1 //===-- SIMCCodeEmitter.cpp - SI Code Emitter -----------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// The SI code emitter produces machine code that can be executed 11 /// directly on the GPU device. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "MCTargetDesc/AMDGPUFixupKinds.h" 16 #include "MCTargetDesc/AMDGPUMCCodeEmitter.h" 17 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 18 #include "SIDefines.h" 19 #include "Utils/AMDGPUBaseInfo.h" 20 #include "llvm/MC/MCContext.h" 21 #include "llvm/MC/MCExpr.h" 22 #include "llvm/MC/MCInstrInfo.h" 23 #include "llvm/MC/MCRegisterInfo.h" 24 25 using namespace llvm; 26 27 namespace { 28 29 class SIMCCodeEmitter : public AMDGPUMCCodeEmitter { 30 const MCRegisterInfo &MRI; 31 32 /// Encode an fp or int literal 33 uint32_t getLitEncoding(const MCOperand &MO, const MCOperandInfo &OpInfo, 34 const MCSubtargetInfo &STI) const; 35 36 public: 37 SIMCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri, 38 MCContext &ctx) 39 : AMDGPUMCCodeEmitter(mcii), MRI(mri) {} 40 SIMCCodeEmitter(const SIMCCodeEmitter &) = delete; 41 SIMCCodeEmitter &operator=(const SIMCCodeEmitter &) = delete; 42 43 /// Encode the instruction and write it to the OS. 44 void encodeInstruction(const MCInst &MI, raw_ostream &OS, 45 SmallVectorImpl<MCFixup> &Fixups, 46 const MCSubtargetInfo &STI) const override; 47 48 /// \returns the encoding for an MCOperand. 49 uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO, 50 SmallVectorImpl<MCFixup> &Fixups, 51 const MCSubtargetInfo &STI) const override; 52 53 /// Use a fixup to encode the simm16 field for SOPP branch 54 /// instructions. 55 unsigned getSOPPBrEncoding(const MCInst &MI, unsigned OpNo, 56 SmallVectorImpl<MCFixup> &Fixups, 57 const MCSubtargetInfo &STI) const override; 58 59 unsigned getSMEMOffsetEncoding(const MCInst &MI, unsigned OpNo, 60 SmallVectorImpl<MCFixup> &Fixups, 61 const MCSubtargetInfo &STI) const override; 62 63 unsigned getSDWASrcEncoding(const MCInst &MI, unsigned OpNo, 64 SmallVectorImpl<MCFixup> &Fixups, 65 const MCSubtargetInfo &STI) const override; 66 67 unsigned getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo, 68 SmallVectorImpl<MCFixup> &Fixups, 69 const MCSubtargetInfo &STI) const override; 70 71 unsigned getAVOperandEncoding(const MCInst &MI, unsigned OpNo, 72 SmallVectorImpl<MCFixup> &Fixups, 73 const MCSubtargetInfo &STI) const override; 74 75 private: 76 uint64_t getImplicitOpSelHiEncoding(int Opcode) const; 77 }; 78 79 } // end anonymous namespace 80 81 MCCodeEmitter *llvm::createSIMCCodeEmitter(const MCInstrInfo &MCII, 82 const MCRegisterInfo &MRI, 83 MCContext &Ctx) { 84 return new SIMCCodeEmitter(MCII, MRI, Ctx); 85 } 86 87 // Returns the encoding value to use if the given integer is an integer inline 88 // immediate value, or 0 if it is not. 89 template <typename IntTy> 90 static uint32_t getIntInlineImmEncoding(IntTy Imm) { 91 if (Imm >= 0 && Imm <= 64) 92 return 128 + Imm; 93 94 if (Imm >= -16 && Imm <= -1) 95 return 192 + std::abs(Imm); 96 97 return 0; 98 } 99 100 static uint32_t getLit16IntEncoding(uint16_t Val, const MCSubtargetInfo &STI) { 101 uint16_t IntImm = getIntInlineImmEncoding(static_cast<int16_t>(Val)); 102 return IntImm == 0 ? 