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/ADT/StringSwitch.h"
18 #include "llvm/IR/CFG.h"
19 #include "llvm/IR/CallSite.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/DIBuilder.h"
22 #include "llvm/IR/DebugInfo.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/Instruction.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Verifier.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/Regex.h"
33 #include <cstring>
34 using namespace llvm;
35 
36 static void rename(GlobalValue *GV) { GV->setName(GV->getName() + ".old"); }
37 
38 // Upgrade the declarations of the SSE4.1 ptest intrinsics whose arguments have
39 // changed their type from v4f32 to v2i64.
40 static bool UpgradePTESTIntrinsic(Function* F, Intrinsic::ID IID,
41                                   Function *&NewFn) {
42   // Check whether this is an old version of the function, which received
43   // v4f32 arguments.
44   Type *Arg0Type = F->getFunctionType()->getParamType(0);
45   if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
46     return false;
47 
48   // Yes, it's old, replace it with new version.
49   rename(F);
50   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
51   return true;
52 }
53 
54 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
55 // arguments have changed their type from i32 to i8.
56 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
57                                              Function *&NewFn) {
58   // Check that the last argument is an i32.
59   Type *LastArgType = F->getFunctionType()->getParamType(
60      F->getFunctionType()->getNumParams() - 1);
61   if (!LastArgType->isIntegerTy(32))
62     return false;
63 
64   // Move this function aside and map down.
65   rename(F);
66   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
67   return true;
68 }
69 
70 static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
71   // All of the intrinsics matches below should be marked with which llvm
72   // version started autoupgrading them. At some point in the future we would
73   // like to use this information to remove upgrade code for some older
74   // intrinsics. It is currently undecided how we will determine that future
75   // point.
76   if (Name=="ssse3.pabs.b.128" || // Added in 6.0
77       Name=="ssse3.pabs.w.128" || // Added in 6.0
78       Name=="ssse3.pabs.d.128" || // Added in 6.0
79       Name.startswith("avx2.pabs.") || // Added in 6.0
80       Name.startswith("avx512.mask.pabs.") || // Added in 6.0
81       Name.startswith("avx512.broadcastm") || // Added in 6.0
82       Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
83       Name.startswith("sse2.pcmpeq.") || // Added in 3.1
84       Name.startswith("sse2.pcmpgt.") || // Added in 3.1
85       Name.startswith("avx2.pcmpeq.") || // Added in 3.1
86       Name.startswith("avx2.pcmpgt.") || // Added in 3.1
87       Name.startswith("avx512.mask.pcmpeq.") || // Added in 3.9
88       Name.startswith("avx512.mask.pcmpgt.") || // Added in 3.9
89       Name.startswith("avx.vperm2f128.") || // Added in 6.0
90       Name == "avx2.vperm2i128" || // Added in 6.0
91       Name == "sse.add.ss" || // Added in 4.0
92       Name == "sse2.add.sd" || // Added in 4.0
93       Name == "sse.sub.ss" || // Added in 4.0
94       Name == "sse2.sub.sd" || // Added in 4.0
95       Name == "sse.mul.ss" || // Added in 4.0
96       Name == "sse2.mul.sd" || // Added in 4.0
97       Name == "sse.div.ss" || // Added in 4.0
98       Name == "sse2.div.sd" || // Added in 4.0
99       Name == "sse41.pmaxsb" || // Added in 3.9
100       Name == "sse2.pmaxs.w" || // Added in 3.9
101       Name == "sse41.pmaxsd" || // Added in 3.9
102       Name == "sse2.pmaxu.b" || // Added in 3.9
103       Name == "sse41.pmaxuw" || // Added in 3.9
104       Name == "sse41.pmaxud" || // Added in 3.9
105       Name == "sse41.pminsb" || // Added in 3.9
106       Name == "sse2.pmins.w" || // Added in 3.9
107       Name == "sse41.pminsd" || // Added in 3.9
108       Name == "sse2.pminu.b" || // Added in 3.9
109       Name == "sse41.pminuw" || // Added in 3.9
110       Name == "sse41.pminud" || // Added in 3.9
111       Name.startswith("avx512.mask.pshuf.b.") || // Added in 4.0
112       Name.startswith("avx2.pmax") || // Added in 3.9
113       Name.startswith("avx2.pmin") || // Added in 3.9
114       Name.startswith("avx512.mask.pmax") || // Added in 4.0
115       Name.startswith("avx512.mask.pmin") || // Added in 4.0
116       Name.startswith("avx2.vbroadcast") || // Added in 3.8
117       Name.startswith("avx2.pbroadcast") || // Added in 3.8
118       Name.startswith("avx.vpermil.") || // Added in 3.1
119       Name.startswith("sse2.pshuf") || // Added in 3.9
120       Name.startswith("avx512.pbroadcast") || // Added in 3.9
121       Name.startswith("avx512.mask.broadcast.s") || // Added in 3.9
122       Name.startswith("avx512.mask.movddup") || // Added in 3.9
123       Name.startswith("avx512.mask.movshdup") || // Added in 3.9
124       Name.startswith("avx512.mask.movsldup") || // Added in 3.9
125       Name.startswith("avx512.mask.pshuf.d.") || // Added in 3.9
126       Name.startswith("avx512.mask.pshufl.w.") || // Added in 3.9
127       Name.startswith("avx512.mask.pshufh.w.") || // Added in 3.9
128       Name.startswith("avx512.mask.shuf.p") || // Added in 4.0
129       Name.startswith("avx512.mask.vpermil.p") || // Added in 3.9
130       Name.startswith("avx512.mask.perm.df.") || // Added in 3.9
131       Name.startswith("avx512.mask.perm.di.") || // Added in 3.9
132       Name.startswith("avx512.mask.punpckl") || // Added in 3.9
133       Name.startswith("avx512.mask.punpckh") || // Added in 3.9
134       Name.startswith("avx512.mask.unpckl.") || // Added in 3.9
135       Name.startswith("avx512.mask.unpckh.") || // Added in 3.9
136       Name.startswith("avx512.mask.pand.") || // Added in 3.9
137       Name.startswith("avx512.mask.pandn.") || // Added in 3.9
138       Name.startswith("avx512.mask.por.") || // Added in 3.9
139       Name.startswith("avx512.mask.pxor.") || // Added in 3.9
140       Name.startswith("avx512.mask.and.") || // Added in 3.9
141       Name.startswith("avx512.mask.andn.") || // Added in 3.9
142       Name.startswith("avx512.mask.or.") || // Added in 3.9
143       Name.startswith("avx512.mask.xor.") || // Added in 3.9
144       Name.startswith("avx512.mask.padd.") || // Added in 4.0
145       Name.startswith("avx512.mask.psub.") || // Added in 4.0
146       Name.startswith("avx512.mask.pmull.") || // Added in 4.0
147       Name.startswith("avx512.mask.cvtdq2pd.") || // Added in 4.0
148       Name.startswith("avx512.mask.cvtudq2pd.") || // Added in 4.0
149       Name.startswith("avx512.mask.pmul.dq.") || // Added in 4.0
150       Name.startswith("avx512.mask.pmulu.dq.") || // Added in 4.0
151       Name.startswith("avx512.mask.packsswb.") || // Added in 5.0
152       Name.startswith("avx512.mask.packssdw.") || // Added in 5.0
153       Name.startswith("avx512.mask.packuswb.") || // Added in 5.0
154       Name.startswith("avx512.mask.packusdw.") || // Added in 5.0
155       Name.startswith("avx512.mask.cmp.b") || // Added in 5.0
156       Name.startswith("avx512.mask.cmp.d") || // Added in 5.0
157       Name.startswith("avx512.mask.cmp.q") || // Added in 5.0
158       Name.startswith("avx512.mask.cmp.w") || // Added in 5.0
159       Name.startswith("avx512.mask.ucmp.") || // Added in 5.0
160       Name == "avx512.mask.add.pd.128" || // Added in 4.0
161       Name == "avx512.mask.add.pd.256" || // Added in 4.0
162       Name == "avx512.mask.add.ps.128" || // Added in 4.0
163       Name == "avx512.mask.add.ps.256" || // Added in 4.0
164       Name == "avx512.mask.div.pd.128" || // Added in 4.0
165       Name == "avx512.mask.div.pd.256" || // Added in 4.0
166       Name == "avx512.mask.div.ps.128" || // Added in 4.0
167       Name == "avx512.mask.div.ps.256" || // Added in 4.0
168       Name == "avx512.mask.mul.pd.128" || // Added in 4.0
169       Name == "avx512.mask.mul.pd.256" || // Added in 4.0
170       Name == "avx512.mask.mul.ps.128" || // Added in 4.0
171       Name == "avx512.mask.mul.ps.256" || // Added in 4.0
172       Name == "avx512.mask.sub.pd.128" || // Added in 4.0
173       Name == "avx512.mask.sub.pd.256" || // Added in 4.0
174       Name == "avx512.mask.sub.ps.128" || // Added in 4.0
175       Name == "avx512.mask.sub.ps.256" || // Added in 4.0
176       Name == "avx512.mask.max.pd.128" || // Added in 5.0
177       Name == "avx512.mask.max.pd.256" || // Added in 5.0
178       Name == "avx512.mask.max.ps.128" || // Added in 5.0
179       Name == "avx512.mask.max.ps.256" || // Added in 5.0
180       Name == "avx512.mask.min.pd.128" || // Added in 5.0
181       Name == "avx512.mask.min.pd.256" || // Added in 5.0
182       Name == "avx512.mask.min.ps.128" || // Added in 5.0
183       Name == "avx512.mask.min.ps.256" || // Added in 5.0
184       Name.startswith("avx512.mask.vpermilvar.") || // Added in 4.0
185       Name.startswith("avx512.mask.psll.d") || // Added in 4.0
186       Name.startswith("avx512.mask.psll.q") || // Added in 4.0
187       Name.startswith("avx512.mask.psll.w") || // Added in 4.0
188       Name.startswith("avx512.mask.psra.d") || // Added in 4.0
189       Name.startswith("avx512.mask.psra.q") || // Added in 4.0
190       Name.startswith("avx512.mask.psra.w") || // Added in 4.0
191       Name.startswith("avx512.mask.psrl.d") || // Added in 4.0
192       Name.startswith("avx512.mask.psrl.q") || // Added in 4.0
193       Name.startswith("avx512.mask.psrl.w") || // Added in 4.0
194       Name.startswith("avx512.mask.pslli") || // Added in 4.0
195       Name.startswith("avx512.mask.psrai") || // Added in 4.0
196       Name.startswith("avx512.mask.psrli") || // Added in 4.0
197       Name.startswith("avx512.mask.psllv") || // Added in 4.0
198       Name.startswith("avx512.mask.psrav") || // Added in 4.0
199       Name.startswith("avx512.mask.psrlv") || // Added in 4.0
200       Name.startswith("sse41.pmovsx") || // Added in 3.8
201       Name.startswith("sse41.pmovzx") || // Added in 3.9
202       Name.startswith("avx2.pmovsx") || // Added in 3.9
203       Name.startswith("avx2.pmovzx") || // Added in 3.9
204       Name.startswith("avx512.mask.pmovsx") || // Added in 4.0
205       Name.startswith("avx512.mask.pmovzx") || // Added in 4.0
206       Name.startswith("avx512.mask.lzcnt.") || // Added in 5.0
207       Name == "sse2.cvtdq2pd" || // Added in 3.9
208       Name == "sse2.cvtps2pd" || // Added in 3.9
209       Name == "avx.cvtdq2.pd.256" || // Added in 3.9
210       Name == "avx.cvt.ps2.pd.256" || // Added in 3.9
211       Name.startswith("avx.vinsertf128.") || // Added in 3.7
212       Name == "avx2.vinserti128" || // Added in 3.7
213       Name.startswith("avx512.mask.insert") || // Added in 4.0
214       Name.startswith("avx.vextractf128.") || // Added in 3.7
215       Name == "avx2.vextracti128" || // Added in 3.7
216       Name.startswith("avx512.mask.vextract") || // Added in 4.0
217       Name.startswith("sse4a.movnt.") || // Added in 3.9
218       Name.startswith("avx.movnt.") || // Added in 3.2
219       Name.startswith("avx512.storent.") || // Added in 3.9
220       Name == "sse41.movntdqa" || // Added in 5.0
221       Name == "avx2.movntdqa" || // Added in 5.0
222       Name == "avx512.movntdqa" || // Added in 5.0
223       Name == "sse2.storel.dq" || // Added in 3.9
224       Name.startswith("sse.storeu.") || // Added in 3.9
225       Name.startswith("sse2.storeu.") || // Added in 3.9
226       Name.startswith("avx.storeu.") || // Added in 3.9
227       Name.startswith("avx512.mask.storeu.") || // Added in 3.9
228       Name.startswith("avx512.mask.store.p") || // Added in 3.9
229       Name.startswith("avx512.mask.store.b.") || // Added in 3.9
230       Name.startswith("avx512.mask.store.w.") || // Added in 3.9
231       Name.startswith("avx512.mask.store.d.") || // Added in 3.9
232       Name.startswith("avx512.mask.store.q.") || // Added in 3.9
233       Name.startswith("avx512.mask.loadu.") || // Added in 3.9
234       Name.startswith("avx512.mask.load.") || // Added in 3.9
235       Name == "sse42.crc32.64.8" || // Added in 3.4
236       Name.startswith("avx.vbroadcast.s") || // Added in 3.5
237       Name.startswith("avx512.mask.palignr.") || // Added in 3.9
238       Name.startswith("avx512.mask.valign.") || // Added in 4.0
239       Name.startswith("sse2.psll.dq") || // Added in 3.7
240       Name.startswith("sse2.psrl.dq") || // Added in 3.7
241       Name.startswith("avx2.psll.dq") || // Added in 3.7
242       Name.startswith("avx2.psrl.dq") || // Added in 3.7
243       Name.startswith("avx512.psll.dq") || // Added in 3.9
244       Name.startswith("avx512.psrl.dq") || // Added in 3.9
245       Name == "sse41.pblendw" || // Added in 3.7
246       Name.startswith("sse41.blendp") || // Added in 3.7
247       Name.startswith("avx.blend.p") || // Added in 3.7
248       Name == "avx2.pblendw" || // Added in 3.7
249       Name.startswith("avx2.pblendd.") || // Added in 3.7
250       Name.startswith("avx.vbroadcastf128") || // Added in 4.0
251       Name == "avx2.vbroadcasti128" || // Added in 3.7
252       Name.startswith("avx512.mask.broadcastf") || // Added in 6.0
253       Name.startswith("avx512.mask.broadcasti") || // Added in 6.0
254       Name == "xop.vpcmov" || // Added in 3.8
255       Name == "xop.vpcmov.256" || // Added in 5.0
256       Name.startswith("avx512.mask.move.s") || // Added in 4.0
257       Name.startswith("avx512.cvtmask2") || // Added in 5.0
258       (Name.startswith("xop.vpcom") && // Added in 3.2
259        F->arg_size() == 2) ||
260       Name.startswith("sse2.pavg") || // Added in 6.0
261       Name.startswith("avx2.pavg") || // Added in 6.0
262       Name.startswith("avx512.mask.pavg")) // Added in 6.0
263     return true;
264 
265   return false;
266 }
267 
268 static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
269                                         Function *&NewFn) {
270   // Only handle intrinsics that start with "x86.".
