xref: /aosp_15_r20/external/swiftshader/third_party/llvm-16.0/llvm/lib/IR/AutoUpgrade.cpp (revision 03ce13f70fcc45d86ee91b7ee4cab1936a95046e)
1 //===-- AutoUpgrade.cpp - Implement auto-upgrade helper functions ---------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the auto-upgrade helper functions.
10 // This is where deprecated IR intrinsics and other IR features are updated to
11 // current specifications.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/IR/AutoUpgrade.h"
16 #include "llvm/ADT/StringSwitch.h"
17 #include "llvm/ADT/Triple.h"
18 #include "llvm/IR/Constants.h"
19 #include "llvm/IR/DebugInfo.h"
20 #include "llvm/IR/DiagnosticInfo.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/IR/InstVisitor.h"
24 #include "llvm/IR/Instruction.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/IR/Intrinsics.h"
27 #include "llvm/IR/IntrinsicsAArch64.h"
28 #include "llvm/IR/IntrinsicsARM.h"
29 #include "llvm/IR/IntrinsicsX86.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/IR/Verifier.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/Regex.h"
35 #include <cstring>
36 using namespace llvm;
37 
rename(GlobalValue * GV)38 static void rename(GlobalValue *GV) { GV->setName(GV->getName() + ".old"); }
39 
40 // Upgrade the declarations of the SSE4.1 ptest intrinsics whose arguments have
41 // changed their type from v4f32 to v2i64.
UpgradePTESTIntrinsic(Function * F,Intrinsic::ID IID,Function * & NewFn)42 static bool UpgradePTESTIntrinsic(Function* F, Intrinsic::ID IID,
43                                   Function *&NewFn) {
44   // Check whether this is an old version of the function, which received
45   // v4f32 arguments.
46   Type *Arg0Type = F->getFunctionType()->getParamType(0);
47   if (Arg0Type != FixedVectorType::get(Type::getFloatTy(F->getContext()), 4))
48     return false;
49 
50   // Yes, it's old, replace it with new version.
51   rename(F);
52   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
53   return true;
54 }
55 
56 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
57 // arguments have changed their type from i32 to i8.
UpgradeX86IntrinsicsWith8BitMask(Function * F,Intrinsic::ID IID,Function * & NewFn)58 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
59                                              Function *&NewFn) {
60   // Check that the last argument is an i32.
61   Type *LastArgType = F->getFunctionType()->getParamType(
62      F->getFunctionType()->getNumParams() - 1);
63   if (!LastArgType->isIntegerTy(32))
64     return false;
65 
66   // Move this function aside and map down.
67   rename(F);
68   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
69   return true;
70 }
71 
72 // Upgrade the declaration of fp compare intrinsics that change return type
73 // from scalar to vXi1 mask.
UpgradeX86MaskedFPCompare(Function * F,Intrinsic::ID IID,Function * & NewFn)74 static bool UpgradeX86MaskedFPCompare(Function *F, Intrinsic::ID IID,
75                                       Function *&NewFn) {
76   // Check if the return type is a vector.
77   if (F->getReturnType()->isVectorTy())
78     return false;
79 
80   rename(F);
81   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
82   return true;
83 }
84 
UpgradeX86BF16Intrinsic(Function * F,Intrinsic::ID IID,Function * & NewFn)85 static bool UpgradeX86BF16Intrinsic(Function *F, Intrinsic::ID IID,
86                                     Function *&NewFn) {
87   if (F->getReturnType()->getScalarType()->isBFloatTy())
88     return false;
89 
90   rename(F);
91   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
92   return true;
93 }
94 
UpgradeX86BF16DPIntrinsic(Function * F,Intrinsic::ID IID,Function * & NewFn)95 static bool UpgradeX86BF16DPIntrinsic(Function *F, Intrinsic::ID IID,
96                                       Function *&NewFn) {
97   if (F->getFunctionType()->getParamType(1)->getScalarType()->isBFloatTy())
98     return false;
99 
100   rename(F);
101   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
102   return true;
103 }
104 
ShouldUpgradeX86Intrinsic(Function * F,StringRef Name)105 static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
106   // All of the intrinsics matches below should be marked with which llvm
107   // version started autoupgrading them. At some point in the future we would
108   // like to use this information to remove upgrade code for some older
109   // intrinsics. It is currently undecided how we will determine that future
110   // point.
111   if (Name == "addcarryx.u32" || // Added in 8.0
112       Name == "addcarryx.u64" || // Added in 8.0
113       Name == "addcarry.u32" || // Added in 8.0
114       Name == "addcarry.u64" || // Added in 8.0
115       Name == "subborrow.u32" || // Added in 8.0
116       Name == "subborrow.u64" || // Added in 8.0
117       Name.startswith("sse2.padds.") || // Added in 8.0
118       Name.startswith("sse2.psubs.") || // Added in 8.0
119       Name.startswith("sse2.paddus.") || // Added in 8.0
120       Name.startswith("sse2.psubus.") || // Added in 8.0
121       Name.startswith("avx2.padds.") || // Added in 8.0
122       Name.startswith("avx2.psubs.") || // Added in 8.0
123       Name.startswith("avx2.paddus.") || // Added in 8.0
124       Name.startswith("avx2.psubus.") || // Added in 8.0
125       Name.startswith("avx512.padds.") || // Added in 8.0
126       Name.startswith("avx512.psubs.") || // Added in 8.0
127       Name.startswith("avx512.mask.padds.") || // Added in 8.0
128       Name.startswith("avx512.mask.psubs.") || // Added in 8.0
129       Name.startswith("avx512.mask.paddus.") || // Added in 8.0
130       Name.startswith("avx512.mask.psubus.") || // Added in 8.0
131       Name=="ssse3.pabs.b.128" || // Added in 6.0
132       Name=="ssse3.pabs.w.128" || // Added in 6.0
133       Name=="ssse3.pabs.d.128" || // Added in 6.0
134       Name.startswith("fma4.vfmadd.s") || // Added in 7.0
135       Name.startswith("fma.vfmadd.") || // Added in 7.0
136       Name.startswith("fma.vfmsub.") || // Added in 7.0
137       Name.startswith("fma.vfmsubadd.") || // Added in 7.0
138       Name.startswith("fma.vfnmadd.") || // Added in 7.0
139       Name.startswith("fma.vfnmsub.") || // Added in 7.0
140       Name.startswith("avx512.mask.vfmadd.") || // Added in 7.0
141       Name.startswith("avx512.mask.vfnmadd.") || // Added in 7.0
142       Name.startswith("avx512.mask.vfnmsub.") || // Added in 7.0
143       Name.startswith("avx512.mask3.vfmadd.") || // Added in 7.0
144       Name.startswith("avx512.maskz.vfmadd.") || // Added in 7.0
145       Name.startswith("avx512.mask3.vfmsub.") || // Added in 7.0
146       Name.startswith("avx512.mask3.vfnmsub.") || // Added in 7.0
147       Name.startswith("avx512.mask.vfmaddsub.") || // Added in 7.0
148       Name.startswith("avx512.maskz.vfmaddsub.") || // Added in 7.0
149       Name.startswith("avx512.mask3.vfmaddsub.") || // Added in 7.0
150       Name.startswith("avx512.mask3.vfmsubadd.") || // Added in 7.0
151       Name.startswith("avx512.mask.shuf.i") || // Added in 6.0
152       Name.startswith("avx512.mask.shuf.f") || // Added in 6.0
153       Name.startswith("avx512.kunpck") || //added in 6.0
154       Name.startswith("avx2.pabs.") || // Added in 6.0
155       Name.startswith("avx512.mask.pabs.") || // Added in 6.0
156       Name.startswith("avx512.broadcastm") || // Added in 6.0
157       Name == "sse.sqrt.ss" || // Added in 7.0
158       Name == "sse2.sqrt.sd" || // Added in 7.0
159       Name.startswith("avx512.mask.sqrt.p") || // Added in 7.0
160       Name.startswith("avx.sqrt.p") || // Added in 7.0
161       Name.startswith("sse2.sqrt.p") || // Added in 7.0
162       Name.startswith("sse.sqrt.p") || // Added in 7.0
163       Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
164       Name.startswith("sse2.pcmpeq.") || // Added in 3.1
165       Name.startswith("sse2.pcmpgt.") || // Added in 3.1
166       Name.startswith("avx2.pcmpeq.") || // Added in 3.1
167       Name.startswith("avx2.pcmpgt.") || // Added in 3.1
168       Name.startswith("avx512.mask.pcmpeq.") || // Added in 3.9
169       Name.startswith("avx512.mask.pcmpgt.") || // Added in 3.9
170       Name.startswith("avx.vperm2f128.") || // Added in 6.0
171       Name == "avx2.vperm2i128" || // Added in 6.0
172       Name == "sse.add.ss" || // Added in 4.0
173       Name == "sse2.add.sd" || // Added in 4.0
174       Name == "sse.sub.ss" || // Added in 4.0
175       Name == "sse2.sub.sd" || // Added in 4.0
176       Name == "sse.mul.ss" || // Added in 4.0
177       Name == "sse2.mul.sd" || // Added in 4.0
178       Name == "sse.div.ss" || // Added in 4.0
179       Name == "sse2.div.sd" || // Added in 4.0
180       Name == "sse41.pmaxsb" || // Added in 3.9
181       Name == "sse2.pmaxs.w" || // Added in 3.9
182       Name == "sse41.pmaxsd" || // Added in 3.9
183       Name == "sse2.pmaxu.b" || // Added in 3.9
184       Name == "sse41.pmaxuw" || // Added in 3.9
185       Name == "sse41.pmaxud" || // Added in 3.9
186       Name == "sse41.pminsb" || // Added in 3.9
187       Name == "sse2.pmins.w" || // Added in 3.9
188       Name == "sse41.pminsd" || // Added in 3.9
189       Name == "sse2.pminu.b" || // Added in 3.9
190       Name == "sse41.pminuw" || // Added in 3.9
191       Name == "sse41.pminud" || // Added in 3.9
192       Name == "avx512.kand.w" || // Added in 7.0
193       Name == "avx512.kandn.w" || // Added in 7.0
194       Name == "avx512.knot.w" || // Added in 7.0
195       Name == "avx512.kor.w" || // Added in 7.0
196       Name == "avx512.kxor.w" || // Added in 7.0
197       Name == "avx512.kxnor.w" || // Added in 7.0
198       Name == "avx512.kortestc.w" || // Added in 7.0
199       Name == "avx512.kortestz.w" || // Added in 7.0
200       Name.startswith("avx512.mask.pshuf.b.") || // Added in 4.0
201       Name.startswith("avx2.pmax") || // Added in 3.9
202       Name.startswith("avx2.pmin") || // Added in 3.9
203       Name.startswith("avx512.mask.pmax") || // Added in 4.0
204       Name.startswith("avx512.mask.pmin") || // Added in 4.0
205       Name.startswith("avx2.vbroadcast") || // Added in 3.8
206       Name.startswith("avx2.pbroadcast") || // Added in 3.8
207       Name.startswith("avx.vpermil.") || // Added in 3.1
208       Name.startswith("sse2.pshuf") || // Added in 3.9
209       Name.startswith("avx512.pbroadcast") || // Added in 3.9
210       Name.startswith("avx512.mask.broadcast.s") || // Added in 3.9
211       Name.startswith("avx512.mask.movddup") || // Added in 3.9
212       Name.startswith("avx512.mask.movshdup") || // Added in 3.9
213       Name.startswith("avx512.mask.movsldup") || // Added in 3.9
214       Name.startswith("avx512.mask.pshuf.d.") || // Added in 3.9
215       Name.startswith("avx512.mask.pshufl.w.") || // Added in 3.9
216       Name.startswith("avx512.mask.pshufh.w.") || // Added in 3.9
217       Name.startswith("avx512.mask.shuf.p") || // Added in 4.0
218       Name.startswith("avx512.mask.vpermil.p") || // Added in 3.9
219       Name.startswith("avx512.mask.perm.df.") || // Added in 3.9
220       Name.startswith("avx512.mask.perm.di.") || // Added in 3.9
221       Name.startswith("avx512.mask.punpckl") || // Added in 3.9
222       Name.startswith("avx512.mask.punpckh") || // Added in 3.9
223       Name.startswith("avx512.mask.unpckl.") || // Added in 3.9
224       Name.startswith("avx512.mask.unpckh.") || // Added in 3.9
225       Name.startswith("avx512.mask.pand.") || // Added in 3.9
226       Name.startswith("avx512.mask.pandn.") || // Added in 3.9
227       Name.startswith("avx512.mask.por.") || // Added in 3.9
228       Name.startswith("avx512.mask.pxor.") || // Added in 3.9
229       Name.startswith("avx512.mask.and.") || // Added in 3.9
230       Name.startswith("avx512.mask.andn.") || // Added in 3.9
231       Name.startswith("avx512.mask.or.") || // Added in 3.9
232       Name.startswith("avx512.mask.xor.") || // Added in 3.9
233       Name.startswith("avx512.mask.padd.") || // Added in 4.0
234       Name.startswith("avx512.mask.psub.") || // Added in 4.0
235       Name.startswith("avx512.mask.pmull.") || // Added in 4.0
236       Name.startswith("avx512.mask.cvtdq2pd.") || // Added in 4.0
237       Name.startswith("avx512.mask.cvtudq2pd.") || // Added in 4.0
238       Name.startswith("avx512.mask.cvtudq2ps.") || // Added in 7.0 updated 9.0
239       Name.startswith("avx512.mask.cvtqq2pd.") || // Added in 7.0 updated 9.0
240       Name.startswith("avx512.mask.cvtuqq2pd.") || // Added in 7.0 updated 9.0
241       Name.startswith("avx512.mask.cvtdq2ps.") || // Added in 7.0 updated 9.0
242       Name == "avx512.mask.vcvtph2ps.128" || // Added in 11.0
243       Name == "avx512.mask.vcvtph2ps.256" || // Added in 11.0
244       Name == "avx512.mask.cvtqq2ps.256" || // Added in 9.0
245       Name == "avx512.mask.cvtqq2ps.512" || // Added in 9.0
246       Name == "avx512.mask.cvtuqq2ps.256" || // Added in 9.0
247       Name == "avx512.mask.cvtuqq2ps.512" || // Added in 9.0
248       Name == "avx512.mask.cvtpd2dq.256" || // Added in 7.0
249       Name == "avx512.mask.cvtpd2ps.256" || // Added in 7.0
250       Name == "avx512.mask.cvttpd2dq.256" || // Added in 7.0
251       Name == "avx512.mask.cvttps2dq.128" || // Added in 7.0
252       Name == "avx512.mask.cvttps2dq.256" || // Added in 7.0
253       Name == "avx512.mask.cvtps2pd.128" || // Added in 7.0
