1 //=== AMDGPUPrintfRuntimeBinding.cpp - OpenCL printf implementation -------===//
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
7 //===----------------------------------------------------------------------===//
10 // The pass bind printfs to a kernel arg pointer that will be bound to a buffer
11 // later by the runtime.
13 // This pass traverses the functions in the module and converts
14 // each call to printf to a sequence of operations that
15 // store the following into the printf buffer:
16 // - format string (passed as a module's metadata unique ID)
17 // - bitwise copies of printf arguments
18 // The backend passes will need to store metadata in the kernel
19 //===----------------------------------------------------------------------===//
22 #include "llvm/ADT/SmallString.h"
23 #include "llvm/ADT/StringExtras.h"
24 #include "llvm/ADT/Triple.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/TargetLibraryInfo.h"
27 #include "llvm/CodeGen/Passes.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/IRBuilder.h"
33 #include "llvm/IR/Instructions.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/Type.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #define DEBUG_TYPE "printfToRuntime"
47 class LLVM_LIBRARY_VISIBILITY AMDGPUPrintfRuntimeBinding final
53 explicit AMDGPUPrintfRuntimeBinding();
56 bool runOnModule(Module &M) override;
57 void getConversionSpecifiers(SmallVectorImpl<char> &OpConvSpecifiers,
58 StringRef fmt, size_t num_ops) const;
60 bool shouldPrintAsStr(char Specifier, Type *OpType) const;
62 lowerPrintfForGpu(Module &M,
63 function_ref<const TargetLibraryInfo &(Function &)> GetTLI);
65 void getAnalysisUsage(AnalysisUsage &AU) const override {
66 AU.addRequired<TargetLibraryInfoWrapperPass>();
67 AU.addRequired<DominatorTreeWrapperPass>();
70 Value *simplify(Instruction *I, const TargetLibraryInfo *TLI) {
71 return SimplifyInstruction(I, {*TD, TLI, DT});
75 const DominatorTree *DT;
76 SmallVector<CallInst *, 32> Printfs;
80 char AMDGPUPrintfRuntimeBinding::ID = 0;
82 INITIALIZE_PASS_BEGIN(AMDGPUPrintfRuntimeBinding,
83 "amdgpu-printf-runtime-binding", "AMDGPU Printf lowering",
85 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
86 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
87 INITIALIZE_PASS_END(AMDGPUPrintfRuntimeBinding, "amdgpu-printf-runtime-binding",
88 "AMDGPU Printf lowering", false, false)
90 char &llvm::AMDGPUPrintfRuntimeBindingID = AMDGPUPrintfRuntimeBinding::ID;
93 ModulePass *createAMDGPUPrintfRuntimeBinding() {
94 return new AMDGPUPrintfRuntimeBinding();
98 AMDGPUPrintfRuntimeBinding::AMDGPUPrintfRuntimeBinding()
99 : ModulePass(ID), TD(nullptr), DT(nullptr) {
100 initializeAMDGPUPrintfRuntimeBindingPass(*PassRegistry::getPassRegistry());
103 void AMDGPUPrintfRuntimeBinding::getConversionSpecifiers(
104 SmallVectorImpl<char> &OpConvSpecifiers, StringRef Fmt,
105 size_t NumOps) const {
106 // not all format characters are collected.
107 // At this time the format characters of interest
108 // are %p and %s, which use to know if we
109 // are either storing a literal string or a
110 // pointer to the printf buffer.
