 0x80000;
6394 return (u_int *)((u_char *)psxR + (addr&0x7ffff));
6396 else if (addr >= 0x80000000 && addr < 0x80000000+RAM_SIZE) {
6397 *limit = (addr & 0x80600000) + 0x00200000;
6398 return (u_int *)(rdram + (addr&0x1fffff));
6403 static u_int scan_for_ret(u_int addr)
6408 mem = get_source_start(addr, &limit);
6412 if (limit > addr + 0x1000)
6413 limit = addr + 0x1000;
6414 for (; addr < limit; addr += 4, mem++) {
6415 if (*mem == 0x03e00008) // jr $ra
6421 struct savestate_block {
6426 static int addr_cmp(const void *p1_, const void *p2_)
6428 const struct savestate_block *p1 = p1_, *p2 = p2_;
6429 return p1->addr - p2->addr;
6432 int new_dynarec_save_blocks(void *save, int size)
6434 struct savestate_block *blocks = save;
6435 int maxcount = size / sizeof(blocks[0]);
6436 struct savestate_block tmp_blocks[1024];
6437 struct ll_entry *head;
6438 int p, s, d, o, bcnt;
6442 for (p = 0; p < ARRAY_SIZE(jump_in); p++) {
6444 for (head = jump_in[p]; head != NULL; head = head->next) {
6445 tmp_blocks[bcnt].addr = head->vaddr;
6446 tmp_blocks[bcnt].regflags = head->reg_sv_flags;
6451 qsort(tmp_blocks, bcnt, sizeof(tmp_blocks[0]), addr_cmp);
6453 addr = tmp_blocks[0].addr;
6454 for (s = d = 0; s < bcnt; s++) {
6455 if (tmp_blocks[s].addr < addr)
6457 if (d == 0 || tmp_blocks[d-1].addr != tmp_blocks[s].addr)
6458 tmp_blocks[d++] = tmp_blocks[s];
6459 addr = scan_for_ret(tmp_blocks[s].addr);
6462 if (o + d > maxcount)
6464 memcpy(&blocks[o], tmp_blocks, d * sizeof(blocks[0]));
6468 return o * sizeof(blocks[0]);
6471 void new_dynarec_load_blocks(const void *save, int size)
6473 const struct savestate_block *blocks = save;
6474 int count = size / sizeof(blocks[0]);
6475 u_int regs_save[32];
6479 get_addr(psxRegs.pc);
6481 // change GPRs for speculation to at least partially work..
6482 memcpy(regs_save, &psxRegs.GPR, sizeof(regs_save));
6483 for (i = 1; i < 32; i++)
6484 psxRegs.GPR.r[i] = 0x80000000;
6486 for (b = 0; b < count; b++) {
6487 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6489 psxRegs.GPR.r[i] = 0x1f800000;
6492 get_addr(blocks[b].addr);
6494 for (f = blocks[b].regflags, i = 0; f; f >>= 1, i++) {
6496 psxRegs.GPR.r[i] = 0x80000000;
6500 memcpy(&psxRegs.GPR, regs_save, sizeof(regs_save));
6503 int new_recompile_block(int addr)
6505 u_int pagelimit = 0;
6506 u_int state_rflags = 0;
6509 assem_debug("NOTCOMPILED: addr = %x -> %p\n", addr, out);
6510 //printf("TRACE: count=%d next=%d (compile %x)\n",Count,next_interupt,addr);
6512 //printf("fpu mapping=%x enabled=%x\n",(Status & 0x04000000)>>26,(Status & 0x20000000)>>29);
6514 // this is just for speculation
6515 for (i = 1; i < 32; i++) {
6516 if ((psxRegs.GPR.r[i] & 0xffff0000) == 0x1f800000)
6517 state_rflags |= 1 << i;
6520 start = (u_int)addr&~3;
6521 //assert(((u_int)addr&1)==0);
6522 new_dynarec_did_compile=1;
6523 if (Config.HLE && start == 0x80001000) // hlecall
6525 // XXX: is this enough? Maybe check hleSoftCall?
6526 void *beginning=start_block();
6527 u_int page=get_page(start);
6529 invalid_code[start>>12]=0;
6530 emit_movimm(start,0);
6531 emit_writeword(0,&pcaddr);
6532 emit_jmp(new_dyna_leave);
6534 end_block(beginning);
6535 ll_add_flags(jump_in+page,start,state_rflags,(void *)beginning);
6539 source = get_source_start(start, &pagelimit);
6540 if (source == NULL) {
6541 SysPrintf("Compile at bogus memory address: %08x\n", addr);
6545 /* Pass 1: disassemble */
6546 /* Pass 2: register dependencies, branch targets */
6547 /* Pass 3: register allocation */
6548 /* Pass 4: branch dependencies */
6549 /* Pass 5: pre-alloc */
6550 /* Pass 6: optimize clean/dirty state */
6551 /* Pass 7: flag 32-bit registers */
6552 /* Pass 8: assembly */
6553 /* Pass 9: linker */
6554 /* Pass 10: garbage collection / free memory */
6558 unsigned int type,op,op2;
6560 //printf("addr = %x source = %x %x\n", addr,source,source[0]);
6562 /* Pass 1 disassembly */
6564 for(i=0;!done;i++) {
6565 bt[i]=0;likely[i]=0;ooo[i]=0;op2=0;
6566 minimum_free_regs[i]=0;
6567 opcode[i]=op=source[i]>>26;
6570 case 0x00: strcpy(insn[i],"special"); type=NI;
6574 case 0x00: strcpy(insn[i],"SLL"); type=SHIFTIMM; break;
6575 case 0x02: strcpy(insn[i],"SRL"); type=SHIFTIMM; break;
6576 case 0x03: strcpy(insn[i],"SRA"); type=SHIFTIMM; break;
6577 case 0x04: strcpy(insn[i],"SLLV"); type=SHIFT; break;
6578 case 0x06: strcpy(insn[i],"SRLV"); type=SHIFT; break;
6579 case 0x07: strcpy(insn[i],"SRAV"); type=SHIFT; break;
6580 case 0x08: strcpy(insn[i],"JR"); type=RJUMP; break;
6581 case 0x09: strcpy(insn[i],"JALR"); type=RJUMP; break;
6582 case 0x0C: strcpy(insn[i],"SYSCALL"); type=SYSCALL; break;
6583 case 0x0D: strcpy(insn[i],"BREAK"); type=OTHER; break;
6584 case 0x0F: strcpy(insn[i],"SYNC"); type=OTHER; break;
6585 case 0x10: strcpy(insn[i],"MFHI"); type=MOV; break;
6586 case 0x11: strcpy(insn[i],"MTHI"); type=MOV; break;
6587 case 0x12: strcpy(insn[i],"MFLO"); type=MOV; break;
6588 case 0x13: strcpy(insn[i],"MTLO"); type=MOV; break;
6589 case 0x18: strcpy(insn[i],"MULT"); type=MULTDIV; break;
6590 case 0x19: strcpy(insn[i],"MULTU"); type=MULTDIV; break;
6591 case 0x1A: strcpy(insn[i],"DIV"); type=MULTDIV; break;
6592 case 0x1B: strcpy(insn[i],"DIVU"); type=MULTDIV; break;
6593 case 0x20: strcpy(insn[i],"ADD"); type=ALU; break;
6594 case 0x21: strcpy(insn[i],"ADDU"); type=ALU; break;
6595 case 0x22: strcpy(insn[i],"SUB"); type=ALU; break;
6596 case 0x23: strcpy(insn[i],"SUBU"); type=ALU; break;
6597 case 0x24: strcpy(insn[i],"AND"); type=ALU; break;
6598 case 0x25: strcpy(insn[i],"OR"); type=ALU; break;
6599 case 0x26: strcpy(insn[i],"XOR"); type=ALU; break;
6600 case 0x27: strcpy(insn[i],"NOR"); type=ALU; break;
6601 case 0x2A: strcpy(insn[i],"SLT"); type=ALU; break;
6602 case 0x2B: strcpy(insn[i],"SLTU"); type=ALU; break;
6603 case 0x30: strcpy(insn[i],"TGE"); type=NI; break;
6604 case 0x31: strcpy(insn[i],"TGEU"); type=NI; break;
6605 case 0x32: strcpy(insn[i],"TLT"); type=NI; break;
6606 case 0x33: strcpy(insn[i],"TLTU"); type=NI; break;
6607 case 0x34: strcpy(insn[i],"TEQ"); type=NI; break;
6608 case 0x36: strcpy(insn[i],"TNE"); type=NI; break;
6610 case 0x14: strcpy(insn[i],"DSLLV"); type=SHIFT; break;
6611 case 0x16: strcpy(insn[i],"DSRLV"); type=SHIFT; break;
6612 case 0x17: strcpy(insn[i],"DSRAV"); type=SHIFT; break;
6613 case 0x1C: strcpy(insn[i],"DMULT"); type=MULTDIV; break;
6614 case 0x1D: strcpy(insn[i],"DMULTU"); type=MULTDIV; break;
6615 case 0x1E: strcpy(insn[i],"DDIV"); type=MULTDIV; break;
6616 case 0x1F: strcpy(insn[i],"DDIVU"); type=MULTDIV; break;
6617 case 0x2C: strcpy(insn[i],"DADD"); type=ALU; break;
6618 case 0x2D: strcpy(insn[i],"DADDU"); type=ALU; break;
6619 case 0x2E: strcpy(insn[i],"DSUB"); type=ALU; break;
6620 case 0x2F: strcpy(insn[i],"DSUBU"); type=ALU; break;
6621 case 0x38: strcpy(insn[i],"DSLL"); type=SHIFTIMM; break;
6622 case 0x3A: strcpy(insn[i],"DSRL"); type=SHIFTIMM; break;
6623 case 0x3B: strcpy(insn[i],"DSRA"); type=SHIFTIMM; break;
6624 case 0x3C: strcpy(insn[i],"DSLL32"); type=SHIFTIMM; break;
6625 case 0x3E: strcpy(insn[i],"DSRL32"); type=SHIFTIMM; break;
6626 case 0x3F: strcpy(insn[i],"DSRA32"); type=SHIFTIMM; break;
6630 case 0x01: strcpy(insn[i],"regimm"); type=NI;
6631 op2=(source[i]>>16)&0x1f;
6634 case 0x00: strcpy(insn[i],"BLTZ"); type=SJUMP; break;
6635 case 0x01: strcpy(insn[i],"BGEZ"); type=SJUMP; break;
6636 case 0x02: strcpy(insn[i],"BLTZL"); type=SJUMP; break;
6637 case 0x03: strcpy(insn[i],"BGEZL"); type=SJUMP; break;
6638 case 0x08: strcpy(insn[i],"TGEI"); type=NI; break;
6639 case 0x09: strcpy(insn[i],"TGEIU"); type=NI; break;
6640 case 0x0A: strcpy(insn[i],"TLTI"); type=NI; break;
6641 case 0x0B: strcpy(insn[i],"TLTIU"); type=NI; break;
6642 case 0x0C: strcpy(insn[i],"TEQI"); type=NI; break;
6643 case 0x0E: strcpy(insn[i],"TNEI"); type=NI; break;
6644 case 0x10: strcpy(insn[i],"BLTZAL"); type=SJUMP; break;
6645 case 0x11: strcpy(insn[i],"BGEZAL"); type=SJUMP; break;
6646 case 0x12: strcpy(insn[i],"BLTZALL"); type=SJUMP; break;
6647 case 0x13: strcpy(insn[i],"BGEZALL"); type=SJUMP; break;
6650 case 0x02: strcpy(insn[i],"J"); type=UJUMP; break;
6651 case 0x03: strcpy(insn[i],"JAL"); type=UJUMP; break;
6652 case 0x04: strcpy(insn[i],"BEQ"); type=CJUMP; break;
6653 case 0x05: strcpy(insn[i],"BNE"); type=CJUMP; break;
6654 case 0x06: strcpy(insn[i],"BLEZ"); type=CJUMP; break;
6655 case 0x07: strcpy(insn[i],"BGTZ"); type=CJUMP; break;
6656 case 0x08: strcpy(insn[i],"ADDI"); type=IMM16; break;
6657 case 0x09: strcpy(insn[i],"ADDIU"); type=IMM16; break;
