1 //===-- AMDGPUAsmPrinter.cpp - AMDGPU Assebly printer --------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 /// \file 11 /// 12 /// The AMDGPUAsmPrinter is used to print both assembly string and also binary 13 /// code. When passed an MCAsmStreamer it prints assembly and when passed 14 /// an MCObjectStreamer it outputs binary code. 15 // 16 //===----------------------------------------------------------------------===// 17 // 18 19 #include "AMDGPUAsmPrinter.h" 20 #include "MCTargetDesc/AMDGPUTargetStreamer.h" 21 #include "InstPrinter/AMDGPUInstPrinter.h" 22 #include "Utils/AMDGPUBaseInfo.h" 23 #include "AMDGPU.h" 24 #include "AMDKernelCodeT.h" 25 #include "AMDGPUSubtarget.h" 26 #include "R600Defines.h" 27 #include "R600MachineFunctionInfo.h" 28 #include "R600RegisterInfo.h" 29 #include "SIDefines.h" 30 #include "SIMachineFunctionInfo.h" 31 #include "SIRegisterInfo.h" 32 #include "llvm/CodeGen/MachineFrameInfo.h" 33 #include "llvm/MC/MCContext.h" 34 #include "llvm/MC/MCSectionELF.h" 35 #include "llvm/MC/MCStreamer.h" 36 #include "llvm/Support/ELF.h" 37 #include "llvm/Support/MathExtras.h" 38 #include "llvm/Support/TargetRegistry.h" 39 #include "llvm/Target/TargetLoweringObjectFile.h" 40 41 using namespace llvm; 42 43 // TODO: This should get the default rounding mode from the kernel. We just set 44 // the default here, but this could change if the OpenCL rounding mode pragmas 45 // are used. 46 // 47 // The denormal mode here should match what is reported by the OpenCL runtime 48 // for the CL_FP_DENORM bit from CL_DEVICE_{HALF|SINGLE|DOUBLE}_FP_CONFIG, but 49 // can also be override to flush with the -cl-denorms-are-zero compiler flag. 50 // 51 // AMD OpenCL only sets flush none and reports CL_FP_DENORM for double 52 // precision, and leaves single precision to flush all and does not report 53 // CL_FP_DENORM for CL_DEVICE_SINGLE_FP_CONFIG. Mesa's OpenCL currently reports 54 // CL_FP_DENORM for both. 55 // 56 // FIXME: It seems some instructions do not support single precision denormals 57 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32, 58 // and sin_f32, cos_f32 on most parts). 59 60 // We want to use these instructions, and using fp32 denormals also causes 61 // instructions to run at the double precision rate for the device so it's 62 // probably best to just report no single precision denormals. 63 static uint32_t getFPMode(const MachineFunction &F) { 64 const AMDGPUSubtarget& ST = F.getSubtarget<AMDGPUSubtarget>(); 65 // TODO: Is there any real use for the flush in only / flush out only modes? 