| 1 | //=====-- GCNSubtarget.h - Define GCN Subtarget for AMDGPU ------*- C++ -*-===// |
| 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 | /// AMD GCN specific subclass of TargetSubtarget. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_LIB_TARGET_AMDGPU_GCNSUBTARGET_H |
| 15 | #define LLVM_LIB_TARGET_AMDGPU_GCNSUBTARGET_H |
| 16 | |
| 17 | #include "AMDGPUCallLowering.h" |
| 18 | #include "AMDGPURegisterBankInfo.h" |
| 19 | #include "AMDGPUSubtarget.h" |
| 20 | #include "SIFrameLowering.h" |
| 21 | #include "SIISelLowering.h" |
| 22 | #include "SIInstrInfo.h" |
| 23 | #include "Utils/AMDGPUBaseInfo.h" |
| 24 | #include "llvm/Support/AMDHSAKernelDescriptor.h" |
| 25 | #include "llvm/Support/ErrorHandling.h" |
| 26 | #include <optional> |
| 27 | |
| 28 | #define |
| 29 | #include "AMDGPUGenSubtargetInfo.inc" |
| 30 | |
| 31 | namespace llvm { |
| 32 | |
| 33 | class GCNTargetMachine; |
| 34 | |
| 35 | /// Module flag names controlling out-of-bounds buffer access semantics. |
| 36 | /// Each flag is an i32 with Module::Max merge behaviour and tri-state values: |
| 37 | /// 0 = any (absent/default - backend currently treats as strict) |
| 38 | /// 1 = relaxed |
| 39 | /// 2 = strict |
| 40 | namespace AMDGPUOOBMode { |
| 41 | inline constexpr StringLiteral BufferFlag("amdgpu.buffer.oob.mode" ); |
| 42 | inline constexpr StringLiteral TBufferFlag("amdgpu.tbuffer.oob.mode" ); |
| 43 | } // namespace AMDGPUOOBMode |
| 44 | |
| 45 | class GCNSubtarget final : public AMDGPUGenSubtargetInfo, |
| 46 | public AMDGPUSubtarget { |
| 47 | public: |
| 48 | using AMDGPUSubtarget::getMaxWavesPerEU; |
| 49 | |
| 50 | // Following 2 enums are documented at: |
| 51 | // - https://llvm.org/docs/AMDGPUUsage.html#trap-handler-abi |
| 52 | enum class TrapHandlerAbi { |
| 53 | NONE = 0x00, |
| 54 | AMDHSA = 0x01, |
| 55 | }; |
| 56 | |
| 57 | enum class TrapID { |
| 58 | LLVMAMDHSATrap = 0x02, |
| 59 | LLVMAMDHSADebugTrap = 0x03, |
| 60 | }; |
| 61 | |
| 62 | private: |
| 63 | /// SelectionDAGISel related APIs. |
| 64 | std::unique_ptr<const SelectionDAGTargetInfo> TSInfo; |
| 65 | |
| 66 | /// GlobalISel related APIs. |
| 67 | std::unique_ptr<AMDGPUCallLowering> CallLoweringInfo; |
| 68 | std::unique_ptr<InlineAsmLowering> InlineAsmLoweringInfo; |
| 69 | std::unique_ptr<InstructionSelector> InstSelector; |
| 70 | std::unique_ptr<LegalizerInfo> Legalizer; |
| 71 | std::unique_ptr<AMDGPURegisterBankInfo> RegBankInfo; |
| 72 | |
| 73 | protected: |
| 74 | // Basic subtarget description. |
| 75 | AMDGPU::TargetID TargetID; |
| 76 | unsigned Gen = INVALID; |
| 77 | InstrItineraryData InstrItins; |
| 78 | int LDSBankCount = 0; |
| 79 | unsigned MaxPrivateElementSize = 0; |
| 80 | |
| 81 | // Instruction cache line size in bytes; set from TableGen subtarget features. |
| 82 | unsigned InstCacheLineSize = 0; |
| 83 | |
| 84 | // Data (VMEM) cache line size in bytes; set from TableGen subtarget features. |
| 85 | unsigned DataCacheLineSize = 0; |
| 86 | |
| 87 | /// The width, in bits, of the num_records field of a buffer resource (V#), |
| 88 | /// set from tablegen subtarget features, 0 is unknown. |
| 89 | unsigned BufferResourceNumRecordsWidth = 0; |
| 90 | |
| 91 | // Dynamically set bits that enable features. |
| 92 | bool ScalarizeGlobal = false; |
| 93 | const bool BufferOOBRelaxed; |
| 94 | const bool TBufferOOBRelaxed; |
| 95 | |
| 96 | /// The maximum number of instructions that may be placed within an S_CLAUSE, |
| 97 | /// which is one greater than the maximum argument to S_CLAUSE. A value of 0 |
| 98 | /// indicates a lack of S_CLAUSE support. |
| 99 | unsigned MaxHardClauseLength = 0; |
| 100 | |
| 101 | #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER) \ |
| 102 | bool ATTRIBUTE = DEFAULT; |
| 103 | #include "AMDGPUGenSubtargetInfo.inc" |
| 104 | |
| 105 | private: |
| 106 | SIInstrInfo InstrInfo; |
| 107 | SITargetLowering TLInfo; |
| 108 | SIFrameLowering FrameLowering; |
| 109 | |
| 110 | /// Get the register that represents the actual dependency between the |
| 111 | /// definition and the use. The definition might only affect a subregister |
| 112 | /// that is not actually used. Works for both virtual and physical registers. |
| 113 | /// Note: Currently supports VOP3P instructions (without WMMA an SWMMAC). |
| 114 | /// Returns the definition register if there is a real dependency and no |
| 115 | /// better match is found. |
| 116 | Register getRealSchedDependency(const MachineInstr &DefI, int DefOpIdx, |
| 117 | const MachineInstr &UseI, int UseOpIdx) const; |
| 118 | |
| 119 | public: |
| 120 | GCNSubtarget( |
| 121 | const Triple &TT, StringRef GPU, StringRef FS, const GCNTargetMachine &TM, |
| 122 | bool BufferOOBRelaxed = false, bool TBufferOOBRelaxed = false, |
| 123 | AMDGPU::TargetIDSetting XnackSetting = AMDGPU::TargetIDSetting::Any, |
| 124 | AMDGPU::TargetIDSetting SramEccSetting = AMDGPU::TargetIDSetting::Any); |
| 125 | ~GCNSubtarget() override; |
| 126 | |
| 127 | GCNSubtarget &initializeSubtargetDependencies(const Triple &TT, StringRef GPU, |
| 128 | StringRef FS); |
| 129 | |
| 130 | /// Diagnose inconsistent subtarget features before attempting to codegen |
| 131 | /// function \p F. |
| 132 | void checkSubtargetFeatures(const Function &F) const; |
| 133 | |
| 134 | const SIInstrInfo *getInstrInfo() const override { return &InstrInfo; } |
| 135 | |
| 136 | const SIFrameLowering *getFrameLowering() const override { |
| 137 | return &FrameLowering; |
| 138 | } |
| 139 | |
| 140 | const SITargetLowering *getTargetLowering() const override { return &TLInfo; } |
| 141 | |
| 142 | const SIRegisterInfo *getRegisterInfo() const override { |
| 143 | return &InstrInfo.getRegisterInfo(); |
| 144 | } |
| 145 | |
| 146 | const SelectionDAGTargetInfo *getSelectionDAGInfo() const override; |
| 147 | |
| 148 | const CallLowering *getCallLowering() const override { |
| 149 | return CallLoweringInfo.get(); |
| 150 | } |
| 151 | |
| 152 | const InlineAsmLowering *getInlineAsmLowering() const override { |
| 153 | return InlineAsmLoweringInfo.get(); |
| 154 | } |
| 155 | |
| 156 | InstructionSelector *getInstructionSelector() const override { |
| 157 | return InstSelector.get(); |
| 158 | } |
| 159 | |
| 160 | const LegalizerInfo *getLegalizerInfo() const override { |
| 161 | return Legalizer.get(); |