6658 case 0x0A: strcpy(insn[i],"SLTI"); type=IMM16; break;
6659 case 0x0B: strcpy(insn[i],"SLTIU"); type=IMM16; break;
6660 case 0x0C: strcpy(insn[i],"ANDI"); type=IMM16; break;
6661 case 0x0D: strcpy(insn[i],"ORI"); type=IMM16; break;
6662 case 0x0E: strcpy(insn[i],"XORI"); type=IMM16; break;
6663 case 0x0F: strcpy(insn[i],"LUI"); type=IMM16; break;
6664 case 0x10: strcpy(insn[i],"cop0"); type=NI;
6665 op2=(source[i]>>21)&0x1f;
6668 case 0x00: strcpy(insn[i],"MFC0"); type=COP0; break;
6669 case 0x02: strcpy(insn[i],"CFC0"); type=COP0; break;
6670 case 0x04: strcpy(insn[i],"MTC0"); type=COP0; break;
6671 case 0x06: strcpy(insn[i],"CTC0"); type=COP0; break;
6672 case 0x10: strcpy(insn[i],"RFE"); type=COP0; break;
6675 case 0x11: strcpy(insn[i],"cop1"); type=COP1;
6676 op2=(source[i]>>21)&0x1f;
6679 case 0x14: strcpy(insn[i],"BEQL"); type=CJUMP; break;
6680 case 0x15: strcpy(insn[i],"BNEL"); type=CJUMP; break;
6681 case 0x16: strcpy(insn[i],"BLEZL"); type=CJUMP; break;
6682 case 0x17: strcpy(insn[i],"BGTZL"); type=CJUMP; break;
6683 case 0x18: strcpy(insn[i],"DADDI"); type=IMM16; break;
6684 case 0x19: strcpy(insn[i],"DADDIU"); type=IMM16; break;
6685 case 0x1A: strcpy(insn[i],"LDL"); type=LOADLR; break;
6686 case 0x1B: strcpy(insn[i],"LDR"); type=LOADLR; break;
6688 case 0x20: strcpy(insn[i],"LB"); type=LOAD; break;
6689 case 0x21: strcpy(insn[i],"LH"); type=LOAD; break;
6690 case 0x22: strcpy(insn[i],"LWL"); type=LOADLR; break;
6691 case 0x23: strcpy(insn[i],"LW"); type=LOAD; break;
6692 case 0x24: strcpy(insn[i],"LBU"); type=LOAD; break;
6693 case 0x25: strcpy(insn[i],"LHU"); type=LOAD; break;
6694 case 0x26: strcpy(insn[i],"LWR"); type=LOADLR; break;
6696 case 0x27: strcpy(insn[i],"LWU"); type=LOAD; break;
6698 case 0x28: strcpy(insn[i],"SB"); type=STORE; break;
6699 case 0x29: strcpy(insn[i],"SH"); type=STORE; break;
6700 case 0x2A: strcpy(insn[i],"SWL"); type=STORELR; break;
6701 case 0x2B: strcpy(insn[i],"SW"); type=STORE; break;
6703 case 0x2C: strcpy(insn[i],"SDL"); type=STORELR; break;
6704 case 0x2D: strcpy(insn[i],"SDR"); type=STORELR; break;
6706 case 0x2E: strcpy(insn[i],"SWR"); type=STORELR; break;
6707 case 0x2F: strcpy(insn[i],"CACHE"); type=NOP; break;
6708 case 0x30: strcpy(insn[i],"LL"); type=NI; break;
6709 case 0x31: strcpy(insn[i],"LWC1"); type=C1LS; break;
6711 case 0x34: strcpy(insn[i],"LLD"); type=NI; break;
6712 case 0x35: strcpy(insn[i],"LDC1"); type=C1LS; break;
6713 case 0x37: strcpy(insn[i],"LD"); type=LOAD; break;
6715 case 0x38: strcpy(insn[i],"SC"); type=NI; break;
6716 case 0x39: strcpy(insn[i],"SWC1"); type=C1LS; break;
6718 case 0x3C: strcpy(insn[i],"SCD"); type=NI; break;
6719 case 0x3D: strcpy(insn[i],"SDC1"); type=C1LS; break;
6720 case 0x3F: strcpy(insn[i],"SD"); type=STORE; break;
6722 case 0x12: strcpy(insn[i],"COP2"); type=NI;
6723 op2=(source[i]>>21)&0x1f;
6725 if (source[i]&0x3f) { // use this hack to support old savestates with patched gte insns
6726 if (gte_handlers[source[i]&0x3f]!=NULL) {
6727 if (gte_regnames[source[i]&0x3f]!=NULL)
6728 strcpy(insn[i],gte_regnames[source[i]&0x3f]);
6730 snprintf(insn[i], sizeof(insn[i]), "COP2 %x", source[i]&0x3f);
6736 case 0x00: strcpy(insn[i],"MFC2"); type=COP2; break;
6737 case 0x02: strcpy(insn[i],"CFC2"); type=COP2; break;
6738 case 0x04: strcpy(insn[i],"MTC2"); type=COP2; break;
6739 case 0x06: strcpy(insn[i],"CTC2"); type=COP2; break;
6742 case 0x32: strcpy(insn[i],"LWC2"); type=C2LS; break;
6743 case 0x3A: strcpy(insn[i],"SWC2"); type=C2LS; break;
6744 case 0x3B: strcpy(insn[i],"HLECALL"); type=HLECALL; break;
6745 default: strcpy(insn[i],"???"); type=NI;
6746 SysPrintf("NI %08x @%08x (%08x)\n", source[i], addr + i*4, addr);
6751 /* Get registers/immediates */
6757 gte_rs[i]=gte_rt[i]=0;
6760 rs1[i]=(source[i]>>21)&0x1f;
6762 rt1[i]=(source[i]>>16)&0x1f;
6764 imm[i]=(short)source[i];
6768 rs1[i]=(source[i]>>21)&0x1f;
6769 rs2[i]=(source[i]>>16)&0x1f;
6772 imm[i]=(short)source[i];
6773 if(op==0x2c||op==0x2d||op==0x3f) us1[i]=rs2[i]; // 64-bit SDL/SDR/SD
6776 // LWL/LWR only load part of the register,
6777 // therefore the target register must be treated as a source too
6778 rs1[i]=(source[i]>>21)&0x1f;
6779 rs2[i]=(source[i]>>16)&0x1f;
6780 rt1[i]=(source[i]>>16)&0x1f;
6782 imm[i]=(short)source[i];
6783 if(op==0x1a||op==0x1b) us1[i]=rs2[i]; // LDR/LDL
6784 if(op==0x26) dep1[i]=rt1[i]; // LWR
6787 if (op==0x0f) rs1[i]=0; // LUI instruction has no source register
6788 else rs1[i]=(source[i]>>21)&0x1f;
6790 rt1[i]=(source[i]>>16)&0x1f;
6792 if(op>=0x0c&&op<=0x0e) { // ANDI/ORI/XORI
6793 imm[i]=(unsigned short)source[i];
6795 imm[i]=(short)source[i];
6797 if(op==0x18||op==0x19) us1[i]=rs1[i]; // DADDI/DADDIU
6798 if(op==0x0a||op==0x0b) us1[i]=rs1[i]; // SLTI/SLTIU
6799 if(op==0x0d||op==0x0e) dep1[i]=rs1[i]; // ORI/XORI
6806 // The JAL instruction writes to r31.
6813 rs1[i]=(source[i]>>21)&0x1f;
6817 // The JALR instruction writes to rd.
6819 rt1[i]=(source[i]>>11)&0x1f;
6824 rs1[i]=(source[i]>>21)&0x1f;
6825 rs2[i]=(source[i]>>16)&0x1f;
6828 if(op&2) { // BGTZ/BLEZ
6836 rs1[i]=(source[i]>>21)&0x1f;
6841 if(op2&0x10) { // BxxAL
6843 // NOTE: If the branch is not taken, r31 is still overwritten
6845 likely[i]=(op2&2)>>1;
6848 rs1[i]=(source[i]>>21)&0x1f; // source
6849 rs2[i]=(source[i]>>16)&0x1f; // subtract amount
6850 rt1[i]=(source[i]>>11)&0x1f; // destination
6852 if(op2==0x2a||op2==0x2b) { // SLT/SLTU
6853 us1[i]=rs1[i];us2[i]=rs2[i];
6855 else if(op2>=0x24&&op2<=0x27) { // AND/OR/XOR/NOR
6856 dep1[i]=rs1[i];dep2[i]=rs2[i];
6858 else if(op2>=0x2c&&op2<=0x2f) { // DADD/DSUB
6859 dep1[i]=rs1[i];dep2[i]=rs2[i];
6863 rs1[i]=(source[i]>>21)&0x1f; // source
6864 rs2[i]=(source[i]>>16)&0x1f; // divisor
6867 if (op2>=0x1c&&op2<=0x1f) { // DMULT/DMULTU/DDIV/DDIVU
6868 us1[i]=rs1[i];us2[i]=rs2[i];
6876 if(op2==0x10) rs1[i]=HIREG; // MFHI
6877 if(op2==0x11) rt1[i]=HIREG; // MTHI
6878 if(op2==0x12) rs1[i]=LOREG; // MFLO
6879 if(op2==0x13) rt1[i]=LOREG; // MTLO
6880 if((op2&0x1d)==0x10) rt1[i]=(source[i]>>11)&0x1f; // MFxx
6881 if((op2&0x1d)==0x11) rs1[i]=(source[i]>>21)&0x1f; // MTxx
6885 rs1[i]=(source[i]>>16)&0x1f; // target of shift
6886 rs2[i]=(source[i]>>21)&0x1f; // shift amount
6887 rt1[i]=(source[i]>>11)&0x1f; // destination
6889 // DSLLV/DSRLV/DSRAV are 64-bit
6890 if(op2>=0x14&&op2<=0x17) us1[i]=rs1[i];
6893 rs1[i]=(source[i]>>16)&0x1f;
6895 rt1[i]=(source[i]>>11)&0x1f;
6897 imm[i]=(source[i]>>6)&0x1f;
6898 // DSxx32 instructions
6899 if(op2>=0x3c) imm[i]|=0x20;
6900 // DSLL/DSRL/DSRA/DSRA32/DSRL32 but not DSLL32 require 64-bit source
6901 if(op2>=0x38&&op2!=0x3c) us1[i]=rs1[i];
6908 if(op2==0||op2==2) rt1[i]=(source[i]>>16)&0x1F; // MFC0/CFC0
6909 if(op2==4||op2==6) rs1[i]=(source[i]>>16)&0x1F; // MTC0/CTC0
6910 if(op2==4&&((source[i]>>11)&0x1f)==12) rt2[i]=CSREG; // Status
6911 if(op2==16) if((source[i]&0x3f)==0x18) rs2[i]=CCREG; // ERET
6918 if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC1/DMFC1/CFC1
6919 if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC1/DMTC1/CTC1
6920 if(op2==5) us1[i]=rs1[i]; // DMTC1
6928 if(op2<3) rt1[i]=(source[i]>>16)&0x1F; // MFC2/CFC2
6929 if(op2>3) rs1[i]=(source[i]>>16)&0x1F; // MTC2/CTC2
6931 int gr=(source[i]>>11)&0x1F;
6934 case 0x00: gte_rs[i]=1ll<<gr; break; // MFC2
6935 case 0x04: gte_rt[i]=1ll<<gr; break; // MTC2
6936 case 0x02: gte_rs[i]=1ll<<(gr+32); break; // CFC2
6937 case 0x06: gte_rt[i]=1ll<<(gr+32); break; // CTC2
6941 rs1[i]=(source[i]>>21)&0x1F;
6945 imm[i]=(short)source[i];
6948 rs1[i]=(source[i]>>21)&0x1F;
6952 imm[i]=(short)source[i];
6953 if(op==0x32) gte_rt[i]=1ll<<((source[i]>>16)&0x1F); // LWC2
6954 else gte_rs[i]=1ll<<((source[i]>>16)&0x1F); // SWC2
6961 gte_rs[i]=gte_reg_reads[source[i]&0x3f];
6962 gte_rt[i]=gte_reg_writes[source[i]&0x3f];
6963 gte_rt[i]|=1ll<<63; // every op changes flags
6964 if((source[i]&0x3f)==GTE_MVMVA) {
6965 int v = (source[i] >> 15) & 3;
6966 gte_rs[i]&=~0xe3fll;
6967 if(v==3) gte_rs[i]|=0xe00ll;
6968 else gte_rs[i]|=3ll<<(v*2);
6985 /* Calculate branch target addresses */
6987 ba[i]=((start+i*4+4)&0xF0000000)|(((unsigned int)source[i]<<6)>>4);
6988 else if(type==CJUMP&&rs1[i]==rs2[i]&&(op&1))
6989 ba[i]=start+i*4+8; // Ignore never taken branch
6990 else if(type==SJUMP&&rs1[i]==0&&!(op2&1))
6991 ba[i]=start+i*4+8; // Ignore never taken branch
6992 else if(type==CJUMP||type==SJUMP)
6993 ba[i]=start+i*4+4+((signed int)((unsigned int)source[i]<<16)>>14);
6995 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP)) {
6997 // branch in delay slot?
