Home | History | Annotate | Download | only in NVPTX
      1 //===- NVVMReflect.cpp - NVVM Emulate conditional compilation -------------===//
      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 // This pass replaces occurrences of __nvvm_reflect("foo") and llvm.nvvm.reflect
     11 // with an integer.
     12 //
     13 // We choose the value we use by looking, in this order, at:
     14 //
     15 //  * the -nvvm-reflect-list flag, which has the format "foo=1,bar=42",
     16 //  * the StringMap passed to the pass's constructor, and
     17 //  * metadata in the module itself.
     18 //
     19 // If we see an unknown string, we replace its call with 0.
     20 //
     21 //===----------------------------------------------------------------------===//
     22 
     23 #include "NVPTX.h"
     24 #include "llvm/ADT/SmallVector.h"
     25 #include "llvm/ADT/StringMap.h"
     26 #include "llvm/IR/Constants.h"
     27 #include "llvm/IR/DerivedTypes.h"
     28 #include "llvm/IR/Function.h"
     29 #include "llvm/IR/InstIterator.h"
     30 #include "llvm/IR/Instructions.h"
     31 #include "llvm/IR/Intrinsics.h"
     32 #include "llvm/IR/Module.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 private:
     53   StringMap<int> VarMap;
     54 
     55 public:
     56   static char ID;
     57   NVVMReflect() : NVVMReflect(StringMap<int>()) {}
     58 
     59   NVVMReflect(const StringMap<int> &Mapping)
     60       : FunctionPass(ID), VarMap(Mapping) {
     61     initializeNVVMReflectPass(*PassRegistry::getPassRegistry());
     62     setVarMap();
     63   }
     64 
     65   bool runOnFunction(Function &) override;
     66 
     67 private:
     68   bool handleFunction(Function *ReflectFunction);
     69   void setVarMap();
     70 };
     71 }
     72 
     73 FunctionPass *llvm::createNVVMReflectPass() { return new NVVMReflect(); }
     74 FunctionPass *llvm::createNVVMReflectPass(const StringMap<int> &Mapping) {
     75   return new NVVMReflect(Mapping);
     76 }
     77 
     78 static cl::opt<bool>
     79 NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden,
     80                    cl::desc("NVVM reflection, enabled by default"));
     81 
     82 char NVVMReflect::ID = 0;
     83 INITIALIZE_PASS(NVVMReflect, "nvvm-reflect",
     84                 "Replace occurrences of __nvvm_reflect() calls with 0/1", false,
     85                 false)
     86 
     87 static cl::list<std::string>
     88 ReflectList("nvvm-reflect-list", cl::value_desc("name=<int>"), cl::Hidden,
     89             cl::desc("A list of string=num assignments"),
     90             cl::ValueRequired);
     91 
     92 /// The command line can look as follows :
     93 /// -nvvm-reflect-list a=1,b=2 -nvvm-reflect-list c=3,d=0 -R e=2
     94 /// The strings "a=1,b=2", "c=3,d=0", "e=2" are available in the
     95 /// ReflectList vector. First, each of ReflectList[i] is 'split'
     96 /// using "," as the delimiter. Then each of this part is split
     97 /// using "=" as the delimiter.
     98 void NVVMReflect::setVarMap() {
     99   for (unsigned i = 0, e = ReflectList.size(); i != e; ++i) {
    100     DEBUG(dbgs() << "Option : "  << ReflectList[i] << "\n");
    101     SmallVector<StringRef, 4> NameValList;
    102     StringRef(ReflectList[i]).split(NameValList, ',');
    103     for (unsigned j = 0, ej = NameValList.size(); j != ej; ++j) {
    104       SmallVector<StringRef, 2> NameValPair;
    105       NameValList[j].split(NameValPair, '=');
    106       assert(NameValPair.size() == 2 && "name=val expected");
    107       std::stringstream ValStream(NameValPair[1]);
    108       int Val;
    109       ValStream >> Val;
    110       assert((!(ValStream.fail())) && "integer value expected");
    111       VarMap[NameValPair[0]] = Val;
    112     }
    113   }
    114 }
    115 
    116 bool NVVMReflect::runOnFunction(Function &F) {
    117   if (!NVVMReflectEnabled)
    118     return false;
    119 
    120   if (F.getName() == NVVM_REFLECT_FUNCTION) {
    121     assert(F.isDeclaration() && "_reflect function should not have a body");
    122     assert(F.getReturnType()->isIntegerTy() &&
    123            "_reflect's return type should be integer");
    124     return false;
    125   }
    126 
    127   SmallVector<Instruction *, 4> ToRemove;
    128 
    129   // Go through the calls in this function.  Each call to __nvvm_reflect or
    130   // llvm.nvvm.reflect should be a CallInst with a ConstantArray argument.
