1 | //===------ CGGPUBuiltin.cpp - Codegen for GPU builtins -------------------===// |
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 | // Generates code for built-in GPU calls which are not runtime-specific. |
10 | // (Runtime-specific codegen lives in programming model specific files.) |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "CodeGenFunction.h" |
15 | #include "clang/Basic/Builtins.h" |
16 | #include "llvm/IR/DataLayout.h" |
17 | #include "llvm/IR/Instruction.h" |
18 | #include "llvm/Support/MathExtras.h" |
19 | #include "llvm/Transforms/Utils/AMDGPUEmitPrintf.h" |
20 | |
21 | using namespace clang; |
22 | using namespace CodeGen; |
23 | |
24 | namespace { |
25 | llvm::Function *GetVprintfDeclaration(llvm::Module &M) { |
26 | llvm::Type *ArgTypes[] = {llvm::PointerType::getUnqual(C&: M.getContext()), |
27 | llvm::PointerType::getUnqual(C&: M.getContext())}; |
28 | llvm::FunctionType *VprintfFuncType = llvm::FunctionType::get( |
29 | Result: llvm::Type::getInt32Ty(C&: M.getContext()), Params: ArgTypes, isVarArg: false); |
30 | |
31 | if (auto *F = M.getFunction(Name: "vprintf" )) { |
32 | // Our CUDA system header declares vprintf with the right signature, so |
33 | // nobody else should have been able to declare vprintf with a bogus |
34 | // signature. |
35 | assert(F->getFunctionType() == VprintfFuncType); |
36 | return F; |
37 | } |
38 | |
39 | // vprintf doesn't already exist; create a declaration and insert it into the |
40 | // module. |
41 | return llvm::Function::Create( |
42 | Ty: VprintfFuncType, Linkage: llvm::GlobalVariable::ExternalLinkage, N: "vprintf" , M: &M); |
43 | } |
44 | |
45 | llvm::Function *GetOpenMPVprintfDeclaration(CodeGenModule &CGM) { |
46 | const char *Name = "__llvm_omp_vprintf" ; |
47 | llvm::Module &M = CGM.getModule(); |
48 | llvm::Type *ArgTypes[] = {llvm::PointerType::getUnqual(C&: M.getContext()), |
49 | llvm::PointerType::getUnqual(C&: M.getContext()), |
50 | llvm::Type::getInt32Ty(C&: M.getContext())}; |
51 | llvm::FunctionType *VprintfFuncType = llvm::FunctionType::get( |
52 | Result: llvm::Type::getInt32Ty(C&: M.getContext()), Params: ArgTypes, isVarArg: false); |
53 | |
54 | if (auto *F = M.getFunction(Name)) { |
55 | if (F->getFunctionType() != VprintfFuncType) { |
56 | CGM.Error(loc: SourceLocation(), |
57 | error: "Invalid type declaration for __llvm_omp_vprintf" ); |
58 | return nullptr; |
59 | } |
60 | return F; |
61 | } |
62 | |
63 | return llvm::Function::Create( |
64 | Ty: VprintfFuncType, Linkage: llvm::GlobalVariable::ExternalLinkage, N: Name, M: &M); |
65 | } |
66 | |
67 | // Transforms a call to printf into a call to the NVPTX vprintf syscall (which |
68 | // isn't particularly special; it's invoked just like a regular function). |
69 | // vprintf takes two args: A format string, and a pointer to a buffer containing |
70 | // the varargs. |
71 | // |
72 | // For example, the call |
73 | // |
74 | // printf("format string", arg1, arg2, arg3); |
75 | // |
76 | // is converted into something resembling |
77 | // |
78 | // struct Tmp { |
79 | // Arg1 a1; |
80 | // Arg2 a2; |
81 | // Arg3 a3; |
82 | // }; |
83 | // char* buf = alloca(sizeof(Tmp)); |
84 | // *(Tmp*)buf = {a1, a2, a3}; |
85 | // vprintf("format string", buf); |
86 | // |