| 162 | } |
| 163 | |
| 164 | const AMDGPURegisterBankInfo *getRegBankInfo() const override { |
| 165 | return RegBankInfo.get(); |
| 166 | } |
| 167 | |
| 168 | const AMDGPU::TargetID &getTargetID() const { return TargetID; } |
| 169 | |
| 170 | const InstrItineraryData *getInstrItineraryData() const override { |
| 171 | return &InstrItins; |
| 172 | } |
| 173 | |
| 174 | void ParseSubtargetFeatures(StringRef CPU, StringRef TuneCPU, StringRef FS); |
| 175 | |
| 176 | Generation getGeneration() const { return (Generation)Gen; } |
| 177 | |
| 178 | bool isGFX11Plus() const { return getGeneration() >= GFX11; } |
| 179 | |
| 180 | #define GET_SUBTARGETINFO_MACRO(ATTRIBUTE, DEFAULT, GETTER) \ |
| 181 | bool GETTER() const override { return ATTRIBUTE; } |
| 182 | #include "AMDGPUGenSubtargetInfo.inc" |
| 183 | |
| 184 | unsigned getMaxWaveScratchSize() const { |
| 185 | // See COMPUTE_TMPRING_SIZE.WAVESIZE. |
| 186 | if (getGeneration() >= GFX12) { |
| 187 | // 18-bit field in units of 64-dword. |
| 188 | return (64 * 4) * ((1 << 18) - 1); |
| 189 | } |
| 190 | if (getGeneration() == GFX11) { |
| 191 | // 15-bit field in units of 64-dword. |
| 192 | return (64 * 4) * ((1 << 15) - 1); |
| 193 | } |
| 194 | // 13-bit field in units of 256-dword. |
| 195 | return (256 * 4) * ((1 << 13) - 1); |
| 196 | } |
| 197 | |
| 198 | /// Return the number of high bits known to be zero for a frame index. |
| 199 | unsigned getKnownHighZeroBitsForFrameIndex() const { |
| 200 | return llvm::countl_zero(Val: getMaxWaveScratchSize()) + getWavefrontSizeLog2(); |
| 201 | } |
| 202 | |
| 203 | int getLDSBankCount() const { return LDSBankCount; } |
| 204 | |
| 205 | /// Instruction cache line size in bytes (64 for pre-GFX11, 128 for GFX11+). |
| 206 | unsigned getInstCacheLineSize() const { return InstCacheLineSize; } |
| 207 | |
| 208 | /// Data (VMEM) cache line size in bytes (128 for gfx12), has no use before |
| 209 | /// GFX12. |
| 210 | unsigned getDataCacheLineSize() const { return DataCacheLineSize; } |
| 211 | |
| 212 | unsigned getMaxPrivateElementSize(bool ForBufferRSrc = false) const { |
| 213 | return (ForBufferRSrc || !hasFlatScratchEnabled()) ? MaxPrivateElementSize |
| 214 | : 16; |
| 215 | } |
| 216 | |
| 217 | unsigned getConstantBusLimit(unsigned Opcode) const; |
| 218 | |
| 219 | /// Returns if the result of this instruction with a 16-bit result returned in |
| 220 | /// a 32-bit register implicitly zeroes the high 16-bits, rather than preserve |
| 221 | /// the original value. |
| 222 | bool zeroesHigh16BitsOfDest(unsigned Opcode) const; |
| 223 | |
| 224 | bool hasHWFP64() const { return HasFP64; } |
| 225 | |
| 226 | bool hasAddr64() const { |
| 227 | return (getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS); |
| 228 | } |
| 229 | |
| 230 | bool hasFlat() const { |
| 231 | return (getGeneration() > AMDGPUSubtarget::SOUTHERN_ISLANDS); |
| 232 | } |
| 233 | |
| 234 | // Return true if the target only has the reverse operand versions of VALU |
| 235 | // shift instructions (e.g. v_lshrrev_b32, and no v_lshr_b32). |
| 236 | bool hasOnlyRevVALUShifts() const { |
| 237 | return getGeneration() >= VOLCANIC_ISLANDS; |
| 238 | } |
| 239 | |
| 240 | bool hasFractBug() const { return getGeneration() == SOUTHERN_ISLANDS; } |
| 241 | |
| 242 | bool hasMed3_16() const { return getGeneration() >= AMDGPUSubtarget::GFX9; } |
| 243 | |
| 244 | bool hasMin3Max3_16() const { |
| 245 | return getGeneration() >= AMDGPUSubtarget::GFX9; |
| 246 | } |
| 247 | |
| 248 | bool hasSwap() const { return HasGFX9Insts; } |
| 249 | |
| 250 | bool hasScalarPackInsts() const { return HasGFX9Insts; } |
| 251 | |
| 252 | bool hasScalarMulHiInsts() const { return HasGFX9Insts; } |
| 253 | |
| 254 | bool hasScalarSubwordLoads() const { return getGeneration() >= GFX12; } |
| 255 | |
| 256 | bool hasAsyncMark() const { return hasVMemToLDSLoad() || HasAsynccnt; } |
| 257 | |
| 258 | TrapHandlerAbi getTrapHandlerAbi() const { |
| 259 | return isAmdHsaOS() ? TrapHandlerAbi::AMDHSA : TrapHandlerAbi::NONE; |
| 260 | } |
| 261 | |
| 262 | bool supportsGetDoorbellID() const { |
| 263 | // The S_GETREG DOORBELL_ID is supported by all GFX9 onward targets. |
| 264 | return getGeneration() >= GFX9; |
| 265 | } |
| 266 | |
| 267 | /// True if the offset field of DS instructions works as expected. On SI, the |
| 268 | /// offset uses a 16-bit adder and does not always wrap properly. |
| 269 | bool hasUsableDSOffset() const { return getGeneration() >= SEA_ISLANDS; } |
| 270 | |
| 271 | bool unsafeDSOffsetFoldingEnabled() const { |
| 272 | return EnableUnsafeDSOffsetFolding; |
| 273 | } |
| 274 | |
| 275 | /// Condition output from div_scale is usable. |
| 276 | bool hasUsableDivScaleConditionOutput() const { |
| 277 | return getGeneration() != SOUTHERN_ISLANDS; |
| 278 | } |
| 279 | |
| 280 | /// Extra wait hazard is needed in some cases before |
| 281 | /// s_cbranch_vccnz/s_cbranch_vccz. |
| 282 | bool hasReadVCCZBug() const { return getGeneration() <= SEA_ISLANDS; } |
| 283 | |
| 284 | /// Writes to VCC_LO/VCC_HI update the VCCZ flag. |
| 285 | bool partialVCCWritesUpdateVCCZ() const { return getGeneration() >= GFX10; } |
| 286 | |
| 287 | /// A read of an SGPR by SMRD instruction requires 4 wait states when the SGPR |
| 288 | /// was written by a VALU instruction. |
| 289 | bool hasSMRDReadVALUDefHazard() const { |
| 290 | return getGeneration() == SOUTHERN_ISLANDS; |
| 291 | } |
| 292 | |
| 293 | /// A read of an SGPR by a VMEM instruction requires 5 wait states when the |
| 294 | /// SGPR was written by a VALU Instruction. |
| 295 | bool hasVMEMReadSGPRVALUDefHazard() const { |
| 296 | return getGeneration() >= VOLCANIC_ISLANDS; |
| 297 | } |
| 298 | |
| 299 | bool hasRFEHazards() const { return getGeneration() >= VOLCANIC_ISLANDS; } |
| 300 | |
| 301 | /// Number of hazard wait states for s_setreg_b32/s_setreg_imm32_b32. |
| 302 | unsigned getSetRegWaitStates() const { |
| 303 | return getGeneration() <= SEA_ISLANDS ? 1 : 2; |
| 304 | } |
| 305 | |
| 306 | /// Return the amount of LDS that can be used that will not restrict the |
| 307 | /// occupancy lower than WaveCount. |
| 308 | unsigned getMaxLocalMemSizeWithWaveCount(unsigned WaveCount, |
| 309 | const Function &) const; |
| 310 | |
| 311 | bool supportsMinMaxDenormModes() const { |