254       Name == "avx512.mask.cvtps2pd.256" || // Added in 7.0
255       Name == "avx512.cvtusi2sd" || // Added in 7.0
256       Name.startswith("avx512.mask.permvar.") || // Added in 7.0
257       Name == "sse2.pmulu.dq" || // Added in 7.0
258       Name == "sse41.pmuldq" || // Added in 7.0
259       Name == "avx2.pmulu.dq" || // Added in 7.0
260       Name == "avx2.pmul.dq" || // Added in 7.0
261       Name == "avx512.pmulu.dq.512" || // Added in 7.0
262       Name == "avx512.pmul.dq.512" || // Added in 7.0
263       Name.startswith("avx512.mask.pmul.dq.") || // Added in 4.0
264       Name.startswith("avx512.mask.pmulu.dq.") || // Added in 4.0
265       Name.startswith("avx512.mask.pmul.hr.sw.") || // Added in 7.0
266       Name.startswith("avx512.mask.pmulh.w.") || // Added in 7.0
267       Name.startswith("avx512.mask.pmulhu.w.") || // Added in 7.0
268       Name.startswith("avx512.mask.pmaddw.d.") || // Added in 7.0
269       Name.startswith("avx512.mask.pmaddubs.w.") || // Added in 7.0
270       Name.startswith("avx512.mask.packsswb.") || // Added in 5.0
271       Name.startswith("avx512.mask.packssdw.") || // Added in 5.0
272       Name.startswith("avx512.mask.packuswb.") || // Added in 5.0
273       Name.startswith("avx512.mask.packusdw.") || // Added in 5.0
274       Name.startswith("avx512.mask.cmp.b") || // Added in 5.0
275       Name.startswith("avx512.mask.cmp.d") || // Added in 5.0
276       Name.startswith("avx512.mask.cmp.q") || // Added in 5.0
277       Name.startswith("avx512.mask.cmp.w") || // Added in 5.0
278       Name.startswith("avx512.cmp.p") || // Added in 12.0
279       Name.startswith("avx512.mask.ucmp.") || // Added in 5.0
280       Name.startswith("avx512.cvtb2mask.") || // Added in 7.0
281       Name.startswith("avx512.cvtw2mask.") || // Added in 7.0
282       Name.startswith("avx512.cvtd2mask.") || // Added in 7.0
283       Name.startswith("avx512.cvtq2mask.") || // Added in 7.0
284       Name.startswith("avx512.mask.vpermilvar.") || // Added in 4.0
285       Name.startswith("avx512.mask.psll.d") || // Added in 4.0
286       Name.startswith("avx512.mask.psll.q") || // Added in 4.0
287       Name.startswith("avx512.mask.psll.w") || // Added in 4.0
288       Name.startswith("avx512.mask.psra.d") || // Added in 4.0
289       Name.startswith("avx512.mask.psra.q") || // Added in 4.0
290       Name.startswith("avx512.mask.psra.w") || // Added in 4.0
291       Name.startswith("avx512.mask.psrl.d") || // Added in 4.0
292       Name.startswith("avx512.mask.psrl.q") || // Added in 4.0
293       Name.startswith("avx512.mask.psrl.w") || // Added in 4.0
294       Name.startswith("avx512.mask.pslli") || // Added in 4.0
295       Name.startswith("avx512.mask.psrai") || // Added in 4.0
296       Name.startswith("avx512.mask.psrli") || // Added in 4.0
297       Name.startswith("avx512.mask.psllv") || // Added in 4.0
298       Name.startswith("avx512.mask.psrav") || // Added in 4.0
299       Name.startswith("avx512.mask.psrlv") || // Added in 4.0
300       Name.startswith("sse41.pmovsx") || // Added in 3.8
301       Name.startswith("sse41.pmovzx") || // Added in 3.9
302       Name.startswith("avx2.pmovsx") || // Added in 3.9
303       Name.startswith("avx2.pmovzx") || // Added in 3.9
304       Name.startswith("avx512.mask.pmovsx") || // Added in 4.0
305       Name.startswith("avx512.mask.pmovzx") || // Added in 4.0
306       Name.startswith("avx512.mask.lzcnt.") || // Added in 5.0
307       Name.startswith("avx512.mask.pternlog.") || // Added in 7.0
308       Name.startswith("avx512.maskz.pternlog.") || // Added in 7.0
309       Name.startswith("avx512.mask.vpmadd52") || // Added in 7.0
310       Name.startswith("avx512.maskz.vpmadd52") || // Added in 7.0
311       Name.startswith("avx512.mask.vpermi2var.") || // Added in 7.0
312       Name.startswith("avx512.mask.vpermt2var.") || // Added in 7.0
313       Name.startswith("avx512.maskz.vpermt2var.") || // Added in 7.0
314       Name.startswith("avx512.mask.vpdpbusd.") || // Added in 7.0
315       Name.startswith("avx512.maskz.vpdpbusd.") || // Added in 7.0
316       Name.startswith("avx512.mask.vpdpbusds.") || // Added in 7.0
317       Name.startswith("avx512.maskz.vpdpbusds.") || // Added in 7.0
318       Name.startswith("avx512.mask.vpdpwssd.") || // Added in 7.0
319       Name.startswith("avx512.maskz.vpdpwssd.") || // Added in 7.0
320       Name.startswith("avx512.mask.vpdpwssds.") || // Added in 7.0
321       Name.startswith("avx512.maskz.vpdpwssds.") || // Added in 7.0
322       Name.startswith("avx512.mask.dbpsadbw.") || // Added in 7.0
323       Name.startswith("avx512.mask.vpshld.") || // Added in 7.0
324       Name.startswith("avx512.mask.vpshrd.") || // Added in 7.0
325       Name.startswith("avx512.mask.vpshldv.") || // Added in 8.0
326       Name.startswith("avx512.mask.vpshrdv.") || // Added in 8.0
327       Name.startswith("avx512.maskz.vpshldv.") || // Added in 8.0
328       Name.startswith("avx512.maskz.vpshrdv.") || // Added in 8.0
329       Name.startswith("avx512.vpshld.") || // Added in 8.0
330       Name.startswith("avx512.vpshrd.") || // Added in 8.0
331       Name.startswith("avx512.mask.add.p") || // Added in 7.0. 128/256 in 4.0
332       Name.startswith("avx512.mask.sub.p") || // Added in 7.0. 128/256 in 4.0
333       Name.startswith("avx512.mask.mul.p") || // Added in 7.0. 128/256 in 4.0
334       Name.startswith("avx512.mask.div.p") || // Added in 7.0. 128/256 in 4.0
335       Name.startswith("avx512.mask.max.p") || // Added in 7.0. 128/256 in 5.0
336       Name.startswith("avx512.mask.min.p") || // Added in 7.0. 128/256 in 5.0
337       Name.startswith("avx512.mask.fpclass.p") || // Added in 7.0
338       Name.startswith("avx512.mask.vpshufbitqmb.") || // Added in 8.0
339       Name.startswith("avx512.mask.pmultishift.qb.") || // Added in 8.0
340       Name.startswith("avx512.mask.conflict.") || // Added in 9.0
341       Name == "avx512.mask.pmov.qd.256" || // Added in 9.0
342       Name == "avx512.mask.pmov.qd.512" || // Added in 9.0
343       Name == "avx512.mask.pmov.wb.256" || // Added in 9.0
344       Name == "avx512.mask.pmov.wb.512" || // Added in 9.0
345       Name == "sse.cvtsi2ss" || // Added in 7.0
346       Name == "sse.cvtsi642ss" || // Added in 7.0
347       Name == "sse2.cvtsi2sd" || // Added in 7.0
348       Name == "sse2.cvtsi642sd" || // Added in 7.0
349       Name == "sse2.cvtss2sd" || // Added in 7.0
350       Name == "sse2.cvtdq2pd" || // Added in 3.9
351       Name == "sse2.cvtdq2ps" || // Added in 7.0
352       Name == "sse2.cvtps2pd" || // Added in 3.9
353       Name == "avx.cvtdq2.pd.256" || // Added in 3.9
354       Name == "avx.cvtdq2.ps.256" || // Added in 7.0
355       Name == "avx.cvt.ps2.pd.256" || // Added in 3.9
356       Name.startswith("vcvtph2ps.") || // Added in 11.0
357       Name.startswith("avx.vinsertf128.") || // Added in 3.7
358       Name == "avx2.vinserti128" || // Added in 3.7
359       Name.startswith("avx512.mask.insert") || // Added in 4.0
360       Name.startswith("avx.vextractf128.") || // Added in 3.7
361       Name == "avx2.vextracti128" || // Added in 3.7
362       Name.startswith("avx512.mask.vextract") || // Added in 4.0
363       Name.startswith("sse4a.movnt.") || // Added in 3.9
364       Name.startswith("avx.movnt.") || // Added in 3.2
365       Name.startswith("avx512.storent.") || // Added in 3.9
366       Name == "sse41.movntdqa" || // Added in 5.0
367       Name == "avx2.movntdqa" || // Added in 5.0
368       Name == "avx512.movntdqa" || // Added in 5.0
369       Name == "sse2.storel.dq" || // Added in 3.9
370       Name.startswith("sse.storeu.") || // Added in 3.9
371       Name.startswith("sse2.storeu.") || // Added in 3.9
372       Name.startswith("avx.storeu.") || // Added in 3.9
373       Name.startswith("avx512.mask.storeu.") || // Added in 3.9
374       Name.startswith("avx512.mask.store.p") || // Added in 3.9
375       Name.startswith("avx512.mask.store.b.") || // Added in 3.9
376       Name.startswith("avx512.mask.store.w.") || // Added in 3.9
377       Name.startswith("avx512.mask.store.d.") || // Added in 3.9
378       Name.startswith("avx512.mask.store.q.") || // Added in 3.9
379       Name == "avx512.mask.store.ss" || // Added in 7.0
380       Name.startswith("avx512.mask.loadu.") || // Added in 3.9
381       Name.startswith("avx512.mask.load.") || // Added in 3.9
382       Name.startswith("avx512.mask.expand.load.") || // Added in 7.0
383       Name.startswith("avx512.mask.compress.store.") || // Added in 7.0
384       Name.startswith("avx512.mask.expand.b") || // Added in 9.0
385       Name.startswith("avx512.mask.expand.w") || // Added in 9.0
386       Name.startswith("avx512.mask.expand.d") || // Added in 9.0
387       Name.startswith("avx512.mask.expand.q") || // Added in 9.0
388       Name.startswith("avx512.mask.expand.p") || // Added in 9.0
389       Name.startswith("avx512.mask.compress.b") || // Added in 9.0
390       Name.startswith("avx512.mask.compress.w") || // Added in 9.0
391       Name.startswith("avx512.mask.compress.d") || // Added in 9.0
392       Name.startswith("avx512.mask.compress.q") || // Added in 9.0
393       Name.startswith("avx512.mask.compress.p") || // Added in 9.0
394       Name == "sse42.crc32.64.8" || // Added in 3.4
395       Name.startswith("avx.vbroadcast.s") || // Added in 3.5
396       Name.startswith("avx512.vbroadcast.s") || // Added in 7.0
397       Name.startswith("avx512.mask.palignr.") || // Added in 3.9
398       Name.startswith("avx512.mask.valign.") || // Added in 4.0
399       Name.startswith("sse2.psll.dq") || // Added in 3.7
400       Name.startswith("sse2.psrl.dq") || // Added in 3.7
401       Name.startswith("avx2.psll.dq") || // Added in 3.7
402       Name.startswith("avx2.psrl.dq") || // Added in 3.7
403       Name.startswith("avx512.psll.dq") || // Added in 3.9
404       Name.startswith("avx512.psrl.dq") || // Added in 3.9
405       Name == "sse41.pblendw" || // Added in 3.7
406       Name.startswith("sse41.blendp") || // Added in 3.7
407       Name.startswith("avx.blend.p") || // Added in 3.7
408       Name == "avx2.pblendw" || // Added in 3.7
409       Name.startswith("avx2.pblendd.") || // Added in 3.7
410       Name.startswith("avx.vbroadcastf128") || // Added in 4.0
411       Name == "avx2.vbroadcasti128" || // Added in 3.7
412       Name.startswith("avx512.mask.broadcastf32x4.") || // Added in 6.0
413       Name.startswith("avx512.mask.broadcastf64x2.") || // Added in 6.0
414       Name.startswith("avx512.mask.broadcastf32x8.") || // Added in 6.0
415       Name.startswith("avx512.mask.broadcastf64x4.") || // Added in 6.0
416       Name.startswith("avx512.mask.broadcasti32x4.") || // Added in 6.0
417       Name.startswith("avx512.mask.broadcasti64x2.") || // Added in 6.0
418       Name.startswith("avx512.mask.broadcasti32x8.") || // Added in 6.0
419       Name.startswith("avx512.mask.broadcasti64x4.") || // Added in 6.0
420       Name == "xop.vpcmov" || // Added in 3.8
421       Name == "xop.vpcmov.256" || // Added in 5.0
422       Name.startswith("avx512.mask.move.s") || // Added in 4.0
423       Name.startswith("avx512.cvtmask2") || // Added in 5.0
424       Name.startswith("xop.vpcom") || // Added in 3.2, Updated in 9.0
425       Name.startswith("xop.vprot") || // Added in 8.0
426       Name.startswith("avx512.prol") || // Added in 8.0
427       Name.startswith("avx512.pror") || // Added in 8.0
428       Name.startswith("avx512.mask.prorv.") || // Added in 8.0
429       Name.startswith("avx512.mask.pror.") ||  // Added in 8.0
430       Name.startswith("avx512.mask.prolv.") || // Added in 8.0
431       Name.startswith("avx512.mask.prol.") ||  // Added in 8.0
432       Name.startswith("avx512.ptestm") || //Added in 6.0
433       Name.startswith("avx512.ptestnm") || //Added in 6.0
434       Name.startswith("avx512.mask.pavg")) // Added in 6.0
435     return true;
436 
437   return false;
438 }
439 
UpgradeX86IntrinsicFunction(Function * F,StringRef Name,Function * & NewFn)440 static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
441                                         Function *&NewFn) {
442   // Only handle intrinsics that start with "x86.".
443   if (!Name.startswith("x86."))
444     return false;
445   // Remove "x86." prefix.
446   Name = Name.substr(4);
447 
448   if (ShouldUpgradeX86Intrinsic(F, Name)) {
449     NewFn = nullptr;
450     return true;
451   }
452 
453   if (Name == "rdtscp") { // Added in 8.0
454     // If this intrinsic has 0 operands, it's the new version.
455     if (F->getFunctionType()->getNumParams() == 0)
456       return false;
457 
458     rename(F);
459     NewFn = Intrinsic::getDeclaration(F->getParent(),
460                                       Intrinsic::x86_rdtscp);
461     return true;
462   }
463 
464   // SSE4.1 ptest functions may have an old signature.
465   if (Name.startswith("sse41.ptest")) { // Added in 3.2
466     if (Name.substr(11) == "c")
467       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestc, NewFn);
468     if (Name.substr(11) == "z")
469       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestz, NewFn);
470     if (Name.substr(11) == "nzc")
471       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
472   }
473   // Several blend and other instructions with masks used the wrong number of
474   // bits.