87 | // buf is aligned to the max of {alignof(Arg1), ...}. Furthermore, each of the |
88 | // args is itself aligned to its preferred alignment. |
89 | // |
90 | // Note that by the time this function runs, E's args have already undergone the |
91 | // standard C vararg promotion (short -> int, float -> double, etc.). |
92 | |
93 | std::pair<llvm::Value *, llvm::TypeSize> |
94 | packArgsIntoNVPTXFormatBuffer(CodeGenFunction *CGF, const CallArgList &Args) { |
95 | const llvm::DataLayout &DL = CGF->CGM.getDataLayout(); |
96 | llvm::LLVMContext &Ctx = CGF->CGM.getLLVMContext(); |
97 | CGBuilderTy &Builder = CGF->Builder; |
98 | |
99 | // Construct and fill the args buffer that we'll pass to vprintf. |
100 | if (Args.size() <= 1) { |
101 | // If there are no args, pass a null pointer and size 0 |
102 | llvm::Value *BufferPtr = |
103 | llvm::ConstantPointerNull::get(T: llvm::PointerType::getUnqual(C&: Ctx)); |
104 | return {BufferPtr, llvm::TypeSize::getFixed(ExactSize: 0)}; |
105 | } else { |
106 | llvm::SmallVector<llvm::Type *, 8> ArgTypes; |
107 | for (unsigned I = 1, NumArgs = Args.size(); I < NumArgs; ++I) |
108 | ArgTypes.push_back(Elt: Args[I].getRValue(CGF&: *CGF).getScalarVal()->getType()); |
109 | |
110 | // Using llvm::StructType is correct only because printf doesn't accept |
111 | // aggregates. If we had to handle aggregates here, we'd have to manually |
112 | // compute the offsets within the alloca -- we wouldn't be able to assume |
113 | // that the alignment of the llvm type was the same as the alignment of the |
114 | // clang type. |
115 | llvm::Type *AllocaTy = llvm::StructType::create(Elements: ArgTypes, Name: "printf_args" ); |
116 | llvm::Value *Alloca = CGF->CreateTempAlloca(Ty: AllocaTy); |
117 | |
118 | for (unsigned I = 1, NumArgs = Args.size(); I < NumArgs; ++I) { |
119 | llvm::Value *P = Builder.CreateStructGEP(Ty: AllocaTy, Ptr: Alloca, Idx: I - 1); |
120 | llvm::Value *Arg = Args[I].getRValue(CGF&: *CGF).getScalarVal(); |
121 | Builder.CreateAlignedStore(Val: Arg, Ptr: P, Align: DL.getPrefTypeAlign(Ty: Arg->getType())); |
122 | } |
123 | llvm::Value *BufferPtr = |
124 | Builder.CreatePointerCast(V: Alloca, DestTy: llvm::PointerType::getUnqual(C&: Ctx)); |
125 | return {BufferPtr, DL.getTypeAllocSize(Ty: AllocaTy)}; |
126 | } |
127 | } |
128 | |
129 | bool containsNonScalarVarargs(CodeGenFunction *CGF, const CallArgList &Args) { |
130 | return llvm::any_of(Range: llvm::drop_begin(RangeOrContainer: Args), P: [&](const CallArg &A) { |
131 | return !A.getRValue(CGF&: *CGF).isScalar(); |
132 | }); |
133 | } |
134 | |
135 | RValue EmitDevicePrintfCallExpr(const CallExpr *E, CodeGenFunction *CGF, |
136 | llvm::Function *Decl, bool WithSizeArg) { |
137 | CodeGenModule &CGM = CGF->CGM; |
138 | CGBuilderTy &Builder = CGF->Builder; |
139 | assert(E->getBuiltinCallee() == Builtin::BIprintf || |
140 | E->getBuiltinCallee() == Builtin::BI__builtin_printf); |
141 | assert(E->getNumArgs() >= 1); // printf always has at least one arg. |
142 | |
143 | // Uses the same format as nvptx for the argument packing, but also passes |
144 | // an i32 for the total size of the passed pointer |
145 | CallArgList Args; |
146 | CGF->EmitCallArgs(Args, |
147 | Prototype: E->getDirectCallee()->getType()->getAs<FunctionProtoType>(), |