| 312 | return getGeneration() >= AMDGPUSubtarget::GFX9; |
| 313 | } |
| 314 | |
| 315 | /// \returns If target supports S_DENORM_MODE. |
| 316 | bool hasDenormModeInst() const { |
| 317 | return getGeneration() >= AMDGPUSubtarget::GFX10; |
| 318 | } |
| 319 | |
| 320 | /// \returns If target supports ds_read/write_b128 and user enables generation |
| 321 | /// of ds_read/write_b128. |
| 322 | bool useDS128() const { return HasCIInsts && EnableDS128; } |
| 323 | |
| 324 | /// \return If target supports ds_read/write_b96/128. |
| 325 | bool hasDS96AndDS128() const { return HasCIInsts; } |
| 326 | |
| 327 | /// Have v_trunc_f64, v_ceil_f64, v_rndne_f64 |
| 328 | bool haveRoundOpsF64() const { return HasCIInsts; } |
| 329 | |
| 330 | /// \returns If MUBUF instructions always perform range checking, even for |
| 331 | /// buffer resources used for private memory access. |
| 332 | bool privateMemoryResourceIsRangeChecked() const { |
| 333 | return getGeneration() < AMDGPUSubtarget::GFX9; |
| 334 | } |
| 335 | |
| 336 | /// \returns If target requires PRT Struct NULL support (zero result registers |
| 337 | /// for sparse texture support). |
| 338 | bool usePRTStrictNull() const { return EnablePRTStrictNull; } |
| 339 | |
| 340 | bool hasUnalignedBufferAccessEnabled() const { |
| 341 | return HasUnalignedBufferAccess && HasUnalignedAccessMode; |
| 342 | } |
| 343 | |
| 344 | bool hasUnalignedDSAccessEnabled() const { |
| 345 | return HasUnalignedDSAccess && HasUnalignedAccessMode; |
| 346 | } |
| 347 | |
| 348 | bool hasUnalignedScratchAccessEnabled() const { |
| 349 | return HasUnalignedScratchAccess && HasUnalignedAccessMode; |
| 350 | } |
| 351 | |
| 352 | bool isXNACKEnabled() const { |
| 353 | return enableXNACK() || TargetID.isXnackOnOrAny(); |
| 354 | } |
| 355 | |
| 356 | bool hasRelaxedBufferOOBMode() const { return BufferOOBRelaxed; } |
| 357 | bool hasRelaxedTBufferOOBMode() const { return TBufferOOBRelaxed; } |
| 358 | |
| 359 | /// Return the width, in bits, of the num_records field of a buffer resource |
| 360 | /// (V#) on this subtarget, or std::nullopt if not yet known. |
| 361 | std::optional<unsigned> getBufferResourceNumRecordsWidth() const { |
| 362 | if (BufferResourceNumRecordsWidth == 0) |
| 363 | return std::nullopt; |
| 364 | return BufferResourceNumRecordsWidth; |
| 365 | } |
| 366 | |
| 367 | bool isCuModeEnabled() const { return EnableCuMode; } |
| 368 | |
| 369 | bool isPreciseMemoryEnabled() const { return EnablePreciseMemory; } |
| 370 | |
| 371 | bool hasFlatScrRegister() const { return hasFlatAddressSpace(); } |
| 372 | |
| 373 | // Check if target supports ST addressing mode with FLAT scratch instructions. |
| 374 | // The ST addressing mode means no registers are used, either VGPR or SGPR, |
| 375 | // but only immediate offset is swizzled and added to the FLAT scratch base. |
| 376 | bool hasFlatScratchSTMode() const { |
| 377 | return hasFlatScratchInsts() && (hasGFX10_3Insts() || hasGFX940Insts()); |
| 378 | } |
| 379 | |
| 380 | bool hasFlatScratchSVSMode() const { return HasGFX940Insts || HasGFX11Insts; } |
| 381 | |
| 382 | bool hasFlatScratchEnabled() const { |
| 383 | return hasArchitectedFlatScratch() || |
| 384 | (EnableFlatScratch && hasFlatScratchInsts()); |
| 385 | } |
| 386 | |
| 387 | bool hasGlobalAddTidInsts() const { return HasGFX10_BEncoding; } |
| 388 | |
| 389 | bool hasAtomicCSub() const { return HasGFX10_BEncoding; } |
| 390 | |
| 391 | bool hasExportInsts() const { |
| 392 | return !hasGFX940Insts() && !hasGFX1250Insts(); |
| 393 | } |
| 394 | |
| 395 | bool hasVINTERPEncoding() const { |
| 396 | return HasGFX11Insts && !hasGFX1250Insts(); |
| 397 | } |
| 398 | |
| 399 | bool hasMultiDwordFlatScratchAddressing() const { |
| 400 | return getGeneration() >= GFX9; |
| 401 | } |
| 402 | |
| 403 | bool hasFlatLgkmVMemCountInOrder() const { return getGeneration() > GFX9; } |
| 404 | |
| 405 | bool hasD16LoadStore() const { return getGeneration() >= GFX9; } |
| 406 | |
| 407 | bool d16PreservesUnusedBits() const { |
| 408 | return hasD16LoadStore() && !TargetID.isSramEccOnOrAny(); |
| 409 | } |
| 410 | |
| 411 | bool hasD16Images() const { return getGeneration() >= VOLCANIC_ISLANDS; } |
| 412 | |
| 413 | /// Return if most LDS instructions have an m0 use that require m0 to be |
| 414 | /// initialized. |
| 415 | bool ldsRequiresM0Init() const { return getGeneration() < GFX9; } |
| 416 | |
| 417 | // True if the hardware rewinds and replays GWS operations if a wave is |
| 418 | // preempted. |
| 419 | // |
| 420 | // If this is false, a GWS operation requires testing if a nack set the |
| 421 | // MEM_VIOL bit, and repeating if so. |
| 422 | bool hasGWSAutoReplay() const { return getGeneration() >= GFX9; } |
| 423 | |
| 424 | /// \returns if target has ds_gws_sema_release_all instruction. |
| 425 | bool hasGWSSemaReleaseAll() const { return HasCIInsts; } |
| 426 | |
| 427 | bool hasScalarAddSub64() const { return getGeneration() >= GFX12; } |
| 428 | |
| 429 | bool hasScalarSMulU64() const { return getGeneration() >= GFX12; } |
| 430 | |
| 431 | // Covers VS/PS/CS graphics shaders |
| 432 | bool isMesaGfxShader(const Function &F) const { |
| 433 | return isMesa3DOS() && AMDGPU::isShader(CC: F.getCallingConv()); |
| 434 | } |
| 435 | |
| 436 | bool hasMad64_32() const { return getGeneration() >= SEA_ISLANDS; } |
| 437 | |
| 438 | bool hasAtomicFaddInsts() const { |
| 439 | return HasAtomicFaddRtnInsts || HasAtomicFaddNoRtnInsts; |
| 440 | } |
| 441 | |
| 442 | bool vmemWriteNeedsExpWaitcnt() const { |
| 443 | return getGeneration() < SEA_ISLANDS; |
| 444 | } |
| 445 | |
| 446 | bool hasInstPrefetch() const { |
| 447 | return getGeneration() == GFX10 || getGeneration() == GFX11; |
| 448 | } |
| 449 | |
| 450 | bool hasPrefetch() const { return HasGFX12Insts; } |
| 451 | |
| 452 | bool hasInstPrefSize() const { return isGFX11Plus(); } |
| 453 | |
| 454 | void getInstPrefSizeArgs(uint32_t &Mask, uint32_t &Shift, uint32_t &Width, |
| 455 | uint32_t &CacheLineSize) const { |
| 456 | assert(isGFX11Plus()); |
| 457 | CacheLineSize = getInstCacheLineSize(); |
| 458 | if (getGeneration() == GFX11) { |
| 459 | Mask = amdhsa::COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE; |
| 460 | Shift = amdhsa::COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE_SHIFT; |
