1 //===-- AutoUpgrade.cpp - Implement auto-upgrade helper functions ---------===//
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 file implements the auto-upgrade helper functions.
11 // This is where deprecated IR intrinsics and other IR features are updated to
12 // current specifications.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/IR/AutoUpgrade.h"
17 #include "llvm/IR/CFG.h"
18 #include "llvm/IR/CallSite.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/DIBuilder.h"
21 #include "llvm/IR/DebugInfo.h"
22 #include "llvm/IR/DiagnosticInfo.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/Instruction.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/Regex.h"
31 #include <cstring>
32 using namespace llvm;
33 
34 // Upgrade the declarations of the SSE4.1 functions whose arguments have
35 // changed their type from v4f32 to v2i64.
36 static bool UpgradeSSE41Function(Function* F, Intrinsic::ID IID,
37                                  Function *&NewFn) {
38   // Check whether this is an old version of the function, which received
39   // v4f32 arguments.
40   Type *Arg0Type = F->getFunctionType()->getParamType(0);
41   if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
42     return false;
43 
44   // Yes, it's old, replace it with new version.
45   F->setName(F->getName() + ".old");
46   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
47   return true;
48 }
49 
50 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
51 // arguments have changed their type from i32 to i8.
52 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
53                                              Function *&NewFn) {
54   // Check that the last argument is an i32.
55   Type *LastArgType = F->getFunctionType()->getParamType(
56      F->getFunctionType()->getNumParams() - 1);
57   if (!LastArgType->isIntegerTy(32))
58     return false;
59 
60   // Move this function aside and map down.
61   F->setName(F->getName() + ".old");
62   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
63   return true;
64 }
65 
66 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
67   assert(F && "Illegal to upgrade a non-existent Function.");
68 
69   // Quickly eliminate it, if it's not a candidate.
70   StringRef Name = F->getName();
71   if (Name.size() <= 8 || !Name.startswith("llvm."))
72     return false;
73   Name = Name.substr(5); // Strip off "llvm."
74 
75   switch (Name[0]) {
76   default: break;
77   case 'a': {
78     if (Name.startswith("arm.neon.vclz")) {
79       Type* args[2] = {
80         F->arg_begin()->getType(),
81         Type::getInt1Ty(F->getContext())
82       };
83       // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
84       // the end of the name. Change name from llvm.arm.neon.vclz.* to
85       //  llvm.ctlz.*
86       FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
87       NewFn = Function::Create(fType, F->getLinkage(),
88                                "llvm.ctlz." + Name.substr(14), F->getParent());
89       return true;
90     }
91     if (Name.startswith("arm.neon.vcnt")) {
92       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
93                                         F->arg_begin()->getType());
94       return true;
95     }
96     Regex vldRegex("^arm\\.neon\\.vld([1234]|[234]lane)\\.v[a-z0-9]*$");
97     if (vldRegex.match(Name)) {
98       auto fArgs = F->getFunctionType()->params();
99       SmallVector<Type *, 4> Tys(fArgs.begin(), fArgs.end());
100       // Can't use Intrinsic::getDeclaration here as the return types might
101       // then only be structurally equal.
102       FunctionType* fType = FunctionType::get(F->getReturnType(), Tys, false);
103       NewFn = Function::Create(fType, F->getLinkage(),
104                                "llvm." + Name + ".p0i8", F->getParent());
105       return true;
106     }
107     Regex vstRegex("^arm\\.neon\\.vst([1234]|[234]lane)\\.v[a-z0-9]*$");
108     if (vstRegex.match(Name)) {
109       static const Intrinsic::ID StoreInts[] = {Intrinsic::arm_neon_vst1,
110                                                 Intrinsic::arm_neon_vst2,
111                                                 Intrinsic::arm_neon_vst3,
112                                                 Intrinsic::arm_neon_vst4};
113 
114       static const Intrinsic::ID StoreLaneInts[] = {
115         Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
116         Intrinsic::arm_neon_vst4lane
117       };
118 
119       auto fArgs = F->getFunctionType()->params();
120       Type *Tys[] = {fArgs[0], fArgs[1]};
121       if (Name.find("lane") == StringRef::npos)
122         NewFn = Intrinsic::getDeclaration(F->getParent(),
123                                           StoreInts[fArgs.size() - 3], Tys);
124       else
125         NewFn = Intrinsic::getDeclaration(F->getParent(),
126                                           StoreLaneInts[fArgs.size() - 5], Tys);
127       return true;
128     }
129     if (Name == "aarch64.thread.pointer" || Name == "arm.thread.pointer") {
130       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
131       return true;
132     }
133     break;
134   }
135 
136   case 'c': {
137     if (Name.startswith("ctlz.") && F->arg_size() == 1) {
138       F->setName(Name + ".old");
139       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
140                                         F->arg_begin()->getType());
141       return true;
142     }
143     if (Name.startswith("cttz.") && F->arg_size() == 1) {
144       F->setName(Name + ".old");
145       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
146                                         F->arg_begin()->getType());
147       return true;
148     }
149     break;
150   }
151 
152   case 'o':
153     // We only need to change the name to match the mangling including the
154     // address space.
