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/Constants.h"
19 #include "llvm/IR/DIBuilder.h"
20 #include "llvm/IR/DebugInfo.h"
21 #include "llvm/IR/DiagnosticInfo.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/IR/IRBuilder.h"
24 #include "llvm/IR/Instruction.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/IR/LLVMContext.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/IR/Verifier.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/Regex.h"
31 #include <cstring>
32 using namespace llvm;
33 
34 static void rename(GlobalValue *GV) { GV->setName(GV->getName() + ".old"); }
35 
36 // Upgrade the declarations of the SSE4.1 ptest intrinsics whose arguments have
37 // changed their type from v4f32 to v2i64.
38 static bool UpgradePTESTIntrinsic(Function* F, Intrinsic::ID IID,
39                                   Function *&NewFn) {
40   // Check whether this is an old version of the function, which received
41   // v4f32 arguments.
42   Type *Arg0Type = F->getFunctionType()->getParamType(0);
43   if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
44     return false;
45 
46   // Yes, it's old, replace it with new version.
47   rename(F);
48   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
49   return true;
50 }
51 
52 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
53 // arguments have changed their type from i32 to i8.
54 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
55                                              Function *&NewFn) {
56   // Check that the last argument is an i32.
57   Type *LastArgType = F->getFunctionType()->getParamType(
58      F->getFunctionType()->getNumParams() - 1);
59   if (!LastArgType->isIntegerTy(32))
60     return false;
61 
62   // Move this function aside and map down.
63   rename(F);
64   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
65   return true;
66 }
67 
68 // Upgrade the declaration of fp compare intrinsics that change return type
69 // from scalar to vXi1 mask.
70 static bool UpgradeX86MaskedFPCompare(Function *F, Intrinsic::ID IID,
71                                       Function *&NewFn) {
72   // Check if the return type is a vector.
73   if (F->getReturnType()->isVectorTy())
74     return false;
75 
76   rename(F);
77   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
78   return true;
79 }
80 
81 static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
82   // All of the intrinsics matches below should be marked with which llvm
83   // version started autoupgrading them. At some point in the future we would
84   // like to use this information to remove upgrade code for some older
85   // intrinsics. It is currently undecided how we will determine that future
86   // point.
87   if (Name=="ssse3.pabs.b.128" || // Added in 6.0
88       Name=="ssse3.pabs.w.128" || // Added in 6.0
89       Name=="ssse3.pabs.d.128" || // Added in 6.0
90       Name.startswith("avx512.mask.shuf.i") || // Added in 6.0
91       Name.startswith("avx512.mask.shuf.f") || // Added in 6.0
92       Name.startswith("avx512.kunpck") || //added in 6.0
93       Name.startswith("avx2.pabs.") || // Added in 6.0
94       Name.startswith("avx512.mask.pabs.") || // Added in 6.0
95       Name.startswith("avx512.broadcastm") || // Added in 6.0
96       Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
97       Name.startswith("sse2.pcmpeq.") || // Added in 3.1
98       Name.startswith("sse2.pcmpgt.") || // Added in 3.1
99       Name.startswith("avx2.pcmpeq.") || // Added in 3.1
100       Name.startswith("avx2.pcmpgt.") || // Added in 3.1
101       Name.startswith("avx512.mask.pcmpeq.") || // Added in 3.9
102       Name.startswith("avx512.mask.pcmpgt.") || // Added in 3.9
103       Name.startswith("avx.vperm2f128.") || // Added in 6.0
104       Name == "avx2.vperm2i128" || // Added in 6.0
105       Name == "sse.add.ss" || // Added in 4.0
106       Name == "sse2.add.sd" || // Added in 4.0
107       Name == "sse.sub.ss" || // Added in 4.0
108       Name == "sse2.sub.sd" || // Added in 4.0
109       Name == "sse.mul.ss" || // Added in 4.0
110       Name == "sse2.mul.sd" || // Added in 4.0
111       Name == "sse.div.ss" || // Added in 4.0
112       Name == "sse2.div.sd" || // Added in 4.0
113       Name == "sse41.pmaxsb" || // Added in 3.9
114       Name == "sse2.pmaxs.w" || // Added in 3.9
115       Name == "sse41.pmaxsd" || // Added in 3.9
116       Name == "sse2.pmaxu.b" || // Added in 3.9
117       Name == "sse41.pmaxuw" || // Added in 3.9
118       Name == "sse41.pmaxud" || // Added in 3.9
119       Name == "sse41.pminsb" || // Added in 3.9
120       Name == "sse2.pmins.w" || // Added in 3.9
121       Name == "sse41.pminsd" || // Added in 3.9
122       Name == "sse2.pminu.b" || // Added in 3.9
123       Name == "sse41.pminuw" || // Added in 3.9
124       Name == "sse41.pminud" || // Added in 3.9
125       Name == "avx512.kand.w" || // Added in 7.0
126       Name == "avx512.kandn.w" || // Added in 7.0
127       Name == "avx512.knot.w" || // Added in 7.0
128       Name == "avx512.kor.w" || // Added in 7.0
129       Name == "avx512.kxor.w" || // Added in 7.0
130       Name == "avx512.kxnor.w" || // Added in 7.0
131       Name == "avx512.kortestc.w" || // Added in 7.0
132       Name == "avx512.kortestz.w" || // Added in 7.0
133       Name.startswith("avx512.mask.pshuf.b.") || // Added in 4.0
134       Name.startswith("avx2.pmax") || // Added in 3.9
135       Name.startswith("avx2.pmin") || // Added in 3.9
136       Name.startswith("avx512.mask.pmax") || // Added in 4.0
137       Name.startswith("avx512.mask.pmin") || // Added in 4.0
138       Name.startswith("avx2.vbroadcast") || // Added in 3.8
139       Name.startswith("avx2.pbroadcast") || // Added in 3.8
140       Name.startswith("avx.vpermil.") || // Added in 3.1
141       Name.startswith("sse2.pshuf") || // Added in 3.9
142       Name.startswith("avx512.pbroadcast") || // Added in 3.9
143       Name.startswith("avx512.mask.broadcast.s") || // Added in 3.9
144       Name.startswith("avx512.mask.movddup") || // Added in 3.9
145       Name.startswith("avx512.mask.movshdup") || // Added in 3.9
146       Name.startswith("avx512.mask.movsldup") || // Added in 3.9
147       Name.startswith("avx512.mask.pshuf.d.") || // Added in 3.9
148       Name.startswith("avx512.mask.pshufl.w.") || // Added in 3.9
149       Name.startswith("avx512.mask.pshufh.w.") || // Added in 3.9
150       Name.startswith("avx512.mask.shuf.p") || // Added in 4.0
151       Name.startswith("avx512.mask.vpermil.p") || // Added in 3.9
152       Name.startswith("avx512.mask.perm.df.") || // Added in 3.9
153       Name.startswith("avx512.mask.perm.di.") || // Added in 3.9
154       Name.startswith("avx512.mask.punpckl") || // Added in 3.9
155       Name.startswith("avx512.mask.punpckh") || // Added in 3.9
156       Name.startswith("avx512.mask.unpckl.") || // Added in 3.9
157       Name.startswith("avx512.mask.unpckh.") || // Added in 3.9
158       Name.startswith("avx512.mask.pand.") || // Added in 3.9
159       Name.startswith("avx512.mask.pandn.") || // Added in 3.9
160       Name.startswith("avx512.mask.por.") || // Added in 3.9
161       Name.startswith("avx512.mask.pxor.") || // Added in 3.9
162       Name.startswith("avx512.mask.and.") || // Added in 3.9
163       Name.startswith("avx512.mask.andn.") || // Added in 3.9
164       Name.startswith("avx512.mask.or.") || // Added in 3.9
165       Name.startswith("avx512.mask.xor.") || // Added in 3.9
166       Name.startswith("avx512.mask.padd.") || // Added in 4.0
167       Name.startswith("avx512.mask.psub.") || // Added in 4.0
168       Name.startswith("avx512.mask.pmull.") || // Added in 4.0
169       Name.startswith("avx512.mask.cvtdq2pd.") || // Added in 4.0
170       Name.startswith("avx512.mask.cvtudq2pd.") || // Added in 4.0
171       Name.startswith("avx512.mask.pmul.dq.") || // Added in 4.0
172       Name.startswith("avx512.mask.pmulu.dq.") || // Added in 4.0
173       Name.startswith("avx512.mask.pmul.hr.sw.") || // Added in 7.0
174       Name.startswith("avx512.mask.pmulh.w.") || // Added in 7.0
175       Name.startswith("avx512.mask.pmulhu.w.") || // Added in 7.0
176       Name.startswith("avx512.mask.packsswb.") || // Added in 5.0
177       Name.startswith("avx512.mask.packssdw.") || // Added in 5.0
178       Name.startswith("avx512.mask.packuswb.") || // Added in 5.0
179       Name.startswith("avx512.mask.packusdw.") || // Added in 5.0
180       Name.startswith("avx512.mask.cmp.b") || // Added in 5.0
181       Name.startswith("avx512.mask.cmp.d") || // Added in 5.0
182       Name.startswith("avx512.mask.cmp.q") || // Added in 5.0
183       Name.startswith("avx512.mask.cmp.w") || // Added in 5.0
184       Name.startswith("avx512.mask.ucmp.") || // Added in 5.0
185       Name.startswith("avx512.cvtb2mask.") || // Added in 7.0
186       Name.startswith("avx512.cvtw2mask.") || // Added in 7.0
187       Name.startswith("avx512.cvtd2mask.") || // Added in 7.0
188       Name.startswith("avx512.cvtq2mask.") || // Added in 7.0
189       Name == "avx512.mask.add.pd.128" || // Added in 4.0
190       Name == "avx512.mask.add.pd.256" || // Added in 4.0
191       Name == "avx512.mask.add.ps.128" || // Added in 4.0
192       Name == "avx512.mask.add.ps.256" || // Added in 4.0
193       Name == "avx512.mask.div.pd.128" || // Added in 4.0
194       Name == "avx512.mask.div.pd.256" || // Added in 4.0
195       Name == "avx512.mask.div.ps.128" || // Added in 4.0
196       Name == "avx512.mask.div.ps.256" || // Added in 4.0
197       Name == "avx512.mask.mul.pd.128" || // Added in 4.0
198       Name == "avx512.mask.mul.pd.256" || // Added in 4.0
199       Name == "avx512.mask.mul.ps.128" || // Added in 4.0
200       Name == "avx512.mask.mul.ps.256" || // Added in 4.0
201       Name == "avx512.mask.sub.pd.128" || // Added in 4.0
202       Name == "avx512.mask.sub.pd.256" || // Added in 4.0
203       Name == "avx512.mask.sub.ps.128" || // Added in 4.0
204       Name == "avx512.mask.sub.ps.256" || // Added in 4.0
205       Name == "avx512.mask.max.pd.128" || // Added in 5.0
206       Name == "avx512.mask.max.pd.256" || // Added in 5.0
207       Name == "avx512.mask.max.ps.128" || // Added in 5.0
208       Name == "avx512.mask.max.ps.256" || // Added in 5.0
209       Name == "avx512.mask.min.pd.128" || // Added in 5.0
210       Name == "avx512.mask.min.pd.256" || // Added in 5.0
211       Name == "avx512.mask.min.ps.128" || // Added in 5.0
212       Name == "avx512.mask.min.ps.256" || // Added in 5.0
213       Name.startswith("avx512.mask.vpermilvar.") || // Added in 4.0
214       Name.startswith("avx512.mask.psll.d") || // Added in 4.0
215       Name.startswith("avx512.mask.psll.q") || // Added in 4.0
216       Name.startswith("avx512.mask.psll.w") || // Added in 4.0
217       Name.startswith("avx512.mask.psra.d") || // Added in 4.0
218       Name.startswith("avx512.mask.psra.q") || // Added in 4.0
219       Name.startswith("avx512.mask.psra.w") || // Added in 4.0
220       Name.startswith("avx512.mask.psrl.d") || // Added in 4.0
221       Name.startswith("avx512.mask.psrl.q") || // Added in 4.0
222       Name.startswith("avx512.mask.psrl.w") || // Added in 4.0
223       Name.startswith("avx512.mask.pslli") || // Added in 4.0
224       Name.startswith("avx512.mask.psrai") || // Added in 4.0
225       Name.startswith("avx512.mask.psrli") || // Added in 4.0
226       Name.startswith("avx512.mask.psllv") || // Added in 4.0
227       Name.startswith("avx512.mask.psrav") || // Added in 4.0
228       Name.startswith("avx512.mask.psrlv") || // Added in 4.0
229       Name.startswith("sse41.pmovsx") || // Added in 3.8
230       Name.startswith("sse41.pmovzx") || // Added in 3.9
231       Name.startswith("avx2.pmovsx") || // Added in 3.9
232       Name.startswith("avx2.pmovzx") || // Added in 3.9
233       Name.startswith("avx512.mask.pmovsx") || // Added in 4.0
234       Name.startswith("avx512.mask.pmovzx") || // Added in 4.0
235       Name.startswith("avx512.mask.lzcnt.") || // Added in 5.0
236       Name == "sse2.cvtdq2pd" || // Added in 3.9
237       Name == "sse2.cvtps2pd" || // Added in 3.9
238       Name == "avx.cvtdq2.pd.256" || // Added in 3.9
239       Name == "avx.cvt.ps2.pd.256" || // Added in 3.9
240       Name.startswith("avx.vinsertf128.") || // Added in 3.7
241       Name == "avx2.vinserti128" || // Added in 3.7
242       Name.startswith("avx512.mask.insert") || // Added in 4.0
243       Name.startswith("avx.vextractf128.") || // Added in 3.7
244       Name == "avx2.vextracti128" || // Added in 3.7
245       Name.startswith("avx512.mask.vextract") || // Added in 4.0
246       Name.startswith("sse4a.movnt.") || // Added in 3.9
247       Name.startswith("avx.movnt.") || // Added in 3.2
248       Name.startswith("avx512.storent.") || // Added in 3.9
249       Name == "sse41.movntdqa" || // Added in 5.0
250       Name == "avx2.movntdqa" || // Added in 5.0
251       Name == "avx512.movntdqa" || // Added in 5.0
252       Name == "sse2.storel.dq" || // Added in 3.9
253       Name.startswith("sse.storeu.") || // Added in 3.9
254       Name.startswith("sse2.storeu.") || // Added in 3.9
255       Name.startswith("avx.storeu.") || // Added in 3.9
256       Name.startswith("avx512.mask.storeu.") || // Added in 3.9
257       Name.startswith("avx512.mask.store.p") || // Added in 3.9
258       Name.startswith("avx512.mask.store.b.") || // Added in 3.9
259       Name.startswith("avx512.mask.store.w.") || // Added in 3.9
260       Name.startswith("avx512.mask.store.d.") || // Added in 3.9
261       Name.startswith("avx512.mask.store.q.") || // Added in 3.9
262       Name.startswith("avx512.mask.loadu.") || // Added in 3.9
263       Name.startswith("avx512.mask.load.") || // Added in 3.9
264       Name == "sse42.crc32.64.8" || // Added in 3.4
265       Name.startswith("avx.vbroadcast.s") || // Added in 3.5
266       Name.startswith("avx512.mask.palignr.") || // Added in 3.9
267       Name.startswith("avx512.mask.valign.") || // Added in 4.0
268       Name.startswith("sse2.psll.dq") || // Added in 3.7
269       Name.startswith("sse2.psrl.dq") || // Added in 3.7
270       Name.startswith("avx2.psll.dq") || // Added in 3.7
271       Name.startswith("avx2.psrl.dq") || // Added in 3.7
272       Name.startswith("avx512.psll.dq") || // Added in 3.9
273       Name.startswith("avx512.psrl.dq") || // Added in 3.9
274       Name == "sse41.pblendw" || // Added in 3.7
275       Name.startswith("sse41.blendp") || // Added in 3.7
276       Name.startswith("avx.blend.p") || // Added in 3.7
277       Name == "avx2.pblendw" || // Added in 3.7
278       Name.startswith("avx2.pblendd.") || // Added in 3.7
279       Name.startswith("avx.vbroadcastf128") || // Added in 4.0
280       Name == "avx2.vbroadcasti128" || // Added in 3.7
281       Name.startswith("avx512.mask.broadcastf") || // Added in 6.0
282       Name.startswith("avx512.mask.broadcasti") || // Added in 6.0
283       Name == "xop.vpcmov" || // Added in 3.8
284       Name == "xop.vpcmov.256" || // Added in 5.0
285       Name.startswith("avx512.mask.move.s") || // Added in 4.0
286       Name.startswith("avx512.cvtmask2") || // Added in 5.0
287       (Name.startswith("xop.vpcom") && // Added in 3.2
288        F->arg_size() == 2) ||
289       Name.startswith("avx512.ptestm") || //Added in 6.0
290       Name.startswith("avx512.ptestnm") || //Added in 6.0
291       Name.startswith("sse2.pavg") || // Added in 6.0
292       Name.startswith("avx2.pavg") || // Added in 6.0
293       Name.startswith("avx512.mask.pavg")) // Added in 6.0
294     return true;
295 
296   return false;
297 }
298 
299 static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
300                                         Function *&NewFn) {
301   // Only handle intrinsics that start with "x86.".
