1 //===-- NVPTXLowerKernelArgs.cpp - Lower kernel arguments -----------------===//
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 // Pointer arguments to kernel functions need to be lowered specially.
11 //
12 // 1. Copy byval struct args to local memory. This is a preparation for handling
13 // cases like
14 //
15 // kernel void foo(struct A arg, ...)
16 // {
17 // struct A *p = &arg;
18 // ...
19 // ... = p->filed1 ... (this is no generic address for .param)
20 // p->filed2 = ... (this is no write access to .param)
21 // }
22 //
23 // 2. Convert non-byval pointer arguments of CUDA kernels to pointers in the
24 // global address space. This allows later optimizations to emit
25 // ld.global.*/st.global.* for accessing these pointer arguments. For
26 // example,
27 //
28 // define void @foo(float* %input) {
29 // %v = load float, float* %input, align 4
30 // ...
31 // }
32 //
33 // becomes
34 //
35 // define void @foo(float* %input) {
36 // %input2 = addrspacecast float* %input to float addrspace(1)*
37 // %input3 = addrspacecast float addrspace(1)* %input2 to float*
38 // %v = load float, float* %input3, align 4
39 // ...
40 // }
41 //
42 // Later, NVPTXFavorNonGenericAddrSpaces will optimize it to
43 //
44 // define void @foo(float* %input) {
45 // %input2 = addrspacecast float* %input to float addrspace(1)*
46 // %v = load float, float addrspace(1)* %input2, align 4
47 // ...
48 // }
49 //
50 // TODO: merge this pass with NVPTXFavorNonGenericAddrSpace so that other passes
51 // don't cancel the addrspacecast pair this pass emits.
52 //===----------------------------------------------------------------------===//
53
54 #include "NVPTX.h"
55 #include "NVPTXUtilities.h"
56 #include "NVPTXTargetMachine.h"
57 #include "llvm/IR/Function.h"
58 #include "llvm/IR/Instructions.h"
59 #include "llvm/IR/Module.h"
60 #include "llvm/IR/Type.h"
61 #include "llvm/Pass.h"
62
63 using namespace llvm;
64
65 namespace llvm {
66 void initializeNVPTXLowerKernelArgsPass(PassRegistry &);
67 }
68
69 namespace {
70 class NVPTXLowerKernelArgs : public FunctionPass {
71 bool runOnFunction(Function &F) override;
72
73 // handle byval parameters
74 void handleByValParam(Argument *);
75 // handle non-byval pointer parameters
76 void handlePointerParam(Argument *);
77
78 public:
79 static char ID; // Pass identification, replacement for typeid
NVPTXLowerKernelArgs(const NVPTXTargetMachine * TM=nullptr)80 NVPTXLowerKernelArgs(const NVPTXTargetMachine *TM = nullptr)
81 : FunctionPass(ID), TM(TM) {}
getPassName() const82 const char *getPassName() const override {
83 return "Lower pointer arguments of CUDA kernels";
84 }
85
86 private:
87 const NVPTXTargetMachine *TM;
88 };
89 } // namespace
90
91 char NVPTXLowerKernelArgs::ID = 1;
92
93 INITIALIZE_PASS(NVPTXLowerKernelArgs, "nvptx-lower-kernel-args",
94 "Lower kernel arguments (NVPTX)", false, false)
95
96 // =============================================================================
97 // If the function had a byval struct ptr arg, say foo(%struct.x *byval %d),
98 // then add the following instructions to the first basic block:
99 //
100 // %temp = alloca %struct.x, align 8
101 // %tempd = addrspacecast %struct.x* %d to %struct.x addrspace(101)*
102 // %tv = load %struct.x addrspace(101)* %tempd
103 // store %struct.x %tv, %struct.x* %temp, align 8
104 //
105 // The above code allocates some space in the stack and copies the incoming
106 // struct from param space to local space.
107 // Then replace all occurences of %d by %temp.
108 // =============================================================================
handleByValParam(Argument * Arg)109 void NVPTXLowerKernelArgs::handleByValParam(Argument *Arg) {
110 Function *Func = Arg->getParent();
111 Instruction *FirstInst = &(Func->getEntryBlock().front());
112 PointerType *PType = dyn_cast<PointerType>(Arg->getType());
113
114 assert(PType && "Expecting pointer type in handleByValParam");
115
116 Type *StructType = PType->getElementType();
117 AllocaInst *AllocA = new AllocaInst(StructType, Arg->getName(), FirstInst);
118 // Set the alignment to alignment of the byval parameter. This is because,
119 // later load/stores assume that alignment, and we are going to replace
120 // the use of the byval parameter with this alloca instruction.
121 AllocA->setAlignment(Func->getParamAlignment(Arg->getArgNo() + 1));
122 Arg->replaceAllUsesWith(AllocA);
123
124 Value *ArgInParam = new AddrSpaceCastInst(
125 Arg, PointerType::get(StructType, ADDRESS_SPACE_PARAM), Arg->getName(),
126 FirstInst);
127 LoadInst *LI = new LoadInst(ArgInParam, Arg->getName(), FirstInst);
128 new StoreInst(LI, AllocA, FirstInst);
129 }
130
handlePointerParam(Argument * Arg)131 void NVPTXLowerKernelArgs::handlePointerParam(Argument *Arg) {
132 assert(!Arg->hasByValAttr() &&
133 "byval params should be handled by handleByValParam");
134
135 // Do nothing if the argument already points to the global address space.
136 if (Arg->getType()->getPointerAddressSpace() == ADDRESS_SPACE_GLOBAL)
137 return;
138
139 Instruction *FirstInst = Arg->getParent()->getEntryBlock().begin();
140 Instruction *ArgInGlobal = new AddrSpaceCastInst(
141 Arg, PointerType::get(Arg->getType()->getPointerElementType(),
142 ADDRESS_SPACE_GLOBAL),
143 Arg->getName(), FirstInst);
144 Value *ArgInGeneric = new AddrSpaceCastInst(ArgInGlobal, Arg->getType(),
145 Arg->getName(), FirstInst);
146 // Replace with ArgInGeneric all uses of Args except ArgInGlobal.
147 Arg->replaceAllUsesWith(ArgInGeneric);
148 ArgInGlobal->setOperand(0, Arg);
149 }
150
151
152 // =============================================================================
153 // Main function for this pass.
154 // =============================================================================
runOnFunction(Function & F)155 bool NVPTXLowerKernelArgs::runOnFunction(Function &F) {
156 // Skip non-kernels. See the comments at the top of this file.
157 if (!isKernelFunction(F))
158 return false;
159
160 for (Argument &Arg : F.args()) {
161 if (Arg.getType()->isPointerTy()) {
162 if (Arg.hasByValAttr())
163 handleByValParam(&Arg);
164 else if (TM && TM->getDrvInterface() == NVPTX::CUDA)
165 handlePointerParam(&Arg);
166 }
167 }
168 return true;
169 }
170
171 FunctionPass *
createNVPTXLowerKernelArgsPass(const NVPTXTargetMachine * TM)172 llvm::createNVPTXLowerKernelArgsPass(const NVPTXTargetMachine *TM) {
173 return new NVPTXLowerKernelArgs(TM);
174 }
175