xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUPrintfRuntimeBinding.cpp (revision e40139ff33b48b56a24c808b166b04b8ee6f5b21)
1 //=== AMDGPUPrintfRuntimeBinding.cpp - OpenCL printf implementation -------===//
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 // \file
9 //
10 // The pass bind printfs to a kernel arg pointer that will be bound to a buffer
11 // later by the runtime.
12 //
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 //===----------------------------------------------------------------------===//
20 
21 #include "AMDGPU.h"
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/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
40 using namespace llvm;
41 
42 #define DEBUG_TYPE "printfToRuntime"
43 #define DWORD_ALIGN 4
44 
45 namespace {
46 class LLVM_LIBRARY_VISIBILITY AMDGPUPrintfRuntimeBinding final
47     : public ModulePass {
48 
49 public:
50   static char ID;
51 
52   explicit AMDGPUPrintfRuntimeBinding();
53 
54 private:
55   bool runOnModule(Module &M) override;
56   void getConversionSpecifiers(SmallVectorImpl<char> &OpConvSpecifiers,
57                                StringRef fmt, size_t num_ops) const;
58 
59   bool shouldPrintAsStr(char Specifier, Type *OpType) const;
60   bool
61   lowerPrintfForGpu(Module &M,
62                     function_ref<const TargetLibraryInfo &(Function &)> GetTLI);
63 
64   void getAnalysisUsage(AnalysisUsage &AU) const override {
65     AU.addRequired<TargetLibraryInfoWrapperPass>();
66     AU.addRequired<DominatorTreeWrapperPass>();
67   }
68 
69   Value *simplify(Instruction *I, const TargetLibraryInfo *TLI) {
70     return SimplifyInstruction(I, {*TD, TLI, DT});
71   }
72 
73   const DataLayout *TD;
74   const DominatorTree *DT;
75   SmallVector<CallInst *, 32> Printfs;
76 };
77 } // namespace
78 
79 char AMDGPUPrintfRuntimeBinding::ID = 0;
80 
81 INITIALIZE_PASS_BEGIN(AMDGPUPrintfRuntimeBinding,
82                       "amdgpu-printf-runtime-binding", "AMDGPU Printf lowering",
83                       false, false)
84 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
85 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
86 INITIALIZE_PASS_END(AMDGPUPrintfRuntimeBinding, "amdgpu-printf-runtime-binding",
87                     "AMDGPU Printf lowering", false, false)
88 
89 char &llvm::AMDGPUPrintfRuntimeBindingID = AMDGPUPrintfRuntimeBinding::ID;
90 
91 namespace llvm {
92 ModulePass *createAMDGPUPrintfRuntimeBinding() {
93   return new AMDGPUPrintfRuntimeBinding();
94 }
95 } // namespace llvm
96 
97 AMDGPUPrintfRuntimeBinding::AMDGPUPrintfRuntimeBinding()
98     : ModulePass(ID), TD(nullptr), DT(nullptr) {
99   initializeAMDGPUPrintfRuntimeBindingPass(*PassRegistry::getPassRegistry());
100 }
101 
102 void AMDGPUPrintfRuntimeBinding::getConversionSpecifiers(
103     SmallVectorImpl<char> &OpConvSpecifiers, StringRef Fmt,
104     size_t NumOps) const {
105   // not all format characters are collected.
106   // At this time the format characters of interest
107   // are %p and %s, which use to know if we
108   // are either storing a literal string or a
109   // pointer to the printf buffer.
