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/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   unsigned int SmVersion;
54   NVVMReflect() : NVVMReflect(0) {}
55   explicit NVVMReflect(unsigned int Sm) : FunctionPass(ID), SmVersion(Sm) {
56     initializeNVVMReflectPass(*PassRegistry::getPassRegistry());
57   }
58 
59   bool runOnFunction(Function &) override;
60 };
61 }
62 
63 FunctionPass *llvm::createNVVMReflectPass(unsigned int SmVersion) {
64   return new NVVMReflect(SmVersion);
65 }
66 
67 static cl::opt<bool>
68 NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden,
69                    cl::desc("NVVM reflection, enabled by default"));
70 
71 char NVVMReflect::ID = 0;
72 INITIALIZE_PASS(NVVMReflect, "nvvm-reflect",
73                 "Replace occurrences of __nvvm_reflect() calls with 0/1", false,
74                 false)
75 
76 bool NVVMReflect::runOnFunction(Function &F) {
77   if (!NVVMReflectEnabled)
78     return false;
79 
80   if (F.getName() == NVVM_REFLECT_FUNCTION) {
81     assert(F.isDeclaration() && "_reflect function should not have a body");
82     assert(F.getReturnType()->isIntegerTy() &&
83            "_reflect's return type should be integer");
84     return false;
85   }
86 
87   SmallVector<Instruction *, 4> ToRemove;
88 
89   // Go through the calls in this function.  Each call to __nvvm_reflect or
90   // llvm.nvvm.reflect should be a CallInst with a ConstantArray argument.
91   // First validate that. If the c-string corresponding to the ConstantArray can
92   // be found successfully, see if it can be found in VarMap. If so, replace the
93   // uses of CallInst with the value found in VarMap. If not, replace the use
94   // with value 0.
95 
96   // The IR for __nvvm_reflect calls differs between CUDA versions.
97   //
98   // CUDA 6.5 and earlier uses this sequence:
99   //    %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8
100   //        (i8 addrspace(4)* getelementptr inbounds
101   //           ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0))
102   //    %reflect = tail call i32 @__nvvm_reflect(i8* %ptr)
103   //
104   // The value returned by Sym->getOperand(0) is a Constant with a
105   // ConstantDataSequential operand which can be converted to string and used
106   // for lookup.
107   //
108   // CUDA 7.0 does it slightly differently:
109   //   %reflect = call i32 @__nvvm_reflect(i8* addrspacecast
110   //        (i8 addrspace(1)* getelementptr inbounds
111   //           ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*))
112   //
113   // In this case, we get a Constant with a GlobalVariable operand and we need
114   // to dig deeper to find its initializer with the string we'll use for lookup.
115   for (Instruction &I : instructions(F)) {
116     CallInst *Call = dyn_cast<CallInst>(&I);
117     if (!Call)
118       continue;
119     Function *Callee = Call->getCalledFunction();
120     if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION &&
121                     Callee->getIntrinsicID() != Intrinsic::nvvm_reflect))
122       continue;
123 
124     // FIXME: Improve error handling here and elsewhere in this pass.
125     assert(Call->getNumOperands() == 2 &&
126            "Wrong number of operands to __nvvm_reflect function");
127 
128     // In cuda 6.5 and earlier, we will have an extra constant-to-generic
129     // conversion of the string.
130     const Value *Str = Call->getArgOperand(0);
131     if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) {
132       // FIXME: Add assertions about ConvCall.
133       Str = ConvCall->getArgOperand(0);
134     }
135     assert(isa<ConstantExpr>(Str) &&
136            "Format of __nvvm__reflect function not recognized");
137     const ConstantExpr *GEP = cast<ConstantExpr>(Str);
138 
139     const Value *Sym = GEP->getOperand(0);
140     assert(isa<Constant>(Sym) &&
141            "Format of __nvvm_reflect function not recognized");
142 
143     const Value *Operand = cast<Constant>(Sym)->getOperand(0);
144     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) {
145       // For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's
146       // initializer.
147       assert(GV->hasInitializer() &&
148              "Format of _reflect function not recognized");
149       const Constant *Initializer = GV->getInitializer();
150       Operand = Initializer;
151     }
152 
153     assert(isa<ConstantDataSequential>(Operand) &&
154            "Format of _reflect function not recognized");
155     assert(cast<ConstantDataSequential>(Operand)->isCString() &&
156            "Format of _reflect function not recognized");
157 
158     StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString();
159     ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
160     LLVM_DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
161 
162     int ReflectVal = 0; // The default value is 0
163     if (ReflectArg == "__CUDA_FTZ") {
164       // Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag.  Our
165       // choice here must be kept in sync with AutoUpgrade, which uses the same
166       // technique to detect whether ftz is enabled.
167       if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
168               F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
169         ReflectVal = Flag->getSExtValue();
170     } else if (ReflectArg == "__CUDA_ARCH") {
171       ReflectVal = SmVersion * 10;
172     }
173     Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
174     ToRemove.push_back(Call);
175   }
176 
177   for (Instruction *I : ToRemove)
178     I->eraseFromParent();
179 
180   return ToRemove.size() > 0;
181 }
182