111 static const char ConvSpecifiers[] = "cdieEfgGaosuxXp";
112 size_t CurFmtSpecifierIdx = 0;
113 size_t PrevFmtSpecifierIdx = 0;
115 while ((CurFmtSpecifierIdx = Fmt.find_first_of(
116 ConvSpecifiers, CurFmtSpecifierIdx)) != StringRef::npos) {
117 bool ArgDump = false;
118 StringRef CurFmt = Fmt.substr(PrevFmtSpecifierIdx,
119 CurFmtSpecifierIdx - PrevFmtSpecifierIdx);
120 size_t pTag = CurFmt.find_last_of("%");
121 if (pTag != StringRef::npos) {
123 while (pTag && CurFmt[--pTag] == '%') {
129 OpConvSpecifiers.push_back(Fmt[CurFmtSpecifierIdx]);
131 PrevFmtSpecifierIdx = ++CurFmtSpecifierIdx;
135 bool AMDGPUPrintfRuntimeBinding::shouldPrintAsStr(char Specifier,
136 Type *OpType) const {
137 if (Specifier != 's')
139 const PointerType *PT = dyn_cast<PointerType>(OpType);
140 if (!PT || PT->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
142 Type *ElemType = PT->getContainedType(0);
143 if (ElemType->getTypeID() != Type::IntegerTyID)
145 IntegerType *ElemIType = cast<IntegerType>(ElemType);
146 return ElemIType->getBitWidth() == 8;
149 bool AMDGPUPrintfRuntimeBinding::lowerPrintfForGpu(
150 Module &M, function_ref<const TargetLibraryInfo &(Function &)> GetTLI) {
151 LLVMContext &Ctx = M.getContext();
152 IRBuilder<> Builder(Ctx);
153 Type *I32Ty = Type::getInt32Ty(Ctx);
155 // NB: This is important for this string size to be divizable by 4
156 const char NonLiteralStr[4] = "???";
158 for (auto CI : Printfs) {
159 unsigned NumOps = CI->getNumArgOperands();
161 SmallString<16> OpConvSpecifiers;
162 Value *Op = CI->getArgOperand(0);
164 if (auto LI = dyn_cast<LoadInst>(Op)) {
165 Op = LI->getPointerOperand();
166 for (auto Use : Op->users()) {
167 if (auto SI = dyn_cast<StoreInst>(Use)) {
168 Op = SI->getValueOperand();
174 if (auto I = dyn_cast<Instruction>(Op)) {
175 Value *Op_simplified = simplify(I, &GetTLI(*I->getFunction()));
180 ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Op);
183 GlobalVariable *GVar = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
185 StringRef Str("unknown");
186 if (GVar && GVar->hasInitializer()) {
187 auto Init = GVar->getInitializer();
188 if (auto CA = dyn_cast<ConstantDataArray>(Init)) {
190 Str = CA->getAsCString();
191 } else if (isa<ConstantAggregateZero>(Init)) {
195 // we need this call to ascertain
196 // that we are printing a string
197 // or a pointer. It takes out the
198 // specifiers and fills up the first
200 getConversionSpecifiers(OpConvSpecifiers, Str, NumOps - 1);
202 // Add metadata for the string
203 std::string AStreamHolder;
204 raw_string_ostream Sizes(AStreamHolder);
205 int Sum = DWORD_ALIGN;
206 Sizes << CI->getNumArgOperands() - 1;
208 for (unsigned ArgCount = 1; ArgCount < CI->getNumArgOperands() &&
209 ArgCount <= OpConvSpecifiers.size();
211 Value *Arg = CI->getArgOperand(ArgCount);
212 Type *ArgType = Arg->getType();
213 unsigned ArgSize = TD->getTypeAllocSizeInBits(ArgType);
214 ArgSize = ArgSize / 8;
216 // ArgSize by design should be a multiple of DWORD_ALIGN,
217 // expand the arguments that do not follow this rule.
