| | 1 | /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * |
| | 2 | * Mupen64plus/PCSX - assem_arm.c * |
| | 3 | * Copyright (C) 2009-2011 Ari64 * |
| | 4 | * Copyright (C) 2010-2011 GraÅžvydas "notaz" Ignotas * |
| | 5 | * * |
| | 6 | * This program is free software; you can redistribute it and/or modify * |
| | 7 | * it under the terms of the GNU General Public License as published by * |
| | 8 | * the Free Software Foundation; either version 2 of the License, or * |
| | 9 | * (at your option) any later version. * |
| | 10 | * * |
| | 11 | * This program is distributed in the hope that it will be useful, * |
| | 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of * |
| | 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * |
| | 14 | * GNU General Public License for more details. * |
| | 15 | * * |
| | 16 | * You should have received a copy of the GNU General Public License * |
| | 17 | * along with this program; if not, write to the * |
| | 18 | * Free Software Foundation, Inc., * |
| | 19 | * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. * |
| | 20 | * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ |
| | 21 | |
| | 22 | extern int cycle_count; |
| | 23 | extern int last_count; |
| | 24 | extern int pcaddr; |
| | 25 | extern int pending_exception; |
| | 26 | extern int branch_target; |
| | 27 | extern uint64_t readmem_dword; |
| | 28 | #ifdef MUPEN64 |
| | 29 | extern precomp_instr fake_pc; |
| | 30 | #endif |
| | 31 | extern void *dynarec_local; |
| | 32 | extern u_int memory_map[1048576]; |
| | 33 | extern u_int mini_ht[32][2]; |
| | 34 | extern u_int rounding_modes[4]; |
| | 35 | |
| | 36 | void indirect_jump_indexed(); |
| | 37 | void indirect_jump(); |
| | 38 | void do_interrupt(); |
| | 39 | void jump_vaddr_r0(); |
| | 40 | void jump_vaddr_r1(); |
| | 41 | void jump_vaddr_r2(); |
| | 42 | void jump_vaddr_r3(); |
| | 43 | void jump_vaddr_r4(); |
| | 44 | void jump_vaddr_r5(); |
| | 45 | void jump_vaddr_r6(); |
| | 46 | void jump_vaddr_r7(); |
| | 47 | void jump_vaddr_r8(); |
| | 48 | void jump_vaddr_r9(); |
| | 49 | void jump_vaddr_r10(); |
| | 50 | void jump_vaddr_r12(); |
| | 51 | |
| | 52 | const u_int jump_vaddr_reg[16] = { |
| | 53 | (int)jump_vaddr_r0, |
| | 54 | (int)jump_vaddr_r1, |
| | 55 | (int)jump_vaddr_r2, |
| | 56 | (int)jump_vaddr_r3, |
| | 57 | (int)jump_vaddr_r4, |
| | 58 | (int)jump_vaddr_r5, |
| | 59 | (int)jump_vaddr_r6, |
| | 60 | (int)jump_vaddr_r7, |
| | 61 | (int)jump_vaddr_r8, |
| | 62 | (int)jump_vaddr_r9, |
| | 63 | (int)jump_vaddr_r10, |
| | 64 | 0, |
| | 65 | (int)jump_vaddr_r12, |
| | 66 | 0, |
| | 67 | 0, |
| | 68 | 0}; |
| | 69 | |
| | 70 | void invalidate_addr_r0(); |
| | 71 | void invalidate_addr_r1(); |
| | 72 | void invalidate_addr_r2(); |
| | 73 | void invalidate_addr_r3(); |
| | 74 | void invalidate_addr_r4(); |
| | 75 | void invalidate_addr_r5(); |
| | 76 | void invalidate_addr_r6(); |
| | 77 | void invalidate_addr_r7(); |
| | 78 | void invalidate_addr_r8(); |
| | 79 | void invalidate_addr_r9(); |
| | 80 | void invalidate_addr_r10(); |
| | 81 | void invalidate_addr_r12(); |
| | 82 | |
| | 83 | const u_int invalidate_addr_reg[16] = { |
| | 84 | (int)invalidate_addr_r0, |
| | 85 | (int)invalidate_addr_r1, |
| | 86 | (int)invalidate_addr_r2, |
| | 87 | (int)invalidate_addr_r3, |
| | 88 | (int)invalidate_addr_r4, |
| | 89 | (int)invalidate_addr_r5, |
| | 90 | (int)invalidate_addr_r6, |
| | 91 | (int)invalidate_addr_r7, |
| | 92 | (int)invalidate_addr_r8, |
| | 93 | (int)invalidate_addr_r9, |
| | 94 | (int)invalidate_addr_r10, |
| | 95 | 0, |
| | 96 | (int)invalidate_addr_r12, |
| | 97 | 0, |
| | 98 | 0, |
| | 99 | 0}; |
| | 100 | |
| | 101 | #include "fpu.h" |
| | 102 | |
| | 103 | unsigned int needs_clear_cache[1<<(TARGET_SIZE_2-17)]; |
| | 104 | |
| | 105 | /* Linker */ |
| | 106 | |
| | 107 | void set_jump_target(int addr,u_int target) |
| | 108 | { |
| | 109 | u_char *ptr=(u_char *)addr; |
| | 110 | u_int *ptr2=(u_int *)ptr; |
| | 111 | if(ptr[3]==0xe2) { |
| | 112 | assert((target-(u_int)ptr2-8)<1024); |
| | 113 | assert((addr&3)==0); |
| | 114 | assert((target&3)==0); |
| | 115 | *ptr2=(*ptr2&0xFFFFF000)|((target-(u_int)ptr2-8)>>2)|0xF00; |
| | 116 | //printf("target=%x addr=%x insn=%x\n",target,addr,*ptr2); |
| | 117 | } |
| | 118 | else if(ptr[3]==0x72) { |
| | 119 | // generated by emit_jno_unlikely |
| | 120 | if((target-(u_int)ptr2-8)<1024) { |
| | 121 | assert((addr&3)==0); |
| | 122 | assert((target&3)==0); |
| | 123 | *ptr2=(*ptr2&0xFFFFF000)|((target-(u_int)ptr2-8)>>2)|0xF00; |
| | 124 | } |
| | 125 | else if((target-(u_int)ptr2-8)<4096&&!((target-(u_int)ptr2-8)&15)) { |
| | 126 | assert((addr&3)==0); |
| | 127 | assert((target&3)==0); |
| | 128 | *ptr2=(*ptr2&0xFFFFF000)|((target-(u_int)ptr2-8)>>4)|0xE00; |
| | 129 | } |
| | 130 | else *ptr2=(0x7A000000)|(((target-(u_int)ptr2-8)<<6)>>8); |
| | 131 | } |
| | 132 | else { |
| | 133 | assert((ptr[3]&0x0e)==0xa); |
| | 134 | *ptr2=(*ptr2&0xFF000000)|(((target-(u_int)ptr2-8)<<6)>>8); |
| | 135 | } |
| | 136 | } |
| | 137 | |
| | 138 | // This optionally copies the instruction from the target of the branch into |
| | 139 | // the space before the branch. Works, but the difference in speed is |
| | 140 | // usually insignificant. |
| | 141 | void set_jump_target_fillslot(int addr,u_int target,int copy) |
| | 142 | { |
| | 143 | u_char *ptr=(u_char *)addr; |
| | 144 | u_int *ptr2=(u_int *)ptr; |
| | 145 | assert(!copy||ptr2[-1]==0xe28dd000); |
| | 146 | if(ptr[3]==0xe2) { |
| | 147 | assert(!copy); |
| | 148 | assert((target-(u_int)ptr2-8)<4096); |
| | 149 | *ptr2=(*ptr2&0xFFFFF000)|(target-(u_int)ptr2-8); |
| | 150 | } |
| | 151 | else { |
| | 152 | assert((ptr[3]&0x0e)==0xa); |
| | 153 | u_int target_insn=*(u_int *)target; |
| | 154 | if((target_insn&0x0e100000)==0) { // ALU, no immediate, no flags |
| | 155 | copy=0; |
| | 156 | } |
| | 157 | if((target_insn&0x0c100000)==0x04100000) { // Load |
| | 158 | copy=0; |
| | 159 | } |
| | 160 | if(target_insn&0x08000000) { |
| | 161 | copy=0; |
| | 162 | } |
| | 163 | if(copy) { |
| | 164 | ptr2[-1]=target_insn; |
| | 165 | target+=4; |
| | 166 | } |
| | 167 | *ptr2=(*ptr2&0xFF000000)|(((target-(u_int)ptr2-8)<<6)>>8); |
| | 168 | } |
| | 169 | } |
| | 170 | |
| | 171 | /* Literal pool */ |
| | 172 | add_literal(int addr,int val) |
| | 173 | { |
| | 174 | assert(literalcount<sizeof(literals)/sizeof(literals[0])); |
| | 175 | literals[literalcount][0]=addr; |
| | 176 | literals[literalcount][1]=val; |
| | 177 | literalcount++; |
| | 178 | } |
| | 179 | |
| | 180 | void *kill_pointer(void *stub) |
| | 181 | { |
| | 182 | int *ptr=(int *)(stub+4); |
| | 183 | assert((*ptr&0x0ff00000)==0x05900000); |
| | 184 | u_int offset=*ptr&0xfff; |
| | 185 | int **l_ptr=(void *)ptr+offset+8; |
| | 186 | int *i_ptr=*l_ptr; |
| | 187 | set_jump_target((int)i_ptr,(int)stub); |
| | 188 | return i_ptr; |
| | 189 | } |
| | 190 | |
| | 191 | // find where external branch is liked to using addr of it's stub: |
| | 192 | // get address that insn one after stub loads (dyna_linker arg1), |
| | 193 | // treat it as a pointer to branch insn, |
| | 194 | // return addr where that branch jumps to |
| | 195 | int get_pointer(void *stub) |
| | 196 | { |
| | 197 | //printf("get_pointer(%x)\n",(int)stub); |
| | 198 | int *ptr=(int *)(stub+4); |
| | 199 | assert((*ptr&0x0fff0000)==0x059f0000); |
| | 200 | u_int offset=*ptr&0xfff; |
| | 201 | int **l_ptr=(void *)ptr+offset+8; |
| | 202 | int *i_ptr=*l_ptr; |
| | 203 | assert((*i_ptr&0x0f000000)==0x0a000000); |
| | 204 | return (int)i_ptr+((*i_ptr<<8)>>6)+8; |
| | 205 | } |
| | 206 | |
| | 207 | // Find the "clean" entry point from a "dirty" entry point |
| | 208 | // by skipping past the call to verify_code |
| | 209 | u_int get_clean_addr(int addr) |
| | 210 | { |
| | 211 | int *ptr=(int *)addr; |
| | 212 | #ifdef ARMv5_ONLY |
| | 213 | ptr+=4; |
| | 214 | #else |
| | 215 | ptr+=6; |
| | 216 | #endif |
| | 217 | if((*ptr&0xFF000000)!=0xeb000000) ptr++; |
| | 218 | assert((*ptr&0xFF000000)==0xeb000000); // bl instruction |
| | 219 | ptr++; |
| | 220 | if((*ptr&0xFF000000)==0xea000000) { |
| | 221 | return (int)ptr+((*ptr<<8)>>6)+8; // follow jump |
| | 222 | } |
| | 223 | return (u_int)ptr; |
| | 224 | } |
| | 225 | |
| | 226 | int verify_dirty(int addr) |
| | 227 | { |
| | 228 | u_int *ptr=(u_int *)addr; |
| | 229 | #ifdef ARMv5_ONLY |
| | 230 | // get from literal pool |
| | 231 | assert((*ptr&0xFFFF0000)==0xe59f0000); |
| | 232 | u_int offset=*ptr&0xfff; |
| | 233 | u_int *l_ptr=(void *)ptr+offset+8; |
| | 234 | u_int source=l_ptr[0]; |
| | 235 | u_int copy=l_ptr[1]; |
| | 236 | u_int len=l_ptr[2]; |
| | 237 | ptr+=4; |
| | 238 | #else |
| | 239 | // ARMv7 movw/movt |
| | 240 | assert((*ptr&0xFFF00000)==0xe3000000); |
| | 241 | u_int source=(ptr[0]&0xFFF)+((ptr[0]>>4)&0xF000)+((ptr[2]<<16)&0xFFF0000)+((ptr[2]<<12)&0xF0000000); |
| | 242 | u_int copy=(ptr[1]&0xFFF)+((ptr[1]>>4)&0xF000)+((ptr[3]<<16)&0xFFF0000)+((ptr[3]<<12)&0xF0000000); |
| | 243 | u_int len=(ptr[4]&0xFFF)+((ptr[4]>>4)&0xF000); |
| | 244 | ptr+=6; |
| | 245 | #endif |
| | 246 | if((*ptr&0xFF000000)!=0xeb000000) ptr++; |
| | 247 | assert((*ptr&0xFF000000)==0xeb000000); // bl instruction |
| | 248 | u_int verifier=(int)ptr+((signed int)(*ptr<<8)>>6)+8; // get target of bl |
| | 249 | if(verifier==(u_int)verify_code_vm||verifier==(u_int)verify_code_ds) { |
| | 250 | unsigned int page=source>>12; |
| | 251 | unsigned int map_value=memory_map[page]; |
| | 252 | if(map_value>=0x80000000) return 0; |
| | 253 | while(page<((source+len-1)>>12)) { |
| | 254 | if((memory_map[++page]<<2)!=(map_value<<2)) return 0; |
| | 255 | } |
| | 256 | source = source+(map_value<<2); |
| | 257 | } |
| | 258 | //printf("verify_dirty: %x %x %x\n",source,copy,len); |
| | 259 | return !memcmp((void *)source,(void *)copy,len); |
| | 260 | } |
| | 261 | |
| | 262 | // This doesn't necessarily find all clean entry points, just |
| | 263 | // guarantees that it's not dirty |
| | 264 | int isclean(int addr) |
| | 265 | { |
| | 266 | #ifdef ARMv5_ONLY |
| | 267 | int *ptr=((u_int *)addr)+4; |
| | 268 | #else |
| | 269 | int *ptr=((u_int *)addr)+6; |
| | 270 | #endif |
| | 271 | if((*ptr&0xFF000000)!=0xeb000000) ptr++; |
| | 272 | if((*ptr&0xFF000000)!=0xeb000000) return 1; // bl instruction |
| | 273 | if((int)ptr+((*ptr<<8)>>6)+8==(int)verify_code) return 0; |
| | 274 | if((int)ptr+((*ptr<<8)>>6)+8==(int)verify_code_vm) return 0; |
| | 275 | if((int)ptr+((*ptr<<8)>>6)+8==(int)verify_code_ds) return 0; |
| | 276 | return 1; |
| | 277 | } |
| | 278 | |
| | 279 | void get_bounds(int addr,u_int *start,u_int *end) |
| | 280 | { |
| | 281 | u_int *ptr=(u_int *)addr; |
| | 282 | #ifdef ARMv5_ONLY |
| | 283 | // get from literal pool |
| | 284 | assert((*ptr&0xFFFF0000)==0xe59f0000); |
| | 285 | u_int offset=*ptr&0xfff; |
| | 286 | u_int *l_ptr=(void *)ptr+offset+8; |
| | 287 | u_int source=l_ptr[0]; |
| | 288 | //u_int copy=l_ptr[1]; |
| | 289 | u_int len=l_ptr[2]; |
| | 290 | ptr+=4; |
| | 291 | #else |
| | 292 | // ARMv7 movw/movt |
| | 293 | assert((*ptr&0xFFF00000)==0xe3000000); |
| | 294 | u_int source=(ptr[0]&0xFFF)+((ptr[0]>>4)&0xF000)+((ptr[2]<<16)&0xFFF0000)+((ptr[2]<<12)&0xF0000000); |
| | 295 | //u_int copy=(ptr[1]&0xFFF)+((ptr[1]>>4)&0xF000)+((ptr[3]<<16)&0xFFF0000)+((ptr[3]<<12)&0xF0000000); |
| | 296 | u_int len=(ptr[4]&0xFFF)+((ptr[4]>>4)&0xF000); |
| | 297 | ptr+=6; |
| | 298 | #endif |
| | 299 | if((*ptr&0xFF000000)!=0xeb000000) ptr++; |
| | 300 | assert((*ptr&0xFF000000)==0xeb000000); // bl instruction |
| | 301 | u_int verifier=(int)ptr+((signed int)(*ptr<<8)>>6)+8; // get target of bl |
| | 302 | if(verifier==(u_int)verify_code_vm||verifier==(u_int)verify_code_ds) { |
| | 303 | if(memory_map[source>>12]>=0x80000000) source = 0; |
| | 304 | else source = source+(memory_map[source>>12]<<2); |
| | 305 | } |
| | 306 | *start=source; |
| | 307 | *end=source+len; |
| | 308 | } |
| | 309 | |
| | 310 | /* Register allocation */ |
| | 311 | |
| | 312 | // Note: registers are allocated clean (unmodified state) |
| | 313 | // if you intend to modify the register, you must call dirty_reg(). |
| | 314 | void alloc_reg(struct regstat *cur,int i,signed char reg) |
| | 315 | { |
| | 316 | int r,hr; |
| | 317 | int preferred_reg = (reg&7); |
| | 318 | if(reg==CCREG) preferred_reg=HOST_CCREG; |
| | 319 | if(reg==PTEMP||reg==FTEMP) preferred_reg=12; |
| | 320 | |
| | 321 | // Don't allocate unused registers |
| | 322 | if((cur->u>>reg)&1) return; |
| | 323 | |
| | 324 | // see if it's already allocated |
| | 325 | for(hr=0;hr<HOST_REGS;hr++) |
| | 326 | { |
| | 327 | if(cur->regmap[hr]==reg) return; |
| | 328 | } |
| | 329 | |
| | 330 | // Keep the same mapping if the register was already allocated in a loop |
| | 331 | preferred_reg = loop_reg(i,reg,preferred_reg); |
| | 332 | |
| | 333 | // Try to allocate the preferred register |
| | 334 | if(cur->regmap[preferred_reg]==-1) { |
| | 335 | cur->regmap[preferred_reg]=reg; |
| | 336 | cur->dirty&=~(1<<preferred_reg); |
| | 337 | cur->isconst&=~(1<<preferred_reg); |
| | 338 | return; |
| | 339 | } |
| | 340 | r=cur->regmap[preferred_reg]; |
| | 341 | if(r<64&&((cur->u>>r)&1)) { |
| | 342 | cur->regmap[preferred_reg]=reg; |
| | 343 | cur->dirty&=~(1<<preferred_reg); |
| | 344 | cur->isconst&=~(1<<preferred_reg); |
| | 345 | return; |
| | 346 | } |
| | 347 | if(r>=64&&((cur->uu>>(r&63))&1)) { |
| | 348 | cur->regmap[preferred_reg]=reg; |
| | 349 | cur->dirty&=~(1<<preferred_reg); |
| | 350 | cur->isconst&=~(1<<preferred_reg); |
| | 351 | return; |
| | 352 | } |
| | 353 | |
| | 354 | // Clear any unneeded registers |
| | 355 | // We try to keep the mapping consistent, if possible, because it |
| | 356 | // makes branches easier (especially loops). So we try to allocate |
| | 357 | // first (see above) before removing old mappings. If this is not |
| | 358 | // possible then go ahead and clear out the registers that are no |
| | 359 | // longer needed. |
| | 360 | for(hr=0;hr<HOST_REGS;hr++) |
| | 361 | { |
| | 362 | r=cur->regmap[hr]; |
| | 363 | if(r>=0) { |
| | 364 | if(r<64) { |
| | 365 | if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;} |
| | 366 | } |
| | 367 | else |
| | 368 | { |
| | 369 | if((cur->uu>>(r&63))&1) {cur->regmap[hr]=-1;break;} |
| | 370 | } |
| | 371 | } |
| | 372 | } |
| | 373 | // Try to allocate any available register, but prefer |
| | 374 | // registers that have not been used recently. |
| | 375 | if(i>0) { |
| | 376 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 377 | if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) { |
| | 378 | 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]) { |
| | 379 | cur->regmap[hr]=reg; |
| | 380 | cur->dirty&=~(1<<hr); |
| | 381 | cur->isconst&=~(1<<hr); |
| | 382 | return; |
| | 383 | } |
| | 384 | } |
| | 385 | } |
| | 386 | } |
| | 387 | // Try to allocate any available register |
| | 388 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 389 | if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) { |
| | 390 | cur->regmap[hr]=reg; |
| | 391 | cur->dirty&=~(1<<hr); |
| | 392 | cur->isconst&=~(1<<hr); |
| | 393 | return; |
| | 394 | } |
| | 395 | } |
| | 396 | |
| | 397 | // Ok, now we have to evict someone |
| | 398 | // Pick a register we hopefully won't need soon |
| | 399 | u_char hsn[MAXREG+1]; |
| | 400 | memset(hsn,10,sizeof(hsn)); |
| | 401 | int j; |
| | 402 | lsn(hsn,i,&preferred_reg); |
| | 403 | //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]); |
| | 404 | //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]); |
| | 405 | if(i>0) { |
| | 406 | // Don't evict the cycle count at entry points, otherwise the entry |
| | 407 | // stub will have to write it. |
| | 408 | if(bt[i]&&hsn[CCREG]>2) hsn[CCREG]=2; |
| | 409 | if(i>1&&hsn[CCREG]>2&&(itype[i-2]==RJUMP||itype[i-2]==UJUMP||itype[i-2]==CJUMP||itype[i-2]==SJUMP||itype[i-2]==FJUMP)) hsn[CCREG]=2; |
| | 410 | for(j=10;j>=3;j--) |
| | 411 | { |
| | 412 | // Alloc preferred register if available |
| | 413 | if(hsn[r=cur->regmap[preferred_reg]&63]==j) { |
| | 414 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 415 | // Evict both parts of a 64-bit register |
| | 416 | if((cur->regmap[hr]&63)==r) { |
| | 417 | cur->regmap[hr]=-1; |
| | 418 | cur->dirty&=~(1<<hr); |
| | 419 | cur->isconst&=~(1<<hr); |
| | 420 | } |
| | 421 | } |
| | 422 | cur->regmap[preferred_reg]=reg; |
| | 423 | return; |
| | 424 | } |
| | 425 | for(r=1;r<=MAXREG;r++) |
| | 426 | { |
| | 427 | if(hsn[r]==j&&r!=rs1[i-1]&&r!=rs2[i-1]&&r!=rt1[i-1]&&r!=rt2[i-1]) { |
| | 428 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 429 | if(hr!=HOST_CCREG||j<hsn[CCREG]) { |
| | 430 | if(cur->regmap[hr]==r+64) { |
| | 431 | cur->regmap[hr]=reg; |
| | 432 | cur->dirty&=~(1<<hr); |
| | 433 | cur->isconst&=~(1<<hr); |
| | 434 | return; |
| | 435 | } |
| | 436 | } |
| | 437 | } |
| | 438 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 439 | if(hr!=HOST_CCREG||j<hsn[CCREG]) { |
| | 440 | if(cur->regmap[hr]==r) { |
| | 441 | cur->regmap[hr]=reg; |
| | 442 | cur->dirty&=~(1<<hr); |
| | 443 | cur->isconst&=~(1<<hr); |
| | 444 | return; |
| | 445 | } |
| | 446 | } |
| | 447 | } |
| | 448 | } |
| | 449 | } |
| | 450 | } |
| | 451 | } |
| | 452 | for(j=10;j>=0;j--) |
| | 453 | { |
| | 454 | for(r=1;r<=MAXREG;r++) |
| | 455 | { |
| | 456 | if(hsn[r]==j) { |
| | 457 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 458 | if(cur->regmap[hr]==r+64) { |
| | 459 | cur->regmap[hr]=reg; |
| | 460 | cur->dirty&=~(1<<hr); |
| | 461 | cur->isconst&=~(1<<hr); |
| | 462 | return; |
| | 463 | } |
| | 464 | } |
| | 465 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 466 | if(cur->regmap[hr]==r) { |
| | 467 | cur->regmap[hr]=reg; |
| | 468 | cur->dirty&=~(1<<hr); |
| | 469 | cur->isconst&=~(1<<hr); |
| | 470 | return; |
| | 471 | } |
| | 472 | } |
| | 473 | } |
| | 474 | } |
| | 475 | } |
| | 476 | printf("This shouldn't happen (alloc_reg)");exit(1); |
| | 477 | } |
| | 478 | |
| | 479 | void alloc_reg64(struct regstat *cur,int i,signed char reg) |
| | 480 | { |
| | 481 | int preferred_reg = 8+(reg&1); |
| | 482 | int r,hr; |
| | 483 | |
| | 484 | // allocate the lower 32 bits |
| | 485 | alloc_reg(cur,i,reg); |
| | 486 | |
| | 487 | // Don't allocate unused registers |
| | 488 | if((cur->uu>>reg)&1) return; |
| | 489 | |
| | 490 | // see if the upper half is already allocated |
| | 491 | for(hr=0;hr<HOST_REGS;hr++) |
| | 492 | { |
| | 493 | if(cur->regmap[hr]==reg+64) return; |
| | 494 | } |
| | 495 | |
| | 496 | // Keep the same mapping if the register was already allocated in a loop |
| | 497 | preferred_reg = loop_reg(i,reg,preferred_reg); |
| | 498 | |
| | 499 | // Try to allocate the preferred register |
| | 500 | if(cur->regmap[preferred_reg]==-1) { |
| | 501 | cur->regmap[preferred_reg]=reg|64; |
| | 502 | cur->dirty&=~(1<<preferred_reg); |
| | 503 | cur->isconst&=~(1<<preferred_reg); |
| | 504 | return; |
| | 505 | } |
| | 506 | r=cur->regmap[preferred_reg]; |
| | 507 | if(r<64&&((cur->u>>r)&1)) { |
| | 508 | cur->regmap[preferred_reg]=reg|64; |
| | 509 | cur->dirty&=~(1<<preferred_reg); |
| | 510 | cur->isconst&=~(1<<preferred_reg); |
| | 511 | return; |
| | 512 | } |
| | 513 | if(r>=64&&((cur->uu>>(r&63))&1)) { |
| | 514 | cur->regmap[preferred_reg]=reg|64; |
| | 515 | cur->dirty&=~(1<<preferred_reg); |
| | 516 | cur->isconst&=~(1<<preferred_reg); |
| | 517 | return; |
| | 518 | } |
| | 519 | |
| | 520 | // Clear any unneeded registers |
| | 521 | // We try to keep the mapping consistent, if possible, because it |
| | 522 | // makes branches easier (especially loops). So we try to allocate |
| | 523 | // first (see above) before removing old mappings. If this is not |
| | 524 | // possible then go ahead and clear out the registers that are no |
| | 525 | // longer needed. |
| | 526 | for(hr=HOST_REGS-1;hr>=0;hr--) |
| | 527 | { |
| | 528 | r=cur->regmap[hr]; |
| | 529 | if(r>=0) { |
| | 530 | if(r<64) { |
| | 531 | if((cur->u>>r)&1) {cur->regmap[hr]=-1;break;} |
| | 532 | } |
| | 533 | else |
| | 534 | { |
| | 535 | if((cur->uu>>(r&63))&1) {cur->regmap[hr]=-1;break;} |
| | 536 | } |
| | 537 | } |
| | 538 | } |
| | 539 | // Try to allocate any available register, but prefer |
| | 540 | // registers that have not been used recently. |
| | 541 | if(i>0) { |
| | 542 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 543 | if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) { |
| | 544 | 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]) { |
| | 545 | cur->regmap[hr]=reg|64; |
| | 546 | cur->dirty&=~(1<<hr); |
| | 547 | cur->isconst&=~(1<<hr); |
| | 548 | return; |
| | 549 | } |
| | 550 | } |
| | 551 | } |
| | 552 | } |
| | 553 | // Try to allocate any available register |
| | 554 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 555 | if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) { |
| | 556 | cur->regmap[hr]=reg|64; |
| | 557 | cur->dirty&=~(1<<hr); |
| | 558 | cur->isconst&=~(1<<hr); |
| | 559 | return; |
| | 560 | } |
| | 561 | } |
| | 562 | |
| | 563 | // Ok, now we have to evict someone |
| | 564 | // Pick a register we hopefully won't need soon |
| | 565 | u_char hsn[MAXREG+1]; |
| | 566 | memset(hsn,10,sizeof(hsn)); |
| | 567 | int j; |
| | 568 | lsn(hsn,i,&preferred_reg); |
| | 569 | //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]); |
| | 570 | //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]); |
| | 571 | if(i>0) { |
| | 572 | // Don't evict the cycle count at entry points, otherwise the entry |
| | 573 | // stub will have to write it. |
| | 574 | if(bt[i]&&hsn[CCREG]>2) hsn[CCREG]=2; |
| | 575 | if(i>1&&hsn[CCREG]>2&&(itype[i-2]==RJUMP||itype[i-2]==UJUMP||itype[i-2]==CJUMP||itype[i-2]==SJUMP||itype[i-2]==FJUMP)) hsn[CCREG]=2; |
| | 576 | for(j=10;j>=3;j--) |
| | 577 | { |
| | 578 | // Alloc preferred register if available |
| | 579 | if(hsn[r=cur->regmap[preferred_reg]&63]==j) { |
| | 580 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 581 | // Evict both parts of a 64-bit register |
| | 582 | if((cur->regmap[hr]&63)==r) { |
| | 583 | cur->regmap[hr]=-1; |
| | 584 | cur->dirty&=~(1<<hr); |
| | 585 | cur->isconst&=~(1<<hr); |
| | 586 | } |
| | 587 | } |
| | 588 | cur->regmap[preferred_reg]=reg|64; |
| | 589 | return; |
| | 590 | } |
| | 591 | for(r=1;r<=MAXREG;r++) |
| | 592 | { |
| | 593 | if(hsn[r]==j&&r!=rs1[i-1]&&r!=rs2[i-1]&&r!=rt1[i-1]&&r!=rt2[i-1]) { |
| | 594 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 595 | if(hr!=HOST_CCREG||j<hsn[CCREG]) { |
| | 596 | if(cur->regmap[hr]==r+64) { |
| | 597 | cur->regmap[hr]=reg|64; |
| | 598 | cur->dirty&=~(1<<hr); |
| | 599 | cur->isconst&=~(1<<hr); |
| | 600 | return; |
| | 601 | } |
| | 602 | } |
| | 603 | } |
| | 604 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 605 | if(hr!=HOST_CCREG||j<hsn[CCREG]) { |
| | 606 | if(cur->regmap[hr]==r) { |
| | 607 | cur->regmap[hr]=reg|64; |
| | 608 | cur->dirty&=~(1<<hr); |
| | 609 | cur->isconst&=~(1<<hr); |
| | 610 | return; |
| | 611 | } |
| | 612 | } |
| | 613 | } |
| | 614 | } |
| | 615 | } |
| | 616 | } |
| | 617 | } |
| | 618 | for(j=10;j>=0;j--) |
| | 619 | { |
| | 620 | for(r=1;r<=MAXREG;r++) |
| | 621 | { |
| | 622 | if(hsn[r]==j) { |
| | 623 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 624 | if(cur->regmap[hr]==r+64) { |
| | 625 | cur->regmap[hr]=reg|64; |
| | 626 | cur->dirty&=~(1<<hr); |
| | 627 | cur->isconst&=~(1<<hr); |
| | 628 | return; |
| | 629 | } |
| | 630 | } |
| | 631 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 632 | if(cur->regmap[hr]==r) { |
| | 633 | cur->regmap[hr]=reg|64; |
| | 634 | cur->dirty&=~(1<<hr); |
| | 635 | cur->isconst&=~(1<<hr); |
| | 636 | return; |
| | 637 | } |
| | 638 | } |
| | 639 | } |
| | 640 | } |
| | 641 | } |
| | 642 | printf("This shouldn't happen");exit(1); |
| | 643 | } |
| | 644 | |
| | 645 | // Allocate a temporary register. This is done without regard to |
| | 646 | // dirty status or whether the register we request is on the unneeded list |
| | 647 | // Note: This will only allocate one register, even if called multiple times |
| | 648 | void alloc_reg_temp(struct regstat *cur,int i,signed char reg) |
| | 649 | { |
| | 650 | int r,hr; |
| | 651 | int preferred_reg = -1; |
| | 652 | |
| | 653 | // see if it's already allocated |
| | 654 | for(hr=0;hr<HOST_REGS;hr++) |
| | 655 | { |
| | 656 | if(hr!=EXCLUDE_REG&&cur->regmap[hr]==reg) return; |
| | 657 | } |
| | 658 | |
| | 659 | // Try to allocate any available register |
| | 660 | for(hr=HOST_REGS-1;hr>=0;hr--) { |
| | 661 | if(hr!=EXCLUDE_REG&&cur->regmap[hr]==-1) { |
| | 662 | cur->regmap[hr]=reg; |
| | 663 | cur->dirty&=~(1<<hr); |
| | 664 | cur->isconst&=~(1<<hr); |
| | 665 | return; |
| | 666 | } |
| | 667 | } |
| | 668 | |
| | 669 | // Find an unneeded register |
| | 670 | for(hr=HOST_REGS-1;hr>=0;hr--) |
| | 671 | { |
| | 672 | r=cur->regmap[hr]; |
| | 673 | if(r>=0) { |
| | 674 | if(r<64) { |
| | 675 | if((cur->u>>r)&1) { |
| | 676 | if(i==0||((unneeded_reg[i-1]>>r)&1)) { |
| | 677 | cur->regmap[hr]=reg; |
| | 678 | cur->dirty&=~(1<<hr); |
| | 679 | cur->isconst&=~(1<<hr); |
| | 680 | return; |
| | 681 | } |
| | 682 | } |
| | 683 | } |
| | 684 | else |
| | 685 | { |
| | 686 | if((cur->uu>>(r&63))&1) { |
| | 687 | if(i==0||((unneeded_reg_upper[i-1]>>(r&63))&1)) { |
| | 688 | cur->regmap[hr]=reg; |
| | 689 | cur->dirty&=~(1<<hr); |
| | 690 | cur->isconst&=~(1<<hr); |
| | 691 | return; |
| | 692 | } |
| | 693 | } |
| | 694 | } |
| | 695 | } |
| | 696 | } |
| | 697 | |
| | 698 | // Ok, now we have to evict someone |
| | 699 | // Pick a register we hopefully won't need soon |
| | 700 | // TODO: we might want to follow unconditional jumps here |
| | 701 | // TODO: get rid of dupe code and make this into a function |
| | 702 | u_char hsn[MAXREG+1]; |
| | 703 | memset(hsn,10,sizeof(hsn)); |
| | 704 | int j; |
| | 705 | lsn(hsn,i,&preferred_reg); |
| | 706 | //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]); |
| | 707 | if(i>0) { |
| | 708 | // Don't evict the cycle count at entry points, otherwise the entry |
| | 709 | // stub will have to write it. |
| | 710 | if(bt[i]&&hsn[CCREG]>2) hsn[CCREG]=2; |
| | 711 | if(i>1&&hsn[CCREG]>2&&(itype[i-2]==RJUMP||itype[i-2]==UJUMP||itype[i-2]==CJUMP||itype[i-2]==SJUMP||itype[i-2]==FJUMP)) hsn[CCREG]=2; |
| | 712 | for(j=10;j>=3;j--) |
| | 713 | { |
| | 714 | for(r=1;r<=MAXREG;r++) |
| | 715 | { |
| | 716 | if(hsn[r]==j&&r!=rs1[i-1]&&r!=rs2[i-1]&&r!=rt1[i-1]&&r!=rt2[i-1]) { |
| | 717 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 718 | if(hr!=HOST_CCREG||hsn[CCREG]>2) { |
