| 1 | //===- AMDGPUAttributor.cpp -----------------------------------------------===// |
| 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 This pass uses Attributor framework to deduce AMDGPU attributes. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "AMDGPU.h" |
| 14 | #include "GCNSubtarget.h" |
| 15 | #include "Utils/AMDGPUBaseInfo.h" |
| 16 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 17 | #include "llvm/IR/IntrinsicsR600.h" |
| 18 | #include "llvm/Target/TargetMachine.h" |
| 19 | #include "llvm/Transforms/IPO/Attributor.h" |
| 20 | |
| 21 | #define DEBUG_TYPE "amdgpu-attributor" |
| 22 | |
| 23 | using namespace llvm; |
| 24 | |
| 25 | static cl::opt<unsigned> IndirectCallSpecializationThreshold( |
| 26 | "amdgpu-indirect-call-specialization-threshold" , |
| 27 | cl::desc( |
| 28 | "A threshold controls whether an indirect call will be specialized" ), |
| 29 | cl::init(Val: 3)); |
| 30 | |
| 31 | #define AMDGPU_ATTRIBUTE(Name, Str) Name##_POS, |
| 32 | |
| 33 | enum ImplicitArgumentPositions { |
| 34 | #include "AMDGPUAttributes.def" |
| 35 | LAST_ARG_POS |
| 36 | }; |
| 37 | |
| 38 | #define AMDGPU_ATTRIBUTE(Name, Str) Name = 1 << Name##_POS, |
| 39 | |
| 40 | enum ImplicitArgumentMask { |
| 41 | UNKNOWN_INTRINSIC = 0, |
| 42 | #include "AMDGPUAttributes.def" |
| 43 | ALL_ARGUMENT_MASK = (1 << LAST_ARG_POS) - 1, |
| 44 | NOT_IMPLICIT_INPUT |
| 45 | }; |
| 46 | |
| 47 | #define AMDGPU_ATTRIBUTE(Name, Str) {Name, Str}, |
| 48 | static constexpr std::pair<ImplicitArgumentMask, StringLiteral> |
| 49 | ImplicitAttrs[] = { |
| 50 | #include "AMDGPUAttributes.def" |
| 51 | }; |
| 52 | |
| 53 | // We do not need to note the x workitem or workgroup id because they are always |
| 54 | // initialized. |
| 55 | // |
| 56 | // TODO: We should not add the attributes if the known compile time workgroup |
| 57 | // size is 1 for y/z. |
| 58 | static ImplicitArgumentMask |
| 59 | intrinsicToAttrMask(Intrinsic::ID ID, bool &NonKernelOnly, bool &NeedsImplicit, |
| 60 | bool HasApertureRegs, bool SupportsGetDoorBellID, |
| 61 | unsigned CodeObjectVersion) { |
| 62 | switch (ID) { |
| 63 | case Intrinsic::amdgcn_workitem_id_x: |
| 64 | NonKernelOnly = true; |
| 65 | return WORKITEM_ID_X; |
| 66 | case Intrinsic::amdgcn_workgroup_id_x: |
| 67 | NonKernelOnly = true; |
| 68 | return WORKGROUP_ID_X; |
| 69 | case Intrinsic::amdgcn_workitem_id_y: |
| 70 | case Intrinsic::r600_read_tidig_y: |
| 71 | return WORKITEM_ID_Y; |
| 72 | case Intrinsic::amdgcn_workitem_id_z: |
| 73 | case Intrinsic::r600_read_tidig_z: |
| 74 | return WORKITEM_ID_Z; |
| 75 | case Intrinsic::amdgcn_workgroup_id_y: |
| 76 | case Intrinsic::r600_read_tgid_y: |
| 77 | return WORKGROUP_ID_Y; |
| 78 | case Intrinsic::amdgcn_workgroup_id_z: |
| 79 | case Intrinsic::r600_read_tgid_z: |
| 80 | return WORKGROUP_ID_Z; |
| 81 | case Intrinsic::amdgcn_cluster_id_x: |
| 82 | NonKernelOnly = true; |
| 83 | return CLUSTER_ID_X; |
| 84 | case Intrinsic::amdgcn_cluster_id_y: |
| 85 | return CLUSTER_ID_Y; |
| 86 | case Intrinsic::amdgcn_cluster_id_z: |
| 87 | return CLUSTER_ID_Z; |
| 88 | case Intrinsic::amdgcn_lds_kernel_id: |
| 89 | return LDS_KERNEL_ID; |
| 90 | case Intrinsic::amdgcn_dispatch_ptr: |
| 91 | return DISPATCH_PTR; |
| 92 | case Intrinsic::amdgcn_dispatch_id: |
| 93 | return DISPATCH_ID; |
| 94 | case Intrinsic::amdgcn_implicitarg_ptr: |
| 95 | return IMPLICIT_ARG_PTR; |
| 96 | // Need queue_ptr anyway. But under V5, we also need implicitarg_ptr to access |
| 97 | // queue_ptr. |
| 98 | case Intrinsic::amdgcn_queue_ptr: |
| 99 | NeedsImplicit = (CodeObjectVersion >= AMDGPU::AMDHSA_COV5); |
| 100 | return QUEUE_PTR; |
| 101 | case Intrinsic::amdgcn_is_shared: |
| 102 | case Intrinsic::amdgcn_is_private: |
| 103 | if (HasApertureRegs) |
| 104 | return NOT_IMPLICIT_INPUT; |
| 105 | // Under V5, we need implicitarg_ptr + offsets to access private_base or |
| 106 | // shared_base. For pre-V5, however, need to access them through queue_ptr + |
| 107 | // offsets. |
| 108 | return CodeObjectVersion >= AMDGPU::AMDHSA_COV5 ? IMPLICIT_ARG_PTR |
| 109 | : QUEUE_PTR; |
| 110 | case Intrinsic::trap: |
| 111 | case Intrinsic::debugtrap: |
| 112 | case Intrinsic::ubsantrap: |
| 113 | if (SupportsGetDoorBellID) // GetDoorbellID support implemented since V4. |
| 114 | return CodeObjectVersion >= AMDGPU::AMDHSA_COV4 ? NOT_IMPLICIT_INPUT |
| 115 | : QUEUE_PTR; |
| 116 | NeedsImplicit = (CodeObjectVersion >= AMDGPU::AMDHSA_COV5); |
| 117 | return QUEUE_PTR; |
| 118 | default: |
| 119 | return UNKNOWN_INTRINSIC; |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | static bool castRequiresQueuePtr(unsigned SrcAS) { |
| 124 | return SrcAS == AMDGPUAS::LOCAL_ADDRESS || SrcAS == AMDGPUAS::PRIVATE_ADDRESS; |
| 125 | } |
| 126 | |
| 127 | static bool isDSAddress(const Constant *C) { |
| 128 | const GlobalValue *GV = dyn_cast<GlobalValue>(Val: C); |
| 129 | if (!GV) |
| 130 | return false; |
| 131 | unsigned AS = GV->getAddressSpace(); |
| 132 | return AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS; |
| 133 | } |
| 134 | |
| 135 | /// Returns true if sanitizer attributes are present on a function. |
| 136 | static bool hasSanitizerAttributes(const Function &F) { |
| 137 | return F.hasFnAttribute(Kind: Attribute::SanitizeAddress) || |
| 138 | F.hasFnAttribute(Kind: Attribute::SanitizeThread) || |
| 139 | F.hasFnAttribute(Kind: Attribute::SanitizeMemory) || |
| 140 | F.hasFnAttribute(Kind: Attribute::SanitizeHWAddress) || |
| 141 | F.hasFnAttribute(Kind: Attribute::SanitizeMemTag); |
| 142 | } |
| 143 | |
| 144 | namespace { |
| 145 | class AMDGPUInformationCache : public InformationCache { |
| 146 | public: |
| 147 | AMDGPUInformationCache(const Module &M, AnalysisGetter &AG, |
| 148 | BumpPtrAllocator &Allocator, |
| 149 | SetVector<Function *> *CGSCC, TargetMachine &TM) |
| 150 | : InformationCache(M, AG, Allocator, CGSCC), TM(TM), |
| 151 | CodeObjectVersion(AMDGPU::getAMDHSACodeObjectVersion(M)) {} |
| 152 | |
| 153 | TargetMachine &TM; |
| 154 | |
| 155 | enum ConstantStatus : uint8_t { |
| 156 | NONE = 0, |
| 157 | DS_GLOBAL = 1 << 0, |
| 158 | ADDR_SPACE_CAST_PRIVATE_TO_FLAT = 1 << 1, |
| 159 | ADDR_SPACE_CAST_LOCAL_TO_FLAT = 1 << 2, |
| 160 | ADDR_SPACE_CAST_BOTH_TO_FLAT = |
| 161 | ADDR_SPACE_CAST_PRIVATE_TO_FLAT | ADDR_SPACE_CAST_LOCAL_TO_FLAT |
| 162 | }; |
| 163 | |
| 164 | /// Check if the subtarget has aperture regs. |
| 165 | bool hasApertureRegs(Function &F) { |
| 166 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 167 | return ST.hasApertureRegs(); |
| 168 | } |
| 169 | |
| 170 | /// Check if the subtarget supports GetDoorbellID. |
| 171 | bool supportsGetDoorbellID(Function &F) { |
| 172 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 173 | return ST.supportsGetDoorbellID(); |
| 174 | } |
| 175 | |
| 176 | std::optional<std::pair<unsigned, unsigned>> |
| 177 | getFlatWorkGroupSizeAttr(const Function &F) const { |
| 178 | auto R = AMDGPU::getIntegerPairAttribute(F, Name: "amdgpu-flat-work-group-size" ); |
| 179 | if (!R) |
| 180 | return std::nullopt; |
| 181 | return std::make_pair(x&: R->first, y&: *(R->second)); |
| 182 | } |
| 183 | |
| 184 | std::pair<unsigned, unsigned> |
| 185 | getDefaultFlatWorkGroupSize(const Function &F) const { |
| 186 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 187 | return ST.getDefaultFlatWorkGroupSize(CC: F.getCallingConv()); |
| 188 | } |
| 189 | |
| 190 | std::pair<unsigned, unsigned> |
| 191 | getMaximumFlatWorkGroupRange(const Function &F) { |
| 192 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 193 | return {ST.getMinFlatWorkGroupSize(), ST.getMaxFlatWorkGroupSize()}; |
| 194 | } |
| 195 | |
| 196 | SmallVector<unsigned> getMaxNumWorkGroups(const Function &F) { |
| 197 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 198 | return ST.getMaxNumWorkGroups(F); |
| 199 | } |
| 200 | |
| 201 | /// Get code object version. |
| 202 | unsigned getCodeObjectVersion() const { return CodeObjectVersion; } |
| 203 | |
| 204 | /// Get the effective value of "amdgpu-waves-per-eu" for the function, |
| 205 | /// accounting for the interaction with the passed value to use for |
| 206 | /// "amdgpu-flat-work-group-size". |
| 207 | std::pair<unsigned, unsigned> |
| 208 | getWavesPerEU(const Function &F, |
| 209 | std::pair<unsigned, unsigned> FlatWorkGroupSize) { |
| 210 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 211 | return ST.getWavesPerEU(FlatWorkGroupSizes: FlatWorkGroupSize, LDSBytes: getLDSSize(F), F); |
| 212 | } |
| 213 | |
| 214 | std::optional<std::pair<unsigned, unsigned>> |
| 215 | getWavesPerEUAttr(const Function &F) { |
| 216 | auto Val = AMDGPU::getIntegerPairAttribute(F, Name: "amdgpu-waves-per-eu" , |
| 217 | /*OnlyFirstRequired=*/true); |
| 218 | if (!Val) |
| 219 | return std::nullopt; |
| 220 | if (!Val->second) { |