110   static const char ConvSpecifiers[] = "cdieEfgGaosuxXp";
111   size_t CurFmtSpecifierIdx = 0;
112   size_t PrevFmtSpecifierIdx = 0;
113 
114   while ((CurFmtSpecifierIdx = Fmt.find_first_of(
115               ConvSpecifiers, CurFmtSpecifierIdx)) != StringRef::npos) {
116     bool ArgDump = false;
117     StringRef CurFmt = Fmt.substr(PrevFmtSpecifierIdx,
118                                   CurFmtSpecifierIdx - PrevFmtSpecifierIdx);
119     size_t pTag = CurFmt.find_last_of("%");
120     if (pTag != StringRef::npos) {
121       ArgDump = true;
122       while (pTag && CurFmt[--pTag] == '%') {
123         ArgDump = !ArgDump;
124       }
125     }
126 
127     if (ArgDump)
128       OpConvSpecifiers.push_back(Fmt[CurFmtSpecifierIdx]);
129 
130     PrevFmtSpecifierIdx = ++CurFmtSpecifierIdx;
131   }
132 }
133 
134 bool AMDGPUPrintfRuntimeBinding::shouldPrintAsStr(char Specifier,
135                                                   Type *OpType) const {
136   if (Specifier != 's')
137     return false;
138   const PointerType *PT = dyn_cast<PointerType>(OpType);
139   if (!PT || PT->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
140     return false;
141   Type *ElemType = PT->getContainedType(0);
142   if (ElemType->getTypeID() != Type::IntegerTyID)
143     return false;
144   IntegerType *ElemIType = cast<IntegerType>(ElemType);
145   return ElemIType->getBitWidth() == 8;
146 }
147 
148 bool AMDGPUPrintfRuntimeBinding::lowerPrintfForGpu(
149     Module &M, function_ref<const TargetLibraryInfo &(Function &)> GetTLI) {
150   LLVMContext &Ctx = M.getContext();
151   IRBuilder<> Builder(Ctx);
152   Type *I32Ty = Type::getInt32Ty(Ctx);
153   unsigned UniqID = 0;
154   // NB: This is important for this string size to be divizable by 4
155   const char NonLiteralStr[4] = "???";
156 
157   for (auto CI : Printfs) {
158     unsigned NumOps = CI->getNumArgOperands();
159 
160     SmallString<16> OpConvSpecifiers;
161     Value *Op = CI->getArgOperand(0);
162 
163     if (auto LI = dyn_cast<LoadInst>(Op)) {
164       Op = LI->getPointerOperand();
165       for (auto Use : Op->users()) {
166         if (auto SI = dyn_cast<StoreInst>(Use)) {
167           Op = SI->getValueOperand();
168           break;
169         }
170       }
171     }
172 
173     if (auto I = dyn_cast<Instruction>(Op)) {
174       Value *Op_simplified = simplify(I, &GetTLI(*I->getFunction()));
175       if (Op_simplified)
176         Op = Op_simplified;
177     }
178 
179     ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Op);
180 
181     if (ConstExpr) {
182       GlobalVariable *GVar = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
183 
184       StringRef Str("unknown");
185       if (GVar && GVar->hasInitializer()) {
186         auto Init = GVar->getInitializer();
187         if (auto CA = dyn_cast<ConstantDataArray>(Init)) {
188           if (CA->isString())
189             Str = CA->getAsCString();
190         } else if (isa<ConstantAggregateZero>(Init)) {
191           Str = "";
192         }
193         //
194         // we need this call to ascertain
195         // that we are printing a string
196         // or a pointer. It takes out the
197         // specifiers and fills up the first
198         // arg
199         getConversionSpecifiers(OpConvSpecifiers, Str, NumOps - 1);
200       }
201       // Add metadata for the string
202       std::string AStreamHolder;
203       raw_string_ostream Sizes(AStreamHolder);
204       int Sum = DWORD_ALIGN;
205       Sizes << CI->getNumArgOperands() - 1;
206       Sizes << ':';
207       for (unsigned ArgCount = 1; ArgCount < CI->getNumArgOperands() &&
208                                   ArgCount <= OpConvSpecifiers.size();
209            ArgCount++) {
210         Value *Arg = CI->getArgOperand(ArgCount);
211         Type *ArgType = Arg->getType();
212         unsigned ArgSize = TD->getTypeAllocSizeInBits(ArgType);
213         ArgSize = ArgSize / 8;
214         //
215         // ArgSize by design should be a multiple of DWORD_ALIGN,
216         // expand the arguments that do not follow this rule.