| | 719 | if(cur->regmap[hr]==r+64) { |
| | 720 | cur->regmap[hr]=reg; |
| | 721 | cur->dirty&=~(1<<hr); |
| | 722 | cur->isconst&=~(1<<hr); |
| | 723 | return; |
| | 724 | } |
| | 725 | } |
| | 726 | } |
| | 727 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 728 | if(hr!=HOST_CCREG||hsn[CCREG]>2) { |
| | 729 | if(cur->regmap[hr]==r) { |
| | 730 | cur->regmap[hr]=reg; |
| | 731 | cur->dirty&=~(1<<hr); |
| | 732 | cur->isconst&=~(1<<hr); |
| | 733 | return; |
| | 734 | } |
| | 735 | } |
| | 736 | } |
| | 737 | } |
| | 738 | } |
| | 739 | } |
| | 740 | } |
| | 741 | for(j=10;j>=0;j--) |
| | 742 | { |
| | 743 | for(r=1;r<=MAXREG;r++) |
| | 744 | { |
| | 745 | if(hsn[r]==j) { |
| | 746 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 747 | if(cur->regmap[hr]==r+64) { |
| | 748 | cur->regmap[hr]=reg; |
| | 749 | cur->dirty&=~(1<<hr); |
| | 750 | cur->isconst&=~(1<<hr); |
| | 751 | return; |
| | 752 | } |
| | 753 | } |
| | 754 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 755 | if(cur->regmap[hr]==r) { |
| | 756 | cur->regmap[hr]=reg; |
| | 757 | cur->dirty&=~(1<<hr); |
| | 758 | cur->isconst&=~(1<<hr); |
| | 759 | return; |
| | 760 | } |
| | 761 | } |
| | 762 | } |
| | 763 | } |
| | 764 | } |
| | 765 | printf("This shouldn't happen");exit(1); |
| | 766 | } |
| | 767 | // Allocate a specific ARM register. |
| | 768 | void alloc_arm_reg(struct regstat *cur,int i,signed char reg,char hr) |
| | 769 | { |
| | 770 | int n; |
| | 771 | int dirty=0; |
| | 772 | |
| | 773 | // see if it's already allocated (and dealloc it) |
| | 774 | for(n=0;n<HOST_REGS;n++) |
| | 775 | { |
| | 776 | if(n!=EXCLUDE_REG&&cur->regmap[n]==reg) { |
| | 777 | dirty=(cur->dirty>>n)&1; |
| | 778 | cur->regmap[n]=-1; |
| | 779 | } |
| | 780 | } |
| | 781 | |
| | 782 | cur->regmap[hr]=reg; |
| | 783 | cur->dirty&=~(1<<hr); |
| | 784 | cur->dirty|=dirty<<hr; |
| | 785 | cur->isconst&=~(1<<hr); |
| | 786 | } |
| | 787 | |
| | 788 | // Alloc cycle count into dedicated register |
| | 789 | alloc_cc(struct regstat *cur,int i) |
| | 790 | { |
| | 791 | alloc_arm_reg(cur,i,CCREG,HOST_CCREG); |
| | 792 | } |
| | 793 | |
| | 794 | /* Special alloc */ |
| | 795 | |
| | 796 | |
| | 797 | /* Assembler */ |
| | 798 | |
| | 799 | char regname[16][4] = { |
| | 800 | "r0", |
| | 801 | "r1", |
| | 802 | "r2", |
| | 803 | "r3", |
| | 804 | "r4", |
| | 805 | "r5", |
| | 806 | "r6", |
| | 807 | "r7", |
| | 808 | "r8", |
| | 809 | "r9", |
| | 810 | "r10", |
| | 811 | "fp", |
| | 812 | "r12", |
| | 813 | "sp", |
| | 814 | "lr", |
| | 815 | "pc"}; |
| | 816 | |
| | 817 | void output_byte(u_char byte) |
| | 818 | { |
| | 819 | *(out++)=byte; |
| | 820 | } |
| | 821 | void output_modrm(u_char mod,u_char rm,u_char ext) |
| | 822 | { |
| | 823 | assert(mod<4); |
| | 824 | assert(rm<8); |
| | 825 | assert(ext<8); |
| | 826 | u_char byte=(mod<<6)|(ext<<3)|rm; |
| | 827 | *(out++)=byte; |
| | 828 | } |
| | 829 | void output_sib(u_char scale,u_char index,u_char base) |
| | 830 | { |
| | 831 | assert(scale<4); |
| | 832 | assert(index<8); |
| | 833 | assert(base<8); |
| | 834 | u_char byte=(scale<<6)|(index<<3)|base; |
| | 835 | *(out++)=byte; |
| | 836 | } |
| | 837 | void output_w32(u_int word) |
| | 838 | { |
| | 839 | *((u_int *)out)=word; |
| | 840 | out+=4; |
| | 841 | } |
| | 842 | u_int rd_rn_rm(u_int rd, u_int rn, u_int rm) |
| | 843 | { |
| | 844 | assert(rd<16); |
| | 845 | assert(rn<16); |
| | 846 | assert(rm<16); |
| | 847 | return((rn<<16)|(rd<<12)|rm); |
| | 848 | } |
| | 849 | u_int rd_rn_imm_shift(u_int rd, u_int rn, u_int imm, u_int shift) |
| | 850 | { |
| | 851 | assert(rd<16); |
| | 852 | assert(rn<16); |
| | 853 | assert(imm<256); |
| | 854 | assert((shift&1)==0); |
| | 855 | return((rn<<16)|(rd<<12)|(((32-shift)&30)<<7)|imm); |
| | 856 | } |
| | 857 | u_int genimm(u_int imm,u_int *encoded) |
| | 858 | { |
| | 859 | *encoded=0; |
| | 860 | if(imm==0) return 1; |
| | 861 | int i=32; |
| | 862 | while(i>0) |
| | 863 | { |
| | 864 | if(imm<256) { |
| | 865 | *encoded=((i&30)<<7)|imm; |
| | 866 | return 1; |
| | 867 | } |
| | 868 | imm=(imm>>2)|(imm<<30);i-=2; |
| | 869 | } |
| | 870 | return 0; |
| | 871 | } |
| | 872 | void genimm_checked(u_int imm,u_int *encoded) |
| | 873 | { |
| | 874 | u_int ret=genimm(imm,encoded); |
| | 875 | assert(ret); |
| | 876 | } |
| | 877 | u_int genjmp(u_int addr) |
| | 878 | { |
| | 879 | int offset=addr-(int)out-8; |
| | 880 | if(offset<-33554432||offset>=33554432) { |
| | 881 | if (addr>2) { |
| | 882 | printf("genjmp: out of range: %08x\n", offset); |
| | 883 | exit(1); |
| | 884 | } |
| | 885 | return 0; |
| | 886 | } |
| | 887 | return ((u_int)offset>>2)&0xffffff; |
| | 888 | } |
| | 889 | |
| | 890 | void emit_mov(int rs,int rt) |
| | 891 | { |
| | 892 | assem_debug("mov %s,%s\n",regname[rt],regname[rs]); |
| | 893 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)); |
| | 894 | } |
| | 895 | |
| | 896 | void emit_movs(int rs,int rt) |
| | 897 | { |
| | 898 | assem_debug("movs %s,%s\n",regname[rt],regname[rs]); |
| | 899 | output_w32(0xe1b00000|rd_rn_rm(rt,0,rs)); |
| | 900 | } |
| | 901 | |
| | 902 | void emit_add(int rs1,int rs2,int rt) |
| | 903 | { |
| | 904 | assem_debug("add %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 905 | output_w32(0xe0800000|rd_rn_rm(rt,rs1,rs2)); |
| | 906 | } |
| | 907 | |
| | 908 | void emit_adds(int rs1,int rs2,int rt) |
| | 909 | { |
| | 910 | assem_debug("adds %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 911 | output_w32(0xe0900000|rd_rn_rm(rt,rs1,rs2)); |
| | 912 | } |
| | 913 | |
| | 914 | void emit_adcs(int rs1,int rs2,int rt) |
| | 915 | { |
| | 916 | assem_debug("adcs %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 917 | output_w32(0xe0b00000|rd_rn_rm(rt,rs1,rs2)); |
| | 918 | } |
| | 919 | |
| | 920 | void emit_sbc(int rs1,int rs2,int rt) |
| | 921 | { |
| | 922 | assem_debug("sbc %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 923 | output_w32(0xe0c00000|rd_rn_rm(rt,rs1,rs2)); |
| | 924 | } |
| | 925 | |
| | 926 | void emit_sbcs(int rs1,int rs2,int rt) |
| | 927 | { |
| | 928 | assem_debug("sbcs %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 929 | output_w32(0xe0d00000|rd_rn_rm(rt,rs1,rs2)); |
| | 930 | } |
| | 931 | |
| | 932 | void emit_neg(int rs, int rt) |
| | 933 | { |
| | 934 | assem_debug("rsb %s,%s,#0\n",regname[rt],regname[rs]); |
| | 935 | output_w32(0xe2600000|rd_rn_rm(rt,rs,0)); |
| | 936 | } |
| | 937 | |
| | 938 | void emit_negs(int rs, int rt) |
| | 939 | { |
| | 940 | assem_debug("rsbs %s,%s,#0\n",regname[rt],regname[rs]); |
| | 941 | output_w32(0xe2700000|rd_rn_rm(rt,rs,0)); |
| | 942 | } |
| | 943 | |
| | 944 | void emit_sub(int rs1,int rs2,int rt) |
| | 945 | { |
| | 946 | assem_debug("sub %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 947 | output_w32(0xe0400000|rd_rn_rm(rt,rs1,rs2)); |
| | 948 | } |
| | 949 | |
| | 950 | void emit_subs(int rs1,int rs2,int rt) |
| | 951 | { |
| | 952 | assem_debug("subs %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 953 | output_w32(0xe0500000|rd_rn_rm(rt,rs1,rs2)); |
| | 954 | } |
| | 955 | |
| | 956 | void emit_zeroreg(int rt) |
| | 957 | { |
| | 958 | assem_debug("mov %s,#0\n",regname[rt]); |
| | 959 | output_w32(0xe3a00000|rd_rn_rm(rt,0,0)); |
| | 960 | } |
| | 961 | |
| | 962 | void emit_loadlp(u_int imm,u_int rt) |
| | 963 | { |
| | 964 | add_literal((int)out,imm); |
| | 965 | assem_debug("ldr %s,pc+? [=%x]\n",regname[rt],imm); |
| | 966 | output_w32(0xe5900000|rd_rn_rm(rt,15,0)); |
| | 967 | } |
| | 968 | void emit_movw(u_int imm,u_int rt) |
| | 969 | { |
| | 970 | assert(imm<65536); |
| | 971 | assem_debug("movw %s,#%d (0x%x)\n",regname[rt],imm,imm); |
| | 972 | output_w32(0xe3000000|rd_rn_rm(rt,0,0)|(imm&0xfff)|((imm<<4)&0xf0000)); |
| | 973 | } |
| | 974 | void emit_movt(u_int imm,u_int rt) |
| | 975 | { |
| | 976 | assem_debug("movt %s,#%d (0x%x)\n",regname[rt],imm&0xffff0000,imm&0xffff0000); |
| | 977 | output_w32(0xe3400000|rd_rn_rm(rt,0,0)|((imm>>16)&0xfff)|((imm>>12)&0xf0000)); |
| | 978 | } |
| | 979 | void emit_movimm(u_int imm,u_int rt) |
| | 980 | { |
| | 981 | u_int armval; |
| | 982 | if(genimm(imm,&armval)) { |
| | 983 | assem_debug("mov %s,#%d\n",regname[rt],imm); |
| | 984 | output_w32(0xe3a00000|rd_rn_rm(rt,0,0)|armval); |
| | 985 | }else if(genimm(~imm,&armval)) { |
| | 986 | assem_debug("mvn %s,#%d\n",regname[rt],imm); |
| | 987 | output_w32(0xe3e00000|rd_rn_rm(rt,0,0)|armval); |
| | 988 | }else if(imm<65536) { |
| | 989 | #ifdef ARMv5_ONLY |
| | 990 | assem_debug("mov %s,#%d\n",regname[rt],imm&0xFF00); |
| | 991 | output_w32(0xe3a00000|rd_rn_imm_shift(rt,0,imm>>8,8)); |
| | 992 | assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF); |
| | 993 | output_w32(0xe2800000|rd_rn_imm_shift(rt,rt,imm&0xff,0)); |
| | 994 | #else |
| | 995 | emit_movw(imm,rt); |
| | 996 | #endif |
| | 997 | }else{ |
| | 998 | #ifdef ARMv5_ONLY |
| | 999 | emit_loadlp(imm,rt); |
| | 1000 | #else |
| | 1001 | emit_movw(imm&0x0000FFFF,rt); |
| | 1002 | emit_movt(imm&0xFFFF0000,rt); |
| | 1003 | #endif |
| | 1004 | } |
| | 1005 | } |
| | 1006 | void emit_pcreladdr(u_int rt) |
| | 1007 | { |
| | 1008 | assem_debug("add %s,pc,#?\n",regname[rt]); |
| | 1009 | output_w32(0xe2800000|rd_rn_rm(rt,15,0)); |
| | 1010 | } |
| | 1011 | |
| | 1012 | void emit_loadreg(int r, int hr) |
| | 1013 | { |
| | 1014 | #ifdef FORCE32 |
| | 1015 | if(r&64) { |
| | 1016 | printf("64bit load in 32bit mode!\n"); |
| | 1017 | assert(0); |
| | 1018 | return; |
| | 1019 | } |
| | 1020 | #endif |
| | 1021 | if((r&63)==0) |
| | 1022 | emit_zeroreg(hr); |
| | 1023 | else { |
| | 1024 | int addr=((int)reg)+((r&63)<<REG_SHIFT)+((r&64)>>4); |
| | 1025 | if((r&63)==HIREG) addr=(int)&hi+((r&64)>>4); |
| | 1026 | if((r&63)==LOREG) addr=(int)&lo+((r&64)>>4); |
| | 1027 | if(r==CCREG) addr=(int)&cycle_count; |
| | 1028 | if(r==CSREG) addr=(int)&Status; |
| | 1029 | if(r==FSREG) addr=(int)&FCR31; |
| | 1030 | if(r==INVCP) addr=(int)&invc_ptr; |
| | 1031 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1032 | assert(offset<4096); |
| | 1033 | assem_debug("ldr %s,fp+%d\n",regname[hr],offset); |
| | 1034 | output_w32(0xe5900000|rd_rn_rm(hr,FP,0)|offset); |
| | 1035 | } |
| | 1036 | } |
| | 1037 | void emit_storereg(int r, int hr) |
| | 1038 | { |
| | 1039 | #ifdef FORCE32 |
| | 1040 | if(r&64) { |
| | 1041 | printf("64bit store in 32bit mode!\n"); |
| | 1042 | assert(0); |
| | 1043 | return; |
| | 1044 | } |
| | 1045 | #endif |
| | 1046 | int addr=((int)reg)+((r&63)<<REG_SHIFT)+((r&64)>>4); |
| | 1047 | if((r&63)==HIREG) addr=(int)&hi+((r&64)>>4); |
| | 1048 | if((r&63)==LOREG) addr=(int)&lo+((r&64)>>4); |
| | 1049 | if(r==CCREG) addr=(int)&cycle_count; |
| | 1050 | if(r==FSREG) addr=(int)&FCR31; |
| | 1051 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1052 | assert(offset<4096); |
| | 1053 | assem_debug("str %s,fp+%d\n",regname[hr],offset); |
| | 1054 | output_w32(0xe5800000|rd_rn_rm(hr,FP,0)|offset); |
| | 1055 | } |
| | 1056 | |
| | 1057 | void emit_test(int rs, int rt) |
| | 1058 | { |
| | 1059 | assem_debug("tst %s,%s\n",regname[rs],regname[rt]); |
| | 1060 | output_w32(0xe1100000|rd_rn_rm(0,rs,rt)); |
| | 1061 | } |
| | 1062 | |
| | 1063 | void emit_testimm(int rs,int imm) |
| | 1064 | { |
| | 1065 | u_int armval; |
| | 1066 | assem_debug("tst %s,#%d\n",regname[rs],imm); |
| | 1067 | genimm_checked(imm,&armval); |
| | 1068 | output_w32(0xe3100000|rd_rn_rm(0,rs,0)|armval); |
| | 1069 | } |
| | 1070 | |
| | 1071 | void emit_testeqimm(int rs,int imm) |
| | 1072 | { |
| | 1073 | u_int armval; |
| | 1074 | assem_debug("tsteq %s,$%d\n",regname[rs],imm); |
| | 1075 | genimm_checked(imm,&armval); |
| | 1076 | output_w32(0x03100000|rd_rn_rm(0,rs,0)|armval); |
| | 1077 | } |
| | 1078 | |
| | 1079 | void emit_not(int rs,int rt) |
| | 1080 | { |
| | 1081 | assem_debug("mvn %s,%s\n",regname[rt],regname[rs]); |
| | 1082 | output_w32(0xe1e00000|rd_rn_rm(rt,0,rs)); |
| | 1083 | } |
| | 1084 | |
| | 1085 | void emit_mvnmi(int rs,int rt) |
| | 1086 | { |
| | 1087 | assem_debug("mvnmi %s,%s\n",regname[rt],regname[rs]); |
| | 1088 | output_w32(0x41e00000|rd_rn_rm(rt,0,rs)); |
| | 1089 | } |
| | 1090 | |
| | 1091 | void emit_and(u_int rs1,u_int rs2,u_int rt) |
| | 1092 | { |
| | 1093 | assem_debug("and %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1094 | output_w32(0xe0000000|rd_rn_rm(rt,rs1,rs2)); |
| | 1095 | } |
| | 1096 | |
| | 1097 | void emit_or(u_int rs1,u_int rs2,u_int rt) |
| | 1098 | { |
| | 1099 | assem_debug("orr %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1100 | output_w32(0xe1800000|rd_rn_rm(rt,rs1,rs2)); |
| | 1101 | } |
| | 1102 | void emit_or_and_set_flags(int rs1,int rs2,int rt) |
| | 1103 | { |
| | 1104 | assem_debug("orrs %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1105 | output_w32(0xe1900000|rd_rn_rm(rt,rs1,rs2)); |
| | 1106 | } |
| | 1107 | |
| | 1108 | void emit_orrshl_imm(u_int rs,u_int imm,u_int rt) |
| | 1109 | { |
| | 1110 | assert(rs<16); |
| | 1111 | assert(rt<16); |
| | 1112 | assert(imm<32); |
| | 1113 | assem_debug("orr %s,%s,%s,lsl #%d\n",regname[rt],regname[rt],regname[rs],imm); |
| | 1114 | output_w32(0xe1800000|rd_rn_rm(rt,rt,rs)|(imm<<7)); |
| | 1115 | } |
| | 1116 | |
| | 1117 | void emit_orrshr_imm(u_int rs,u_int imm,u_int rt) |
| | 1118 | { |
| | 1119 | assert(rs<16); |
| | 1120 | assert(rt<16); |
| | 1121 | assert(imm<32); |
| | 1122 | assem_debug("orr %s,%s,%s,lsr #%d\n",regname[rt],regname[rt],regname[rs],imm); |
| | 1123 | output_w32(0xe1800020|rd_rn_rm(rt,rt,rs)|(imm<<7)); |
| | 1124 | } |
| | 1125 | |
| | 1126 | void emit_xor(u_int rs1,u_int rs2,u_int rt) |
| | 1127 | { |
| | 1128 | assem_debug("eor %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1129 | output_w32(0xe0200000|rd_rn_rm(rt,rs1,rs2)); |
| | 1130 | } |
| | 1131 | |
| | 1132 | void emit_addimm(u_int rs,int imm,u_int rt) |
| | 1133 | { |
| | 1134 | assert(rs<16); |
| | 1135 | assert(rt<16); |
| | 1136 | if(imm!=0) { |
| | 1137 | u_int armval; |
| | 1138 | if(genimm(imm,&armval)) { |
| | 1139 | assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1140 | output_w32(0xe2800000|rd_rn_rm(rt,rs,0)|armval); |
| | 1141 | }else if(genimm(-imm,&armval)) { |
| | 1142 | assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1143 | output_w32(0xe2400000|rd_rn_rm(rt,rs,0)|armval); |
| | 1144 | }else if(imm<0) { |
| | 1145 | assert(imm>-65536); |
| | 1146 | assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rs],(-imm)&0xFF00); |
| | 1147 | assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rt],(-imm)&0xFF); |
| | 1148 | output_w32(0xe2400000|rd_rn_imm_shift(rt,rs,(-imm)>>8,8)); |
| | 1149 | output_w32(0xe2400000|rd_rn_imm_shift(rt,rt,(-imm)&0xff,0)); |
| | 1150 | }else{ |
| | 1151 | assert(imm<65536); |
| | 1152 | assem_debug("add %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF00); |
| | 1153 | assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF); |
| | 1154 | output_w32(0xe2800000|rd_rn_imm_shift(rt,rs,imm>>8,8)); |
| | 1155 | output_w32(0xe2800000|rd_rn_imm_shift(rt,rt,imm&0xff,0)); |
| | 1156 | } |
| | 1157 | } |
| | 1158 | else if(rs!=rt) emit_mov(rs,rt); |
| | 1159 | } |
| | 1160 | |
| | 1161 | void emit_addimm_and_set_flags(int imm,int rt) |
| | 1162 | { |
| | 1163 | assert(imm>-65536&&imm<65536); |
| | 1164 | u_int armval; |
| | 1165 | if(genimm(imm,&armval)) { |
| | 1166 | assem_debug("adds %s,%s,#%d\n",regname[rt],regname[rt],imm); |
| | 1167 | output_w32(0xe2900000|rd_rn_rm(rt,rt,0)|armval); |
| | 1168 | }else if(genimm(-imm,&armval)) { |
| | 1169 | assem_debug("subs %s,%s,#%d\n",regname[rt],regname[rt],imm); |
| | 1170 | output_w32(0xe2500000|rd_rn_rm(rt,rt,0)|armval); |
| | 1171 | }else if(imm<0) { |
| | 1172 | assem_debug("sub %s,%s,#%d\n",regname[rt],regname[rt],(-imm)&0xFF00); |
| | 1173 | assem_debug("subs %s,%s,#%d\n",regname[rt],regname[rt],(-imm)&0xFF); |
| | 1174 | output_w32(0xe2400000|rd_rn_imm_shift(rt,rt,(-imm)>>8,8)); |
| | 1175 | output_w32(0xe2500000|rd_rn_imm_shift(rt,rt,(-imm)&0xff,0)); |
| | 1176 | }else{ |
| | 1177 | assem_debug("add %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF00); |
| | 1178 | assem_debug("adds %s,%s,#%d\n",regname[rt],regname[rt],imm&0xFF); |
| | 1179 | output_w32(0xe2800000|rd_rn_imm_shift(rt,rt,imm>>8,8)); |
| | 1180 | output_w32(0xe2900000|rd_rn_imm_shift(rt,rt,imm&0xff,0)); |
| | 1181 | } |
| | 1182 | } |
| | 1183 | void emit_addimm_no_flags(u_int imm,u_int rt) |
| | 1184 | { |
| | 1185 | emit_addimm(rt,imm,rt); |
| | 1186 | } |
| | 1187 | |
| | 1188 | void emit_addnop(u_int r) |
| | 1189 | { |
| | 1190 | assert(r<16); |
| | 1191 | assem_debug("add %s,%s,#0 (nop)\n",regname[r],regname[r]); |
| | 1192 | output_w32(0xe2800000|rd_rn_rm(r,r,0)); |
| | 1193 | } |
| | 1194 | |
| | 1195 | void emit_adcimm(u_int rs,int imm,u_int rt) |
| | 1196 | { |
| | 1197 | u_int armval; |
| | 1198 | genimm_checked(imm,&armval); |
| | 1199 | assem_debug("adc %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1200 | output_w32(0xe2a00000|rd_rn_rm(rt,rs,0)|armval); |
| | 1201 | } |
| | 1202 | /*void emit_sbcimm(int imm,u_int rt) |
| | 1203 | { |
| | 1204 | u_int armval; |
| | 1205 | genimm_checked(imm,&armval); |
| | 1206 | assem_debug("sbc %s,%s,#%d\n",regname[rt],regname[rt],imm); |
| | 1207 | output_w32(0xe2c00000|rd_rn_rm(rt,rt,0)|armval); |
| | 1208 | }*/ |
| | 1209 | void emit_sbbimm(int imm,u_int rt) |
| | 1210 | { |
| | 1211 | assem_debug("sbb $%d,%%%s\n",imm,regname[rt]); |
| | 1212 | assert(rt<8); |
| | 1213 | if(imm<128&&imm>=-128) { |
| | 1214 | output_byte(0x83); |
| | 1215 | output_modrm(3,rt,3); |
| | 1216 | output_byte(imm); |
| | 1217 | } |
| | 1218 | else |
| | 1219 | { |
| | 1220 | output_byte(0x81); |
| | 1221 | output_modrm(3,rt,3); |
| | 1222 | output_w32(imm); |
| | 1223 | } |
| | 1224 | } |
| | 1225 | void emit_rscimm(int rs,int imm,u_int rt) |
| | 1226 | { |
| | 1227 | assert(0); |
| | 1228 | u_int armval; |
| | 1229 | genimm_checked(imm,&armval); |
| | 1230 | assem_debug("rsc %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1231 | output_w32(0xe2e00000|rd_rn_rm(rt,rs,0)|armval); |
| | 1232 | } |
| | 1233 | |
| | 1234 | void emit_addimm64_32(int rsh,int rsl,int imm,int rth,int rtl) |
| | 1235 | { |
| | 1236 | // TODO: if(genimm(imm,&armval)) ... |
| | 1237 | // else |
| | 1238 | emit_movimm(imm,HOST_TEMPREG); |
| | 1239 | emit_adds(HOST_TEMPREG,rsl,rtl); |
| | 1240 | emit_adcimm(rsh,0,rth); |
| | 1241 | } |
| | 1242 | |
| | 1243 | void emit_sbb(int rs1,int rs2) |
| | 1244 | { |
| | 1245 | assem_debug("sbb %%%s,%%%s\n",regname[rs2],regname[rs1]); |
| | 1246 | output_byte(0x19); |
| | 1247 | output_modrm(3,rs1,rs2); |
| | 1248 | } |
| | 1249 | |
| | 1250 | void emit_andimm(int rs,int imm,int rt) |
| | 1251 | { |
| | 1252 | u_int armval; |
| | 1253 | if(imm==0) { |
| | 1254 | emit_zeroreg(rt); |
| | 1255 | }else if(genimm(imm,&armval)) { |
| | 1256 | assem_debug("and %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1257 | output_w32(0xe2000000|rd_rn_rm(rt,rs,0)|armval); |
| | 1258 | }else if(genimm(~imm,&armval)) { |
| | 1259 | assem_debug("bic %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1260 | output_w32(0xe3c00000|rd_rn_rm(rt,rs,0)|armval); |
| | 1261 | }else if(imm==65535) { |
| | 1262 | #ifdef ARMv5_ONLY |
| | 1263 | assem_debug("bic %s,%s,#FF000000\n",regname[rt],regname[rs]); |
| | 1264 | output_w32(0xe3c00000|rd_rn_rm(rt,rs,0)|0x4FF); |
| | 1265 | assem_debug("bic %s,%s,#00FF0000\n",regname[rt],regname[rt]); |
| | 1266 | output_w32(0xe3c00000|rd_rn_rm(rt,rt,0)|0x8FF); |
| | 1267 | #else |
| | 1268 | assem_debug("uxth %s,%s\n",regname[rt],regname[rs]); |
| | 1269 | output_w32(0xe6ff0070|rd_rn_rm(rt,0,rs)); |
| | 1270 | #endif |
| | 1271 | }else{ |
| | 1272 | assert(imm>0&&imm<65535); |
| | 1273 | #ifdef ARMv5_ONLY |
| | 1274 | assem_debug("mov r14,#%d\n",imm&0xFF00); |
| | 1275 | output_w32(0xe3a00000|rd_rn_imm_shift(HOST_TEMPREG,0,imm>>8,8)); |
| | 1276 | assem_debug("add r14,r14,#%d\n",imm&0xFF); |
| | 1277 | output_w32(0xe2800000|rd_rn_imm_shift(HOST_TEMPREG,HOST_TEMPREG,imm&0xff,0)); |
| | 1278 | #else |
| | 1279 | emit_movw(imm,HOST_TEMPREG); |
| | 1280 | #endif |
| | 1281 | assem_debug("and %s,%s,r14\n",regname[rt],regname[rs]); |
| | 1282 | output_w32(0xe0000000|rd_rn_rm(rt,rs,HOST_TEMPREG)); |
| | 1283 | } |
| | 1284 | } |
| | 1285 | |
| | 1286 | void emit_orimm(int rs,int imm,int rt) |
| | 1287 | { |
| | 1288 | u_int armval; |
| | 1289 | if(imm==0) { |
| | 1290 | if(rs!=rt) emit_mov(rs,rt); |
| | 1291 | }else if(genimm(imm,&armval)) { |
| | 1292 | assem_debug("orr %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1293 | output_w32(0xe3800000|rd_rn_rm(rt,rs,0)|armval); |
| | 1294 | }else{ |
| | 1295 | assert(imm>0&&imm<65536); |
| | 1296 | assem_debug("orr %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF00); |
| | 1297 | assem_debug("orr %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF); |
| | 1298 | output_w32(0xe3800000|rd_rn_imm_shift(rt,rs,imm>>8,8)); |
| | 1299 | output_w32(0xe3800000|rd_rn_imm_shift(rt,rt,imm&0xff,0)); |
| | 1300 | } |
| | 1301 | } |
| | 1302 | |
| | 1303 | void emit_xorimm(int rs,int imm,int rt) |
| | 1304 | { |
| | 1305 | u_int armval; |
| | 1306 | if(imm==0) { |
| | 1307 | if(rs!=rt) emit_mov(rs,rt); |
| | 1308 | }else if(genimm(imm,&armval)) { |
| | 1309 | assem_debug("eor %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1310 | output_w32(0xe2200000|rd_rn_rm(rt,rs,0)|armval); |
| | 1311 | }else{ |
| | 1312 | assert(imm>0&&imm<65536); |
| | 1313 | assem_debug("eor %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF00); |
| | 1314 | assem_debug("eor %s,%s,#%d\n",regname[rt],regname[rs],imm&0xFF); |
| | 1315 | output_w32(0xe2200000|rd_rn_imm_shift(rt,rs,imm>>8,8)); |
| | 1316 | output_w32(0xe2200000|rd_rn_imm_shift(rt,rt,imm&0xff,0)); |
| | 1317 | } |
| | 1318 | } |
| | 1319 | |
| | 1320 | void emit_shlimm(int rs,u_int imm,int rt) |
| | 1321 | { |
| | 1322 | assert(imm>0); |
| | 1323 | assert(imm<32); |
| | 1324 | //if(imm==1) ... |
| | 1325 | assem_debug("lsl %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1326 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|(imm<<7)); |
| | 1327 | } |
| | 1328 | |
| | 1329 | void emit_lsls_imm(int rs,int imm,int rt) |
| | 1330 | { |
| | 1331 | assert(imm>0); |
| | 1332 | assert(imm<32); |
| | 1333 | assem_debug("lsls %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1334 | output_w32(0xe1b00000|rd_rn_rm(rt,0,rs)|(imm<<7)); |
| | 1335 | } |
| | 1336 | |
| | 1337 | void emit_shrimm(int rs,u_int imm,int rt) |
| | 1338 | { |
| | 1339 | assert(imm>0); |
| | 1340 | assert(imm<32); |
| | 1341 | assem_debug("lsr %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1342 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|0x20|(imm<<7)); |
| | 1343 | } |
| | 1344 | |
| | 1345 | void emit_sarimm(int rs,u_int imm,int rt) |
| | 1346 | { |
| | 1347 | assert(imm>0); |
| | 1348 | assert(imm<32); |
| | 1349 | assem_debug("asr %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1350 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|0x40|(imm<<7)); |
| | 1351 | } |
| | 1352 | |
| | 1353 | void emit_rorimm(int rs,u_int imm,int rt) |
| | 1354 | { |
| | 1355 | assert(imm>0); |
| | 1356 | assert(imm<32); |
| | 1357 | assem_debug("ror %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1358 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|0x60|(imm<<7)); |
| | 1359 | } |
| | 1360 | |
| | 1361 | void emit_shldimm(int rs,int rs2,u_int imm,int rt) |
| | 1362 | { |
| | 1363 | assem_debug("shld %%%s,%%%s,%d\n",regname[rt],regname[rs2],imm); |
| | 1364 | assert(imm>0); |
| | 1365 | assert(imm<32); |
| | 1366 | //if(imm==1) ... |
| | 1367 | assem_debug("lsl %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1368 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|(imm<<7)); |
| | 1369 | assem_debug("orr %s,%s,%s,lsr #%d\n",regname[rt],regname[rt],regname[rs2],32-imm); |
| | 1370 | output_w32(0xe1800020|rd_rn_rm(rt,rt,rs2)|((32-imm)<<7)); |
| | 1371 | } |
| | 1372 | |
| | 1373 | void emit_shrdimm(int rs,int rs2,u_int imm,int rt) |
| | 1374 | { |
| | 1375 | assem_debug("shrd %%%s,%%%s,%d\n",regname[rt],regname[rs2],imm); |
| | 1376 | assert(imm>0); |
| | 1377 | assert(imm<32); |
| | 1378 | //if(imm==1) ... |
| | 1379 | assem_debug("lsr %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 1380 | output_w32(0xe1a00020|rd_rn_rm(rt,0,rs)|(imm<<7)); |
| | 1381 | assem_debug("orr %s,%s,%s,lsl #%d\n",regname[rt],regname[rt],regname[rs2],32-imm); |
| | 1382 | output_w32(0xe1800000|rd_rn_rm(rt,rt,rs2)|((32-imm)<<7)); |
| | 1383 | } |
| | 1384 | |
| | 1385 | void emit_signextend16(int rs,int rt) |
| | 1386 | { |
| | 1387 | #ifdef ARMv5_ONLY |
| | 1388 | emit_shlimm(rs,16,rt); |
| | 1389 | emit_sarimm(rt,16,rt); |
| | 1390 | #else |
| | 1391 | assem_debug("sxth %s,%s\n",regname[rt],regname[rs]); |
| | 1392 | output_w32(0xe6bf0070|rd_rn_rm(rt,0,rs)); |
| | 1393 | #endif |
| | 1394 | } |
| | 1395 | |
| | 1396 | void emit_signextend8(int rs,int rt) |
| | 1397 | { |
| | 1398 | #ifdef ARMv5_ONLY |
| | 1399 | emit_shlimm(rs,24,rt); |
| | 1400 | emit_sarimm(rt,24,rt); |
| | 1401 | #else |
| | 1402 | assem_debug("sxtb %s,%s\n",regname[rt],regname[rs]); |
| | 1403 | output_w32(0xe6af0070|rd_rn_rm(rt,0,rs)); |
| | 1404 | #endif |
| | 1405 | } |
| | 1406 | |
| | 1407 | void emit_shl(u_int rs,u_int shift,u_int rt) |
| | 1408 | { |
| | 1409 | assert(rs<16); |
| | 1410 | assert(rt<16); |
| | 1411 | assert(shift<16); |
| | 1412 | //if(imm==1) ... |
| | 1413 | assem_debug("lsl %s,%s,%s\n",regname[rt],regname[rs],regname[shift]); |
| | 1414 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|0x10|(shift<<8)); |
| | 1415 | } |
| | 1416 | void emit_shr(u_int rs,u_int shift,u_int rt) |
| | 1417 | { |
| | 1418 | assert(rs<16); |
| | 1419 | assert(rt<16); |
| | 1420 | assert(shift<16); |
| | 1421 | assem_debug("lsr %s,%s,%s\n",regname[rt],regname[rs],regname[shift]); |
| | 1422 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|0x30|(shift<<8)); |
| | 1423 | } |
| | 1424 | void emit_sar(u_int rs,u_int shift,u_int rt) |
| | 1425 | { |
| | 1426 | assert(rs<16); |
| | 1427 | assert(rt<16); |
| | 1428 | assert(shift<16); |
| | 1429 | assem_debug("asr %s,%s,%s\n",regname[rt],regname[rs],regname[shift]); |
| | 1430 | output_w32(0xe1a00000|rd_rn_rm(rt,0,rs)|0x50|(shift<<8)); |
| | 1431 | } |
| | 1432 | void emit_shlcl(int r) |
| | 1433 | { |
| | 1434 | assem_debug("shl %%%s,%%cl\n",regname[r]); |
| | 1435 | assert(0); |
| | 1436 | } |
| | 1437 | void emit_shrcl(int r) |
| | 1438 | { |