66 67 uint32_t FP32Denormals = 68 ST.hasFP32Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT; 69 70 uint32_t FP64Denormals = 71 ST.hasFP64Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT; 72 73 return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) | 74 FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) | 75 FP_DENORM_MODE_SP(FP32Denormals) | 76 FP_DENORM_MODE_DP(FP64Denormals); 77 } 78 79 static AsmPrinter * 80 createAMDGPUAsmPrinterPass(TargetMachine &tm, 81 std::unique_ptr<MCStreamer> &&Streamer) { 82 return new AMDGPUAsmPrinter(tm, std::move(Streamer)); 83 } 84 85 extern "C" void LLVMInitializeAMDGPUAsmPrinter() { 86 TargetRegistry::RegisterAsmPrinter(TheAMDGPUTarget, createAMDGPUAsmPrinterPass); 87 TargetRegistry::RegisterAsmPrinter(TheGCNTarget, createAMDGPUAsmPrinterPass); 88 } 89 90 AMDGPUAsmPrinter::AMDGPUAsmPrinter(TargetMachine &TM, 91 std::unique_ptr<MCStreamer> Streamer) 92 : AsmPrinter(TM, std::move(Streamer)) {} 93 94 void AMDGPUAsmPrinter::EmitStartOfAsmFile(Module &M) { 95 if (TM.getTargetTriple().getOS() != Triple::AMDHSA) 96 return; 97 98 // Need to construct an MCSubtargetInfo here in case we have no functions 99 // in the module. 100 std::unique_ptr<MCSubtargetInfo> STI(TM.getTarget().createMCSubtargetInfo( 101 TM.getTargetTriple().str(), TM.getTargetCPU(), 102 TM.getTargetFeatureString())); 103 104 AMDGPUTargetStreamer *TS = 105 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer()); 106 107 TS->EmitDirectiveHSACodeObjectVersion(1, 0); 108 AMDGPU::IsaVersion ISA = AMDGPU::getIsaVersion(STI->getFeatureBits()); 109 TS->EmitDirectiveHSACodeObjectISA(ISA.Major, ISA.Minor, ISA.Stepping, 110 "AMD", "AMDGPU"); 111 } 112 113 void AMDGPUAsmPrinter::EmitFunctionBodyStart() { 114 const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>(); 115 SIProgramInfo KernelInfo; 116 if (STM.isAmdHsaOS()) { 117 getSIProgramInfo(KernelInfo, *MF); 118 EmitAmdKernelCodeT(*MF, KernelInfo); 119 } 120 } 121 122 void AMDGPUAsmPrinter::EmitFunctionEntryLabel() { 123 const SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>(); 124 const AMDGPUSubtarget &STM = MF->getSubtarget<AMDGPUSubtarget>(); 125 if (MFI->isKernel() && STM.isAmdHsaOS()) { 126 AMDGPUTargetStreamer *TS = 127 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer()); 128 TS->EmitAMDGPUSymbolType(CurrentFnSym->getName(), 129 ELF::STT_AMDGPU_HSA_KERNEL); 130 } 131 132 AsmPrinter::EmitFunctionEntryLabel(); 133 } 134 135 static bool isModuleLinkage(const GlobalValue *GV) { 136 switch (GV->getLinkage()) { 137 case GlobalValue::LinkOnceODRLinkage: 138 case GlobalValue::LinkOnceAnyLinkage: 139 case GlobalValue::InternalLinkage: 140 case GlobalValue::CommonLinkage: 141 return true; 142 case GlobalValue::ExternalLinkage: 143 return false; 144 default: llvm_unreachable("unknown linkage type"); 145 } 146 } 147 148 void AMDGPUAsmPrinter::EmitGlobalVariable(const GlobalVariable *GV) { 149 150 if (TM.getTargetTriple().getOS() != Triple::AMDHSA) { 151 AsmPrinter::EmitGlobalVariable(GV); 152 return; 153 } 154 155 if (GV->isDeclaration() || GV->getLinkage() == GlobalValue::PrivateLinkage) { 156 AsmPrinter::EmitGlobalVariable(GV); 157 return; 158 } 159 160 // Group segment variables aren't emitted in HSA. 161 if (AMDGPU::isGroupSegment(GV)) 162 return; 163 164 AMDGPUTargetStreamer *TS = 165 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer()); 166 if (isModuleLinkage(GV)) { 167 