271   if (!Name.startswith("x86."))
272     return false;
273   // Remove "x86." prefix.
274   Name = Name.substr(4);
275 
276   if (ShouldUpgradeX86Intrinsic(F, Name)) {
277     NewFn = nullptr;
278     return true;
279   }
280 
281   // SSE4.1 ptest functions may have an old signature.
282   if (Name.startswith("sse41.ptest")) { // Added in 3.2
283     if (Name.substr(11) == "c")
284       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestc, NewFn);
285     if (Name.substr(11) == "z")
286       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestz, NewFn);
287     if (Name.substr(11) == "nzc")
288       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
289   }
290   // Several blend and other instructions with masks used the wrong number of
291   // bits.
292   if (Name == "sse41.insertps") // Added in 3.6
293     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
294                                             NewFn);
295   if (Name == "sse41.dppd") // Added in 3.6
296     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
297                                             NewFn);
298   if (Name == "sse41.dpps") // Added in 3.6
299     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
300                                             NewFn);
301   if (Name == "sse41.mpsadbw") // Added in 3.6
302     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
303                                             NewFn);
304   if (Name == "avx.dp.ps.256") // Added in 3.6
305     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
306                                             NewFn);
307   if (Name == "avx2.mpsadbw") // Added in 3.6
308     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
309                                             NewFn);
310 
311   // frcz.ss/sd may need to have an argument dropped. Added in 3.2
312   if (Name.startswith("xop.vfrcz.ss") && F->arg_size() == 2) {
313     rename(F);
314     NewFn = Intrinsic::getDeclaration(F->getParent(),
315                                       Intrinsic::x86_xop_vfrcz_ss);
316     return true;
317   }
318   if (Name.startswith("xop.vfrcz.sd") && F->arg_size() == 2) {
319     rename(F);
320     NewFn = Intrinsic::getDeclaration(F->getParent(),
321                                       Intrinsic::x86_xop_vfrcz_sd);
322     return true;
323   }
324   // Upgrade any XOP PERMIL2 index operand still using a float/double vector.
325   if (Name.startswith("xop.vpermil2")) { // Added in 3.9
326     auto Idx = F->getFunctionType()->getParamType(2);
327     if (Idx->isFPOrFPVectorTy()) {
328       rename(F);
329       unsigned IdxSize = Idx->getPrimitiveSizeInBits();
330       unsigned EltSize = Idx->getScalarSizeInBits();
331       Intrinsic::ID Permil2ID;
332       if (EltSize == 64 && IdxSize == 128)
333         Permil2ID = Intrinsic::x86_xop_vpermil2pd;
334       else if (EltSize == 32 && IdxSize == 128)
335         Permil2ID = Intrinsic::x86_xop_vpermil2ps;
336       else if (EltSize == 64 && IdxSize == 256)
337         Permil2ID = Intrinsic::x86_xop_vpermil2pd_256;
338       else
339         Permil2ID = Intrinsic::x86_xop_vpermil2ps_256;
340       NewFn = Intrinsic::getDeclaration(F->getParent(), Permil2ID);
341       return true;
342     }
343   }
344 
345   return false;
346 }
347 
348 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
349   assert(F && "Illegal to upgrade a non-existent Function.");
350 
351   // Quickly eliminate it, if it's not a candidate.
352   StringRef Name = F->getName();
353   if (Name.size() <= 8 || !Name.startswith("llvm."))
354     return false;
355   Name = Name.substr(5); // Strip off "llvm."
356 
357   switch (Name[0]) {
358   default: break;
359   case 'a': {
360     if (Name.startswith("arm.rbit") || Name.startswith("aarch64.rbit")) {
361       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
362                                         F->arg_begin()->getType());
363       return true;
364     }
365     if (Name.startswith("arm.neon.vclz")) {
366       Type* args[2] = {
367         F->arg_begin()->getType(),
368         Type::getInt1Ty(F->getContext())
369       };
370       // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
371       // the end of the name. Change name from llvm.arm.neon.vclz.* to
372       //  llvm.ctlz.*
373       FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
374       NewFn = Function::Create(fType, F->getLinkage(),
375                                "llvm.ctlz." + Name.substr(14), F->getParent());
376       return true;
377     }
378     if (Name.startswith("arm.neon.vcnt")) {
379       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
380                                         F->arg_begin()->getType());
381       return true;
382     }
383     Regex vldRegex("^arm\\.neon\\.vld([1234]|[234]lane)\\.v[a-z0-9]*$");
384     if (vldRegex.match(Name)) {
385       auto fArgs = F->getFunctionType()->params();
386       SmallVector<Type *, 4> Tys(fArgs.begin(), fArgs.end());
387       // Can't use Intrinsic::getDeclaration here as the return types might
388       // then only be structurally equal.
389       FunctionType* fType = FunctionType::get(F->getReturnType(), Tys, false);
390       NewFn = Function::Create(fType, F->getLinkage(),
391                                "llvm." + Name + ".p0i8", F->getParent());
392       return true;
393     }
394     Regex vstRegex("^arm\\.neon\\.vst([1234]|[234]lane)\\.v[a-z0-9]*$");
395     if (vstRegex.match(Name)) {
396       static const Intrinsic::ID StoreInts[] = {Intrinsic::arm_neon_vst1,
397                                                 Intrinsic::arm_neon_vst2,
398                                                 Intrinsic::arm_neon_vst3,
399                                                 Intrinsic::arm_neon_vst4};
400 
401       static const Intrinsic::ID StoreLaneInts[] = {
402         Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
403         Intrinsic::arm_neon_vst4lane
404       };
405 
406       auto fArgs = F->getFunctionType()->params();
407       Type *Tys[] = {fArgs[0], fArgs[1]};
408       if (Name.find("lane") == StringRef::npos)
409         NewFn = Intrinsic::getDeclaration(F->getParent(),
410                                           StoreInts[fArgs.size() - 3], Tys);
411       else
412         NewFn = Intrinsic::getDeclaration(F->getParent(),
413                                           StoreLaneInts[fArgs.size() - 5], Tys);
414       return true;
415     }
416     if (Name == "aarch64.thread.pointer" || Name == "arm.thread.pointer") {
417       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
418       return true;
419     }
420     break;
421   }
422 
423   case 'c': {
424     if (Name.startswith("ctlz.") && F->arg_size() == 1) {
425       rename(F);
426       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
427                                         F->arg_begin()->getType());
428       return true;
429     }
430     if (Name.startswith("cttz.") && F->arg_size() == 1) {
431       rename(F);
432       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
433                                         F->arg_begin()->getType());
434       return true;
435     }
436     break;
437   }
438   case 'd': {
439     if (Name == "dbg.value" && F->arg_size() == 4) {
440       rename(F);
441       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_value);
442       return true;
443     }
444     break;
445   }
446   case 'i':
447   case 'l': {
448     bool IsLifetimeStart = Name.startswith("lifetime.start");
449     if (IsLifetimeStart || Name.startswith("invariant.start")) {
450       Intrinsic::ID ID = IsLifetimeStart ?
451         Intrinsic::lifetime_start : Intrinsic::invariant_start;
452       auto Args = F->getFunctionType()->params();
453       Type* ObjectPtr[1] = {Args[1]};
454       if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
455         rename(F);
456         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
457         return true;
458       }
459     }
460 
461     bool IsLifetimeEnd = Name.startswith("lifetime.end");
462     if (IsLifetimeEnd || Name.startswith("invariant.end")) {
463       Intrinsic::ID ID = IsLifetimeEnd ?
464         Intrinsic::lifetime_end : Intrinsic::invariant_end;
465 
466       auto Args = F->getFunctionType()->params();
467       Type* ObjectPtr[1] = {Args[IsLifetimeEnd ? 1 : 2]};
468       if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
469         rename(F);
470         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
471         return true;
472       }
473     }
474     break;
475   }
476   case 'm': {
477     if (Name.startswith("masked.load.")) {
478       Type *Tys[] = { F->getReturnType(), F->arg_begin()->getType() };
479       if (F->getName() != Intrinsic::getName(Intrinsic::masked_load, Tys)) {
480         rename(F);
481         NewFn = Intrinsic::getDeclaration(F->getParent(),
482                                           Intrinsic::masked_load,
483                                           Tys);
484         return true;
485       }
486     }
487     if (Name.startswith("masked.store.")) {
488       auto Args = F->getFunctionType()->params();
489       Type *Tys[] = { Args[0], Args[1] };
490       if (F->getName() != Intrinsic::getName(Intrinsic::masked_store, Tys)) {
491         rename(F);
492         NewFn = Intrinsic::getDeclaration(F->getParent(),
493                                           Intrinsic::masked_store,
494                                           Tys);
495         return true;
496       }
497     }
498     // Renaming gather/scatter intrinsics with no address space overloading
499     // to the new overload which includes an address space
500     if (Name.startswith("masked.gather.")) {
501       Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
502       if (F->getName() != Intrinsic::getName(Intrinsic::masked_gather, Tys)) {
503         rename(F);
504         NewFn = Intrinsic::getDeclaration(F->getParent(),
505                                           Intrinsic::masked_gather, Tys);
506         return true;
507       }
508     }
509     if (Name.startswith("masked.scatter.")) {
510       auto Args = F->getFunctionType()->params();
511       Type *Tys[] = {Args[0], Args[1]};
512       if (F->getName() != Intrinsic::getName(Intrinsic::masked_scatter, Tys)) {
513         rename(F);
514         NewFn = Intrinsic::getDeclaration(F->getParent(),
515                                           Intrinsic::masked_scatter, Tys);
516         return true;
517       }
518     }
519     break;
520   }
521   case 'n': {
522     if (Name.startswith("nvvm.")) {
523       Name = Name.substr(5);
524 
525       // The following nvvm intrinsics correspond exactly to an LLVM intrinsic.
526       Intrinsic::ID IID = StringSwitch<Intrinsic::ID>(Name)
527                               .Cases("brev32", "brev64", Intrinsic::bitreverse)
528                               .Case("clz.i", Intrinsic::ctlz)
529                               .Case("popc.i", Intrinsic::ctpop)
530                               .Default(Intrinsic::not_intrinsic);
531       if (IID != Intrinsic::not_intrinsic && F->arg_size() == 1) {
532         NewFn = Intrinsic::getDeclaration(F->getParent(), IID,
533                                           {F->getReturnType()});
534         return true;
535       }
536 
537       // The following nvvm intrinsics correspond exactly to an LLVM idiom, but
538       // not to an intrinsic alone.  We expand them in UpgradeIntrinsicCall.
539       //
540       // TODO: We could add lohi.i2d.
541       bool Expand = StringSwitch<bool>(Name)
542                         .Cases("abs.i", "abs.ll", true)
543                         .Cases("clz.ll", "popc.ll", "h2f", true)
544                         .Cases("max.i", "max.ll", "max.ui", "max.ull", true)
545                         .Cases("min.i", "min.ll", "min.ui", "min.ull", true)
546                         .Default(false);
547       if (Expand) {
548         NewFn = nullptr;
549         return true;
550       }
551     }
552     break;
553   }
554   case 'o':
555     // We only need to change the name to match the mangling including the
556     // address space.
557     if (Name.startswith("objectsize.")) {
558       Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
559       if (F->arg_size() == 2 ||
560           F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
561         rename(F);
562         NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::objectsize,
563                                           Tys);
564         return true;
565       }
566     }
567     break;
568 
569   case 's':
570     if (Name == "stackprotectorcheck") {
571       NewFn = nullptr;
572       return true;
573     }
574     break;
575 
576   case 'x':
577     if (UpgradeX86IntrinsicFunction(F, Name, NewFn))
578       return true;
579   }
580   // Remangle our intrinsic since we upgrade the mangling
581   auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F);
582   if (Result != None) {
583     NewFn = Result.getValue();
584     return true;
585   }
586 
587   //  This may not belong here. This function is effectively being overloaded
588   //  to both detect an intrinsic which needs upgrading, and to provide the
589   //  upgraded form of the intrinsic. We should perhaps have two separate
590   //  functions for this.
591   return false;
592 }
593 
594 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
595   NewFn = nullptr;
596   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
597   assert(F != NewFn && "Intrinsic function upgraded to the same function");
598 
599   // Upgrade intrinsic attributes.  This does not change the function.
600   if (NewFn)
601     F = NewFn;
602   if (Intrinsic::ID id = F->getIntrinsicID())
603     F->setAttributes(Intrinsic::getAttributes(F->getContext(), id));
604   return Upgraded;
605 }
606 
607 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
608   // Nothing to do yet.
609   return false;
610 }
611 
612 // Handles upgrading SSE2/AVX2/AVX512BW PSLLDQ intrinsics by converting them
613 // to byte shuffles.
614 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder,
615                                          Value *Op, unsigned Shift) {
616   Type *ResultTy = Op->getType();
617   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
618 
619   // Bitcast from a 64-bit element type to a byte element type.
620   Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
621   Op = Builder.CreateBitCast(Op, VecTy, "cast");
622 
623   // We'll be shuffling in zeroes.
624   Value *Res = Constant::getNullValue(VecTy);
625 
626   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
627   // we'll just return the zero vector.
628   if (Shift < 16) {
629     uint32_t Idxs[64];
630     // 256/512-bit version is split into 2/4 16-byte lanes.
631     for (unsigned l = 0; l != NumElts; l += 16)
632       for (unsigned i = 0; i != 16; ++i) {
633         unsigned Idx = NumElts + i - Shift;
634         if (Idx < NumElts)
635           Idx -= NumElts - 16; // end of lane, switch operand.
636         Idxs[l + i] = Idx + l;
637       }
638 
639     Res = Builder.CreateShuffleVector(Res, Op, makeArrayRef(Idxs, NumElts));
640   }
641 
642   // Bitcast back to a 64-bit element type.
643   return Builder.CreateBitCast(Res, ResultTy, "cast");
644 }
645 
646 // Handles upgrading SSE2/AVX2/AVX512BW PSRLDQ intrinsics by converting them
647 // to byte shuffles.
648 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, Value *Op,
649                                          unsigned Shift) {
650   Type *ResultTy = Op->getType();
651   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
652 
653   // Bitcast from a 64-bit element type to a byte element type.
654   Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
655   Op = Builder.CreateBitCast(Op, VecTy, "cast");
656 
657   // We'll be shuffling in zeroes.
658   Value *Res = Constant::getNullValue(VecTy);
659 
660   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
661   // we'll just return the zero vector.
