| 1 | //===-- AMDGPULowerBufferFatPointers.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 | // This pass lowers operations on buffer fat pointers (addrspace 7) to |
| 10 | // operations on buffer resources (addrspace 8) and is needed for correct |
| 11 | // codegen. |
| 12 | // |
| 13 | // # Background |
| 14 | // |
| 15 | // Address space 7 (the buffer fat pointer) is a 160-bit pointer that consists |
| 16 | // of a 128-bit buffer descriptor and a 32-bit offset into that descriptor. |
| 17 | // The buffer resource part needs to be it needs to be a "raw" buffer resource |
| 18 | // (it must have a stride of 0 and bounds checks must be in raw buffer mode |
| 19 | // or disabled). |
| 20 | // |
| 21 | // When these requirements are met, a buffer resource can be treated as a |
| 22 | // typical (though quite wide) pointer that follows typical LLVM pointer |
| 23 | // semantics. This allows the frontend to reason about such buffers (which are |
| 24 | // often encountered in the context of SPIR-V kernels). |
| 25 | // |
| 26 | // However, because of their non-power-of-2 size, these fat pointers cannot be |
| 27 | // present during translation to MIR (though this restriction may be lifted |
| 28 | // during the transition to GlobalISel). Therefore, this pass is needed in order |
| 29 | // to correctly implement these fat pointers. |
| 30 | // |
| 31 | // The resource intrinsics take the resource part (the address space 8 pointer) |
| 32 | // and the offset part (the 32-bit integer) as separate arguments. In addition, |
| 33 | // many users of these buffers manipulate the offset while leaving the resource |
| 34 | // part alone. For these reasons, we want to typically separate the resource |
| 35 | // and offset parts into separate variables, but combine them together when |
| 36 | // encountering cases where this is required, such as by inserting these values |
| 37 | // into aggretates or moving them to memory. |
| 38 | // |
| 39 | // Therefore, at a high level, `ptr addrspace(7) %x` becomes `ptr addrspace(8) |
| 40 | // %x.rsrc` and `i32 %x.off`, which will be combined into `{ptr addrspace(8), |
| 41 | // i32} %x = {%x.rsrc, %x.off}` if needed. Similarly, `vector<Nxp7>` becomes |
| 42 | // `{vector<Nxp8>, vector<Nxi32 >}` and its component parts. |
| 43 | // |
| 44 | // # Implementation |
| 45 | // |
| 46 | // This pass proceeds in three main phases: |
| 47 | // |
| 48 | // ## Rewriting loads and stores of p7 and memcpy()-like handling |
| 49 | // |
| 50 | // The first phase is to rewrite away all loads and stors of `ptr addrspace(7)`, |
| 51 | // including aggregates containing such pointers, to ones that use `i160`. This |
| 52 | // is handled by `StoreFatPtrsAsIntsAndExpandMemcpyVisitor` , which visits |
| 53 | // loads, stores, and allocas and, if the loaded or stored type contains `ptr |
| 54 | // addrspace(7)`, rewrites it to use i160, `ptrtoint`ing before stores and |
| 55 | // `inttoptr`ing after loads. Vectors of pointers work the same way. Since i160 |
| 56 | // and p7 differ in size and alignment, aggregates are split into one access per |
| 57 | // leaf, and allocas and GEPs use byte offsets and sizes from the original type. |
| 58 | // |
| 59 | // Such a transformation allows the later phases of the pass to not need |
| 60 | // to handle buffer fat pointers moving to and from memory, where we load |
| 61 | // have to handle the incompatibility between a `{Nxp8, Nxi32}` representation |
| 62 | // and `Nxi60` directly. Instead, that transposing action (where the vectors |
| 63 | // of resources and vectors of offsets are concatentated before being stored to |
| 64 | // memory) are handled through implementing `inttoptr` and `ptrtoint` only. |
| 65 | // |
| 66 | // Atomics operations on `ptr addrspace(7)` values are not suppported, as the |
| 67 | // hardware does not include a 160-bit atomic. |
| 68 | // |
| 69 | // In order to save on O(N) work and to ensure that the contents type |
| 70 | // legalizer correctly splits up wide loads, also unconditionally lower |
| 71 | // memcpy-like intrinsics into loops here. |
| 72 | // |
| 73 | // ## Buffer contents type legalization |
| 74 | // |
| 75 | // The underlying buffer intrinsics only support types up to 128 bits long, |
| 76 | // and don't support complex types. If buffer operations were |
| 77 | // standard pointer operations that could be represented as MIR-level loads, |
| 78 | // this would be handled by the various legalization schemes in instruction |
| 79 | // selection. However, because we have to do the conversion from `load` and |
| 80 | // `store` to intrinsics at LLVM IR level, we must perform that legalization |
| 81 | // ourselves. |
| 82 | // |
| 83 | // This involves a combination of |
| 84 | // - Converting arrays to vectors where possible |
| 85 | // - Otherwise, splitting loads and stores of aggregates into loads/stores of |
| 86 | // each component. |
| 87 | // - Zero-extending things to fill a whole number of bytes |
| 88 | // - Casting values of types that don't neatly correspond to supported machine |
| 89 | // value |
| 90 | // (for example, an i96 or i256) into ones that would work ( |
| 91 | // like <3 x i32> and <8 x i32>, respectively) |
| 92 | // - Splitting values that are too long (such as aforementioned <8 x i32>) into |
| 93 | // multiple operations. |
| 94 | // |
| 95 | // ## Type remapping |
| 96 | // |
| 97 | // We use a `ValueMapper` to mangle uses of [vectors of] buffer fat pointers |
| 98 | // to the corresponding struct type, which has a resource part and an offset |
| 99 | // part. |
| 100 | // |
| 101 | // This uses a `BufferFatPtrToStructTypeMap` and a `FatPtrConstMaterializer` |
| 102 | // to, usually by way of `setType`ing values. Constants are handled here |
| 103 | // because there isn't a good way to fix them up later. |
| 104 | // |
| 105 | // This has the downside of leaving the IR in an invalid state (for example, |
| 106 | // the instruction `getelementptr {ptr addrspace(8), i32} %p, ...` will exist), |
| 107 | // but all such invalid states will be resolved by the third phase. |
| 108 | // |
| 109 | // Functions that don't take buffer fat pointers are modified in place. Those |
| 110 | // that do take such pointers have their basic blocks moved to a new function |
| 111 | // with arguments that are {ptr addrspace(8), i32} arguments and return values. |
| 112 | // This phase also records intrinsics so that they can be remangled or deleted |
| 113 | // later. |
| 114 | // |
| 115 | // ## Splitting pointer structs |
| 116 | // |
| 117 | // The meat of this pass consists of defining semantics for operations that |
| 118 | // produce or consume [vectors of] buffer fat pointers in terms of their |
| 119 | // resource and offset parts. This is accomplished throgh the `SplitPtrStructs` |
| 120 | // visitor. |
| 121 | // |
| 122 | // In the first pass through each function that is being lowered, the splitter |
| 123 | // inserts new instructions to implement the split-structures behavior, which is |
| 124 | // needed for correctness and performance. It records a list of "split users", |
| 125 | // instructions that are being replaced by operations on the resource and offset |
| 126 | // parts. |
| 127 | // |
| 128 | // Split users do not necessarily need to produce parts themselves ( |
| 129 | // a `load float, ptr addrspace(7)` does not, for example), but, if they do not |
| 130 | // generate fat buffer pointers, they must RAUW in their replacement |
| 131 | // instructions during the initial visit. |
| 132 | // |
| 133 | // When these new instructions are created, they use the split parts recorded |
| 134 | // for their initial arguments in order to generate their replacements, creating |
| 135 | // a parallel set of instructions that does not refer to the original fat |
| 136 | // pointer values but instead to their resource and offset components. |
| 137 | // |
| 138 | // Instructions, such as `extractvalue`, that produce buffer fat pointers from |
| 139 | // sources that do not have split parts, have such parts generated using |
| 140 | // `extractvalue`. This is also the initial handling of PHI nodes, which |
| 141 | // are then cleaned up. |
| 142 | // |
| 143 | // ### Conditionals |
| 144 | // |
| 145 | // PHI nodes are initially given resource parts via `extractvalue`. However, |
| 146 | // this is not an efficient rewrite of such nodes, as, in most cases, the |
| 147 | // resource part in a conditional or loop remains constant throughout the loop |
| 148 | // and only the offset varies. Failing to optimize away these constant resources |
| 149 | // would cause additional registers to be sent around loops and might lead to |
| 150 | // waterfall loops being generated for buffer operations due to the |
| 151 | // "non-uniform" resource argument. |
| 152 | // |
| 153 | // Therefore, after all instructions have been visited, the pointer splitter |
| 154 | // post-processes all encountered conditionals. Given a PHI node or select, |
| 155 | // getPossibleRsrcRoots() collects all values that the resource parts of that |
| 156 | // conditional's input could come from as well as collecting all conditional |
| 157 | // instructions encountered during the search. If, after filtering out the |
| 158 | // initial node itself, the set of encountered conditionals is a subset of the |
| 159 | // potential roots and there is a single potential resource that isn't in the |
| 160 | // conditional set, that value is the only possible value the resource argument |
| 161 | // could have throughout the control flow. |
| 162 | // |
| 163 | // If that condition is met, then a PHI node can have its resource part changed |
| 164 | // to the singleton value and then be replaced by a PHI on the offsets. |
| 165 | // Otherwise, each PHI node is split into two, one for the resource part and one |
| 166 | // for the offset part, which replace the temporary `extractvalue` instructions |
| 167 | // that were added during the first pass. |
| 168 | // |
| 169 | // Similar logic applies to `select`, where |
| 170 | // `%z = select i1 %cond, %cond, ptr addrspace(7) %x, ptr addrspace(7) %y` |
| 171 | // can be split into `%z.rsrc = %x.rsrc` and |
| 172 | // `%z.off = select i1 %cond, ptr i32 %x.off, i32 %y.off` |
| 173 | // if both `%x` and `%y` have the same resource part, but two `select` |
| 174 | // operations will be needed if they do not. |
| 175 | // |
| 176 | // ### Final processing |
| 177 | // |
| 178 | // After conditionals have been cleaned up, the IR for each function is |
| 179 | // rewritten to remove all the old instructions that have been split up. |
| 180 | // |
| 181 | // Any instruction that used to produce a buffer fat pointer (and therefore now |
| 182 | // produces a resource-and-offset struct after type remapping) is |
| 183 | // replaced as follows: |
| 184 | // 1. All debug value annotations are cloned to reflect that the resource part |
| 185 | // and offset parts are computed separately and constitute different |
| 186 | // fragments of the underlying source language variable. |
| 187 | // 2. All uses that were themselves split are replaced by a `poison` of the |
| 188 | // struct type, as they will themselves be erased soon. This rule, combined |
| 189 | // with debug handling, should leave the use lists of split instructions |
| 190 | // empty in almost all cases. |
| 191 | // 3. If a user of the original struct-valued result remains, the structure |
| 192 | // needed for the new types to work is constructed out of the newly-defined |
| 193 | // parts, and the original instruction is replaced by this structure |
| 194 | // before being erased. Instructions requiring this construction include |
| 195 | // `ret` and `insertvalue`. |
| 196 | // |
| 197 | // # Consequences |
| 198 | // |
| 199 | // This pass does not alter the CFG. |
| 200 | // |
| 201 | // Alias analysis information will become coarser, as the LLVM alias analyzer |
| 202 | // cannot handle the buffer intrinsics. Specifically, while we can determine |
| 203 | // that the following two loads do not alias: |
| 204 | // ``` |
| 205 | // %y = getelementptr i32, ptr addrspace(7) %x, i32 1 |
| 206 | // %a = load i32, ptr addrspace(7) %x |
| 207 | // %b = load i32, ptr addrspace(7) %y |
| 208 | // ``` |
| 209 | // we cannot (except through some code that runs during scheduling) determine |
| 210 | // that the rewritten loads below do not alias. |
| 211 | // ``` |
| 212 | // %y.off = add i32 %x.off, 1 |
| 213 | // %a = call @llvm.amdgcn.raw.ptr.buffer.load(ptr addrspace(8) %x.rsrc, i32 |
| 214 | // %x.off, ...) |
| 215 | // %b = call @llvm.amdgcn.raw.ptr.buffer.load(ptr addrspace(8) |
| 216 | // %x.rsrc, i32 %y.off, ...) |
| 217 | // ``` |
| 218 | // However, existing alias information is preserved. |
| 219 | //===----------------------------------------------------------------------===// |
| 220 | |
| 221 | #include "AMDGPU.h" |
| 222 | #include "AMDGPUTargetMachine.h" |
| 223 | #include "GCNSubtarget.h" |
| 224 | #include "SIDefines.h" |
| 225 | #include "llvm/ADT/SetOperations.h" |
| 226 | #include "llvm/ADT/SmallVector.h" |
| 227 | #include "llvm/Analysis/InstSimplifyFolder.h" |
| 228 | #include "llvm/Analysis/ScalarEvolution.h" |
| 229 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 230 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 231 | #include "llvm/Analysis/Utils/Local.h" |
| 232 | #include "llvm/CodeGen/TargetPassConfig.h" |
| 233 | #include "llvm/IR/AttributeMask.h" |
| 234 | #include "llvm/IR/Constants.h" |
| 235 | #include "llvm/IR/DebugInfo.h" |
| 236 | #include "llvm/IR/DerivedTypes.h" |
| 237 | #include "llvm/IR/IRBuilder.h" |
| 238 | #include "llvm/IR/InstIterator.h" |
| 239 | #include "llvm/IR/InstVisitor.h" |
| 240 | #include "llvm/IR/Instructions.h" |
| 241 | #include "llvm/IR/IntrinsicInst.h" |
| 242 | #include "llvm/IR/Intrinsics.h" |
| 243 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 244 | #include "llvm/IR/Metadata.h" |
| 245 | #include "llvm/IR/PassManager.h" |
| 246 | #include "llvm/IR/PatternMatch.h" |
| 247 | #include "llvm/IR/ReplaceConstant.h" |
| 248 | #include "llvm/InitializePasses.h" |
| 249 | #include "llvm/Pass.h" |
| 250 | #include "llvm/Support/AMDGPUAddrSpace.h" |
| 251 | #include "llvm/Support/Alignment.h" |
| 252 | #include "llvm/Support/AtomicOrdering.h" |
| 253 | #include "llvm/Support/Debug.h" |
| 254 | #include "llvm/Support/ErrorHandling.h" |
| 255 | #include "llvm/Support/MathExtras.h" |
| 256 | #include "llvm/Transforms/Utils/Cloning.h" |
| 257 | #include "llvm/Transforms/Utils/Local.h" |
| 258 | #include "llvm/Transforms/Utils/LowerMemIntrinsics.h" |
| 259 | #include "llvm/Transforms/Utils/ValueMapper.h" |
| 260 | |
| 261 | #define DEBUG_TYPE "amdgpu-lower-buffer-fat-pointers" |
| 262 | |
| 263 | using namespace llvm; |
| 264 | |
| 265 | using GetTTIFn = function_ref<const TargetTransformInfo *(Function &)>; |
| 266 | using GetSEFn = function_ref<ScalarEvolution *(Function &)>; |
| 267 | |
| 268 | static constexpr unsigned BufferOffsetWidth = 32; |
| 269 | |
| 270 | namespace { |
| 271 | /// Recursively replace instances of ptr addrspace(7) and vector<Nxptr |
| 272 | /// addrspace(7)> with some other type as defined by the relevant subclass. |
| 273 | class BufferFatPtrTypeLoweringBase : public ValueMapTypeRemapper { |
| 274 | DenseMap<Type *, Type *> Map; |
| 275 | |
| 276 | Type *remapTypeImpl(Type *Ty); |
| 277 | |
| 278 | protected: |
| 279 | virtual Type *remapScalar(PointerType *PT) = 0; |
| 280 | virtual Type *remapVector(VectorType *VT) = 0; |
| 281 | |
| 282 | const DataLayout &DL; |
| 283 | |
| 284 | public: |
| 285 | BufferFatPtrTypeLoweringBase(const DataLayout &DL) : DL(DL) {} |
| 286 | Type *remapType(Type *SrcTy) override; |
| 287 | void clear() { Map.clear(); } |
| 288 | }; |
| 289 | |
| 290 | /// Remap ptr addrspace(7) to i160 and vector<Nxptr addrspace(7)> to |
| 291 | /// vector<Nxi60> in order to correctly handling loading/storing these values |
| 292 | /// from memory. |
| 293 | class BufferFatPtrToIntTypeMap : public BufferFatPtrTypeLoweringBase { |
| 294 | using BufferFatPtrTypeLoweringBase::BufferFatPtrTypeLoweringBase; |
| 295 | |
| 296 | protected: |
| 297 | Type *remapScalar(PointerType *PT) override { return DL.getIntPtrType(PT); } |
| 298 | Type *remapVector(VectorType *VT) override { return DL.getIntPtrType(VT); } |
| 299 | }; |
| 300 | |
| 301 | /// Remap ptr addrspace(7) to {ptr addrspace(8), i32} (the resource and offset |
| 302 | /// parts of the pointer) so that we can easily rewrite operations on these |
| 303 | /// values that aren't loading them from or storing them to memory. |
| 304 | class BufferFatPtrToStructTypeMap : public BufferFatPtrTypeLoweringBase { |
| 305 | using BufferFatPtrTypeLoweringBase::BufferFatPtrTypeLoweringBase; |
| 306 | |
| 307 | protected: |
| 308 | Type *remapScalar(PointerType *PT) override; |
| 309 | Type *remapVector(VectorType *VT) override; |
| 310 | }; |
| 311 | } // namespace |
| 312 | |
| 313 | // This code is adapted from the type remapper in lib/Linker/IRMover.cpp |
| 314 | Type *BufferFatPtrTypeLoweringBase::remapTypeImpl(Type *Ty) { |
| 315 | Type **Entry = &Map[Ty]; |
| 316 | if (*Entry) |
| 317 | return *Entry; |
| 318 | if (auto *PT = dyn_cast<PointerType>(Val: Ty)) { |
| 319 | if (PT->getAddressSpace() == AMDGPUAS::BUFFER_FAT_POINTER) { |
| 320 | return *Entry = remapScalar(PT); |
| 321 | } |
| 322 | } |
| 323 | if (auto *VT = dyn_cast<VectorType>(Val: Ty)) { |
| 324 | auto *PT = dyn_cast<PointerType>(Val: VT->getElementType()); |
| 325 | if (PT && PT->getAddressSpace() == AMDGPUAS::BUFFER_FAT_POINTER) { |
