lib/Target/NVPTX/NVVMReflect.cpp

   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/Module.h"
  31 #include "llvm/IR/Type.h"
  32 #include "llvm/Pass.h"
  33 #include "llvm/Support/CommandLine.h"
  34 #include "llvm/Support/Debug.h"
  35 #include "llvm/Support/raw_os_ostream.h"
  36 #include "llvm/Support/raw_ostream.h"
  37 #include "llvm/Transforms/Scalar.h"
  38 #include <sstream>
  39 #include <string>
  40 #define NVVM_REFLECT_FUNCTION "__nvvm_reflect"
  41
  42 using namespace llvm;
  43
  44 #define DEBUG_TYPE "nvptx-reflect"
  45
  46 namespace llvm { void initializeNVVMReflectPass(PassRegistry &); }
  47
  48 namespace {
  49 class NVVMReflect : public FunctionPass {
  50 public:
  51   static char ID;
  52   unsigned int SmVersion;
  53   NVVMReflect() : NVVMReflect(0) {}
  54   explicit NVVMReflect(unsigned int Sm) : FunctionPass(ID), SmVersion(Sm) {
  55     initializeNVVMReflectPass(*PassRegistry::getPassRegistry());
  56   }
  57
  58   bool runOnFunction(Function &) override;
  59 };
  60 }
  61
  62 FunctionPass *llvm::createNVVMReflectPass(unsigned int SmVersion) {
  63   return new NVVMReflect(SmVersion);
  64 }
  65
  66 static cl::opt<bool>
  67 NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden,
  68                    cl::desc("NVVM reflection, enabled by default"));
  69
  70 char NVVMReflect::ID = 0;
  71 INITIALIZE_PASS(NVVMReflect, "nvvm-reflect",
  72                 "Replace occurrences of __nvvm_reflect() calls with 0/1", false,
  73                 false)
  74
  75 bool NVVMReflect::runOnFunction(Function &F) {
  76   if (!NVVMReflectEnabled)
  77     return false;
  78
  79   if (F.getName() == NVVM_REFLECT_FUNCTION) {
  80     assert(F.isDeclaration() && "_reflect function should not have a body");
  81     assert(F.getReturnType()->isIntegerTy() &&
  82            "_reflect's return type should be integer");
  83     return false;
  84   }
  85
  86   SmallVector<Instruction *, 4> ToRemove;
  87
  88   // Go through the calls in this function.  Each call to __nvvm_reflect or
  89   // llvm.nvvm.reflect should be a CallInst with a ConstantArray argument.
  90   // First validate that. If the c-string corresponding to the ConstantArray can
  91   // be found successfully, see if it can be found in VarMap. If so, replace the
  92   // uses of CallInst with the value found in VarMap. If not, replace the use
  93   // with value 0.
  94
  95   // The IR for __nvvm_reflect calls differs between CUDA versions.
  96   //
  97   // CUDA 6.5 and earlier uses this sequence:
  98   //    %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8
  99   //        (i8 addrspace(4)* getelementptr inbounds
 100   //           ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0))
 101   //    %reflect = tail call i32 @__nvvm_reflect(i8* %ptr)
 102   //
 103   // The value returned by Sym->getOperand(0) is a Constant with a
 104   // ConstantDataSequential operand which can be converted to string and used
 105   // for lookup.
 106   //
 107   // CUDA 7.0 does it slightly differently:
 108   //   %reflect = call i32 @__nvvm_reflect(i8* addrspacecast
 109   //        (i8 addrspace(1)* getelementptr inbounds
 110   //           ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*))
 111   //
 112   // In this case, we get a Constant with a GlobalVariable operand and we need
 113   // to dig deeper to find its initializer with the string we'll use for lookup.
 114   for (Instruction &I : instructions(F)) {
 115     CallInst *Call = dyn_cast<CallInst>(&I);
 116     if (!Call)
 117       continue;
 118     Function *Callee = Call->getCalledFunction();
 119     if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION &&
 120                     Callee->getIntrinsicID() != Intrinsic::nvvm_reflect))
 121       continue;
 122
 123     // FIXME: Improve error handling here and elsewhere in this pass.
 124     assert(Call->getNumOperands() == 2 &&
 125            "Wrong number of operands to __nvvm_reflect function");
 126
 127     // In cuda 6.5 and earlier, we will have an extra constant-to-generic
 128     // conversion of the string.
 129     const Value *Str = Call->getArgOperand(0);
 130     if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) {
 131       // FIXME: Add assertions about ConvCall.
 132       Str = ConvCall->getArgOperand(0);
 133     }
 134     assert(isa<ConstantExpr>(Str) &&
 135            "Format of __nvvm__reflect function not recognized");
 136     const ConstantExpr *GEP = cast<ConstantExpr>(Str);
 137
 138     const Value *Sym = GEP->getOperand(0);
 139     assert(isa<Constant>(Sym) &&
 140            "Format of __nvvm_reflect function not recognized");
 141
 142     const Value *Operand = cast<Constant>(Sym)->getOperand(0);
 143     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) {
 144       // For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's
 145       // initializer.
 146       assert(GV->hasInitializer() &&
 147              "Format of _reflect function not recognized");
 148       const Constant *Initializer = GV->getInitializer();
 149       Operand = Initializer;
 150     }
 151
 152     assert(isa<ConstantDataSequential>(Operand) &&
 153            "Format of _reflect function not recognized");
 154     assert(cast<ConstantDataSequential>(Operand)->isCString() &&
 155            "Format of _reflect function not recognized");
 156
 157     StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString();
 158     ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
 159     LLVM_DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
 160
 161     int ReflectVal = 0; // The default value is 0
 162     if (ReflectArg == "__CUDA_FTZ") {
 163       // Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag.  Our
 164       // choice here must be kept in sync with AutoUpgrade, which uses the same
 165       // technique to detect whether ftz is enabled.
 166       if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
 167               F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
 168         ReflectVal = Flag->getSExtValue();
 169     } else if (ReflectArg == "__CUDA_ARCH") {
 170       ReflectVal = SmVersion * 10;
 171     }
 172     Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
 173     ToRemove.push_back(Call);
 174   }
 175
 176   for (Instruction *I : ToRemove)
 177     I->eraseFromParent();
 178
 179   return ToRemove.size() > 0;
 180 }