| 461 | Width = amdhsa::COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE_WIDTH; |
| 462 | } else { |
| 463 | Mask = amdhsa::COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE; |
| 464 | Shift = amdhsa::COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE_SHIFT; |
| 465 | Width = amdhsa::COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE_WIDTH; |
| 466 | } |
| 467 | } |
| 468 | |
| 469 | // Has s_cmpk_* instructions. |
| 470 | bool hasSCmpK() const { return getGeneration() < GFX12; } |
| 471 | |
| 472 | // Scratch is allocated in 256 dword per wave blocks for the entire |
| 473 | // wavefront. When viewed from the perspective of an arbitrary workitem, this |
| 474 | // is 4-byte aligned. |
| 475 | // |
| 476 | // Only 4-byte alignment is really needed to access anything. Transformations |
| 477 | // on the pointer value itself may rely on the alignment / known low bits of |
| 478 | // the pointer. Set this to something above the minimum to avoid needing |
| 479 | // dynamic realignment in common cases. |
| 480 | Align getStackAlignment() const { return Align(16); } |
| 481 | |
| 482 | bool enableMachineScheduler() const override { return true; } |
| 483 | |
| 484 | bool useAA() const override; |
| 485 | |
| 486 | bool enableSubRegLiveness() const override { return true; } |
| 487 | |
| 488 | void setScalarizeGlobalBehavior(bool b) { ScalarizeGlobal = b; } |
| 489 | bool getScalarizeGlobalBehavior() const { return ScalarizeGlobal; } |
| 490 | |
| 491 | // XXX - Why is this here if it isn't in the default pass set? |
| 492 | bool enableEarlyIfConversion() const override { return true; } |
| 493 | |
| 494 | void overrideSchedPolicy(MachineSchedPolicy &Policy, |
| 495 | const SchedRegion &Region) const override; |
| 496 | |
| 497 | void overridePostRASchedPolicy(MachineSchedPolicy &Policy, |
| 498 | const SchedRegion &Region) const override; |
| 499 | |
| 500 | void overridePipelinerPolicy(MachinePipelinerPolicy &Policy) const override; |
| 501 | |
| 502 | void mirFileLoaded(MachineFunction &MF) const override; |
| 503 | |
| 504 | unsigned getMaxNumUserSGPRs() const { |
| 505 | return AMDGPU::getMaxNumUserSGPRs(STI: *this); |
| 506 | } |
| 507 | |
| 508 | bool useVGPRIndexMode() const; |
| 509 | |
| 510 | bool hasScalarCompareEq64() const { |
| 511 | return getGeneration() >= VOLCANIC_ISLANDS; |
| 512 | } |
| 513 | |
| 514 | bool hasLDSFPAtomicAddF32() const { return HasGFX8Insts; } |
| 515 | bool hasLDSFPAtomicAddF64() const { |
| 516 | return HasGFX90AInsts || HasGFX1250Insts; |
| 517 | } |
| 518 | |
| 519 | /// \returns true if the subtarget has the v_permlane64_b32 instruction. |
| 520 | bool hasPermLane64() const { return getGeneration() >= GFX11; } |
| 521 | |
| 522 | /// \returns true if the subtarget supports the ds_swizzle rotate and FFT |
| 523 | /// swizzle modes (GFX9+). |
| 524 | bool hasDsSwizzleRotateMode() const { return getGeneration() >= GFX9; } |
| 525 | |
| 526 | bool hasDPPRowShare() const { |
| 527 | return HasDPP && (HasGFX90AInsts || getGeneration() >= GFX10); |
| 528 | } |
| 529 | |
| 530 | // Has V_PK_MOV_B32 opcode |
| 531 | bool hasPkMovB32() const { return HasGFX90AInsts; } |
| 532 | |
| 533 | bool hasFmaakFmamkF32Insts() const { |
| 534 | return getGeneration() >= GFX10 || hasGFX940Insts(); |
| 535 | } |
| 536 | |
| 537 | bool hasFmaakFmamkF64Insts() const { return hasGFX1250Insts(); } |
| 538 | |
| 539 | bool hasNonNSAEncoding() const { return getGeneration() < GFX12; } |
| 540 | |
| 541 | unsigned getNSAMaxSize(bool HasSampler = false) const { |
| 542 | return AMDGPU::getNSAMaxSize(STI: *this, HasSampler); |
| 543 | } |
| 544 | |
| 545 | bool hasMadF16() const; |
| 546 | |
| 547 | // Scalar and global loads support scale_offset bit. |
| 548 | bool hasScaleOffset() const { return HasGFX1250Insts; } |
| 549 | |
| 550 | // FLAT GLOBAL VOffset is signed |
| 551 | bool hasSignedGVSOffset() const { return HasGFX1250Insts; } |
| 552 | |
| 553 | bool loadStoreOptEnabled() const { return EnableLoadStoreOpt; } |
| 554 | |
| 555 | bool hasUserSGPRInit16BugInWave32() const { |
| 556 | return HasUserSGPRInit16Bug && isWave32(); |
| 557 | } |
| 558 | |
| 559 | bool has12DWordStoreHazard() const { |
| 560 | return getGeneration() != AMDGPUSubtarget::SOUTHERN_ISLANDS; |
| 561 | } |
| 562 | |
| 563 | // \returns true if the subtarget supports DWORDX3 load/store instructions. |
| 564 | bool hasDwordx3LoadStores() const { return HasCIInsts; } |
| 565 | |
| 566 | bool hasReadM0MovRelInterpHazard() const { |
| 567 | return getGeneration() == AMDGPUSubtarget::GFX9; |
| 568 | } |
| 569 | |
| 570 | bool hasReadM0SendMsgHazard() const { |
| 571 | return getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS && |
| 572 | getGeneration() <= AMDGPUSubtarget::GFX9; |
| 573 | } |
| 574 | |
| 575 | bool hasReadM0LdsDmaHazard() const { |
| 576 | return getGeneration() == AMDGPUSubtarget::GFX9; |
| 577 | } |
| 578 | |
| 579 | bool hasReadM0LdsDirectHazard() const { |
| 580 | return getGeneration() == AMDGPUSubtarget::GFX9; |
| 581 | } |
| 582 | |
| 583 | bool hasLDSMisalignedBugInWGPMode() const { |
| 584 | return HasLDSMisalignedBug && !EnableCuMode; |
| 585 | } |
| 586 | |
| 587 | // Shift amount of a 64 bit shift cannot be a highest allocated register |
| 588 | // if also at the end of the allocation block. |
| 589 | bool hasShift64HighRegBug() const { return HasGFX90AInsts; } |
| 590 | |
| 591 | // v_dot2c_f32_f16 unconditionally flushes f16 subnormal inputs to zero |
| 592 | // regardless of the MODE register, unlike v_fma_mix_f32 which respects it. |
| 593 | bool dot2UnconditionalFlush() const { |
| 594 | return HasGFX90AInsts && !HasGFX940Insts; |
| 595 | } |
| 596 | |
| 597 | // Has one cycle hazard on transcendental instruction feeding a |
| 598 | // non transcendental VALU. |
| 599 | bool hasTransForwardingHazard() const { return HasGFX940Insts; } |
| 600 | |
| 601 | // Has one cycle hazard on a VALU instruction partially writing dst with |
| 602 | // a shift of result bits feeding another VALU instruction. |
| 603 | bool hasDstSelForwardingHazard() const { return HasGFX940Insts; } |
| 604 | |
| 605 | // Cannot use op_sel with v_dot instructions. |
| 606 | bool hasDOTOpSelHazard() const { return HasGFX940Insts || HasGFX11Insts; } |
| 607 | |