| 326 | return *Entry = remapVector(VT); |
| 327 | } |
| 328 | return *Entry = Ty; |
| 329 | } |
| 330 | // Whether the type is one that is structurally uniqued - that is, if it is |
| 331 | // not a named struct (the only kind of type where multiple structurally |
| 332 | // identical types that have a distinct `Type*`) |
| 333 | StructType *TyAsStruct = dyn_cast<StructType>(Val: Ty); |
| 334 | bool IsUniqued = !TyAsStruct || TyAsStruct->isLiteral(); |
| 335 | // Base case for ints, floats, opaque pointers, and so on, which don't |
| 336 | // require recursion. |
| 337 | if (Ty->getNumContainedTypes() == 0 && IsUniqued) |
| 338 | return *Entry = Ty; |
| 339 | bool Changed = false; |
| 340 | SmallVector<Type *> ElementTypes(Ty->getNumContainedTypes(), nullptr); |
| 341 | for (unsigned int I = 0, E = Ty->getNumContainedTypes(); I < E; ++I) { |
| 342 | Type *OldElem = Ty->getContainedType(i: I); |
| 343 | Type *NewElem = remapTypeImpl(Ty: OldElem); |
| 344 | ElementTypes[I] = NewElem; |
| 345 | Changed |= (OldElem != NewElem); |
| 346 | } |
| 347 | // Recursive calls to remapTypeImpl() may have invalidated pointer. |
| 348 | Entry = &Map[Ty]; |
| 349 | if (!Changed) { |
| 350 | return *Entry = Ty; |
| 351 | } |
| 352 | if (auto *ArrTy = dyn_cast<ArrayType>(Val: Ty)) |
| 353 | return *Entry = ArrayType::get(ElementType: ElementTypes[0], NumElements: ArrTy->getNumElements()); |
| 354 | if (auto *FnTy = dyn_cast<FunctionType>(Val: Ty)) |
| 355 | return *Entry = FunctionType::get(Result: ElementTypes[0], |
| 356 | Params: ArrayRef(ElementTypes).slice(N: 1), |
| 357 | isVarArg: FnTy->isVarArg()); |
| 358 | if (auto *STy = dyn_cast<StructType>(Val: Ty)) { |
| 359 | // Genuine opaque types don't have a remapping. |
| 360 | if (STy->isOpaque()) |
| 361 | return *Entry = Ty; |
| 362 | bool IsPacked = STy->isPacked(); |
| 363 | if (IsUniqued) |
| 364 | return *Entry = StructType::get(Context&: Ty->getContext(), Elements: ElementTypes, isPacked: IsPacked); |
| 365 | SmallString<16> Name(STy->getName()); |
| 366 | STy->setName(""); |
| 367 | return *Entry = StructType::create(Context&: Ty->getContext(), Elements: ElementTypes, Name, |
| 368 | isPacked: IsPacked); |
| 369 | } |
| 370 | llvm_unreachable("Unknown type of type that contains elements"); |
| 371 | } |
| 372 | |
| 373 | Type *BufferFatPtrTypeLoweringBase::remapType(Type *SrcTy) { |
| 374 | return remapTypeImpl(Ty: SrcTy); |
| 375 | } |
| 376 | |
| 377 | Type *BufferFatPtrToStructTypeMap::remapScalar(PointerType *PT) { |
| 378 | LLVMContext &Ctx = PT->getContext(); |
| 379 | return StructType::get(elt1: PointerType::get(C&: Ctx, AddressSpace: AMDGPUAS::BUFFER_RESOURCE), |
| 380 | elts: IntegerType::get(C&: Ctx, NumBits: BufferOffsetWidth)); |
| 381 | } |
| 382 | |
| 383 | Type *BufferFatPtrToStructTypeMap::remapVector(VectorType *VT) { |
| 384 | ElementCount EC = VT->getElementCount(); |
| 385 | LLVMContext &Ctx = VT->getContext(); |
| 386 | Type *RsrcVec = |
| 387 | VectorType::get(ElementType: PointerType::get(C&: Ctx, AddressSpace: AMDGPUAS::BUFFER_RESOURCE), EC); |
| 388 | Type *OffVec = VectorType::get(ElementType: IntegerType::get(C&: Ctx, NumBits: BufferOffsetWidth), EC); |
| 389 | return StructType::get(elt1: RsrcVec, elts: OffVec); |
| 390 | } |
| 391 | |
| 392 | static bool isBufferFatPtrOrVector(Type *Ty) { |
| 393 | if (auto *PT = dyn_cast<PointerType>(Val: Ty->getScalarType())) |
| 394 | return PT->getAddressSpace() == AMDGPUAS::BUFFER_FAT_POINTER; |
| 395 | return false; |
| 396 | } |
| 397 | |
| 398 | // True if the type is {ptr addrspace(8), i32} or a struct containing vectors of |
| 399 | // those types. Used to quickly skip instructions we don't need to process. |
| 400 | static bool isSplitFatPtr(Type *Ty) { |
| 401 | auto *ST = dyn_cast<StructType>(Val: Ty); |
| 402 | if (!ST) |
| 403 | return false; |
| 404 | if (!ST->isLiteral() || ST->getNumElements() != 2) |
| 405 | return false; |
| 406 | auto *MaybeRsrc = |
| 407 | dyn_cast<PointerType>(Val: ST->getElementType(N: 0)->getScalarType()); |
| 408 | auto *MaybeOff = |
| 409 | dyn_cast<IntegerType>(Val: ST->getElementType(N: 1)->getScalarType()); |
| 410 | return MaybeRsrc && MaybeOff && |
| 411 | MaybeRsrc->getAddressSpace() == AMDGPUAS::BUFFER_RESOURCE && |
| 412 | MaybeOff->getBitWidth() == BufferOffsetWidth; |
| 413 | } |
| 414 | |
| 415 | // True if the result type or any argument types are buffer fat pointers. |
| 416 | static bool isBufferFatPtrConst(Constant *C) { |
| 417 | Type *T = C->getType(); |
| 418 | return isBufferFatPtrOrVector(Ty: T) || any_of(Range: C->operands(), P: [](const Use &U) { |
| 419 | return isBufferFatPtrOrVector(Ty: U.get()->getType()); |
| 420 | }); |
| 421 | } |
| 422 | |
| 423 | namespace { |
| 424 | /// Convert [vectors of] buffer fat pointers to integers when they are read from |
| 425 | /// or stored to memory. This ensures that these pointers will have the same |
| 426 | /// memory layout as before they are lowered, even though they will no longer |
| 427 | /// have their previous layout in registers/in the program (they'll be broken |
| 428 | /// down into resource and offset parts). This has the downside of imposing |
| 429 | /// marshalling costs when reading or storing these values, but since placing |
| 430 | /// such pointers into memory is an uncommon operation at best, we feel that |
| 431 | /// this cost is acceptable for better performance in the common case. |
| 432 | class StoreFatPtrsAsIntsAndExpandMemcpyVisitor |
| 433 | : public InstVisitor<StoreFatPtrsAsIntsAndExpandMemcpyVisitor, bool> { |
| 434 | BufferFatPtrToIntTypeMap *TypeMap; |
| 435 | |
| 436 | IRBuilder<InstSimplifyFolder> IRB; |
| 437 | |
| 438 | const DataLayout &DL; |
| 439 | |
| 440 | // Used for memcpy() lowering. |
| 441 | const TargetTransformInfo *TTI; |
| 442 | ScalarEvolution *SE; |
| 443 | |
| 444 | Value *applyOffset(Value *Ptr, uint64_t Off); |
| 445 | // Visits each maximal subtree of `Ty` that is fat-ptr-free or is itself a |
| 446 | // [vector of] fat pointer(s), at its offset in `Ty`'s original layout. |
| 447 | void forEachAggLeaf( |
| 448 | Type *Ty, SmallVectorImpl<unsigned> &AggIdxs, uint64_t Off, |
| 449 | const Twine &Name, |
| 450 | function_ref<void(Type *LeafTy, Type *IntLeafTy, ArrayRef<unsigned> Idxs, |
| 451 | uint64_t Off, const Twine &Name)> |
| 452 | Visit); |
| 453 | |
| 454 | public: |
| 455 | StoreFatPtrsAsIntsAndExpandMemcpyVisitor(BufferFatPtrToIntTypeMap *TypeMap, |
| 456 | Module &M) |
| 457 | : TypeMap(TypeMap), IRB(M, InstSimplifyFolder(M.getDataLayout())), |
| 458 | DL(M.getDataLayout()) {} |
| 459 | bool processFunction(Function &F, const TargetTransformInfo *TTI, |
| 460 | ScalarEvolution *SE); |
| 461 | |
| 462 | bool visitInstruction(Instruction &I) { return false; } |
| 463 | bool visitAllocaInst(AllocaInst &I); |
| 464 | bool visitLoadInst(LoadInst &LI); |
| 465 | bool visitStoreInst(StoreInst &SI); |
| 466 | bool visitGetElementPtrInst(GetElementPtrInst &I); |
| 467 | |
| 468 | bool visitMemCpyInst(MemCpyInst &MCI); |
| 469 | bool visitMemMoveInst(MemMoveInst &MMI); |
| 470 | bool visitMemSetInst(MemSetInst &MSI); |
| 471 | bool visitMemSetPatternInst(MemSetPatternInst &MSPI); |
| 472 | }; |
| 473 | } // namespace |
| 474 | |
| 475 | Value *StoreFatPtrsAsIntsAndExpandMemcpyVisitor::applyOffset(Value *Ptr, |
| 476 | uint64_t Off) { |
| 477 | // The InstSimplifyFolder gives back `Ptr` itself when `Off` is 0. |
| 478 | return IRB.CreatePtrAdd( |
| 479 | Ptr, Offset: ConstantInt::get(Ty: DL.getIndexType(PtrTy: Ptr->getType()), V: Off), |
| 480 | Name: Ptr->getName() + ".off."+ Twine(Off), NW: GEPNoWrapFlags::noUnsignedWrap()); |
| 481 | } |
| 482 | |
| 483 | void StoreFatPtrsAsIntsAndExpandMemcpyVisitor::forEachAggLeaf( |
| 484 | Type *Ty, SmallVectorImpl<unsigned> &AggIdxs, uint64_t Off, |
| 485 | const Twine &Name, |
| 486 | function_ref<void(Type *LeafTy, Type *IntLeafTy, ArrayRef<unsigned> Idxs, |
| 487 | uint64_t Off, const Twine &Name)> |
| 488 | Visit) { |
| 489 | Type *IntTy = TypeMap->remapType(SrcTy: Ty); |
| 490 | if (isBufferFatPtrOrVector(Ty) || Ty == IntTy) { |
| 491 | // Zero-sized leaves ({} or [0 x T]) access no bytes; skip them. |
| 492 | if (DL.getTypeStoreSize(Ty) != 0) |
| 493 | Visit(Ty, IntTy, AggIdxs, Off, Name); |
| 494 | return; |
| 495 | } |
| 496 | auto Recurse = [&](unsigned I, Type *ElemTy, uint64_t ElemOff) { |
| 497 | AggIdxs.push_back(Elt: I); |
| 498 | forEachAggLeaf(Ty: ElemTy, AggIdxs, Off: Off + ElemOff, Name: Name + "."+ Twine(I), |
| 499 | Visit); |
| 500 | AggIdxs.pop_back(); |
| 501 | }; |
| 502 | if (auto *ST = dyn_cast<StructType>(Val: Ty)) { |
| 503 | const StructLayout *Layout = DL.getStructLayout(Ty: ST); |
| 504 | for (auto [I, ElemTy, ElemOff] : |
| 505 | enumerate(First: ST->elements(), Rest: Layout->getMemberOffsets())) |
| 506 | Recurse(I, ElemTy, ElemOff.getFixedValue()); |
| 507 | return; |
| 508 | } |
| 509 | auto *AT = cast<ArrayType>(Val: Ty); |
| 510 | Type *ElemTy = AT->getElementType(); |
| 511 | uint64_t Stride = DL.getTypeAllocSize(Ty: ElemTy).getFixedValue(); |
| 512 | for (unsigned I : seq<unsigned>(Size: AT->getNumElements())) |
| 513 | Recurse(I, ElemTy, I * Stride); |
| 514 | } |
| 515 | |
| 516 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::processFunction( |
| 517 | Function &F, const TargetTransformInfo *TTI, ScalarEvolution *SE) { |
| 518 | this->TTI = TTI; |
| 519 | this->SE = SE; |
| 520 | bool Changed = false; |
| 521 | // Process memcpy-like instructions after the main iteration because they can |
| 522 | // invalidate iterators. |
| 523 | SmallVector<WeakTrackingVH> CanBecomeLoops; |
| 524 | for (Instruction &I : make_early_inc_range(Range: instructions(F))) { |
| 525 | if (isa<MemTransferInst, MemSetInst, MemSetPatternInst>(Val: I)) |
| 526 | CanBecomeLoops.push_back(Elt: &I); |
| 527 | else |
| 528 | Changed |= visit(I); |
| 529 | } |
| 530 | for (WeakTrackingVH VH : make_early_inc_range(Range&: CanBecomeLoops)) { |
| 531 | Changed |= visit(I: cast<Instruction>(Val&: VH)); |
| 532 | } |
| 533 | this->TTI = nullptr; |
| 534 | this->SE = nullptr; |
| 535 | return Changed; |
| 536 | } |
| 537 | |
| 538 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitAllocaInst(AllocaInst &I) { |
| 539 | Type *Ty = I.getAllocatedType(); |
| 540 | Type *NewTy = TypeMap->remapType(SrcTy: Ty); |
| 541 | if (Ty == NewTy) |
| 542 | return false; |
| 543 | // i160 is smaller than ptr addrspace(7) (24 bytes vs. 32); fall back to a |
| 544 | // byte array of the original size so sizes computed from Ty stay in bounds. |
| 545 | TypeSize AllocSize = DL.getTypeAllocSize(Ty); |
| 546 | if (AllocSize.isFixed() && DL.getTypeAllocSize(Ty: NewTy) != AllocSize) |
| 547 | NewTy = ArrayType::get(ElementType: IRB.getInt8Ty(), NumElements: AllocSize.getFixedValue()); |
| 548 | I.setAllocatedType(NewTy); |
| 549 | return true; |
| 550 | } |
| 551 | |
| 552 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitGetElementPtrInst( |
| 553 | GetElementPtrInst &I) { |
| 554 | Type *Ty = I.getSourceElementType(); |
| 555 | if (Ty == TypeMap->remapType(SrcTy: Ty)) |
| 556 | return false; |
| 557 | // Lower to a byte offset now, before remapping changes p7's layout (see file |
| 558 | // header). |
| 559 | IRB.SetInsertPoint(&I); |
| 560 | Value *Off = emitGEPOffset(Builder: &IRB, DL, GEP: &I); |
| 561 | Value *NewGEP = IRB.CreatePtrAdd(Ptr: I.getPointerOperand(), Offset: Off, Name: I.getName(), |
| 562 | NW: I.getNoWrapFlags()); |
| 563 | I.replaceAllUsesWith(V: NewGEP); |
| 564 | I.eraseFromParent(); |
| 565 | return true; |
| 566 | } |
| 567 | |
| 568 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitLoadInst(LoadInst &LI) { |
| 569 | Type *Ty = LI.getType(); |
| 570 | Type *IntTy = TypeMap->remapType(SrcTy: Ty); |
| 571 | if (Ty == IntTy) |
| 572 | return false; |
| 573 | |
| 574 | IRB.SetInsertPoint(&LI); |
| 575 | if (!isBufferFatPtrOrVector(Ty)) { |
| 576 | // i160 has the same 20-byte store size as p7, so loading each leaf at |
| 577 | // its original-layout offset accesses the same bytes as the unlowered load. |
| 578 | Value *Agg = PoisonValue::get(T: Ty); |
| 579 | AAMDNodes AATags = LI.getAAMetadata(); |
| 580 | SmallVector<unsigned> AggIdxs; |
| 581 | forEachAggLeaf( |
| 582 | Ty, AggIdxs, Off: 0, Name: LI.getName(), |
| 583 | Visit: [&](Type *LeafTy, Type *IntLeafTy, ArrayRef<unsigned> Idxs, |
| 584 | uint64_t Off, const Twine &Name) { |
| 585 | Value *Ptr = applyOffset(Ptr: LI.getPointerOperand(), Off); |
| 586 | LoadInst *NewLI = IRB.CreateAlignedLoad( |
| 587 | Ty: IntLeafTy, Ptr, Align: commonAlignment(A: LI.getAlign(), Offset: Off), Name); |
| 588 | NewLI->setVolatile(LI.isVolatile()); |
| 589 | copyMetadataForLoad(Dest&: *NewLI, Source: LI); |
| 590 | NewLI->setAAMetadata(AATags.adjustForAccess(Offset: Off, AccessTy: IntLeafTy, DL)); |
| 591 | Value *V = NewLI; |
| 592 | if (LeafTy != IntLeafTy) |
| 593 | V = IRB.CreateIntToPtr(V: NewLI, DestTy: LeafTy, Name: Name + ".ptr"); |
| 594 | Agg = IRB.CreateInsertValue(Agg, Val: V, Idxs, Name: Name + ".agg"); |
| 595 | }); |
| 596 | LI.replaceAllUsesWith(V: Agg); |
| 597 | LI.eraseFromParent(); |
| 598 | return true; |
| 599 | } |
| 600 | auto *NLI = cast<LoadInst>(Val: LI.clone()); |
| 601 | NLI->mutateType(Ty: IntTy); |
| 602 | NLI = IRB.Insert(I: NLI); |
| 603 | NLI->takeName(V: &LI); |
| 604 | |
| 605 | Value *CastBack = IRB.CreateIntToPtr(V: NLI, DestTy: Ty, Name: NLI->getName() + ".ptr"); |
| 606 | LI.replaceAllUsesWith(V: CastBack); |
| 607 | LI.eraseFromParent(); |
| 608 | return true; |
| 609 | } |
| 610 | |
| 611 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitStoreInst(StoreInst &SI) { |
| 612 | Value *V = SI.getValueOperand(); |
| 613 | Type *Ty = V->getType(); |
| 614 | Type *IntTy = TypeMap->remapType(SrcTy: Ty); |
| 615 | if (Ty == IntTy) |
| 616 | return false; |
| 617 | |
| 618 | IRB.SetInsertPoint(&SI); |
| 619 | if (!isBufferFatPtrOrVector(Ty)) { |
| 620 | // Store each leaf at its byte offset in the original layout; see |
| 621 | // visitLoadInst. |
| 622 | AAMDNodes AATags = SI.getAAMetadata(); |
| 623 | SmallVector<unsigned> AggIdxs; |
| 624 | forEachAggLeaf( |
| 625 | Ty, AggIdxs, Off: 0, Name: V->getName(), |
| 626 | Visit: [&](Type *LeafTy, Type *IntLeafTy, ArrayRef<unsigned> Idxs, |
| 627 | uint64_t Off, const Twine &Name) { |
| 628 | Value *Leaf = IRB.CreateExtractValue(Agg: V, Idxs, Name); |
| 629 | if (LeafTy != IntLeafTy) |
| 630 | Leaf = IRB.CreatePtrToInt(V: Leaf, DestTy: IntLeafTy, Name: Name + ".int"); |
| 631 | auto *NewSI = cast<StoreInst>(Val: SI.clone()); |
| 632 | NewSI->setAlignment(commonAlignment(A: SI.getAlign(), Offset: Off)); |
| 633 | NewSI->setOperand(i_nocapture: 0, Val_nocapture: Leaf); |
| 634 | NewSI->setOperand(i_nocapture: 1, Val_nocapture: applyOffset(Ptr: SI.getPointerOperand(), Off)); |
| 635 | // Each leaf covers only part of the original assignment. |
| 636 | NewSI->setMetadata(KindID: LLVMContext::MD_DIAssignID, Node: nullptr); |
| 637 | IRB.Insert(I: NewSI); |
| 638 | NewSI->setAAMetadata(AATags.adjustForAccess(Offset: Off, AccessTy: IntLeafTy, DL)); |
| 639 | }); |
| 640 | SI.eraseFromParent(); |
| 641 | return true; |
| 642 | } |
| 643 | Value *IntV = IRB.CreatePtrToInt(V, DestTy: IntTy, Name: V->getName() + ".int"); |
| 644 | for (auto *Dbg : at::getDVRAssignmentMarkers(Inst: &SI)) |
| 645 | Dbg->setRawLocation(ValueAsMetadata::get(V: IntV)); |
| 646 | |
| 647 | SI.setOperand(i_nocapture: 0, Val_nocapture: IntV); |
| 648 | return true; |
| 649 | } |
| 650 | |
| 651 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitMemCpyInst( |
| 652 | MemCpyInst &MCI) { |
| 653 | // TODO: Allow memcpy.p7.p3 as a synonym for the direct-to-LDS copy, which'll |
| 654 | // need loop expansion here. |
| 655 | if (MCI.getSourceAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER && |
| 656 | MCI.getDestAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 657 | return false; |
| 658 | llvm::expandMemCpyAsLoop(MemCpy: &MCI, TTI: *TTI, SE); |
| 659 | MCI.eraseFromParent(); |
| 660 | return true; |
| 661 | } |
| 662 | |
| 663 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitMemMoveInst( |
| 664 | MemMoveInst &MMI) { |
| 665 | if (MMI.getSourceAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER && |
| 666 | MMI.getDestAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 667 | return false; |
| 668 | reportFatalUsageError( |
| 669 | reason: "memmove() on buffer descriptors is not implemented because pointer " |
| 670 | "comparison on buffer descriptors isn't implemented\n"); |
| 671 | } |
| 672 | |
| 673 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitMemSetInst( |
| 674 | MemSetInst &MSI) { |
| 675 | if (MSI.getDestAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 676 | return false; |
| 677 | llvm::expandMemSetAsLoop(MemSet: &MSI, TTI); |
| 678 | MSI.eraseFromParent(); |
| 679 | return true; |
| 680 | } |
| 681 | |
| 682 | bool StoreFatPtrsAsIntsAndExpandMemcpyVisitor::visitMemSetPatternInst( |
| 683 | MemSetPatternInst &MSPI) { |
| 684 | if (MSPI.getDestAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 685 | return false; |
| 686 | llvm::expandMemSetPatternAsLoop(MemSet: &MSPI, TTI: *TTI); |
| 687 | MSPI.eraseFromParent(); |
| 688 | return true; |
| 689 | } |
| 690 | |
| 691 | namespace { |
| 692 | /// Convert loads/stores of types that the buffer intrinsics can't handle into |
| 693 | /// one ore more such loads/stores that consist of legal types. |
| 694 | /// |
| 695 | /// Do this by |
| 696 | /// 1. Recursing into structs (and arrays that don't share a memory layout with |
| 697 | /// vectors) since the intrinsics can't handle complex types. |