155     if (F->arg_size() == 2 && Name.startswith("objectsize.")) {
156       Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
157       if (F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
158         F->setName(Name + ".old");
159         NewFn = Intrinsic::getDeclaration(F->getParent(),
160                                           Intrinsic::objectsize, Tys);
161         return true;
162       }
163     }
164     break;
165 
166   case 's':
167     if (Name == "stackprotectorcheck") {
168       NewFn = nullptr;
169       return true;
170     }
171 
172   case 'x': {
173     if (Name.startswith("x86.sse2.pcmpeq.") ||
174         Name.startswith("x86.sse2.pcmpgt.") ||
175         Name.startswith("x86.avx2.pcmpeq.") ||
176         Name.startswith("x86.avx2.pcmpgt.") ||
177         Name.startswith("x86.avx2.vbroadcast") ||
178         Name.startswith("x86.avx2.pbroadcast") ||
179         Name.startswith("x86.avx.vpermil.") ||
180         Name.startswith("x86.sse41.pmovsx") ||
181         Name.startswith("x86.sse41.pmovzx") ||
182         Name.startswith("x86.avx2.pmovsx") ||
183         Name.startswith("x86.avx2.pmovzx") ||
184         Name == "x86.sse2.cvtdq2pd" ||
185         Name == "x86.sse2.cvtps2pd" ||
186         Name == "x86.avx.cvtdq2.pd.256" ||
187         Name == "x86.avx.cvt.ps2.pd.256" ||
188         Name.startswith("x86.avx.vinsertf128.") ||
189         Name == "x86.avx2.vinserti128" ||
190         Name.startswith("x86.avx.vextractf128.") ||
191         Name == "x86.avx2.vextracti128" ||
192         Name.startswith("x86.avx.movnt.") ||
193         Name == "x86.sse2.storel.dq" ||
194         Name.startswith("x86.sse.storeu.") ||
195         Name.startswith("x86.sse2.storeu.") ||
196         Name.startswith("x86.avx.storeu.") ||
197         Name.startswith("x86.avx512.mask.storeu.p") ||
198         Name.startswith("x86.avx512.mask.storeu.b.") ||
199         Name.startswith("x86.avx512.mask.storeu.w.") ||
200         Name.startswith("x86.avx512.mask.storeu.d.") ||
201         Name.startswith("x86.avx512.mask.storeu.q.") ||
202         Name.startswith("x86.avx512.mask.store.p") ||
203         Name.startswith("x86.avx512.mask.store.b.") ||
204         Name.startswith("x86.avx512.mask.store.w.") ||
205         Name.startswith("x86.avx512.mask.store.d.") ||
206         Name.startswith("x86.avx512.mask.store.q.") ||
207         Name == "x86.sse42.crc32.64.8" ||
208         Name.startswith("x86.avx.vbroadcast.s") ||
209         Name.startswith("x86.sse2.psll.dq") ||
210         Name.startswith("x86.sse2.psrl.dq") ||
211         Name.startswith("x86.avx2.psll.dq") ||
212         Name.startswith("x86.avx2.psrl.dq") ||
213         Name == "x86.sse41.pblendw" ||
214         Name.startswith("x86.sse41.blendp") ||
215         Name.startswith("x86.avx.blend.p") ||
216         Name == "x86.avx2.pblendw" ||
217         Name.startswith("x86.avx2.pblendd.") ||
218         Name == "x86.avx2.vbroadcasti128" ||
219         Name == "x86.xop.vpcmov" ||
220         (Name.startswith("x86.xop.vpcom") && F->arg_size() == 2)) {
221       NewFn = nullptr;
222       return true;
223     }
224     // SSE4.1 ptest functions may have an old signature.
225     if (Name.startswith("x86.sse41.ptest")) {
226       if (Name == "x86.sse41.ptestc")
227         return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestc, NewFn);
228       if (Name == "x86.sse41.ptestz")
229         return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestz, NewFn);
230       if (Name == "x86.sse41.ptestnzc")
231         return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
232     }
233     // Several blend and other instructions with masks used the wrong number of
234     // bits.
235     if (Name == "x86.sse41.insertps")
236       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
237                                               NewFn);
238     if (Name == "x86.sse41.dppd")
239       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
240                                               NewFn);
241     if (Name == "x86.sse41.dpps")
242       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
243                                               NewFn);
244     if (Name == "x86.sse41.mpsadbw")
245       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
246                                               NewFn);
247     if (Name == "x86.avx.dp.ps.256")
248       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
249                                               NewFn);
250     if (Name == "x86.avx2.mpsadbw")
251       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
252                                               NewFn);
253 
254     // frcz.ss/sd may need to have an argument dropped
255     if (Name.startswith("x86.xop.vfrcz.ss") && F->arg_size() == 2) {
256       F->setName(Name + ".old");
257       NewFn = Intrinsic::getDeclaration(F->getParent(),
258                                         Intrinsic::x86_xop_vfrcz_ss);
259       return true;
260     }
261     if (Name.startswith("x86.xop.vfrcz.sd") && F->arg_size() == 2) {
262       F->setName(Name + ".old");
263       NewFn = Intrinsic::getDeclaration(F->getParent(),
264                                         Intrinsic::x86_xop_vfrcz_sd);
265       return true;
266     }
267     // Fix the FMA4 intrinsics to remove the 4
268     if (Name.startswith("x86.fma4.")) {
269       F->setName("llvm.x86.fma" + Name.substr(8));
270       NewFn = F;
271       return true;
272     }
273     break;
274   }
275   }
276 
277   //  This may not belong here. This function is effectively being overloaded
278   //  to both detect an intrinsic which needs upgrading, and to provide the
279   //  upgraded form of the intrinsic. We should perhaps have two separate
280   //  functions for this.
281   return false;
282 }
283 
284 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
285   NewFn = nullptr;
286   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
287   assert(F != NewFn && "Intrinsic function upgraded to the same function");
288 
289   // Upgrade intrinsic attributes.  This does not change the function.
290   if (NewFn)
291     F = NewFn;
292   if (Intrinsic::ID id = F->getIntrinsicID())
293     F->setAttributes(Intrinsic::getAttributes(F->getContext(), id));
294   return Upgraded;
295 }
296 
297 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
298   // Nothing to do yet.
299   return false;
300 }
301 
302 // Handles upgrading SSE2 and AVX2 PSLLDQ intrinsics by converting them
303 // to byte shuffles.
304 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder, LLVMContext &C,
305                                          Value *Op, unsigned Shift) {
306   Type *ResultTy = Op->getType();
307   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
308 
309   // Bitcast from a 64-bit element type to a byte element type.
