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/PassManager.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/InitializePasses.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 using namespace llvm;
44 
45 #define DEBUG_TYPE "printfToRuntime"
46 #define DWORD_ALIGN 4
47 
48 namespace {
49 class AMDGPUPrintfRuntimeBinding final : public ModulePass {
50 
51 public:
52   static char ID;
53 
54   explicit AMDGPUPrintfRuntimeBinding();
55 
56 private:
57   bool runOnModule(Module &M) override;
58 
59   void getAnalysisUsage(AnalysisUsage &AU) const override {
60     AU.addRequired<TargetLibraryInfoWrapperPass>();
61     AU.addRequired<DominatorTreeWrapperPass>();
62   }
63 };
64 
65 class AMDGPUPrintfRuntimeBindingImpl {
66 public:
67   AMDGPUPrintfRuntimeBindingImpl(
68       function_ref<const DominatorTree &(Function &)> GetDT,
69       function_ref<const TargetLibraryInfo &(Function &)> GetTLI)
70       : GetDT(GetDT), GetTLI(GetTLI) {}
71   bool run(Module &M);
72 
73 private:
74   void getConversionSpecifiers(SmallVectorImpl<char> &OpConvSpecifiers,
75                                StringRef fmt, size_t num_ops) const;
76 
77   bool shouldPrintAsStr(char Specifier, Type *OpType) const;
78   bool lowerPrintfForGpu(Module &M);
79 
80   Value *simplify(Instruction *I, const TargetLibraryInfo *TLI,
81                   const DominatorTree *DT) {
82     return SimplifyInstruction(I, {*TD, TLI, DT});
83   }
84 
85   const DataLayout *TD;
86   function_ref<const DominatorTree &(Function &)> GetDT;
87   function_ref<const TargetLibraryInfo &(Function &)> GetTLI;
88   SmallVector<CallInst *, 32> Printfs;
89 };
90 } // namespace
91 
92 char AMDGPUPrintfRuntimeBinding::ID = 0;
93 
94 INITIALIZE_PASS_BEGIN(AMDGPUPrintfRuntimeBinding,
95                       "amdgpu-printf-runtime-binding", "AMDGPU Printf lowering",
96                       false, false)
97 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
98 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
99 INITIALIZE_PASS_END(AMDGPUPrintfRuntimeBinding, "amdgpu-printf-runtime-binding",
100                     "AMDGPU Printf lowering", false, false)
101 
102 char &llvm::AMDGPUPrintfRuntimeBindingID = AMDGPUPrintfRuntimeBinding::ID;
103 
104 namespace llvm {
105 ModulePass *createAMDGPUPrintfRuntimeBinding() {
106   return new AMDGPUPrintfRuntimeBinding();
107 }
108 } // namespace llvm
109 
110 AMDGPUPrintfRuntimeBinding::AMDGPUPrintfRuntimeBinding() : ModulePass(ID) {
111   initializeAMDGPUPrintfRuntimeBindingPass(*PassRegistry::getPassRegistry());
112 }
113 
114 void AMDGPUPrintfRuntimeBindingImpl::getConversionSpecifiers(
115     SmallVectorImpl<char> &OpConvSpecifiers, StringRef Fmt,
116     size_t NumOps) const {
117   // not all format characters are collected.
118   // At this time the format characters of interest
119   // are %p and %s, which use to know if we
120   // are either storing a literal string or a
121   // pointer to the printf buffer.