475   if (Name == "sse41.insertps") // Added in 3.6
476     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
477                                             NewFn);
478   if (Name == "sse41.dppd") // Added in 3.6
479     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
480                                             NewFn);
481   if (Name == "sse41.dpps") // Added in 3.6
482     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
483                                             NewFn);
484   if (Name == "sse41.mpsadbw") // Added in 3.6
485     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
486                                             NewFn);
487   if (Name == "avx.dp.ps.256") // Added in 3.6
488     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
489                                             NewFn);
490   if (Name == "avx2.mpsadbw") // Added in 3.6
491     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
492                                             NewFn);
493   if (Name == "avx512.mask.cmp.pd.128") // Added in 7.0
494     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_128,
495                                      NewFn);
496   if (Name == "avx512.mask.cmp.pd.256") // Added in 7.0
497     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_256,
498                                      NewFn);
499   if (Name == "avx512.mask.cmp.pd.512") // Added in 7.0
500     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_512,
501                                      NewFn);
502   if (Name == "avx512.mask.cmp.ps.128") // Added in 7.0
503     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_128,
504                                      NewFn);
505   if (Name == "avx512.mask.cmp.ps.256") // Added in 7.0
506     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_256,
507                                      NewFn);
508   if (Name == "avx512.mask.cmp.ps.512") // Added in 7.0
509     return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_512,
510                                      NewFn);
511   if (Name == "avx512bf16.cvtne2ps2bf16.128") // Added in 9.0
512     return UpgradeX86BF16Intrinsic(
513         F, Intrinsic::x86_avx512bf16_cvtne2ps2bf16_128, NewFn);
514   if (Name == "avx512bf16.cvtne2ps2bf16.256") // Added in 9.0
515     return UpgradeX86BF16Intrinsic(
516         F, Intrinsic::x86_avx512bf16_cvtne2ps2bf16_256, NewFn);
517   if (Name == "avx512bf16.cvtne2ps2bf16.512") // Added in 9.0
518     return UpgradeX86BF16Intrinsic(
519         F, Intrinsic::x86_avx512bf16_cvtne2ps2bf16_512, NewFn);
520   if (Name == "avx512bf16.mask.cvtneps2bf16.128") // Added in 9.0
521     return UpgradeX86BF16Intrinsic(
522         F, Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128, NewFn);
523   if (Name == "avx512bf16.cvtneps2bf16.256") // Added in 9.0
524     return UpgradeX86BF16Intrinsic(
525         F, Intrinsic::x86_avx512bf16_cvtneps2bf16_256, NewFn);
526   if (Name == "avx512bf16.cvtneps2bf16.512") // Added in 9.0
527     return UpgradeX86BF16Intrinsic(
528         F, Intrinsic::x86_avx512bf16_cvtneps2bf16_512, NewFn);
529   if (Name == "avx512bf16.dpbf16ps.128") // Added in 9.0
530     return UpgradeX86BF16DPIntrinsic(
531         F, Intrinsic::x86_avx512bf16_dpbf16ps_128, NewFn);
532   if (Name == "avx512bf16.dpbf16ps.256") // Added in 9.0
533     return UpgradeX86BF16DPIntrinsic(
534         F, Intrinsic::x86_avx512bf16_dpbf16ps_256, NewFn);
535   if (Name == "avx512bf16.dpbf16ps.512") // Added in 9.0
536     return UpgradeX86BF16DPIntrinsic(
537         F, Intrinsic::x86_avx512bf16_dpbf16ps_512, NewFn);
538 
539   // frcz.ss/sd may need to have an argument dropped. Added in 3.2
540   if (Name.startswith("xop.vfrcz.ss") && F->arg_size() == 2) {
541     rename(F);
542     NewFn = Intrinsic::getDeclaration(F->getParent(),
543                                       Intrinsic::x86_xop_vfrcz_ss);
544     return true;
545   }
546   if (Name.startswith("xop.vfrcz.sd") && F->arg_size() == 2) {
547     rename(F);
548     NewFn = Intrinsic::getDeclaration(F->getParent(),
549                                       Intrinsic::x86_xop_vfrcz_sd);
550     return true;
551   }
552   // Upgrade any XOP PERMIL2 index operand still using a float/double vector.
553   if (Name.startswith("xop.vpermil2")) { // Added in 3.9
554     auto Idx = F->getFunctionType()->getParamType(2);
555     if (Idx->isFPOrFPVectorTy()) {
556       rename(F);
557       unsigned IdxSize = Idx->getPrimitiveSizeInBits();
558       unsigned EltSize = Idx->getScalarSizeInBits();
559       Intrinsic::ID Permil2ID;
560       if (EltSize == 64 && IdxSize == 128)
561         Permil2ID = Intrinsic::x86_xop_vpermil2pd;
562       else if (EltSize == 32 && IdxSize == 128)
563         Permil2ID = Intrinsic::x86_xop_vpermil2ps;
564       else if (EltSize == 64 && IdxSize == 256)
565         Permil2ID = Intrinsic::x86_xop_vpermil2pd_256;
566       else
567         Permil2ID = Intrinsic::x86_xop_vpermil2ps_256;
568       NewFn = Intrinsic::getDeclaration(F->getParent(), Permil2ID);
569       return true;
570     }
571   }
572 
573   if (Name == "seh.recoverfp") {
574     NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::eh_recoverfp);
575     return true;
576   }
577 
578   return false;
579 }
580 
UpgradeIntrinsicFunction1(Function * F,Function * & NewFn)581 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
582   assert(F && "Illegal to upgrade a non-existent Function.");
583 
584   // Quickly eliminate it, if it's not a candidate.
585   StringRef Name = F->getName();
586   if (Name.size() <= 7 || !Name.startswith("llvm."))
587     return false;
588   Name = Name.substr(5); // Strip off "llvm."
589 
590   switch (Name[0]) {
591   default: break;
592   case 'a': {
593     if (Name.startswith("arm.rbit") || Name.startswith("aarch64.rbit")) {
594       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
595                                         F->arg_begin()->getType());
596       return true;
597     }
598     if (Name.startswith("aarch64.neon.frintn")) {
599       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::roundeven,
600                                         F->arg_begin()->getType());
601       return true;
602     }
603     if (Name.startswith("aarch64.neon.rbit")) {
604       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
605                                         F->arg_begin()->getType());
606       return true;
607     }
608     if (Name == "aarch64.sve.bfdot.lane") {
609       NewFn = Intrinsic::getDeclaration(F->getParent(),
610                                         Intrinsic::aarch64_sve_bfdot_lane_v2);
611       return true;
612     }
613     if (Name == "aarch64.sve.bfmlalb.lane") {
614       NewFn = Intrinsic::getDeclaration(F->getParent(),
615                                         Intrinsic::aarch64_sve_bfmlalb_lane_v2);
616       return true;
617     }
618     if (Name == "aarch64.sve.bfmlalt.lane") {
619       NewFn = Intrinsic::getDeclaration(F->getParent(),
620                                         Intrinsic::aarch64_sve_bfmlalt_lane_v2);
621       return true;
622     }
623     static const Regex LdRegex("^aarch64\\.sve\\.ld[234](.nxv[a-z0-9]+|$)");
624     if (LdRegex.match(Name)) {
625       Type *ScalarTy =
626           dyn_cast<VectorType>(F->getReturnType())->getElementType();
627       ElementCount EC =
628           dyn_cast<VectorType>(F->arg_begin()->getType())->getElementCount();
629       Type *Ty = VectorType::get(ScalarTy, EC);
630       Intrinsic::ID ID =
631           StringSwitch<Intrinsic::ID>(Name)
632               .StartsWith("aarch64.sve.ld2", Intrinsic::aarch64_sve_ld2_sret)
633               .StartsWith("aarch64.sve.ld3", Intrinsic::aarch64_sve_ld3_sret)
634               .StartsWith("aarch64.sve.ld4", Intrinsic::aarch64_sve_ld4_sret)
635               .Default(Intrinsic::not_intrinsic);
636       NewFn = Intrinsic::getDeclaration(F->getParent(), ID, Ty);
637       return true;
638     }
639     if (Name.startswith("aarch64.sve.tuple.get")) {
640       Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
641       NewFn = Intrinsic::getDeclaration(F->getParent(),
642                                         Intrinsic::vector_extract, Tys);
643       return true;
644     }
645     if (Name.startswith("aarch64.sve.tuple.set")) {
646       auto Args = F->getFunctionType()->params();
647       Type *Tys[] = {Args[0], Args[2], Args[1]};
648       NewFn = Intrinsic::getDeclaration(F->getParent(),
649                                         Intrinsic::vector_insert, Tys);
650       return true;
651     }
652     static const Regex CreateTupleRegex(
653         "^aarch64\\.sve\\.tuple\\.create[234](.nxv[a-z0-9]+|$)");
654     if (CreateTupleRegex.match(Name)) {
655       auto Args = F->getFunctionType()->params();
656       Type *Tys[] = {F->getReturnType(), Args[1]};
657       NewFn = Intrinsic::getDeclaration(F->getParent(),
658                                         Intrinsic::vector_insert, Tys);
659       return true;
660     }
661     if (Name.startswith("arm.neon.vclz")) {
662       Type* args[2] = {
663         F->arg_begin()->getType(),
664         Type::getInt1Ty(F->getContext())
665       };
666       // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
667       // the end of the name. Change name from llvm.arm.neon.vclz.* to
668       //  llvm.ctlz.*
669       FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
670       NewFn = Function::Create(fType, F->getLinkage(), F->getAddressSpace(),
671                                "llvm.ctlz." + Name.substr(14), F->getParent());
672       return true;
673     }
674     if (Name.startswith("arm.neon.vcnt")) {
675       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
676                                         F->arg_begin()->getType());
677       return true;
678     }
679     static const Regex vstRegex("^arm\\.neon\\.vst([1234]|[234]lane)\\.v[a-z0-9]*$");
680     if (vstRegex.match(Name)) {
681       static const Intrinsic::ID StoreInts[] = {Intrinsic::arm_neon_vst1,
682                                                 Intrinsic::arm_neon_vst2,
683                                                 Intrinsic::arm_neon_vst3,
684                                                 Intrinsic::arm_neon_vst4};
685 
686       static const Intrinsic::ID StoreLaneInts[] = {
687         Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
688         Intrinsic::arm_neon_vst4lane
689       };
690 
691       auto fArgs = F->getFunctionType()->params();
692       Type *Tys[] = {fArgs[0], fArgs[1]};
693       if (!Name.contains("lane"))
694         NewFn = Intrinsic::getDeclaration(F->getParent(),
695                                           StoreInts[fArgs.size() - 3], Tys);
696       else
697         NewFn = Intrinsic::getDeclaration(F->getParent(),
698                                           StoreLaneInts[fArgs.size() - 5], Tys);
699       return true;
700     }
701     if (Name == "aarch64.thread.pointer" || Name == "arm.thread.pointer") {
702       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
703       return true;
704     }
705     if (Name.startswith("arm.neon.vqadds.")) {
706       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::sadd_sat,
707                                         F->arg_begin()->getType());
708       return true;
709     }
710     if (Name.startswith("arm.neon.vqaddu.")) {
711       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::uadd_sat,
712                                         F->arg_begin()->getType());
713       return true;
714     }
715     if (Name.startswith("arm.neon.vqsubs.")) {
716       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ssub_sat,
717                                         F->arg_begin()->getType());
718       return true;
719     }
720     if (Name.startswith("arm.neon.vqsubu.")) {
721       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::usub_sat,
722                                         F->arg_begin()->getType());
723       return true;
724     }
725     if (Name.startswith("aarch64.neon.addp")) {
726       if (F->arg_size() != 2)
727         break; // Invalid IR.
728       VectorType *Ty = dyn_cast<VectorType>(F->getReturnType());
729       if (Ty && Ty->getElementType()->isFloatingPointTy()) {
730         NewFn = Intrinsic::getDeclaration(F->getParent(),
731                                           Intrinsic::aarch64_neon_faddp, Ty);
732         return true;
733       }
734     }
735 
736     // Changed in 12.0: bfdot accept v4bf16 and v8bf16 instead of v8i8 and v16i8
737     // respectively
738     if ((Name.startswith("arm.neon.bfdot.") ||
739          Name.startswith("aarch64.neon.bfdot.")) &&
740         Name.endswith("i8")) {
741       Intrinsic::ID IID =
742           StringSwitch<Intrinsic::ID>(Name)
743               .Cases("arm.neon.bfdot.v2f32.v8i8",
744                      "arm.neon.bfdot.v4f32.v16i8",
745                      Intrinsic::arm_neon_bfdot)
746               .Cases("aarch64.neon.bfdot.v2f32.v8i8",
747                      "aarch64.neon.bfdot.v4f32.v16i8",
748                      Intrinsic::aarch64_neon_bfdot)
749               .Default(Intrinsic::not_intrinsic);
750       if (IID == Intrinsic::not_intrinsic)
751         break;
752 
753       size_t OperandWidth = F->getReturnType()->getPrimitiveSizeInBits();
754       assert((OperandWidth == 64 || OperandWidth == 128) &&
755              "Unexpected operand width");
756       LLVMContext &Ctx = F->getParent()->getContext();
757       std::array<Type *, 2> Tys {{
758         F->getReturnType(),
759         FixedVectorType::get(Type::getBFloatTy(Ctx), OperandWidth / 16)
760       }};
761       NewFn = Intrinsic::getDeclaration(F->getParent(), IID, Tys);
762       return true;
763     }
764 
765     // Changed in 12.0: bfmmla, bfmlalb and bfmlalt are not polymorphic anymore
766     // and accept v8bf16 instead of v16i8
767     if ((Name.startswith("arm.neon.bfm") ||
768          Name.startswith("aarch64.neon.bfm")) &&
769         Name.endswith(".v4f32.v16i8")) {
770       Intrinsic::ID IID =
771           StringSwitch<Intrinsic::ID>(Name)
772               .Case("arm.neon.bfmmla.v4f32.v16i8",
773                     Intrinsic::arm_neon_bfmmla)
774               .Case("arm.neon.bfmlalb.v4f32.v16i8",
775                     Intrinsic::arm_neon_bfmlalb)
776               .Case("arm.neon.bfmlalt.v4f32.v16i8",
777                     Intrinsic::arm_neon_bfmlalt)
778               .Case("aarch64.neon.bfmmla.v4f32.v16i8",
779                     Intrinsic::aarch64_neon_bfmmla)
780               .Case("aarch64.neon.bfmlalb.v4f32.v16i8",
781                     Intrinsic::aarch64_neon_bfmlalb)
782               .Case("aarch64.neon.bfmlalt.v4f32.v16i8",
783                     Intrinsic::aarch64_neon_bfmlalt)
784               .Default(Intrinsic::not_intrinsic);
785       if (IID == Intrinsic::not_intrinsic)
786         break;
787 
788       std::array<Type *, 0> Tys;
789       NewFn = Intrinsic::getDeclaration(F->getParent(), IID, Tys);
790       return true;
791     }
792 
793     if (Name == "arm.mve.vctp64" &&
794         cast<FixedVectorType>(F->getReturnType())->getNumElements() == 4) {
795       // A vctp64 returning a v4i1 is converted to return a v2i1. Rename the
796       // function and deal with it below in UpgradeIntrinsicCall.
