1 /* 2 * Copyright 2011 Christoph Bumiller 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 */ 22 23 #ifndef __NV50_IR_H__ 24 #define __NV50_IR_H__ 25 26 #include <stdio.h> 27 #include <stdlib.h> 28 #include <stdint.h> 29 #include <deque> 30 #include <list> 31 #include <unordered_set> 32 #include <vector> 33 34 #include "nv50_ir_util.h" 35 #include "nv50_ir_graph.h" 36 37 #include "nv50_ir_driver.h" 38 39 namespace nv50_ir { 40 41 enum operation 42 { 43 OP_NOP = 0, 44 OP_PHI, 45 OP_UNION, // unify a new definition and several source values 46 OP_SPLIT, // $r0d -> { $r0, $r1 } ($r0d and $r0/$r1 will be coalesced) 47 OP_MERGE, // opposite of split, e.g. combine 2 32 bit into a 64 bit value 48 OP_CONSTRAINT, // copy values into consecutive registers 49 OP_MOV, // simple copy, no modifiers allowed 50 OP_LOAD, 51 OP_STORE, 52 OP_ADD, // NOTE: add u64 + u32 is legal for targets w/o 64-bit integer adds 53 OP_SUB, 54 OP_MUL, 55 OP_DIV, 56 OP_MOD, 57 OP_MAD, 58 OP_FMA, 59 OP_SAD, // abs(src0 - src1) + src2 60 OP_SHLADD, 61 // extended multiply-add (GM107+), does a lot of things. 62 // see envytools for detailed documentation 63 OP_XMAD, 64 OP_ABS, 65 OP_NEG, 66 OP_NOT, 67 OP_AND, 68 OP_OR, 69 OP_XOR, 70 OP_LOP3_LUT, 71 OP_SHL, 72 OP_SHR, 73 OP_SHF, 74 OP_MAX, 75 OP_MIN, 76 OP_SAT, // CLAMP(f32, 0.0, 1.0) 77 OP_CEIL, 78 OP_FLOOR, 79 OP_TRUNC, 80 OP_CVT, 81 OP_SET_AND, // dst = (src0 CMP src1) & src2 82 OP_SET_OR, 83 OP_SET_XOR, 84 OP_SET, 85 OP_SELP, // dst = src2 ? src0 : src1 86 OP_SLCT, // dst = (src2 CMP 0) ? src0 : src1 87 OP_RCP, 88 OP_RSQ, 89 OP_LG2, 90 OP_SIN, 91 OP_COS, 92 OP_EX2, 93 OP_EXP, // exponential (base M_E) 94 OP_LOG, // natural logarithm 95 OP_PRESIN, 96 OP_PREEX2, 97 OP_SQRT, 98 OP_POW, 99 OP_BRA, 100 OP_CALL, 101 OP_RET, 102 OP_CONT, 103 OP_BREAK, 104 OP_PRERET, 105 OP_PRECONT, 106 OP_PREBREAK, 107 OP_BRKPT, // breakpoint (not related to loops) 108 OP_JOINAT, // push control flow convergence point 109 OP_JOIN, // converge 110 OP_DISCARD, 111 OP_EXIT, 112 OP_MEMBAR, // memory barrier (mfence, lfence, sfence) 113 OP_VFETCH, // indirection 0 in attribute space, indirection 1 is vertex base 114 OP_PFETCH, // fetch base address of vertex src0 (immediate) [+ src1] 115 OP_AFETCH, // fetch base address of shader input (a[%r1+0x10]) 116 OP_EXPORT, 117 OP_LINTERP, 118 OP_PINTERP, 119 OP_EMIT, // emit vertex 120 OP_RESTART, // restart primitive 121 OP_FINAL, // finish emitting primitives 122 OP_TEX, 123 OP_TXB, // texture bias 124 OP_TXL, // texture lod 125 OP_TXF, // texel fetch 126 OP_TXQ, // texture size query 127 OP_TXD, // texture derivatives 128 OP_TXG, // texture gather 129 OP_TXLQ, // texture query lod 130 OP_TEXCSAA, // texture op for coverage sampling 131 OP_TEXPREP, // turn cube map array into 2d array coordinates 132 OP_SULDB, // surface load (raw) 133 OP_SULDP, // surface load (formatted) 134 OP_SUSTB, // surface store (raw) 135 OP_SUSTP, // surface store (formatted) 136 OP_SUREDB, 137 OP_SUREDP, // surface reduction (atomic op) 138 OP_SULEA, // surface load effective address 139 OP_SUBFM, // surface bitfield manipulation 140 OP_SUCLAMP, // clamp surface coordinates 141 OP_SUEAU, // surface effective address 142 OP_SUQ, // surface query 143 OP_MADSP, // special integer multiply-add 144 OP_TEXBAR, // texture dependency barrier 145 OP_DFDX, 146 OP_DFDY, 147 OP_RDSV, // read system value 148 OP_WRSV, // write system value 149 OP_PIXLD, // get info about raster object or surfaces 150 OP_QUADOP, 151 OP_QUADON, 152 OP_QUADPOP, 153 OP_POPCNT, // bitcount(src0 & src1) 154 OP_INSBF, // insert first src1[8:15] bits of src0 into src2 at src1[0:7] 155 OP_EXTBF, // place bits [K,K+N) of src0 into dst, src1 = 0xNNKK 156 OP_BFIND, // find highest/lowest set bit 157 OP_BREV, // bitfield reverse 158 OP_BMSK, // bitfield mask 159 OP_PERMT, // dst = bytes from src2,src0 selected by src1 (nvc0's src order) 160 OP_SGXT, 161 OP_ATOM, 162 OP_BAR, // execution barrier, sources = { id, thread count, predicate } 163 OP_VADD, // byte/word vector operations 164 OP_VAVG, 165 OP_VMIN, 166 OP_VMAX, 167 OP_VSAD, 168 OP_VSET, 169 OP_VSHR, 170 OP_VSHL, 171 OP_VSEL, 172 OP_CCTL, // cache control 173 OP_SHFL, // warp shuffle 174 OP_VOTE, 175 OP_BUFQ, // buffer query 176 OP_WARPSYNC, 177 OP_LAST 178 }; 179 180 // various instruction-specific modifier definitions Instruction::subOp 181 // MOV_FINAL marks a MOV originating from an EXPORT (used for placing TEXBARs) 182 #define NV50_IR_SUBOP_MUL_HIGH 1 183 #define NV50_IR_SUBOP_EMIT_RESTART 1 184 #define NV50_IR_SUBOP_LDC_IL 1 185 #define NV50_IR_SUBOP_LDC_IS 2 186 #define NV50_IR_SUBOP_LDC_ISL 3 187 #define NV50_IR_SUBOP_SHIFT_WRAP 1 188 #define NV50_IR_SUBOP_SHIFT_HIGH 2 189 #define NV50_IR_SUBOP_EMU_PRERET 1 190 #define NV50_IR_SUBOP_TEXBAR(n) n 191 #define NV50_IR_SUBOP_MOV_FINAL 1 192 #define NV50_IR_SUBOP_EXTBF_REV 1 193 #define NV50_IR_SUBOP_BFIND_SAMT 1 194 #define NV50_IR_SUBOP_RCPRSQ_64H 1 195 #define NV50_IR_SUBOP_PERMT_F4E 1 196 #define NV50_IR_SUBOP_PERMT_B4E 2 197 #define NV50_IR_SUBOP_PERMT_RC8 3 198 #define NV50_IR_SUBOP_PERMT_ECL 4 