6998 if(type==RJUMP||type==UJUMP||type==CJUMP||type==SJUMP) {
6999 // don't handle first branch and call interpreter if it's hit
7000 SysPrintf("branch in delay slot @%08x (%08x)\n", addr + i*4, addr);
7003 // basic load delay detection
7004 else if((type==LOAD||type==LOADLR||type==COP0||type==COP2||type==C2LS)&&rt1[i]!=0) {
7005 int t=(ba[i-1]-start)/4;
7006 if(0 <= t && t < i &&(rt1[i]==rs1[t]||rt1[i]==rs2[t])&&itype[t]!=CJUMP&&itype[t]!=SJUMP) {
7007 // jump target wants DS result - potential load delay effect
7008 SysPrintf("load delay @%08x (%08x)\n", addr + i*4, addr);
7010 bt[t+1]=1; // expected return from interpreter
7012 else if(i>=2&&rt1[i-2]==2&&rt1[i]==2&&rs1[i]!=2&&rs2[i]!=2&&rs1[i-1]!=2&&rs2[i-1]!=2&&
7013 !(i>=3&&(itype[i-3]==RJUMP||itype[i-3]==UJUMP||itype[i-3]==CJUMP||itype[i-3]==SJUMP))) {
7014 // v0 overwrite like this is a sign of trouble, bail out
7015 SysPrintf("v0 overwrite @%08x (%08x)\n", addr + i*4, addr);
7021 rs2[i-1]=rt1[i-1]=rt2[i-1]=0;
7025 i--; // don't compile the DS
7028 /* Is this the end of the block? */
7029 if(i>0&&(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)) {
7030 if(rt1[i-1]==0) { // Continue past subroutine call (JAL)
7034 if(stop_after_jal) done=1;
7036 if((source[i+1]&0xfc00003f)==0x0d) done=1;
7038 // Don't recompile stuff that's already compiled
7039 if(check_addr(start+i*4+4)) done=1;
7040 // Don't get too close to the limit
7041 if(i>MAXBLOCK/2) done=1;
7043 if(itype[i]==SYSCALL&&stop_after_jal) done=1;
7044 if(itype[i]==HLECALL||itype[i]==INTCALL) done=2;
7046 // Does the block continue due to a branch?
7049 if(ba[j]==start+i*4) done=j=0; // Branch into delay slot
7050 if(ba[j]==start+i*4+4) done=j=0;
7051 if(ba[j]==start+i*4+8) done=j=0;
7054 //assert(i<MAXBLOCK-1);
7055 if(start+i*4==pagelimit-4) done=1;
7056 assert(start+i*4<pagelimit);
7057 if (i==MAXBLOCK-1) done=1;
7058 // Stop if we're compiling junk
7059 if(itype[i]==NI&&opcode[i]==0x11) {
7060 done=stop_after_jal=1;
7061 SysPrintf("Disabled speculative precompilation\n");
7065 if(itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i-1]==RJUMP) {
7066 if(start+i*4==pagelimit) {
7072 /* Pass 2 - Register dependencies and branch targets */
7074 unneeded_registers(0,slen-1,0);
7076 /* Pass 3 - Register allocation */
7078 struct regstat current; // Current register allocations/status
7080 current.u=unneeded_reg[0];
7081 clear_all_regs(current.regmap);
7082 alloc_reg(¤t,0,CCREG);
7083 dirty_reg(¤t,CCREG);
7086 current.waswritten=0;
7092 // First instruction is delay slot
7097 current.regmap[HOST_BTREG]=BTREG;
7105 for(hr=0;hr<HOST_REGS;hr++)
7107 // Is this really necessary?
7108 if(current.regmap[hr]==0) current.regmap[hr]=-1;
7111 current.waswritten=0;
7115 if((opcode[i-2]&0x2f)==0x05) // BNE/BNEL
7117 if(rs1[i-2]==0||rs2[i-2]==0)
7120 int hr=get_reg(current.regmap,rs1[i-2]|64);
7121 if(hr>=0) current.regmap[hr]=-1;
7124 int hr=get_reg(current.regmap,rs2[i-2]|64);
7125 if(hr>=0) current.regmap[hr]=-1;
7131 memcpy(regmap_pre[i],current.regmap,sizeof(current.regmap));
7132 regs[i].wasconst=current.isconst;
7133 regs[i].wasdirty=current.dirty;
7134 regs[i].loadedconst=0;
7135 if(itype[i]!=UJUMP&&itype[i]!=CJUMP&&itype[i]!=SJUMP&&itype[i]!=RJUMP) {
7137 current.u=unneeded_reg[i+1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
7144 current.u=branch_unneeded_reg[i]&~((1LL<<rs1[i+1])|(1LL<<rs2[i+1]));
7145 current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
7147 } else { SysPrintf("oops, branch at end of block with no delay slot\n");exit(1); }
7151 ds=0; // Skip delay slot, already allocated as part of branch
7152 // ...but we need to alloc it in case something jumps here
7154 current.u=branch_unneeded_reg[i-1]&unneeded_reg[i+1];
7156 current.u=branch_unneeded_reg[i-1];
7158 current.u&=~((1LL<<rs1[i])|(1LL<<rs2[i]));
7160 struct regstat temp;
7161 memcpy(&temp,¤t,sizeof(current));
7162 temp.wasdirty=temp.dirty;
7163 // TODO: Take into account unconditional branches, as below
7164 delayslot_alloc(&temp,i);
7165 memcpy(regs[i].regmap,temp.regmap,sizeof(temp.regmap));
7166 regs[i].wasdirty=temp.wasdirty;
7167 regs[i].dirty=temp.dirty;
7171 // Create entry (branch target) regmap
7172 for(hr=0;hr<HOST_REGS;hr++)
7174 int r=temp.regmap[hr];
7176 if(r!=regmap_pre[i][hr]) {
7177 regs[i].regmap_entry[hr]=-1;
7182 if((current.u>>r)&1) {
7183 regs[i].regmap_entry[hr]=-1;
7184 regs[i].regmap[hr]=-1;
7185 //Don't clear regs in the delay slot as the branch might need them
7186 //current.regmap[hr]=-1;
7188 regs[i].regmap_entry[hr]=r;
7195 // First instruction expects CCREG to be allocated
7196 if(i==0&&hr==HOST_CCREG)
7197 regs[i].regmap_entry[hr]=CCREG;
7199 regs[i].regmap_entry[hr]=-1;
7203 else { // Not delay slot
7206 //current.isconst=0; // DEBUG
7207 //current.wasconst=0; // DEBUG
7208 //regs[i].wasconst=0; // DEBUG
7209 clear_const(¤t,rt1[i]);
7210 alloc_cc(¤t,i);
7211 dirty_reg(¤t,CCREG);
7213 alloc_reg(¤t,i,31);
7214 dirty_reg(¤t,31);
7215 //assert(rs1[i+1]!=31&&rs2[i+1]!=31);
7216 //assert(rt1[i+1]!=rt1[i]);
7218 alloc_reg(¤t,i,PTEMP);
7222 delayslot_alloc(¤t,i+1);
7223 //current.isconst=0; // DEBUG
7225 //printf("i=%d, isconst=%x\n",i,current.isconst);
7228 //current.isconst=0;
7229 //current.wasconst=0;
7230 //regs[i].wasconst=0;
7231 clear_const(¤t,rs1[i]);
7232 clear_const(¤t,rt1[i]);
7233 alloc_cc(¤t,i);
7234 dirty_reg(¤t,CCREG);
7235 if(rs1[i]!=rt1[i+1]&&rs1[i]!=rt2[i+1]) {
7236 alloc_reg(¤t,i,rs1[i]);
7238 alloc_reg(¤t,i,rt1[i]);
7239 dirty_reg(¤t,rt1[i]);
7240 assert(rs1[i+1]!=rt1[i]&&rs2[i+1]!=rt1[i]);
7241 assert(rt1[i+1]!=rt1[i]);
7243 alloc_reg(¤t,i,PTEMP);
7247 if(rs1[i]==31) { // JALR
7248 alloc_reg(¤t,i,RHASH);
7249 alloc_reg(¤t,i,RHTBL);
7252 delayslot_alloc(¤t,i+1);
7254 // The delay slot overwrites our source register,
7255 // allocate a temporary register to hold the old value.
7259 delayslot_alloc(¤t,i+1);
7261 alloc_reg(¤t,i,RTEMP);
7263 //current.isconst=0; // DEBUG
7268 //current.isconst=0;
7269 //current.wasconst=0;
7270 //regs[i].wasconst=0;
7271 clear_const(¤t,rs1[i]);
7272 clear_const(¤t,rs2[i]);
7273 if((opcode[i]&0x3E)==4) // BEQ/BNE
7275 alloc_cc(¤t,i);
7276 dirty_reg(¤t,CCREG);
7277 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
7278 if(rs2[i]) alloc_reg(¤t,i,rs2[i]);
7279 if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1]))||
7280 (rs2[i]&&(rs2[i]==rt1[i+1]||rs2[i]==rt2[i+1]))) {
7281 // The delay slot overwrites one of our conditions.
7282 // Allocate the branch condition registers instead.
7286 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
7287 if(rs2[i]) alloc_reg(¤t,i,rs2[i]);
7292 delayslot_alloc(¤t,i+1);
7296 if((opcode[i]&0x3E)==6) // BLEZ/BGTZ
7298 alloc_cc(¤t,i);
7299 dirty_reg(¤t,CCREG);
7300 alloc_reg(¤t,i,rs1[i]);
7301 if(rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) {
7302 // The delay slot overwrites one of our conditions.
7303 // Allocate the branch condition registers instead.
7307 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
7312 delayslot_alloc(¤t,i+1);
7316 // Don't alloc the delay slot yet because we might not execute it
7317 if((opcode[i]&0x3E)==0x14) // BEQL/BNEL
7322 alloc_cc(¤t,i);
7323 dirty_reg(¤t,CCREG);
7324 alloc_reg(¤t,i,rs1[i]);
7325 alloc_reg(¤t,i,rs2[i]);
7328 if((opcode[i]&0x3E)==0x16) // BLEZL/BGTZL
7333 alloc_cc(¤t,i);
7334 dirty_reg(¤t,CCREG);
7335 alloc_reg(¤t,i,rs1[i]);
7338 //current.isconst=0;
7341 //current.isconst=0;
7342 //current.wasconst=0;
7343 //regs[i].wasconst=0;
7344 clear_const(¤t,rs1[i]);
7345 clear_const(¤t,rt1[i]);
7346 //if((opcode2[i]&0x1E)==0x0) // BLTZ/BGEZ
7347 if((opcode2[i]&0x0E)==0x0) // BLTZ/BGEZ
7349 alloc_cc(¤t,i);
7350 dirty_reg(¤t,CCREG);
7351 alloc_reg(¤t,i,rs1[i]);
7352 if (rt1[i]==31) { // BLTZAL/BGEZAL
7353 alloc_reg(¤t,i,31);
7354 dirty_reg(¤t,31);
7355 //#ifdef REG_PREFETCH
7356 //alloc_reg(¤t,i,PTEMP);
7359 if((rs1[i]&&(rs1[i]==rt1[i+1]||rs1[i]==rt2[i+1])) // The delay slot overwrites the branch condition.
7360 ||(rt1[i]==31&&(rs1[i+1]==31||rs2[i+1]==31||rt1[i+1]==31||rt2[i+1]==31))) { // DS touches $ra
7361 // Allocate the branch condition registers instead.
7365 if(rs1[i]) alloc_reg(¤t,i,rs1[i]);
7370 delayslot_alloc(¤t,i+1);
7374 // Don't alloc the delay slot yet because we might not execute it
7375 if((opcode2[i]&0x1E)==0x2) // BLTZL/BGEZL
7380 alloc_cc(¤t,i);
7381 dirty_reg(¤t,CCREG);
7382 alloc_reg(¤t,i,rs1[i]);
7385 //current.isconst=0;
7388 imm16_alloc(¤t,i);
7392 load_alloc(¤t,i);
7396 store_alloc(¤t,i);
7399 alu_alloc(¤t,i);
7402 shift_alloc(¤t,i);
7405 multdiv_alloc(¤t,i);
7408 shiftimm_alloc(¤t,i);
7411 mov_alloc(¤t,i);
7414 cop0_alloc(¤t,i);
7418 cop12_alloc(¤t,i);
7421 c1ls_alloc(¤t,i);
7424 c2ls_alloc(¤t,i);
7427 c2op_alloc(¤t,i);
7432 syscall_alloc(¤t,i);
7435 pagespan_alloc(¤t,i);
7439 // Create entry (branch target) regmap
7440 for(hr=0;hr<HOST_REGS;hr++)
7443 r=current.regmap[hr];
7445 if(r!=regmap_pre[i][hr]) {
7446 // TODO: delay slot (?)