| 698 | /// 2. Converting arrays of non-aggregate, byte-sized types into their |
| 699 | /// corresponding vectors |
| 700 | /// 3. Bitcasting unsupported types, namely overly-long scalars and byte |
| 701 | /// vectors, into vectors of supported types. |
| 702 | /// 4. Splitting up excessively long reads/writes into multiple operations. |
| 703 | /// |
| 704 | /// Note that this doesn't handle complex data strucures, but, in the future, |
| 705 | /// the aggregate load splitter from SROA could be refactored to allow for that |
| 706 | /// case. |
| 707 | /// |
| 708 | /// Note that, if we can prove that the initial value of the pointer offset is 0 |
| 709 | /// and that the load/store won't wrap from the left or won't have bounds checks |
| 710 | /// that straddle a word boundary, we can emit some of the strict bounds |
| 711 | /// checking pessimizations even in strict OOB mode, and we attempt to do so. |
| 712 | class LegalizeBufferContentTypesVisitor |
| 713 | : public InstVisitor<LegalizeBufferContentTypesVisitor, bool> { |
| 714 | friend class InstVisitor<LegalizeBufferContentTypesVisitor, bool>; |
| 715 | |
| 716 | IRBuilder<InstSimplifyFolder> IRB; |
| 717 | |
| 718 | const DataLayout &DL; |
| 719 | |
| 720 | ScalarEvolution *SE = nullptr; |
| 721 | |
| 722 | // Map base (non-GEP'd) pointers to the number of records they have, if known. |
| 723 | // If a pointer is known to have a starting offset of 0 but it wasn't known to |
| 724 | // have a number of records (ex. it was `addrspacecast` from a buffer |
| 725 | // resource), it will be present in this map, but the key will be null. |
| 726 | // Otherwise, there will be no map entry. |
| 727 | ValueToValueMapTy ZeroBasePointerToNumRecords; |
| 728 | |
| 729 | // Subtarget info, needed for determining what cache control bits to set. |
| 730 | const TargetMachine *TM; |
| 731 | const GCNSubtarget *ST = nullptr; |
| 732 | |
| 733 | /// If T is [N x U], where U is a scalar type, return the vector type |
| 734 | /// <N x U>, otherwise, return T. |
| 735 | Type *scalarArrayTypeAsVector(Type *MaybeArrayType); |
| 736 | Value *arrayToVector(Value *V, Type *TargetType, const Twine &Name); |
| 737 | Value *vectorToArray(Value *V, Type *OrigType, const Twine &Name); |
| 738 | |
| 739 | /// Analyze how a given buffer access could be out of bounds. Used to optimize |
| 740 | /// the strict splitting used in strict bounds checking mode. |
| 741 | struct OobProperties { |
| 742 | // Offset is far enough from all-1s that we won't get wrapping around to 0. |
| 743 | bool NoWrapFromMax = false; |
| 744 | // Offset is either entirely in-bounds or entirely out of bounds. |
| 745 | bool NoPartialOOB = false; |
| 746 | |
| 747 | OobProperties() = delete; |
| 748 | // Needed for some Clangs. |
| 749 | OobProperties(bool NoWrapFromMax, bool NoPartialOOB) |
| 750 | : NoWrapFromMax(NoWrapFromMax), NoPartialOOB(NoPartialOOB) {} |
| 751 | }; |
| 752 | OobProperties analyzeOobProperties(Value *Ptr, Type *Ty, uint64_t ByteOffset); |
| 753 | |
| 754 | /// Break up the loads of a struct into the loads of its components |
| 755 | |
| 756 | /// Return the maximum allowed load/store width for the given type and |
| 757 | /// alignment combination based on subtarget flags. |
| 758 | /// 1. If unaligned accesses are not enabled, then any load/store that is less |
| 759 | /// than word-aligned has to be handled one byte or ushort at a time. |
| 760 | /// 2. If relaxed OOB mode is not set, we must ensure that the in-bounds |
| 761 | /// part of a partially out of bounds read/write is performed correctly. This |
| 762 | /// means that any load that isn't naturally aligned has to be split into |
| 763 | /// parts that are naturally aligned, so that, after bitcasting, we don't have |
| 764 | /// unaligned loads that could discard valid data. |
| 765 | /// |
| 766 | /// For example, if we're loading a <8 x i8>, that's actually a load of a <2 x |
| 767 | /// i32>, and if we load from an align(2) address, that address might be 2 |
| 768 | /// bytes from the end of the buffer. The hardware will, when performing the |
| 769 | /// <2 x i32> load, mask off the entire first word, causing the two in-bounds |
| 770 | /// bytes to be masked off. However,if we know the offset can't be too close |
| 771 | /// to the number of records in the buffer (if known), we can skip this |
| 772 | /// expansion. |
| 773 | /// |
| 774 | /// Unlike the complete disablement of unaligned accesses from point 1, |
| 775 | /// this does not apply to unaligned scalars, but will apply to cases like |
| 776 | /// `load <2 x i32>, align 4` since the left elemenvt might be out of bounds. |
| 777 | /// Note that if the we know that the base offset is known to be |
| 778 | /// less than `uint32_max - byte_size(Ty)`, we can skip these alignment |
| 779 | /// checks. |
| 780 | uint64_t maxIntrinsicWidth(Type *Ty, Align A, OobProperties OobProps); |
| 781 | |
| 782 | /// Convert a vector or scalar type that can't be operated on by buffer |
| 783 | /// intrinsics to one that would be legal through bitcasts and/or truncation. |
| 784 | /// Uses the wider of i32, i16, or i8 where possible, clamping to the maximum |
| 785 | /// allowed width under the alignment rules and subtarget flags. |
| 786 | Type *legalNonAggregateForMemOp(Type *T, uint64_t MaxWidth); |
| 787 | Value *makeLegalNonAggregate(Value *V, Type *TargetType, const Twine &Name); |
| 788 | Value *makeIllegalNonAggregate(Value *V, Type *OrigType, const Twine &Name); |
| 789 | |
| 790 | struct VecSlice { |
| 791 | uint64_t Index = 0; |
| 792 | uint64_t Length = 0; |
| 793 | VecSlice() = delete; |
| 794 | // Needed for some Clangs |
| 795 | VecSlice(uint64_t Index, uint64_t Length) : Index(Index), Length(Length) {} |
| 796 | }; |
| 797 | /// Return the [index, length] pairs into which `T` needs to be cut to form |
| 798 | /// legal buffer load or store operations. Clears `Slices`. Creates an empty |
| 799 | /// `Slices` for non-vector inputs and creates one slice if no slicing will be |
| 800 | /// needed. No slice may be larger than `MaxWidth`. |
| 801 | void getVecSlices(Type *T, uint64_t MaxWidth, |
| 802 | SmallVectorImpl<VecSlice> &Slices); |
| 803 | |
| 804 | Value *extractSlice(Value *Vec, VecSlice S, const Twine &Name); |
| 805 | Value *insertSlice(Value *Whole, Value *Part, VecSlice S, const Twine &Name); |
| 806 | |
| 807 | /// In most cases, return `LegalType`. However, when given an input that would |
| 808 | /// normally be a legal type for the buffer intrinsics to return but that |
| 809 | /// isn't hooked up through SelectionDAG, return a type of the same width that |
| 810 | /// can be used with the relevant intrinsics. Specifically, handle the cases: |
| 811 | /// - <1 x T> => T for all T |
| 812 | /// - <N x i8> <=> i16, i32, 2xi32, 4xi32 (as needed) |
| 813 | /// - <N x T> where T is under 32 bits and the total size is 96 bits <=> <3 x |
| 814 | /// i32> |
| 815 | Type *intrinsicTypeFor(Type *LegalType); |
| 816 | |
| 817 | bool visitLoadImpl(LoadInst &OrigLI, Type *PartType, |
| 818 | SmallVectorImpl<uint32_t> &AggIdxs, uint64_t AggByteOffset, |
| 819 | Value *&Result, const Twine &Name); |
| 820 | /// Return value is (Changed, ModifiedInPlace) |
| 821 | std::pair<bool, bool> visitStoreImpl(StoreInst &OrigSI, Type *PartType, |
| 822 | SmallVectorImpl<uint32_t> &AggIdxs, |
| 823 | uint64_t AggByteOffset, |
| 824 | const Twine &Name); |
| 825 | |
| 826 | bool visitInstruction(Instruction &I) { return false; } |
| 827 | bool visitLoadInst(LoadInst &LI); |
| 828 | bool visitStoreInst(StoreInst &SI); |
| 829 | |
| 830 | // Record base pointer data and num_records (if known). |
| 831 | bool visitIntrinsicInst(IntrinsicInst &II); |
| 832 | bool visitAddrSpaceCastInst(AddrSpaceCastInst &ASCI); |
| 833 | |
| 834 | public: |
| 835 | LegalizeBufferContentTypesVisitor(Module &M, const TargetMachine *TM) |
| 836 | : IRB(M, InstSimplifyFolder(M.getDataLayout())), DL(M.getDataLayout()), |
| 837 | TM(TM) {} |
| 838 | bool processFunction(Function &F, ScalarEvolution *SE); |
| 839 | }; |
| 840 | } // namespace |
| 841 | |
| 842 | Type *LegalizeBufferContentTypesVisitor::scalarArrayTypeAsVector(Type *T) { |
| 843 | ArrayType *AT = dyn_cast<ArrayType>(Val: T); |
| 844 | if (!AT) |
| 845 | return T; |
| 846 | Type *ET = AT->getElementType(); |
| 847 | if (!ET->isSingleValueType() || isa<VectorType>(Val: ET)) |
| 848 | reportFatalUsageError(reason: "loading non-scalar arrays from buffer fat pointers " |
| 849 | "should have recursed"); |
| 850 | if (!DL.typeSizeEqualsStoreSize(Ty: AT)) |
| 851 | reportFatalUsageError( |
| 852 | reason: "loading padded arrays from buffer fat pinters should have recursed"); |
| 853 | return FixedVectorType::get(ElementType: ET, NumElts: AT->getNumElements()); |
| 854 | } |
| 855 | |
| 856 | Value *LegalizeBufferContentTypesVisitor::arrayToVector(Value *V, |
| 857 | Type *TargetType, |
| 858 | const Twine &Name) { |
| 859 | Value *VectorRes = PoisonValue::get(T: TargetType); |
| 860 | auto *VT = cast<FixedVectorType>(Val: TargetType); |
| 861 | unsigned EC = VT->getNumElements(); |
| 862 | for (auto I : iota_range<unsigned>(0, EC, /*Inclusive=*/false)) { |
| 863 | Value *Elem = IRB.CreateExtractValue(Agg: V, Idxs: I, Name: Name + ".elem."+ Twine(I)); |
| 864 | VectorRes = IRB.CreateInsertElement(Vec: VectorRes, NewElt: Elem, Idx: I, |
| 865 | Name: Name + ".as.vec."+ Twine(I)); |
| 866 | } |
| 867 | return VectorRes; |
| 868 | } |
| 869 | |
| 870 | Value *LegalizeBufferContentTypesVisitor::vectorToArray(Value *V, |
| 871 | Type *OrigType, |
| 872 | const Twine &Name) { |
| 873 | Value *ArrayRes = PoisonValue::get(T: OrigType); |
| 874 | ArrayType *AT = cast<ArrayType>(Val: OrigType); |
| 875 | unsigned EC = AT->getNumElements(); |
| 876 | for (auto I : iota_range<unsigned>(0, EC, /*Inclusive=*/false)) { |
| 877 | Value *Elem = IRB.CreateExtractElement(Vec: V, Idx: I, Name: Name + ".elem."+ Twine(I)); |
| 878 | ArrayRes = IRB.CreateInsertValue(Agg: ArrayRes, Val: Elem, Idxs: I, |
| 879 | Name: Name + ".as.array."+ Twine(I)); |
| 880 | } |
| 881 | return ArrayRes; |
| 882 | } |
| 883 | |
| 884 | LegalizeBufferContentTypesVisitor::OobProperties |
| 885 | LegalizeBufferContentTypesVisitor::analyzeOobProperties(Value *Ptr, Type *Ty, |
| 886 | uint64_t ByteOffset) { |
| 887 | OobProperties Result(false, false); |
| 888 | |
| 889 | if (ST->hasRelaxedBufferOOBMode()) |
| 890 | return OobProperties(true, true); |
| 891 | |
| 892 | if (!SE) |
| 893 | return Result; |
| 894 | if (!SE->isSCEVable(Ty: Ptr->getType())) |
| 895 | return Result; |
| 896 | const SCEV *PtrOp = SE->getSCEV(V: Ptr); |
| 897 | if (ByteOffset > 0) |
| 898 | PtrOp = SE->getAddExpr(LHS: PtrOp, RHS: SE->getConstant(V: IRB.getInt32(C: ByteOffset))); |
| 899 | const auto *PtrBase = dyn_cast<SCEVUnknown>(Val: SE->getPointerBase(V: PtrOp)); |
| 900 | if (!PtrBase) |
| 901 | return Result; |
| 902 | Value *PtrBaseVal = PtrBase->getValue(); |
| 903 | // We don't know if the offset field started at 0, so there's no safe analysis |
| 904 | // we can do. If it weren't for the fact that nuw / inbounds / ... are |
| 905 | // properties of the pointer, we might be able to use hem, but loads where the |
| 906 | // address computation for sub-parts of the loaded type wraps the address |
| 907 | // space are explicitly in scope here so there's not much we can do inside |
| 908 | // functions that can't "see" the fat pointer creation. |
| 909 | auto NumRecordsIfKnown = ZeroBasePointerToNumRecords.find(Val: PtrBaseVal); |
| 910 | if (NumRecordsIfKnown == ZeroBasePointerToNumRecords.end()) |
| 911 | return Result; |
| 912 | |
| 913 | unsigned TypeSize = DL.getTypeStoreSize(Ty).getKnownMinValue(); |
| 914 | const SCEV *PtrDiff = SE->getMinusSCEV(LHS: PtrOp, RHS: PtrBase); |
| 915 | APInt MaxNoWrapOffset = APInt::getAllOnes(numBits: BufferOffsetWidth) - TypeSize; |
| 916 | if (SE->isKnownNonNegative(S: PtrDiff) || |
| 917 | SE->getUnsignedRangeMax(S: PtrDiff).ule(RHS: MaxNoWrapOffset)) |
| 918 | Result.NoWrapFromMax = true; |
| 919 | |
| 920 | // If we know that the pointer is zero-based but not what its upper bound is, |
| 921 | // we'll need to split up underaligned loads of small types. |
| 922 | if (!NumRecordsIfKnown->second) |
| 923 | return Result; |
| 924 | const SCEV *NumRecords = SE->getSCEV(V: NumRecordsIfKnown->second); |
| 925 | |
| 926 | // We'll normalize all bounds to the num_records width on the hardware. |
| 927 | std::optional<unsigned> MaybeNumRecordsWidth = |
| 928 | ST->getBufferResourceNumRecordsWidth(); |
| 929 | if (!MaybeNumRecordsWidth) |
| 930 | return Result; |
| 931 | unsigned NumRecordsWidth = *MaybeNumRecordsWidth; |
| 932 | Type *NumRecordsTy = IRB.getIntNTy(N: NumRecordsWidth); |
| 933 | // Compare in i64 so wraparound is visible as a negative. |
| 934 | Type *CompareTy = IRB.getInt64Ty(); |
| 935 | const SCEV *Bound = SE->getNoopOrZeroExtend( |
| 936 | V: SE->getTruncateOrZeroExtend(V: NumRecords, Ty: NumRecordsTy), Ty: CompareTy); |
| 937 | |
| 938 | // All-1s is (per ISA or as a consequence of the bounds check rules, depending |
| 939 | // on architecture) no bounds check. |
| 940 | if (Bound == SE->getConstant(Val: APInt::getMaxValue(numBits: NumRecordsWidth) |
| 941 | .zext(width: CompareTy->getIntegerBitWidth()))) |
| 942 | Result.NoPartialOOB = true; |
| 943 | |
| 944 | const SCEV *BoundsDiff = |
| 945 | SE->getMinusSCEV(LHS: Bound, RHS: SE->getNoopOrZeroExtend(V: PtrDiff, Ty: CompareTy)); |
| 946 | |
| 947 | if (SE->getSignedRangeMin(S: BoundsDiff).sge(RHS: TypeSize) || |
| 948 | SE->isKnownNonPositive(S: BoundsDiff)) |
| 949 | Result.NoPartialOOB = true; |
| 950 | return Result; |
| 951 | } |
| 952 | |
| 953 | uint64_t |
| 954 | LegalizeBufferContentTypesVisitor::maxIntrinsicWidth(Type *T, Align A, |
| 955 | OobProperties OobProps) { |
| 956 | Align Result(16); |
| 957 | if (!ST->hasUnalignedBufferAccessEnabled() && A < Align(4)) |
| 958 | Result = A; |
| 959 | auto *VT = dyn_cast<VectorType>(Val: T); |
| 960 | if (!ST->hasRelaxedBufferOOBMode() && VT) { |
| 961 | TypeSize ElemBits = DL.getTypeSizeInBits(Ty: VT->getElementType()); |
| 962 | if (ElemBits.isKnownMultipleOf(RHS: 32)) { |
| 963 | // Word-sized operations are bounds-checked per word. So, the only case we |
| 964 | // have to worry about is stores that start out of bounds and then go in, |
| 965 | // and those can only become in-bounds on a multiple of their alignment. |
| 966 | // Therefore, we can use the declared alignment of the operation as the |
| 967 | // maximum width, rounding up to 4. |
| 968 | if (!OobProps.NoWrapFromMax) |
| 969 | Result = std::min(a: Result, b: std::max(a: A, b: Align(4))); |
| 970 | } else if ((ElemBits.isKnownMultipleOf(RHS: 8) || |
| 971 | isPowerOf2_64(Value: ElemBits.getKnownMinValue()))) { |
| 972 | // To ensure correct behavior for sub-word types, we must always scalarize |
| 973 | // unaligned loads of sub-word types. For example, if you load |
| 974 | // a <4 x i8> from offset 7 in an 8-byte buffer, expecting the vector |
| 975 | // to be padded out with 0s after that last byte, you'll get all 0s |
| 976 | // instead. To prevent this behavior when not requested, de-vectorize such |
| 977 | // loads. |
| 978 | // |
| 979 | // If we knew that the value that triggers bounds checks was a multiple of |
| 980 | // 4 along with the access being word-aligned, we could avoid the |
| 981 | // scalarization here, as the bitcast wouldn't change any check behavior, |
| 982 | // but we don't currently try to analyze this. |
| 983 | // |
| 984 | // Strict OOB checking isn't supported if the size of each element is a |
| 985 | // non-power-of-2 value less than 8, since there's no feasible way to |
| 986 | // apply such a strict bounds check. |
| 987 | if (!OobProps.NoPartialOOB) |
| 988 | Result = |
| 989 | commonAlignment(A: Result, Offset: divideCeil(Numerator: ElemBits.getKnownMinValue(), Denominator: 8)); |
| 990 | } |
| 991 | } |
| 992 | return Result.value() * 8; |
| 993 | } |
| 994 | |
| 995 | Type *LegalizeBufferContentTypesVisitor::legalNonAggregateForMemOp( |
| 996 | Type *T, uint64_t MaxWidth) { |
| 997 | TypeSize Size = DL.getTypeStoreSizeInBits(Ty: T); |
| 998 | // Implicitly zero-extend to the next byte if needed. |
| 999 | if (!DL.typeSizeEqualsStoreSize(Ty: T)) |
| 1000 | T = IRB.getIntNTy(N: Size.getFixedValue()); |
| 1001 | Type *ElemTy = T->getScalarType(); |
| 1002 | if (isa<PointerType, ScalableVectorType>(Val: ElemTy)) { |
| 1003 | // Pointers are always big enough, and we'll let scalable vectors through to |
| 1004 | // fail in codegen. |
| 1005 | return T; |
| 1006 | } |
| 1007 | unsigned ElemSize = DL.getTypeSizeInBits(Ty: ElemTy).getFixedValue(); |
| 1008 | if (isPowerOf2_32(Value: ElemSize) && ElemSize >= 16 && ElemSize <= MaxWidth) { |
| 1009 | // [vectors of] anything that's 16/32/64/128 bits can be cast and split into |
| 1010 | // legal buffer operations, except that we might need to cut them into |
| 1011 | // smaller values if we're not allowed to do unaligned vector loads. |
| 1012 | return T; |
| 1013 | } |
| 1014 | Type *BestVectorElemType = nullptr; |
| 1015 | if (Size.isKnownMultipleOf(RHS: 32) && MaxWidth >= 32) |
| 1016 | BestVectorElemType = IRB.getInt32Ty(); |
| 1017 | else if (Size.isKnownMultipleOf(RHS: 16) && MaxWidth >= 16) |
| 1018 | BestVectorElemType = IRB.getInt16Ty(); |
| 1019 | else |
| 1020 | BestVectorElemType = IRB.getInt8Ty(); |
| 1021 | unsigned NumCastElems = |
| 1022 | Size.getFixedValue() / BestVectorElemType->getIntegerBitWidth(); |
| 1023 | if (NumCastElems == 1) |
| 1024 | return BestVectorElemType; |
| 1025 | return FixedVectorType::get(ElementType: BestVectorElemType, NumElts: NumCastElems); |