310   Type *VecTy = VectorType::get(Type::getInt8Ty(C), NumElts);
311   Op = Builder.CreateBitCast(Op, VecTy, "cast");
312 
313   // We'll be shuffling in zeroes.
314   Value *Res = Constant::getNullValue(VecTy);
315 
316   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
317   // we'll just return the zero vector.
318   if (Shift < 16) {
319     int Idxs[32];
320     // 256-bit version is split into two 16-byte lanes.
321     for (unsigned l = 0; l != NumElts; l += 16)
322       for (unsigned i = 0; i != 16; ++i) {
323         unsigned Idx = NumElts + i - Shift;
324         if (Idx < NumElts)
325           Idx -= NumElts - 16; // end of lane, switch operand.
326         Idxs[l + i] = Idx + l;
327       }
328 
329     Res = Builder.CreateShuffleVector(Res, Op, makeArrayRef(Idxs, NumElts));
330   }
331 
332   // Bitcast back to a 64-bit element type.
333   return Builder.CreateBitCast(Res, ResultTy, "cast");
334 }
335 
336 // Handles upgrading SSE2 and AVX2 PSRLDQ intrinsics by converting them
337 // to byte shuffles.
338 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, LLVMContext &C,
339                                          Value *Op,
340                                          unsigned Shift) {
341   Type *ResultTy = Op->getType();
342   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
343 
344   // Bitcast from a 64-bit element type to a byte element type.
345   Type *VecTy = VectorType::get(Type::getInt8Ty(C), NumElts);
346   Op = Builder.CreateBitCast(Op, VecTy, "cast");
347 
348   // We'll be shuffling in zeroes.
349   Value *Res = Constant::getNullValue(VecTy);
350 
351   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
352   // we'll just return the zero vector.
353   if (Shift < 16) {
354     int Idxs[32];
355     // 256-bit version is split into two 16-byte lanes.
356     for (unsigned l = 0; l != NumElts; l += 16)
357       for (unsigned i = 0; i != 16; ++i) {
358         unsigned Idx = i + Shift;
359         if (Idx >= 16)
360           Idx += NumElts - 16; // end of lane, switch operand.
361         Idxs[l + i] = Idx + l;
362       }
363 
364     Res = Builder.CreateShuffleVector(Op, Res, makeArrayRef(Idxs, NumElts));
365   }
366 
367   // Bitcast back to a 64-bit element type.
368   return Builder.CreateBitCast(Res, ResultTy, "cast");
369 }
370 
371 static Value *UpgradeMaskedStore(IRBuilder<> &Builder, LLVMContext &C,
372                                  Value *Ptr, Value *Data, Value *Mask,
373                                  bool Aligned) {
374   // Cast the pointer to the right type.
375   Ptr = Builder.CreateBitCast(Ptr,
376                               llvm::PointerType::getUnqual(Data->getType()));
377   unsigned Align =
378     Aligned ? cast<VectorType>(Data->getType())->getBitWidth() / 8 : 1;
379 
380   // If the mask is all ones just emit a regular store.
381   if (const auto *C = dyn_cast<Constant>(Mask))
382     if (C->isAllOnesValue())
383       return Builder.CreateAlignedStore(Data, Ptr, Align);
384 
385   // Convert the mask from an integer type to a vector of i1.
386   unsigned NumElts = Data->getType()->getVectorNumElements();
387   llvm::VectorType *MaskTy = llvm::VectorType::get(Builder.getInt1Ty(),
388                              cast<IntegerType>(Mask->getType())->getBitWidth());
389   Mask = Builder.CreateBitCast(Mask, MaskTy);
390 
391   // If we have less than 8 elements, then the starting mask was an i8 and
392   // we need to extract down to the right number of elements.
393   if (NumElts < 8) {
394     int Indices[4];
395     for (unsigned i = 0; i != NumElts; ++i)
396       Indices[i] = i;
397     Mask = Builder.CreateShuffleVector(Mask, Mask,
398                                            makeArrayRef(Indices, NumElts),
399                                            "extract");
400   }
401 
402   return Builder.CreateMaskedStore(Data, Ptr, Align, Mask);
403 }
404 
405 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
406 // upgraded intrinsic. All argument and return casting must be provided in
407 // order to seamlessly integrate with existing context.
408 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
409   Function *F = CI->getCalledFunction();
410   LLVMContext &C = CI->getContext();
411   IRBuilder<> Builder(C);
412   Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
413 
414   assert(F && "Intrinsic call is not direct?");
415 
416   if (!NewFn) {
417     // Get the Function's name.
418     StringRef Name = F->getName();
419 
420     Value *Rep;
421     // Upgrade packed integer vector compares intrinsics to compare instructions
422     if (Name.startswith("llvm.x86.sse2.pcmpeq.") ||
423         Name.startswith("llvm.x86.avx2.pcmpeq.")) {
424       Rep = Builder.CreateICmpEQ(CI->getArgOperand(0), CI->getArgOperand(1),
425                                  "pcmpeq");
426       // need to sign extend since icmp returns vector of i1
427       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
428     } else if (Name.startswith("llvm.x86.sse2.pcmpgt.") ||
429                Name.startswith("llvm.x86.avx2.pcmpgt.")) {
430       Rep = Builder.CreateICmpSGT(CI->getArgOperand(0), CI->getArgOperand(1),
431                                   "pcmpgt");
432       // need to sign extend since icmp returns vector of i1
433       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
434     } else if (Name == "llvm.x86.sse2.cvtdq2pd" ||
435                Name == "llvm.x86.sse2.cvtps2pd" ||
436                Name == "llvm.x86.avx.cvtdq2.pd.256" ||
437                Name == "llvm.x86.avx.cvt.ps2.pd.256") {
438       // Lossless i32/float to double conversion.
439       // Extract the bottom elements if necessary and convert to double vector.