7447 or=get_reg(regmap_pre[i],r); // Get old mapping for this register
7448 if(or<0||(r&63)>=TEMPREG){
7449 regs[i].regmap_entry[hr]=-1;
7453 // Just move it to a different register
7454 regs[i].regmap_entry[hr]=r;
7455 // If it was dirty before, it's still dirty
7456 if((regs[i].wasdirty>>or)&1) dirty_reg(¤t,r&63);
7463 regs[i].regmap_entry[hr]=0;
7467 if((current.u>>r)&1) {
7468 regs[i].regmap_entry[hr]=-1;
7469 //regs[i].regmap[hr]=-1;
7470 current.regmap[hr]=-1;
7472 regs[i].regmap_entry[hr]=r;
7479 // Branches expect CCREG to be allocated at the target
7480 if(regmap_pre[i][hr]==CCREG)
7481 regs[i].regmap_entry[hr]=CCREG;
7483 regs[i].regmap_entry[hr]=-1;
7486 memcpy(regs[i].regmap,current.regmap,sizeof(current.regmap));
7489 if(i>0&&(itype[i-1]==STORE||itype[i-1]==STORELR||(itype[i-1]==C2LS&&opcode[i-1]==0x3a))&&(u_int)imm[i-1]<0x800)
7490 current.waswritten|=1<<rs1[i-1];
7491 current.waswritten&=~(1<<rt1[i]);
7492 current.waswritten&=~(1<<rt2[i]);
7493 if((itype[i]==STORE||itype[i]==STORELR||(itype[i]==C2LS&&opcode[i]==0x3a))&&(u_int)imm[i]>=0x800)
7494 current.waswritten&=~(1<<rs1[i]);
7496 /* Branch post-alloc */
7499 current.wasdirty=current.dirty;
7500 switch(itype[i-1]) {
7502 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7503 branch_regs[i-1].isconst=0;
7504 branch_regs[i-1].wasconst=0;
7505 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
7506 alloc_cc(&branch_regs[i-1],i-1);
7507 dirty_reg(&branch_regs[i-1],CCREG);
7508 if(rt1[i-1]==31) { // JAL
7509 alloc_reg(&branch_regs[i-1],i-1,31);
7510 dirty_reg(&branch_regs[i-1],31);
7512 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7513 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
7516 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7517 branch_regs[i-1].isconst=0;
7518 branch_regs[i-1].wasconst=0;
7519 branch_regs[i-1].u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
7520 alloc_cc(&branch_regs[i-1],i-1);
7521 dirty_reg(&branch_regs[i-1],CCREG);
7522 alloc_reg(&branch_regs[i-1],i-1,rs1[i-1]);
7523 if(rt1[i-1]!=0) { // JALR
7524 alloc_reg(&branch_regs[i-1],i-1,rt1[i-1]);
7525 dirty_reg(&branch_regs[i-1],rt1[i-1]);
7528 if(rs1[i-1]==31) { // JALR
7529 alloc_reg(&branch_regs[i-1],i-1,RHASH);
7530 alloc_reg(&branch_regs[i-1],i-1,RHTBL);
7533 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7534 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
7537 if((opcode[i-1]&0x3E)==4) // BEQ/BNE
7539 alloc_cc(¤t,i-1);
7540 dirty_reg(¤t,CCREG);
7541 if((rs1[i-1]&&(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]))||
7542 (rs2[i-1]&&(rs2[i-1]==rt1[i]||rs2[i-1]==rt2[i]))) {
7543 // The delay slot overwrote one of our conditions
7544 // Delay slot goes after the test (in order)
7545 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
7547 delayslot_alloc(¤t,i);
7552 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i-1])|(1LL<<rs2[i-1]));
7553 // Alloc the branch condition registers
7554 if(rs1[i-1]) alloc_reg(¤t,i-1,rs1[i-1]);
7555 if(rs2[i-1]) alloc_reg(¤t,i-1,rs2[i-1]);
7557 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7558 branch_regs[i-1].isconst=0;
7559 branch_regs[i-1].wasconst=0;
7560 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7561 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
7564 if((opcode[i-1]&0x3E)==6) // BLEZ/BGTZ
7566 alloc_cc(¤t,i-1);
7567 dirty_reg(¤t,CCREG);
7568 if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
7569 // The delay slot overwrote the branch condition
7570 // Delay slot goes after the test (in order)
7571 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
7573 delayslot_alloc(¤t,i);
7578 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
7579 // Alloc the branch condition register
7580 alloc_reg(¤t,i-1,rs1[i-1]);
7582 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7583 branch_regs[i-1].isconst=0;
7584 branch_regs[i-1].wasconst=0;
7585 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7586 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
7589 // Alloc the delay slot in case the branch is taken
7590 if((opcode[i-1]&0x3E)==0x14) // BEQL/BNEL
7592 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7593 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
7594 alloc_cc(&branch_regs[i-1],i);
7595 dirty_reg(&branch_regs[i-1],CCREG);
7596 delayslot_alloc(&branch_regs[i-1],i);
7597 branch_regs[i-1].isconst=0;
7598 alloc_reg(¤t,i,CCREG); // Not taken path
7599 dirty_reg(¤t,CCREG);
7600 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7603 if((opcode[i-1]&0x3E)==0x16) // BLEZL/BGTZL
7605 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7606 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
7607 alloc_cc(&branch_regs[i-1],i);
7608 dirty_reg(&branch_regs[i-1],CCREG);
7609 delayslot_alloc(&branch_regs[i-1],i);
7610 branch_regs[i-1].isconst=0;
7611 alloc_reg(¤t,i,CCREG); // Not taken path
7612 dirty_reg(¤t,CCREG);
7613 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7617 //if((opcode2[i-1]&0x1E)==0) // BLTZ/BGEZ
7618 if((opcode2[i-1]&0x0E)==0) // BLTZ/BGEZ
7620 alloc_cc(¤t,i-1);
7621 dirty_reg(¤t,CCREG);
7622 if(rs1[i-1]==rt1[i]||rs1[i-1]==rt2[i]) {
7623 // The delay slot overwrote the branch condition
7624 // Delay slot goes after the test (in order)
7625 current.u=branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i]));
7627 delayslot_alloc(¤t,i);
7632 current.u=branch_unneeded_reg[i-1]&~(1LL<<rs1[i-1]);
7633 // Alloc the branch condition register
7634 alloc_reg(¤t,i-1,rs1[i-1]);
7636 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7637 branch_regs[i-1].isconst=0;
7638 branch_regs[i-1].wasconst=0;
7639 memcpy(&branch_regs[i-1].regmap_entry,¤t.regmap,sizeof(current.regmap));
7640 memcpy(constmap[i],constmap[i-1],sizeof(current_constmap));
7643 // Alloc the delay slot in case the branch is taken
7644 if((opcode2[i-1]&0x1E)==2) // BLTZL/BGEZL
7646 memcpy(&branch_regs[i-1],¤t,sizeof(current));
7647 branch_regs[i-1].u=(branch_unneeded_reg[i-1]&~((1LL<<rs1[i])|(1LL<<rs2[i])|(1LL<<rt1[i])|(1LL<<rt2[i])))|1;
7648 alloc_cc(&branch_regs[i-1],i);
7649 dirty_reg(&branch_regs[i-1],CCREG);
7650 delayslot_alloc(&branch_regs[i-1],i);
7651 branch_regs[i-1].isconst=0;
7652 alloc_reg(¤t,i,CCREG); // Not taken path
7653 dirty_reg(¤t,CCREG);
7654 memcpy(&branch_regs[i-1].regmap_entry,&branch_regs[i-1].regmap,sizeof(current.regmap));
7656 // FIXME: BLTZAL/BGEZAL
7657 if(opcode2[i-1]&0x10) { // BxxZAL
7658 alloc_reg(&branch_regs[i-1],i-1,31);
7659 dirty_reg(&branch_regs[i-1],31);
7664 if(itype[i-1]==UJUMP||itype[i-1]==RJUMP||(source[i-1]>>16)==0x1000)
7666 if(rt1[i-1]==31) // JAL/JALR
7668 // Subroutine call will return here, don't alloc any registers
7670 clear_all_regs(current.regmap);
7671 alloc_reg(¤t,i,CCREG);
7672 dirty_reg(¤t,CCREG);
7676 // Internal branch will jump here, match registers to caller
7678 clear_all_regs(current.regmap);
7679 alloc_reg(¤t,i,CCREG);
7680 dirty_reg(¤t,CCREG);
7683 if(ba[j]==start+i*4+4) {
7684 memcpy(current.regmap,branch_regs[j].regmap,sizeof(current.regmap));
7685 current.dirty=branch_regs[j].dirty;
7690 if(ba[j]==start+i*4+4) {
7691 for(hr=0;hr<HOST_REGS;hr++) {
7692 if(current.regmap[hr]!=branch_regs[j].regmap[hr]) {
7693 current.regmap[hr]=-1;
7695 current.dirty&=branch_regs[j].dirty;
7704 // Count cycles in between branches
7706 if(i>0&&(itype[i-1]==RJUMP||itype[i-1]==UJUMP||itype[i-1]==CJUMP||itype[i-1]==SJUMP||itype[i]==SYSCALL||itype[i]==HLECALL))
7710 #if !defined(DRC_DBG)
7711 else if(itype[i]==C2OP&>e_cycletab[source[i]&0x3f]>2)
7713 // GTE runs in parallel until accessed, divide by 2 for a rough guess
7714 cc+=gte_cycletab[source[i]&0x3f]/2;
7716 else if(/*itype[i]==LOAD||itype[i]==STORE||*/itype[i]==C1LS) // load,store causes weird timing issues
7718 cc+=2; // 2 cycle penalty (after CLOCK_DIVIDER)
7720 else if(i>1&&itype[i]==STORE&&itype[i-1]==STORE&&itype[i-2]==STORE&&!bt[i])
7724 else if(itype[i]==C2LS)
7735 regs[i].dirty=current.dirty;
7736 regs[i].isconst=current.isconst;
7737 memcpy(constmap[i],current_constmap,sizeof(current_constmap));
7739 for(hr=0;hr<HOST_REGS;hr++) {
7740 if(hr!=EXCLUDE_REG&®s[i].regmap[hr]>=0) {
7741 if(regmap_pre[i][hr]!=regs[i].regmap[hr]) {
7742 regs[i].wasconst&=~(1<<hr);
7746 if(current.regmap[HOST_BTREG]==BTREG) current.regmap[HOST_BTREG]=-1;
7747 regs[i].waswritten=current.waswritten;
7750 /* Pass 4 - Cull unused host registers */
7754 for (i=slen-1;i>=0;i--)
7757 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
7759 if(ba[i]<start || ba[i]>=(start+slen*4))
7761 // Branch out of this block, don't need anything
7767 // Need whatever matches the target
7769 int t=(ba[i]-start)>>2;
7770 for(hr=0;hr<HOST_REGS;hr++)
7772 if(regs[i].regmap_entry[hr]>=0) {
7773 if(regs[i].regmap_entry[hr]==regs[t].regmap_entry[hr]) nr|=1<<hr;
7777 // Conditional branch may need registers for following instructions
7778 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
7781 nr|=needed_reg[i+2];
7782 for(hr=0;hr<HOST_REGS;hr++)
7784 if(regmap_pre[i+2][hr]>=0&&get_reg(regs[i+2].regmap_entry,regmap_pre[i+2][hr])<0) nr&=~(1<<hr);
7785 //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]);
7789 // Don't need stuff which is overwritten
7790 //if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7791 //if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7792 // Merge in delay slot
7793 for(hr=0;hr<HOST_REGS;hr++)
7796 // These are overwritten unless the branch is "likely"
7797 // and the delay slot is nullified if not taken
7798 if(rt1[i+1]&&rt1[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
7799 if(rt2[i+1]&&rt2[i+1]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
7801 if(us1[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
7802 if(us2[i+1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
7803 if(rs1[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
7804 if(rs2[i+1]==regmap_pre[i][hr]) nr|=1<<hr;
7805 if(us1[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
7806 if(us2[i+1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
7807 if(rs1[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
7808 if(rs2[i+1]==regs[i].regmap_entry[hr]) nr|=1<<hr;
7809 if(itype[i+1]==STORE || itype[i+1]==STORELR || (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) {
7810 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
7811 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
7815 else if(itype[i]==SYSCALL||itype[i]==HLECALL||itype[i]==INTCALL)
7817 // SYSCALL instruction (software interrupt)
7820 else if(itype[i]==COP0 && (source[i]&0x3f)==0x18)
7822 // ERET instruction (return from interrupt)
7828 for(hr=0;hr<HOST_REGS;hr++) {
7829 if(regmap_pre[i+1][hr]>=0&&get_reg(regs[i+1].regmap_entry,regmap_pre[i+1][hr])<0) nr&=~(1<<hr);
7830 if(regs[i].regmap[hr]!=regmap_pre[i+1][hr]) nr&=~(1<<hr);
7831 if(regs[i].regmap[hr]!=regmap_pre[i][hr]) nr&=~(1<<hr);
7832 if(regs[i].regmap[hr]<0) nr&=~(1<<hr);
7836 for(hr=0;hr<HOST_REGS;hr++)
7838 // Overwritten registers are not needed
7839 if(rt1[i]&&rt1[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
7840 if(rt2[i]&&rt2[i]==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
7841 if(FTEMP==(regs[i].regmap[hr]&63)) nr&=~(1<<hr);
7842 // Source registers are needed
7843 if(us1[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
7844 if(us2[i]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
7845 if(rs1[i]==regmap_pre[i][hr]) nr|=1<<hr;
7846 if(rs2[i]==regmap_pre[i][hr]) nr|=1<<hr;
7847 if(us1[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
7848 if(us2[i]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
7849 if(rs1[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
7850 if(rs2[i]==regs[i].regmap_entry[hr]) nr|=1<<hr;
7851 if(itype[i]==STORE || itype[i]==STORELR || (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) {
7852 if(regmap_pre[i][hr]==INVCP) nr|=1<<hr;
7853 if(regs[i].regmap_entry[hr]==INVCP) nr|=1<<hr;
7855 // Don't store a register immediately after writing it,
7856 // may prevent dual-issue.