122   static const char ConvSpecifiers[] = "cdieEfgGaosuxXp";
123   size_t CurFmtSpecifierIdx = 0;
124   size_t PrevFmtSpecifierIdx = 0;
125 
126   while ((CurFmtSpecifierIdx = Fmt.find_first_of(
127               ConvSpecifiers, CurFmtSpecifierIdx)) != StringRef::npos) {
128     bool ArgDump = false;
129     StringRef CurFmt = Fmt.substr(PrevFmtSpecifierIdx,
130                                   CurFmtSpecifierIdx - PrevFmtSpecifierIdx);
131     size_t pTag = CurFmt.find_last_of("%");
132     if (pTag != StringRef::npos) {
133       ArgDump = true;
134       while (pTag && CurFmt[--pTag] == '%') {
135         ArgDump = !ArgDump;
136       }
137     }
138 
139     if (ArgDump)
140       OpConvSpecifiers.push_back(Fmt[CurFmtSpecifierIdx]);
141 
142     PrevFmtSpecifierIdx = ++CurFmtSpecifierIdx;
143   }
144 }
145 
146 bool AMDGPUPrintfRuntimeBindingImpl::shouldPrintAsStr(char Specifier,
147                                                       Type *OpType) const {
148   if (Specifier != 's')
149     return false;
150   const PointerType *PT = dyn_cast<PointerType>(OpType);
151   if (!PT || PT->getAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
152     return false;
153   Type *ElemType = PT->getContainedType(0);
154   if (ElemType->getTypeID() != Type::IntegerTyID)
155     return false;
156   IntegerType *ElemIType = cast<IntegerType>(ElemType);
157   return ElemIType->getBitWidth() == 8;
158 }
159 
160 bool AMDGPUPrintfRuntimeBindingImpl::lowerPrintfForGpu(Module &M) {
161   LLVMContext &Ctx = M.getContext();
162   IRBuilder<> Builder(Ctx);
163   Type *I32Ty = Type::getInt32Ty(Ctx);
164   unsigned UniqID = 0;
165   // NB: This is important for this string size to be divizable by 4
166   const char NonLiteralStr[4] = "???";
167 
168   for (auto CI : Printfs) {
169     unsigned NumOps = CI->getNumArgOperands();
170 
171     SmallString<16> OpConvSpecifiers;
172     Value *Op = CI->getArgOperand(0);
173 
174     if (auto LI = dyn_cast<LoadInst>(Op)) {
175       Op = LI->getPointerOperand();
176       for (auto Use : Op->users()) {
177         if (auto SI = dyn_cast<StoreInst>(Use)) {
178           Op = SI->getValueOperand();
179           break;
180         }
181       }
182     }
183 
184     if (auto I = dyn_cast<Instruction>(Op)) {
185       Value *Op_simplified =
186           simplify(I, &GetTLI(*I->getFunction()), &GetDT(*I->getFunction()));
187       if (Op_simplified)
188         Op = Op_simplified;
189     }
190 
191     ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Op);
192 
193     if (ConstExpr) {
194       GlobalVariable *GVar = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
195 
196       StringRef Str("unknown");
197       if (GVar && GVar->hasInitializer()) {
198         auto Init = GVar->getInitializer();
199         if (auto CA = dyn_cast<ConstantDataArray>(Init)) {
200           if (CA->isString())
201             Str = CA->getAsCString();
202         } else if (isa<ConstantAggregateZero>(Init)) {
203           Str = "";
204         }
205         //
206         // we need this call to ascertain
207         // that we are printing a string
208         // or a pointer. It takes out the
209         // specifiers and fills up the first
210         // arg
211         getConversionSpecifiers(OpConvSpecifiers, Str, NumOps - 1);
212       }
213       // Add metadata for the string
214       std::string AStreamHolder;
215       raw_string_ostream Sizes(AStreamHolder);
216       int Sum = DWORD_ALIGN;
217       Sizes << CI->getNumArgOperands() - 1;
218       Sizes << ':';
219       for (unsigned ArgCount = 1; ArgCount < CI->getNumArgOperands() &&
220                                   ArgCount <= OpConvSpecifiers.size();
221            ArgCount++) {
222         Value *Arg = CI->getArgOperand(ArgCount);
223         Type *ArgType = Arg->getType();
224         unsigned ArgSize = TD->getTypeAllocSizeInBits(ArgType);
225         ArgSize = ArgSize / 8;
226         //
227         // ArgSize by design should be a multiple of DWORD_ALIGN,
228         // expand the arguments that do not follow this rule.