| | 1439 | assem_debug("shr %%%s,%%cl\n",regname[r]); |
| | 1440 | assert(0); |
| | 1441 | } |
| | 1442 | void emit_sarcl(int r) |
| | 1443 | { |
| | 1444 | assem_debug("sar %%%s,%%cl\n",regname[r]); |
| | 1445 | assert(0); |
| | 1446 | } |
| | 1447 | |
| | 1448 | void emit_shldcl(int r1,int r2) |
| | 1449 | { |
| | 1450 | assem_debug("shld %%%s,%%%s,%%cl\n",regname[r1],regname[r2]); |
| | 1451 | assert(0); |
| | 1452 | } |
| | 1453 | void emit_shrdcl(int r1,int r2) |
| | 1454 | { |
| | 1455 | assem_debug("shrd %%%s,%%%s,%%cl\n",regname[r1],regname[r2]); |
| | 1456 | assert(0); |
| | 1457 | } |
| | 1458 | void emit_orrshl(u_int rs,u_int shift,u_int rt) |
| | 1459 | { |
| | 1460 | assert(rs<16); |
| | 1461 | assert(rt<16); |
| | 1462 | assert(shift<16); |
| | 1463 | assem_debug("orr %s,%s,%s,lsl %s\n",regname[rt],regname[rt],regname[rs],regname[shift]); |
| | 1464 | output_w32(0xe1800000|rd_rn_rm(rt,rt,rs)|0x10|(shift<<8)); |
| | 1465 | } |
| | 1466 | void emit_orrshr(u_int rs,u_int shift,u_int rt) |
| | 1467 | { |
| | 1468 | assert(rs<16); |
| | 1469 | assert(rt<16); |
| | 1470 | assert(shift<16); |
| | 1471 | assem_debug("orr %s,%s,%s,lsr %s\n",regname[rt],regname[rt],regname[rs],regname[shift]); |
| | 1472 | output_w32(0xe1800000|rd_rn_rm(rt,rt,rs)|0x30|(shift<<8)); |
| | 1473 | } |
| | 1474 | |
| | 1475 | void emit_cmpimm(int rs,int imm) |
| | 1476 | { |
| | 1477 | u_int armval; |
| | 1478 | if(genimm(imm,&armval)) { |
| | 1479 | assem_debug("cmp %s,#%d\n",regname[rs],imm); |
| | 1480 | output_w32(0xe3500000|rd_rn_rm(0,rs,0)|armval); |
| | 1481 | }else if(genimm(-imm,&armval)) { |
| | 1482 | assem_debug("cmn %s,#%d\n",regname[rs],imm); |
| | 1483 | output_w32(0xe3700000|rd_rn_rm(0,rs,0)|armval); |
| | 1484 | }else if(imm>0) { |
| | 1485 | assert(imm<65536); |
| | 1486 | #ifdef ARMv5_ONLY |
| | 1487 | emit_movimm(imm,HOST_TEMPREG); |
| | 1488 | #else |
| | 1489 | emit_movw(imm,HOST_TEMPREG); |
| | 1490 | #endif |
| | 1491 | assem_debug("cmp %s,r14\n",regname[rs]); |
| | 1492 | output_w32(0xe1500000|rd_rn_rm(0,rs,HOST_TEMPREG)); |
| | 1493 | }else{ |
| | 1494 | assert(imm>-65536); |
| | 1495 | #ifdef ARMv5_ONLY |
| | 1496 | emit_movimm(-imm,HOST_TEMPREG); |
| | 1497 | #else |
| | 1498 | emit_movw(-imm,HOST_TEMPREG); |
| | 1499 | #endif |
| | 1500 | assem_debug("cmn %s,r14\n",regname[rs]); |
| | 1501 | output_w32(0xe1700000|rd_rn_rm(0,rs,HOST_TEMPREG)); |
| | 1502 | } |
| | 1503 | } |
| | 1504 | |
| | 1505 | void emit_cmovne(u_int *addr,int rt) |
| | 1506 | { |
| | 1507 | assem_debug("cmovne %x,%%%s",(int)addr,regname[rt]); |
| | 1508 | assert(0); |
| | 1509 | } |
| | 1510 | void emit_cmovl(u_int *addr,int rt) |
| | 1511 | { |
| | 1512 | assem_debug("cmovl %x,%%%s",(int)addr,regname[rt]); |
| | 1513 | assert(0); |
| | 1514 | } |
| | 1515 | void emit_cmovs(u_int *addr,int rt) |
| | 1516 | { |
| | 1517 | assem_debug("cmovs %x,%%%s",(int)addr,regname[rt]); |
| | 1518 | assert(0); |
| | 1519 | } |
| | 1520 | void emit_cmovne_imm(int imm,int rt) |
| | 1521 | { |
| | 1522 | assem_debug("movne %s,#%d\n",regname[rt],imm); |
| | 1523 | u_int armval; |
| | 1524 | genimm_checked(imm,&armval); |
| | 1525 | output_w32(0x13a00000|rd_rn_rm(rt,0,0)|armval); |
| | 1526 | } |
| | 1527 | void emit_cmovl_imm(int imm,int rt) |
| | 1528 | { |
| | 1529 | assem_debug("movlt %s,#%d\n",regname[rt],imm); |
| | 1530 | u_int armval; |
| | 1531 | genimm_checked(imm,&armval); |
| | 1532 | output_w32(0xb3a00000|rd_rn_rm(rt,0,0)|armval); |
| | 1533 | } |
| | 1534 | void emit_cmovb_imm(int imm,int rt) |
| | 1535 | { |
| | 1536 | assem_debug("movcc %s,#%d\n",regname[rt],imm); |
| | 1537 | u_int armval; |
| | 1538 | genimm_checked(imm,&armval); |
| | 1539 | output_w32(0x33a00000|rd_rn_rm(rt,0,0)|armval); |
| | 1540 | } |
| | 1541 | void emit_cmovs_imm(int imm,int rt) |
| | 1542 | { |
| | 1543 | assem_debug("movmi %s,#%d\n",regname[rt],imm); |
| | 1544 | u_int armval; |
| | 1545 | genimm_checked(imm,&armval); |
| | 1546 | output_w32(0x43a00000|rd_rn_rm(rt,0,0)|armval); |
| | 1547 | } |
| | 1548 | void emit_cmove_reg(int rs,int rt) |
| | 1549 | { |
| | 1550 | assem_debug("moveq %s,%s\n",regname[rt],regname[rs]); |
| | 1551 | output_w32(0x01a00000|rd_rn_rm(rt,0,rs)); |
| | 1552 | } |
| | 1553 | void emit_cmovne_reg(int rs,int rt) |
| | 1554 | { |
| | 1555 | assem_debug("movne %s,%s\n",regname[rt],regname[rs]); |
| | 1556 | output_w32(0x11a00000|rd_rn_rm(rt,0,rs)); |
| | 1557 | } |
| | 1558 | void emit_cmovl_reg(int rs,int rt) |
| | 1559 | { |
| | 1560 | assem_debug("movlt %s,%s\n",regname[rt],regname[rs]); |
| | 1561 | output_w32(0xb1a00000|rd_rn_rm(rt,0,rs)); |
| | 1562 | } |
| | 1563 | void emit_cmovs_reg(int rs,int rt) |
| | 1564 | { |
| | 1565 | assem_debug("movmi %s,%s\n",regname[rt],regname[rs]); |
| | 1566 | output_w32(0x41a00000|rd_rn_rm(rt,0,rs)); |
| | 1567 | } |
| | 1568 | |
| | 1569 | void emit_slti32(int rs,int imm,int rt) |
| | 1570 | { |
| | 1571 | if(rs!=rt) emit_zeroreg(rt); |
| | 1572 | emit_cmpimm(rs,imm); |
| | 1573 | if(rs==rt) emit_movimm(0,rt); |
| | 1574 | emit_cmovl_imm(1,rt); |
| | 1575 | } |
| | 1576 | void emit_sltiu32(int rs,int imm,int rt) |
| | 1577 | { |
| | 1578 | if(rs!=rt) emit_zeroreg(rt); |
| | 1579 | emit_cmpimm(rs,imm); |
| | 1580 | if(rs==rt) emit_movimm(0,rt); |
| | 1581 | emit_cmovb_imm(1,rt); |
| | 1582 | } |
| | 1583 | void emit_slti64_32(int rsh,int rsl,int imm,int rt) |
| | 1584 | { |
| | 1585 | assert(rsh!=rt); |
| | 1586 | emit_slti32(rsl,imm,rt); |
| | 1587 | if(imm>=0) |
| | 1588 | { |
| | 1589 | emit_test(rsh,rsh); |
| | 1590 | emit_cmovne_imm(0,rt); |
| | 1591 | emit_cmovs_imm(1,rt); |
| | 1592 | } |
| | 1593 | else |
| | 1594 | { |
| | 1595 | emit_cmpimm(rsh,-1); |
| | 1596 | emit_cmovne_imm(0,rt); |
| | 1597 | emit_cmovl_imm(1,rt); |
| | 1598 | } |
| | 1599 | } |
| | 1600 | void emit_sltiu64_32(int rsh,int rsl,int imm,int rt) |
| | 1601 | { |
| | 1602 | assert(rsh!=rt); |
| | 1603 | emit_sltiu32(rsl,imm,rt); |
| | 1604 | if(imm>=0) |
| | 1605 | { |
| | 1606 | emit_test(rsh,rsh); |
| | 1607 | emit_cmovne_imm(0,rt); |
| | 1608 | } |
| | 1609 | else |
| | 1610 | { |
| | 1611 | emit_cmpimm(rsh,-1); |
| | 1612 | emit_cmovne_imm(1,rt); |
| | 1613 | } |
| | 1614 | } |
| | 1615 | |
| | 1616 | void emit_cmp(int rs,int rt) |
| | 1617 | { |
| | 1618 | assem_debug("cmp %s,%s\n",regname[rs],regname[rt]); |
| | 1619 | output_w32(0xe1500000|rd_rn_rm(0,rs,rt)); |
| | 1620 | } |
| | 1621 | void emit_set_gz32(int rs, int rt) |
| | 1622 | { |
| | 1623 | //assem_debug("set_gz32\n"); |
| | 1624 | emit_cmpimm(rs,1); |
| | 1625 | emit_movimm(1,rt); |
| | 1626 | emit_cmovl_imm(0,rt); |
| | 1627 | } |
| | 1628 | void emit_set_nz32(int rs, int rt) |
| | 1629 | { |
| | 1630 | //assem_debug("set_nz32\n"); |
| | 1631 | if(rs!=rt) emit_movs(rs,rt); |
| | 1632 | else emit_test(rs,rs); |
| | 1633 | emit_cmovne_imm(1,rt); |
| | 1634 | } |
| | 1635 | void emit_set_gz64_32(int rsh, int rsl, int rt) |
| | 1636 | { |
| | 1637 | //assem_debug("set_gz64\n"); |
| | 1638 | emit_set_gz32(rsl,rt); |
| | 1639 | emit_test(rsh,rsh); |
| | 1640 | emit_cmovne_imm(1,rt); |
| | 1641 | emit_cmovs_imm(0,rt); |
| | 1642 | } |
| | 1643 | void emit_set_nz64_32(int rsh, int rsl, int rt) |
| | 1644 | { |
| | 1645 | //assem_debug("set_nz64\n"); |
| | 1646 | emit_or_and_set_flags(rsh,rsl,rt); |
| | 1647 | emit_cmovne_imm(1,rt); |
| | 1648 | } |
| | 1649 | void emit_set_if_less32(int rs1, int rs2, int rt) |
| | 1650 | { |
| | 1651 | //assem_debug("set if less (%%%s,%%%s),%%%s\n",regname[rs1],regname[rs2],regname[rt]); |
| | 1652 | if(rs1!=rt&&rs2!=rt) emit_zeroreg(rt); |
| | 1653 | emit_cmp(rs1,rs2); |
| | 1654 | if(rs1==rt||rs2==rt) emit_movimm(0,rt); |
| | 1655 | emit_cmovl_imm(1,rt); |
| | 1656 | } |
| | 1657 | void emit_set_if_carry32(int rs1, int rs2, int rt) |
| | 1658 | { |
| | 1659 | //assem_debug("set if carry (%%%s,%%%s),%%%s\n",regname[rs1],regname[rs2],regname[rt]); |
| | 1660 | if(rs1!=rt&&rs2!=rt) emit_zeroreg(rt); |
| | 1661 | emit_cmp(rs1,rs2); |
| | 1662 | if(rs1==rt||rs2==rt) emit_movimm(0,rt); |
| | 1663 | emit_cmovb_imm(1,rt); |
| | 1664 | } |
| | 1665 | void emit_set_if_less64_32(int u1, int l1, int u2, int l2, int rt) |
| | 1666 | { |
| | 1667 | //assem_debug("set if less64 (%%%s,%%%s,%%%s,%%%s),%%%s\n",regname[u1],regname[l1],regname[u2],regname[l2],regname[rt]); |
| | 1668 | assert(u1!=rt); |
| | 1669 | assert(u2!=rt); |
| | 1670 | emit_cmp(l1,l2); |
| | 1671 | emit_movimm(0,rt); |
| | 1672 | emit_sbcs(u1,u2,HOST_TEMPREG); |
| | 1673 | emit_cmovl_imm(1,rt); |
| | 1674 | } |
| | 1675 | void emit_set_if_carry64_32(int u1, int l1, int u2, int l2, int rt) |
| | 1676 | { |
| | 1677 | //assem_debug("set if carry64 (%%%s,%%%s,%%%s,%%%s),%%%s\n",regname[u1],regname[l1],regname[u2],regname[l2],regname[rt]); |
| | 1678 | assert(u1!=rt); |
| | 1679 | assert(u2!=rt); |
| | 1680 | emit_cmp(l1,l2); |
| | 1681 | emit_movimm(0,rt); |
| | 1682 | emit_sbcs(u1,u2,HOST_TEMPREG); |
| | 1683 | emit_cmovb_imm(1,rt); |
| | 1684 | } |
| | 1685 | |
| | 1686 | void emit_call(int a) |
| | 1687 | { |
| | 1688 | assem_debug("bl %x (%x+%x)\n",a,(int)out,a-(int)out-8); |
| | 1689 | u_int offset=genjmp(a); |
| | 1690 | output_w32(0xeb000000|offset); |
| | 1691 | } |
| | 1692 | void emit_jmp(int a) |
| | 1693 | { |
| | 1694 | assem_debug("b %x (%x+%x)\n",a,(int)out,a-(int)out-8); |
| | 1695 | u_int offset=genjmp(a); |
| | 1696 | output_w32(0xea000000|offset); |
| | 1697 | } |
| | 1698 | void emit_jne(int a) |
| | 1699 | { |
| | 1700 | assem_debug("bne %x\n",a); |
| | 1701 | u_int offset=genjmp(a); |
| | 1702 | output_w32(0x1a000000|offset); |
| | 1703 | } |
| | 1704 | void emit_jeq(int a) |
| | 1705 | { |
| | 1706 | assem_debug("beq %x\n",a); |
| | 1707 | u_int offset=genjmp(a); |
| | 1708 | output_w32(0x0a000000|offset); |
| | 1709 | } |
| | 1710 | void emit_js(int a) |
| | 1711 | { |
| | 1712 | assem_debug("bmi %x\n",a); |
| | 1713 | u_int offset=genjmp(a); |
| | 1714 | output_w32(0x4a000000|offset); |
| | 1715 | } |
| | 1716 | void emit_jns(int a) |
| | 1717 | { |
| | 1718 | assem_debug("bpl %x\n",a); |
| | 1719 | u_int offset=genjmp(a); |
| | 1720 | output_w32(0x5a000000|offset); |
| | 1721 | } |
| | 1722 | void emit_jl(int a) |
| | 1723 | { |
| | 1724 | assem_debug("blt %x\n",a); |
| | 1725 | u_int offset=genjmp(a); |
| | 1726 | output_w32(0xba000000|offset); |
| | 1727 | } |
| | 1728 | void emit_jge(int a) |
| | 1729 | { |
| | 1730 | assem_debug("bge %x\n",a); |
| | 1731 | u_int offset=genjmp(a); |
| | 1732 | output_w32(0xaa000000|offset); |
| | 1733 | } |
| | 1734 | void emit_jno(int a) |
| | 1735 | { |
| | 1736 | assem_debug("bvc %x\n",a); |
| | 1737 | u_int offset=genjmp(a); |
| | 1738 | output_w32(0x7a000000|offset); |
| | 1739 | } |
| | 1740 | void emit_jc(int a) |
| | 1741 | { |
| | 1742 | assem_debug("bcs %x\n",a); |
| | 1743 | u_int offset=genjmp(a); |
| | 1744 | output_w32(0x2a000000|offset); |
| | 1745 | } |
| | 1746 | void emit_jcc(int a) |
| | 1747 | { |
| | 1748 | assem_debug("bcc %x\n",a); |
| | 1749 | u_int offset=genjmp(a); |
| | 1750 | output_w32(0x3a000000|offset); |
| | 1751 | } |
| | 1752 | |
| | 1753 | void emit_pushimm(int imm) |
| | 1754 | { |
| | 1755 | assem_debug("push $%x\n",imm); |
| | 1756 | assert(0); |
| | 1757 | } |
| | 1758 | void emit_pusha() |
| | 1759 | { |
| | 1760 | assem_debug("pusha\n"); |
| | 1761 | assert(0); |
| | 1762 | } |
| | 1763 | void emit_popa() |
| | 1764 | { |
| | 1765 | assem_debug("popa\n"); |
| | 1766 | assert(0); |
| | 1767 | } |
| | 1768 | void emit_pushreg(u_int r) |
| | 1769 | { |
| | 1770 | assem_debug("push %%%s\n",regname[r]); |
| | 1771 | assert(0); |
| | 1772 | } |
| | 1773 | void emit_popreg(u_int r) |
| | 1774 | { |
| | 1775 | assem_debug("pop %%%s\n",regname[r]); |
| | 1776 | assert(0); |
| | 1777 | } |
| | 1778 | void emit_callreg(u_int r) |
| | 1779 | { |
| | 1780 | assert(r<15); |
| | 1781 | assem_debug("blx %s\n",regname[r]); |
| | 1782 | output_w32(0xe12fff30|r); |
| | 1783 | } |
| | 1784 | void emit_jmpreg(u_int r) |
| | 1785 | { |
| | 1786 | assem_debug("mov pc,%s\n",regname[r]); |
| | 1787 | output_w32(0xe1a00000|rd_rn_rm(15,0,r)); |
| | 1788 | } |
| | 1789 | |
| | 1790 | void emit_readword_indexed(int offset, int rs, int rt) |
| | 1791 | { |
| | 1792 | assert(offset>-4096&&offset<4096); |
| | 1793 | assem_debug("ldr %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 1794 | if(offset>=0) { |
| | 1795 | output_w32(0xe5900000|rd_rn_rm(rt,rs,0)|offset); |
| | 1796 | }else{ |
| | 1797 | output_w32(0xe5100000|rd_rn_rm(rt,rs,0)|(-offset)); |
| | 1798 | } |
| | 1799 | } |
| | 1800 | void emit_readword_dualindexedx4(int rs1, int rs2, int rt) |
| | 1801 | { |
| | 1802 | assem_debug("ldr %s,%s,%s lsl #2\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1803 | output_w32(0xe7900000|rd_rn_rm(rt,rs1,rs2)|0x100); |
| | 1804 | } |
| | 1805 | void emit_ldrcc_dualindexed(int rs1, int rs2, int rt) |
| | 1806 | { |
| | 1807 | assem_debug("ldrcc %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1808 | output_w32(0x37900000|rd_rn_rm(rt,rs1,rs2)); |
| | 1809 | } |
| | 1810 | void emit_ldrccb_dualindexed(int rs1, int rs2, int rt) |
| | 1811 | { |
| | 1812 | assem_debug("ldrccb %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1813 | output_w32(0x37d00000|rd_rn_rm(rt,rs1,rs2)); |
| | 1814 | } |
| | 1815 | void emit_ldrccsb_dualindexed(int rs1, int rs2, int rt) |
| | 1816 | { |
| | 1817 | assem_debug("ldrccsb %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1818 | output_w32(0x319000d0|rd_rn_rm(rt,rs1,rs2)); |
| | 1819 | } |
| | 1820 | void emit_ldrcch_dualindexed(int rs1, int rs2, int rt) |
| | 1821 | { |
| | 1822 | assem_debug("ldrcch %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1823 | output_w32(0x319000b0|rd_rn_rm(rt,rs1,rs2)); |
| | 1824 | } |
| | 1825 | void emit_ldrccsh_dualindexed(int rs1, int rs2, int rt) |
| | 1826 | { |
| | 1827 | assem_debug("ldrccsh %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1828 | output_w32(0x319000f0|rd_rn_rm(rt,rs1,rs2)); |
| | 1829 | } |
| | 1830 | void emit_readword_indexed_tlb(int addr, int rs, int map, int rt) |
| | 1831 | { |
| | 1832 | if(map<0) emit_readword_indexed(addr, rs, rt); |
| | 1833 | else { |
| | 1834 | assert(addr==0); |
| | 1835 | emit_readword_dualindexedx4(rs, map, rt); |
| | 1836 | } |
| | 1837 | } |
| | 1838 | void emit_readdword_indexed_tlb(int addr, int rs, int map, int rh, int rl) |
| | 1839 | { |
| | 1840 | if(map<0) { |
| | 1841 | if(rh>=0) emit_readword_indexed(addr, rs, rh); |
| | 1842 | emit_readword_indexed(addr+4, rs, rl); |
| | 1843 | }else{ |
| | 1844 | assert(rh!=rs); |
| | 1845 | if(rh>=0) emit_readword_indexed_tlb(addr, rs, map, rh); |
| | 1846 | emit_addimm(map,1,map); |
| | 1847 | emit_readword_indexed_tlb(addr, rs, map, rl); |
| | 1848 | } |
| | 1849 | } |
| | 1850 | void emit_movsbl_indexed(int offset, int rs, int rt) |
| | 1851 | { |
| | 1852 | assert(offset>-256&&offset<256); |
| | 1853 | assem_debug("ldrsb %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 1854 | if(offset>=0) { |
| | 1855 | output_w32(0xe1d000d0|rd_rn_rm(rt,rs,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 1856 | }else{ |
| | 1857 | output_w32(0xe15000d0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf)); |
| | 1858 | } |
| | 1859 | } |
| | 1860 | void emit_movsbl_indexed_tlb(int addr, int rs, int map, int rt) |
| | 1861 | { |
| | 1862 | if(map<0) emit_movsbl_indexed(addr, rs, rt); |
| | 1863 | else { |
| | 1864 | if(addr==0) { |
| | 1865 | emit_shlimm(map,2,map); |
| | 1866 | assem_debug("ldrsb %s,%s+%s\n",regname[rt],regname[rs],regname[map]); |
| | 1867 | output_w32(0xe19000d0|rd_rn_rm(rt,rs,map)); |
| | 1868 | }else{ |
| | 1869 | assert(addr>-256&&addr<256); |
| | 1870 | assem_debug("add %s,%s,%s,lsl #2\n",regname[rt],regname[rs],regname[map]); |
| | 1871 | output_w32(0xe0800000|rd_rn_rm(rt,rs,map)|(2<<7)); |
| | 1872 | emit_movsbl_indexed(addr, rt, rt); |
| | 1873 | } |
| | 1874 | } |
| | 1875 | } |
| | 1876 | void emit_movswl_indexed(int offset, int rs, int rt) |
| | 1877 | { |
| | 1878 | assert(offset>-256&&offset<256); |
| | 1879 | assem_debug("ldrsh %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 1880 | if(offset>=0) { |
| | 1881 | output_w32(0xe1d000f0|rd_rn_rm(rt,rs,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 1882 | }else{ |
| | 1883 | output_w32(0xe15000f0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf)); |
| | 1884 | } |
| | 1885 | } |
| | 1886 | void emit_movzbl_indexed(int offset, int rs, int rt) |
| | 1887 | { |
| | 1888 | assert(offset>-4096&&offset<4096); |
| | 1889 | assem_debug("ldrb %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 1890 | if(offset>=0) { |
| | 1891 | output_w32(0xe5d00000|rd_rn_rm(rt,rs,0)|offset); |
| | 1892 | }else{ |
| | 1893 | output_w32(0xe5500000|rd_rn_rm(rt,rs,0)|(-offset)); |
| | 1894 | } |
| | 1895 | } |
| | 1896 | void emit_movzbl_dualindexedx4(int rs1, int rs2, int rt) |
| | 1897 | { |
| | 1898 | assem_debug("ldrb %s,%s,%s lsl #2\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1899 | output_w32(0xe7d00000|rd_rn_rm(rt,rs1,rs2)|0x100); |
| | 1900 | } |
| | 1901 | void emit_movzbl_indexed_tlb(int addr, int rs, int map, int rt) |
| | 1902 | { |
| | 1903 | if(map<0) emit_movzbl_indexed(addr, rs, rt); |
| | 1904 | else { |
| | 1905 | if(addr==0) { |
| | 1906 | emit_movzbl_dualindexedx4(rs, map, rt); |
| | 1907 | }else{ |
| | 1908 | emit_addimm(rs,addr,rt); |
| | 1909 | emit_movzbl_dualindexedx4(rt, map, rt); |
| | 1910 | } |
| | 1911 | } |
| | 1912 | } |
| | 1913 | void emit_movzwl_indexed(int offset, int rs, int rt) |
| | 1914 | { |
| | 1915 | assert(offset>-256&&offset<256); |
| | 1916 | assem_debug("ldrh %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 1917 | if(offset>=0) { |
| | 1918 | output_w32(0xe1d000b0|rd_rn_rm(rt,rs,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 1919 | }else{ |
| | 1920 | output_w32(0xe15000b0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf)); |
| | 1921 | } |
| | 1922 | } |
| | 1923 | void emit_readword(int addr, int rt) |
| | 1924 | { |
| | 1925 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1926 | assert(offset<4096); |
| | 1927 | assem_debug("ldr %s,fp+%d\n",regname[rt],offset); |
| | 1928 | output_w32(0xe5900000|rd_rn_rm(rt,FP,0)|offset); |
| | 1929 | } |
| | 1930 | void emit_movsbl(int addr, int rt) |
| | 1931 | { |
| | 1932 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1933 | assert(offset<256); |
| | 1934 | assem_debug("ldrsb %s,fp+%d\n",regname[rt],offset); |
| | 1935 | output_w32(0xe1d000d0|rd_rn_rm(rt,FP,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 1936 | } |
| | 1937 | void emit_movswl(int addr, int rt) |
| | 1938 | { |
| | 1939 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1940 | assert(offset<256); |
| | 1941 | assem_debug("ldrsh %s,fp+%d\n",regname[rt],offset); |
| | 1942 | output_w32(0xe1d000f0|rd_rn_rm(rt,FP,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 1943 | } |
| | 1944 | void emit_movzbl(int addr, int rt) |
| | 1945 | { |
| | 1946 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1947 | assert(offset<4096); |
| | 1948 | assem_debug("ldrb %s,fp+%d\n",regname[rt],offset); |
| | 1949 | output_w32(0xe5d00000|rd_rn_rm(rt,FP,0)|offset); |
| | 1950 | } |
| | 1951 | void emit_movzwl(int addr, int rt) |
| | 1952 | { |
| | 1953 | u_int offset = addr-(u_int)&dynarec_local; |
| | 1954 | assert(offset<256); |
| | 1955 | assem_debug("ldrh %s,fp+%d\n",regname[rt],offset); |
| | 1956 | output_w32(0xe1d000b0|rd_rn_rm(rt,FP,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 1957 | } |
| | 1958 | void emit_movzwl_reg(int rs, int rt) |
| | 1959 | { |
| | 1960 | assem_debug("movzwl %%%s,%%%s\n",regname[rs]+1,regname[rt]); |
| | 1961 | assert(0); |
| | 1962 | } |
| | 1963 | |
| | 1964 | void emit_xchg(int rs, int rt) |
| | 1965 | { |
| | 1966 | assem_debug("xchg %%%s,%%%s\n",regname[rs],regname[rt]); |
| | 1967 | assert(0); |
| | 1968 | } |
| | 1969 | void emit_writeword_indexed(int rt, int offset, int rs) |
| | 1970 | { |
| | 1971 | assert(offset>-4096&&offset<4096); |
| | 1972 | assem_debug("str %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 1973 | if(offset>=0) { |
| | 1974 | output_w32(0xe5800000|rd_rn_rm(rt,rs,0)|offset); |
| | 1975 | }else{ |
| | 1976 | output_w32(0xe5000000|rd_rn_rm(rt,rs,0)|(-offset)); |
| | 1977 | } |
| | 1978 | } |
| | 1979 | void emit_writeword_dualindexedx4(int rt, int rs1, int rs2) |
| | 1980 | { |
| | 1981 | assem_debug("str %s,%s,%s lsl #2\n",regname[rt],regname[rs1],regname[rs2]); |
| | 1982 | output_w32(0xe7800000|rd_rn_rm(rt,rs1,rs2)|0x100); |
| | 1983 | } |
| | 1984 | void emit_writeword_indexed_tlb(int rt, int addr, int rs, int map, int temp) |
| | 1985 | { |
| | 1986 | if(map<0) emit_writeword_indexed(rt, addr, rs); |
| | 1987 | else { |
| | 1988 | assert(addr==0); |
| | 1989 | emit_writeword_dualindexedx4(rt, rs, map); |
| | 1990 | } |
| | 1991 | } |
| | 1992 | void emit_writedword_indexed_tlb(int rh, int rl, int addr, int rs, int map, int temp) |
| | 1993 | { |
| | 1994 | if(map<0) { |
| | 1995 | if(rh>=0) emit_writeword_indexed(rh, addr, rs); |
| | 1996 | emit_writeword_indexed(rl, addr+4, rs); |
| | 1997 | }else{ |
| | 1998 | assert(rh>=0); |
| | 1999 | if(temp!=rs) emit_addimm(map,1,temp); |
| | 2000 | emit_writeword_indexed_tlb(rh, addr, rs, map, temp); |
| | 2001 | if(temp!=rs) emit_writeword_indexed_tlb(rl, addr, rs, temp, temp); |
| | 2002 | else { |
| | 2003 | emit_addimm(rs,4,rs); |
| | 2004 | emit_writeword_indexed_tlb(rl, addr, rs, map, temp); |
| | 2005 | } |
| | 2006 | } |
| | 2007 | } |
| | 2008 | void emit_writehword_indexed(int rt, int offset, int rs) |
| | 2009 | { |
| | 2010 | assert(offset>-256&&offset<256); |
| | 2011 | assem_debug("strh %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 2012 | if(offset>=0) { |
| | 2013 | output_w32(0xe1c000b0|rd_rn_rm(rt,rs,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 2014 | }else{ |
| | 2015 | output_w32(0xe14000b0|rd_rn_rm(rt,rs,0)|(((-offset)<<4)&0xf00)|((-offset)&0xf)); |
| | 2016 | } |
| | 2017 | } |
| | 2018 | void emit_writebyte_indexed(int rt, int offset, int rs) |
| | 2019 | { |
| | 2020 | assert(offset>-4096&&offset<4096); |
| | 2021 | assem_debug("strb %s,%s+%d\n",regname[rt],regname[rs],offset); |
| | 2022 | if(offset>=0) { |
| | 2023 | output_w32(0xe5c00000|rd_rn_rm(rt,rs,0)|offset); |
| | 2024 | }else{ |
| | 2025 | output_w32(0xe5400000|rd_rn_rm(rt,rs,0)|(-offset)); |
| | 2026 | } |
| | 2027 | } |
| | 2028 | void emit_writebyte_dualindexedx4(int rt, int rs1, int rs2) |
| | 2029 | { |
| | 2030 | assem_debug("strb %s,%s,%s lsl #2\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2031 | output_w32(0xe7c00000|rd_rn_rm(rt,rs1,rs2)|0x100); |
| | 2032 | } |
| | 2033 | void emit_writebyte_indexed_tlb(int rt, int addr, int rs, int map, int temp) |
| | 2034 | { |
| | 2035 | if(map<0) emit_writebyte_indexed(rt, addr, rs); |
| | 2036 | else { |
| | 2037 | if(addr==0) { |
| | 2038 | emit_writebyte_dualindexedx4(rt, rs, map); |
| | 2039 | }else{ |
| | 2040 | emit_addimm(rs,addr,temp); |
| | 2041 | emit_writebyte_dualindexedx4(rt, temp, map); |
| | 2042 | } |
| | 2043 | } |
| | 2044 | } |
| | 2045 | void emit_strcc_dualindexed(int rs1, int rs2, int rt) |
| | 2046 | { |
| | 2047 | assem_debug("strcc %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2048 | output_w32(0x37800000|rd_rn_rm(rt,rs1,rs2)); |
| | 2049 | } |
| | 2050 | void emit_strccb_dualindexed(int rs1, int rs2, int rt) |
| | 2051 | { |
| | 2052 | assem_debug("strccb %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2053 | output_w32(0x37c00000|rd_rn_rm(rt,rs1,rs2)); |
| | 2054 | } |
| | 2055 | void emit_strcch_dualindexed(int rs1, int rs2, int rt) |
| | 2056 | { |
| | 2057 | assem_debug("strcch %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2058 | output_w32(0x318000b0|rd_rn_rm(rt,rs1,rs2)); |
| | 2059 | } |
| | 2060 | void emit_writeword(int rt, int addr) |
| | 2061 | { |
| | 2062 | u_int offset = addr-(u_int)&dynarec_local; |
| | 2063 | assert(offset<4096); |
| | 2064 | assem_debug("str %s,fp+%d\n",regname[rt],offset); |
| | 2065 | output_w32(0xe5800000|rd_rn_rm(rt,FP,0)|offset); |
| | 2066 | } |
| | 2067 | void emit_writehword(int rt, int addr) |
| | 2068 | { |
| | 2069 | u_int offset = addr-(u_int)&dynarec_local; |
| | 2070 | assert(offset<256); |
| | 2071 | assem_debug("strh %s,fp+%d\n",regname[rt],offset); |
| | 2072 | output_w32(0xe1c000b0|rd_rn_rm(rt,FP,0)|((offset<<4)&0xf00)|(offset&0xf)); |
| | 2073 | } |
| | 2074 | void emit_writebyte(int rt, int addr) |
| | 2075 | { |
| | 2076 | u_int offset = addr-(u_int)&dynarec_local; |
| | 2077 | assert(offset<4096); |
| | 2078 | assem_debug("strb %s,fp+%d\n",regname[rt],offset); |
| | 2079 | output_w32(0xe5c00000|rd_rn_rm(rt,FP,0)|offset); |
| | 2080 | } |
| | 2081 | void emit_writeword_imm(int imm, int addr) |
| | 2082 | { |
| | 2083 | assem_debug("movl $%x,%x\n",imm,addr); |
| | 2084 | assert(0); |
| | 2085 | } |
| | 2086 | void emit_writebyte_imm(int imm, int addr) |
| | 2087 | { |
| | 2088 | assem_debug("movb $%x,%x\n",imm,addr); |
| | 2089 | assert(0); |
| | 2090 | } |
| | 2091 | |
| | 2092 | void emit_mul(int rs) |
| | 2093 | { |
| | 2094 | assem_debug("mul %%%s\n",regname[rs]); |
| | 2095 | assert(0); |
| | 2096 | } |
| | 2097 | void emit_imul(int rs) |
| | 2098 | { |
| | 2099 | assem_debug("imul %%%s\n",regname[rs]); |
| | 2100 | assert(0); |
| | 2101 | } |
| | 2102 | void emit_umull(u_int rs1, u_int rs2, u_int hi, u_int lo) |
| | 2103 | { |
| | 2104 | assem_debug("umull %s, %s, %s, %s\n",regname[lo],regname[hi],regname[rs1],regname[rs2]); |
| | 2105 | assert(rs1<16); |
| | 2106 | assert(rs2<16); |
| | 2107 | assert(hi<16); |
| | 2108 | assert(lo<16); |
| | 2109 | output_w32(0xe0800090|(hi<<16)|(lo<<12)|(rs2<<8)|rs1); |
| | 2110 | } |
| | 2111 | void emit_smull(u_int rs1, u_int rs2, u_int hi, u_int lo) |
| | 2112 | { |
| | 2113 | assem_debug("smull %s, %s, %s, %s\n",regname[lo],regname[hi],regname[rs1],regname[rs2]); |
| | 2114 | assert(rs1<16); |
| | 2115 | assert(rs2<16); |
| | 2116 | assert(hi<16); |
| | 2117 | assert(lo<16); |
| | 2118 | output_w32(0xe0c00090|(hi<<16)|(lo<<12)|(rs2<<8)|rs1); |
| | 2119 | } |
| | 2120 | |
| | 2121 | void emit_div(int rs) |
| | 2122 | { |
| | 2123 | assem_debug("div %%%s\n",regname[rs]); |
| | 2124 | assert(0); |
| | 2125 | } |
| | 2126 | void emit_idiv(int rs) |
| | 2127 | { |
| | 2128 | assem_debug("idiv %%%s\n",regname[rs]); |
| | 2129 | assert(0); |
| | 2130 | } |
| | 2131 | void emit_cdq() |
| | 2132 | { |
| | 2133 | assem_debug("cdq\n"); |
| | 2134 | assert(0); |
| | 2135 | } |
