Matt Arsenault | 7016f13 | 2017-08-03 22:30:46 +0000 | [diff] [blame] | 1 | //==- AMDGPUArgumentrUsageInfo.h - Function Arg Usage Info -------*- C++ -*-==// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //===----------------------------------------------------------------------===// |
| 9 | |
| 10 | #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUARGUMENTUSAGEINFO_H |
| 11 | #define LLVM_LIB_TARGET_AMDGPU_AMDGPUARGUMENTUSAGEINFO_H |
| 12 | |
| 13 | #include "llvm/ADT/DenseMap.h" |
| 14 | #include "llvm/IR/Function.h" |
| 15 | #include "llvm/Pass.h" |
| 16 | |
| 17 | namespace llvm { |
| 18 | |
| 19 | class Function; |
| 20 | class raw_ostream; |
| 21 | class SISubtarget; |
| 22 | class TargetMachine; |
| 23 | class TargetRegisterClass; |
| 24 | class TargetRegisterInfo; |
| 25 | |
| 26 | struct ArgDescriptor { |
| 27 | private: |
| 28 | friend struct AMDGPUFunctionArgInfo; |
Florian Gross | 2feb105 | 2017-08-04 10:53:07 +0000 | [diff] [blame^] | 29 | friend class AMDGPUArgumentUsageInfo; |
Matt Arsenault | 7016f13 | 2017-08-03 22:30:46 +0000 | [diff] [blame] | 30 | |
| 31 | union { |
| 32 | unsigned Register; |
| 33 | unsigned StackOffset; |
| 34 | }; |
| 35 | |
| 36 | bool IsStack : 1; |
| 37 | bool IsSet : 1; |
| 38 | |
| 39 | ArgDescriptor(unsigned Val = 0, bool IsStack = false, bool IsSet = false) |
| 40 | : Register(Val), IsStack(IsStack), IsSet(IsSet) {} |
| 41 | public: |
| 42 | static ArgDescriptor createRegister(unsigned Reg) { |
| 43 | return ArgDescriptor(Reg, false, true); |
| 44 | } |
| 45 | |
| 46 | static ArgDescriptor createStack(unsigned Reg) { |
| 47 | return ArgDescriptor(Reg, true, true); |
| 48 | } |
| 49 | |
| 50 | bool isSet() const { |
| 51 | return IsSet; |
| 52 | } |
| 53 | |
| 54 | explicit operator bool() const { |
| 55 | return isSet(); |
| 56 | } |
| 57 | |
| 58 | bool isRegister() const { |
| 59 | return !IsStack; |
| 60 | } |
| 61 | |
| 62 | unsigned getRegister() const { |
| 63 | assert(!IsStack); |
| 64 | return Register; |
| 65 | } |
| 66 | |
| 67 | unsigned getStackOffset() const { |
| 68 | assert(IsStack); |
| 69 | return StackOffset; |
| 70 | } |
| 71 | |
| 72 | void print(raw_ostream &OS, const TargetRegisterInfo *TRI = nullptr) const; |
| 73 | }; |
| 74 | |
| 75 | inline raw_ostream &operator<<(raw_ostream &OS, const ArgDescriptor &Arg) { |
| 76 | Arg.print(OS); |
| 77 | return OS; |
| 78 | } |
| 79 | |
| 80 | struct AMDGPUFunctionArgInfo { |
| 81 | enum PreloadedValue { |
| 82 | // SGPRS: |
| 83 | PRIVATE_SEGMENT_BUFFER = 0, |
| 84 | DISPATCH_PTR = 1, |
| 85 | QUEUE_PTR = 2, |
| 86 | KERNARG_SEGMENT_PTR = 3, |
| 87 | DISPATCH_ID = 4, |
| 88 | FLAT_SCRATCH_INIT = 5, |
| 89 | WORKGROUP_ID_X = 10, |
| 90 | WORKGROUP_ID_Y = 11, |
| 91 | WORKGROUP_ID_Z = 12, |
| 92 | PRIVATE_SEGMENT_WAVE_BYTE_OFFSET = 14, |
| 93 | IMPLICIT_BUFFER_PTR = 15, |
Matt Arsenault | 817c253 | 2017-08-03 23:12:44 +0000 | [diff] [blame] | 94 | IMPLICIT_ARG_PTR = 16, |
Matt Arsenault | 7016f13 | 2017-08-03 22:30:46 +0000 | [diff] [blame] | 95 | |
| 96 | // VGPRS: |