| 608 | // Does not have HW interlocs for VALU writing and then reading SGPRs. |
| 609 | bool hasVDecCoExecHazard() const { return HasGFX940Insts; } |
| 610 | |
| 611 | bool hasHardClauses() const { return MaxHardClauseLength > 0; } |
| 612 | |
| 613 | bool hasFPAtomicToDenormModeHazard() const { |
| 614 | return getGeneration() == GFX10; |
| 615 | } |
| 616 | |
| 617 | bool hasVOP3DPP() const { return getGeneration() >= GFX11; } |
| 618 | |
| 619 | bool hasLdsDirect() const { return getGeneration() >= GFX11; } |
| 620 | |
| 621 | bool hasLdsWaitVMSRC() const { return getGeneration() >= GFX12; } |
| 622 | |
| 623 | bool hasVALUPartialForwardingHazard() const { |
| 624 | return getGeneration() == GFX11; |
| 625 | } |
| 626 | |
| 627 | bool hasCvtScaleForwardingHazard() const { return HasGFX950Insts; } |
| 628 | |
| 629 | // All GFX9 targets experience a fetch delay when an instruction at the start |
| 630 | // of a loop header is split by a 32-byte fetch window boundary, but GFX950 |
| 631 | // is uniquely sensitive to this: the delay triggers further performance |
| 632 | // degradation beyond the fetch latency itself. |
| 633 | bool hasLoopHeadInstSplitSensitivity() const { return HasGFX950Insts; } |
| 634 | |
| 635 | bool requiresCodeObjectV6() const { return RequiresCOV6; } |
| 636 | |
| 637 | bool useVGPRBlockOpsForCSR() const { return UseBlockVGPROpsForCSR; } |
| 638 | |
| 639 | bool hasVALUMaskWriteHazard() const { return getGeneration() == GFX11; } |
| 640 | |
| 641 | bool hasVALUReadSGPRHazard() const { |
| 642 | return HasGFX12Insts && !HasGFX1250Insts; |
| 643 | } |
| 644 | |
| 645 | bool setRegModeNeedsVNOPs() const { |
| 646 | return HasGFX1250Insts && getGeneration() == GFX12; |
| 647 | } |
| 648 | |
| 649 | /// Return if operations acting on VGPR tuples require even alignment. |
| 650 | bool needsAlignedVGPRs() const { return RequiresAlignVGPR; } |
| 651 | |
| 652 | /// Return true if the target has the S_PACK_HL_B32_B16 instruction. |
| 653 | bool hasSPackHL() const { return HasGFX11Insts; } |
| 654 | |
| 655 | /// Return true if the target has the V_CVT_PK_I16_F32/V_CVT_PK_U16_F32 |
| 656 | /// instructions. |
| 657 | bool hasVCvtPkIU16F32() const { return HasGFX11Insts; } |
| 658 | |
| 659 | /// Return true if the target's EXP instruction supports the NULL export |
| 660 | /// target. |
| 661 | bool hasNullExportTarget() const { return !HasGFX11Insts; } |
| 662 | |
| 663 | bool hasFlatScratchSVSSwizzleBug() const { return getGeneration() == GFX11; } |
| 664 | |
| 665 | /// Return true if the target has the S_DELAY_ALU instruction. |
| 666 | bool hasDelayAlu() const { return HasGFX11Insts; } |
| 667 | |
| 668 | /// Returns true if the target supports |
| 669 | /// global_load_lds_dwordx3/global_load_lds_dwordx4 or |
| 670 | /// buffer_load_dwordx3/buffer_load_dwordx4 with the lds bit. |
| 671 | bool hasLDSLoadB96_B128() const { return hasGFX950Insts(); } |
| 672 | |
| 673 | /// \returns true if the target uses LOADcnt/SAMPLEcnt/BVHcnt, DScnt/KMcnt |
| 674 | /// and STOREcnt rather than VMcnt, LGKMcnt and VScnt respectively. |
| 675 | bool hasExtendedWaitCounts() const { return getGeneration() >= GFX12; } |
| 676 | |
| 677 | /// \returns true if the target has packed f32 instructions that only read 32 |
| 678 | /// bits from a scalar operand (SGPR or literal) and replicates the bits to |
| 679 | /// both channels. |
| 680 | bool hasPKF32InstsReplicatingLower32BitsOfScalarInput() const { |
| 681 | return getGeneration() == GFX12 && HasGFX1250Insts; |
| 682 | } |
| 683 | |
| 684 | bool hasAddPC64Inst() const { return HasGFX1250Insts; } |
| 685 | |
| 686 | /// \returns true if the target supports expert scheduling mode 2 which relies |
| 687 | /// on the compiler to insert waits to avoid hazards between VMEM and VALU |
| 688 | /// instructions in some instances. |
| 689 | bool hasExpertSchedulingMode() const { return getGeneration() >= GFX12; } |
| 690 | |
| 691 | /// \returns The maximum number of instructions that can be enclosed in an |
| 692 | /// S_CLAUSE on the given subtarget, or 0 for targets that do not support that |
| 693 | /// instruction. |
| 694 | unsigned maxHardClauseLength() const { return MaxHardClauseLength; } |
| 695 | |
| 696 | /// Return the maximum number of waves per SIMD for kernels using \p SGPRs |
| 697 | /// SGPRs |
| 698 | unsigned getOccupancyWithNumSGPRs(unsigned SGPRs) const; |
| 699 | |
| 700 | /// Return the maximum number of waves per SIMD for kernels using \p VGPRs |
| 701 | /// VGPRs |
| 702 | unsigned getOccupancyWithNumVGPRs(unsigned VGPRs, |
| 703 | unsigned DynamicVGPRBlockSize) const; |
| 704 | |
| 705 | /// Subtarget's minimum/maximum occupancy, in number of waves per EU, that can |
| 706 | /// be achieved when the only function running on a CU is \p F, each workgroup |
| 707 | /// uses \p LDSSize bytes of LDS, and each wave uses \p NumSGPRs SGPRs and \p |
| 708 | /// NumVGPRs VGPRs. The flat workgroup sizes associated to the function are a |
| 709 | /// range, so this returns a range as well. |
| 710 | /// |
| 711 | /// Note that occupancy can be affected by the scratch allocation as well, but |
| 712 | /// we do not have enough information to compute it. |
| 713 | std::pair<unsigned, unsigned> computeOccupancy(const Function &F, |
| 714 | unsigned LDSSize = 0, |
| 715 | unsigned NumSGPRs = 0, |
| 716 | unsigned NumVGPRs = 0) const; |
| 717 | |
| 718 | /// \returns true if the flat_scratch register should be initialized with the |
| 719 | /// pointer to the wave's scratch memory rather than a size and offset. |
| 720 | bool flatScratchIsPointer() const { |
| 721 | return getGeneration() >= AMDGPUSubtarget::GFX9; |
| 722 | } |
| 723 | |
| 724 | /// \returns true if the machine has merged shaders in which s0-s7 are |
| 725 | /// reserved by the hardware and user SGPRs start at s8 |
| 726 | bool hasMergedShaders() const { return getGeneration() >= GFX9; } |
| 727 | |
| 728 | // \returns true if the target supports the pre-NGG legacy geometry path. |
| 729 | bool hasLegacyGeometry() const { return getGeneration() < GFX11; } |
| 730 | |
| 731 | // \returns true if the target has split barriers feature |
| 732 | bool hasSplitBarriers() const { return getGeneration() >= GFX12; } |
| 733 | |
| 734 | // \returns true if the target has WG_RR_MODE kernel descriptor mode bit |