219 if (ArgSize % DWORD_ALIGN != 0) {
220 llvm::Type *ResType = llvm::Type::getInt32Ty(Ctx);
221 auto *LLVMVecType = llvm::dyn_cast<llvm::FixedVectorType>(ArgType);
222 int NumElem = LLVMVecType ? LLVMVecType->getNumElements() : 1;
223 if (LLVMVecType && NumElem > 1)
224 ResType = llvm::FixedVectorType::get(ResType, NumElem);
225 Builder.SetInsertPoint(CI);
226 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
227 if (OpConvSpecifiers[ArgCount - 1] == 'x' ||
228 OpConvSpecifiers[ArgCount - 1] == 'X' ||
229 OpConvSpecifiers[ArgCount - 1] == 'u' ||
230 OpConvSpecifiers[ArgCount - 1] == 'o')
231 Arg = Builder.CreateZExt(Arg, ResType);
233 Arg = Builder.CreateSExt(Arg, ResType);
234 ArgType = Arg->getType();
235 ArgSize = TD->getTypeAllocSizeInBits(ArgType);
236 ArgSize = ArgSize / 8;
237 CI->setOperand(ArgCount, Arg);
239 if (OpConvSpecifiers[ArgCount - 1] == 'f') {
240 ConstantFP *FpCons = dyn_cast<ConstantFP>(Arg);
244 FPExtInst *FpExt = dyn_cast<FPExtInst>(Arg);
245 if (FpExt && FpExt->getType()->isDoubleTy() &&
246 FpExt->getOperand(0)->getType()->isFloatTy())
250 if (shouldPrintAsStr(OpConvSpecifiers[ArgCount - 1], ArgType)) {
251 if (ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Arg)) {
253 dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
254 if (GV && GV->hasInitializer()) {
255 Constant *Init = GV->getInitializer();
256 ConstantDataArray *CA = dyn_cast<ConstantDataArray>(Init);
257 if (Init->isZeroValue() || CA->isString()) {
258 size_t SizeStr = Init->isZeroValue()
260 : (strlen(CA->getAsCString().data()) + 1);
261 size_t Rem = SizeStr % DWORD_ALIGN;
263 LLVM_DEBUG(dbgs() << "Printf string original size = " << SizeStr
266 NSizeStr = SizeStr + (DWORD_ALIGN - Rem);
273 ArgSize = sizeof(NonLiteralStr);
276 ArgSize = sizeof(NonLiteralStr);
279 LLVM_DEBUG(dbgs() << "Printf ArgSize (in buffer) = " << ArgSize
280 << " for type: " << *ArgType << '\n');
281 Sizes << ArgSize << ':';
284 LLVM_DEBUG(dbgs() << "Printf format string in source = " << Str.str()
286 for (size_t I = 0; I < Str.size(); ++I) {
287 // Rest of the C escape sequences (e.g. \') are handled correctly
309 // ':' cannot be scanned by Flex, as it is defined as a delimiter
310 // Replace it with it's octal representation \72
319 // Insert the printf_alloc call
320 Builder.SetInsertPoint(CI);
321 Builder.SetCurrentDebugLocation(CI->getDebugLoc());
323 AttributeList Attr = AttributeList::get(Ctx, AttributeList::FunctionIndex,
324 Attribute::NoUnwind);
326 Type *SizetTy = Type::getInt32Ty(Ctx);
328 Type *Tys_alloc[1] = {SizetTy};
329 Type *I8Ptr = PointerType::get(Type::getInt8Ty(Ctx), 1);
330 FunctionType *FTy_alloc = FunctionType::get(I8Ptr, Tys_alloc, false);
331 FunctionCallee PrintfAllocFn =
332 M.getOrInsertFunction(StringRef("__printf_alloc"), FTy_alloc, Attr);
334 LLVM_DEBUG(dbgs() << "Printf metadata = " << Sizes.str() << '\n');
335 std::string fmtstr = itostr(++UniqID) + ":" + Sizes.str().c_str();
336 MDString *fmtStrArray = MDString::get(Ctx, fmtstr);
338 // Instead of creating global variables, the
339 // printf format strings are extracted
340 // and passed as metadata. This avoids
341 // polluting llvm's symbol tables in this module.
342 // Metadata is going to be extracted
343 // by the backend passes and inserted
344 // into the OpenCL binary as appropriate.
345 StringRef amd("llvm.printf.fmts");
346 NamedMDNode *metaD = M.getOrInsertNamedMetadata(amd);
347 MDNode *myMD = MDNode::get(Ctx, fmtStrArray);
348 metaD->addOperand(myMD);
349 Value *sumC = ConstantInt::get(SizetTy, Sum, false);
350 SmallVector<Value *, 1> alloc_args;
351 alloc_args.push_back(sumC);
353 CallInst::Create(PrintfAllocFn, alloc_args, "printf_alloc_fn", CI);
356 // Insert code to split basicblock with a
357 // piece of hammock code.