229         //
230         if (ArgSize % DWORD_ALIGN != 0) {
231           llvm::Type *ResType = llvm::Type::getInt32Ty(Ctx);
232           auto *LLVMVecType = llvm::dyn_cast<llvm::FixedVectorType>(ArgType);
233           int NumElem = LLVMVecType ? LLVMVecType->getNumElements() : 1;
234           if (LLVMVecType && NumElem > 1)
235             ResType = llvm::FixedVectorType::get(ResType, NumElem);
236           Builder.SetInsertPoint(CI);
237           Builder.SetCurrentDebugLocation(CI->getDebugLoc());
238           if (OpConvSpecifiers[ArgCount - 1] == 'x' ||
239               OpConvSpecifiers[ArgCount - 1] == 'X' ||
240               OpConvSpecifiers[ArgCount - 1] == 'u' ||
241               OpConvSpecifiers[ArgCount - 1] == 'o')
242             Arg = Builder.CreateZExt(Arg, ResType);
243           else
244             Arg = Builder.CreateSExt(Arg, ResType);
245           ArgType = Arg->getType();
246           ArgSize = TD->getTypeAllocSizeInBits(ArgType);
247           ArgSize = ArgSize / 8;
248           CI->setOperand(ArgCount, Arg);
249         }
250         if (OpConvSpecifiers[ArgCount - 1] == 'f') {
251           ConstantFP *FpCons = dyn_cast<ConstantFP>(Arg);
252           if (FpCons)
253             ArgSize = 4;
254           else {
255             FPExtInst *FpExt = dyn_cast<FPExtInst>(Arg);
256             if (FpExt && FpExt->getType()->isDoubleTy() &&
257                 FpExt->getOperand(0)->getType()->isFloatTy())
258               ArgSize = 4;
259           }
260         }
261         if (shouldPrintAsStr(OpConvSpecifiers[ArgCount - 1], ArgType)) {
262           if (ConstantExpr *ConstExpr = dyn_cast<ConstantExpr>(Arg)) {
263             GlobalVariable *GV =
264                 dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
265             if (GV && GV->hasInitializer()) {
266               Constant *Init = GV->getInitializer();
267               ConstantDataArray *CA = dyn_cast<ConstantDataArray>(Init);
268               if (Init->isZeroValue() || CA->isString()) {
269                 size_t SizeStr = Init->isZeroValue()
270                                      ? 1
271                                      : (strlen(CA->getAsCString().data()) + 1);
272                 size_t Rem = SizeStr % DWORD_ALIGN;
273                 size_t NSizeStr = 0;
274                 LLVM_DEBUG(dbgs() << "Printf string original size = " << SizeStr
275                                   << '\n');
276                 if (Rem) {
277                   NSizeStr = SizeStr + (DWORD_ALIGN - Rem);
278                 } else {
279                   NSizeStr = SizeStr;
280                 }
281                 ArgSize = NSizeStr;
282               }
283             } else {
284               ArgSize = sizeof(NonLiteralStr);
285             }
286           } else {
287             ArgSize = sizeof(NonLiteralStr);
288           }
289         }
290         LLVM_DEBUG(dbgs() << "Printf ArgSize (in buffer) = " << ArgSize
291                           << " for type: " << *ArgType << '\n');
292         Sizes << ArgSize << ':';
293         Sum += ArgSize;
294       }
295       LLVM_DEBUG(dbgs() << "Printf format string in source = " << Str.str()
296                         << '\n');
297       for (size_t I = 0; I < Str.size(); ++I) {
298         // Rest of the C escape sequences (e.g. \') are handled correctly
299         // by the MDParser
300         switch (Str[I]) {
301         case '\a':
302           Sizes << "\\a";
303           break;
304         case '\b':
305           Sizes << "\\b";
306           break;
307         case '\f':
308           Sizes << "\\f";
309           break;
310         case '\n':
311           Sizes << "\\n";
312           break;
313         case '\r':
314           Sizes << "\\r";
315           break;
316         case '\v':
317           Sizes << "\\v";
318           break;
319         case ':':
320           // ':' cannot be scanned by Flex, as it is defined as a delimiter
321           // Replace it with it's octal representation \72
322           Sizes << "\\72";
323           break;
324         default:
325           Sizes << Str[I];
326           break;
327         }
328       }
329 
330       // Insert the printf_alloc call
331       Builder.SetInsertPoint(CI);
332       Builder.SetCurrentDebugLocation(CI->getDebugLoc());
333 
334       AttributeList Attr = AttributeList::get(Ctx, AttributeList::FunctionIndex,
335                                               Attribute::NoUnwind);
336 
337       Type *SizetTy = Type::getInt32Ty(Ctx);
338 
339       Type *Tys_alloc[1] = {SizetTy};
340       Type *I8Ptr = PointerType::get(Type::getInt8Ty(Ctx), 1);
341       FunctionType *FTy_alloc = FunctionType::get(I8Ptr, Tys_alloc, false);
342       FunctionCallee PrintfAllocFn =
343           M.getOrInsertFunction(StringRef("__printf_alloc"), FTy_alloc, Attr);
344 
345       LLVM_DEBUG(dbgs() << "Printf metadata = " << Sizes.str() << '\n');
346       std::string fmtstr = itostr(++UniqID) + ":" + Sizes.str().c_str();
347       MDString *fmtStrArray = MDString::get(Ctx, fmtstr);
348 
349       // Instead of creating global variables, the
350       // printf format strings are extracted
351       // and passed as metadata. This avoids
352       // polluting llvm's symbol tables in this module.
