| 1 | //===- AMDGPUMemoryUtils.h - Memory related helper functions -*- C++ -*----===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #ifndef LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUMEMORYUTILS_H |
| 10 | #define LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUMEMORYUTILS_H |
| 11 | |
| 12 | #include "llvm/ADT/ArrayRef.h" |
| 13 | #include "llvm/ADT/BitmaskEnum.h" |
| 14 | #include "llvm/ADT/DenseMap.h" |
| 15 | #include "llvm/ADT/DenseSet.h" |
| 16 | #include "llvm/ADT/MapVector.h" |
| 17 | #include "llvm/IR/DebugLoc.h" |
| 18 | |
| 19 | namespace llvm { |
| 20 | |
| 21 | struct Align; |
| 22 | class AAResults; |
| 23 | class DataLayout; |
| 24 | class GlobalVariable; |
| 25 | class Instruction; |
| 26 | class LoadInst; |
| 27 | class MemoryDef; |
| 28 | class MemorySSA; |
| 29 | class MMRAMetadata; |
| 30 | class Value; |
| 31 | class Function; |
| 32 | class CallGraph; |
| 33 | class Module; |
| 34 | class TargetExtType; |
| 35 | |
| 36 | namespace AMDGPU { |
| 37 | |
| 38 | static constexpr unsigned NamedBarrierTypeSizeInBytes = 1; |
| 39 | |
| 40 | using FunctionVariableMap = MapVector<Function *, DenseSet<GlobalVariable *>>; |
| 41 | using VariableFunctionMap = DenseMap<GlobalVariable *, DenseSet<Function *>>; |
| 42 | |
| 43 | Align getAlign(const DataLayout &DL, const GlobalVariable *GV); |
| 44 | |
| 45 | // Get the synthetic aperture number for the given address space, or None (0) |
| 46 | // if the address space does not have one. |
| 47 | unsigned getSyntheticApertureNumber(unsigned AS); |
| 48 | |
| 49 | // Copy metadata onto a load widened to read a superset of Source's bytes. Only |
| 50 | // value-independent metadata is copied; metadata describing the loaded value |
| 51 | // (!range, !noundef, !nofpclass, !tbaa, ...) is dropped. |
| 52 | void copyMetadataForWidenedLoad(LoadInst &Dest, const LoadInst &Source); |
| 53 | |
| 54 | // If GV is a named-barrier return its type. Otherwise return nullptr. |
| 55 | TargetExtType *isNamedBarrier(const GlobalVariable &GV); |
| 56 | |
| 57 | /// \returns how many named barriers are declared by \p GV. |
| 58 | unsigned getNumNamedBarriersDeclared(const DataLayout &DL, |
| 59 | const GlobalVariable &GV); |
| 60 | |
| 61 | bool isDynamicLDS(const GlobalVariable &GV); |
| 62 | bool isLDSVariableToLower(const GlobalVariable &GV); |
| 63 | |
| 64 | struct GVUsesInfoTy { |
| 65 | FunctionVariableMap DirectAccess; |
| 66 | FunctionVariableMap IndirectAccess; |
| 67 | }; |
| 68 | |
| 69 | /// Iterates over all GlobalVariables in \p M, and whenever \p Filter returns |
| 70 | /// true, replace all constant users of the GV with instructions. |
| 71 | bool eliminateGVConstantExprUsesFromAllInstructions( |
| 72 | Module &M, function_ref<bool(const GlobalVariable &)> Filter); |
| 73 | |
| 74 | /// Finds uses of Global Variables on a per-function basis. |
| 75 | /// \param CG \p M Call Graph |
| 76 | /// \param M Module |
| 77 | /// \param Filter Function that returns true for GVs that need to be considered. |
| 78 | /// \param Kernels[out] Maps kernels to global variables used by that kernel. |
| 79 | /// \param Functions[out] Maps functions to global variables used by that |
| 80 | /// function. |
| 81 | void getUsesOfGVByFunction(const CallGraph &CG, Module &M, |
| 82 | function_ref<bool(const GlobalVariable &)> Filter, |
| 83 | FunctionVariableMap &Kernels, |
| 84 | FunctionVariableMap &Functions); |
| 85 | |
| 86 | /// Collects all uses of Global Variables in \p M using |
| 87 | /// \ref getUsesOfGVByFunction. |
| 88 | /// \param CG \p M Call Graph |
| 89 | /// \param M Module |