TS->EmitAMDGPUHsaModuleScopeGlobal(GV->getName()); 168 } else { 169 TS->EmitAMDGPUHsaProgramScopeGlobal(GV->getName()); 170 } 171 172 MCSymbolELF *GVSym = cast<MCSymbolELF>(getSymbol(GV)); 173 const DataLayout &DL = getDataLayout(); 174 175 // Emit the size 176 uint64_t Size = DL.getTypeAllocSize(GV->getType()->getElementType()); 177 OutStreamer->emitELFSize(GVSym, MCConstantExpr::create(Size, OutContext)); 178 OutStreamer->PushSection(); 179 OutStreamer->SwitchSection( 180 getObjFileLowering().SectionForGlobal(GV, *Mang, TM)); 181 const Constant *C = GV->getInitializer(); 182 OutStreamer->EmitLabel(GVSym); 183 EmitGlobalConstant(DL, C); 184 OutStreamer->PopSection(); 185 } 186 187 bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) { 188 189 // The starting address of all shader programs must be 256 bytes aligned. 190 MF.setAlignment(8); 191 192 SetupMachineFunction(MF); 193 194 MCContext &Context = getObjFileLowering().getContext(); 195 MCSectionELF *ConfigSection = 196 Context.getELFSection(".AMDGPU.config", ELF::SHT_PROGBITS, 0); 197 OutStreamer->SwitchSection(ConfigSection); 198 199 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>(); 200 SIProgramInfo KernelInfo; 201 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) { 202 getSIProgramInfo(KernelInfo, MF); 203 if (!STM.isAmdHsaOS()) { 204 EmitProgramInfoSI(MF, KernelInfo); 205 } 206 } else { 207 EmitProgramInfoR600(MF); 208 } 209 210 DisasmLines.clear(); 211 HexLines.clear(); 212 DisasmLineMaxLen = 0; 213 214 EmitFunctionBody(); 215 216 if (isVerbose()) { 217 MCSectionELF *CommentSection = 218 Context.getELFSection(".AMDGPU.csdata", ELF::SHT_PROGBITS, 0); 219 OutStreamer->SwitchSection(CommentSection); 220 221 if (STM.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) { 222 OutStreamer->emitRawComment(" Kernel info:", false); 223 OutStreamer->emitRawComment(" codeLenInByte = " + Twine(KernelInfo.CodeLen), 224 false); 225 OutStreamer->emitRawComment(" NumSgprs: " + Twine(KernelInfo.NumSGPR), 226 false); 227 OutStreamer->emitRawComment(" NumVgprs: " + Twine(KernelInfo.NumVGPR), 228 false); 229 OutStreamer->emitRawComment(" FloatMode: " + Twine(KernelInfo.FloatMode), 230 false); 231 OutStreamer->emitRawComment(" IeeeMode: " + Twine(KernelInfo.IEEEMode), 232 false); 233 OutStreamer->emitRawComment(" ScratchSize: " + Twine(KernelInfo.ScratchSize), 234 false); 235 236 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:USER_SGPR: " + 237 Twine(G_00B84C_USER_SGPR(KernelInfo.ComputePGMRSrc2)), 238 false); 239 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_X_EN: " + 240 Twine(G_00B84C_TGID_X_EN(KernelInfo.ComputePGMRSrc2)), 241 false); 242 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Y_EN: " + 243 Twine(G_00B84C_TGID_Y_EN(KernelInfo.ComputePGMRSrc2)), 244 false); 245 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TGID_Z_EN: " + 246 Twine(G_00B84C_TGID_Z_EN(KernelInfo.ComputePGMRSrc2)), 247 false); 248 OutStreamer->emitRawComment(" COMPUTE_PGM_RSRC2:TIDIG_COMP_CNT: " + 249 