440       Value *Src = CI->getArgOperand(0);
441       VectorType *SrcTy = cast<VectorType>(Src->getType());
442       VectorType *DstTy = cast<VectorType>(CI->getType());
443       Rep = CI->getArgOperand(0);
444 
445       unsigned NumDstElts = DstTy->getNumElements();
446       if (NumDstElts < SrcTy->getNumElements()) {
447         assert(NumDstElts == 2 && "Unexpected vector size");
448         const int ShuffleMask[2] = { 0, 1 };
449         Rep = Builder.CreateShuffleVector(Rep, UndefValue::get(SrcTy), ShuffleMask);
450       }
451 
452       bool Int2Double = (StringRef::npos != Name.find("cvtdq2"));
453       if (Int2Double)
454         Rep = Builder.CreateSIToFP(Rep, DstTy, "cvtdq2pd");
455       else
456         Rep = Builder.CreateFPExt(Rep, DstTy, "cvtps2pd");
457     } else if (Name.startswith("llvm.x86.avx.movnt.")) {
458       Module *M = F->getParent();
459       SmallVector<Metadata *, 1> Elts;
460       Elts.push_back(
461           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
462       MDNode *Node = MDNode::get(C, Elts);
463 
464       Value *Arg0 = CI->getArgOperand(0);
465       Value *Arg1 = CI->getArgOperand(1);
466 
467       // Convert the type of the pointer to a pointer to the stored type.
468       Value *BC = Builder.CreateBitCast(Arg0,
469                                         PointerType::getUnqual(Arg1->getType()),
470                                         "cast");
471       StoreInst *SI = Builder.CreateAlignedStore(Arg1, BC, 32);
472       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
473 
474       // Remove intrinsic.
475       CI->eraseFromParent();
476       return;
477     } else if (Name == "llvm.x86.sse2.storel.dq") {
478       Value *Arg0 = CI->getArgOperand(0);
479       Value *Arg1 = CI->getArgOperand(1);
480 
481       Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
482       Value *BC0 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
483       Value *Elt = Builder.CreateExtractElement(BC0, (uint64_t)0);
484       Value *BC = Builder.CreateBitCast(Arg0,
485                                         PointerType::getUnqual(Elt->getType()),
486                                         "cast");
487       Builder.CreateAlignedStore(Elt, BC, 1);
488 
489       // Remove intrinsic.
490       CI->eraseFromParent();
491       return;
492     } else if (Name.startswith("llvm.x86.sse.storeu.") ||
493                Name.startswith("llvm.x86.sse2.storeu.") ||
494                Name.startswith("llvm.x86.avx.storeu.")) {
495       Value *Arg0 = CI->getArgOperand(0);
496       Value *Arg1 = CI->getArgOperand(1);
497 
498       Arg0 = Builder.CreateBitCast(Arg0,
499                                    PointerType::getUnqual(Arg1->getType()),
500                                    "cast");
501       Builder.CreateAlignedStore(Arg1, Arg0, 1);
502 
503       // Remove intrinsic.
504       CI->eraseFromParent();
505       return;
506     } else if (Name.startswith("llvm.x86.avx512.mask.storeu.p") ||
507                Name.startswith("llvm.x86.avx512.mask.storeu.b.") ||
508                Name.startswith("llvm.x86.avx512.mask.storeu.w.") ||
509                Name.startswith("llvm.x86.avx512.mask.storeu.d.") ||
510                Name.startswith("llvm.x86.avx512.mask.storeu.q.")) {
511       UpgradeMaskedStore(Builder, C, CI->getArgOperand(0), CI->getArgOperand(1),
512                          CI->getArgOperand(2), /*Aligned*/false);
513 
514       // Remove intrinsic.
515       CI->eraseFromParent();
516       return;
517     } else if (Name.startswith("llvm.x86.avx512.mask.store.p") ||
518                Name.startswith("llvm.x86.avx512.mask.store.b.") ||
519                Name.startswith("llvm.x86.avx512.mask.store.w.") ||
520                Name.startswith("llvm.x86.avx512.mask.store.d.") ||
521                Name.startswith("llvm.x86.avx512.mask.store.q.")) {
522       UpgradeMaskedStore(Builder, C, CI->getArgOperand(0), CI->getArgOperand(1),
523                          CI->getArgOperand(2), /*Aligned*/true);
524 
525       // Remove intrinsic.