| | 2136 | |
| | 2137 | void emit_clz(int rs,int rt) |
| | 2138 | { |
| | 2139 | assem_debug("clz %s,%s\n",regname[rt],regname[rs]); |
| | 2140 | output_w32(0xe16f0f10|rd_rn_rm(rt,0,rs)); |
| | 2141 | } |
| | 2142 | |
| | 2143 | void emit_subcs(int rs1,int rs2,int rt) |
| | 2144 | { |
| | 2145 | assem_debug("subcs %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2146 | output_w32(0x20400000|rd_rn_rm(rt,rs1,rs2)); |
| | 2147 | } |
| | 2148 | |
| | 2149 | void emit_shrcc_imm(int rs,u_int imm,int rt) |
| | 2150 | { |
| | 2151 | assert(imm>0); |
| | 2152 | assert(imm<32); |
| | 2153 | assem_debug("lsrcc %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2154 | output_w32(0x31a00000|rd_rn_rm(rt,0,rs)|0x20|(imm<<7)); |
| | 2155 | } |
| | 2156 | |
| | 2157 | void emit_negmi(int rs, int rt) |
| | 2158 | { |
| | 2159 | assem_debug("rsbmi %s,%s,#0\n",regname[rt],regname[rs]); |
| | 2160 | output_w32(0x42600000|rd_rn_rm(rt,rs,0)); |
| | 2161 | } |
| | 2162 | |
| | 2163 | void emit_negsmi(int rs, int rt) |
| | 2164 | { |
| | 2165 | assem_debug("rsbsmi %s,%s,#0\n",regname[rt],regname[rs]); |
| | 2166 | output_w32(0x42700000|rd_rn_rm(rt,rs,0)); |
| | 2167 | } |
| | 2168 | |
| | 2169 | void emit_orreq(u_int rs1,u_int rs2,u_int rt) |
| | 2170 | { |
| | 2171 | assem_debug("orreq %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2172 | output_w32(0x01800000|rd_rn_rm(rt,rs1,rs2)); |
| | 2173 | } |
| | 2174 | |
| | 2175 | void emit_orrne(u_int rs1,u_int rs2,u_int rt) |
| | 2176 | { |
| | 2177 | assem_debug("orrne %s,%s,%s\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2178 | output_w32(0x11800000|rd_rn_rm(rt,rs1,rs2)); |
| | 2179 | } |
| | 2180 | |
| | 2181 | void emit_bic_lsl(u_int rs1,u_int rs2,u_int shift,u_int rt) |
| | 2182 | { |
| | 2183 | assem_debug("bic %s,%s,%s lsl %s\n",regname[rt],regname[rs1],regname[rs2],regname[shift]); |
| | 2184 | output_w32(0xe1C00000|rd_rn_rm(rt,rs1,rs2)|0x10|(shift<<8)); |
| | 2185 | } |
| | 2186 | |
| | 2187 | void emit_biceq_lsl(u_int rs1,u_int rs2,u_int shift,u_int rt) |
| | 2188 | { |
| | 2189 | assem_debug("biceq %s,%s,%s lsl %s\n",regname[rt],regname[rs1],regname[rs2],regname[shift]); |
| | 2190 | output_w32(0x01C00000|rd_rn_rm(rt,rs1,rs2)|0x10|(shift<<8)); |
| | 2191 | } |
| | 2192 | |
| | 2193 | void emit_bicne_lsl(u_int rs1,u_int rs2,u_int shift,u_int rt) |
| | 2194 | { |
| | 2195 | assem_debug("bicne %s,%s,%s lsl %s\n",regname[rt],regname[rs1],regname[rs2],regname[shift]); |
| | 2196 | output_w32(0x11C00000|rd_rn_rm(rt,rs1,rs2)|0x10|(shift<<8)); |
| | 2197 | } |
| | 2198 | |
| | 2199 | void emit_bic_lsr(u_int rs1,u_int rs2,u_int shift,u_int rt) |
| | 2200 | { |
| | 2201 | assem_debug("bic %s,%s,%s lsr %s\n",regname[rt],regname[rs1],regname[rs2],regname[shift]); |
| | 2202 | output_w32(0xe1C00000|rd_rn_rm(rt,rs1,rs2)|0x30|(shift<<8)); |
| | 2203 | } |
| | 2204 | |
| | 2205 | void emit_biceq_lsr(u_int rs1,u_int rs2,u_int shift,u_int rt) |
| | 2206 | { |
| | 2207 | assem_debug("biceq %s,%s,%s lsr %s\n",regname[rt],regname[rs1],regname[rs2],regname[shift]); |
| | 2208 | output_w32(0x01C00000|rd_rn_rm(rt,rs1,rs2)|0x30|(shift<<8)); |
| | 2209 | } |
| | 2210 | |
| | 2211 | void emit_bicne_lsr(u_int rs1,u_int rs2,u_int shift,u_int rt) |
| | 2212 | { |
| | 2213 | assem_debug("bicne %s,%s,%s lsr %s\n",regname[rt],regname[rs1],regname[rs2],regname[shift]); |
| | 2214 | output_w32(0x11C00000|rd_rn_rm(rt,rs1,rs2)|0x30|(shift<<8)); |
| | 2215 | } |
| | 2216 | |
| | 2217 | void emit_teq(int rs, int rt) |
| | 2218 | { |
| | 2219 | assem_debug("teq %s,%s\n",regname[rs],regname[rt]); |
| | 2220 | output_w32(0xe1300000|rd_rn_rm(0,rs,rt)); |
| | 2221 | } |
| | 2222 | |
| | 2223 | void emit_rsbimm(int rs, int imm, int rt) |
| | 2224 | { |
| | 2225 | u_int armval; |
| | 2226 | genimm_checked(imm,&armval); |
| | 2227 | assem_debug("rsb %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2228 | output_w32(0xe2600000|rd_rn_rm(rt,rs,0)|armval); |
| | 2229 | } |
| | 2230 | |
| | 2231 | // Load 2 immediates optimizing for small code size |
| | 2232 | void emit_mov2imm_compact(int imm1,u_int rt1,int imm2,u_int rt2) |
| | 2233 | { |
| | 2234 | emit_movimm(imm1,rt1); |
| | 2235 | u_int armval; |
| | 2236 | if(genimm(imm2-imm1,&armval)) { |
| | 2237 | assem_debug("add %s,%s,#%d\n",regname[rt2],regname[rt1],imm2-imm1); |
| | 2238 | output_w32(0xe2800000|rd_rn_rm(rt2,rt1,0)|armval); |
| | 2239 | }else if(genimm(imm1-imm2,&armval)) { |
| | 2240 | assem_debug("sub %s,%s,#%d\n",regname[rt2],regname[rt1],imm1-imm2); |
| | 2241 | output_w32(0xe2400000|rd_rn_rm(rt2,rt1,0)|armval); |
| | 2242 | } |
| | 2243 | else emit_movimm(imm2,rt2); |
| | 2244 | } |
| | 2245 | |
| | 2246 | // Conditionally select one of two immediates, optimizing for small code size |
| | 2247 | // This will only be called if HAVE_CMOV_IMM is defined |
| | 2248 | void emit_cmov2imm_e_ne_compact(int imm1,int imm2,u_int rt) |
| | 2249 | { |
| | 2250 | u_int armval; |
| | 2251 | if(genimm(imm2-imm1,&armval)) { |
| | 2252 | emit_movimm(imm1,rt); |
| | 2253 | assem_debug("addne %s,%s,#%d\n",regname[rt],regname[rt],imm2-imm1); |
| | 2254 | output_w32(0x12800000|rd_rn_rm(rt,rt,0)|armval); |
| | 2255 | }else if(genimm(imm1-imm2,&armval)) { |
| | 2256 | emit_movimm(imm1,rt); |
| | 2257 | assem_debug("subne %s,%s,#%d\n",regname[rt],regname[rt],imm1-imm2); |
| | 2258 | output_w32(0x12400000|rd_rn_rm(rt,rt,0)|armval); |
| | 2259 | } |
| | 2260 | else { |
| | 2261 | #ifdef ARMv5_ONLY |
| | 2262 | emit_movimm(imm1,rt); |
| | 2263 | add_literal((int)out,imm2); |
| | 2264 | assem_debug("ldrne %s,pc+? [=%x]\n",regname[rt],imm2); |
| | 2265 | output_w32(0x15900000|rd_rn_rm(rt,15,0)); |
| | 2266 | #else |
| | 2267 | emit_movw(imm1&0x0000FFFF,rt); |
| | 2268 | if((imm1&0xFFFF)!=(imm2&0xFFFF)) { |
| | 2269 | assem_debug("movwne %s,#%d (0x%x)\n",regname[rt],imm2&0xFFFF,imm2&0xFFFF); |
| | 2270 | output_w32(0x13000000|rd_rn_rm(rt,0,0)|(imm2&0xfff)|((imm2<<4)&0xf0000)); |
| | 2271 | } |
| | 2272 | emit_movt(imm1&0xFFFF0000,rt); |
| | 2273 | if((imm1&0xFFFF0000)!=(imm2&0xFFFF0000)) { |
| | 2274 | assem_debug("movtne %s,#%d (0x%x)\n",regname[rt],imm2&0xffff0000,imm2&0xffff0000); |
| | 2275 | output_w32(0x13400000|rd_rn_rm(rt,0,0)|((imm2>>16)&0xfff)|((imm2>>12)&0xf0000)); |
| | 2276 | } |
| | 2277 | #endif |
| | 2278 | } |
| | 2279 | } |
| | 2280 | |
| | 2281 | // special case for checking invalid_code |
| | 2282 | void emit_cmpmem_indexedsr12_imm(int addr,int r,int imm) |
| | 2283 | { |
| | 2284 | assert(0); |
| | 2285 | } |
| | 2286 | |
| | 2287 | // special case for checking invalid_code |
| | 2288 | void emit_cmpmem_indexedsr12_reg(int base,int r,int imm) |
| | 2289 | { |
| | 2290 | assert(imm<128&&imm>=0); |
| | 2291 | assert(r>=0&&r<16); |
| | 2292 | assem_debug("ldrb lr,%s,%s lsr #12\n",regname[base],regname[r]); |
| | 2293 | output_w32(0xe7d00000|rd_rn_rm(HOST_TEMPREG,base,r)|0x620); |
| | 2294 | emit_cmpimm(HOST_TEMPREG,imm); |
| | 2295 | } |
| | 2296 | |
| | 2297 | // special case for tlb mapping |
| | 2298 | void emit_addsr12(int rs1,int rs2,int rt) |
| | 2299 | { |
| | 2300 | assem_debug("add %s,%s,%s lsr #12\n",regname[rt],regname[rs1],regname[rs2]); |
| | 2301 | output_w32(0xe0800620|rd_rn_rm(rt,rs1,rs2)); |
| | 2302 | } |
| | 2303 | |
| | 2304 | void emit_callne(int a) |
| | 2305 | { |
| | 2306 | assem_debug("blne %x\n",a); |
| | 2307 | u_int offset=genjmp(a); |
| | 2308 | output_w32(0x1b000000|offset); |
| | 2309 | } |
| | 2310 | |
| | 2311 | // Used to preload hash table entries |
| | 2312 | void emit_prefetch(void *addr) |
| | 2313 | { |
| | 2314 | assem_debug("prefetch %x\n",(int)addr); |
| | 2315 | output_byte(0x0F); |
| | 2316 | output_byte(0x18); |
| | 2317 | output_modrm(0,5,1); |
| | 2318 | output_w32((int)addr); |
| | 2319 | } |
| | 2320 | void emit_prefetchreg(int r) |
| | 2321 | { |
| | 2322 | assem_debug("pld %s\n",regname[r]); |
| | 2323 | output_w32(0xf5d0f000|rd_rn_rm(0,r,0)); |
| | 2324 | } |
| | 2325 | |
| | 2326 | // Special case for mini_ht |
| | 2327 | void emit_ldreq_indexed(int rs, u_int offset, int rt) |
| | 2328 | { |
| | 2329 | assert(offset<4096); |
| | 2330 | assem_debug("ldreq %s,[%s, #%d]\n",regname[rt],regname[rs],offset); |
| | 2331 | output_w32(0x05900000|rd_rn_rm(rt,rs,0)|offset); |
| | 2332 | } |
| | 2333 | |
| | 2334 | void emit_flds(int r,int sr) |
| | 2335 | { |
| | 2336 | assem_debug("flds s%d,[%s]\n",sr,regname[r]); |
| | 2337 | output_w32(0xed900a00|((sr&14)<<11)|((sr&1)<<22)|(r<<16)); |
| | 2338 | } |
| | 2339 | |
| | 2340 | void emit_vldr(int r,int vr) |
| | 2341 | { |
| | 2342 | assem_debug("vldr d%d,[%s]\n",vr,regname[r]); |
| | 2343 | output_w32(0xed900b00|(vr<<12)|(r<<16)); |
| | 2344 | } |
| | 2345 | |
| | 2346 | void emit_fsts(int sr,int r) |
| | 2347 | { |
| | 2348 | assem_debug("fsts s%d,[%s]\n",sr,regname[r]); |
| | 2349 | output_w32(0xed800a00|((sr&14)<<11)|((sr&1)<<22)|(r<<16)); |
| | 2350 | } |
| | 2351 | |
| | 2352 | void emit_vstr(int vr,int r) |
| | 2353 | { |
| | 2354 | assem_debug("vstr d%d,[%s]\n",vr,regname[r]); |
| | 2355 | output_w32(0xed800b00|(vr<<12)|(r<<16)); |
| | 2356 | } |
| | 2357 | |
| | 2358 | void emit_ftosizs(int s,int d) |
| | 2359 | { |
| | 2360 | assem_debug("ftosizs s%d,s%d\n",d,s); |
| | 2361 | output_w32(0xeebd0ac0|((d&14)<<11)|((d&1)<<22)|((s&14)>>1)|((s&1)<<5)); |
| | 2362 | } |
| | 2363 | |
| | 2364 | void emit_ftosizd(int s,int d) |
| | 2365 | { |
| | 2366 | assem_debug("ftosizd s%d,d%d\n",d,s); |
| | 2367 | output_w32(0xeebd0bc0|((d&14)<<11)|((d&1)<<22)|(s&7)); |
| | 2368 | } |
| | 2369 | |
| | 2370 | void emit_fsitos(int s,int d) |
| | 2371 | { |
| | 2372 | assem_debug("fsitos s%d,s%d\n",d,s); |
| | 2373 | output_w32(0xeeb80ac0|((d&14)<<11)|((d&1)<<22)|((s&14)>>1)|((s&1)<<5)); |
| | 2374 | } |
| | 2375 | |
| | 2376 | void emit_fsitod(int s,int d) |
| | 2377 | { |
| | 2378 | assem_debug("fsitod d%d,s%d\n",d,s); |
| | 2379 | output_w32(0xeeb80bc0|((d&7)<<12)|((s&14)>>1)|((s&1)<<5)); |
| | 2380 | } |
| | 2381 | |
| | 2382 | void emit_fcvtds(int s,int d) |
| | 2383 | { |
| | 2384 | assem_debug("fcvtds d%d,s%d\n",d,s); |
| | 2385 | output_w32(0xeeb70ac0|((d&7)<<12)|((s&14)>>1)|((s&1)<<5)); |
| | 2386 | } |
| | 2387 | |
| | 2388 | void emit_fcvtsd(int s,int d) |
| | 2389 | { |
| | 2390 | assem_debug("fcvtsd s%d,d%d\n",d,s); |
| | 2391 | output_w32(0xeeb70bc0|((d&14)<<11)|((d&1)<<22)|(s&7)); |
| | 2392 | } |
| | 2393 | |
| | 2394 | void emit_fsqrts(int s,int d) |
| | 2395 | { |
| | 2396 | assem_debug("fsqrts d%d,s%d\n",d,s); |
| | 2397 | output_w32(0xeeb10ac0|((d&14)<<11)|((d&1)<<22)|((s&14)>>1)|((s&1)<<5)); |
| | 2398 | } |
| | 2399 | |
| | 2400 | void emit_fsqrtd(int s,int d) |
| | 2401 | { |
| | 2402 | assem_debug("fsqrtd s%d,d%d\n",d,s); |
| | 2403 | output_w32(0xeeb10bc0|((d&7)<<12)|(s&7)); |
| | 2404 | } |
| | 2405 | |
| | 2406 | void emit_fabss(int s,int d) |
| | 2407 | { |
| | 2408 | assem_debug("fabss d%d,s%d\n",d,s); |
| | 2409 | output_w32(0xeeb00ac0|((d&14)<<11)|((d&1)<<22)|((s&14)>>1)|((s&1)<<5)); |
| | 2410 | } |
| | 2411 | |
| | 2412 | void emit_fabsd(int s,int d) |
| | 2413 | { |
| | 2414 | assem_debug("fabsd s%d,d%d\n",d,s); |
| | 2415 | output_w32(0xeeb00bc0|((d&7)<<12)|(s&7)); |
| | 2416 | } |
| | 2417 | |
| | 2418 | void emit_fnegs(int s,int d) |
| | 2419 | { |
| | 2420 | assem_debug("fnegs d%d,s%d\n",d,s); |
| | 2421 | output_w32(0xeeb10a40|((d&14)<<11)|((d&1)<<22)|((s&14)>>1)|((s&1)<<5)); |
| | 2422 | } |
| | 2423 | |
| | 2424 | void emit_fnegd(int s,int d) |
| | 2425 | { |
| | 2426 | assem_debug("fnegd s%d,d%d\n",d,s); |
| | 2427 | output_w32(0xeeb10b40|((d&7)<<12)|(s&7)); |
| | 2428 | } |
| | 2429 | |
| | 2430 | void emit_fadds(int s1,int s2,int d) |
| | 2431 | { |
| | 2432 | assem_debug("fadds s%d,s%d,s%d\n",d,s1,s2); |
| | 2433 | output_w32(0xee300a00|((d&14)<<11)|((d&1)<<22)|((s1&14)<<15)|((s1&1)<<7)|((s2&14)>>1)|((s2&1)<<5)); |
| | 2434 | } |
| | 2435 | |
| | 2436 | void emit_faddd(int s1,int s2,int d) |
| | 2437 | { |
| | 2438 | assem_debug("faddd d%d,d%d,d%d\n",d,s1,s2); |
| | 2439 | output_w32(0xee300b00|((d&7)<<12)|((s1&7)<<16)|(s2&7)); |
| | 2440 | } |
| | 2441 | |
| | 2442 | void emit_fsubs(int s1,int s2,int d) |
| | 2443 | { |
| | 2444 | assem_debug("fsubs s%d,s%d,s%d\n",d,s1,s2); |
| | 2445 | output_w32(0xee300a40|((d&14)<<11)|((d&1)<<22)|((s1&14)<<15)|((s1&1)<<7)|((s2&14)>>1)|((s2&1)<<5)); |
| | 2446 | } |
| | 2447 | |
| | 2448 | void emit_fsubd(int s1,int s2,int d) |
| | 2449 | { |
| | 2450 | assem_debug("fsubd d%d,d%d,d%d\n",d,s1,s2); |
| | 2451 | output_w32(0xee300b40|((d&7)<<12)|((s1&7)<<16)|(s2&7)); |
| | 2452 | } |
| | 2453 | |
| | 2454 | void emit_fmuls(int s1,int s2,int d) |
| | 2455 | { |
| | 2456 | assem_debug("fmuls s%d,s%d,s%d\n",d,s1,s2); |
| | 2457 | output_w32(0xee200a00|((d&14)<<11)|((d&1)<<22)|((s1&14)<<15)|((s1&1)<<7)|((s2&14)>>1)|((s2&1)<<5)); |
| | 2458 | } |
| | 2459 | |
| | 2460 | void emit_fmuld(int s1,int s2,int d) |
| | 2461 | { |
| | 2462 | assem_debug("fmuld d%d,d%d,d%d\n",d,s1,s2); |
| | 2463 | output_w32(0xee200b00|((d&7)<<12)|((s1&7)<<16)|(s2&7)); |
| | 2464 | } |
| | 2465 | |
| | 2466 | void emit_fdivs(int s1,int s2,int d) |
| | 2467 | { |
| | 2468 | assem_debug("fdivs s%d,s%d,s%d\n",d,s1,s2); |
| | 2469 | output_w32(0xee800a00|((d&14)<<11)|((d&1)<<22)|((s1&14)<<15)|((s1&1)<<7)|((s2&14)>>1)|((s2&1)<<5)); |
| | 2470 | } |
| | 2471 | |
| | 2472 | void emit_fdivd(int s1,int s2,int d) |
| | 2473 | { |
| | 2474 | assem_debug("fdivd d%d,d%d,d%d\n",d,s1,s2); |
| | 2475 | output_w32(0xee800b00|((d&7)<<12)|((s1&7)<<16)|(s2&7)); |
| | 2476 | } |
| | 2477 | |
| | 2478 | void emit_fcmps(int x,int y) |
| | 2479 | { |
| | 2480 | assem_debug("fcmps s14, s15\n"); |
| | 2481 | output_w32(0xeeb47a67); |
| | 2482 | } |
| | 2483 | |
| | 2484 | void emit_fcmpd(int x,int y) |
| | 2485 | { |
| | 2486 | assem_debug("fcmpd d6, d7\n"); |
| | 2487 | output_w32(0xeeb46b47); |
| | 2488 | } |
| | 2489 | |
| | 2490 | void emit_fmstat() |
| | 2491 | { |
| | 2492 | assem_debug("fmstat\n"); |
| | 2493 | output_w32(0xeef1fa10); |
| | 2494 | } |
| | 2495 | |
| | 2496 | void emit_bicne_imm(int rs,int imm,int rt) |
| | 2497 | { |
| | 2498 | u_int armval; |
| | 2499 | genimm_checked(imm,&armval); |
| | 2500 | assem_debug("bicne %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2501 | output_w32(0x13c00000|rd_rn_rm(rt,rs,0)|armval); |
| | 2502 | } |
| | 2503 | |
| | 2504 | void emit_biccs_imm(int rs,int imm,int rt) |
| | 2505 | { |
| | 2506 | u_int armval; |
| | 2507 | genimm_checked(imm,&armval); |
| | 2508 | assem_debug("biccs %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2509 | output_w32(0x23c00000|rd_rn_rm(rt,rs,0)|armval); |
| | 2510 | } |
| | 2511 | |
| | 2512 | void emit_bicvc_imm(int rs,int imm,int rt) |
| | 2513 | { |
| | 2514 | u_int armval; |
| | 2515 | genimm_checked(imm,&armval); |
| | 2516 | assem_debug("bicvc %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2517 | output_w32(0x73c00000|rd_rn_rm(rt,rs,0)|armval); |
| | 2518 | } |
| | 2519 | |
| | 2520 | void emit_bichi_imm(int rs,int imm,int rt) |
| | 2521 | { |
| | 2522 | u_int armval; |
| | 2523 | genimm_checked(imm,&armval); |
| | 2524 | assem_debug("bichi %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2525 | output_w32(0x83c00000|rd_rn_rm(rt,rs,0)|armval); |
| | 2526 | } |
| | 2527 | |
| | 2528 | void emit_orrvs_imm(int rs,int imm,int rt) |
| | 2529 | { |
| | 2530 | u_int armval; |
| | 2531 | genimm_checked(imm,&armval); |
| | 2532 | assem_debug("orrvs %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2533 | output_w32(0x63800000|rd_rn_rm(rt,rs,0)|armval); |
| | 2534 | } |
| | 2535 | |
| | 2536 | void emit_orrne_imm(int rs,int imm,int rt) |
| | 2537 | { |
| | 2538 | u_int armval; |
| | 2539 | genimm_checked(imm,&armval); |
| | 2540 | assem_debug("orrne %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2541 | output_w32(0x13800000|rd_rn_rm(rt,rs,0)|armval); |
| | 2542 | } |
| | 2543 | |
| | 2544 | void emit_andne_imm(int rs,int imm,int rt) |
| | 2545 | { |
| | 2546 | u_int armval; |
| | 2547 | genimm_checked(imm,&armval); |
| | 2548 | assem_debug("andne %s,%s,#%d\n",regname[rt],regname[rs],imm); |
| | 2549 | output_w32(0x12000000|rd_rn_rm(rt,rs,0)|armval); |
| | 2550 | } |
| | 2551 | |
| | 2552 | void emit_jno_unlikely(int a) |
| | 2553 | { |
| | 2554 | //emit_jno(a); |
| | 2555 | assem_debug("addvc pc,pc,#? (%x)\n",/*a-(int)out-8,*/a); |
| | 2556 | output_w32(0x72800000|rd_rn_rm(15,15,0)); |
| | 2557 | } |
| | 2558 | |
| | 2559 | // Save registers before function call |
| | 2560 | void save_regs(u_int reglist) |
| | 2561 | { |
| | 2562 | reglist&=0x100f; // only save the caller-save registers, r0-r3, r12 |
| | 2563 | if(!reglist) return; |
| | 2564 | assem_debug("stmia fp,{"); |
| | 2565 | if(reglist&1) assem_debug("r0, "); |
| | 2566 | if(reglist&2) assem_debug("r1, "); |
| | 2567 | if(reglist&4) assem_debug("r2, "); |
| | 2568 | if(reglist&8) assem_debug("r3, "); |
| | 2569 | if(reglist&0x1000) assem_debug("r12"); |
| | 2570 | assem_debug("}\n"); |
| | 2571 | output_w32(0xe88b0000|reglist); |
| | 2572 | } |
| | 2573 | // Restore registers after function call |
| | 2574 | void restore_regs(u_int reglist) |
| | 2575 | { |
| | 2576 | reglist&=0x100f; // only restore the caller-save registers, r0-r3, r12 |
| | 2577 | if(!reglist) return; |
| | 2578 | assem_debug("ldmia fp,{"); |
| | 2579 | if(reglist&1) assem_debug("r0, "); |
| | 2580 | if(reglist&2) assem_debug("r1, "); |
| | 2581 | if(reglist&4) assem_debug("r2, "); |
| | 2582 | if(reglist&8) assem_debug("r3, "); |
| | 2583 | if(reglist&0x1000) assem_debug("r12"); |
| | 2584 | assem_debug("}\n"); |
| | 2585 | output_w32(0xe89b0000|reglist); |
| | 2586 | } |
| | 2587 | |
| | 2588 | // Write back consts using r14 so we don't disturb the other registers |
| | 2589 | void wb_consts(signed char i_regmap[],uint64_t i_is32,u_int i_dirty,int i) |
| | 2590 | { |
| | 2591 | int hr; |
| | 2592 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 2593 | if(hr!=EXCLUDE_REG&&i_regmap[hr]>=0&&((i_dirty>>hr)&1)) { |
| | 2594 | if(((regs[i].isconst>>hr)&1)&&i_regmap[hr]>0) { |
| | 2595 | if(i_regmap[hr]<64 || !((i_is32>>(i_regmap[hr]&63))&1) ) { |
| | 2596 | int value=constmap[i][hr]; |
| | 2597 | if(value==0) { |
| | 2598 | emit_zeroreg(HOST_TEMPREG); |
| | 2599 | } |
| | 2600 | else { |
| | 2601 | emit_movimm(value,HOST_TEMPREG); |
| | 2602 | } |
| | 2603 | emit_storereg(i_regmap[hr],HOST_TEMPREG); |
| | 2604 | #ifndef FORCE32 |
| | 2605 | if((i_is32>>i_regmap[hr])&1) { |
| | 2606 | if(value!=-1&&value!=0) emit_sarimm(HOST_TEMPREG,31,HOST_TEMPREG); |
| | 2607 | emit_storereg(i_regmap[hr]|64,HOST_TEMPREG); |
| | 2608 | } |
| | 2609 | #endif |
| | 2610 | } |
| | 2611 | } |
| | 2612 | } |
| | 2613 | } |
| | 2614 | } |
| | 2615 | |
| | 2616 | /* Stubs/epilogue */ |
| | 2617 | |
| | 2618 | void literal_pool(int n) |
| | 2619 | { |
| | 2620 | if(!literalcount) return; |
| | 2621 | if(n) { |
| | 2622 | if((int)out-literals[0][0]<4096-n) return; |
| | 2623 | } |
| | 2624 | u_int *ptr; |
| | 2625 | int i; |
| | 2626 | for(i=0;i<literalcount;i++) |
| | 2627 | { |
| | 2628 | ptr=(u_int *)literals[i][0]; |
| | 2629 | u_int offset=(u_int)out-(u_int)ptr-8; |
| | 2630 | assert(offset<4096); |
| | 2631 | assert(!(offset&3)); |
| | 2632 | *ptr|=offset; |
| | 2633 | output_w32(literals[i][1]); |
| | 2634 | } |
| | 2635 | literalcount=0; |
| | 2636 | } |
| | 2637 | |
| | 2638 | void literal_pool_jumpover(int n) |
| | 2639 | { |
| | 2640 | if(!literalcount) return; |
| | 2641 | if(n) { |
| | 2642 | if((int)out-literals[0][0]<4096-n) return; |
| | 2643 | } |
| | 2644 | int jaddr=(int)out; |
| | 2645 | emit_jmp(0); |
| | 2646 | literal_pool(0); |
| | 2647 | set_jump_target(jaddr,(int)out); |
| | 2648 | } |
| | 2649 | |
| | 2650 | emit_extjump2(int addr, int target, int linker) |
| | 2651 | { |
| | 2652 | u_char *ptr=(u_char *)addr; |
| | 2653 | assert((ptr[3]&0x0e)==0xa); |
| | 2654 | emit_loadlp(target,0); |
| | 2655 | emit_loadlp(addr,1); |
| | 2656 | assert(addr>=BASE_ADDR&&addr<(BASE_ADDR+(1<<TARGET_SIZE_2))); |
| | 2657 | //assert((target>=0x80000000&&target<0x80800000)||(target>0xA4000000&&target<0xA4001000)); |
| | 2658 | //DEBUG > |
| | 2659 | #ifdef DEBUG_CYCLE_COUNT |
| | 2660 | emit_readword((int)&last_count,ECX); |
| | 2661 | emit_add(HOST_CCREG,ECX,HOST_CCREG); |
| | 2662 | emit_readword((int)&next_interupt,ECX); |
| | 2663 | emit_writeword(HOST_CCREG,(int)&Count); |
| | 2664 | emit_sub(HOST_CCREG,ECX,HOST_CCREG); |
| | 2665 | emit_writeword(ECX,(int)&last_count); |
| | 2666 | #endif |
| | 2667 | //DEBUG < |
| | 2668 | emit_jmp(linker); |
| | 2669 | } |
| | 2670 | |
| | 2671 | emit_extjump(int addr, int target) |
| | 2672 | { |
| | 2673 | emit_extjump2(addr, target, (int)dyna_linker); |
| | 2674 | } |
| | 2675 | emit_extjump_ds(int addr, int target) |
| | 2676 | { |
| | 2677 | emit_extjump2(addr, target, (int)dyna_linker_ds); |
| | 2678 | } |
| | 2679 | |
| | 2680 | // put rt_val into rt, potentially making use of rs with value rs_val |
| | 2681 | static void emit_movimm_from(u_int rs_val,int rs,u_int rt_val,int rt) |
| | 2682 | { |
| | 2683 | u_int xor=rs_val^rt_val; |
| | 2684 | u_int xs; |
| | 2685 | for(xs=xor;xs!=0&&(xs&3)==0;xs>>=2) |
| | 2686 | ; |
| | 2687 | if(xs<0x100) |
| | 2688 | emit_xorimm(rs,xor,rt); |
| | 2689 | else |
| | 2690 | emit_movimm(rt_val,rt); |
| | 2691 | } |
| | 2692 | |
| | 2693 | // trashes r2 |
| | 2694 | static void pass_args(int a0, int a1) |
| | 2695 | { |
| | 2696 | if(a0==1&&a1==0) { |
| | 2697 | // must swap |
| | 2698 | emit_mov(a0,2); emit_mov(a1,1); emit_mov(2,0); |
| | 2699 | } |
| | 2700 | else if(a0!=0&&a1==0) { |
| | 2701 | emit_mov(a1,1); |
| | 2702 | if (a0>=0) emit_mov(a0,0); |
| | 2703 | } |
| | 2704 | else { |
| | 2705 | if(a0>=0&&a0!=0) emit_mov(a0,0); |
| | 2706 | if(a1>=0&&a1!=1) emit_mov(a1,1); |
| | 2707 | } |
| | 2708 | } |
| | 2709 | |
| | 2710 | do_readstub(int n) |
| | 2711 | { |
| | 2712 | assem_debug("do_readstub %x\n",start+stubs[n][3]*4); |
| | 2713 | literal_pool(256); |
| | 2714 | set_jump_target(stubs[n][1],(int)out); |
| | 2715 | int type=stubs[n][0]; |
| | 2716 | int i=stubs[n][3]; |
| | 2717 | int rs=stubs[n][4]; |
| | 2718 | struct regstat *i_regs=(struct regstat *)stubs[n][5]; |
| | 2719 | u_int reglist=stubs[n][7]; |
| | 2720 | signed char *i_regmap=i_regs->regmap; |
| | 2721 | int addr=get_reg(i_regmap,AGEN1+(i&1)); |
| | 2722 | int rth,rt; |
| | 2723 | int ds; |
| | 2724 | if(itype[i]==C1LS||itype[i]==C2LS||itype[i]==LOADLR) { |
| | 2725 | rth=get_reg(i_regmap,FTEMP|64); |
| | 2726 | rt=get_reg(i_regmap,FTEMP); |
| | 2727 | }else{ |
| | 2728 | rth=get_reg(i_regmap,rt1[i]|64); |
| | 2729 | rt=get_reg(i_regmap,rt1[i]); |
| | 2730 | } |
| | 2731 | assert(rs>=0); |
| | 2732 | #ifdef PCSX |
| | 2733 | int r,temp=-1,temp2=HOST_TEMPREG,regs_saved=0,restore_jump=0; |
| | 2734 | reglist|=(1<<rs); |
| | 2735 | for(r=0;r<=12;r++) { |
| | 2736 | if(((1<<r)&0x13ff)&&((1<<r)®list)==0) { |
| | 2737 | temp=r; break; |
| | 2738 | } |
| | 2739 | } |
| | 2740 | if(rt>=0) |
| | 2741 | reglist&=~(1<<rt); |
| | 2742 | if(temp==-1) { |
| | 2743 | save_regs(reglist); |
| | 2744 | regs_saved=1; |
| | 2745 | temp=(rs==0)?2:0; |
| | 2746 | } |
| | 2747 | if((regs_saved||(reglist&2)==0)&&temp!=1&&rs!=1) |
| | 2748 | temp2=1; |
| | 2749 | emit_readword((int)&mem_rtab,temp); |
| | 2750 | emit_shrimm(rs,12,temp2); |
| | 2751 | emit_readword_dualindexedx4(temp,temp2,temp2); |
| | 2752 | emit_lsls_imm(temp2,1,temp2); |
| | 2753 | if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) { |
| | 2754 | switch(type) { |
| | 2755 | case LOADB_STUB: emit_ldrccsb_dualindexed(temp2,rs,rt); break; |
| | 2756 | case LOADBU_STUB: emit_ldrccb_dualindexed(temp2,rs,rt); break; |
| | 2757 | case LOADH_STUB: emit_ldrccsh_dualindexed(temp2,rs,rt); break; |
| | 2758 | case LOADHU_STUB: emit_ldrcch_dualindexed(temp2,rs,rt); break; |
| | 2759 | case LOADW_STUB: emit_ldrcc_dualindexed(temp2,rs,rt); break; |
| | 2760 | } |
| | 2761 | } |
| | 2762 | if(regs_saved) { |
| | 2763 | restore_jump=(int)out; |
| | 2764 | emit_jcc(0); // jump to reg restore |
| | 2765 | } |
| | 2766 | else |
| | 2767 | emit_jcc(stubs[n][2]); // return address |
| | 2768 | |
| | 2769 | if(!regs_saved) |
| | 2770 | save_regs(reglist); |
| | 2771 | int handler=0; |
| | 2772 | if(type==LOADB_STUB||type==LOADBU_STUB) |
| | 2773 | handler=(int)jump_handler_read8; |
| | 2774 | if(type==LOADH_STUB||type==LOADHU_STUB) |
| | 2775 | handler=(int)jump_handler_read16; |
| | 2776 | if(type==LOADW_STUB) |
| | 2777 | handler=(int)jump_handler_read32; |
| | 2778 | assert(handler!=0); |
| | 2779 | pass_args(rs,temp2); |
| | 2780 | int cc=get_reg(i_regmap,CCREG); |
| | 2781 | if(cc<0) |
| | 2782 | emit_loadreg(CCREG,2); |
| | 2783 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*stubs[n][6]+2,2); |
| | 2784 | emit_call(handler); |
| | 2785 | if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) { |
| | 2786 | switch(type) { |
| | 2787 | case LOADB_STUB: emit_signextend8(0,rt); break; |
| | 2788 | case LOADBU_STUB: emit_andimm(0,0xff,rt); break; |
| | 2789 | case LOADH_STUB: emit_signextend16(0,rt); break; |
| | 2790 | case LOADHU_STUB: emit_andimm(0,0xffff,rt); break; |
| | 2791 | case LOADW_STUB: if(rt!=0) emit_mov(0,rt); break; |
| | 2792 | } |
| | 2793 | } |
| | 2794 | if(restore_jump) |
| | 2795 | set_jump_target(restore_jump,(int)out); |
| | 2796 | restore_regs(reglist); |
| | 2797 | emit_jmp(stubs[n][2]); // return address |
| | 2798 | #else // !PCSX |
| | 2799 | if(addr<0) addr=rt; |
| | 2800 | if(addr<0&&itype[i]!=C1LS&&itype[i]!=C2LS&&itype[i]!=LOADLR) addr=get_reg(i_regmap,-1); |
| | 2801 | assert(addr>=0); |
| | 2802 | int ftable=0; |
| | 2803 | if(type==LOADB_STUB||type==LOADBU_STUB) |
| | 2804 | ftable=(int)readmemb; |
| | 2805 | if(type==LOADH_STUB||type==LOADHU_STUB) |
| | 2806 | ftable=(int)readmemh; |
| | 2807 | if(type==LOADW_STUB) |
| | 2808 | ftable=(int)readmem; |
| | 2809 | #ifndef FORCE32 |
| | 2810 | if(type==LOADD_STUB) |
| | 2811 | ftable=(int)readmemd; |
| | 2812 | #endif |
| | 2813 | assert(ftable!=0); |
| | 2814 | emit_writeword(rs,(int)&address); |
| | 2815 | //emit_pusha(); |
| | 2816 | save_regs(reglist); |
| | 2817 | #ifndef PCSX |
| | 2818 | ds=i_regs!=®s[i]; |
| | 2819 | int real_rs=(itype[i]==LOADLR)?-1:get_reg(i_regmap,rs1[i]); |
| | 2820 | u_int cmask=ds?-1:(0x100f|~i_regs->wasconst); |
| | 2821 | if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i); |
| | 2822 | wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))); |
| | 2823 | if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i); |
| | 2824 | #endif |
