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