7857 // But do so if this is a branch target, otherwise we
7858 // might have to load the register before the branch.
7859 if(i>0&&!bt[i]&&((regs[i].wasdirty>>hr)&1)) {
7860 if((regmap_pre[i][hr]>0&®map_pre[i][hr]<64&&!((unneeded_reg[i]>>regmap_pre[i][hr])&1))) {
7861 if(rt1[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
7862 if(rt2[i-1]==(regmap_pre[i][hr]&63)) nr|=1<<hr;
7864 if((regs[i].regmap_entry[hr]>0&®s[i].regmap_entry[hr]<64&&!((unneeded_reg[i]>>regs[i].regmap_entry[hr])&1))) {
7865 if(rt1[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
7866 if(rt2[i-1]==(regs[i].regmap_entry[hr]&63)) nr|=1<<hr;
7870 // Cycle count is needed at branches. Assume it is needed at the target too.
7871 if(i==0||bt[i]||itype[i]==CJUMP||itype[i]==SPAN) {
7872 if(regmap_pre[i][HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7873 if(regs[i].regmap_entry[HOST_CCREG]==CCREG) nr|=1<<HOST_CCREG;
7878 // Deallocate unneeded registers
7879 for(hr=0;hr<HOST_REGS;hr++)
7882 if(regs[i].regmap_entry[hr]!=CCREG) regs[i].regmap_entry[hr]=-1;
7883 if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
7884 (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
7885 (regs[i].regmap[hr]&63)!=PTEMP && (regs[i].regmap[hr]&63)!=CCREG)
7887 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
7890 regs[i].regmap[hr]=-1;
7891 regs[i].isconst&=~(1<<hr);
7893 regmap_pre[i+2][hr]=-1;
7894 regs[i+2].wasconst&=~(1<<hr);
7899 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
7901 int d1=0,d2=0,map=0,temp=0;
7902 if(get_reg(regs[i].regmap,rt1[i+1]|64)>=0||get_reg(branch_regs[i].regmap,rt1[i+1]|64)>=0)
7907 if(itype[i+1]==STORE || itype[i+1]==STORELR ||
7908 (opcode[i+1]&0x3b)==0x39 || (opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
7911 if(itype[i+1]==LOADLR || itype[i+1]==STORELR ||
7912 itype[i+1]==C1LS || itype[i+1]==C2LS)
7914 if((regs[i].regmap[hr]&63)!=rs1[i] && (regs[i].regmap[hr]&63)!=rs2[i] &&
7915 (regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
7916 (regs[i].regmap[hr]&63)!=rt1[i+1] && (regs[i].regmap[hr]&63)!=rt2[i+1] &&
7917 (regs[i].regmap[hr]^64)!=us1[i+1] && (regs[i].regmap[hr]^64)!=us2[i+1] &&
7918 (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
7919 regs[i].regmap[hr]!=rs1[i+1] && regs[i].regmap[hr]!=rs2[i+1] &&
7920 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=PTEMP &&
7921 regs[i].regmap[hr]!=RHASH && regs[i].regmap[hr]!=RHTBL &&
7922 regs[i].regmap[hr]!=RTEMP && regs[i].regmap[hr]!=CCREG &&
7923 regs[i].regmap[hr]!=map )
7925 regs[i].regmap[hr]=-1;
7926 regs[i].isconst&=~(1<<hr);
7927 if((branch_regs[i].regmap[hr]&63)!=rs1[i] && (branch_regs[i].regmap[hr]&63)!=rs2[i] &&
7928 (branch_regs[i].regmap[hr]&63)!=rt1[i] && (branch_regs[i].regmap[hr]&63)!=rt2[i] &&
7929 (branch_regs[i].regmap[hr]&63)!=rt1[i+1] && (branch_regs[i].regmap[hr]&63)!=rt2[i+1] &&
7930 (branch_regs[i].regmap[hr]^64)!=us1[i+1] && (branch_regs[i].regmap[hr]^64)!=us2[i+1] &&
7931 (branch_regs[i].regmap[hr]^64)!=d1 && (branch_regs[i].regmap[hr]^64)!=d2 &&
7932 branch_regs[i].regmap[hr]!=rs1[i+1] && branch_regs[i].regmap[hr]!=rs2[i+1] &&
7933 (branch_regs[i].regmap[hr]&63)!=temp && branch_regs[i].regmap[hr]!=PTEMP &&
7934 branch_regs[i].regmap[hr]!=RHASH && branch_regs[i].regmap[hr]!=RHTBL &&
7935 branch_regs[i].regmap[hr]!=RTEMP && branch_regs[i].regmap[hr]!=CCREG &&
7936 branch_regs[i].regmap[hr]!=map)
7938 branch_regs[i].regmap[hr]=-1;
7939 branch_regs[i].regmap_entry[hr]=-1;
7940 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000)
7942 if(!likely[i]&&i<slen-2) {
7943 regmap_pre[i+2][hr]=-1;
7944 regs[i+2].wasconst&=~(1<<hr);
7955 int d1=0,d2=0,map=-1,temp=-1;
7956 if(get_reg(regs[i].regmap,rt1[i]|64)>=0)
7961 if(itype[i]==STORE || itype[i]==STORELR ||
7962 (opcode[i]&0x3b)==0x39 || (opcode[i]&0x3b)==0x3a) { // SWC1/SDC1 || SWC2/SDC2
7965 if(itype[i]==LOADLR || itype[i]==STORELR ||
7966 itype[i]==C1LS || itype[i]==C2LS)
7968 if((regs[i].regmap[hr]&63)!=rt1[i] && (regs[i].regmap[hr]&63)!=rt2[i] &&
7969 (regs[i].regmap[hr]^64)!=us1[i] && (regs[i].regmap[hr]^64)!=us2[i] &&
7970 (regs[i].regmap[hr]^64)!=d1 && (regs[i].regmap[hr]^64)!=d2 &&
7971 regs[i].regmap[hr]!=rs1[i] && regs[i].regmap[hr]!=rs2[i] &&
7972 (regs[i].regmap[hr]&63)!=temp && regs[i].regmap[hr]!=map &&
7973 (itype[i]!=SPAN||regs[i].regmap[hr]!=CCREG))
7975 if(i<slen-1&&!is_ds[i]) {
7976 assert(regs[i].regmap[hr]<64);
7977 if(regmap_pre[i+1][hr]!=-1 || regs[i].regmap[hr]!=-1)
7978 if(regmap_pre[i+1][hr]!=regs[i].regmap[hr])
7980 SysPrintf("fail: %x (%d %d!=%d)\n",start+i*4,hr,regmap_pre[i+1][hr],regs[i].regmap[hr]);
7981 assert(regmap_pre[i+1][hr]==regs[i].regmap[hr]);
7983 regmap_pre[i+1][hr]=-1;
7984 if(regs[i+1].regmap_entry[hr]==CCREG) regs[i+1].regmap_entry[hr]=-1;
7985 regs[i+1].wasconst&=~(1<<hr);
7987 regs[i].regmap[hr]=-1;
7988 regs[i].isconst&=~(1<<hr);
7996 /* Pass 5 - Pre-allocate registers */
7998 // If a register is allocated during a loop, try to allocate it for the
7999 // entire loop, if possible. This avoids loading/storing registers
8000 // inside of the loop.
8002 signed char f_regmap[HOST_REGS];
8003 clear_all_regs(f_regmap);
8004 for(i=0;i<slen-1;i++)
8006 if(itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
8008 if(ba[i]>=start && ba[i]<(start+i*4))
8009 if(itype[i+1]==NOP||itype[i+1]==MOV||itype[i+1]==ALU
8010 ||itype[i+1]==SHIFTIMM||itype[i+1]==IMM16||itype[i+1]==LOAD
8011 ||itype[i+1]==STORE||itype[i+1]==STORELR||itype[i+1]==C1LS
8012 ||itype[i+1]==SHIFT||itype[i+1]==COP1
8013 ||itype[i+1]==COP2||itype[i+1]==C2LS||itype[i+1]==C2OP)
8015 int t=(ba[i]-start)>>2;
8016 if(t>0&&(itype[t-1]!=UJUMP&&itype[t-1]!=RJUMP&&itype[t-1]!=CJUMP&&itype[t-1]!=SJUMP)) // loop_preload can't handle jumps into delay slots
8017 if(t<2||(itype[t-2]!=UJUMP&&itype[t-2]!=RJUMP)||rt1[t-2]!=31) // call/ret assumes no registers allocated
8018 for(hr=0;hr<HOST_REGS;hr++)
8020 if(regs[i].regmap[hr]>64) {
8021 if(!((regs[i].dirty>>hr)&1))
8022 f_regmap[hr]=regs[i].regmap[hr];
8023 else f_regmap[hr]=-1;
8025 else if(regs[i].regmap[hr]>=0) {
8026 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8027 // dealloc old register
8029 for(n=0;n<HOST_REGS;n++)
8031 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8033 // and alloc new one
8034 f_regmap[hr]=regs[i].regmap[hr];
8037 if(branch_regs[i].regmap[hr]>64) {
8038 if(!((branch_regs[i].dirty>>hr)&1))
8039 f_regmap[hr]=branch_regs[i].regmap[hr];
8040 else f_regmap[hr]=-1;
8042 else if(branch_regs[i].regmap[hr]>=0) {
8043 if(f_regmap[hr]!=branch_regs[i].regmap[hr]) {
8044 // dealloc old register
8046 for(n=0;n<HOST_REGS;n++)
8048 if(f_regmap[n]==branch_regs[i].regmap[hr]) {f_regmap[n]=-1;}
8050 // and alloc new one
8051 f_regmap[hr]=branch_regs[i].regmap[hr];
8055 if(count_free_regs(regs[i].regmap)<=minimum_free_regs[i+1])
8056 f_regmap[hr]=branch_regs[i].regmap[hr];
8058 if(count_free_regs(branch_regs[i].regmap)<=minimum_free_regs[i+1])
8059 f_regmap[hr]=branch_regs[i].regmap[hr];
8061 // Avoid dirty->clean transition
8062 #ifdef DESTRUCTIVE_WRITEBACK
8063 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;
8065 // This check is only strictly required in the DESTRUCTIVE_WRITEBACK
8066 // case above, however it's always a good idea. We can't hoist the
8067 // load if the register was already allocated, so there's no point
8068 // wasting time analyzing most of these cases. It only "succeeds"
8069 // when the mapping was different and the load can be replaced with
8070 // a mov, which is of negligible benefit. So such cases are
8072 if(f_regmap[hr]>0) {
8073 if(regs[t].regmap[hr]==f_regmap[hr]||(regs[t].regmap_entry[hr]<0&&get_reg(regmap_pre[t],f_regmap[hr])<0)) {
8077 //printf("Test %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8078 if(r<34&&((unneeded_reg[j]>>r)&1)) break;
8080 if(regs[j].regmap[hr]==f_regmap[hr]&&(f_regmap[hr]&63)<TEMPREG) {
8081 //printf("Hit %x -> %x, %x %d/%d\n",start+i*4,ba[i],start+j*4,hr,r);
8083 if(regs[i].regmap[hr]==-1&&branch_regs[i].regmap[hr]==-1) {
8084 if(get_reg(regs[i+2].regmap,f_regmap[hr])>=0) break;
8086 if(get_reg(regs[i].regmap,r&63)<0) break;
8087 if(get_reg(branch_regs[i].regmap,r&63)<0) break;
8090 while(k>1&®s[k-1].regmap[hr]==-1) {
8091 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8092 //printf("no free regs for store %x\n",start+(k-1)*4);
8095 if(get_reg(regs[k-1].regmap,f_regmap[hr])>=0) {
8096 //printf("no-match due to different register\n");
8099 if(itype[k-2]==UJUMP||itype[k-2]==RJUMP||itype[k-2]==CJUMP||itype[k-2]==SJUMP) {
8100 //printf("no-match due to branch\n");
8103 // call/ret fast path assumes no registers allocated
8104 if(k>2&&(itype[k-3]==UJUMP||itype[k-3]==RJUMP)&&rt1[k-3]==31) {
8110 if(regs[k-1].regmap[hr]==f_regmap[hr]&®map_pre[k][hr]==f_regmap[hr]) {
8111 //printf("Extend r%d, %x ->\n",hr,start+k*4);
8113 regs[k].regmap_entry[hr]=f_regmap[hr];
8114 regs[k].regmap[hr]=f_regmap[hr];
8115 regmap_pre[k+1][hr]=f_regmap[hr];
8116 regs[k].wasdirty&=~(1<<hr);