| 735 | bool hasRrWGMode() const { return getGeneration() >= GFX12; } |
| 736 | |
| 737 | /// \returns true if VADDR and SADDR fields in VSCRATCH can use negative |
| 738 | /// values. |
| 739 | bool hasSignedScratchOffsets() const { return getGeneration() >= GFX12; } |
| 740 | |
| 741 | bool hasINVWBL2WaitCntRequirement() const { return HasGFX1250Insts; } |
| 742 | |
| 743 | bool hasVOPD3() const { return HasGFX1250Insts; } |
| 744 | |
| 745 | // \returns true if the target has V_PK_{MIN|MAX}3_{I|U}16 instructions. |
| 746 | bool hasPkMinMax3Insts() const { return HasGFX1250Insts; } |
| 747 | |
| 748 | // \returns ture if target has S_GET_SHADER_CYCLES_U64 instruction. |
| 749 | bool hasSGetShaderCyclesInst() const { return HasGFX1250Insts; } |
| 750 | |
| 751 | // \returns true if S_GETPC_B64 zero-extends the result from 48 bits instead |
| 752 | // of sign-extending. Note that GFX1250 has not only fixed the bug but also |
| 753 | // extended VA to 57 bits. |
| 754 | bool hasGetPCZeroExtension() const { |
| 755 | return HasGFX12Insts && !HasGFX1250Insts; |
| 756 | } |
| 757 | |
| 758 | // \returns true if the target needs to create a prolog for backward |
| 759 | // compatibility when preloading kernel arguments. |
| 760 | bool needsKernArgPreloadProlog() const { |
| 761 | return hasKernargPreload() && !HasGFX1250Insts; |
| 762 | } |
| 763 | |
| 764 | bool hasCondSubInsts() const { return HasGFX12Insts; } |
| 765 | |
| 766 | bool hasSubClampInsts() const { return hasGFX10_3Insts(); } |
| 767 | |
| 768 | bool hasAnyPackedFP32Ops() const { |
| 769 | return hasPackedFP32Ops() || hasPackedFP32SingleSGPROps(); |
| 770 | }; |
| 771 | |
| 772 | bool hasAnyPackedFP64Ops() const { return hasPackedFP64SingleSGPROps(); }; |
| 773 | |
| 774 | bool hasAnyPackedU64Ops() const { return hasPackedU64SingleSGPROps(); }; |
| 775 | |
| 776 | /// \returns SGPR allocation granularity supported by the subtarget. |
| 777 | unsigned getSGPRAllocGranule() const { |
| 778 | return AMDGPU::getSGPRAllocGranule(AK: getTargetID().getGPUKind()); |
| 779 | } |
| 780 | |
| 781 | /// \returns SGPR encoding granularity supported by the subtarget. |
| 782 | unsigned getSGPREncodingGranule() const { |
| 783 | return AMDGPU::IsaInfo::getSGPREncodingGranule(STI: *this); |
| 784 | } |
| 785 | |
| 786 | /// \returns Total number of SGPRs supported by the subtarget. |
| 787 | unsigned getTotalNumSGPRs() const { |
| 788 | return AMDGPU::getTotalNumSGPRs(AK: getTargetID().getGPUKind()); |
| 789 | } |
| 790 | |
| 791 | /// \returns Addressable number of SGPRs supported by the subtarget. |
| 792 | unsigned getAddressableNumSGPRs() const { |
| 793 | return AMDGPU::getAddressableNumSGPRs(AK: getTargetID().getGPUKind()); |
| 794 | } |
| 795 | |
| 796 | /// \returns Minimum number of SGPRs that meets the given number of waves per |
| 797 | /// execution unit requirement supported by the subtarget. |
| 798 | unsigned getMinNumSGPRs(unsigned WavesPerEU) const { |
| 799 | return AMDGPU::IsaInfo::getMinNumSGPRs(STI: *this, WavesPerEU); |
| 800 | } |
| 801 | |
| 802 | /// \returns Maximum number of SGPRs that meets the given number of waves per |
| 803 | /// execution unit requirement supported by the subtarget. |
| 804 | unsigned getMaxNumSGPRs(unsigned WavesPerEU, bool Addressable) const { |
| 805 | return AMDGPU::IsaInfo::getMaxNumSGPRs(STI: *this, WavesPerEU, Addressable); |
| 806 | } |
| 807 | |
| 808 | /// \returns Reserved number of SGPRs. This is common |
| 809 | /// utility function called by MachineFunction and |
| 810 | /// Function variants of getReservedNumSGPRs. |
| 811 | unsigned getBaseReservedNumSGPRs(const bool HasFlatScratch) const; |
| 812 | /// \returns Reserved number of SGPRs for given machine function \p MF. |
| 813 | unsigned getReservedNumSGPRs(const MachineFunction &MF) const; |
| 814 | |
| 815 | /// \returns Reserved number of SGPRs for given function \p F. |
| 816 | unsigned getReservedNumSGPRs(const Function &F) const; |
| 817 | |
| 818 | /// \returns Maximum number of preloaded SGPRs for the subtarget. |
| 819 | unsigned getMaxNumPreloadedSGPRs() const; |
| 820 | |
| 821 | /// \returns max num SGPRs. This is the common utility |
| 822 | /// function called by MachineFunction and Function |
| 823 | /// variants of getMaxNumSGPRs. |
| 824 | unsigned getBaseMaxNumSGPRs(const Function &F, |
| 825 | std::pair<unsigned, unsigned> WavesPerEU, |
| 826 | unsigned PreloadedSGPRs, |
| 827 | unsigned ReservedNumSGPRs) const; |
| 828 | |
| 829 | /// \returns Maximum number of SGPRs that meets number of waves per execution |
| 830 | /// unit requirement for function \p MF, or number of SGPRs explicitly |
| 831 | /// requested using "amdgpu-num-sgpr" attribute attached to function \p MF. |
| 832 | /// |
| 833 | /// \returns Value that meets number of waves per execution unit requirement |
| 834 | /// if explicitly requested value cannot be converted to integer, violates |
| 835 | /// subtarget's specifications, or does not meet number of waves per execution |
| 836 | /// unit requirement. |
| 837 | unsigned getMaxNumSGPRs(const MachineFunction &MF) const; |
| 838 | |
| 839 | /// \returns Maximum number of SGPRs that meets number of waves per execution |
| 840 | /// unit requirement for function \p F, or number of SGPRs explicitly |
| 841 | /// requested using "amdgpu-num-sgpr" attribute attached to function \p F. |
| 842 | /// |
| 843 | /// \returns Value that meets number of waves per execution unit requirement |
| 844 | /// if explicitly requested value cannot be converted to integer, violates |
| 845 | /// subtarget's specifications, or does not meet number of waves per execution |
| 846 | /// unit requirement. |
| 847 | unsigned getMaxNumSGPRs(const Function &F) const; |
| 848 | |
| 849 | /// \returns VGPR allocation granularity supported by the subtarget. |
| 850 | unsigned getVGPRAllocGranule(unsigned DynamicVGPRBlockSize) const { |
| 851 | return AMDGPU::IsaInfo::getVGPRAllocGranule(STI: *this, DynamicVGPRBlockSize); |
| 852 | } |
| 853 | |
| 854 | /// \returns VGPR encoding granularity supported by the subtarget. |
| 855 | unsigned getVGPREncodingGranule() const { |
| 856 | return AMDGPU::IsaInfo::getVGPREncodingGranule(STI: *this); |
| 857 | } |
| 858 | |
| 859 | /// \returns Total number of VGPRs supported by the subtarget. |