| | 2825 | emit_shrimm(rs,16,1); |
| | 2826 | int cc=get_reg(i_regmap,CCREG); |
| | 2827 | if(cc<0) { |
| | 2828 | emit_loadreg(CCREG,2); |
| | 2829 | } |
| | 2830 | emit_movimm(ftable,0); |
| | 2831 | emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2); |
| | 2832 | #ifndef PCSX |
| | 2833 | emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3); |
| | 2834 | #endif |
| | 2835 | //emit_readword((int)&last_count,temp); |
| | 2836 | //emit_add(cc,temp,cc); |
| | 2837 | //emit_writeword(cc,(int)&Count); |
| | 2838 | //emit_mov(15,14); |
| | 2839 | emit_call((int)&indirect_jump_indexed); |
| | 2840 | //emit_callreg(rs); |
| | 2841 | //emit_readword_dualindexedx4(rs,HOST_TEMPREG,15); |
| | 2842 | #ifndef PCSX |
| | 2843 | // We really shouldn't need to update the count here, |
| | 2844 | // but not doing so causes random crashes... |
| | 2845 | emit_readword((int)&Count,HOST_TEMPREG); |
| | 2846 | emit_readword((int)&next_interupt,2); |
| | 2847 | emit_addimm(HOST_TEMPREG,-2*stubs[n][6]-2,HOST_TEMPREG); |
| | 2848 | emit_writeword(2,(int)&last_count); |
| | 2849 | emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc); |
| | 2850 | if(cc<0) { |
| | 2851 | emit_storereg(CCREG,HOST_TEMPREG); |
| | 2852 | } |
| | 2853 | #endif |
| | 2854 | //emit_popa(); |
| | 2855 | restore_regs(reglist); |
| | 2856 | //if((cc=get_reg(regmap,CCREG))>=0) { |
| | 2857 | // emit_loadreg(CCREG,cc); |
| | 2858 | //} |
| | 2859 | if(itype[i]==C1LS||itype[i]==C2LS||(rt>=0&&rt1[i]!=0)) { |
| | 2860 | assert(rt>=0); |
| | 2861 | if(type==LOADB_STUB) |
| | 2862 | emit_movsbl((int)&readmem_dword,rt); |
| | 2863 | if(type==LOADBU_STUB) |
| | 2864 | emit_movzbl((int)&readmem_dword,rt); |
| | 2865 | if(type==LOADH_STUB) |
| | 2866 | emit_movswl((int)&readmem_dword,rt); |
| | 2867 | if(type==LOADHU_STUB) |
| | 2868 | emit_movzwl((int)&readmem_dword,rt); |
| | 2869 | if(type==LOADW_STUB) |
| | 2870 | emit_readword((int)&readmem_dword,rt); |
| | 2871 | if(type==LOADD_STUB) { |
| | 2872 | emit_readword((int)&readmem_dword,rt); |
| | 2873 | if(rth>=0) emit_readword(((int)&readmem_dword)+4,rth); |
| | 2874 | } |
| | 2875 | } |
| | 2876 | emit_jmp(stubs[n][2]); // return address |
| | 2877 | #endif // !PCSX |
| | 2878 | } |
| | 2879 | |
| | 2880 | #ifdef PCSX |
| | 2881 | // return memhandler, or get directly accessable address and return 0 |
| | 2882 | u_int get_direct_memhandler(void *table,u_int addr,int type,u_int *addr_host) |
| | 2883 | { |
| | 2884 | u_int l1,l2=0; |
| | 2885 | l1=((u_int *)table)[addr>>12]; |
| | 2886 | if((l1&(1<<31))==0) { |
| | 2887 | u_int v=l1<<1; |
| | 2888 | *addr_host=v+addr; |
| | 2889 | return 0; |
| | 2890 | } |
| | 2891 | else { |
| | 2892 | l1<<=1; |
| | 2893 | if(type==LOADB_STUB||type==LOADBU_STUB||type==STOREB_STUB) |
| | 2894 | l2=((u_int *)l1)[0x1000/4 + 0x1000/2 + (addr&0xfff)]; |
| | 2895 | else if(type==LOADH_STUB||type==LOADHU_STUB||type==STOREH_STUB) |
| | 2896 | l2=((u_int *)l1)[0x1000/4 + (addr&0xfff)/2]; |
| | 2897 | else |
| | 2898 | l2=((u_int *)l1)[(addr&0xfff)/4]; |
| | 2899 | if((l2&(1<<31))==0) { |
| | 2900 | u_int v=l2<<1; |
| | 2901 | *addr_host=v+(addr&0xfff); |
| | 2902 | return 0; |
| | 2903 | } |
| | 2904 | return l2<<1; |
| | 2905 | } |
| | 2906 | } |
| | 2907 | #endif |
| | 2908 | |
| | 2909 | inline_readstub(int type, int i, u_int addr, signed char regmap[], int target, int adj, u_int reglist) |
| | 2910 | { |
| | 2911 | int rs=get_reg(regmap,target); |
| | 2912 | int rth=get_reg(regmap,target|64); |
| | 2913 | int rt=get_reg(regmap,target); |
| | 2914 | if(rs<0) rs=get_reg(regmap,-1); |
| | 2915 | assert(rs>=0); |
| | 2916 | #ifdef PCSX |
| | 2917 | u_int handler,host_addr=0; |
| | 2918 | handler=get_direct_memhandler(mem_rtab,addr,type,&host_addr); |
| | 2919 | if (handler==0) { |
| | 2920 | if(rt<0) |
| | 2921 | return; |
| | 2922 | if(addr!=host_addr) |
| | 2923 | emit_movimm_from(addr,rs,host_addr,rs); |
| | 2924 | switch(type) { |
| | 2925 | case LOADB_STUB: emit_movsbl_indexed(0,rs,rt); break; |
| | 2926 | case LOADBU_STUB: emit_movzbl_indexed(0,rs,rt); break; |
| | 2927 | case LOADH_STUB: emit_movswl_indexed(0,rs,rt); break; |
| | 2928 | case LOADHU_STUB: emit_movzwl_indexed(0,rs,rt); break; |
| | 2929 | case LOADW_STUB: emit_readword_indexed(0,rs,rt); break; |
| | 2930 | default: assert(0); |
| | 2931 | } |
| | 2932 | return; |
| | 2933 | } |
| | 2934 | |
| | 2935 | // call a memhandler |
| | 2936 | if(rt>=0) |
| | 2937 | reglist&=~(1<<rt); |
| | 2938 | save_regs(reglist); |
| | 2939 | if(target==0) |
| | 2940 | emit_movimm(addr,0); |
| | 2941 | else if(rs!=0) |
| | 2942 | emit_mov(rs,0); |
| | 2943 | int cc=get_reg(regmap,CCREG); |
| | 2944 | if(cc<0) |
| | 2945 | emit_loadreg(CCREG,2); |
| | 2946 | emit_readword((int)&last_count,3); |
| | 2947 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2); |
| | 2948 | emit_add(2,3,3); |
| | 2949 | emit_writeword(3,(int)&Count); |
| | 2950 | |
| | 2951 | int offset=(int)handler-(int)out-8; |
| | 2952 | if(offset<-33554432||offset>=33554432) { |
| | 2953 | // unreachable memhandler, a plugin func perhaps |
| | 2954 | emit_movimm(handler,1); |
| | 2955 | emit_callreg(1); |
| | 2956 | } |
| | 2957 | else |
| | 2958 | emit_call(handler); |
| | 2959 | if(rt>=0) { |
| | 2960 | switch(type) { |
| | 2961 | case LOADB_STUB: emit_signextend8(0,rt); break; |
| | 2962 | case LOADBU_STUB: emit_andimm(0,0xff,rt); break; |
| | 2963 | case LOADH_STUB: emit_signextend16(0,rt); break; |
| | 2964 | case LOADHU_STUB: emit_andimm(0,0xffff,rt); break; |
| | 2965 | case LOADW_STUB: if(rt!=0) emit_mov(0,rt); break; |
| | 2966 | default: assert(0); |
| | 2967 | } |
| | 2968 | } |
| | 2969 | restore_regs(reglist); |
| | 2970 | #else // if !PCSX |
| | 2971 | int ftable=0; |
| | 2972 | if(type==LOADB_STUB||type==LOADBU_STUB) |
| | 2973 | ftable=(int)readmemb; |
| | 2974 | if(type==LOADH_STUB||type==LOADHU_STUB) |
| | 2975 | ftable=(int)readmemh; |
| | 2976 | if(type==LOADW_STUB) |
| | 2977 | ftable=(int)readmem; |
| | 2978 | #ifndef FORCE32 |
| | 2979 | if(type==LOADD_STUB) |
| | 2980 | ftable=(int)readmemd; |
| | 2981 | #endif |
| | 2982 | assert(ftable!=0); |
| | 2983 | if(target==0) |
| | 2984 | emit_movimm(addr,rs); |
| | 2985 | emit_writeword(rs,(int)&address); |
| | 2986 | //emit_pusha(); |
| | 2987 | save_regs(reglist); |
| | 2988 | #ifndef PCSX |
| | 2989 | if((signed int)addr>=(signed int)0xC0000000) { |
| | 2990 | // Theoretically we can have a pagefault here, if the TLB has never |
| | 2991 | // been enabled and the address is outside the range 80000000..BFFFFFFF |
| | 2992 | // Write out the registers so the pagefault can be handled. This is |
| | 2993 | // a very rare case and likely represents a bug. |
| | 2994 | int ds=regmap!=regs[i].regmap; |
| | 2995 | if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i); |
| | 2996 | if(!ds) wb_dirtys(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty); |
| | 2997 | else wb_dirtys(branch_regs[i-1].regmap_entry,branch_regs[i-1].was32,branch_regs[i-1].wasdirty); |
| | 2998 | } |
| | 2999 | #endif |
| | 3000 | //emit_shrimm(rs,16,1); |
| | 3001 | int cc=get_reg(regmap,CCREG); |
| | 3002 | if(cc<0) { |
| | 3003 | emit_loadreg(CCREG,2); |
| | 3004 | } |
| | 3005 | //emit_movimm(ftable,0); |
| | 3006 | emit_movimm(((u_int *)ftable)[addr>>16],0); |
| | 3007 | //emit_readword((int)&last_count,12); |
| | 3008 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2); |
| | 3009 | #ifndef PCSX |
| | 3010 | if((signed int)addr>=(signed int)0xC0000000) { |
| | 3011 | // Pagefault address |
| | 3012 | int ds=regmap!=regs[i].regmap; |
| | 3013 | emit_movimm(start+i*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3); |
| | 3014 | } |
| | 3015 | #endif |
| | 3016 | //emit_add(12,2,2); |
| | 3017 | //emit_writeword(2,(int)&Count); |
| | 3018 | //emit_call(((u_int *)ftable)[addr>>16]); |
| | 3019 | emit_call((int)&indirect_jump); |
| | 3020 | #ifndef PCSX |
| | 3021 | // We really shouldn't need to update the count here, |
| | 3022 | // but not doing so causes random crashes... |
| | 3023 | emit_readword((int)&Count,HOST_TEMPREG); |
| | 3024 | emit_readword((int)&next_interupt,2); |
| | 3025 | emit_addimm(HOST_TEMPREG,-CLOCK_DIVIDER*(adj+1),HOST_TEMPREG); |
| | 3026 | emit_writeword(2,(int)&last_count); |
| | 3027 | emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc); |
| | 3028 | if(cc<0) { |
| | 3029 | emit_storereg(CCREG,HOST_TEMPREG); |
| | 3030 | } |
| | 3031 | #endif |
| | 3032 | //emit_popa(); |
| | 3033 | restore_regs(reglist); |
| | 3034 | if(rt>=0) { |
| | 3035 | if(type==LOADB_STUB) |
| | 3036 | emit_movsbl((int)&readmem_dword,rt); |
| | 3037 | if(type==LOADBU_STUB) |
| | 3038 | emit_movzbl((int)&readmem_dword,rt); |
| | 3039 | if(type==LOADH_STUB) |
| | 3040 | emit_movswl((int)&readmem_dword,rt); |
| | 3041 | if(type==LOADHU_STUB) |
| | 3042 | emit_movzwl((int)&readmem_dword,rt); |
| | 3043 | if(type==LOADW_STUB) |
| | 3044 | emit_readword((int)&readmem_dword,rt); |
| | 3045 | if(type==LOADD_STUB) { |
| | 3046 | emit_readword((int)&readmem_dword,rt); |
| | 3047 | if(rth>=0) emit_readword(((int)&readmem_dword)+4,rth); |
| | 3048 | } |
| | 3049 | } |
| | 3050 | #endif // !PCSX |
| | 3051 | } |
| | 3052 | |
| | 3053 | do_writestub(int n) |
| | 3054 | { |
| | 3055 | assem_debug("do_writestub %x\n",start+stubs[n][3]*4); |
| | 3056 | literal_pool(256); |
| | 3057 | set_jump_target(stubs[n][1],(int)out); |
| | 3058 | int type=stubs[n][0]; |
| | 3059 | int i=stubs[n][3]; |
| | 3060 | int rs=stubs[n][4]; |
| | 3061 | struct regstat *i_regs=(struct regstat *)stubs[n][5]; |
| | 3062 | u_int reglist=stubs[n][7]; |
| | 3063 | signed char *i_regmap=i_regs->regmap; |
| | 3064 | int addr=get_reg(i_regmap,AGEN1+(i&1)); |
| | 3065 | int rth,rt,r; |
| | 3066 | int ds; |
| | 3067 | if(itype[i]==C1LS||itype[i]==C2LS) { |
| | 3068 | rth=get_reg(i_regmap,FTEMP|64); |
| | 3069 | rt=get_reg(i_regmap,r=FTEMP); |
| | 3070 | }else{ |
| | 3071 | rth=get_reg(i_regmap,rs2[i]|64); |
| | 3072 | rt=get_reg(i_regmap,r=rs2[i]); |
| | 3073 | } |
| | 3074 | assert(rs>=0); |
| | 3075 | assert(rt>=0); |
| | 3076 | #ifdef PCSX |
| | 3077 | int rtmp,temp=-1,temp2=HOST_TEMPREG,regs_saved=0,restore_jump=0,ra; |
| | 3078 | int reglist2=reglist|(1<<rs)|(1<<rt); |
| | 3079 | for(rtmp=0;rtmp<=12;rtmp++) { |
| | 3080 | if(((1<<rtmp)&0x13ff)&&((1<<rtmp)®list2)==0) { |
| | 3081 | temp=rtmp; break; |
| | 3082 | } |
| | 3083 | } |
| | 3084 | if(temp==-1) { |
| | 3085 | save_regs(reglist); |
| | 3086 | regs_saved=1; |
| | 3087 | for(rtmp=0;rtmp<=3;rtmp++) |
| | 3088 | if(rtmp!=rs&&rtmp!=rt) |
| | 3089 | {temp=rtmp;break;} |
| | 3090 | } |
| | 3091 | if((regs_saved||(reglist2&8)==0)&&temp!=3&&rs!=3&&rt!=3) |
| | 3092 | temp2=3; |
| | 3093 | emit_readword((int)&mem_wtab,temp); |
| | 3094 | emit_shrimm(rs,12,temp2); |
| | 3095 | emit_readword_dualindexedx4(temp,temp2,temp2); |
| | 3096 | emit_lsls_imm(temp2,1,temp2); |
| | 3097 | switch(type) { |
| | 3098 | case STOREB_STUB: emit_strccb_dualindexed(temp2,rs,rt); break; |
| | 3099 | case STOREH_STUB: emit_strcch_dualindexed(temp2,rs,rt); break; |
| | 3100 | case STOREW_STUB: emit_strcc_dualindexed(temp2,rs,rt); break; |
| | 3101 | default: assert(0); |
| | 3102 | } |
| | 3103 | if(regs_saved) { |
| | 3104 | restore_jump=(int)out; |
| | 3105 | emit_jcc(0); // jump to reg restore |
| | 3106 | } |
| | 3107 | else |
| | 3108 | emit_jcc(stubs[n][2]); // return address (invcode check) |
| | 3109 | |
| | 3110 | if(!regs_saved) |
| | 3111 | save_regs(reglist); |
| | 3112 | int handler=0; |
| | 3113 | switch(type) { |
| | 3114 | case STOREB_STUB: handler=(int)jump_handler_write8; break; |
| | 3115 | case STOREH_STUB: handler=(int)jump_handler_write16; break; |
| | 3116 | case STOREW_STUB: handler=(int)jump_handler_write32; break; |
| | 3117 | } |
| | 3118 | assert(handler!=0); |
| | 3119 | pass_args(rs,rt); |
| | 3120 | if(temp2!=3) |
| | 3121 | emit_mov(temp2,3); |
| | 3122 | int cc=get_reg(i_regmap,CCREG); |
| | 3123 | if(cc<0) |
| | 3124 | emit_loadreg(CCREG,2); |
| | 3125 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*stubs[n][6]+2,2); |
| | 3126 | // returns new cycle_count |
| | 3127 | emit_call(handler); |
| | 3128 | emit_addimm(0,-CLOCK_DIVIDER*stubs[n][6]-2,cc<0?2:cc); |
| | 3129 | if(cc<0) |
| | 3130 | emit_storereg(CCREG,2); |
| | 3131 | if(restore_jump) |
| | 3132 | set_jump_target(restore_jump,(int)out); |
| | 3133 | restore_regs(reglist); |
| | 3134 | ra=stubs[n][2]; |
| | 3135 | if(!restore_jump) ra+=4*3; // skip invcode check |
| | 3136 | emit_jmp(ra); |
| | 3137 | #else // if !PCSX |
| | 3138 | if(addr<0) addr=get_reg(i_regmap,-1); |
| | 3139 | assert(addr>=0); |
| | 3140 | int ftable=0; |
| | 3141 | if(type==STOREB_STUB) |
| | 3142 | ftable=(int)writememb; |
| | 3143 | if(type==STOREH_STUB) |
| | 3144 | ftable=(int)writememh; |
| | 3145 | if(type==STOREW_STUB) |
| | 3146 | ftable=(int)writemem; |
| | 3147 | #ifndef FORCE32 |
| | 3148 | if(type==STORED_STUB) |
| | 3149 | ftable=(int)writememd; |
| | 3150 | #endif |
| | 3151 | assert(ftable!=0); |
| | 3152 | emit_writeword(rs,(int)&address); |
| | 3153 | //emit_shrimm(rs,16,rs); |
| | 3154 | //emit_movmem_indexedx4(ftable,rs,rs); |
| | 3155 | if(type==STOREB_STUB) |
| | 3156 | emit_writebyte(rt,(int)&byte); |
| | 3157 | if(type==STOREH_STUB) |
| | 3158 | emit_writehword(rt,(int)&hword); |
| | 3159 | if(type==STOREW_STUB) |
| | 3160 | emit_writeword(rt,(int)&word); |
| | 3161 | if(type==STORED_STUB) { |
| | 3162 | #ifndef FORCE32 |
| | 3163 | emit_writeword(rt,(int)&dword); |
| | 3164 | emit_writeword(r?rth:rt,(int)&dword+4); |
| | 3165 | #else |
| | 3166 | printf("STORED_STUB\n"); |
| | 3167 | #endif |
| | 3168 | } |
| | 3169 | //emit_pusha(); |
| | 3170 | save_regs(reglist); |
| | 3171 | #ifndef PCSX |
| | 3172 | ds=i_regs!=®s[i]; |
| | 3173 | int real_rs=get_reg(i_regmap,rs1[i]); |
| | 3174 | u_int cmask=ds?-1:(0x100f|~i_regs->wasconst); |
| | 3175 | if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i); |
| | 3176 | wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))); |
| | 3177 | if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i); |
| | 3178 | #endif |
| | 3179 | emit_shrimm(rs,16,1); |
| | 3180 | int cc=get_reg(i_regmap,CCREG); |
| | 3181 | if(cc<0) { |
| | 3182 | emit_loadreg(CCREG,2); |
| | 3183 | } |
| | 3184 | emit_movimm(ftable,0); |
| | 3185 | emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2); |
| | 3186 | #ifndef PCSX |
| | 3187 | emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3); |
| | 3188 | #endif |
| | 3189 | //emit_readword((int)&last_count,temp); |
| | 3190 | //emit_addimm(cc,2*stubs[n][5]+2,cc); |
| | 3191 | //emit_add(cc,temp,cc); |
| | 3192 | //emit_writeword(cc,(int)&Count); |
| | 3193 | emit_call((int)&indirect_jump_indexed); |
| | 3194 | //emit_callreg(rs); |
| | 3195 | emit_readword((int)&Count,HOST_TEMPREG); |
| | 3196 | emit_readword((int)&next_interupt,2); |
| | 3197 | emit_addimm(HOST_TEMPREG,-2*stubs[n][6]-2,HOST_TEMPREG); |
| | 3198 | emit_writeword(2,(int)&last_count); |
| | 3199 | emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc); |
| | 3200 | if(cc<0) { |
| | 3201 | emit_storereg(CCREG,HOST_TEMPREG); |
| | 3202 | } |
| | 3203 | //emit_popa(); |
| | 3204 | restore_regs(reglist); |
| | 3205 | //if((cc=get_reg(regmap,CCREG))>=0) { |
| | 3206 | // emit_loadreg(CCREG,cc); |
| | 3207 | //} |
| | 3208 | emit_jmp(stubs[n][2]); // return address |
| | 3209 | #endif // !PCSX |
| | 3210 | } |
| | 3211 | |
| | 3212 | inline_writestub(int type, int i, u_int addr, signed char regmap[], int target, int adj, u_int reglist) |
| | 3213 | { |
| | 3214 | int rs=get_reg(regmap,-1); |
| | 3215 | int rth=get_reg(regmap,target|64); |
| | 3216 | int rt=get_reg(regmap,target); |
| | 3217 | assert(rs>=0); |
| | 3218 | assert(rt>=0); |
| | 3219 | #ifdef PCSX |
| | 3220 | u_int handler,host_addr=0; |
| | 3221 | handler=get_direct_memhandler(mem_wtab,addr,type,&host_addr); |
| | 3222 | if (handler==0) { |
| | 3223 | if(addr!=host_addr) |
| | 3224 | emit_movimm_from(addr,rs,host_addr,rs); |
| | 3225 | switch(type) { |
| | 3226 | case STOREB_STUB: emit_writebyte_indexed(rt,0,rs); break; |
| | 3227 | case STOREH_STUB: emit_writehword_indexed(rt,0,rs); break; |
| | 3228 | case STOREW_STUB: emit_writeword_indexed(rt,0,rs); break; |
| | 3229 | default: assert(0); |
| | 3230 | } |
| | 3231 | return; |
| | 3232 | } |
| | 3233 | |
| | 3234 | // call a memhandler |
| | 3235 | save_regs(reglist); |
| | 3236 | pass_args(rs,rt); |
| | 3237 | int cc=get_reg(regmap,CCREG); |
| | 3238 | if(cc<0) |
| | 3239 | emit_loadreg(CCREG,2); |
| | 3240 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2); |
| | 3241 | emit_movimm(handler,3); |
| | 3242 | // returns new cycle_count |
| | 3243 | emit_call((int)jump_handler_write_h); |
| | 3244 | emit_addimm(0,-CLOCK_DIVIDER*(adj+1),cc<0?2:cc); |
| | 3245 | if(cc<0) |
| | 3246 | emit_storereg(CCREG,2); |
| | 3247 | restore_regs(reglist); |
| | 3248 | #else // if !pcsx |
| | 3249 | int ftable=0; |
| | 3250 | if(type==STOREB_STUB) |
| | 3251 | ftable=(int)writememb; |
| | 3252 | if(type==STOREH_STUB) |
| | 3253 | ftable=(int)writememh; |
| | 3254 | if(type==STOREW_STUB) |
| | 3255 | ftable=(int)writemem; |
| | 3256 | #ifndef FORCE32 |
| | 3257 | if(type==STORED_STUB) |
| | 3258 | ftable=(int)writememd; |
| | 3259 | #endif |
| | 3260 | assert(ftable!=0); |
| | 3261 | emit_writeword(rs,(int)&address); |
| | 3262 | //emit_shrimm(rs,16,rs); |
| | 3263 | //emit_movmem_indexedx4(ftable,rs,rs); |
| | 3264 | if(type==STOREB_STUB) |
| | 3265 | emit_writebyte(rt,(int)&byte); |
| | 3266 | if(type==STOREH_STUB) |
| | 3267 | emit_writehword(rt,(int)&hword); |
| | 3268 | if(type==STOREW_STUB) |
| | 3269 | emit_writeword(rt,(int)&word); |
| | 3270 | if(type==STORED_STUB) { |
| | 3271 | #ifndef FORCE32 |
| | 3272 | emit_writeword(rt,(int)&dword); |
| | 3273 | emit_writeword(target?rth:rt,(int)&dword+4); |
| | 3274 | #else |
| | 3275 | printf("STORED_STUB\n"); |
| | 3276 | #endif |
| | 3277 | } |
| | 3278 | //emit_pusha(); |
| | 3279 | save_regs(reglist); |
| | 3280 | #ifndef PCSX |
| | 3281 | // rearmed note: load_all_consts prevents BIOS boot, some bug? |
| | 3282 | if((signed int)addr>=(signed int)0xC0000000) { |
| | 3283 | // Theoretically we can have a pagefault here, if the TLB has never |
| | 3284 | // been enabled and the address is outside the range 80000000..BFFFFFFF |
| | 3285 | // Write out the registers so the pagefault can be handled. This is |
| | 3286 | // a very rare case and likely represents a bug. |
| | 3287 | int ds=regmap!=regs[i].regmap; |
| | 3288 | if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i); |
| | 3289 | if(!ds) wb_dirtys(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty); |
| | 3290 | else wb_dirtys(branch_regs[i-1].regmap_entry,branch_regs[i-1].was32,branch_regs[i-1].wasdirty); |
| | 3291 | } |
| | 3292 | #endif |
| | 3293 | //emit_shrimm(rs,16,1); |
| | 3294 | int cc=get_reg(regmap,CCREG); |
| | 3295 | if(cc<0) { |
| | 3296 | emit_loadreg(CCREG,2); |
| | 3297 | } |
| | 3298 | //emit_movimm(ftable,0); |
| | 3299 | emit_movimm(((u_int *)ftable)[addr>>16],0); |
| | 3300 | //emit_readword((int)&last_count,12); |
| | 3301 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*(adj+1),2); |
| | 3302 | #ifndef PCSX |
| | 3303 | if((signed int)addr>=(signed int)0xC0000000) { |
| | 3304 | // Pagefault address |
| | 3305 | int ds=regmap!=regs[i].regmap; |
| | 3306 | emit_movimm(start+i*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3); |
| | 3307 | } |
| | 3308 | #endif |
| | 3309 | //emit_add(12,2,2); |
| | 3310 | //emit_writeword(2,(int)&Count); |
| | 3311 | //emit_call(((u_int *)ftable)[addr>>16]); |
| | 3312 | emit_call((int)&indirect_jump); |
| | 3313 | emit_readword((int)&Count,HOST_TEMPREG); |
| | 3314 | emit_readword((int)&next_interupt,2); |
| | 3315 | emit_addimm(HOST_TEMPREG,-CLOCK_DIVIDER*(adj+1),HOST_TEMPREG); |
| | 3316 | emit_writeword(2,(int)&last_count); |
| | 3317 | emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc); |
| | 3318 | if(cc<0) { |
| | 3319 | emit_storereg(CCREG,HOST_TEMPREG); |
| | 3320 | } |
| | 3321 | //emit_popa(); |
| | 3322 | restore_regs(reglist); |
| | 3323 | #endif |
| | 3324 | } |
| | 3325 | |
| | 3326 | do_unalignedwritestub(int n) |
| | 3327 | { |
| | 3328 | assem_debug("do_unalignedwritestub %x\n",start+stubs[n][3]*4); |
| | 3329 | literal_pool(256); |
| | 3330 | set_jump_target(stubs[n][1],(int)out); |
| | 3331 | |
| | 3332 | int i=stubs[n][3]; |
| | 3333 | struct regstat *i_regs=(struct regstat *)stubs[n][4]; |
| | 3334 | int addr=stubs[n][5]; |
| | 3335 | u_int reglist=stubs[n][7]; |
| | 3336 | signed char *i_regmap=i_regs->regmap; |
| | 3337 | int temp2=get_reg(i_regmap,FTEMP); |
| | 3338 | int rt; |
| | 3339 | int ds, real_rs; |
| | 3340 | rt=get_reg(i_regmap,rs2[i]); |
| | 3341 | assert(rt>=0); |
| | 3342 | assert(addr>=0); |
| | 3343 | assert(opcode[i]==0x2a||opcode[i]==0x2e); // SWL/SWR only implemented |
| | 3344 | reglist|=(1<<addr); |
| | 3345 | reglist&=~(1<<temp2); |
| | 3346 | |
| | 3347 | #if 1 |
| | 3348 | // don't bother with it and call write handler |
| | 3349 | save_regs(reglist); |
| | 3350 | pass_args(addr,rt); |
| | 3351 | int cc=get_reg(i_regmap,CCREG); |
| | 3352 | if(cc<0) |
| | 3353 | emit_loadreg(CCREG,2); |
| | 3354 | emit_addimm(cc<0?2:cc,CLOCK_DIVIDER*stubs[n][6]+2,2); |
| | 3355 | emit_call((int)(opcode[i]==0x2a?jump_handle_swl:jump_handle_swr)); |
| | 3356 | emit_addimm(0,-CLOCK_DIVIDER*stubs[n][6]-2,cc<0?2:cc); |
| | 3357 | if(cc<0) |
| | 3358 | emit_storereg(CCREG,2); |
| | 3359 | restore_regs(reglist); |
| | 3360 | emit_jmp(stubs[n][2]); // return address |
| | 3361 | #else |
| | 3362 | emit_andimm(addr,0xfffffffc,temp2); |
| | 3363 | emit_writeword(temp2,(int)&address); |
| | 3364 | |
| | 3365 | save_regs(reglist); |
| | 3366 | #ifndef PCSX |
| | 3367 | ds=i_regs!=®s[i]; |
| | 3368 | real_rs=get_reg(i_regmap,rs1[i]); |
| | 3369 | u_int cmask=ds?-1:(0x100f|~i_regs->wasconst); |
| | 3370 | if(!ds) load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&0x100f,i); |
| | 3371 | wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty&cmask&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))); |
| | 3372 | if(!ds) wb_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty&~(1<<addr)&(real_rs<0?-1:~(1<<real_rs))&~0x100f,i); |
| | 3373 | #endif |
| | 3374 | emit_shrimm(addr,16,1); |
| | 3375 | int cc=get_reg(i_regmap,CCREG); |
| | 3376 | if(cc<0) { |
| | 3377 | emit_loadreg(CCREG,2); |
| | 3378 | } |
| | 3379 | emit_movimm((u_int)readmem,0); |
| | 3380 | emit_addimm(cc<0?2:cc,2*stubs[n][6]+2,2); |
| | 3381 | #ifndef PCSX |
| | 3382 | // pagefault address |
| | 3383 | emit_movimm(start+stubs[n][3]*4+(((regs[i].was32>>rs1[i])&1)<<1)+ds,3); |
| | 3384 | #endif |
| | 3385 | emit_call((int)&indirect_jump_indexed); |
| | 3386 | restore_regs(reglist); |
| | 3387 | |
| | 3388 | emit_readword((int)&readmem_dword,temp2); |
| | 3389 | int temp=addr; //hmh |
| | 3390 | emit_shlimm(addr,3,temp); |
| | 3391 | emit_andimm(temp,24,temp); |
| | 3392 | #ifdef BIG_ENDIAN_MIPS |
| | 3393 | if (opcode[i]==0x2e) // SWR |
| | 3394 | #else |
| | 3395 | if (opcode[i]==0x2a) // SWL |
| | 3396 | #endif |
| | 3397 | emit_xorimm(temp,24,temp); |
| | 3398 | emit_movimm(-1,HOST_TEMPREG); |
| | 3399 | if (opcode[i]==0x2a) { // SWL |
| | 3400 | emit_bic_lsr(temp2,HOST_TEMPREG,temp,temp2); |
| | 3401 | emit_orrshr(rt,temp,temp2); |
| | 3402 | }else{ |
| | 3403 | emit_bic_lsl(temp2,HOST_TEMPREG,temp,temp2); |
| | 3404 | emit_orrshl(rt,temp,temp2); |
| | 3405 | } |
| | 3406 | emit_readword((int)&address,addr); |
| | 3407 | emit_writeword(temp2,(int)&word); |
| | 3408 | //save_regs(reglist); // don't need to, no state changes |
| | 3409 | emit_shrimm(addr,16,1); |
| | 3410 | emit_movimm((u_int)writemem,0); |
| | 3411 | //emit_call((int)&indirect_jump_indexed); |
| | 3412 | emit_mov(15,14); |
| | 3413 | emit_readword_dualindexedx4(0,1,15); |
| | 3414 | emit_readword((int)&Count,HOST_TEMPREG); |
| | 3415 | emit_readword((int)&next_interupt,2); |
| | 3416 | emit_addimm(HOST_TEMPREG,-2*stubs[n][6]-2,HOST_TEMPREG); |
| | 3417 | emit_writeword(2,(int)&last_count); |
| | 3418 | emit_sub(HOST_TEMPREG,2,cc<0?HOST_TEMPREG:cc); |
| | 3419 | if(cc<0) { |
| | 3420 | emit_storereg(CCREG,HOST_TEMPREG); |
| | 3421 | } |
| | 3422 | restore_regs(reglist); |
| | 3423 | emit_jmp(stubs[n][2]); // return address |
| | 3424 | #endif |
| | 3425 | } |
| | 3426 | |
| | 3427 | void printregs(int edi,int esi,int ebp,int esp,int b,int d,int c,int a) |
| | 3428 | { |
| | 3429 | printf("regs: %x %x %x %x %x %x %x (%x)\n",a,b,c,d,ebp,esi,edi,(&edi)[-1]); |
| | 3430 | } |
| | 3431 | |
| | 3432 | do_invstub(int n) |
| | 3433 | { |
| | 3434 | literal_pool(20); |
| | 3435 | u_int reglist=stubs[n][3]; |
| | 3436 | set_jump_target(stubs[n][1],(int)out); |
| | 3437 | save_regs(reglist); |
| | 3438 | if(stubs[n][4]!=0) emit_mov(stubs[n][4],0); |
| | 3439 | emit_call((int)&invalidate_addr); |
| | 3440 | restore_regs(reglist); |
| | 3441 | emit_jmp(stubs[n][2]); // return address |
| | 3442 | } |
| | 3443 | |
| | 3444 | int do_dirty_stub(int i) |
| | 3445 | { |
| | 3446 | assem_debug("do_dirty_stub %x\n",start+i*4); |
| | 3447 | u_int addr=(int)start<(int)0xC0000000?(u_int)source:(u_int)start; |
| | 3448 | #ifdef PCSX |
| | 3449 | addr=(u_int)source; |
| | 3450 | #endif |
| | 3451 | // Careful about the code output here, verify_dirty needs to parse it. |
| | 3452 | #ifdef ARMv5_ONLY |
| | 3453 | emit_loadlp(addr,1); |
| | 3454 | emit_loadlp((int)copy,2); |
| | 3455 | emit_loadlp(slen*4,3); |
| | 3456 | #else |
| | 3457 | emit_movw(addr&0x0000FFFF,1); |
| | 3458 | emit_movw(((u_int)copy)&0x0000FFFF,2); |
| | 3459 | emit_movt(addr&0xFFFF0000,1); |
| | 3460 | emit_movt(((u_int)copy)&0xFFFF0000,2); |
| | 3461 | emit_movw(slen*4,3); |
| | 3462 | #endif |
| | 3463 | emit_movimm(start+i*4,0); |
| | 3464 | emit_call((int)start<(int)0xC0000000?(int)&verify_code:(int)&verify_code_vm); |
| | 3465 | int entry=(int)out; |
| | 3466 | load_regs_entry(i); |
| | 3467 | if(entry==(int)out) entry=instr_addr[i]; |
| | 3468 | emit_jmp(instr_addr[i]); |
| | 3469 | return entry; |
| | 3470 | } |
| | 3471 | |
| | 3472 | void do_dirty_stub_ds() |
| | 3473 | { |
| | 3474 | // Careful about the code output here, verify_dirty needs to parse it. |
| | 3475 | #ifdef ARMv5_ONLY |
| | 3476 | emit_loadlp((int)start<(int)0xC0000000?(int)source:(int)start,1); |
| | 3477 | emit_loadlp((int)copy,2); |
| | 3478 | emit_loadlp(slen*4,3); |
| | 3479 | #else |
| | 3480 | emit_movw(((int)start<(int)0xC0000000?(u_int)source:(u_int)start)&0x0000FFFF,1); |