199 #define NV50_IR_SUBOP_PERMT_ECR 5 200 #define NV50_IR_SUBOP_PERMT_RC16 6 201 #define NV50_IR_SUBOP_BAR_SYNC 0 202 #define NV50_IR_SUBOP_BAR_ARRIVE 1 203 #define NV50_IR_SUBOP_BAR_RED_AND 2 204 #define NV50_IR_SUBOP_BAR_RED_OR 3 205 #define NV50_IR_SUBOP_BAR_RED_POPC 4 206 #define NV50_IR_SUBOP_MEMBAR_L 1 207 #define NV50_IR_SUBOP_MEMBAR_S 2 208 #define NV50_IR_SUBOP_MEMBAR_M 3 209 #define NV50_IR_SUBOP_MEMBAR_CTA (0 << 2) 210 #define NV50_IR_SUBOP_MEMBAR_GL (1 << 2) 211 #define NV50_IR_SUBOP_MEMBAR_SYS (2 << 2) 212 #define NV50_IR_SUBOP_MEMBAR_DIR(m) ((m) & 0x3) 213 #define NV50_IR_SUBOP_MEMBAR_SCOPE(m) ((m) & ~0x3) 214 #define NV50_IR_SUBOP_MEMBAR(d,s) \ 215 (NV50_IR_SUBOP_MEMBAR_##d | NV50_IR_SUBOP_MEMBAR_##s) 216 #define NV50_IR_SUBOP_ATOM_ADD 0 217 #define NV50_IR_SUBOP_ATOM_MIN 1 218 #define NV50_IR_SUBOP_ATOM_MAX 2 219 #define NV50_IR_SUBOP_ATOM_INC 3 220 #define NV50_IR_SUBOP_ATOM_DEC 4 221 #define NV50_IR_SUBOP_ATOM_AND 5 222 #define NV50_IR_SUBOP_ATOM_OR 6 223 #define NV50_IR_SUBOP_ATOM_XOR 7 224 #define NV50_IR_SUBOP_ATOM_CAS 8 225 #define NV50_IR_SUBOP_ATOM_EXCH 9 226 #define NV50_IR_SUBOP_CCTL_IV 5 227 #define NV50_IR_SUBOP_CCTL_IVALL 6 228 #define NV50_IR_SUBOP_SUST_IGN 0 229 #define NV50_IR_SUBOP_SUST_TRAP 1 230 #define NV50_IR_SUBOP_SUST_SDCL 3 231 #define NV50_IR_SUBOP_SULD_ZERO 0 232 #define NV50_IR_SUBOP_SULD_TRAP 1 233 #define NV50_IR_SUBOP_SULD_SDCL 3 234 #define NV50_IR_SUBOP_SUBFM_3D 1 235 #define NV50_IR_SUBOP_SUCLAMP_2D 0x10 236 #define NV50_IR_SUBOP_SUCLAMP_SD(r, d) (( 0 + (r)) | ((d == 2) ? 0x10 : 0)) 237 #define NV50_IR_SUBOP_SUCLAMP_PL(r, d) (( 5 + (r)) | ((d == 2) ? 0x10 : 0)) 238 #define NV50_IR_SUBOP_SUCLAMP_BL(r, d) ((10 + (r)) | ((d == 2) ? 0x10 : 0)) 239 #define NV50_IR_SUBOP_PIXLD_COUNT 0 240 #define NV50_IR_SUBOP_PIXLD_COVMASK 1 241 #define NV50_IR_SUBOP_PIXLD_COVERED 2 242 #define NV50_IR_SUBOP_PIXLD_OFFSET 3 243 #define NV50_IR_SUBOP_PIXLD_CENT_OFFSET 4 244 #define NV50_IR_SUBOP_PIXLD_SAMPLEID 5 245 #define NV50_IR_SUBOP_SHFL_IDX 0 246 #define NV50_IR_SUBOP_SHFL_UP 1 247 #define NV50_IR_SUBOP_SHFL_DOWN 2 248 #define NV50_IR_SUBOP_SHFL_BFLY 3 249 #define NV50_IR_SUBOP_LOAD_LOCKED 1 250 #define NV50_IR_SUBOP_STORE_UNLOCKED 2 251 #define NV50_IR_SUBOP_MADSP_SD 0xffff 252 // Yes, we could represent those with DataType. 253 // Or put the type into operation and have a couple 1000 values in that enum. 254 // This will have to do for now. 255 // The bitfields are supposed to correspond to nve4 ISA. 256 #define NV50_IR_SUBOP_MADSP(a,b,c) (((c) << 8) | ((b) << 4) | (a)) 257 #define NV50_IR_SUBOP_V1(d,a,b) (((d) << 10) | ((b) << 5) | (a) | 0x0000) 258 #define NV50_IR_SUBOP_V2(d,a,b) (((d) << 10) | ((b) << 5) | (a) | 0x4000) 259 #define NV50_IR_SUBOP_V4(d,a,b) (((d) << 10) | ((b) << 5) | (a) | 0x8000) 260 #define NV50_IR_SUBOP_Vn(n) ((n) >> 14) 261 #define NV50_IR_SUBOP_VOTE_ALL 0 262 #define NV50_IR_SUBOP_VOTE_ANY 1 263 #define NV50_IR_SUBOP_VOTE_UNI 2 264 #define NV50_IR_SUBOP_LOP3_LUT_SRC0 0xf0 265 #define NV50_IR_SUBOP_LOP3_LUT_SRC1 0xcc 266 #define NV50_IR_SUBOP_LOP3_LUT_SRC2 0xaa 267 #define NV50_IR_SUBOP_LOP3_LUT(exp) ({ \ 268 uint8_t a = NV50_IR_SUBOP_LOP3_LUT_SRC0; \ 269 uint8_t b = NV50_IR_SUBOP_LOP3_LUT_SRC1; \ 270 uint8_t c = NV50_IR_SUBOP_LOP3_LUT_SRC2; \ 271 (uint8_t)(exp); \ 272 }) 273 #define NV50_IR_SUBOP_BMSK_C (0 << 0) 274 #define NV50_IR_SUBOP_BMSK_W (1 << 0) 275 276 #define NV50_IR_SUBOP_MINMAX_LOW 1 277 #define NV50_IR_SUBOP_MINMAX_MED 2 278 #define NV50_IR_SUBOP_MINMAX_HIGH 3 279 280 #define NV50_IR_SUBOP_SHF_L (0 << 0) 281 #define NV50_IR_SUBOP_SHF_R (1 << 0) 282 #define NV50_IR_SUBOP_SHF_LO (0 << 1) 283 #define NV50_IR_SUBOP_SHF_HI (1 << 1) 284 #define NV50_IR_SUBOP_SHF_C (0 << 2) 285 #define NV50_IR_SUBOP_SHF_W (1 << 2) 286 287 #define NV50_IR_SUBOP_VFETCH_PHYS 1 288 289 // xmad(src0, src1, 0) << 16 + src2 290 #define NV50_IR_SUBOP_XMAD_PSL (1 << 0) 291 // (xmad(src0, src1, src2) & 0xffff) | (src1 << 16) 292 #define NV50_IR_SUBOP_XMAD_MRG (1 << 1) 293 // xmad(src0, src1, src2.lo) 294 #define NV50_IR_SUBOP_XMAD_CLO (1 << 2) 295 // xmad(src0, src1, src2.hi) 296 #define NV50_IR_SUBOP_XMAD_CHI (2 << 2) 297 // if both operands to the multiplication are non-zero, subtract 65536 for each 298 // negative operand 299 #define NV50_IR_SUBOP_XMAD_CSFU (3 << 2) 300 // xmad(src0, src1, src2) + src1 << 16 301 #define NV50_IR_SUBOP_XMAD_CBCC (4 << 2) 302 #define NV50_IR_SUBOP_XMAD_CMODE_SHIFT 2 303 #define NV50_IR_SUBOP_XMAD_CMODE_MASK (0x7 << NV50_IR_SUBOP_XMAD_CMODE_SHIFT) 304 305 // use the high 16 bits instead of the low 16 bits for the multiplication. 306 // if the instruction's sType is signed, sign extend the operand from 16 bits 307 // to 32 before multiplication. 