353       // Metadata is going to be extracted
354       // by the backend passes and inserted
355       // into the OpenCL binary as appropriate.
356       StringRef amd("llvm.printf.fmts");
357       NamedMDNode *metaD = M.getOrInsertNamedMetadata(amd);
358       MDNode *myMD = MDNode::get(Ctx, fmtStrArray);
359       metaD->addOperand(myMD);
360       Value *sumC = ConstantInt::get(SizetTy, Sum, false);
361       SmallVector<Value *, 1> alloc_args;
362       alloc_args.push_back(sumC);
363       CallInst *pcall =
364           CallInst::Create(PrintfAllocFn, alloc_args, "printf_alloc_fn", CI);
365 
366       //
367       // Insert code to split basicblock with a
368       // piece of hammock code.
369       // basicblock splits after buffer overflow check
370       //
371       ConstantPointerNull *zeroIntPtr =
372           ConstantPointerNull::get(PointerType::get(Type::getInt8Ty(Ctx), 1));
373       ICmpInst *cmp =
374           dyn_cast<ICmpInst>(Builder.CreateICmpNE(pcall, zeroIntPtr, ""));
375       if (!CI->use_empty()) {
376         Value *result =
377             Builder.CreateSExt(Builder.CreateNot(cmp), I32Ty, "printf_res");
378         CI->replaceAllUsesWith(result);
379       }
380       SplitBlock(CI->getParent(), cmp);
381       Instruction *Brnch =
382           SplitBlockAndInsertIfThen(cmp, cmp->getNextNode(), false);
383 
384       Builder.SetInsertPoint(Brnch);
385 
386       // store unique printf id in the buffer
387       //
388       SmallVector<Value *, 1> ZeroIdxList;
389       ConstantInt *zeroInt =
390           ConstantInt::get(Ctx, APInt(32, StringRef("0"), 10));
391       ZeroIdxList.push_back(zeroInt);
392 
393       GetElementPtrInst *BufferIdx = GetElementPtrInst::Create(
394           nullptr, pcall, ZeroIdxList, "PrintBuffID", Brnch);
395 
396       Type *idPointer = PointerType::get(I32Ty, AMDGPUAS::GLOBAL_ADDRESS);
397       Value *id_gep_cast =
398           new BitCastInst(BufferIdx, idPointer, "PrintBuffIdCast", Brnch);
399 
400       new StoreInst(ConstantInt::get(I32Ty, UniqID), id_gep_cast, Brnch);
401 
402       SmallVector<Value *, 2> FourthIdxList;
403       ConstantInt *fourInt =
404           ConstantInt::get(Ctx, APInt(32, StringRef("4"), 10));
405 
406       FourthIdxList.push_back(fourInt); // 1st 4 bytes hold the printf_id
407       // the following GEP is the buffer pointer
408       BufferIdx = GetElementPtrInst::Create(nullptr, pcall, FourthIdxList,
409                                             "PrintBuffGep", Brnch);
410 
411       Type *Int32Ty = Type::getInt32Ty(Ctx);
412       Type *Int64Ty = Type::getInt64Ty(Ctx);
413       for (unsigned ArgCount = 1; ArgCount < CI->getNumArgOperands() &&
414                                   ArgCount <= OpConvSpecifiers.size();
415            ArgCount++) {
416         Value *Arg = CI->getArgOperand(ArgCount);
417         Type *ArgType = Arg->getType();
418         SmallVector<Value *, 32> WhatToStore;
419         if (ArgType->isFPOrFPVectorTy() && !isa<VectorType>(ArgType)) {
420           Type *IType = (ArgType->isFloatTy()) ? Int32Ty : Int64Ty;
421           if (OpConvSpecifiers[ArgCount - 1] == 'f') {
422             if (auto *FpCons = dyn_cast<ConstantFP>(Arg)) {
423               APFloat Val(FpCons->getValueAPF());
424               bool Lost = false;
425               Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
426                           &Lost);
427               Arg = ConstantFP::get(Ctx, Val);
428               IType = Int32Ty;
429             } else if (auto *FpExt = dyn_cast<FPExtInst>(Arg)) {
430               if (FpExt->getType()->isDoubleTy() &&
431                   FpExt->getOperand(0)->getType()->isFloatTy()) {
432                 Arg = FpExt->getOperand(0);
433                 IType = Int32Ty;
434               }
