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