| 1026 | } |
| 1027 | |
| 1028 | Value *LegalizeBufferContentTypesVisitor::makeLegalNonAggregate( |
| 1029 | Value *V, Type *TargetType, const Twine &Name) { |
| 1030 | Type *SourceType = V->getType(); |
| 1031 | TypeSize SourceSize = DL.getTypeSizeInBits(Ty: SourceType); |
| 1032 | TypeSize TargetSize = DL.getTypeSizeInBits(Ty: TargetType); |
| 1033 | if (SourceSize != TargetSize) { |
| 1034 | Type *ShortScalarTy = IRB.getIntNTy(N: SourceSize.getFixedValue()); |
| 1035 | Type *ByteScalarTy = IRB.getIntNTy(N: TargetSize.getFixedValue()); |
| 1036 | Value *AsScalar = IRB.CreateBitCast(V, DestTy: ShortScalarTy, Name: Name + ".as.scalar"); |
| 1037 | Value *Zext = IRB.CreateZExt(V: AsScalar, DestTy: ByteScalarTy, Name: Name + ".zext"); |
| 1038 | V = Zext; |
| 1039 | SourceType = ByteScalarTy; |
| 1040 | } |
| 1041 | return IRB.CreateBitCast(V, DestTy: TargetType, Name: Name + ".legal"); |
| 1042 | } |
| 1043 | |
| 1044 | Value *LegalizeBufferContentTypesVisitor::makeIllegalNonAggregate( |
| 1045 | Value *V, Type *OrigType, const Twine &Name) { |
| 1046 | Type *LegalType = V->getType(); |
| 1047 | TypeSize LegalSize = DL.getTypeSizeInBits(Ty: LegalType); |
| 1048 | TypeSize OrigSize = DL.getTypeSizeInBits(Ty: OrigType); |
| 1049 | if (LegalSize != OrigSize) { |
| 1050 | Type *ShortScalarTy = IRB.getIntNTy(N: OrigSize.getFixedValue()); |
| 1051 | Type *ByteScalarTy = IRB.getIntNTy(N: LegalSize.getFixedValue()); |
| 1052 | Value *AsScalar = IRB.CreateBitCast(V, DestTy: ByteScalarTy, Name: Name + ".bytes.cast"); |
| 1053 | Value *Trunc = IRB.CreateTrunc(V: AsScalar, DestTy: ShortScalarTy, Name: Name + ".trunc"); |
| 1054 | return IRB.CreateBitCast(V: Trunc, DestTy: OrigType, Name: Name + ".orig"); |
| 1055 | } |
| 1056 | return IRB.CreateBitCast(V, DestTy: OrigType, Name: Name + ".real.ty"); |
| 1057 | } |
| 1058 | |
| 1059 | Type *LegalizeBufferContentTypesVisitor::intrinsicTypeFor(Type *LegalType) { |
| 1060 | auto *VT = dyn_cast<FixedVectorType>(Val: LegalType); |
| 1061 | if (!VT) |
| 1062 | return LegalType; |
| 1063 | Type *ET = VT->getElementType(); |
| 1064 | // Explicitly return the element type of 1-element vectors because the |
| 1065 | // underlying intrinsics don't like <1 x T> even though it's a synonym for T. |
| 1066 | if (VT->getNumElements() == 1) |
| 1067 | return ET; |
| 1068 | if (DL.getTypeSizeInBits(Ty: LegalType) == 96 && DL.getTypeSizeInBits(Ty: ET) < 32) |
| 1069 | return FixedVectorType::get(ElementType: IRB.getInt32Ty(), NumElts: 3); |
| 1070 | if (ET->isIntegerTy(BitWidth: 8)) { |
| 1071 | switch (VT->getNumElements()) { |
| 1072 | default: |
| 1073 | return LegalType; // Let it crash later |
| 1074 | case 1: |
| 1075 | return IRB.getInt8Ty(); |
| 1076 | case 2: |
| 1077 | return IRB.getInt16Ty(); |
| 1078 | case 4: |
| 1079 | return IRB.getInt32Ty(); |
| 1080 | case 8: |
| 1081 | return FixedVectorType::get(ElementType: IRB.getInt32Ty(), NumElts: 2); |
| 1082 | case 16: |
| 1083 | return FixedVectorType::get(ElementType: IRB.getInt32Ty(), NumElts: 4); |
| 1084 | } |
| 1085 | } |
| 1086 | return LegalType; |
| 1087 | } |
| 1088 | |
| 1089 | void LegalizeBufferContentTypesVisitor::getVecSlices( |
| 1090 | Type *T, uint64_t MaxWidth, SmallVectorImpl<VecSlice> &Slices) { |
| 1091 | Slices.clear(); |
| 1092 | auto *VT = dyn_cast<FixedVectorType>(Val: T); |
| 1093 | if (!VT) |
| 1094 | return; |
| 1095 | |
| 1096 | uint64_t ElemBitWidth = |
| 1097 | DL.getTypeSizeInBits(Ty: VT->getElementType()).getFixedValue(); |
| 1098 | |
| 1099 | uint64_t ElemsPer4Words = 128 / ElemBitWidth; |
| 1100 | uint64_t ElemsPer2Words = ElemsPer4Words / 2; |
| 1101 | uint64_t ElemsPerWord = ElemsPer2Words / 2; |
| 1102 | uint64_t ElemsPerShort = ElemsPerWord / 2; |
| 1103 | uint64_t ElemsPerByte = ElemsPerShort / 2; |
| 1104 | // If the elements evenly pack into 32-bit words, we can use 3-word stores, |
| 1105 | // such as for <6 x bfloat> or <3 x i32>, but we can't dot his for, for |
| 1106 | // example, <3 x i64>, since that's not slicing. |
| 1107 | uint64_t ElemsPer3Words = ElemsPerWord * 3; |
| 1108 | |
| 1109 | uint64_t TotalElems = VT->getNumElements(); |
| 1110 | uint64_t Index = 0; |
| 1111 | auto TrySlice = [&](unsigned MaybeLen, unsigned Width) { |
| 1112 | if (MaybeLen > 0 && Width <= MaxWidth && Index + MaybeLen <= TotalElems) { |
| 1113 | VecSlice Slice{/*Index=*/Index, /*Length=*/MaybeLen}; |
| 1114 | Slices.push_back(Elt: Slice); |
| 1115 | Index += MaybeLen; |
| 1116 | return true; |
| 1117 | } |
| 1118 | return false; |
| 1119 | }; |
| 1120 | while (Index < TotalElems) { |
| 1121 | TrySlice(ElemsPer4Words, 128) || TrySlice(ElemsPer3Words, 96) || |
| 1122 | TrySlice(ElemsPer2Words, 64) || TrySlice(ElemsPerWord, 32) || |
| 1123 | TrySlice(ElemsPerShort, 16) || TrySlice(ElemsPerByte, 8); |
| 1124 | } |
| 1125 | } |
| 1126 | |
| 1127 | Value *LegalizeBufferContentTypesVisitor::extractSlice(Value *Vec, VecSlice S, |
| 1128 | const Twine &Name) { |
| 1129 | auto *VecVT = dyn_cast<FixedVectorType>(Val: Vec->getType()); |
| 1130 | if (!VecVT) |
| 1131 | return Vec; |
| 1132 | if (S.Length == VecVT->getNumElements() && S.Index == 0) |
| 1133 | return Vec; |
| 1134 | if (S.Length == 1) |
| 1135 | return IRB.CreateExtractElement(Vec, Idx: S.Index, |
| 1136 | Name: Name + ".slice."+ Twine(S.Index)); |
| 1137 | SmallVector<int> Mask = llvm::to_vector( |
| 1138 | Range: llvm::iota_range<int>(S.Index, S.Index + S.Length, /*Inclusive=*/false)); |
| 1139 | return IRB.CreateShuffleVector(V: Vec, Mask, Name: Name + ".slice."+ Twine(S.Index)); |
| 1140 | } |
| 1141 | |
| 1142 | Value *LegalizeBufferContentTypesVisitor::insertSlice(Value *Whole, Value *Part, |
| 1143 | VecSlice S, |
| 1144 | const Twine &Name) { |
| 1145 | auto *WholeVT = dyn_cast<FixedVectorType>(Val: Whole->getType()); |
| 1146 | if (!WholeVT) |
| 1147 | return Part; |
| 1148 | if (S.Length == WholeVT->getNumElements() && S.Index == 0) |
| 1149 | return Part; |
| 1150 | if (S.Length == 1) { |
| 1151 | return IRB.CreateInsertElement(Vec: Whole, NewElt: Part, Idx: S.Index, |
| 1152 | Name: Name + ".slice."+ Twine(S.Index)); |
| 1153 | } |
| 1154 | int NumElems = cast<FixedVectorType>(Val: Whole->getType())->getNumElements(); |
| 1155 | |
| 1156 | // Extend the slice with poisons to make the main shufflevector happy. |
| 1157 | SmallVector<int> ExtPartMask(NumElems, -1); |
| 1158 | for (auto [I, E] : llvm::enumerate( |
| 1159 | First: MutableArrayRef<int>(ExtPartMask).take_front(N: S.Length))) { |
| 1160 | E = I; |
| 1161 | } |
| 1162 | Value *ExtPart = IRB.CreateShuffleVector(V: Part, Mask: ExtPartMask, |
| 1163 | Name: Name + ".ext."+ Twine(S.Index)); |
| 1164 | |
| 1165 | SmallVector<int> Mask = |
| 1166 | llvm::to_vector(Range: llvm::iota_range<int>(0, NumElems, /*Inclusive=*/false)); |
| 1167 | for (auto [I, E] : |
| 1168 | llvm::enumerate(First: MutableArrayRef<int>(Mask).slice(N: S.Index, M: S.Length))) |
| 1169 | E = I + NumElems; |
| 1170 | return IRB.CreateShuffleVector(V1: Whole, V2: ExtPart, Mask, |
| 1171 | Name: Name + ".parts."+ Twine(S.Index)); |
| 1172 | } |
| 1173 | |
| 1174 | bool LegalizeBufferContentTypesVisitor::visitLoadImpl( |
| 1175 | LoadInst &OrigLI, Type *PartType, SmallVectorImpl<uint32_t> &AggIdxs, |
| 1176 | uint64_t AggByteOff, Value *&Result, const Twine &Name) { |
| 1177 | if (auto *ST = dyn_cast<StructType>(Val: PartType)) { |
| 1178 | const StructLayout *Layout = DL.getStructLayout(Ty: ST); |
| 1179 | bool Changed = false; |
| 1180 | for (auto [I, ElemTy, Offset] : |
| 1181 | llvm::enumerate(First: ST->elements(), Rest: Layout->getMemberOffsets())) { |
| 1182 | AggIdxs.push_back(Elt: I); |
| 1183 | Changed |= visitLoadImpl(OrigLI, PartType: ElemTy, AggIdxs, |
| 1184 | AggByteOff: AggByteOff + Offset.getFixedValue(), Result, |
| 1185 | Name: Name + "."+ Twine(I)); |
| 1186 | AggIdxs.pop_back(); |
| 1187 | } |
| 1188 | return Changed; |
| 1189 | } |
| 1190 | if (auto *AT = dyn_cast<ArrayType>(Val: PartType)) { |
| 1191 | Type *ElemTy = AT->getElementType(); |
| 1192 | if (!ElemTy->isSingleValueType() || !DL.typeSizeEqualsStoreSize(Ty: ElemTy) || |
| 1193 | ElemTy->isVectorTy()) { |
| 1194 | TypeSize ElemAllocSize = DL.getTypeAllocSize(Ty: ElemTy); |
| 1195 | bool Changed = false; |
| 1196 | for (auto I : llvm::iota_range<uint32_t>(0, AT->getNumElements(), |
| 1197 | /*Inclusive=*/false)) { |
| 1198 | AggIdxs.push_back(Elt: I); |
| 1199 | Changed |= visitLoadImpl(OrigLI, PartType: ElemTy, AggIdxs, |
| 1200 | AggByteOff: AggByteOff + I * ElemAllocSize.getFixedValue(), |
| 1201 | Result, Name: Name + Twine(I)); |
| 1202 | AggIdxs.pop_back(); |
| 1203 | } |
| 1204 | return Changed; |
| 1205 | } |
| 1206 | } |
| 1207 | |
| 1208 | // Typical case |
| 1209 | |
| 1210 | Align PartAlign = commonAlignment(A: OrigLI.getAlign(), Offset: AggByteOff); |
| 1211 | Type *ArrayAsVecType = scalarArrayTypeAsVector(T: PartType); |
| 1212 | OobProperties OobProps = |
| 1213 | analyzeOobProperties(Ptr: OrigLI.getPointerOperand(), Ty: PartType, ByteOffset: AggByteOff); |
| 1214 | uint64_t MaxWidth = maxIntrinsicWidth(T: ArrayAsVecType, A: PartAlign, OobProps); |
| 1215 | Type *LegalType = legalNonAggregateForMemOp(T: ArrayAsVecType, MaxWidth); |
| 1216 | |
| 1217 | SmallVector<VecSlice> Slices; |
| 1218 | getVecSlices(T: LegalType, MaxWidth, Slices); |
| 1219 | bool HasSlices = Slices.size() > 1; |
| 1220 | bool IsAggPart = !AggIdxs.empty(); |
| 1221 | Value *LoadsRes; |
| 1222 | if (!HasSlices && !IsAggPart) { |
| 1223 | Type *LoadableType = intrinsicTypeFor(LegalType); |
| 1224 | if (LoadableType == PartType) |
| 1225 | return false; |
| 1226 | |
| 1227 | IRB.SetInsertPoint(&OrigLI); |
| 1228 | auto *NLI = cast<LoadInst>(Val: OrigLI.clone()); |
| 1229 | NLI->mutateType(Ty: LoadableType); |
| 1230 | NLI = IRB.Insert(I: NLI); |
| 1231 | NLI->setName(Name + ".loadable"); |
| 1232 | |
| 1233 | LoadsRes = IRB.CreateBitCast(V: NLI, DestTy: LegalType, Name: Name + ".from.loadable"); |
| 1234 | } else { |
| 1235 | IRB.SetInsertPoint(&OrigLI); |
| 1236 | LoadsRes = PoisonValue::get(T: LegalType); |
| 1237 | Value *OrigPtr = OrigLI.getPointerOperand(); |
| 1238 | // If we're needing to spill something into more than one load, its legal |
| 1239 | // type will be a vector (ex. an i256 load will have LegalType = <8 x i32>). |
| 1240 | // But if we're already a scalar (which can happen if we're splitting up a |
| 1241 | // struct), the element type will be the legal type itself. |
| 1242 | Type *ElemType = LegalType->getScalarType(); |
| 1243 | unsigned ElemBytes = DL.getTypeStoreSize(Ty: ElemType); |
| 1244 | AAMDNodes AANodes = OrigLI.getAAMetadata(); |
| 1245 | if (IsAggPart && Slices.empty()) |
| 1246 | Slices.push_back(Elt: VecSlice{/*Index=*/0, /*Length=*/1}); |
| 1247 | for (VecSlice S : Slices) { |
| 1248 | Type *SliceType = |
| 1249 | S.Length != 1 ? FixedVectorType::get(ElementType: ElemType, NumElts: S.Length) : ElemType; |
| 1250 | int64_t ByteOffset = AggByteOff + S.Index * ElemBytes; |
| 1251 | // You can't reasonably expect loads to wrap around the edge of memory. |
| 1252 | Value *NewPtr = IRB.CreateGEP( |
| 1253 | Ty: IRB.getInt8Ty(), Ptr: OrigLI.getPointerOperand(), IdxList: IRB.getInt32(C: ByteOffset), |
| 1254 | Name: OrigPtr->getName() + ".off.ptr."+ Twine(ByteOffset), |
| 1255 | NW: ST->hasRelaxedBufferOOBMode() ? GEPNoWrapFlags::noUnsignedWrap() |
| 1256 | : GEPNoWrapFlags::none()); |
| 1257 | Type *LoadableType = intrinsicTypeFor(LegalType: SliceType); |
| 1258 | LoadInst *NewLI = IRB.CreateAlignedLoad( |
| 1259 | Ty: LoadableType, Ptr: NewPtr, Align: commonAlignment(A: OrigLI.getAlign(), Offset: ByteOffset), |
| 1260 | Name: Name + ".off."+ Twine(ByteOffset)); |
| 1261 | copyMetadataForLoad(Dest&: *NewLI, Source: OrigLI); |
| 1262 | NewLI->setAAMetadata( |
| 1263 | AANodes.adjustForAccess(Offset: ByteOffset, AccessTy: LoadableType, DL)); |
| 1264 | NewLI->setAtomic(Ordering: OrigLI.getOrdering(), SSID: OrigLI.getSyncScopeID()); |
| 1265 | NewLI->setVolatile(OrigLI.isVolatile()); |
| 1266 | Value *Loaded = IRB.CreateBitCast(V: NewLI, DestTy: SliceType, |
| 1267 | Name: NewLI->getName() + ".from.loadable"); |
| 1268 | LoadsRes = insertSlice(Whole: LoadsRes, Part: Loaded, S, Name); |
| 1269 | } |
| 1270 | } |
| 1271 | if (LegalType != ArrayAsVecType) |
| 1272 | LoadsRes = makeIllegalNonAggregate(V: LoadsRes, OrigType: ArrayAsVecType, Name); |
| 1273 | if (ArrayAsVecType != PartType) |
| 1274 | LoadsRes = vectorToArray(V: LoadsRes, OrigType: PartType, Name); |
| 1275 | |
| 1276 | if (IsAggPart) |
| 1277 | Result = IRB.CreateInsertValue(Agg: Result, Val: LoadsRes, Idxs: AggIdxs, Name); |
| 1278 | else |
| 1279 | Result = LoadsRes; |
| 1280 | return true; |
| 1281 | } |
| 1282 | |
| 1283 | bool LegalizeBufferContentTypesVisitor::visitLoadInst(LoadInst &LI) { |
| 1284 | if (LI.getPointerAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 1285 | return false; |
| 1286 | |
| 1287 | SmallVector<uint32_t> AggIdxs; |
| 1288 | Type *OrigType = LI.getType(); |
| 1289 | Value *Result = PoisonValue::get(T: OrigType); |
| 1290 | bool Changed = visitLoadImpl(OrigLI&: LI, PartType: OrigType, AggIdxs, AggByteOff: 0, Result, Name: LI.getName()); |
| 1291 | if (!Changed) |
| 1292 | return false; |
| 1293 | Result->takeName(V: &LI); |
| 1294 | LI.replaceAllUsesWith(V: Result); |
| 1295 | LI.eraseFromParent(); |
| 1296 | return Changed; |
| 1297 | } |
| 1298 | |
| 1299 | std::pair<bool, bool> LegalizeBufferContentTypesVisitor::visitStoreImpl( |
| 1300 | StoreInst &OrigSI, Type *PartType, SmallVectorImpl<uint32_t> &AggIdxs, |
| 1301 | uint64_t AggByteOff, const Twine &Name) { |
| 1302 | if (auto *ST = dyn_cast<StructType>(Val: PartType)) { |
| 1303 | const StructLayout *Layout = DL.getStructLayout(Ty: ST); |
| 1304 | bool Changed = false; |
| 1305 | for (auto [I, ElemTy, Offset] : |
| 1306 | llvm::enumerate(First: ST->elements(), Rest: Layout->getMemberOffsets())) { |
| 1307 | AggIdxs.push_back(Elt: I); |
| 1308 | Changed |= std::get<0>(in: visitStoreImpl(OrigSI, PartType: ElemTy, AggIdxs, |
| 1309 | AggByteOff: AggByteOff + Offset.getFixedValue(), |
| 1310 | Name: Name + "."+ Twine(I))); |
| 1311 | AggIdxs.pop_back(); |
| 1312 | } |
| 1313 | return std::make_pair(x&: Changed, /*ModifiedInPlace=*/y: false); |
| 1314 | } |
| 1315 | if (auto *AT = dyn_cast<ArrayType>(Val: PartType)) { |
| 1316 | Type *ElemTy = AT->getElementType(); |
| 1317 | if (!ElemTy->isSingleValueType() || !DL.typeSizeEqualsStoreSize(Ty: ElemTy) || |
| 1318 | ElemTy->isVectorTy()) { |
| 1319 | TypeSize ElemAllocSize = DL.getTypeAllocSize(Ty: ElemTy); |
| 1320 | bool Changed = false; |
| 1321 | for (auto I : llvm::iota_range<uint32_t>(0, AT->getNumElements(), |
| 1322 | /*Inclusive=*/false)) { |
| 1323 | AggIdxs.push_back(Elt: I); |
| 1324 | Changed |= std::get<0>(in: visitStoreImpl( |
| 1325 | OrigSI, PartType: ElemTy, AggIdxs, |
| 1326 | AggByteOff: AggByteOff + I * ElemAllocSize.getFixedValue(), Name: Name + Twine(I))); |
| 1327 | AggIdxs.pop_back(); |
| 1328 | } |
| 1329 | return std::make_pair(x&: Changed, /*ModifiedInPlace=*/y: false); |
| 1330 | } |
| 1331 | } |
| 1332 | |
| 1333 | Value *OrigData = OrigSI.getValueOperand(); |
| 1334 | Value *NewData = OrigData; |
| 1335 | |
| 1336 | bool IsAggPart = !AggIdxs.empty(); |
| 1337 | if (IsAggPart) |
| 1338 | NewData = IRB.CreateExtractValue(Agg: NewData, Idxs: AggIdxs, Name); |
| 1339 | |
| 1340 | Type *ArrayAsVecType = scalarArrayTypeAsVector(T: PartType); |
| 1341 | if (ArrayAsVecType != PartType) { |
| 1342 | NewData = arrayToVector(V: NewData, TargetType: ArrayAsVecType, Name); |
| 1343 | } |
| 1344 | |
| 1345 | Align PartAlign = commonAlignment(A: OrigSI.getAlign(), Offset: AggByteOff); |
| 1346 | OobProperties OobProps = |
| 1347 | analyzeOobProperties(Ptr: OrigSI.getPointerOperand(), Ty: PartType, ByteOffset: AggByteOff); |
| 1348 | uint64_t MaxWidth = maxIntrinsicWidth(T: ArrayAsVecType, A: PartAlign, OobProps); |
| 1349 | Type *LegalType = legalNonAggregateForMemOp(T: ArrayAsVecType, MaxWidth); |
| 1350 | if (LegalType != ArrayAsVecType) { |
| 1351 | NewData = makeLegalNonAggregate(V: NewData, TargetType: LegalType, Name); |
| 1352 | } |
| 1353 | |
| 1354 | SmallVector<VecSlice> Slices; |
| 1355 | getVecSlices(T: LegalType, MaxWidth, Slices); |
| 1356 | bool NeedToSplit = Slices.size() > 1 || IsAggPart; |
| 1357 | if (!NeedToSplit) { |
| 1358 | Type *StorableType = intrinsicTypeFor(LegalType); |
| 1359 | if (StorableType == PartType) |
| 1360 | return std::make_pair(/*Changed=*/x: false, /*ModifiedInPlace=*/y: false); |
| 1361 | NewData = IRB.CreateBitCast(V: NewData, DestTy: StorableType, Name: Name + ".storable"); |
| 1362 | OrigSI.setOperand(i_nocapture: 0, Val_nocapture: NewData); |
| 1363 | return std::make_pair(/*Changed=*/x: true, /*ModifiedInPlace=*/y: true); |
| 1364 | } |
| 1365 | |
| 1366 | Value *OrigPtr = OrigSI.getPointerOperand(); |
| 1367 | Type *ElemType = LegalType->getScalarType(); |
| 1368 | if (IsAggPart && Slices.empty()) |
| 1369 | Slices.push_back(Elt: VecSlice{/*Index=*/0, /*Length=*/1}); |
| 1370 | unsigned ElemBytes = DL.getTypeStoreSize(Ty: ElemType); |
| 1371 | AAMDNodes AANodes = OrigSI.getAAMetadata(); |
| 1372 | for (VecSlice S : Slices) { |
| 1373 | Type *SliceType = |
| 1374 | S.Length != 1 ? FixedVectorType::get(ElementType: ElemType, NumElts: S.Length) : ElemType; |
| 1375 | int64_t ByteOffset = AggByteOff + S.Index * ElemBytes; |
| 1376 | Value *NewPtr = IRB.CreateGEP( |
| 1377 | Ty: IRB.getInt8Ty(), Ptr: OrigPtr, IdxList: IRB.getInt32(C: ByteOffset), |
| 1378 | Name: OrigPtr->getName() + ".part."+ Twine(S.Index), |
| 1379 | NW: ST->hasRelaxedBufferOOBMode() ? GEPNoWrapFlags::noUnsignedWrap() |