| 221 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 222 | Val->second = ST.getMaxWavesPerEU(); |
| 223 | } |
| 224 | return std::make_pair(x&: Val->first, y&: *(Val->second)); |
| 225 | } |
| 226 | |
| 227 | unsigned getMaxWavesPerEU(const Function &F) { |
| 228 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F); |
| 229 | return ST.getMaxWavesPerEU(); |
| 230 | } |
| 231 | |
| 232 | unsigned getMaxAddrSpace() const override { |
| 233 | return AMDGPUAS::MAX_AMDGPU_ADDRESS; |
| 234 | } |
| 235 | |
| 236 | private: |
| 237 | /// Check if the ConstantExpr \p CE uses an addrspacecast from private or |
| 238 | /// local to flat. These casts may require the queue pointer. |
| 239 | static uint8_t visitConstExpr(const ConstantExpr *CE) { |
| 240 | uint8_t Status = NONE; |
| 241 | |
| 242 | if (CE->getOpcode() == Instruction::AddrSpaceCast) { |
| 243 | unsigned SrcAS = CE->getOperand(i_nocapture: 0)->getType()->getPointerAddressSpace(); |
| 244 | if (SrcAS == AMDGPUAS::PRIVATE_ADDRESS) |
| 245 | Status |= ADDR_SPACE_CAST_PRIVATE_TO_FLAT; |
| 246 | else if (SrcAS == AMDGPUAS::LOCAL_ADDRESS) |
| 247 | Status |= ADDR_SPACE_CAST_LOCAL_TO_FLAT; |
| 248 | } |
| 249 | |
| 250 | return Status; |
| 251 | } |
| 252 | |
| 253 | /// Returns the minimum amount of LDS space used by a workgroup running |
| 254 | /// function \p F. |
| 255 | static unsigned getLDSSize(const Function &F) { |
| 256 | return AMDGPU::getIntegerPairAttribute(F, Name: "amdgpu-lds-size" , |
| 257 | Default: {0, UINT32_MAX}, OnlyFirstRequired: true) |
| 258 | .first; |
| 259 | } |
| 260 | |
| 261 | /// Get the constant access bitmap for \p C. |
| 262 | uint8_t getConstantAccess(const Constant *C, |
| 263 | SmallPtrSetImpl<const Constant *> &Visited) { |
| 264 | auto It = ConstantStatus.find(Val: C); |
| 265 | if (It != ConstantStatus.end()) |
| 266 | return It->second; |
| 267 | |
| 268 | uint8_t Result = 0; |
| 269 | if (isDSAddress(C)) |
| 270 | Result = DS_GLOBAL; |
| 271 | |
| 272 | if (const auto *CE = dyn_cast<ConstantExpr>(Val: C)) |
| 273 | Result |= visitConstExpr(CE); |
| 274 | |
| 275 | for (const Use &U : C->operands()) { |
| 276 | const auto *OpC = dyn_cast<Constant>(Val: U); |
| 277 | if (!OpC || !Visited.insert(Ptr: OpC).second) |
| 278 | continue; |
| 279 | |
| 280 | Result |= getConstantAccess(C: OpC, Visited); |
| 281 | } |
| 282 | return Result; |
| 283 | } |
| 284 | |
| 285 | public: |
| 286 | /// Returns true if \p Fn needs the queue pointer because of \p C. |
| 287 | bool needsQueuePtr(const Constant *C, Function &Fn) { |
| 288 | bool IsNonEntryFunc = !AMDGPU::isEntryFunctionCC(CC: Fn.getCallingConv()); |
| 289 | bool HasAperture = hasApertureRegs(F&: Fn); |
| 290 | |
| 291 | // No need to explore the constants. |
| 292 | if (!IsNonEntryFunc && HasAperture) |
| 293 | return false; |
| 294 | |
| 295 | SmallPtrSet<const Constant *, 8> Visited; |
| 296 | uint8_t Access = getConstantAccess(C, Visited); |
| 297 | |
| 298 | // We need to trap on DS globals in non-entry functions. |
| 299 | if (IsNonEntryFunc && (Access & DS_GLOBAL)) |
| 300 | return true; |
| 301 | |
| 302 | return !HasAperture && (Access & ADDR_SPACE_CAST_BOTH_TO_FLAT); |
| 303 | } |
| 304 | |
| 305 | bool checkConstForAddrSpaceCastFromPrivate(const Constant *C) { |
| 306 | SmallPtrSet<const Constant *, 8> Visited; |
| 307 | uint8_t Access = getConstantAccess(C, Visited); |
| 308 | return Access & ADDR_SPACE_CAST_PRIVATE_TO_FLAT; |
| 309 | } |
| 310 | |
| 311 | private: |
| 312 | /// Used to determine if the Constant needs the queue pointer. |
| 313 | DenseMap<const Constant *, uint8_t> ConstantStatus; |
| 314 | const unsigned CodeObjectVersion; |
| 315 | }; |
| 316 | |
| 317 | struct AAAMDAttributes |
| 318 | : public StateWrapper<BitIntegerState<uint32_t, ALL_ARGUMENT_MASK, 0>, |
| 319 | AbstractAttribute> { |
| 320 | using Base = StateWrapper<BitIntegerState<uint32_t, ALL_ARGUMENT_MASK, 0>, |
| 321 | AbstractAttribute>; |
| 322 | |
| 323 | AAAMDAttributes(const IRPosition &IRP, Attributor &A) : Base(IRP) {} |
| 324 | |
| 325 | /// Create an abstract attribute view for the position \p IRP. |
| 326 | static AAAMDAttributes &createForPosition(const IRPosition &IRP, |
| 327 | Attributor &A); |
| 328 | |
| 329 | /// See AbstractAttribute::getName(). |
| 330 | StringRef getName() const override { return "AAAMDAttributes" ; } |
| 331 | |
| 332 | /// See AbstractAttribute::getIdAddr(). |
| 333 | const char *getIdAddr() const override { return &ID; } |
| 334 | |
| 335 | /// This function should return true if the type of the \p AA is |
| 336 | /// AAAMDAttributes. |
| 337 | static bool classof(const AbstractAttribute *AA) { |
| 338 | return (AA->getIdAddr() == &ID); |
| 339 | } |
| 340 | |
| 341 | /// Unique ID (due to the unique address) |
| 342 | static const char ID; |
| 343 | }; |
| 344 | const char AAAMDAttributes::ID = 0; |
| 345 | |
| 346 | struct AAUniformWorkGroupSize |
| 347 | : public StateWrapper<BooleanState, AbstractAttribute> { |
| 348 | using Base = StateWrapper<BooleanState, AbstractAttribute>; |
| 349 | AAUniformWorkGroupSize(const IRPosition &IRP, Attributor &A) : Base(IRP) {} |
| 350 | |
| 351 | /// Create an abstract attribute view for the position \p IRP. |
| 352 | static AAUniformWorkGroupSize &createForPosition(const IRPosition &IRP, |
| 353 | Attributor &A); |
| 354 | |
| 355 | /// See AbstractAttribute::getName(). |
| 356 | StringRef getName() const override { return "AAUniformWorkGroupSize" ; } |
| 357 | |
| 358 | /// See AbstractAttribute::getIdAddr(). |
| 359 | const char *getIdAddr() const override { return &ID; } |
| 360 | |
| 361 | /// This function should return true if the type of the \p AA is |
| 362 | /// AAAMDAttributes. |
| 363 | static bool classof(const AbstractAttribute *AA) { |
| 364 | return (AA->getIdAddr() == &ID); |
| 365 | } |
| 366 | |
| 367 | /// Unique ID (due to the unique address) |
| 368 | static const char ID; |
| 369 | }; |
| 370 | const char AAUniformWorkGroupSize::ID = 0; |
| 371 | |
| 372 | struct AAUniformWorkGroupSizeFunction : public AAUniformWorkGroupSize { |
| 373 | AAUniformWorkGroupSizeFunction(const IRPosition &IRP, Attributor &A) |
| 374 | : AAUniformWorkGroupSize(IRP, A) {} |
| 375 | |
| 376 | void initialize(Attributor &A) override { |
| 377 | Function *F = getAssociatedFunction(); |
| 378 | CallingConv::ID CC = F->getCallingConv(); |
| 379 | |
| 380 | if (CC != CallingConv::AMDGPU_KERNEL) |
| 381 | return; |
| 382 | |
| 383 | bool InitialValue = false; |
| 384 | if (F->hasFnAttribute(Kind: "uniform-work-group-size" )) |
| 385 | InitialValue = |
| 386 | F->getFnAttribute(Kind: "uniform-work-group-size" ).getValueAsString() == |
| 387 | "true" ; |
| 388 | |
| 389 | if (InitialValue) |
| 390 | indicateOptimisticFixpoint(); |
| 391 | else |
| 392 | indicatePessimisticFixpoint(); |
| 393 | } |
| 394 | |
| 395 | ChangeStatus updateImpl(Attributor &A) override { |
| 396 | ChangeStatus Change = ChangeStatus::UNCHANGED; |
| 397 | |
| 398 | auto CheckCallSite = [&](AbstractCallSite CS) { |
| 399 | Function *Caller = CS.getInstruction()->getFunction(); |
| 400 | LLVM_DEBUG(dbgs() << "[AAUniformWorkGroupSize] Call " << Caller->getName() |
| 401 | << "->" << getAssociatedFunction()->getName() << "\n" ); |
| 402 | |
| 403 | const auto *CallerInfo = A.getAAFor<AAUniformWorkGroupSize>( |
| 404 | QueryingAA: *this, IRP: IRPosition::function(F: *Caller), DepClass: DepClassTy::REQUIRED); |
| 405 | if (!CallerInfo || !CallerInfo->isValidState()) |
| 406 | return false; |
| 407 | |
| 408 | Change = Change | clampStateAndIndicateChange(S&: this->getState(), |
| 409 | R: CallerInfo->getState()); |
| 410 | |
| 411 | return true; |
| 412 | }; |
| 413 | |
| 414 | bool AllCallSitesKnown = true; |
| 415 | if (!A.checkForAllCallSites(Pred: CheckCallSite, QueryingAA: *this, RequireAllCallSites: true, UsedAssumedInformation&: AllCallSitesKnown)) |
| 416 | return indicatePessimisticFixpoint(); |
| 417 | |
| 418 | return Change; |
| 419 | } |
| 420 | |
| 421 | ChangeStatus manifest(Attributor &A) override { |
| 422 | SmallVector<Attribute, 8> AttrList; |
| 423 | LLVMContext &Ctx = getAssociatedFunction()->getContext(); |
| 424 | |
| 425 | AttrList.push_back(Elt: Attribute::get(Context&: Ctx, Kind: "uniform-work-group-size" , |
| 426 | Val: getAssumed() ? "true" : "false" )); |
| 427 | return A.manifestAttrs(IRP: getIRPosition(), DeducedAttrs: AttrList, |
| 428 | /* ForceReplace */ true); |
| 429 | } |
| 430 | |
| 431 | bool isValidState() const override { |
| 432 | // This state is always valid, even when the state is false. |
| 433 | return true; |
| 434 | } |
| 435 | |
| 436 | const std::string getAsStr(Attributor *) const override { |
| 437 | return "AMDWorkGroupSize[" + std::to_string(val: getAssumed()) + "]" ; |
| 438 | } |
| 439 | |
| 440 | /// See AbstractAttribute::trackStatistics() |
| 441 | void trackStatistics() const override {} |