217         //
218         if (ArgSize % DWORD_ALIGN != 0) {
219           llvm::Type *ResType = llvm::Type::getInt32Ty(Ctx);
220           VectorType *LLVMVecType = llvm::dyn_cast<llvm::VectorType>(ArgType);
221           int NumElem = LLVMVecType ? LLVMVecType->getNumElements() : 1;
222           if (LLVMVecType && NumElem > 1)
223             ResType = llvm::VectorType::get(ResType, NumElem);
224           Builder.SetInsertPoint(CI);
225           Builder.SetCurrentDebugLocation(CI->getDebugLoc());
226           if (OpConvSpecifiers[ArgCount - 1] == 'x' ||
227               OpConvSpecifiers[ArgCount - 1] == 'X' ||
228               OpConvSpecifiers[ArgCount - 1] == 'u' ||
229               OpConvSpecifiers[ArgCount - 1] == 'o')
230             Arg = Builder.CreateZExt(Arg, ResType);
231           else
232             Arg = Builder.CreateSExt(Arg, ResType);
233           ArgType = Arg->getType();
234           ArgSize = TD->getTypeAllocSizeInBits(ArgType);
235           ArgSize = ArgSize / 8;
236           CI->setOperand(ArgCount, Arg);
237         }
238         if (OpConvSpecifiers[ArgCount - 1] == 'f') {
239           ConstantFP *FpCons = dyn_cast<ConstantFP>(Arg);
240           if (FpCons)
241             ArgSize = 4;
242           else {
243             FPExtInst *FpExt = dyn_cast<FPExtInst>(Arg);
244             if (FpExt && FpExt->getType()->isDoubleTy() &&
245                 FpExt->getOperand(0)->getType()->isFloatTy())
246               ArgSize = 4;
247           }
248         }
249         if (shouldPrintAsStr(OpConvSpecifiers[ArgCount - 1], ArgType)) {
250           if (ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Arg)) {
251             GlobalVariable *GV =
252                 dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
253             if (GV && GV->hasInitializer()) {
254               Constant *Init = GV->getInitializer();
255               ConstantDataArray *CA = dyn_cast<ConstantDataArray>(Init);
256               if (Init->isZeroValue() || CA->isString()) {
257                 size_t SizeStr = Init->isZeroValue()
258                                      ? 1
259                                      : (strlen(CA->getAsCString().data()) + 1);
260                 size_t Rem = SizeStr % DWORD_ALIGN;
261                 size_t NSizeStr = 0;
262                 LLVM_DEBUG(dbgs() << "Printf string original size = " << SizeStr
263                                   << '\n');
264                 if (Rem) {
265                   NSizeStr = SizeStr + (DWORD_ALIGN - Rem);
266                 } else {
267                   NSizeStr = SizeStr;
268                 }
269                 ArgSize = NSizeStr;
270               }
271             } else {
272               ArgSize = sizeof(NonLiteralStr);
273             }
274           } else {
275             ArgSize = sizeof(NonLiteralStr);
276           }
277         }
278         LLVM_DEBUG(dbgs() << "Printf ArgSize (in buffer) = " << ArgSize
279                           << " for type: " << *ArgType << '\n');
280         Sizes << ArgSize << ':';
281         Sum += ArgSize;
282       }
283       LLVM_DEBUG(dbgs() << "Printf format string in source = " << Str.str()
284                         << '\n');
285       for (size_t I = 0; I < Str.size(); ++I) {
286         // Rest of the C escape sequences (e.g. \') are handled correctly
287         // by the MDParser
288         switch (Str[I]) {
289         case '\a':
290           Sizes << "\\a";
291           break;
292         case '\b':
293           Sizes << "\\b";
294           break;
295         case '\f':
296           Sizes << "\\f";
297           break;
298         case '\n':
299           Sizes << "\\n";
300           break;
301         case '\r':
302           Sizes << "\\r";
303           break;
304         case '\v':
305           Sizes << "\\v";
306           break;
307         case ':':
308           // ':' cannot be scanned by Flex, as it is defined as a delimiter
309           // Replace it with it's octal representation \72
310           Sizes << "\\72";
311           break;
312         default:
313           Sizes << Str[I];
314           break;
315         }
316       }
317 
318       // Insert the printf_alloc call
319       Builder.SetInsertPoint(CI);
320       Builder.SetCurrentDebugLocation(CI->getDebugLoc());
321 
322       AttributeList Attr = AttributeList::get(Ctx, AttributeList::FunctionIndex,
323                                               Attribute::NoUnwind);
324 
325       Type *SizetTy = Type::getInt32Ty(Ctx);
326 
327       Type *Tys_alloc[1] = {SizetTy};
328       Type *I8Ptr = PointerType::get(Type::getInt8Ty(Ctx), 1);
329       FunctionType *FTy_alloc = FunctionType::get(I8Ptr, Tys_alloc, false);
330       FunctionCallee PrintfAllocFn =
331           M.getOrInsertFunction(StringRef("__printf_alloc"), FTy_alloc, Attr);
332 
333       LLVM_DEBUG(dbgs() << "Printf metadata = " << Sizes.str() << '\n');
334       std::string fmtstr = itostr(++UniqID) + ":" + Sizes.str().c_str();
335       MDString *fmtStrArray = MDString::get(Ctx, fmtstr);
336 
337       // Instead of creating global variables, the
338       // printf format strings are extracted
339       // and passed as metadata. This avoids
340       // polluting llvm's symbol tables in this module.