797       rename(F);
798       return true;
799     }
800     // These too are changed to accept a v2i1 insteead of the old v4i1.
801     if (Name == "arm.mve.mull.int.predicated.v2i64.v4i32.v4i1" ||
802         Name == "arm.mve.vqdmull.predicated.v2i64.v4i32.v4i1" ||
803         Name == "arm.mve.vldr.gather.base.predicated.v2i64.v2i64.v4i1" ||
804         Name == "arm.mve.vldr.gather.base.wb.predicated.v2i64.v2i64.v4i1" ||
805         Name == "arm.mve.vldr.gather.offset.predicated.v2i64.p0i64.v2i64.v4i1" ||
806         Name == "arm.mve.vstr.scatter.base.predicated.v2i64.v2i64.v4i1" ||
807         Name == "arm.mve.vstr.scatter.base.wb.predicated.v2i64.v2i64.v4i1" ||
808         Name == "arm.mve.vstr.scatter.offset.predicated.p0i64.v2i64.v2i64.v4i1" ||
809         Name == "arm.cde.vcx1q.predicated.v2i64.v4i1" ||
810         Name == "arm.cde.vcx1qa.predicated.v2i64.v4i1" ||
811         Name == "arm.cde.vcx2q.predicated.v2i64.v4i1" ||
812         Name == "arm.cde.vcx2qa.predicated.v2i64.v4i1" ||
813         Name == "arm.cde.vcx3q.predicated.v2i64.v4i1" ||
814         Name == "arm.cde.vcx3qa.predicated.v2i64.v4i1")
815       return true;
816 
817     if (Name == "amdgcn.alignbit") {
818       // Target specific intrinsic became redundant
819       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::fshr,
820                                         {F->getReturnType()});
821       return true;
822     }
823 
824     break;
825   }
826 
827   case 'c': {
828     if (Name.startswith("ctlz.") && F->arg_size() == 1) {
829       rename(F);
830       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
831                                         F->arg_begin()->getType());
832       return true;
833     }
834     if (Name.startswith("cttz.") && F->arg_size() == 1) {
835       rename(F);
836       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
837                                         F->arg_begin()->getType());
838       return true;
839     }
840     break;
841   }
842   case 'd': {
843     if (Name == "dbg.value" && F->arg_size() == 4) {
844       rename(F);
845       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_value);
846       return true;
847     }
848     break;
849   }
850   case 'e': {
851     if (Name.startswith("experimental.vector.extract.")) {
852       rename(F);
853       Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
854       NewFn = Intrinsic::getDeclaration(F->getParent(),
855                                         Intrinsic::vector_extract, Tys);
856       return true;
857     }
858 
859     if (Name.startswith("experimental.vector.insert.")) {
860       rename(F);
861       auto Args = F->getFunctionType()->params();
862       Type *Tys[] = {Args[0], Args[1]};
863       NewFn = Intrinsic::getDeclaration(F->getParent(),
864                                         Intrinsic::vector_insert, Tys);
865       return true;
866     }
867 
868     SmallVector<StringRef, 2> Groups;
869     static const Regex R("^experimental.vector.reduce.([a-z]+)\\.[a-z][0-9]+");
870     if (R.match(Name, &Groups)) {
871       Intrinsic::ID ID;
872       ID = StringSwitch<Intrinsic::ID>(Groups[1])
873                .Case("add", Intrinsic::vector_reduce_add)
874                .Case("mul", Intrinsic::vector_reduce_mul)
875                .Case("and", Intrinsic::vector_reduce_and)
876                .Case("or", Intrinsic::vector_reduce_or)
877                .Case("xor", Intrinsic::vector_reduce_xor)
878                .Case("smax", Intrinsic::vector_reduce_smax)
879                .Case("smin", Intrinsic::vector_reduce_smin)
880                .Case("umax", Intrinsic::vector_reduce_umax)
881                .Case("umin", Intrinsic::vector_reduce_umin)
882                .Case("fmax", Intrinsic::vector_reduce_fmax)
883                .Case("fmin", Intrinsic::vector_reduce_fmin)
884                .Default(Intrinsic::not_intrinsic);
885       if (ID != Intrinsic::not_intrinsic) {
886         rename(F);
887         auto Args = F->getFunctionType()->params();
888         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, {Args[0]});
889         return true;
890       }
891     }
892     static const Regex R2(
893         "^experimental.vector.reduce.v2.([a-z]+)\\.[fi][0-9]+");
894     Groups.clear();
895     if (R2.match(Name, &Groups)) {
896       Intrinsic::ID ID = Intrinsic::not_intrinsic;
897       if (Groups[1] == "fadd")
898         ID = Intrinsic::vector_reduce_fadd;
899       if (Groups[1] == "fmul")
900         ID = Intrinsic::vector_reduce_fmul;
901       if (ID != Intrinsic::not_intrinsic) {
902         rename(F);
903         auto Args = F->getFunctionType()->params();
904         Type *Tys[] = {Args[1]};
905         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, Tys);
906         return true;
907       }
908     }
909     break;
910   }
911   case 'f':
912     if (Name.startswith("flt.rounds")) {
913       rename(F);
914       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::get_rounding);
915       return true;
916     }
917     break;
918   case 'i':
919   case 'l': {
920     bool IsLifetimeStart = Name.startswith("lifetime.start");
921     if (IsLifetimeStart || Name.startswith("invariant.start")) {
922       Intrinsic::ID ID = IsLifetimeStart ?
923         Intrinsic::lifetime_start : Intrinsic::invariant_start;
924       auto Args = F->getFunctionType()->params();
925       Type* ObjectPtr[1] = {Args[1]};
926       if (F->getName() != Intrinsic::getName(ID, ObjectPtr, F->getParent())) {
927         rename(F);
928         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
929         return true;
930       }
931     }
932 
933     bool IsLifetimeEnd = Name.startswith("lifetime.end");
934     if (IsLifetimeEnd || Name.startswith("invariant.end")) {
935       Intrinsic::ID ID = IsLifetimeEnd ?
936         Intrinsic::lifetime_end : Intrinsic::invariant_end;
937 
938       auto Args = F->getFunctionType()->params();
939       Type* ObjectPtr[1] = {Args[IsLifetimeEnd ? 1 : 2]};
940       if (F->getName() != Intrinsic::getName(ID, ObjectPtr, F->getParent())) {
941         rename(F);
942         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
943         return true;
944       }
945     }
946     if (Name.startswith("invariant.group.barrier")) {
947       // Rename invariant.group.barrier to launder.invariant.group
948       auto Args = F->getFunctionType()->params();
949       Type* ObjectPtr[1] = {Args[0]};
950       rename(F);
951       NewFn = Intrinsic::getDeclaration(F->getParent(),
952           Intrinsic::launder_invariant_group, ObjectPtr);
953       return true;
954 
955     }
956 
957     break;
958   }
959   case 'm': {
960     if (Name.startswith("masked.load.")) {
961       Type *Tys[] = { F->getReturnType(), F->arg_begin()->getType() };
962       if (F->getName() !=
963           Intrinsic::getName(Intrinsic::masked_load, Tys, F->getParent())) {
964         rename(F);
965         NewFn = Intrinsic::getDeclaration(F->getParent(),
966                                           Intrinsic::masked_load,
967                                           Tys);
968         return true;
969       }
970     }
971     if (Name.startswith("masked.store.")) {
972       auto Args = F->getFunctionType()->params();
973       Type *Tys[] = { Args[0], Args[1] };
974       if (F->getName() !=
975           Intrinsic::getName(Intrinsic::masked_store, Tys, F->getParent())) {
976         rename(F);
977         NewFn = Intrinsic::getDeclaration(F->getParent(),
978                                           Intrinsic::masked_store,
979                                           Tys);
980         return true;
981       }
982     }
983     // Renaming gather/scatter intrinsics with no address space overloading
984     // to the new overload which includes an address space
985     if (Name.startswith("masked.gather.")) {
986       Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
987       if (F->getName() !=
988           Intrinsic::getName(Intrinsic::masked_gather, Tys, F->getParent())) {
989         rename(F);
990         NewFn = Intrinsic::getDeclaration(F->getParent(),
991                                           Intrinsic::masked_gather, Tys);
992         return true;
993       }
994     }
995     if (Name.startswith("masked.scatter.")) {
996       auto Args = F->getFunctionType()->params();
997       Type *Tys[] = {Args[0], Args[1]};
998       if (F->getName() !=
999           Intrinsic::getName(Intrinsic::masked_scatter, Tys, F->getParent())) {
1000         rename(F);
1001         NewFn = Intrinsic::getDeclaration(F->getParent(),
1002                                           Intrinsic::masked_scatter, Tys);
1003         return true;
1004       }
1005     }
1006     // Updating the memory intrinsics (memcpy/memmove/memset) that have an
1007     // alignment parameter to embedding the alignment as an attribute of
1008     // the pointer args.
1009     if (Name.startswith("memcpy.") && F->arg_size() == 5) {
1010       rename(F);
1011       // Get the types of dest, src, and len
1012       ArrayRef<Type *> ParamTypes = F->getFunctionType()->params().slice(0, 3);
1013       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memcpy,
1014                                         ParamTypes);
1015       return true;
1016     }
1017     if (Name.startswith("memmove.") && F->arg_size() == 5) {
1018       rename(F);
1019       // Get the types of dest, src, and len
1020       ArrayRef<Type *> ParamTypes = F->getFunctionType()->params().slice(0, 3);
1021       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memmove,
1022                                         ParamTypes);
1023       return true;
1024     }
1025     if (Name.startswith("memset.") && F->arg_size() == 5) {
1026       rename(F);
1027       // Get the types of dest, and len
1028       const auto *FT = F->getFunctionType();
1029       Type *ParamTypes[2] = {
1030           FT->getParamType(0), // Dest
1031           FT->getParamType(2)  // len
1032       };
1033       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memset,
1034                                         ParamTypes);
1035       return true;
1036     }
1037     break;
1038   }
1039   case 'n': {
1040     if (Name.startswith("nvvm.")) {
1041       Name = Name.substr(5);
1042 
1043       // The following nvvm intrinsics correspond exactly to an LLVM intrinsic.
1044       Intrinsic::ID IID = StringSwitch<Intrinsic::ID>(Name)
1045                               .Cases("brev32", "brev64", Intrinsic::bitreverse)
1046                               .Case("clz.i", Intrinsic::ctlz)
1047                               .Case("popc.i", Intrinsic::ctpop)
1048                               .Default(Intrinsic::not_intrinsic);
1049       if (IID != Intrinsic::not_intrinsic && F->arg_size() == 1) {
1050         NewFn = Intrinsic::getDeclaration(F->getParent(), IID,
1051                                           {F->getReturnType()});
1052         return true;
1053       }
1054 
1055       // The following nvvm intrinsics correspond exactly to an LLVM idiom, but
1056       // not to an intrinsic alone.  We expand them in UpgradeIntrinsicCall.
1057       //
1058       // TODO: We could add lohi.i2d.
1059       bool Expand = StringSwitch<bool>(Name)
1060                         .Cases("abs.i", "abs.ll", true)
1061                         .Cases("clz.ll", "popc.ll", "h2f", true)
1062                         .Cases("max.i", "max.ll", "max.ui", "max.ull", true)
1063                         .Cases("min.i", "min.ll", "min.ui", "min.ull", true)
1064                         .StartsWith("atomic.load.add.f32.p", true)
1065                         .StartsWith("atomic.load.add.f64.p", true)
1066                         .Default(false);
1067       if (Expand) {
1068         NewFn = nullptr;
1069         return true;
1070       }
1071     }
1072     break;
1073   }
1074   case 'o':
1075     // We only need to change the name to match the mangling including the
1076     // address space.
1077     if (Name.startswith("objectsize.")) {
1078       Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
1079       if (F->arg_size() == 2 || F->arg_size() == 3 ||
1080           F->getName() !=
1081               Intrinsic::getName(Intrinsic::objectsize, Tys, F->getParent())) {
1082         rename(F);
1083         NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::objectsize,
1084                                           Tys);
1085         return true;
1086       }
1087     }
1088     break;
1089 
1090   case 'p':
1091     if (Name == "prefetch") {
1092       // Handle address space overloading.
1093       Type *Tys[] = {F->arg_begin()->getType()};
1094       if (F->getName() !=
1095           Intrinsic::getName(Intrinsic::prefetch, Tys, F->getParent())) {
1096         rename(F);
1097         NewFn =
1098             Intrinsic::getDeclaration(F->getParent(), Intrinsic::prefetch, Tys);
1099         return true;
1100       }
1101     } else if (Name.startswith("ptr.annotation.") && F->arg_size() == 4) {
1102       rename(F);
1103       NewFn = Intrinsic::getDeclaration(
1104           F->getParent(), Intrinsic::ptr_annotation,
1105           {F->arg_begin()->getType(), F->getArg(1)->getType()});
1106       return true;
1107     }
1108     break;
1109 
1110   case 's':
1111     if (Name == "stackprotectorcheck") {
1112       NewFn = nullptr;
1113       return true;
1114     }
1115     break;
1116 
1117   case 'v': {
1118     if (Name == "var.annotation" && F->arg_size() == 4) {
1119       rename(F);
1120       NewFn = Intrinsic::getDeclaration(
1121           F->getParent(), Intrinsic::var_annotation,
1122           {{F->arg_begin()->getType(), F->getArg(1)->getType()}});
1123       return true;
1124     }
1125     break;
1126   }
1127 
1128   case 'x':
1129     if (UpgradeX86IntrinsicFunction(F, Name, NewFn))
1130       return true;
1131   }
1132 
1133   auto *ST = dyn_cast<StructType>(F->getReturnType());
1134   if (ST && (!ST->isLiteral() || ST->isPacked())) {
1135     // Replace return type with literal non-packed struct. Only do this for
1136     // intrinsics declared to return a struct, not for intrinsics with
1137     // overloaded return type, in which case the exact struct type will be
1138     // mangled into the name.
1139     SmallVector<Intrinsic::IITDescriptor> Desc;
1140     Intrinsic::getIntrinsicInfoTableEntries(F->getIntrinsicID(), Desc);
1141     if (Desc.front().Kind == Intrinsic::IITDescriptor::Struct) {
1142       auto *FT = F->getFunctionType();
1143       auto *NewST = StructType::get(ST->getContext(), ST->elements());
1144       auto *NewFT = FunctionType::get(NewST, FT->params(), FT->isVarArg());
1145       std::string Name = F->getName().str();
1146       rename(F);
1147       NewFn = Function::Create(NewFT, F->getLinkage(), F->getAddressSpace(),
1148                                Name, F->getParent());
1149 
1150       // The new function may also need remangling.
