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 at metadata in the module itself.  Note
14 // that we intentionally only have one way to choose these values, because other
15 // parts of LLVM (particularly, InstCombineCall) rely on being able to predict
16 // the values chosen by this pass.
17 //
18 // If we see an unknown string, we replace its call with 0.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "NVPTX.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/StringMap.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/InstIterator.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/Type.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_os_ostream.h"
37 #include "llvm/Support/raw_ostream.h"
38 #include "llvm/Transforms/Scalar.h"
39 #include <sstream>
40 #include <string>
41 #define NVVM_REFLECT_FUNCTION "__nvvm_reflect"
42 
43 using namespace llvm;
44 
45 #define DEBUG_TYPE "nvptx-reflect"
46 
47 namespace llvm { void initializeNVVMReflectPass(PassRegistry &); }
48 
49 namespace {
50 class NVVMReflect : public FunctionPass {
51 public:
52   static char ID;
53   NVVMReflect() : FunctionPass(ID) {
54     initializeNVVMReflectPass(*PassRegistry::getPassRegistry());
55   }
56 
57   bool runOnFunction(Function &) override;
58 };
59 }
60 
61 FunctionPass *llvm::createNVVMReflectPass() { return new NVVMReflect(); }
62 
63 static cl::opt<bool>
64 NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden,
65                    cl::desc("NVVM reflection, enabled by default"));
66 
67 char NVVMReflect::ID = 0;
68 INITIALIZE_PASS(NVVMReflect, "nvvm-reflect",
69                 "Replace occurrences of __nvvm_reflect() calls with 0/1", false,
70                 false)
71 
72 bool NVVMReflect::runOnFunction(Function &F) {
73   if (!NVVMReflectEnabled)
74     return false;
75 
76   if (F.getName() == NVVM_REFLECT_FUNCTION) {
77     assert(F.isDeclaration() && "_reflect function should not have a body");
78     assert(F.getReturnType()->isIntegerTy() &&
79            "_reflect's return type should be integer");
80     return false;
81   }
82 
83   SmallVector<Instruction *, 4> ToRemove;
84 
85   // Go through the calls in this function.  Each call to __nvvm_reflect or
86   // llvm.nvvm.reflect should be a CallInst with a ConstantArray argument.
87   // First validate that. If the c-string corresponding to the ConstantArray can
88   // be found successfully, see if it can be found in VarMap. If so, replace the
89   // uses of CallInst with the value found in VarMap. If not, replace the use
90   // with value 0.
91 
92   // The IR for __nvvm_reflect calls differs between CUDA versions.
93   //
94   // CUDA 6.5 and earlier uses this sequence:
95   //    %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8
96   //        (i8 addrspace(4)* getelementptr inbounds
97   //           ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0))
98   //    %reflect = tail call i32 @__nvvm_reflect(i8* %ptr)
99   //
100   // The value returned by Sym->getOperand(0) is a Constant with a
101   // ConstantDataSequential operand which can be converted to string and used
102   // for lookup.
103   //
104   // CUDA 7.0 does it slightly differently:
105   //   %reflect = call i32 @__nvvm_reflect(i8* addrspacecast
106   //        (i8 addrspace(1)* getelementptr inbounds
107   //           ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*))
108   //
109   // In this case, we get a Constant with a GlobalVariable operand and we need
110   // to dig deeper to find its initializer with the string we'll use for lookup.
111   for (Instruction &I : instructions(F)) {
112     CallInst *Call = dyn_cast<CallInst>(&I);
113     if (!Call)
114       continue;
115     Function *Callee = Call->getCalledFunction();
116     if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION &&
117                     Callee->getIntrinsicID() != Intrinsic::nvvm_reflect))
118       continue;
119 
120     // FIXME: Improve error handling here and elsewhere in this pass.
121     assert(Call->getNumOperands() == 2 &&
122            "Wrong number of operands to __nvvm_reflect function");
123 
124     // In cuda 6.5 and earlier, we will have an extra constant-to-generic
125     // conversion of the string.
126     const Value *Str = Call->getArgOperand(0);
127     if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) {
128       // FIXME: Add assertions about ConvCall.
129       Str = ConvCall->getArgOperand(0);
130     }
131     assert(isa<ConstantExpr>(Str) &&
132            "Format of __nvvm__reflect function not recognized");
133     const ConstantExpr *GEP = cast<ConstantExpr>(Str);
134 
135     const Value *Sym = GEP->getOperand(0);
136     assert(isa<Constant>(Sym) &&
137            "Format of __nvvm_reflect function not recognized");
138 
139     const Value *Operand = cast<Constant>(Sym)->getOperand(0);
140     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) {
141       // For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's
142       // initializer.
143       assert(GV->hasInitializer() &&
144              "Format of _reflect function not recognized");
145       const Constant *Initializer = GV->getInitializer();
146       Operand = Initializer;
147     }
148 
149     assert(isa<ConstantDataSequential>(Operand) &&
150            "Format of _reflect function not recognized");
151     assert(cast<ConstantDataSequential>(Operand)->isCString() &&
152            "Format of _reflect function not recognized");
153 
154     StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString();
155     ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
156     DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
157 
158     int ReflectVal = 0; // The default value is 0
159     if (ReflectArg == "__CUDA_FTZ") {
160       // Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag.  Our
161       // choice here must be kept in sync with AutoUpgrade, which uses the same
162       // technique to detect whether ftz is enabled.
163       if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
164               F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
165         ReflectVal = Flag->getSExtValue();
166     }
167     Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
168     ToRemove.push_back(Call);
169   }
170 
171   for (Instruction *I : ToRemove)
172     I->eraseFromParent();
173 
174   return ToRemove.size() > 0;
175 }
176