| 1380 | : GEPNoWrapFlags::none()); |
| 1381 | Value *DataSlice = extractSlice(Vec: NewData, S, Name); |
| 1382 | Type *StorableType = intrinsicTypeFor(LegalType: SliceType); |
| 1383 | DataSlice = IRB.CreateBitCast(V: DataSlice, DestTy: StorableType, |
| 1384 | Name: DataSlice->getName() + ".storable"); |
| 1385 | auto *NewSI = cast<StoreInst>(Val: OrigSI.clone()); |
| 1386 | NewSI->setAlignment(commonAlignment(A: OrigSI.getAlign(), Offset: ByteOffset)); |
| 1387 | IRB.Insert(I: NewSI); |
| 1388 | NewSI->setOperand(i_nocapture: 0, Val_nocapture: DataSlice); |
| 1389 | NewSI->setOperand(i_nocapture: 1, Val_nocapture: NewPtr); |
| 1390 | NewSI->setAAMetadata(AANodes.adjustForAccess(Offset: ByteOffset, AccessTy: StorableType, DL)); |
| 1391 | } |
| 1392 | return std::make_pair(/*Changed=*/x: true, /*ModifiedInPlace=*/y: false); |
| 1393 | } |
| 1394 | |
| 1395 | bool LegalizeBufferContentTypesVisitor::visitStoreInst(StoreInst &SI) { |
| 1396 | if (SI.getPointerAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 1397 | return false; |
| 1398 | IRB.SetInsertPoint(&SI); |
| 1399 | SmallVector<uint32_t> AggIdxs; |
| 1400 | Value *OrigData = SI.getValueOperand(); |
| 1401 | auto [Changed, ModifiedInPlace] = |
| 1402 | visitStoreImpl(OrigSI&: SI, PartType: OrigData->getType(), AggIdxs, AggByteOff: 0, Name: OrigData->getName()); |
| 1403 | if (Changed && !ModifiedInPlace) |
| 1404 | SI.eraseFromParent(); |
| 1405 | return Changed; |
| 1406 | } |
| 1407 | |
| 1408 | bool LegalizeBufferContentTypesVisitor::visitAddrSpaceCastInst( |
| 1409 | AddrSpaceCastInst &AI) { |
| 1410 | if (AI.getSrcAddressSpace() != AMDGPUAS::BUFFER_RESOURCE || |
| 1411 | AI.getDestAddressSpace() != AMDGPUAS::BUFFER_FAT_POINTER) |
| 1412 | return false; |
| 1413 | Value *Src = AI.getPointerOperand(); |
| 1414 | auto Record = ZeroBasePointerToNumRecords.find(Val: Src); |
| 1415 | if (Record != ZeroBasePointerToNumRecords.end()) |
| 1416 | ZeroBasePointerToNumRecords.insert(KV: {&AI, Record->second}); |
| 1417 | else |
| 1418 | ZeroBasePointerToNumRecords.insert(KV: {&AI, nullptr}); |
| 1419 | return false; |
| 1420 | } |
| 1421 | |
| 1422 | bool LegalizeBufferContentTypesVisitor::visitIntrinsicInst(IntrinsicInst &II) { |
| 1423 | if (II.getIntrinsicID() != Intrinsic::amdgcn_make_buffer_rsrc) |
| 1424 | return false; |
| 1425 | ZeroBasePointerToNumRecords.insert(KV: {&II, II.getOperand(i_nocapture: 2)}); |
| 1426 | return false; |
| 1427 | } |
| 1428 | |
| 1429 | bool LegalizeBufferContentTypesVisitor::processFunction(Function &F, |
| 1430 | ScalarEvolution *SE) { |
| 1431 | this->SE = SE; |
| 1432 | ST = &TM->getSubtarget<GCNSubtarget>(F); |
| 1433 | bool Changed = false; |
| 1434 | for (Instruction &I : make_early_inc_range(Range: instructions(F))) { |
| 1435 | Changed |= visit(I); |
| 1436 | } |
| 1437 | ZeroBasePointerToNumRecords.clear(); |
| 1438 | this->SE = nullptr; |
| 1439 | return Changed; |
| 1440 | } |
| 1441 | |
| 1442 | /// Return the ptr addrspace(8) and i32 (resource and offset parts) in a lowered |
| 1443 | /// buffer fat pointer constant. |
| 1444 | static std::pair<Constant *, Constant *> |
| 1445 | splitLoweredFatBufferConst(Constant *C) { |
| 1446 | assert(isSplitFatPtr(C->getType()) && "Not a split fat buffer pointer"); |
| 1447 | return std::make_pair(x: C->getAggregateElement(Elt: 0u), y: C->getAggregateElement(Elt: 1u)); |
| 1448 | } |
| 1449 | |
| 1450 | namespace { |
| 1451 | /// Handle the remapping of ptr addrspace(7) constants. |
| 1452 | class FatPtrConstMaterializer final : public ValueMaterializer { |
| 1453 | BufferFatPtrToStructTypeMap *TypeMap; |
| 1454 | // An internal mapper that is used to recurse into the arguments of constants. |
| 1455 | // While the documentation for `ValueMapper` specifies not to use it |
| 1456 | // recursively, examination of the logic in mapValue() shows that it can |
| 1457 | // safely be used recursively when handling constants, like it does in its own |
| 1458 | // logic. |
| 1459 | ValueMapper InternalMapper; |
| 1460 | |
| 1461 | Constant *materializeBufferFatPtrConst(Constant *C); |
| 1462 | |
| 1463 | public: |
| 1464 | // UnderlyingMap is the value map this materializer will be filling. |
| 1465 | FatPtrConstMaterializer(BufferFatPtrToStructTypeMap *TypeMap, |
| 1466 | ValueToValueMapTy &UnderlyingMap) |
| 1467 | : TypeMap(TypeMap), |
| 1468 | InternalMapper(UnderlyingMap, RF_None, TypeMap, this) {} |
| 1469 | ~FatPtrConstMaterializer() = default; |
| 1470 | |
| 1471 | Value *materialize(Value *V) override; |
| 1472 | }; |
| 1473 | } // namespace |
| 1474 | |
| 1475 | Constant *FatPtrConstMaterializer::materializeBufferFatPtrConst(Constant *C) { |
| 1476 | Type *SrcTy = C->getType(); |
| 1477 | auto *NewTy = dyn_cast<StructType>(Val: TypeMap->remapType(SrcTy)); |
| 1478 | if (C->isNullValue()) |
| 1479 | return ConstantAggregateZero::getNullValue(Ty: NewTy); |
| 1480 | if (isa<PoisonValue>(Val: C)) { |
| 1481 | return ConstantStruct::get(T: NewTy, |
| 1482 | V: {PoisonValue::get(T: NewTy->getElementType(N: 0)), |
| 1483 | PoisonValue::get(T: NewTy->getElementType(N: 1))}); |
| 1484 | } |
| 1485 | if (isa<UndefValue>(Val: C)) { |
| 1486 | return ConstantStruct::get(T: NewTy, |
| 1487 | V: {UndefValue::get(T: NewTy->getElementType(N: 0)), |
| 1488 | UndefValue::get(T: NewTy->getElementType(N: 1))}); |
| 1489 | } |
| 1490 | |
| 1491 | if (auto *VC = dyn_cast<ConstantVector>(Val: C)) { |
| 1492 | if (Constant *S = VC->getSplatValue()) { |
| 1493 | Constant *NewS = InternalMapper.mapConstant(C: *S); |
| 1494 | if (!NewS) |
| 1495 | return nullptr; |
| 1496 | auto [Rsrc, Off] = splitLoweredFatBufferConst(C: NewS); |
| 1497 | auto EC = VC->getType()->getElementCount(); |
| 1498 | return ConstantStruct::get(T: NewTy, V: {ConstantVector::getSplat(EC, Elt: Rsrc), |
| 1499 | ConstantVector::getSplat(EC, Elt: Off)}); |
| 1500 | } |
| 1501 | SmallVector<Constant *> Rsrcs; |
| 1502 | SmallVector<Constant *> Offs; |
| 1503 | for (Value *Op : VC->operand_values()) { |
| 1504 | auto *NewOp = dyn_cast_or_null<Constant>(Val: InternalMapper.mapValue(V: *Op)); |
| 1505 | if (!NewOp) |
| 1506 | return nullptr; |
| 1507 | auto [Rsrc, Off] = splitLoweredFatBufferConst(C: NewOp); |
| 1508 | Rsrcs.push_back(Elt: Rsrc); |
| 1509 | Offs.push_back(Elt: Off); |
| 1510 | } |
| 1511 | Constant *RsrcVec = ConstantVector::get(V: Rsrcs); |
| 1512 | Constant *OffVec = ConstantVector::get(V: Offs); |
| 1513 | return ConstantStruct::get(T: NewTy, V: {RsrcVec, OffVec}); |
| 1514 | } |
| 1515 | |
| 1516 | if (isa<GlobalValue>(Val: C)) |
| 1517 | reportFatalUsageError(reason: "global values containing ptr addrspace(7) (buffer " |
| 1518 | "fat pointer) values are not supported"); |
| 1519 | |
| 1520 | if (isa<ConstantExpr>(Val: C)) |
| 1521 | reportFatalUsageError( |
| 1522 | reason: "constant exprs containing ptr addrspace(7) (buffer " |
| 1523 | "fat pointer) values should have been expanded earlier"); |
| 1524 | |
| 1525 | return nullptr; |
| 1526 | } |
| 1527 | |
| 1528 | Value *FatPtrConstMaterializer::materialize(Value *V) { |
| 1529 | Constant *C = dyn_cast<Constant>(Val: V); |
| 1530 | if (!C) |
| 1531 | return nullptr; |
| 1532 | // Structs and other types that happen to contain fat pointers get remapped |
| 1533 | // by the mapValue() logic. |
| 1534 | if (!isBufferFatPtrConst(C)) |
| 1535 | return nullptr; |
| 1536 | return materializeBufferFatPtrConst(C); |
| 1537 | } |
| 1538 | |
| 1539 | using PtrParts = std::pair<Value *, Value *>; |
| 1540 | namespace { |
| 1541 | // The visitor returns the resource and offset parts for an instruction if they |
| 1542 | // can be computed, or (nullptr, nullptr) for cases that don't have a meaningful |
| 1543 | // value mapping. |
| 1544 | class SplitPtrStructs : public InstVisitor<SplitPtrStructs, PtrParts> { |
| 1545 | ValueToValueMapTy RsrcParts; |
| 1546 | ValueToValueMapTy OffParts; |
| 1547 | |
| 1548 | // Track instructions that have been rewritten into a user of the component |
| 1549 | // parts of their ptr addrspace(7) input. Instructions that produced |
| 1550 | // ptr addrspace(7) parts should **not** be RAUW'd before being added to this |
| 1551 | // set, as that replacement will be handled in a post-visit step. However, |
| 1552 | // instructions that yield values that aren't fat pointers (ex. ptrtoint) |
| 1553 | // should RAUW themselves with new instructions that use the split parts |
| 1554 | // of their arguments during processing. |
| 1555 | DenseSet<Instruction *> SplitUsers; |
| 1556 | |
| 1557 | // Nodes that need a second look once we've computed the parts for all other |
| 1558 | // instructions to see if, for example, we really need to phi on the resource |
| 1559 | // part. |
| 1560 | SmallVector<Instruction *> Conditionals; |
| 1561 | // Temporary instructions produced while lowering conditionals that should be |
| 1562 | // killed. |
| 1563 | SmallVector<Instruction *> ConditionalTemps; |
| 1564 | |
| 1565 | // Subtarget info, needed for determining what cache control bits to set. |
| 1566 | const TargetMachine *TM; |
| 1567 | const GCNSubtarget *ST = nullptr; |
| 1568 | |
| 1569 | IRBuilder<InstSimplifyFolder> IRB; |
| 1570 | |
| 1571 | // Copy metadata between instructions if applicable. |
| 1572 | void copyMetadata(Value *Dest, Value *Src); |
| 1573 | |
| 1574 | // Get the resource and offset parts of the value V, inserting appropriate |
| 1575 | // extractvalue calls if needed. |
| 1576 | PtrParts getPtrParts(Value *V); |
| 1577 | |
| 1578 | // Given an instruction that could produce multiple resource parts (a PHI or |
| 1579 | // select), collect the set of possible instructions that could have provided |
| 1580 | // its resource parts that it could have (the `Roots`) and the set of |
| 1581 | // conditional instructions visited during the search (`Seen`). If, after |
| 1582 | // removing the root of the search from `Seen` and `Roots`, `Seen` is a subset |
| 1583 | // of `Roots` and `Roots - Seen` contains one element, the resource part of |
| 1584 | // that element can replace the resource part of all other elements in `Seen`. |
| 1585 | void getPossibleRsrcRoots(Instruction *I, SmallPtrSetImpl<Value *> &Roots, |
| 1586 | SmallPtrSetImpl<Value *> &Seen); |
| 1587 | void processConditionals(); |
| 1588 | |
| 1589 | // If an instruction hav been split into resource and offset parts, |
| 1590 | // delete that instruction. If any of its uses have not themselves been split |
| 1591 | // into parts (for example, an insertvalue), construct the structure |
| 1592 | // that the type rewrites declared should be produced by the dying instruction |
| 1593 | // and use that. |
| 1594 | // Also, kill the temporary extractvalue operations produced by the two-stage |
| 1595 | // lowering of PHIs and conditionals. |
| 1596 | void killAndReplaceSplitInstructions(SmallVectorImpl<Instruction *> &Origs); |
| 1597 | |
| 1598 | void setAlign(CallInst *Intr, Align A, unsigned RsrcArgIdx); |
| 1599 | void insertPreMemOpFence(AtomicOrdering Order, SyncScope::ID SSID); |
| 1600 | void insertPostMemOpFence(AtomicOrdering Order, SyncScope::ID SSID); |
| 1601 | Value *handleMemoryInst(Instruction *I, Value *Arg, Value *Ptr, Type *Ty, |
| 1602 | Align Alignment, AtomicOrdering Order, |
| 1603 | bool IsVolatile, SyncScope::ID SSID); |
| 1604 | |
| 1605 | public: |
| 1606 | SplitPtrStructs(Module &M, const TargetMachine *TM) |
| 1607 | : TM(TM), IRB(M, InstSimplifyFolder(M.getDataLayout())) {} |
| 1608 | |
| 1609 | void processFunction(Function &F); |
| 1610 | |
| 1611 | PtrParts visitInstruction(Instruction &I); |
| 1612 | PtrParts visitLoadInst(LoadInst &LI); |
| 1613 | PtrParts visitStoreInst(StoreInst &SI); |
| 1614 | PtrParts visitAtomicRMWInst(AtomicRMWInst &AI); |
| 1615 | PtrParts visitAtomicCmpXchgInst(AtomicCmpXchgInst &AI); |
| 1616 | PtrParts visitGetElementPtrInst(GetElementPtrInst &GEP); |
| 1617 | |
| 1618 | PtrParts visitPtrToAddrInst(PtrToAddrInst &PA); |
| 1619 | PtrParts visitPtrToIntInst(PtrToIntInst &PI); |
| 1620 | PtrParts visitIntToPtrInst(IntToPtrInst &IP); |
| 1621 | PtrParts visitAddrSpaceCastInst(AddrSpaceCastInst &I); |
| 1622 | PtrParts visitICmpInst(ICmpInst &Cmp); |
| 1623 | PtrParts visitFreezeInst(FreezeInst &I); |
| 1624 | |
| 1625 | PtrParts visitExtractElementInst(ExtractElementInst &I); |
| 1626 | PtrParts visitInsertElementInst(InsertElementInst &I); |
| 1627 | PtrParts visitShuffleVectorInst(ShuffleVectorInst &I); |
| 1628 | |
| 1629 | PtrParts visitPHINode(PHINode &PHI); |
| 1630 | PtrParts visitSelectInst(SelectInst &SI); |
| 1631 | |
| 1632 | PtrParts visitIntrinsicInst(IntrinsicInst &II); |
| 1633 | }; |
| 1634 | } // namespace |
| 1635 | |
| 1636 | void SplitPtrStructs::copyMetadata(Value *Dest, Value *Src) { |
| 1637 | auto *DestI = dyn_cast<Instruction>(Val: Dest); |
| 1638 | auto *SrcI = dyn_cast<Instruction>(Val: Src); |
| 1639 | |
| 1640 | if (!DestI || !SrcI) |
| 1641 | return; |
| 1642 | |
| 1643 | DestI->copyMetadata(SrcInst: *SrcI); |
| 1644 | } |
| 1645 | |
| 1646 | PtrParts SplitPtrStructs::getPtrParts(Value *V) { |
| 1647 | assert(isSplitFatPtr(V->getType()) && "it's not meaningful to get the parts " |
| 1648 | "of something that wasn't rewritten"); |
| 1649 | auto *RsrcEntry = &RsrcParts[V]; |
| 1650 | auto *OffEntry = &OffParts[V]; |
| 1651 | if (*RsrcEntry && *OffEntry) |
| 1652 | return {*RsrcEntry, *OffEntry}; |
| 1653 | |
| 1654 | if (auto *C = dyn_cast<Constant>(Val: V)) { |
| 1655 | auto [Rsrc, Off] = splitLoweredFatBufferConst(C); |
| 1656 | return {*RsrcEntry = Rsrc, *OffEntry = Off}; |
| 1657 | } |
| 1658 | |
| 1659 | IRBuilder<InstSimplifyFolder>::InsertPointGuard Guard(IRB); |
| 1660 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 1661 | LLVM_DEBUG(dbgs() << "Recursing to split parts of "<< *I << "\n"); |
| 1662 | auto [Rsrc, Off] = visit(I&: *I); |
| 1663 | if (Rsrc && Off) |
| 1664 | return {*RsrcEntry = Rsrc, *OffEntry = Off}; |
| 1665 | // We'll be creating the new values after the relevant instruction. |
| 1666 | // This instruction generates a value and so isn't a terminator. |
| 1667 | IRB.SetInsertPoint(*I->getInsertionPointAfterDef()); |
| 1668 | IRB.SetCurrentDebugLocation(I->getDebugLoc()); |
| 1669 | } else if (auto *A = dyn_cast<Argument>(Val: V)) { |
| 1670 | IRB.SetInsertPointPastAllocas(A->getParent()); |
| 1671 | IRB.SetCurrentDebugLocation(DebugLoc()); |
| 1672 | } |
| 1673 | Value *Rsrc = IRB.CreateExtractValue(Agg: V, Idxs: 0, Name: V->getName() + ".rsrc"); |
| 1674 | Value *Off = IRB.CreateExtractValue(Agg: V, Idxs: 1, Name: V->getName() + ".off"); |
| 1675 | return {*RsrcEntry = Rsrc, *OffEntry = Off}; |
| 1676 | } |
| 1677 | |
| 1678 | /// Returns the instruction that defines the resource part of the value V. |
| 1679 | /// Note that this is not getUnderlyingObject(), since that looks through |
| 1680 | /// operations like ptrmask which might modify the resource part. |
| 1681 | /// |
| 1682 | /// We can limit ourselves to just looking through GEPs followed by looking |
| 1683 | /// through addrspacecasts because only those two operations preserve the |
| 1684 | /// resource part, and because operations on an `addrspace(8)` (which is the |
| 1685 | /// legal input to this addrspacecast) would produce a different resource part. |
| 1686 | static Value *rsrcPartRoot(Value *V) { |
| 1687 | while (auto *GEP = dyn_cast<GEPOperator>(Val: V)) |
| 1688 | V = GEP->getPointerOperand(); |
| 1689 | while (auto *ASC = dyn_cast<AddrSpaceCastOperator>(Val: V)) |
| 1690 | V = ASC->getPointerOperand(); |
| 1691 | return V; |
| 1692 | } |
| 1693 | |
| 1694 | void SplitPtrStructs::getPossibleRsrcRoots(Instruction *I, |
| 1695 | SmallPtrSetImpl<Value *> &Roots, |
| 1696 | SmallPtrSetImpl<Value *> &Seen) { |
| 1697 | if (auto *PHI = dyn_cast<PHINode>(Val: I)) { |
| 1698 | if (!Seen.insert(Ptr: I).second) |
| 1699 | return; |
| 1700 | for (Value *In : PHI->incoming_values()) { |
| 1701 | In = rsrcPartRoot(V: In); |
| 1702 | Roots.insert(Ptr: In); |
| 1703 | if (isa<PHINode, SelectInst>(Val: In)) |
| 1704 | getPossibleRsrcRoots(I: cast<Instruction>(Val: In), Roots, Seen); |
| 1705 | } |
| 1706 | } else if (auto *SI = dyn_cast<SelectInst>(Val: I)) { |
| 1707 | if (!Seen.insert(Ptr: SI).second) |
| 1708 | return; |
| 1709 | Value *TrueVal = rsrcPartRoot(V: SI->getTrueValue()); |
| 1710 | Value *FalseVal = rsrcPartRoot(V: SI->getFalseValue()); |
| 1711 | Roots.insert(Ptr: TrueVal); |
| 1712 | Roots.insert(Ptr: FalseVal); |
| 1713 | if (isa<PHINode, SelectInst>(Val: TrueVal)) |
| 1714 | getPossibleRsrcRoots(I: cast<Instruction>(Val: TrueVal), Roots, Seen); |
| 1715 | if (isa<PHINode, SelectInst>(Val: FalseVal)) |
| 1716 | getPossibleRsrcRoots(I: cast<Instruction>(Val: FalseVal), Roots, Seen); |
| 1717 | } else { |
| 1718 | llvm_unreachable("getPossibleRsrcParts() only works on phi and select"); |
| 1719 | } |
| 1720 | } |
| 1721 | |
| 1722 | void SplitPtrStructs::processConditionals() { |
| 1723 | SmallDenseMap<Value *, Value *> FoundRsrcs; |
| 1724 | SmallPtrSet<Value *, 4> Roots; |
| 1725 | SmallPtrSet<Value *, 4> Seen; |
| 1726 | for (Instruction *I : Conditionals) { |
| 1727 | // These have to exist by now because we've visited these nodes. |
| 1728 | Value *Rsrc = RsrcParts[I]; |
| 1729 | Value *Off = OffParts[I]; |
| 1730 | assert(Rsrc && Off && "must have visited conditionals by now"); |
| 1731 | |
| 1732 | std::optional<Value *> MaybeRsrc; |
| 1733 | auto MaybeFoundRsrc = FoundRsrcs.find(Val: I); |
| 1734 | if (MaybeFoundRsrc != FoundRsrcs.end()) { |
| 1735 | MaybeRsrc = MaybeFoundRsrc->second; |
| 1736 | } else { |
| 1737 | IRBuilder<InstSimplifyFolder>::InsertPointGuard Guard(IRB); |
| 1738 | Roots.clear(); |
