1 | //===- AMDGPUAnnotateKernelFeaturesPass.cpp -------------------------------===// |
---|---|
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | /// \file This pass propagates the uniform-work-group-size attribute from |
10 | /// kernels to leaf functions when possible. It also adds additional attributes |
11 | /// to hint ABI lowering optimizations later. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "AMDGPU.h" |
16 | #include "GCNSubtarget.h" |
17 | #include "llvm/Analysis/CallGraph.h" |
18 | #include "llvm/Analysis/CallGraphSCCPass.h" |
19 | #include "llvm/CodeGen/TargetPassConfig.h" |
20 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
21 | #include "llvm/IR/IntrinsicsR600.h" |
22 | #include "llvm/Target/TargetMachine.h" |
23 | |
24 | #define DEBUG_TYPE "amdgpu-annotate-kernel-features" |
25 | |
26 | using namespace llvm; |
27 | |
28 | namespace { |
29 | class AMDGPUAnnotateKernelFeatures : public CallGraphSCCPass { |
30 | private: |
31 | const TargetMachine *TM = nullptr; |
32 | |
33 | bool addFeatureAttributes(Function &F); |
34 | |
35 | public: |
36 | static char ID; |
37 | |
38 | AMDGPUAnnotateKernelFeatures() : CallGraphSCCPass(ID) {} |
39 | |
40 | bool doInitialization(CallGraph &CG) override; |
41 | bool runOnSCC(CallGraphSCC &SCC) override; |
42 | |
43 | StringRef getPassName() const override { |
44 | return "AMDGPU Annotate Kernel Features"; |
45 | } |
46 | |
47 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
48 | AU.setPreservesAll(); |
49 | CallGraphSCCPass::getAnalysisUsage(Info&: AU); |
50 | } |
51 | }; |
52 | |
53 | } // end anonymous namespace |
54 | |
55 | char AMDGPUAnnotateKernelFeatures::ID = 0; |
56 | |
57 | char &llvm::AMDGPUAnnotateKernelFeaturesID = AMDGPUAnnotateKernelFeatures::ID; |
58 | |
59 | INITIALIZE_PASS(AMDGPUAnnotateKernelFeatures, DEBUG_TYPE, |
60 | "Add AMDGPU function attributes", false, false) |
61 | |
62 | bool AMDGPUAnnotateKernelFeatures::addFeatureAttributes(Function &F) { |
63 | bool HaveStackObjects = false; |
64 | bool Changed = false; |
65 | bool HaveCall = false; |
66 | bool IsFunc = !AMDGPU::isEntryFunctionCC(CC: F.getCallingConv()); |
67 | |
68 | for (BasicBlock &BB : F) { |
69 | for (Instruction &I : BB) { |
70 | if (isa<AllocaInst>(Val: I)) { |
71 | HaveStackObjects = true; |
72 | continue; |
73 | } |
74 | |
75 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
76 | const Function *Callee = |
77 | dyn_cast<Function>(Val: CB->getCalledOperand()->stripPointerCasts()); |
78 | |
79 | // Note the occurrence of indirect call. |
80 | if (!Callee) { |
81 | if (!CB->isInlineAsm()) |
82 | HaveCall = true; |
83 | |
84 | continue; |
85 | } |
86 | |
87 | Intrinsic::ID IID = Callee->getIntrinsicID(); |
88 | if (IID == Intrinsic::not_intrinsic) { |
89 | HaveCall = true; |
90 | Changed = true; |
91 | } |
92 | } |
93 | } |
94 | } |
95 | |
96 | // TODO: We could refine this to captured pointers that could possibly be |
97 | // accessed by flat instructions. For now this is mostly a poor way of |
98 | // estimating whether there are calls before argument lowering. |
99 | if (!IsFunc && HaveCall) { |
100 | F.addFnAttr(Kind: "amdgpu-calls"); |
101 | Changed = true; |
102 | } |
103 | |
104 | if (HaveStackObjects) { |
105 | F.addFnAttr(Kind: "amdgpu-stack-objects"); |
106 | Changed = true; |
107 | } |
108 | |
109 | return Changed; |
110 | } |
111 | |
112 | bool AMDGPUAnnotateKernelFeatures::runOnSCC(CallGraphSCC &SCC) { |
113 | bool Changed = false; |
114 | |
115 | for (CallGraphNode *I : SCC) { |
116 | Function *F = I->getFunction(); |
117 | // Ignore functions with graphics calling conventions, these are currently |
118 | // not allowed to have kernel arguments. |
119 | if (!F || F->isDeclaration() || AMDGPU::isGraphics(CC: F->getCallingConv())) |
120 | continue; |
121 | // Add feature attributes |
122 | Changed |= addFeatureAttributes(F&: *F); |
123 | } |
124 | |
125 | return Changed; |
126 | } |
127 | |
128 | bool AMDGPUAnnotateKernelFeatures::doInitialization(CallGraph &CG) { |
129 | auto *TPC = getAnalysisIfAvailable<TargetPassConfig>(); |
130 | if (!TPC) |
131 | report_fatal_error(reason: "TargetMachine is required"); |
132 | |
133 | TM = &TPC->getTM<TargetMachine>(); |
134 | return false; |
135 | } |
136 | |
137 | Pass *llvm::createAMDGPUAnnotateKernelFeaturesPass() { |
138 | return new AMDGPUAnnotateKernelFeatures(); |
139 | } |
140 |