1 | //===----------------------------------------------------------------------===// |
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 | #include "AMDGPUArgumentUsageInfo.h" |
10 | #include "AMDGPU.h" |
11 | #include "AMDGPUTargetMachine.h" |
12 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
13 | #include "SIRegisterInfo.h" |
14 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
15 | #include "llvm/IR/Function.h" |
16 | #include "llvm/Support/NativeFormatting.h" |
17 | #include "llvm/Support/raw_ostream.h" |
18 | |
19 | using namespace llvm; |
20 | |
21 | #define DEBUG_TYPE "amdgpu-argument-reg-usage-info" |
22 | |
23 | INITIALIZE_PASS(AMDGPUArgumentUsageInfo, DEBUG_TYPE, |
24 | "Argument Register Usage Information Storage" , false, true) |
25 | |
26 | void ArgDescriptor::print(raw_ostream &OS, |
27 | const TargetRegisterInfo *TRI) const { |
28 | if (!isSet()) { |
29 | OS << "<not set>\n" ; |
30 | return; |
31 | } |
32 | |
33 | if (isRegister()) |
34 | OS << "Reg " << printReg(Reg: getRegister(), TRI); |
35 | else |
36 | OS << "Stack offset " << getStackOffset(); |
37 | |
38 | if (isMasked()) { |
39 | OS << " & " ; |
40 | llvm::write_hex(S&: OS, N: Mask, Style: llvm::HexPrintStyle::PrefixLower); |
41 | } |
42 | |
43 | OS << '\n'; |
44 | } |
45 | |
46 | char AMDGPUArgumentUsageInfo::ID = 0; |
47 | |
48 | const AMDGPUFunctionArgInfo AMDGPUArgumentUsageInfo::ExternFunctionInfo{}; |
49 | |
50 | // Hardcoded registers from fixed function ABI |
51 | const AMDGPUFunctionArgInfo AMDGPUArgumentUsageInfo::FixedABIFunctionInfo |
52 | = AMDGPUFunctionArgInfo::fixedABILayout(); |
53 | |
54 | bool AMDGPUArgumentUsageInfo::doInitialization(Module &M) { |
55 | return false; |
56 | } |
57 | |
58 | bool AMDGPUArgumentUsageInfo::doFinalization(Module &M) { |
59 | ArgInfoMap.clear(); |
60 | return false; |
61 | } |
62 | |
63 | // TODO: Print preload kernargs? |
64 | void AMDGPUArgumentUsageInfo::print(raw_ostream &OS, const Module *M) const { |
65 | for (const auto &FI : ArgInfoMap) { |
66 | OS << "Arguments for " << FI.first->getName() << '\n' |
67 | << " PrivateSegmentBuffer: " << FI.second.PrivateSegmentBuffer |
68 | << " DispatchPtr: " << FI.second.DispatchPtr |
69 | << " QueuePtr: " << FI.second.QueuePtr |
70 | << " KernargSegmentPtr: " << FI.second.KernargSegmentPtr |
71 | << " DispatchID: " << FI.second.DispatchID |
72 | << " FlatScratchInit: " << FI.second.FlatScratchInit |
73 | << " PrivateSegmentSize: " << FI.second.PrivateSegmentSize |
74 | << " WorkGroupIDX: " << FI.second.WorkGroupIDX |
75 | << " WorkGroupIDY: " << FI.second.WorkGroupIDY |
76 | << " WorkGroupIDZ: " << FI.second.WorkGroupIDZ |
77 | << " WorkGroupInfo: " << FI.second.WorkGroupInfo |
78 | << " LDSKernelId: " << FI.second.LDSKernelId |
79 | << " PrivateSegmentWaveByteOffset: " |
80 | << FI.second.PrivateSegmentWaveByteOffset |
81 | << " ImplicitBufferPtr: " << FI.second.ImplicitBufferPtr |
82 | << " ImplicitArgPtr: " << FI.second.ImplicitArgPtr |
83 | << " WorkItemIDX " << FI.second.WorkItemIDX |
84 | << " WorkItemIDY " << FI.second.WorkItemIDY |
85 | << " WorkItemIDZ " << FI.second.WorkItemIDZ |
86 | << '\n'; |
87 | } |
88 | } |
89 | |
90 | std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT> |
91 | AMDGPUFunctionArgInfo::getPreloadedValue( |
92 | AMDGPUFunctionArgInfo::PreloadedValue Value) const { |
93 | switch (Value) { |
94 | case AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_BUFFER: { |
95 | return std::tuple(PrivateSegmentBuffer ? &PrivateSegmentBuffer : nullptr, |
96 | &AMDGPU::SGPR_128RegClass, LLT::fixed_vector(NumElements: 4, ScalarSizeInBits: 32)); |
97 | } |
98 | case AMDGPUFunctionArgInfo::IMPLICIT_BUFFER_PTR: |
99 | return std::tuple(ImplicitBufferPtr ? &ImplicitBufferPtr : nullptr, |
100 | &AMDGPU::SGPR_64RegClass, |
101 | LLT::pointer(AddressSpace: AMDGPUAS::CONSTANT_ADDRESS, SizeInBits: 64)); |
102 | case AMDGPUFunctionArgInfo::WORKGROUP_ID_X: |
103 | return std::tuple(WorkGroupIDX ? &WorkGroupIDX : nullptr, |
104 | &AMDGPU::SGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
105 | case AMDGPUFunctionArgInfo::WORKGROUP_ID_Y: |
106 | return std::tuple(WorkGroupIDY ? &WorkGroupIDY : nullptr, |
107 | &AMDGPU::SGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
108 | case AMDGPUFunctionArgInfo::WORKGROUP_ID_Z: |
109 | return std::tuple(WorkGroupIDZ ? &WorkGroupIDZ : nullptr, |