341       // Metadata is going to be extracted
342       // by the backend passes and inserted
343       // into the OpenCL binary as appropriate.
344       StringRef amd("llvm.printf.fmts");
345       NamedMDNode *metaD = M.getOrInsertNamedMetadata(amd);
346       MDNode *myMD = MDNode::get(Ctx, fmtStrArray);
347       metaD->addOperand(myMD);
348       Value *sumC = ConstantInt::get(SizetTy, Sum, false);
349       SmallVector<Value *, 1> alloc_args;
350       alloc_args.push_back(sumC);
351       CallInst *pcall =
352           CallInst::Create(PrintfAllocFn, alloc_args, "printf_alloc_fn", CI);
353 
354       //
355       // Insert code to split basicblock with a
356       // piece of hammock code.
357       // basicblock splits after buffer overflow check
358       //
359       ConstantPointerNull *zeroIntPtr =
360           ConstantPointerNull::get(PointerType::get(Type::getInt8Ty(Ctx), 1));
361       ICmpInst *cmp =
362           dyn_cast<ICmpInst>(Builder.CreateICmpNE(pcall, zeroIntPtr, ""));
363       if (!CI->use_empty()) {
364         Value *result =
365             Builder.CreateSExt(Builder.CreateNot(cmp), I32Ty, "printf_res");
366         CI->replaceAllUsesWith(result);
367       }
368       SplitBlock(CI->getParent(), cmp);
369       Instruction *Brnch =
370           SplitBlockAndInsertIfThen(cmp, cmp->getNextNode(), false);
371 
372       Builder.SetInsertPoint(Brnch);
373 
374       // store unique printf id in the buffer
375       //
376       SmallVector<Value *, 1> ZeroIdxList;
377       ConstantInt *zeroInt =
378           ConstantInt::get(Ctx, APInt(32, StringRef("0"), 10));
379       ZeroIdxList.push_back(zeroInt);
380 
381       GetElementPtrInst *BufferIdx =
382           dyn_cast<GetElementPtrInst>(GetElementPtrInst::Create(
383               nullptr, pcall, ZeroIdxList, "PrintBuffID", Brnch));
384 
385       Type *idPointer = PointerType::get(I32Ty, AMDGPUAS::GLOBAL_ADDRESS);
386       Value *id_gep_cast =
387           new BitCastInst(BufferIdx, idPointer, "PrintBuffIdCast", Brnch);
388 
389       StoreInst *stbuff =
390           new StoreInst(ConstantInt::get(I32Ty, UniqID), id_gep_cast);
391       stbuff->insertBefore(Brnch); // to Remove unused variable warning
392 
393       SmallVector<Value *, 2> FourthIdxList;
394       ConstantInt *fourInt =
395           ConstantInt::get(Ctx, APInt(32, StringRef("4"), 10));
396 
397       FourthIdxList.push_back(fourInt); // 1st 4 bytes hold the printf_id
398       // the following GEP is the buffer pointer
399       BufferIdx = cast<GetElementPtrInst>(GetElementPtrInst::Create(
400           nullptr, pcall, FourthIdxList, "PrintBuffGep", Brnch));
401 
402       Type *Int32Ty = Type::getInt32Ty(Ctx);
403       Type *Int64Ty = Type::getInt64Ty(Ctx);
404       for (unsigned ArgCount = 1; ArgCount < CI->getNumArgOperands() &&
405                                   ArgCount <= OpConvSpecifiers.size();
406            ArgCount++) {
407         Value *Arg = CI->getArgOperand(ArgCount);
408         Type *ArgType = Arg->getType();
409         SmallVector<Value *, 32> WhatToStore;
410         if (ArgType->isFPOrFPVectorTy() &&
411             (ArgType->getTypeID() != Type::VectorTyID)) {
412           Type *IType = (ArgType->isFloatTy()) ? Int32Ty : Int64Ty;
413           if (OpConvSpecifiers[ArgCount - 1] == 'f') {
414             ConstantFP *fpCons = dyn_cast<ConstantFP>(Arg);
415             if (fpCons) {
416               APFloat Val(fpCons->getValueAPF());
417               bool Lost = false;
418               Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
419                           &Lost);
420               Arg = ConstantFP::get(Ctx, Val);