662   if (Shift < 16) {
663     uint32_t Idxs[64];
664     // 256/512-bit version is split into 2/4 16-byte lanes.
665     for (unsigned l = 0; l != NumElts; l += 16)
666       for (unsigned i = 0; i != 16; ++i) {
667         unsigned Idx = i + Shift;
668         if (Idx >= 16)
669           Idx += NumElts - 16; // end of lane, switch operand.
670         Idxs[l + i] = Idx + l;
671       }
672 
673     Res = Builder.CreateShuffleVector(Op, Res, makeArrayRef(Idxs, NumElts));
674   }
675 
676   // Bitcast back to a 64-bit element type.
677   return Builder.CreateBitCast(Res, ResultTy, "cast");
678 }
679 
680 static Value *getX86MaskVec(IRBuilder<> &Builder, Value *Mask,
681                             unsigned NumElts) {
682   llvm::VectorType *MaskTy = llvm::VectorType::get(Builder.getInt1Ty(),
683                              cast<IntegerType>(Mask->getType())->getBitWidth());
684   Mask = Builder.CreateBitCast(Mask, MaskTy);
685 
686   // If we have less than 8 elements, then the starting mask was an i8 and
687   // we need to extract down to the right number of elements.
688   if (NumElts < 8) {
689     uint32_t Indices[4];
690     for (unsigned i = 0; i != NumElts; ++i)
691       Indices[i] = i;
692     Mask = Builder.CreateShuffleVector(Mask, Mask,
693                                        makeArrayRef(Indices, NumElts),
694                                        "extract");
695   }
696 
697   return Mask;
698 }
699 
700 static Value *EmitX86Select(IRBuilder<> &Builder, Value *Mask,
701                             Value *Op0, Value *Op1) {
702   // If the mask is all ones just emit the align operation.
703   if (const auto *C = dyn_cast<Constant>(Mask))
704     if (C->isAllOnesValue())
705       return Op0;
706 
707   Mask = getX86MaskVec(Builder, Mask, Op0->getType()->getVectorNumElements());
708   return Builder.CreateSelect(Mask, Op0, Op1);
709 }
710 
711 // Handle autoupgrade for masked PALIGNR and VALIGND/Q intrinsics.
712 // PALIGNR handles large immediates by shifting while VALIGN masks the immediate
713 // so we need to handle both cases. VALIGN also doesn't have 128-bit lanes.
714 static Value *UpgradeX86ALIGNIntrinsics(IRBuilder<> &Builder, Value *Op0,
715                                         Value *Op1, Value *Shift,
716                                         Value *Passthru, Value *Mask,
717                                         bool IsVALIGN) {
718   unsigned ShiftVal = cast<llvm::ConstantInt>(Shift)->getZExtValue();
719 
720   unsigned NumElts = Op0->getType()->getVectorNumElements();
721   assert((IsVALIGN || NumElts % 16 == 0) && "Illegal NumElts for PALIGNR!");
722   assert((!IsVALIGN || NumElts <= 16) && "NumElts too large for VALIGN!");
723   assert(isPowerOf2_32(NumElts) && "NumElts not a power of 2!");
724 
725   // Mask the immediate for VALIGN.
726   if (IsVALIGN)
727     ShiftVal &= (NumElts - 1);
728 
729   // If palignr is shifting the pair of vectors more than the size of two
730   // lanes, emit zero.
731   if (ShiftVal >= 32)
732     return llvm::Constant::getNullValue(Op0->getType());
733 
734   // If palignr is shifting the pair of input vectors more than one lane,
735   // but less than two lanes, convert to shifting in zeroes.
736   if (ShiftVal > 16) {
737     ShiftVal -= 16;
738     Op1 = Op0;
739     Op0 = llvm::Constant::getNullValue(Op0->getType());
740   }
741 
742   uint32_t Indices[64];
743   // 256-bit palignr operates on 128-bit lanes so we need to handle that
744   for (unsigned l = 0; l < NumElts; l += 16) {
745     for (unsigned i = 0; i != 16; ++i) {
746       unsigned Idx = ShiftVal + i;
747       if (!IsVALIGN && Idx >= 16) // Disable wrap for VALIGN.
748         Idx += NumElts - 16; // End of lane, switch operand.
749       Indices[l + i] = Idx + l;
750     }
751   }
752 
753   Value *Align = Builder.CreateShuffleVector(Op1, Op0,
754                                              makeArrayRef(Indices, NumElts),
755                                              "palignr");
756 
757   return EmitX86Select(Builder, Mask, Align, Passthru);
758 }
759 
760 static Value *UpgradeMaskedStore(IRBuilder<> &Builder,
761                                  Value *Ptr, Value *Data, Value *Mask,
762                                  bool Aligned) {
763   // Cast the pointer to the right type.
764   Ptr = Builder.CreateBitCast(Ptr,
765                               llvm::PointerType::getUnqual(Data->getType()));
766   unsigned Align =
767     Aligned ? cast<VectorType>(Data->getType())->getBitWidth() / 8 : 1;
768 
769   // If the mask is all ones just emit a regular store.
770   if (const auto *C = dyn_cast<Constant>(Mask))
771     if (C->isAllOnesValue())
772       return Builder.CreateAlignedStore(Data, Ptr, Align);
773 
774   // Convert the mask from an integer type to a vector of i1.
775   unsigned NumElts = Data->getType()->getVectorNumElements();
776   Mask = getX86MaskVec(Builder, Mask, NumElts);
777   return Builder.CreateMaskedStore(Data, Ptr, Align, Mask);
778 }
779 
780 static Value *UpgradeMaskedLoad(IRBuilder<> &Builder,
781                                 Value *Ptr, Value *Passthru, Value *Mask,
782                                 bool Aligned) {
783   // Cast the pointer to the right type.
784   Ptr = Builder.CreateBitCast(Ptr,
785                              llvm::PointerType::getUnqual(Passthru->getType()));
786   unsigned Align =
787     Aligned ? cast<VectorType>(Passthru->getType())->getBitWidth() / 8 : 1;
788 
789   // If the mask is all ones just emit a regular store.
790   if (const auto *C = dyn_cast<Constant>(Mask))
791     if (C->isAllOnesValue())
792       return Builder.CreateAlignedLoad(Ptr, Align);
793 
794   // Convert the mask from an integer type to a vector of i1.
795   unsigned NumElts = Passthru->getType()->getVectorNumElements();
796   Mask = getX86MaskVec(Builder, Mask, NumElts);
797   return Builder.CreateMaskedLoad(Ptr, Align, Mask, Passthru);
798 }
799 
800 static Value *upgradeAbs(IRBuilder<> &Builder, CallInst &CI) {
801   Value *Op0 = CI.getArgOperand(0);
802   llvm::Type *Ty = Op0->getType();
803   Value *Zero = llvm::Constant::getNullValue(Ty);
804   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_SGT, Op0, Zero);
805   Value *Neg = Builder.CreateNeg(Op0);
806   Value *Res = Builder.CreateSelect(Cmp, Op0, Neg);
807 
808   if (CI.getNumArgOperands() == 3)
809     Res = EmitX86Select(Builder,CI.getArgOperand(2), Res, CI.getArgOperand(1));
810 
811   return Res;
812 }
813 
814 static Value *upgradeIntMinMax(IRBuilder<> &Builder, CallInst &CI,
815                                ICmpInst::Predicate Pred) {
816   Value *Op0 = CI.getArgOperand(0);
817   Value *Op1 = CI.getArgOperand(1);
818   Value *Cmp = Builder.CreateICmp(Pred, Op0, Op1);
819   Value *Res = Builder.CreateSelect(Cmp, Op0, Op1);
820 
821   if (CI.getNumArgOperands() == 4)
822     Res = EmitX86Select(Builder, CI.getArgOperand(3), Res, CI.getArgOperand(2));
823 
824   return Res;
825 }
826 
827 static Value *upgradeMaskedCompare(IRBuilder<> &Builder, CallInst &CI,
828                                    unsigned CC, bool Signed) {
829   Value *Op0 = CI.getArgOperand(0);
830   unsigned NumElts = Op0->getType()->getVectorNumElements();
831 
832   Value *Cmp;
833   if (CC == 3) {
834     Cmp = Constant::getNullValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
835   } else if (CC == 7) {
836     Cmp = Constant::getAllOnesValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
837   } else {
838     ICmpInst::Predicate Pred;
839     switch (CC) {
840     default: llvm_unreachable("Unknown condition code");
841     case 0: Pred = ICmpInst::ICMP_EQ;  break;
842     case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
843     case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
844     case 4: Pred = ICmpInst::ICMP_NE;  break;
845     case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
846     case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
847     }
848     Cmp = Builder.CreateICmp(Pred, Op0, CI.getArgOperand(1));
849   }
850 
851   Value *Mask = CI.getArgOperand(CI.getNumArgOperands() - 1);
852   const auto *C = dyn_cast<Constant>(Mask);
853   if (!C || !C->isAllOnesValue())
854     Cmp = Builder.CreateAnd(Cmp, getX86MaskVec(Builder, Mask, NumElts));
855 
856   if (NumElts < 8) {
857     uint32_t Indices[8];
858     for (unsigned i = 0; i != NumElts; ++i)
859       Indices[i] = i;
860     for (unsigned i = NumElts; i != 8; ++i)
861       Indices[i] = NumElts + i % NumElts;
862     Cmp = Builder.CreateShuffleVector(Cmp,
863                                       Constant::getNullValue(Cmp->getType()),
864                                       Indices);
865   }
866   return Builder.CreateBitCast(Cmp, IntegerType::get(CI.getContext(),
867                                                      std::max(NumElts, 8U)));
868 }
869 
870 // Replace a masked intrinsic with an older unmasked intrinsic.
871 static Value *UpgradeX86MaskedShift(IRBuilder<> &Builder, CallInst &CI,
872                                     Intrinsic::ID IID) {
873   Function *F = CI.getCalledFunction();
874   Function *Intrin = Intrinsic::getDeclaration(F->getParent(), IID);
875   Value *Rep = Builder.CreateCall(Intrin,
876                                  { CI.getArgOperand(0), CI.getArgOperand(1) });
877   return EmitX86Select(Builder, CI.getArgOperand(3), Rep, CI.getArgOperand(2));
878 }
879 
880 static Value* upgradeMaskedMove(IRBuilder<> &Builder, CallInst &CI) {
881   Value* A = CI.getArgOperand(0);
882   Value* B = CI.getArgOperand(1);
883   Value* Src = CI.getArgOperand(2);
884   Value* Mask = CI.getArgOperand(3);
885 
886   Value* AndNode = Builder.CreateAnd(Mask, APInt(8, 1));
887   Value* Cmp = Builder.CreateIsNotNull(AndNode);
888   Value* Extract1 = Builder.CreateExtractElement(B, (uint64_t)0);
889   Value* Extract2 = Builder.CreateExtractElement(Src, (uint64_t)0);
890   Value* Select = Builder.CreateSelect(Cmp, Extract1, Extract2);
891   return Builder.CreateInsertElement(A, Select, (uint64_t)0);
892 }
893 
894 
895 static Value* UpgradeMaskToInt(IRBuilder<> &Builder, CallInst &CI) {
896   Value* Op = CI.getArgOperand(0);
897   Type* ReturnOp = CI.getType();
898   unsigned NumElts = CI.getType()->getVectorNumElements();
899   Value *Mask = getX86MaskVec(Builder, Op, NumElts);
900   return Builder.CreateSExt(Mask, ReturnOp, "vpmovm2");
901 }
902 
903 /// Upgrade a call to an old intrinsic. All argument and return casting must be
904 /// provided to seamlessly integrate with existing context.
905 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
906   Function *F = CI->getCalledFunction();
907   LLVMContext &C = CI->getContext();
908   IRBuilder<> Builder(C);
909   Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
910 
911   assert(F && "Intrinsic call is not direct?");
912 
913   if (!NewFn) {
914     // Get the Function's name.
915     StringRef Name = F->getName();
916 
917     assert(Name.startswith("llvm.") && "Intrinsic doesn't start with 'llvm.'");
918     Name = Name.substr(5);
919 
920     bool IsX86 = Name.startswith("x86.");
921     if (IsX86)
922       Name = Name.substr(4);
923     bool IsNVVM = Name.startswith("nvvm.");
924     if (IsNVVM)
925       Name = Name.substr(5);
926 
927     if (IsX86 && Name.startswith("sse4a.movnt.")) {
928       Module *M = F->getParent();
929       SmallVector<Metadata *, 1> Elts;
930       Elts.push_back(
931           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
932       MDNode *Node = MDNode::get(C, Elts);
933 
934       Value *Arg0 = CI->getArgOperand(0);
935       Value *Arg1 = CI->getArgOperand(1);
936 
937       // Nontemporal (unaligned) store of the 0'th element of the float/double
938       // vector.
939       Type *SrcEltTy = cast<VectorType>(Arg1->getType())->getElementType();
940       PointerType *EltPtrTy = PointerType::getUnqual(SrcEltTy);
941       Value *Addr = Builder.CreateBitCast(Arg0, EltPtrTy, "cast");
942       Value *Extract =
943           Builder.CreateExtractElement(Arg1, (uint64_t)0, "extractelement");
944 
945       StoreInst *SI = Builder.CreateAlignedStore(Extract, Addr, 1);
946       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
947 
948       // Remove intrinsic.
949       CI->eraseFromParent();
950       return;
951     }
952 
953     if (IsX86 && (Name.startswith("avx.movnt.") ||
954                   Name.startswith("avx512.storent."))) {
955       Module *M = F->getParent();
956       SmallVector<Metadata *, 1> Elts;
957       Elts.push_back(
958           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
959       MDNode *Node = MDNode::get(C, Elts);
960 
961       Value *Arg0 = CI->getArgOperand(0);
962       Value *Arg1 = CI->getArgOperand(1);
963 
964       // Convert the type of the pointer to a pointer to the stored type.
965       Value *BC = Builder.CreateBitCast(Arg0,
966                                         PointerType::getUnqual(Arg1->getType()),
967                                         "cast");
968       VectorType *VTy = cast<VectorType>(Arg1->getType());
969       StoreInst *SI = Builder.CreateAlignedStore(Arg1, BC,
970                                                  VTy->getBitWidth() / 8);
971       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
972 
973       // Remove intrinsic.
974       CI->eraseFromParent();
975       return;
976     }
977 
978     if (IsX86 && Name == "sse2.storel.dq") {
979       Value *Arg0 = CI->getArgOperand(0);
980       Value *Arg1 = CI->getArgOperand(1);
981 
982       Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
983       Value *BC0 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
984       Value *Elt = Builder.CreateExtractElement(BC0, (uint64_t)0);
985       Value *BC = Builder.CreateBitCast(Arg0,
986                                         PointerType::getUnqual(Elt->getType()),
987                                         "cast");
988       Builder.CreateAlignedStore(Elt, BC, 1);
989 
990       // Remove intrinsic.