526       CI->eraseFromParent();
527       return;
528     } else if (Name.startswith("llvm.x86.xop.vpcom")) {
529       Intrinsic::ID intID;
530       if (Name.endswith("ub"))
531         intID = Intrinsic::x86_xop_vpcomub;
532       else if (Name.endswith("uw"))
533         intID = Intrinsic::x86_xop_vpcomuw;
534       else if (Name.endswith("ud"))
535         intID = Intrinsic::x86_xop_vpcomud;
536       else if (Name.endswith("uq"))
537         intID = Intrinsic::x86_xop_vpcomuq;
538       else if (Name.endswith("b"))
539         intID = Intrinsic::x86_xop_vpcomb;
540       else if (Name.endswith("w"))
541         intID = Intrinsic::x86_xop_vpcomw;
542       else if (Name.endswith("d"))
543         intID = Intrinsic::x86_xop_vpcomd;
544       else if (Name.endswith("q"))
545         intID = Intrinsic::x86_xop_vpcomq;
546       else
547         llvm_unreachable("Unknown suffix");
548 
549       Name = Name.substr(18); // strip off "llvm.x86.xop.vpcom"
550       unsigned Imm;
551       if (Name.startswith("lt"))
552         Imm = 0;
553       else if (Name.startswith("le"))
554         Imm = 1;
555       else if (Name.startswith("gt"))
556         Imm = 2;
557       else if (Name.startswith("ge"))
558         Imm = 3;
559       else if (Name.startswith("eq"))
560         Imm = 4;
561       else if (Name.startswith("ne"))
562         Imm = 5;
563       else if (Name.startswith("false"))
564         Imm = 6;
565       else if (Name.startswith("true"))
566         Imm = 7;
567       else
568         llvm_unreachable("Unknown condition");
569 
570       Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
571       Rep =
572           Builder.CreateCall(VPCOM, {CI->getArgOperand(0), CI->getArgOperand(1),
573                                      Builder.getInt8(Imm)});
574     } else if (Name == "llvm.x86.xop.vpcmov") {
575       Value *Arg0 = CI->getArgOperand(0);
576       Value *Arg1 = CI->getArgOperand(1);
577       Value *Sel = CI->getArgOperand(2);
578       unsigned NumElts = CI->getType()->getVectorNumElements();
579       Constant *MinusOne = ConstantVector::getSplat(NumElts, Builder.getInt64(-1));
580       Value *NotSel = Builder.CreateXor(Sel, MinusOne);
581       Value *Sel0 = Builder.CreateAnd(Arg0, Sel);
582       Value *Sel1 = Builder.CreateAnd(Arg1, NotSel);
583       Rep = Builder.CreateOr(Sel0, Sel1);
584     } else if (Name == "llvm.x86.sse42.crc32.64.8") {
585       Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
586                                                Intrinsic::x86_sse42_crc32_32_8);
587       Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
588       Rep = Builder.CreateCall(CRC32, {Trunc0, CI->getArgOperand(1)});
589       Rep = Builder.CreateZExt(Rep, CI->getType(), "");
590     } else if (Name.startswith("llvm.x86.avx.vbroadcast")) {
591       // Replace broadcasts with a series of insertelements.
592       Type *VecTy = CI->getType();
593       Type *EltTy = VecTy->getVectorElementType();
594       unsigned EltNum = VecTy->getVectorNumElements();
595       Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
596                                           EltTy->getPointerTo());
597       Value *Load = Builder.CreateLoad(EltTy, Cast);
598       Type *I32Ty = Type::getInt32Ty(C);
599       Rep = UndefValue::get(VecTy);
600       for (unsigned I = 0; I < EltNum; ++I)
601         Rep = Builder.CreateInsertElement(Rep, Load,
602                                           ConstantInt::get(I32Ty, I));
603     } else if (Name.startswith("llvm.x86.sse41.pmovsx") ||
604                Name.startswith("llvm.x86.sse41.pmovzx") ||
605                Name.startswith("llvm.x86.avx2.pmovsx") ||
606                Name.startswith("llvm.x86.avx2.pmovzx")) {
607       VectorType *SrcTy = cast<VectorType>(CI->getArgOperand(0)->getType());
608       VectorType *DstTy = cast<VectorType>(CI->getType());
609       unsigned NumDstElts = DstTy->getNumElements();
610 
611       // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
612       SmallVector<int, 8> ShuffleMask;
613       for (int i = 0; i != (int)NumDstElts; ++i)
614         ShuffleMask.push_back(i);
615 
616       Value *SV = Builder.CreateShuffleVector(
617           CI->getArgOperand(0), UndefValue::get(SrcTy), ShuffleMask);
618 
619       bool DoSext = (StringRef::npos != Name.find("pmovsx"));
620       Rep = DoSext ? Builder.CreateSExt(SV, DstTy)
621                    : Builder.CreateZExt(SV, DstTy);
622     } else if (Name == "llvm.x86.avx2.vbroadcasti128") {
623       // Replace vbroadcasts with a vector shuffle.
624       Type *VT = VectorType::get(Type::getInt64Ty(C), 2);
625       Value *Op = Builder.CreatePointerCast(CI->getArgOperand(0),
626                                             PointerType::getUnqual(VT));
627       Value *Load = Builder.CreateLoad(VT, Op);
628       const int Idxs[4] = { 0, 1, 0, 1 };
629       Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
630                                         Idxs);
631     } else if (Name.startswith("llvm.x86.avx2.pbroadcast") ||
632                Name.startswith("llvm.x86.avx2.vbroadcast")) {
633       // Replace vp?broadcasts with a vector shuffle.
634       Value *Op = CI->getArgOperand(0);
635       unsigned NumElts = CI->getType()->getVectorNumElements();
636       Type *MaskTy = VectorType::get(Type::getInt32Ty(C), NumElts);
637       Rep = Builder.CreateShuffleVector(Op, UndefValue::get(Op->getType()),
638                                         Constant::getNullValue(MaskTy));
639     } else if (Name == "llvm.x86.sse2.psll.dq" ||
640                Name == "llvm.x86.avx2.psll.dq") {
641       // 128/256-bit shift left specified in bits.
642       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
643       Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0),
644                                        Shift / 8); // Shift is in bits.
645     } else if (Name == "llvm.x86.sse2.psrl.dq" ||
646                Name == "llvm.x86.avx2.psrl.dq") {
647       // 128/256-bit shift right specified in bits.
648       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
649       Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0),
650                                        Shift / 8); // Shift is in bits.
651     } else if (Name == "llvm.x86.sse2.psll.dq.bs" ||
652                Name == "llvm.x86.avx2.psll.dq.bs") {
653       // 128/256-bit shift left specified in bytes.
654       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
655       Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), Shift);
656     } else if (Name == "llvm.x86.sse2.psrl.dq.bs" ||
657                Name == "llvm.x86.avx2.psrl.dq.bs") {
658       // 128/256-bit shift right specified in bytes.