302   if (!Name.startswith("x86."))
303     return false;
304   // Remove "x86." prefix.
305   Name = Name.substr(4);
306 
307   if (ShouldUpgradeX86Intrinsic(F, Name)) {
308     NewFn = nullptr;
309     return true;
310   }
311 
312   // SSE4.1 ptest functions may have an old signature.
313   if (Name.startswith("sse41.ptest")) { // Added in 3.2
314     if (Name.substr(11) == "c")
315       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestc, NewFn);
316     if (Name.substr(11) == "z")
317       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestz, NewFn);
318     if (Name.substr(11) == "nzc")
319       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
320   }
321   // Several blend and other instructions with masks used the wrong number of
322   // bits.
323   if (Name == "sse41.insertps") // Added in 3.6
324     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
325                                             NewFn);
326   if (Name == "sse41.dppd") // Added in 3.6
327     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
328                                             NewFn);
329   if (Name == "sse41.dpps") // Added in 3.6
330     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
331                                             NewFn);
332   if (Name == "sse41.mpsadbw") // Added in 3.6
333     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
334                                             NewFn);
335   if (Name == "avx.dp.ps.256") // Added in 3.6
336     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
337                                             NewFn);
338   if (Name == "avx2.mpsadbw") // Added in 3.6
339     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
340                                             NewFn);
341   if (Name == "avx512.mask.cmp.pd.128") // Added in 7.0
342     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_128,
343                                      NewFn);
344   if (Name == "avx512.mask.cmp.pd.256") // Added in 7.0
345     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_256,
346                                      NewFn);
347   if (Name == "avx512.mask.cmp.pd.512") // Added in 7.0
348     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_512,
349                                      NewFn);
350   if (Name == "avx512.mask.cmp.ps.128") // Added in 7.0
351     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_128,
352                                      NewFn);
353   if (Name == "avx512.mask.cmp.ps.256") // Added in 7.0
354     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_256,
355                                      NewFn);
356   if (Name == "avx512.mask.cmp.ps.512") // Added in 7.0
357     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_512,
358                                      NewFn);
359 
360   // frcz.ss/sd may need to have an argument dropped. Added in 3.2
361   if (Name.startswith("xop.vfrcz.ss") && F->arg_size() == 2) {
362     rename(F);
363     NewFn = Intrinsic::getDeclaration(F->getParent(),
364                                       Intrinsic::x86_xop_vfrcz_ss);
365     return true;
366   }
367   if (Name.startswith("xop.vfrcz.sd") && F->arg_size() == 2) {
368     rename(F);
369     NewFn = Intrinsic::getDeclaration(F->getParent(),
370                                       Intrinsic::x86_xop_vfrcz_sd);
371     return true;
372   }
373   // Upgrade any XOP PERMIL2 index operand still using a float/double vector.
374   if (Name.startswith("xop.vpermil2")) { // Added in 3.9
375     auto Idx = F->getFunctionType()->getParamType(2);
376     if (Idx->isFPOrFPVectorTy()) {
377       rename(F);
378       unsigned IdxSize = Idx->getPrimitiveSizeInBits();
379       unsigned EltSize = Idx->getScalarSizeInBits();
380       Intrinsic::ID Permil2ID;
381       if (EltSize == 64 && IdxSize == 128)
382         Permil2ID = Intrinsic::x86_xop_vpermil2pd;
383       else if (EltSize == 32 && IdxSize == 128)
384         Permil2ID = Intrinsic::x86_xop_vpermil2ps;
385       else if (EltSize == 64 && IdxSize == 256)
386         Permil2ID = Intrinsic::x86_xop_vpermil2pd_256;
387       else
388         Permil2ID = Intrinsic::x86_xop_vpermil2ps_256;
389       NewFn = Intrinsic::getDeclaration(F->getParent(), Permil2ID);
390       return true;
391     }
392   }
393 
394   return false;
395 }
396 
397 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
398   assert(F && "Illegal to upgrade a non-existent Function.");
399 
400   // Quickly eliminate it, if it's not a candidate.
401   StringRef Name = F->getName();
402   if (Name.size() <= 8 || !Name.startswith("llvm."))
403     return false;
404   Name = Name.substr(5); // Strip off "llvm."
405 
406   switch (Name[0]) {
407   default: break;
408   case 'a': {
409     if (Name.startswith("arm.rbit") || Name.startswith("aarch64.rbit")) {
410       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
411                                         F->arg_begin()->getType());
412       return true;
413     }
414     if (Name.startswith("arm.neon.vclz")) {
415       Type* args[2] = {
416         F->arg_begin()->getType(),
417         Type::getInt1Ty(F->getContext())
418       };
419       // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
420       // the end of the name. Change name from llvm.arm.neon.vclz.* to
421       //  llvm.ctlz.*
422       FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
423       NewFn = Function::Create(fType, F->getLinkage(),
424                                "llvm.ctlz." + Name.substr(14), F->getParent());
425       return true;
426     }
427     if (Name.startswith("arm.neon.vcnt")) {
428       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
429                                         F->arg_begin()->getType());
430       return true;
431     }
432     Regex vldRegex("^arm\\.neon\\.vld([1234]|[234]lane)\\.v[a-z0-9]*$");
433     if (vldRegex.match(Name)) {
434       auto fArgs = F->getFunctionType()->params();
435       SmallVector<Type *, 4> Tys(fArgs.begin(), fArgs.end());
436       // Can't use Intrinsic::getDeclaration here as the return types might
437       // then only be structurally equal.
438       FunctionType* fType = FunctionType::get(F->getReturnType(), Tys, false);
439       NewFn = Function::Create(fType, F->getLinkage(),
440                                "llvm." + Name + ".p0i8", F->getParent());
441       return true;
442     }
443     Regex vstRegex("^arm\\.neon\\.vst([1234]|[234]lane)\\.v[a-z0-9]*$");
444     if (vstRegex.match(Name)) {
445       static const Intrinsic::ID StoreInts[] = {Intrinsic::arm_neon_vst1,
446                                                 Intrinsic::arm_neon_vst2,
447                                                 Intrinsic::arm_neon_vst3,
448                                                 Intrinsic::arm_neon_vst4};
449 
450       static const Intrinsic::ID StoreLaneInts[] = {
451         Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
452         Intrinsic::arm_neon_vst4lane
453       };
454 
455       auto fArgs = F->getFunctionType()->params();
456       Type *Tys[] = {fArgs[0], fArgs[1]};
457       if (Name.find("lane") == StringRef::npos)
458         NewFn = Intrinsic::getDeclaration(F->getParent(),
459                                           StoreInts[fArgs.size() - 3], Tys);
460       else
461         NewFn = Intrinsic::getDeclaration(F->getParent(),
462                                           StoreLaneInts[fArgs.size() - 5], Tys);
463       return true;
464     }
465     if (Name == "aarch64.thread.pointer" || Name == "arm.thread.pointer") {
466       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
467       return true;
468     }
469     break;
470   }
471 
472   case 'c': {
473     if (Name.startswith("ctlz.") && F->arg_size() == 1) {
474       rename(F);
475       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
476                                         F->arg_begin()->getType());
477       return true;
478     }
479     if (Name.startswith("cttz.") && F->arg_size() == 1) {
480       rename(F);
481       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
482                                         F->arg_begin()->getType());
483       return true;
484     }
485     break;
486   }
487   case 'd': {
488     if (Name == "dbg.value" && F->arg_size() == 4) {
489       rename(F);
490       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_value);
491       return true;
492     }
493     break;
494   }
495   case 'i':
496   case 'l': {
497     bool IsLifetimeStart = Name.startswith("lifetime.start");
498     if (IsLifetimeStart || Name.startswith("invariant.start")) {
499       Intrinsic::ID ID = IsLifetimeStart ?
500         Intrinsic::lifetime_start : Intrinsic::invariant_start;
501       auto Args = F->getFunctionType()->params();
502       Type* ObjectPtr[1] = {Args[1]};
503       if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
504         rename(F);
505         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
506         return true;
507       }
508     }
509 
510     bool IsLifetimeEnd = Name.startswith("lifetime.end");
511     if (IsLifetimeEnd || Name.startswith("invariant.end")) {
512       Intrinsic::ID ID = IsLifetimeEnd ?
513         Intrinsic::lifetime_end : Intrinsic::invariant_end;
514 
515       auto Args = F->getFunctionType()->params();
516       Type* ObjectPtr[1] = {Args[IsLifetimeEnd ? 1 : 2]};
517       if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
518         rename(F);
519         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
520         return true;
521       }
522     }
523     break;
524   }
525   case 'm': {
526     if (Name.startswith("masked.load.")) {
527       Type *Tys[] = { F->getReturnType(), F->arg_begin()->getType() };
528       if (F->getName() != Intrinsic::getName(Intrinsic::masked_load, Tys)) {
529         rename(F);
530         NewFn = Intrinsic::getDeclaration(F->getParent(),
531                                           Intrinsic::masked_load,
532                                           Tys);
533         return true;
534       }
535     }
536     if (Name.startswith("masked.store.")) {
537       auto Args = F->getFunctionType()->params();
538       Type *Tys[] = { Args[0], Args[1] };
539       if (F->getName() != Intrinsic::getName(Intrinsic::masked_store, Tys)) {
540         rename(F);
541         NewFn = Intrinsic::getDeclaration(F->getParent(),
542                                           Intrinsic::masked_store,
543                                           Tys);
544         return true;
545       }
546     }
547     // Renaming gather/scatter intrinsics with no address space overloading
548     // to the new overload which includes an address space
549     if (Name.startswith("masked.gather.")) {
550       Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
551       if (F->getName() != Intrinsic::getName(Intrinsic::masked_gather, Tys)) {
552         rename(F);
553         NewFn = Intrinsic::getDeclaration(F->getParent(),
554                                           Intrinsic::masked_gather, Tys);
555         return true;
556       }
557     }
558     if (Name.startswith("masked.scatter.")) {
559       auto Args = F->getFunctionType()->params();
560       Type *Tys[] = {Args[0], Args[1]};
561       if (F->getName() != Intrinsic::getName(Intrinsic::masked_scatter, Tys)) {
562         rename(F);
563         NewFn = Intrinsic::getDeclaration(F->getParent(),
564                                           Intrinsic::masked_scatter, Tys);
565         return true;
566       }
567     }
568     // Updating the memory intrinsics (memcpy/memmove/memset) that have an
569     // alignment parameter to embedding the alignment as an attribute of
570     // the pointer args.
571     if (Name.startswith("memcpy.") && F->arg_size() == 5) {
572       rename(F);
573       // Get the types of dest, src, and len
574       ArrayRef<Type *> ParamTypes = F->getFunctionType()->params().slice(0, 3);
575       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memcpy,
576                                         ParamTypes);
577       return true;
578     }
579     if (Name.startswith("memmove.") && F->arg_size() == 5) {
580       rename(F);
581       // Get the types of dest, src, and len
582       ArrayRef<Type *> ParamTypes = F->getFunctionType()->params().slice(0, 3);
583       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memmove,
584                                         ParamTypes);
585       return true;
586     }
587     if (Name.startswith("memset.") && F->arg_size() == 5) {
588       rename(F);
589       // Get the types of dest, and len
590       const auto *FT = F->getFunctionType();
591       Type *ParamTypes[2] = {
592           FT->getParamType(0), // Dest
593           FT->getParamType(2)  // len
594       };
595       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memset,
596                                         ParamTypes);
597       return true;
598     }
599     break;
600   }
601   case 'n': {
602     if (Name.startswith("nvvm.")) {
603       Name = Name.substr(5);
604 
605       // The following nvvm intrinsics correspond exactly to an LLVM intrinsic.
606       Intrinsic::ID IID = StringSwitch<Intrinsic::ID>(Name)
607                               .Cases("brev32", "brev64", Intrinsic::bitreverse)
608                               .Case("clz.i", Intrinsic::ctlz)
609                               .Case("popc.i", Intrinsic::ctpop)
610                               .Default(Intrinsic::not_intrinsic);
611       if (IID != Intrinsic::not_intrinsic && F->arg_size() == 1) {
612         NewFn = Intrinsic::getDeclaration(F->getParent(), IID,
613                                           {F->getReturnType()});
614         return true;
615       }
616 
617       // The following nvvm intrinsics correspond exactly to an LLVM idiom, but
618       // not to an intrinsic alone.  We expand them in UpgradeIntrinsicCall.
619       //
620       // TODO: We could add lohi.i2d.
621       bool Expand = StringSwitch<bool>(Name)
622                         .Cases("abs.i", "abs.ll", true)
623                         .Cases("clz.ll", "popc.ll", "h2f", true)
624                         .Cases("max.i", "max.ll", "max.ui", "max.ull", true)
625                         .Cases("min.i", "min.ll", "min.ui", "min.ull", true)
626                         .Default(false);
627       if (Expand) {
628         NewFn = nullptr;
629         return true;
630       }
631     }
632     break;
633   }
634   case 'o':
635     // We only need to change the name to match the mangling including the
636     // address space.
637     if (Name.startswith("objectsize.")) {
638       Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
639       if (F->arg_size() == 2 ||
640           F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
641         rename(F);
642         NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::objectsize,
643                                           Tys);
644         return true;
645       }
646     }
647     break;
648 
649   case 's':
650     if (Name == "stackprotectorcheck") {
651       NewFn = nullptr;
652       return true;
653     }
654     break;
655 
656   case 'x':
657     if (UpgradeX86IntrinsicFunction(F, Name, NewFn))
658       return true;
659   }
660   // Remangle our intrinsic since we upgrade the mangling
661   auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F);
662   if (Result != None) {
663     NewFn = Result.getValue();
664     return true;
665   }
666 
667   //  This may not belong here. This function is effectively being overloaded
668   //  to both detect an intrinsic which needs upgrading, and to provide the
669   //  upgraded form of the intrinsic. We should perhaps have two separate
670   //  functions for this.
671   return false;
672 }
673 
674 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
675   NewFn = nullptr;
676   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
677   assert(F != NewFn && "Intrinsic function upgraded to the same function");
678 
679   // Upgrade intrinsic attributes.  This does not change the function.
680   if (NewFn)
681     F = NewFn;
682   if (Intrinsic::ID id = F->getIntrinsicID())
683     F->setAttributes(Intrinsic::getAttributes(F->getContext(), id));
684   return Upgraded;
685 }
686 
687 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
688   // Nothing to do yet.
689   return false;
690 }
691 
692 // Handles upgrading SSE2/AVX2/AVX512BW PSLLDQ intrinsics by converting them
693 // to byte shuffles.
694 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder,
695                                          Value *Op, unsigned Shift) {
696   Type *ResultTy = Op->getType();
697   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
698 
699   // Bitcast from a 64-bit element type to a byte element type.
700   Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
701   Op = Builder.CreateBitCast(Op, VecTy, "cast");
702 
703   // We'll be shuffling in zeroes.
704   Value *Res = Constant::getNullValue(VecTy);
705 
706   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
707   // we'll just return the zero vector.
708   if (Shift < 16) {
709     uint32_t Idxs[64];
710     // 256/512-bit version is split into 2/4 16-byte lanes.
711     for (unsigned l = 0; l != NumElts; l += 16)
712       for (unsigned i = 0; i != 16; ++i) {
713         unsigned Idx = NumElts + i - Shift;
714         if (Idx < NumElts)
715           Idx -= NumElts - 16; // end of lane, switch operand.
716         Idxs[l + i] = Idx + l;
717       }
718 
719     Res = Builder.CreateShuffleVector(Res, Op, makeArrayRef(Idxs, NumElts));
720   }
721 
722   // Bitcast back to a 64-bit element type.
723   return Builder.CreateBitCast(Res, ResultTy, "cast");
724 }
725 
726 // Handles upgrading SSE2/AVX2/AVX512BW PSRLDQ intrinsics by converting them
727 // to byte shuffles.
728 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, Value *Op,
729                                          unsigned Shift) {
730   Type *ResultTy = Op->getType();
731   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
732 
733   // Bitcast from a 64-bit element type to a byte element type.
734   Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
735   Op = Builder.CreateBitCast(Op, VecTy, "cast");
736 
737   // We'll be shuffling in zeroes.
738   Value *Res = Constant::getNullValue(VecTy);
739 
740   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
741   // we'll just return the zero vector.
