LLVM 18.0.0git
NVVMReflect.cpp
Go to the documentation of this file.
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"
23#include "llvm/IR/Constants.h"
25#include "llvm/IR/Function.h"
28#include "llvm/IR/Intrinsics.h"
29#include "llvm/IR/IntrinsicsNVPTX.h"
30#include "llvm/IR/Module.h"
31#include "llvm/IR/PassManager.h"
32#include "llvm/IR/Type.h"
33#include "llvm/Pass.h"
35#include "llvm/Support/Debug.h"
39#include <sstream>
40#include <string>
41#define NVVM_REFLECT_FUNCTION "__nvvm_reflect"
42#define NVVM_REFLECT_OCL_FUNCTION "__nvvm_reflect_ocl"
43
44using namespace llvm;
45
46#define DEBUG_TYPE "nvptx-reflect"
47
48namespace llvm { void initializeNVVMReflectPass(PassRegistry &); }
49
50namespace {
51class NVVMReflect : public FunctionPass {
52public:
53 static char ID;
54 unsigned int SmVersion;
55 NVVMReflect() : NVVMReflect(0) {}
56 explicit NVVMReflect(unsigned int Sm) : FunctionPass(ID), SmVersion(Sm) {
58 }
59
60 bool runOnFunction(Function &) override;
61};
62}
63
65 return new NVVMReflect(SmVersion);
66}
67
68static cl::opt<bool>
69NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden,
70 cl::desc("NVVM reflection, enabled by default"));
71
72char NVVMReflect::ID = 0;
73INITIALIZE_PASS(NVVMReflect, "nvvm-reflect",
74 "Replace occurrences of __nvvm_reflect() calls with 0/1", false,
75 false)
76
77static bool runNVVMReflect(Function &F, unsigned SmVersion) {
79 return false;
80
81 if (F.getName() == NVVM_REFLECT_FUNCTION ||
82 F.getName() == NVVM_REFLECT_OCL_FUNCTION) {
83 assert(F.isDeclaration() && "_reflect function should not have a body");
84 assert(F.getReturnType()->isIntegerTy() &&
85 "_reflect's return type should be integer");
86 return false;
87 }
88
90
91 // Go through the calls in this function. Each call to __nvvm_reflect or
92 // llvm.nvvm.reflect should be a CallInst with a ConstantArray argument.
93 // First validate that. If the c-string corresponding to the ConstantArray can
94 // be found successfully, see if it can be found in VarMap. If so, replace the
95 // uses of CallInst with the value found in VarMap. If not, replace the use
96 // with value 0.
97
98 // The IR for __nvvm_reflect calls differs between CUDA versions.
99 //
100 // CUDA 6.5 and earlier uses this sequence:
101 // %ptr = tail call i8* @llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8
102 // (i8 addrspace(4)* getelementptr inbounds
103 // ([8 x i8], [8 x i8] addrspace(4)* @str, i32 0, i32 0))
104 // %reflect = tail call i32 @__nvvm_reflect(i8* %ptr)
105 //
106 // The value returned by Sym->getOperand(0) is a Constant with a
107 // ConstantDataSequential operand which can be converted to string and used
108 // for lookup.
109 //
110 // CUDA 7.0 does it slightly differently:
111 // %reflect = call i32 @__nvvm_reflect(i8* addrspacecast
112 // (i8 addrspace(1)* getelementptr inbounds
113 // ([8 x i8], [8 x i8] addrspace(1)* @str, i32 0, i32 0) to i8*))
114 //
115 // In this case, we get a Constant with a GlobalVariable operand and we need
116 // to dig deeper to find its initializer with the string we'll use for lookup.
118 CallInst *Call = dyn_cast<CallInst>(&I);
119 if (!Call)
120 continue;
121 Function *Callee = Call->getCalledFunction();
122 if (!Callee || (Callee->getName() != NVVM_REFLECT_FUNCTION &&
123 Callee->getName() != NVVM_REFLECT_OCL_FUNCTION &&
124 Callee->getIntrinsicID() != Intrinsic::nvvm_reflect))
125 continue;
126
127 // FIXME: Improve error handling here and elsewhere in this pass.
128 assert(Call->getNumOperands() == 2 &&
129 "Wrong number of operands to __nvvm_reflect function");
130
131 // In cuda 6.5 and earlier, we will have an extra constant-to-generic
132 // conversion of the string.