| 1739 | Seen.clear(); |
| 1740 | getPossibleRsrcRoots(I, Roots, Seen); |
| 1741 | LLVM_DEBUG(dbgs() << "Processing conditional: "<< *I << "\n"); |
| 1742 | #ifndef NDEBUG |
| 1743 | for (Value *V : Roots) |
| 1744 | LLVM_DEBUG(dbgs() << "Root: "<< *V << "\n"); |
| 1745 | for (Value *V : Seen) |
| 1746 | LLVM_DEBUG(dbgs() << "Seen: "<< *V << "\n"); |
| 1747 | #endif |
| 1748 | // If we are our own possible root, then we shouldn't block our |
| 1749 | // replacement with a valid incoming value. |
| 1750 | Roots.erase(Ptr: I); |
| 1751 | // We don't want to block the optimization for conditionals that don't |
| 1752 | // refer to themselves but did see themselves during the traversal. |
| 1753 | Seen.erase(Ptr: I); |
| 1754 | |
| 1755 | if (set_is_subset(S1: Seen, S2: Roots)) { |
| 1756 | auto Diff = set_difference(S1: Roots, S2: Seen); |
| 1757 | if (Diff.size() == 1) { |
| 1758 | Value *RootVal = *Diff.begin(); |
| 1759 | // Handle the case where previous loops already looked through |
| 1760 | // an addrspacecast. |
| 1761 | if (isSplitFatPtr(Ty: RootVal->getType())) |
| 1762 | MaybeRsrc = std::get<0>(in: getPtrParts(V: RootVal)); |
| 1763 | else |
| 1764 | MaybeRsrc = RootVal; |
| 1765 | } |
| 1766 | } |
| 1767 | } |
| 1768 | |
| 1769 | if (auto *PHI = dyn_cast<PHINode>(Val: I)) { |
| 1770 | Value *NewRsrc; |
| 1771 | StructType *PHITy = cast<StructType>(Val: PHI->getType()); |
| 1772 | IRB.SetInsertPoint(*PHI->getInsertionPointAfterDef()); |
| 1773 | IRB.SetCurrentDebugLocation(PHI->getDebugLoc()); |
| 1774 | if (MaybeRsrc) { |
| 1775 | NewRsrc = *MaybeRsrc; |
| 1776 | } else { |
| 1777 | Type *RsrcTy = PHITy->getElementType(N: 0); |
| 1778 | auto *RsrcPHI = IRB.CreatePHI(Ty: RsrcTy, NumReservedValues: PHI->getNumIncomingValues()); |
| 1779 | RsrcPHI->takeName(V: Rsrc); |
| 1780 | for (auto [V, BB] : llvm::zip(t: PHI->incoming_values(), u: PHI->blocks())) { |
| 1781 | Value *VRsrc = std::get<0>(in: getPtrParts(V)); |
| 1782 | RsrcPHI->addIncoming(V: VRsrc, BB); |
| 1783 | } |
| 1784 | copyMetadata(Dest: RsrcPHI, Src: PHI); |
| 1785 | NewRsrc = RsrcPHI; |
| 1786 | } |
| 1787 | |
| 1788 | Type *OffTy = PHITy->getElementType(N: 1); |
| 1789 | auto *NewOff = IRB.CreatePHI(Ty: OffTy, NumReservedValues: PHI->getNumIncomingValues()); |
| 1790 | NewOff->takeName(V: Off); |
| 1791 | for (auto [V, BB] : llvm::zip(t: PHI->incoming_values(), u: PHI->blocks())) { |
| 1792 | assert(OffParts.count(V) && "An offset part had to be created by now"); |
| 1793 | Value *VOff = std::get<1>(in: getPtrParts(V)); |
| 1794 | NewOff->addIncoming(V: VOff, BB); |
| 1795 | } |
| 1796 | copyMetadata(Dest: NewOff, Src: PHI); |
| 1797 | |
| 1798 | // Note: We don't eraseFromParent() the temporaries because we don't want |
| 1799 | // to put the corrections maps in an inconstent state. That'll be handed |
| 1800 | // during the rest of the killing. Also, `ValueToValueMapTy` guarantees |
| 1801 | // that references in that map will be updated as well. |
| 1802 | // Note that if the temporary instruction got `InstSimplify`'d away, it |
| 1803 | // might be something like a block argument. |
| 1804 | if (auto *RsrcInst = dyn_cast<Instruction>(Val: Rsrc)) { |
| 1805 | ConditionalTemps.push_back(Elt: RsrcInst); |
| 1806 | RsrcInst->replaceAllUsesWith(V: NewRsrc); |
| 1807 | } |
| 1808 | if (auto *OffInst = dyn_cast<Instruction>(Val: Off)) { |
| 1809 | ConditionalTemps.push_back(Elt: OffInst); |
| 1810 | OffInst->replaceAllUsesWith(V: NewOff); |
| 1811 | } |
| 1812 | |
| 1813 | // Save on recomputing the cycle traversals in known-root cases. |
| 1814 | if (MaybeRsrc) |
| 1815 | for (Value *V : Seen) |
| 1816 | FoundRsrcs[V] = NewRsrc; |
| 1817 | } else if (isa<SelectInst>(Val: I)) { |
| 1818 | if (MaybeRsrc) { |
| 1819 | if (auto *RsrcInst = dyn_cast<Instruction>(Val: Rsrc)) { |
| 1820 | // Guard against conditionals that were already folded away. |
| 1821 | if (RsrcInst != *MaybeRsrc) { |
| 1822 | ConditionalTemps.push_back(Elt: RsrcInst); |
| 1823 | RsrcInst->replaceAllUsesWith(V: *MaybeRsrc); |
| 1824 | } |
| 1825 | } |
| 1826 | for (Value *V : Seen) |
| 1827 | FoundRsrcs[V] = *MaybeRsrc; |
| 1828 | } |
| 1829 | } else { |
| 1830 | llvm_unreachable("Only PHIs and selects go in the conditionals list"); |
| 1831 | } |
| 1832 | } |
| 1833 | } |
| 1834 | |
| 1835 | void SplitPtrStructs::killAndReplaceSplitInstructions( |
| 1836 | SmallVectorImpl<Instruction *> &Origs) { |
| 1837 | for (Instruction *I : ConditionalTemps) |
| 1838 | I->eraseFromParent(); |
| 1839 | |
| 1840 | for (Instruction *I : Origs) { |
| 1841 | if (!SplitUsers.contains(V: I)) |
| 1842 | continue; |
| 1843 | |
| 1844 | SmallVector<DbgVariableRecord *> Dbgs; |
| 1845 | findDbgValues(V: I, DbgVariableRecords&: Dbgs); |
| 1846 | for (DbgVariableRecord *Dbg : Dbgs) { |
| 1847 | auto &DL = I->getDataLayout(); |
| 1848 | assert(isSplitFatPtr(I->getType()) && |
| 1849 | "We should've RAUW'd away loads, stores, etc. at this point"); |
| 1850 | DbgVariableRecord *OffDbg = Dbg->clone(); |
| 1851 | auto [Rsrc, Off] = getPtrParts(V: I); |
| 1852 | |
| 1853 | int64_t RsrcSz = DL.getTypeSizeInBits(Ty: Rsrc->getType()); |
| 1854 | int64_t OffSz = DL.getTypeSizeInBits(Ty: Off->getType()); |
| 1855 | |
| 1856 | std::optional<DIExpression *> RsrcExpr = |
| 1857 | DIExpression::createFragmentExpression(Expr: Dbg->getExpression(), OffsetInBits: 0, |
| 1858 | SizeInBits: RsrcSz); |
| 1859 | std::optional<DIExpression *> OffExpr = |
| 1860 | DIExpression::createFragmentExpression(Expr: Dbg->getExpression(), OffsetInBits: RsrcSz, |
| 1861 | SizeInBits: OffSz); |
| 1862 | if (OffExpr) { |
| 1863 | OffDbg->setExpression(*OffExpr); |
| 1864 | OffDbg->replaceVariableLocationOp(OldValue: I, NewValue: Off); |
| 1865 | OffDbg->insertBefore(InsertBefore: Dbg); |
| 1866 | } else { |
| 1867 | OffDbg->eraseFromParent(); |
| 1868 | } |
| 1869 | if (RsrcExpr) { |
| 1870 | Dbg->setExpression(*RsrcExpr); |
| 1871 | Dbg->replaceVariableLocationOp(OldValue: I, NewValue: Rsrc); |
| 1872 | } else { |
| 1873 | Dbg->replaceVariableLocationOp(OldValue: I, NewValue: PoisonValue::get(T: I->getType())); |
| 1874 | } |
| 1875 | } |
| 1876 | |
| 1877 | Value *Poison = PoisonValue::get(T: I->getType()); |
| 1878 | I->replaceUsesWithIf(New: Poison, ShouldReplace: [&](const Use &U) -> bool { |
| 1879 | if (const auto *UI = dyn_cast<Instruction>(Val: U.getUser())) |
| 1880 | return SplitUsers.contains(V: UI); |
| 1881 | return false; |
| 1882 | }); |
| 1883 | |
| 1884 | if (I->use_empty()) { |
| 1885 | I->eraseFromParent(); |
| 1886 | continue; |
| 1887 | } |
| 1888 | IRB.SetInsertPoint(*I->getInsertionPointAfterDef()); |
| 1889 | IRB.SetCurrentDebugLocation(I->getDebugLoc()); |
| 1890 | auto [Rsrc, Off] = getPtrParts(V: I); |
| 1891 | Value *Struct = PoisonValue::get(T: I->getType()); |
| 1892 | Struct = IRB.CreateInsertValue(Agg: Struct, Val: Rsrc, Idxs: 0); |
| 1893 | Struct = IRB.CreateInsertValue(Agg: Struct, Val: Off, Idxs: 1); |
| 1894 | copyMetadata(Dest: Struct, Src: I); |
| 1895 | Struct->takeName(V: I); |
| 1896 | I->replaceAllUsesWith(V: Struct); |
| 1897 | I->eraseFromParent(); |
| 1898 | } |
| 1899 | } |
| 1900 | |
| 1901 | void SplitPtrStructs::setAlign(CallInst *Intr, Align A, unsigned RsrcArgIdx) { |
| 1902 | LLVMContext &Ctx = Intr->getContext(); |
| 1903 | Intr->addParamAttr(ArgNo: RsrcArgIdx, Attr: Attribute::getWithAlignment(Context&: Ctx, Alignment: A)); |
| 1904 | } |
| 1905 | |
| 1906 | void SplitPtrStructs::insertPreMemOpFence(AtomicOrdering Order, |
| 1907 | SyncScope::ID SSID) { |
| 1908 | switch (Order) { |
| 1909 | case AtomicOrdering::Release: |
| 1910 | case AtomicOrdering::AcquireRelease: |
| 1911 | case AtomicOrdering::SequentiallyConsistent: |
| 1912 | IRB.CreateFence(Ordering: AtomicOrdering::Release, SSID); |
| 1913 | break; |
| 1914 | default: |
| 1915 | break; |
| 1916 | } |
| 1917 | } |
| 1918 | |
| 1919 | void SplitPtrStructs::insertPostMemOpFence(AtomicOrdering Order, |
| 1920 | SyncScope::ID SSID) { |
| 1921 | switch (Order) { |
| 1922 | case AtomicOrdering::Acquire: |
| 1923 | case AtomicOrdering::AcquireRelease: |
| 1924 | case AtomicOrdering::SequentiallyConsistent: |
| 1925 | IRB.CreateFence(Ordering: AtomicOrdering::Acquire, SSID); |
| 1926 | break; |
| 1927 | default: |
| 1928 | break; |
| 1929 | } |
| 1930 | } |
| 1931 | |
| 1932 | Value *SplitPtrStructs::handleMemoryInst(Instruction *I, Value *Arg, Value *Ptr, |
| 1933 | Type *Ty, Align Alignment, |
| 1934 | AtomicOrdering Order, bool IsVolatile, |
| 1935 | SyncScope::ID SSID) { |
| 1936 | IRB.SetInsertPoint(I); |
| 1937 | |
| 1938 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 1939 | SmallVector<Value *, 5> Args; |
| 1940 | if (Arg) |
| 1941 | Args.push_back(Elt: Arg); |
| 1942 | Args.push_back(Elt: Rsrc); |
| 1943 | Args.push_back(Elt: Off); |
| 1944 | insertPreMemOpFence(Order, SSID); |
| 1945 | // soffset is always 0 for these cases, where we always want any offset to be |
| 1946 | // part of bounds checking and we don't know which parts of the GEPs is |
| 1947 | // uniform. |
| 1948 | Args.push_back(Elt: IRB.getInt32(C: 0)); |
| 1949 | |
| 1950 | uint32_t Aux = 0; |
| 1951 | if (IsVolatile) |
| 1952 | Aux |= AMDGPU::CPol::VOLATILE; |
| 1953 | Args.push_back(Elt: IRB.getInt32(C: Aux)); |
| 1954 | |
| 1955 | Intrinsic::ID IID = Intrinsic::not_intrinsic; |
| 1956 | if (isa<LoadInst>(Val: I)) |
| 1957 | IID = Order == AtomicOrdering::NotAtomic |
| 1958 | ? Intrinsic::amdgcn_raw_ptr_buffer_load |
| 1959 | : Intrinsic::amdgcn_raw_ptr_atomic_buffer_load; |
| 1960 | else if (isa<StoreInst>(Val: I)) |
| 1961 | IID = Intrinsic::amdgcn_raw_ptr_buffer_store; |
| 1962 | else if (auto *RMW = dyn_cast<AtomicRMWInst>(Val: I)) { |
| 1963 | switch (RMW->getOperation()) { |
| 1964 | case AtomicRMWInst::Xchg: |
| 1965 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_swap; |
| 1966 | break; |
| 1967 | case AtomicRMWInst::Add: |
| 1968 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_add; |
| 1969 | break; |
| 1970 | case AtomicRMWInst::Sub: |
| 1971 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_sub; |
| 1972 | break; |
| 1973 | case AtomicRMWInst::And: |
| 1974 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_and; |
| 1975 | break; |
| 1976 | case AtomicRMWInst::Or: |
| 1977 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_or; |
| 1978 | break; |
| 1979 | case AtomicRMWInst::Xor: |
| 1980 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_xor; |
| 1981 | break; |
| 1982 | case AtomicRMWInst::Max: |
| 1983 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_smax; |
| 1984 | break; |
| 1985 | case AtomicRMWInst::Min: |
| 1986 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_smin; |
| 1987 | break; |
| 1988 | case AtomicRMWInst::UMax: |
| 1989 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_umax; |
| 1990 | break; |
| 1991 | case AtomicRMWInst::UMin: |
| 1992 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_umin; |
| 1993 | break; |
| 1994 | case AtomicRMWInst::FAdd: |
| 1995 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_fadd; |
| 1996 | break; |
| 1997 | case AtomicRMWInst::FMax: |
| 1998 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_fmax; |
| 1999 | break; |
| 2000 | case AtomicRMWInst::FMin: |
| 2001 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_fmin; |
| 2002 | break; |
| 2003 | case AtomicRMWInst::USubCond: |
| 2004 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_cond_sub_u32; |
| 2005 | break; |
| 2006 | case AtomicRMWInst::USubSat: |
| 2007 | IID = Intrinsic::amdgcn_raw_ptr_buffer_atomic_sub_clamp_u32; |
| 2008 | break; |
| 2009 | case AtomicRMWInst::FSub: { |
| 2010 | reportFatalUsageError( |
| 2011 | reason: "atomic floating point subtraction not supported for " |
| 2012 | "buffer resources and should've been expanded away"); |
| 2013 | break; |
| 2014 | } |
| 2015 | case AtomicRMWInst::FMaximum: { |
| 2016 | reportFatalUsageError( |
| 2017 | reason: "atomic floating point fmaximum not supported for " |
| 2018 | "buffer resources and should've been expanded away"); |
| 2019 | break; |
| 2020 | } |
| 2021 | case AtomicRMWInst::FMinimum: { |
| 2022 | reportFatalUsageError( |
| 2023 | reason: "atomic floating point fminimum not supported for " |
| 2024 | "buffer resources and should've been expanded away"); |
| 2025 | break; |
| 2026 | } |
| 2027 | case AtomicRMWInst::FMaximumNum: { |
| 2028 | reportFatalUsageError( |
| 2029 | reason: "atomic floating point fmaximumnum not supported for " |
| 2030 | "buffer resources and should've been expanded away"); |
| 2031 | break; |
| 2032 | } |
| 2033 | case AtomicRMWInst::FMinimumNum: { |
| 2034 | reportFatalUsageError( |
| 2035 | reason: "atomic floating point fminimumnum not supported for " |
| 2036 | "buffer resources and should've been expanded away"); |
| 2037 | break; |
| 2038 | } |
| 2039 | case AtomicRMWInst::Nand: |
| 2040 | reportFatalUsageError( |
| 2041 | reason: "atomic nand not supported for buffer resources and " |
| 2042 | "should've been expanded away"); |
| 2043 | break; |
| 2044 | case AtomicRMWInst::UIncWrap: |
| 2045 | case AtomicRMWInst::UDecWrap: |
| 2046 | reportFatalUsageError( |
| 2047 | reason: "wrapping increment/decrement not supported for " |
| 2048 | "buffer resources and should've been expanded away"); |
| 2049 | break; |
| 2050 | case AtomicRMWInst::BAD_BINOP: |
| 2051 | llvm_unreachable("Not sure how we got a bad binop"); |
| 2052 | } |
| 2053 | } |
| 2054 | |
| 2055 | CallInst *Call = IRB.CreateIntrinsicWithoutFolding(ID: IID, OverloadTypes: Ty, Args); |
| 2056 | copyMetadata(Dest: Call, Src: I); |
| 2057 | setAlign(Intr: Call, A: Alignment, RsrcArgIdx: Arg ? 1 : 0); |
| 2058 | Call->takeName(V: I); |
| 2059 | |
| 2060 | insertPostMemOpFence(Order, SSID); |
| 2061 | // The "no moving p7 directly" rewrites ensure that this load or store won't |
| 2062 | // itself need to be split into parts. |
| 2063 | SplitUsers.insert(V: I); |
| 2064 | I->replaceAllUsesWith(V: Call); |
| 2065 | return Call; |
| 2066 | } |
| 2067 | |
| 2068 | PtrParts SplitPtrStructs::visitInstruction(Instruction &I) { |
| 2069 | return {nullptr, nullptr}; |
| 2070 | } |
| 2071 | |
| 2072 | PtrParts SplitPtrStructs::visitLoadInst(LoadInst &LI) { |
| 2073 | if (!isSplitFatPtr(Ty: LI.getPointerOperandType())) |
| 2074 | return {nullptr, nullptr}; |
| 2075 | handleMemoryInst(I: &LI, Arg: nullptr, Ptr: LI.getPointerOperand(), Ty: LI.getType(), |
| 2076 | Alignment: LI.getAlign(), Order: LI.getOrdering(), IsVolatile: LI.isVolatile(), |
| 2077 | SSID: LI.getSyncScopeID()); |
| 2078 | return {nullptr, nullptr}; |
| 2079 | } |
| 2080 | |
| 2081 | PtrParts SplitPtrStructs::visitStoreInst(StoreInst &SI) { |
| 2082 | if (!isSplitFatPtr(Ty: SI.getPointerOperandType())) |
| 2083 | return {nullptr, nullptr}; |
| 2084 | Value *Arg = SI.getValueOperand(); |
| 2085 | handleMemoryInst(I: &SI, Arg, Ptr: SI.getPointerOperand(), Ty: Arg->getType(), |
| 2086 | Alignment: SI.getAlign(), Order: SI.getOrdering(), IsVolatile: SI.isVolatile(), |
| 2087 | SSID: SI.getSyncScopeID()); |
| 2088 | return {nullptr, nullptr}; |
| 2089 | } |
| 2090 | |
| 2091 | PtrParts SplitPtrStructs::visitAtomicRMWInst(AtomicRMWInst &AI) { |
| 2092 | if (!isSplitFatPtr(Ty: AI.getPointerOperand()->getType())) |
| 2093 | return {nullptr, nullptr}; |
| 2094 | Value *Arg = AI.getValOperand(); |
| 2095 | handleMemoryInst(I: &AI, Arg, Ptr: AI.getPointerOperand(), Ty: Arg->getType(), |
| 2096 | Alignment: AI.getAlign(), Order: AI.getOrdering(), IsVolatile: AI.isVolatile(), |
| 2097 | SSID: AI.getSyncScopeID()); |
| 2098 | return {nullptr, nullptr}; |
| 2099 | } |
| 2100 | |
| 2101 | // Unlike load, store, and RMW, cmpxchg needs special handling to account |
| 2102 | // for the boolean argument. |
| 2103 | PtrParts SplitPtrStructs::visitAtomicCmpXchgInst(AtomicCmpXchgInst &AI) { |
| 2104 | Value *Ptr = AI.getPointerOperand(); |
| 2105 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2106 | return {nullptr, nullptr}; |
| 2107 | IRB.SetInsertPoint(&AI); |
| 2108 | |
| 2109 | Type *Ty = AI.getNewValOperand()->getType(); |
| 2110 | AtomicOrdering Order = AI.getMergedOrdering(); |
| 2111 | SyncScope::ID SSID = AI.getSyncScopeID(); |
| 2112 | bool IsNonTemporal = AI.getMetadata(KindID: LLVMContext::MD_nontemporal); |
| 2113 | |
| 2114 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2115 | insertPreMemOpFence(Order, SSID); |
| 2116 | |
| 2117 | uint32_t Aux = 0; |
| 2118 | if (IsNonTemporal) |
| 2119 | Aux |= AMDGPU::CPol::SLC; |
| 2120 | if (AI.isVolatile()) |
| 2121 | Aux |= AMDGPU::CPol::VOLATILE; |
| 2122 | CallInst *Call = IRB.CreateIntrinsicWithoutFolding( |
| 2123 | ID: Intrinsic::amdgcn_raw_ptr_buffer_atomic_cmpswap, OverloadTypes: Ty, |
| 2124 | Args: {AI.getNewValOperand(), AI.getCompareOperand(), Rsrc, Off, |
| 2125 | IRB.getInt32(C: 0), IRB.getInt32(C: Aux)}); |
| 2126 | copyMetadata(Dest: Call, Src: &AI); |
| 2127 | setAlign(Intr: Call, A: AI.getAlign(), RsrcArgIdx: 2); |
| 2128 | Call->takeName(V: &AI); |
| 2129 | insertPostMemOpFence(Order, SSID); |
| 2130 | |
| 2131 | Value *Res = PoisonValue::get(T: AI.getType()); |
| 2132 | Res = IRB.CreateInsertValue(Agg: Res, Val: Call, Idxs: 0); |
| 2133 | Value *Succeeded = IRB.CreateICmpEQ(LHS: Call, RHS: AI.getCompareOperand()); |
| 2134 | Res = IRB.CreateInsertValue(Agg: Res, Val: Succeeded, Idxs: 1); |
| 2135 | SplitUsers.insert(V: &AI); |
| 2136 | AI.replaceAllUsesWith(V: Res); |
| 2137 | return {nullptr, nullptr}; |
| 2138 | } |
| 2139 | |
| 2140 | PtrParts SplitPtrStructs::visitGetElementPtrInst(GetElementPtrInst &GEP) { |
| 2141 | using namespace llvm::PatternMatch; |
| 2142 | Value *Ptr = GEP.getPointerOperand(); |