Twine(G_00B84C_TIDIG_COMP_CNT(KernelInfo.ComputePGMRSrc2)), 250 false); 251 252 } else { 253 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>(); 254 OutStreamer->emitRawComment( 255 Twine("SQ_PGM_RESOURCES:STACK_SIZE = " + Twine(MFI->StackSize))); 256 } 257 } 258 259 if (STM.dumpCode()) { 260 261 OutStreamer->SwitchSection( 262 Context.getELFSection(".AMDGPU.disasm", ELF::SHT_NOTE, 0)); 263 264 for (size_t i = 0; i < DisasmLines.size(); ++i) { 265 std::string Comment(DisasmLineMaxLen - DisasmLines[i].size(), ' '); 266 Comment += " ; " + HexLines[i] + "\n"; 267 268 OutStreamer->EmitBytes(StringRef(DisasmLines[i])); 269 OutStreamer->EmitBytes(StringRef(Comment)); 270 } 271 } 272 273 return false; 274 } 275 276 void AMDGPUAsmPrinter::EmitProgramInfoR600(const MachineFunction &MF) { 277 unsigned MaxGPR = 0; 278 bool killPixel = false; 279 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>(); 280 const R600RegisterInfo *RI = 281 static_cast<const R600RegisterInfo *>(STM.getRegisterInfo()); 282 const R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>(); 283 284 for (const MachineBasicBlock &MBB : MF) { 285 for (const MachineInstr &MI : MBB) { 286 if (MI.getOpcode() == AMDGPU::KILLGT) 287 killPixel = true; 288 unsigned numOperands = MI.getNumOperands(); 289 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) { 290 const MachineOperand &MO = MI.getOperand(op_idx); 291 if (!MO.isReg()) 292 continue; 293 unsigned HWReg = RI->getEncodingValue(MO.getReg()) & 0xff; 294 295 // Register with value > 127 aren't GPR 296 if (HWReg > 127) 297 continue; 298 MaxGPR = std::max(MaxGPR, HWReg); 299 } 300 } 301 } 302 303 unsigned RsrcReg; 304 if (STM.getGeneration() >= AMDGPUSubtarget::EVERGREEN) { 305 // Evergreen / Northern Islands 306 switch (MF.getFunction()->getCallingConv()) { 307 default: // Fall through 308 case CallingConv::AMDGPU_CS: RsrcReg = R_0288D4_SQ_PGM_RESOURCES_LS; break; 309 case CallingConv::AMDGPU_GS: RsrcReg = R_028878_SQ_PGM_RESOURCES_GS; break; 310 case CallingConv::AMDGPU_PS: RsrcReg = R_028844_SQ_PGM_RESOURCES_PS; break; 311 case CallingConv::AMDGPU_VS: RsrcReg = R_028860_SQ_PGM_RESOURCES_VS; break; 312 } 313 } else { 314 // R600 / R700 315 switch (MF.getFunction()->getCallingConv()) { 316 default: // Fall through 317 case CallingConv::AMDGPU_GS: // Fall through 318 case CallingConv::AMDGPU_CS: // Fall through 319 case CallingConv::AMDGPU_VS: RsrcReg = R_028868_SQ_PGM_RESOURCES_VS; break; 320 case CallingConv::AMDGPU_PS: RsrcReg = R_028850_SQ_PGM_RESOURCES_PS; break; 321 } 322 } 323 324 OutStreamer->EmitIntValue(RsrcReg, 4); 325 OutStreamer->EmitIntValue(S_NUM_GPRS(MaxGPR + 1) | 326 S_STACK_SIZE(MFI->StackSize), 4); 327 OutStreamer->EmitIntValue(R_02880C_DB_SHADER_CONTROL, 4); 328 OutStreamer->EmitIntValue(S_02880C_KILL_ENABLE(killPixel), 4); 329 330 if (AMDGPU::isCompute(MF.getFunction()->getCallingConv())) { 331 OutStreamer->EmitIntValue(R_0288E8_SQ_LDS_ALLOC, 4); 332 OutStreamer->EmitIntValue(alignTo(MFI->LDSSize, 4) >> 2, 4); 333 } 