742   if (Shift < 16) {
743     uint32_t Idxs[64];
744     // 256/512-bit version is split into 2/4 16-byte lanes.
745     for (unsigned l = 0; l != NumElts; l += 16)
746       for (unsigned i = 0; i != 16; ++i) {
747         unsigned Idx = i + Shift;
748         if (Idx >= 16)
749           Idx += NumElts - 16; // end of lane, switch operand.
750         Idxs[l + i] = Idx + l;
751       }
752 
753     Res = Builder.CreateShuffleVector(Op, Res, makeArrayRef(Idxs, NumElts));
754   }
755 
756   // Bitcast back to a 64-bit element type.
757   return Builder.CreateBitCast(Res, ResultTy, "cast");
758 }
759 
760 static Value *getX86MaskVec(IRBuilder<> &Builder, Value *Mask,
761                             unsigned NumElts) {
762   llvm::VectorType *MaskTy = llvm::VectorType::get(Builder.getInt1Ty(),
763                              cast<IntegerType>(Mask->getType())->getBitWidth());
764   Mask = Builder.CreateBitCast(Mask, MaskTy);
765 
766   // If we have less than 8 elements, then the starting mask was an i8 and
767   // we need to extract down to the right number of elements.
768   if (NumElts < 8) {
769     uint32_t Indices[4];
770     for (unsigned i = 0; i != NumElts; ++i)
771       Indices[i] = i;
772     Mask = Builder.CreateShuffleVector(Mask, Mask,
773                                        makeArrayRef(Indices, NumElts),
774                                        "extract");
775   }
776 
777   return Mask;
778 }
779 
780 static Value *EmitX86Select(IRBuilder<> &Builder, Value *Mask,
781                             Value *Op0, Value *Op1) {
782   // If the mask is all ones just emit the align operation.
783   if (const auto *C = dyn_cast<Constant>(Mask))
784     if (C->isAllOnesValue())
785       return Op0;
786 
787   Mask = getX86MaskVec(Builder, Mask, Op0->getType()->getVectorNumElements());
788   return Builder.CreateSelect(Mask, Op0, Op1);
789 }
790 
791 // Handle autoupgrade for masked PALIGNR and VALIGND/Q intrinsics.
792 // PALIGNR handles large immediates by shifting while VALIGN masks the immediate
793 // so we need to handle both cases. VALIGN also doesn't have 128-bit lanes.
794 static Value *UpgradeX86ALIGNIntrinsics(IRBuilder<> &Builder, Value *Op0,
795                                         Value *Op1, Value *Shift,
796                                         Value *Passthru, Value *Mask,
797                                         bool IsVALIGN) {
798   unsigned ShiftVal = cast<llvm::ConstantInt>(Shift)->getZExtValue();
799 
800   unsigned NumElts = Op0->getType()->getVectorNumElements();
801   assert((IsVALIGN || NumElts % 16 == 0) && "Illegal NumElts for PALIGNR!");
802   assert((!IsVALIGN || NumElts <= 16) && "NumElts too large for VALIGN!");
803   assert(isPowerOf2_32(NumElts) && "NumElts not a power of 2!");
804 
805   // Mask the immediate for VALIGN.
806   if (IsVALIGN)
807     ShiftVal &= (NumElts - 1);
808 
809   // If palignr is shifting the pair of vectors more than the size of two
810   // lanes, emit zero.
811   if (ShiftVal >= 32)
812     return llvm::Constant::getNullValue(Op0->getType());
813 
814   // If palignr is shifting the pair of input vectors more than one lane,
815   // but less than two lanes, convert to shifting in zeroes.
816   if (ShiftVal > 16) {
817     ShiftVal -= 16;
818     Op1 = Op0;
819     Op0 = llvm::Constant::getNullValue(Op0->getType());
820   }
821 
822   uint32_t Indices[64];
823   // 256-bit palignr operates on 128-bit lanes so we need to handle that
824   for (unsigned l = 0; l < NumElts; l += 16) {
825     for (unsigned i = 0; i != 16; ++i) {
826       unsigned Idx = ShiftVal + i;
827       if (!IsVALIGN && Idx >= 16) // Disable wrap for VALIGN.
828         Idx += NumElts - 16; // End of lane, switch operand.
829       Indices[l + i] = Idx + l;
830     }
831   }
832 
833   Value *Align = Builder.CreateShuffleVector(Op1, Op0,
834                                              makeArrayRef(Indices, NumElts),
835                                              "palignr");
836 
837   return EmitX86Select(Builder, Mask, Align, Passthru);
838 }
839 
840 static Value *UpgradeMaskedStore(IRBuilder<> &Builder,
841                                  Value *Ptr, Value *Data, Value *Mask,
842                                  bool Aligned) {
843   // Cast the pointer to the right type.
844   Ptr = Builder.CreateBitCast(Ptr,
845                               llvm::PointerType::getUnqual(Data->getType()));
846   unsigned Align =
847     Aligned ? cast<VectorType>(Data->getType())->getBitWidth() / 8 : 1;
848 
849   // If the mask is all ones just emit a regular store.
850   if (const auto *C = dyn_cast<Constant>(Mask))
851     if (C->isAllOnesValue())
852       return Builder.CreateAlignedStore(Data, Ptr, Align);
853 
854   // Convert the mask from an integer type to a vector of i1.
855   unsigned NumElts = Data->getType()->getVectorNumElements();
856   Mask = getX86MaskVec(Builder, Mask, NumElts);
857   return Builder.CreateMaskedStore(Data, Ptr, Align, Mask);
858 }
859 
860 static Value *UpgradeMaskedLoad(IRBuilder<> &Builder,
861                                 Value *Ptr, Value *Passthru, Value *Mask,
862                                 bool Aligned) {
863   // Cast the pointer to the right type.
864   Ptr = Builder.CreateBitCast(Ptr,
865                              llvm::PointerType::getUnqual(Passthru->getType()));
866   unsigned Align =
867     Aligned ? cast<VectorType>(Passthru->getType())->getBitWidth() / 8 : 1;
868 
869   // If the mask is all ones just emit a regular store.
870   if (const auto *C = dyn_cast<Constant>(Mask))
871     if (C->isAllOnesValue())
872       return Builder.CreateAlignedLoad(Ptr, Align);
873 
874   // Convert the mask from an integer type to a vector of i1.
875   unsigned NumElts = Passthru->getType()->getVectorNumElements();
876   Mask = getX86MaskVec(Builder, Mask, NumElts);
877   return Builder.CreateMaskedLoad(Ptr, Align, Mask, Passthru);
878 }
879 
880 static Value *upgradeAbs(IRBuilder<> &Builder, CallInst &CI) {
881   Value *Op0 = CI.getArgOperand(0);
882   llvm::Type *Ty = Op0->getType();
883   Value *Zero = llvm::Constant::getNullValue(Ty);
884   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_SGT, Op0, Zero);
885   Value *Neg = Builder.CreateNeg(Op0);
886   Value *Res = Builder.CreateSelect(Cmp, Op0, Neg);
887 
888   if (CI.getNumArgOperands() == 3)
889     Res = EmitX86Select(Builder,CI.getArgOperand(2), Res, CI.getArgOperand(1));
890 
891   return Res;
892 }
893 
894 static Value *upgradeIntMinMax(IRBuilder<> &Builder, CallInst &CI,
895                                ICmpInst::Predicate Pred) {
896   Value *Op0 = CI.getArgOperand(0);
897   Value *Op1 = CI.getArgOperand(1);
898   Value *Cmp = Builder.CreateICmp(Pred, Op0, Op1);
899   Value *Res = Builder.CreateSelect(Cmp, Op0, Op1);
900 
901   if (CI.getNumArgOperands() == 4)
902     Res = EmitX86Select(Builder, CI.getArgOperand(3), Res, CI.getArgOperand(2));
903 
904   return Res;
905 }
906 
907 // Applying mask on vector of i1's and make sure result is at least 8 bits wide.
908 static Value *ApplyX86MaskOn1BitsVec(IRBuilder<> &Builder,Value *Vec, Value *Mask,
909                                      unsigned NumElts) {
910   if (Mask) {
911     const auto *C = dyn_cast<Constant>(Mask);
912     if (!C || !C->isAllOnesValue())
913       Vec = Builder.CreateAnd(Vec, getX86MaskVec(Builder, Mask, NumElts));
914   }
915 
916   if (NumElts < 8) {
917     uint32_t Indices[8];
918     for (unsigned i = 0; i != NumElts; ++i)
919       Indices[i] = i;
920     for (unsigned i = NumElts; i != 8; ++i)
921       Indices[i] = NumElts + i % NumElts;
922     Vec = Builder.CreateShuffleVector(Vec,
923                                       Constant::getNullValue(Vec->getType()),
924                                       Indices);
925   }
926   return Builder.CreateBitCast(Vec, Builder.getIntNTy(std::max(NumElts, 8U)));
927 }
928 
929 static Value *upgradeMaskedCompare(IRBuilder<> &Builder, CallInst &CI,
930                                    unsigned CC, bool Signed) {
931   Value *Op0 = CI.getArgOperand(0);
932   unsigned NumElts = Op0->getType()->getVectorNumElements();
933 
934   Value *Cmp;
935   if (CC == 3) {
936     Cmp = Constant::getNullValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
937   } else if (CC == 7) {
938     Cmp = Constant::getAllOnesValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
939   } else {
940     ICmpInst::Predicate Pred;
941     switch (CC) {
942     default: llvm_unreachable("Unknown condition code");
943     case 0: Pred = ICmpInst::ICMP_EQ;  break;
944     case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
945     case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
946     case 4: Pred = ICmpInst::ICMP_NE;  break;
947     case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
948     case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
949     }
950     Cmp = Builder.CreateICmp(Pred, Op0, CI.getArgOperand(1));
951   }
952 
953   Value *Mask = CI.getArgOperand(CI.getNumArgOperands() - 1);
954 
955   return ApplyX86MaskOn1BitsVec(Builder, Cmp, Mask, NumElts);
956 }
957 
958 // Replace a masked intrinsic with an older unmasked intrinsic.
959 static Value *UpgradeX86MaskedShift(IRBuilder<> &Builder, CallInst &CI,
960                                     Intrinsic::ID IID) {
961   Function *F = CI.getCalledFunction();
962   Function *Intrin = Intrinsic::getDeclaration(F->getParent(), IID);
963   Value *Rep = Builder.CreateCall(Intrin,
964                                  { CI.getArgOperand(0), CI.getArgOperand(1) });
965   return EmitX86Select(Builder, CI.getArgOperand(3), Rep, CI.getArgOperand(2));
966 }
967 
968 static Value* upgradeMaskedMove(IRBuilder<> &Builder, CallInst &CI) {
969   Value* A = CI.getArgOperand(0);
970   Value* B = CI.getArgOperand(1);
971   Value* Src = CI.getArgOperand(2);
972   Value* Mask = CI.getArgOperand(3);
973 
974   Value* AndNode = Builder.CreateAnd(Mask, APInt(8, 1));
975   Value* Cmp = Builder.CreateIsNotNull(AndNode);
976   Value* Extract1 = Builder.CreateExtractElement(B, (uint64_t)0);
977   Value* Extract2 = Builder.CreateExtractElement(Src, (uint64_t)0);
978   Value* Select = Builder.CreateSelect(Cmp, Extract1, Extract2);
979   return Builder.CreateInsertElement(A, Select, (uint64_t)0);
980 }
981 
982 
983 static Value* UpgradeMaskToInt(IRBuilder<> &Builder, CallInst &CI) {
984   Value* Op = CI.getArgOperand(0);
985   Type* ReturnOp = CI.getType();
986   unsigned NumElts = CI.getType()->getVectorNumElements();
987   Value *Mask = getX86MaskVec(Builder, Op, NumElts);
988   return Builder.CreateSExt(Mask, ReturnOp, "vpmovm2");
989 }
990 
991 /// Upgrade a call to an old intrinsic. All argument and return casting must be
992 /// provided to seamlessly integrate with existing context.
993 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
994   Function *F = CI->getCalledFunction();
995   LLVMContext &C = CI->getContext();
996   IRBuilder<> Builder(C);
997   Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
998 
999   assert(F && "Intrinsic call is not direct?");
1000 
1001   if (!NewFn) {
1002     // Get the Function's name.
1003     StringRef Name = F->getName();
1004 
1005     assert(Name.startswith("llvm.") && "Intrinsic doesn't start with 'llvm.'");
1006     Name = Name.substr(5);
1007 
1008     bool IsX86 = Name.startswith("x86.");
1009     if (IsX86)
1010       Name = Name.substr(4);
1011     bool IsNVVM = Name.startswith("nvvm.");
1012     if (IsNVVM)
1013       Name = Name.substr(5);
1014 
1015     if (IsX86 && Name.startswith("sse4a.movnt.")) {
1016       Module *M = F->getParent();
1017       SmallVector<Metadata *, 1> Elts;
1018       Elts.push_back(
1019           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
1020       MDNode *Node = MDNode::get(C, Elts);
1021 
1022       Value *Arg0 = CI->getArgOperand(0);
1023       Value *Arg1 = CI->getArgOperand(1);
1024 
1025       // Nontemporal (unaligned) store of the 0'th element of the float/double
1026       // vector.
1027       Type *SrcEltTy = cast<VectorType>(Arg1->getType())->getElementType();
1028       PointerType *EltPtrTy = PointerType::getUnqual(SrcEltTy);
1029       Value *Addr = Builder.CreateBitCast(Arg0, EltPtrTy, "cast");
1030       Value *Extract =
1031           Builder.CreateExtractElement(Arg1, (uint64_t)0, "extractelement");
1032 
1033       StoreInst *SI = Builder.CreateAlignedStore(Extract, Addr, 1);
1034       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
1035 
1036       // Remove intrinsic.
1037       CI->eraseFromParent();
1038       return;
1039     }
1040 
1041     if (IsX86 && (Name.startswith("avx.movnt.") ||
1042                   Name.startswith("avx512.storent."))) {
1043       Module *M = F->getParent();
1044       SmallVector<Metadata *, 1> Elts;
1045       Elts.push_back(
1046           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
1047       MDNode *Node = MDNode::get(C, Elts);
1048 
1049       Value *Arg0 = CI->getArgOperand(0);
1050       Value *Arg1 = CI->getArgOperand(1);
1051 
1052       // Convert the type of the pointer to a pointer to the stored type.
1053       Value *BC = Builder.CreateBitCast(Arg0,
1054                                         PointerType::getUnqual(Arg1->getType()),
1055                                         "cast");
1056       VectorType *VTy = cast<VectorType>(Arg1->getType());
1057       StoreInst *SI = Builder.CreateAlignedStore(Arg1, BC,
1058                                                  VTy->getBitWidth() / 8);
1059       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
1060 
1061       // Remove intrinsic.
1062       CI->eraseFromParent();
1063       return;
1064     }
1065 
1066     if (IsX86 && Name == "sse2.storel.dq") {
1067       Value *Arg0 = CI->getArgOperand(0);
1068       Value *Arg1 = CI->getArgOperand(1);
1069 
1070       Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
1071       Value *BC0 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
1072       Value *Elt = Builder.CreateExtractElement(BC0, (uint64_t)0);
1073       Value *BC = Builder.CreateBitCast(Arg0,
1074                                         PointerType::getUnqual(Elt->getType()),
1075                                         "cast");
1076       Builder.CreateAlignedStore(Elt, BC, 1);
1077 
1078       // Remove intrinsic.
1079       CI->eraseFromParent();
1080       return;
1081     }
1082 
1083     if (IsX86 && (Name.startswith("sse.storeu.") ||
1084                   Name.startswith("sse2.storeu.") ||
1085                   Name.startswith("avx.storeu."))) {
1086       Value *Arg0 = CI->getArgOperand(0);
1087       Value *Arg1 = CI->getArgOperand(1);
1088 
1089       Arg0 = Builder.CreateBitCast(Arg0,
1090                                    PointerType::getUnqual(Arg1->getType()),
1091                                    "cast");
1092       Builder.CreateAlignedStore(Arg1, Arg0, 1);
1093 
1094       // Remove intrinsic.
1095       CI->eraseFromParent();
1096       return;
1097     }
1098 
1099     if (IsX86 && (Name.startswith("avx512.mask.store"))) {
1100       // "avx512.mask.storeu." or "avx512.mask.store."
1101       bool Aligned = Name[17] != 'u'; // "avx512.mask.storeu".
1102       UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
1103                          CI->getArgOperand(2), Aligned);
1104 
1105       // Remove intrinsic.
1106       CI->eraseFromParent();
1107       return;
1108     }
1109 
1110     Value *Rep;
1111     // Upgrade packed integer vector compare intrinsics to compare instructions.