991       CI->eraseFromParent();
992       return;
993     }
994 
995     if (IsX86 && (Name.startswith("sse.storeu.") ||
996                   Name.startswith("sse2.storeu.") ||
997                   Name.startswith("avx.storeu."))) {
998       Value *Arg0 = CI->getArgOperand(0);
999       Value *Arg1 = CI->getArgOperand(1);
1000 
1001       Arg0 = Builder.CreateBitCast(Arg0,
1002                                    PointerType::getUnqual(Arg1->getType()),
1003                                    "cast");
1004       Builder.CreateAlignedStore(Arg1, Arg0, 1);
1005 
1006       // Remove intrinsic.
1007       CI->eraseFromParent();
1008       return;
1009     }
1010 
1011     if (IsX86 && (Name.startswith("avx512.mask.store"))) {
1012       // "avx512.mask.storeu." or "avx512.mask.store."
1013       bool Aligned = Name[17] != 'u'; // "avx512.mask.storeu".
1014       UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
1015                          CI->getArgOperand(2), Aligned);
1016 
1017       // Remove intrinsic.
1018       CI->eraseFromParent();
1019       return;
1020     }
1021 
1022     Value *Rep;
1023     // Upgrade packed integer vector compare intrinsics to compare instructions.
1024     if (IsX86 && (Name.startswith("sse2.pcmp") ||
1025                   Name.startswith("avx2.pcmp"))) {
1026       // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt."
1027       bool CmpEq = Name[9] == 'e';
1028       Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
1029                                CI->getArgOperand(0), CI->getArgOperand(1));
1030       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
1031     } else if (IsX86 && (Name.startswith("avx512.broadcastm"))) {
1032       Type *ExtTy = Type::getInt32Ty(C);
1033       if (CI->getOperand(0)->getType()->isIntegerTy(8))
1034         ExtTy = Type::getInt64Ty(C);
1035       unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() /
1036                          ExtTy->getPrimitiveSizeInBits();
1037       Rep = Builder.CreateZExt(CI->getArgOperand(0), ExtTy);
1038       Rep = Builder.CreateVectorSplat(NumElts, Rep);
1039     } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))) {
1040       unsigned NumElts =
1041           CI->getArgOperand(1)->getType()->getVectorNumElements();
1042       Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
1043       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1044                           CI->getArgOperand(1));
1045     } else if (IsX86 && (Name == "sse.add.ss" || Name == "sse2.add.sd")) {
1046       Type *I32Ty = Type::getInt32Ty(C);
1047       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1048                                                  ConstantInt::get(I32Ty, 0));
1049       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1050                                                  ConstantInt::get(I32Ty, 0));
1051       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1052                                         Builder.CreateFAdd(Elt0, Elt1),
1053                                         ConstantInt::get(I32Ty, 0));
1054     } else if (IsX86 && (Name == "sse.sub.ss" || Name == "sse2.sub.sd")) {
1055       Type *I32Ty = Type::getInt32Ty(C);
1056       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1057                                                  ConstantInt::get(I32Ty, 0));
1058       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1059                                                  ConstantInt::get(I32Ty, 0));
1060       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1061                                         Builder.CreateFSub(Elt0, Elt1),
1062                                         ConstantInt::get(I32Ty, 0));
1063     } else if (IsX86 && (Name == "sse.mul.ss" || Name == "sse2.mul.sd")) {
1064       Type *I32Ty = Type::getInt32Ty(C);
1065       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1066                                                  ConstantInt::get(I32Ty, 0));
1067       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1068                                                  ConstantInt::get(I32Ty, 0));
1069       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1070                                         Builder.CreateFMul(Elt0, Elt1),
1071                                         ConstantInt::get(I32Ty, 0));
1072     } else if (IsX86 && (Name == "sse.div.ss" || Name == "sse2.div.sd")) {
1073       Type *I32Ty = Type::getInt32Ty(C);
1074       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1075                                                  ConstantInt::get(I32Ty, 0));
1076       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1077                                                  ConstantInt::get(I32Ty, 0));
1078       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1079                                         Builder.CreateFDiv(Elt0, Elt1),
1080                                         ConstantInt::get(I32Ty, 0));
1081     } else if (IsX86 && Name.startswith("avx512.mask.pcmp")) {
1082       // "avx512.mask.pcmpeq." or "avx512.mask.pcmpgt."
1083       bool CmpEq = Name[16] == 'e';
1084       Rep = upgradeMaskedCompare(Builder, *CI, CmpEq ? 0 : 6, true);
1085     } else if (IsX86 && Name.startswith("avx512.mask.cmp")) {
1086       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1087       Rep = upgradeMaskedCompare(Builder, *CI, Imm, true);
1088     } else if (IsX86 && Name.startswith("avx512.mask.ucmp")) {
1089       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1090       Rep = upgradeMaskedCompare(Builder, *CI, Imm, false);
1091     } else if(IsX86 && (Name == "ssse3.pabs.b.128" ||
1092                         Name == "ssse3.pabs.w.128" ||
1093                         Name == "ssse3.pabs.d.128" ||
1094                         Name.startswith("avx2.pabs") ||
1095                         Name.startswith("avx512.mask.pabs"))) {
1096       Rep = upgradeAbs(Builder, *CI);
1097     } else if (IsX86 && (Name == "sse41.pmaxsb" ||
1098                          Name == "sse2.pmaxs.w" ||
1099                          Name == "sse41.pmaxsd" ||
1100                          Name.startswith("avx2.pmaxs") ||
1101                          Name.startswith("avx512.mask.pmaxs"))) {
1102       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SGT);
1103     } else if (IsX86 && (Name == "sse2.pmaxu.b" ||
1104                          Name == "sse41.pmaxuw" ||
1105                          Name == "sse41.pmaxud" ||
1106                          Name.startswith("avx2.pmaxu") ||
1107                          Name.startswith("avx512.mask.pmaxu"))) {
1108       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_UGT);
1109     } else if (IsX86 && (Name == "sse41.pminsb" ||
1110                          Name == "sse2.pmins.w" ||
1111                          Name == "sse41.pminsd" ||
1112                          Name.startswith("avx2.pmins") ||
1113                          Name.startswith("avx512.mask.pmins"))) {
1114       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SLT);
1115     } else if (IsX86 && (Name == "sse2.pminu.b" ||
1116                          Name == "sse41.pminuw" ||
1117                          Name == "sse41.pminud" ||
1118                          Name.startswith("avx2.pminu") ||
1119                          Name.startswith("avx512.mask.pminu"))) {
1120       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_ULT);
1121     } else if (IsX86 && (Name == "sse2.cvtdq2pd" ||
1122                          Name == "sse2.cvtps2pd" ||
1123                          Name == "avx.cvtdq2.pd.256" ||
1124                          Name == "avx.cvt.ps2.pd.256" ||
1125                          Name.startswith("avx512.mask.cvtdq2pd.") ||
1126                          Name.startswith("avx512.mask.cvtudq2pd."))) {
1127       // Lossless i32/float to double conversion.
1128       // Extract the bottom elements if necessary and convert to double vector.
1129       Value *Src = CI->getArgOperand(0);
1130       VectorType *SrcTy = cast<VectorType>(Src->getType());
1131       VectorType *DstTy = cast<VectorType>(CI->getType());
1132       Rep = CI->getArgOperand(0);
1133 
1134       unsigned NumDstElts = DstTy->getNumElements();
1135       if (NumDstElts < SrcTy->getNumElements()) {
1136         assert(NumDstElts == 2 && "Unexpected vector size");
1137         uint32_t ShuffleMask[2] = { 0, 1 };
1138         Rep = Builder.CreateShuffleVector(Rep, UndefValue::get(SrcTy),
1139                                           ShuffleMask);
1140       }
1141 
1142       bool SInt2Double = (StringRef::npos != Name.find("cvtdq2"));
1143       bool UInt2Double = (StringRef::npos != Name.find("cvtudq2"));
1144       if (SInt2Double)
1145         Rep = Builder.CreateSIToFP(Rep, DstTy, "cvtdq2pd");
1146       else if (UInt2Double)
1147         Rep = Builder.CreateUIToFP(Rep, DstTy, "cvtudq2pd");
1148       else
1149         Rep = Builder.CreateFPExt(Rep, DstTy, "cvtps2pd");
1150 
1151       if (CI->getNumArgOperands() == 3)
1152         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1153                             CI->getArgOperand(1));
1154     } else if (IsX86 && (Name.startswith("avx512.mask.loadu."))) {
1155       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
1156                               CI->getArgOperand(1), CI->getArgOperand(2),
1157                               /*Aligned*/false);
1158     } else if (IsX86 && (Name.startswith("avx512.mask.load."))) {
1159       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
1160                               CI->getArgOperand(1),CI->getArgOperand(2),
1161                               /*Aligned*/true);
1162     } else if (IsX86 && Name.startswith("xop.vpcom")) {
1163       Intrinsic::ID intID;
1164       if (Name.endswith("ub"))
1165         intID = Intrinsic::x86_xop_vpcomub;
1166       else if (Name.endswith("uw"))
1167         intID = Intrinsic::x86_xop_vpcomuw;
1168       else if (Name.endswith("ud"))
1169         intID = Intrinsic::x86_xop_vpcomud;
1170       else if (Name.endswith("uq"))
1171         intID = Intrinsic::x86_xop_vpcomuq;
1172       else if (Name.endswith("b"))
1173         intID = Intrinsic::x86_xop_vpcomb;
1174       else if (Name.endswith("w"))
1175         intID = Intrinsic::x86_xop_vpcomw;
1176       else if (Name.endswith("d"))
1177         intID = Intrinsic::x86_xop_vpcomd;
1178       else if (Name.endswith("q"))
1179         intID = Intrinsic::x86_xop_vpcomq;
1180       else
1181         llvm_unreachable("Unknown suffix");
1182 
1183       Name = Name.substr(9); // strip off "xop.vpcom"
1184       unsigned Imm;
1185       if (Name.startswith("lt"))
1186         Imm = 0;
1187       else if (Name.startswith("le"))
1188         Imm = 1;
1189       else if (Name.startswith("gt"))
1190         Imm = 2;
1191       else if (Name.startswith("ge"))
1192         Imm = 3;
1193       else if (Name.startswith("eq"))
1194         Imm = 4;
1195       else if (Name.startswith("ne"))
1196         Imm = 5;
1197       else if (Name.startswith("false"))
1198         Imm = 6;
1199       else if (Name.startswith("true"))
1200         Imm = 7;
1201       else
1202         llvm_unreachable("Unknown condition");
1203 
1204       Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
1205       Rep =
1206           Builder.CreateCall(VPCOM, {CI->getArgOperand(0), CI->getArgOperand(1),
1207                                      Builder.getInt8(Imm)});
1208     } else if (IsX86 && Name.startswith("xop.vpcmov")) {
1209       Value *Sel = CI->getArgOperand(2);
1210       Value *NotSel = Builder.CreateNot(Sel);
1211       Value *Sel0 = Builder.CreateAnd(CI->getArgOperand(0), Sel);
1212       Value *Sel1 = Builder.CreateAnd(CI->getArgOperand(1), NotSel);
1213       Rep = Builder.CreateOr(Sel0, Sel1);
1214     } else if (IsX86 && Name == "sse42.crc32.64.8") {
1215       Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
1216                                                Intrinsic::x86_sse42_crc32_32_8);
1217       Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
1218       Rep = Builder.CreateCall(CRC32, {Trunc0, CI->getArgOperand(1)});
1219       Rep = Builder.CreateZExt(Rep, CI->getType(), "");
1220     } else if (IsX86 && Name.startswith("avx.vbroadcast.s")) {
1221       // Replace broadcasts with a series of insertelements.
1222       Type *VecTy = CI->getType();
1223       Type *EltTy = VecTy->getVectorElementType();
1224       unsigned EltNum = VecTy->getVectorNumElements();
1225       Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
1226                                           EltTy->getPointerTo());
1227       Value *Load = Builder.CreateLoad(EltTy, Cast);
1228       Type *I32Ty = Type::getInt32Ty(C);
1229       Rep = UndefValue::get(VecTy);
1230       for (unsigned I = 0; I < EltNum; ++I)
1231         Rep = Builder.CreateInsertElement(Rep, Load,
1232                                           ConstantInt::get(I32Ty, I));
1233     } else if (IsX86 && (Name.startswith("sse41.pmovsx") ||
1234                          Name.startswith("sse41.pmovzx") ||
1235                          Name.startswith("avx2.pmovsx") ||
1236                          Name.startswith("avx2.pmovzx") ||
1237                          Name.startswith("avx512.mask.pmovsx") ||
1238                          Name.startswith("avx512.mask.pmovzx"))) {
1239       VectorType *SrcTy = cast<VectorType>(CI->getArgOperand(0)->getType());
1240       VectorType *DstTy = cast<VectorType>(CI->getType());
1241       unsigned NumDstElts = DstTy->getNumElements();
1242 
1243       // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
1244       SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
1245       for (unsigned i = 0; i != NumDstElts; ++i)
1246         ShuffleMask[i] = i;
1247 
1248       Value *SV = Builder.CreateShuffleVector(
1249           CI->getArgOperand(0), UndefValue::get(SrcTy), ShuffleMask);
1250 
1251       bool DoSext = (StringRef::npos != Name.find("pmovsx"));
1252       Rep = DoSext ? Builder.CreateSExt(SV, DstTy)
1253                    : Builder.CreateZExt(SV, DstTy);
1254       // If there are 3 arguments, it's a masked intrinsic so we need a select.
1255       if (CI->getNumArgOperands() == 3)
1256         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1257                             CI->getArgOperand(1));
1258     } else if (IsX86 && (Name.startswith("avx.vbroadcastf128") ||
1259                          Name == "avx2.vbroadcasti128")) {
1260       // Replace vbroadcastf128/vbroadcasti128 with a vector load+shuffle.
1261       Type *EltTy = CI->getType()->getVectorElementType();
1262       unsigned NumSrcElts = 128 / EltTy->getPrimitiveSizeInBits();
1263       Type *VT = VectorType::get(EltTy, NumSrcElts);
1264       Value *Op = Builder.CreatePointerCast(CI->getArgOperand(0),
1265                                             PointerType::getUnqual(VT));
1266       Value *Load = Builder.CreateAlignedLoad(Op, 1);
1267       if (NumSrcElts == 2)
1268         Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
1269                                           { 0, 1, 0, 1 });
1270       else
1271         Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
1272                                           { 0, 1, 2, 3, 0, 1, 2, 3 });
1273     } else if (IsX86 && (Name.startswith("avx512.mask.broadcastf") ||
1274                          Name.startswith("avx512.mask.broadcasti"))) {
1275       unsigned NumSrcElts =
1276                         CI->getArgOperand(0)->getType()->getVectorNumElements();
1277       unsigned NumDstElts = CI->getType()->getVectorNumElements();
1278 
1279       SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
1280       for (unsigned i = 0; i != NumDstElts; ++i)
1281         ShuffleMask[i] = i % NumSrcElts;
1282 
1283       Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
1284                                         CI->getArgOperand(0),
1285                                         ShuffleMask);
1286       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1287                           CI->getArgOperand(1));
1288     } else if (IsX86 && (Name.startswith("avx2.pbroadcast") ||
1289                          Name.startswith("avx2.vbroadcast") ||
1290                          Name.startswith("avx512.pbroadcast") ||
1291                          Name.startswith("avx512.mask.broadcast.s"))) {
1292       // Replace vp?broadcasts with a vector shuffle.