8117 regs[k].dirty&=~(1<<hr);
8118 regs[k].wasdirty|=(1<<hr)®s[k-1].dirty;
8119 regs[k].dirty|=(1<<hr)®s[k].wasdirty;
8120 regs[k].wasconst&=~(1<<hr);
8121 regs[k].isconst&=~(1<<hr);
8126 //printf("Fail Extend r%d, %x ->\n",hr,start+k*4);
8129 assert(regs[i-1].regmap[hr]==f_regmap[hr]);
8130 if(regs[i-1].regmap[hr]==f_regmap[hr]&®map_pre[i][hr]==f_regmap[hr]) {
8131 //printf("OK fill %x (r%d)\n",start+i*4,hr);
8132 regs[i].regmap_entry[hr]=f_regmap[hr];
8133 regs[i].regmap[hr]=f_regmap[hr];
8134 regs[i].wasdirty&=~(1<<hr);
8135 regs[i].dirty&=~(1<<hr);
8136 regs[i].wasdirty|=(1<<hr)®s[i-1].dirty;
8137 regs[i].dirty|=(1<<hr)®s[i-1].dirty;
8138 regs[i].wasconst&=~(1<<hr);
8139 regs[i].isconst&=~(1<<hr);
8140 branch_regs[i].regmap_entry[hr]=f_regmap[hr];
8141 branch_regs[i].wasdirty&=~(1<<hr);
8142 branch_regs[i].wasdirty|=(1<<hr)®s[i].dirty;
8143 branch_regs[i].regmap[hr]=f_regmap[hr];
8144 branch_regs[i].dirty&=~(1<<hr);
8145 branch_regs[i].dirty|=(1<<hr)®s[i].dirty;
8146 branch_regs[i].wasconst&=~(1<<hr);
8147 branch_regs[i].isconst&=~(1<<hr);
8148 if(itype[i]!=RJUMP&&itype[i]!=UJUMP&&(source[i]>>16)!=0x1000) {
8149 regmap_pre[i+2][hr]=f_regmap[hr];
8150 regs[i+2].wasdirty&=~(1<<hr);
8151 regs[i+2].wasdirty|=(1<<hr)®s[i].dirty;
8156 // Alloc register clean at beginning of loop,
8157 // but may dirty it in pass 6
8158 regs[k].regmap_entry[hr]=f_regmap[hr];
8159 regs[k].regmap[hr]=f_regmap[hr];
8160 regs[k].dirty&=~(1<<hr);
8161 regs[k].wasconst&=~(1<<hr);
8162 regs[k].isconst&=~(1<<hr);
8163 if(itype[k]==UJUMP||itype[k]==RJUMP||itype[k]==CJUMP||itype[k]==SJUMP) {
8164 branch_regs[k].regmap_entry[hr]=f_regmap[hr];
8165 branch_regs[k].regmap[hr]=f_regmap[hr];
8166 branch_regs[k].dirty&=~(1<<hr);
8167 branch_regs[k].wasconst&=~(1<<hr);
8168 branch_regs[k].isconst&=~(1<<hr);
8169 if(itype[k]!=RJUMP&&itype[k]!=UJUMP&&(source[k]>>16)!=0x1000) {
8170 regmap_pre[k+2][hr]=f_regmap[hr];
8171 regs[k+2].wasdirty&=~(1<<hr);
8176 regmap_pre[k+1][hr]=f_regmap[hr];
8177 regs[k+1].wasdirty&=~(1<<hr);
8180 if(regs[j].regmap[hr]==f_regmap[hr])
8181 regs[j].regmap_entry[hr]=f_regmap[hr];
8185 if(regs[j].regmap[hr]>=0)
8187 if(get_reg(regs[j].regmap,f_regmap[hr])>=0) {
8188 //printf("no-match due to different register\n");
8191 if(itype[j]==UJUMP||itype[j]==RJUMP||(source[j]>>16)==0x1000)
8193 // Stop on unconditional branch
8196 if(itype[j]==CJUMP||itype[j]==SJUMP)
8199 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j+1])
8202 if(count_free_regs(branch_regs[j].regmap)<=minimum_free_regs[j+1])
8205 if(get_reg(branch_regs[j].regmap,f_regmap[hr])>=0) {
8206 //printf("no-match due to different register (branch)\n");
8210 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8211 //printf("No free regs for store %x\n",start+j*4);
8214 assert(f_regmap[hr]<64);
8221 // Non branch or undetermined branch target
8222 for(hr=0;hr<HOST_REGS;hr++)
8224 if(hr!=EXCLUDE_REG) {
8225 if(regs[i].regmap[hr]>64) {
8226 if(!((regs[i].dirty>>hr)&1))
8227 f_regmap[hr]=regs[i].regmap[hr];
8229 else if(regs[i].regmap[hr]>=0) {
8230 if(f_regmap[hr]!=regs[i].regmap[hr]) {
8231 // dealloc old register
8233 for(n=0;n<HOST_REGS;n++)
8235 if(f_regmap[n]==regs[i].regmap[hr]) {f_regmap[n]=-1;}
8237 // and alloc new one
8238 f_regmap[hr]=regs[i].regmap[hr];
8243 // Try to restore cycle count at branch targets
8245 for(j=i;j<slen-1;j++) {
8246 if(regs[j].regmap[HOST_CCREG]!=-1) break;
8247 if(count_free_regs(regs[j].regmap)<=minimum_free_regs[j]) {
8248 //printf("no free regs for store %x\n",start+j*4);
8252 if(regs[j].regmap[HOST_CCREG]==CCREG) {
8254 //printf("Extend CC, %x -> %x\n",start+k*4,start+j*4);
8256 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8257 regs[k].regmap[HOST_CCREG]=CCREG;
8258 regmap_pre[k+1][HOST_CCREG]=CCREG;
8259 regs[k+1].wasdirty|=1<<HOST_CCREG;
8260 regs[k].dirty|=1<<HOST_CCREG;
8261 regs[k].wasconst&=~(1<<HOST_CCREG);
8262 regs[k].isconst&=~(1<<HOST_CCREG);
8265 regs[j].regmap_entry[HOST_CCREG]=CCREG;
8267 // Work backwards from the branch target
8268 if(j>i&&f_regmap[HOST_CCREG]==CCREG)
8270 //printf("Extend backwards\n");
8273 while(regs[k-1].regmap[HOST_CCREG]==-1) {
8274 if(count_free_regs(regs[k-1].regmap)<=minimum_free_regs[k-1]) {
8275 //printf("no free regs for store %x\n",start+(k-1)*4);
8280 if(regs[k-1].regmap[HOST_CCREG]==CCREG) {
8281 //printf("Extend CC, %x ->\n",start+k*4);
8283 regs[k].regmap_entry[HOST_CCREG]=CCREG;
8284 regs[k].regmap[HOST_CCREG]=CCREG;
8285 regmap_pre[k+1][HOST_CCREG]=CCREG;
8286 regs[k+1].wasdirty|=1<<HOST_CCREG;
8287 regs[k].dirty|=1<<HOST_CCREG;
8288 regs[k].wasconst&=~(1<<HOST_CCREG);
8289 regs[k].isconst&=~(1<<HOST_CCREG);
8294 //printf("Fail Extend CC, %x ->\n",start+k*4);
8298 if(itype[i]!=STORE&&itype[i]!=STORELR&&itype[i]!=C1LS&&itype[i]!=SHIFT&&
8299 itype[i]!=NOP&&itype[i]!=MOV&&itype[i]!=ALU&&itype[i]!=SHIFTIMM&&
8300 itype[i]!=IMM16&&itype[i]!=LOAD&&itype[i]!=COP1)
8302 memcpy(f_regmap,regs[i].regmap,sizeof(f_regmap));
8307 // This allocates registers (if possible) one instruction prior
8308 // to use, which can avoid a load-use penalty on certain CPUs.
8309 for(i=0;i<slen-1;i++)
8311 if(!i||(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP))
8315 if(itype[i]==ALU||itype[i]==MOV||itype[i]==LOAD||itype[i]==SHIFTIMM||itype[i]==IMM16
8316 ||((itype[i]==COP1||itype[i]==COP2)&&opcode2[i]<3))
8319 if((hr=get_reg(regs[i+1].regmap,rs1[i+1]))>=0)
8321 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8323 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8324 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8325 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8326 regs[i].isconst&=~(1<<hr);
8327 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8328 constmap[i][hr]=constmap[i+1][hr];
8329 regs[i+1].wasdirty&=~(1<<hr);
8330 regs[i].dirty&=~(1<<hr);
8335 if((hr=get_reg(regs[i+1].regmap,rs2[i+1]))>=0)
8337 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8339 regs[i].regmap[hr]=regs[i+1].regmap[hr];
8340 regmap_pre[i+1][hr]=regs[i+1].regmap[hr];
8341 regs[i+1].regmap_entry[hr]=regs[i+1].regmap[hr];
8342 regs[i].isconst&=~(1<<hr);
8343 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8344 constmap[i][hr]=constmap[i+1][hr];
8345 regs[i+1].wasdirty&=~(1<<hr);
8346 regs[i].dirty&=~(1<<hr);
8350 // Preload target address for load instruction (non-constant)
8351 if(itype[i+1]==LOAD&&rs1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
8352 if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
8354 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8356 regs[i].regmap[hr]=rs1[i+1];
8357 regmap_pre[i+1][hr]=rs1[i+1];
8358 regs[i+1].regmap_entry[hr]=rs1[i+1];
8359 regs[i].isconst&=~(1<<hr);
8360 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8361 constmap[i][hr]=constmap[i+1][hr];
8362 regs[i+1].wasdirty&=~(1<<hr);
8363 regs[i].dirty&=~(1<<hr);
8367 // Load source into target register
8368 if(lt1[i+1]&&get_reg(regs[i+1].regmap,rs1[i+1])<0) {
8369 if((hr=get_reg(regs[i+1].regmap,rt1[i+1]))>=0)
8371 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8373 regs[i].regmap[hr]=rs1[i+1];
8374 regmap_pre[i+1][hr]=rs1[i+1];
8375 regs[i+1].regmap_entry[hr]=rs1[i+1];
8376 regs[i].isconst&=~(1<<hr);
8377 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8378 constmap[i][hr]=constmap[i+1][hr];
8379 regs[i+1].wasdirty&=~(1<<hr);
8380 regs[i].dirty&=~(1<<hr);
8384 // Address for store instruction (non-constant)
8385 if(itype[i+1]==STORE||itype[i+1]==STORELR
8386 ||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SB/SH/SW/SD/SWC1/SDC1/SWC2/SDC2
8387 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
8388 hr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1);
8389 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8390 else {regs[i+1].regmap[hr]=AGEN1+((i+1)&1);regs[i+1].isconst&=~(1<<hr);}
8392 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8394 regs[i].regmap[hr]=rs1[i+1];
8395 regmap_pre[i+1][hr]=rs1[i+1];
8396 regs[i+1].regmap_entry[hr]=rs1[i+1];
8397 regs[i].isconst&=~(1<<hr);
8398 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8399 constmap[i][hr]=constmap[i+1][hr];
8400 regs[i+1].wasdirty&=~(1<<hr);
8401 regs[i].dirty&=~(1<<hr);
8405 if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) { // LWC1/LDC1, LWC2/LDC2
8406 if(get_reg(regs[i+1].regmap,rs1[i+1])<0) {
8408 hr=get_reg(regs[i+1].regmap,FTEMP);
8410 if(regs[i].regmap[hr]<0&®s[i+1].regmap_entry[hr]<0)
8412 regs[i].regmap[hr]=rs1[i+1];
8413 regmap_pre[i+1][hr]=rs1[i+1];
8414 regs[i+1].regmap_entry[hr]=rs1[i+1];
8415 regs[i].isconst&=~(1<<hr);
8416 regs[i].isconst|=regs[i+1].isconst&(1<<hr);
8417 constmap[i][hr]=constmap[i+1][hr];
8418 regs[i+1].wasdirty&=~(1<<hr);
8419 regs[i].dirty&=~(1<<hr);
8421 else if((nr=get_reg2(regs[i].regmap,regs[i+1].regmap,-1))>=0)
8423 // move it to another register
8424 regs[i+1].regmap[hr]=-1;
8425 regmap_pre[i+2][hr]=-1;
8426 regs[i+1].regmap[nr]=FTEMP;