334 } 335 336 void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo, 337 const MachineFunction &MF) const { 338 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>(); 339 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 340 uint64_t CodeSize = 0; 341 unsigned MaxSGPR = 0; 342 unsigned MaxVGPR = 0; 343 bool VCCUsed = false; 344 bool FlatUsed = false; 345 const SIRegisterInfo *RI = 346 static_cast<const SIRegisterInfo *>(STM.getRegisterInfo()); 347 348 for (const MachineBasicBlock &MBB : MF) { 349 for (const MachineInstr &MI : MBB) { 350 // TODO: CodeSize should account for multiple functions. 351 352 // TODO: Should we count size of debug info? 353 if (MI.isDebugValue()) 354 continue; 355 356 // FIXME: This is reporting 0 for many instructions. 357 CodeSize += MI.getDesc().Size; 358 359 unsigned numOperands = MI.getNumOperands(); 360 for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) { 361 const MachineOperand &MO = MI.getOperand(op_idx); 362 unsigned width = 0; 363 bool isSGPR = false; 364 365 if (!MO.isReg()) 366 continue; 367 368 unsigned reg = MO.getReg(); 369 switch (reg) { 370 case AMDGPU::EXEC: 371 case AMDGPU::SCC: 372 case AMDGPU::M0: 373 continue; 374 375 case AMDGPU::VCC: 376 case AMDGPU::VCC_LO: 377 case AMDGPU::VCC_HI: 378 VCCUsed = true; 379 continue; 380 381 case AMDGPU::FLAT_SCR: 382 case AMDGPU::FLAT_SCR_LO: 383 case AMDGPU::FLAT_SCR_HI: 384 FlatUsed = true; 385 continue; 386 387 default: 388 break; 389 } 390 391 if (AMDGPU::SReg_32RegClass.contains(reg)) { 392 isSGPR = true; 393 width = 1; 394 } else if (AMDGPU::VGPR_32RegClass.contains(reg)) { 395 isSGPR = false; 396 width = 1; 397 } else if (AMDGPU::SReg_64RegClass.contains(reg)) { 398 isSGPR = true; 399 width = 2; 400 } else if (AMDGPU::VReg_64RegClass.contains(reg)) { 401 isSGPR = false; 402 width = 2; 403 } else if (AMDGPU::VReg_96RegClass.contains(reg)) { 404 isSGPR = false; 405 width = 3; 406 } else if (AMDGPU::SReg_128RegClass.contains(reg)) { 407 isSGPR = true; 408 width = 4; 409 } else if (AMDGPU::VReg_128RegClass.contains(reg)) { 410 isSGPR = false; 411 width = 4; 412 } else if (AMDGPU::SReg_256RegClass.contains(reg)) { 413 isSGPR = true; 414 width = 8; 415 } else if (AMDGPU::VReg_256RegClass.contains(reg)) { 416 isSGPR = false; 417 width = 8; 418 } else if (AMDGPU::SReg_512RegClass.contains(reg)) { 419 isSGPR = true; 420 width = 16; 421 } else if (AMDGPU::VReg_512RegClass.contains(reg)) { 422 isSGPR = false; 423 width = 16; 424 } else { 425 llvm_unreachable("Unknown register class"); 426 } 427 unsigned hwReg = RI->getEncodingValue(reg) & 0xff; 428 unsigned maxUsed = hwReg + width - 1; 429 if (isSGPR) { 430 MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR; 431 } else { 432 MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR; 433 } 434 } 435 } 436 } 437 438 unsigned ExtraSGPRs = 0; 439 440 if (VCCUsed) 441 ExtraSGPRs = 2; 442 443 if (STM.getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS) { 444 if (FlatUsed) 445 ExtraSGPRs = 4; 446 } else { 447 if (STM.isXNACKEnabled()) 448 ExtraSGPRs = 4; 449 450 if (FlatUsed) 451 ExtraSGPRs = 6; 452 } 453 454 MaxSGPR += ExtraSGPRs; 455 456 // We found the maximum register index. They start at 0, so add one to get the 457 // number of registers. 