| 860 | unsigned getTotalNumVGPRs() const { |
| 861 | return AMDGPU::IsaInfo::getTotalNumVGPRs(STI: *this); |
| 862 | } |
| 863 | |
| 864 | /// \returns Addressable number of architectural VGPRs supported by the |
| 865 | /// subtarget. |
| 866 | unsigned getAddressableNumArchVGPRs() const { |
| 867 | return AMDGPU::IsaInfo::getAddressableNumArchVGPRs(STI: *this); |
| 868 | } |
| 869 | |
| 870 | /// \returns Addressable number of VGPRs supported by the subtarget. |
| 871 | unsigned getAddressableNumVGPRs(unsigned DynamicVGPRBlockSize) const { |
| 872 | return AMDGPU::IsaInfo::getAddressableNumVGPRs(STI: *this, DynamicVGPRBlockSize); |
| 873 | } |
| 874 | |
| 875 | /// \returns the minimum number of VGPRs that will prevent achieving more than |
| 876 | /// the specified number of waves \p WavesPerEU. |
| 877 | unsigned getMinNumVGPRs(unsigned WavesPerEU, |
| 878 | unsigned DynamicVGPRBlockSize) const { |
| 879 | return AMDGPU::IsaInfo::getMinNumVGPRs(STI: *this, WavesPerEU, |
| 880 | DynamicVGPRBlockSize); |
| 881 | } |
| 882 | |
| 883 | /// \returns the maximum number of VGPRs that can be used and still achieved |
| 884 | /// at least the specified number of waves \p WavesPerEU. |
| 885 | unsigned getMaxNumVGPRs(unsigned WavesPerEU, |
| 886 | unsigned DynamicVGPRBlockSize) const { |
| 887 | return AMDGPU::IsaInfo::getMaxNumVGPRs(STI: *this, WavesPerEU, |
| 888 | DynamicVGPRBlockSize); |
| 889 | } |
| 890 | |
| 891 | /// \returns max num VGPRs. This is the common utility function |
| 892 | /// called by MachineFunction and Function variants of getMaxNumVGPRs. |
| 893 | unsigned |
| 894 | getBaseMaxNumVGPRs(const Function &F, |
| 895 | std::pair<unsigned, unsigned> NumVGPRBounds) const; |
| 896 | |
| 897 | /// \returns Maximum number of VGPRs that meets number of waves per execution |
| 898 | /// unit requirement for function \p F, or number of VGPRs explicitly |
| 899 | /// requested using "amdgpu-num-vgpr" attribute attached to function \p F. |
| 900 | /// |
| 901 | /// \returns Value that meets number of waves per execution unit requirement |
| 902 | /// if explicitly requested value cannot be converted to integer, violates |
| 903 | /// subtarget's specifications, or does not meet number of waves per execution |
| 904 | /// unit requirement. |
| 905 | unsigned getMaxNumVGPRs(const Function &F) const; |
| 906 | |
| 907 | unsigned getMaxNumAGPRs(const Function &F) const { return getMaxNumVGPRs(F); } |
| 908 | |
| 909 | /// Return a pair of maximum numbers of VGPRs and AGPRs that meet the number |
| 910 | /// of waves per execution unit required for the function \p MF. |
| 911 | std::pair<unsigned, unsigned> getMaxNumVectorRegs(const Function &F) const; |
| 912 | |
| 913 | /// \returns Maximum number of VGPRs that meets number of waves per execution |
| 914 | /// unit requirement for function \p MF, or number of VGPRs explicitly |
| 915 | /// requested using "amdgpu-num-vgpr" attribute attached to function \p MF. |
| 916 | /// |
| 917 | /// \returns Value that meets number of waves per execution unit requirement |
| 918 | /// if explicitly requested value cannot be converted to integer, violates |
| 919 | /// subtarget's specifications, or does not meet number of waves per execution |
| 920 | /// unit requirement. |
| 921 | unsigned getMaxNumVGPRs(const MachineFunction &MF) const; |
| 922 | |
| 923 | bool isWave32() const { return getWavefrontSize() == 32; } |
| 924 | |
| 925 | bool isWave64() const { return getWavefrontSize() == 64; } |
| 926 | |
| 927 | /// Returns if the wavesize of this subtarget is known reliable. This is false |
| 928 | /// only for the a default target-cpu that does not have an explicit |
| 929 | /// +wavefrontsize target feature. |
| 930 | bool isWaveSizeKnown() const { |
| 931 | return hasFeature(Feature: AMDGPU::FeatureWavefrontSize32) || |
| 932 | hasFeature(Feature: AMDGPU::FeatureWavefrontSize64); |
| 933 | } |
| 934 | |
| 935 | const TargetRegisterClass *getBoolRC() const { |
| 936 | return getRegisterInfo()->getBoolRC(); |
| 937 | } |
| 938 | |
| 939 | /// \returns Maximum number of work groups per compute unit supported by the |
| 940 | /// subtarget and limited by given \p FlatWorkGroupSize. |
| 941 | unsigned getMaxWorkGroupsPerCU(unsigned FlatWorkGroupSize) const override { |
| 942 | return AMDGPU::IsaInfo::getMaxWorkGroupsPerCU(STI: *this, FlatWorkGroupSize); |
| 943 | } |
| 944 | |
| 945 | /// \returns Minimum flat work group size supported by the subtarget. |
| 946 | unsigned getMinFlatWorkGroupSize() const override { |
| 947 | return AMDGPU::IsaInfo::getMinFlatWorkGroupSize(STI: *this); |
| 948 | } |
| 949 | |
| 950 | /// \returns Maximum flat work group size supported by the subtarget. |
| 951 | unsigned getMaxFlatWorkGroupSize() const override { |
| 952 | return AMDGPU::IsaInfo::getMaxFlatWorkGroupSize(); |
| 953 | } |
| 954 | |
| 955 | /// \returns Number of waves per execution unit required to support the given |
| 956 | /// \p FlatWorkGroupSize. |
| 957 | unsigned |
| 958 | getWavesPerEUForWorkGroup(unsigned FlatWorkGroupSize) const override { |
| 959 | return AMDGPU::IsaInfo::getWavesPerEUForWorkGroup(STI: *this, FlatWorkGroupSize); |
| 960 | } |
| 961 | |
| 962 | /// \returns Minimum number of waves per execution unit supported by the |
| 963 | /// subtarget. |
| 964 | unsigned getMinWavesPerEU() const override { |
| 965 | return AMDGPU::getMinWavesPerEU(); |
| 966 | } |
| 967 | |
| 968 | void adjustSchedDependency(SUnit *Def, int DefOpIdx, SUnit *Use, int UseOpIdx, |
| 969 | SDep &Dep, |
| 970 | const TargetSchedModel *SchedModel) const override; |
| 971 | |
| 972 | // \returns true if it's beneficial on this subtarget for the scheduler to |
| 973 | // cluster stores as well as loads. |
| 974 | bool shouldClusterStores() const { return getGeneration() >= GFX11; } |
| 975 | |
| 976 | // \returns the number of address arguments from which to enable MIMG NSA |
| 977 | // on supported architectures. |
| 978 | unsigned getNSAThreshold(const MachineFunction &MF) const; |
| 979 | |
| 980 | // \returns true if the subtarget has a hazard requiring an "s_nop 0" |
| 981 | // instruction before "s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)". |
| 982 | bool requiresNopBeforeDeallocVGPRs() const { return !HasGFX1250Insts; } |