1112     if (IsX86 && (Name.startswith("sse2.pcmp") ||
1113                   Name.startswith("avx2.pcmp"))) {
1114       // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt."
1115       bool CmpEq = Name[9] == 'e';
1116       Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
1117                                CI->getArgOperand(0), CI->getArgOperand(1));
1118       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
1119     } else if (IsX86 && (Name.startswith("avx512.broadcastm"))) {
1120       Type *ExtTy = Type::getInt32Ty(C);
1121       if (CI->getOperand(0)->getType()->isIntegerTy(8))
1122         ExtTy = Type::getInt64Ty(C);
1123       unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() /
1124                          ExtTy->getPrimitiveSizeInBits();
1125       Rep = Builder.CreateZExt(CI->getArgOperand(0), ExtTy);
1126       Rep = Builder.CreateVectorSplat(NumElts, Rep);
1127     } else if (IsX86 && (Name.startswith("avx512.ptestm") ||
1128                          Name.startswith("avx512.ptestnm"))) {
1129       Value *Op0 = CI->getArgOperand(0);
1130       Value *Op1 = CI->getArgOperand(1);
1131       Value *Mask = CI->getArgOperand(2);
1132       Rep = Builder.CreateAnd(Op0, Op1);
1133       llvm::Type *Ty = Op0->getType();
1134       Value *Zero = llvm::Constant::getNullValue(Ty);
1135       ICmpInst::Predicate Pred =
1136         Name.startswith("avx512.ptestm") ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
1137       Rep = Builder.CreateICmp(Pred, Rep, Zero);
1138       unsigned NumElts = Op0->getType()->getVectorNumElements();
1139       Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, Mask, NumElts);
1140     } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))){
1141       unsigned NumElts =
1142           CI->getArgOperand(1)->getType()->getVectorNumElements();
1143       Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
1144       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1145                           CI->getArgOperand(1));
1146     } else if (IsX86 && (Name.startswith("avx512.kunpck"))) {
1147       unsigned NumElts = CI->getType()->getScalarSizeInBits();
1148       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), NumElts);
1149       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), NumElts);
1150       uint32_t Indices[64];
1151       for (unsigned i = 0; i != NumElts; ++i)
1152         Indices[i] = i;
1153 
1154       // First extract half of each vector. This gives better codegen than
1155       // doing it in a single shuffle.
1156       LHS = Builder.CreateShuffleVector(LHS, LHS,
1157                                         makeArrayRef(Indices, NumElts / 2));
1158       RHS = Builder.CreateShuffleVector(RHS, RHS,
1159                                         makeArrayRef(Indices, NumElts / 2));
1160       // Concat the vectors.
1161       // NOTE: Operands have to be swapped to match intrinsic definition.
1162       Rep = Builder.CreateShuffleVector(RHS, LHS,
1163                                         makeArrayRef(Indices, NumElts));
1164       Rep = Builder.CreateBitCast(Rep, CI->getType());
1165     } else if (IsX86 && Name == "avx512.kand.w") {
1166       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1167       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1168       Rep = Builder.CreateAnd(LHS, RHS);
1169       Rep = Builder.CreateBitCast(Rep, CI->getType());
1170     } else if (IsX86 && Name == "avx512.kandn.w") {
1171       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1172       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1173       LHS = Builder.CreateNot(LHS);
1174       Rep = Builder.CreateAnd(LHS, RHS);
1175       Rep = Builder.CreateBitCast(Rep, CI->getType());
1176     } else if (IsX86 && Name == "avx512.kor.w") {
1177       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1178       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1179       Rep = Builder.CreateOr(LHS, RHS);
1180       Rep = Builder.CreateBitCast(Rep, CI->getType());
1181     } else if (IsX86 && Name == "avx512.kxor.w") {
1182       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1183       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1184       Rep = Builder.CreateXor(LHS, RHS);
1185       Rep = Builder.CreateBitCast(Rep, CI->getType());
1186     } else if (IsX86 && Name == "avx512.kxnor.w") {
1187       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1188       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1189       LHS = Builder.CreateNot(LHS);
1190       Rep = Builder.CreateXor(LHS, RHS);
1191       Rep = Builder.CreateBitCast(Rep, CI->getType());
1192     } else if (IsX86 && Name == "avx512.knot.w") {
1193       Rep = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1194       Rep = Builder.CreateNot(Rep);
1195       Rep = Builder.CreateBitCast(Rep, CI->getType());
1196     } else if (IsX86 &&
1197                (Name == "avx512.kortestz.w" || Name == "avx512.kortestc.w")) {
1198       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1199       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1200       Rep = Builder.CreateOr(LHS, RHS);
1201       Rep = Builder.CreateBitCast(Rep, Builder.getInt16Ty());
1202       Value *C;
1203       if (Name[14] == 'c')
1204         C = ConstantInt::getAllOnesValue(Builder.getInt16Ty());
1205       else
1206         C = ConstantInt::getNullValue(Builder.getInt16Ty());
1207       Rep = Builder.CreateICmpEQ(Rep, C);
1208       Rep = Builder.CreateZExt(Rep, Builder.getInt32Ty());
1209     } else if (IsX86 && (Name == "sse.add.ss" || Name == "sse2.add.sd")) {
1210       Type *I32Ty = Type::getInt32Ty(C);
1211       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1212                                                  ConstantInt::get(I32Ty, 0));
1213       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1214                                                  ConstantInt::get(I32Ty, 0));
1215       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1216                                         Builder.CreateFAdd(Elt0, Elt1),
1217                                         ConstantInt::get(I32Ty, 0));
1218     } else if (IsX86 && (Name == "sse.sub.ss" || Name == "sse2.sub.sd")) {
1219       Type *I32Ty = Type::getInt32Ty(C);
1220       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1221                                                  ConstantInt::get(I32Ty, 0));
1222       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1223                                                  ConstantInt::get(I32Ty, 0));
1224       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1225                                         Builder.CreateFSub(Elt0, Elt1),
1226                                         ConstantInt::get(I32Ty, 0));
1227     } else if (IsX86 && (Name == "sse.mul.ss" || Name == "sse2.mul.sd")) {
1228       Type *I32Ty = Type::getInt32Ty(C);
1229       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1230                                                  ConstantInt::get(I32Ty, 0));
1231       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1232                                                  ConstantInt::get(I32Ty, 0));
1233       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1234                                         Builder.CreateFMul(Elt0, Elt1),
1235                                         ConstantInt::get(I32Ty, 0));
1236     } else if (IsX86 && (Name == "sse.div.ss" || Name == "sse2.div.sd")) {
1237       Type *I32Ty = Type::getInt32Ty(C);
1238       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1239                                                  ConstantInt::get(I32Ty, 0));
1240       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1241                                                  ConstantInt::get(I32Ty, 0));
1242       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1243                                         Builder.CreateFDiv(Elt0, Elt1),
1244                                         ConstantInt::get(I32Ty, 0));
1245     } else if (IsX86 && Name.startswith("avx512.mask.pcmp")) {
1246       // "avx512.mask.pcmpeq." or "avx512.mask.pcmpgt."
1247       bool CmpEq = Name[16] == 'e';
1248       Rep = upgradeMaskedCompare(Builder, *CI, CmpEq ? 0 : 6, true);
1249     } else if (IsX86 && Name.startswith("avx512.mask.cmp")) {
1250       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1251       Rep = upgradeMaskedCompare(Builder, *CI, Imm, true);
1252     } else if (IsX86 && Name.startswith("avx512.mask.ucmp")) {
1253       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1254       Rep = upgradeMaskedCompare(Builder, *CI, Imm, false);
1255     } else if (IsX86 && (Name.startswith("avx512.cvtb2mask.") ||
1256                          Name.startswith("avx512.cvtw2mask.") ||
1257                          Name.startswith("avx512.cvtd2mask.") ||
1258                          Name.startswith("avx512.cvtq2mask."))) {
1259       Value *Op = CI->getArgOperand(0);
1260       Value *Zero = llvm::Constant::getNullValue(Op->getType());
1261       Rep = Builder.CreateICmp(ICmpInst::ICMP_SLT, Op, Zero);
1262       Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, nullptr,
1263                                    Op->getType()->getVectorNumElements());
1264     } else if(IsX86 && (Name == "ssse3.pabs.b.128" ||
1265                         Name == "ssse3.pabs.w.128" ||
1266                         Name == "ssse3.pabs.d.128" ||
1267                         Name.startswith("avx2.pabs") ||
1268                         Name.startswith("avx512.mask.pabs"))) {
1269       Rep = upgradeAbs(Builder, *CI);
1270     } else if (IsX86 && (Name == "sse41.pmaxsb" ||
1271                          Name == "sse2.pmaxs.w" ||
1272                          Name == "sse41.pmaxsd" ||
1273                          Name.startswith("avx2.pmaxs") ||
1274                          Name.startswith("avx512.mask.pmaxs"))) {
1275       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SGT);
1276     } else if (IsX86 && (Name == "sse2.pmaxu.b" ||
1277                          Name == "sse41.pmaxuw" ||
1278                          Name == "sse41.pmaxud" ||
1279                          Name.startswith("avx2.pmaxu") ||
1280                          Name.startswith("avx512.mask.pmaxu"))) {
1281       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_UGT);
1282     } else if (IsX86 && (Name == "sse41.pminsb" ||
1283                          Name == "sse2.pmins.w" ||
1284                          Name == "sse41.pminsd" ||
1285                          Name.startswith("avx2.pmins") ||
1286                          Name.startswith("avx512.mask.pmins"))) {
1287       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SLT);
1288     } else if (IsX86 && (Name == "sse2.pminu.b" ||
1289                          Name == "sse41.pminuw" ||
1290                          Name == "sse41.pminud" ||
1291                          Name.startswith("avx2.pminu") ||
1292                          Name.startswith("avx512.mask.pminu"))) {
1293       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_ULT);
1294     } else if (IsX86 && (Name == "sse2.cvtdq2pd" ||
1295                          Name == "sse2.cvtps2pd" ||
1296                          Name == "avx.cvtdq2.pd.256" ||
1297                          Name == "avx.cvt.ps2.pd.256" ||
1298                          Name.startswith("avx512.mask.cvtdq2pd.") ||
1299                          Name.startswith("avx512.mask.cvtudq2pd."))) {
1300       // Lossless i32/float to double conversion.
1301       // Extract the bottom elements if necessary and convert to double vector.
1302       Value *Src = CI->getArgOperand(0);
1303       VectorType *SrcTy = cast<VectorType>(Src->getType());
1304       VectorType *DstTy = cast<VectorType>(CI->getType());
1305       Rep = CI->getArgOperand(0);
1306 
1307       unsigned NumDstElts = DstTy->getNumElements();
1308       if (NumDstElts < SrcTy->getNumElements()) {
1309         assert(NumDstElts == 2 && "Unexpected vector size");
1310         uint32_t ShuffleMask[2] = { 0, 1 };
1311         Rep = Builder.CreateShuffleVector(Rep, UndefValue::get(SrcTy),
1312                                           ShuffleMask);
1313       }
1314 
1315       bool SInt2Double = (StringRef::npos != Name.find("cvtdq2"));
1316       bool UInt2Double = (StringRef::npos != Name.find("cvtudq2"));
1317       if (SInt2Double)
1318         Rep = Builder.CreateSIToFP(Rep, DstTy, "cvtdq2pd");
1319       else if (UInt2Double)
1320         Rep = Builder.CreateUIToFP(Rep, DstTy, "cvtudq2pd");
1321       else
1322         Rep = Builder.CreateFPExt(Rep, DstTy, "cvtps2pd");
1323 
1324       if (CI->getNumArgOperands() == 3)
1325         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1326                             CI->getArgOperand(1));
1327     } else if (IsX86 && (Name.startswith("avx512.mask.loadu."))) {
1328       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
1329                               CI->getArgOperand(1), CI->getArgOperand(2),
1330                               /*Aligned*/false);
1331     } else if (IsX86 && (Name.startswith("avx512.mask.load."))) {
1332       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
1333                               CI->getArgOperand(1),CI->getArgOperand(2),
1334                               /*Aligned*/true);
1335     } else if (IsX86 && Name.startswith("xop.vpcom")) {
1336       Intrinsic::ID intID;
1337       if (Name.endswith("ub"))
1338         intID = Intrinsic::x86_xop_vpcomub;
1339       else if (Name.endswith("uw"))
1340         intID = Intrinsic::x86_xop_vpcomuw;
1341       else if (Name.endswith("ud"))
1342         intID = Intrinsic::x86_xop_vpcomud;
1343       else if (Name.endswith("uq"))
1344         intID = Intrinsic::x86_xop_vpcomuq;
1345       else if (Name.endswith("b"))
1346         intID = Intrinsic::x86_xop_vpcomb;
1347       else if (Name.endswith("w"))
1348         intID = Intrinsic::x86_xop_vpcomw;
1349       else if (Name.endswith("d"))
1350         intID = Intrinsic::x86_xop_vpcomd;
1351       else if (Name.endswith("q"))
1352         intID = Intrinsic::x86_xop_vpcomq;
1353       else
1354         llvm_unreachable("Unknown suffix");
1355 
1356       Name = Name.substr(9); // strip off "xop.vpcom"
1357       unsigned Imm;
1358       if (Name.startswith("lt"))
1359         Imm = 0;
1360       else if (Name.startswith("le"))
1361         Imm = 1;
1362       else if (Name.startswith("gt"))
1363         Imm = 2;
1364       else if (Name.startswith("ge"))
1365         Imm = 3;
1366       else if (Name.startswith("eq"))
1367         Imm = 4;
1368       else if (Name.startswith("ne"))
1369         Imm = 5;
1370       else if (Name.startswith("false"))
1371         Imm = 6;
1372       else if (Name.startswith("true"))
1373         Imm = 7;
1374       else
1375         llvm_unreachable("Unknown condition");
1376 
1377       Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
1378       Rep =
1379           Builder.CreateCall(VPCOM, {CI->getArgOperand(0), CI->getArgOperand(1),
1380                                      Builder.getInt8(Imm)});
1381     } else if (IsX86 && Name.startswith("xop.vpcmov")) {
1382       Value *Sel = CI->getArgOperand(2);
1383       Value *NotSel = Builder.CreateNot(Sel);
1384       Value *Sel0 = Builder.CreateAnd(CI->getArgOperand(0), Sel);
1385       Value *Sel1 = Builder.CreateAnd(CI->getArgOperand(1), NotSel);
1386       Rep = Builder.CreateOr(Sel0, Sel1);
1387     } else if (IsX86 && Name == "sse42.crc32.64.8") {
1388       Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
1389                                                Intrinsic::x86_sse42_crc32_32_8);
1390       Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
1391       Rep = Builder.CreateCall(CRC32, {Trunc0, CI->getArgOperand(1)});
1392       Rep = Builder.CreateZExt(Rep, CI->getType(), "");
1393     } else if (IsX86 && Name.startswith("avx.vbroadcast.s")) {
1394       // Replace broadcasts with a series of insertelements.
1395       Type *VecTy = CI->getType();
1396       Type *EltTy = VecTy->getVectorElementType();
1397       unsigned EltNum = VecTy->getVectorNumElements();
1398       Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
1399                                           EltTy->getPointerTo());
1400       Value *Load = Builder.CreateLoad(EltTy, Cast);
1401       Type *I32Ty = Type::getInt32Ty(C);
1402       Rep = UndefValue::get(VecTy);
1403       for (unsigned I = 0; I < EltNum; ++I)
1404         Rep = Builder.CreateInsertElement(Rep, Load,
1405                                           ConstantInt::get(I32Ty, I));
1406     } else if (IsX86 && (Name.startswith("sse41.pmovsx") ||
1407                          Name.startswith("sse41.pmovzx") ||
1408                          Name.startswith("avx2.pmovsx") ||
1409                          Name.startswith("avx2.pmovzx") ||
1410                          Name.startswith("avx512.mask.pmovsx") ||
1411                          Name.startswith("avx512.mask.pmovzx"))) {
1412       VectorType *SrcTy = cast<VectorType>(CI->getArgOperand(0)->getType());
1413       VectorType *DstTy = cast<VectorType>(CI->getType());
1414       unsigned NumDstElts = DstTy->getNumElements();
1415 
1416       // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
1417       SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
1418       for (unsigned i = 0; i != NumDstElts; ++i)
1419         ShuffleMask[i] = i;
1420 
1421       Value *SV = Builder.CreateShuffleVector(
1422           CI->getArgOperand(0), UndefValue::get(SrcTy), ShuffleMask);
1423 
1424       bool DoSext = (StringRef::npos != Name.find("pmovsx"));
1425       Rep = DoSext ? Builder.CreateSExt(SV, DstTy)
1426                    : Builder.CreateZExt(SV, DstTy);
1427       // If there are 3 arguments, it's a masked intrinsic so we need a select.