659       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
660       Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), Shift);
661     } else if (Name == "llvm.x86.sse41.pblendw" ||
662                Name.startswith("llvm.x86.sse41.blendp") ||
663                Name.startswith("llvm.x86.avx.blend.p") ||
664                Name == "llvm.x86.avx2.pblendw" ||
665                Name.startswith("llvm.x86.avx2.pblendd.")) {
666       Value *Op0 = CI->getArgOperand(0);
667       Value *Op1 = CI->getArgOperand(1);
668       unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
669       VectorType *VecTy = cast<VectorType>(CI->getType());
670       unsigned NumElts = VecTy->getNumElements();
671 
672       SmallVector<Constant*, 16> Idxs;
673       for (unsigned i = 0; i != NumElts; ++i) {
674         unsigned Idx = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
675         Idxs.push_back(Builder.getInt32(Idx));
676       }
677 
678       Rep = Builder.CreateShuffleVector(Op0, Op1, ConstantVector::get(Idxs));
679     } else if (Name.startswith("llvm.x86.avx.vinsertf128.") ||
680                Name == "llvm.x86.avx2.vinserti128") {
681       Value *Op0 = CI->getArgOperand(0);
682       Value *Op1 = CI->getArgOperand(1);
683       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
684       VectorType *VecTy = cast<VectorType>(CI->getType());
685       unsigned NumElts = VecTy->getNumElements();
686 
687       // Mask off the high bits of the immediate value; hardware ignores those.
688       Imm = Imm & 1;
689 
690       // Extend the second operand into a vector that is twice as big.
691       Value *UndefV = UndefValue::get(Op1->getType());
692       SmallVector<Constant*, 8> Idxs;
693       for (unsigned i = 0; i != NumElts; ++i) {
694         Idxs.push_back(Builder.getInt32(i));
695       }
696       Rep = Builder.CreateShuffleVector(Op1, UndefV, ConstantVector::get(Idxs));
697 
698       // Insert the second operand into the first operand.
699 
700       // Note that there is no guarantee that instruction lowering will actually
701       // produce a vinsertf128 instruction for the created shuffles. In
702       // particular, the 0 immediate case involves no lane changes, so it can
703       // be handled as a blend.
704 
705       // Example of shuffle mask for 32-bit elements:
706       // Imm = 1  <i32 0, i32 1, i32 2,  i32 3,  i32 8, i32 9, i32 10, i32 11>
707       // Imm = 0  <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6,  i32 7 >
708 
709       SmallVector<Constant*, 8> Idxs2;
710       // The low half of the result is either the low half of the 1st operand
711       // or the low half of the 2nd operand (the inserted vector).
712       for (unsigned i = 0; i != NumElts / 2; ++i) {
713         unsigned Idx = Imm ? i : (i + NumElts);
714         Idxs2.push_back(Builder.getInt32(Idx));
715       }
716       // The high half of the result is either the low half of the 2nd operand
717       // (the inserted vector) or the high half of the 1st operand.
718       for (unsigned i = NumElts / 2; i != NumElts; ++i) {
719         unsigned Idx = Imm ? (i + NumElts / 2) : i;
720         Idxs2.push_back(Builder.getInt32(Idx));
721       }
722       Rep = Builder.CreateShuffleVector(Op0, Rep, ConstantVector::get(Idxs2));
723     } else if (Name.startswith("llvm.x86.avx.vextractf128.") ||
724                Name == "llvm.x86.avx2.vextracti128") {
725       Value *Op0 = CI->getArgOperand(0);
726       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
727       VectorType *VecTy = cast<VectorType>(CI->getType());
728       unsigned NumElts = VecTy->getNumElements();
729 
730       // Mask off the high bits of the immediate value; hardware ignores those.
731       Imm = Imm & 1;
732 
733       // Get indexes for either the high half or low half of the input vector.
734       SmallVector<Constant*, 4> Idxs(NumElts);
735       for (unsigned i = 0; i != NumElts; ++i) {
736         unsigned Idx = Imm ? (i + NumElts) : i;
737         Idxs[i] = Builder.getInt32(Idx);
738       }
739 
740       Value *UndefV = UndefValue::get(Op0->getType());
741       Rep = Builder.CreateShuffleVector(Op0, UndefV, ConstantVector::get(Idxs));
742     } else if (Name == "llvm.stackprotectorcheck") {
743       Rep = nullptr;
744     } else {
745       bool PD128 = false, PD256 = false, PS128 = false, PS256 = false;
746       if (Name == "llvm.x86.avx.vpermil.pd.256")
747         PD256 = true;
748       else if (Name == "llvm.x86.avx.vpermil.pd")
749         PD128 = true;
750       else if (Name == "llvm.x86.avx.vpermil.ps.256")
751         PS256 = true;
752       else if (Name == "llvm.x86.avx.vpermil.ps")
753         PS128 = true;
754 
755       if (PD256 || PD128 || PS256 || PS128) {
756         Value *Op0 = CI->getArgOperand(0);
757         unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
758         SmallVector<Constant*, 8> Idxs;
759 
760         if (PD128)
761           for (unsigned i = 0; i != 2; ++i)
762             Idxs.push_back(Builder.getInt32((Imm >> i) & 0x1));
763         else if (PD256)
764           for (unsigned l = 0; l != 4; l+=2)
765             for (unsigned i = 0; i != 2; ++i)
766               Idxs.push_back(Builder.getInt32(((Imm >> (l+i)) & 0x1) + l));
767         else if (PS128)
768           for (unsigned i = 0; i != 4; ++i)
769             Idxs.push_back(Builder.getInt32((Imm >> (2 * i)) & 0x3));
770         else if (PS256)
771           for (unsigned l = 0; l != 8; l+=4)