358 // basicblock splits after buffer overflow check
360 ConstantPointerNull *zeroIntPtr =
361 ConstantPointerNull::get(PointerType::get(Type::getInt8Ty(Ctx), 1));
363 dyn_cast<ICmpInst>(Builder.CreateICmpNE(pcall, zeroIntPtr, ""));
364 if (!CI->use_empty()) {
366 Builder.CreateSExt(Builder.CreateNot(cmp), I32Ty, "printf_res");
367 CI->replaceAllUsesWith(result);
369 SplitBlock(CI->getParent(), cmp);
371 SplitBlockAndInsertIfThen(cmp, cmp->getNextNode(), false);
373 Builder.SetInsertPoint(Brnch);
375 // store unique printf id in the buffer
377 SmallVector<Value *, 1> ZeroIdxList;
378 ConstantInt *zeroInt =
379 ConstantInt::get(Ctx, APInt(32, StringRef("0"), 10));
380 ZeroIdxList.push_back(zeroInt);
382 GetElementPtrInst *BufferIdx =
383 dyn_cast<GetElementPtrInst>(GetElementPtrInst::Create(
384 nullptr, pcall, ZeroIdxList, "PrintBuffID", Brnch));
386 Type *idPointer = PointerType::get(I32Ty, AMDGPUAS::GLOBAL_ADDRESS);
388 new BitCastInst(BufferIdx, idPointer, "PrintBuffIdCast", Brnch);
390 new StoreInst(ConstantInt::get(I32Ty, UniqID), id_gep_cast, Brnch);
392 SmallVector<Value *, 2> FourthIdxList;
393 ConstantInt *fourInt =
394 ConstantInt::get(Ctx, APInt(32, StringRef("4"), 10));
396 FourthIdxList.push_back(fourInt); // 1st 4 bytes hold the printf_id
397 // the following GEP is the buffer pointer
398 BufferIdx = cast<GetElementPtrInst>(GetElementPtrInst::Create(
399 nullptr, pcall, FourthIdxList, "PrintBuffGep", Brnch));
401 Type *Int32Ty = Type::getInt32Ty(Ctx);
402 Type *Int64Ty = Type::getInt64Ty(Ctx);
403 for (unsigned ArgCount = 1; ArgCount < CI->getNumArgOperands() &&
404 ArgCount <= OpConvSpecifiers.size();
406 Value *Arg = CI->getArgOperand(ArgCount);
407 Type *ArgType = Arg->getType();
408 SmallVector<Value *, 32> WhatToStore;
409 if (ArgType->isFPOrFPVectorTy() && !isa<VectorType>(ArgType)) {
410 Type *IType = (ArgType->isFloatTy()) ? Int32Ty : Int64Ty;
411 if (OpConvSpecifiers[ArgCount - 1] == 'f') {
412 ConstantFP *fpCons = dyn_cast<ConstantFP>(Arg);
414 APFloat Val(fpCons->getValueAPF());
416 Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
418 Arg = ConstantFP::get(Ctx, Val);
421 FPExtInst *FpExt = dyn_cast<FPExtInst>(Arg);
422 if (FpExt && FpExt->getType()->isDoubleTy() &&
423 FpExt->getOperand(0)->getType()->isFloatTy()) {
424 Arg = FpExt->getOperand(0);
429 Arg = new BitCastInst(Arg, IType, "PrintArgFP", Brnch);
430 WhatToStore.push_back(Arg);
431 } else if (ArgType->getTypeID() == Type::PointerTyID) {
432 if (shouldPrintAsStr(OpConvSpecifiers[ArgCount - 1], ArgType)) {
433 const char *S = NonLiteralStr;
434 if (ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Arg)) {
436 dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
437 if (GV && GV->hasInitializer()) {
438 Constant *Init = GV->getInitializer();
439 ConstantDataArray *CA = dyn_cast<ConstantDataArray>(Init);
440 if (Init->isZeroValue() || CA->isString()) {
441 S = Init->isZeroValue() ? "" : CA->getAsCString().data();
445 size_t SizeStr = strlen(S) + 1;
446 size_t Rem = SizeStr % DWORD_ALIGN;
449 NSizeStr = SizeStr + (DWORD_ALIGN - Rem);
454 char *MyNewStr = new char[NSizeStr]();
456 int NumInts = NSizeStr / 4;
459 int ANum = *(int *)(MyNewStr + CharC);
462 Value *ANumV = ConstantInt::get(Int32Ty, ANum, false);
463 WhatToStore.push_back(ANumV);
467 // Empty string, give a hint to RT it is no NULL