110 | &AMDGPU::SGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
111 | case AMDGPUFunctionArgInfo::LDS_KERNEL_ID: |
112 | return std::tuple(LDSKernelId ? &LDSKernelId : nullptr, |
113 | &AMDGPU::SGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
114 | case AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_WAVE_BYTE_OFFSET: |
115 | return std::tuple( |
116 | PrivateSegmentWaveByteOffset ? &PrivateSegmentWaveByteOffset : nullptr, |
117 | &AMDGPU::SGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
118 | case AMDGPUFunctionArgInfo::PRIVATE_SEGMENT_SIZE: |
119 | return {PrivateSegmentSize ? &PrivateSegmentSize : nullptr, |
120 | &AMDGPU::SGPR_32RegClass, LLT::scalar(SizeInBits: 32)}; |
121 | case AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR: |
122 | return std::tuple(KernargSegmentPtr ? &KernargSegmentPtr : nullptr, |
123 | &AMDGPU::SGPR_64RegClass, |
124 | LLT::pointer(AddressSpace: AMDGPUAS::CONSTANT_ADDRESS, SizeInBits: 64)); |
125 | case AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR: |
126 | return std::tuple(ImplicitArgPtr ? &ImplicitArgPtr : nullptr, |
127 | &AMDGPU::SGPR_64RegClass, |
128 | LLT::pointer(AddressSpace: AMDGPUAS::CONSTANT_ADDRESS, SizeInBits: 64)); |
129 | case AMDGPUFunctionArgInfo::DISPATCH_ID: |
130 | return std::tuple(DispatchID ? &DispatchID : nullptr, |
131 | &AMDGPU::SGPR_64RegClass, LLT::scalar(SizeInBits: 64)); |
132 | case AMDGPUFunctionArgInfo::FLAT_SCRATCH_INIT: |
133 | return std::tuple(FlatScratchInit ? &FlatScratchInit : nullptr, |
134 | &AMDGPU::SGPR_64RegClass, LLT::scalar(SizeInBits: 64)); |
135 | case AMDGPUFunctionArgInfo::DISPATCH_PTR: |
136 | return std::tuple(DispatchPtr ? &DispatchPtr : nullptr, |
137 | &AMDGPU::SGPR_64RegClass, |
138 | LLT::pointer(AddressSpace: AMDGPUAS::CONSTANT_ADDRESS, SizeInBits: 64)); |
139 | case AMDGPUFunctionArgInfo::QUEUE_PTR: |
140 | return std::tuple(QueuePtr ? &QueuePtr : nullptr, &AMDGPU::SGPR_64RegClass, |
141 | LLT::pointer(AddressSpace: AMDGPUAS::CONSTANT_ADDRESS, SizeInBits: 64)); |
142 | case AMDGPUFunctionArgInfo::WORKITEM_ID_X: |
143 | return std::tuple(WorkItemIDX ? &WorkItemIDX : nullptr, |
144 | &AMDGPU::VGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
145 | case AMDGPUFunctionArgInfo::WORKITEM_ID_Y: |
146 | return std::tuple(WorkItemIDY ? &WorkItemIDY : nullptr, |
147 | &AMDGPU::VGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
148 | case AMDGPUFunctionArgInfo::WORKITEM_ID_Z: |
149 | return std::tuple(WorkItemIDZ ? &WorkItemIDZ : nullptr, |
150 | &AMDGPU::VGPR_32RegClass, LLT::scalar(SizeInBits: 32)); |
151 | } |
152 | llvm_unreachable("unexpected preloaded value type" ); |
153 | } |
154 | |
155 | AMDGPUFunctionArgInfo AMDGPUFunctionArgInfo::fixedABILayout() { |
156 | AMDGPUFunctionArgInfo AI; |
157 | AI.PrivateSegmentBuffer |
158 | = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3); |
159 | AI.DispatchPtr = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR4_SGPR5); |
160 | AI.QueuePtr = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR6_SGPR7); |
161 | |
162 | // Do not pass kernarg segment pointer, only pass increment version in its |
163 | // place. |
164 | AI.ImplicitArgPtr = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR8_SGPR9); |
165 | AI.DispatchID = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR10_SGPR11); |
166 | |
167 | // Skip FlatScratchInit/PrivateSegmentSize |
168 | AI.WorkGroupIDX = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR12); |
169 | AI.WorkGroupIDY = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR13); |
170 | AI.WorkGroupIDZ = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR14); |
171 | AI.LDSKernelId = ArgDescriptor::createRegister(Reg: AMDGPU::SGPR15); |
172 | |
173 | const unsigned Mask = 0x3ff; |
174 | AI.WorkItemIDX = ArgDescriptor::createRegister(Reg: AMDGPU::VGPR31, Mask); |
175 | AI.WorkItemIDY = ArgDescriptor::createRegister(Reg: AMDGPU::VGPR31, Mask: Mask << 10); |
176 | AI.WorkItemIDZ = ArgDescriptor::createRegister(Reg: AMDGPU::VGPR31, Mask: Mask << 20); |
177 | return AI; |
178 | } |
179 | |
180 | const AMDGPUFunctionArgInfo & |
181 | AMDGPUArgumentUsageInfo::lookupFuncArgInfo(const Function &F) const { |
182 | auto I = ArgInfoMap.find(Val: &F); |
183 | if (I == ArgInfoMap.end()) |
184 | return FixedABIFunctionInfo; |
185 | return I->second; |
186 | } |
187 | |