772             for (unsigned i = 0; i != 4; ++i)
773               Idxs.push_back(Builder.getInt32(((Imm >> (2 * i)) & 0x3) + l));
774         else
775           llvm_unreachable("Unexpected function");
776 
777         Rep = Builder.CreateShuffleVector(Op0, Op0, ConstantVector::get(Idxs));
778       } else {
779         llvm_unreachable("Unknown function for CallInst upgrade.");
780       }
781     }
782 
783     if (Rep)
784       CI->replaceAllUsesWith(Rep);
785     CI->eraseFromParent();
786     return;
787   }
788 
789   std::string Name = CI->getName();
790   if (!Name.empty())
791     CI->setName(Name + ".old");
792 
793   switch (NewFn->getIntrinsicID()) {
794   default:
795     llvm_unreachable("Unknown function for CallInst upgrade.");
796 
797   case Intrinsic::arm_neon_vld1:
798   case Intrinsic::arm_neon_vld2:
799   case Intrinsic::arm_neon_vld3:
800   case Intrinsic::arm_neon_vld4:
801   case Intrinsic::arm_neon_vld2lane:
802   case Intrinsic::arm_neon_vld3lane:
803   case Intrinsic::arm_neon_vld4lane:
804   case Intrinsic::arm_neon_vst1:
805   case Intrinsic::arm_neon_vst2:
806   case Intrinsic::arm_neon_vst3:
807   case Intrinsic::arm_neon_vst4:
808   case Intrinsic::arm_neon_vst2lane:
809   case Intrinsic::arm_neon_vst3lane:
810   case Intrinsic::arm_neon_vst4lane: {
811     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
812                                  CI->arg_operands().end());
813     CI->replaceAllUsesWith(Builder.CreateCall(NewFn, Args));
814     CI->eraseFromParent();
815     return;
816   }
817 
818   case Intrinsic::ctlz:
819   case Intrinsic::cttz:
820     assert(CI->getNumArgOperands() == 1 &&
821            "Mismatch between function args and call args");
822     CI->replaceAllUsesWith(Builder.CreateCall(
823         NewFn, {CI->getArgOperand(0), Builder.getFalse()}, Name));
824     CI->eraseFromParent();
825     return;
826 
827   case Intrinsic::objectsize:
828     CI->replaceAllUsesWith(Builder.CreateCall(
829         NewFn, {CI->getArgOperand(0), CI->getArgOperand(1)}, Name));
830     CI->eraseFromParent();
831     return;
832 
833   case Intrinsic::ctpop: {
834     CI->replaceAllUsesWith(Builder.CreateCall(NewFn, {CI->getArgOperand(0)}));
835     CI->eraseFromParent();
836     return;
837   }
838 
839   case Intrinsic::x86_xop_vfrcz_ss:
840   case Intrinsic::x86_xop_vfrcz_sd:
841     CI->replaceAllUsesWith(
842         Builder.CreateCall(NewFn, {CI->getArgOperand(1)}, Name));
843     CI->eraseFromParent();
844     return;
845 
846   case Intrinsic::x86_sse41_ptestc:
847   case Intrinsic::x86_sse41_ptestz:
848   case Intrinsic::x86_sse41_ptestnzc: {
849     // The arguments for these intrinsics used to be v4f32, and changed
850     // to v2i64. This is purely a nop, since those are bitwise intrinsics.
851     // So, the only thing required is a bitcast for both arguments.
852     // First, check the arguments have the old type.
853     Value *Arg0 = CI->getArgOperand(0);
854     if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
855       return;
856 
857     // Old intrinsic, add bitcasts
858     Value *Arg1 = CI->getArgOperand(1);
859 
860     Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
861 
862     Value *BC0 = Builder.CreateBitCast(Arg0, NewVecTy, "cast");
863     Value *BC1 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
864 
865     CallInst *NewCall = Builder.CreateCall(NewFn, {BC0, BC1}, Name);
866     CI->replaceAllUsesWith(NewCall);
867     CI->eraseFromParent();
868     return;
869   }
870 
871   case Intrinsic::x86_sse41_insertps:
872   case Intrinsic::x86_sse41_dppd:
873   case Intrinsic::x86_sse41_dpps:
874   case Intrinsic::x86_sse41_mpsadbw:
875   case Intrinsic::x86_avx_dp_ps_256:
876   case Intrinsic::x86_avx2_mpsadbw: {
877     // Need to truncate the last argument from i32 to i8 -- this argument models
878     // an inherently 8-bit immediate operand to these x86 instructions.
879     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
880                                  CI->arg_operands().end());
881 
882     // Replace the last argument with a trunc.
883     Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
884 
885     CallInst *NewCall = Builder.CreateCall(NewFn, Args);
886     CI->replaceAllUsesWith(NewCall);
887     CI->eraseFromParent();
888     return;
889   }
890 
891   case Intrinsic::thread_pointer: {
892     CI->replaceAllUsesWith(Builder.CreateCall(NewFn, {}));
893     CI->eraseFromParent();
894     return;
895   }
896   }
897 }
898 
899 void llvm::UpgradeCallsToIntrinsic(Function *F) {
900   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
901 
902   // Check if this function should be upgraded and get the replacement function
903   // if there is one.
904   Function *NewFn;
905   if (UpgradeIntrinsicFunction(F, NewFn)) {
906     // Replace all users of the old function with the new function or new
907     // instructions. This is not a range loop because the call is deleted.
908     for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; )
909       if (CallInst *CI = dyn_cast<CallInst>(*UI++))
910         UpgradeIntrinsicCall(CI, NewFn);
911 
912     // Remove old function, no longer used, from the module.
913     F->eraseFromParent();
914   }
915 }
916 
917 void llvm::UpgradeInstWithTBAATag(Instruction *I) {
918   MDNode *MD = I->getMetadata(LLVMContext::MD_tbaa);
919   assert(MD && "UpgradeInstWithTBAATag should have a TBAA tag");
920   // Check if the tag uses struct-path aware TBAA format.