1428       if (CI->getNumArgOperands() == 3)
1429         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1430                             CI->getArgOperand(1));
1431     } else if (IsX86 && (Name.startswith("avx.vbroadcastf128") ||
1432                          Name == "avx2.vbroadcasti128")) {
1433       // Replace vbroadcastf128/vbroadcasti128 with a vector load+shuffle.
1434       Type *EltTy = CI->getType()->getVectorElementType();
1435       unsigned NumSrcElts = 128 / EltTy->getPrimitiveSizeInBits();
1436       Type *VT = VectorType::get(EltTy, NumSrcElts);
1437       Value *Op = Builder.CreatePointerCast(CI->getArgOperand(0),
1438                                             PointerType::getUnqual(VT));
1439       Value *Load = Builder.CreateAlignedLoad(Op, 1);
1440       if (NumSrcElts == 2)
1441         Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
1442                                           { 0, 1, 0, 1 });
1443       else
1444         Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
1445                                           { 0, 1, 2, 3, 0, 1, 2, 3 });
1446     } else if (IsX86 && (Name.startswith("avx512.mask.shuf.i") ||
1447                          Name.startswith("avx512.mask.shuf.f"))) {
1448       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1449       Type *VT = CI->getType();
1450       unsigned NumLanes = VT->getPrimitiveSizeInBits() / 128;
1451       unsigned NumElementsInLane = 128 / VT->getScalarSizeInBits();
1452       unsigned ControlBitsMask = NumLanes - 1;
1453       unsigned NumControlBits = NumLanes / 2;
1454       SmallVector<uint32_t, 8> ShuffleMask(0);
1455 
1456       for (unsigned l = 0; l != NumLanes; ++l) {
1457         unsigned LaneMask = (Imm >> (l * NumControlBits)) & ControlBitsMask;
1458         // We actually need the other source.
1459         if (l >= NumLanes / 2)
1460           LaneMask += NumLanes;
1461         for (unsigned i = 0; i != NumElementsInLane; ++i)
1462           ShuffleMask.push_back(LaneMask * NumElementsInLane + i);
1463       }
1464       Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
1465                                         CI->getArgOperand(1), ShuffleMask);
1466       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1467                           CI->getArgOperand(3));
1468     }else if (IsX86 && (Name.startswith("avx512.mask.broadcastf") ||
1469                          Name.startswith("avx512.mask.broadcasti"))) {
1470       unsigned NumSrcElts =
1471                         CI->getArgOperand(0)->getType()->getVectorNumElements();
1472       unsigned NumDstElts = CI->getType()->getVectorNumElements();
1473 
1474       SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
1475       for (unsigned i = 0; i != NumDstElts; ++i)
1476         ShuffleMask[i] = i % NumSrcElts;
1477 
1478       Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
1479                                         CI->getArgOperand(0),
1480                                         ShuffleMask);
1481       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1482                           CI->getArgOperand(1));
1483     } else if (IsX86 && (Name.startswith("avx2.pbroadcast") ||
1484                          Name.startswith("avx2.vbroadcast") ||
1485                          Name.startswith("avx512.pbroadcast") ||
1486                          Name.startswith("avx512.mask.broadcast.s"))) {
1487       // Replace vp?broadcasts with a vector shuffle.
1488       Value *Op = CI->getArgOperand(0);
1489       unsigned NumElts = CI->getType()->getVectorNumElements();
1490       Type *MaskTy = VectorType::get(Type::getInt32Ty(C), NumElts);
1491       Rep = Builder.CreateShuffleVector(Op, UndefValue::get(Op->getType()),
1492                                         Constant::getNullValue(MaskTy));
1493 
1494       if (CI->getNumArgOperands() == 3)
1495         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1496                             CI->getArgOperand(1));
1497     } else if (IsX86 && Name.startswith("avx512.mask.palignr.")) {
1498       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
1499                                       CI->getArgOperand(1),
1500                                       CI->getArgOperand(2),
1501                                       CI->getArgOperand(3),
1502                                       CI->getArgOperand(4),
1503                                       false);
1504     } else if (IsX86 && Name.startswith("avx512.mask.valign.")) {
1505       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
1506                                       CI->getArgOperand(1),
1507                                       CI->getArgOperand(2),
1508                                       CI->getArgOperand(3),
1509                                       CI->getArgOperand(4),
1510                                       true);
1511     } else if (IsX86 && (Name == "sse2.psll.dq" ||
1512                          Name == "avx2.psll.dq")) {
1513       // 128/256-bit shift left specified in bits.
1514       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1515       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0),
1516                                        Shift / 8); // Shift is in bits.
1517     } else if (IsX86 && (Name == "sse2.psrl.dq" ||
1518                          Name == "avx2.psrl.dq")) {
1519       // 128/256-bit shift right specified in bits.
1520       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1521       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0),
1522                                        Shift / 8); // Shift is in bits.
1523     } else if (IsX86 && (Name == "sse2.psll.dq.bs" ||
1524                          Name == "avx2.psll.dq.bs" ||
1525                          Name == "avx512.psll.dq.512")) {
1526       // 128/256/512-bit shift left specified in bytes.
1527       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1528       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
1529     } else if (IsX86 && (Name == "sse2.psrl.dq.bs" ||
1530                          Name == "avx2.psrl.dq.bs" ||
1531                          Name == "avx512.psrl.dq.512")) {
1532       // 128/256/512-bit shift right specified in bytes.
1533       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1534       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
1535     } else if (IsX86 && (Name == "sse41.pblendw" ||
1536                          Name.startswith("sse41.blendp") ||
1537                          Name.startswith("avx.blend.p") ||
1538                          Name == "avx2.pblendw" ||
1539                          Name.startswith("avx2.pblendd."))) {
1540       Value *Op0 = CI->getArgOperand(0);
1541       Value *Op1 = CI->getArgOperand(1);
1542       unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1543       VectorType *VecTy = cast<VectorType>(CI->getType());
1544       unsigned NumElts = VecTy->getNumElements();
1545 
1546       SmallVector<uint32_t, 16> Idxs(NumElts);
1547       for (unsigned i = 0; i != NumElts; ++i)
1548         Idxs[i] = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
1549 
1550       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1551     } else if (IsX86 && (Name.startswith("avx.vinsertf128.") ||
1552                          Name == "avx2.vinserti128" ||
1553                          Name.startswith("avx512.mask.insert"))) {
1554       Value *Op0 = CI->getArgOperand(0);
1555       Value *Op1 = CI->getArgOperand(1);
1556       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1557       unsigned DstNumElts = CI->getType()->getVectorNumElements();
1558       unsigned SrcNumElts = Op1->getType()->getVectorNumElements();
1559       unsigned Scale = DstNumElts / SrcNumElts;
1560 
1561       // Mask off the high bits of the immediate value; hardware ignores those.
1562       Imm = Imm % Scale;
1563 
1564       // Extend the second operand into a vector the size of the destination.
1565       Value *UndefV = UndefValue::get(Op1->getType());
1566       SmallVector<uint32_t, 8> Idxs(DstNumElts);
1567       for (unsigned i = 0; i != SrcNumElts; ++i)
1568         Idxs[i] = i;
1569       for (unsigned i = SrcNumElts; i != DstNumElts; ++i)
1570         Idxs[i] = SrcNumElts;
1571       Rep = Builder.CreateShuffleVector(Op1, UndefV, Idxs);
1572 
1573       // Insert the second operand into the first operand.
1574 
1575       // Note that there is no guarantee that instruction lowering will actually
1576       // produce a vinsertf128 instruction for the created shuffles. In
1577       // particular, the 0 immediate case involves no lane changes, so it can
1578       // be handled as a blend.
1579 
1580       // Example of shuffle mask for 32-bit elements:
1581       // Imm = 1  <i32 0, i32 1, i32 2,  i32 3,  i32 8, i32 9, i32 10, i32 11>
1582       // Imm = 0  <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6,  i32 7 >
1583 
1584       // First fill with identify mask.
1585       for (unsigned i = 0; i != DstNumElts; ++i)
1586         Idxs[i] = i;
1587       // Then replace the elements where we need to insert.
1588       for (unsigned i = 0; i != SrcNumElts; ++i)
1589         Idxs[i + Imm * SrcNumElts] = i + DstNumElts;
1590       Rep = Builder.CreateShuffleVector(Op0, Rep, Idxs);
1591 
1592       // If the intrinsic has a mask operand, handle that.
1593       if (CI->getNumArgOperands() == 5)
1594         Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1595                             CI->getArgOperand(3));
1596     } else if (IsX86 && (Name.startswith("avx.vextractf128.") ||
1597                          Name == "avx2.vextracti128" ||
1598                          Name.startswith("avx512.mask.vextract"))) {
1599       Value *Op0 = CI->getArgOperand(0);
1600       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1601       unsigned DstNumElts = CI->getType()->getVectorNumElements();
1602       unsigned SrcNumElts = Op0->getType()->getVectorNumElements();
1603       unsigned Scale = SrcNumElts / DstNumElts;
1604 
1605       // Mask off the high bits of the immediate value; hardware ignores those.
1606       Imm = Imm % Scale;
1607 
1608       // Get indexes for the subvector of the input vector.
1609       SmallVector<uint32_t, 8> Idxs(DstNumElts);
1610       for (unsigned i = 0; i != DstNumElts; ++i) {
1611         Idxs[i] = i + (Imm * DstNumElts);
1612       }
1613       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1614 
1615       // If the intrinsic has a mask operand, handle that.
1616       if (CI->getNumArgOperands() == 4)
1617         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1618                             CI->getArgOperand(2));
1619     } else if (!IsX86 && Name == "stackprotectorcheck") {
1620       Rep = nullptr;
1621     } else if (IsX86 && (Name.startswith("avx512.mask.perm.df.") ||
1622                          Name.startswith("avx512.mask.perm.di."))) {
1623       Value *Op0 = CI->getArgOperand(0);
1624       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1625       VectorType *VecTy = cast<VectorType>(CI->getType());
1626       unsigned NumElts = VecTy->getNumElements();
1627 
1628       SmallVector<uint32_t, 8> Idxs(NumElts);
1629       for (unsigned i = 0; i != NumElts; ++i)
1630         Idxs[i] = (i & ~0x3) + ((Imm >> (2 * (i & 0x3))) & 3);
1631 
1632       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1633 
1634       if (CI->getNumArgOperands() == 4)
1635         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1636                             CI->getArgOperand(2));
1637     } else if (IsX86 && (Name.startswith("avx.vperm2f128.") ||
1638                          Name == "avx2.vperm2i128")) {
1639       // The immediate permute control byte looks like this:
1640       //    [1:0] - select 128 bits from sources for low half of destination
1641       //    [2]   - ignore
1642       //    [3]   - zero low half of destination
1643       //    [5:4] - select 128 bits from sources for high half of destination
1644       //    [6]   - ignore
1645       //    [7]   - zero high half of destination
1646 
1647       uint8_t Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1648 
1649       unsigned NumElts = CI->getType()->getVectorNumElements();
1650       unsigned HalfSize = NumElts / 2;
1651       SmallVector<uint32_t, 8> ShuffleMask(NumElts);
1652 
1653       // Determine which operand(s) are actually in use for this instruction.
1654       Value *V0 = (Imm & 0x02) ? CI->getArgOperand(1) : CI->getArgOperand(0);
1655       Value *V1 = (Imm & 0x20) ? CI->getArgOperand(1) : CI->getArgOperand(0);
1656 
1657       // If needed, replace operands based on zero mask.
1658       V0 = (Imm & 0x08) ? ConstantAggregateZero::get(CI->getType()) : V0;
1659       V1 = (Imm & 0x80) ? ConstantAggregateZero::get(CI->getType()) : V1;
1660 
1661       // Permute low half of result.
1662       unsigned StartIndex = (Imm & 0x01) ? HalfSize : 0;
1663       for (unsigned i = 0; i < HalfSize; ++i)
1664         ShuffleMask[i] = StartIndex + i;
1665 
1666       // Permute high half of result.
1667       StartIndex = (Imm & 0x10) ? HalfSize : 0;
1668       for (unsigned i = 0; i < HalfSize; ++i)
1669         ShuffleMask[i + HalfSize] = NumElts + StartIndex + i;
1670 
1671       Rep = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
1672 
1673     } else if (IsX86 && (Name.startswith("avx.vpermil.") ||
1674                          Name == "sse2.pshuf.d" ||
1675                          Name.startswith("avx512.mask.vpermil.p") ||
1676                          Name.startswith("avx512.mask.pshuf.d."))) {
1677       Value *Op0 = CI->getArgOperand(0);
1678       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1679       VectorType *VecTy = cast<VectorType>(CI->getType());
1680       unsigned NumElts = VecTy->getNumElements();
1681       // Calculate the size of each index in the immediate.
1682       unsigned IdxSize = 64 / VecTy->getScalarSizeInBits();
1683       unsigned IdxMask = ((1 << IdxSize) - 1);
1684 
1685       SmallVector<uint32_t, 8> Idxs(NumElts);
1686       // Lookup the bits for this element, wrapping around the immediate every
1687       // 8-bits. Elements are grouped into sets of 2 or 4 elements so we need
1688       // to offset by the first index of each group.
1689       for (unsigned i = 0; i != NumElts; ++i)
1690         Idxs[i] = ((Imm >> ((i * IdxSize) % 8)) & IdxMask) | (i & ~IdxMask);
1691 
1692       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1693 
1694       if (CI->getNumArgOperands() == 4)
1695         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1696                             CI->getArgOperand(2));
1697     } else if (IsX86 && (Name == "sse2.pshufl.w" ||
1698                          Name.startswith("avx512.mask.pshufl.w."))) {
1699       Value *Op0 = CI->getArgOperand(0);
1700       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1701       unsigned NumElts = CI->getType()->getVectorNumElements();
1702 
1703       SmallVector<uint32_t, 16> Idxs(NumElts);
1704       for (unsigned l = 0; l != NumElts; l += 8) {
1705         for (unsigned i = 0; i != 4; ++i)
1706           Idxs[i + l] = ((Imm >> (2 * i)) & 0x3) + l;
1707         for (unsigned i = 4; i != 8; ++i)
1708           Idxs[i + l] = i + l;
1709       }
1710 
1711       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1712 
1713       if (CI->getNumArgOperands() == 4)
1714         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1715                             CI->getArgOperand(2));
1716     } else if (IsX86 && (Name == "sse2.pshufh.w" ||
1717                          Name.startswith("avx512.mask.pshufh.w."))) {
1718       Value *Op0 = CI->getArgOperand(0);
1719       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1720       unsigned NumElts = CI->getType()->getVectorNumElements();
1721 
1722       SmallVector<uint32_t, 16> Idxs(NumElts);
1723       for (unsigned l = 0; l != NumElts; l += 8) {
1724         for (unsigned i = 0; i != 4; ++i)
1725           Idxs[i + l] = i + l;
1726         for (unsigned i = 0; i != 4; ++i)
1727           Idxs[i + l + 4] = ((Imm >> (2 * i)) & 0x3) + 4 + l;
1728       }
1729 
1730       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1731 
1732       if (CI->getNumArgOperands() == 4)
1733         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1734                             CI->getArgOperand(2));
1735     } else if (IsX86 && Name.startswith("avx512.mask.shuf.p")) {
1736       Value *Op0 = CI->getArgOperand(0);
1737       Value *Op1 = CI->getArgOperand(1);
1738       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1739       unsigned NumElts = CI->getType()->getVectorNumElements();
1740 
1741       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1742       unsigned HalfLaneElts = NumLaneElts / 2;
1743 
1744       SmallVector<uint32_t, 16> Idxs(NumElts);
1745       for (unsigned i = 0; i != NumElts; ++i) {
1746         // Base index is the starting element of the lane.
1747         Idxs[i] = i - (i % NumLaneElts);
1748         // If we are half way through the lane switch to the other source.
1749         if ((i % NumLaneElts) >= HalfLaneElts)
1750           Idxs[i] += NumElts;
1751         // Now select the specific element. By adding HalfLaneElts bits from
1752         // the immediate. Wrapping around the immediate every 8-bits.
1753         Idxs[i] += (Imm >> ((i * HalfLaneElts) % 8)) & ((1 << HalfLaneElts) - 1);
1754       }
1755 
1756       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1757 
1758       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1759                           CI->getArgOperand(3));
1760     } else if (IsX86 && (Name.startswith("avx512.mask.movddup") ||
1761                          Name.startswith("avx512.mask.movshdup") ||
1762                          Name.startswith("avx512.mask.movsldup"))) {
1763       Value *Op0 = CI->getArgOperand(0);
1764       unsigned NumElts = CI->getType()->getVectorNumElements();
1765       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1766 
1767       unsigned Offset = 0;
1768       if (Name.startswith("avx512.mask.movshdup."))