| 442 | }; |
| 443 | |
| 444 | AAUniformWorkGroupSize & |
| 445 | AAUniformWorkGroupSize::createForPosition(const IRPosition &IRP, |
| 446 | Attributor &A) { |
| 447 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 448 | return *new (A.Allocator) AAUniformWorkGroupSizeFunction(IRP, A); |
| 449 | llvm_unreachable( |
| 450 | "AAUniformWorkGroupSize is only valid for function position" ); |
| 451 | } |
| 452 | |
| 453 | struct AAAMDAttributesFunction : public AAAMDAttributes { |
| 454 | AAAMDAttributesFunction(const IRPosition &IRP, Attributor &A) |
| 455 | : AAAMDAttributes(IRP, A) {} |
| 456 | |
| 457 | void initialize(Attributor &A) override { |
| 458 | Function *F = getAssociatedFunction(); |
| 459 | |
| 460 | // If the function requires the implicit arg pointer due to sanitizers, |
| 461 | // assume it's needed even if explicitly marked as not requiring it. |
| 462 | // Flat scratch initialization is needed because `asan_malloc_impl` |
| 463 | // calls introduced later in pipeline will have flat scratch accesses. |
| 464 | // FIXME: FLAT_SCRATCH_INIT will not be required here if device-libs |
| 465 | // implementation for `asan_malloc_impl` is updated. |
| 466 | const bool HasSanitizerAttrs = hasSanitizerAttributes(F: *F); |
| 467 | if (HasSanitizerAttrs) { |
| 468 | removeAssumedBits(BitsEncoding: IMPLICIT_ARG_PTR); |
| 469 | removeAssumedBits(BitsEncoding: HOSTCALL_PTR); |
| 470 | removeAssumedBits(BitsEncoding: FLAT_SCRATCH_INIT); |
| 471 | } |
| 472 | |
| 473 | for (auto Attr : ImplicitAttrs) { |
| 474 | if (HasSanitizerAttrs && |
| 475 | (Attr.first == IMPLICIT_ARG_PTR || Attr.first == HOSTCALL_PTR || |
| 476 | Attr.first == FLAT_SCRATCH_INIT)) |
| 477 | continue; |
| 478 | |
| 479 | if (F->hasFnAttribute(Kind: Attr.second)) |
| 480 | addKnownBits(Bits: Attr.first); |
| 481 | } |
| 482 | |
| 483 | if (F->isDeclaration()) |
| 484 | return; |
| 485 | |
| 486 | // Ignore functions with graphics calling conventions, these are currently |
| 487 | // not allowed to have kernel arguments. |
| 488 | if (AMDGPU::isGraphics(CC: F->getCallingConv())) { |
| 489 | indicatePessimisticFixpoint(); |
| 490 | return; |
| 491 | } |
| 492 | } |
| 493 | |
| 494 | ChangeStatus updateImpl(Attributor &A) override { |
| 495 | Function *F = getAssociatedFunction(); |
| 496 | // The current assumed state used to determine a change. |
| 497 | auto OrigAssumed = getAssumed(); |
| 498 | |
| 499 | // Check for Intrinsics and propagate attributes. |
| 500 | const AACallEdges *AAEdges = A.getAAFor<AACallEdges>( |
| 501 | QueryingAA: *this, IRP: this->getIRPosition(), DepClass: DepClassTy::REQUIRED); |
| 502 | if (!AAEdges || !AAEdges->isValidState() || |
| 503 | AAEdges->hasNonAsmUnknownCallee()) |
| 504 | return indicatePessimisticFixpoint(); |
| 505 | |
| 506 | bool IsNonEntryFunc = !AMDGPU::isEntryFunctionCC(CC: F->getCallingConv()); |
| 507 | |
| 508 | bool NeedsImplicit = false; |
| 509 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 510 | bool HasApertureRegs = InfoCache.hasApertureRegs(F&: *F); |
| 511 | bool SupportsGetDoorbellID = InfoCache.supportsGetDoorbellID(F&: *F); |
| 512 | unsigned COV = InfoCache.getCodeObjectVersion(); |
| 513 | |
| 514 | for (Function *Callee : AAEdges->getOptimisticEdges()) { |
| 515 | Intrinsic::ID IID = Callee->getIntrinsicID(); |
| 516 | if (IID == Intrinsic::not_intrinsic) { |
| 517 | const AAAMDAttributes *AAAMD = A.getAAFor<AAAMDAttributes>( |
| 518 | QueryingAA: *this, IRP: IRPosition::function(F: *Callee), DepClass: DepClassTy::REQUIRED); |
| 519 | if (!AAAMD || !AAAMD->isValidState()) |
| 520 | return indicatePessimisticFixpoint(); |
| 521 | *this &= *AAAMD; |
| 522 | continue; |
| 523 | } |
| 524 | |
| 525 | bool NonKernelOnly = false; |
| 526 | ImplicitArgumentMask AttrMask = |
| 527 | intrinsicToAttrMask(ID: IID, NonKernelOnly, NeedsImplicit, |
| 528 | HasApertureRegs, SupportsGetDoorBellID: SupportsGetDoorbellID, CodeObjectVersion: COV); |
| 529 | |
| 530 | if (AttrMask == UNKNOWN_INTRINSIC) { |
| 531 | // Assume not-nocallback intrinsics may invoke a function which accesses |
| 532 | // implicit arguments. |
| 533 | // |
| 534 | // FIXME: This isn't really the correct check. We want to ensure it |
| 535 | // isn't calling any function that may use implicit arguments regardless |
| 536 | // of whether it's internal to the module or not. |
| 537 | // |
| 538 | // TODO: Ignoring callsite attributes. |
| 539 | if (!Callee->hasFnAttribute(Kind: Attribute::NoCallback)) |
| 540 | return indicatePessimisticFixpoint(); |
| 541 | continue; |
| 542 | } |
| 543 | |
| 544 | if (AttrMask != NOT_IMPLICIT_INPUT) { |
| 545 | if ((IsNonEntryFunc || !NonKernelOnly)) |
| 546 | removeAssumedBits(BitsEncoding: AttrMask); |
| 547 | } |
| 548 | } |
| 549 | |
| 550 | // Need implicitarg_ptr to acess queue_ptr, private_base, and shared_base. |
| 551 | if (NeedsImplicit) |
| 552 | removeAssumedBits(BitsEncoding: IMPLICIT_ARG_PTR); |
| 553 | |
| 554 | if (isAssumed(BitsEncoding: QUEUE_PTR) && checkForQueuePtr(A)) { |
| 555 | // Under V5, we need implicitarg_ptr + offsets to access private_base or |
| 556 | // shared_base. We do not actually need queue_ptr. |
| 557 | if (COV >= 5) |
| 558 | removeAssumedBits(BitsEncoding: IMPLICIT_ARG_PTR); |
| 559 | else |
| 560 | removeAssumedBits(BitsEncoding: QUEUE_PTR); |
| 561 | } |
| 562 | |
| 563 | if (funcRetrievesMultigridSyncArg(A, COV)) { |
| 564 | assert(!isAssumed(IMPLICIT_ARG_PTR) && |
| 565 | "multigrid_sync_arg needs implicitarg_ptr" ); |
| 566 | removeAssumedBits(BitsEncoding: MULTIGRID_SYNC_ARG); |
| 567 | } |
| 568 | |
| 569 | if (funcRetrievesHostcallPtr(A, COV)) { |
| 570 | assert(!isAssumed(IMPLICIT_ARG_PTR) && "hostcall needs implicitarg_ptr" ); |
| 571 | removeAssumedBits(BitsEncoding: HOSTCALL_PTR); |
| 572 | } |
| 573 | |
| 574 | if (funcRetrievesHeapPtr(A, COV)) { |
| 575 | assert(!isAssumed(IMPLICIT_ARG_PTR) && "heap_ptr needs implicitarg_ptr" ); |
| 576 | removeAssumedBits(BitsEncoding: HEAP_PTR); |
| 577 | } |
| 578 | |
| 579 | if (isAssumed(BitsEncoding: QUEUE_PTR) && funcRetrievesQueuePtr(A, COV)) { |
| 580 | assert(!isAssumed(IMPLICIT_ARG_PTR) && "queue_ptr needs implicitarg_ptr" ); |
| 581 | removeAssumedBits(BitsEncoding: QUEUE_PTR); |
| 582 | } |
| 583 | |
| 584 | if (isAssumed(BitsEncoding: LDS_KERNEL_ID) && funcRetrievesLDSKernelId(A)) { |
| 585 | removeAssumedBits(BitsEncoding: LDS_KERNEL_ID); |
| 586 | } |
| 587 | |
| 588 | if (isAssumed(BitsEncoding: DEFAULT_QUEUE) && funcRetrievesDefaultQueue(A, COV)) |
| 589 | removeAssumedBits(BitsEncoding: DEFAULT_QUEUE); |
| 590 | |
| 591 | if (isAssumed(BitsEncoding: COMPLETION_ACTION) && funcRetrievesCompletionAction(A, COV)) |
| 592 | removeAssumedBits(BitsEncoding: COMPLETION_ACTION); |
| 593 | |
| 594 | if (isAssumed(BitsEncoding: FLAT_SCRATCH_INIT) && needFlatScratchInit(A)) |
| 595 | removeAssumedBits(BitsEncoding: FLAT_SCRATCH_INIT); |
| 596 | |
| 597 | return getAssumed() != OrigAssumed ? ChangeStatus::CHANGED |
| 598 | : ChangeStatus::UNCHANGED; |
| 599 | } |
| 600 | |
| 601 | ChangeStatus manifest(Attributor &A) override { |
| 602 | SmallVector<Attribute, 8> AttrList; |
| 603 | LLVMContext &Ctx = getAssociatedFunction()->getContext(); |
| 604 | |
| 605 | for (auto Attr : ImplicitAttrs) { |
| 606 | if (isKnown(BitsEncoding: Attr.first)) |
| 607 | AttrList.push_back(Elt: Attribute::get(Context&: Ctx, Kind: Attr.second)); |
| 608 | } |
| 609 | |
| 610 | return A.manifestAttrs(IRP: getIRPosition(), DeducedAttrs: AttrList, |
| 611 | /* ForceReplace */ true); |
| 612 | } |
| 613 | |
| 614 | const std::string getAsStr(Attributor *) const override { |
| 615 | std::string Str; |
| 616 | raw_string_ostream OS(Str); |
| 617 | OS << "AMDInfo[" ; |
| 618 | for (auto Attr : ImplicitAttrs) |
| 619 | if (isAssumed(BitsEncoding: Attr.first)) |
| 620 | OS << ' ' << Attr.second; |
| 621 | OS << " ]" ; |
| 622 | return OS.str(); |
| 623 | } |
| 624 | |
| 625 | /// See AbstractAttribute::trackStatistics() |
| 626 | void trackStatistics() const override {} |
| 627 | |
| 628 | private: |
| 629 | bool checkForQueuePtr(Attributor &A) { |
| 630 | Function *F = getAssociatedFunction(); |
| 631 | bool IsNonEntryFunc = !AMDGPU::isEntryFunctionCC(CC: F->getCallingConv()); |
| 632 | |
| 633 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 634 | |
| 635 | bool NeedsQueuePtr = false; |
| 636 | |
| 637 | auto CheckAddrSpaceCasts = [&](Instruction &I) { |
| 638 | unsigned SrcAS = static_cast<AddrSpaceCastInst &>(I).getSrcAddressSpace(); |
| 639 | if (castRequiresQueuePtr(SrcAS)) { |
| 640 | NeedsQueuePtr = true; |
| 641 | return false; |
| 642 | } |
| 643 | return true; |
| 644 | }; |
| 645 | |
| 646 | bool HasApertureRegs = InfoCache.hasApertureRegs(F&: *F); |
| 647 | |
| 648 | // `checkForAllInstructions` is much more cheaper than going through all |
| 649 | // instructions, try it first. |