8427 regmap_pre[i+2][nr]=FTEMP;
8428 regs[i].regmap[nr]=rs1[i+1];
8429 regmap_pre[i+1][nr]=rs1[i+1];
8430 regs[i+1].regmap_entry[nr]=rs1[i+1];
8431 regs[i].isconst&=~(1<<nr);
8432 regs[i+1].isconst&=~(1<<nr);
8433 regs[i].dirty&=~(1<<nr);
8434 regs[i+1].wasdirty&=~(1<<nr);
8435 regs[i+1].dirty&=~(1<<nr);
8436 regs[i+2].wasdirty&=~(1<<nr);
8440 if(itype[i+1]==LOAD||itype[i+1]==LOADLR||itype[i+1]==STORE||itype[i+1]==STORELR/*||itype[i+1]==C1LS||||itype[i+1]==C2LS*/) {
8441 if(itype[i+1]==LOAD)
8442 hr=get_reg(regs[i+1].regmap,rt1[i+1]);
8443 if(itype[i+1]==LOADLR||(opcode[i+1]&0x3b)==0x31||(opcode[i+1]&0x3b)==0x32) // LWC1/LDC1, LWC2/LDC2
8444 hr=get_reg(regs[i+1].regmap,FTEMP);
8445 if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a) { // SWC1/SDC1/SWC2/SDC2
8446 hr=get_reg(regs[i+1].regmap,AGEN1+((i+1)&1));
8447 if(hr<0) hr=get_reg(regs[i+1].regmap,-1);
8449 if(hr>=0&®s[i].regmap[hr]<0) {
8450 int rs=get_reg(regs[i+1].regmap,rs1[i+1]);
8451 if(rs>=0&&((regs[i+1].wasconst>>rs)&1)) {
8452 regs[i].regmap[hr]=AGEN1+((i+1)&1);
8453 regmap_pre[i+1][hr]=AGEN1+((i+1)&1);
8454 regs[i+1].regmap_entry[hr]=AGEN1+((i+1)&1);
8455 regs[i].isconst&=~(1<<hr);
8456 regs[i+1].wasdirty&=~(1<<hr);
8457 regs[i].dirty&=~(1<<hr);
8466 /* Pass 6 - Optimize clean/dirty state */
8467 clean_registers(0,slen-1,1);
8469 /* Pass 7 - Identify 32-bit registers */
8470 for (i=slen-1;i>=0;i--)
8472 if(itype[i]==CJUMP||itype[i]==SJUMP)
8474 // Conditional branch
8475 if((source[i]>>16)!=0x1000&&i<slen-2) {
8476 // Mark this address as a branch target since it may be called
8477 // upon return from interrupt
8483 if(itype[slen-1]==SPAN) {
8484 bt[slen-1]=1; // Mark as a branch target so instruction can restart after exception
8488 /* Debug/disassembly */
8493 for(r=1;r<=CCREG;r++) {
8494 if((unneeded_reg[i]>>r)&1) {
8495 if(r==HIREG) printf(" HI");
8496 else if(r==LOREG) printf(" LO");
8497 else printf(" r%d",r);
8501 #if defined(__i386__) || defined(__x86_64__)
8502 printf("pre: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7]);
8505 printf("pre: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regmap_pre[i][0],regmap_pre[i][1],regmap_pre[i][2],regmap_pre[i][3],regmap_pre[i][4],regmap_pre[i][5],regmap_pre[i][6],regmap_pre[i][7],regmap_pre[i][8],regmap_pre[i][9],regmap_pre[i][10],regmap_pre[i][12]);
8508 if(needed_reg[i]&1) printf("eax ");
8509 if((needed_reg[i]>>1)&1) printf("ecx ");
8510 if((needed_reg[i]>>2)&1) printf("edx ");
8511 if((needed_reg[i]>>3)&1) printf("ebx ");
8512 if((needed_reg[i]>>5)&1) printf("ebp ");
8513 if((needed_reg[i]>>6)&1) printf("esi ");
8514 if((needed_reg[i]>>7)&1) printf("edi ");
8516 #if defined(__i386__) || defined(__x86_64__)
8517 printf("entry: eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7]);
8519 if(regs[i].wasdirty&1) printf("eax ");
8520 if((regs[i].wasdirty>>1)&1) printf("ecx ");
8521 if((regs[i].wasdirty>>2)&1) printf("edx ");
8522 if((regs[i].wasdirty>>3)&1) printf("ebx ");
8523 if((regs[i].wasdirty>>5)&1) printf("ebp ");
8524 if((regs[i].wasdirty>>6)&1) printf("esi ");
8525 if((regs[i].wasdirty>>7)&1) printf("edi ");
8528 printf("entry: r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d\n",regs[i].regmap_entry[0],regs[i].regmap_entry[1],regs[i].regmap_entry[2],regs[i].regmap_entry[3],regs[i].regmap_entry[4],regs[i].regmap_entry[5],regs[i].regmap_entry[6],regs[i].regmap_entry[7],regs[i].regmap_entry[8],regs[i].regmap_entry[9],regs[i].regmap_entry[10],regs[i].regmap_entry[12]);
8530 if(regs[i].wasdirty&1) printf("r0 ");
8531 if((regs[i].wasdirty>>1)&1) printf("r1 ");
8532 if((regs[i].wasdirty>>2)&1) printf("r2 ");
8533 if((regs[i].wasdirty>>3)&1) printf("r3 ");
8534 if((regs[i].wasdirty>>4)&1) printf("r4 ");
8535 if((regs[i].wasdirty>>5)&1) printf("r5 ");
8536 if((regs[i].wasdirty>>6)&1) printf("r6 ");
8537 if((regs[i].wasdirty>>7)&1) printf("r7 ");
8538 if((regs[i].wasdirty>>8)&1) printf("r8 ");
8539 if((regs[i].wasdirty>>9)&1) printf("r9 ");
8540 if((regs[i].wasdirty>>10)&1) printf("r10 ");
8541 if((regs[i].wasdirty>>12)&1) printf("r12 ");
8544 disassemble_inst(i);
8545 //printf ("ccadj[%d] = %d\n",i,ccadj[i]);
8546 #if defined(__i386__) || defined(__x86_64__)
8547 printf("eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7]);
8548 if(regs[i].dirty&1) printf("eax ");
8549 if((regs[i].dirty>>1)&1) printf("ecx ");
8550 if((regs[i].dirty>>2)&1) printf("edx ");
8551 if((regs[i].dirty>>3)&1) printf("ebx ");
8552 if((regs[i].dirty>>5)&1) printf("ebp ");
8553 if((regs[i].dirty>>6)&1) printf("esi ");
8554 if((regs[i].dirty>>7)&1) printf("edi ");
8557 printf("r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",regs[i].regmap[0],regs[i].regmap[1],regs[i].regmap[2],regs[i].regmap[3],regs[i].regmap[4],regs[i].regmap[5],regs[i].regmap[6],regs[i].regmap[7],regs[i].regmap[8],regs[i].regmap[9],regs[i].regmap[10],regs[i].regmap[12]);
8558 if(regs[i].dirty&1) printf("r0 ");
8559 if((regs[i].dirty>>1)&1) printf("r1 ");
8560 if((regs[i].dirty>>2)&1) printf("r2 ");
8561 if((regs[i].dirty>>3)&1) printf("r3 ");
8562 if((regs[i].dirty>>4)&1) printf("r4 ");
8563 if((regs[i].dirty>>5)&1) printf("r5 ");
8564 if((regs[i].dirty>>6)&1) printf("r6 ");
8565 if((regs[i].dirty>>7)&1) printf("r7 ");
8566 if((regs[i].dirty>>8)&1) printf("r8 ");
8567 if((regs[i].dirty>>9)&1) printf("r9 ");
8568 if((regs[i].dirty>>10)&1) printf("r10 ");
8569 if((regs[i].dirty>>12)&1) printf("r12 ");
8572 if(regs[i].isconst) {
8573 printf("constants: ");
8574 #if defined(__i386__) || defined(__x86_64__)
8575 if(regs[i].isconst&1) printf("eax=%x ",(u_int)constmap[i][0]);
8576 if((regs[i].isconst>>1)&1) printf("ecx=%x ",(u_int)constmap[i][1]);
8577 if((regs[i].isconst>>2)&1) printf("edx=%x ",(u_int)constmap[i][2]);
8578 if((regs[i].isconst>>3)&1) printf("ebx=%x ",(u_int)constmap[i][3]);
8579 if((regs[i].isconst>>5)&1) printf("ebp=%x ",(u_int)constmap[i][5]);
8580 if((regs[i].isconst>>6)&1) printf("esi=%x ",(u_int)constmap[i][6]);
8581 if((regs[i].isconst>>7)&1) printf("edi=%x ",(u_int)constmap[i][7]);
8585 for (r = 0; r < ARRAY_SIZE(constmap[i]); r++)
8586 if ((regs[i].isconst >> r) & 1)
8587 printf(" r%d=%x", r, (u_int)constmap[i][r]);
8591 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP) {
8592 #if defined(__i386__) || defined(__x86_64__)
8593 printf("branch(%d): eax=%d ecx=%d edx=%d ebx=%d ebp=%d esi=%d edi=%d dirty: ",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7]);
8594 if(branch_regs[i].dirty&1) printf("eax ");
8595 if((branch_regs[i].dirty>>1)&1) printf("ecx ");
8596 if((branch_regs[i].dirty>>2)&1) printf("edx ");
8597 if((branch_regs[i].dirty>>3)&1) printf("ebx ");
8598 if((branch_regs[i].dirty>>5)&1) printf("ebp ");
8599 if((branch_regs[i].dirty>>6)&1) printf("esi ");
8600 if((branch_regs[i].dirty>>7)&1) printf("edi ");
8603 printf("branch(%d): r0=%d r1=%d r2=%d r3=%d r4=%d r5=%d r6=%d r7=%d r8=%d r9=%d r10=%d r12=%d dirty: ",i,branch_regs[i].regmap[0],branch_regs[i].regmap[1],branch_regs[i].regmap[2],branch_regs[i].regmap[3],branch_regs[i].regmap[4],branch_regs[i].regmap[5],branch_regs[i].regmap[6],branch_regs[i].regmap[7],branch_regs[i].regmap[8],branch_regs[i].regmap[9],branch_regs[i].regmap[10],branch_regs[i].regmap[12]);
8604 if(branch_regs[i].dirty&1) printf("r0 ");
8605 if((branch_regs[i].dirty>>1)&1) printf("r1 ");
8606 if((branch_regs[i].dirty>>2)&1) printf("r2 ");
8607 if((branch_regs[i].dirty>>3)&1) printf("r3 ");
8608 if((branch_regs[i].dirty>>4)&1) printf("r4 ");
8609 if((branch_regs[i].dirty>>5)&1) printf("r5 ");
8610 if((branch_regs[i].dirty>>6)&1) printf("r6 ");
8611 if((branch_regs[i].dirty>>7)&1) printf("r7 ");
8612 if((branch_regs[i].dirty>>8)&1) printf("r8 ");
8613 if((branch_regs[i].dirty>>9)&1) printf("r9 ");
8614 if((branch_regs[i].dirty>>10)&1) printf("r10 ");
8615 if((branch_regs[i].dirty>>12)&1) printf("r12 ");
8621 /* Pass 8 - Assembly */
8622 linkcount=0;stubcount=0;
8623 ds=0;is_delayslot=0;
8625 void *beginning=start_block();
8630 void *instr_addr0_override = NULL;
8632 if (start == 0x80030000) {
8633 // nasty hack for fastbios thing
8634 // override block entry to this code
8635 instr_addr0_override = out;
8636 emit_movimm(start,0);
8637 // abuse io address var as a flag that we
8638 // have already returned here once
8639 emit_readword(&address,1);
8640 emit_writeword(0,&pcaddr);
8641 emit_writeword(0,&address);
8643 emit_jne(new_dyna_leave);
8647 //if(ds) printf("ds: ");
8648 disassemble_inst(i);
8650 ds=0; // Skip delay slot
8651 if(bt[i]) assem_debug("OOPS - branch into delay slot\n");
8652 instr_addr[i] = NULL;
8654 speculate_register_values(i);
8655 #ifndef DESTRUCTIVE_WRITEBACK
8656 if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
8658 wb_valid(regmap_pre[i],regs[i].regmap_entry,dirty_pre,regs[i].wasdirty,unneeded_reg[i]);
8660 if((itype[i]==CJUMP||itype[i]==SJUMP)&&!likely[i]) {
8661 dirty_pre=branch_regs[i].dirty;
8663 dirty_pre=regs[i].dirty;
8667 if(i<2||(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000))