421               IType = Int32Ty;
422             } else {
423               FPExtInst *FpExt = dyn_cast<FPExtInst>(Arg);
424               if (FpExt && FpExt->getType()->isDoubleTy() &&
425                   FpExt->getOperand(0)->getType()->isFloatTy()) {
426                 Arg = FpExt->getOperand(0);
427                 IType = Int32Ty;
428               }
429             }
430           }
431           Arg = new BitCastInst(Arg, IType, "PrintArgFP", Brnch);
432           WhatToStore.push_back(Arg);
433         } else if (ArgType->getTypeID() == Type::PointerTyID) {
434           if (shouldPrintAsStr(OpConvSpecifiers[ArgCount - 1], ArgType)) {
435             const char *S = NonLiteralStr;
436             if (ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Arg)) {
437               GlobalVariable *GV =
438                   dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
439               if (GV && GV->hasInitializer()) {
440                 Constant *Init = GV->getInitializer();
441                 ConstantDataArray *CA = dyn_cast<ConstantDataArray>(Init);
442                 if (Init->isZeroValue() || CA->isString()) {
443                   S = Init->isZeroValue() ? "" : CA->getAsCString().data();
444                 }
445               }
446             }
447             size_t SizeStr = strlen(S) + 1;
448             size_t Rem = SizeStr % DWORD_ALIGN;
449             size_t NSizeStr = 0;
450             if (Rem) {
451               NSizeStr = SizeStr + (DWORD_ALIGN - Rem);
452             } else {
453               NSizeStr = SizeStr;
454             }
455             if (S[0]) {
456               char *MyNewStr = new char[NSizeStr]();
457               strcpy(MyNewStr, S);
458               int NumInts = NSizeStr / 4;
459               int CharC = 0;
460               while (NumInts) {
461                 int ANum = *(int *)(MyNewStr + CharC);
462                 CharC += 4;
463                 NumInts--;
464                 Value *ANumV = ConstantInt::get(Int32Ty, ANum, false);
465                 WhatToStore.push_back(ANumV);
466               }
467               delete[] MyNewStr;
468             } else {
469               // Empty string, give a hint to RT it is no NULL
470               Value *ANumV = ConstantInt::get(Int32Ty, 0xFFFFFF00, false);
471               WhatToStore.push_back(ANumV);
472             }
473           } else {
474             uint64_t Size = TD->getTypeAllocSizeInBits(ArgType);
475             assert((Size == 32 || Size == 64) && "unsupported size");
476             Type *DstType = (Size == 32) ? Int32Ty : Int64Ty;
477             Arg = new PtrToIntInst(Arg, DstType, "PrintArgPtr", Brnch);
478             WhatToStore.push_back(Arg);
479           }
480         } else if (ArgType->getTypeID() == Type::VectorTyID) {
481           Type *IType = NULL;
482           uint32_t EleCount = cast<VectorType>(ArgType)->getNumElements();
483           uint32_t EleSize = ArgType->getScalarSizeInBits();
484           uint32_t TotalSize = EleCount * EleSize;
485           if (EleCount == 3) {
486             IntegerType *Int32Ty = Type::getInt32Ty(ArgType->getContext());
487             Constant *Indices[4] = {
488                 ConstantInt::get(Int32Ty, 0), ConstantInt::get(Int32Ty, 1),
489                 ConstantInt::get(Int32Ty, 2), ConstantInt::get(Int32Ty, 2)};
490             Constant *Mask = ConstantVector::get(Indices);
491             ShuffleVectorInst *Shuffle = new ShuffleVectorInst(Arg, Arg, Mask);
492             Shuffle->insertBefore(Brnch);
493             Arg = Shuffle;
494             ArgType = Arg->getType();
495             TotalSize += EleSize;