1769         Offset = 1;
1770 
1771       SmallVector<uint32_t, 16> Idxs(NumElts);
1772       for (unsigned l = 0; l != NumElts; l += NumLaneElts)
1773         for (unsigned i = 0; i != NumLaneElts; i += 2) {
1774           Idxs[i + l + 0] = i + l + Offset;
1775           Idxs[i + l + 1] = i + l + Offset;
1776         }
1777 
1778       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1779 
1780       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1781                           CI->getArgOperand(1));
1782     } else if (IsX86 && (Name.startswith("avx512.mask.punpckl") ||
1783                          Name.startswith("avx512.mask.unpckl."))) {
1784       Value *Op0 = CI->getArgOperand(0);
1785       Value *Op1 = CI->getArgOperand(1);
1786       int NumElts = CI->getType()->getVectorNumElements();
1787       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1788 
1789       SmallVector<uint32_t, 64> Idxs(NumElts);
1790       for (int l = 0; l != NumElts; l += NumLaneElts)
1791         for (int i = 0; i != NumLaneElts; ++i)
1792           Idxs[i + l] = l + (i / 2) + NumElts * (i % 2);
1793 
1794       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1795 
1796       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1797                           CI->getArgOperand(2));
1798     } else if (IsX86 && (Name.startswith("avx512.mask.punpckh") ||
1799                          Name.startswith("avx512.mask.unpckh."))) {
1800       Value *Op0 = CI->getArgOperand(0);
1801       Value *Op1 = CI->getArgOperand(1);
1802       int NumElts = CI->getType()->getVectorNumElements();
1803       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1804 
1805       SmallVector<uint32_t, 64> Idxs(NumElts);
1806       for (int l = 0; l != NumElts; l += NumLaneElts)
1807         for (int i = 0; i != NumLaneElts; ++i)
1808           Idxs[i + l] = (NumLaneElts / 2) + l + (i / 2) + NumElts * (i % 2);
1809 
1810       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1811 
1812       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1813                           CI->getArgOperand(2));
1814     } else if (IsX86 && Name.startswith("avx512.mask.pand.")) {
1815       Rep = Builder.CreateAnd(CI->getArgOperand(0), CI->getArgOperand(1));
1816       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1817                           CI->getArgOperand(2));
1818     } else if (IsX86 && Name.startswith("avx512.mask.pandn.")) {
1819       Rep = Builder.CreateAnd(Builder.CreateNot(CI->getArgOperand(0)),
1820                               CI->getArgOperand(1));
1821       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1822                           CI->getArgOperand(2));
1823     } else if (IsX86 && Name.startswith("avx512.mask.por.")) {
1824       Rep = Builder.CreateOr(CI->getArgOperand(0), CI->getArgOperand(1));
1825       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1826                           CI->getArgOperand(2));
1827     } else if (IsX86 && Name.startswith("avx512.mask.pxor.")) {
1828       Rep = Builder.CreateXor(CI->getArgOperand(0), CI->getArgOperand(1));
1829       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1830                           CI->getArgOperand(2));
1831     } else if (IsX86 && Name.startswith("avx512.mask.and.")) {
1832       VectorType *FTy = cast<VectorType>(CI->getType());
1833       VectorType *ITy = VectorType::getInteger(FTy);
1834       Rep = Builder.CreateAnd(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1835                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1836       Rep = Builder.CreateBitCast(Rep, FTy);
1837       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1838                           CI->getArgOperand(2));
1839     } else if (IsX86 && Name.startswith("avx512.mask.andn.")) {
1840       VectorType *FTy = cast<VectorType>(CI->getType());
1841       VectorType *ITy = VectorType::getInteger(FTy);
1842       Rep = Builder.CreateNot(Builder.CreateBitCast(CI->getArgOperand(0), ITy));
1843       Rep = Builder.CreateAnd(Rep,
1844                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1845       Rep = Builder.CreateBitCast(Rep, FTy);
1846       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1847                           CI->getArgOperand(2));
1848     } else if (IsX86 && Name.startswith("avx512.mask.or.")) {
1849       VectorType *FTy = cast<VectorType>(CI->getType());
1850       VectorType *ITy = VectorType::getInteger(FTy);
1851       Rep = Builder.CreateOr(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1852                              Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1853       Rep = Builder.CreateBitCast(Rep, FTy);
1854       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1855                           CI->getArgOperand(2));
1856     } else if (IsX86 && Name.startswith("avx512.mask.xor.")) {
1857       VectorType *FTy = cast<VectorType>(CI->getType());
1858       VectorType *ITy = VectorType::getInteger(FTy);
1859       Rep = Builder.CreateXor(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1860                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1861       Rep = Builder.CreateBitCast(Rep, FTy);
1862       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1863                           CI->getArgOperand(2));
1864     } else if (IsX86 && Name.startswith("avx512.mask.padd.")) {
1865       Rep = Builder.CreateAdd(CI->getArgOperand(0), CI->getArgOperand(1));
1866       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1867                           CI->getArgOperand(2));
1868     } else if (IsX86 && Name.startswith("avx512.mask.psub.")) {
1869       Rep = Builder.CreateSub(CI->getArgOperand(0), CI->getArgOperand(1));
1870       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1871                           CI->getArgOperand(2));
1872     } else if (IsX86 && Name.startswith("avx512.mask.pmull.")) {
1873       Rep = Builder.CreateMul(CI->getArgOperand(0), CI->getArgOperand(1));
1874       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1875                           CI->getArgOperand(2));
1876     } else if (IsX86 && (Name.startswith("avx512.mask.add.p"))) {
1877       Rep = Builder.CreateFAdd(CI->getArgOperand(0), CI->getArgOperand(1));
1878       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1879                           CI->getArgOperand(2));
1880     } else if (IsX86 && Name.startswith("avx512.mask.div.p")) {
1881       Rep = Builder.CreateFDiv(CI->getArgOperand(0), CI->getArgOperand(1));
1882       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1883                           CI->getArgOperand(2));
1884     } else if (IsX86 && Name.startswith("avx512.mask.mul.p")) {
1885       Rep = Builder.CreateFMul(CI->getArgOperand(0), CI->getArgOperand(1));
1886       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1887                           CI->getArgOperand(2));
1888     } else if (IsX86 && Name.startswith("avx512.mask.sub.p")) {
1889       Rep = Builder.CreateFSub(CI->getArgOperand(0), CI->getArgOperand(1));
1890       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1891                           CI->getArgOperand(2));
1892     } else if (IsX86 && Name.startswith("avx512.mask.lzcnt.")) {
1893       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
1894                                                          Intrinsic::ctlz,
1895                                                          CI->getType()),
1896                                { CI->getArgOperand(0), Builder.getInt1(false) });
1897       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1898                           CI->getArgOperand(1));
1899     } else if (IsX86 && (Name.startswith("avx512.mask.max.p") ||
1900                          Name.startswith("avx512.mask.min.p"))) {
1901       bool IsMin = Name[13] == 'i';
1902       VectorType *VecTy = cast<VectorType>(CI->getType());
1903       unsigned VecWidth = VecTy->getPrimitiveSizeInBits();
1904       unsigned EltWidth = VecTy->getScalarSizeInBits();
1905       Intrinsic::ID IID;
1906       if (!IsMin && VecWidth == 128 && EltWidth == 32)
1907         IID = Intrinsic::x86_sse_max_ps;
1908       else if (!IsMin && VecWidth == 128 && EltWidth == 64)
1909         IID = Intrinsic::x86_sse2_max_pd;
1910       else if (!IsMin && VecWidth == 256 && EltWidth == 32)
1911         IID = Intrinsic::x86_avx_max_ps_256;
1912       else if (!IsMin && VecWidth == 256 && EltWidth == 64)
1913         IID = Intrinsic::x86_avx_max_pd_256;
1914       else if (IsMin && VecWidth == 128 && EltWidth == 32)
1915         IID = Intrinsic::x86_sse_min_ps;
1916       else if (IsMin && VecWidth == 128 && EltWidth == 64)
1917         IID = Intrinsic::x86_sse2_min_pd;
1918       else if (IsMin && VecWidth == 256 && EltWidth == 32)
1919         IID = Intrinsic::x86_avx_min_ps_256;
1920       else if (IsMin && VecWidth == 256 && EltWidth == 64)
1921         IID = Intrinsic::x86_avx_min_pd_256;
1922       else
1923         llvm_unreachable("Unexpected intrinsic");
1924 
1925       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1926                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1927       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1928                           CI->getArgOperand(2));
1929     } else if (IsX86 && Name.startswith("avx512.mask.pshuf.b.")) {
1930       VectorType *VecTy = cast<VectorType>(CI->getType());
1931       Intrinsic::ID IID;
1932       if (VecTy->getPrimitiveSizeInBits() == 128)
1933         IID = Intrinsic::x86_ssse3_pshuf_b_128;
1934       else if (VecTy->getPrimitiveSizeInBits() == 256)
1935         IID = Intrinsic::x86_avx2_pshuf_b;
1936       else if (VecTy->getPrimitiveSizeInBits() == 512)
1937         IID = Intrinsic::x86_avx512_pshuf_b_512;
1938       else
1939         llvm_unreachable("Unexpected intrinsic");
1940 
1941       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1942                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1943       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1944                           CI->getArgOperand(2));
1945     } else if (IsX86 && (Name.startswith("avx512.mask.pmul.dq.") ||
1946                          Name.startswith("avx512.mask.pmulu.dq."))) {
1947       bool IsUnsigned = Name[16] == 'u';
1948       VectorType *VecTy = cast<VectorType>(CI->getType());
1949       Intrinsic::ID IID;
1950       if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1951         IID = Intrinsic::x86_sse41_pmuldq;
1952       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1953         IID = Intrinsic::x86_avx2_pmul_dq;
1954       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1955         IID = Intrinsic::x86_avx512_pmul_dq_512;
1956       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1957         IID = Intrinsic::x86_sse2_pmulu_dq;
1958       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1959         IID = Intrinsic::x86_avx2_pmulu_dq;
1960       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1961         IID = Intrinsic::x86_avx512_pmulu_dq_512;
1962       else
1963         llvm_unreachable("Unexpected intrinsic");
1964 
1965       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1966                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1967       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1968                           CI->getArgOperand(2));
1969     } else if (IsX86 && Name.startswith("avx512.mask.pmul.hr.sw.")) {
1970       VectorType *VecTy = cast<VectorType>(CI->getType());
1971       Intrinsic::ID IID;
1972       if (VecTy->getPrimitiveSizeInBits() == 128)
1973         IID = Intrinsic::x86_ssse3_pmul_hr_sw_128;
1974       else if (VecTy->getPrimitiveSizeInBits() == 256)
1975         IID = Intrinsic::x86_avx2_pmul_hr_sw;
1976       else if (VecTy->getPrimitiveSizeInBits() == 512)
1977         IID = Intrinsic::x86_avx512_pmul_hr_sw_512;
1978       else
1979         llvm_unreachable("Unexpected intrinsic");
1980 
1981       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1982                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1983       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1984                           CI->getArgOperand(2));
1985     } else if (IsX86 && (Name.startswith("avx512.mask.pmulh.w.") ||
1986                          Name.startswith("avx512.mask.pmulhu.w."))) {
1987       bool IsUnsigned = Name[17] == 'u';
1988       VectorType *VecTy = cast<VectorType>(CI->getType());
1989       Intrinsic::ID IID;
1990       if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1991         IID = Intrinsic::x86_sse2_pmulh_w;
1992       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1993         IID = Intrinsic::x86_avx2_pmulh_w;
1994       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1995         IID = Intrinsic::x86_avx512_pmulh_w_512;
1996       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1997         IID = Intrinsic::x86_sse2_pmulhu_w;
1998       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1999         IID = Intrinsic::x86_avx2_pmulhu_w;
2000       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
2001         IID = Intrinsic::x86_avx512_pmulhu_w_512;
2002       else
2003         llvm_unreachable("Unexpected intrinsic");
2004 
2005       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2006                                { CI->getArgOperand(0), CI->getArgOperand(1) });
2007       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2008                           CI->getArgOperand(2));
2009     } else if (IsX86 && Name.startswith("avx512.mask.pack")) {
2010       bool IsUnsigned = Name[16] == 'u';
2011       bool IsDW = Name[18] == 'd';
2012       VectorType *VecTy = cast<VectorType>(CI->getType());
2013       Intrinsic::ID IID;
2014       if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 128)
2015         IID = Intrinsic::x86_sse2_packsswb_128;
2016       else if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 256)
2017         IID = Intrinsic::x86_avx2_packsswb;
2018       else if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 512)
2019         IID = Intrinsic::x86_avx512_packsswb_512;
2020       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 128)
2021         IID = Intrinsic::x86_sse2_packssdw_128;
2022       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 256)
2023         IID = Intrinsic::x86_avx2_packssdw;
2024       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 512)
2025         IID = Intrinsic::x86_avx512_packssdw_512;
2026       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 128)
2027         IID = Intrinsic::x86_sse2_packuswb_128;
2028       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 256)
2029         IID = Intrinsic::x86_avx2_packuswb;
2030       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 512)
2031         IID = Intrinsic::x86_avx512_packuswb_512;
2032       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 128)
2033         IID = Intrinsic::x86_sse41_packusdw;
2034       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 256)
2035         IID = Intrinsic::x86_avx2_packusdw;
2036       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 512)
2037         IID = Intrinsic::x86_avx512_packusdw_512;
2038       else
2039         llvm_unreachable("Unexpected intrinsic");
2040 
2041       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2042                                { CI->getArgOperand(0), CI->getArgOperand(1) });
2043       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2044                           CI->getArgOperand(2));
2045     } else if (IsX86 && Name.startswith("avx512.mask.psll")) {
2046       bool IsImmediate = Name[16] == 'i' ||
2047                          (Name.size() > 18 && Name[18] == 'i');
2048       bool IsVariable = Name[16] == 'v';
2049       char Size = Name[16] == '.' ? Name[17] :
2050                   Name[17] == '.' ? Name[18] :
2051                   Name[18] == '.' ? Name[19] :
2052                                     Name[20];
2053 
2054       Intrinsic::ID IID;
2055       if (IsVariable && Name[17] != '.') {
2056         if (Size == 'd' && Name[17] == '2') // avx512.mask.psllv2.di
2057           IID = Intrinsic::x86_avx2_psllv_q;
2058         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psllv4.di
2059           IID = Intrinsic::x86_avx2_psllv_q_256;
2060         else if (Size == 's' && Name[17] == '4') // avx512.mask.psllv4.si
2061           IID = Intrinsic::x86_avx2_psllv_d;
2062         else if (Size == 's' && Name[17] == '8') // avx512.mask.psllv8.si
2063           IID = Intrinsic::x86_avx2_psllv_d_256;
2064         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psllv8.hi
2065           IID = Intrinsic::x86_avx512_psllv_w_128;
2066         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psllv16.hi
2067           IID = Intrinsic::x86_avx512_psllv_w_256;
2068         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psllv32hi
2069           IID = Intrinsic::x86_avx512_psllv_w_512;
2070         else
2071           llvm_unreachable("Unexpected size");
2072       } else if (Name.endswith(".128")) {
2073         if (Size == 'd') // avx512.mask.psll.d.128, avx512.mask.psll.di.128
2074           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_d
2075                             : Intrinsic::x86_sse2_psll_d;
2076         else if (Size == 'q') // avx512.mask.psll.q.128, avx512.mask.psll.qi.128
2077           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_q
2078                             : Intrinsic::x86_sse2_psll_q;
2079         else if (Size == 'w') // avx512.mask.psll.w.128, avx512.mask.psll.wi.128
2080           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_w
2081                             : Intrinsic::x86_sse2_psll_w;
2082         else
2083           llvm_unreachable("Unexpected size");
2084       } else if (Name.endswith(".256")) {
2085         if (Size == 'd') // avx512.mask.psll.d.256, avx512.mask.psll.di.256