    131   // First validate that. If the c-string corresponding to the ConstantArray can
    132   // be found successfully, see if it can be found in VarMap. If so, replace the
    133   // uses of CallInst with the value found in VarMap. If not, replace the use
    134   // with value 0.
    135 
    136   // The IR for __nvvm_reflect calls differs between CUDA versions.
    137   //
    138   // CUDA 6.5 and earlier uses this sequence:
    139   //    %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8
    140   //        (i8 addrspace(4)* getelementptr inbounds
    141   //           ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0))
    142   //    %reflect = tail call i32 @__nvvm_reflect(i8* %ptr)
    143   //
    144   // The value returned by Sym->getOperand(0) is a Constant with a
    145   // ConstantDataSequential operand which can be converted to string and used
    146   // for lookup.
    147   //
    148   // CUDA 7.0 does it slightly differently:
    149   //   %reflect = call i32 @__nvvm_reflect(i8* addrspacecast
    150   //        (i8 addrspace(1)* getelementptr inbounds
    151   //           ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*))
    152   //
    153   // In this case, we get a Constant with a GlobalVariable operand and we need
    154   // to dig deeper to find its initializer with the string we'll use for lookup.
    155   for (Instruction &I : instructions(F)) {
    156     CallInst *Call = dyn_cast<CallInst>(&I);
    157     if (!Call)
    158       continue;
    159     Function *Callee = Call->getCalledFunction();
    160     if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION &&
    161                     Callee->getIntrinsicID() != Intrinsic::nvvm_reflect))
    162       continue;
    163 
    164     // FIXME: Improve error handling here and elsewhere in this pass.
    165     assert(Call->getNumOperands() == 2 &&
    166            "Wrong number of operands to __nvvm_reflect function");
    167 
    168     // In cuda 6.5 and earlier, we will have an extra constant-to-generic
    169     // conversion of the string.
    170     const Value *Str = Call->getArgOperand(0);
    171     if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) {
    172       // FIXME: Add assertions about ConvCall.
    173       Str = ConvCall->getArgOperand(0);
    174     }
    175     assert(isa<ConstantExpr>(Str) &&
    176            "Format of __nvvm__reflect function not recognized");
    177     const ConstantExpr *GEP = cast<ConstantExpr>(Str);
    178 
    179     const Value *Sym = GEP->getOperand(0);
    180     assert(isa<Constant>(Sym) &&
    181            "Format of __nvvm_reflect function not recognized");
    182 
    183     const Value *Operand = cast<Constant>(Sym)->getOperand(0);
    184     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) {
    185       // For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's
    186       // initializer.
    187       assert(GV->hasInitializer() &&
    188              "Format of _reflect function not recognized");
    189       const Constant *Initializer = GV->getInitializer();
    190       Operand = Initializer;
    191     }
    192 
    193     assert(isa<ConstantDataSequential>(Operand) &&
    194            "Format of _reflect function not recognized");
    195     assert(cast<ConstantDataSequential>(Operand)->isCString() &&
    196            "Format of _reflect function not recognized");
    197 
    198     StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString();
    199     ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
    200     DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
    201 
    202     int ReflectVal = 0; // The default value is 0
    203     auto Iter = VarMap.find(ReflectArg);
    204     if (Iter != VarMap.end())
    205       ReflectVal = Iter->second;
    206     else if (ReflectArg == "__CUDA_FTZ") {
    207       // Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag.
    208       if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
    209               F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
    210         ReflectVal = Flag->getSExtValue();
    211     }
    212     Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
    213     ToRemove.push_back(Call);
    214   }
    215 
    216   for (Instruction *I : ToRemove)
    217     I->eraseFromParent();
    218 
    219   return ToRemove.size() > 0;
    220 }
    221