255 : IntImm; 103 } 104 105 static uint32_t getLit16Encoding(uint16_t Val, const MCSubtargetInfo &STI) { 106 uint16_t IntImm = getIntInlineImmEncoding(static_cast<int16_t>(Val)); 107 if (IntImm != 0) 108 return IntImm; 109 110 if (Val == 0x3800) // 0.5 111 return 240; 112 113 if (Val == 0xB800) // -0.5 114 return 241; 115 116 if (Val == 0x3C00) // 1.0 117 return 242; 118 119 if (Val == 0xBC00) // -1.0 120 return 243; 121 122 if (Val == 0x4000) // 2.0 123 return 244; 124 125 if (Val == 0xC000) // -2.0 126 return 245; 127 128 if (Val == 0x4400) // 4.0 129 return 246; 130 131 if (Val == 0xC400) // -4.0 132 return 247; 133 134 if (Val == 0x3118 && // 1.0 / (2.0 * pi) 135 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]) 136 return 248; 137 138 return 255; 139 } 140 141 static uint32_t getLit32Encoding(uint32_t Val, const MCSubtargetInfo &STI) { 142 uint32_t IntImm = getIntInlineImmEncoding(static_cast<int32_t>(Val)); 143 if (IntImm != 0) 144 return IntImm; 145 146 if (Val == FloatToBits(0.5f)) 147 return 240; 148 149 if (Val == FloatToBits(-0.5f)) 150 return 241; 151 152 if (Val == FloatToBits(1.0f)) 153 return 242; 154 155 if (Val == FloatToBits(-1.0f)) 156 return 243; 157 158 if (Val == FloatToBits(2.0f)) 159 return 244; 160 161 if (Val == FloatToBits(-2.0f)) 162 return 245; 163 164 if (Val == FloatToBits(4.0f)) 165 return 246; 166 167 if (Val == FloatToBits(-4.0f)) 168 return 247; 169 170 if (Val == 0x3e22f983 && // 1.0 / (2.0 * pi) 171 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]) 172 return 248; 173 174 return 255; 175 } 176 177 static uint32_t getLit64Encoding(uint64_t Val, const MCSubtargetInfo &STI) { 178 uint32_t IntImm = getIntInlineImmEncoding(static_cast<int64_t>(Val)); 179 if (IntImm != 0) 180 return IntImm; 181 182 if (Val == DoubleToBits(0.5)) 183 return 240; 184 185 if (Val == DoubleToBits(-0.5)) 186 return 241; 187 188 if (Val == DoubleToBits(1.0)) 189 return 242; 190 191 if (Val == DoubleToBits(-1.0)) 192 return 243; 193 194 if (Val == DoubleToBits(2.0)) 195 return 244; 196 197 if (Val == DoubleToBits(-2.0)) 198 return 245; 199 200 if (Val == DoubleToBits(4.0)) 201 return 246; 202 203 if (Val == DoubleToBits(-4.0)) 204 return 247; 205 206 if (Val == 0x3fc45f306dc9c882 && // 1.0 / (2.0 * pi) 207 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]) 208 return 248; 209 210 return 255; 211 } 212 213 uint32_t SIMCCodeEmitter::getLitEncoding(const MCOperand &MO, 214 const MCOperandInfo &OpInfo, 215 const MCSubtargetInfo &STI) const { 216 int64_t Imm; 217 if (MO.isExpr()) { 218 const auto *C = dyn_cast<MCConstantExpr>(MO.getExpr()); 219 if (!C) 220 return 255; 221 222 Imm = C->getValue(); 223 } else { 224 225 assert(!MO.isDFPImm()); 226 227 if (!MO.isImm()) 228 return ~0; 229 230 Imm = MO.getImm(); 231 } 232 233 switch (OpInfo.OperandType) { 234 case AMDGPU::OPERAND_REG_IMM_INT32: 235 case AMDGPU::OPERAND_REG_IMM_FP32: 236 case AMDGPU::OPERAND_REG_INLINE_C_INT32: 237 case AMDGPU::OPERAND_REG_INLINE_C_FP32: 238 case AMDGPU::OPERAND_REG_INLINE_AC_INT32: 239 case AMDGPU::OPERAND_REG_INLINE_AC_FP32: 240 case AMDGPU::OPERAND_REG_IMM_V2INT32: 241 case AMDGPU::OPERAND_REG_IMM_V2FP32: 242 case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: 243 case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: 244 return getLit32Encoding(static_cast<uint32_t>(Imm), STI); 245 246 case AMDGPU::OPERAND_REG_IMM_INT64: 247 case AMDGPU::OPERAND_REG_IMM_FP64: 248 case AMDGPU::OPERAND_REG_INLINE_C_INT64: 249 case AMDGPU::OPERAND_REG_INLINE_C_FP64: 250 case AMDGPU::OPERAND_REG_INLINE_AC_FP64: 251 return getLit64Encoding(static_cast<uint64_t>(Imm), STI); 252 253 case AMDGPU::OPERAND_REG_IMM_INT16: 254 case AMDGPU::OPERAND_REG_INLINE_C_INT16: 255 case AMDGPU::OPERAND_REG_INLINE_AC_INT16: 256 return getLit16IntEncoding(static_cast<uint16_t>(Imm), STI); 257 case AMDGPU::OPERAND_REG_IMM_FP16: 258 case AMDGPU::OPERAND_REG_INLINE_C_FP16: 259 case AMDGPU::OPERAND_REG_INLINE_AC_FP16: 260 // FIXME Is this correct? What do inline immediates do on SI for f16 src 261 // which does not have f16 support? 262 return getLit16Encoding(static_cast<uint16_t>(Imm), STI); 263 case AMDGPU::OPERAND_REG_IMM_V2INT16: 264 case AMDGPU::OPERAND_REG_IMM_V2FP16: { 265 if (!isUInt<16>(Imm) && STI.getFeatureBits()[AMDGPU::FeatureVOP3Literal]) 266 return getLit32Encoding(static_cast<uint32_t>(Imm), STI); 267 if (OpInfo.OperandType == AMDGPU::OPERAND_REG_IMM_V2FP16) 268 return getLit16Encoding(static_cast<uint16_t>(Imm), STI); 269 LLVM_FALLTHROUGH; 270 } 271 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: 272 case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: 273 return getLit16IntEncoding(static_cast<uint16_t>(Imm), STI); 274 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: 275 case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: { 276 uint16_t Lo16 = static_cast<uint16_t>(Imm); 277 uint32_t Encoding = getLit16Encoding(Lo16, STI); 278 return Encoding; 279 } 280 default: 281 llvm_unreachable("invalid operand size"); 282 } 283 } 284 285 uint64_t SIMCCodeEmitter::getImplicitOpSelHiEncoding(int Opcode) const { 286 using namespace AMDGPU::VOP3PEncoding; 287 using namespace AMDGPU::OpName; 288 289 if (AMDGPU::getNamedOperandIdx(Opcode, op_sel_hi) != -1) { 290 if (AMDGPU::getNamedOperandIdx(Opcode, src2) != -1) 291 return 0; 292 if (AMDGPU::getNamedOperandIdx(Opcode, src1) != -1) 293 return OP_SEL_HI_2; 294 if (AMDGPU::getNamedOperandIdx(Opcode, src0) != -1) 295 return OP_SEL_HI_1 | OP_SEL_HI_2; 296 } 297 return OP_SEL_HI_0 | OP_SEL_HI_1 | OP_SEL_HI_2; 298 } 299 300 void SIMCCodeEmitter::encodeInstruction(const MCInst &MI, raw_ostream &OS, 301 SmallVectorImpl<MCFixup> &Fixups, 302 const MCSubtargetInfo &STI) const { 303 verifyInstructionPredicates(MI, 304 computeAvailableFeatures(STI.getFeatureBits())); 305 306 int Opcode = MI.getOpcode(); 307 uint64_t Encoding = getBinaryCodeForInstr(MI, Fixups, STI); 308 const MCInstrDesc &Desc = MCII.get(Opcode); 309 unsigned bytes = Desc.getSize(); 310 311 // Set unused op_sel_hi bits to 1 for VOP3P and MAI instructions. 312 // Note that accvgpr_read/write are MAI, have src0, but do not use op_sel. 313 if ((Desc.TSFlags & SIInstrFlags::VOP3P) || 314 Opcode == AMDGPU::V_ACCVGPR_READ_B32_vi || 315 Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_vi) { 316 Encoding |= getImplicitOpSelHiEncoding(Opcode); 317 } 318 319 for (unsigned i = 0; i < bytes; i++) { 320 OS.write((uint8_t) ((Encoding >> (8 * i)) & 0xff)); 321 } 322 323 // NSA encoding. 324 if (AMDGPU::isGFX10Plus(STI) && Desc.TSFlags & SIInstrFlags::MIMG) { 325 int vaddr0 = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 326 AMDGPU::OpName::vaddr0); 327 int srsrc = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 328 AMDGPU::OpName::srsrc); 329 assert(vaddr0 >= 0 && srsrc > vaddr0); 330 unsigned NumExtraAddrs = srsrc - vaddr0 - 1; 331 unsigned NumPadding = (-NumExtraAddrs) & 3; 332 333 for (unsigned i = 0; i < NumExtraAddrs; ++i) 334 OS.write((uint8_t)getMachineOpValue(MI, MI.getOperand(vaddr0 + 1 + i), 335 Fixups, STI)); 336 for (unsigned i = 0; i < NumPadding; ++i) 337 OS.write(0); 338 } 339 340 if ((bytes > 8 && STI.getFeatureBits()[AMDGPU::FeatureVOP3Literal]) || 341 (bytes > 4 && !STI.getFeatureBits()[AMDGPU::FeatureVOP3Literal])) 342 return; 343 344 // Check for additional literals in SRC0/1/2 (Op 1/2/3) 345 for (unsigned i = 0, e = Desc.getNumOperands(); i < e; ++i) { 346 347 // Check if this operand should be encoded as [SV]Src 348 if (!AMDGPU::isSISrcOperand(Desc, i)) 349 continue; 350 351 // Is this operand a literal immediate? 352 const MCOperand &Op = MI.getOperand(i); 353 if (getLitEncoding(Op, Desc.OpInfo[i], STI) != 255) 354 continue; 355 356 // Yes! Encode it 357 int64_t Imm = 0; 358 359 if (Op.isImm()) 360 Imm = Op.getImm(); 361 else if (Op.isExpr()) { 362 if (const auto *C = dyn_cast<MCConstantExpr>(Op.getExpr())) 363 Imm = C->getValue(); 364 365 } else if (!Op.isExpr()) // Exprs will be replaced with a fixup value. 