458 ProgInfo.NumVGPR = MaxVGPR + 1; 459 ProgInfo.NumSGPR = MaxSGPR + 1; 460 461 if (STM.hasSGPRInitBug()) { 462 if (ProgInfo.NumSGPR > AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG) { 463 LLVMContext &Ctx = MF.getFunction()->getContext(); 464 Ctx.emitError("too many SGPRs used with the SGPR init bug"); 465 } 466 467 ProgInfo.NumSGPR = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG; 468 } 469 470 if (MFI->NumUserSGPRs > STM.getMaxNumUserSGPRs()) { 471 LLVMContext &Ctx = MF.getFunction()->getContext(); 472 Ctx.emitError("too many user SGPRs used"); 473 } 474 475 ProgInfo.VGPRBlocks = (ProgInfo.NumVGPR - 1) / 4; 476 ProgInfo.SGPRBlocks = (ProgInfo.NumSGPR - 1) / 8; 477 // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode 478 // register. 479 ProgInfo.FloatMode = getFPMode(MF); 480 481 ProgInfo.IEEEMode = 0; 482 483 // Make clamp modifier on NaN input returns 0. 484 ProgInfo.DX10Clamp = 1; 485 486 const MachineFrameInfo *FrameInfo = MF.getFrameInfo(); 487 ProgInfo.ScratchSize = FrameInfo->getStackSize(); 488 489 ProgInfo.FlatUsed = FlatUsed; 490 ProgInfo.VCCUsed = VCCUsed; 491 ProgInfo.CodeLen = CodeSize; 492 493 unsigned LDSAlignShift; 494 if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) { 495 // LDS is allocated in 64 dword blocks. 496 LDSAlignShift = 8; 497 } else { 498 // LDS is allocated in 128 dword blocks. 499 LDSAlignShift = 9; 500 } 501 502 unsigned LDSSpillSize = MFI->LDSWaveSpillSize * 503 MFI->getMaximumWorkGroupSize(MF); 504 505 ProgInfo.LDSSize = MFI->LDSSize + LDSSpillSize; 506 ProgInfo.LDSBlocks = 507 alignTo(ProgInfo.LDSSize, 1ULL << LDSAlignShift) >> LDSAlignShift; 508 509 // Scratch is allocated in 256 dword blocks. 510 unsigned ScratchAlignShift = 10; 511 // We need to program the hardware with the amount of scratch memory that 512 // is used by the entire wave. ProgInfo.ScratchSize is the amount of 513 // scratch memory used per thread. 514 ProgInfo.ScratchBlocks = 515 alignTo(ProgInfo.ScratchSize * STM.getWavefrontSize(), 516 1ULL << ScratchAlignShift) >> 517 ScratchAlignShift; 518 519 ProgInfo.ComputePGMRSrc1 = 520 S_00B848_VGPRS(ProgInfo.VGPRBlocks) | 521 S_00B848_SGPRS(ProgInfo.SGPRBlocks) | 522 S_00B848_PRIORITY(ProgInfo.Priority) | 523 S_00B848_FLOAT_MODE(ProgInfo.FloatMode) | 524 S_00B848_PRIV(ProgInfo.Priv) | 525 S_00B848_DX10_CLAMP(ProgInfo.DX10Clamp) | 526 S_00B848_DEBUG_MODE(ProgInfo.DebugMode) | 527 S_00B848_IEEE_MODE(ProgInfo.IEEEMode); 528 529 // 0 = X, 1 = XY, 2 = XYZ 530 unsigned TIDIGCompCnt = 0; 531 if (MFI->hasWorkItemIDZ()) 532 TIDIGCompCnt = 2; 533 else if (MFI->hasWorkItemIDY()) 534 TIDIGCompCnt = 1; 535 536 ProgInfo.ComputePGMRSrc2 = 537 S_00B84C_SCRATCH_EN(ProgInfo.ScratchBlocks > 0) | 538 S_00B84C_USER_SGPR(MFI->getNumUserSGPRs()) | 