496           }
497           switch (EleSize) {
498           default:
499             EleCount = TotalSize / 64;
500             IType = dyn_cast<Type>(Type::getInt64Ty(ArgType->getContext()));
501             break;
502           case 8:
503             if (EleCount >= 8) {
504               EleCount = TotalSize / 64;
505               IType = dyn_cast<Type>(Type::getInt64Ty(ArgType->getContext()));
506             } else if (EleCount >= 3) {
507               EleCount = 1;
508               IType = dyn_cast<Type>(Type::getInt32Ty(ArgType->getContext()));
509             } else {
510               EleCount = 1;
511               IType = dyn_cast<Type>(Type::getInt16Ty(ArgType->getContext()));
512             }
513             break;
514           case 16:
515             if (EleCount >= 3) {
516               EleCount = TotalSize / 64;
517               IType = dyn_cast<Type>(Type::getInt64Ty(ArgType->getContext()));
518             } else {
519               EleCount = 1;
520               IType = dyn_cast<Type>(Type::getInt32Ty(ArgType->getContext()));
521             }
522             break;
523           }
524           if (EleCount > 1) {
525             IType = dyn_cast<Type>(VectorType::get(IType, EleCount));
526           }
527           Arg = new BitCastInst(Arg, IType, "PrintArgVect", Brnch);
528           WhatToStore.push_back(Arg);
529         } else {
530           WhatToStore.push_back(Arg);
531         }
532         for (unsigned I = 0, E = WhatToStore.size(); I != E; ++I) {
533           Value *TheBtCast = WhatToStore[I];
534           unsigned ArgSize =
535               TD->getTypeAllocSizeInBits(TheBtCast->getType()) / 8;
536           SmallVector<Value *, 1> BuffOffset;
537           BuffOffset.push_back(ConstantInt::get(I32Ty, ArgSize));
538 
539           Type *ArgPointer = PointerType::get(TheBtCast->getType(), 1);
540           Value *CastedGEP =
541               new BitCastInst(BufferIdx, ArgPointer, "PrintBuffPtrCast", Brnch);
542           StoreInst *StBuff = new StoreInst(TheBtCast, CastedGEP, Brnch);
543           LLVM_DEBUG(dbgs() << "inserting store to printf buffer:\n"
544                             << *StBuff << '\n');
545           (void)StBuff;
546           if (I + 1 == E && ArgCount + 1 == CI->getNumArgOperands())
547             break;
548           BufferIdx = dyn_cast<GetElementPtrInst>(GetElementPtrInst::Create(
549               nullptr, BufferIdx, BuffOffset, "PrintBuffNextPtr", Brnch));
550           LLVM_DEBUG(dbgs() << "inserting gep to the printf buffer:\n"
551                             << *BufferIdx << '\n');
552         }
553       }
554     }
555   }
556 
557   // erase the printf calls
558   for (auto CI : Printfs)
559     CI->eraseFromParent();
560 
561   Printfs.clear();
562   return true;
563 }
564 
565 bool AMDGPUPrintfRuntimeBinding::runOnModule(Module &M) {
566   Triple TT(M.getTargetTriple());
567   if (TT.getArch() == Triple::r600)
568     return false;
569 
570   auto PrintfFunction = M.getFunction("printf");
571   if (!PrintfFunction)
572     return false;
573 
574   for (auto &U : PrintfFunction->uses()) {
575     if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
576       if (CI->isCallee(&U))
577         Printfs.push_back(CI);
578     }
579   }
580 
581   if (Printfs.empty())
582     return false;
583 
584   TD = &M.getDataLayout();
585   auto DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
586   DT = DTWP ? &DTWP->getDomTree() : nullptr;
587   auto GetTLI = [this](Function &F) -> TargetLibraryInfo & {
588     return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
589   };
590 
591   return lowerPrintfForGpu(M, GetTLI);
592 }
593