| | 3481 | emit_movw(((u_int)copy)&0x0000FFFF,2); |
| | 3482 | emit_movt(((int)start<(int)0xC0000000?(u_int)source:(u_int)start)&0xFFFF0000,1); |
| | 3483 | emit_movt(((u_int)copy)&0xFFFF0000,2); |
| | 3484 | emit_movw(slen*4,3); |
| | 3485 | #endif |
| | 3486 | emit_movimm(start+1,0); |
| | 3487 | emit_call((int)&verify_code_ds); |
| | 3488 | } |
| | 3489 | |
| | 3490 | do_cop1stub(int n) |
| | 3491 | { |
| | 3492 | literal_pool(256); |
| | 3493 | assem_debug("do_cop1stub %x\n",start+stubs[n][3]*4); |
| | 3494 | set_jump_target(stubs[n][1],(int)out); |
| | 3495 | int i=stubs[n][3]; |
| | 3496 | // int rs=stubs[n][4]; |
| | 3497 | struct regstat *i_regs=(struct regstat *)stubs[n][5]; |
| | 3498 | int ds=stubs[n][6]; |
| | 3499 | if(!ds) { |
| | 3500 | load_all_consts(regs[i].regmap_entry,regs[i].was32,regs[i].wasdirty,i); |
| | 3501 | //if(i_regs!=®s[i]) printf("oops: regs[i]=%x i_regs=%x",(int)®s[i],(int)i_regs); |
| | 3502 | } |
| | 3503 | //else {printf("fp exception in delay slot\n");} |
| | 3504 | wb_dirtys(i_regs->regmap_entry,i_regs->was32,i_regs->wasdirty); |
| | 3505 | if(regs[i].regmap_entry[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG); |
| | 3506 | emit_movimm(start+(i-ds)*4,EAX); // Get PC |
| | 3507 | emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); // CHECK: is this right? There should probably be an extra cycle... |
| | 3508 | emit_jmp(ds?(int)fp_exception_ds:(int)fp_exception); |
| | 3509 | } |
| | 3510 | |
| | 3511 | /* TLB */ |
| | 3512 | |
| | 3513 | int do_tlb_r(int s,int ar,int map,int x,int a,int shift,int c,u_int addr) |
| | 3514 | { |
| | 3515 | if(c) { |
| | 3516 | if((signed int)addr>=(signed int)0xC0000000) { |
| | 3517 | // address_generation already loaded the const |
| | 3518 | emit_readword_dualindexedx4(FP,map,map); |
| | 3519 | } |
| | 3520 | else |
| | 3521 | return -1; // No mapping |
| | 3522 | } |
| | 3523 | else { |
| | 3524 | assert(s!=map); |
| | 3525 | emit_movimm(((int)memory_map-(int)&dynarec_local)>>2,map); |
| | 3526 | emit_addsr12(map,s,map); |
| | 3527 | // Schedule this while we wait on the load |
| | 3528 | //if(x) emit_xorimm(s,x,ar); |
| | 3529 | if(shift>=0) emit_shlimm(s,3,shift); |
| | 3530 | if(~a) emit_andimm(s,a,ar); |
| | 3531 | emit_readword_dualindexedx4(FP,map,map); |
| | 3532 | } |
| | 3533 | return map; |
| | 3534 | } |
| | 3535 | int do_tlb_r_branch(int map, int c, u_int addr, int *jaddr) |
| | 3536 | { |
| | 3537 | if(!c||(signed int)addr>=(signed int)0xC0000000) { |
| | 3538 | emit_test(map,map); |
| | 3539 | *jaddr=(int)out; |
| | 3540 | emit_js(0); |
| | 3541 | } |
| | 3542 | return map; |
| | 3543 | } |
| | 3544 | |
| | 3545 | int gen_tlb_addr_r(int ar, int map) { |
| | 3546 | if(map>=0) { |
| | 3547 | assem_debug("add %s,%s,%s lsl #2\n",regname[ar],regname[ar],regname[map]); |
| | 3548 | output_w32(0xe0800100|rd_rn_rm(ar,ar,map)); |
| | 3549 | } |
| | 3550 | } |
| | 3551 | |
| | 3552 | int do_tlb_w(int s,int ar,int map,int x,int c,u_int addr) |
| | 3553 | { |
| | 3554 | if(c) { |
| | 3555 | if(addr<0x80800000||addr>=0xC0000000) { |
| | 3556 | // address_generation already loaded the const |
| | 3557 | emit_readword_dualindexedx4(FP,map,map); |
| | 3558 | } |
| | 3559 | else |
| | 3560 | return -1; // No mapping |
| | 3561 | } |
| | 3562 | else { |
| | 3563 | assert(s!=map); |
| | 3564 | emit_movimm(((int)memory_map-(int)&dynarec_local)>>2,map); |
| | 3565 | emit_addsr12(map,s,map); |
| | 3566 | // Schedule this while we wait on the load |
| | 3567 | //if(x) emit_xorimm(s,x,ar); |
| | 3568 | emit_readword_dualindexedx4(FP,map,map); |
| | 3569 | } |
| | 3570 | return map; |
| | 3571 | } |
| | 3572 | int do_tlb_w_branch(int map, int c, u_int addr, int *jaddr) |
| | 3573 | { |
| | 3574 | if(!c||addr<0x80800000||addr>=0xC0000000) { |
| | 3575 | emit_testimm(map,0x40000000); |
| | 3576 | *jaddr=(int)out; |
| | 3577 | emit_jne(0); |
| | 3578 | } |
| | 3579 | } |
| | 3580 | |
| | 3581 | int gen_tlb_addr_w(int ar, int map) { |
| | 3582 | if(map>=0) { |
| | 3583 | assem_debug("add %s,%s,%s lsl #2\n",regname[ar],regname[ar],regname[map]); |
| | 3584 | output_w32(0xe0800100|rd_rn_rm(ar,ar,map)); |
| | 3585 | } |
| | 3586 | } |
| | 3587 | |
| | 3588 | // Generate the address of the memory_map entry, relative to dynarec_local |
| | 3589 | generate_map_const(u_int addr,int reg) { |
| | 3590 | //printf("generate_map_const(%x,%s)\n",addr,regname[reg]); |
| | 3591 | emit_movimm((addr>>12)+(((u_int)memory_map-(u_int)&dynarec_local)>>2),reg); |
| | 3592 | } |
| | 3593 | |
| | 3594 | /* Special assem */ |
| | 3595 | |
| | 3596 | void shift_assemble_arm(int i,struct regstat *i_regs) |
| | 3597 | { |
| | 3598 | if(rt1[i]) { |
| | 3599 | if(opcode2[i]<=0x07) // SLLV/SRLV/SRAV |
| | 3600 | { |
| | 3601 | signed char s,t,shift; |
| | 3602 | t=get_reg(i_regs->regmap,rt1[i]); |
| | 3603 | s=get_reg(i_regs->regmap,rs1[i]); |
| | 3604 | shift=get_reg(i_regs->regmap,rs2[i]); |
| | 3605 | if(t>=0){ |
| | 3606 | if(rs1[i]==0) |
| | 3607 | { |
| | 3608 | emit_zeroreg(t); |
| | 3609 | } |
| | 3610 | else if(rs2[i]==0) |
| | 3611 | { |
| | 3612 | assert(s>=0); |
| | 3613 | if(s!=t) emit_mov(s,t); |
| | 3614 | } |
| | 3615 | else |
| | 3616 | { |
| | 3617 | emit_andimm(shift,31,HOST_TEMPREG); |
| | 3618 | if(opcode2[i]==4) // SLLV |
| | 3619 | { |
| | 3620 | emit_shl(s,HOST_TEMPREG,t); |
| | 3621 | } |
| | 3622 | if(opcode2[i]==6) // SRLV |
| | 3623 | { |
| | 3624 | emit_shr(s,HOST_TEMPREG,t); |
| | 3625 | } |
| | 3626 | if(opcode2[i]==7) // SRAV |
| | 3627 | { |
| | 3628 | emit_sar(s,HOST_TEMPREG,t); |
| | 3629 | } |
| | 3630 | } |
| | 3631 | } |
| | 3632 | } else { // DSLLV/DSRLV/DSRAV |
| | 3633 | signed char sh,sl,th,tl,shift; |
| | 3634 | th=get_reg(i_regs->regmap,rt1[i]|64); |
| | 3635 | tl=get_reg(i_regs->regmap,rt1[i]); |
| | 3636 | sh=get_reg(i_regs->regmap,rs1[i]|64); |
| | 3637 | sl=get_reg(i_regs->regmap,rs1[i]); |
| | 3638 | shift=get_reg(i_regs->regmap,rs2[i]); |
| | 3639 | if(tl>=0){ |
| | 3640 | if(rs1[i]==0) |
| | 3641 | { |
| | 3642 | emit_zeroreg(tl); |
| | 3643 | if(th>=0) emit_zeroreg(th); |
| | 3644 | } |
| | 3645 | else if(rs2[i]==0) |
| | 3646 | { |
| | 3647 | assert(sl>=0); |
| | 3648 | if(sl!=tl) emit_mov(sl,tl); |
| | 3649 | if(th>=0&&sh!=th) emit_mov(sh,th); |
| | 3650 | } |
| | 3651 | else |
| | 3652 | { |
| | 3653 | // FIXME: What if shift==tl ? |
| | 3654 | assert(shift!=tl); |
| | 3655 | int temp=get_reg(i_regs->regmap,-1); |
| | 3656 | int real_th=th; |
| | 3657 | if(th<0&&opcode2[i]!=0x14) {th=temp;} // DSLLV doesn't need a temporary register |
| | 3658 | assert(sl>=0); |
| | 3659 | assert(sh>=0); |
| | 3660 | emit_andimm(shift,31,HOST_TEMPREG); |
| | 3661 | if(opcode2[i]==0x14) // DSLLV |
| | 3662 | { |
| | 3663 | if(th>=0) emit_shl(sh,HOST_TEMPREG,th); |
| | 3664 | emit_rsbimm(HOST_TEMPREG,32,HOST_TEMPREG); |
| | 3665 | emit_orrshr(sl,HOST_TEMPREG,th); |
| | 3666 | emit_andimm(shift,31,HOST_TEMPREG); |
| | 3667 | emit_testimm(shift,32); |
| | 3668 | emit_shl(sl,HOST_TEMPREG,tl); |
| | 3669 | if(th>=0) emit_cmovne_reg(tl,th); |
| | 3670 | emit_cmovne_imm(0,tl); |
| | 3671 | } |
| | 3672 | if(opcode2[i]==0x16) // DSRLV |
| | 3673 | { |
| | 3674 | assert(th>=0); |
| | 3675 | emit_shr(sl,HOST_TEMPREG,tl); |
| | 3676 | emit_rsbimm(HOST_TEMPREG,32,HOST_TEMPREG); |
| | 3677 | emit_orrshl(sh,HOST_TEMPREG,tl); |
| | 3678 | emit_andimm(shift,31,HOST_TEMPREG); |
| | 3679 | emit_testimm(shift,32); |
| | 3680 | emit_shr(sh,HOST_TEMPREG,th); |
| | 3681 | emit_cmovne_reg(th,tl); |
| | 3682 | if(real_th>=0) emit_cmovne_imm(0,th); |
| | 3683 | } |
| | 3684 | if(opcode2[i]==0x17) // DSRAV |
| | 3685 | { |
| | 3686 | assert(th>=0); |
| | 3687 | emit_shr(sl,HOST_TEMPREG,tl); |
| | 3688 | emit_rsbimm(HOST_TEMPREG,32,HOST_TEMPREG); |
| | 3689 | if(real_th>=0) { |
| | 3690 | assert(temp>=0); |
| | 3691 | emit_sarimm(th,31,temp); |
| | 3692 | } |
| | 3693 | emit_orrshl(sh,HOST_TEMPREG,tl); |
| | 3694 | emit_andimm(shift,31,HOST_TEMPREG); |
| | 3695 | emit_testimm(shift,32); |
| | 3696 | emit_sar(sh,HOST_TEMPREG,th); |
| | 3697 | emit_cmovne_reg(th,tl); |
| | 3698 | if(real_th>=0) emit_cmovne_reg(temp,th); |
| | 3699 | } |
| | 3700 | } |
| | 3701 | } |
| | 3702 | } |
| | 3703 | } |
| | 3704 | } |
| | 3705 | |
| | 3706 | #ifdef PCSX |
| | 3707 | static void speculate_mov(int rs,int rt) |
| | 3708 | { |
| | 3709 | if(rt!=0) { |
| | 3710 | smrv_strong_next|=1<<rt; |
| | 3711 | smrv[rt]=smrv[rs]; |
| | 3712 | } |
| | 3713 | } |
| | 3714 | |
| | 3715 | static void speculate_mov_weak(int rs,int rt) |
| | 3716 | { |
| | 3717 | if(rt!=0) { |
| | 3718 | smrv_weak_next|=1<<rt; |
| | 3719 | smrv[rt]=smrv[rs]; |
| | 3720 | } |
| | 3721 | } |
| | 3722 | |
| | 3723 | static void speculate_register_values(int i) |
| | 3724 | { |
| | 3725 | if(i==0) { |
| | 3726 | memcpy(smrv,psxRegs.GPR.r,sizeof(smrv)); |
| | 3727 | // gp,sp are likely to stay the same throughout the block |
| | 3728 | smrv_strong_next=(1<<28)|(1<<29)|(1<<30); |
| | 3729 | smrv_weak_next=~smrv_strong_next; |
| | 3730 | //printf(" llr %08x\n", smrv[4]); |
| | 3731 | } |
| | 3732 | smrv_strong=smrv_strong_next; |
| | 3733 | smrv_weak=smrv_weak_next; |
| | 3734 | switch(itype[i]) { |
| | 3735 | case ALU: |
| | 3736 | if ((smrv_strong>>rs1[i])&1) speculate_mov(rs1[i],rt1[i]); |
| | 3737 | else if((smrv_strong>>rs2[i])&1) speculate_mov(rs2[i],rt1[i]); |
| | 3738 | else if((smrv_weak>>rs1[i])&1) speculate_mov_weak(rs1[i],rt1[i]); |
| | 3739 | else if((smrv_weak>>rs2[i])&1) speculate_mov_weak(rs2[i],rt1[i]); |
| | 3740 | else { |
| | 3741 | smrv_strong_next&=~(1<<rt1[i]); |
| | 3742 | smrv_weak_next&=~(1<<rt1[i]); |
| | 3743 | } |
| | 3744 | break; |
| | 3745 | case SHIFTIMM: |
| | 3746 | smrv_strong_next&=~(1<<rt1[i]); |
| | 3747 | smrv_weak_next&=~(1<<rt1[i]); |
| | 3748 | // fallthrough |
| | 3749 | case IMM16: |
| | 3750 | if(rt1[i]&&is_const(®s[i],rt1[i])) { |
| | 3751 | int value,hr=get_reg(regs[i].regmap,rt1[i]); |
| | 3752 | if(hr>=0) { |
| | 3753 | if(get_final_value(hr,i,&value)) |
| | 3754 | smrv[rt1[i]]=value; |
| | 3755 | else smrv[rt1[i]]=constmap[i][hr]; |
| | 3756 | smrv_strong_next|=1<<rt1[i]; |
| | 3757 | } |
| | 3758 | } |
| | 3759 | else { |
| | 3760 | if ((smrv_strong>>rs1[i])&1) speculate_mov(rs1[i],rt1[i]); |
| | 3761 | else if((smrv_weak>>rs1[i])&1) speculate_mov_weak(rs1[i],rt1[i]); |
| | 3762 | } |
| | 3763 | break; |
| | 3764 | case LOAD: |
| | 3765 | if(start<0x2000&&(rt1[i]==26||(smrv[rt1[i]]>>24)==0xa0)) { |
| | 3766 | // special case for BIOS |
| | 3767 | smrv[rt1[i]]=0xa0000000; |
| | 3768 | smrv_strong_next|=1<<rt1[i]; |
| | 3769 | break; |
| | 3770 | } |
| | 3771 | // fallthrough |
| | 3772 | case SHIFT: |
| | 3773 | case LOADLR: |
| | 3774 | case MOV: |
| | 3775 | smrv_strong_next&=~(1<<rt1[i]); |
| | 3776 | smrv_weak_next&=~(1<<rt1[i]); |
| | 3777 | break; |
| | 3778 | case COP0: |
| | 3779 | case COP2: |
| | 3780 | if(opcode2[i]==0||opcode2[i]==2) { // MFC/CFC |
| | 3781 | smrv_strong_next&=~(1<<rt1[i]); |
| | 3782 | smrv_weak_next&=~(1<<rt1[i]); |
| | 3783 | } |
| | 3784 | break; |
| | 3785 | case C2LS: |
| | 3786 | if (opcode[i]==0x32) { // LWC2 |
| | 3787 | smrv_strong_next&=~(1<<rt1[i]); |
| | 3788 | smrv_weak_next&=~(1<<rt1[i]); |
| | 3789 | } |
| | 3790 | break; |
| | 3791 | } |
| | 3792 | #if 0 |
| | 3793 | int r=4; |
| | 3794 | printf("x %08x %08x %d %d c %08x %08x\n",smrv[r],start+i*4, |
| | 3795 | ((smrv_strong>>r)&1),(smrv_weak>>r)&1,regs[i].isconst,regs[i].wasconst); |
| | 3796 | #endif |
| | 3797 | } |
| | 3798 | |
| | 3799 | enum { |
| | 3800 | MTYPE_8000 = 0, |
| | 3801 | MTYPE_8020, |
| | 3802 | MTYPE_0000, |
| | 3803 | MTYPE_A000, |
| | 3804 | MTYPE_1F80, |
| | 3805 | }; |
| | 3806 | |
| | 3807 | static int get_ptr_mem_type(u_int a) |
| | 3808 | { |
| | 3809 | if(a < 0x00200000) { |
| | 3810 | if(a<0x1000&&((start>>20)==0xbfc||(start>>24)==0xa0)) |
| | 3811 | // return wrong, must use memhandler for BIOS self-test to pass |
| | 3812 | // 007 does similar stuff from a00 mirror, weird stuff |
| | 3813 | return MTYPE_8000; |
| | 3814 | return MTYPE_0000; |
| | 3815 | } |
| | 3816 | if(0x1f800000 <= a && a < 0x1f801000) |
| | 3817 | return MTYPE_1F80; |
| | 3818 | if(0x80200000 <= a && a < 0x80800000) |
| | 3819 | return MTYPE_8020; |
| | 3820 | if(0xa0000000 <= a && a < 0xa0200000) |
| | 3821 | return MTYPE_A000; |
| | 3822 | return MTYPE_8000; |
| | 3823 | } |
| | 3824 | #endif |
| | 3825 | |
| | 3826 | static int emit_fastpath_cmp_jump(int i,int addr,int *addr_reg_override) |
| | 3827 | { |
| | 3828 | int jaddr,type=0; |
| | 3829 | |
| | 3830 | #ifdef PCSX |
| | 3831 | int mr=rs1[i]; |
| | 3832 | if(((smrv_strong|smrv_weak)>>mr)&1) { |
| | 3833 | type=get_ptr_mem_type(smrv[mr]); |
| | 3834 | //printf("set %08x @%08x r%d %d\n", smrv[mr], start+i*4, mr, type); |
| | 3835 | } |
| | 3836 | else { |
| | 3837 | // use the mirror we are running on |
| | 3838 | type=get_ptr_mem_type(start); |
| | 3839 | //printf("set nospec @%08x r%d %d\n", start+i*4, mr, type); |
| | 3840 | } |
| | 3841 | |
| | 3842 | if(type==MTYPE_8020) { // RAM 80200000+ mirror |
| | 3843 | emit_andimm(addr,~0x00e00000,HOST_TEMPREG); |
| | 3844 | addr=*addr_reg_override=HOST_TEMPREG; |
| | 3845 | type=0; |
| | 3846 | } |
| | 3847 | else if(type==MTYPE_0000) { // RAM 0 mirror |
| | 3848 | emit_orimm(addr,0x80000000,HOST_TEMPREG); |
| | 3849 | addr=*addr_reg_override=HOST_TEMPREG; |
| | 3850 | type=0; |
| | 3851 | } |
| | 3852 | else if(type==MTYPE_A000) { // RAM A mirror |
| | 3853 | emit_andimm(addr,~0x20000000,HOST_TEMPREG); |
| | 3854 | addr=*addr_reg_override=HOST_TEMPREG; |
| | 3855 | type=0; |
| | 3856 | } |
| | 3857 | else if(type==MTYPE_1F80) { // scratchpad |
| | 3858 | emit_addimm(addr,-0x1f800000,HOST_TEMPREG); |
| | 3859 | emit_cmpimm(HOST_TEMPREG,0x1000); |
| | 3860 | jaddr=(int)out; |
| | 3861 | emit_jc(0); |
| | 3862 | } |
| | 3863 | #endif |
| | 3864 | |
| | 3865 | if(type==0) |
| | 3866 | { |
| | 3867 | emit_cmpimm(addr,RAM_SIZE); |
| | 3868 | jaddr=(int)out; |
| | 3869 | #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK |
| | 3870 | // Hint to branch predictor that the branch is unlikely to be taken |
| | 3871 | if(rs1[i]>=28) |
| | 3872 | emit_jno_unlikely(0); |
| | 3873 | else |
| | 3874 | #endif |
| | 3875 | emit_jno(0); |
| | 3876 | } |
| | 3877 | |
| | 3878 | return jaddr; |
| | 3879 | } |
| | 3880 | |
| | 3881 | #define shift_assemble shift_assemble_arm |
| | 3882 | |
| | 3883 | void loadlr_assemble_arm(int i,struct regstat *i_regs) |
| | 3884 | { |
| | 3885 | int s,th,tl,temp,temp2,addr,map=-1; |
| | 3886 | int offset; |
| | 3887 | int jaddr=0; |
| | 3888 | int memtarget=0,c=0; |
| | 3889 | int fastload_reg_override=0; |
| | 3890 | u_int hr,reglist=0; |
| | 3891 | th=get_reg(i_regs->regmap,rt1[i]|64); |
| | 3892 | tl=get_reg(i_regs->regmap,rt1[i]); |
| | 3893 | s=get_reg(i_regs->regmap,rs1[i]); |
| | 3894 | temp=get_reg(i_regs->regmap,-1); |
| | 3895 | temp2=get_reg(i_regs->regmap,FTEMP); |
| | 3896 | addr=get_reg(i_regs->regmap,AGEN1+(i&1)); |
| | 3897 | assert(addr<0); |
| | 3898 | offset=imm[i]; |
| | 3899 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 3900 | if(i_regs->regmap[hr]>=0) reglist|=1<<hr; |
| | 3901 | } |
| | 3902 | reglist|=1<<temp; |
| | 3903 | if(offset||s<0||c) addr=temp2; |
| | 3904 | else addr=s; |
| | 3905 | if(s>=0) { |
| | 3906 | c=(i_regs->wasconst>>s)&1; |
| | 3907 | if(c) { |
| | 3908 | memtarget=((signed int)(constmap[i][s]+offset))<(signed int)0x80000000+RAM_SIZE; |
| | 3909 | if(using_tlb&&((signed int)(constmap[i][s]+offset))>=(signed int)0xC0000000) memtarget=1; |
| | 3910 | } |
| | 3911 | } |
| | 3912 | if(!using_tlb) { |
| | 3913 | if(!c) { |
| | 3914 | #ifdef RAM_OFFSET |
| | 3915 | map=get_reg(i_regs->regmap,ROREG); |
| | 3916 | if(map<0) emit_loadreg(ROREG,map=HOST_TEMPREG); |
| | 3917 | #endif |
| | 3918 | emit_shlimm(addr,3,temp); |
| | 3919 | if (opcode[i]==0x22||opcode[i]==0x26) { |
| | 3920 | emit_andimm(addr,0xFFFFFFFC,temp2); // LWL/LWR |
| | 3921 | }else{ |
| | 3922 | emit_andimm(addr,0xFFFFFFF8,temp2); // LDL/LDR |
| | 3923 | } |
| | 3924 | jaddr=emit_fastpath_cmp_jump(i,temp2,&fastload_reg_override); |
| | 3925 | } |
| | 3926 | else { |
| | 3927 | if (opcode[i]==0x22||opcode[i]==0x26) { |
| | 3928 | emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR |
| | 3929 | }else{ |
| | 3930 | emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR |
| | 3931 | } |
| | 3932 | } |
| | 3933 | }else{ // using tlb |
| | 3934 | int a; |
| | 3935 | if(c) { |
| | 3936 | a=-1; |
| | 3937 | }else if (opcode[i]==0x22||opcode[i]==0x26) { |
| | 3938 | a=0xFFFFFFFC; // LWL/LWR |
| | 3939 | }else{ |
| | 3940 | a=0xFFFFFFF8; // LDL/LDR |
| | 3941 | } |
| | 3942 | map=get_reg(i_regs->regmap,TLREG); |
| | 3943 | assert(map>=0); |
| | 3944 | reglist&=~(1<<map); |
| | 3945 | map=do_tlb_r(addr,temp2,map,0,a,c?-1:temp,c,constmap[i][s]+offset); |
| | 3946 | if(c) { |
| | 3947 | if (opcode[i]==0x22||opcode[i]==0x26) { |
| | 3948 | emit_movimm(((constmap[i][s]+offset)<<3)&24,temp); // LWL/LWR |
| | 3949 | }else{ |
| | 3950 | emit_movimm(((constmap[i][s]+offset)<<3)&56,temp); // LDL/LDR |
| | 3951 | } |
| | 3952 | } |
| | 3953 | do_tlb_r_branch(map,c,constmap[i][s]+offset,&jaddr); |
| | 3954 | } |
| | 3955 | if (opcode[i]==0x22||opcode[i]==0x26) { // LWL/LWR |
| | 3956 | if(!c||memtarget) { |
| | 3957 | int a=temp2; |
| | 3958 | if(fastload_reg_override) a=fastload_reg_override; |
| | 3959 | //emit_readword_indexed((int)rdram-0x80000000,temp2,temp2); |
| | 3960 | emit_readword_indexed_tlb(0,a,map,temp2); |
| | 3961 | if(jaddr) add_stub(LOADW_STUB,jaddr,(int)out,i,temp2,(int)i_regs,ccadj[i],reglist); |
| | 3962 | } |
| | 3963 | else |
| | 3964 | inline_readstub(LOADW_STUB,i,(constmap[i][s]+offset)&0xFFFFFFFC,i_regs->regmap,FTEMP,ccadj[i],reglist); |
| | 3965 | if(rt1[i]) { |
| | 3966 | assert(tl>=0); |
| | 3967 | emit_andimm(temp,24,temp); |
| | 3968 | #ifdef BIG_ENDIAN_MIPS |
| | 3969 | if (opcode[i]==0x26) // LWR |
| | 3970 | #else |
| | 3971 | if (opcode[i]==0x22) // LWL |
| | 3972 | #endif |
| | 3973 | emit_xorimm(temp,24,temp); |
| | 3974 | emit_movimm(-1,HOST_TEMPREG); |
| | 3975 | if (opcode[i]==0x26) { |
| | 3976 | emit_shr(temp2,temp,temp2); |
| | 3977 | emit_bic_lsr(tl,HOST_TEMPREG,temp,tl); |
| | 3978 | }else{ |
| | 3979 | emit_shl(temp2,temp,temp2); |
| | 3980 | emit_bic_lsl(tl,HOST_TEMPREG,temp,tl); |
| | 3981 | } |
| | 3982 | emit_or(temp2,tl,tl); |
| | 3983 | } |
| | 3984 | //emit_storereg(rt1[i],tl); // DEBUG |
| | 3985 | } |
| | 3986 | if (opcode[i]==0x1A||opcode[i]==0x1B) { // LDL/LDR |
| | 3987 | // FIXME: little endian, fastload_reg_override |
| | 3988 | int temp2h=get_reg(i_regs->regmap,FTEMP|64); |
| | 3989 | if(!c||memtarget) { |
| | 3990 | //if(th>=0) emit_readword_indexed((int)rdram-0x80000000,temp2,temp2h); |
| | 3991 | //emit_readword_indexed((int)rdram-0x7FFFFFFC,temp2,temp2); |
| | 3992 | emit_readdword_indexed_tlb(0,temp2,map,temp2h,temp2); |
| | 3993 | if(jaddr) add_stub(LOADD_STUB,jaddr,(int)out,i,temp2,(int)i_regs,ccadj[i],reglist); |
| | 3994 | } |
| | 3995 | else |
| | 3996 | inline_readstub(LOADD_STUB,i,(constmap[i][s]+offset)&0xFFFFFFF8,i_regs->regmap,FTEMP,ccadj[i],reglist); |
| | 3997 | if(rt1[i]) { |
| | 3998 | assert(th>=0); |
| | 3999 | assert(tl>=0); |
| | 4000 | emit_testimm(temp,32); |
| | 4001 | emit_andimm(temp,24,temp); |
| | 4002 | if (opcode[i]==0x1A) { // LDL |
| | 4003 | emit_rsbimm(temp,32,HOST_TEMPREG); |
| | 4004 | emit_shl(temp2h,temp,temp2h); |
| | 4005 | emit_orrshr(temp2,HOST_TEMPREG,temp2h); |
| | 4006 | emit_movimm(-1,HOST_TEMPREG); |
| | 4007 | emit_shl(temp2,temp,temp2); |
| | 4008 | emit_cmove_reg(temp2h,th); |
| | 4009 | emit_biceq_lsl(tl,HOST_TEMPREG,temp,tl); |
| | 4010 | emit_bicne_lsl(th,HOST_TEMPREG,temp,th); |
| | 4011 | emit_orreq(temp2,tl,tl); |
| | 4012 | emit_orrne(temp2,th,th); |
| | 4013 | } |
| | 4014 | if (opcode[i]==0x1B) { // LDR |
| | 4015 | emit_xorimm(temp,24,temp); |
| | 4016 | emit_rsbimm(temp,32,HOST_TEMPREG); |
| | 4017 | emit_shr(temp2,temp,temp2); |
| | 4018 | emit_orrshl(temp2h,HOST_TEMPREG,temp2); |
| | 4019 | emit_movimm(-1,HOST_TEMPREG); |
| | 4020 | emit_shr(temp2h,temp,temp2h); |
| | 4021 | emit_cmovne_reg(temp2,tl); |
| | 4022 | emit_bicne_lsr(th,HOST_TEMPREG,temp,th); |
| | 4023 | emit_biceq_lsr(tl,HOST_TEMPREG,temp,tl); |
| | 4024 | emit_orrne(temp2h,th,th); |
| | 4025 | emit_orreq(temp2h,tl,tl); |
| | 4026 | } |
| | 4027 | } |
| | 4028 | } |
| | 4029 | } |
| | 4030 | #define loadlr_assemble loadlr_assemble_arm |
| | 4031 | |
| | 4032 | void cop0_assemble(int i,struct regstat *i_regs) |
| | 4033 | { |
| | 4034 | if(opcode2[i]==0) // MFC0 |
| | 4035 | { |
| | 4036 | signed char t=get_reg(i_regs->regmap,rt1[i]); |
| | 4037 | char copr=(source[i]>>11)&0x1f; |
| | 4038 | //assert(t>=0); // Why does this happen? OOT is weird |
| | 4039 | if(t>=0&&rt1[i]!=0) { |
| | 4040 | #ifdef MUPEN64 |
| | 4041 | emit_addimm(FP,(int)&fake_pc-(int)&dynarec_local,0); |
| | 4042 | emit_movimm((source[i]>>11)&0x1f,1); |
| | 4043 | emit_writeword(0,(int)&PC); |
| | 4044 | emit_writebyte(1,(int)&(fake_pc.f.r.nrd)); |
| | 4045 | if(copr==9) { |
| | 4046 | emit_readword((int)&last_count,ECX); |
| | 4047 | emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc |
| | 4048 | emit_add(HOST_CCREG,ECX,HOST_CCREG); |
| | 4049 | emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); |
| | 4050 | emit_writeword(HOST_CCREG,(int)&Count); |
| | 4051 | } |
| | 4052 | emit_call((int)MFC0); |
| | 4053 | emit_readword((int)&readmem_dword,t); |
| | 4054 | #else |
| | 4055 | emit_readword((int)®_cop0+copr*4,t); |
| | 4056 | #endif |
| | 4057 | } |
| | 4058 | } |
| | 4059 | else if(opcode2[i]==4) // MTC0 |
| | 4060 | { |
| | 4061 | signed char s=get_reg(i_regs->regmap,rs1[i]); |
| | 4062 | char copr=(source[i]>>11)&0x1f; |
| | 4063 | assert(s>=0); |
| | 4064 | emit_writeword(s,(int)&readmem_dword); |
| | 4065 | wb_register(rs1[i],i_regs->regmap,i_regs->dirty,i_regs->is32); |
| | 4066 | #ifdef MUPEN64 |
| | 4067 | emit_addimm(FP,(int)&fake_pc-(int)&dynarec_local,0); |
| | 4068 | emit_movimm((source[i]>>11)&0x1f,1); |
| | 4069 | emit_writeword(0,(int)&PC); |
| | 4070 | emit_writebyte(1,(int)&(fake_pc.f.r.nrd)); |
| | 4071 | #endif |
| | 4072 | if(copr==9||copr==11||copr==12||copr==13) { |
| | 4073 | emit_readword((int)&last_count,ECX); |
| | 4074 | emit_loadreg(CCREG,HOST_CCREG); // TODO: do proper reg alloc |
| | 4075 | emit_add(HOST_CCREG,ECX,HOST_CCREG); |
| | 4076 | emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); |
| | 4077 | emit_writeword(HOST_CCREG,(int)&Count); |
| | 4078 | } |
| | 4079 | // What a mess. The status register (12) can enable interrupts, |
| | 4080 | // so needs a special case to handle a pending interrupt. |
| | 4081 | // The interrupt must be taken immediately, because a subsequent |
| | 4082 | // instruction might disable interrupts again. |
| | 4083 | if(copr==12||copr==13) { |
| | 4084 | #ifdef PCSX |
| | 4085 | if (is_delayslot) { |
| | 4086 | // burn cycles to cause cc_interrupt, which will |
| | 4087 | // reschedule next_interupt. Relies on CCREG from above. |
| | 4088 | assem_debug("MTC0 DS %d\n", copr); |
| | 4089 | emit_writeword(HOST_CCREG,(int)&last_count); |
| | 4090 | emit_movimm(0,HOST_CCREG); |
| | 4091 | emit_storereg(CCREG,HOST_CCREG); |
| | 4092 | emit_movimm(copr,0); |
| | 4093 | emit_call((int)pcsx_mtc0_ds); |
| | 4094 | return; |
| | 4095 | } |
| | 4096 | #endif |
| | 4097 | emit_movimm(start+i*4+4,0); |
| | 4098 | emit_movimm(0,1); |
| | 4099 | emit_writeword(0,(int)&pcaddr); |
| | 4100 | emit_writeword(1,(int)&pending_exception); |
| | 4101 | } |
| | 4102 | //else if(copr==12&&is_delayslot) emit_call((int)MTC0_R12); |
| | 4103 | //else |
| | 4104 | #ifdef PCSX |
| | 4105 | emit_movimm(copr,0); |
| | 4106 | emit_call((int)pcsx_mtc0); |
| | 4107 | #else |
| | 4108 | emit_call((int)MTC0); |
| | 4109 | #endif |
| | 4110 | if(copr==9||copr==11||copr==12||copr==13) { |
| | 4111 | emit_readword((int)&Count,HOST_CCREG); |
| | 4112 | emit_readword((int)&next_interupt,ECX); |
| | 4113 | emit_addimm(HOST_CCREG,-CLOCK_DIVIDER*ccadj[i],HOST_CCREG); |
| | 4114 | emit_sub(HOST_CCREG,ECX,HOST_CCREG); |
| | 4115 | emit_writeword(ECX,(int)&last_count); |
| | 4116 | emit_storereg(CCREG,HOST_CCREG); |
| | 4117 | } |
| | 4118 | if(copr==12||copr==13) { |
| | 4119 | assert(!is_delayslot); |
| | 4120 | emit_readword((int)&pending_exception,14); |
| | 4121 | } |
| | 4122 | emit_loadreg(rs1[i],s); |
| | 4123 | if(get_reg(i_regs->regmap,rs1[i]|64)>=0) |
| | 4124 | emit_loadreg(rs1[i]|64,get_reg(i_regs->regmap,rs1[i]|64)); |
| | 4125 | if(copr==12||copr==13) { |
| | 4126 | emit_test(14,14); |
| | 4127 | emit_jne((int)&do_interrupt); |
| | 4128 | } |
| | 4129 | cop1_usable=0; |
| | 4130 | } |
| | 4131 | else |
| | 4132 | { |
| | 4133 | assert(opcode2[i]==0x10); |
| | 4134 | #ifndef DISABLE_TLB |
| | 4135 | if((source[i]&0x3f)==0x01) // TLBR |
| | 4136 | emit_call((int)TLBR); |
| | 4137 | if((source[i]&0x3f)==0x02) // TLBWI |
| | 4138 | emit_call((int)TLBWI_new); |
| | 4139 | if((source[i]&0x3f)==0x06) { // TLBWR |
| | 4140 | // The TLB entry written by TLBWR is dependent on the count, |
| | 4141 | // so update the cycle count |
| | 4142 | emit_readword((int)&last_count,ECX); |
| | 4143 | if(i_regs->regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG); |
| | 4144 | emit_add(HOST_CCREG,ECX,HOST_CCREG); |
| | 4145 | emit_addimm(HOST_CCREG,CLOCK_DIVIDER*ccadj[i],HOST_CCREG); |
| | 4146 | emit_writeword(HOST_CCREG,(int)&Count); |
| | 4147 | emit_call((int)TLBWR_new); |
| | 4148 | } |
| | 4149 | if((source[i]&0x3f)==0x08) // TLBP |
| | 4150 | emit_call((int)TLBP); |
| | 4151 | #endif |
| | 4152 | #ifdef PCSX |
| | 4153 | if((source[i]&0x3f)==0x10) // RFE |
| | 4154 | { |
| | 4155 | emit_readword((int)&Status,0); |
| | 4156 | emit_andimm(0,0x3c,1); |
| | 4157 | emit_andimm(0,~0xf,0); |
| | 4158 | emit_orrshr_imm(1,2,0); |
| | 4159 | emit_writeword(0,(int)&Status); |
| | 4160 | } |
| | 4161 | #else |
| | 4162 | if((source[i]&0x3f)==0x18) // ERET |
| | 4163 | { |
| | 4164 | int count=ccadj[i]; |
| | 4165 | if(i_regs->regmap[HOST_CCREG]!=CCREG) emit_loadreg(CCREG,HOST_CCREG); |
| | 4166 | emit_addimm(HOST_CCREG,CLOCK_DIVIDER*count,HOST_CCREG); // TODO: Should there be an extra cycle here? |
| | 4167 | emit_jmp((int)jump_eret); |
| | 4168 | } |
| | 4169 | #endif |
| | 4170 | } |
| | 4171 | } |
| | 4172 | |
| | 4173 | static void cop2_get_dreg(u_int copr,signed char tl,signed char temp) |
| | 4174 | { |
| | 4175 | switch (copr) { |
| | 4176 | case 1: |