921   if (isa<MDNode>(MD->getOperand(0)) && MD->getNumOperands() >= 3)
922     return;
923 
924   if (MD->getNumOperands() == 3) {
925     Metadata *Elts[] = {MD->getOperand(0), MD->getOperand(1)};
926     MDNode *ScalarType = MDNode::get(I->getContext(), Elts);
927     // Create a MDNode <ScalarType, ScalarType, offset 0, const>
928     Metadata *Elts2[] = {ScalarType, ScalarType,
929                          ConstantAsMetadata::get(Constant::getNullValue(
930                              Type::getInt64Ty(I->getContext()))),
931                          MD->getOperand(2)};
932     I->setMetadata(LLVMContext::MD_tbaa, MDNode::get(I->getContext(), Elts2));
933   } else {
934     // Create a MDNode <MD, MD, offset 0>
935     Metadata *Elts[] = {MD, MD, ConstantAsMetadata::get(Constant::getNullValue(
936                                     Type::getInt64Ty(I->getContext())))};
937     I->setMetadata(LLVMContext::MD_tbaa, MDNode::get(I->getContext(), Elts));
938   }
939 }
940 
941 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
942                                       Instruction *&Temp) {
943   if (Opc != Instruction::BitCast)
944     return nullptr;
945 
946   Temp = nullptr;
947   Type *SrcTy = V->getType();
948   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
949       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
950     LLVMContext &Context = V->getContext();
951 
952     // We have no information about target data layout, so we assume that
953     // the maximum pointer size is 64bit.
954     Type *MidTy = Type::getInt64Ty(Context);
955     Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
956 
957     return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
958   }
959 
960   return nullptr;
961 }
962 
963 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
964   if (Opc != Instruction::BitCast)
965     return nullptr;
966 
967   Type *SrcTy = C->getType();
968   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
969       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
970     LLVMContext &Context = C->getContext();
971 
972     // We have no information about target data layout, so we assume that
973     // the maximum pointer size is 64bit.
974     Type *MidTy = Type::getInt64Ty(Context);
975 
976     return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
977                                      DestTy);
978   }
979 
980   return nullptr;
981 }
982 
983 /// Check the debug info version number, if it is out-dated, drop the debug
984 /// info. Return true if module is modified.
985 bool llvm::UpgradeDebugInfo(Module &M) {
986   unsigned Version = getDebugMetadataVersionFromModule(M);
987   if (Version == DEBUG_METADATA_VERSION)
988     return false;
989 
990   bool RetCode = StripDebugInfo(M);
991   if (RetCode) {
992     DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
993     M.getContext().diagnose(DiagVersion);
994   }
995   return RetCode;
996 }
997 
998 bool llvm::UpgradeModuleFlags(Module &M) {
999   const NamedMDNode *ModFlags = M.getModuleFlagsMetadata();
1000   if (!ModFlags)
1001     return false;
1002 
1003   bool HasObjCFlag = false, HasClassProperties = false;
1004   for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) {
1005     MDNode *Op = ModFlags->getOperand(I);
1006     if (Op->getNumOperands() < 2)
1007       continue;
1008     MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(1));
1009     if (!ID)
1010       continue;
1011     if (ID->getString() == "Objective-C Image Info Version")
1012       HasObjCFlag = true;
1013     if (ID->getString() == "Objective-C Class Properties")
1014       HasClassProperties = true;
1015   }
1016   // "Objective-C Class Properties" is recently added for Objective-C. We
1017   // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module
1018   // flag of value 0, so we can correclty report error when trying to link
1019   // an ObjC bitcode without this module flag with an ObjC bitcode with this
1020   // module flag.
1021   if (HasObjCFlag && !HasClassProperties) {
1022     M.addModuleFlag(llvm::Module::Error, "Objective-C Class Properties",
1023                     (uint32_t)0);
1024     return true;
1025   }
1026   return false;
1027 }
1028 
1029 static bool isOldLoopArgument(Metadata *MD) {
1030   auto *T = dyn_cast_or_null<MDTuple>(MD);
1031   if (!T)
1032     return false;
1033   if (T->getNumOperands() < 1)
1034     return false;
1035   auto *S = dyn_cast_or_null<MDString>(T->getOperand(0));
1036   if (!S)
1037     return false;
1038   return S->getString().startswith("llvm.vectorizer.");
1039 }
1040 
1041 static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) {
1042   StringRef OldPrefix = "llvm.vectorizer.";
1043   assert(OldTag.startswith(OldPrefix) && "Expected old prefix");
1044 
1045   if (OldTag == "llvm.vectorizer.unroll")
1046     return MDString::get(C, "llvm.loop.interleave.count");
1047 
1048   return MDString::get(
1049       C, (Twine("llvm.loop.vectorize.") + OldTag.drop_front(OldPrefix.size()))
1050              .str());
1051 }
1052 
1053 static Metadata *upgradeLoopArgument(Metadata *MD) {
1054   auto *T = dyn_cast_or_null<MDTuple>(MD);
1055   if (!T)
1056     return MD;
1057   if (T->getNumOperands() < 1)
1058     return MD;
1059   auto *OldTag = dyn_cast_or_null<MDString>(T->getOperand(0));
1060   if (!OldTag)
1061     return MD;
1062   if (!OldTag->getString().startswith("llvm.vectorizer."))
1063     return MD;
1064 
1065   // This has an old tag.  Upgrade it.
1066   SmallVector<Metadata *, 8> Ops;
1067   Ops.reserve(T->getNumOperands());
1068   Ops.push_back(upgradeLoopTag(T->getContext(), OldTag->getString()));
1069   for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I)
1070     Ops.push_back(T->getOperand(I));
1071 
1072   return MDTuple::get(T->getContext(), Ops);
1073 }
1074 
1075 MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) {
1076   auto *T = dyn_cast<MDTuple>(&N);
1077   if (!T)
1078     return &N;
1079 
1080   if (!llvm::any_of(T->operands(), isOldLoopArgument))
1081     return &N;
1082 
1083   SmallVector<Metadata *, 8> Ops;
1084   Ops.reserve(T->getNumOperands());
1085   for (Metadata *MD : T->operands())
1086     Ops.push_back(upgradeLoopArgument(MD));
1087 
1088   return MDTuple::get(T->getContext(), Ops);
1089 }
1090