1 //===----- CGCUDABuiltin.cpp - Codegen for CUDA builtins ------------------===//
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 // Generates code for built-in CUDA calls which are not runtime-specific.
11 // (Runtime-specific codegen lives in CGCUDARuntime.)
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "CodeGenFunction.h"
16 #include "clang/Basic/Builtins.h"
17 #include "llvm/IR/DataLayout.h"
18 #include "llvm/IR/Instruction.h"
19 #include "llvm/Support/MathExtras.h"
20 
21 using namespace clang;
22 using namespace CodeGen;
23 
GetVprintfDeclaration(llvm::Module & M)24 static llvm::Function *GetVprintfDeclaration(llvm::Module &M) {
25   llvm::Type *ArgTypes[] = {llvm::Type::getInt8PtrTy(M.getContext()),
26                             llvm::Type::getInt8PtrTy(M.getContext())};
27   llvm::FunctionType *VprintfFuncType = llvm::FunctionType::get(
28       llvm::Type::getInt32Ty(M.getContext()), ArgTypes, false);
29 
30   if (auto* F = M.getFunction("vprintf")) {
31     // Our CUDA system header declares vprintf with the right signature, so
32     // nobody else should have been able to declare vprintf with a bogus
33     // signature.
34     assert(F->getFunctionType() == VprintfFuncType);
35     return F;
36   }
37 
38   // vprintf doesn't already exist; create a declaration and insert it into the
39   // module.
40   return llvm::Function::Create(
41       VprintfFuncType, llvm::GlobalVariable::ExternalLinkage, "vprintf", &M);
42 }
43 
44 // Transforms a call to printf into a call to the NVPTX vprintf syscall (which
45 // isn't particularly special; it's invoked just like a regular function).
46 // vprintf takes two args: A format string, and a pointer to a buffer containing
47 // the varargs.
48 //
49 // For example, the call
50 //
51 //   printf("format string", arg1, arg2, arg3);
52 //
53 // is converted into something resembling
54 //
55 //   struct Tmp {
56 //     Arg1 a1;
57 //     Arg2 a2;
58 //     Arg3 a3;
59 //   };
60 //   char* buf = alloca(sizeof(Tmp));
61 //   *(Tmp*)buf = {a1, a2, a3};
62 //   vprintf("format string", buf);
63 //
64 // buf is aligned to the max of {alignof(Arg1), ...}.  Furthermore, each of the
65 // args is itself aligned to its preferred alignment.
66 //
67 // Note that by the time this function runs, E's args have already undergone the
68 // standard C vararg promotion (short -> int, float -> double, etc.).
69 RValue
EmitCUDADevicePrintfCallExpr(const CallExpr * E,ReturnValueSlot ReturnValue)70 CodeGenFunction::EmitCUDADevicePrintfCallExpr(const CallExpr *E,
71                                               ReturnValueSlot ReturnValue) {
72   assert(getLangOpts().CUDA);
73   assert(getLangOpts().CUDAIsDevice);
74   assert(E->getBuiltinCallee() == Builtin::BIprintf);
75   assert(E->getNumArgs() >= 1); // printf always has at least one arg.
76 
77   const llvm::DataLayout &DL = CGM.getDataLayout();
78   llvm::LLVMContext &Ctx = CGM.getLLVMContext();
79 
80   CallArgList Args;
81   EmitCallArgs(Args,
82                E->getDirectCallee()->getType()->getAs<FunctionProtoType>(),
83                E->arguments(), E->getDirectCallee(),
84                /* ParamsToSkip = */ 0);
85 
86   // We don't know how to emit non-scalar varargs.
87   if (std::any_of(Args.begin() + 1, Args.end(),
88                   [](const CallArg &A) { return !A.RV.isScalar(); })) {
89     CGM.ErrorUnsupported(E, "non-scalar arg to printf");
90     return RValue::get(llvm::ConstantInt::get(IntTy, 0));
91   }
92 
93   // Construct and fill the args buffer that we'll pass to vprintf.
94   llvm::Value *BufferPtr;
95   if (Args.size() <= 1) {
96     // If there are no args, pass a null pointer to vprintf.
97     BufferPtr = llvm::ConstantPointerNull::get(llvm::Type::getInt8PtrTy(Ctx));
98   } else {
99     llvm::SmallVector<llvm::Type *, 8> ArgTypes;
100     for (unsigned I = 1, NumArgs = Args.size(); I < NumArgs; ++I)
101       ArgTypes.push_back(Args[I].RV.getScalarVal()->getType());
102     llvm::Type *AllocaTy = llvm::StructType::create(ArgTypes, "printf_args");
103     llvm::Value *Alloca = CreateTempAlloca(AllocaTy);
104 
105     for (unsigned I = 1, NumArgs = Args.size(); I < NumArgs; ++I) {
106       llvm::Value *P = Builder.CreateStructGEP(AllocaTy, Alloca, I - 1);
107       llvm::Value *Arg = Args[I].RV.getScalarVal();
108       Builder.CreateAlignedStore(Arg, P, DL.getPrefTypeAlignment(Arg->getType()));
109     }
110     BufferPtr = Builder.CreatePointerCast(Alloca, llvm::Type::getInt8PtrTy(Ctx));
111   }
112 
113   // Invoke vprintf and return.
114   llvm::Function* VprintfFunc = GetVprintfDeclaration(CGM.getModule());
115   return RValue::get(
116       Builder.CreateCall(VprintfFunc, {Args[0].RV.getScalarVal(), BufferPtr}));
117 }
118