Matt Arsenault | 817c253 | 2017-08-03 23:12:44 +0000 | [diff] [blame] | 97 | WORKITEM_ID_X = 17, |
| 98 | WORKITEM_ID_Y = 18, |
| 99 | WORKITEM_ID_Z = 19, |
| 100 | FIRST_VGPR_VALUE = WORKITEM_ID_X |
Matt Arsenault | 7016f13 | 2017-08-03 22:30:46 +0000 | [diff] [blame] | 101 | }; |
| 102 | |
| 103 | // Kernel input registers setup for the HSA ABI in allocation order. |
| 104 | |
| 105 | // User SGPRs in kernels |
| 106 | // XXX - Can these require argument spills? |
| 107 | ArgDescriptor PrivateSegmentBuffer; |
| 108 | ArgDescriptor DispatchPtr; |
| 109 | ArgDescriptor QueuePtr; |
| 110 | ArgDescriptor KernargSegmentPtr; |
| 111 | ArgDescriptor DispatchID; |
| 112 | ArgDescriptor FlatScratchInit; |
| 113 | ArgDescriptor PrivateSegmentSize; |
| 114 | ArgDescriptor GridWorkGroupCountX; |
| 115 | ArgDescriptor GridWorkGroupCountY; |
| 116 | ArgDescriptor GridWorkGroupCountZ; |
| 117 | |
| 118 | // System SGPRs in kernels. |
| 119 | ArgDescriptor WorkGroupIDX; |
| 120 | ArgDescriptor WorkGroupIDY; |
| 121 | ArgDescriptor WorkGroupIDZ; |
| 122 | ArgDescriptor WorkGroupInfo; |
| 123 | ArgDescriptor PrivateSegmentWaveByteOffset; |
| 124 | |
Matt Arsenault | 817c253 | 2017-08-03 23:12:44 +0000 | [diff] [blame] | 125 | // Pointer with offset from kernargsegmentptr to where special ABI arguments |
| 126 | // are passed to callable functions. |
| 127 | ArgDescriptor ImplicitArgPtr; |
| 128 | |
Matt Arsenault | 7016f13 | 2017-08-03 22:30:46 +0000 | [diff] [blame] | 129 | // Input registers for non-HSA ABI |
| 130 | ArgDescriptor ImplicitBufferPtr = 0; |
| 131 | |
| 132 | // VGPRs inputs. These are always v0, v1 and v2 for entry functions. |
| 133 | ArgDescriptor WorkItemIDX; |
| 134 | ArgDescriptor WorkItemIDY; |
| 135 | ArgDescriptor WorkItemIDZ; |
| 136 | |
| 137 | std::pair<const ArgDescriptor *, const TargetRegisterClass *> |
| 138 | getPreloadedValue(PreloadedValue Value) const; |
| 139 | }; |
| 140 | |
| 141 | class AMDGPUArgumentUsageInfo : public ImmutablePass { |
| 142 | private: |
| 143 | static const AMDGPUFunctionArgInfo ExternFunctionInfo; |
| 144 | DenseMap<const Function *, AMDGPUFunctionArgInfo> ArgInfoMap; |
| 145 | |
| 146 | public: |
| 147 | static char ID; |
| 148 | |
| 149 | AMDGPUArgumentUsageInfo() : ImmutablePass(ID) { } |
| 150 | |
| 151 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 152 | AU.setPreservesAll(); |
| 153 | } |
| 154 | |
| 155 | bool doInitialization(Module &M) override; |
| 156 | bool doFinalization(Module &M) override; |
| 157 | |
| 158 | void print(raw_ostream &OS, const Module *M = nullptr) const override; |
| 159 | |
| 160 | void setFuncArgInfo(const Function &F, const AMDGPUFunctionArgInfo &ArgInfo) { |
| 161 | ArgInfoMap[&F] = ArgInfo; |
| 162 | } |
| 163 | |
| 164 | const AMDGPUFunctionArgInfo &lookupFuncArgInfo(const Function &F) const { |
| 165 | auto I = ArgInfoMap.find(&F); |
| 166 | if (I == ArgInfoMap.end()) { |
| 167 | assert(F.isDeclaration()); |
| 168 | return ExternFunctionInfo; |
| 169 | } |
| 170 | |
| 171 | return I->second; |
| 172 | } |
| 173 | }; |
| 174 | |
| 175 | } // end namespace llvm |
| 176 | |
| 177 | #endif |