1151       if (auto Result = llvm::Intrinsic::remangleIntrinsicFunction(NewFn))
1152         NewFn = *Result;
1153       return true;
1154     }
1155   }
1156 
1157   // Remangle our intrinsic since we upgrade the mangling
1158   auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F);
1159   if (Result != std::nullopt) {
1160     NewFn = *Result;
1161     return true;
1162   }
1163 
1164   //  This may not belong here. This function is effectively being overloaded
1165   //  to both detect an intrinsic which needs upgrading, and to provide the
1166   //  upgraded form of the intrinsic. We should perhaps have two separate
1167   //  functions for this.
1168   return false;
1169 }
1170 
UpgradeIntrinsicFunction(Function * F,Function * & NewFn)1171 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
1172   NewFn = nullptr;
1173   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
1174   assert(F != NewFn && "Intrinsic function upgraded to the same function");
1175 
1176   // Upgrade intrinsic attributes.  This does not change the function.
1177   if (NewFn)
1178     F = NewFn;
1179   if (Intrinsic::ID id = F->getIntrinsicID())
1180     F->setAttributes(Intrinsic::getAttributes(F->getContext(), id));
1181   return Upgraded;
1182 }
1183 
UpgradeGlobalVariable(GlobalVariable * GV)1184 GlobalVariable *llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
1185   if (!(GV->hasName() && (GV->getName() == "llvm.global_ctors" ||
1186                           GV->getName() == "llvm.global_dtors")) ||
1187       !GV->hasInitializer())
1188     return nullptr;
1189   ArrayType *ATy = dyn_cast<ArrayType>(GV->getValueType());
1190   if (!ATy)
1191     return nullptr;
1192   StructType *STy = dyn_cast<StructType>(ATy->getElementType());
1193   if (!STy || STy->getNumElements() != 2)
1194     return nullptr;
1195 
1196   LLVMContext &C = GV->getContext();
1197   IRBuilder<> IRB(C);
1198   auto EltTy = StructType::get(STy->getElementType(0), STy->getElementType(1),
1199                                IRB.getInt8PtrTy());
1200   Constant *Init = GV->getInitializer();
1201   unsigned N = Init->getNumOperands();
1202   std::vector<Constant *> NewCtors(N);
1203   for (unsigned i = 0; i != N; ++i) {
1204     auto Ctor = cast<Constant>(Init->getOperand(i));
1205     NewCtors[i] = ConstantStruct::get(
1206         EltTy, Ctor->getAggregateElement(0u), Ctor->getAggregateElement(1),
1207         Constant::getNullValue(IRB.getInt8PtrTy()));
1208   }
1209   Constant *NewInit = ConstantArray::get(ArrayType::get(EltTy, N), NewCtors);
1210 
1211   return new GlobalVariable(NewInit->getType(), false, GV->getLinkage(),
1212                             NewInit, GV->getName());
1213 }
1214 
1215 // Handles upgrading SSE2/AVX2/AVX512BW PSLLDQ intrinsics by converting them
1216 // to byte shuffles.
UpgradeX86PSLLDQIntrinsics(IRBuilder<> & Builder,Value * Op,unsigned Shift)1217 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder,
1218                                          Value *Op, unsigned Shift) {
1219   auto *ResultTy = cast<FixedVectorType>(Op->getType());
1220   unsigned NumElts = ResultTy->getNumElements() * 8;
1221 
1222   // Bitcast from a 64-bit element type to a byte element type.
1223   Type *VecTy = FixedVectorType::get(Builder.getInt8Ty(), NumElts);
1224   Op = Builder.CreateBitCast(Op, VecTy, "cast");
1225 
1226   // We'll be shuffling in zeroes.
1227   Value *Res = Constant::getNullValue(VecTy);
1228 
1229   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
1230   // we'll just return the zero vector.
1231   if (Shift < 16) {
1232     int Idxs[64];
1233     // 256/512-bit version is split into 2/4 16-byte lanes.
1234     for (unsigned l = 0; l != NumElts; l += 16)
1235       for (unsigned i = 0; i != 16; ++i) {
1236         unsigned Idx = NumElts + i - Shift;
1237         if (Idx < NumElts)
1238           Idx -= NumElts - 16; // end of lane, switch operand.
1239         Idxs[l + i] = Idx + l;
1240       }
1241 
1242     Res = Builder.CreateShuffleVector(Res, Op, ArrayRef(Idxs, NumElts));
1243   }
1244 
1245   // Bitcast back to a 64-bit element type.
1246   return Builder.CreateBitCast(Res, ResultTy, "cast");
1247 }
1248 
1249 // Handles upgrading SSE2/AVX2/AVX512BW PSRLDQ intrinsics by converting them
1250 // to byte shuffles.
UpgradeX86PSRLDQIntrinsics(IRBuilder<> & Builder,Value * Op,unsigned Shift)1251 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, Value *Op,
1252                                          unsigned Shift) {
1253   auto *ResultTy = cast<FixedVectorType>(Op->getType());
1254   unsigned NumElts = ResultTy->getNumElements() * 8;
1255 
1256   // Bitcast from a 64-bit element type to a byte element type.
1257   Type *VecTy = FixedVectorType::get(Builder.getInt8Ty(), NumElts);
1258   Op = Builder.CreateBitCast(Op, VecTy, "cast");
1259 
1260   // We'll be shuffling in zeroes.
1261   Value *Res = Constant::getNullValue(VecTy);
1262 
1263   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
1264   // we'll just return the zero vector.
1265   if (Shift < 16) {
1266     int Idxs[64];
1267     // 256/512-bit version is split into 2/4 16-byte lanes.
1268     for (unsigned l = 0; l != NumElts; l += 16)
1269       for (unsigned i = 0; i != 16; ++i) {
1270         unsigned Idx = i + Shift;
1271         if (Idx >= 16)
1272           Idx += NumElts - 16; // end of lane, switch operand.
1273         Idxs[l + i] = Idx + l;
1274       }
1275 
1276     Res = Builder.CreateShuffleVector(Op, Res, ArrayRef(Idxs, NumElts));
1277   }
1278 
1279   // Bitcast back to a 64-bit element type.
1280   return Builder.CreateBitCast(Res, ResultTy, "cast");
1281 }
1282 
getX86MaskVec(IRBuilder<> & Builder,Value * Mask,unsigned NumElts)1283 static Value *getX86MaskVec(IRBuilder<> &Builder, Value *Mask,
1284                             unsigned NumElts) {
1285   assert(isPowerOf2_32(NumElts) && "Expected power-of-2 mask elements");
1286   llvm::VectorType *MaskTy = FixedVectorType::get(
1287       Builder.getInt1Ty(), cast<IntegerType>(Mask->getType())->getBitWidth());
1288   Mask = Builder.CreateBitCast(Mask, MaskTy);
1289 
1290   // If we have less than 8 elements (1, 2 or 4), then the starting mask was an
1291   // i8 and we need to extract down to the right number of elements.
1292   if (NumElts <= 4) {
1293     int Indices[4];
1294     for (unsigned i = 0; i != NumElts; ++i)
1295       Indices[i] = i;
1296     Mask = Builder.CreateShuffleVector(Mask, Mask, ArrayRef(Indices, NumElts),
1297                                        "extract");
1298   }
1299 
1300   return Mask;
1301 }
1302 
EmitX86Select(IRBuilder<> & Builder,Value * Mask,Value * Op0,Value * Op1)1303 static Value *EmitX86Select(IRBuilder<> &Builder, Value *Mask,
1304                             Value *Op0, Value *Op1) {
1305   // If the mask is all ones just emit the first operation.
1306   if (const auto *C = dyn_cast<Constant>(Mask))
1307     if (C->isAllOnesValue())
1308       return Op0;
1309 
1310   Mask = getX86MaskVec(Builder, Mask,
1311                        cast<FixedVectorType>(Op0->getType())->getNumElements());
1312   return Builder.CreateSelect(Mask, Op0, Op1);
1313 }
1314 
EmitX86ScalarSelect(IRBuilder<> & Builder,Value * Mask,Value * Op0,Value * Op1)1315 static Value *EmitX86ScalarSelect(IRBuilder<> &Builder, Value *Mask,
1316                                   Value *Op0, Value *Op1) {
1317   // If the mask is all ones just emit the first operation.
1318   if (const auto *C = dyn_cast<Constant>(Mask))
1319     if (C->isAllOnesValue())
1320       return Op0;
1321 
1322   auto *MaskTy = FixedVectorType::get(Builder.getInt1Ty(),
1323                                       Mask->getType()->getIntegerBitWidth());
1324   Mask = Builder.CreateBitCast(Mask, MaskTy);
1325   Mask = Builder.CreateExtractElement(Mask, (uint64_t)0);
1326   return Builder.CreateSelect(Mask, Op0, Op1);
1327 }
1328 
1329 // Handle autoupgrade for masked PALIGNR and VALIGND/Q intrinsics.
1330 // PALIGNR handles large immediates by shifting while VALIGN masks the immediate
1331 // so we need to handle both cases. VALIGN also doesn't have 128-bit lanes.
UpgradeX86ALIGNIntrinsics(IRBuilder<> & Builder,Value * Op0,Value * Op1,Value * Shift,Value * Passthru,Value * Mask,bool IsVALIGN)1332 static Value *UpgradeX86ALIGNIntrinsics(IRBuilder<> &Builder, Value *Op0,
1333                                         Value *Op1, Value *Shift,
1334                                         Value *Passthru, Value *Mask,
1335                                         bool IsVALIGN) {
1336   unsigned ShiftVal = cast<llvm::ConstantInt>(Shift)->getZExtValue();
1337 
1338   unsigned NumElts = cast<FixedVectorType>(Op0->getType())->getNumElements();
1339   assert((IsVALIGN || NumElts % 16 == 0) && "Illegal NumElts for PALIGNR!");
1340   assert((!IsVALIGN || NumElts <= 16) && "NumElts too large for VALIGN!");
1341   assert(isPowerOf2_32(NumElts) && "NumElts not a power of 2!");
1342 
1343   // Mask the immediate for VALIGN.
1344   if (IsVALIGN)
1345     ShiftVal &= (NumElts - 1);
1346 
1347   // If palignr is shifting the pair of vectors more than the size of two
1348   // lanes, emit zero.
1349   if (ShiftVal >= 32)
1350     return llvm::Constant::getNullValue(Op0->getType());
1351 
1352   // If palignr is shifting the pair of input vectors more than one lane,
1353   // but less than two lanes, convert to shifting in zeroes.
1354   if (ShiftVal > 16) {
1355     ShiftVal -= 16;
1356     Op1 = Op0;
1357     Op0 = llvm::Constant::getNullValue(Op0->getType());
1358   }
1359 
1360   int Indices[64];
1361   // 256-bit palignr operates on 128-bit lanes so we need to handle that
1362   for (unsigned l = 0; l < NumElts; l += 16) {
1363     for (unsigned i = 0; i != 16; ++i) {
1364       unsigned Idx = ShiftVal + i;
1365       if (!IsVALIGN && Idx >= 16) // Disable wrap for VALIGN.
1366         Idx += NumElts - 16; // End of lane, switch operand.
1367       Indices[l + i] = Idx + l;
1368     }
1369   }
1370 
1371   Value *Align = Builder.CreateShuffleVector(
1372       Op1, Op0, ArrayRef(Indices, NumElts), "palignr");
1373 
1374   return EmitX86Select(Builder, Mask, Align, Passthru);
1375 }
1376 
UpgradeX86VPERMT2Intrinsics(IRBuilder<> & Builder,CallBase & CI,bool ZeroMask,bool IndexForm)1377 static Value *UpgradeX86VPERMT2Intrinsics(IRBuilder<> &Builder, CallBase &CI,
1378                                           bool ZeroMask, bool IndexForm) {
1379   Type *Ty = CI.getType();
1380   unsigned VecWidth = Ty->getPrimitiveSizeInBits();
1381   unsigned EltWidth = Ty->getScalarSizeInBits();
1382   bool IsFloat = Ty->isFPOrFPVectorTy();
1383   Intrinsic::ID IID;
1384   if (VecWidth == 128 && EltWidth == 32 && IsFloat)
1385     IID = Intrinsic::x86_avx512_vpermi2var_ps_128;
1386   else if (VecWidth == 128 && EltWidth == 32 && !IsFloat)
1387     IID = Intrinsic::x86_avx512_vpermi2var_d_128;
1388   else if (VecWidth == 128 && EltWidth == 64 && IsFloat)
1389     IID = Intrinsic::x86_avx512_vpermi2var_pd_128;
1390   else if (VecWidth == 128 && EltWidth == 64 && !IsFloat)
1391     IID = Intrinsic::x86_avx512_vpermi2var_q_128;
1392   else if (VecWidth == 256 && EltWidth == 32 && IsFloat)
1393     IID = Intrinsic::x86_avx512_vpermi2var_ps_256;
1394   else if (VecWidth == 256 && EltWidth == 32 && !IsFloat)
1395     IID = Intrinsic::x86_avx512_vpermi2var_d_256;
1396   else if (VecWidth == 256 && EltWidth == 64 && IsFloat)
1397     IID = Intrinsic::x86_avx512_vpermi2var_pd_256;
1398   else if (VecWidth == 256 && EltWidth == 64 && !IsFloat)
1399     IID = Intrinsic::x86_avx512_vpermi2var_q_256;
1400   else if (VecWidth == 512 && EltWidth == 32 && IsFloat)
1401     IID = Intrinsic::x86_avx512_vpermi2var_ps_512;
1402   else if (VecWidth == 512 && EltWidth == 32 && !IsFloat)
1403     IID = Intrinsic::x86_avx512_vpermi2var_d_512;
1404   else if (VecWidth == 512 && EltWidth == 64 && IsFloat)
1405     IID = Intrinsic::x86_avx512_vpermi2var_pd_512;
1406   else if (VecWidth == 512 && EltWidth == 64 && !IsFloat)
1407     IID = Intrinsic::x86_avx512_vpermi2var_q_512;
1408   else if (VecWidth == 128 && EltWidth == 16)
1409     IID = Intrinsic::x86_avx512_vpermi2var_hi_128;
1410   else if (VecWidth == 256 && EltWidth == 16)
1411     IID = Intrinsic::x86_avx512_vpermi2var_hi_256;
1412   else if (VecWidth == 512 && EltWidth == 16)
1413     IID = Intrinsic::x86_avx512_vpermi2var_hi_512;
1414   else if (VecWidth == 128 && EltWidth == 8)
1415     IID = Intrinsic::x86_avx512_vpermi2var_qi_128;
1416   else if (VecWidth == 256 && EltWidth == 8)
1417     IID = Intrinsic::x86_avx512_vpermi2var_qi_256;
1418   else if (VecWidth == 512 && EltWidth == 8)
1419     IID = Intrinsic::x86_avx512_vpermi2var_qi_512;
1420   else
1421     llvm_unreachable("Unexpected intrinsic");
1422 
1423   Value *Args[] = { CI.getArgOperand(0) , CI.getArgOperand(1),
1424                     CI.getArgOperand(2) };
1425 
1426   // If this isn't index form we need to swap operand 0 and 1.