1293       Value *Op = CI->getArgOperand(0);
1294       unsigned NumElts = CI->getType()->getVectorNumElements();
1295       Type *MaskTy = VectorType::get(Type::getInt32Ty(C), NumElts);
1296       Rep = Builder.CreateShuffleVector(Op, UndefValue::get(Op->getType()),
1297                                         Constant::getNullValue(MaskTy));
1298 
1299       if (CI->getNumArgOperands() == 3)
1300         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1301                             CI->getArgOperand(1));
1302     } else if (IsX86 && Name.startswith("avx512.mask.palignr.")) {
1303       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
1304                                       CI->getArgOperand(1),
1305                                       CI->getArgOperand(2),
1306                                       CI->getArgOperand(3),
1307                                       CI->getArgOperand(4),
1308                                       false);
1309     } else if (IsX86 && Name.startswith("avx512.mask.valign.")) {
1310       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
1311                                       CI->getArgOperand(1),
1312                                       CI->getArgOperand(2),
1313                                       CI->getArgOperand(3),
1314                                       CI->getArgOperand(4),
1315                                       true);
1316     } else if (IsX86 && (Name == "sse2.psll.dq" ||
1317                          Name == "avx2.psll.dq")) {
1318       // 128/256-bit shift left specified in bits.
1319       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1320       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0),
1321                                        Shift / 8); // Shift is in bits.
1322     } else if (IsX86 && (Name == "sse2.psrl.dq" ||
1323                          Name == "avx2.psrl.dq")) {
1324       // 128/256-bit shift right specified in bits.
1325       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1326       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0),
1327                                        Shift / 8); // Shift is in bits.
1328     } else if (IsX86 && (Name == "sse2.psll.dq.bs" ||
1329                          Name == "avx2.psll.dq.bs" ||
1330                          Name == "avx512.psll.dq.512")) {
1331       // 128/256/512-bit shift left specified in bytes.
1332       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1333       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
1334     } else if (IsX86 && (Name == "sse2.psrl.dq.bs" ||
1335                          Name == "avx2.psrl.dq.bs" ||
1336                          Name == "avx512.psrl.dq.512")) {
1337       // 128/256/512-bit shift right specified in bytes.
1338       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1339       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
1340     } else if (IsX86 && (Name == "sse41.pblendw" ||
1341                          Name.startswith("sse41.blendp") ||
1342                          Name.startswith("avx.blend.p") ||
1343                          Name == "avx2.pblendw" ||
1344                          Name.startswith("avx2.pblendd."))) {
1345       Value *Op0 = CI->getArgOperand(0);
1346       Value *Op1 = CI->getArgOperand(1);
1347       unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1348       VectorType *VecTy = cast<VectorType>(CI->getType());
1349       unsigned NumElts = VecTy->getNumElements();
1350 
1351       SmallVector<uint32_t, 16> Idxs(NumElts);
1352       for (unsigned i = 0; i != NumElts; ++i)
1353         Idxs[i] = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
1354 
1355       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1356     } else if (IsX86 && (Name.startswith("avx.vinsertf128.") ||
1357                          Name == "avx2.vinserti128" ||
1358                          Name.startswith("avx512.mask.insert"))) {
1359       Value *Op0 = CI->getArgOperand(0);
1360       Value *Op1 = CI->getArgOperand(1);
1361       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1362       unsigned DstNumElts = CI->getType()->getVectorNumElements();
1363       unsigned SrcNumElts = Op1->getType()->getVectorNumElements();
1364       unsigned Scale = DstNumElts / SrcNumElts;
1365 
1366       // Mask off the high bits of the immediate value; hardware ignores those.
1367       Imm = Imm % Scale;
1368 
1369       // Extend the second operand into a vector the size of the destination.
1370       Value *UndefV = UndefValue::get(Op1->getType());
1371       SmallVector<uint32_t, 8> Idxs(DstNumElts);
1372       for (unsigned i = 0; i != SrcNumElts; ++i)
1373         Idxs[i] = i;
1374       for (unsigned i = SrcNumElts; i != DstNumElts; ++i)
1375         Idxs[i] = SrcNumElts;
1376       Rep = Builder.CreateShuffleVector(Op1, UndefV, Idxs);
1377 
1378       // Insert the second operand into the first operand.
1379 
1380       // Note that there is no guarantee that instruction lowering will actually
1381       // produce a vinsertf128 instruction for the created shuffles. In
1382       // particular, the 0 immediate case involves no lane changes, so it can
1383       // be handled as a blend.
1384 
1385       // Example of shuffle mask for 32-bit elements:
1386       // Imm = 1  <i32 0, i32 1, i32 2,  i32 3,  i32 8, i32 9, i32 10, i32 11>
1387       // Imm = 0  <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6,  i32 7 >
1388 
1389       // First fill with identify mask.
1390       for (unsigned i = 0; i != DstNumElts; ++i)
1391         Idxs[i] = i;
1392       // Then replace the elements where we need to insert.
1393       for (unsigned i = 0; i != SrcNumElts; ++i)
1394         Idxs[i + Imm * SrcNumElts] = i + DstNumElts;
1395       Rep = Builder.CreateShuffleVector(Op0, Rep, Idxs);
1396 
1397       // If the intrinsic has a mask operand, handle that.
1398       if (CI->getNumArgOperands() == 5)
1399         Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1400                             CI->getArgOperand(3));
1401     } else if (IsX86 && (Name.startswith("avx.vextractf128.") ||
1402                          Name == "avx2.vextracti128" ||
1403                          Name.startswith("avx512.mask.vextract"))) {
1404       Value *Op0 = CI->getArgOperand(0);
1405       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1406       unsigned DstNumElts = CI->getType()->getVectorNumElements();
1407       unsigned SrcNumElts = Op0->getType()->getVectorNumElements();
1408       unsigned Scale = SrcNumElts / DstNumElts;
1409 
1410       // Mask off the high bits of the immediate value; hardware ignores those.
1411       Imm = Imm % Scale;
1412 
1413       // Get indexes for the subvector of the input vector.
1414       SmallVector<uint32_t, 8> Idxs(DstNumElts);
1415       for (unsigned i = 0; i != DstNumElts; ++i) {
1416         Idxs[i] = i + (Imm * DstNumElts);
1417       }
1418       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1419 
1420       // If the intrinsic has a mask operand, handle that.
1421       if (CI->getNumArgOperands() == 4)
1422         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1423                             CI->getArgOperand(2));
1424     } else if (!IsX86 && Name == "stackprotectorcheck") {
1425       Rep = nullptr;
1426     } else if (IsX86 && (Name.startswith("avx512.mask.perm.df.") ||
1427                          Name.startswith("avx512.mask.perm.di."))) {
1428       Value *Op0 = CI->getArgOperand(0);
1429       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1430       VectorType *VecTy = cast<VectorType>(CI->getType());
1431       unsigned NumElts = VecTy->getNumElements();
1432 
1433       SmallVector<uint32_t, 8> Idxs(NumElts);
1434       for (unsigned i = 0; i != NumElts; ++i)
1435         Idxs[i] = (i & ~0x3) + ((Imm >> (2 * (i & 0x3))) & 3);
1436 
1437       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1438 
1439       if (CI->getNumArgOperands() == 4)
1440         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1441                             CI->getArgOperand(2));
1442     } else if (IsX86 && (Name.startswith("avx.vperm2f128.") ||
1443                          Name == "avx2.vperm2i128")) {
1444       // The immediate permute control byte looks like this:
1445       //    [1:0] - select 128 bits from sources for low half of destination
1446       //    [2]   - ignore
1447       //    [3]   - zero low half of destination
1448       //    [5:4] - select 128 bits from sources for high half of destination
1449       //    [6]   - ignore
1450       //    [7]   - zero high half of destination
1451 
1452       uint8_t Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1453 
1454       unsigned NumElts = CI->getType()->getVectorNumElements();
1455       unsigned HalfSize = NumElts / 2;
1456       SmallVector<uint32_t, 8> ShuffleMask(NumElts);
1457 
1458       // Determine which operand(s) are actually in use for this instruction.
1459       Value *V0 = (Imm & 0x02) ? CI->getArgOperand(1) : CI->getArgOperand(0);
1460       Value *V1 = (Imm & 0x20) ? CI->getArgOperand(1) : CI->getArgOperand(0);
1461 
1462       // If needed, replace operands based on zero mask.
1463       V0 = (Imm & 0x08) ? ConstantAggregateZero::get(CI->getType()) : V0;
1464       V1 = (Imm & 0x80) ? ConstantAggregateZero::get(CI->getType()) : V1;
1465 
1466       // Permute low half of result.
1467       unsigned StartIndex = (Imm & 0x01) ? HalfSize : 0;
1468       for (unsigned i = 0; i < HalfSize; ++i)
1469         ShuffleMask[i] = StartIndex + i;
1470 
1471       // Permute high half of result.
1472       StartIndex = (Imm & 0x10) ? HalfSize : 0;
1473       for (unsigned i = 0; i < HalfSize; ++i)
1474         ShuffleMask[i + HalfSize] = NumElts + StartIndex + i;
1475 
1476       Rep = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
1477 
1478     } else if (IsX86 && (Name.startswith("avx.vpermil.") ||
1479                          Name == "sse2.pshuf.d" ||
1480                          Name.startswith("avx512.mask.vpermil.p") ||
1481                          Name.startswith("avx512.mask.pshuf.d."))) {
1482       Value *Op0 = CI->getArgOperand(0);
1483       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1484       VectorType *VecTy = cast<VectorType>(CI->getType());
1485       unsigned NumElts = VecTy->getNumElements();
1486       // Calculate the size of each index in the immediate.
1487       unsigned IdxSize = 64 / VecTy->getScalarSizeInBits();
1488       unsigned IdxMask = ((1 << IdxSize) - 1);
1489 
1490       SmallVector<uint32_t, 8> Idxs(NumElts);
1491       // Lookup the bits for this element, wrapping around the immediate every
1492       // 8-bits. Elements are grouped into sets of 2 or 4 elements so we need
1493       // to offset by the first index of each group.
1494       for (unsigned i = 0; i != NumElts; ++i)
1495         Idxs[i] = ((Imm >> ((i * IdxSize) % 8)) & IdxMask) | (i & ~IdxMask);
1496 
1497       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1498 
1499       if (CI->getNumArgOperands() == 4)
1500         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1501                             CI->getArgOperand(2));
1502     } else if (IsX86 && (Name == "sse2.pshufl.w" ||
1503                          Name.startswith("avx512.mask.pshufl.w."))) {
1504       Value *Op0 = CI->getArgOperand(0);
1505       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1506       unsigned NumElts = CI->getType()->getVectorNumElements();
1507 
1508       SmallVector<uint32_t, 16> Idxs(NumElts);
1509       for (unsigned l = 0; l != NumElts; l += 8) {
1510         for (unsigned i = 0; i != 4; ++i)
1511           Idxs[i + l] = ((Imm >> (2 * i)) & 0x3) + l;
1512         for (unsigned i = 4; i != 8; ++i)
1513           Idxs[i + l] = i + l;
1514       }
1515 
1516       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1517 
1518       if (CI->getNumArgOperands() == 4)
1519         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1520                             CI->getArgOperand(2));
1521     } else if (IsX86 && (Name == "sse2.pshufh.w" ||
1522                          Name.startswith("avx512.mask.pshufh.w."))) {
1523       Value *Op0 = CI->getArgOperand(0);
1524       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1525       unsigned NumElts = CI->getType()->getVectorNumElements();
1526 
1527       SmallVector<uint32_t, 16> Idxs(NumElts);
1528       for (unsigned l = 0; l != NumElts; l += 8) {
1529         for (unsigned i = 0; i != 4; ++i)
1530           Idxs[i + l] = i + l;
1531         for (unsigned i = 0; i != 4; ++i)
1532           Idxs[i + l + 4] = ((Imm >> (2 * i)) & 0x3) + 4 + l;
1533       }
1534 
1535       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1536 
1537       if (CI->getNumArgOperands() == 4)
1538         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1539                             CI->getArgOperand(2));
1540     } else if (IsX86 && Name.startswith("avx512.mask.shuf.p")) {
1541       Value *Op0 = CI->getArgOperand(0);
1542       Value *Op1 = CI->getArgOperand(1);
1543       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1544       unsigned NumElts = CI->getType()->getVectorNumElements();
1545 
1546       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1547       unsigned HalfLaneElts = NumLaneElts / 2;
1548 
1549       SmallVector<uint32_t, 16> Idxs(NumElts);
1550       for (unsigned i = 0; i != NumElts; ++i) {
1551         // Base index is the starting element of the lane.
1552         Idxs[i] = i - (i % NumLaneElts);
1553         // If we are half way through the lane switch to the other source.
1554         if ((i % NumLaneElts) >= HalfLaneElts)
1555           Idxs[i] += NumElts;
1556         // Now select the specific element. By adding HalfLaneElts bits from
1557         // the immediate. Wrapping around the immediate every 8-bits.
1558         Idxs[i] += (Imm >> ((i * HalfLaneElts) % 8)) & ((1 << HalfLaneElts) - 1);
1559       }
1560 
1561       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1562 
1563       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1564                           CI->getArgOperand(3));
1565     } else if (IsX86 && (Name.startswith("avx512.mask.movddup") ||
1566                          Name.startswith("avx512.mask.movshdup") ||
1567                          Name.startswith("avx512.mask.movsldup"))) {
1568       Value *Op0 = CI->getArgOperand(0);
1569       unsigned NumElts = CI->getType()->getVectorNumElements();
1570       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1571 
1572       unsigned Offset = 0;
1573       if (Name.startswith("avx512.mask.movshdup."))