468 Value *ANumV = ConstantInt::get(Int32Ty, 0xFFFFFF00, false);
469 WhatToStore.push_back(ANumV);
472 uint64_t Size = TD->getTypeAllocSizeInBits(ArgType);
473 assert((Size == 32 || Size == 64) && "unsupported size");
474 Type *DstType = (Size == 32) ? Int32Ty : Int64Ty;
475 Arg = new PtrToIntInst(Arg, DstType, "PrintArgPtr", Brnch);
476 WhatToStore.push_back(Arg);
478 } else if (isa<FixedVectorType>(ArgType)) {
480 uint32_t EleCount = cast<FixedVectorType>(ArgType)->getNumElements();
481 uint32_t EleSize = ArgType->getScalarSizeInBits();
482 uint32_t TotalSize = EleCount * EleSize;
484 ShuffleVectorInst *Shuffle =
485 new ShuffleVectorInst(Arg, Arg, ArrayRef<int>{0, 1, 2, 2});
486 Shuffle->insertBefore(Brnch);
488 ArgType = Arg->getType();
489 TotalSize += EleSize;
493 EleCount = TotalSize / 64;
494 IType = dyn_cast<Type>(Type::getInt64Ty(ArgType->getContext()));
498 EleCount = TotalSize / 64;
499 IType = dyn_cast<Type>(Type::getInt64Ty(ArgType->getContext()));
500 } else if (EleCount >= 3) {
502 IType = dyn_cast<Type>(Type::getInt32Ty(ArgType->getContext()));
505 IType = dyn_cast<Type>(Type::getInt16Ty(ArgType->getContext()));
510 EleCount = TotalSize / 64;
511 IType = dyn_cast<Type>(Type::getInt64Ty(ArgType->getContext()));
514 IType = dyn_cast<Type>(Type::getInt32Ty(ArgType->getContext()));
519 IType = FixedVectorType::get(IType, EleCount);
521 Arg = new BitCastInst(Arg, IType, "PrintArgVect", Brnch);
522 WhatToStore.push_back(Arg);
524 WhatToStore.push_back(Arg);
526 for (unsigned I = 0, E = WhatToStore.size(); I != E; ++I) {
527 Value *TheBtCast = WhatToStore[I];
529 TD->getTypeAllocSizeInBits(TheBtCast->getType()) / 8;
530 SmallVector<Value *, 1> BuffOffset;
531 BuffOffset.push_back(ConstantInt::get(I32Ty, ArgSize));
533 Type *ArgPointer = PointerType::get(TheBtCast->getType(), 1);
535 new BitCastInst(BufferIdx, ArgPointer, "PrintBuffPtrCast", Brnch);
536 StoreInst *StBuff = new StoreInst(TheBtCast, CastedGEP, Brnch);
537 LLVM_DEBUG(dbgs() << "inserting store to printf buffer:\n"
540 if (I + 1 == E && ArgCount + 1 == CI->getNumArgOperands())
542 BufferIdx = dyn_cast<GetElementPtrInst>(GetElementPtrInst::Create(
543 nullptr, BufferIdx, BuffOffset, "PrintBuffNextPtr", Brnch));
544 LLVM_DEBUG(dbgs() << "inserting gep to the printf buffer:\n"
545 << *BufferIdx << '\n');
551 // erase the printf calls
552 for (auto CI : Printfs)
553 CI->eraseFromParent();
559 bool AMDGPUPrintfRuntimeBinding::runOnModule(Module &M) {
560 Triple TT(M.getTargetTriple());
561 if (TT.getArch() == Triple::r600)
564 auto PrintfFunction = M.getFunction("printf");
568 for (auto &U : PrintfFunction->uses()) {
569 if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
570 if (CI->isCallee(&U))
571 Printfs.push_back(CI);
578 if (auto HostcallFunction = M.getFunction("__ockl_hostcall_internal")) {
579 for (auto &U : HostcallFunction->uses()) {
580 if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
581 M.getContext().emitError(
582 CI, "Cannot use both printf and hostcall in the same module");
587 TD = &M.getDataLayout();
588 auto DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
589 DT = DTWP ? &DTWP->getDomTree() : nullptr;
590 auto GetTLI = [this](Function &F) -> TargetLibraryInfo & {
591 return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
594 return lowerPrintfForGpu(M, GetTLI);