1427   if (!IndexForm)
1428     std::swap(Args[0], Args[1]);
1429 
1430   Value *V = Builder.CreateCall(Intrinsic::getDeclaration(CI.getModule(), IID),
1431                                 Args);
1432   Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty)
1433                              : Builder.CreateBitCast(CI.getArgOperand(1),
1434                                                      Ty);
1435   return EmitX86Select(Builder, CI.getArgOperand(3), V, PassThru);
1436 }
1437 
UpgradeX86BinaryIntrinsics(IRBuilder<> & Builder,CallBase & CI,Intrinsic::ID IID)1438 static Value *UpgradeX86BinaryIntrinsics(IRBuilder<> &Builder, CallBase &CI,
1439                                          Intrinsic::ID IID) {
1440   Type *Ty = CI.getType();
1441   Value *Op0 = CI.getOperand(0);
1442   Value *Op1 = CI.getOperand(1);
1443   Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID, Ty);
1444   Value *Res = Builder.CreateCall(Intrin, {Op0, Op1});
1445 
1446   if (CI.arg_size() == 4) { // For masked intrinsics.
1447     Value *VecSrc = CI.getOperand(2);
1448     Value *Mask = CI.getOperand(3);
1449     Res = EmitX86Select(Builder, Mask, Res, VecSrc);
1450   }
1451   return Res;
1452 }
1453 
upgradeX86Rotate(IRBuilder<> & Builder,CallBase & CI,bool IsRotateRight)1454 static Value *upgradeX86Rotate(IRBuilder<> &Builder, CallBase &CI,
1455                                bool IsRotateRight) {
1456   Type *Ty = CI.getType();
1457   Value *Src = CI.getArgOperand(0);
1458   Value *Amt = CI.getArgOperand(1);
1459 
1460   // Amount may be scalar immediate, in which case create a splat vector.
1461   // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
1462   // we only care about the lowest log2 bits anyway.
1463   if (Amt->getType() != Ty) {
1464     unsigned NumElts = cast<FixedVectorType>(Ty)->getNumElements();
1465     Amt = Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
1466     Amt = Builder.CreateVectorSplat(NumElts, Amt);
1467   }
1468 
1469   Intrinsic::ID IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl;
1470   Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID, Ty);
1471   Value *Res = Builder.CreateCall(Intrin, {Src, Src, Amt});
1472 
1473   if (CI.arg_size() == 4) { // For masked intrinsics.
1474     Value *VecSrc = CI.getOperand(2);
1475     Value *Mask = CI.getOperand(3);
1476     Res = EmitX86Select(Builder, Mask, Res, VecSrc);
1477   }
1478   return Res;
1479 }
1480 
upgradeX86vpcom(IRBuilder<> & Builder,CallBase & CI,unsigned Imm,bool IsSigned)1481 static Value *upgradeX86vpcom(IRBuilder<> &Builder, CallBase &CI, unsigned Imm,
1482                               bool IsSigned) {
1483   Type *Ty = CI.getType();
1484   Value *LHS = CI.getArgOperand(0);
1485   Value *RHS = CI.getArgOperand(1);
1486 
1487   CmpInst::Predicate Pred;
1488   switch (Imm) {
1489   case 0x0:
1490     Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1491     break;
1492   case 0x1:
1493     Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1494     break;
1495   case 0x2:
1496     Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1497     break;
1498   case 0x3:
1499     Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1500     break;
1501   case 0x4:
1502     Pred = ICmpInst::ICMP_EQ;
1503     break;
1504   case 0x5:
1505     Pred = ICmpInst::ICMP_NE;
1506     break;
1507   case 0x6:
1508     return Constant::getNullValue(Ty); // FALSE
1509   case 0x7:
1510     return Constant::getAllOnesValue(Ty); // TRUE
1511   default:
1512     llvm_unreachable("Unknown XOP vpcom/vpcomu predicate");
1513   }
1514 
1515   Value *Cmp = Builder.CreateICmp(Pred, LHS, RHS);
1516   Value *Ext = Builder.CreateSExt(Cmp, Ty);
1517   return Ext;
1518 }
1519 
upgradeX86ConcatShift(IRBuilder<> & Builder,CallBase & CI,bool IsShiftRight,bool ZeroMask)1520 static Value *upgradeX86ConcatShift(IRBuilder<> &Builder, CallBase &CI,
1521                                     bool IsShiftRight, bool ZeroMask) {
1522   Type *Ty = CI.getType();
1523   Value *Op0 = CI.getArgOperand(0);
1524   Value *Op1 = CI.getArgOperand(1);
1525   Value *Amt = CI.getArgOperand(2);
1526 
1527   if (IsShiftRight)
1528     std::swap(Op0, Op1);
1529 
1530   // Amount may be scalar immediate, in which case create a splat vector.
1531   // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
1532   // we only care about the lowest log2 bits anyway.
1533   if (Amt->getType() != Ty) {
1534     unsigned NumElts = cast<FixedVectorType>(Ty)->getNumElements();
1535     Amt = Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
1536     Amt = Builder.CreateVectorSplat(NumElts, Amt);
1537   }
1538 
1539   Intrinsic::ID IID = IsShiftRight ? Intrinsic::fshr : Intrinsic::fshl;
1540   Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID, Ty);
1541   Value *Res = Builder.CreateCall(Intrin, {Op0, Op1, Amt});
1542 
1543   unsigned NumArgs = CI.arg_size();
1544   if (NumArgs >= 4) { // For masked intrinsics.
1545     Value *VecSrc = NumArgs == 5 ? CI.getArgOperand(3) :
1546                     ZeroMask     ? ConstantAggregateZero::get(CI.getType()) :
1547                                    CI.getArgOperand(0);
1548     Value *Mask = CI.getOperand(NumArgs - 1);
1549     Res = EmitX86Select(Builder, Mask, Res, VecSrc);
1550   }
1551   return Res;
1552 }
1553 
UpgradeMaskedStore(IRBuilder<> & Builder,Value * Ptr,Value * Data,Value * Mask,bool Aligned)1554 static Value *UpgradeMaskedStore(IRBuilder<> &Builder,
1555                                  Value *Ptr, Value *Data, Value *Mask,
1556                                  bool Aligned) {
1557   // Cast the pointer to the right type.
1558   Ptr = Builder.CreateBitCast(Ptr,
1559                               llvm::PointerType::getUnqual(Data->getType()));
1560   const Align Alignment =
1561       Aligned
1562           ? Align(Data->getType()->getPrimitiveSizeInBits().getFixedValue() / 8)
1563           : Align(1);
1564 
1565   // If the mask is all ones just emit a regular store.
1566   if (const auto *C = dyn_cast<Constant>(Mask))
1567     if (C->isAllOnesValue())
1568       return Builder.CreateAlignedStore(Data, Ptr, Alignment);
1569 
1570   // Convert the mask from an integer type to a vector of i1.
1571   unsigned NumElts = cast<FixedVectorType>(Data->getType())->getNumElements();
1572   Mask = getX86MaskVec(Builder, Mask, NumElts);
1573   return Builder.CreateMaskedStore(Data, Ptr, Alignment, Mask);
1574 }
1575 
UpgradeMaskedLoad(IRBuilder<> & Builder,Value * Ptr,Value * Passthru,Value * Mask,bool Aligned)1576 static Value *UpgradeMaskedLoad(IRBuilder<> &Builder,
1577                                 Value *Ptr, Value *Passthru, Value *Mask,
1578                                 bool Aligned) {
1579   Type *ValTy = Passthru->getType();
1580   // Cast the pointer to the right type.
1581   Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(ValTy));
1582   const Align Alignment =
1583       Aligned
1584           ? Align(
1585                 Passthru->getType()->getPrimitiveSizeInBits().getFixedValue() /
1586                 8)
1587           : Align(1);
1588 
1589   // If the mask is all ones just emit a regular store.
1590   if (const auto *C = dyn_cast<Constant>(Mask))
1591     if (C->isAllOnesValue())
1592       return Builder.CreateAlignedLoad(ValTy, Ptr, Alignment);
1593 
1594   // Convert the mask from an integer type to a vector of i1.
1595   unsigned NumElts = cast<FixedVectorType>(ValTy)->getNumElements();
1596   Mask = getX86MaskVec(Builder, Mask, NumElts);
1597   return Builder.CreateMaskedLoad(ValTy, Ptr, Alignment, Mask, Passthru);
1598 }
1599 
upgradeAbs(IRBuilder<> & Builder,CallBase & CI)1600 static Value *upgradeAbs(IRBuilder<> &Builder, CallBase &CI) {
1601   Type *Ty = CI.getType();
1602   Value *Op0 = CI.getArgOperand(0);
1603   Function *F = Intrinsic::getDeclaration(CI.getModule(), Intrinsic::abs, Ty);
1604   Value *Res = Builder.CreateCall(F, {Op0, Builder.getInt1(false)});
1605   if (CI.arg_size() == 3)
1606     Res = EmitX86Select(Builder, CI.getArgOperand(2), Res, CI.getArgOperand(1));
1607   return Res;
1608 }
1609 
upgradePMULDQ(IRBuilder<> & Builder,CallBase & CI,bool IsSigned)1610 static Value *upgradePMULDQ(IRBuilder<> &Builder, CallBase &CI, bool IsSigned) {
1611   Type *Ty = CI.getType();
1612 
1613   // Arguments have a vXi32 type so cast to vXi64.
1614   Value *LHS = Builder.CreateBitCast(CI.getArgOperand(0), Ty);
1615   Value *RHS = Builder.CreateBitCast(CI.getArgOperand(1), Ty);
1616 
1617   if (IsSigned) {
1618     // Shift left then arithmetic shift right.
1619     Constant *ShiftAmt = ConstantInt::get(Ty, 32);
1620     LHS = Builder.CreateShl(LHS, ShiftAmt);
1621     LHS = Builder.CreateAShr(LHS, ShiftAmt);
1622     RHS = Builder.CreateShl(RHS, ShiftAmt);
1623     RHS = Builder.CreateAShr(RHS, ShiftAmt);
1624   } else {
1625     // Clear the upper bits.
1626     Constant *Mask = ConstantInt::get(Ty, 0xffffffff);
1627     LHS = Builder.CreateAnd(LHS, Mask);
1628     RHS = Builder.CreateAnd(RHS, Mask);
1629   }
1630 
1631   Value *Res = Builder.CreateMul(LHS, RHS);
1632 
1633   if (CI.arg_size() == 4)
1634     Res = EmitX86Select(Builder, CI.getArgOperand(3), Res, CI.getArgOperand(2));
1635 
1636   return Res;
1637 }
1638 
1639 // Applying mask on vector of i1's and make sure result is at least 8 bits wide.
ApplyX86MaskOn1BitsVec(IRBuilder<> & Builder,Value * Vec,Value * Mask)1640 static Value *ApplyX86MaskOn1BitsVec(IRBuilder<> &Builder, Value *Vec,
1641                                      Value *Mask) {
1642   unsigned NumElts = cast<FixedVectorType>(Vec->getType())->getNumElements();
1643   if (Mask) {
1644     const auto *C = dyn_cast<Constant>(Mask);
1645     if (!C || !C->isAllOnesValue())
1646       Vec = Builder.CreateAnd(Vec, getX86MaskVec(Builder, Mask, NumElts));
1647   }
1648 
1649   if (NumElts < 8) {
1650     int Indices[8];
1651     for (unsigned i = 0; i != NumElts; ++i)
1652       Indices[i] = i;
1653     for (unsigned i = NumElts; i != 8; ++i)
1654       Indices[i] = NumElts + i % NumElts;
1655     Vec = Builder.CreateShuffleVector(Vec,
1656                                       Constant::getNullValue(Vec->getType()),
1657                                       Indices);
1658   }
1659   return Builder.CreateBitCast(Vec, Builder.getIntNTy(std::max(NumElts, 8U)));
1660 }
1661 
upgradeMaskedCompare(IRBuilder<> & Builder,CallBase & CI,unsigned CC,bool Signed)1662 static Value *upgradeMaskedCompare(IRBuilder<> &Builder, CallBase &CI,
1663                                    unsigned CC, bool Signed) {
1664   Value *Op0 = CI.getArgOperand(0);
1665   unsigned NumElts = cast<FixedVectorType>(Op0->getType())->getNumElements();
1666 
1667   Value *Cmp;
1668   if (CC == 3) {
1669     Cmp = Constant::getNullValue(
1670         FixedVectorType::get(Builder.getInt1Ty(), NumElts));
1671   } else if (CC == 7) {
1672     Cmp = Constant::getAllOnesValue(
1673         FixedVectorType::get(Builder.getInt1Ty(), NumElts));
1674   } else {
1675     ICmpInst::Predicate Pred;
1676     switch (CC) {
1677     default: llvm_unreachable("Unknown condition code");
1678     case 0: Pred = ICmpInst::ICMP_EQ;  break;
1679     case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
1680     case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
1681     case 4: Pred = ICmpInst::ICMP_NE;  break;
1682     case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
1683     case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
1684     }
1685     Cmp = Builder.CreateICmp(Pred, Op0, CI.getArgOperand(1));
1686   }
1687 
1688   Value *Mask = CI.getArgOperand(CI.arg_size() - 1);
1689 
1690   return ApplyX86MaskOn1BitsVec(Builder, Cmp, Mask);
1691 }
1692 
1693 // Replace a masked intrinsic with an older unmasked intrinsic.