2086           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_d
2087                             : Intrinsic::x86_avx2_psll_d;
2088         else if (Size == 'q') // avx512.mask.psll.q.256, avx512.mask.psll.qi.256
2089           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_q
2090                             : Intrinsic::x86_avx2_psll_q;
2091         else if (Size == 'w') // avx512.mask.psll.w.256, avx512.mask.psll.wi.256
2092           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_w
2093                             : Intrinsic::x86_avx2_psll_w;
2094         else
2095           llvm_unreachable("Unexpected size");
2096       } else {
2097         if (Size == 'd') // psll.di.512, pslli.d, psll.d, psllv.d.512
2098           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_d_512 :
2099                 IsVariable  ? Intrinsic::x86_avx512_psllv_d_512 :
2100                               Intrinsic::x86_avx512_psll_d_512;
2101         else if (Size == 'q') // psll.qi.512, pslli.q, psll.q, psllv.q.512
2102           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_q_512 :
2103                 IsVariable  ? Intrinsic::x86_avx512_psllv_q_512 :
2104                               Intrinsic::x86_avx512_psll_q_512;
2105         else if (Size == 'w') // psll.wi.512, pslli.w, psll.w
2106           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_w_512
2107                             : Intrinsic::x86_avx512_psll_w_512;
2108         else
2109           llvm_unreachable("Unexpected size");
2110       }
2111 
2112       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2113     } else if (IsX86 && Name.startswith("avx512.mask.psrl")) {
2114       bool IsImmediate = Name[16] == 'i' ||
2115                          (Name.size() > 18 && Name[18] == 'i');
2116       bool IsVariable = Name[16] == 'v';
2117       char Size = Name[16] == '.' ? Name[17] :
2118                   Name[17] == '.' ? Name[18] :
2119                   Name[18] == '.' ? Name[19] :
2120                                     Name[20];
2121 
2122       Intrinsic::ID IID;
2123       if (IsVariable && Name[17] != '.') {
2124         if (Size == 'd' && Name[17] == '2') // avx512.mask.psrlv2.di
2125           IID = Intrinsic::x86_avx2_psrlv_q;
2126         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psrlv4.di
2127           IID = Intrinsic::x86_avx2_psrlv_q_256;
2128         else if (Size == 's' && Name[17] == '4') // avx512.mask.psrlv4.si
2129           IID = Intrinsic::x86_avx2_psrlv_d;
2130         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrlv8.si
2131           IID = Intrinsic::x86_avx2_psrlv_d_256;
2132         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrlv8.hi
2133           IID = Intrinsic::x86_avx512_psrlv_w_128;
2134         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrlv16.hi
2135           IID = Intrinsic::x86_avx512_psrlv_w_256;
2136         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrlv32hi
2137           IID = Intrinsic::x86_avx512_psrlv_w_512;
2138         else
2139           llvm_unreachable("Unexpected size");
2140       } else if (Name.endswith(".128")) {
2141         if (Size == 'd') // avx512.mask.psrl.d.128, avx512.mask.psrl.di.128
2142           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_d
2143                             : Intrinsic::x86_sse2_psrl_d;
2144         else if (Size == 'q') // avx512.mask.psrl.q.128, avx512.mask.psrl.qi.128
2145           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_q
2146                             : Intrinsic::x86_sse2_psrl_q;
2147         else if (Size == 'w') // avx512.mask.psrl.w.128, avx512.mask.psrl.wi.128
2148           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_w
2149                             : Intrinsic::x86_sse2_psrl_w;
2150         else
2151           llvm_unreachable("Unexpected size");
2152       } else if (Name.endswith(".256")) {
2153         if (Size == 'd') // avx512.mask.psrl.d.256, avx512.mask.psrl.di.256
2154           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_d
2155                             : Intrinsic::x86_avx2_psrl_d;
2156         else if (Size == 'q') // avx512.mask.psrl.q.256, avx512.mask.psrl.qi.256
2157           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_q
2158                             : Intrinsic::x86_avx2_psrl_q;
2159         else if (Size == 'w') // avx512.mask.psrl.w.256, avx512.mask.psrl.wi.256
2160           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_w
2161                             : Intrinsic::x86_avx2_psrl_w;
2162         else
2163           llvm_unreachable("Unexpected size");
2164       } else {
2165         if (Size == 'd') // psrl.di.512, psrli.d, psrl.d, psrl.d.512
2166           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_d_512 :
2167                 IsVariable  ? Intrinsic::x86_avx512_psrlv_d_512 :
2168                               Intrinsic::x86_avx512_psrl_d_512;
2169         else if (Size == 'q') // psrl.qi.512, psrli.q, psrl.q, psrl.q.512
2170           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_q_512 :
2171                 IsVariable  ? Intrinsic::x86_avx512_psrlv_q_512 :
2172                               Intrinsic::x86_avx512_psrl_q_512;
2173         else if (Size == 'w') // psrl.wi.512, psrli.w, psrl.w)
2174           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_w_512
2175                             : Intrinsic::x86_avx512_psrl_w_512;
2176         else
2177           llvm_unreachable("Unexpected size");
2178       }
2179 
2180       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2181     } else if (IsX86 && Name.startswith("avx512.mask.psra")) {
2182       bool IsImmediate = Name[16] == 'i' ||
2183                          (Name.size() > 18 && Name[18] == 'i');
2184       bool IsVariable = Name[16] == 'v';
2185       char Size = Name[16] == '.' ? Name[17] :
2186                   Name[17] == '.' ? Name[18] :
2187                   Name[18] == '.' ? Name[19] :
2188                                     Name[20];
2189 
2190       Intrinsic::ID IID;
2191       if (IsVariable && Name[17] != '.') {
2192         if (Size == 's' && Name[17] == '4') // avx512.mask.psrav4.si
2193           IID = Intrinsic::x86_avx2_psrav_d;
2194         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrav8.si
2195           IID = Intrinsic::x86_avx2_psrav_d_256;
2196         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrav8.hi
2197           IID = Intrinsic::x86_avx512_psrav_w_128;
2198         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrav16.hi
2199           IID = Intrinsic::x86_avx512_psrav_w_256;
2200         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrav32hi
2201           IID = Intrinsic::x86_avx512_psrav_w_512;
2202         else
2203           llvm_unreachable("Unexpected size");
2204       } else if (Name.endswith(".128")) {
2205         if (Size == 'd') // avx512.mask.psra.d.128, avx512.mask.psra.di.128
2206           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_d
2207                             : Intrinsic::x86_sse2_psra_d;
2208         else if (Size == 'q') // avx512.mask.psra.q.128, avx512.mask.psra.qi.128
2209           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_128 :
2210                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_128 :
2211                               Intrinsic::x86_avx512_psra_q_128;
2212         else if (Size == 'w') // avx512.mask.psra.w.128, avx512.mask.psra.wi.128
2213           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_w
2214                             : Intrinsic::x86_sse2_psra_w;
2215         else
2216           llvm_unreachable("Unexpected size");
2217       } else if (Name.endswith(".256")) {
2218         if (Size == 'd') // avx512.mask.psra.d.256, avx512.mask.psra.di.256
2219           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_d
2220                             : Intrinsic::x86_avx2_psra_d;
2221         else if (Size == 'q') // avx512.mask.psra.q.256, avx512.mask.psra.qi.256
2222           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_256 :
2223                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_256 :
2224                               Intrinsic::x86_avx512_psra_q_256;
2225         else if (Size == 'w') // avx512.mask.psra.w.256, avx512.mask.psra.wi.256
2226           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_w
2227                             : Intrinsic::x86_avx2_psra_w;
2228         else
2229           llvm_unreachable("Unexpected size");
2230       } else {
2231         if (Size == 'd') // psra.di.512, psrai.d, psra.d, psrav.d.512
2232           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_d_512 :
2233                 IsVariable  ? Intrinsic::x86_avx512_psrav_d_512 :
2234                               Intrinsic::x86_avx512_psra_d_512;
2235         else if (Size == 'q') // psra.qi.512, psrai.q, psra.q
2236           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_512 :
2237                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_512 :
2238                               Intrinsic::x86_avx512_psra_q_512;
2239         else if (Size == 'w') // psra.wi.512, psrai.w, psra.w
2240           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_w_512
2241                             : Intrinsic::x86_avx512_psra_w_512;
2242         else
2243           llvm_unreachable("Unexpected size");
2244       }
2245 
2246       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2247     } else if (IsX86 && Name.startswith("avx512.mask.move.s")) {
2248       Rep = upgradeMaskedMove(Builder, *CI);
2249     } else if (IsX86 && Name.startswith("avx512.cvtmask2")) {
2250       Rep = UpgradeMaskToInt(Builder, *CI);
2251     } else if (IsX86 && Name.startswith("avx512.mask.vpermilvar.")) {
2252       Intrinsic::ID IID;
2253       if (Name.endswith("ps.128"))
2254         IID = Intrinsic::x86_avx_vpermilvar_ps;
2255       else if (Name.endswith("pd.128"))
2256         IID = Intrinsic::x86_avx_vpermilvar_pd;
2257       else if (Name.endswith("ps.256"))
2258         IID = Intrinsic::x86_avx_vpermilvar_ps_256;
2259       else if (Name.endswith("pd.256"))
2260         IID = Intrinsic::x86_avx_vpermilvar_pd_256;
2261       else if (Name.endswith("ps.512"))
2262         IID = Intrinsic::x86_avx512_vpermilvar_ps_512;
2263       else if (Name.endswith("pd.512"))
2264         IID = Intrinsic::x86_avx512_vpermilvar_pd_512;
2265       else
2266         llvm_unreachable("Unexpected vpermilvar intrinsic");
2267 
2268       Function *Intrin = Intrinsic::getDeclaration(F->getParent(), IID);
2269       Rep = Builder.CreateCall(Intrin,
2270                                { CI->getArgOperand(0), CI->getArgOperand(1) });
2271       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2272                           CI->getArgOperand(2));
2273     } else if (IsX86 && Name.endswith(".movntdqa")) {
2274       Module *M = F->getParent();
2275       MDNode *Node = MDNode::get(
2276           C, ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
2277 
2278       Value *Ptr = CI->getArgOperand(0);
2279       VectorType *VTy = cast<VectorType>(CI->getType());
2280 
2281       // Convert the type of the pointer to a pointer to the stored type.
2282       Value *BC =
2283           Builder.CreateBitCast(Ptr, PointerType::getUnqual(VTy), "cast");
2284       LoadInst *LI = Builder.CreateAlignedLoad(BC, VTy->getBitWidth() / 8);
2285       LI->setMetadata(M->getMDKindID("nontemporal"), Node);
2286       Rep = LI;
2287     } else if (IsX86 &&
2288                (Name.startswith("sse2.pavg") || Name.startswith("avx2.pavg") ||
2289                 Name.startswith("avx512.mask.pavg"))) {
2290       // llvm.x86.sse2.pavg.b/w, llvm.x86.avx2.pavg.b/w,
2291       // llvm.x86.avx512.mask.pavg.b/w
2292       Value *A = CI->getArgOperand(0);
2293       Value *B = CI->getArgOperand(1);
2294       VectorType *ZextType = VectorType::getExtendedElementVectorType(
2295           cast<VectorType>(A->getType()));
2296       Value *ExtendedA = Builder.CreateZExt(A, ZextType);
2297       Value *ExtendedB = Builder.CreateZExt(B, ZextType);
2298       Value *Sum = Builder.CreateAdd(ExtendedA, ExtendedB);
2299       Value *AddOne = Builder.CreateAdd(Sum, ConstantInt::get(ZextType, 1));
2300       Value *ShiftR = Builder.CreateLShr(AddOne, ConstantInt::get(ZextType, 1));
2301       Rep = Builder.CreateTrunc(ShiftR, A->getType());
2302       if (CI->getNumArgOperands() > 2) {
2303         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2304                             CI->getArgOperand(2));
2305       }
2306     } else if (IsNVVM && (Name == "abs.i" || Name == "abs.ll")) {
2307       Value *Arg = CI->getArgOperand(0);
2308       Value *Neg = Builder.CreateNeg(Arg, "neg");
2309       Value *Cmp = Builder.CreateICmpSGE(
2310           Arg, llvm::Constant::getNullValue(Arg->getType()), "abs.cond");
2311       Rep = Builder.CreateSelect(Cmp, Arg, Neg, "abs");
2312     } else if (IsNVVM && (Name == "max.i" || Name == "max.ll" ||
2313                           Name == "max.ui" || Name == "max.ull")) {
2314       Value *Arg0 = CI->getArgOperand(0);
2315       Value *Arg1 = CI->getArgOperand(1);
2316       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
2317                        ? Builder.CreateICmpUGE(Arg0, Arg1, "max.cond")
2318                        : Builder.CreateICmpSGE(Arg0, Arg1, "max.cond");
2319       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "max");
2320     } else if (IsNVVM && (Name == "min.i" || Name == "min.ll" ||
2321                           Name == "min.ui" || Name == "min.ull")) {
2322       Value *Arg0 = CI->getArgOperand(0);
2323       Value *Arg1 = CI->getArgOperand(1);
2324       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
2325                        ? Builder.CreateICmpULE(Arg0, Arg1, "min.cond")
2326                        : Builder.CreateICmpSLE(Arg0, Arg1, "min.cond");
2327       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "min");
2328     } else if (IsNVVM && Name == "clz.ll") {
2329       // llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 and returns an i64.
2330       Value *Arg = CI->getArgOperand(0);
2331       Value *Ctlz = Builder.CreateCall(
2332           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
2333                                     {Arg->getType()}),
2334           {Arg, Builder.getFalse()}, "ctlz");
2335       Rep = Builder.CreateTrunc(Ctlz, Builder.getInt32Ty(), "ctlz.trunc");
2336     } else if (IsNVVM && Name == "popc.ll") {
2337       // llvm.nvvm.popc.ll returns an i32, but llvm.ctpop.i64 and returns an
2338       // i64.
2339       Value *Arg = CI->getArgOperand(0);
2340       Value *Popc = Builder.CreateCall(
2341           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
2342                                     {Arg->getType()}),
2343           Arg, "ctpop");
2344       Rep = Builder.CreateTrunc(Popc, Builder.getInt32Ty(), "ctpop.trunc");
2345     } else if (IsNVVM && Name == "h2f") {
2346       Rep = Builder.CreateCall(Intrinsic::getDeclaration(
2347                                    F->getParent(), Intrinsic::convert_from_fp16,
2348                                    {Builder.getFloatTy()}),
2349                                CI->getArgOperand(0), "h2f");
2350     } else {
2351       llvm_unreachable("Unknown function for CallInst upgrade.");
2352     }
2353 
2354     if (Rep)
2355       CI->replaceAllUsesWith(Rep);
2356     CI->eraseFromParent();
2357     return;
2358   }
2359 
2360   const auto &DefaultCase = [&NewFn, &CI]() -> void {
2361     // Handle generic mangling change, but nothing else
2362     assert(
2363         (CI->getCalledFunction()->getName() != NewFn->getName()) &&
2364         "Unknown function for CallInst upgrade and isn't just a name change");
2365     CI->setCalledFunction(NewFn);
2366   };
2367   CallInst *NewCall = nullptr;
2368   switch (NewFn->getIntrinsicID()) {
2369   default: {
2370     DefaultCase();
2371     return;
2372   }
2373 
2374   case Intrinsic::arm_neon_vld1:
2375   case Intrinsic::arm_neon_vld2:
2376   case Intrinsic::arm_neon_vld3:
2377   case Intrinsic::arm_neon_vld4:
2378   case Intrinsic::arm_neon_vld2lane:
2379   case Intrinsic::arm_neon_vld3lane:
2380   case Intrinsic::arm_neon_vld4lane:
2381   case Intrinsic::arm_neon_vst1:
2382   case Intrinsic::arm_neon_vst2:
2383   case Intrinsic::arm_neon_vst3:
2384   case Intrinsic::arm_neon_vst4:
2385   case Intrinsic::arm_neon_vst2lane:
2386   case Intrinsic::arm_neon_vst3lane:
2387   case Intrinsic::arm_neon_vst4lane: {
2388     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2389                                  CI->arg_operands().end());
2390     NewCall = Builder.CreateCall(NewFn, Args);
2391     break;
2392   }
2393 
2394   case Intrinsic::bitreverse:
2395     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2396     break;
2397 
2398   case Intrinsic::ctlz:
2399   case Intrinsic::cttz:
2400     assert(CI->getNumArgOperands() == 1 &&
2401            "Mismatch between function args and call args");
2402     NewCall =
2403         Builder.CreateCall(NewFn, {CI->getArgOperand(0), Builder.getFalse()});
2404     break;
2405 
2406   case Intrinsic::objectsize: {
2407     Value *NullIsUnknownSize = CI->getNumArgOperands() == 2
2408                                    ? Builder.getFalse()
2409                                    : CI->getArgOperand(2);
2410     NewCall = Builder.CreateCall(
2411         NewFn, {CI->getArgOperand(0), CI->getArgOperand(1), NullIsUnknownSize});
2412     break;
2413   }
2414 
2415   case Intrinsic::ctpop:
2416     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2417     break;
2418 
2419   case Intrinsic::convert_from_fp16:
2420     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2421     break;
2422 
2423   case Intrinsic::dbg_value:
2424     // Upgrade from the old version that had an extra offset argument.
2425     assert(CI->getNumArgOperands() == 4);
2426     // Drop nonzero offsets instead of attempting to upgrade them.