308 #define NV50_IR_SUBOP_XMAD_H1_SHIFT 5 309 #define NV50_IR_SUBOP_XMAD_H1(i) (1 << (NV50_IR_SUBOP_XMAD_H1_SHIFT + (i))) 310 #define NV50_IR_SUBOP_XMAD_H1_MASK (0x3 << NV50_IR_SUBOP_XMAD_H1_SHIFT) 311 312 enum DataType 313 { 314 TYPE_NONE, 315 TYPE_U8, 316 TYPE_S8, 317 TYPE_U16, 318 TYPE_S16, 319 TYPE_U32, 320 TYPE_S32, 321 TYPE_U64, // 64 bit operations are only lowered after register allocation 322 TYPE_S64, 323 TYPE_F16, 324 TYPE_F32, 325 TYPE_F64, 326 TYPE_B96, 327 TYPE_B128 328 }; 329 330 enum CondCode 331 { 332 CC_FL = 0, 333 CC_NEVER = CC_FL, // when used with FILE_FLAGS 334 CC_LT = 1, 335 CC_EQ = 2, 336 CC_NOT_P = CC_EQ, // when used with FILE_PREDICATE 337 CC_LE = 3, 338 CC_GT = 4, 339 CC_NE = 5, 340 CC_P = CC_NE, 341 CC_GE = 6, 342 CC_TR = 7, 343 CC_ALWAYS = CC_TR, 344 CC_U = 8, 345 CC_LTU = 9, 346 CC_EQU = 10, 347 CC_LEU = 11, 348 CC_GTU = 12, 349 CC_NEU = 13, 350 CC_GEU = 14, 351 CC_NO = 0x10, 352 CC_NC = 0x11, 353 CC_NS = 0x12, 354 CC_NA = 0x13, 355 CC_A = 0x14, 356 CC_S = 0x15, 357 CC_C = 0x16, 358 CC_O = 0x17 359 }; 360 361 enum RoundMode 362 { 363 ROUND_N, // nearest 364 ROUND_M, // towards -inf 365 ROUND_Z, // towards 0 366 ROUND_P, // towards +inf 367 ROUND_NI, // nearest integer 368 ROUND_MI, // to integer towards -inf 369 ROUND_ZI, // to integer towards 0 370 ROUND_PI, // to integer towards +inf 371 }; 372 373 enum CacheMode 374 { 375 CACHE_CA, // cache at all levels 376 CACHE_WB = CACHE_CA, // cache write back 377 CACHE_CG, // cache at global level 378 CACHE_CS, // cache streaming 379 CACHE_CV, // cache as volatile 380 CACHE_WT = CACHE_CV // cache write-through 381 }; 382 383 enum DataFile 384 { 385 FILE_NULL = 0, 386 FILE_GPR, 387 FILE_PREDICATE, // boolean predicate 388 FILE_FLAGS, // zero/sign/carry/overflow bits 389 FILE_ADDRESS, 390 FILE_BARRIER, 391 LAST_REGISTER_FILE = FILE_BARRIER, 392 FILE_IMMEDIATE, 393 FILE_MEMORY_CONST, 394 FILE_SHADER_INPUT, 395 FILE_SHADER_OUTPUT, 396 FILE_MEMORY_BUFFER, 397 FILE_MEMORY_GLOBAL, 398 FILE_MEMORY_SHARED, 399 FILE_MEMORY_LOCAL, 400 FILE_SYSTEM_VALUE, 401 FILE_THREAD_STATE, // "special" barrier registers 402 DATA_FILE_COUNT 403 }; 404 405 enum TexTarget 406 { 407 TEX_TARGET_1D, 408 TEX_TARGET_2D, 409 TEX_TARGET_2D_MS, 410 TEX_TARGET_3D, 411 TEX_TARGET_CUBE, 412 TEX_TARGET_1D_SHADOW, 413 TEX_TARGET_2D_SHADOW, 414 TEX_TARGET_CUBE_SHADOW, 415 TEX_TARGET_1D_ARRAY, 416 TEX_TARGET_2D_ARRAY, 417 TEX_TARGET_2D_MS_ARRAY, 418 TEX_TARGET_CUBE_ARRAY, 419 TEX_TARGET_1D_ARRAY_SHADOW, 420 TEX_TARGET_2D_ARRAY_SHADOW, 421 TEX_TARGET_RECT, 422 TEX_TARGET_RECT_SHADOW, 423 TEX_TARGET_CUBE_ARRAY_SHADOW, 424 TEX_TARGET_BUFFER, 425 TEX_TARGET_COUNT 426 }; 427 428 enum ImgFormat 429 { 430 FMT_NONE, 431 432 FMT_RGBA32F, 433 FMT_RGBA16F, 434 FMT_RG32F, 435 FMT_RG16F, 436 FMT_R11G11B10F, 437 FMT_R32F, 438 FMT_R16F, 439 440 FMT_RGBA32UI, 441 FMT_RGBA16UI, 442 FMT_RGB10A2UI, 443 FMT_RGBA8UI, 444 FMT_RG32UI, 445 FMT_RG16UI, 446 FMT_RG8UI, 447 FMT_R32UI, 448 FMT_R16UI, 449 FMT_R8UI, 450 451 FMT_RGBA32I, 452 FMT_RGBA16I, 453 FMT_RGBA8I, 454 FMT_RG32I, 455 FMT_RG16I, 456 FMT_RG8I, 457 FMT_R32I, 458 FMT_R16I, 459 FMT_R8I, 460 461 FMT_RGBA16, 462 FMT_RGB10A2, 463 FMT_RGBA8, 464 FMT_RG16, 465 FMT_RG8, 466 FMT_R16, 467 FMT_R8, 468 469 FMT_RGBA16_SNORM, 470 FMT_RGBA8_SNORM, 471 FMT_RG16_SNORM, 472 FMT_RG8_SNORM, 473 FMT_R16_SNORM, 474 FMT_R8_SNORM, 475 476 FMT_BGRA8, 477 478 IMG_FORMAT_COUNT, 479 }; 480 481 enum ImgType { 482 UINT, 483 SINT, 484 UNORM, 485 SNORM, 486 FLOAT, 487 }; 488 489 enum SVSemantic 490 { 491 SV_POSITION, // WPOS 492 SV_VERTEX_ID, 493 SV_INSTANCE_ID, 494 SV_INVOCATION_ID, 495 SV_PRIMITIVE_ID, 496 SV_VERTEX_COUNT, // gl_PatchVerticesIn 497 SV_LAYER, 498 SV_VIEWPORT_INDEX, 499 SV_VIEWPORT_MASK, 500 SV_YDIR, 501 SV_FACE, 502 SV_POINT_SIZE, 503 SV_POINT_COORD, 504 SV_CLIP_DISTANCE, 505 SV_SAMPLE_INDEX, 506 SV_SAMPLE_POS, 507 SV_SAMPLE_MASK, 508 SV_TESS_OUTER, 509 SV_TESS_INNER, 510 SV_TESS_COORD, 511 SV_TID, 512 SV_COMBINED_TID, 513 SV_CTAID, 514 SV_NTID, 515 SV_GRIDID, 516 SV_NCTAID, 517 SV_LANEID, 518 SV_PHYSID, 519 SV_NPHYSID, 520 SV_CLOCK, 521 SV_LBASE, 522 SV_SBASE, 523 SV_VERTEX_STRIDE, 524 SV_INVOCATION_INFO, 525 SV_THREAD_KILL, 526 SV_BASEVERTEX, 527 SV_BASEINSTANCE, 528 SV_DRAWID, 529 SV_WORK_DIM, 530 SV_LANEMASK_EQ, 531 SV_LANEMASK_LT, 532 SV_LANEMASK_LE, 533 SV_LANEMASK_GT, 534 SV_LANEMASK_GE, 535 SV_UNDEFINED, 536 SV_LAST 537 }; 538 539 enum TSSemantic 540 { 541 // 0-15 are fixed ones on Volta/Turing 542 TS_THREAD_STATE_ENUM0 = 0, 543 TS_THREAD_STATE_ENUM1 = 1, 544 TS_THREAD_STATE_ENUM2 = 2, 545 TS_THREAD_STATE_ENUM3 = 3, 546 TS_THREAD_STATE_ENUM4 = 4, 547 TS_TRAP_RETURN_PC_LO = 5, 548 TS_TRAP_RETURN_PC_HI = 6, 549 TS_TRAP_RETURN_MASK = 7, 550 TS_MEXITED = 8, 551 TS_MKILL = 9, 552 TS_MACTIVE = 10, 553 TS_MATEXIT = 11, 554 TS_OPT_STACK = 12, 555 TS_API_CALL_DEPTH = 13, 556 TS_ATEXIT_PC_LO = 14, 557 TS_ATEXIT_PC_HI = 15, 558 // special ones to make our life easier 559 TS_PQUAD_MACTIVE, 560 }; 561 562 class Program; 563 class Function; 564 class BasicBlock; 565 566 class Target; 567 568 class Instruction; 569 class CmpInstruction; 570 class TexInstruction; 571 class FlowInstruction; 572 573 class Value; 574 class LValue; 575 class Symbol; 576 class ImmediateValue; 577 578 struct Storage 579 { 580 DataFile file; 581 int8_t fileIndex; // signed, may be indirect for CONST[] 582 uint8_t size; // this should match the Instruction type's size 583 DataType type; // mainly for pretty printing 584 union { 585 uint64_t u64; // immediate values 586 uint32_t u32; 587 uint16_t u16; 588 uint8_t u8; 589 int64_t s64; 590 int32_t s32; 591 int16_t s16; 592 int8_t s8; 593 float f32; 594 double f64; 595 int32_t offset; // offset from 0 (base of address space) 596 int32_t id; // register id (< 0 if virtual/unassigned, in units <= 4) 597 struct { 598 SVSemantic sv; 599 int index; 600 } sv; 601 TSSemantic ts; 602 } data; 603 }; 604 605 // precedence: NOT after SAT after NEG after ABS 606 #define NV50_IR_MOD_ABS (1 << 0) 607 #define NV50_IR_MOD_NEG (1 << 1) 608 #define NV50_IR_MOD_SAT (1 << 2) 609 #define NV50_IR_MOD_NOT (1 << 3) 610 #define NV50_IR_MOD_NEG_ABS (NV50_IR_MOD_NEG | NV50_IR_MOD_ABS) 611 612 #define NV50_IR_INTERP_MODE_MASK 0x3 613 #define NV50_IR_INTERP_LINEAR (0 << 0) 614 #define NV50_IR_INTERP_PERSPECTIVE (1 << 0) 615 #define NV50_IR_INTERP_FLAT (2 << 0) 616 #define NV50_IR_INTERP_SC (3 << 0) // what exactly is that ? 617 #define NV50_IR_INTERP_SAMPLE_MASK 0xc 618 #define NV50_IR_INTERP_DEFAULT (0 << 2) 619 #define NV50_IR_INTERP_CENTROID (1 << 2) 620 #define NV50_IR_INTERP_OFFSET (2 << 2) 621 #define NV50_IR_INTERP_SAMPLEID (3 << 2) 622 623 // do we really want this to be a class ? 624 class Modifier 625 { 626 public: Modifier()627 Modifier() : bits(0) { } Modifier(unsigned int m)628 Modifier(unsigned int m) : bits(m) { } 629 Modifier(operation op); 630 631 // @return new Modifier applying a after b (asserts if unrepresentable) 632 Modifier operator*(const Modifier) const; 633 Modifier operator*=(const Modifier m) { *this = *this * m; return *this; } 634 Modifier operator==(const Modifier m) const { return m.bits == bits; } 635 Modifier operator!=(const Modifier m) const { return m.bits != bits; } 636 637 inline Modifier operator&(const Modifier m) const { return bits & m.bits; } 638 inline Modifier operator|(const Modifier m) const { return bits | m.bits; } 639 inline Modifier operator^(const Modifier m) const { return bits ^ m.bits; } 640 641 operation getOp() const; 642 neg()643 inline int neg() const { return (bits & NV50_IR_MOD_NEG) ? 1 : 0; } abs()644 inline int abs() const { return (bits & NV50_IR_MOD_ABS) ? 1 : 0; } 645 646 inline operator bool() const { return bits ? true : false; } 647 648 void applyTo(ImmediateValue &imm) const; 649 650 int print(char *buf, size_t size) const; 651 652 private: 653 uint8_t bits; 654 }; 655 656 class ValueRef 657 { 658 public: 659 ValueRef(Value * = NULL); 660 ValueRef(const ValueRef&); 661 ~ValueRef(); 662 exists()663 inline bool exists() const { return value != NULL; } 664 665 void set(Value *); 666 void set(const ValueRef&); get()667 inline Value *get() const { return value; } 668 inline Value *rep() const; 669 getInsn()670 inline Instruction *getInsn() const { return insn; } setInsn(Instruction * inst)671 inline void setInsn(Instruction *inst) { insn = inst; } 672 isIndirect(int dim)673 inline bool isIndirect(int dim) const { return indirect[dim] >= 0; } 674 inline const ValueRef *getIndirect(int dim) const; 675 676 inline DataFile getFile() const; 677 inline unsigned getSize() const; 678 679 // SSA: return eventual (traverse MOVs) literal value, if it exists 680 bool getImmediate(ImmediateValue&) const; 681 682 public: 683 Modifier mod; 684 int8_t indirect[2]; // >= 0 if relative to lvalue in insn->src(indirect[i]) 685 686 bool usedAsPtr; // for printing 687 688 private: 689 Value *value; 690 Instruction *insn; 691 }; 692 693 class ValueDef 694 { 695 public: 696 ValueDef(Value * = NULL); 697 ValueDef(const ValueDef&); 698 ~ValueDef(); 699 exists()700 inline bool exists() const { return value != NULL; } 701 get()702 inline Value *get() const { return value; } 703 inline Value *rep() const; 704 void set(Value *); 705 bool mayReplace(const ValueRef &); 706 void replace(const ValueRef &, bool doSet); // replace all uses of the old value 707 getInsn()708 inline Instruction *getInsn() const { return insn; } setInsn(Instruction * inst)709 inline void setInsn(Instruction *inst) { insn = inst; } 710 711 inline DataFile getFile() const; 712 inline unsigned getSize() const; 713 714 inline void setSSA(LValue *); 715 inline const LValue *preSSA() const; 716 717 private: 718 Value *value; // should make this LValue * ... 719 LValue *origin; // pre SSA value 720 Instruction *insn; 721 }; 722 723 class Value 724 { 725 public: 726 Value(); ~Value()727 virtual ~Value() { } 728 729 virtual Value *clone(ClonePolicy<Function>&) const = 0; 730 731 virtual int print(char *, size_t, DataType ty = TYPE_NONE) const = 0; 732 733 virtual bool equals(const Value *, bool strict = false) const; 734 virtual bool interfers(const Value *) const; isUniform()735 virtual bool isUniform() const { return true; } 736 rep()737 inline Value *rep() const { return join; } 738 739 inline Instruction *getUniqueInsn() const; 740 inline Instruction *getInsn() const; // use when uniqueness is certain 741 refCount()742 inline int refCount() { return uses.size(); } 743 744 inline LValue *asLValue(); 745 inline Symbol *asSym(); 746 inline ImmediateValue *asImm(); 747 inline const