1574         Offset = 1;
1575 
1576       SmallVector<uint32_t, 16> Idxs(NumElts);
1577       for (unsigned l = 0; l != NumElts; l += NumLaneElts)
1578         for (unsigned i = 0; i != NumLaneElts; i += 2) {
1579           Idxs[i + l + 0] = i + l + Offset;
1580           Idxs[i + l + 1] = i + l + Offset;
1581         }
1582 
1583       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1584 
1585       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1586                           CI->getArgOperand(1));
1587     } else if (IsX86 && (Name.startswith("avx512.mask.punpckl") ||
1588                          Name.startswith("avx512.mask.unpckl."))) {
1589       Value *Op0 = CI->getArgOperand(0);
1590       Value *Op1 = CI->getArgOperand(1);
1591       int NumElts = CI->getType()->getVectorNumElements();
1592       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1593 
1594       SmallVector<uint32_t, 64> Idxs(NumElts);
1595       for (int l = 0; l != NumElts; l += NumLaneElts)
1596         for (int i = 0; i != NumLaneElts; ++i)
1597           Idxs[i + l] = l + (i / 2) + NumElts * (i % 2);
1598 
1599       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1600 
1601       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1602                           CI->getArgOperand(2));
1603     } else if (IsX86 && (Name.startswith("avx512.mask.punpckh") ||
1604                          Name.startswith("avx512.mask.unpckh."))) {
1605       Value *Op0 = CI->getArgOperand(0);
1606       Value *Op1 = CI->getArgOperand(1);
1607       int NumElts = CI->getType()->getVectorNumElements();
1608       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1609 
1610       SmallVector<uint32_t, 64> Idxs(NumElts);
1611       for (int l = 0; l != NumElts; l += NumLaneElts)
1612         for (int i = 0; i != NumLaneElts; ++i)
1613           Idxs[i + l] = (NumLaneElts / 2) + l + (i / 2) + NumElts * (i % 2);
1614 
1615       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1616 
1617       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1618                           CI->getArgOperand(2));
1619     } else if (IsX86 && Name.startswith("avx512.mask.pand.")) {
1620       Rep = Builder.CreateAnd(CI->getArgOperand(0), CI->getArgOperand(1));
1621       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1622                           CI->getArgOperand(2));
1623     } else if (IsX86 && Name.startswith("avx512.mask.pandn.")) {
1624       Rep = Builder.CreateAnd(Builder.CreateNot(CI->getArgOperand(0)),
1625                               CI->getArgOperand(1));
1626       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1627                           CI->getArgOperand(2));
1628     } else if (IsX86 && Name.startswith("avx512.mask.por.")) {
1629       Rep = Builder.CreateOr(CI->getArgOperand(0), CI->getArgOperand(1));
1630       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1631                           CI->getArgOperand(2));
1632     } else if (IsX86 && Name.startswith("avx512.mask.pxor.")) {
1633       Rep = Builder.CreateXor(CI->getArgOperand(0), CI->getArgOperand(1));
1634       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1635                           CI->getArgOperand(2));
1636     } else if (IsX86 && Name.startswith("avx512.mask.and.")) {
1637       VectorType *FTy = cast<VectorType>(CI->getType());
1638       VectorType *ITy = VectorType::getInteger(FTy);
1639       Rep = Builder.CreateAnd(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1640                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1641       Rep = Builder.CreateBitCast(Rep, FTy);
1642       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1643                           CI->getArgOperand(2));
1644     } else if (IsX86 && Name.startswith("avx512.mask.andn.")) {
1645       VectorType *FTy = cast<VectorType>(CI->getType());
1646       VectorType *ITy = VectorType::getInteger(FTy);
1647       Rep = Builder.CreateNot(Builder.CreateBitCast(CI->getArgOperand(0), ITy));
1648       Rep = Builder.CreateAnd(Rep,
1649                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1650       Rep = Builder.CreateBitCast(Rep, FTy);
1651       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1652                           CI->getArgOperand(2));
1653     } else if (IsX86 && Name.startswith("avx512.mask.or.")) {
1654       VectorType *FTy = cast<VectorType>(CI->getType());
1655       VectorType *ITy = VectorType::getInteger(FTy);
1656       Rep = Builder.CreateOr(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1657                              Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1658       Rep = Builder.CreateBitCast(Rep, FTy);
1659       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1660                           CI->getArgOperand(2));
1661     } else if (IsX86 && Name.startswith("avx512.mask.xor.")) {
1662       VectorType *FTy = cast<VectorType>(CI->getType());
1663       VectorType *ITy = VectorType::getInteger(FTy);
1664       Rep = Builder.CreateXor(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1665                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1666       Rep = Builder.CreateBitCast(Rep, FTy);
1667       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1668                           CI->getArgOperand(2));
1669     } else if (IsX86 && Name.startswith("avx512.mask.padd.")) {
1670       Rep = Builder.CreateAdd(CI->getArgOperand(0), CI->getArgOperand(1));
1671       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1672                           CI->getArgOperand(2));
1673     } else if (IsX86 && Name.startswith("avx512.mask.psub.")) {
1674       Rep = Builder.CreateSub(CI->getArgOperand(0), CI->getArgOperand(1));
1675       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1676                           CI->getArgOperand(2));
1677     } else if (IsX86 && Name.startswith("avx512.mask.pmull.")) {
1678       Rep = Builder.CreateMul(CI->getArgOperand(0), CI->getArgOperand(1));
1679       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1680                           CI->getArgOperand(2));
1681     } else if (IsX86 && (Name.startswith("avx512.mask.add.p"))) {
1682       Rep = Builder.CreateFAdd(CI->getArgOperand(0), CI->getArgOperand(1));
1683       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1684                           CI->getArgOperand(2));
1685     } else if (IsX86 && Name.startswith("avx512.mask.div.p")) {
1686       Rep = Builder.CreateFDiv(CI->getArgOperand(0), CI->getArgOperand(1));
1687       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1688                           CI->getArgOperand(2));
1689     } else if (IsX86 && Name.startswith("avx512.mask.mul.p")) {
1690       Rep = Builder.CreateFMul(CI->getArgOperand(0), CI->getArgOperand(1));
1691       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1692                           CI->getArgOperand(2));
1693     } else if (IsX86 && Name.startswith("avx512.mask.sub.p")) {
1694       Rep = Builder.CreateFSub(CI->getArgOperand(0), CI->getArgOperand(1));
1695       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1696                           CI->getArgOperand(2));
1697     } else if (IsX86 && Name.startswith("avx512.mask.lzcnt.")) {
1698       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
1699                                                          Intrinsic::ctlz,
1700                                                          CI->getType()),
1701                                { CI->getArgOperand(0), Builder.getInt1(false) });
1702       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1703                           CI->getArgOperand(1));
1704     } else if (IsX86 && (Name.startswith("avx512.mask.max.p") ||
1705                          Name.startswith("avx512.mask.min.p"))) {
1706       bool IsMin = Name[13] == 'i';
1707       VectorType *VecTy = cast<VectorType>(CI->getType());
1708       unsigned VecWidth = VecTy->getPrimitiveSizeInBits();
1709       unsigned EltWidth = VecTy->getScalarSizeInBits();
1710       Intrinsic::ID IID;
1711       if (!IsMin && VecWidth == 128 && EltWidth == 32)
1712         IID = Intrinsic::x86_sse_max_ps;
1713       else if (!IsMin && VecWidth == 128 && EltWidth == 64)
1714         IID = Intrinsic::x86_sse2_max_pd;
1715       else if (!IsMin && VecWidth == 256 && EltWidth == 32)
1716         IID = Intrinsic::x86_avx_max_ps_256;
1717       else if (!IsMin && VecWidth == 256 && EltWidth == 64)
1718         IID = Intrinsic::x86_avx_max_pd_256;
1719       else if (IsMin && VecWidth == 128 && EltWidth == 32)
1720         IID = Intrinsic::x86_sse_min_ps;
1721       else if (IsMin && VecWidth == 128 && EltWidth == 64)
1722         IID = Intrinsic::x86_sse2_min_pd;
1723       else if (IsMin && VecWidth == 256 && EltWidth == 32)
1724         IID = Intrinsic::x86_avx_min_ps_256;
1725       else if (IsMin && VecWidth == 256 && EltWidth == 64)
1726         IID = Intrinsic::x86_avx_min_pd_256;
1727       else
1728         llvm_unreachable("Unexpected intrinsic");
1729 
1730       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1731                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1732       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1733                           CI->getArgOperand(2));
1734     } else if (IsX86 && Name.startswith("avx512.mask.pshuf.b.")) {
1735       VectorType *VecTy = cast<VectorType>(CI->getType());
1736       Intrinsic::ID IID;
1737       if (VecTy->getPrimitiveSizeInBits() == 128)
1738         IID = Intrinsic::x86_ssse3_pshuf_b_128;
1739       else if (VecTy->getPrimitiveSizeInBits() == 256)
1740         IID = Intrinsic::x86_avx2_pshuf_b;
1741       else if (VecTy->getPrimitiveSizeInBits() == 512)
1742         IID = Intrinsic::x86_avx512_pshuf_b_512;
1743       else
1744         llvm_unreachable("Unexpected intrinsic");
1745 
1746       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1747                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1748       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1749                           CI->getArgOperand(2));
1750     } else if (IsX86 && (Name.startswith("avx512.mask.pmul.dq.") ||
1751                          Name.startswith("avx512.mask.pmulu.dq."))) {
1752       bool IsUnsigned = Name[16] == 'u';
1753       VectorType *VecTy = cast<VectorType>(CI->getType());
1754       Intrinsic::ID IID;
1755       if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1756         IID = Intrinsic::x86_sse41_pmuldq;
1757       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1758         IID = Intrinsic::x86_avx2_pmul_dq;
1759       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1760         IID = Intrinsic::x86_avx512_pmul_dq_512;
1761       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1762         IID = Intrinsic::x86_sse2_pmulu_dq;
1763       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1764         IID = Intrinsic::x86_avx2_pmulu_dq;
1765       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1766         IID = Intrinsic::x86_avx512_pmulu_dq_512;
1767       else
1768         llvm_unreachable("Unexpected intrinsic");
1769 
1770       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1771                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1772       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1773                           CI->getArgOperand(2));
1774     } else if (IsX86 && Name.startswith("avx512.mask.pack")) {
1775       bool IsUnsigned = Name[16] == 'u';
1776       bool IsDW = Name[18] == 'd';
1777       VectorType *VecTy = cast<VectorType>(CI->getType());
1778       Intrinsic::ID IID;
1779       if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1780         IID = Intrinsic::x86_sse2_packsswb_128;
1781       else if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1782         IID = Intrinsic::x86_avx2_packsswb;
1783       else if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1784         IID = Intrinsic::x86_avx512_packsswb_512;
1785       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1786         IID = Intrinsic::x86_sse2_packssdw_128;
1787       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1788         IID = Intrinsic::x86_avx2_packssdw;
1789       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1790         IID = Intrinsic::x86_avx512_packssdw_512;
1791       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1792         IID = Intrinsic::x86_sse2_packuswb_128;
1793       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1794         IID = Intrinsic::x86_avx2_packuswb;
1795       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1796         IID = Intrinsic::x86_avx512_packuswb_512;
1797       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1798         IID = Intrinsic::x86_sse41_packusdw;
1799       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1800         IID = Intrinsic::x86_avx2_packusdw;
1801       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1802         IID = Intrinsic::x86_avx512_packusdw_512;
1803       else
1804         llvm_unreachable("Unexpected intrinsic");
1805 
1806       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1807                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1808       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1809                           CI->getArgOperand(2));
1810     } else if (IsX86 && Name.startswith("avx512.mask.psll")) {
1811       bool IsImmediate = Name[16] == 'i' ||
1812                          (Name.size() > 18 && Name[18] == 'i');
1813       bool IsVariable = Name[16] == 'v';
1814       char Size = Name[16] == '.' ? Name[17] :
1815                   Name[17] == '.' ? Name[18] :
1816                   Name[18] == '.' ? Name[19] :
1817                                     Name[20];
1818 
1819       Intrinsic::ID IID;
1820       if (IsVariable && Name[17] != '.') {
1821         if (Size == 'd' && Name[17] == '2') // avx512.mask.psllv2.di
1822           IID = Intrinsic::x86_avx2_psllv_q;
1823         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psllv4.di
1824           IID = Intrinsic::x86_avx2_psllv_q_256;
1825         else if (Size == 's' && Name[17] == '4') // avx512.mask.psllv4.si
1826           IID = Intrinsic::x86_avx2_psllv_d;
1827         else if (Size == 's' && Name[17] == '8') // avx512.mask.psllv8.si
1828           IID = Intrinsic::x86_avx2_psllv_d_256;
1829         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psllv8.hi
1830           IID = Intrinsic::x86_avx512_psllv_w_128;
1831         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psllv16.hi
1832           IID = Intrinsic::x86_avx512_psllv_w_256;
1833         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psllv32hi
1834           IID = Intrinsic::x86_avx512_psllv_w_512;
1835         else
1836           llvm_unreachable("Unexpected size");
1837       } else if (Name.endswith(".128")) {
1838         if (Size == 'd') // avx512.mask.psll.d.128, avx512.mask.psll.di.128
1839           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_d
1840                             : Intrinsic::x86_sse2_psll_d;
1841         else if (Size == 'q') // avx512.mask.psll.q.128, avx512.mask.psll.qi.128
1842           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_q
1843                             : Intrinsic::x86_sse2_psll_q;
1844         else if (Size == 'w') // avx512.mask.psll.w.128, avx512.mask.psll.wi.128
1845           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_w
1846                             : Intrinsic::x86_sse2_psll_w;
1847         else
1848           llvm_unreachable("Unexpected size");
1849       } else if (Name.endswith(".256")) {
1850         if (Size == 'd') // avx512.mask.psll.d.256, avx512.mask.psll.di.256
1851           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_d
1852                             : Intrinsic::x86_avx2_psll_d;
1853         else if (Size == 'q') // avx512.mask.psll.q.256, avx512.mask.psll.qi.256
1854           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_q
1855                             : Intrinsic::x86_avx2_psll_q;
1856         else if (Size == 'w') // avx512.mask.psll.w.256, avx512.mask.psll.wi.256
1857           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_w
1858                             : Intrinsic::x86_avx2_psll_w;
1859         else
1860           llvm_unreachable("Unexpected size");
1861       } else {
1862         if (Size == 'd') // psll.di.512, pslli.d, psll.d, psllv.d.512
1863           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_d_512 :
1864                 IsVariable  ? Intrinsic::x86_avx512_psllv_d_512 :
1865                               Intrinsic::x86_avx512_psll_d_512;
1866         else if (Size == 'q') // psll.qi.512, pslli.q, psll.q, psllv.q.512