| 650 | |
| 651 | // The queue pointer is not needed if aperture regs is present. |
| 652 | if (!HasApertureRegs) { |
| 653 | bool UsedAssumedInformation = false; |
| 654 | A.checkForAllInstructions(Pred: CheckAddrSpaceCasts, QueryingAA: *this, |
| 655 | Opcodes: {Instruction::AddrSpaceCast}, |
| 656 | UsedAssumedInformation); |
| 657 | } |
| 658 | |
| 659 | // If we found that we need the queue pointer, nothing else to do. |
| 660 | if (NeedsQueuePtr) |
| 661 | return true; |
| 662 | |
| 663 | if (!IsNonEntryFunc && HasApertureRegs) |
| 664 | return false; |
| 665 | |
| 666 | for (BasicBlock &BB : *F) { |
| 667 | for (Instruction &I : BB) { |
| 668 | for (const Use &U : I.operands()) { |
| 669 | if (const auto *C = dyn_cast<Constant>(Val: U)) { |
| 670 | if (InfoCache.needsQueuePtr(C, Fn&: *F)) |
| 671 | return true; |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | } |
| 676 | |
| 677 | return false; |
| 678 | } |
| 679 | |
| 680 | bool funcRetrievesMultigridSyncArg(Attributor &A, unsigned COV) { |
| 681 | auto Pos = llvm::AMDGPU::getMultigridSyncArgImplicitArgPosition(COV); |
| 682 | AA::RangeTy Range(Pos, 8); |
| 683 | return funcRetrievesImplicitKernelArg(A, Range); |
| 684 | } |
| 685 | |
| 686 | bool funcRetrievesHostcallPtr(Attributor &A, unsigned COV) { |
| 687 | auto Pos = llvm::AMDGPU::getHostcallImplicitArgPosition(COV); |
| 688 | AA::RangeTy Range(Pos, 8); |
| 689 | return funcRetrievesImplicitKernelArg(A, Range); |
| 690 | } |
| 691 | |
| 692 | bool funcRetrievesDefaultQueue(Attributor &A, unsigned COV) { |
| 693 | auto Pos = llvm::AMDGPU::getDefaultQueueImplicitArgPosition(COV); |
| 694 | AA::RangeTy Range(Pos, 8); |
| 695 | return funcRetrievesImplicitKernelArg(A, Range); |
| 696 | } |
| 697 | |
| 698 | bool funcRetrievesCompletionAction(Attributor &A, unsigned COV) { |
| 699 | auto Pos = llvm::AMDGPU::getCompletionActionImplicitArgPosition(COV); |
| 700 | AA::RangeTy Range(Pos, 8); |
| 701 | return funcRetrievesImplicitKernelArg(A, Range); |
| 702 | } |
| 703 | |
| 704 | bool funcRetrievesHeapPtr(Attributor &A, unsigned COV) { |
| 705 | if (COV < 5) |
| 706 | return false; |
| 707 | AA::RangeTy Range(AMDGPU::ImplicitArg::HEAP_PTR_OFFSET, 8); |
| 708 | return funcRetrievesImplicitKernelArg(A, Range); |
| 709 | } |
| 710 | |
| 711 | bool funcRetrievesQueuePtr(Attributor &A, unsigned COV) { |
| 712 | if (COV < 5) |
| 713 | return false; |
| 714 | AA::RangeTy Range(AMDGPU::ImplicitArg::QUEUE_PTR_OFFSET, 8); |
| 715 | return funcRetrievesImplicitKernelArg(A, Range); |
| 716 | } |
| 717 | |
| 718 | bool funcRetrievesImplicitKernelArg(Attributor &A, AA::RangeTy Range) { |
| 719 | // Check if this is a call to the implicitarg_ptr builtin and it |
| 720 | // is used to retrieve the hostcall pointer. The implicit arg for |
| 721 | // hostcall is not used only if every use of the implicitarg_ptr |
| 722 | // is a load that clearly does not retrieve any byte of the |
| 723 | // hostcall pointer. We check this by tracing all the uses of the |
| 724 | // initial call to the implicitarg_ptr intrinsic. |
| 725 | auto DoesNotLeadToKernelArgLoc = [&](Instruction &I) { |
| 726 | auto &Call = cast<CallBase>(Val&: I); |
| 727 | if (Call.getIntrinsicID() != Intrinsic::amdgcn_implicitarg_ptr) |
| 728 | return true; |
| 729 | |
| 730 | const auto *PointerInfoAA = A.getAAFor<AAPointerInfo>( |
| 731 | QueryingAA: *this, IRP: IRPosition::callsite_returned(CB: Call), DepClass: DepClassTy::REQUIRED); |
| 732 | if (!PointerInfoAA || !PointerInfoAA->getState().isValidState()) |
| 733 | return false; |
| 734 | |
| 735 | return PointerInfoAA->forallInterferingAccesses( |
| 736 | Range, CB: [](const AAPointerInfo::Access &Acc, bool IsExact) { |
| 737 | return Acc.getRemoteInst()->isDroppable(); |
| 738 | }); |
| 739 | }; |
| 740 | |
| 741 | bool UsedAssumedInformation = false; |
| 742 | return !A.checkForAllCallLikeInstructions(Pred: DoesNotLeadToKernelArgLoc, QueryingAA: *this, |
| 743 | UsedAssumedInformation); |
| 744 | } |
| 745 | |
| 746 | bool funcRetrievesLDSKernelId(Attributor &A) { |
| 747 | auto DoesNotRetrieve = [&](Instruction &I) { |
| 748 | auto &Call = cast<CallBase>(Val&: I); |
| 749 | return Call.getIntrinsicID() != Intrinsic::amdgcn_lds_kernel_id; |
| 750 | }; |
| 751 | bool UsedAssumedInformation = false; |
| 752 | return !A.checkForAllCallLikeInstructions(Pred: DoesNotRetrieve, QueryingAA: *this, |
| 753 | UsedAssumedInformation); |
| 754 | } |
| 755 | |
| 756 | // Returns true if FlatScratchInit is needed, i.e., no-flat-scratch-init is |
| 757 | // not to be set. |
| 758 | bool needFlatScratchInit(Attributor &A) { |
| 759 | assert(isAssumed(FLAT_SCRATCH_INIT)); // only called if the bit is still set |
| 760 | |
| 761 | // Check all AddrSpaceCast instructions. FlatScratchInit is needed if |
| 762 | // there is a cast from PRIVATE_ADDRESS. |
| 763 | auto AddrSpaceCastNotFromPrivate = [](Instruction &I) { |
| 764 | return cast<AddrSpaceCastInst>(Val&: I).getSrcAddressSpace() != |
| 765 | AMDGPUAS::PRIVATE_ADDRESS; |
| 766 | }; |
| 767 | |
| 768 | bool UsedAssumedInformation = false; |
| 769 | if (!A.checkForAllInstructions(Pred: AddrSpaceCastNotFromPrivate, QueryingAA: *this, |
| 770 | Opcodes: {Instruction::AddrSpaceCast}, |
| 771 | UsedAssumedInformation)) |
| 772 | return true; |
| 773 | |
| 774 | // Check for addrSpaceCast from PRIVATE_ADDRESS in constant expressions |
| 775 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 776 | |
| 777 | Function *F = getAssociatedFunction(); |
| 778 | for (Instruction &I : instructions(F)) { |
| 779 | for (const Use &U : I.operands()) { |
| 780 | if (const auto *C = dyn_cast<Constant>(Val: U)) { |
| 781 | if (InfoCache.checkConstForAddrSpaceCastFromPrivate(C)) |
| 782 | return true; |
| 783 | } |
| 784 | } |
| 785 | } |
| 786 | |
| 787 | // Finally check callees. |
| 788 | |
| 789 | // This is called on each callee; false means callee shouldn't have |
| 790 | // no-flat-scratch-init. |
| 791 | auto CheckForNoFlatScratchInit = [&](Instruction &I) { |
| 792 | const auto &CB = cast<CallBase>(Val&: I); |
| 793 | const Function *Callee = CB.getCalledFunction(); |
| 794 | |
| 795 | // Callee == 0 for inline asm or indirect call with known callees. |
| 796 | // In the latter case, updateImpl() already checked the callees and we |
| 797 | // know their FLAT_SCRATCH_INIT bit is set. |
| 798 | // If function has indirect call with unknown callees, the bit is |
| 799 | // already removed in updateImpl() and execution won't reach here. |
| 800 | if (!Callee) |
| 801 | return true; |
| 802 | |
| 803 | return Callee->getIntrinsicID() != |
| 804 | Intrinsic::amdgcn_addrspacecast_nonnull; |
| 805 | }; |
| 806 | |
| 807 | UsedAssumedInformation = false; |
| 808 | // If any callee is false (i.e. need FlatScratchInit), |
| 809 | // checkForAllCallLikeInstructions returns false, in which case this |
| 810 | // function returns true. |
| 811 | return !A.checkForAllCallLikeInstructions(Pred: CheckForNoFlatScratchInit, QueryingAA: *this, |
| 812 | UsedAssumedInformation); |
| 813 | } |
| 814 | }; |
| 815 | |
| 816 | AAAMDAttributes &AAAMDAttributes::createForPosition(const IRPosition &IRP, |
| 817 | Attributor &A) { |
| 818 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 819 | return *new (A.Allocator) AAAMDAttributesFunction(IRP, A); |
| 820 | llvm_unreachable("AAAMDAttributes is only valid for function position" ); |
| 821 | } |
| 822 | |
| 823 | /// Base class to derive different size ranges. |
| 824 | struct AAAMDSizeRangeAttribute |
| 825 | : public StateWrapper<IntegerRangeState, AbstractAttribute, uint32_t> { |
| 826 | using Base = StateWrapper<IntegerRangeState, AbstractAttribute, uint32_t>; |
| 827 | |
| 828 | StringRef AttrName; |
| 829 | |
| 830 | AAAMDSizeRangeAttribute(const IRPosition &IRP, Attributor &A, |
| 831 | StringRef AttrName) |
| 832 | : Base(IRP, 32), AttrName(AttrName) {} |
| 833 | |
| 834 | /// See AbstractAttribute::trackStatistics() |
| 835 | void trackStatistics() const override {} |
| 836 | |
| 837 | template <class AttributeImpl> ChangeStatus updateImplImpl(Attributor &A) { |
| 838 | ChangeStatus Change = ChangeStatus::UNCHANGED; |
| 839 | |
| 840 | auto CheckCallSite = [&](AbstractCallSite CS) { |
| 841 | Function *Caller = CS.getInstruction()->getFunction(); |
| 842 | LLVM_DEBUG(dbgs() << '[' << getName() << "] Call " << Caller->getName() |
| 843 | << "->" << getAssociatedFunction()->getName() << '\n'); |
| 844 | |
| 845 | const auto *CallerInfo = A.getAAFor<AttributeImpl>( |
| 846 | *this, IRPosition::function(F: *Caller), DepClassTy::REQUIRED); |
| 847 | if (!CallerInfo || !CallerInfo->isValidState()) |
| 848 | return false; |
| 849 | |
| 850 | Change |= |
| 851 | clampStateAndIndicateChange(this->getState(), CallerInfo->getState()); |
| 852 | |
| 853 | return true; |
| 854 | }; |
| 855 | |
| 856 | bool AllCallSitesKnown = true; |