Symbol *asSym() const; 748 inline const ImmediateValue *asImm() const; 749 inFile(DataFile f)750 inline bool inFile(DataFile f) const { return reg.file == f; } 751 752 static inline Value *get(Iterator&); 753 754 std::unordered_set<ValueRef *> uses; 755 std::list<ValueDef *> defs; 756 typedef std::unordered_set<ValueRef *>::iterator UseIterator; 757 typedef std::unordered_set<ValueRef *>::const_iterator UseCIterator; 758 typedef std::list<ValueDef *>::iterator DefIterator; 759 typedef std::list<ValueDef *>::const_iterator DefCIterator; 760 761 int id; 762 Storage reg; 763 764 // TODO: these should be in LValue: 765 Interval livei; 766 Value *join; 767 }; 768 769 class LValue : public Value 770 { 771 public: 772 LValue(Function *, DataFile file); 773 LValue(Function *, LValue *); ~LValue()774 ~LValue() { } 775 776 virtual bool isUniform() const; 777 778 virtual LValue *clone(ClonePolicy<Function>&) const; 779 780 virtual int print(char *, size_t, DataType ty = TYPE_NONE) const; 781 782 public: 783 unsigned compMask : 8; // compound/component mask 784 unsigned compound : 1; // used by RA, value involved in split/merge 785 unsigned ssa : 1; 786 unsigned fixedReg : 1; // set & used by RA, earlier just use (id < 0) 787 unsigned noSpill : 1; // do not spill (e.g. if spill temporary already) 788 }; 789 790 class Symbol : public Value 791 { 792 public: 793 Symbol(Program *, DataFile file = FILE_MEMORY_CONST, uint8_t fileIdx = 0); ~Symbol()794 ~Symbol() { } 795 796 virtual Symbol *clone(ClonePolicy<Function>&) const; 797 798 virtual bool equals(const Value *that, bool strict) const; 799 800 virtual bool isUniform() const; 801 802 virtual int print(char *, size_t, DataType ty = TYPE_NONE) const; 803 804 // print with indirect values 805 int print(char *, size_t, Value *, Value *, DataType ty = TYPE_NONE) const; 806 807 inline void setFile(DataFile file, uint8_t fileIndex = 0) 808 { 809 reg.file = file; 810 reg.fileIndex = fileIndex; 811 } 812 813 inline void setOffset(int32_t offset); 814 inline void setAddress(Symbol *base, int32_t offset); 815 inline void setSV(SVSemantic sv, uint32_t idx = 0); 816 getBase()817 inline const Symbol *getBase() const { return baseSym; } 818 819 private: 820 Symbol *baseSym; // array base for Symbols representing array elements 821 }; 822 823 class ImmediateValue : public Value 824 { 825 public: ImmediateValue()826 ImmediateValue() { } 827 ImmediateValue(Program *, uint32_t); 828 ImmediateValue(Program *, float); 829 ImmediateValue(Program *, double); 830 // NOTE: not added to program with 831 ImmediateValue(const ImmediateValue *, DataType ty); ~ImmediateValue()832 ~ImmediateValue() { }; 833 834 virtual ImmediateValue *clone(ClonePolicy<Function>&) const; 835 836 virtual bool equals(const Value *that, bool strict) const; 837 838 // these only work if 'type' is valid (we mostly use untyped literals): 839 bool isInteger(const int ival) const; // ival is cast to this' type 840 bool isNegative() const; 841 bool isPow2() const; 842 843 void applyLog2(); 844 845 // for constant folding: 846 ImmediateValue operator+(const ImmediateValue&) const; 847 ImmediateValue operator-(const ImmediateValue&) const; 848 ImmediateValue operator*(const ImmediateValue&) const; 849 ImmediateValue operator/(const ImmediateValue&) const; 850 851 ImmediateValue& operator=(const ImmediateValue&); // only sets value ! 852 853 bool compare(CondCode cc, float fval) const; 854 855 virtual int print(char *, size_t, DataType ty = TYPE_NONE) const; 856 }; 857 858 class Instruction 859 { 860 public: 861 Instruction(); 862 Instruction(Function *, operation, DataType); 863 virtual ~Instruction(); 864 865 virtual Instruction *clone(ClonePolicy<Function>&, 866 Instruction * = NULL) const; 867 868 void setDef(int i, Value *); 869 void setSrc(int s, Value *); 870 void setSrc(int s, const ValueRef&); 871 void swapSources(int a, int b); 872 void moveSources(int s, int delta); 873 bool setIndirect(int s, int dim, Value *); 874 src(int s)875 inline ValueRef& src(int s) { return srcs[s]; } def(int s)876 inline ValueDef& def(int s) { return defs[s]; } src(int s)877 inline const ValueRef& src(int s) const { return srcs[s]; } def(int s)878 inline const ValueDef& def(int s) const { return defs[s]; } 879 getDef(int d)880 inline Value *getDef(int d) const { return defs[d].get(); } getSrc(int s)881 inline Value *getSrc(int s) const { return srcs[s].get(); } 882 inline Value *getIndirect(int s, int dim) const; 883 defExists(unsigned d)884 inline bool defExists(unsigned d) const 885 { 886 return d < defs.size() && defs[d].exists(); 887 } srcExists(unsigned s)888 inline bool srcExists(unsigned s) const 889 { 890 return s < srcs.size() && srcs[s].exists(); 891 } 892 893 inline bool constrainedDefs() const; 894 895 bool setPredicate(CondCode ccode, Value *); 896 inline Value *getPredicate() const; 897 bool writesPredicate() const; isPredicated()898 inline bool isPredicated() const { return predSrc >= 0; } 899 900 inline void setFlagsSrc(int s, Value *); 901 inline void setFlagsDef(int d, Value *); usesFlags()902 