| 2143 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2144 | return {nullptr, nullptr}; |
| 2145 | IRB.SetInsertPoint(&GEP); |
| 2146 | |
| 2147 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2148 | const DataLayout &DL = GEP.getDataLayout(); |
| 2149 | bool IsNUW = GEP.hasNoUnsignedWrap(); |
| 2150 | bool IsNUSW = GEP.hasNoUnsignedSignedWrap(); |
| 2151 | |
| 2152 | StructType *ResTy = cast<StructType>(Val: GEP.getType()); |
| 2153 | Type *ResRsrcTy = ResTy->getElementType(N: 0); |
| 2154 | VectorType *ResRsrcVecTy = dyn_cast<VectorType>(Val: ResRsrcTy); |
| 2155 | bool BroadcastsPtr = ResRsrcVecTy && !isa<VectorType>(Val: Off->getType()); |
| 2156 | |
| 2157 | // In order to call emitGEPOffset() and thus not have to reimplement it, |
| 2158 | // we need the GEP result to have ptr addrspace(7) type. |
| 2159 | Type *FatPtrTy = |
| 2160 | ResRsrcTy->getWithNewType(EltTy: IRB.getPtrTy(AddrSpace: AMDGPUAS::BUFFER_FAT_POINTER)); |
| 2161 | GEP.mutateType(Ty: FatPtrTy); |
| 2162 | Value *OffAccum = emitGEPOffset(Builder: &IRB, DL, GEP: &GEP); |
| 2163 | GEP.mutateType(Ty: ResTy); |
| 2164 | |
| 2165 | if (BroadcastsPtr) { |
| 2166 | Rsrc = IRB.CreateVectorSplat(EC: ResRsrcVecTy->getElementCount(), V: Rsrc, |
| 2167 | Name: Rsrc->getName()); |
| 2168 | Off = IRB.CreateVectorSplat(EC: ResRsrcVecTy->getElementCount(), V: Off, |
| 2169 | Name: Off->getName()); |
| 2170 | } |
| 2171 | if (match(V: OffAccum, P: m_Zero())) { // Constant-zero offset |
| 2172 | SplitUsers.insert(V: &GEP); |
| 2173 | return {Rsrc, Off}; |
| 2174 | } |
| 2175 | |
| 2176 | bool HasNonNegativeOff = false; |
| 2177 | if (auto *CI = dyn_cast<ConstantInt>(Val: OffAccum)) { |
| 2178 | HasNonNegativeOff = !CI->isNegative(); |
| 2179 | } |
| 2180 | Value *NewOff; |
| 2181 | if (match(V: Off, P: m_Zero())) { |
| 2182 | NewOff = OffAccum; |
| 2183 | } else { |
| 2184 | NewOff = IRB.CreateAdd(LHS: Off, RHS: OffAccum, Name: "", |
| 2185 | /*hasNUW=*/HasNUW: IsNUW || (IsNUSW && HasNonNegativeOff), |
| 2186 | /*hasNSW=*/HasNSW: false); |
| 2187 | } |
| 2188 | copyMetadata(Dest: NewOff, Src: &GEP); |
| 2189 | NewOff->takeName(V: &GEP); |
| 2190 | SplitUsers.insert(V: &GEP); |
| 2191 | return {Rsrc, NewOff}; |
| 2192 | } |
| 2193 | |
| 2194 | PtrParts SplitPtrStructs::visitPtrToIntInst(PtrToIntInst &PI) { |
| 2195 | Value *Ptr = PI.getPointerOperand(); |
| 2196 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2197 | return {nullptr, nullptr}; |
| 2198 | IRB.SetInsertPoint(&PI); |
| 2199 | |
| 2200 | Type *ResTy = PI.getType(); |
| 2201 | unsigned Width = ResTy->getScalarSizeInBits(); |
| 2202 | |
| 2203 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2204 | const DataLayout &DL = PI.getDataLayout(); |
| 2205 | unsigned FatPtrWidth = DL.getPointerSizeInBits(AS: AMDGPUAS::BUFFER_FAT_POINTER); |
| 2206 | |
| 2207 | Value *Res; |
| 2208 | if (Width <= BufferOffsetWidth) { |
| 2209 | Res = IRB.CreateIntCast(V: Off, DestTy: ResTy, /*isSigned=*/false, |
| 2210 | Name: PI.getName() + ".off"); |
| 2211 | } else { |
| 2212 | Value *RsrcInt = IRB.CreatePtrToInt(V: Rsrc, DestTy: ResTy, Name: PI.getName() + ".rsrc"); |
| 2213 | Value *Shl = IRB.CreateShl( |
| 2214 | LHS: RsrcInt, |
| 2215 | RHS: ConstantExpr::getIntegerValue(Ty: ResTy, V: APInt(Width, BufferOffsetWidth)), |
| 2216 | Name: "", HasNUW: Width >= FatPtrWidth, HasNSW: Width > FatPtrWidth); |
| 2217 | Value *OffCast = IRB.CreateIntCast(V: Off, DestTy: ResTy, /*isSigned=*/false, |
| 2218 | Name: PI.getName() + ".off"); |
| 2219 | Res = IRB.CreateOr(LHS: Shl, RHS: OffCast); |
| 2220 | } |
| 2221 | |
| 2222 | copyMetadata(Dest: Res, Src: &PI); |
| 2223 | Res->takeName(V: &PI); |
| 2224 | SplitUsers.insert(V: &PI); |
| 2225 | PI.replaceAllUsesWith(V: Res); |
| 2226 | return {nullptr, nullptr}; |
| 2227 | } |
| 2228 | |
| 2229 | PtrParts SplitPtrStructs::visitPtrToAddrInst(PtrToAddrInst &PA) { |
| 2230 | Value *Ptr = PA.getPointerOperand(); |
| 2231 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2232 | return {nullptr, nullptr}; |
| 2233 | IRB.SetInsertPoint(&PA); |
| 2234 | |
| 2235 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2236 | Value *Res = IRB.CreateIntCast(V: Off, DestTy: PA.getType(), /*isSigned=*/false); |
| 2237 | copyMetadata(Dest: Res, Src: &PA); |
| 2238 | Res->takeName(V: &PA); |
| 2239 | SplitUsers.insert(V: &PA); |
| 2240 | PA.replaceAllUsesWith(V: Res); |
| 2241 | return {nullptr, nullptr}; |
| 2242 | } |
| 2243 | |
| 2244 | PtrParts SplitPtrStructs::visitIntToPtrInst(IntToPtrInst &IP) { |
| 2245 | if (!isSplitFatPtr(Ty: IP.getType())) |
| 2246 | return {nullptr, nullptr}; |
| 2247 | IRB.SetInsertPoint(&IP); |
| 2248 | const DataLayout &DL = IP.getDataLayout(); |
| 2249 | unsigned RsrcPtrWidth = DL.getPointerSizeInBits(AS: AMDGPUAS::BUFFER_RESOURCE); |
| 2250 | Value *Int = IP.getOperand(i_nocapture: 0); |
| 2251 | Type *IntTy = Int->getType(); |
| 2252 | Type *RsrcIntTy = IntTy->getWithNewBitWidth(NewBitWidth: RsrcPtrWidth); |
| 2253 | unsigned Width = IntTy->getScalarSizeInBits(); |
| 2254 | |
| 2255 | auto *RetTy = cast<StructType>(Val: IP.getType()); |
| 2256 | Type *RsrcTy = RetTy->getElementType(N: 0); |
| 2257 | Type *OffTy = RetTy->getElementType(N: 1); |
| 2258 | // inttoptr zero-extends, so narrow inputs contribute nothing to the resource |
| 2259 | // part. |
| 2260 | Value *RsrcInt; |
| 2261 | if (Width <= BufferOffsetWidth) { |
| 2262 | RsrcInt = Constant::getNullValue(Ty: RsrcIntTy); |
| 2263 | } else { |
| 2264 | Value *RsrcPart = |
| 2265 | IRB.CreateLShr(LHS: Int, RHS: ConstantInt::get(Ty: IntTy, V: BufferOffsetWidth)); |
| 2266 | RsrcInt = IRB.CreateIntCast(V: RsrcPart, DestTy: RsrcIntTy, /*isSigned=*/false); |
| 2267 | } |
| 2268 | Value *Rsrc = IRB.CreateIntToPtr(V: RsrcInt, DestTy: RsrcTy, Name: IP.getName() + ".rsrc"); |
| 2269 | Value *Off = |
| 2270 | IRB.CreateIntCast(V: Int, DestTy: OffTy, /*IsSigned=*/isSigned: false, Name: IP.getName() + ".off"); |
| 2271 | |
| 2272 | copyMetadata(Dest: Rsrc, Src: &IP); |
| 2273 | SplitUsers.insert(V: &IP); |
| 2274 | return {Rsrc, Off}; |
| 2275 | } |
| 2276 | |
| 2277 | PtrParts SplitPtrStructs::visitAddrSpaceCastInst(AddrSpaceCastInst &I) { |
| 2278 | // TODO(krzysz00): handle casts from ptr addrspace(7) to global pointers |
| 2279 | // by computing the effective address. |
| 2280 | if (!isSplitFatPtr(Ty: I.getType())) |
| 2281 | return {nullptr, nullptr}; |
| 2282 | IRB.SetInsertPoint(&I); |
| 2283 | Value *In = I.getPointerOperand(); |
| 2284 | // No-op casts preserve parts |
| 2285 | if (In->getType() == I.getType()) { |
| 2286 | auto [Rsrc, Off] = getPtrParts(V: In); |
| 2287 | SplitUsers.insert(V: &I); |
| 2288 | return {Rsrc, Off}; |
| 2289 | } |
| 2290 | |
| 2291 | auto *ResTy = cast<StructType>(Val: I.getType()); |
| 2292 | Type *RsrcTy = ResTy->getElementType(N: 0); |
| 2293 | Type *OffTy = ResTy->getElementType(N: 1); |
| 2294 | Value *ZeroOff = Constant::getNullValue(Ty: OffTy); |
| 2295 | |
| 2296 | // Special case for null pointers, undef, and poison, which can be created by |
| 2297 | // address space propagation. |
| 2298 | auto *InConst = dyn_cast<Constant>(Val: In); |
| 2299 | if (InConst && InConst->isNullValue()) { |
| 2300 | Value *NullRsrc = Constant::getNullValue(Ty: RsrcTy); |
| 2301 | SplitUsers.insert(V: &I); |
| 2302 | return {NullRsrc, ZeroOff}; |
| 2303 | } |
| 2304 | if (isa<PoisonValue>(Val: In)) { |
| 2305 | Value *PoisonRsrc = PoisonValue::get(T: RsrcTy); |
| 2306 | Value *PoisonOff = PoisonValue::get(T: OffTy); |
| 2307 | SplitUsers.insert(V: &I); |
| 2308 | return {PoisonRsrc, PoisonOff}; |
| 2309 | } |
| 2310 | if (isa<UndefValue>(Val: In)) { |
| 2311 | Value *UndefRsrc = UndefValue::get(T: RsrcTy); |
| 2312 | Value *UndefOff = UndefValue::get(T: OffTy); |
| 2313 | SplitUsers.insert(V: &I); |
| 2314 | return {UndefRsrc, UndefOff}; |
| 2315 | } |
| 2316 | |
| 2317 | if (I.getSrcAddressSpace() != AMDGPUAS::BUFFER_RESOURCE) |
| 2318 | reportFatalUsageError( |
| 2319 | reason: "only buffer resources (addrspace 8) and null/poison pointers can be " |
| 2320 | "cast to buffer fat pointers (addrspace 7)"); |
| 2321 | SplitUsers.insert(V: &I); |
| 2322 | return {In, ZeroOff}; |
| 2323 | } |
| 2324 | |
| 2325 | PtrParts SplitPtrStructs::visitICmpInst(ICmpInst &Cmp) { |
| 2326 | Value *Lhs = Cmp.getOperand(i_nocapture: 0); |
| 2327 | if (!isSplitFatPtr(Ty: Lhs->getType())) |
| 2328 | return {nullptr, nullptr}; |
| 2329 | Value *Rhs = Cmp.getOperand(i_nocapture: 1); |
| 2330 | IRB.SetInsertPoint(&Cmp); |
| 2331 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
| 2332 | |
| 2333 | assert((Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE) && |
| 2334 | "Pointer comparison is only equal or unequal"); |
| 2335 | auto [LhsRsrc, LhsOff] = getPtrParts(V: Lhs); |
| 2336 | auto [RhsRsrc, RhsOff] = getPtrParts(V: Rhs); |
| 2337 | Value *Res = IRB.CreateICmp(P: Pred, LHS: LhsOff, RHS: RhsOff); |
| 2338 | copyMetadata(Dest: Res, Src: &Cmp); |
| 2339 | Res->takeName(V: &Cmp); |
| 2340 | SplitUsers.insert(V: &Cmp); |
| 2341 | Cmp.replaceAllUsesWith(V: Res); |
| 2342 | return {nullptr, nullptr}; |
| 2343 | } |
| 2344 | |
| 2345 | PtrParts SplitPtrStructs::visitFreezeInst(FreezeInst &I) { |
| 2346 | if (!isSplitFatPtr(Ty: I.getType())) |
| 2347 | return {nullptr, nullptr}; |
| 2348 | IRB.SetInsertPoint(&I); |
| 2349 | auto [Rsrc, Off] = getPtrParts(V: I.getOperand(i_nocapture: 0)); |
| 2350 | |
| 2351 | Value *RsrcRes = IRB.CreateFreeze(V: Rsrc, Name: I.getName() + ".rsrc"); |
| 2352 | copyMetadata(Dest: RsrcRes, Src: &I); |
| 2353 | Value *OffRes = IRB.CreateFreeze(V: Off, Name: I.getName() + ".off"); |
| 2354 | copyMetadata(Dest: OffRes, Src: &I); |
| 2355 | SplitUsers.insert(V: &I); |
| 2356 | return {RsrcRes, OffRes}; |
| 2357 | } |
| 2358 | |
| 2359 | PtrParts SplitPtrStructs::visitExtractElementInst(ExtractElementInst &I) { |
| 2360 | if (!isSplitFatPtr(Ty: I.getType())) |
| 2361 | return {nullptr, nullptr}; |
| 2362 | IRB.SetInsertPoint(&I); |
| 2363 | Value *Vec = I.getVectorOperand(); |
| 2364 | Value *Idx = I.getIndexOperand(); |
| 2365 | auto [Rsrc, Off] = getPtrParts(V: Vec); |
| 2366 | |
| 2367 | Value *RsrcRes = IRB.CreateExtractElement(Vec: Rsrc, Idx, Name: I.getName() + ".rsrc"); |
| 2368 | copyMetadata(Dest: RsrcRes, Src: &I); |
| 2369 | Value *OffRes = IRB.CreateExtractElement(Vec: Off, Idx, Name: I.getName() + ".off"); |
| 2370 | copyMetadata(Dest: OffRes, Src: &I); |
| 2371 | SplitUsers.insert(V: &I); |
| 2372 | return {RsrcRes, OffRes}; |
| 2373 | } |
| 2374 | |
| 2375 | PtrParts SplitPtrStructs::visitInsertElementInst(InsertElementInst &I) { |
| 2376 | // The mutated instructions temporarily don't return vectors, and so |
| 2377 | // we need the generic getType() here to avoid crashes. |
| 2378 | if (!isSplitFatPtr(Ty: cast<Instruction>(Val&: I).getType())) |
| 2379 | return {nullptr, nullptr}; |
| 2380 | IRB.SetInsertPoint(&I); |
| 2381 | Value *Vec = I.getOperand(i_nocapture: 0); |
| 2382 | Value *Elem = I.getOperand(i_nocapture: 1); |
| 2383 | Value *Idx = I.getOperand(i_nocapture: 2); |
| 2384 | auto [VecRsrc, VecOff] = getPtrParts(V: Vec); |
| 2385 | auto [ElemRsrc, ElemOff] = getPtrParts(V: Elem); |
| 2386 | |
| 2387 | Value *RsrcRes = |
| 2388 | IRB.CreateInsertElement(Vec: VecRsrc, NewElt: ElemRsrc, Idx, Name: I.getName() + ".rsrc"); |
| 2389 | copyMetadata(Dest: RsrcRes, Src: &I); |
| 2390 | Value *OffRes = |
| 2391 | IRB.CreateInsertElement(Vec: VecOff, NewElt: ElemOff, Idx, Name: I.getName() + ".off"); |
| 2392 | copyMetadata(Dest: OffRes, Src: &I); |
| 2393 | SplitUsers.insert(V: &I); |
| 2394 | return {RsrcRes, OffRes}; |
| 2395 | } |
| 2396 | |
| 2397 | PtrParts SplitPtrStructs::visitShuffleVectorInst(ShuffleVectorInst &I) { |
| 2398 | // Cast is needed for the same reason as insertelement's. |
| 2399 | if (!isSplitFatPtr(Ty: cast<Instruction>(Val&: I).getType())) |
| 2400 | return {nullptr, nullptr}; |
| 2401 | IRB.SetInsertPoint(&I); |
| 2402 | |
| 2403 | Value *V1 = I.getOperand(i_nocapture: 0); |
| 2404 | Value *V2 = I.getOperand(i_nocapture: 1); |
| 2405 | ArrayRef<int> Mask = I.getShuffleMask(); |
| 2406 | auto [V1Rsrc, V1Off] = getPtrParts(V: V1); |
| 2407 | auto [V2Rsrc, V2Off] = getPtrParts(V: V2); |
| 2408 | |
| 2409 | Value *RsrcRes = |
| 2410 | IRB.CreateShuffleVector(V1: V1Rsrc, V2: V2Rsrc, Mask, Name: I.getName() + ".rsrc"); |
| 2411 | copyMetadata(Dest: RsrcRes, Src: &I); |
| 2412 | Value *OffRes = |
| 2413 | IRB.CreateShuffleVector(V1: V1Off, V2: V2Off, Mask, Name: I.getName() + ".off"); |
| 2414 | copyMetadata(Dest: OffRes, Src: &I); |
| 2415 | SplitUsers.insert(V: &I); |
| 2416 | return {RsrcRes, OffRes}; |
| 2417 | } |
| 2418 | |
| 2419 | PtrParts SplitPtrStructs::visitPHINode(PHINode &PHI) { |
| 2420 | if (!isSplitFatPtr(Ty: PHI.getType())) |
| 2421 | return {nullptr, nullptr}; |
| 2422 | IRB.SetInsertPoint(*PHI.getInsertionPointAfterDef()); |
| 2423 | // Phi nodes will be handled in post-processing after we've visited every |
| 2424 | // instruction. However, instead of just returning {nullptr, nullptr}, |
| 2425 | // we explicitly create the temporary extractvalue operations that are our |
| 2426 | // temporary results so that they end up at the beginning of the block with |
| 2427 | // the PHIs. |
| 2428 | Value *TmpRsrc = IRB.CreateExtractValue(Agg: &PHI, Idxs: 0, Name: PHI.getName() + ".rsrc"); |
| 2429 | Value *TmpOff = IRB.CreateExtractValue(Agg: &PHI, Idxs: 1, Name: PHI.getName() + ".off"); |
| 2430 | Conditionals.push_back(Elt: &PHI); |
| 2431 | SplitUsers.insert(V: &PHI); |
| 2432 | return {TmpRsrc, TmpOff}; |
| 2433 | } |
| 2434 | |
| 2435 | PtrParts SplitPtrStructs::visitSelectInst(SelectInst &SI) { |
| 2436 | if (!isSplitFatPtr(Ty: SI.getType())) |
| 2437 | return {nullptr, nullptr}; |
| 2438 | IRB.SetInsertPoint(&SI); |
| 2439 | |
| 2440 | Value *Cond = SI.getCondition(); |
| 2441 | Value *True = SI.getTrueValue(); |
| 2442 | Value *False = SI.getFalseValue(); |
| 2443 | auto [TrueRsrc, TrueOff] = getPtrParts(V: True); |
| 2444 | auto [FalseRsrc, FalseOff] = getPtrParts(V: False); |
| 2445 | |
| 2446 | Value *RsrcRes = |
| 2447 | IRB.CreateSelect(C: Cond, True: TrueRsrc, False: FalseRsrc, Name: SI.getName() + ".rsrc", MDFrom: &SI); |
| 2448 | copyMetadata(Dest: RsrcRes, Src: &SI); |
| 2449 | Conditionals.push_back(Elt: &SI); |
| 2450 | Value *OffRes = |
| 2451 | IRB.CreateSelect(C: Cond, True: TrueOff, False: FalseOff, Name: SI.getName() + ".off", MDFrom: &SI); |
| 2452 | copyMetadata(Dest: OffRes, Src: &SI); |
| 2453 | SplitUsers.insert(V: &SI); |
| 2454 | return {RsrcRes, OffRes}; |
| 2455 | } |
| 2456 | |
| 2457 | /// Returns true if this intrinsic needs to be removed when it is |
| 2458 | /// applied to `ptr addrspace(7)` values. Calls to these intrinsics are |
| 2459 | /// rewritten into calls to versions of that intrinsic on the resource |
| 2460 | /// descriptor. |
| 2461 | static bool isRemovablePointerIntrinsic(Intrinsic::ID IID) { |
| 2462 | switch (IID) { |
| 2463 | default: |
| 2464 | return false; |
| 2465 | case Intrinsic::amdgcn_make_buffer_rsrc: |
| 2466 | case Intrinsic::ptrmask: |
| 2467 | case Intrinsic::invariant_start: |
| 2468 | case Intrinsic::invariant_end: |
| 2469 | case Intrinsic::launder_invariant_group: |
| 2470 | case Intrinsic::memcpy: |
| 2471 | case Intrinsic::memcpy_inline: |
| 2472 | case Intrinsic::memmove: |
| 2473 | case Intrinsic::memset: |
| 2474 | case Intrinsic::memset_inline: |
| 2475 | case Intrinsic::experimental_memset_pattern: |
| 2476 | case Intrinsic::amdgcn_load_to_lds: |
| 2477 | case Intrinsic::amdgcn_load_async_to_lds: |
| 2478 | return true; |
| 2479 | } |
| 2480 | } |
| 2481 | |
| 2482 | PtrParts SplitPtrStructs::visitIntrinsicInst(IntrinsicInst &I) { |
| 2483 | Intrinsic::ID IID = I.getIntrinsicID(); |
| 2484 | switch (IID) { |
| 2485 | default: |
| 2486 | break; |
| 2487 | case Intrinsic::amdgcn_make_buffer_rsrc: { |
| 2488 | if (!isSplitFatPtr(Ty: I.getType())) |
| 2489 | return {nullptr, nullptr}; |
| 2490 | Value *Base = I.getArgOperand(i: 0); |
| 2491 | Value *Stride = I.getArgOperand(i: 1); |
| 2492 | Value *NumRecords = I.getArgOperand(i: 2); |
| 2493 | Value *Flags = I.getArgOperand(i: 3); |
| 2494 | auto *SplitType = cast<StructType>(Val: I.getType()); |
| 2495 | Type *RsrcType = SplitType->getElementType(N: 0); |
| 2496 | Type *OffType = SplitType->getElementType(N: 1); |
| 2497 | IRB.SetInsertPoint(&I); |
| 2498 | Value *Rsrc = IRB.CreateIntrinsic( |
| 2499 | ID: IID, OverloadTypes: {RsrcType, Base->getType(), NumRecords->getType()}, |
| 2500 | Args: {Base, Stride, NumRecords, Flags}); |
| 2501 | copyMetadata(Dest: Rsrc, Src: &I); |
| 2502 | Rsrc->takeName(V: &I); |
| 2503 | Value *Zero = Constant::getNullValue(Ty: OffType); |
| 2504 | SplitUsers.insert(V: &I); |
| 2505 | return {Rsrc, Zero}; |
| 2506 | } |
| 2507 | case Intrinsic::ptrmask: { |
| 2508 | Value *Ptr = I.getArgOperand(i: 0); |
| 2509 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2510 | return {nullptr, nullptr}; |
| 2511 | Value *Mask = I.getArgOperand(i: 1); |
| 2512 | IRB.SetInsertPoint(&I); |
| 2513 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2514 | if (Mask->getType() != Off->getType()) |
| 2515 | reportFatalUsageError(reason: "offset width is not equal to index width of fat " |
| 2516 | "pointer (data layout not set up correctly?)"); |
| 2517 | Value *OffRes = IRB.CreateAnd(LHS: Off, RHS: Mask, Name: I.getName() + ".off"); |
| 2518 | copyMetadata(Dest: OffRes, Src: &I); |
| 2519 | SplitUsers.insert(V: &I); |
| 2520 | return {Rsrc, OffRes}; |
| 2521 | } |
| 2522 | // Pointer annotation intrinsics that, given their object-wide nature |
| 2523 | // operate on the resource part. |
| 2524 | case Intrinsic::invariant_start: { |