366 llvm_unreachable("Must be immediate or expr"); 367 368 for (unsigned j = 0; j < 4; j++) { 369 OS.write((uint8_t) ((Imm >> (8 * j)) & 0xff)); 370 } 371 372 // Only one literal value allowed 373 break; 374 } 375 } 376 377 unsigned SIMCCodeEmitter::getSOPPBrEncoding(const MCInst &MI, unsigned OpNo, 378 SmallVectorImpl<MCFixup> &Fixups, 379 const MCSubtargetInfo &STI) const { 380 const MCOperand &MO = MI.getOperand(OpNo); 381 382 if (MO.isExpr()) { 383 const MCExpr *Expr = MO.getExpr(); 384 MCFixupKind Kind = (MCFixupKind)AMDGPU::fixup_si_sopp_br; 385 Fixups.push_back(MCFixup::create(0, Expr, Kind, MI.getLoc())); 386 return 0; 387 } 388 389 return getMachineOpValue(MI, MO, Fixups, STI); 390 } 391 392 unsigned SIMCCodeEmitter::getSMEMOffsetEncoding(const MCInst &MI, unsigned OpNo, 393 SmallVectorImpl<MCFixup> &Fixups, 394 const MCSubtargetInfo &STI) const { 395 auto Offset = MI.getOperand(OpNo).getImm(); 396 // VI only supports 20-bit unsigned offsets. 397 assert(!AMDGPU::isVI(STI) || isUInt<20>(Offset)); 398 return Offset; 399 } 400 401 unsigned 402 SIMCCodeEmitter::getSDWASrcEncoding(const MCInst &MI, unsigned OpNo, 403 SmallVectorImpl<MCFixup> &Fixups, 404 const MCSubtargetInfo &STI) const { 405 using namespace AMDGPU::SDWA; 406 407 uint64_t RegEnc = 0; 408 409 const MCOperand &MO = MI.getOperand(OpNo); 410 411 if (MO.isReg()) { 412 unsigned Reg = MO.getReg(); 413 RegEnc |= MRI.getEncodingValue(Reg); 414 RegEnc &= SDWA9EncValues::SRC_VGPR_MASK; 415 if (AMDGPU::isSGPR(AMDGPU::mc2PseudoReg(Reg), &MRI)) { 416 RegEnc |= SDWA9EncValues::SRC_SGPR_MASK; 417 } 418 return RegEnc; 419 } else { 420 const MCInstrDesc &Desc = MCII.get(MI.getOpcode()); 421 uint32_t Enc = getLitEncoding(MO, Desc.OpInfo[OpNo], STI); 422 if (Enc != ~0U && Enc != 255) { 423 return Enc | SDWA9EncValues::SRC_SGPR_MASK; 424 } 425 } 426 427 llvm_unreachable("Unsupported operand kind"); 428 return 0; 429 } 430 431 unsigned 432 SIMCCodeEmitter::getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo, 433 SmallVectorImpl<MCFixup> &Fixups, 434 const MCSubtargetInfo &STI) const { 435 using namespace AMDGPU::SDWA; 436 437 uint64_t RegEnc = 0; 438 439 const MCOperand &MO = MI.getOperand(OpNo); 440 441 unsigned Reg = MO.getReg(); 442 if (Reg != AMDGPU::VCC && Reg != AMDGPU::VCC_LO) { 443 RegEnc |= MRI.getEncodingValue(Reg); 444 RegEnc &= SDWA9EncValues::VOPC_DST_SGPR_MASK; 445 RegEnc |= SDWA9EncValues::VOPC_DST_VCC_MASK; 446 } 447 return RegEnc; 448 } 449 450 unsigned 451 SIMCCodeEmitter::getAVOperandEncoding(const MCInst &MI, unsigned OpNo, 452 SmallVectorImpl<MCFixup> &Fixups, 453 const MCSubtargetInfo &STI) const { 454 unsigned Reg = MI.getOperand(OpNo).getReg(); 455 uint64_t Enc = MRI.getEncodingValue(Reg); 456 457 // VGPR and AGPR have the same encoding, but SrcA and SrcB operands of mfma 458 // instructions use acc[0:1] modifier bits to distinguish. These bits are 459 // encoded as a virtual 9th bit of the register for these operands. 