148 | ArgRange: E->arguments(), AC: E->getDirectCallee(), |
149 | /* ParamsToSkip = */ 0); |
150 | |
151 | // We don't know how to emit non-scalar varargs. |
152 | if (containsNonScalarVarargs(CGF, Args)) { |
153 | CGM.ErrorUnsupported(E, "non-scalar arg to printf" ); |
154 | return RValue::get(V: llvm::ConstantInt::get(Ty: CGF->IntTy, V: 0)); |
155 | } |
156 | |
157 | auto r = packArgsIntoNVPTXFormatBuffer(CGF, Args); |
158 | llvm::Value *BufferPtr = r.first; |
159 | |
160 | llvm::SmallVector<llvm::Value *, 3> Vec = { |
161 | Args[0].getRValue(CGF&: *CGF).getScalarVal(), BufferPtr}; |
162 | if (WithSizeArg) { |
163 | // Passing > 32bit of data as a local alloca doesn't work for nvptx or |
164 | // amdgpu |
165 | llvm::Constant *Size = |
166 | llvm::ConstantInt::get(Ty: llvm::Type::getInt32Ty(C&: CGM.getLLVMContext()), |
167 | V: static_cast<uint32_t>(r.second.getFixedValue())); |
168 | |
169 | Vec.push_back(Elt: Size); |
170 | } |
171 | return RValue::get(V: Builder.CreateCall(Callee: Decl, Args: Vec)); |
172 | } |
173 | } // namespace |
174 | |
175 | RValue CodeGenFunction::EmitNVPTXDevicePrintfCallExpr(const CallExpr *E) { |
176 | assert(getTarget().getTriple().isNVPTX()); |
177 | return EmitDevicePrintfCallExpr( |
178 | E, CGF: this, Decl: GetVprintfDeclaration(M&: CGM.getModule()), WithSizeArg: false); |
179 | } |
180 | |
181 | RValue CodeGenFunction::EmitAMDGPUDevicePrintfCallExpr(const CallExpr *E) { |
182 | assert(getTarget().getTriple().getArch() == llvm::Triple::amdgcn); |
183 | assert(E->getBuiltinCallee() == Builtin::BIprintf || |
184 | E->getBuiltinCallee() == Builtin::BI__builtin_printf); |
185 | assert(E->getNumArgs() >= 1); // printf always has at least one arg. |
186 | |
187 | CallArgList CallArgs; |
188 | EmitCallArgs(Args&: CallArgs, |
189 | Prototype: E->getDirectCallee()->getType()->getAs<FunctionProtoType>(), |
190 | ArgRange: E->arguments(), AC: E->getDirectCallee(), |
191 | /* ParamsToSkip = */ 0); |
192 | |
193 | SmallVector<llvm::Value *, 8> Args; |
194 | for (const auto &A : CallArgs) { |
195 | // We don't know how to emit non-scalar varargs. |
196 | if (!A.getRValue(CGF&: *this).isScalar()) { |
197 | CGM.ErrorUnsupported(E, "non-scalar arg to printf" ); |
198 | return RValue::get(V: llvm::ConstantInt::get(Ty: IntTy, V: -1)); |
199 | } |
200 | |
201 | llvm::Value *Arg = A.getRValue(CGF&: *this).getScalarVal(); |
202 | Args.push_back(Elt: Arg); |
203 | } |
204 | |
205 | llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint()); |
206 | IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation()); |
207 | |
208 | bool isBuffered = (CGM.getTarget().getTargetOpts().AMDGPUPrintfKindVal == |
209 | clang::TargetOptions::AMDGPUPrintfKind::Buffered); |
210 | auto Printf = llvm::emitAMDGPUPrintfCall(Builder&: IRB, Args, isBuffered); |
211 | Builder.SetInsertPoint(TheBB: IRB.GetInsertBlock(), IP: IRB.GetInsertPoint()); |
212 | return RValue::get(V: Printf); |
213 | } |
214 | |
215 | RValue CodeGenFunction::EmitOpenMPDevicePrintfCallExpr(const CallExpr *E) { |
216 | assert(getTarget().getTriple().isNVPTX() || |
217 | getTarget().getTriple().isAMDGCN()); |
218 | return EmitDevicePrintfCallExpr(E, CGF: this, Decl: GetOpenMPVprintfDeclaration(CGM), |
219 | WithSizeArg: true); |
220 | } |
221 | |