1867           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_q_512 :
1868                 IsVariable  ? Intrinsic::x86_avx512_psllv_q_512 :
1869                               Intrinsic::x86_avx512_psll_q_512;
1870         else if (Size == 'w') // psll.wi.512, pslli.w, psll.w
1871           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_w_512
1872                             : Intrinsic::x86_avx512_psll_w_512;
1873         else
1874           llvm_unreachable("Unexpected size");
1875       }
1876 
1877       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
1878     } else if (IsX86 && Name.startswith("avx512.mask.psrl")) {
1879       bool IsImmediate = Name[16] == 'i' ||
1880                          (Name.size() > 18 && Name[18] == 'i');
1881       bool IsVariable = Name[16] == 'v';
1882       char Size = Name[16] == '.' ? Name[17] :
1883                   Name[17] == '.' ? Name[18] :
1884                   Name[18] == '.' ? Name[19] :
1885                                     Name[20];
1886 
1887       Intrinsic::ID IID;
1888       if (IsVariable && Name[17] != '.') {
1889         if (Size == 'd' && Name[17] == '2') // avx512.mask.psrlv2.di
1890           IID = Intrinsic::x86_avx2_psrlv_q;
1891         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psrlv4.di
1892           IID = Intrinsic::x86_avx2_psrlv_q_256;
1893         else if (Size == 's' && Name[17] == '4') // avx512.mask.psrlv4.si
1894           IID = Intrinsic::x86_avx2_psrlv_d;
1895         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrlv8.si
1896           IID = Intrinsic::x86_avx2_psrlv_d_256;
1897         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrlv8.hi
1898           IID = Intrinsic::x86_avx512_psrlv_w_128;
1899         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrlv16.hi
1900           IID = Intrinsic::x86_avx512_psrlv_w_256;
1901         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrlv32hi
1902           IID = Intrinsic::x86_avx512_psrlv_w_512;
1903         else
1904           llvm_unreachable("Unexpected size");
1905       } else if (Name.endswith(".128")) {
1906         if (Size == 'd') // avx512.mask.psrl.d.128, avx512.mask.psrl.di.128
1907           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_d
1908                             : Intrinsic::x86_sse2_psrl_d;
1909         else if (Size == 'q') // avx512.mask.psrl.q.128, avx512.mask.psrl.qi.128
1910           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_q
1911                             : Intrinsic::x86_sse2_psrl_q;
1912         else if (Size == 'w') // avx512.mask.psrl.w.128, avx512.mask.psrl.wi.128
1913           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_w
1914                             : Intrinsic::x86_sse2_psrl_w;
1915         else
1916           llvm_unreachable("Unexpected size");
1917       } else if (Name.endswith(".256")) {
1918         if (Size == 'd') // avx512.mask.psrl.d.256, avx512.mask.psrl.di.256
1919           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_d
1920                             : Intrinsic::x86_avx2_psrl_d;
1921         else if (Size == 'q') // avx512.mask.psrl.q.256, avx512.mask.psrl.qi.256
1922           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_q
1923                             : Intrinsic::x86_avx2_psrl_q;
1924         else if (Size == 'w') // avx512.mask.psrl.w.256, avx512.mask.psrl.wi.256
1925           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_w
1926                             : Intrinsic::x86_avx2_psrl_w;
1927         else
1928           llvm_unreachable("Unexpected size");
1929       } else {
1930         if (Size == 'd') // psrl.di.512, psrli.d, psrl.d, psrl.d.512
1931           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_d_512 :
1932                 IsVariable  ? Intrinsic::x86_avx512_psrlv_d_512 :
1933                               Intrinsic::x86_avx512_psrl_d_512;
1934         else if (Size == 'q') // psrl.qi.512, psrli.q, psrl.q, psrl.q.512
1935           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_q_512 :
1936                 IsVariable  ? Intrinsic::x86_avx512_psrlv_q_512 :
1937                               Intrinsic::x86_avx512_psrl_q_512;
1938         else if (Size == 'w') // psrl.wi.512, psrli.w, psrl.w)
1939           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_w_512
1940                             : Intrinsic::x86_avx512_psrl_w_512;
1941         else
1942           llvm_unreachable("Unexpected size");
1943       }
1944 
1945       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
1946     } else if (IsX86 && Name.startswith("avx512.mask.psra")) {
1947       bool IsImmediate = Name[16] == 'i' ||
1948                          (Name.size() > 18 && Name[18] == 'i');
1949       bool IsVariable = Name[16] == 'v';
1950       char Size = Name[16] == '.' ? Name[17] :
1951                   Name[17] == '.' ? Name[18] :
1952                   Name[18] == '.' ? Name[19] :
1953                                     Name[20];
1954 
1955       Intrinsic::ID IID;
1956       if (IsVariable && Name[17] != '.') {
1957         if (Size == 's' && Name[17] == '4') // avx512.mask.psrav4.si
1958           IID = Intrinsic::x86_avx2_psrav_d;
1959         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrav8.si
1960           IID = Intrinsic::x86_avx2_psrav_d_256;
1961         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrav8.hi
1962           IID = Intrinsic::x86_avx512_psrav_w_128;
1963         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrav16.hi
1964           IID = Intrinsic::x86_avx512_psrav_w_256;
1965         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrav32hi
1966           IID = Intrinsic::x86_avx512_psrav_w_512;
1967         else
1968           llvm_unreachable("Unexpected size");
1969       } else if (Name.endswith(".128")) {
1970         if (Size == 'd') // avx512.mask.psra.d.128, avx512.mask.psra.di.128
1971           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_d
1972                             : Intrinsic::x86_sse2_psra_d;
1973         else if (Size == 'q') // avx512.mask.psra.q.128, avx512.mask.psra.qi.128
1974           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_128 :
1975                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_128 :
1976                               Intrinsic::x86_avx512_psra_q_128;
1977         else if (Size == 'w') // avx512.mask.psra.w.128, avx512.mask.psra.wi.128
1978           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_w
1979                             : Intrinsic::x86_sse2_psra_w;
1980         else
1981           llvm_unreachable("Unexpected size");
1982       } else if (Name.endswith(".256")) {
1983         if (Size == 'd') // avx512.mask.psra.d.256, avx512.mask.psra.di.256
1984           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_d
1985                             : Intrinsic::x86_avx2_psra_d;
1986         else if (Size == 'q') // avx512.mask.psra.q.256, avx512.mask.psra.qi.256
1987           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_256 :
1988                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_256 :
1989                               Intrinsic::x86_avx512_psra_q_256;
1990         else if (Size == 'w') // avx512.mask.psra.w.256, avx512.mask.psra.wi.256
1991           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_w
1992                             : Intrinsic::x86_avx2_psra_w;
1993         else
1994           llvm_unreachable("Unexpected size");
1995       } else {
1996         if (Size == 'd') // psra.di.512, psrai.d, psra.d, psrav.d.512
1997           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_d_512 :
1998                 IsVariable  ? Intrinsic::x86_avx512_psrav_d_512 :
1999                               Intrinsic::x86_avx512_psra_d_512;
2000         else if (Size == 'q') // psra.qi.512, psrai.q, psra.q
2001           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_512 :
2002                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_512 :
2003                               Intrinsic::x86_avx512_psra_q_512;
2004         else if (Size == 'w') // psra.wi.512, psrai.w, psra.w
2005           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_w_512
2006                             : Intrinsic::x86_avx512_psra_w_512;
2007         else
2008           llvm_unreachable("Unexpected size");
2009       }
2010 
2011       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2012     } else if (IsX86 && Name.startswith("avx512.mask.move.s")) {
2013       Rep = upgradeMaskedMove(Builder, *CI);
2014     } else if (IsX86 && Name.startswith("avx512.cvtmask2")) {
2015       Rep = UpgradeMaskToInt(Builder, *CI);
2016     } else if (IsX86 && Name.startswith("avx512.mask.vpermilvar.")) {
2017       Intrinsic::ID IID;
2018       if (Name.endswith("ps.128"))
2019         IID = Intrinsic::x86_avx_vpermilvar_ps;
2020       else if (Name.endswith("pd.128"))
2021         IID = Intrinsic::x86_avx_vpermilvar_pd;
2022       else if (Name.endswith("ps.256"))
2023         IID = Intrinsic::x86_avx_vpermilvar_ps_256;
2024       else if (Name.endswith("pd.256"))
2025         IID = Intrinsic::x86_avx_vpermilvar_pd_256;
2026       else if (Name.endswith("ps.512"))
2027         IID = Intrinsic::x86_avx512_vpermilvar_ps_512;
2028       else if (Name.endswith("pd.512"))
2029         IID = Intrinsic::x86_avx512_vpermilvar_pd_512;
2030       else
2031         llvm_unreachable("Unexpected vpermilvar intrinsic");
2032 
2033       Function *Intrin = Intrinsic::getDeclaration(F->getParent(), IID);
2034       Rep = Builder.CreateCall(Intrin,
2035                                { CI->getArgOperand(0), CI->getArgOperand(1) });
2036       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2037                           CI->getArgOperand(2));
2038     } else if (IsX86 && Name.endswith(".movntdqa")) {
2039       Module *M = F->getParent();
2040       MDNode *Node = MDNode::get(
2041           C, ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
2042 
2043       Value *Ptr = CI->getArgOperand(0);
2044       VectorType *VTy = cast<VectorType>(CI->getType());
2045 
2046       // Convert the type of the pointer to a pointer to the stored type.
2047       Value *BC =
2048           Builder.CreateBitCast(Ptr, PointerType::getUnqual(VTy), "cast");
2049       LoadInst *LI = Builder.CreateAlignedLoad(BC, VTy->getBitWidth() / 8);
2050       LI->setMetadata(M->getMDKindID("nontemporal"), Node);
2051       Rep = LI;
2052     } else if (IsX86 &&
2053                (Name.startswith("sse2.pavg") || Name.startswith("avx2.pavg") ||
2054                 Name.startswith("avx512.mask.pavg"))) {
2055       // llvm.x86.sse2.pavg.b/w, llvm.x86.avx2.pavg.b/w,
2056       // llvm.x86.avx512.mask.pavg.b/w
2057       Value *A = CI->getArgOperand(0);
2058       Value *B = CI->getArgOperand(1);
2059       VectorType *ZextType = VectorType::getExtendedElementVectorType(
2060           cast<VectorType>(A->getType()));
2061       Value *ExtendedA = Builder.CreateZExt(A, ZextType);
2062       Value *ExtendedB = Builder.CreateZExt(B, ZextType);
2063       Value *Sum = Builder.CreateAdd(ExtendedA, ExtendedB);
2064       Value *AddOne = Builder.CreateAdd(Sum, ConstantInt::get(ZextType, 1));
2065       Value *ShiftR = Builder.CreateLShr(AddOne, ConstantInt::get(ZextType, 1));
2066       Rep = Builder.CreateTrunc(ShiftR, A->getType());
2067       if (CI->getNumArgOperands() > 2) {
2068         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2069                             CI->getArgOperand(2));
2070       }
2071     } else if (IsNVVM && (Name == "abs.i" || Name == "abs.ll")) {
2072       Value *Arg = CI->getArgOperand(0);
2073       Value *Neg = Builder.CreateNeg(Arg, "neg");
2074       Value *Cmp = Builder.CreateICmpSGE(
2075           Arg, llvm::Constant::getNullValue(Arg->getType()), "abs.cond");
2076       Rep = Builder.CreateSelect(Cmp, Arg, Neg, "abs");
2077     } else if (IsNVVM && (Name == "max.i" || Name == "max.ll" ||
2078                           Name == "max.ui" || Name == "max.ull")) {
2079       Value *Arg0 = CI->getArgOperand(0);
2080       Value *Arg1 = CI->getArgOperand(1);
2081       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
2082                        ? Builder.CreateICmpUGE(Arg0, Arg1, "max.cond")
2083                        : Builder.CreateICmpSGE(Arg0, Arg1, "max.cond");
2084       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "max");
2085     } else if (IsNVVM && (Name == "min.i" || Name == "min.ll" ||
2086                           Name == "min.ui" || Name == "min.ull")) {
2087       Value *Arg0 = CI->getArgOperand(0);
2088       Value *Arg1 = CI->getArgOperand(1);
2089       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
2090                        ? Builder.CreateICmpULE(Arg0, Arg1, "min.cond")
2091                        : Builder.CreateICmpSLE(Arg0, Arg1, "min.cond");
2092       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "min");
2093     } else if (IsNVVM && Name == "clz.ll") {
2094       // llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 and returns an i64.
2095       Value *Arg = CI->getArgOperand(0);
2096       Value *Ctlz = Builder.CreateCall(
2097           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
2098                                     {Arg->getType()}),
2099           {Arg, Builder.getFalse()}, "ctlz");
2100       Rep = Builder.CreateTrunc(Ctlz, Builder.getInt32Ty(), "ctlz.trunc");
2101     } else if (IsNVVM && Name == "popc.ll") {
2102       // llvm.nvvm.popc.ll returns an i32, but llvm.ctpop.i64 and returns an
2103       // i64.
2104       Value *Arg = CI->getArgOperand(0);
2105       Value *Popc = Builder.CreateCall(
2106           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
2107                                     {Arg->getType()}),
2108           Arg, "ctpop");
2109       Rep = Builder.CreateTrunc(Popc, Builder.getInt32Ty(), "ctpop.trunc");
2110     } else if (IsNVVM && Name == "h2f") {
2111       Rep = Builder.CreateCall(Intrinsic::getDeclaration(
2112                                    F->getParent(), Intrinsic::convert_from_fp16,
2113                                    {Builder.getFloatTy()}),
2114                                CI->getArgOperand(0), "h2f");
2115     } else {
2116       llvm_unreachable("Unknown function for CallInst upgrade.");
2117     }
2118 
2119     if (Rep)
2120       CI->replaceAllUsesWith(Rep);
2121     CI->eraseFromParent();
2122     return;
2123   }
2124 
2125   CallInst *NewCall = nullptr;
2126   switch (NewFn->getIntrinsicID()) {
2127   default: {
2128     // Handle generic mangling change, but nothing else
2129     assert(
2130         (CI->getCalledFunction()->getName() != NewFn->getName()) &&
2131         "Unknown function for CallInst upgrade and isn't just a name change");
2132     CI->setCalledFunction(NewFn);
2133     return;
2134   }
2135 
2136   case Intrinsic::arm_neon_vld1:
2137   case Intrinsic::arm_neon_vld2:
2138   case Intrinsic::arm_neon_vld3:
2139   case Intrinsic::arm_neon_vld4:
2140   case Intrinsic::arm_neon_vld2lane:
2141   case Intrinsic::arm_neon_vld3lane:
2142   case Intrinsic::arm_neon_vld4lane:
2143   case Intrinsic::arm_neon_vst1:
2144   case Intrinsic::arm_neon_vst2:
2145   case Intrinsic::arm_neon_vst3:
2146   case Intrinsic::arm_neon_vst4:
2147   case Intrinsic::arm_neon_vst2lane:
2148   case Intrinsic::arm_neon_vst3lane:
2149   case Intrinsic::arm_neon_vst4lane: {
2150     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2151                                  CI->arg_operands().end());
2152     NewCall = Builder.CreateCall(NewFn, Args);
2153     break;
2154   }
2155 
2156   case Intrinsic::bitreverse:
2157     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2158     break;
2159 
2160   case Intrinsic::ctlz:
2161   case Intrinsic::cttz:
2162     assert(CI->getNumArgOperands() == 1 &&
2163            "Mismatch between function args and call args");
2164     NewCall =
2165         Builder.CreateCall(NewFn, {CI->getArgOperand(0), Builder.getFalse()});
2166     break;
2167 
2168   case Intrinsic::objectsize: {
2169     Value *NullIsUnknownSize = CI->getNumArgOperands() == 2
2170                                    ? Builder.getFalse()
2171                                    : CI->getArgOperand(2);
2172     NewCall = Builder.CreateCall(
2173         NewFn, {CI->getArgOperand(0), CI->getArgOperand(1), NullIsUnknownSize});
2174     break;
2175   }
2176 
2177   case Intrinsic::ctpop:
2178     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2179     break;
2180 
2181   case Intrinsic::convert_from_fp16:
2182     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2183     break;
2184 
2185   case Intrinsic::dbg_value:
2186     // Upgrade from the old version that had an extra offset argument.