| 857 | if (!A.checkForAllCallSites(CheckCallSite, *this, |
| 858 | /*RequireAllCallSites=*/true, |
| 859 | AllCallSitesKnown)) |
| 860 | return indicatePessimisticFixpoint(); |
| 861 | |
| 862 | return Change; |
| 863 | } |
| 864 | |
| 865 | /// Clamp the assumed range to the default value ([Min, Max]) and emit the |
| 866 | /// attribute if it is not same as default. |
| 867 | ChangeStatus |
| 868 | emitAttributeIfNotDefaultAfterClamp(Attributor &A, |
| 869 | std::pair<unsigned, unsigned> Default) { |
| 870 | auto [Min, Max] = Default; |
| 871 | unsigned Lower = getAssumed().getLower().getZExtValue(); |
| 872 | unsigned Upper = getAssumed().getUpper().getZExtValue(); |
| 873 | |
| 874 | // Clamp the range to the default value. |
| 875 | if (Lower < Min) |
| 876 | Lower = Min; |
| 877 | if (Upper > Max + 1) |
| 878 | Upper = Max + 1; |
| 879 | |
| 880 | // No manifest if the value is invalid or same as default after clamp. |
| 881 | if ((Lower == Min && Upper == Max + 1) || (Upper < Lower)) |
| 882 | return ChangeStatus::UNCHANGED; |
| 883 | |
| 884 | Function *F = getAssociatedFunction(); |
| 885 | LLVMContext &Ctx = F->getContext(); |
| 886 | SmallString<10> Buffer; |
| 887 | raw_svector_ostream OS(Buffer); |
| 888 | OS << Lower << ',' << Upper - 1; |
| 889 | return A.manifestAttrs(IRP: getIRPosition(), |
| 890 | DeducedAttrs: {Attribute::get(Context&: Ctx, Kind: AttrName, Val: OS.str())}, |
| 891 | /*ForceReplace=*/true); |
| 892 | } |
| 893 | |
| 894 | const std::string getAsStr(Attributor *) const override { |
| 895 | std::string Str; |
| 896 | raw_string_ostream OS(Str); |
| 897 | OS << getName() << '['; |
| 898 | OS << getAssumed().getLower() << ',' << getAssumed().getUpper() - 1; |
| 899 | OS << ']'; |
| 900 | return OS.str(); |
| 901 | } |
| 902 | }; |
| 903 | |
| 904 | /// Propagate amdgpu-flat-work-group-size attribute. |
| 905 | struct AAAMDFlatWorkGroupSize : public AAAMDSizeRangeAttribute { |
| 906 | AAAMDFlatWorkGroupSize(const IRPosition &IRP, Attributor &A) |
| 907 | : AAAMDSizeRangeAttribute(IRP, A, "amdgpu-flat-work-group-size" ) {} |
| 908 | |
| 909 | void initialize(Attributor &A) override { |
| 910 | Function *F = getAssociatedFunction(); |
| 911 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 912 | |
| 913 | bool HasAttr = false; |
| 914 | auto Range = InfoCache.getDefaultFlatWorkGroupSize(F: *F); |
| 915 | auto MaxRange = InfoCache.getMaximumFlatWorkGroupRange(F: *F); |
| 916 | |
| 917 | if (auto Attr = InfoCache.getFlatWorkGroupSizeAttr(F: *F)) { |
| 918 | // We only consider an attribute that is not max range because the front |
| 919 | // end always emits the attribute, unfortunately, and sometimes it emits |
| 920 | // the max range. |
| 921 | if (*Attr != MaxRange) { |
| 922 | Range = *Attr; |
| 923 | HasAttr = true; |
| 924 | } |
| 925 | } |
| 926 | |
| 927 | // We don't want to directly clamp the state if it's the max range because |
| 928 | // that is basically the worst state. |
| 929 | if (Range == MaxRange) |
| 930 | return; |
| 931 | |
| 932 | auto [Min, Max] = Range; |
| 933 | ConstantRange CR(APInt(32, Min), APInt(32, Max + 1)); |
| 934 | IntegerRangeState IRS(CR); |
| 935 | clampStateAndIndicateChange(S&: this->getState(), R: IRS); |
| 936 | |
| 937 | if (HasAttr || AMDGPU::isEntryFunctionCC(CC: F->getCallingConv())) |
| 938 | indicateOptimisticFixpoint(); |
| 939 | } |
| 940 | |
| 941 | ChangeStatus updateImpl(Attributor &A) override { |
| 942 | return updateImplImpl<AAAMDFlatWorkGroupSize>(A); |
| 943 | } |
| 944 | |
| 945 | /// Create an abstract attribute view for the position \p IRP. |
| 946 | static AAAMDFlatWorkGroupSize &createForPosition(const IRPosition &IRP, |
| 947 | Attributor &A); |
| 948 | |
| 949 | ChangeStatus manifest(Attributor &A) override { |
| 950 | Function *F = getAssociatedFunction(); |
| 951 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 952 | return emitAttributeIfNotDefaultAfterClamp( |
| 953 | A, Default: InfoCache.getMaximumFlatWorkGroupRange(F: *F)); |
| 954 | } |
| 955 | |
| 956 | /// See AbstractAttribute::getName() |
| 957 | StringRef getName() const override { return "AAAMDFlatWorkGroupSize" ; } |
| 958 | |
| 959 | /// See AbstractAttribute::getIdAddr() |
| 960 | const char *getIdAddr() const override { return &ID; } |
| 961 | |
| 962 | /// This function should return true if the type of the \p AA is |
| 963 | /// AAAMDFlatWorkGroupSize |
| 964 | static bool classof(const AbstractAttribute *AA) { |
| 965 | return (AA->getIdAddr() == &ID); |
| 966 | } |
| 967 | |
| 968 | /// Unique ID (due to the unique address) |
| 969 | static const char ID; |
| 970 | }; |
| 971 | |
| 972 | const char AAAMDFlatWorkGroupSize::ID = 0; |
| 973 | |
| 974 | AAAMDFlatWorkGroupSize & |
| 975 | AAAMDFlatWorkGroupSize::createForPosition(const IRPosition &IRP, |
| 976 | Attributor &A) { |
| 977 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 978 | return *new (A.Allocator) AAAMDFlatWorkGroupSize(IRP, A); |
| 979 | llvm_unreachable( |
| 980 | "AAAMDFlatWorkGroupSize is only valid for function position" ); |
| 981 | } |
| 982 | |
| 983 | struct TupleDecIntegerRangeState : public AbstractState { |
| 984 | DecIntegerState<uint32_t> X, Y, Z; |
| 985 | |
| 986 | bool isValidState() const override { |
| 987 | return X.isValidState() && Y.isValidState() && Z.isValidState(); |
| 988 | } |
| 989 | |
| 990 | bool isAtFixpoint() const override { |
| 991 | return X.isAtFixpoint() && Y.isAtFixpoint() && Z.isAtFixpoint(); |
| 992 | } |
| 993 | |
| 994 | ChangeStatus indicateOptimisticFixpoint() override { |
| 995 | return X.indicateOptimisticFixpoint() | Y.indicateOptimisticFixpoint() | |
| 996 | Z.indicateOptimisticFixpoint(); |
| 997 | } |
| 998 | |
| 999 | ChangeStatus indicatePessimisticFixpoint() override { |
| 1000 | return X.indicatePessimisticFixpoint() | Y.indicatePessimisticFixpoint() | |
| 1001 | Z.indicatePessimisticFixpoint(); |
| 1002 | } |
| 1003 | |
| 1004 | TupleDecIntegerRangeState operator^=(const TupleDecIntegerRangeState &Other) { |
| 1005 | X ^= Other.X; |
| 1006 | Y ^= Other.Y; |
| 1007 | Z ^= Other.Z; |
| 1008 | return *this; |
| 1009 | } |
| 1010 | |
| 1011 | bool operator==(const TupleDecIntegerRangeState &Other) const { |
| 1012 | return X == Other.X && Y == Other.Y && Z == Other.Z; |
| 1013 | } |
| 1014 | |
| 1015 | TupleDecIntegerRangeState &getAssumed() { return *this; } |
| 1016 | const TupleDecIntegerRangeState &getAssumed() const { return *this; } |
| 1017 | }; |
| 1018 | |
| 1019 | using AAAMDMaxNumWorkgroupsState = |
| 1020 | StateWrapper<TupleDecIntegerRangeState, AbstractAttribute, uint32_t>; |
| 1021 | |
| 1022 | /// Propagate amdgpu-max-num-workgroups attribute. |
| 1023 | struct AAAMDMaxNumWorkgroups |
| 1024 | : public StateWrapper<TupleDecIntegerRangeState, AbstractAttribute> { |
| 1025 | using Base = StateWrapper<TupleDecIntegerRangeState, AbstractAttribute>; |
| 1026 | |
| 1027 | AAAMDMaxNumWorkgroups(const IRPosition &IRP, Attributor &A) : Base(IRP) {} |
| 1028 | |
| 1029 | void initialize(Attributor &A) override { |
| 1030 | Function *F = getAssociatedFunction(); |
| 1031 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 1032 | |
| 1033 | SmallVector<unsigned> MaxNumWorkgroups = InfoCache.getMaxNumWorkGroups(F: *F); |
| 1034 | |
| 1035 | X.takeKnownMinimum(Value: MaxNumWorkgroups[0]); |
| 1036 | Y.takeKnownMinimum(Value: MaxNumWorkgroups[1]); |
| 1037 | Z.takeKnownMinimum(Value: MaxNumWorkgroups[2]); |
| 1038 | |
| 1039 | if (AMDGPU::isEntryFunctionCC(CC: F->getCallingConv())) |
| 1040 | indicatePessimisticFixpoint(); |
| 1041 | } |
| 1042 | |
| 1043 | ChangeStatus updateImpl(Attributor &A) override { |
| 1044 | ChangeStatus Change = ChangeStatus::UNCHANGED; |
| 1045 | |
| 1046 | auto CheckCallSite = [&](AbstractCallSite CS) { |
| 1047 | Function *Caller = CS.getInstruction()->getFunction(); |
| 1048 | LLVM_DEBUG(dbgs() << "[AAAMDMaxNumWorkgroups] Call " << Caller->getName() |
| 1049 | << "->" << getAssociatedFunction()->getName() << '\n'); |
| 1050 | |
| 1051 | const auto *CallerInfo = A.getAAFor<AAAMDMaxNumWorkgroups>( |
| 1052 | QueryingAA: *this, IRP: IRPosition::function(F: *Caller), DepClass: DepClassTy::REQUIRED); |
| 1053 | if (!CallerInfo || !CallerInfo->isValidState()) |
| 1054 | return false; |
| 1055 | |
| 1056 | Change |= |
| 1057 | clampStateAndIndicateChange(S&: this->getState(), R: CallerInfo->getState()); |
| 1058 | return true; |
| 1059 | }; |
| 1060 | |
| 1061 | bool AllCallSitesKnown = true; |
| 1062 | if (!A.checkForAllCallSites(Pred: CheckCallSite, QueryingAA: *this, |
| 1063 | /*RequireAllCallSites=*/true, |
| 1064 | UsedAssumedInformation&: AllCallSitesKnown)) |
| 1065 | return indicatePessimisticFixpoint(); |
| 1066 | |
| 1067 | return Change; |
| 1068 | } |
| 1069 | |
| 1070 | /// Create an abstract attribute view for the position \p IRP. |