| 2525 | Value *Ptr = I.getArgOperand(i: 1); |
| 2526 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2527 | return {nullptr, nullptr}; |
| 2528 | IRB.SetInsertPoint(&I); |
| 2529 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2530 | Type *NewTy = PointerType::get(C&: I.getContext(), AddressSpace: AMDGPUAS::BUFFER_RESOURCE); |
| 2531 | auto *NewRsrc = IRB.CreateIntrinsic(ID: IID, OverloadTypes: {NewTy}, Args: {I.getOperand(i_nocapture: 0), Rsrc}); |
| 2532 | copyMetadata(Dest: NewRsrc, Src: &I); |
| 2533 | NewRsrc->takeName(V: &I); |
| 2534 | SplitUsers.insert(V: &I); |
| 2535 | I.replaceAllUsesWith(V: NewRsrc); |
| 2536 | return {nullptr, nullptr}; |
| 2537 | } |
| 2538 | case Intrinsic::invariant_end: { |
| 2539 | Value *RealPtr = I.getArgOperand(i: 2); |
| 2540 | if (!isSplitFatPtr(Ty: RealPtr->getType())) |
| 2541 | return {nullptr, nullptr}; |
| 2542 | IRB.SetInsertPoint(&I); |
| 2543 | Value *RealRsrc = getPtrParts(V: RealPtr).first; |
| 2544 | Value *InvPtr = I.getArgOperand(i: 0); |
| 2545 | Value *Size = I.getArgOperand(i: 1); |
| 2546 | Value *NewRsrc = IRB.CreateIntrinsic(ID: IID, OverloadTypes: {RealRsrc->getType()}, |
| 2547 | Args: {InvPtr, Size, RealRsrc}); |
| 2548 | copyMetadata(Dest: NewRsrc, Src: &I); |
| 2549 | NewRsrc->takeName(V: &I); |
| 2550 | SplitUsers.insert(V: &I); |
| 2551 | I.replaceAllUsesWith(V: NewRsrc); |
| 2552 | return {nullptr, nullptr}; |
| 2553 | } |
| 2554 | case Intrinsic::launder_invariant_group: { |
| 2555 | Value *Ptr = I.getArgOperand(i: 0); |
| 2556 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2557 | return {nullptr, nullptr}; |
| 2558 | IRB.SetInsertPoint(&I); |
| 2559 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2560 | Value *NewRsrc = IRB.CreateIntrinsic(ID: IID, OverloadTypes: {Rsrc->getType()}, Args: {Rsrc}); |
| 2561 | copyMetadata(Dest: NewRsrc, Src: &I); |
| 2562 | NewRsrc->takeName(V: &I); |
| 2563 | SplitUsers.insert(V: &I); |
| 2564 | return {NewRsrc, Off}; |
| 2565 | } |
| 2566 | case Intrinsic::amdgcn_load_to_lds: |
| 2567 | case Intrinsic::amdgcn_load_async_to_lds: { |
| 2568 | Value *Ptr = I.getArgOperand(i: 0); |
| 2569 | if (!isSplitFatPtr(Ty: Ptr->getType())) |
| 2570 | return {nullptr, nullptr}; |
| 2571 | IRB.SetInsertPoint(&I); |
| 2572 | auto [Rsrc, Off] = getPtrParts(V: Ptr); |
| 2573 | Value *LDSPtr = I.getArgOperand(i: 1); |
| 2574 | Value *LoadSize = I.getArgOperand(i: 2); |
| 2575 | Value *ImmOff = I.getArgOperand(i: 3); |
| 2576 | Value *Aux = I.getArgOperand(i: 4); |
| 2577 | Value *SOffset = IRB.getInt32(C: 0); |
| 2578 | Intrinsic::ID NewIntr = |
| 2579 | IID == Intrinsic::amdgcn_load_to_lds |
| 2580 | ? Intrinsic::amdgcn_raw_ptr_buffer_load_lds |
| 2581 | : Intrinsic::amdgcn_raw_ptr_buffer_load_async_lds; |
| 2582 | Instruction *NewLoad = IRB.CreateIntrinsicWithoutFolding( |
| 2583 | ID: NewIntr, OverloadTypes: {}, Args: {Rsrc, LDSPtr, LoadSize, Off, SOffset, ImmOff, Aux}); |
| 2584 | copyMetadata(Dest: NewLoad, Src: &I); |
| 2585 | SplitUsers.insert(V: &I); |
| 2586 | I.replaceAllUsesWith(V: NewLoad); |
| 2587 | return {nullptr, nullptr}; |
| 2588 | } |
| 2589 | } |
| 2590 | return {nullptr, nullptr}; |
| 2591 | } |
| 2592 | |
| 2593 | void SplitPtrStructs::processFunction(Function &F) { |
| 2594 | ST = &TM->getSubtarget<GCNSubtarget>(F); |
| 2595 | SmallVector<Instruction *, 0> Originals( |
| 2596 | llvm::make_pointer_range(Range: instructions(F))); |
| 2597 | LLVM_DEBUG(dbgs() << "Splitting pointer structs in function: "<< F.getName() |
| 2598 | << "\n"); |
| 2599 | for (Instruction *I : Originals) { |
| 2600 | // In some cases, instruction order doesn't reflect program order, |
| 2601 | // so the visit() call will have already visited coertain instructions |
| 2602 | // by the time this loop gets to them. Avoid re-visiting these so as to, |
| 2603 | // for example, avoid processing the same conditional twice. |
| 2604 | if (SplitUsers.contains(V: I)) |
| 2605 | continue; |
| 2606 | auto [Rsrc, Off] = visit(I); |
| 2607 | assert(((Rsrc && Off) || (!Rsrc && !Off)) && |
| 2608 | "Can't have a resource but no offset"); |
| 2609 | if (Rsrc) |
| 2610 | RsrcParts[I] = Rsrc; |
| 2611 | if (Off) |
| 2612 | OffParts[I] = Off; |
| 2613 | } |
| 2614 | processConditionals(); |
| 2615 | killAndReplaceSplitInstructions(Origs&: Originals); |
| 2616 | |
| 2617 | // Clean up after ourselves to save on memory. |
| 2618 | RsrcParts.clear(); |
| 2619 | OffParts.clear(); |
| 2620 | SplitUsers.clear(); |
| 2621 | Conditionals.clear(); |
| 2622 | ConditionalTemps.clear(); |
| 2623 | } |
| 2624 | |
| 2625 | namespace { |
| 2626 | class AMDGPULowerBufferFatPointers : public ModulePass { |
| 2627 | public: |
| 2628 | static char ID; |
| 2629 | |
| 2630 | AMDGPULowerBufferFatPointers() : ModulePass(ID) {} |
| 2631 | |
| 2632 | bool run(Module &M, const TargetMachine &TM, GetTTIFn GetTTI, GetSEFn GetSE); |
| 2633 | bool runOnModule(Module &M) override; |
| 2634 | |
| 2635 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 2636 | }; |
| 2637 | } // namespace |
| 2638 | |
| 2639 | /// Returns true if there are values that have a buffer fat pointer in them, |
| 2640 | /// which means we'll need to perform rewrites on this function. As a side |
| 2641 | /// effect, this will populate the type remapping cache. |
| 2642 | static bool containsBufferFatPointers(const Function &F, |
| 2643 | BufferFatPtrToStructTypeMap *TypeMap) { |
| 2644 | bool HasFatPointers = false; |
| 2645 | for (const BasicBlock &BB : F) |
| 2646 | for (const Instruction &I : BB) { |
| 2647 | HasFatPointers |= (I.getType() != TypeMap->remapType(SrcTy: I.getType())); |
| 2648 | // Catch null pointer constants in loads, stores, etc. |
| 2649 | for (const Value *V : I.operand_values()) |
| 2650 | HasFatPointers |= (V->getType() != TypeMap->remapType(SrcTy: V->getType())); |
| 2651 | } |
| 2652 | return HasFatPointers; |
| 2653 | } |
| 2654 | |
| 2655 | static bool hasFatPointerInterface(const Function &F, |
| 2656 | BufferFatPtrToStructTypeMap *TypeMap) { |
| 2657 | Type *Ty = F.getFunctionType(); |
| 2658 | return Ty != TypeMap->remapType(SrcTy: Ty); |
| 2659 | } |
| 2660 | |
| 2661 | /// Move the body of `OldF` into a new function, returning it. |
| 2662 | static Function *moveFunctionAdaptingType(Function *OldF, FunctionType *NewTy, |
| 2663 | ValueToValueMapTy &CloneMap) { |
| 2664 | bool IsIntrinsic = OldF->isIntrinsic(); |
| 2665 | Function *NewF = |
| 2666 | Function::Create(Ty: NewTy, Linkage: OldF->getLinkage(), AddrSpace: OldF->getAddressSpace()); |
| 2667 | NewF->copyAttributesFrom(Src: OldF); |
| 2668 | NewF->copyMetadata(Src: OldF, Offset: 0); |
| 2669 | NewF->takeName(V: OldF); |
| 2670 | NewF->updateAfterNameChange(); |
| 2671 | NewF->setDLLStorageClass(OldF->getDLLStorageClass()); |
| 2672 | OldF->getParent()->getFunctionList().insertAfter(where: OldF->getIterator(), New: NewF); |
| 2673 | |
| 2674 | while (!OldF->empty()) { |
| 2675 | BasicBlock *BB = &OldF->front(); |
| 2676 | BB->removeFromParent(); |
| 2677 | BB->insertInto(Parent: NewF); |
| 2678 | CloneMap[BB] = BB; |
| 2679 | for (Instruction &I : *BB) { |
| 2680 | CloneMap[&I] = &I; |
| 2681 | } |
| 2682 | } |
| 2683 | |
| 2684 | SmallVector<AttributeSet> ArgAttrs; |
| 2685 | AttributeList OldAttrs = OldF->getAttributes(); |
| 2686 | |
| 2687 | for (auto [I, OldArg, NewArg] : enumerate(First: OldF->args(), Rest: NewF->args())) { |
| 2688 | CloneMap[&NewArg] = &OldArg; |
| 2689 | NewArg.takeName(V: &OldArg); |
| 2690 | Type *OldArgTy = OldArg.getType(), *NewArgTy = NewArg.getType(); |
| 2691 | // Temporarily mutate type of `NewArg` to allow RAUW to work. |
| 2692 | NewArg.mutateType(Ty: OldArgTy); |
| 2693 | OldArg.replaceAllUsesWith(V: &NewArg); |
| 2694 | NewArg.mutateType(Ty: NewArgTy); |
| 2695 | |
| 2696 | AttributeSet ArgAttr = OldAttrs.getParamAttrs(ArgNo: I); |
| 2697 | // Intrinsics get their attributes fixed later. |
| 2698 | if (OldArgTy != NewArgTy && !IsIntrinsic) |
| 2699 | ArgAttr = ArgAttr.removeAttributes( |
| 2700 | C&: NewF->getContext(), |
| 2701 | AttrsToRemove: AttributeFuncs::typeIncompatible(Ty: NewArgTy, AS: ArgAttr)); |
| 2702 | ArgAttrs.push_back(Elt: ArgAttr); |
| 2703 | } |
| 2704 | AttributeSet RetAttrs = OldAttrs.getRetAttrs(); |
| 2705 | if (OldF->getReturnType() != NewF->getReturnType() && !IsIntrinsic) |
| 2706 | RetAttrs = RetAttrs.removeAttributes( |
| 2707 | C&: NewF->getContext(), |
| 2708 | AttrsToRemove: AttributeFuncs::typeIncompatible(Ty: NewF->getReturnType(), AS: RetAttrs)); |
| 2709 | NewF->setAttributes(AttributeList::get( |
| 2710 | C&: NewF->getContext(), FnAttrs: OldAttrs.getFnAttrs(), RetAttrs, ArgAttrs)); |
| 2711 | return NewF; |
| 2712 | } |
| 2713 | |
| 2714 | static void makeCloneInPraceMap(Function *F, ValueToValueMapTy &CloneMap) { |
| 2715 | for (Argument &A : F->args()) |
| 2716 | CloneMap[&A] = &A; |
| 2717 | for (BasicBlock &BB : *F) { |
| 2718 | CloneMap[&BB] = &BB; |
| 2719 | for (Instruction &I : BB) |
| 2720 | CloneMap[&I] = &I; |
| 2721 | } |
| 2722 | } |
| 2723 | |
| 2724 | bool AMDGPULowerBufferFatPointers::run(Module &M, const TargetMachine &TM, |
| 2725 | GetTTIFn GetTTI, GetSEFn GetSE) { |
| 2726 | bool Changed = false; |
| 2727 | const DataLayout &DL = M.getDataLayout(); |
| 2728 | // Record the functions which need to be remapped. |
| 2729 | // The second element of the pair indicates whether the function has to have |
| 2730 | // its arguments or return types adjusted. |
| 2731 | SmallVector<std::pair<Function *, bool>> NeedsRemap; |
| 2732 | |
| 2733 | LLVMContext &Ctx = M.getContext(); |
| 2734 | |
| 2735 | BufferFatPtrToStructTypeMap StructTM(DL); |
| 2736 | BufferFatPtrToIntTypeMap IntTM(DL); |
| 2737 | for (GlobalVariable &GV : make_early_inc_range(Range: M.globals())) { |
| 2738 | if (GV.getAddressSpace() == AMDGPUAS::BUFFER_FAT_POINTER) { |
| 2739 | // FIXME: Use DiagnosticInfo unsupported but it requires a Function |
| 2740 | Ctx.emitError(ErrorStr: "global variables with a buffer fat pointer address " |
| 2741 | "space (7) are not supported"); |
| 2742 | GV.replaceAllUsesWith(V: PoisonValue::get(T: GV.getType())); |
| 2743 | GV.eraseFromParent(); |
| 2744 | Changed = true; |
| 2745 | continue; |
| 2746 | } |
| 2747 | |
| 2748 | Type *VT = GV.getValueType(); |
| 2749 | if (VT != StructTM.remapType(SrcTy: VT)) { |
| 2750 | // FIXME: Use DiagnosticInfo unsupported but it requires a Function |
| 2751 | Ctx.emitError(ErrorStr: "global variables that contain buffer fat pointers " |
| 2752 | "(address space 7 pointers) are unsupported. Use " |
| 2753 | "buffer resource pointers (address space 8) instead"); |
| 2754 | GV.replaceAllUsesWith(V: PoisonValue::get(T: GV.getType())); |
| 2755 | GV.eraseFromParent(); |
| 2756 | Changed = true; |
| 2757 | continue; |
| 2758 | } |
| 2759 | } |
| 2760 | |
| 2761 | { |
| 2762 | // Collect all constant exprs and aggregates referenced by any function. |
| 2763 | SmallVector<Constant *, 8> Worklist; |
| 2764 | for (Function &F : M.functions()) |
| 2765 | for (Instruction &I : instructions(F)) |
| 2766 | for (Value *Op : I.operands()) |
| 2767 | if (isa<ConstantExpr, ConstantAggregate>(Val: Op)) |
| 2768 | Worklist.push_back(Elt: cast<Constant>(Val: Op)); |
| 2769 | |
| 2770 | // Recursively look for any referenced buffer pointer constants. |
| 2771 | SmallPtrSet<Constant *, 8> Visited; |
| 2772 | SetVector<Constant *> BufferFatPtrConsts; |
| 2773 | while (!Worklist.empty()) { |
| 2774 | Constant *C = Worklist.pop_back_val(); |
| 2775 | if (!Visited.insert(Ptr: C).second) |
| 2776 | continue; |
| 2777 | if (isBufferFatPtrOrVector(Ty: C->getType())) |
| 2778 | BufferFatPtrConsts.insert(X: C); |
| 2779 | for (Value *Op : C->operands()) |
| 2780 | if (isa<ConstantExpr, ConstantAggregate>(Val: Op)) |
| 2781 | Worklist.push_back(Elt: cast<Constant>(Val: Op)); |
| 2782 | } |
| 2783 | |
| 2784 | // Expand all constant expressions using fat buffer pointers to |
| 2785 | // instructions. |
| 2786 | Changed |= convertUsersOfConstantsToInstructions( |
| 2787 | Consts: BufferFatPtrConsts.getArrayRef(), /*RestrictToFunc=*/nullptr, |
| 2788 | /*RemoveDeadConstants=*/false, /*IncludeSelf=*/true); |
| 2789 | } |
| 2790 | |
| 2791 | StoreFatPtrsAsIntsAndExpandMemcpyVisitor MemOpsRewrite(&IntTM, M); |
| 2792 | LegalizeBufferContentTypesVisitor BufferContentsTypeRewrite(M, &TM); |
| 2793 | for (Function &F : M.functions()) { |
| 2794 | bool InterfaceChange = hasFatPointerInterface(F, TypeMap: &StructTM); |
| 2795 | bool BodyChanges = containsBufferFatPointers(F, TypeMap: &StructTM); |
| 2796 | const TargetTransformInfo *TTI = GetTTI(F); |
| 2797 | ScalarEvolution *SE = GetSE(F); |
| 2798 | Changed |= MemOpsRewrite.processFunction(F, TTI, SE); |
| 2799 | if (InterfaceChange || BodyChanges) { |
| 2800 | NeedsRemap.push_back(Elt: std::make_pair(x: &F, y&: InterfaceChange)); |
| 2801 | Changed |= BufferContentsTypeRewrite.processFunction(F, SE); |
| 2802 | } |
| 2803 | } |
| 2804 | if (NeedsRemap.empty()) |
| 2805 | return Changed; |
| 2806 | |
| 2807 | SmallVector<Function *> NeedsPostProcess; |
| 2808 | SmallVector<Function *> Intrinsics; |
| 2809 | // Keep one big map so as to memoize constants across functions. |
| 2810 | ValueToValueMapTy CloneMap; |
| 2811 | FatPtrConstMaterializer Materializer(&StructTM, CloneMap); |
| 2812 | |
| 2813 | ValueMapper LowerInFuncs(CloneMap, RF_None, &StructTM, &Materializer); |
| 2814 | for (auto [F, InterfaceChange] : NeedsRemap) { |
| 2815 | Function *NewF = F; |
| 2816 | if (InterfaceChange) |
| 2817 | NewF = moveFunctionAdaptingType( |
| 2818 | OldF: F, NewTy: cast<FunctionType>(Val: StructTM.remapType(SrcTy: F->getFunctionType())), |
| 2819 | CloneMap); |
| 2820 | else |
| 2821 | makeCloneInPraceMap(F, CloneMap); |
| 2822 | LowerInFuncs.remapFunction(F&: *NewF); |
| 2823 | if (NewF->isIntrinsic()) |
| 2824 | Intrinsics.push_back(Elt: NewF); |
| 2825 | else |
| 2826 | NeedsPostProcess.push_back(Elt: NewF); |
| 2827 | if (InterfaceChange) { |
| 2828 | F->replaceAllUsesWith(V: NewF); |
| 2829 | F->eraseFromParent(); |
| 2830 | } |
| 2831 | Changed = true; |
| 2832 | } |
| 2833 | StructTM.clear(); |
| 2834 | IntTM.clear(); |
| 2835 | CloneMap.clear(); |
| 2836 | |
| 2837 | SplitPtrStructs Splitter(M, &TM); |
| 2838 | for (Function *F : NeedsPostProcess) |
| 2839 | Splitter.processFunction(F&: *F); |
| 2840 | for (Function *F : Intrinsics) { |
| 2841 | // use_empty() can also occur with cases like masked load, which will |
| 2842 | // have been rewritten out of the module by now but not erased. |
| 2843 | if (F->use_empty() || isRemovablePointerIntrinsic(IID: F->getIntrinsicID())) { |
| 2844 | F->eraseFromParent(); |
| 2845 | } else { |
| 2846 | std::optional<Function *> NewF = Intrinsic::remangleIntrinsicFunction(F); |
| 2847 | if (NewF) |
| 2848 | F->replaceAllUsesWith(V: *NewF); |
| 2849 | } |
| 2850 | } |
| 2851 | return Changed; |
| 2852 | } |
| 2853 | |
| 2854 | bool AMDGPULowerBufferFatPointers::runOnModule(Module &M) { |
| 2855 | TargetPassConfig &TPC = getAnalysis<TargetPassConfig>(); |
| 2856 | const TargetMachine &TM = TPC.getTM<TargetMachine>(); |
| 2857 | auto GetTTI = [&](Function &F) -> const TargetTransformInfo * { |
| 2858 | if (F.isDeclaration()) |
| 2859 | return nullptr; |
| 2860 | return &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); |
| 2861 | }; |
| 2862 | auto GetSE = [&](Function &F) -> ScalarEvolution * { |
| 2863 | if (F.isDeclaration()) |
| 2864 | return nullptr; |
| 2865 | return &getAnalysis<ScalarEvolutionWrapperPass>(F).getSE(); |
| 2866 | }; |
| 2867 | return run(M, TM, GetTTI, GetSE); |
| 2868 | } |
| 2869 | |
| 2870 | char AMDGPULowerBufferFatPointers::ID = 0; |
| 2871 | |
| 2872 | char &llvm::AMDGPULowerBufferFatPointersID = AMDGPULowerBufferFatPointers::ID; |
| 2873 | |
| 2874 | void AMDGPULowerBufferFatPointers::getAnalysisUsage(AnalysisUsage &AU) const { |
| 2875 | AU.addRequired<TargetPassConfig>(); |
| 2876 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
| 2877 | AU.addRequired<ScalarEvolutionWrapperPass>(); |
| 2878 | } |
| 2879 | |
| 2880 | #define PASS_DESC "Lower buffer fat pointer operations to buffer resources" |
| 2881 | INITIALIZE_PASS_BEGIN(AMDGPULowerBufferFatPointers, DEBUG_TYPE, PASS_DESC, |
| 2882 | false, false) |
| 2883 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) |
| 2884 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
| 2885 | INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) |
| 2886 | INITIALIZE_PASS_END(AMDGPULowerBufferFatPointers, DEBUG_TYPE, PASS_DESC, false, |
| 2887 | false) |
| 2888 | #undef PASS_DESC |
| 2889 | |
| 2890 | ModulePass *llvm::createAMDGPULowerBufferFatPointersPass() { |
| 2891 | return new AMDGPULowerBufferFatPointers(); |
| 2892 | } |
| 2893 | |
| 2894 | PreservedAnalyses |
| 2895 | AMDGPULowerBufferFatPointersPass::run(Module &M, ModuleAnalysisManager &MA) { |
| 2896 | auto &FA = MA.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 2897 | auto GetTTI = [&](Function &F) -> const TargetTransformInfo * { |
| 2898 | if (F.isDeclaration()) |
| 2899 | return nullptr; |
| 2900 | return &FA.getResult<TargetIRAnalysis>(IR&: F); |
| 2901 | }; |
| 2902 | auto GetSE = [&](Function &F) -> ScalarEvolution * { |
| 2903 | if (F.isDeclaration()) |
| 2904 | return nullptr; |
| 2905 | return &FA.getResult<ScalarEvolutionAnalysis>(IR&: F); |
| 2906 | }; |
| 2907 | return AMDGPULowerBufferFatPointers().run(M, TM, GetTTI, GetSE) |
| 2908 | ? PreservedAnalyses::none() |
| 2909 | : PreservedAnalyses::all(); |
| 2910 | } |
| 2911 |
Generated on 2026-Oct-10 from project llvm_projects revision b1dd93c6a
Powered by 
Code Browser 2.1
Generator usage only permitted with license.