| | 4177 | case 3: |
| | 4178 | case 5: |
| | 4179 | case 8: |
| | 4180 | case 9: |
| | 4181 | case 10: |
| | 4182 | case 11: |
| | 4183 | emit_readword((int)®_cop2d[copr],tl); |
| | 4184 | emit_signextend16(tl,tl); |
| | 4185 | emit_writeword(tl,(int)®_cop2d[copr]); // hmh |
| | 4186 | break; |
| | 4187 | case 7: |
| | 4188 | case 16: |
| | 4189 | case 17: |
| | 4190 | case 18: |
| | 4191 | case 19: |
| | 4192 | emit_readword((int)®_cop2d[copr],tl); |
| | 4193 | emit_andimm(tl,0xffff,tl); |
| | 4194 | emit_writeword(tl,(int)®_cop2d[copr]); |
| | 4195 | break; |
| | 4196 | case 15: |
| | 4197 | emit_readword((int)®_cop2d[14],tl); // SXY2 |
| | 4198 | emit_writeword(tl,(int)®_cop2d[copr]); |
| | 4199 | break; |
| | 4200 | case 28: |
| | 4201 | case 29: |
| | 4202 | emit_readword((int)®_cop2d[9],temp); |
| | 4203 | emit_testimm(temp,0x8000); // do we need this? |
| | 4204 | emit_andimm(temp,0xf80,temp); |
| | 4205 | emit_andne_imm(temp,0,temp); |
| | 4206 | emit_shrimm(temp,7,tl); |
| | 4207 | emit_readword((int)®_cop2d[10],temp); |
| | 4208 | emit_testimm(temp,0x8000); |
| | 4209 | emit_andimm(temp,0xf80,temp); |
| | 4210 | emit_andne_imm(temp,0,temp); |
| | 4211 | emit_orrshr_imm(temp,2,tl); |
| | 4212 | emit_readword((int)®_cop2d[11],temp); |
| | 4213 | emit_testimm(temp,0x8000); |
| | 4214 | emit_andimm(temp,0xf80,temp); |
| | 4215 | emit_andne_imm(temp,0,temp); |
| | 4216 | emit_orrshl_imm(temp,3,tl); |
| | 4217 | emit_writeword(tl,(int)®_cop2d[copr]); |
| | 4218 | break; |
| | 4219 | default: |
| | 4220 | emit_readword((int)®_cop2d[copr],tl); |
| | 4221 | break; |
| | 4222 | } |
| | 4223 | } |
| | 4224 | |
| | 4225 | static void cop2_put_dreg(u_int copr,signed char sl,signed char temp) |
| | 4226 | { |
| | 4227 | switch (copr) { |
| | 4228 | case 15: |
| | 4229 | emit_readword((int)®_cop2d[13],temp); // SXY1 |
| | 4230 | emit_writeword(sl,(int)®_cop2d[copr]); |
| | 4231 | emit_writeword(temp,(int)®_cop2d[12]); // SXY0 |
| | 4232 | emit_readword((int)®_cop2d[14],temp); // SXY2 |
| | 4233 | emit_writeword(sl,(int)®_cop2d[14]); |
| | 4234 | emit_writeword(temp,(int)®_cop2d[13]); // SXY1 |
| | 4235 | break; |
| | 4236 | case 28: |
| | 4237 | emit_andimm(sl,0x001f,temp); |
| | 4238 | emit_shlimm(temp,7,temp); |
| | 4239 | emit_writeword(temp,(int)®_cop2d[9]); |
| | 4240 | emit_andimm(sl,0x03e0,temp); |
| | 4241 | emit_shlimm(temp,2,temp); |
| | 4242 | emit_writeword(temp,(int)®_cop2d[10]); |
| | 4243 | emit_andimm(sl,0x7c00,temp); |
| | 4244 | emit_shrimm(temp,3,temp); |
| | 4245 | emit_writeword(temp,(int)®_cop2d[11]); |
| | 4246 | emit_writeword(sl,(int)®_cop2d[28]); |
| | 4247 | break; |
| | 4248 | case 30: |
| | 4249 | emit_movs(sl,temp); |
| | 4250 | emit_mvnmi(temp,temp); |
| | 4251 | emit_clz(temp,temp); |
| | 4252 | emit_writeword(sl,(int)®_cop2d[30]); |
| | 4253 | emit_writeword(temp,(int)®_cop2d[31]); |
| | 4254 | break; |
| | 4255 | case 31: |
| | 4256 | break; |
| | 4257 | default: |
| | 4258 | emit_writeword(sl,(int)®_cop2d[copr]); |
| | 4259 | break; |
| | 4260 | } |
| | 4261 | } |
| | 4262 | |
| | 4263 | void cop2_assemble(int i,struct regstat *i_regs) |
| | 4264 | { |
| | 4265 | u_int copr=(source[i]>>11)&0x1f; |
| | 4266 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4267 | if (opcode2[i]==0) { // MFC2 |
| | 4268 | signed char tl=get_reg(i_regs->regmap,rt1[i]); |
| | 4269 | if(tl>=0&&rt1[i]!=0) |
| | 4270 | cop2_get_dreg(copr,tl,temp); |
| | 4271 | } |
| | 4272 | else if (opcode2[i]==4) { // MTC2 |
| | 4273 | signed char sl=get_reg(i_regs->regmap,rs1[i]); |
| | 4274 | cop2_put_dreg(copr,sl,temp); |
| | 4275 | } |
| | 4276 | else if (opcode2[i]==2) // CFC2 |
| | 4277 | { |
| | 4278 | signed char tl=get_reg(i_regs->regmap,rt1[i]); |
| | 4279 | if(tl>=0&&rt1[i]!=0) |
| | 4280 | emit_readword((int)®_cop2c[copr],tl); |
| | 4281 | } |
| | 4282 | else if (opcode2[i]==6) // CTC2 |
| | 4283 | { |
| | 4284 | signed char sl=get_reg(i_regs->regmap,rs1[i]); |
| | 4285 | switch(copr) { |
| | 4286 | case 4: |
| | 4287 | case 12: |
| | 4288 | case 20: |
| | 4289 | case 26: |
| | 4290 | case 27: |
| | 4291 | case 29: |
| | 4292 | case 30: |
| | 4293 | emit_signextend16(sl,temp); |
| | 4294 | break; |
| | 4295 | case 31: |
| | 4296 | //value = value & 0x7ffff000; |
| | 4297 | //if (value & 0x7f87e000) value |= 0x80000000; |
| | 4298 | emit_shrimm(sl,12,temp); |
| | 4299 | emit_shlimm(temp,12,temp); |
| | 4300 | emit_testimm(temp,0x7f000000); |
| | 4301 | emit_testeqimm(temp,0x00870000); |
| | 4302 | emit_testeqimm(temp,0x0000e000); |
| | 4303 | emit_orrne_imm(temp,0x80000000,temp); |
| | 4304 | break; |
| | 4305 | default: |
| | 4306 | temp=sl; |
| | 4307 | break; |
| | 4308 | } |
| | 4309 | emit_writeword(temp,(int)®_cop2c[copr]); |
| | 4310 | assert(sl>=0); |
| | 4311 | } |
| | 4312 | } |
| | 4313 | |
| | 4314 | void c2op_assemble(int i,struct regstat *i_regs) |
| | 4315 | { |
| | 4316 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4317 | u_int c2op=source[i]&0x3f; |
| | 4318 | u_int hr,reglist=0; |
| | 4319 | int need_flags; |
| | 4320 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 4321 | if(i_regs->regmap[hr]>=0) reglist|=1<<hr; |
| | 4322 | } |
| | 4323 | if(i==0||itype[i-1]!=C2OP) |
| | 4324 | save_regs(reglist); |
| | 4325 | |
| | 4326 | if (gte_handlers[c2op]!=NULL) { |
| | 4327 | int cc=get_reg(i_regs->regmap,CCREG); |
| | 4328 | emit_movimm(source[i],1); // opcode |
| | 4329 | if (cc>=0&>e_cycletab[c2op]) |
| | 4330 | emit_addimm(cc,gte_cycletab[c2op]/2,cc); // XXX: could just adjust ccadj? |
| | 4331 | emit_addimm(FP,(int)&psxRegs.CP2D.r[0]-(int)&dynarec_local,0); // cop2 regs |
| | 4332 | emit_writeword(1,(int)&psxRegs.code); |
| | 4333 | need_flags=!(gte_unneeded[i+1]>>63); // +1 because of how liveness detection works |
| | 4334 | assem_debug("gte unneeded %016llx, need_flags %d\n",gte_unneeded[i+1],need_flags); |
| | 4335 | #ifdef ARMv5_ONLY |
| | 4336 | // let's take more risk here |
| | 4337 | need_flags=need_flags&>e_reads_flags; |
| | 4338 | #endif |
| | 4339 | emit_call((int)(need_flags?gte_handlers[c2op]:gte_handlers_nf[c2op])); |
| | 4340 | } |
| | 4341 | |
| | 4342 | if(i>=slen-1||itype[i+1]!=C2OP) |
| | 4343 | restore_regs(reglist); |
| | 4344 | } |
| | 4345 | |
| | 4346 | void cop1_unusable(int i,struct regstat *i_regs) |
| | 4347 | { |
| | 4348 | // XXX: should just just do the exception instead |
| | 4349 | if(!cop1_usable) { |
| | 4350 | int jaddr=(int)out; |
| | 4351 | emit_jmp(0); |
| | 4352 | add_stub(FP_STUB,jaddr,(int)out,i,0,(int)i_regs,is_delayslot,0); |
| | 4353 | cop1_usable=1; |
| | 4354 | } |
| | 4355 | } |
| | 4356 | |
| | 4357 | void cop1_assemble(int i,struct regstat *i_regs) |
| | 4358 | { |
| | 4359 | #ifndef DISABLE_COP1 |
| | 4360 | // Check cop1 unusable |
| | 4361 | if(!cop1_usable) { |
| | 4362 | signed char rs=get_reg(i_regs->regmap,CSREG); |
| | 4363 | assert(rs>=0); |
| | 4364 | emit_testimm(rs,0x20000000); |
| | 4365 | int jaddr=(int)out; |
| | 4366 | emit_jeq(0); |
| | 4367 | add_stub(FP_STUB,jaddr,(int)out,i,rs,(int)i_regs,is_delayslot,0); |
| | 4368 | cop1_usable=1; |
| | 4369 | } |
| | 4370 | if (opcode2[i]==0) { // MFC1 |
| | 4371 | signed char tl=get_reg(i_regs->regmap,rt1[i]); |
| | 4372 | if(tl>=0) { |
| | 4373 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],tl); |
| | 4374 | emit_readword_indexed(0,tl,tl); |
| | 4375 | } |
| | 4376 | } |
| | 4377 | else if (opcode2[i]==1) { // DMFC1 |
| | 4378 | signed char tl=get_reg(i_regs->regmap,rt1[i]); |
| | 4379 | signed char th=get_reg(i_regs->regmap,rt1[i]|64); |
| | 4380 | if(tl>=0) { |
| | 4381 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],tl); |
| | 4382 | if(th>=0) emit_readword_indexed(4,tl,th); |
| | 4383 | emit_readword_indexed(0,tl,tl); |
| | 4384 | } |
| | 4385 | } |
| | 4386 | else if (opcode2[i]==4) { // MTC1 |
| | 4387 | signed char sl=get_reg(i_regs->regmap,rs1[i]); |
| | 4388 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4389 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4390 | emit_writeword_indexed(sl,0,temp); |
| | 4391 | } |
| | 4392 | else if (opcode2[i]==5) { // DMTC1 |
| | 4393 | signed char sl=get_reg(i_regs->regmap,rs1[i]); |
| | 4394 | signed char sh=rs1[i]>0?get_reg(i_regs->regmap,rs1[i]|64):sl; |
| | 4395 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4396 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4397 | emit_writeword_indexed(sh,4,temp); |
| | 4398 | emit_writeword_indexed(sl,0,temp); |
| | 4399 | } |
| | 4400 | else if (opcode2[i]==2) // CFC1 |
| | 4401 | { |
| | 4402 | signed char tl=get_reg(i_regs->regmap,rt1[i]); |
| | 4403 | if(tl>=0) { |
| | 4404 | u_int copr=(source[i]>>11)&0x1f; |
| | 4405 | if(copr==0) emit_readword((int)&FCR0,tl); |
| | 4406 | if(copr==31) emit_readword((int)&FCR31,tl); |
| | 4407 | } |
| | 4408 | } |
| | 4409 | else if (opcode2[i]==6) // CTC1 |
| | 4410 | { |
| | 4411 | signed char sl=get_reg(i_regs->regmap,rs1[i]); |
| | 4412 | u_int copr=(source[i]>>11)&0x1f; |
| | 4413 | assert(sl>=0); |
| | 4414 | if(copr==31) |
| | 4415 | { |
| | 4416 | emit_writeword(sl,(int)&FCR31); |
| | 4417 | // Set the rounding mode |
| | 4418 | //FIXME |
| | 4419 | //char temp=get_reg(i_regs->regmap,-1); |
| | 4420 | //emit_andimm(sl,3,temp); |
| | 4421 | //emit_fldcw_indexed((int)&rounding_modes,temp); |
| | 4422 | } |
| | 4423 | } |
| | 4424 | #else |
| | 4425 | cop1_unusable(i, i_regs); |
| | 4426 | #endif |
| | 4427 | } |
| | 4428 | |
| | 4429 | void fconv_assemble_arm(int i,struct regstat *i_regs) |
| | 4430 | { |
| | 4431 | #ifndef DISABLE_COP1 |
| | 4432 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4433 | assert(temp>=0); |
| | 4434 | // Check cop1 unusable |
| | 4435 | if(!cop1_usable) { |
| | 4436 | signed char rs=get_reg(i_regs->regmap,CSREG); |
| | 4437 | assert(rs>=0); |
| | 4438 | emit_testimm(rs,0x20000000); |
| | 4439 | int jaddr=(int)out; |
| | 4440 | emit_jeq(0); |
| | 4441 | add_stub(FP_STUB,jaddr,(int)out,i,rs,(int)i_regs,is_delayslot,0); |
| | 4442 | cop1_usable=1; |
| | 4443 | } |
| | 4444 | |
| | 4445 | #if(defined(__VFP_FP__) && !defined(__SOFTFP__)) |
| | 4446 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0d) { // trunc_w_s |
| | 4447 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4448 | emit_flds(temp,15); |
| | 4449 | emit_ftosizs(15,15); // float->int, truncate |
| | 4450 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) |
| | 4451 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4452 | emit_fsts(15,temp); |
| | 4453 | return; |
| | 4454 | } |
| | 4455 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0d) { // trunc_w_d |
| | 4456 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4457 | emit_vldr(temp,7); |
| | 4458 | emit_ftosizd(7,13); // double->int, truncate |
| | 4459 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4460 | emit_fsts(13,temp); |
| | 4461 | return; |
| | 4462 | } |
| | 4463 | |
| | 4464 | if(opcode2[i]==0x14&&(source[i]&0x3f)==0x20) { // cvt_s_w |
| | 4465 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4466 | emit_flds(temp,13); |
| | 4467 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) |
| | 4468 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4469 | emit_fsitos(13,15); |
| | 4470 | emit_fsts(15,temp); |
| | 4471 | return; |
| | 4472 | } |
| | 4473 | if(opcode2[i]==0x14&&(source[i]&0x3f)==0x21) { // cvt_d_w |
| | 4474 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4475 | emit_flds(temp,13); |
| | 4476 | emit_readword((int)®_cop1_double[(source[i]>>6)&0x1f],temp); |
| | 4477 | emit_fsitod(13,7); |
| | 4478 | emit_vstr(7,temp); |
| | 4479 | return; |
| | 4480 | } |
| | 4481 | |
| | 4482 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x21) { // cvt_d_s |
| | 4483 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4484 | emit_flds(temp,13); |
| | 4485 | emit_readword((int)®_cop1_double[(source[i]>>6)&0x1f],temp); |
| | 4486 | emit_fcvtds(13,7); |
| | 4487 | emit_vstr(7,temp); |
| | 4488 | return; |
| | 4489 | } |
| | 4490 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x20) { // cvt_s_d |
| | 4491 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4492 | emit_vldr(temp,7); |
| | 4493 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4494 | emit_fcvtsd(7,13); |
| | 4495 | emit_fsts(13,temp); |
| | 4496 | return; |
| | 4497 | } |
| | 4498 | #endif |
| | 4499 | |
| | 4500 | // C emulation code |
| | 4501 | |
| | 4502 | u_int hr,reglist=0; |
| | 4503 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 4504 | if(i_regs->regmap[hr]>=0) reglist|=1<<hr; |
| | 4505 | } |
| | 4506 | save_regs(reglist); |
| | 4507 | |
| | 4508 | if(opcode2[i]==0x14&&(source[i]&0x3f)==0x20) { |
| | 4509 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4510 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4511 | emit_call((int)cvt_s_w); |
| | 4512 | } |
| | 4513 | if(opcode2[i]==0x14&&(source[i]&0x3f)==0x21) { |
| | 4514 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4515 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4516 | emit_call((int)cvt_d_w); |
| | 4517 | } |
| | 4518 | if(opcode2[i]==0x15&&(source[i]&0x3f)==0x20) { |
| | 4519 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4520 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4521 | emit_call((int)cvt_s_l); |
| | 4522 | } |
| | 4523 | if(opcode2[i]==0x15&&(source[i]&0x3f)==0x21) { |
| | 4524 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4525 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4526 | emit_call((int)cvt_d_l); |
| | 4527 | } |
| | 4528 | |
| | 4529 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x21) { |
| | 4530 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4531 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4532 | emit_call((int)cvt_d_s); |
| | 4533 | } |
| | 4534 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x24) { |
| | 4535 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4536 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4537 | emit_call((int)cvt_w_s); |
| | 4538 | } |
| | 4539 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x25) { |
| | 4540 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4541 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4542 | emit_call((int)cvt_l_s); |
| | 4543 | } |
| | 4544 | |
| | 4545 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x20) { |
| | 4546 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4547 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4548 | emit_call((int)cvt_s_d); |
| | 4549 | } |
| | 4550 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x24) { |
| | 4551 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4552 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4553 | emit_call((int)cvt_w_d); |
| | 4554 | } |
| | 4555 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x25) { |
| | 4556 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4557 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4558 | emit_call((int)cvt_l_d); |
| | 4559 | } |
| | 4560 | |
| | 4561 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x08) { |
| | 4562 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4563 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4564 | emit_call((int)round_l_s); |
| | 4565 | } |
| | 4566 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x09) { |
| | 4567 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4568 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4569 | emit_call((int)trunc_l_s); |
| | 4570 | } |
| | 4571 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0a) { |
| | 4572 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4573 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4574 | emit_call((int)ceil_l_s); |
| | 4575 | } |
| | 4576 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0b) { |
| | 4577 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4578 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4579 | emit_call((int)floor_l_s); |
| | 4580 | } |
| | 4581 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0c) { |
| | 4582 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4583 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4584 | emit_call((int)round_w_s); |
| | 4585 | } |
| | 4586 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0d) { |
| | 4587 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4588 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4589 | emit_call((int)trunc_w_s); |
| | 4590 | } |
| | 4591 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0e) { |
| | 4592 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4593 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4594 | emit_call((int)ceil_w_s); |
| | 4595 | } |
| | 4596 | if(opcode2[i]==0x10&&(source[i]&0x3f)==0x0f) { |
| | 4597 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4598 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4599 | emit_call((int)floor_w_s); |
| | 4600 | } |
| | 4601 | |
| | 4602 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x08) { |
| | 4603 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4604 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4605 | emit_call((int)round_l_d); |
| | 4606 | } |
| | 4607 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x09) { |
| | 4608 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4609 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4610 | emit_call((int)trunc_l_d); |
| | 4611 | } |
| | 4612 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0a) { |
| | 4613 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4614 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4615 | emit_call((int)ceil_l_d); |
| | 4616 | } |
| | 4617 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0b) { |
| | 4618 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4619 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4620 | emit_call((int)floor_l_d); |
| | 4621 | } |
| | 4622 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0c) { |
| | 4623 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4624 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4625 | emit_call((int)round_w_d); |
| | 4626 | } |
| | 4627 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0d) { |
| | 4628 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4629 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4630 | emit_call((int)trunc_w_d); |
| | 4631 | } |
| | 4632 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0e) { |
| | 4633 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4634 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4635 | emit_call((int)ceil_w_d); |
| | 4636 | } |
| | 4637 | if(opcode2[i]==0x11&&(source[i]&0x3f)==0x0f) { |
| | 4638 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4639 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4640 | emit_call((int)floor_w_d); |
| | 4641 | } |
| | 4642 | |
| | 4643 | restore_regs(reglist); |
| | 4644 | #else |
| | 4645 | cop1_unusable(i, i_regs); |
| | 4646 | #endif |
| | 4647 | } |
| | 4648 | #define fconv_assemble fconv_assemble_arm |
| | 4649 | |
| | 4650 | void fcomp_assemble(int i,struct regstat *i_regs) |
| | 4651 | { |
| | 4652 | #ifndef DISABLE_COP1 |
| | 4653 | signed char fs=get_reg(i_regs->regmap,FSREG); |
| | 4654 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4655 | assert(temp>=0); |
| | 4656 | // Check cop1 unusable |
| | 4657 | if(!cop1_usable) { |
| | 4658 | signed char cs=get_reg(i_regs->regmap,CSREG); |
| | 4659 | assert(cs>=0); |
| | 4660 | emit_testimm(cs,0x20000000); |
| | 4661 | int jaddr=(int)out; |
| | 4662 | emit_jeq(0); |
| | 4663 | add_stub(FP_STUB,jaddr,(int)out,i,cs,(int)i_regs,is_delayslot,0); |
| | 4664 | cop1_usable=1; |
| | 4665 | } |
| | 4666 | |
| | 4667 | if((source[i]&0x3f)==0x30) { |
| | 4668 | emit_andimm(fs,~0x800000,fs); |
| | 4669 | return; |
| | 4670 | } |
| | 4671 | |
| | 4672 | if((source[i]&0x3e)==0x38) { |
| | 4673 | // sf/ngle - these should throw exceptions for NaNs |
| | 4674 | emit_andimm(fs,~0x800000,fs); |
| | 4675 | return; |
| | 4676 | } |
| | 4677 | |
| | 4678 | #if(defined(__VFP_FP__) && !defined(__SOFTFP__)) |
| | 4679 | if(opcode2[i]==0x10) { |
| | 4680 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4681 | emit_readword((int)®_cop1_simple[(source[i]>>16)&0x1f],HOST_TEMPREG); |
| | 4682 | emit_orimm(fs,0x800000,fs); |
| | 4683 | emit_flds(temp,14); |
| | 4684 | emit_flds(HOST_TEMPREG,15); |
| | 4685 | emit_fcmps(14,15); |
| | 4686 | emit_fmstat(); |
| | 4687 | if((source[i]&0x3f)==0x31) emit_bicvc_imm(fs,0x800000,fs); // c_un_s |
| | 4688 | if((source[i]&0x3f)==0x32) emit_bicne_imm(fs,0x800000,fs); // c_eq_s |
| | 4689 | if((source[i]&0x3f)==0x33) {emit_bicne_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ueq_s |
| | 4690 | if((source[i]&0x3f)==0x34) emit_biccs_imm(fs,0x800000,fs); // c_olt_s |
| | 4691 | if((source[i]&0x3f)==0x35) {emit_biccs_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ult_s |
| | 4692 | if((source[i]&0x3f)==0x36) emit_bichi_imm(fs,0x800000,fs); // c_ole_s |
| | 4693 | if((source[i]&0x3f)==0x37) {emit_bichi_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ule_s |
| | 4694 | if((source[i]&0x3f)==0x3a) emit_bicne_imm(fs,0x800000,fs); // c_seq_s |
| | 4695 | if((source[i]&0x3f)==0x3b) emit_bicne_imm(fs,0x800000,fs); // c_ngl_s |
| | 4696 | if((source[i]&0x3f)==0x3c) emit_biccs_imm(fs,0x800000,fs); // c_lt_s |
| | 4697 | if((source[i]&0x3f)==0x3d) emit_biccs_imm(fs,0x800000,fs); // c_nge_s |
| | 4698 | if((source[i]&0x3f)==0x3e) emit_bichi_imm(fs,0x800000,fs); // c_le_s |
| | 4699 | if((source[i]&0x3f)==0x3f) emit_bichi_imm(fs,0x800000,fs); // c_ngt_s |
| | 4700 | return; |
| | 4701 | } |
| | 4702 | if(opcode2[i]==0x11) { |
| | 4703 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4704 | emit_readword((int)®_cop1_double[(source[i]>>16)&0x1f],HOST_TEMPREG); |
| | 4705 | emit_orimm(fs,0x800000,fs); |
| | 4706 | emit_vldr(temp,6); |
| | 4707 | emit_vldr(HOST_TEMPREG,7); |
| | 4708 | emit_fcmpd(6,7); |
| | 4709 | emit_fmstat(); |
| | 4710 | if((source[i]&0x3f)==0x31) emit_bicvc_imm(fs,0x800000,fs); // c_un_d |
| | 4711 | if((source[i]&0x3f)==0x32) emit_bicne_imm(fs,0x800000,fs); // c_eq_d |
| | 4712 | if((source[i]&0x3f)==0x33) {emit_bicne_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ueq_d |
| | 4713 | if((source[i]&0x3f)==0x34) emit_biccs_imm(fs,0x800000,fs); // c_olt_d |
| | 4714 | if((source[i]&0x3f)==0x35) {emit_biccs_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ult_d |
| | 4715 | if((source[i]&0x3f)==0x36) emit_bichi_imm(fs,0x800000,fs); // c_ole_d |
| | 4716 | if((source[i]&0x3f)==0x37) {emit_bichi_imm(fs,0x800000,fs);emit_orrvs_imm(fs,0x800000,fs);} // c_ule_d |
| | 4717 | if((source[i]&0x3f)==0x3a) emit_bicne_imm(fs,0x800000,fs); // c_seq_d |
| | 4718 | if((source[i]&0x3f)==0x3b) emit_bicne_imm(fs,0x800000,fs); // c_ngl_d |
| | 4719 | if((source[i]&0x3f)==0x3c) emit_biccs_imm(fs,0x800000,fs); // c_lt_d |
| | 4720 | if((source[i]&0x3f)==0x3d) emit_biccs_imm(fs,0x800000,fs); // c_nge_d |
| | 4721 | if((source[i]&0x3f)==0x3e) emit_bichi_imm(fs,0x800000,fs); // c_le_d |
| | 4722 | if((source[i]&0x3f)==0x3f) emit_bichi_imm(fs,0x800000,fs); // c_ngt_d |
| | 4723 | return; |
| | 4724 | } |
| | 4725 | #endif |
| | 4726 | |
| | 4727 | // C only |
| | 4728 | |
| | 4729 | u_int hr,reglist=0; |
| | 4730 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 4731 | if(i_regs->regmap[hr]>=0) reglist|=1<<hr; |
| | 4732 | } |
| | 4733 | reglist&=~(1<<fs); |
| | 4734 | save_regs(reglist); |
| | 4735 | if(opcode2[i]==0x10) { |
| | 4736 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4737 | emit_readword((int)®_cop1_simple[(source[i]>>16)&0x1f],ARG2_REG); |
| | 4738 | if((source[i]&0x3f)==0x30) emit_call((int)c_f_s); |
| | 4739 | if((source[i]&0x3f)==0x31) emit_call((int)c_un_s); |
| | 4740 | if((source[i]&0x3f)==0x32) emit_call((int)c_eq_s); |
| | 4741 | if((source[i]&0x3f)==0x33) emit_call((int)c_ueq_s); |
| | 4742 | if((source[i]&0x3f)==0x34) emit_call((int)c_olt_s); |
| | 4743 | if((source[i]&0x3f)==0x35) emit_call((int)c_ult_s); |
| | 4744 | if((source[i]&0x3f)==0x36) emit_call((int)c_ole_s); |
| | 4745 | if((source[i]&0x3f)==0x37) emit_call((int)c_ule_s); |
| | 4746 | if((source[i]&0x3f)==0x38) emit_call((int)c_sf_s); |
| | 4747 | if((source[i]&0x3f)==0x39) emit_call((int)c_ngle_s); |
| | 4748 | if((source[i]&0x3f)==0x3a) emit_call((int)c_seq_s); |
| | 4749 | if((source[i]&0x3f)==0x3b) emit_call((int)c_ngl_s); |
| | 4750 | if((source[i]&0x3f)==0x3c) emit_call((int)c_lt_s); |
| | 4751 | if((source[i]&0x3f)==0x3d) emit_call((int)c_nge_s); |
| | 4752 | if((source[i]&0x3f)==0x3e) emit_call((int)c_le_s); |
| | 4753 | if((source[i]&0x3f)==0x3f) emit_call((int)c_ngt_s); |
| | 4754 | } |
| | 4755 | if(opcode2[i]==0x11) { |
| | 4756 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4757 | emit_readword((int)®_cop1_double[(source[i]>>16)&0x1f],ARG2_REG); |
| | 4758 | if((source[i]&0x3f)==0x30) emit_call((int)c_f_d); |
| | 4759 | if((source[i]&0x3f)==0x31) emit_call((int)c_un_d); |
| | 4760 | if((source[i]&0x3f)==0x32) emit_call((int)c_eq_d); |
| | 4761 | if((source[i]&0x3f)==0x33) emit_call((int)c_ueq_d); |
| | 4762 | if((source[i]&0x3f)==0x34) emit_call((int)c_olt_d); |
| | 4763 | if((source[i]&0x3f)==0x35) emit_call((int)c_ult_d); |
| | 4764 | if((source[i]&0x3f)==0x36) emit_call((int)c_ole_d); |
| | 4765 | if((source[i]&0x3f)==0x37) emit_call((int)c_ule_d); |
| | 4766 | if((source[i]&0x3f)==0x38) emit_call((int)c_sf_d); |
| | 4767 | if((source[i]&0x3f)==0x39) emit_call((int)c_ngle_d); |
| | 4768 | if((source[i]&0x3f)==0x3a) emit_call((int)c_seq_d); |
| | 4769 | if((source[i]&0x3f)==0x3b) emit_call((int)c_ngl_d); |
| | 4770 | if((source[i]&0x3f)==0x3c) emit_call((int)c_lt_d); |
| | 4771 | if((source[i]&0x3f)==0x3d) emit_call((int)c_nge_d); |
| | 4772 | if((source[i]&0x3f)==0x3e) emit_call((int)c_le_d); |
| | 4773 | if((source[i]&0x3f)==0x3f) emit_call((int)c_ngt_d); |
| | 4774 | } |
| | 4775 | restore_regs(reglist); |
| | 4776 | emit_loadreg(FSREG,fs); |
| | 4777 | #else |
| | 4778 | cop1_unusable(i, i_regs); |
| | 4779 | #endif |
| | 4780 | } |
| | 4781 | |
| | 4782 | void float_assemble(int i,struct regstat *i_regs) |
| | 4783 | { |
| | 4784 | #ifndef DISABLE_COP1 |
| | 4785 | signed char temp=get_reg(i_regs->regmap,-1); |
| | 4786 | assert(temp>=0); |
| | 4787 | // Check cop1 unusable |
| | 4788 | if(!cop1_usable) { |
| | 4789 | signed char cs=get_reg(i_regs->regmap,CSREG); |
| | 4790 | assert(cs>=0); |
| | 4791 | emit_testimm(cs,0x20000000); |
| | 4792 | int jaddr=(int)out; |
| | 4793 | emit_jeq(0); |