8669 wb_invalidate(regmap_pre[i],regs[i].regmap_entry,regs[i].wasdirty,unneeded_reg[i]);
8670 loop_preload(regmap_pre[i],regs[i].regmap_entry);
8672 // branch target entry point
8673 instr_addr[i] = out;
8674 assem_debug("<->\n");
8675 drc_dbg_emit_do_cmp(i);
8678 if(regs[i].regmap_entry[HOST_CCREG]==CCREG&®s[i].regmap[HOST_CCREG]!=CCREG)
8679 wb_register(CCREG,regs[i].regmap_entry,regs[i].wasdirty);
8680 load_regs(regs[i].regmap_entry,regs[i].regmap,rs1[i],rs2[i]);
8681 address_generation(i,®s[i],regs[i].regmap_entry);
8682 load_consts(regmap_pre[i],regs[i].regmap,i);
8683 if(itype[i]==RJUMP||itype[i]==UJUMP||itype[i]==CJUMP||itype[i]==SJUMP)
8685 // Load the delay slot registers if necessary
8686 if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i]&&(rs1[i+1]!=rt1[i]||rt1[i]==0))
8687 load_regs(regs[i].regmap_entry,regs[i].regmap,rs1[i+1],rs1[i+1]);
8688 if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i]&&(rs2[i+1]!=rt1[i]||rt1[i]==0))
8689 load_regs(regs[i].regmap_entry,regs[i].regmap,rs2[i+1],rs2[i+1]);
8690 if(itype[i+1]==STORE||itype[i+1]==STORELR||(opcode[i+1]&0x3b)==0x39||(opcode[i+1]&0x3b)==0x3a)
8691 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
8695 // Preload registers for following instruction
8696 if(rs1[i+1]!=rs1[i]&&rs1[i+1]!=rs2[i])
8697 if(rs1[i+1]!=rt1[i]&&rs1[i+1]!=rt2[i])
8698 load_regs(regs[i].regmap_entry,regs[i].regmap,rs1[i+1],rs1[i+1]);
8699 if(rs2[i+1]!=rs1[i+1]&&rs2[i+1]!=rs1[i]&&rs2[i+1]!=rs2[i])
8700 if(rs2[i+1]!=rt1[i]&&rs2[i+1]!=rt2[i])
8701 load_regs(regs[i].regmap_entry,regs[i].regmap,rs2[i+1],rs2[i+1]);
8703 // TODO: if(is_ooo(i)) address_generation(i+1);
8705 load_regs(regs[i].regmap_entry,regs[i].regmap,CCREG,CCREG);
8706 if(itype[i]==STORE||itype[i]==STORELR||(opcode[i]&0x3b)==0x39||(opcode[i]&0x3b)==0x3a)
8707 load_regs(regs[i].regmap_entry,regs[i].regmap,INVCP,INVCP);
8711 alu_assemble(i,®s[i]);break;
8713 imm16_assemble(i,®s[i]);break;
8715 shift_assemble(i,®s[i]);break;
8717 shiftimm_assemble(i,®s[i]);break;
8719 load_assemble(i,®s[i]);break;
8721 loadlr_assemble(i,®s[i]);break;
8723 store_assemble(i,®s[i]);break;
8725 storelr_assemble(i,®s[i]);break;
8727 cop0_assemble(i,®s[i]);break;
8729 cop1_assemble(i,®s[i]);break;
8731 c1ls_assemble(i,®s[i]);break;
8733 cop2_assemble(i,®s[i]);break;
8735 c2ls_assemble(i,®s[i]);break;
8737 c2op_assemble(i,®s[i]);break;
8739 multdiv_assemble(i,®s[i]);break;
8741 mov_assemble(i,®s[i]);break;
8743 syscall_assemble(i,®s[i]);break;
8745 hlecall_assemble(i,®s[i]);break;
8747 intcall_assemble(i,®s[i]);break;
8749 ujump_assemble(i,®s[i]);ds=1;break;
8751 rjump_assemble(i,®s[i]);ds=1;break;
8753 cjump_assemble(i,®s[i]);ds=1;break;
8755 sjump_assemble(i,®s[i]);ds=1;break;
8757 pagespan_assemble(i,®s[i]);break;
8759 if(itype[i]==UJUMP||itype[i]==RJUMP||(source[i]>>16)==0x1000)
8762 literal_pool_jumpover(256);
8765 //assert(itype[i-2]==UJUMP||itype[i-2]==RJUMP||(source[i-2]>>16)==0x1000);
8766 // If the block did not end with an unconditional branch,
8767 // add a jump to the next instruction.
8769 if(itype[i-2]!=UJUMP&&itype[i-2]!=RJUMP&&(source[i-2]>>16)!=0x1000&&itype[i-1]!=SPAN) {
8770 assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP);
8772 if(itype[i-2]!=CJUMP&&itype[i-2]!=SJUMP) {
8773 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
8774 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
8775 emit_loadreg(CCREG,HOST_CCREG);
8776 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
8778 else if(!likely[i-2])
8780 store_regs_bt(branch_regs[i-2].regmap,branch_regs[i-2].dirty,start+i*4);
8781 assert(branch_regs[i-2].regmap[HOST_CCREG]==CCREG);
8785 store_regs_bt(regs[i-2].regmap,regs[i-2].dirty,start+i*4);
8786 assert(regs[i-2].regmap[HOST_CCREG]==CCREG);
8788 add_to_linker(out,start+i*4,0);
8795 assert(itype[i-1]!=UJUMP&&itype[i-1]!=CJUMP&&itype[i-1]!=SJUMP&&itype[i-1]!=RJUMP);
8796 store_regs_bt(regs[i-1].regmap,regs[i-1].dirty,start+i*4);
8797 if(regs[i-1].regmap[HOST_CCREG]!=CCREG)
8798 emit_loadreg(CCREG,HOST_CCREG);
8799 emit_addimm(HOST_CCREG,CLOCK_ADJUST(ccadj[i-1]+1),HOST_CCREG);
8800 add_to_linker(out,start+i*4,0);
8804 // TODO: delay slot stubs?
8806 for(i=0;i<stubcount;i++)
8808 switch(stubs[i].type)
8816 do_readstub(i);break;
8821 do_writestub(i);break;
8825 do_invstub(i);break;
8827 do_cop1stub(i);break;
8829 do_unalignedwritestub(i);break;
8833 if (instr_addr0_override)
8834 instr_addr[0] = instr_addr0_override;
8836 /* Pass 9 - Linker */
8837 for(i=0;i<linkcount;i++)
8839 assem_debug("%p -> %8x\n",link_addr[i].addr,link_addr[i].target);
8841 if (!link_addr[i].ext)
8844 void *addr = check_addr(link_addr[i].target);
8845 emit_extjump(link_addr[i].addr, link_addr[i].target);
8847 set_jump_target(link_addr[i].addr, addr);
8848 add_link(link_addr[i].target,stub);
8851 set_jump_target(link_addr[i].addr, stub);
8856 int target=(link_addr[i].target-start)>>2;
8857 assert(target>=0&&target<slen);
8858 assert(instr_addr[target]);
8859 //#ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
8860 //set_jump_target_fillslot(link_addr[i].addr,instr_addr[target],link_addr[i].ext>>1);
8862 set_jump_target(link_addr[i].addr, instr_addr[target]);
8866 // External Branch Targets (jump_in)
8867 if(copy+slen*4>(void *)shadow+sizeof(shadow)) copy=shadow;
8872 if(instr_addr[i]) // TODO - delay slots (=null)
8874 u_int vaddr=start+i*4;
8875 u_int page=get_page(vaddr);
8876 u_int vpage=get_vpage(vaddr);
8879 assem_debug("%p (%d) <- %8x\n",instr_addr[i],i,start+i*4);
8880 assem_debug("jump_in: %x\n",start+i*4);
8881 ll_add(jump_dirty+vpage,vaddr,out);
8882 void *entry_point = do_dirty_stub(i);
8883 ll_add_flags(jump_in+page,vaddr,state_rflags,entry_point);
8884 // If there was an existing entry in the hash table,
8885 // replace it with the new address.
8886 // Don't add new entries. We'll insert the
8887 // ones that actually get used in check_addr().
8888 struct ht_entry *ht_bin = hash_table_get(vaddr);
8889 if (ht_bin->vaddr[0] == vaddr)
8890 ht_bin->tcaddr[0] = entry_point;
8891 if (ht_bin->vaddr[1] == vaddr)
8892 ht_bin->tcaddr[1] = entry_point;
8897 // Write out the literal pool if necessary
8899 #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK
8901 if(((u_int)out)&7) emit_addnop(13);
8903 assert(out - (u_char *)beginning < MAX_OUTPUT_BLOCK_SIZE);
8904 //printf("shadow buffer: %p-%p\n",copy,(u_char *)copy+slen*4);
8905 memcpy(copy,source,slen*4);
8908 end_block(beginning);
8910 // If we're within 256K of the end of the buffer,
8911 // start over from the beginning. (Is 256K enough?)
8912 if (out > translation_cache+(1<<TARGET_SIZE_2)-MAX_OUTPUT_BLOCK_SIZE)
8913 out = translation_cache;
8915 // Trap writes to any of the pages we compiled
8916 for(i=start>>12;i<=(start+slen*4)>>12;i++) {
8919 inv_code_start=inv_code_end=~0;
8921 // for PCSX we need to mark all mirrors too
8922 if(get_page(start)<(RAM_SIZE>>12))
8923 for(i=start>>12;i<=(start+slen*4)>>12;i++)
8924 invalid_code[((u_int)0x00000000>>12)|(i&0x1ff)]=
8925 invalid_code[((u_int)0x80000000>>12)|(i&0x1ff)]=
8926 invalid_code[((u_int)0xa0000000>>12)|(i&0x1ff)]=0;
8928 /* Pass 10 - Free memory by expiring oldest blocks */
8930 int end=(((out-translation_cache)>>(TARGET_SIZE_2-16))+16384)&65535;
8933 int shift=TARGET_SIZE_2-3; // Divide into 8 blocks
8934 uintptr_t base=(uintptr_t)translation_cache+((expirep>>13)<<shift); // Base address of this block
8935 inv_debug("EXP: Phase %d\n",expirep);
8936 switch((expirep>>11)&3)
8939 // Clear jump_in and jump_dirty
8940 ll_remove_matching_addrs(jump_in+(expirep&2047),base,shift);
8941 ll_remove_matching_addrs(jump_dirty+(expirep&2047),base,shift);
8942 ll_remove_matching_addrs(jump_in+2048+(expirep&2047),base,shift);
8943 ll_remove_matching_addrs(jump_dirty+2048+(expirep&2047),base,shift);
8947 ll_kill_pointers(jump_out[expirep&2047],base,shift);
8948 ll_kill_pointers(jump_out[(expirep&2047)+2048],base,shift);
8953 struct ht_entry *ht_bin = &hash_table[((expirep&2047)<<5)+i];
8954 if (((uintptr_t)ht_bin->tcaddr[1]>>shift) == (base>>shift) ||
8955 (((uintptr_t)ht_bin->tcaddr[1]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
8956 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[1],ht_bin->tcaddr[1]);
8957 ht_bin->vaddr[1] = -1;
8958 ht_bin->tcaddr[1] = NULL;
8960 if (((uintptr_t)ht_bin->tcaddr[0]>>shift) == (base>>shift) ||
8961 (((uintptr_t)ht_bin->tcaddr[0]-MAX_OUTPUT_BLOCK_SIZE)>>shift)==(base>>shift)) {
8962 inv_debug("EXP: Remove hash %x -> %p\n",ht_bin->vaddr[0],ht_bin->tcaddr[0]);
8963 ht_bin->vaddr[0] = ht_bin->vaddr[1];
8964 ht_bin->tcaddr[0] = ht_bin->tcaddr[1];
8965 ht_bin->vaddr[1] = -1;
8966 ht_bin->tcaddr[1] = NULL;
8972 #if defined(__arm__) || defined(__aarch64__)
8973 if((expirep&2047)==0)
8976 ll_remove_matching_addrs(jump_out+(expirep&2047),base,shift);
8977 ll_remove_matching_addrs(jump_out+2048+(expirep&2047),base,shift);
8980 expirep=(expirep+1)&65535;
8985 // vim:shiftwidth=2:expandtab