UpgradeX86MaskedShift(IRBuilder<> & Builder,CallBase & CI,Intrinsic::ID IID)1694 static Value *UpgradeX86MaskedShift(IRBuilder<> &Builder, CallBase &CI,
1695                                     Intrinsic::ID IID) {
1696   Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID);
1697   Value *Rep = Builder.CreateCall(Intrin,
1698                                  { CI.getArgOperand(0), CI.getArgOperand(1) });
1699   return EmitX86Select(Builder, CI.getArgOperand(3), Rep, CI.getArgOperand(2));
1700 }
1701 
upgradeMaskedMove(IRBuilder<> & Builder,CallBase & CI)1702 static Value* upgradeMaskedMove(IRBuilder<> &Builder, CallBase &CI) {
1703   Value* A = CI.getArgOperand(0);
1704   Value* B = CI.getArgOperand(1);
1705   Value* Src = CI.getArgOperand(2);
1706   Value* Mask = CI.getArgOperand(3);
1707 
1708   Value* AndNode = Builder.CreateAnd(Mask, APInt(8, 1));
1709   Value* Cmp = Builder.CreateIsNotNull(AndNode);
1710   Value* Extract1 = Builder.CreateExtractElement(B, (uint64_t)0);
1711   Value* Extract2 = Builder.CreateExtractElement(Src, (uint64_t)0);
1712   Value* Select = Builder.CreateSelect(Cmp, Extract1, Extract2);
1713   return Builder.CreateInsertElement(A, Select, (uint64_t)0);
1714 }
1715 
1716 
UpgradeMaskToInt(IRBuilder<> & Builder,CallBase & CI)1717 static Value* UpgradeMaskToInt(IRBuilder<> &Builder, CallBase &CI) {
1718   Value* Op = CI.getArgOperand(0);
1719   Type* ReturnOp = CI.getType();
1720   unsigned NumElts = cast<FixedVectorType>(CI.getType())->getNumElements();
1721   Value *Mask = getX86MaskVec(Builder, Op, NumElts);
1722   return Builder.CreateSExt(Mask, ReturnOp, "vpmovm2");
1723 }
1724 
1725 // Replace intrinsic with unmasked version and a select.
upgradeAVX512MaskToSelect(StringRef Name,IRBuilder<> & Builder,CallBase & CI,Value * & Rep)1726 static bool upgradeAVX512MaskToSelect(StringRef Name, IRBuilder<> &Builder,
1727                                       CallBase &CI, Value *&Rep) {
1728   Name = Name.substr(12); // Remove avx512.mask.
1729 
1730   unsigned VecWidth = CI.getType()->getPrimitiveSizeInBits();
1731   unsigned EltWidth = CI.getType()->getScalarSizeInBits();
1732   Intrinsic::ID IID;
1733   if (Name.startswith("max.p")) {
1734     if (VecWidth == 128 && EltWidth == 32)
1735       IID = Intrinsic::x86_sse_max_ps;
1736     else if (VecWidth == 128 && EltWidth == 64)
1737       IID = Intrinsic::x86_sse2_max_pd;
1738     else if (VecWidth == 256 && EltWidth == 32)
1739       IID = Intrinsic::x86_avx_max_ps_256;
1740     else if (VecWidth == 256 && EltWidth == 64)
1741       IID = Intrinsic::x86_avx_max_pd_256;
1742     else
1743       llvm_unreachable("Unexpected intrinsic");
1744   } else if (Name.startswith("min.p")) {
1745     if (VecWidth == 128 && EltWidth == 32)
1746       IID = Intrinsic::x86_sse_min_ps;
1747     else if (VecWidth == 128 && EltWidth == 64)
1748       IID = Intrinsic::x86_sse2_min_pd;
1749     else if (VecWidth == 256 && EltWidth == 32)
1750       IID = Intrinsic::x86_avx_min_ps_256;
1751     else if (VecWidth == 256 && EltWidth == 64)
1752       IID = Intrinsic::x86_avx_min_pd_256;
1753     else
1754       llvm_unreachable("Unexpected intrinsic");
1755   } else if (Name.startswith("pshuf.b.")) {
1756     if (VecWidth == 128)
1757       IID = Intrinsic::x86_ssse3_pshuf_b_128;
1758     else if (VecWidth == 256)
1759       IID = Intrinsic::x86_avx2_pshuf_b;
1760     else if (VecWidth == 512)
1761       IID = Intrinsic::x86_avx512_pshuf_b_512;
1762     else
1763       llvm_unreachable("Unexpected intrinsic");
1764   } else if (Name.startswith("pmul.hr.sw.")) {
1765     if (VecWidth == 128)
1766       IID = Intrinsic::x86_ssse3_pmul_hr_sw_128;
1767     else if (VecWidth == 256)
1768       IID = Intrinsic::x86_avx2_pmul_hr_sw;
1769     else if (VecWidth == 512)
1770       IID = Intrinsic::x86_avx512_pmul_hr_sw_512;
1771     else
1772       llvm_unreachable("Unexpected intrinsic");
1773   } else if (Name.startswith("pmulh.w.")) {
1774     if (VecWidth == 128)
1775       IID = Intrinsic::x86_sse2_pmulh_w;
1776     else if (VecWidth == 256)
1777       IID = Intrinsic::x86_avx2_pmulh_w;
1778     else if (VecWidth == 512)
1779       IID = Intrinsic::x86_avx512_pmulh_w_512;
1780     else
1781       llvm_unreachable("Unexpected intrinsic");
1782   } else if (Name.startswith("pmulhu.w.")) {
1783     if (VecWidth == 128)
1784       IID = Intrinsic::x86_sse2_pmulhu_w;
1785     else if (VecWidth == 256)
1786       IID = Intrinsic::x86_avx2_pmulhu_w;
1787     else if (VecWidth == 512)
1788       IID = Intrinsic::x86_avx512_pmulhu_w_512;
1789     else
1790       llvm_unreachable("Unexpected intrinsic");
1791   } else if (Name.startswith("pmaddw.d.")) {
1792     if (VecWidth == 128)
1793       IID = Intrinsic::x86_sse2_pmadd_wd;
1794     else if (VecWidth == 256)
1795       IID = Intrinsic::x86_avx2_pmadd_wd;
1796     else if (VecWidth == 512)
1797       IID = Intrinsic::x86_avx512_pmaddw_d_512;
1798     else
1799       llvm_unreachable("Unexpected intrinsic");
1800   } else if (Name.startswith("pmaddubs.w.")) {
1801     if (VecWidth == 128)
1802       IID = Intrinsic::x86_ssse3_pmadd_ub_sw_128;
1803     else if (VecWidth == 256)
1804       IID = Intrinsic::x86_avx2_pmadd_ub_sw;
1805     else if (VecWidth == 512)
1806       IID = Intrinsic::x86_avx512_pmaddubs_w_512;
1807     else
1808       llvm_unreachable("Unexpected intrinsic");
1809   } else if (Name.startswith("packsswb.")) {
1810     if (VecWidth == 128)
1811       IID = Intrinsic::x86_sse2_packsswb_128;
1812     else if (VecWidth == 256)
1813       IID = Intrinsic::x86_avx2_packsswb;
1814     else if (VecWidth == 512)
1815       IID = Intrinsic::x86_avx512_packsswb_512;
1816     else
1817       llvm_unreachable("Unexpected intrinsic");
1818   } else if (Name.startswith("packssdw.")) {
1819     if (VecWidth == 128)
1820       IID = Intrinsic::x86_sse2_packssdw_128;
1821     else if (VecWidth == 256)
1822       IID = Intrinsic::x86_avx2_packssdw;
1823     else if (VecWidth == 512)
1824       IID = Intrinsic::x86_avx512_packssdw_512;
1825     else
1826       llvm_unreachable("Unexpected intrinsic");
1827   } else if (Name.startswith("packuswb.")) {
1828     if (VecWidth == 128)
1829       IID = Intrinsic::x86_sse2_packuswb_128;
1830     else if (VecWidth == 256)
1831       IID = Intrinsic::x86_avx2_packuswb;
1832     else if (VecWidth == 512)
1833       IID = Intrinsic::x86_avx512_packuswb_512;
1834     else
1835       llvm_unreachable("Unexpected intrinsic");
1836   } else if (Name.startswith("packusdw.")) {
1837     if (VecWidth == 128)
1838       IID = Intrinsic::x86_sse41_packusdw;
1839     else if (VecWidth == 256)
1840       IID = Intrinsic::x86_avx2_packusdw;
1841     else if (VecWidth == 512)
1842       IID = Intrinsic::x86_avx512_packusdw_512;
1843     else
1844       llvm_unreachable("Unexpected intrinsic");
1845   } else if (Name.startswith("vpermilvar.")) {
1846     if (VecWidth == 128 && EltWidth == 32)
1847       IID = Intrinsic::x86_avx_vpermilvar_ps;
1848     else if (VecWidth == 128 && EltWidth == 64)
1849       IID = Intrinsic::x86_avx_vpermilvar_pd;
1850     else if (VecWidth == 256 && EltWidth == 32)
1851       IID = Intrinsic::x86_avx_vpermilvar_ps_256;
1852     else if (VecWidth == 256 && EltWidth == 64)
1853       IID = Intrinsic::x86_avx_vpermilvar_pd_256;
1854     else if (VecWidth == 512 && EltWidth == 32)
1855       IID = Intrinsic::x86_avx512_vpermilvar_ps_512;
1856     else if (VecWidth == 512 && EltWidth == 64)
1857       IID = Intrinsic::x86_avx512_vpermilvar_pd_512;
1858     else
1859       llvm_unreachable("Unexpected intrinsic");
1860   } else if (Name == "cvtpd2dq.256") {
1861     IID = Intrinsic::x86_avx_cvt_pd2dq_256;
1862   } else if (Name == "cvtpd2ps.256") {
1863     IID = Intrinsic::x86_avx_cvt_pd2_ps_256;
1864   } else if (Name == "cvttpd2dq.256") {
1865     IID = Intrinsic::x86_avx_cvtt_pd2dq_256;
1866   } else if (Name == "cvttps2dq.128") {
1867     IID = Intrinsic::x86_sse2_cvttps2dq;
1868   } else if (Name == "cvttps2dq.256") {
1869     IID = Intrinsic::x86_avx_cvtt_ps2dq_256;
1870   } else if (Name.startswith("permvar.")) {
1871     bool IsFloat = CI.getType()->isFPOrFPVectorTy();
1872     if (VecWidth == 256 && EltWidth == 32 && IsFloat)
1873       IID = Intrinsic::x86_avx2_permps;
1874     else if (VecWidth == 256 && EltWidth == 32 && !IsFloat)
1875       IID = Intrinsic::x86_avx2_permd;
1876     else if (VecWidth == 256 && EltWidth == 64 && IsFloat)
1877       IID = Intrinsic::x86_avx512_permvar_df_256;
1878     else if (VecWidth == 256 && EltWidth == 64 && !IsFloat)
1879       IID = Intrinsic::x86_avx512_permvar_di_256;
1880     else if (VecWidth == 512 && EltWidth == 32 && IsFloat)
1881       IID = Intrinsic::x86_avx512_permvar_sf_512;
1882     else if (VecWidth == 512 && EltWidth == 32 && !IsFloat)
1883       IID = Intrinsic::x86_avx512_permvar_si_512;
1884     else if (VecWidth == 512 && EltWidth == 64 && IsFloat)
1885       IID = Intrinsic::x86_avx512_permvar_df_512;
1886     else if (VecWidth == 512 && EltWidth == 64 && !IsFloat)
1887       IID = Intrinsic::x86_avx512_permvar_di_512;
1888     else if (VecWidth == 128 && EltWidth == 16)
1889       IID = Intrinsic::x86_avx512_permvar_hi_128;
1890     else if (VecWidth == 256 && EltWidth == 16)
1891       IID = Intrinsic::x86_avx512_permvar_hi_256;
1892     else if (VecWidth == 512 && EltWidth == 16)
1893       IID = Intrinsic::x86_avx512_permvar_hi_512;
1894     else if (VecWidth == 128 && EltWidth == 8)
1895       IID = Intrinsic::x86_avx512_permvar_qi_128;
1896     else if (VecWidth == 256 && EltWidth == 8)
1897       IID = Intrinsic::x86_avx512_permvar_qi_256;
1898     else if (VecWidth == 512 && EltWidth == 8)
1899       IID = Intrinsic::x86_avx512_permvar_qi_512;
1900     else
1901       llvm_unreachable("Unexpected intrinsic");
1902   } else if (Name.startswith("dbpsadbw.")) {
1903     if (VecWidth == 128)
1904       IID = Intrinsic::x86_avx512_dbpsadbw_128;
1905     else if (VecWidth == 256)
1906       IID = Intrinsic::x86_avx512_dbpsadbw_256;
1907     else if (VecWidth == 512)
1908       IID = Intrinsic::x86_avx512_dbpsadbw_512;
1909     else
1910       llvm_unreachable("Unexpected intrinsic");
1911   } else if (Name.startswith("pmultishift.qb.")) {
1912     if (VecWidth == 128)
1913       IID = Intrinsic::x86_avx512_pmultishift_qb_128;
1914     else if (VecWidth == 256)
1915       IID = Intrinsic::x86_avx512_pmultishift_qb_256;
1916     else if (VecWidth == 512)
1917       IID = Intrinsic::x86_avx512_pmultishift_qb_512;
1918     else
1919       llvm_unreachable("Unexpected intrinsic");
1920   } else if (Name.startswith("conflict.")) {
1921     if (Name[9] == 'd' && VecWidth == 128)
1922       IID = Intrinsic::x86_avx512_conflict_d_128;
1923     else if (Name[9] == 'd' && VecWidth == 256)
1924       IID = Intrinsic::x86_avx512_conflict_d_256;
1925     else if (Name[9] == 'd' && VecWidth == 512)
1926       IID = Intrinsic::x86_avx512_conflict_d_512;
1927     else if (Name[9] == 'q' && VecWidth == 128)
1928       IID = Intrinsic::x86_avx512_conflict_q_128;
1929     else if (Name[9] == 'q' && VecWidth == 256)
1930       IID = Intrinsic::x86_avx512_conflict_q_256;
1931     else if (Name[9] == 'q' && VecWidth == 512)
1932       IID = Intrinsic::x86_avx512_conflict_q_512;
1933     else
1934       llvm_unreachable("Unexpected intrinsic");
1935   } else if (Name.startswith("pavg.")) {
1936     if (Name[5] == 'b' && VecWidth == 128)
1937       IID = Intrinsic::x86_sse2_pavg_b;
1938     else if (Name[5] == 'b' && VecWidth == 256)
1939       IID = Intrinsic::x86_avx2_pavg_b;
1940     else if (Name[5] == 'b' && VecWidth == 512)
1941       IID = Intrinsic::x86_avx512_pavg_b_512;
1942     else if (Name[5] == 'w' && VecWidth == 128)
1943       IID = Intrinsic::x86_sse2_pavg_w;
1944     else if (Name[5] == 'w' && VecWidth == 256)
1945       IID = Intrinsic::x86_avx2_pavg_w;
1946     else if (Name[5] == 'w' && VecWidth == 512)
1947       IID = Intrinsic::x86_avx512_pavg_w_512;
1948     else
1949       llvm_unreachable("Unexpected intrinsic");
1950   } else
1951     return false;
1952 
1953   SmallVector<Value *, 4> Args(CI.args());
1954   Args.pop_back();
1955   Args.pop_back();
1956   Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI.getModule(), IID),
1957                            Args);
1958   unsigned NumArgs = CI.arg_size();
1959   Rep = EmitX86Select(Builder, CI.getArgOperand(NumArgs - 1), Rep,
1960                       CI.getArgOperand(NumArgs - 2));
1961   return true;
1962 }
1963 
1964 /// Upgrade comment in call to inline asm that represents an objc retain release
1965 /// marker.
UpgradeInlineAsmString(std::string * AsmStr)1966 void llvm::UpgradeInlineAsmString(std::string *AsmStr) {
1967   size_t Pos;
1968   if (AsmStr->find("mov\tfp") == 0 &&
1969       AsmStr->find("objc_retainAutoreleaseReturnValue") != std::string::npos &&
1970       (Pos = AsmStr->find("# marker")) != std::string::npos) {
1971     AsmStr->replace(Pos, 1, ";");
1972   }
1973 }
1974 
UpgradeARMIntrinsicCall(StringRef Name,CallBase * CI,Function * F,IRBuilder<> & Builder)1975 static Value *UpgradeARMIntrinsicCall(StringRef Name, CallBase *CI, Function *F,
1976                                       IRBuilder<> &Builder) {
1977   if (Name == "mve.vctp64.old") {
1978     // Replace the old v4i1 vctp64 with a v2i1 vctp and predicate-casts to the
1979     // correct type.