2187     assert(CI->getNumArgOperands() == 4);
2188     // Drop nonzero offsets instead of attempting to upgrade them.
2189     if (auto *Offset = dyn_cast_or_null<Constant>(CI->getArgOperand(1)))
2190       if (Offset->isZeroValue()) {
2191         NewCall = Builder.CreateCall(
2192             NewFn,
2193             {CI->getArgOperand(0), CI->getArgOperand(2), CI->getArgOperand(3)});
2194         break;
2195       }
2196     CI->eraseFromParent();
2197     return;
2198 
2199   case Intrinsic::x86_xop_vfrcz_ss:
2200   case Intrinsic::x86_xop_vfrcz_sd:
2201     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(1)});
2202     break;
2203 
2204   case Intrinsic::x86_xop_vpermil2pd:
2205   case Intrinsic::x86_xop_vpermil2ps:
2206   case Intrinsic::x86_xop_vpermil2pd_256:
2207   case Intrinsic::x86_xop_vpermil2ps_256: {
2208     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2209                                  CI->arg_operands().end());
2210     VectorType *FltIdxTy = cast<VectorType>(Args[2]->getType());
2211     VectorType *IntIdxTy = VectorType::getInteger(FltIdxTy);
2212     Args[2] = Builder.CreateBitCast(Args[2], IntIdxTy);
2213     NewCall = Builder.CreateCall(NewFn, Args);
2214     break;
2215   }
2216 
2217   case Intrinsic::x86_sse41_ptestc:
2218   case Intrinsic::x86_sse41_ptestz:
2219   case Intrinsic::x86_sse41_ptestnzc: {
2220     // The arguments for these intrinsics used to be v4f32, and changed
2221     // to v2i64. This is purely a nop, since those are bitwise intrinsics.
2222     // So, the only thing required is a bitcast for both arguments.
2223     // First, check the arguments have the old type.
2224     Value *Arg0 = CI->getArgOperand(0);
2225     if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
2226       return;
2227 
2228     // Old intrinsic, add bitcasts
2229     Value *Arg1 = CI->getArgOperand(1);
2230 
2231     Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
2232 
2233     Value *BC0 = Builder.CreateBitCast(Arg0, NewVecTy, "cast");
2234     Value *BC1 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
2235 
2236     NewCall = Builder.CreateCall(NewFn, {BC0, BC1});
2237     break;
2238   }
2239 
2240   case Intrinsic::x86_sse41_insertps:
2241   case Intrinsic::x86_sse41_dppd:
2242   case Intrinsic::x86_sse41_dpps:
2243   case Intrinsic::x86_sse41_mpsadbw:
2244   case Intrinsic::x86_avx_dp_ps_256:
2245   case Intrinsic::x86_avx2_mpsadbw: {
2246     // Need to truncate the last argument from i32 to i8 -- this argument models
2247     // an inherently 8-bit immediate operand to these x86 instructions.
2248     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2249                                  CI->arg_operands().end());
2250 
2251     // Replace the last argument with a trunc.
2252     Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
2253     NewCall = Builder.CreateCall(NewFn, Args);
2254     break;
2255   }
2256 
2257   case Intrinsic::thread_pointer: {
2258     NewCall = Builder.CreateCall(NewFn, {});
2259     break;
2260   }
2261 
2262   case Intrinsic::invariant_start:
2263   case Intrinsic::invariant_end:
2264   case Intrinsic::masked_load:
2265   case Intrinsic::masked_store:
2266   case Intrinsic::masked_gather:
2267   case Intrinsic::masked_scatter: {
2268     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2269                                  CI->arg_operands().end());
2270     NewCall = Builder.CreateCall(NewFn, Args);
2271     break;
2272   }
2273   }
2274   assert(NewCall && "Should have either set this variable or returned through "
2275                     "the default case");
2276   std::string Name = CI->getName();
2277   if (!Name.empty()) {
2278     CI->setName(Name + ".old");
2279     NewCall->setName(Name);
2280   }
2281   CI->replaceAllUsesWith(NewCall);
2282   CI->eraseFromParent();
2283 }
2284 
2285 void llvm::UpgradeCallsToIntrinsic(Function *F) {
2286   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
2287 
2288   // Check if this function should be upgraded and get the replacement function
2289   // if there is one.
2290   Function *NewFn;
2291   if (UpgradeIntrinsicFunction(F, NewFn)) {
2292     // Replace all users of the old function with the new function or new
2293     // instructions. This is not a range loop because the call is deleted.
2294     for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; )
2295       if (CallInst *CI = dyn_cast<CallInst>(*UI++))
2296         UpgradeIntrinsicCall(CI, NewFn);
2297 
2298     // Remove old function, no longer used, from the module.
2299     F->eraseFromParent();
2300   }
2301 }
2302 
2303 MDNode *llvm::UpgradeTBAANode(MDNode &MD) {
2304   // Check if the tag uses struct-path aware TBAA format.
2305   if (isa<MDNode>(MD.getOperand(0)) && MD.getNumOperands() >= 3)
2306     return &MD;
2307 
2308   auto &Context = MD.getContext();
2309   if (MD.getNumOperands() == 3) {
2310     Metadata *Elts[] = {MD.getOperand(0), MD.getOperand(1)};
2311     MDNode *ScalarType = MDNode::get(Context, Elts);
2312     // Create a MDNode <ScalarType, ScalarType, offset 0, const>
2313     Metadata *Elts2[] = {ScalarType, ScalarType,
2314                          ConstantAsMetadata::get(
2315                              Constant::getNullValue(Type::getInt64Ty(Context))),
2316                          MD.getOperand(2)};
2317     return MDNode::get(Context, Elts2);
2318   }
2319   // Create a MDNode <MD, MD, offset 0>
2320   Metadata *Elts[] = {&MD, &MD, ConstantAsMetadata::get(Constant::getNullValue(
2321                                     Type::getInt64Ty(Context)))};
2322   return MDNode::get(Context, Elts);
2323 }
2324 
2325 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
2326                                       Instruction *&Temp) {
2327   if (Opc != Instruction::BitCast)
2328     return nullptr;
2329 
2330   Temp = nullptr;
2331   Type *SrcTy = V->getType();
2332   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
2333       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
2334     LLVMContext &Context = V->getContext();
2335 
2336     // We have no information about target data layout, so we assume that
2337     // the maximum pointer size is 64bit.
2338     Type *MidTy = Type::getInt64Ty(Context);
2339     Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
2340 
2341     return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
2342   }
2343 
2344   return nullptr;
2345 }
2346 
2347 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
2348   if (Opc != Instruction::BitCast)
2349     return nullptr;
2350 
2351   Type *SrcTy = C->getType();
2352   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
2353       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
2354     LLVMContext &Context = C->getContext();
2355 
2356     // We have no information about target data layout, so we assume that
2357     // the maximum pointer size is 64bit.
2358     Type *MidTy = Type::getInt64Ty(Context);
2359 
2360     return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
2361                                      DestTy);
2362   }
2363 
2364   return nullptr;
2365 }
2366 
2367 /// Check the debug info version number, if it is out-dated, drop the debug
2368 /// info. Return true if module is modified.
2369 bool llvm::UpgradeDebugInfo(Module &M) {
2370   unsigned Version = getDebugMetadataVersionFromModule(M);
2371   if (Version == DEBUG_METADATA_VERSION) {
2372     bool BrokenDebugInfo = false;
2373     if (verifyModule(M, &llvm::errs(), &BrokenDebugInfo))
2374       report_fatal_error("Broken module found, compilation aborted!");
2375     if (!BrokenDebugInfo)
2376       // Everything is ok.
2377       return false;
2378     else {
2379       // Diagnose malformed debug info.
2380       DiagnosticInfoIgnoringInvalidDebugMetadata Diag(M);
2381       M.getContext().diagnose(Diag);
2382     }
2383   }
2384   bool Modified = StripDebugInfo(M);
2385   if (Modified && Version != DEBUG_METADATA_VERSION) {
2386     // Diagnose a version mismatch.
2387     DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
2388     M.getContext().diagnose(DiagVersion);
2389   }
2390   return Modified;
2391 }
2392 
2393 bool llvm::UpgradeModuleFlags(Module &M) {
2394   NamedMDNode *ModFlags = M.getModuleFlagsMetadata();
2395   if (!ModFlags)
2396     return false;
2397 
2398   bool HasObjCFlag = false, HasClassProperties = false, Changed = false;
2399   for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) {
2400     MDNode *Op = ModFlags->getOperand(I);
2401     if (Op->getNumOperands() != 3)
2402       continue;
2403     MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(1));
2404     if (!ID)
2405       continue;
2406     if (ID->getString() == "Objective-C Image Info Version")
2407       HasObjCFlag = true;
2408     if (ID->getString() == "Objective-C Class Properties")
2409       HasClassProperties = true;
2410     // Upgrade PIC/PIE Module Flags. The module flag behavior for these two
2411     // field was Error and now they are Max.
2412     if (ID->getString() == "PIC Level" || ID->getString() == "PIE Level") {
2413       if (auto *Behavior =
2414               mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0))) {
2415         if (Behavior->getLimitedValue() == Module::Error) {
2416           Type *Int32Ty = Type::getInt32Ty(M.getContext());
2417           Metadata *Ops[3] = {
2418               ConstantAsMetadata::get(ConstantInt::get(Int32Ty, Module::Max)),
2419               MDString::get(M.getContext(), ID->getString()),
2420               Op->getOperand(2)};
2421           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
2422           Changed = true;
2423         }
2424       }
2425     }
2426     // Upgrade Objective-C Image Info Section. Removed the whitespce in the
2427     // section name so that llvm-lto will not complain about mismatching
2428     // module flags that is functionally the same.
2429     if (ID->getString() == "Objective-C Image Info Section") {
2430       if (auto *Value = dyn_cast_or_null<MDString>(Op->getOperand(2))) {
2431         SmallVector<StringRef, 4> ValueComp;
2432         Value->getString().split(ValueComp, " ");
2433         if (ValueComp.size() != 1) {
2434           std::string NewValue;
2435           for (auto &S : ValueComp)
2436             NewValue += S.str();
2437           Metadata *Ops[3] = {Op->getOperand(0), Op->getOperand(1),
2438                               MDString::get(M.getContext(), NewValue)};
2439           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
2440           Changed = true;
2441         }
2442       }
2443     }
2444   }
2445 
2446   // "Objective-C Class Properties" is recently added for Objective-C. We
2447   // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module
2448   // flag of value 0, so we can correclty downgrade this flag when trying to
2449   // link an ObjC bitcode without this module flag with an ObjC bitcode with
2450   // this module flag.
2451   if (HasObjCFlag && !HasClassProperties) {
2452     M.addModuleFlag(llvm::Module::Override, "Objective-C Class Properties",
2453                     (uint32_t)0);
2454     Changed = true;
2455   }
2456 
2457   return Changed;
2458 }
2459 
2460 void llvm::UpgradeSectionAttributes(Module &M) {
2461   auto TrimSpaces = [](StringRef Section) -> std::string {
2462     SmallVector<StringRef, 5> Components;
2463     Section.split(Components, ',');
2464 
2465     SmallString<32> Buffer;
2466     raw_svector_ostream OS(Buffer);
2467 
2468     for (auto Component : Components)
2469       OS << ',' << Component.trim();
2470 
2471     return OS.str().substr(1);
2472   };
2473 
2474   for (auto &GV : M.globals()) {
2475     if (!GV.hasSection())
2476       continue;
2477 
2478     StringRef Section = GV.getSection();
2479 
2480     if (!Section.startswith("__DATA, __objc_catlist"))
2481       continue;
2482 
2483     // __DATA, __objc_catlist, regular, no_dead_strip
2484     // __DATA,__objc_catlist,regular,no_dead_strip
2485     GV.setSection(TrimSpaces(Section));
2486   }
2487 }
2488 
2489 static bool isOldLoopArgument(Metadata *MD) {
2490   auto *T = dyn_cast_or_null<MDTuple>(MD);
2491   if (!T)
2492     return false;
2493   if (T->getNumOperands() < 1)
2494     return false;
2495   auto *S = dyn_cast_or_null<MDString>(T->getOperand(0));
2496   if (!S)
2497     return false;
2498   return S->getString().startswith("llvm.vectorizer.");
2499 }
2500 
2501 static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) {
2502   StringRef OldPrefix = "llvm.vectorizer.";
2503   assert(OldTag.startswith(OldPrefix) && "Expected old prefix");
2504 
2505   if (OldTag == "llvm.vectorizer.unroll")
2506     return MDString::get(C, "llvm.loop.interleave.count");
2507 
2508   return MDString::get(
2509       C, (Twine("llvm.loop.vectorize.") + OldTag.drop_front(OldPrefix.size()))
2510              .str());
2511 }
2512 
2513 static Metadata *upgradeLoopArgument(Metadata *MD) {
2514   auto *T = dyn_cast_or_null<MDTuple>(MD);
2515   if (!T)
2516     return MD;
2517   if (T->getNumOperands() < 1)
2518     return MD;
2519   auto *OldTag = dyn_cast_or_null<MDString>(T->getOperand(0));
2520   if (!OldTag)
2521     return MD;
2522   if (!OldTag->getString().startswith("llvm.vectorizer."))
2523     return MD;
2524 
2525   // This has an old tag.  Upgrade it.
2526   SmallVector<Metadata *, 8> Ops;
2527   Ops.reserve(T->getNumOperands());
2528   Ops.push_back(upgradeLoopTag(T->getContext(), OldTag->getString()));
2529   for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I)
2530     Ops.push_back(T->getOperand(I));
2531 
2532   return MDTuple::get(T->getContext(), Ops);
2533 }
2534 
2535 MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) {
2536   auto *T = dyn_cast<MDTuple>(&N);
2537   if (!T)
2538     return &N;
2539 
2540   if (none_of(T->operands(), isOldLoopArgument))
2541     return &N;
2542 
2543   SmallVector<Metadata *, 8> Ops;
2544   Ops.reserve(T->getNumOperands());
2545   for (Metadata *MD : T->operands())
2546     Ops.push_back(upgradeLoopArgument(MD));
2547 
2548   return MDTuple::get(T->getContext(), Ops);
2549 }
2550