| 1071 | static AAAMDMaxNumWorkgroups &createForPosition(const IRPosition &IRP, |
| 1072 | Attributor &A); |
| 1073 | |
| 1074 | ChangeStatus manifest(Attributor &A) override { |
| 1075 | Function *F = getAssociatedFunction(); |
| 1076 | LLVMContext &Ctx = F->getContext(); |
| 1077 | SmallString<32> Buffer; |
| 1078 | raw_svector_ostream OS(Buffer); |
| 1079 | OS << X.getAssumed() << ',' << Y.getAssumed() << ',' << Z.getAssumed(); |
| 1080 | |
| 1081 | // TODO: Should annotate loads of the group size for this to do anything |
| 1082 | // useful. |
| 1083 | return A.manifestAttrs( |
| 1084 | IRP: getIRPosition(), |
| 1085 | DeducedAttrs: {Attribute::get(Context&: Ctx, Kind: "amdgpu-max-num-workgroups" , Val: OS.str())}, |
| 1086 | /* ForceReplace= */ true); |
| 1087 | } |
| 1088 | |
| 1089 | StringRef getName() const override { return "AAAMDMaxNumWorkgroups" ; } |
| 1090 | |
| 1091 | const std::string getAsStr(Attributor *) const override { |
| 1092 | std::string Buffer = "AAAMDMaxNumWorkgroupsState[" ; |
| 1093 | raw_string_ostream OS(Buffer); |
| 1094 | OS << X.getAssumed() << ',' << Y.getAssumed() << ',' << Z.getAssumed() |
| 1095 | << ']'; |
| 1096 | return OS.str(); |
| 1097 | } |
| 1098 | |
| 1099 | const char *getIdAddr() const override { return &ID; } |
| 1100 | |
| 1101 | /// This function should return true if the type of the \p AA is |
| 1102 | /// AAAMDMaxNumWorkgroups |
| 1103 | static bool classof(const AbstractAttribute *AA) { |
| 1104 | return (AA->getIdAddr() == &ID); |
| 1105 | } |
| 1106 | |
| 1107 | void trackStatistics() const override {} |
| 1108 | |
| 1109 | /// Unique ID (due to the unique address) |
| 1110 | static const char ID; |
| 1111 | }; |
| 1112 | |
| 1113 | const char AAAMDMaxNumWorkgroups::ID = 0; |
| 1114 | |
| 1115 | AAAMDMaxNumWorkgroups & |
| 1116 | AAAMDMaxNumWorkgroups::createForPosition(const IRPosition &IRP, Attributor &A) { |
| 1117 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 1118 | return *new (A.Allocator) AAAMDMaxNumWorkgroups(IRP, A); |
| 1119 | llvm_unreachable("AAAMDMaxNumWorkgroups is only valid for function position" ); |
| 1120 | } |
| 1121 | |
| 1122 | /// Propagate amdgpu-waves-per-eu attribute. |
| 1123 | struct AAAMDWavesPerEU : public AAAMDSizeRangeAttribute { |
| 1124 | AAAMDWavesPerEU(const IRPosition &IRP, Attributor &A) |
| 1125 | : AAAMDSizeRangeAttribute(IRP, A, "amdgpu-waves-per-eu" ) {} |
| 1126 | |
| 1127 | void initialize(Attributor &A) override { |
| 1128 | Function *F = getAssociatedFunction(); |
| 1129 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 1130 | |
| 1131 | // If the attribute exists, we will honor it if it is not the default. |
| 1132 | if (auto Attr = InfoCache.getWavesPerEUAttr(F: *F)) { |
| 1133 | std::pair<unsigned, unsigned> MaxWavesPerEURange{ |
| 1134 | 1U, InfoCache.getMaxWavesPerEU(F: *F)}; |
| 1135 | if (*Attr != MaxWavesPerEURange) { |
| 1136 | auto [Min, Max] = *Attr; |
| 1137 | ConstantRange Range(APInt(32, Min), APInt(32, Max + 1)); |
| 1138 | IntegerRangeState RangeState(Range); |
| 1139 | this->getState() = RangeState; |
| 1140 | indicateOptimisticFixpoint(); |
| 1141 | return; |
| 1142 | } |
| 1143 | } |
| 1144 | |
| 1145 | if (AMDGPU::isEntryFunctionCC(CC: F->getCallingConv())) |
| 1146 | indicatePessimisticFixpoint(); |
| 1147 | } |
| 1148 | |
| 1149 | ChangeStatus updateImpl(Attributor &A) override { |
| 1150 | ChangeStatus Change = ChangeStatus::UNCHANGED; |
| 1151 | |
| 1152 | auto CheckCallSite = [&](AbstractCallSite CS) { |
| 1153 | Function *Caller = CS.getInstruction()->getFunction(); |
| 1154 | Function *Func = getAssociatedFunction(); |
| 1155 | LLVM_DEBUG(dbgs() << '[' << getName() << "] Call " << Caller->getName() |
| 1156 | << "->" << Func->getName() << '\n'); |
| 1157 | (void)Func; |
| 1158 | |
| 1159 | const auto *CallerAA = A.getAAFor<AAAMDWavesPerEU>( |
| 1160 | QueryingAA: *this, IRP: IRPosition::function(F: *Caller), DepClass: DepClassTy::REQUIRED); |
| 1161 | if (!CallerAA || !CallerAA->isValidState()) |
| 1162 | return false; |
| 1163 | |
| 1164 | ConstantRange Assumed = getAssumed(); |
| 1165 | unsigned Min = std::max(a: Assumed.getLower().getZExtValue(), |
| 1166 | b: CallerAA->getAssumed().getLower().getZExtValue()); |
| 1167 | unsigned Max = std::max(a: Assumed.getUpper().getZExtValue(), |
| 1168 | b: CallerAA->getAssumed().getUpper().getZExtValue()); |
| 1169 | ConstantRange Range(APInt(32, Min), APInt(32, Max)); |
| 1170 | IntegerRangeState RangeState(Range); |
| 1171 | getState() = RangeState; |
| 1172 | Change |= getState() == Assumed ? ChangeStatus::UNCHANGED |
| 1173 | : ChangeStatus::CHANGED; |
| 1174 | |
| 1175 | return true; |
| 1176 | }; |
| 1177 | |
| 1178 | bool AllCallSitesKnown = true; |
| 1179 | if (!A.checkForAllCallSites(Pred: CheckCallSite, QueryingAA: *this, RequireAllCallSites: true, UsedAssumedInformation&: AllCallSitesKnown)) |
| 1180 | return indicatePessimisticFixpoint(); |
| 1181 | |
| 1182 | return Change; |
| 1183 | } |
| 1184 | |
| 1185 | /// Create an abstract attribute view for the position \p IRP. |
| 1186 | static AAAMDWavesPerEU &createForPosition(const IRPosition &IRP, |
| 1187 | Attributor &A); |
| 1188 | |
| 1189 | ChangeStatus manifest(Attributor &A) override { |
| 1190 | Function *F = getAssociatedFunction(); |
| 1191 | auto &InfoCache = static_cast<AMDGPUInformationCache &>(A.getInfoCache()); |
| 1192 | return emitAttributeIfNotDefaultAfterClamp( |
| 1193 | A, Default: {1U, InfoCache.getMaxWavesPerEU(F: *F)}); |
| 1194 | } |
| 1195 | |
| 1196 | /// See AbstractAttribute::getName() |
| 1197 | StringRef getName() const override { return "AAAMDWavesPerEU" ; } |
| 1198 | |
| 1199 | /// See AbstractAttribute::getIdAddr() |
| 1200 | const char *getIdAddr() const override { return &ID; } |
| 1201 | |
| 1202 | /// This function should return true if the type of the \p AA is |
| 1203 | /// AAAMDWavesPerEU |
| 1204 | static bool classof(const AbstractAttribute *AA) { |
| 1205 | return (AA->getIdAddr() == &ID); |
| 1206 | } |
| 1207 | |
| 1208 | /// Unique ID (due to the unique address) |
| 1209 | static const char ID; |
| 1210 | }; |
| 1211 | |
| 1212 | const char AAAMDWavesPerEU::ID = 0; |
| 1213 | |
| 1214 | AAAMDWavesPerEU &AAAMDWavesPerEU::createForPosition(const IRPosition &IRP, |
| 1215 | Attributor &A) { |
| 1216 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 1217 | return *new (A.Allocator) AAAMDWavesPerEU(IRP, A); |
| 1218 | llvm_unreachable("AAAMDWavesPerEU is only valid for function position" ); |
| 1219 | } |
| 1220 | |
| 1221 | /// Compute the minimum number of AGPRs required to allocate the inline asm. |
| 1222 | static unsigned inlineAsmGetNumRequiredAGPRs(const InlineAsm *IA, |
| 1223 | const CallBase &Call) { |
| 1224 | unsigned ArgNo = 0; |
| 1225 | unsigned ResNo = 0; |
| 1226 | unsigned AGPRDefCount = 0; |
| 1227 | unsigned AGPRUseCount = 0; |
| 1228 | unsigned MaxPhysReg = 0; |
| 1229 | const DataLayout &DL = Call.getFunction()->getParent()->getDataLayout(); |
| 1230 | |
| 1231 | // TODO: Overestimates due to not accounting for tied operands |
| 1232 | for (const InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) { |
| 1233 | Type *Ty = nullptr; |
| 1234 | switch (CI.Type) { |
| 1235 | case InlineAsm::isOutput: { |
| 1236 | Ty = Call.getType(); |
| 1237 | if (auto *STy = dyn_cast<StructType>(Val: Ty)) |
| 1238 | Ty = STy->getElementType(N: ResNo); |
| 1239 | ++ResNo; |
| 1240 | break; |
| 1241 | } |
| 1242 | case InlineAsm::isInput: { |
| 1243 | Ty = Call.getArgOperand(i: ArgNo++)->getType(); |
| 1244 | break; |
| 1245 | } |
| 1246 | case InlineAsm::isLabel: |
| 1247 | continue; |
| 1248 | case InlineAsm::isClobber: |
| 1249 | // Parse the physical register reference. |
| 1250 | break; |
| 1251 | } |
| 1252 | |
| 1253 | for (StringRef Code : CI.Codes) { |
| 1254 | unsigned RegCount = 0; |
| 1255 | if (Code.starts_with(Prefix: "a" )) { |
| 1256 | // Virtual register, compute number of registers based on the type. |
| 1257 | // |
| 1258 | // We ought to be going through TargetLowering to get the number of |
| 1259 | // registers, but we should avoid the dependence on CodeGen here. |
| 1260 | RegCount = divideCeil(Numerator: DL.getTypeSizeInBits(Ty), Denominator: 32); |
| 1261 | } else { |
| 1262 | // Physical register reference |
| 1263 | auto [Kind, RegIdx, NumRegs] = AMDGPU::parseAsmConstraintPhysReg(Constraint: Code); |
| 1264 | if (Kind == 'a') { |
| 1265 | RegCount = NumRegs; |
| 1266 | MaxPhysReg = std::max(a: MaxPhysReg, b: std::min(a: RegIdx + NumRegs, b: 256u)); |
| 1267 | } |
| 1268 | |
| 1269 | continue; |
| 1270 | } |
| 1271 | |
| 1272 | if (CI.Type == InlineAsm::isOutput) { |
| 1273 | // Apply tuple alignment requirement |
| 1274 | // |
| 1275 | // TODO: This is more conservative than necessary. |
| 1276 | AGPRDefCount = alignTo(Value: AGPRDefCount, Align: RegCount); |
| 1277 | |