| | 4794 | add_stub(FP_STUB,jaddr,(int)out,i,cs,(int)i_regs,is_delayslot,0); |
| | 4795 | cop1_usable=1; |
| | 4796 | } |
| | 4797 | |
| | 4798 | #if(defined(__VFP_FP__) && !defined(__SOFTFP__)) |
| | 4799 | if((source[i]&0x3f)==6) // mov |
| | 4800 | { |
| | 4801 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4802 | if(opcode2[i]==0x10) { |
| | 4803 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4804 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],HOST_TEMPREG); |
| | 4805 | emit_readword_indexed(0,temp,temp); |
| | 4806 | emit_writeword_indexed(temp,0,HOST_TEMPREG); |
| | 4807 | } |
| | 4808 | if(opcode2[i]==0x11) { |
| | 4809 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4810 | emit_readword((int)®_cop1_double[(source[i]>>6)&0x1f],HOST_TEMPREG); |
| | 4811 | emit_vldr(temp,7); |
| | 4812 | emit_vstr(7,HOST_TEMPREG); |
| | 4813 | } |
| | 4814 | } |
| | 4815 | return; |
| | 4816 | } |
| | 4817 | |
| | 4818 | if((source[i]&0x3f)>3) |
| | 4819 | { |
| | 4820 | if(opcode2[i]==0x10) { |
| | 4821 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4822 | emit_flds(temp,15); |
| | 4823 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4824 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4825 | } |
| | 4826 | if((source[i]&0x3f)==4) // sqrt |
| | 4827 | emit_fsqrts(15,15); |
| | 4828 | if((source[i]&0x3f)==5) // abs |
| | 4829 | emit_fabss(15,15); |
| | 4830 | if((source[i]&0x3f)==7) // neg |
| | 4831 | emit_fnegs(15,15); |
| | 4832 | emit_fsts(15,temp); |
| | 4833 | } |
| | 4834 | if(opcode2[i]==0x11) { |
| | 4835 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4836 | emit_vldr(temp,7); |
| | 4837 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4838 | emit_readword((int)®_cop1_double[(source[i]>>6)&0x1f],temp); |
| | 4839 | } |
| | 4840 | if((source[i]&0x3f)==4) // sqrt |
| | 4841 | emit_fsqrtd(7,7); |
| | 4842 | if((source[i]&0x3f)==5) // abs |
| | 4843 | emit_fabsd(7,7); |
| | 4844 | if((source[i]&0x3f)==7) // neg |
| | 4845 | emit_fnegd(7,7); |
| | 4846 | emit_vstr(7,temp); |
| | 4847 | } |
| | 4848 | return; |
| | 4849 | } |
| | 4850 | if((source[i]&0x3f)<4) |
| | 4851 | { |
| | 4852 | if(opcode2[i]==0x10) { |
| | 4853 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],temp); |
| | 4854 | } |
| | 4855 | if(opcode2[i]==0x11) { |
| | 4856 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],temp); |
| | 4857 | } |
| | 4858 | if(((source[i]>>11)&0x1f)!=((source[i]>>16)&0x1f)) { |
| | 4859 | if(opcode2[i]==0x10) { |
| | 4860 | emit_readword((int)®_cop1_simple[(source[i]>>16)&0x1f],HOST_TEMPREG); |
| | 4861 | emit_flds(temp,15); |
| | 4862 | emit_flds(HOST_TEMPREG,13); |
| | 4863 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4864 | if(((source[i]>>16)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4865 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4866 | } |
| | 4867 | } |
| | 4868 | if((source[i]&0x3f)==0) emit_fadds(15,13,15); |
| | 4869 | if((source[i]&0x3f)==1) emit_fsubs(15,13,15); |
| | 4870 | if((source[i]&0x3f)==2) emit_fmuls(15,13,15); |
| | 4871 | if((source[i]&0x3f)==3) emit_fdivs(15,13,15); |
| | 4872 | if(((source[i]>>16)&0x1f)==((source[i]>>6)&0x1f)) { |
| | 4873 | emit_fsts(15,HOST_TEMPREG); |
| | 4874 | }else{ |
| | 4875 | emit_fsts(15,temp); |
| | 4876 | } |
| | 4877 | } |
| | 4878 | else if(opcode2[i]==0x11) { |
| | 4879 | emit_readword((int)®_cop1_double[(source[i]>>16)&0x1f],HOST_TEMPREG); |
| | 4880 | emit_vldr(temp,7); |
| | 4881 | emit_vldr(HOST_TEMPREG,6); |
| | 4882 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4883 | if(((source[i]>>16)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4884 | emit_readword((int)®_cop1_double[(source[i]>>6)&0x1f],temp); |
| | 4885 | } |
| | 4886 | } |
| | 4887 | if((source[i]&0x3f)==0) emit_faddd(7,6,7); |
| | 4888 | if((source[i]&0x3f)==1) emit_fsubd(7,6,7); |
| | 4889 | if((source[i]&0x3f)==2) emit_fmuld(7,6,7); |
| | 4890 | if((source[i]&0x3f)==3) emit_fdivd(7,6,7); |
| | 4891 | if(((source[i]>>16)&0x1f)==((source[i]>>6)&0x1f)) { |
| | 4892 | emit_vstr(7,HOST_TEMPREG); |
| | 4893 | }else{ |
| | 4894 | emit_vstr(7,temp); |
| | 4895 | } |
| | 4896 | } |
| | 4897 | } |
| | 4898 | else { |
| | 4899 | if(opcode2[i]==0x10) { |
| | 4900 | emit_flds(temp,15); |
| | 4901 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4902 | emit_readword((int)®_cop1_simple[(source[i]>>6)&0x1f],temp); |
| | 4903 | } |
| | 4904 | if((source[i]&0x3f)==0) emit_fadds(15,15,15); |
| | 4905 | if((source[i]&0x3f)==1) emit_fsubs(15,15,15); |
| | 4906 | if((source[i]&0x3f)==2) emit_fmuls(15,15,15); |
| | 4907 | if((source[i]&0x3f)==3) emit_fdivs(15,15,15); |
| | 4908 | emit_fsts(15,temp); |
| | 4909 | } |
| | 4910 | else if(opcode2[i]==0x11) { |
| | 4911 | emit_vldr(temp,7); |
| | 4912 | if(((source[i]>>11)&0x1f)!=((source[i]>>6)&0x1f)) { |
| | 4913 | emit_readword((int)®_cop1_double[(source[i]>>6)&0x1f],temp); |
| | 4914 | } |
| | 4915 | if((source[i]&0x3f)==0) emit_faddd(7,7,7); |
| | 4916 | if((source[i]&0x3f)==1) emit_fsubd(7,7,7); |
| | 4917 | if((source[i]&0x3f)==2) emit_fmuld(7,7,7); |
| | 4918 | if((source[i]&0x3f)==3) emit_fdivd(7,7,7); |
| | 4919 | emit_vstr(7,temp); |
| | 4920 | } |
| | 4921 | } |
| | 4922 | return; |
| | 4923 | } |
| | 4924 | #endif |
| | 4925 | |
| | 4926 | u_int hr,reglist=0; |
| | 4927 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 4928 | if(i_regs->regmap[hr]>=0) reglist|=1<<hr; |
| | 4929 | } |
| | 4930 | if(opcode2[i]==0x10) { // Single precision |
| | 4931 | save_regs(reglist); |
| | 4932 | emit_readword((int)®_cop1_simple[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4933 | if((source[i]&0x3f)<4) { |
| | 4934 | emit_readword((int)®_cop1_simple[(source[i]>>16)&0x1f],ARG2_REG); |
| | 4935 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG3_REG); |
| | 4936 | }else{ |
| | 4937 | emit_readword((int)®_cop1_simple[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4938 | } |
| | 4939 | switch(source[i]&0x3f) |
| | 4940 | { |
| | 4941 | case 0x00: emit_call((int)add_s);break; |
| | 4942 | case 0x01: emit_call((int)sub_s);break; |
| | 4943 | case 0x02: emit_call((int)mul_s);break; |
| | 4944 | case 0x03: emit_call((int)div_s);break; |
| | 4945 | case 0x04: emit_call((int)sqrt_s);break; |
| | 4946 | case 0x05: emit_call((int)abs_s);break; |
| | 4947 | case 0x06: emit_call((int)mov_s);break; |
| | 4948 | case 0x07: emit_call((int)neg_s);break; |
| | 4949 | } |
| | 4950 | restore_regs(reglist); |
| | 4951 | } |
| | 4952 | if(opcode2[i]==0x11) { // Double precision |
| | 4953 | save_regs(reglist); |
| | 4954 | emit_readword((int)®_cop1_double[(source[i]>>11)&0x1f],ARG1_REG); |
| | 4955 | if((source[i]&0x3f)<4) { |
| | 4956 | emit_readword((int)®_cop1_double[(source[i]>>16)&0x1f],ARG2_REG); |
| | 4957 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG3_REG); |
| | 4958 | }else{ |
| | 4959 | emit_readword((int)®_cop1_double[(source[i]>> 6)&0x1f],ARG2_REG); |
| | 4960 | } |
| | 4961 | switch(source[i]&0x3f) |
| | 4962 | { |
| | 4963 | case 0x00: emit_call((int)add_d);break; |
| | 4964 | case 0x01: emit_call((int)sub_d);break; |
| | 4965 | case 0x02: emit_call((int)mul_d);break; |
| | 4966 | case 0x03: emit_call((int)div_d);break; |
| | 4967 | case 0x04: emit_call((int)sqrt_d);break; |
| | 4968 | case 0x05: emit_call((int)abs_d);break; |
| | 4969 | case 0x06: emit_call((int)mov_d);break; |
| | 4970 | case 0x07: emit_call((int)neg_d);break; |
| | 4971 | } |
| | 4972 | restore_regs(reglist); |
| | 4973 | } |
| | 4974 | #else |
| | 4975 | cop1_unusable(i, i_regs); |
| | 4976 | #endif |
| | 4977 | } |
| | 4978 | |
| | 4979 | void multdiv_assemble_arm(int i,struct regstat *i_regs) |
| | 4980 | { |
| | 4981 | // case 0x18: MULT |
| | 4982 | // case 0x19: MULTU |
| | 4983 | // case 0x1A: DIV |
| | 4984 | // case 0x1B: DIVU |
| | 4985 | // case 0x1C: DMULT |
| | 4986 | // case 0x1D: DMULTU |
| | 4987 | // case 0x1E: DDIV |
| | 4988 | // case 0x1F: DDIVU |
| | 4989 | if(rs1[i]&&rs2[i]) |
| | 4990 | { |
| | 4991 | if((opcode2[i]&4)==0) // 32-bit |
| | 4992 | { |
| | 4993 | if(opcode2[i]==0x18) // MULT |
| | 4994 | { |
| | 4995 | signed char m1=get_reg(i_regs->regmap,rs1[i]); |
| | 4996 | signed char m2=get_reg(i_regs->regmap,rs2[i]); |
| | 4997 | signed char hi=get_reg(i_regs->regmap,HIREG); |
| | 4998 | signed char lo=get_reg(i_regs->regmap,LOREG); |
| | 4999 | assert(m1>=0); |
| | 5000 | assert(m2>=0); |
| | 5001 | assert(hi>=0); |
| | 5002 | assert(lo>=0); |
| | 5003 | emit_smull(m1,m2,hi,lo); |
| | 5004 | } |
| | 5005 | if(opcode2[i]==0x19) // MULTU |
| | 5006 | { |
| | 5007 | signed char m1=get_reg(i_regs->regmap,rs1[i]); |
| | 5008 | signed char m2=get_reg(i_regs->regmap,rs2[i]); |
| | 5009 | signed char hi=get_reg(i_regs->regmap,HIREG); |
| | 5010 | signed char lo=get_reg(i_regs->regmap,LOREG); |
| | 5011 | assert(m1>=0); |
| | 5012 | assert(m2>=0); |
| | 5013 | assert(hi>=0); |
| | 5014 | assert(lo>=0); |
| | 5015 | emit_umull(m1,m2,hi,lo); |
| | 5016 | } |
| | 5017 | if(opcode2[i]==0x1A) // DIV |
| | 5018 | { |
| | 5019 | signed char d1=get_reg(i_regs->regmap,rs1[i]); |
| | 5020 | signed char d2=get_reg(i_regs->regmap,rs2[i]); |
| | 5021 | assert(d1>=0); |
| | 5022 | assert(d2>=0); |
| | 5023 | signed char quotient=get_reg(i_regs->regmap,LOREG); |
| | 5024 | signed char remainder=get_reg(i_regs->regmap,HIREG); |
| | 5025 | assert(quotient>=0); |
| | 5026 | assert(remainder>=0); |
| | 5027 | emit_movs(d1,remainder); |
| | 5028 | emit_movimm(0xffffffff,quotient); |
| | 5029 | emit_negmi(quotient,quotient); // .. quotient and .. |
| | 5030 | emit_negmi(remainder,remainder); // .. remainder for div0 case (will be negated back after jump) |
| | 5031 | emit_movs(d2,HOST_TEMPREG); |
| | 5032 | emit_jeq((int)out+52); // Division by zero |
| | 5033 | emit_negmi(HOST_TEMPREG,HOST_TEMPREG); |
| | 5034 | emit_clz(HOST_TEMPREG,quotient); |
| | 5035 | emit_shl(HOST_TEMPREG,quotient,HOST_TEMPREG); |
| | 5036 | emit_orimm(quotient,1<<31,quotient); |
| | 5037 | emit_shr(quotient,quotient,quotient); |
| | 5038 | emit_cmp(remainder,HOST_TEMPREG); |
| | 5039 | emit_subcs(remainder,HOST_TEMPREG,remainder); |
| | 5040 | emit_adcs(quotient,quotient,quotient); |
| | 5041 | emit_shrimm(HOST_TEMPREG,1,HOST_TEMPREG); |
| | 5042 | emit_jcc((int)out-16); // -4 |
| | 5043 | emit_teq(d1,d2); |
| | 5044 | emit_negmi(quotient,quotient); |
| | 5045 | emit_test(d1,d1); |
| | 5046 | emit_negmi(remainder,remainder); |
| | 5047 | } |
| | 5048 | if(opcode2[i]==0x1B) // DIVU |
| | 5049 | { |
| | 5050 | signed char d1=get_reg(i_regs->regmap,rs1[i]); // dividend |
| | 5051 | signed char d2=get_reg(i_regs->regmap,rs2[i]); // divisor |
| | 5052 | assert(d1>=0); |
| | 5053 | assert(d2>=0); |
| | 5054 | signed char quotient=get_reg(i_regs->regmap,LOREG); |
| | 5055 | signed char remainder=get_reg(i_regs->regmap,HIREG); |
| | 5056 | assert(quotient>=0); |
| | 5057 | assert(remainder>=0); |
| | 5058 | emit_mov(d1,remainder); |
| | 5059 | emit_movimm(0xffffffff,quotient); // div0 case |
| | 5060 | emit_test(d2,d2); |
| | 5061 | emit_jeq((int)out+40); // Division by zero |
| | 5062 | emit_clz(d2,HOST_TEMPREG); |
| | 5063 | emit_movimm(1<<31,quotient); |
| | 5064 | emit_shl(d2,HOST_TEMPREG,d2); |
| | 5065 | emit_shr(quotient,HOST_TEMPREG,quotient); |
| | 5066 | emit_cmp(remainder,d2); |
| | 5067 | emit_subcs(remainder,d2,remainder); |
| | 5068 | emit_adcs(quotient,quotient,quotient); |
| | 5069 | emit_shrcc_imm(d2,1,d2); |
| | 5070 | emit_jcc((int)out-16); // -4 |
| | 5071 | } |
| | 5072 | } |
| | 5073 | else // 64-bit |
| | 5074 | #ifndef FORCE32 |
| | 5075 | { |
| | 5076 | if(opcode2[i]==0x1C) // DMULT |
| | 5077 | { |
| | 5078 | assert(opcode2[i]!=0x1C); |
| | 5079 | signed char m1h=get_reg(i_regs->regmap,rs1[i]|64); |
| | 5080 | signed char m1l=get_reg(i_regs->regmap,rs1[i]); |
| | 5081 | signed char m2h=get_reg(i_regs->regmap,rs2[i]|64); |
| | 5082 | signed char m2l=get_reg(i_regs->regmap,rs2[i]); |
| | 5083 | assert(m1h>=0); |
| | 5084 | assert(m2h>=0); |
| | 5085 | assert(m1l>=0); |
| | 5086 | assert(m2l>=0); |
| | 5087 | emit_pushreg(m2h); |
| | 5088 | emit_pushreg(m2l); |
| | 5089 | emit_pushreg(m1h); |
| | 5090 | emit_pushreg(m1l); |
| | 5091 | emit_call((int)&mult64); |
| | 5092 | emit_popreg(m1l); |
| | 5093 | emit_popreg(m1h); |
| | 5094 | emit_popreg(m2l); |
| | 5095 | emit_popreg(m2h); |
| | 5096 | signed char hih=get_reg(i_regs->regmap,HIREG|64); |
| | 5097 | signed char hil=get_reg(i_regs->regmap,HIREG); |
| | 5098 | if(hih>=0) emit_loadreg(HIREG|64,hih); |
| | 5099 | if(hil>=0) emit_loadreg(HIREG,hil); |
| | 5100 | signed char loh=get_reg(i_regs->regmap,LOREG|64); |
| | 5101 | signed char lol=get_reg(i_regs->regmap,LOREG); |
| | 5102 | if(loh>=0) emit_loadreg(LOREG|64,loh); |
| | 5103 | if(lol>=0) emit_loadreg(LOREG,lol); |
| | 5104 | } |
| | 5105 | if(opcode2[i]==0x1D) // DMULTU |
| | 5106 | { |
| | 5107 | signed char m1h=get_reg(i_regs->regmap,rs1[i]|64); |
| | 5108 | signed char m1l=get_reg(i_regs->regmap,rs1[i]); |
| | 5109 | signed char m2h=get_reg(i_regs->regmap,rs2[i]|64); |
| | 5110 | signed char m2l=get_reg(i_regs->regmap,rs2[i]); |
| | 5111 | assert(m1h>=0); |
| | 5112 | assert(m2h>=0); |
| | 5113 | assert(m1l>=0); |
| | 5114 | assert(m2l>=0); |
| | 5115 | save_regs(0x100f); |
| | 5116 | if(m1l!=0) emit_mov(m1l,0); |
| | 5117 | if(m1h==0) emit_readword((int)&dynarec_local,1); |
| | 5118 | else if(m1h>1) emit_mov(m1h,1); |
| | 5119 | if(m2l<2) emit_readword((int)&dynarec_local+m2l*4,2); |
| | 5120 | else if(m2l>2) emit_mov(m2l,2); |
| | 5121 | if(m2h<3) emit_readword((int)&dynarec_local+m2h*4,3); |
| | 5122 | else if(m2h>3) emit_mov(m2h,3); |
| | 5123 | emit_call((int)&multu64); |
| | 5124 | restore_regs(0x100f); |
| | 5125 | signed char hih=get_reg(i_regs->regmap,HIREG|64); |
| | 5126 | signed char hil=get_reg(i_regs->regmap,HIREG); |
| | 5127 | signed char loh=get_reg(i_regs->regmap,LOREG|64); |
| | 5128 | signed char lol=get_reg(i_regs->regmap,LOREG); |
| | 5129 | /*signed char temp=get_reg(i_regs->regmap,-1); |
| | 5130 | signed char rh=get_reg(i_regs->regmap,HIREG|64); |
| | 5131 | signed char rl=get_reg(i_regs->regmap,HIREG); |
| | 5132 | assert(m1h>=0); |
| | 5133 | assert(m2h>=0); |
| | 5134 | assert(m1l>=0); |
| | 5135 | assert(m2l>=0); |
| | 5136 | assert(temp>=0); |
| | 5137 | //emit_mov(m1l,EAX); |
| | 5138 | //emit_mul(m2l); |
| | 5139 | emit_umull(rl,rh,m1l,m2l); |
| | 5140 | emit_storereg(LOREG,rl); |
| | 5141 | emit_mov(rh,temp); |
| | 5142 | //emit_mov(m1h,EAX); |
| | 5143 | //emit_mul(m2l); |
| | 5144 | emit_umull(rl,rh,m1h,m2l); |
| | 5145 | emit_adds(rl,temp,temp); |
| | 5146 | emit_adcimm(rh,0,rh); |
| | 5147 | emit_storereg(HIREG,rh); |
| | 5148 | //emit_mov(m2h,EAX); |
| | 5149 | //emit_mul(m1l); |
| | 5150 | emit_umull(rl,rh,m1l,m2h); |
| | 5151 | emit_adds(rl,temp,temp); |
| | 5152 | emit_adcimm(rh,0,rh); |
| | 5153 | emit_storereg(LOREG|64,temp); |
| | 5154 | emit_mov(rh,temp); |
| | 5155 | //emit_mov(m2h,EAX); |
| | 5156 | //emit_mul(m1h); |
| | 5157 | emit_umull(rl,rh,m1h,m2h); |
| | 5158 | emit_adds(rl,temp,rl); |
| | 5159 | emit_loadreg(HIREG,temp); |
| | 5160 | emit_adcimm(rh,0,rh); |
| | 5161 | emit_adds(rl,temp,rl); |
| | 5162 | emit_adcimm(rh,0,rh); |
| | 5163 | // DEBUG |
| | 5164 | /* |
| | 5165 | emit_pushreg(m2h); |
| | 5166 | emit_pushreg(m2l); |
| | 5167 | emit_pushreg(m1h); |
| | 5168 | emit_pushreg(m1l); |
| | 5169 | emit_call((int)&multu64); |
| | 5170 | emit_popreg(m1l); |
| | 5171 | emit_popreg(m1h); |
| | 5172 | emit_popreg(m2l); |
| | 5173 | emit_popreg(m2h); |
| | 5174 | signed char hih=get_reg(i_regs->regmap,HIREG|64); |
| | 5175 | signed char hil=get_reg(i_regs->regmap,HIREG); |
| | 5176 | if(hih>=0) emit_loadreg(HIREG|64,hih); // DEBUG |
| | 5177 | if(hil>=0) emit_loadreg(HIREG,hil); // DEBUG |
| | 5178 | */ |
| | 5179 | // Shouldn't be necessary |
| | 5180 | //char loh=get_reg(i_regs->regmap,LOREG|64); |
| | 5181 | //char lol=get_reg(i_regs->regmap,LOREG); |
| | 5182 | //if(loh>=0) emit_loadreg(LOREG|64,loh); |
| | 5183 | //if(lol>=0) emit_loadreg(LOREG,lol); |
| | 5184 | } |
| | 5185 | if(opcode2[i]==0x1E) // DDIV |
| | 5186 | { |
| | 5187 | signed char d1h=get_reg(i_regs->regmap,rs1[i]|64); |
| | 5188 | signed char d1l=get_reg(i_regs->regmap,rs1[i]); |
| | 5189 | signed char d2h=get_reg(i_regs->regmap,rs2[i]|64); |
| | 5190 | signed char d2l=get_reg(i_regs->regmap,rs2[i]); |
| | 5191 | assert(d1h>=0); |
| | 5192 | assert(d2h>=0); |
| | 5193 | assert(d1l>=0); |
| | 5194 | assert(d2l>=0); |
| | 5195 | save_regs(0x100f); |
| | 5196 | if(d1l!=0) emit_mov(d1l,0); |
| | 5197 | if(d1h==0) emit_readword((int)&dynarec_local,1); |
| | 5198 | else if(d1h>1) emit_mov(d1h,1); |
| | 5199 | if(d2l<2) emit_readword((int)&dynarec_local+d2l*4,2); |
| | 5200 | else if(d2l>2) emit_mov(d2l,2); |
| | 5201 | if(d2h<3) emit_readword((int)&dynarec_local+d2h*4,3); |
| | 5202 | else if(d2h>3) emit_mov(d2h,3); |
| | 5203 | emit_call((int)&div64); |
| | 5204 | restore_regs(0x100f); |
| | 5205 | signed char hih=get_reg(i_regs->regmap,HIREG|64); |
| | 5206 | signed char hil=get_reg(i_regs->regmap,HIREG); |
| | 5207 | signed char loh=get_reg(i_regs->regmap,LOREG|64); |
| | 5208 | signed char lol=get_reg(i_regs->regmap,LOREG); |
| | 5209 | if(hih>=0) emit_loadreg(HIREG|64,hih); |
| | 5210 | if(hil>=0) emit_loadreg(HIREG,hil); |
| | 5211 | if(loh>=0) emit_loadreg(LOREG|64,loh); |
| | 5212 | if(lol>=0) emit_loadreg(LOREG,lol); |
| | 5213 | } |
| | 5214 | if(opcode2[i]==0x1F) // DDIVU |
| | 5215 | { |
| | 5216 | //u_int hr,reglist=0; |
| | 5217 | //for(hr=0;hr<HOST_REGS;hr++) { |
| | 5218 | // if(i_regs->regmap[hr]>=0 && (i_regs->regmap[hr]&62)!=HIREG) reglist|=1<<hr; |
| | 5219 | //} |
| | 5220 | signed char d1h=get_reg(i_regs->regmap,rs1[i]|64); |
| | 5221 | signed char d1l=get_reg(i_regs->regmap,rs1[i]); |
| | 5222 | signed char d2h=get_reg(i_regs->regmap,rs2[i]|64); |
| | 5223 | signed char d2l=get_reg(i_regs->regmap,rs2[i]); |
| | 5224 | assert(d1h>=0); |
| | 5225 | assert(d2h>=0); |
| | 5226 | assert(d1l>=0); |
| | 5227 | assert(d2l>=0); |
| | 5228 | save_regs(0x100f); |
| | 5229 | if(d1l!=0) emit_mov(d1l,0); |
| | 5230 | if(d1h==0) emit_readword((int)&dynarec_local,1); |
| | 5231 | else if(d1h>1) emit_mov(d1h,1); |
| | 5232 | if(d2l<2) emit_readword((int)&dynarec_local+d2l*4,2); |
| | 5233 | else if(d2l>2) emit_mov(d2l,2); |
| | 5234 | if(d2h<3) emit_readword((int)&dynarec_local+d2h*4,3); |
| | 5235 | else if(d2h>3) emit_mov(d2h,3); |
| | 5236 | emit_call((int)&divu64); |
| | 5237 | restore_regs(0x100f); |
| | 5238 | signed char hih=get_reg(i_regs->regmap,HIREG|64); |
| | 5239 | signed char hil=get_reg(i_regs->regmap,HIREG); |
| | 5240 | signed char loh=get_reg(i_regs->regmap,LOREG|64); |
| | 5241 | signed char lol=get_reg(i_regs->regmap,LOREG); |
| | 5242 | if(hih>=0) emit_loadreg(HIREG|64,hih); |
| | 5243 | if(hil>=0) emit_loadreg(HIREG,hil); |
| | 5244 | if(loh>=0) emit_loadreg(LOREG|64,loh); |
| | 5245 | if(lol>=0) emit_loadreg(LOREG,lol); |
| | 5246 | } |
| | 5247 | } |
| | 5248 | #else |
| | 5249 | assert(0); |
| | 5250 | #endif |
| | 5251 | } |
| | 5252 | else |
| | 5253 | { |
| | 5254 | // Multiply by zero is zero. |
| | 5255 | // MIPS does not have a divide by zero exception. |
| | 5256 | // The result is undefined, we return zero. |
| | 5257 | signed char hr=get_reg(i_regs->regmap,HIREG); |
| | 5258 | signed char lr=get_reg(i_regs->regmap,LOREG); |
| | 5259 | if(hr>=0) emit_zeroreg(hr); |
| | 5260 | if(lr>=0) emit_zeroreg(lr); |
| | 5261 | } |
| | 5262 | } |
| | 5263 | #define multdiv_assemble multdiv_assemble_arm |
| | 5264 | |
| | 5265 | void do_preload_rhash(int r) { |
| | 5266 | // Don't need this for ARM. On x86, this puts the value 0xf8 into the |
| | 5267 | // register. On ARM the hash can be done with a single instruction (below) |
| | 5268 | } |
| | 5269 | |
| | 5270 | void do_preload_rhtbl(int ht) { |
| | 5271 | emit_addimm(FP,(int)&mini_ht-(int)&dynarec_local,ht); |
| | 5272 | } |
| | 5273 | |
| | 5274 | void do_rhash(int rs,int rh) { |
| | 5275 | emit_andimm(rs,0xf8,rh); |
| | 5276 | } |
| | 5277 | |
| | 5278 | void do_miniht_load(int ht,int rh) { |
| | 5279 | assem_debug("ldr %s,[%s,%s]!\n",regname[rh],regname[ht],regname[rh]); |
| | 5280 | output_w32(0xe7b00000|rd_rn_rm(rh,ht,rh)); |
| | 5281 | } |
| | 5282 | |
| | 5283 | void do_miniht_jump(int rs,int rh,int ht) { |
| | 5284 | emit_cmp(rh,rs); |
| | 5285 | emit_ldreq_indexed(ht,4,15); |
| | 5286 | #ifdef CORTEX_A8_BRANCH_PREDICTION_HACK |
| | 5287 | emit_mov(rs,7); |
| | 5288 | emit_jmp(jump_vaddr_reg[7]); |
| | 5289 | #else |
| | 5290 | emit_jmp(jump_vaddr_reg[rs]); |
| | 5291 | #endif |
| | 5292 | } |
| | 5293 | |
| | 5294 | void do_miniht_insert(u_int return_address,int rt,int temp) { |
| | 5295 | #ifdef ARMv5_ONLY |
| | 5296 | emit_movimm(return_address,rt); // PC into link register |
| | 5297 | add_to_linker((int)out,return_address,1); |
| | 5298 | emit_pcreladdr(temp); |
| | 5299 | emit_writeword(rt,(int)&mini_ht[(return_address&0xFF)>>3][0]); |
| | 5300 | emit_writeword(temp,(int)&mini_ht[(return_address&0xFF)>>3][1]); |
| | 5301 | #else |
| | 5302 | emit_movw(return_address&0x0000FFFF,rt); |
| | 5303 | add_to_linker((int)out,return_address,1); |
| | 5304 | emit_pcreladdr(temp); |
| | 5305 | emit_writeword(temp,(int)&mini_ht[(return_address&0xFF)>>3][1]); |
| | 5306 | emit_movt(return_address&0xFFFF0000,rt); |
| | 5307 | emit_writeword(rt,(int)&mini_ht[(return_address&0xFF)>>3][0]); |
| | 5308 | #endif |
| | 5309 | } |
| | 5310 | |
| | 5311 | // Sign-extend to 64 bits and write out upper half of a register |
| | 5312 | // This is useful where we have a 32-bit value in a register, and want to |
| | 5313 | // keep it in a 32-bit register, but can't guarantee that it won't be read |
| | 5314 | // as a 64-bit value later. |
| | 5315 | void wb_sx(signed char pre[],signed char entry[],uint64_t dirty,uint64_t is32_pre,uint64_t is32,uint64_t u,uint64_t uu) |
| | 5316 | { |
| | 5317 | #ifndef FORCE32 |
| | 5318 | if(is32_pre==is32) return; |
| | 5319 | int hr,reg; |
| | 5320 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 5321 | if(hr!=EXCLUDE_REG) { |
| | 5322 | //if(pre[hr]==entry[hr]) { |
| | 5323 | if((reg=pre[hr])>=0) { |
| | 5324 | if((dirty>>hr)&1) { |
| | 5325 | if( ((is32_pre&~is32&~uu)>>reg)&1 ) { |
| | 5326 | emit_sarimm(hr,31,HOST_TEMPREG); |
| | 5327 | emit_storereg(reg|64,HOST_TEMPREG); |
| | 5328 | } |
| | 5329 | } |
| | 5330 | } |
| | 5331 | //} |
| | 5332 | } |
| | 5333 | } |
| | 5334 | #endif |
| | 5335 | } |
| | 5336 | |
| | 5337 | void wb_valid(signed char pre[],signed char entry[],u_int dirty_pre,u_int dirty,uint64_t is32_pre,uint64_t u,uint64_t uu) |
| | 5338 | { |
| | 5339 | //if(dirty_pre==dirty) return; |
| | 5340 | int hr,reg,new_hr; |
| | 5341 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 5342 | if(hr!=EXCLUDE_REG) { |
| | 5343 | reg=pre[hr]; |
| | 5344 | if(((~u)>>(reg&63))&1) { |
| | 5345 | if(reg>0) { |
| | 5346 | if(((dirty_pre&~dirty)>>hr)&1) { |
| | 5347 | if(reg>0&®<34) { |
| | 5348 | emit_storereg(reg,hr); |
| | 5349 | if( ((is32_pre&~uu)>>reg)&1 ) { |
| | 5350 | emit_sarimm(hr,31,HOST_TEMPREG); |
| | 5351 | emit_storereg(reg|64,HOST_TEMPREG); |
| | 5352 | } |
| | 5353 | } |
| | 5354 | else if(reg>=64) { |
| | 5355 | emit_storereg(reg,hr); |
| | 5356 | } |
| | 5357 | } |
| | 5358 | } |
| | 5359 | } |
| | 5360 | } |
| | 5361 | } |
| | 5362 | } |
| | 5363 | |
| | 5364 | |
| | 5365 | /* using strd could possibly help but you'd have to allocate registers in pairs |
| | 5366 | void wb_invalidate_arm(signed char pre[],signed char entry[],uint64_t dirty,uint64_t is32,uint64_t u,uint64_t uu) |
| | 5367 | { |
| | 5368 | int hr; |
| | 5369 | int wrote=-1; |
| | 5370 | for(hr=HOST_REGS-1;hr>=0;hr--) { |
| | 5371 | if(hr!=EXCLUDE_REG) { |
| | 5372 | if(pre[hr]!=entry[hr]) { |
| | 5373 | if(pre[hr]>=0) { |
| | 5374 | if((dirty>>hr)&1) { |
| | 5375 | if(get_reg(entry,pre[hr])<0) { |
| | 5376 | if(pre[hr]<64) { |
| | 5377 | if(!((u>>pre[hr])&1)) { |
| | 5378 | if(hr<10&&(~hr&1)&&(pre[hr+1]<0||wrote==hr+1)) { |
| | 5379 | if( ((is32>>pre[hr])&1) && !((uu>>pre[hr])&1) ) { |
| | 5380 | emit_sarimm(hr,31,hr+1); |
| | 5381 | emit_strdreg(pre[hr],hr); |
| | 5382 | } |
| | 5383 | else |
| | 5384 | emit_storereg(pre[hr],hr); |
| | 5385 | }else{ |
| | 5386 | emit_storereg(pre[hr],hr); |
| | 5387 | if( ((is32>>pre[hr])&1) && !((uu>>pre[hr])&1) ) { |
| | 5388 | emit_sarimm(hr,31,hr); |
| | 5389 | emit_storereg(pre[hr]|64,hr); |
| | 5390 | } |
| | 5391 | } |
| | 5392 | } |
| | 5393 | }else{ |
| | 5394 | if(!((uu>>(pre[hr]&63))&1) && !((is32>>(pre[hr]&63))&1)) { |
| | 5395 | emit_storereg(pre[hr],hr); |
| | 5396 | } |
| | 5397 | } |
| | 5398 | wrote=hr; |
| | 5399 | } |
| | 5400 | } |
| | 5401 | } |
| | 5402 | } |
| | 5403 | } |
| | 5404 | } |
| | 5405 | for(hr=0;hr<HOST_REGS;hr++) { |
| | 5406 | if(hr!=EXCLUDE_REG) { |
| | 5407 | if(pre[hr]!=entry[hr]) { |
| | 5408 | if(pre[hr]>=0) { |
| | 5409 | int nr; |
| | 5410 | if((nr=get_reg(entry,pre[hr]))>=0) { |
| | 5411 | emit_mov(hr,nr); |
| | 5412 | } |
| | 5413 | } |
| | 5414 | } |
| | 5415 | } |
| | 5416 | } |
| | 5417 | } |
| | 5418 | #define wb_invalidate wb_invalidate_arm |
| | 5419 | */ |
| | 5420 | |
| | 5421 | // Clearing the cache is rather slow on ARM Linux, so mark the areas |
| | 5422 | // that need to be cleared, and then only clear these areas once. |
| | 5423 | void do_clear_cache() |
| | 5424 | { |
| | 5425 | int i,j; |
| | 5426 | for (i=0;i<(1<<(TARGET_SIZE_2-17));i++) |
| | 5427 | { |
| | 5428 | u_int bitmap=needs_clear_cache[i]; |
| | 5429 | if(bitmap) { |
| | 5430 | u_int start,end; |
| | 5431 | for(j=0;j<32;j++) |
| | 5432 | { |
| | 5433 | if(bitmap&(1<<j)) { |
| | 5434 | start=BASE_ADDR+i*131072+j*4096; |
| | 5435 | end=start+4095; |
| | 5436 | j++; |
| | 5437 | while(j<32) { |
| | 5438 | if(bitmap&(1<<j)) { |
| | 5439 | end+=4096; |
| | 5440 | j++; |
| | 5441 | }else{ |
| | 5442 | __clear_cache((void *)start,(void *)end); |
| | 5443 | break; |
| | 5444 | } |
| | 5445 | } |
| | 5446 | } |
| | 5447 | } |
| | 5448 | needs_clear_cache[i]=0; |
| | 5449 | } |
| | 5450 | } |
| | 5451 | } |
| | 5452 | |
| | 5453 | // CPU-architecture-specific initialization |
| | 5454 | void arch_init() { |
| | 5455 | #ifndef DISABLE_COP1 |
| | 5456 | rounding_modes[0]=0x0<<22; // round |
| | 5457 | rounding_modes[1]=0x3<<22; // trunc |
| | 5458 | rounding_modes[2]=0x1<<22; // ceil |
| | 5459 | rounding_modes[3]=0x2<<22; // floor |
| | 5460 | #endif |
| | 5461 | } |
| | 5462 | |
| | 5463 | // vim:shiftwidth=2:expandtab |
notaz.gp2x.de / pcsx_rearmed.git / blame_incremental