539 S_00B84C_TGID_X_EN(MFI->hasWorkGroupIDX()) | 540 S_00B84C_TGID_Y_EN(MFI->hasWorkGroupIDY()) | 541 S_00B84C_TGID_Z_EN(MFI->hasWorkGroupIDZ()) | 542 S_00B84C_TG_SIZE_EN(MFI->hasWorkGroupInfo()) | 543 S_00B84C_TIDIG_COMP_CNT(TIDIGCompCnt) | 544 S_00B84C_EXCP_EN_MSB(0) | 545 S_00B84C_LDS_SIZE(ProgInfo.LDSBlocks) | 546 S_00B84C_EXCP_EN(0); 547 } 548 549 static unsigned getRsrcReg(CallingConv::ID CallConv) { 550 switch (CallConv) { 551 default: // Fall through 552 case CallingConv::AMDGPU_CS: return R_00B848_COMPUTE_PGM_RSRC1; 553 case CallingConv::AMDGPU_GS: return R_00B228_SPI_SHADER_PGM_RSRC1_GS; 554 case CallingConv::AMDGPU_PS: return R_00B028_SPI_SHADER_PGM_RSRC1_PS; 555 case CallingConv::AMDGPU_VS: return R_00B128_SPI_SHADER_PGM_RSRC1_VS; 556 } 557 } 558 559 void AMDGPUAsmPrinter::EmitProgramInfoSI(const MachineFunction &MF, 560 const SIProgramInfo &KernelInfo) { 561 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>(); 562 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 563 unsigned RsrcReg = getRsrcReg(MF.getFunction()->getCallingConv()); 564 565 if (AMDGPU::isCompute(MF.getFunction()->getCallingConv())) { 566 OutStreamer->EmitIntValue(R_00B848_COMPUTE_PGM_RSRC1, 4); 567 568 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc1, 4); 569 570 OutStreamer->EmitIntValue(R_00B84C_COMPUTE_PGM_RSRC2, 4); 571 OutStreamer->EmitIntValue(KernelInfo.ComputePGMRSrc2, 4); 572 573 OutStreamer->EmitIntValue(R_00B860_COMPUTE_TMPRING_SIZE, 4); 574 OutStreamer->EmitIntValue(S_00B860_WAVESIZE(KernelInfo.ScratchBlocks), 4); 575 576 // TODO: Should probably note flat usage somewhere. SC emits a "FlatPtr32 = 577 // 0" comment but I don't see a corresponding field in the register spec. 578 } else { 579 OutStreamer->EmitIntValue(RsrcReg, 4); 580 OutStreamer->EmitIntValue(S_00B028_VGPRS(KernelInfo.VGPRBlocks) | 581 S_00B028_SGPRS(KernelInfo.SGPRBlocks), 4); 582 if (STM.isVGPRSpillingEnabled(*MF.getFunction())) { 583 OutStreamer->EmitIntValue(R_0286E8_SPI_TMPRING_SIZE, 4); 584 OutStreamer->EmitIntValue(S_0286E8_WAVESIZE(KernelInfo.ScratchBlocks), 4); 585 } 586 } 587 588 if (MF.getFunction()->getCallingConv() == CallingConv::AMDGPU_PS) { 589 OutStreamer->EmitIntValue(R_00B02C_SPI_SHADER_PGM_RSRC2_PS, 4); 590 OutStreamer->EmitIntValue(S_00B02C_EXTRA_LDS_SIZE(KernelInfo.LDSBlocks), 4); 591 OutStreamer->EmitIntValue(R_0286CC_SPI_PS_INPUT_ENA, 4); 592 OutStreamer->EmitIntValue(MFI->PSInputEna, 4); 593 OutStreamer->EmitIntValue(R_0286D0_SPI_PS_INPUT_ADDR, 4); 594 OutStreamer->EmitIntValue(MFI->getPSInputAddr(), 4); 595 } 596 } 597 598 // This is supposed to be log2(Size) 599 static amd_element_byte_size_t getElementByteSizeValue(unsigned Size) { 600 switch (Size) { 601 case 4: 602 return AMD_ELEMENT_4_BYTES; 603 case 8: 604 return AMD_ELEMENT_8_BYTES; 605 case 16: 606 return AMD_ELEMENT_16_BYTES; 607 default: 608 llvm_unreachable("invalid private_element_size"); 609 } 610 } 611 612 void AMDGPUAsmPrinter::EmitAmdKernelCodeT(const MachineFunction &MF, 613 const