| 983 | |
| 984 | // \returns true if the subtarget needs S_WAIT_ALU 0 before S_GETREG_B32 on |
| 985 | // STATUS, STATE_PRIV, EXCP_FLAG_PRIV, or EXCP_FLAG_USER. |
| 986 | bool requiresWaitIdleBeforeGetReg() const { return HasGFX1250Insts; } |
| 987 | |
| 988 | bool requiresDisjointEarlyClobberAndUndef() const override { |
| 989 | // AMDGPU doesn't care if early-clobber and undef operands are allocated |
| 990 | // to the same register. |
| 991 | return false; |
| 992 | } |
| 993 | |
| 994 | // DS_ATOMIC_ASYNC_BARRIER_ARRIVE_B64 shall not be claused with anything |
| 995 | // and surronded by S_WAIT_ALU(0xFFE3). |
| 996 | bool hasDsAtomicAsyncBarrierArriveB64PipeBug() const { |
| 997 | return getGeneration() == GFX12; |
| 998 | } |
| 999 | |
| 1000 | // Requires s_wait_alu(0) after s102/s103 write and src_flat_scratch_base |
| 1001 | // read. |
| 1002 | bool hasScratchBaseForwardingHazard() const { |
| 1003 | return HasGFX1250Insts && getGeneration() == GFX12; |
| 1004 | } |
| 1005 | |
| 1006 | // src_flat_scratch_hi cannot be used as a source in SALU producing a 64-bit |
| 1007 | // result. |
| 1008 | bool hasFlatScratchHiInB64InstHazard() const { |
| 1009 | return HasGFX1250Insts && getGeneration() == GFX12; |
| 1010 | } |
| 1011 | |
| 1012 | /// \returns true if the subtarget requires a wait for xcnt before VMEM |
| 1013 | /// accesses that must never be repeated in the event of a page fault/re-try. |
| 1014 | /// Atomic stores/rmw and all volatile accesses fall under this criteria. |
| 1015 | bool requiresWaitXCntForSingleAccessInstructions() const { |
| 1016 | return HasGFX1250Insts; |
| 1017 | } |
| 1018 | |
| 1019 | /// \returns the number of significant bits in the immediate field of the |
| 1020 | /// S_NOP instruction. |
| 1021 | unsigned getSNopBits() const { |
| 1022 | if (getGeneration() >= AMDGPUSubtarget::GFX12) |
| 1023 | return 7; |
| 1024 | if (getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) |
| 1025 | return 4; |
| 1026 | return 3; |
| 1027 | } |
| 1028 | |
| 1029 | bool supportsBPermute() const { |
| 1030 | return getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS; |
| 1031 | } |
| 1032 | |
| 1033 | bool supportsWaveWideBPermute() const { |
| 1034 | return (getGeneration() <= AMDGPUSubtarget::GFX9 || |
| 1035 | getGeneration() == AMDGPUSubtarget::GFX12) || |
| 1036 | isWave32(); |
| 1037 | } |
| 1038 | |
| 1039 | /// Return true if real (non-fake) variants of True16 instructions using |
| 1040 | /// 16-bit registers should be code-generated. Fake True16 instructions are |
| 1041 | /// identical to non-fake ones except that they take 32-bit registers as |
| 1042 | /// operands and always use their low halves. |
| 1043 | // TODO: Remove and use hasTrue16BitInsts() instead once True16 is fully |
| 1044 | // supported and the support for fake True16 instructions is removed. |
| 1045 | bool useRealTrue16Insts() const { |
| 1046 | return hasTrue16BitInsts() && EnableRealTrue16Insts; |
| 1047 | } |
| 1048 | |
| 1049 | bool requiresWaitOnWorkgroupReleaseFence(bool TgSplit) const { |
| 1050 | return getGeneration() >= GFX10 || TgSplit; |
| 1051 | } |
| 1052 | |
| 1053 | bool useDFAforSMS() const override { return false; } |
| 1054 | |
| 1055 | bool enableWindowScheduler() const override { return false; } |
| 1056 | }; |
| 1057 | |
| 1058 | class GCNUserSGPRUsageInfo { |
| 1059 | public: |
| 1060 | bool hasImplicitBufferPtr() const { return ImplicitBufferPtr; } |
| 1061 | |
| 1062 | bool hasPrivateSegmentBuffer() const { return PrivateSegmentBuffer; } |
| 1063 | |
| 1064 | bool hasDispatchPtr() const { return DispatchPtr; } |
| 1065 | |
| 1066 | bool hasQueuePtr() const { return QueuePtr; } |
| 1067 | |
| 1068 | bool hasKernargSegmentPtr() const { return KernargSegmentPtr; } |
| 1069 | |
| 1070 | bool hasDispatchID() const { return DispatchID; } |
| 1071 | |
| 1072 | bool hasFlatScratchInit() const { return FlatScratchInit; } |
| 1073 | |
| 1074 | bool hasPrivateSegmentSize() const { return PrivateSegmentSize; } |
| 1075 | |
| 1076 | unsigned getNumKernargPreloadSGPRs() const { return NumKernargPreloadSGPRs; } |
| 1077 | |
| 1078 | unsigned getNumUsedUserSGPRs() const { return NumUsedUserSGPRs; } |
| 1079 | |
| 1080 | unsigned getNumFreeUserSGPRs(); |
| 1081 | |
| 1082 | void allocKernargPreloadSGPRs(unsigned NumSGPRs); |
| 1083 | |
| 1084 | enum UserSGPRID : unsigned { |
| 1085 | ImplicitBufferPtrID = 0, |
| 1086 | PrivateSegmentBufferID = 1, |
| 1087 | DispatchPtrID = 2, |
| 1088 | QueuePtrID = 3, |
| 1089 | KernargSegmentPtrID = 4, |
| 1090 | DispatchIdID = 5, |
| 1091 | FlatScratchInitID = 6, |
| 1092 | PrivateSegmentSizeID = 7 |
| 1093 | }; |
| 1094 | |
| 1095 | // Returns the size in number of SGPRs for preload user SGPR field. |
| 1096 | static unsigned getNumUserSGPRForField(UserSGPRID ID) { |
| 1097 | switch (ID) { |
| 1098 | case ImplicitBufferPtrID: |
| 1099 | return 2; |
| 1100 | case PrivateSegmentBufferID: |
| 1101 | return 4; |
| 1102 | case DispatchPtrID: |
| 1103 | return 2; |
| 1104 | case QueuePtrID: |
| 1105 | return 2; |
| 1106 | case KernargSegmentPtrID: |
| 1107 | return 2; |
| 1108 | case DispatchIdID: |
| 1109 | return 2; |
| 1110 | case FlatScratchInitID: |
| 1111 | return 2; |
| 1112 | case PrivateSegmentSizeID: |
| 1113 | return 1; |
| 1114 | } |
| 1115 | llvm_unreachable("Unknown UserSGPRID." ); |
| 1116 | } |
| 1117 | |
| 1118 | GCNUserSGPRUsageInfo(const Function &F, const GCNSubtarget &ST); |
| 1119 | |
| 1120 | private: |
| 1121 | const GCNSubtarget &ST; |
| 1122 | |
| 1123 | // Private memory buffer |
| 1124 | // Compute directly in sgpr[0:1] |
| 1125 | // Other shaders indirect 64-bits at sgpr[0:1] |
| 1126 | bool ImplicitBufferPtr = false; |
| 1127 | |
| 1128 | bool PrivateSegmentBuffer = false; |
| 1129 | |
| 1130 | bool DispatchPtr = false; |
| 1131 | |
| 1132 | bool QueuePtr = false; |
| 1133 | |
| 1134 | bool KernargSegmentPtr = false; |
| 1135 | |
| 1136 | bool DispatchID = false; |
| 1137 | |
| 1138 | bool FlatScratchInit = false; |
| 1139 | |
| 1140 | bool PrivateSegmentSize = false; |
| 1141 | |
| 1142 | unsigned NumKernargPreloadSGPRs = 0; |
| 1143 | |
| 1144 | unsigned NumUsedUserSGPRs = 0; |
| 1145 | }; |
| 1146 | |
| 1147 | } // end namespace llvm |
| 1148 | |
| 1149 | #endif // LLVM_LIB_TARGET_AMDGPU_GCNSUBTARGET_H |
| 1150 | |