| 1278 | AGPRDefCount += RegCount; |
| 1279 | if (CI.isEarlyClobber) { |
| 1280 | AGPRUseCount = alignTo(Value: AGPRUseCount, Align: RegCount); |
| 1281 | AGPRUseCount += RegCount; |
| 1282 | } |
| 1283 | } else { |
| 1284 | AGPRUseCount = alignTo(Value: AGPRUseCount, Align: RegCount); |
| 1285 | AGPRUseCount += RegCount; |
| 1286 | } |
| 1287 | } |
| 1288 | } |
| 1289 | |
| 1290 | unsigned MaxVirtReg = std::max(a: AGPRUseCount, b: AGPRDefCount); |
| 1291 | |
| 1292 | // TODO: This is overly conservative. If there are any physical registers, |
| 1293 | // allocate any virtual registers after them so we don't have to solve optimal |
| 1294 | // packing. |
| 1295 | return std::min(a: MaxVirtReg + MaxPhysReg, b: 256u); |
| 1296 | } |
| 1297 | |
| 1298 | struct AAAMDGPUMinAGPRAlloc |
| 1299 | : public StateWrapper<DecIntegerState<>, AbstractAttribute> { |
| 1300 | using Base = StateWrapper<DecIntegerState<>, AbstractAttribute>; |
| 1301 | AAAMDGPUMinAGPRAlloc(const IRPosition &IRP, Attributor &A) : Base(IRP) {} |
| 1302 | |
| 1303 | static AAAMDGPUMinAGPRAlloc &createForPosition(const IRPosition &IRP, |
| 1304 | Attributor &A) { |
| 1305 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 1306 | return *new (A.Allocator) AAAMDGPUMinAGPRAlloc(IRP, A); |
| 1307 | llvm_unreachable( |
| 1308 | "AAAMDGPUMinAGPRAlloc is only valid for function position" ); |
| 1309 | } |
| 1310 | |
| 1311 | void initialize(Attributor &A) override { |
| 1312 | Function *F = getAssociatedFunction(); |
| 1313 | auto [MinNumAGPR, MaxNumAGPR] = |
| 1314 | AMDGPU::getIntegerPairAttribute(F: *F, Name: "amdgpu-agpr-alloc" , Default: {~0u, ~0u}, |
| 1315 | /*OnlyFirstRequired=*/true); |
| 1316 | if (MinNumAGPR == 0) |
| 1317 | indicateOptimisticFixpoint(); |
| 1318 | } |
| 1319 | |
| 1320 | const std::string getAsStr(Attributor *A) const override { |
| 1321 | std::string Str = "amdgpu-agpr-alloc=" ; |
| 1322 | raw_string_ostream OS(Str); |
| 1323 | OS << getAssumed(); |
| 1324 | return OS.str(); |
| 1325 | } |
| 1326 | |
| 1327 | void trackStatistics() const override {} |
| 1328 | |
| 1329 | ChangeStatus updateImpl(Attributor &A) override { |
| 1330 | DecIntegerState<> Maximum; |
| 1331 | |
| 1332 | // Check for cases which require allocation of AGPRs. The only cases where |
| 1333 | // AGPRs are required are if there are direct references to AGPRs, so inline |
| 1334 | // assembly and special intrinsics. |
| 1335 | auto CheckForMinAGPRAllocs = [&](Instruction &I) { |
| 1336 | const auto &CB = cast<CallBase>(Val&: I); |
| 1337 | const Value *CalleeOp = CB.getCalledOperand(); |
| 1338 | |
| 1339 | if (const InlineAsm *IA = dyn_cast<InlineAsm>(Val: CalleeOp)) { |
| 1340 | // Technically, the inline asm could be invoking a call to an unknown |
| 1341 | // external function that requires AGPRs, but ignore that. |
| 1342 | unsigned NumRegs = inlineAsmGetNumRequiredAGPRs(IA, Call: CB); |
| 1343 | Maximum.takeAssumedMaximum(Value: NumRegs); |
| 1344 | return true; |
| 1345 | } |
| 1346 | switch (CB.getIntrinsicID()) { |
| 1347 | case Intrinsic::not_intrinsic: |
| 1348 | break; |
| 1349 | case Intrinsic::write_register: |
| 1350 | case Intrinsic::read_register: |
| 1351 | case Intrinsic::read_volatile_register: { |
| 1352 | const MDString *RegName = cast<MDString>( |
| 1353 | Val: cast<MDNode>( |
| 1354 | Val: cast<MetadataAsValue>(Val: CB.getArgOperand(i: 0))->getMetadata()) |
| 1355 | ->getOperand(I: 0)); |
| 1356 | auto [Kind, RegIdx, NumRegs] = |
| 1357 | AMDGPU::parseAsmPhysRegName(TupleString: RegName->getString()); |
| 1358 | if (Kind == 'a') |
| 1359 | Maximum.takeAssumedMaximum(Value: std::min(a: RegIdx + NumRegs, b: 256u)); |
| 1360 | |
| 1361 | return true; |
| 1362 | } |
| 1363 | // Trap-like intrinsics such as llvm.trap and llvm.debugtrap do not have |
| 1364 | // the nocallback attribute, so the AMDGPU attributor can conservatively |
| 1365 | // drop all implicitly-known inputs and AGPR allocation information. Make |
| 1366 | // sure we still infer that no implicit inputs are required and that the |
| 1367 | // AGPR allocation stays at zero. Trap-like intrinsics may invoke a |
| 1368 | // function which requires AGPRs, so we need to check if the called |
| 1369 | // function has the "trap-func-name" attribute. |
| 1370 | case Intrinsic::trap: |
| 1371 | case Intrinsic::debugtrap: |
| 1372 | case Intrinsic::ubsantrap: |
| 1373 | return CB.hasFnAttr(Kind: Attribute::NoCallback) || |
| 1374 | !CB.hasFnAttr(Kind: "trap-func-name" ); |
| 1375 | default: |
| 1376 | // Some intrinsics may use AGPRs, but if we have a choice, we are not |
| 1377 | // required to use AGPRs. |
| 1378 | // Assume !nocallback intrinsics may call a function which requires |
| 1379 | // AGPRs. |
| 1380 | return CB.hasFnAttr(Kind: Attribute::NoCallback); |
| 1381 | } |
| 1382 | |
| 1383 | // TODO: Handle callsite attributes |
| 1384 | auto *CBEdges = A.getAAFor<AACallEdges>( |
| 1385 | QueryingAA: *this, IRP: IRPosition::callsite_function(CB), DepClass: DepClassTy::REQUIRED); |
| 1386 | if (!CBEdges || CBEdges->hasUnknownCallee()) { |
| 1387 | Maximum.indicatePessimisticFixpoint(); |
| 1388 | return false; |
| 1389 | } |
| 1390 | |
| 1391 | for (const Function *PossibleCallee : CBEdges->getOptimisticEdges()) { |
| 1392 | const auto *CalleeInfo = A.getAAFor<AAAMDGPUMinAGPRAlloc>( |
| 1393 | QueryingAA: *this, IRP: IRPosition::function(F: *PossibleCallee), DepClass: DepClassTy::REQUIRED); |
| 1394 | if (!CalleeInfo || !CalleeInfo->isValidState()) { |
| 1395 | Maximum.indicatePessimisticFixpoint(); |
| 1396 | return false; |
| 1397 | } |
| 1398 | |
| 1399 | Maximum.takeAssumedMaximum(Value: CalleeInfo->getAssumed()); |
| 1400 | } |
| 1401 | |
| 1402 | return true; |
| 1403 | }; |
| 1404 | |
| 1405 | bool UsedAssumedInformation = false; |
| 1406 | if (!A.checkForAllCallLikeInstructions(Pred: CheckForMinAGPRAllocs, QueryingAA: *this, |
| 1407 | UsedAssumedInformation)) |
| 1408 | return indicatePessimisticFixpoint(); |
| 1409 | |
| 1410 | return clampStateAndIndicateChange(S&: getState(), R: Maximum); |
| 1411 | } |
| 1412 | |
| 1413 | ChangeStatus manifest(Attributor &A) override { |
| 1414 | LLVMContext &Ctx = getAssociatedFunction()->getContext(); |
| 1415 | SmallString<4> Buffer; |
| 1416 | raw_svector_ostream OS(Buffer); |
| 1417 | OS << getAssumed(); |
| 1418 | |
| 1419 | return A.manifestAttrs( |
| 1420 | IRP: getIRPosition(), DeducedAttrs: {Attribute::get(Context&: Ctx, Kind: "amdgpu-agpr-alloc" , Val: OS.str())}); |
| 1421 | } |
| 1422 | |
| 1423 | StringRef getName() const override { return "AAAMDGPUMinAGPRAlloc" ; } |
| 1424 | const char *getIdAddr() const override { return &ID; } |
| 1425 | |
| 1426 | /// This function should return true if the type of the \p AA is |
| 1427 | /// AAAMDGPUMinAGPRAllocs |
| 1428 | static bool classof(const AbstractAttribute *AA) { |
| 1429 | return (AA->getIdAddr() == &ID); |
| 1430 | } |
| 1431 | |
| 1432 | static const char ID; |
| 1433 | }; |
| 1434 | |
| 1435 | const char AAAMDGPUMinAGPRAlloc::ID = 0; |
| 1436 | |
| 1437 | /// An abstract attribute to propagate the function attribute |
| 1438 | /// "amdgpu-cluster-dims" from kernel entry functions to device functions. |
| 1439 | struct AAAMDGPUClusterDims |
| 1440 | : public StateWrapper<BooleanState, AbstractAttribute> { |
| 1441 | using Base = StateWrapper<BooleanState, AbstractAttribute>; |
| 1442 | AAAMDGPUClusterDims(const IRPosition &IRP, Attributor &A) : Base(IRP) {} |
| 1443 | |
| 1444 | /// Create an abstract attribute view for the position \p IRP. |
| 1445 | static AAAMDGPUClusterDims &createForPosition(const IRPosition &IRP, |
| 1446 | Attributor &A); |
| 1447 | |
| 1448 | /// See AbstractAttribute::getName(). |
| 1449 | StringRef getName() const override { return "AAAMDGPUClusterDims" ; } |
| 1450 | |
| 1451 | /// See AbstractAttribute::getIdAddr(). |
| 1452 | const char *getIdAddr() const override { return &ID; } |
| 1453 | |
| 1454 | /// This function should return true if the type of the \p AA is |
| 1455 | /// AAAMDGPUClusterDims. |
| 1456 | static bool classof(const AbstractAttribute *AA) { |
| 1457 | return AA->getIdAddr() == &ID; |
| 1458 | } |
| 1459 | |
| 1460 | virtual const AMDGPU::ClusterDimsAttr &getClusterDims() const = 0; |
| 1461 | |
| 1462 | /// Unique ID (due to the unique address) |
| 1463 | static const char ID; |
| 1464 | }; |
| 1465 | |
| 1466 | const char AAAMDGPUClusterDims::ID = 0; |
| 1467 | |
| 1468 | struct AAAMDGPUClusterDimsFunction : public AAAMDGPUClusterDims { |
| 1469 | AAAMDGPUClusterDimsFunction(const IRPosition &IRP, Attributor &A) |
| 1470 | : AAAMDGPUClusterDims(IRP, A) {} |
| 1471 | |
| 1472 | void initialize(Attributor &A) override { |
| 1473 | Function *F = getAssociatedFunction(); |
| 1474 | assert(F && "empty associated function" ); |
| 1475 | |
| 1476 | Attr = AMDGPU::ClusterDimsAttr::get(F: *F); |
| 1477 | |