435             }
436           }
437           Arg = new BitCastInst(Arg, IType, "PrintArgFP", Brnch);
438           WhatToStore.push_back(Arg);
439         } else if (ArgType->getTypeID() == Type::PointerTyID) {
440           if (shouldPrintAsStr(OpConvSpecifiers[ArgCount - 1], ArgType)) {
441             const char *S = NonLiteralStr;
442             if (auto *ConstExpr = dyn_cast<ConstantExpr>(Arg)) {
443               auto *GV = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
444               if (GV && GV->hasInitializer()) {
445                 Constant *Init = GV->getInitializer();
446                 ConstantDataArray *CA = dyn_cast<ConstantDataArray>(Init);
447                 if (Init->isZeroValue() || CA->isString()) {
448                   S = Init->isZeroValue() ? "" : CA->getAsCString().data();
449                 }
450               }
451             }
452             size_t SizeStr = strlen(S) + 1;
453             size_t Rem = SizeStr % DWORD_ALIGN;
454             size_t NSizeStr = 0;
455             if (Rem) {
456               NSizeStr = SizeStr + (DWORD_ALIGN - Rem);
457             } else {
458               NSizeStr = SizeStr;
459             }
460             if (S[0]) {
461               char *MyNewStr = new char[NSizeStr]();
462               strcpy(MyNewStr, S);
463               int NumInts = NSizeStr / 4;
464               int CharC = 0;
465               while (NumInts) {
466                 int ANum = *(int *)(MyNewStr + CharC);
467                 CharC += 4;
468                 NumInts--;
469                 Value *ANumV = ConstantInt::get(Int32Ty, ANum, false);
470                 WhatToStore.push_back(ANumV);
471               }
472               delete[] MyNewStr;
473             } else {
474               // Empty string, give a hint to RT it is no NULL
475               Value *ANumV = ConstantInt::get(Int32Ty, 0xFFFFFF00, false);
476               WhatToStore.push_back(ANumV);
477             }
478           } else {
479             uint64_t Size = TD->getTypeAllocSizeInBits(ArgType);
480             assert((Size == 32 || Size == 64) && "unsupported size");
481             Type *DstType = (Size == 32) ? Int32Ty : Int64Ty;
482             Arg = new PtrToIntInst(Arg, DstType, "PrintArgPtr", Brnch);
483             WhatToStore.push_back(Arg);
484           }
485         } else if (isa<FixedVectorType>(ArgType)) {
486           Type *IType = NULL;
487           uint32_t EleCount = cast<FixedVectorType>(ArgType)->getNumElements();
488           uint32_t EleSize = ArgType->getScalarSizeInBits();
489           uint32_t TotalSize = EleCount * EleSize;
490           if (EleCount == 3) {
491             ShuffleVectorInst *Shuffle =
492                 new ShuffleVectorInst(Arg, Arg, ArrayRef<int>{0, 1, 2, 2});
493             Shuffle->insertBefore(Brnch);
494             Arg = Shuffle;
495             ArgType = Arg->getType();
496             TotalSize += EleSize;
497           }
498           switch (EleSize) {
499           default:
500             EleCount = TotalSize / 64;
501             IType = Type::getInt64Ty(ArgType->getContext());
502             break;
503           case 8:
504             if (EleCount >= 8) {
505               EleCount = TotalSize / 64;
506               IType = Type::getInt64Ty(ArgType->getContext());
507             } else if (EleCount >= 3) {
508               EleCount = 1;
509               IType = Type::getInt32Ty(ArgType->getContext());
510             } else {
511               EleCount = 1;
512               IType = Type::getInt16Ty(ArgType->getContext());
513             }
514             break;
515           case 16:
516             if (EleCount >= 3) {
517               EleCount = TotalSize / 64;
518               IType = Type::getInt64Ty(ArgType->getContext());
519             } else {