| 90 | /// \param Filter Filter for \ref getUsesOfGVByFunction - only GVs for which the |
| 91 | /// filter returns true will be considered. |
| 92 | /// \returns Uses of GVs that were found within each function, sorted by |
| 93 | /// direct and indirect accesses. |
| 94 | GVUsesInfoTy |
| 95 | getTransitiveUsesOfGV(const CallGraph &CG, Module &M, |
| 96 | function_ref<bool(const GlobalVariable &)> Filter); |
| 97 | |
| 98 | /// Collects all uses of LDS Global Variables in \p M using |
| 99 | /// \ref getUsesOfGVByFunction, with \ref isLDSVariableToLower as the filter. |
| 100 | /// \param CG \p M Call Graph |
| 101 | /// \param M Module |
| 102 | /// \returns Uses of LDS GVs that need lowering that were found within each |
| 103 | /// function, sorted by direct and indirect accesses. |
| 104 | GVUsesInfoTy getTransitiveUsesOfLDSForLowering(const CallGraph &CG, Module &M); |
| 105 | |
| 106 | /// Strip FnAttr attribute from any functions where we may have |
| 107 | /// introduced its use. |
| 108 | void removeFnAttrFromReachable(CallGraph &CG, Function *KernelRoot, |
| 109 | ArrayRef<StringRef> FnAttrs); |
| 110 | |
| 111 | /// Given a \p Def clobbering a load from \p Ptr according to the MSSA check |
| 112 | /// if this is actually a memory update or an artificial clobber to facilitate |
| 113 | /// ordering constraints. |
| 114 | bool isReallyAClobber(const Value *Ptr, MemoryDef *Def, AAResults *AA); |
| 115 | |
| 116 | /// Check is a \p Load is clobbered in its function. |
| 117 | bool isClobberedInFunction(const LoadInst *Load, MemorySSA *MSSA, |
| 118 | AAResults *AA); |
| 119 | |
| 120 | /// The distinct address spaces supported by the AMDGPU target for |
| 121 | /// atomic memory operation. Can be ORed together. |
| 122 | enum class SIAtomicAddrSpace { |
| 123 | NONE = 0u, |
| 124 | GLOBAL = 1u << 0, |
| 125 | LDS = 1u << 1, |
| 126 | SCRATCH = 1u << 2, |
| 127 | GDS = 1u << 3, |
| 128 | OTHER = 1u << 4, |
| 129 | |
| 130 | /// The address spaces that can be accessed by a FLAT instruction. |
| 131 | FLAT = GLOBAL | LDS | SCRATCH, |
| 132 | |
| 133 | /// The address spaces that support atomic instructions. |
| 134 | ATOMIC = GLOBAL | LDS | SCRATCH | GDS, |
| 135 | |
| 136 | /// All address spaces. |
| 137 | ALL = GLOBAL | LDS | SCRATCH | GDS | OTHER, |
| 138 | |
| 139 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ ALL) |
| 140 | }; |
| 141 | |
| 142 | /// \return Return a bit set of the address spaces accessed by \p AS. |
| 143 | SIAtomicAddrSpace toSIAtomicAddrSpace(unsigned AS); |
| 144 | |
| 145 | /// Reads \p I's MMRAs to parse the "amdgpu-synchronize-as" MMRA. |
| 146 | /// If this tag isn't present, or if it has no meaningful values, returns |
| 147 | /// \p ATOMIC, otherwise returns the address spaces specified by the MD. |
| 148 | /// Does not diagnose unexpected MMRA tag suffixes. |
| 149 | SIAtomicAddrSpace getSynchronizeAddrSpace(const Instruction &I); |
| 150 | |
| 151 | /// Parses the values for the "amdgpu-synchronize-as" MMRA. |
| 152 | /// If this tag isn't present, or if it has no meaningful values, returns |
| 153 | /// \p ATOMIC, otherwise returns the address spaces specified by the MD. |
| 154 | /// If \p DiagFn is non-null, unexpected suffixes for the MMRA tag are reported |
| 155 | /// as warnings to \p DiagFn's context, located at \p DL. |
| 156 | SIAtomicAddrSpace getSynchronizeAddrSpace(const MMRAMetadata &MMRA, |
| 157 | const Function *DiagFn = nullptr, |
| 158 | const DebugLoc &DL = {}); |
| 159 | |
| 160 | } // end namespace AMDGPU |
| 161 | |
| 162 | } // end namespace llvm |
| 163 | |
| 164 | #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUMEMORYUTILS_H |
| 165 | |