SIProgramInfo &KernelInfo) const { 614 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 615 const AMDGPUSubtarget &STM = MF.getSubtarget<AMDGPUSubtarget>(); 616 amd_kernel_code_t header; 617 618 AMDGPU::initDefaultAMDKernelCodeT(header, STM.getFeatureBits()); 619 620 header.compute_pgm_resource_registers = 621 KernelInfo.ComputePGMRSrc1 | 622 (KernelInfo.ComputePGMRSrc2 << 32); 623 header.code_properties = AMD_CODE_PROPERTY_IS_PTR64; 624 625 626 AMD_HSA_BITS_SET(header.code_properties, 627 AMD_CODE_PROPERTY_PRIVATE_ELEMENT_SIZE, 628 getElementByteSizeValue(STM.getMaxPrivateElementSize())); 629 630 if (MFI->hasPrivateSegmentBuffer()) { 631 header.code_properties |= 632 AMD_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER; 633 } 634 635 if (MFI->hasDispatchPtr()) 636 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR; 637 638 if (MFI->hasQueuePtr()) 639 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR; 640 641 if (MFI->hasKernargSegmentPtr()) 642 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR; 643 644 if (MFI->hasDispatchID()) 645 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID; 646 647 if (MFI->hasFlatScratchInit()) 648 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT; 649 650 // TODO: Private segment size 651 652 if (MFI->hasGridWorkgroupCountX()) { 653 header.code_properties |= 654 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_X; 655 } 656 657 if (MFI->hasGridWorkgroupCountY()) { 658 header.code_properties |= 659 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Y; 660 } 661 662 if (MFI->hasGridWorkgroupCountZ()) { 663 header.code_properties |= 664 AMD_CODE_PROPERTY_ENABLE_SGPR_GRID_WORKGROUP_COUNT_Z; 665 } 666 667 if (MFI->hasDispatchPtr()) 668 header.code_properties |= AMD_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR; 669 670 if (STM.isXNACKEnabled()) 671 header.code_properties |= AMD_CODE_PROPERTY_IS_XNACK_SUPPORTED; 672 673 header.kernarg_segment_byte_size = MFI->ABIArgOffset; 674 header.wavefront_sgpr_count = KernelInfo.NumSGPR; 675 header.workitem_vgpr_count = KernelInfo.NumVGPR; 676 header.workitem_private_segment_byte_size = KernelInfo.ScratchSize; 677 header.workgroup_group_segment_byte_size = KernelInfo.LDSSize; 678 679 AMDGPUTargetStreamer *TS = 680 static_cast<AMDGPUTargetStreamer *>(OutStreamer->getTargetStreamer()); 681 TS->EmitAMDKernelCodeT(header); 682 } 683 684 bool AMDGPUAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNo, 685 unsigned AsmVariant, 686 const char *ExtraCode, raw_ostream &O) { 687 if (ExtraCode && ExtraCode[0]) { 688 if (ExtraCode[1] != 0) 689 return true; // Unknown modifier. 690 691 switch (ExtraCode[0]) { 692 default: 693 // See if this is a generic print operand 694 return AsmPrinter::PrintAsmOperand(MI, OpNo, AsmVariant, ExtraCode, O); 695 case 'r': 696 break; 697 } 698 } 699 700 AMDGPUInstPrinter::printRegOperand(MI->getOperand(OpNo).getReg(), O, 701 *TM.getSubtargetImpl(*MF->getFunction())->getRegisterInfo()); 702 return false; 703 } 704