520               EleCount = 1;
521               IType = Type::getInt32Ty(ArgType->getContext());
522             }
523             break;
524           }
525           if (EleCount > 1) {
526             IType = FixedVectorType::get(IType, EleCount);
527           }
528           Arg = new BitCastInst(Arg, IType, "PrintArgVect", Brnch);
529           WhatToStore.push_back(Arg);
530         } else {
531           WhatToStore.push_back(Arg);
532         }
533         for (unsigned I = 0, E = WhatToStore.size(); I != E; ++I) {
534           Value *TheBtCast = WhatToStore[I];
535           unsigned ArgSize =
536               TD->getTypeAllocSizeInBits(TheBtCast->getType()) / 8;
537           SmallVector<Value *, 1> BuffOffset;
538           BuffOffset.push_back(ConstantInt::get(I32Ty, ArgSize));
539 
540           Type *ArgPointer = PointerType::get(TheBtCast->getType(), 1);
541           Value *CastedGEP =
542               new BitCastInst(BufferIdx, ArgPointer, "PrintBuffPtrCast", Brnch);
543           StoreInst *StBuff = new StoreInst(TheBtCast, CastedGEP, Brnch);
544           LLVM_DEBUG(dbgs() << "inserting store to printf buffer:\n"
545                             << *StBuff << '\n');
546           (void)StBuff;
547           if (I + 1 == E && ArgCount + 1 == CI->getNumArgOperands())
548             break;
549           BufferIdx = GetElementPtrInst::Create(nullptr, BufferIdx, BuffOffset,
550                                                 "PrintBuffNextPtr", Brnch);
551           LLVM_DEBUG(dbgs() << "inserting gep to the printf buffer:\n"
552                             << *BufferIdx << '\n');
553         }
554       }
555     }
556   }
557 
558   // erase the printf calls
559   for (auto CI : Printfs)
560     CI->eraseFromParent();
561 
562   Printfs.clear();
563   return true;
564 }
565 
566 bool AMDGPUPrintfRuntimeBindingImpl::run(Module &M) {
567   Triple TT(M.getTargetTriple());
568   if (TT.getArch() == Triple::r600)
569     return false;
570 
571   auto PrintfFunction = M.getFunction("printf");
572   if (!PrintfFunction)
573     return false;
574 
575   for (auto &U : PrintfFunction->uses()) {
576     if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
577       if (CI->isCallee(&U))
578         Printfs.push_back(CI);
579     }
580   }
581 
582   if (Printfs.empty())
583     return false;
584 
585   if (auto HostcallFunction = M.getFunction("__ockl_hostcall_internal")) {
586     for (auto &U : HostcallFunction->uses()) {
587       if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
588         M.getContext().emitError(
589             CI, "Cannot use both printf and hostcall in the same module");
590       }
591     }
592   }
593 
594   TD = &M.getDataLayout();
595 
596   return lowerPrintfForGpu(M);
597 }
598 
599 bool AMDGPUPrintfRuntimeBinding::runOnModule(Module &M) {
600   auto GetDT = [this](Function &F) -> DominatorTree & {
601     return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
602   };
603   auto GetTLI = [this](Function &F) -> TargetLibraryInfo & {
604     return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
605   };
606 
607   return AMDGPUPrintfRuntimeBindingImpl(GetDT, GetTLI).run(M);
608 }
609 
610 PreservedAnalyses
611 AMDGPUPrintfRuntimeBindingPass::run(Module &M, ModuleAnalysisManager &AM) {
612   FunctionAnalysisManager &FAM =
613       AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
614   auto GetDT = [&FAM](Function &F) -> DominatorTree & {
615     return FAM.getResult<DominatorTreeAnalysis>(F);
616   };
617   auto GetTLI = [&FAM](Function &F) -> TargetLibraryInfo & {
618     return FAM.getResult<TargetLibraryAnalysis>(F);
619   };
620   bool Changed = AMDGPUPrintfRuntimeBindingImpl(GetDT, GetTLI).run(M);
621   return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
622 }
623