133 const Value *Str = Call->getArgOperand(0);
134 if (const CallInst *ConvCall = dyn_cast<CallInst>(Str)) {
135 // FIXME: Add assertions about ConvCall.
136 Str = ConvCall->getArgOperand(0);
137 }
138 // Pre opaque pointers we have a constant expression wrapping the constant
139 // string.
140 Str = Str->stripPointerCasts();
141 assert(isa<Constant>(Str) &&
142 "Format of __nvvm_reflect function not recognized");
143
144 const Value *Operand = cast<Constant>(Str)->getOperand(0);
145 if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Operand)) {
146 // For CUDA-7.0 style __nvvm_reflect calls, we need to find the operand's
147 // initializer.
148 assert(GV->hasInitializer() &&
149 "Format of _reflect function not recognized");
150 const Constant *Initializer = GV->getInitializer();
151 Operand = Initializer;
152 }
153
154 assert(isa<ConstantDataSequential>(Operand) &&
155 "Format of _reflect function not recognized");
156 assert(cast<ConstantDataSequential>(Operand)->isCString() &&
157 "Format of _reflect function not recognized");
158
159 StringRef ReflectArg = cast<ConstantDataSequential>(Operand)->getAsString();
160 ReflectArg = ReflectArg.substr(0, ReflectArg.size() - 1);
161 LLVM_DEBUG(dbgs() << "Arg of _reflect : " << ReflectArg << "\n");
162
163 int ReflectVal = 0; // The default value is 0
164 if (ReflectArg == "__CUDA_FTZ") {
165 // Try to pull __CUDA_FTZ from the nvvm-reflect-ftz module flag. Our
166 // choice here must be kept in sync with AutoUpgrade, which uses the same
167 // technique to detect whether ftz is enabled.
168 if (auto *Flag = mdconst::extract_or_null<ConstantInt>(
169 F.getParent()->getModuleFlag("nvvm-reflect-ftz")))
170 ReflectVal = Flag->getSExtValue();
171 } else if (ReflectArg == "__CUDA_ARCH") {
172 ReflectVal = SmVersion * 10;
173 }
174 Call->replaceAllUsesWith(ConstantInt::get(Call->getType(), ReflectVal));
175 ToRemove.push_back(Call);
176 }
177
179 I->eraseFromParent();
180
181 return ToRemove.size() > 0;
182}
183
184bool NVVMReflect::runOnFunction(Function &F) {
185 return runNVVMReflect(F, SmVersion);
186}
187
189
192 return runNVVMReflect(F, SmVersion) ? PreservedAnalyses::none()
194}
This file contains the declarations for the subclasses of Constant, which represent the different fla...
#define LLVM_DEBUG(X)
Definition: Debug.h:101
Select target instructions out of generic instructions
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
Module.h This file contains the declarations for the Module class.
#define NVVM_REFLECT_OCL_FUNCTION
Definition: NVVMReflect.cpp:42
#define NVVM_REFLECT_FUNCTION
Definition: NVVMReflect.cpp:41
unsigned SmVersion
Definition: NVVMReflect.cpp:77
static cl::opt< bool > NVVMReflectEnabled("nvvm-reflect-enable", cl::init(true), cl::Hidden, cl::desc("NVVM reflection, enabled by default"))
SmallVector< Instruction *, 4 > ToRemove
Definition: NVVMReflect.cpp:89
This header defines various interfaces for pass management in LLVM.
#define INITIALIZE_PASS(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:38
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file defines the SmallVector class.
A container for analyses that lazily runs them and caches their results.
Definition: PassManager.h:649
This class represents a function call, abstracting a target machine's calling convention.
static Constant * get(Type *Ty, uint64_t V, bool IsSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:888
This is an important base class in LLVM.
Definition: Constant.h:41
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:311
virtual bool runOnFunction(Function &F)=0
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass.
PassRegistry - This class manages the registration and intitialization of the pass subsystem as appli...
Definition: PassRegistry.h:37
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:172
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
Definition: PassManager.h:175
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:178
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1200
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
constexpr StringRef substr(size_t Start, size_t N=npos) const
Return a reference to the substring from [Start, Start + N).
Definition: StringRef.h:575
constexpr size_t size() const
size - Get the string size.
Definition: StringRef.h:137
LLVM Value Representation.
Definition: Value.h:74
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:445
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
FunctionPass * createNVVMReflectPass(unsigned int SmVersion)
Definition: NVVMReflect.cpp:64
void initializeNVVMReflectPass(PassRegistry &)
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)