1 //===-- AMDGPULowerIntrinsics.cpp -----------------------------------------===//
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 #include "AMDGPU.h"
11 #include "AMDGPUSubtarget.h"
12 #include "llvm/CodeGen/TargetPassConfig.h"
13 #include "llvm/IR/Constants.h"
14 #include "llvm/IR/Instructions.h"
15 #include "llvm/IR/IntrinsicInst.h"
16 #include "llvm/IR/Module.h"
17 #include "llvm/Transforms/Utils/LowerMemIntrinsics.h"
18 
19 #define DEBUG_TYPE "amdgpu-lower-intrinsics"
20 
21 using namespace llvm;
22 
23 namespace {
24 
25 const unsigned MaxStaticSize = 1024;
26 
27 class AMDGPULowerIntrinsics : public ModulePass {
28 private:
29   bool makeLIDRangeMetadata(Function &F) const;
30 
31 public:
32   static char ID;
33 
34   AMDGPULowerIntrinsics() : ModulePass(ID) {}
35 
36   bool runOnModule(Module &M) override;
37   StringRef getPassName() const override {
38     return "AMDGPU Lower Intrinsics";
39   }
40 };
41 
42 }
43 
44 char AMDGPULowerIntrinsics::ID = 0;
45 
46 char &llvm::AMDGPULowerIntrinsicsID = AMDGPULowerIntrinsics::ID;
47 
48 INITIALIZE_PASS(AMDGPULowerIntrinsics, DEBUG_TYPE, "Lower intrinsics", false,
49                 false)
50 
51 // TODO: Should refine based on estimated number of accesses (e.g. does it
52 // require splitting based on alignment)
53 static bool shouldExpandOperationWithSize(Value *Size) {
54   ConstantInt *CI = dyn_cast<ConstantInt>(Size);
55   return !CI || (CI->getZExtValue() > MaxStaticSize);
56 }
57 
58 static bool expandMemIntrinsicUses(Function &F) {
59   Intrinsic::ID ID = F.getIntrinsicID();
60   bool Changed = false;
61 
62   for (auto I = F.user_begin(), E = F.user_end(); I != E;) {
63     Instruction *Inst = cast<Instruction>(*I);
64     ++I;
65 
66     switch (ID) {
67     case Intrinsic::memcpy: {
68       auto *Memcpy = cast<MemCpyInst>(Inst);
69       if (shouldExpandOperationWithSize(Memcpy->getLength())) {
70         expandMemCpyAsLoop(Memcpy);
71         Changed = true;
72         Memcpy->eraseFromParent();
73       }
74 
75       break;
76     }
77     case Intrinsic::memmove: {
78       auto *Memmove = cast<MemMoveInst>(Inst);
79       if (shouldExpandOperationWithSize(Memmove->getLength())) {
80         expandMemMoveAsLoop(Memmove);
81         Changed = true;
82         Memmove->eraseFromParent();
83       }
84 
85       break;
86     }
87     case Intrinsic::memset: {
88       auto *Memset = cast<MemSetInst>(Inst);
89       if (shouldExpandOperationWithSize(Memset->getLength())) {
90         expandMemSetAsLoop(Memset);
91         Changed = true;
92         Memset->eraseFromParent();
93       }
94 
95       break;
96     }
97     default:
98       break;
99     }
100   }
101 
102   return Changed;
103 }
104 
105 bool AMDGPULowerIntrinsics::makeLIDRangeMetadata(Function &F) const {
106   auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
107   if (!TPC)
108     return false;
109 
110   const TargetMachine &TM = TPC->getTM<TargetMachine>();
111   const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>(F);
112   bool Changed = false;
113 
114   for (auto *U : F.users()) {
115     auto *CI = dyn_cast<CallInst>(U);
116     if (!CI)
117       continue;
118 
119     Changed |= ST.makeLIDRangeMetadata(CI);
120   }
121   return Changed;
122 }
123 
124 bool AMDGPULowerIntrinsics::runOnModule(Module &M) {
125   bool Changed = false;
126 
127   for (Function &F : M) {
128     if (!F.isDeclaration())
129       continue;
130 
131     switch (F.getIntrinsicID()) {
132     case Intrinsic::memcpy:
133     case Intrinsic::memmove:
134     case Intrinsic::memset:
135       if (expandMemIntrinsicUses(F))
136         Changed = true;
137       break;
138 
139     case Intrinsic::amdgcn_workitem_id_x:
140     case Intrinsic::r600_read_tidig_x:
141     case Intrinsic::amdgcn_workitem_id_y:
142     case Intrinsic::r600_read_tidig_y:
143     case Intrinsic::amdgcn_workitem_id_z:
144     case Intrinsic::r600_read_tidig_z:
145     case Intrinsic::r600_read_local_size_x:
146     case Intrinsic::r600_read_local_size_y:
147     case Intrinsic::r600_read_local_size_z:
148       Changed |= makeLIDRangeMetadata(F);
149       break;
150 
151     default:
152       break;
153     }
154   }
155 
156   return Changed;
157 }
158 
159 ModulePass *llvm::createAMDGPULowerIntrinsicsPass() {
160   return new AMDGPULowerIntrinsics();
161 }
162