inline bool usesFlags() const { return flagsSrc >= 0; } 903 defCount()904 unsigned int defCount() const { return defs.size(); }; 905 unsigned int defCount(unsigned int mask, bool singleFile = false) const; srcCount()906 unsigned int srcCount() const { return srcs.size(); }; 907 unsigned int srcCount(unsigned int mask, bool singleFile = false) const; 908 909 // save & remove / set indirect[0,1] and predicate source 910 void takeExtraSources(int s, Value *[3]); 911 void putExtraSources(int s, Value *[3]); 912 setType(DataType type)913 inline void setType(DataType type) { dType = sType = type; } 914 setType(DataType dtype,DataType stype)915 inline void setType(DataType dtype, DataType stype) 916 { 917 dType = dtype; 918 sType = stype; 919 } 920 isPseudo()921 inline bool isPseudo() const { return op < OP_MOV; } 922 bool isDead() const; 923 bool isNop() const; 924 bool isCommutationLegal(const Instruction *) const; // must be adjacent ! 925 bool isActionEqual(const Instruction *) const; 926 bool isResultEqual(const Instruction *) const; 927 928 // check whether the defs interfere with srcs and defs of another instruction 929 bool canCommuteDefDef(const Instruction *) const; 930 bool canCommuteDefSrc(const Instruction *) const; 931 932 void print() const; 933 934 inline CmpInstruction *asCmp(); 935 inline TexInstruction *asTex(); 936 inline FlowInstruction *asFlow(); 937 inline const TexInstruction *asTex() const; 938 inline const CmpInstruction *asCmp() const; 939 inline const FlowInstruction *asFlow() const; 940 941 public: 942 Instruction *next; 943 Instruction *prev; 944 int id; 945 int serial; // CFG order 946 947 operation op; 948 DataType dType; // destination or defining type 949 DataType sType; // source or secondary type 950 CondCode cc; 951 RoundMode rnd; 952 CacheMode cache; 953 954 uint16_t subOp; // quadop, 1 for mul-high, etc. 955 956 unsigned encSize : 5; // encoding size in bytes 957 unsigned saturate : 1; // to [0.0f, 1.0f] 958 unsigned join : 1; // converge control flow (use OP_JOIN until end) 959 unsigned fixed : 1; // prevent dead code elimination 960 unsigned terminator : 1; // end of basic block 961 unsigned ftz : 1; // flush denormal to zero 962 unsigned dnz : 1; // denormals, NaN are zero 963 unsigned ipa : 4; // interpolation mode 964 unsigned lanes : 4; 965 unsigned perPatch : 1; 966 unsigned exit : 1; // terminate program after insn 967 unsigned mask : 4; // for vector ops 968 // prevent algebraic optimisations that aren't bit-for-bit identical 969 unsigned precise : 1; 970 971 int8_t postFactor; // MUL/DIV(if < 0) by 1 << postFactor 972 973 int8_t predSrc; 974 int8_t flagsDef; 975 int8_t flagsSrc; 976 977 uint32_t sched; // scheduling data (NOTE: maybe move to separate storage) 978 979 BasicBlock *bb; 980 981 protected: 982 std::deque<ValueDef> defs; // no gaps ! 983 std::deque<ValueRef> srcs; // no gaps ! 984 985 // instruction specific methods: 986 // (don't want to subclass, would need more constructors and memory pools) 987 public: setInterpolate(unsigned int mode)988 inline void setInterpolate(unsigned int mode) { ipa = mode; } 989 getInterpMode()990 unsigned int getInterpMode() const { return ipa & 0x3; } getSampleMode()991 unsigned int getSampleMode() const { return ipa & 0xc; } 992 993 private: 994 void init(); 995 }; 996 997 enum TexQuery 998 { 999 TXQ_DIMS, /* x, y, z, levels */ 1000 TXQ_TYPE, /* ?, ?, samples, ? */ 1001 TXQ_SAMPLE_POSITION, 1002 TXQ_FILTER, 1003 TXQ_LOD, 1004 TXQ_WRAP, 1005 TXQ_BORDER_COLOUR 1006 }; 1007 1008 class TexInstruction : public Instruction 1009 { 1010 public: 1011 class Target 1012 { 1013 public: target(targ)1014 Target(TexTarget targ = TEX_TARGET_1D) : target(targ) { } 1015 getName()1016 const char *getName() const { return descTable[target].name; } getArgCount()1017 unsigned int getArgCount() const { return descTable[target].argc; } getDim()1018 unsigned int getDim() const { return descTable[target].dim; } isArray()1019 int isArray() const { return descTable[target].array ? 1 : 0; } isCube()1020 int isCube() const { return descTable[target].cube ? 1 : 0; } isShadow()1021 int isShadow() const { return descTable[target].shadow ? 1 : 0; } isMS()1022 int isMS() const { 1023 return target == TEX_TARGET_2D_MS || target == TEX_TARGET_2D_MS_ARRAY; } clearMS()1024 void clearMS() { 1025 if (isMS()) { 1026 if (isArray()) 1027 target = TEX_TARGET_2D_ARRAY; 1028 else 1029 target = TEX_TARGET_2D; 1030 } 1031 } 1032 1033 Target& operator=(TexTarget targ) 1034 { 1035 assert(targ < TEX_TARGET_COUNT); 1036 target = targ; 1037 return *this; 1038 } 1039 1040 inline bool operator==(TexTarget targ) const { return target == targ; } 1041 inline bool operator!