460 if (MRI.getRegClass(AMDGPU::AGPR_32RegClassID).contains(Reg) || 461 MRI.getRegClass(AMDGPU::AReg_64RegClassID).contains(Reg) || 462 MRI.getRegClass(AMDGPU::AReg_96RegClassID).contains(Reg) || 463 MRI.getRegClass(AMDGPU::AReg_128RegClassID).contains(Reg) || 464 MRI.getRegClass(AMDGPU::AReg_160RegClassID).contains(Reg) || 465 MRI.getRegClass(AMDGPU::AReg_192RegClassID).contains(Reg) || 466 MRI.getRegClass(AMDGPU::AReg_256RegClassID).contains(Reg) || 467 MRI.getRegClass(AMDGPU::AGPR_LO16RegClassID).contains(Reg)) 468 Enc |= 512; 469 470 return Enc; 471 } 472 473 static bool needsPCRel(const MCExpr *Expr) { 474 switch (Expr->getKind()) { 475 case MCExpr::SymbolRef: { 476 auto *SE = cast<MCSymbolRefExpr>(Expr); 477 MCSymbolRefExpr::VariantKind Kind = SE->getKind(); 478 return Kind != MCSymbolRefExpr::VK_AMDGPU_ABS32_LO && 479 Kind != MCSymbolRefExpr::VK_AMDGPU_ABS32_HI; 480 } 481 case MCExpr::Binary: { 482 auto *BE = cast<MCBinaryExpr>(Expr); 483 if (BE->getOpcode() == MCBinaryExpr::Sub) 484 return false; 485 return needsPCRel(BE->getLHS()) || needsPCRel(BE->getRHS()); 486 } 487 case MCExpr::Unary: 488 return needsPCRel(cast<MCUnaryExpr>(Expr)->getSubExpr()); 489 case MCExpr::Target: 490 case MCExpr::Constant: 491 return false; 492 } 493 llvm_unreachable("invalid kind"); 494 } 495 496 uint64_t SIMCCodeEmitter::getMachineOpValue(const MCInst &MI, 497 const MCOperand &MO, 498 SmallVectorImpl<MCFixup> &Fixups, 499 const MCSubtargetInfo &STI) const { 500 if (MO.isReg()) 501 return MRI.getEncodingValue(MO.getReg()); 502 503 if (MO.isExpr() && MO.getExpr()->getKind() != MCExpr::Constant) { 504 // FIXME: If this is expression is PCRel or not should not depend on what 505 // the expression looks like. Given that this is just a general expression, 506 // it should probably be FK_Data_4 and whatever is producing 507 // 508 // s_add_u32 s2, s2, (extern_const_addrspace+16 509 // 510 // And expecting a PCRel should instead produce 511 // 512 // .Ltmp1: 513 // s_add_u32 s2, s2, (extern_const_addrspace+16)-.Ltmp1 514 MCFixupKind Kind; 515 if (needsPCRel(MO.getExpr())) 516 Kind = FK_PCRel_4; 517 else 518 Kind = FK_Data_4; 519 520 const MCInstrDesc &Desc = MCII.get(MI.getOpcode()); 521 uint32_t Offset = Desc.getSize(); 522 assert(Offset == 4 || Offset == 8); 523 524 Fixups.push_back( 525 MCFixup::create(Offset, MO.getExpr(), Kind, MI.getLoc())); 526 } 527 528 // Figure out the operand number, needed for isSrcOperand check 529 unsigned OpNo = 0; 530 for (unsigned e = MI.getNumOperands(); OpNo < e; ++OpNo) { 531 if (&MO == &MI.getOperand(OpNo)) 532 break; 533 } 534 535 const MCInstrDesc &Desc = MCII.get(MI.getOpcode()); 536 if (AMDGPU::isSISrcOperand(Desc, OpNo)) { 537 uint32_t Enc = getLitEncoding(MO, Desc.OpInfo[OpNo], STI); 538 if (Enc != ~0U && 539 (Enc != 255 || Desc.getSize() == 4 || Desc.getSize() == 8)) 540 return Enc; 541 542 } else if (MO.isImm()) 543 return MO.getImm(); 544 545 llvm_unreachable("Encoding of this operand type is not supported yet."); 546 return 0; 547 } 548 549 #define ENABLE_INSTR_PREDICATE_VERIFIER 550 #include "AMDGPUGenMCCodeEmitter.inc" 551