1 //===- NVVMReflect.cpp - NVVM Emulate conditional compilation -------------===// 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 // This pass replaces occurrences of __nvvm_reflect("foo") and llvm.nvvm.reflect 10 // with an integer. 11 // 12 // We choose the value we use by looking at metadata in the module itself. Note 13 // that we intentionally only have one way to choose these values, because other 14 // parts of LLVM (particularly, InstCombineCall) rely on being able to predict 15 // the values chosen by this pass. 16 // 17 // If we see an unknown string, we replace its call with 0. 18 // 19 //===----------------------------------------------------------------------===// 20 21 #include "NVPTX.h" 22 #include "llvm/ADT/SmallVector.h" 23 #include "llvm/ADT/StringMap.h" 24 #include "llvm/IR/Constants.h" 25 #include "llvm/IR/DerivedTypes.h" 26 #include "llvm/IR/Function.h" 27 #include "llvm/IR/InstIterator.h" 28 #include "llvm/IR/Instructions.h" 29 #include "llvm/IR/Intrinsics.h" 30 #include "llvm/IR/IntrinsicsNVPTX.h" 31 #include "llvm/IR/Module.h" 32 #include "llvm/IR/PassManager.h" 33 #include "llvm/IR/Type.h" 34 #include "llvm/Pass.h" 35 #include "llvm/Support/CommandLine.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/raw_os_ostream.h" 38 #include "llvm/Support/raw_ostream.h" 39 #include "llvm/Transforms/Scalar.h" 40 #include <sstream> 41 #include <string> 42 #define NVVM_REFLECT_FUNCTION "__nvvm_reflect" 43 44 using namespace llvm; 45 46 #define DEBUG_TYPE "nvptx-reflect" 47 48 namespace llvm { void initializeNVVMReflectPass(PassRegistry &); } 49 50 namespace { 51 class NVVMReflect : public FunctionPass { 52 public: 53 static char ID; 54 unsigned int SmVersion; 55 NVVMReflect() : NVVMReflect(0) {} 56 explicit NVVMReflect(unsigned int Sm) : FunctionPass(ID), SmVersion(Sm) { 57 initializeNVVMReflectPass(*PassRegistry::getPassRegistry()); 58 } 59 60 bool runOnFunction(Function &) override; 61 }; 62 } 63 64 FunctionPass *llvm::createNVVMReflectPass(unsigned int SmVersion) { 65 return new NVVMReflect(SmVersion); 66 } 67 68 static cl::opt<bool> 69 NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden, 70 cl::desc("NVVM reflection, enabled by default")); 71 72 char NVVMReflect::ID = 0; 73 INITIALIZE_PASS(NVVMReflect, "nvvm-reflect", 74 "Replace occurrences of __nvvm_reflect() calls with 0/1", false, 75 false) 76 77 static bool runNVVMReflect(Function &F, unsigned SmVersion) { 78 if (!NVVMReflectEnabled) 79 return false; 80 81 if (F.getName() == NVVM_REFLECT_FUNCTION) { 82 assert(F.isDeclaration() && "_reflect function should not have a body"); 83 assert(F.getReturnType()->isIntegerTy() && 84 "_reflect's return type should be integer"); 85 return false; 86 } 87 88 SmallVector<Instruction *, 4> ToRemove; 89 90 // Go through the calls in this function. Each call to __nvvm_reflect or 91 // llvm.nvvm.reflect should be a CallInst with a ConstantArray argument. 92 // First validate that. If the c-string corresponding to the ConstantArray can 93 // be found successfully, see if it can be found in VarMap. If so, replace the 94 // uses of CallInst with the value found in VarMap. If not, replace the use 95 // with value 0. 96 97 // The IR for __nvvm_reflect calls differs between CUDA versions. 98 // 99 // CUDA 6.5 and earlier uses this sequence: 100 // %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8 101 // (i8 addrspace(4)* getelementptr inbounds 102 // ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0)) 103 // %reflect = tail call i32 @__nvvm_reflect(i8* %ptr) 104 // 105 // The value returned by Sym->getOperand(0) is a Constant with a 106 // ConstantDataSequential operand which can be converted to string and used 107 // for lookup. 108 // 109 // CUDA 7.0 does it slightly differently: 110 // %reflect = call i32 @__nvvm_reflect(i8* addrspacecast 111 // (i8 addrspace(1)* getelementptr inbounds 112 // ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*)) 113 // 114 // In this case, we get a Constant with a GlobalVariable operand and we need 115 // to dig deeper to find its initializer with the string we'll use for lookup. 116 for (Instruction &I : instructions(F)) { 117 CallInst *Call = dyn_cast<CallInst>(&I); 118 if (!Call) 119 continue; 120 Function *Callee = Call->getCalledFunction(); 121 if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION && 122 Callee->getIntrinsicID() != Intrinsic::nvvm_reflect)) 123 continue; 124 125 // FIXME: Improve error handling here and elsewhere in this pass. 126 assert(Call->getNumOperands() == 2 && 127 "Wrong number of operands to __nvvm_reflect function"); 128 129 // In cuda 6.5 and earlier, we will have an extra constant-to-generic 130 // conversion of the string. 131 const Value *Str = Call->getArgOperand(0); 132 if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) { 133 // FIXME: Add assertions about ConvCall. 134 Str = ConvCall->getArgOperand(0); 135 } 136 assert(isa<ConstantExpr>(Str) && 137 "Format of __nvvm__reflect function not recognized"); 138 const ConstantExpr *GEP = cast<ConstantExpr>(Str); 139 140 const Value *Sym = GEP->getOperand(0); 141 assert(isa<Constant>(Sym) && 142 "Format of __nvvm_reflect function not recognized"); 143 144 const Value *Operand = cast<Constant>(Sym)->getOperand(0); 145 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) { 146 // For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's 147 // initializer. 148 assert(GV->hasInitializer() && 149 "Format of _reflect function not recognized"); 150 const Constant *Initializer = GV->getInitializer(); 151 Operand = Initializer; 152 } 153 154 assert(isa<ConstantDataSequential>(Operand) && 155 "Format of _reflect function not recognized"); 156 assert(cast<ConstantDataSequential>(Operand)->isCString() && 157 "Format of _reflect function not recognized"); 158 159 StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString(); 160 ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1); 161 LLVM_DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n"); 162 163 int ReflectVal = 0; // The default value is 0 164 if (ReflectArg == "__CUDA_FTZ") { 165 // Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag. Our 166 // choice here must be kept in sync with AutoUpgrade, which uses the same 167 // technique to detect whether ftz is enabled. 168 if (auto *Flag = mdconst::extract_or_null<ConstantInt>( 169 F.getParent()->getModuleFlag("nvvm-reflect-ftz"))) 170 ReflectVal = Flag->getSExtValue(); 171 } else if (ReflectArg == "__CUDA_ARCH") { 172 ReflectVal = SmVersion * 10; 173 } 174 Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal)); 175 ToRemove.push_back(Call); 176 } 177 178 for (Instruction *I : ToRemove) 179 I->eraseFromParent(); 180 181 return ToRemove.size() > 0; 182 } 183 184 bool NVVMReflect::runOnFunction(Function &F) { 185 return runNVVMReflect(F, SmVersion); 186 } 187 188 NVVMReflectPass::NVVMReflectPass() : NVVMReflectPass(0) {} 189 190 PreservedAnalyses NVVMReflectPass::run(Function &F, 191 FunctionAnalysisManager &AM) { 192 return runNVVMReflect(F, SmVersion) ? PreservedAnalyses::none() 193 : PreservedAnalyses::all(); 194 } 195