=(TexTarget targ) const { return target != targ; } 1042 getEnum()1043 enum TexTarget getEnum() const { return target; } 1044 1045 private: 1046 struct Desc 1047 { 1048 char name[19]; 1049 uint8_t dim; 1050 uint8_t argc; 1051 bool array; 1052 bool cube; 1053 bool shadow; 1054 }; 1055 1056 static const struct Desc descTable[TEX_TARGET_COUNT]; 1057 1058 private: 1059 enum TexTarget target; 1060 }; 1061 1062 public: 1063 struct ImgFormatDesc 1064 { 1065 char name[19]; 1066 uint8_t components; 1067 uint8_t bits[4]; 1068 ImgType type; 1069 bool bgra; 1070 }; 1071 1072 static const struct ImgFormatDesc formatTable[IMG_FORMAT_COUNT]; 1073 static const struct ImgFormatDesc *translateImgFormat( 1074 enum pipe_format format); 1075 1076 public: 1077 TexInstruction(Function *, operation); 1078 virtual ~TexInstruction(); 1079 1080 virtual TexInstruction *clone(ClonePolicy<Function>&, 1081 Instruction * = NULL) const; 1082 setTexture(Target targ,uint8_t r,uint8_t s)1083 inline void setTexture(Target targ, uint8_t r, uint8_t s) 1084 { 1085 tex.r = r; 1086 tex.s = s; 1087 tex.target = targ; 1088 } 1089 1090 void setIndirectR(Value *); 1091 void setIndirectS(Value *); 1092 inline Value *getIndirectR() const; 1093 inline Value *getIndirectS() const; 1094 1095 public: 1096 struct { 1097 Target target; 1098 1099 uint16_t r; 1100 uint16_t s; 1101 int8_t rIndirectSrc; 1102 int8_t sIndirectSrc; 1103 1104 uint8_t mask; 1105 uint8_t gatherComp; 1106 1107 bool liveOnly; // only execute on live pixels of a quad (optimization) 1108 bool levelZero; 1109 bool derivAll; 1110 bool bindless; 1111 1112 int8_t useOffsets; // 0, 1, or 4 for textureGatherOffsets 1113 int8_t offset[3]; // only used on nv50 1114 1115 enum TexQuery query; 1116 const struct ImgFormatDesc *format; 1117 1118 bool scalar; // for GM107s TEXS, TLDS, TLD4S 1119 } tex; 1120 1121 ValueRef dPdx[3]; 1122 ValueRef dPdy[3]; 1123 ValueRef offset[4][3]; 1124 }; 1125 1126 class CmpInstruction : public Instruction 1127 { 1128 public: 1129 CmpInstruction(Function *, operation); 1130 1131 virtual CmpInstruction *clone(ClonePolicy<Function>&, 1132 Instruction * = NULL) const; 1133 setCondition(CondCode cond)1134 void setCondition(CondCode cond) { setCond = cond; } getCondition()1135 CondCode getCondition() const { return setCond; } 1136 1137 public: 1138 CondCode setCond; 1139 }; 1140 1141 class FlowInstruction : public Instruction 1142 { 1143 public: 1144 FlowInstruction(Function *, operation, void *target); 1145 1146 virtual FlowInstruction *clone(ClonePolicy<Function>&, 1147 Instruction * = NULL) const; 1148 1149 public: 1150 unsigned allWarp : 1; 1151 unsigned absolute : 1; 1152 unsigned limit : 1; 1153 unsigned builtin : 1; // true for calls to emulation code 1154 unsigned indirect : 1; // target in src(0) 1155 1156 union { 1157 BasicBlock *bb; 1158 int builtin; 1159 Function *fn; 1160 } target; 1161 }; 1162 1163 class BasicBlock 1164 { 1165 public: 1166 BasicBlock(Function *); 1167 ~BasicBlock(); 1168 1169 BasicBlock *clone(ClonePolicy<Function>&) const; 1170 getId()1171 inline int getId() const { return id; } getInsnCount()1172 inline unsigned int getInsnCount() const { return numInsns; } isTerminated()1173 inline bool isTerminated() const { return exit && exit->terminator; } 1174 1175 bool dominatedBy(BasicBlock *bb); 1176 inline bool reachableBy(const BasicBlock *by, const BasicBlock *term); 1177 1178 // returns mask of conditional out blocks 1179 // e.g. 3 for IF { .. } ELSE { .. } ENDIF, 1 for IF { .. } ENDIF 1180 unsigned int initiatesSimpleConditional() const; 1181 1182 public: getFunction()1183 Function *getFunction() const { return func; } getProgram()1184 Program *getProgram() const { return program; } 1185 getEntry()1186 Instruction *getEntry() const { return entry; } // first non-phi instruction getPhi()1187 Instruction *getPhi() const { return phi; } getFirst()1188 Instruction *getFirst() const { return phi ? phi : entry; } getExit()1189 Instruction *getExit() const { return exit; } 1190 1191 void insertHead(Instruction *); 1192 void insertTail(Instruction *); 1193 void insertBefore(Instruction *, Instruction *); 1194 void insertAfter(Instruction *, Instruction *); 1195 void remove(Instruction *); 1196 void permuteAdjacent(Instruction *, Instruction *); 1197 1198 BasicBlock *idom() const; 1199 1200 // NOTE: currently does not rebuild the dominator tree 1201 BasicBlock *splitBefore(Instruction *, bool attach = true); 1202 BasicBlock *splitAfter(Instruction *, bool attach = true); 1203 getDF()1204 DLList& getDF() { return df; } iterDF()1205 DLList::Iterator iterDF() { return df.iterator(); } 1206 1207 static inline BasicBlock *get(Iterator&); 1208 static inline BasicBlock *get(Graph::Node *); 1209 1210 public: 1211 Graph::Node cfg; // first edge is branch *taken* (the ELSE branch) 1212 Graph::Node dom; 1213 1214 BitSet liveSet; 1215 BitSet defSet; 1216 1217 uint32_t binPos; 1218 uint32_t binSize; 1219 1220 Instruction *joinAt; // for quick reference 1221 1222 bool explicitCont; // loop headers: true if loop contains continue stmts 1223 1224 private: 1225 int id; 1226 DLList df; 1227 1228 Instruction *phi; 1229 Instruction *entry; 1230 Instruction *exit; 1231 1232 unsigned int numInsns; 1233 1234 private: 1235 Function *func; 1236 Program *program; 1237 1238 void splitCommon(Instruction *, BasicBlock *, bool attach); 1239 }; 1240 1241 class Function 1242 { 1243 public: 1244 Function(Program *, const char *name, uint32_t