2427     if (auto *Offset = dyn_cast_or_null<Constant>(CI->getArgOperand(1)))
2428       if (Offset->isZeroValue()) {
2429         NewCall = Builder.CreateCall(
2430             NewFn,
2431             {CI->getArgOperand(0), CI->getArgOperand(2), CI->getArgOperand(3)});
2432         break;
2433       }
2434     CI->eraseFromParent();
2435     return;
2436 
2437   case Intrinsic::x86_xop_vfrcz_ss:
2438   case Intrinsic::x86_xop_vfrcz_sd:
2439     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(1)});
2440     break;
2441 
2442   case Intrinsic::x86_xop_vpermil2pd:
2443   case Intrinsic::x86_xop_vpermil2ps:
2444   case Intrinsic::x86_xop_vpermil2pd_256:
2445   case Intrinsic::x86_xop_vpermil2ps_256: {
2446     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2447                                  CI->arg_operands().end());
2448     VectorType *FltIdxTy = cast<VectorType>(Args[2]->getType());
2449     VectorType *IntIdxTy = VectorType::getInteger(FltIdxTy);
2450     Args[2] = Builder.CreateBitCast(Args[2], IntIdxTy);
2451     NewCall = Builder.CreateCall(NewFn, Args);
2452     break;
2453   }
2454 
2455   case Intrinsic::x86_sse41_ptestc:
2456   case Intrinsic::x86_sse41_ptestz:
2457   case Intrinsic::x86_sse41_ptestnzc: {
2458     // The arguments for these intrinsics used to be v4f32, and changed
2459     // to v2i64. This is purely a nop, since those are bitwise intrinsics.
2460     // So, the only thing required is a bitcast for both arguments.
2461     // First, check the arguments have the old type.
2462     Value *Arg0 = CI->getArgOperand(0);
2463     if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
2464       return;
2465 
2466     // Old intrinsic, add bitcasts
2467     Value *Arg1 = CI->getArgOperand(1);
2468 
2469     Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
2470 
2471     Value *BC0 = Builder.CreateBitCast(Arg0, NewVecTy, "cast");
2472     Value *BC1 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
2473 
2474     NewCall = Builder.CreateCall(NewFn, {BC0, BC1});
2475     break;
2476   }
2477 
2478   case Intrinsic::x86_sse41_insertps:
2479   case Intrinsic::x86_sse41_dppd:
2480   case Intrinsic::x86_sse41_dpps:
2481   case Intrinsic::x86_sse41_mpsadbw:
2482   case Intrinsic::x86_avx_dp_ps_256:
2483   case Intrinsic::x86_avx2_mpsadbw: {
2484     // Need to truncate the last argument from i32 to i8 -- this argument models
2485     // an inherently 8-bit immediate operand to these x86 instructions.
2486     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2487                                  CI->arg_operands().end());
2488 
2489     // Replace the last argument with a trunc.
2490     Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
2491     NewCall = Builder.CreateCall(NewFn, Args);
2492     break;
2493   }
2494 
2495   case Intrinsic::x86_avx512_mask_cmp_pd_128:
2496   case Intrinsic::x86_avx512_mask_cmp_pd_256:
2497   case Intrinsic::x86_avx512_mask_cmp_pd_512:
2498   case Intrinsic::x86_avx512_mask_cmp_ps_128:
2499   case Intrinsic::x86_avx512_mask_cmp_ps_256:
2500   case Intrinsic::x86_avx512_mask_cmp_ps_512: {
2501     SmallVector<Value *, 4> Args;
2502     Args.push_back(CI->getArgOperand(0));
2503     Args.push_back(CI->getArgOperand(1));
2504     Args.push_back(CI->getArgOperand(2));
2505     if (CI->getNumArgOperands() == 5)
2506       Args.push_back(CI->getArgOperand(4));
2507 
2508     NewCall = Builder.CreateCall(NewFn, Args);
2509     unsigned NumElts = Args[0]->getType()->getVectorNumElements();
2510     Value *Res = ApplyX86MaskOn1BitsVec(Builder, NewCall, CI->getArgOperand(3),
2511                                         NumElts);
2512 
2513     std::string Name = CI->getName();
2514     if (!Name.empty()) {
2515       CI->setName(Name + ".old");
2516       NewCall->setName(Name);
2517     }
2518     CI->replaceAllUsesWith(Res);
2519     CI->eraseFromParent();
2520     return;
2521   }
2522 
2523   case Intrinsic::thread_pointer: {
2524     NewCall = Builder.CreateCall(NewFn, {});
2525     break;
2526   }
2527 
2528   case Intrinsic::invariant_start:
2529   case Intrinsic::invariant_end:
2530   case Intrinsic::masked_load:
2531   case Intrinsic::masked_store:
2532   case Intrinsic::masked_gather:
2533   case Intrinsic::masked_scatter: {
2534     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2535                                  CI->arg_operands().end());
2536     NewCall = Builder.CreateCall(NewFn, Args);
2537     break;
2538   }
2539 
2540   case Intrinsic::memcpy:
2541   case Intrinsic::memmove:
2542   case Intrinsic::memset: {
2543     // We have to make sure that the call signature is what we're expecting.
2544     // We only want to change the old signatures by removing the alignment arg:
2545     //  @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i32, i1)
2546     //    -> @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i1)
2547     //  @llvm.memset...(i8*, i8, i[32|64], i32, i1)
2548     //    -> @llvm.memset...(i8*, i8, i[32|64], i1)
2549     // Note: i8*'s in the above can be any pointer type
2550     if (CI->getNumArgOperands() != 5) {
2551       DefaultCase();
2552       return;
2553     }
2554     // Remove alignment argument (3), and add alignment attributes to the
2555     // dest/src pointers.
2556     Value *Args[4] = {CI->getArgOperand(0), CI->getArgOperand(1),
2557                       CI->getArgOperand(2), CI->getArgOperand(4)};
2558     NewCall = Builder.CreateCall(NewFn, Args);
2559     auto *MemCI = cast<MemIntrinsic>(NewCall);
2560     // All mem intrinsics support dest alignment.
2561     const ConstantInt *Align = cast<ConstantInt>(CI->getArgOperand(3));
2562     MemCI->setDestAlignment(Align->getZExtValue());
2563     // Memcpy/Memmove also support source alignment.
2564     if (auto *MTI = dyn_cast<MemTransferInst>(MemCI))
2565       MTI->setSourceAlignment(Align->getZExtValue());
2566     break;
2567   }
2568   }
2569   assert(NewCall && "Should have either set this variable or returned through "
2570                     "the default case");
2571   std::string Name = CI->getName();
2572   if (!Name.empty()) {
2573     CI->setName(Name + ".old");
2574     NewCall->setName(Name);
2575   }
2576   CI->replaceAllUsesWith(NewCall);
2577   CI->eraseFromParent();
2578 }
2579 
2580 void llvm::UpgradeCallsToIntrinsic(Function *F) {
2581   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
2582 
2583   // Check if this function should be upgraded and get the replacement function
2584   // if there is one.
2585   Function *NewFn;
2586   if (UpgradeIntrinsicFunction(F, NewFn)) {
2587     // Replace all users of the old function with the new function or new
2588     // instructions. This is not a range loop because the call is deleted.
2589     for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; )
2590       if (CallInst *CI = dyn_cast<CallInst>(*UI++))
2591         UpgradeIntrinsicCall(CI, NewFn);
2592 
2593     // Remove old function, no longer used, from the module.
2594     F->eraseFromParent();
2595   }
2596 }
2597 
2598 MDNode *llvm::UpgradeTBAANode(MDNode &MD) {
2599   // Check if the tag uses struct-path aware TBAA format.
2600   if (isa<MDNode>(MD.getOperand(0)) && MD.getNumOperands() >= 3)
2601     return &MD;
2602 
2603   auto &Context = MD.getContext();
2604   if (MD.getNumOperands() == 3) {
2605     Metadata *Elts[] = {MD.getOperand(0), MD.getOperand(1)};
2606     MDNode *ScalarType = MDNode::get(Context, Elts);
2607     // Create a MDNode <ScalarType, ScalarType, offset 0, const>
2608     Metadata *Elts2[] = {ScalarType, ScalarType,
2609                          ConstantAsMetadata::get(
2610                              Constant::getNullValue(Type::getInt64Ty(Context))),
2611                          MD.getOperand(2)};
2612     return MDNode::get(Context, Elts2);
2613   }
2614   // Create a MDNode <MD, MD, offset 0>
2615   Metadata *Elts[] = {&MD, &MD, ConstantAsMetadata::get(Constant::getNullValue(
2616                                     Type::getInt64Ty(Context)))};
2617   return MDNode::get(Context, Elts);
2618 }
2619 
2620 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
2621                                       Instruction *&Temp) {
2622   if (Opc != Instruction::BitCast)
2623     return nullptr;
2624 
2625   Temp = nullptr;
2626   Type *SrcTy = V->getType();
2627   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
2628       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
2629     LLVMContext &Context = V->getContext();
2630 
2631     // We have no information about target data layout, so we assume that
2632     // the maximum pointer size is 64bit.
2633     Type *MidTy = Type::getInt64Ty(Context);
2634     Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
2635 
2636     return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
2637   }
2638 
2639   return nullptr;
2640 }
2641 
2642 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
2643   if (Opc != Instruction::BitCast)
2644     return nullptr;
2645 
2646   Type *SrcTy = C->getType();
2647   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
2648       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
2649     LLVMContext &Context = C->getContext();
2650 
2651     // We have no information about target data layout, so we assume that
2652     // the maximum pointer size is 64bit.
2653     Type *MidTy = Type::getInt64Ty(Context);
2654 
2655     return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
2656                                      DestTy);
2657   }
2658 
2659   return nullptr;
2660 }
2661 
2662 /// Check the debug info version number, if it is out-dated, drop the debug
2663 /// info. Return true if module is modified.
2664 bool llvm::UpgradeDebugInfo(Module &M) {
2665   unsigned Version = getDebugMetadataVersionFromModule(M);
2666   if (Version == DEBUG_METADATA_VERSION) {
2667     bool BrokenDebugInfo = false;
2668     if (verifyModule(M, &llvm::errs(), &BrokenDebugInfo))
2669       report_fatal_error("Broken module found, compilation aborted!");
2670     if (!BrokenDebugInfo)
2671       // Everything is ok.
2672       return false;
2673     else {
2674       // Diagnose malformed debug info.
2675       DiagnosticInfoIgnoringInvalidDebugMetadata Diag(M);
2676       M.getContext().diagnose(Diag);
2677     }
2678   }
2679   bool Modified = StripDebugInfo(M);
2680   if (Modified && Version != DEBUG_METADATA_VERSION) {
2681     // Diagnose a version mismatch.
2682     DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
2683     M.getContext().diagnose(DiagVersion);
2684   }
2685   return Modified;
2686 }
2687 
2688 bool llvm::UpgradeModuleFlags(Module &M) {
2689   NamedMDNode *ModFlags = M.getModuleFlagsMetadata();
2690   if (!ModFlags)
2691     return false;
2692 
2693   bool HasObjCFlag = false, HasClassProperties = false, Changed = false;
2694   for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) {
2695     MDNode *Op = ModFlags->getOperand(I);
2696     if (Op->getNumOperands() != 3)
2697       continue;
2698     MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(1));
2699     if (!ID)
2700       continue;
2701     if (ID->getString() == "Objective-C Image Info Version")
2702       HasObjCFlag = true;
2703     if (ID->getString() == "Objective-C Class Properties")
2704       HasClassProperties = true;
2705     // Upgrade PIC/PIE Module Flags. The module flag behavior for these two
2706     // field was Error and now they are Max.
2707     if (ID->getString() == "PIC Level" || ID->getString() == "PIE Level") {
2708       if (auto *Behavior =
2709               mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0))) {
2710         if (Behavior->getLimitedValue() == Module::Error) {
2711           Type *Int32Ty = Type::getInt32Ty(M.getContext());
2712           Metadata *Ops[3] = {
2713               ConstantAsMetadata::get(ConstantInt::get(Int32Ty, Module::Max)),
2714               MDString::get(M.getContext(), ID->getString()),
2715               Op->getOperand(2)};
2716           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
2717           Changed = true;
2718         }
2719       }
2720     }
2721     // Upgrade Objective-C Image Info Section. Removed the whitespce in the
2722     // section name so that llvm-lto will not complain about mismatching
2723     // module flags that is functionally the same.
2724     if (ID->getString() == "Objective-C Image Info Section") {
2725       if (auto *Value = dyn_cast_or_null<MDString>(Op->getOperand(2))) {
2726         SmallVector<StringRef, 4> ValueComp;
2727         Value->getString().split(ValueComp, " ");
2728         if (ValueComp.size() != 1) {
2729           std::string NewValue;
2730           for (auto &S : ValueComp)
2731             NewValue += S.str();
2732           Metadata *Ops[3] = {Op->getOperand(0), Op->getOperand(1),
2733                               MDString::get(M.getContext(), NewValue)};
2734           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
2735           Changed = true;
2736         }
2737       }
2738     }
2739   }
2740 
2741   // "Objective-C Class Properties" is recently added for Objective-C. We
2742   // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module
2743   // flag of value 0, so we can correclty downgrade this flag when trying to
2744   // link an ObjC bitcode without this module flag with an ObjC bitcode with
2745   // this module flag.
2746   if (HasObjCFlag && !HasClassProperties) {
2747     M.addModuleFlag(llvm::Module::Override, "Objective-C Class Properties",
2748                     (uint32_t)0);
2749     Changed = true;
2750   }
2751 
2752   return Changed;
2753 }
2754 
2755 void llvm::UpgradeSectionAttributes(Module &M) {
2756   auto TrimSpaces = [](StringRef Section) -> std::string {
2757     SmallVector<StringRef, 5> Components;
2758     Section.split(Components, ',');
2759 
2760     SmallString<32> Buffer;
2761     raw_svector_ostream OS(Buffer);
2762 
2763     for (auto Component : Components)
2764       OS << ',' << Component.trim();
2765 
2766     return OS.str().substr(1);
2767   };
2768 
2769   for (auto &GV : M.globals()) {
2770     if (!GV.hasSection())
2771       continue;
2772 
2773     StringRef Section = GV.getSection();
2774 
2775     if (!Section.startswith("__DATA, __objc_catlist"))
2776       continue;
2777 
2778     // __DATA, __objc_catlist, regular, no_dead_strip
2779     // __DATA,__objc_catlist,regular,no_dead_strip
2780     GV.setSection(TrimSpaces(Section));
2781   }
2782 }
2783 
2784 static bool isOldLoopArgument(Metadata *MD) {
2785   auto *T = dyn_cast_or_null<MDTuple>(MD);
2786   if (!T)
2787     return false;
2788   if (T->getNumOperands() < 1)
2789     return false;
2790   auto *S = dyn_cast_or_null<MDString>(T->getOperand(0));
2791   if (!S)
2792     return false;
2793   return S->getString().startswith("llvm.vectorizer.");
2794 }
2795 
2796 static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) {
2797   StringRef OldPrefix = "llvm.vectorizer.";
2798   assert(OldTag.startswith(OldPrefix) && "Expected old prefix");
2799 
2800   if (OldTag == "llvm.vectorizer.unroll")
2801     return MDString::get(C, "llvm.loop.interleave.count");
2802 
2803   return MDString::get(
2804       C, (Twine("llvm.loop.vectorize.") + OldTag.drop_front(OldPrefix.size()))
2805              .str());
2806 }
2807 
2808 static Metadata *upgradeLoopArgument(Metadata *MD) {
2809   auto *T = dyn_cast_or_null<MDTuple>(MD);
2810   if (!T)
2811     return MD;
2812   if (T->getNumOperands() < 1)
2813     return MD;
2814   auto *OldTag = dyn_cast_or_null<MDString>(T->getOperand(0));
2815   if (!OldTag)
2816     return MD;
2817   if (!OldTag->getString().startswith("llvm.vectorizer."))
2818     return MD;
2819 
2820   // This has an old tag.  Upgrade it.
2821   SmallVector<Metadata *, 8> Ops;
2822   Ops.reserve(T->getNumOperands());
2823   Ops.push_back(upgradeLoopTag(T->getContext(), OldTag->getString()));
2824   for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I)
2825     Ops.push_back(T->getOperand(I));
2826 
2827   return MDTuple::get(T->getContext(), Ops);
2828 }
2829 
2830 MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) {
2831   auto *T = dyn_cast<MDTuple>(&N);
2832   if (!T)
2833     return &N;
2834 
2835   if (none_of(T->operands(), isOldLoopArgument))
2836     return &N;
2837 
2838   SmallVector<Metadata *, 8> Ops;
2839   Ops.reserve(T->getNumOperands());
2840   for (Metadata *MD : T->operands())
2841     Ops.push_back(upgradeLoopArgument(MD));
2842 
2843   return MDTuple::get(T->getContext(), Ops);
2844 }
2845