| 1478 | // No matter what a kernel function has, it is final. |
| 1479 | if (AMDGPU::isEntryFunctionCC(CC: F->getCallingConv())) { |
| 1480 | if (Attr.isUnknown()) |
| 1481 | indicatePessimisticFixpoint(); |
| 1482 | else |
| 1483 | indicateOptimisticFixpoint(); |
| 1484 | } |
| 1485 | } |
| 1486 | |
| 1487 | const std::string getAsStr(Attributor *A) const override { |
| 1488 | if (!getAssumed() || Attr.isUnknown()) |
| 1489 | return "unknown" ; |
| 1490 | if (Attr.isNoCluster()) |
| 1491 | return "no" ; |
| 1492 | if (Attr.isVariableDims()) |
| 1493 | return "variable" ; |
| 1494 | return Attr.to_string(); |
| 1495 | } |
| 1496 | |
| 1497 | void trackStatistics() const override {} |
| 1498 | |
| 1499 | ChangeStatus updateImpl(Attributor &A) override { |
| 1500 | auto OldState = Attr; |
| 1501 | |
| 1502 | auto CheckCallSite = [&](AbstractCallSite CS) { |
| 1503 | const auto *CallerAA = A.getAAFor<AAAMDGPUClusterDims>( |
| 1504 | QueryingAA: *this, IRP: IRPosition::function(F: *CS.getInstruction()->getFunction()), |
| 1505 | DepClass: DepClassTy::REQUIRED); |
| 1506 | if (!CallerAA || !CallerAA->isValidState()) |
| 1507 | return false; |
| 1508 | |
| 1509 | return merge(Other: CallerAA->getClusterDims()); |
| 1510 | }; |
| 1511 | |
| 1512 | bool UsedAssumedInformation = false; |
| 1513 | if (!A.checkForAllCallSites(Pred: CheckCallSite, QueryingAA: *this, |
| 1514 | /*RequireAllCallSites=*/true, |
| 1515 | UsedAssumedInformation)) |
| 1516 | return indicatePessimisticFixpoint(); |
| 1517 | |
| 1518 | return OldState == Attr ? ChangeStatus::UNCHANGED : ChangeStatus::CHANGED; |
| 1519 | } |
| 1520 | |
| 1521 | ChangeStatus manifest(Attributor &A) override { |
| 1522 | if (Attr.isUnknown()) |
| 1523 | return ChangeStatus::UNCHANGED; |
| 1524 | return A.manifestAttrs( |
| 1525 | IRP: getIRPosition(), |
| 1526 | DeducedAttrs: {Attribute::get(Context&: getAssociatedFunction()->getContext(), Kind: AttrName, |
| 1527 | Val: Attr.to_string())}, |
| 1528 | /*ForceReplace=*/true); |
| 1529 | } |
| 1530 | |
| 1531 | const AMDGPU::ClusterDimsAttr &getClusterDims() const override { |
| 1532 | return Attr; |
| 1533 | } |
| 1534 | |
| 1535 | private: |
| 1536 | bool merge(const AMDGPU::ClusterDimsAttr &Other) { |
| 1537 | // Case 1: Both of them are unknown yet, we do nothing and continue wait for |
| 1538 | // propagation. |
| 1539 | if (Attr.isUnknown() && Other.isUnknown()) |
| 1540 | return true; |
| 1541 | |
| 1542 | // Case 2: The other is determined, but we are unknown yet, we simply take |
| 1543 | // the other's value. |
| 1544 | if (Attr.isUnknown()) { |
| 1545 | Attr = Other; |
| 1546 | return true; |
| 1547 | } |
| 1548 | |
| 1549 | // Case 3: We are determined but the other is unknown yet, we simply keep |
| 1550 | // everything unchanged. |
| 1551 | if (Other.isUnknown()) |
| 1552 | return true; |
| 1553 | |
| 1554 | // After this point, both are determined. |
| 1555 | |
| 1556 | // Case 4: If they are same, we do nothing. |
| 1557 | if (Attr == Other) |
| 1558 | return true; |
| 1559 | |
| 1560 | // Now they are not same. |
| 1561 | |
| 1562 | // Case 5: If either of us uses cluster (but not both; otherwise case 4 |
| 1563 | // would hold), then it is unknown whether cluster will be used, and the |
| 1564 | // state is final, unlike case 1. |
| 1565 | if (Attr.isNoCluster() || Other.isNoCluster()) { |
| 1566 | Attr.setUnknown(); |
| 1567 | return false; |
| 1568 | } |
| 1569 | |
| 1570 | // Case 6: Both of us use cluster, but the dims are different, so the result |
| 1571 | // is, cluster is used, but we just don't have a fixed dims. |
| 1572 | Attr.setVariableDims(); |
| 1573 | return true; |
| 1574 | } |
| 1575 | |
| 1576 | AMDGPU::ClusterDimsAttr Attr; |
| 1577 | |
| 1578 | static constexpr char AttrName[] = "amdgpu-cluster-dims" ; |
| 1579 | }; |
| 1580 | |
| 1581 | AAAMDGPUClusterDims & |
| 1582 | AAAMDGPUClusterDims::createForPosition(const IRPosition &IRP, Attributor &A) { |
| 1583 | if (IRP.getPositionKind() == IRPosition::IRP_FUNCTION) |
| 1584 | return *new (A.Allocator) AAAMDGPUClusterDimsFunction(IRP, A); |
| 1585 | llvm_unreachable("AAAMDGPUClusterDims is only valid for function position" ); |
| 1586 | } |
| 1587 | |
| 1588 | static bool runImpl(Module &M, AnalysisGetter &AG, TargetMachine &TM, |
| 1589 | AMDGPUAttributorOptions Options, |
| 1590 | ThinOrFullLTOPhase LTOPhase) { |
| 1591 | SetVector<Function *> Functions; |
| 1592 | for (Function &F : M) { |
| 1593 | if (!F.isIntrinsic()) |
| 1594 | Functions.insert(X: &F); |
| 1595 | } |
| 1596 | |
| 1597 | CallGraphUpdater CGUpdater; |
| 1598 | BumpPtrAllocator Allocator; |
| 1599 | AMDGPUInformationCache InfoCache(M, AG, Allocator, nullptr, TM); |
| 1600 | DenseSet<const char *> Allowed( |
| 1601 | {&AAAMDAttributes::ID, &AAUniformWorkGroupSize::ID, |
| 1602 | &AAPotentialValues::ID, &AAAMDFlatWorkGroupSize::ID, |
| 1603 | &AAAMDMaxNumWorkgroups::ID, &AAAMDWavesPerEU::ID, |
| 1604 | &AAAMDGPUMinAGPRAlloc::ID, &AACallEdges::ID, &AAPointerInfo::ID, |
| 1605 | &AAPotentialConstantValues::ID, &AAUnderlyingObjects::ID, |
| 1606 | &AANoAliasAddrSpace::ID, &AAAddressSpace::ID, &AAIndirectCallInfo::ID, |
| 1607 | &AAAMDGPUClusterDims::ID, &AAAlign::ID}); |
| 1608 | |
| 1609 | AttributorConfig AC(CGUpdater); |
| 1610 | AC.IsClosedWorldModule = Options.IsClosedWorld; |
| 1611 | AC.Allowed = &Allowed; |
| 1612 | AC.IsModulePass = true; |
| 1613 | AC.DefaultInitializeLiveInternals = false; |
| 1614 | AC.IndirectCalleeSpecializationCallback = |
| 1615 | [](Attributor &A, const AbstractAttribute &AA, CallBase &CB, |
| 1616 | Function &Callee, unsigned NumAssumedCallees) { |
| 1617 | return !AMDGPU::isEntryFunctionCC(CC: Callee.getCallingConv()) && |
| 1618 | (NumAssumedCallees <= IndirectCallSpecializationThreshold); |
| 1619 | }; |
| 1620 | AC.IPOAmendableCB = [](const Function &F) { |
| 1621 | return F.getCallingConv() == CallingConv::AMDGPU_KERNEL; |
| 1622 | }; |
| 1623 | |
| 1624 | Attributor A(Functions, InfoCache, AC); |
| 1625 | |
| 1626 | LLVM_DEBUG({ |
| 1627 | StringRef LTOPhaseStr = to_string(LTOPhase); |
| 1628 | dbgs() << "[AMDGPUAttributor] Running at phase " << LTOPhaseStr << '\n' |
| 1629 | << "[AMDGPUAttributor] Module " << M.getName() << " is " |
| 1630 | << (AC.IsClosedWorldModule ? "" : "not " ) |
| 1631 | << "assumed to be a closed world.\n" ; |
| 1632 | }); |
| 1633 | |
| 1634 | for (auto *F : Functions) { |
| 1635 | A.getOrCreateAAFor<AAAMDAttributes>(IRP: IRPosition::function(F: *F)); |
| 1636 | A.getOrCreateAAFor<AAUniformWorkGroupSize>(IRP: IRPosition::function(F: *F)); |
| 1637 | A.getOrCreateAAFor<AAAMDMaxNumWorkgroups>(IRP: IRPosition::function(F: *F)); |
| 1638 | CallingConv::ID CC = F->getCallingConv(); |
| 1639 | if (!AMDGPU::isEntryFunctionCC(CC)) { |
| 1640 | A.getOrCreateAAFor<AAAMDFlatWorkGroupSize>(IRP: IRPosition::function(F: *F)); |
| 1641 | A.getOrCreateAAFor<AAAMDWavesPerEU>(IRP: IRPosition::function(F: *F)); |
| 1642 | } |
| 1643 | |
| 1644 | const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F: *F); |
| 1645 | if (!F->isDeclaration() && ST.hasClusters()) |
| 1646 | A.getOrCreateAAFor<AAAMDGPUClusterDims>(IRP: IRPosition::function(F: *F)); |
| 1647 | |
| 1648 | if (ST.hasGFX90AInsts()) |
| 1649 | A.getOrCreateAAFor<AAAMDGPUMinAGPRAlloc>(IRP: IRPosition::function(F: *F)); |
| 1650 | |
| 1651 | for (auto &I : instructions(F)) { |
| 1652 | Value *Ptr = nullptr; |
| 1653 | if (auto *LI = dyn_cast<LoadInst>(Val: &I)) |
| 1654 | Ptr = LI->getPointerOperand(); |
| 1655 | else if (auto *SI = dyn_cast<StoreInst>(Val: &I)) |
| 1656 | Ptr = SI->getPointerOperand(); |
| 1657 | else if (auto *RMW = dyn_cast<AtomicRMWInst>(Val: &I)) |
| 1658 | Ptr = RMW->getPointerOperand(); |
| 1659 | else if (auto *CmpX = dyn_cast<AtomicCmpXchgInst>(Val: &I)) |
| 1660 | Ptr = CmpX->getPointerOperand(); |
| 1661 | |
| 1662 | if (Ptr) { |
| 1663 | A.getOrCreateAAFor<AAAddressSpace>(IRP: IRPosition::value(V: *Ptr)); |
| 1664 | A.getOrCreateAAFor<AANoAliasAddrSpace>(IRP: IRPosition::value(V: *Ptr)); |
| 1665 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: Ptr)) { |
| 1666 | if (II->getIntrinsicID() == Intrinsic::amdgcn_make_buffer_rsrc) |
| 1667 | A.getOrCreateAAFor<AAAlign>(IRP: IRPosition::value(V: *Ptr)); |
| 1668 | } |
| 1669 | } |
| 1670 | } |
| 1671 | } |
| 1672 | |
| 1673 | return A.run() == ChangeStatus::CHANGED; |
| 1674 | } |
| 1675 | } // namespace |
| 1676 | |
| 1677 | PreservedAnalyses llvm::AMDGPUAttributorPass::run(Module &M, |
| 1678 | ModuleAnalysisManager &AM) { |
| 1679 | |
| 1680 | FunctionAnalysisManager &FAM = |
| 1681 | AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 1682 | AnalysisGetter AG(FAM); |
| 1683 | |
| 1684 | // TODO: Probably preserves CFG |
| 1685 | return runImpl(M, AG, TM, Options, LTOPhase) ? PreservedAnalyses::none() |
| 1686 | : PreservedAnalyses::all(); |
| 1687 | } |
| 1688 | |