label); 1245 ~Function(); 1246 1247 static inline Function *get(Graph::Node *node); 1248 getProgram()1249 inline Program *getProgram() const { return prog; } getName()1250 inline const char *getName() const { return name; } getId()1251 inline int getId() const { return id; } getLabel()1252 inline uint32_t getLabel() const { return label; } 1253 1254 void print(); 1255 void printLiveIntervals() const; 1256 void printCFGraph(const char *filePath); 1257 1258 bool setEntry(BasicBlock *); 1259 bool setExit(BasicBlock *); 1260 1261 unsigned int orderInstructions(ArrayList&); 1262 add(BasicBlock * bb,int & id)1263 inline void add(BasicBlock *bb, int& id) { allBBlocks.insert(bb, id); } add(Instruction * insn,int & id)1264 inline void add(Instruction *insn, int& id) { allInsns.insert(insn, id); } add(LValue * lval,int & id)1265 inline void add(LValue *lval, int& id) { allLValues.insert(lval, id); } 1266 1267 inline LValue *getLValue(int id); 1268 1269 void buildLiveSets(); 1270 void buildDefSets(); 1271 bool convertToSSA(); 1272 1273 public: 1274 std::deque<ValueDef> ins; 1275 std::deque<ValueRef> outs; 1276 std::deque<Value *> clobbers; 1277 1278 Graph cfg; 1279 Graph::Node *cfgExit; 1280 Graph *domTree; 1281 Graph::Node call; // node in the call graph 1282 1283 BasicBlock **bbArray; // BBs in emission order 1284 int bbCount; 1285 1286 unsigned int loopNestingBound; 1287 int regClobberMax; 1288 1289 uint32_t binPos; 1290 uint32_t binSize; 1291 1292 Value *stackPtr; 1293 1294 uint32_t tlsBase; // base address for l[] space (if no stack pointer is used) 1295 uint32_t tlsSize; 1296 1297 ArrayList allBBlocks; 1298 ArrayList allInsns; 1299 ArrayList allLValues; 1300 1301 private: 1302 void buildLiveSetsPreSSA(BasicBlock *, const int sequence); 1303 void buildDefSetsPreSSA(BasicBlock *bb, const int seq); 1304 1305 private: 1306 uint32_t label; 1307 int id; 1308 const char *const name; 1309 Program *prog; 1310 }; 1311 1312 enum CGStage 1313 { 1314 CG_STAGE_PRE_SSA, 1315 CG_STAGE_SSA, // expected directly before register allocation 1316 CG_STAGE_POST_RA 1317 }; 1318 1319 class Program 1320 { 1321 public: 1322 enum Type 1323 { 1324 TYPE_VERTEX, 1325 TYPE_TESSELLATION_CONTROL, 1326 TYPE_TESSELLATION_EVAL, 1327 TYPE_GEOMETRY, 1328 TYPE_FRAGMENT, 1329 TYPE_COMPUTE 1330 }; 1331 1332 Program(Type type, Target *targ); 1333 ~Program(); 1334 1335 void print(); 1336 getType()1337 Type getType() const { return progType; } 1338 add(Function * fn,int & id)1339 inline void add(Function *fn, int& id) { allFuncs.insert(fn, id); } del(Function * fn,int & id)1340 inline void del(Function *fn, int& id) { allFuncs.remove(id); } add(Value * rval,int & id)1341 inline void add(Value *rval, int& id) { allRValues.insert(rval, id); } 1342 1343 bool makeFromNIR(struct nv50_ir_prog_info *, 1344 struct nv50_ir_prog_info_out *); 1345 bool makeFromTGSI(struct nv50_ir_prog_info *, 1346 struct nv50_ir_prog_info_out *); 1347 bool convertToSSA(); 1348 bool optimizeSSA(int level); 1349 bool optimizePostRA(int level); 1350 bool registerAllocation(); 1351 bool emitBinary(struct nv50_ir_prog_info_out *); 1352 getTarget()1353 const Target *getTarget() const { return target; } 1354 1355 private: 1356 Type progType; 1357 Target *target; 1358 1359 public: 1360 Function *main; 1361 Graph calls; 1362 1363 ArrayList allFuncs; 1364 ArrayList allRValues; 1365 1366 uint32_t *code; 1367 uint32_t binSize; 1368 uint32_t tlsSize; // size required for FILE_MEMORY_LOCAL 1369 1370 int maxGPR; 1371 bool fp64; 1372 bool persampleInvocation; 1373 1374 MemoryPool mem_Instruction; 1375 MemoryPool mem_CmpInstruction; 1376 MemoryPool mem_TexInstruction; 1377 MemoryPool mem_FlowInstruction; 1378 MemoryPool mem_LValue; 1379 MemoryPool mem_Symbol; 1380 MemoryPool mem_ImmediateValue; 1381 1382 uint32_t dbgFlags; 1383 uint8_t optLevel; 1384 1385 void *targetPriv; // e.g. to carry information between passes 1386 1387 const struct nv50_ir_prog_info *driver; // for driver configuration 1388 const struct nv50_ir_prog_info_out *driver_out; // for driver configuration 1389 1390 void releaseInstruction(Instruction *); 1391 void releaseValue(Value *); 1392 }; 1393 1394 // TODO: add const version 1395 class Pass 1396 { 1397 public: 1398 bool run(Program *, bool ordered = false, bool skipPhi = false); 1399 bool run(Function *, bool ordered = false, bool skipPhi = false); 1400 1401 private: 1402 // return false to continue with next entity on next higher level visit(Function *)1403 virtual bool visit(Function *) { return true; } visit(BasicBlock *)1404 virtual bool visit(BasicBlock *) { return true; } visit(Instruction *)1405 virtual bool visit(Instruction *) { return false; } 1406 1407 bool doRun(Program *, bool ordered, bool skipPhi); 1408 bool doRun(Function *, bool ordered, bool skipPhi); 1409 1410 protected: 1411 bool err; 1412 Function *func; 1413 Program *prog; 1414 }; 1415 1416 // ============================================================================= 1417 1418 #include "nv50_ir_inlines.h" 1419 1420 } // namespace nv50_ir 1421 1422 #endif // __NV50_IR_H__ 1423