1980     Value *VCTP = Builder.CreateCall(
1981         Intrinsic::getDeclaration(F->getParent(), Intrinsic::arm_mve_vctp64),
1982         CI->getArgOperand(0), CI->getName());
1983     Value *C1 = Builder.CreateCall(
1984         Intrinsic::getDeclaration(
1985             F->getParent(), Intrinsic::arm_mve_pred_v2i,
1986             {VectorType::get(Builder.getInt1Ty(), 2, false)}),
1987         VCTP);
1988     return Builder.CreateCall(
1989         Intrinsic::getDeclaration(
1990             F->getParent(), Intrinsic::arm_mve_pred_i2v,
1991             {VectorType::get(Builder.getInt1Ty(), 4, false)}),
1992         C1);
1993   } else if (Name == "mve.mull.int.predicated.v2i64.v4i32.v4i1" ||
1994              Name == "mve.vqdmull.predicated.v2i64.v4i32.v4i1" ||
1995              Name == "mve.vldr.gather.base.predicated.v2i64.v2i64.v4i1" ||
1996              Name == "mve.vldr.gather.base.wb.predicated.v2i64.v2i64.v4i1" ||
1997              Name == "mve.vldr.gather.offset.predicated.v2i64.p0i64.v2i64.v4i1" ||
1998              Name == "mve.vstr.scatter.base.predicated.v2i64.v2i64.v4i1" ||
1999              Name == "mve.vstr.scatter.base.wb.predicated.v2i64.v2i64.v4i1" ||
2000              Name == "mve.vstr.scatter.offset.predicated.p0i64.v2i64.v2i64.v4i1" ||
2001              Name == "cde.vcx1q.predicated.v2i64.v4i1" ||
2002              Name == "cde.vcx1qa.predicated.v2i64.v4i1" ||
2003              Name == "cde.vcx2q.predicated.v2i64.v4i1" ||
2004              Name == "cde.vcx2qa.predicated.v2i64.v4i1" ||
2005              Name == "cde.vcx3q.predicated.v2i64.v4i1" ||
2006              Name == "cde.vcx3qa.predicated.v2i64.v4i1") {
2007     std::vector<Type *> Tys;
2008     unsigned ID = CI->getIntrinsicID();
2009     Type *V2I1Ty = FixedVectorType::get(Builder.getInt1Ty(), 2);
2010     switch (ID) {
2011     case Intrinsic::arm_mve_mull_int_predicated:
2012     case Intrinsic::arm_mve_vqdmull_predicated:
2013     case Intrinsic::arm_mve_vldr_gather_base_predicated:
2014       Tys = {CI->getType(), CI->getOperand(0)->getType(), V2I1Ty};
2015       break;
2016     case Intrinsic::arm_mve_vldr_gather_base_wb_predicated:
2017     case Intrinsic::arm_mve_vstr_scatter_base_predicated:
2018     case Intrinsic::arm_mve_vstr_scatter_base_wb_predicated:
2019       Tys = {CI->getOperand(0)->getType(), CI->getOperand(0)->getType(),
2020              V2I1Ty};
2021       break;
2022     case Intrinsic::arm_mve_vldr_gather_offset_predicated:
2023       Tys = {CI->getType(), CI->getOperand(0)->getType(),
2024              CI->getOperand(1)->getType(), V2I1Ty};
2025       break;
2026     case Intrinsic::arm_mve_vstr_scatter_offset_predicated:
2027       Tys = {CI->getOperand(0)->getType(), CI->getOperand(1)->getType(),
2028              CI->getOperand(2)->getType(), V2I1Ty};
2029       break;
2030     case Intrinsic::arm_cde_vcx1q_predicated:
2031     case Intrinsic::arm_cde_vcx1qa_predicated:
2032     case Intrinsic::arm_cde_vcx2q_predicated:
2033     case Intrinsic::arm_cde_vcx2qa_predicated:
2034     case Intrinsic::arm_cde_vcx3q_predicated:
2035     case Intrinsic::arm_cde_vcx3qa_predicated:
2036       Tys = {CI->getOperand(1)->getType(), V2I1Ty};
2037       break;
2038     default:
2039       llvm_unreachable("Unhandled Intrinsic!");
2040     }
2041 
2042     std::vector<Value *> Ops;
2043     for (Value *Op : CI->args()) {
2044       Type *Ty = Op->getType();
2045       if (Ty->getScalarSizeInBits() == 1) {
2046         Value *C1 = Builder.CreateCall(
2047             Intrinsic::getDeclaration(
2048                 F->getParent(), Intrinsic::arm_mve_pred_v2i,
2049                 {VectorType::get(Builder.getInt1Ty(), 4, false)}),
2050             Op);
2051         Op = Builder.CreateCall(
2052             Intrinsic::getDeclaration(F->getParent(),
2053                                       Intrinsic::arm_mve_pred_i2v, {V2I1Ty}),
2054             C1);
2055       }
2056       Ops.push_back(Op);
2057     }
2058 
2059     Function *Fn = Intrinsic::getDeclaration(F->getParent(), ID, Tys);
2060     return Builder.CreateCall(Fn, Ops, CI->getName());
2061   }
2062   llvm_unreachable("Unknown function for ARM CallBase upgrade.");
2063 }
2064 
2065 /// Upgrade a call to an old intrinsic. All argument and return casting must be
2066 /// provided to seamlessly integrate with existing context.
UpgradeIntrinsicCall(CallBase * CI,Function * NewFn)2067 void llvm::UpgradeIntrinsicCall(CallBase *CI, Function *NewFn) {
2068   // Note dyn_cast to Function is not quite the same as getCalledFunction, which
2069   // checks the callee's function type matches. It's likely we need to handle
2070   // type changes here.
2071   Function *F = dyn_cast<Function>(CI->getCalledOperand());
2072   if (!F)
2073     return;
2074 
2075   LLVMContext &C = CI->getContext();
2076   IRBuilder<> Builder(C);
2077   Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
2078 
2079   if (!NewFn) {
2080     // Get the Function's name.
2081     StringRef Name = F->getName();
2082 
2083     assert(Name.startswith("llvm.") && "Intrinsic doesn't start with 'llvm.'");
2084     Name = Name.substr(5);
2085 
2086     bool IsX86 = Name.startswith("x86.");
2087     if (IsX86)
2088       Name = Name.substr(4);
2089     bool IsNVVM = Name.startswith("nvvm.");
2090     if (IsNVVM)
2091       Name = Name.substr(5);
2092     bool IsARM = Name.startswith("arm.");
2093     if (IsARM)
2094       Name = Name.substr(4);
2095 
2096     if (IsX86 && Name.startswith("sse4a.movnt.")) {
2097       Module *M = F->getParent();
2098       SmallVector<Metadata *, 1> Elts;
2099       Elts.push_back(
2100           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
2101       MDNode *Node = MDNode::get(C, Elts);
2102 
2103       Value *Arg0 = CI->getArgOperand(0);
2104       Value *Arg1 = CI->getArgOperand(1);
2105 
2106       // Nontemporal (unaligned) store of the 0'th element of the float/double
2107       // vector.
2108       Type *SrcEltTy = cast<VectorType>(Arg1->getType())->getElementType();
2109       PointerType *EltPtrTy = PointerType::getUnqual(SrcEltTy);
2110       Value *Addr = Builder.CreateBitCast(Arg0, EltPtrTy, "cast");
2111       Value *Extract =
2112           Builder.CreateExtractElement(Arg1, (uint64_t)0, "extractelement");
2113 
2114       StoreInst *SI = Builder.CreateAlignedStore(Extract, Addr, Align(1));
2115       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
2116 
2117       // Remove intrinsic.
2118       CI->eraseFromParent();
2119       return;
2120     }
2121 
2122     if (IsX86 && (Name.startswith("avx.movnt.") ||
2123                   Name.startswith("avx512.storent."))) {
2124       Module *M = F->getParent();
2125       SmallVector<Metadata *, 1> Elts;
2126       Elts.push_back(
2127           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
2128       MDNode *Node = MDNode::get(C, Elts);
2129 
2130       Value *Arg0 = CI->getArgOperand(0);
2131       Value *Arg1 = CI->getArgOperand(1);
2132 
2133       // Convert the type of the pointer to a pointer to the stored type.
2134       Value *BC = Builder.CreateBitCast(Arg0,
2135                                         PointerType::getUnqual(Arg1->getType()),
2136                                         "cast");
2137       StoreInst *SI = Builder.CreateAlignedStore(
2138           Arg1, BC,
2139           Align(Arg1->getType()->getPrimitiveSizeInBits().getFixedValue() / 8));
2140       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
2141 
2142       // Remove intrinsic.
2143       CI->eraseFromParent();
2144       return;
2145     }
2146 
2147     if (IsX86 && Name == "sse2.storel.dq") {
2148       Value *Arg0 = CI->getArgOperand(0);
2149       Value *Arg1 = CI->getArgOperand(1);
2150 
2151       auto *NewVecTy = FixedVectorType::get(Type::getInt64Ty(C), 2);
2152       Value *BC0 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
2153       Value *Elt = Builder.CreateExtractElement(BC0, (uint64_t)0);
2154       Value *BC = Builder.CreateBitCast(Arg0,
2155                                         PointerType::getUnqual(Elt->getType()),
2156                                         "cast");
2157       Builder.CreateAlignedStore(Elt, BC, Align(1));
2158 
2159       // Remove intrinsic.
2160       CI->eraseFromParent();
2161       return;
2162     }
2163 
2164     if (IsX86 && (Name.startswith("sse.storeu.") ||
2165                   Name.startswith("sse2.storeu.") ||
2166                   Name.startswith("avx.storeu."))) {
2167       Value *Arg0 = CI->getArgOperand(0);
2168       Value *Arg1 = CI->getArgOperand(1);
2169 
2170       Arg0 = Builder.CreateBitCast(Arg0,
2171                                    PointerType::getUnqual(Arg1->getType()),
2172                                    "cast");
2173       Builder.CreateAlignedStore(Arg1, Arg0, Align(1));
2174 
2175       // Remove intrinsic.
2176       CI->eraseFromParent();
2177       return;
2178     }
2179 
2180     if (IsX86 && Name == "avx512.mask.store.ss") {
2181       Value *Mask = Builder.CreateAnd(CI->getArgOperand(2), Builder.getInt8(1));
2182       UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
2183                          Mask, false);
2184 
2185       // Remove intrinsic.
2186       CI->eraseFromParent();
2187       return;
2188     }
2189 
2190     if (IsX86 && (Name.startswith("avx512.mask.store"))) {
2191       // "avx512.mask.storeu." or "avx512.mask.store."
2192       bool Aligned = Name[17] != 'u'; // "avx512.mask.storeu".
2193       UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
2194                          CI->getArgOperand(2), Aligned);
2195 
2196       // Remove intrinsic.
2197       CI->eraseFromParent();
2198       return;
2199     }
2200 
2201     Value *Rep;
2202     // Upgrade packed integer vector compare intrinsics to compare instructions.
2203     if (IsX86 && (Name.startswith("sse2.pcmp") ||
2204                   Name.startswith("avx2.pcmp"))) {
2205       // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt."
2206       bool CmpEq = Name[9] == 'e';
2207       Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
2208                                CI->getArgOperand(0), CI->getArgOperand(1));
2209       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
2210     } else if (IsX86 && (Name.startswith("avx512.broadcastm"))) {
2211       Type *ExtTy = Type::getInt32Ty(C);
2212       if (CI->getOperand(0)->getType()->isIntegerTy(8))
2213         ExtTy = Type::getInt64Ty(C);
2214       unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() /
2215                          ExtTy->getPrimitiveSizeInBits();
2216       Rep = Builder.CreateZExt(CI->getArgOperand(0), ExtTy);
2217       Rep = Builder.CreateVectorSplat(NumElts, Rep);
2218     } else if (IsX86 && (Name == "sse.sqrt.ss" ||
2219                          Name == "sse2.sqrt.sd")) {
2220       Value *Vec = CI->getArgOperand(0);
2221       Value *Elt0 = Builder.CreateExtractElement(Vec, (uint64_t)0);
2222       Function *Intr = Intrinsic::getDeclaration(F->getParent(),
2223                                                  Intrinsic::sqrt, Elt0->getType());
2224       Elt0 = Builder.CreateCall(Intr, Elt0);
2225       Rep = Builder.CreateInsertElement(Vec, Elt0, (uint64_t)0);
2226     } else if (IsX86 && (Name.startswith("avx.sqrt.p") ||
2227                          Name.startswith("sse2.sqrt.p") ||
2228                          Name.startswith("sse.sqrt.p"))) {
2229       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
2230                                                          Intrinsic::sqrt,
2231                                                          CI->getType()),
2232                                {CI->getArgOperand(0)});
2233     } else if (IsX86 && (Name.startswith("avx512.mask.sqrt.p"))) {
2234       if (CI->arg_size() == 4 &&
2235           (!isa<ConstantInt>(CI->getArgOperand(3)) ||
2236            cast<ConstantInt>(CI->getArgOperand(3))->getZExtValue() != 4)) {
2237         Intrinsic::ID IID = Name[18] == 's' ? Intrinsic::x86_avx512_sqrt_ps_512
2238                                             : Intrinsic::x86_avx512_sqrt_pd_512;
2239 
2240         Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(3) };
2241         Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(),
2242                                                            IID), Args);
2243       } else {
2244         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
2245                                                            Intrinsic::sqrt,
2246                                                            CI->getType()),
2247                                  {CI->getArgOperand(0)});
2248       }
2249       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2250                           CI->getArgOperand(1));
2251     } else if (IsX86 && (Name.startswith("avx512.ptestm") ||
2252                          Name.startswith("avx512.ptestnm"))) {
2253       Value *Op0 = CI->getArgOperand(0);
2254       Value *Op1 = CI->getArgOperand(1);
2255       Value *Mask = CI->getArgOperand(2);
2256       Rep = Builder.CreateAnd(Op0, Op1);
2257       llvm::Type *Ty = Op0->getType();
2258       Value *Zero = llvm::Constant::getNullValue(Ty);
2259       ICmpInst::Predicate Pred =
2260         Name.startswith("avx512.ptestm") ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
2261       Rep = Builder.CreateICmp(Pred, Rep, Zero);
2262       Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, Mask);
2263     } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))){
2264       unsigned NumElts = cast<FixedVectorType>(CI->getArgOperand(1)->getType())
2265                              ->getNumElements();
2266       Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
2267       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2268                           CI->getArgOperand(1));
2269     } else if (IsX86 && (Name.startswith("avx512.kunpck"))) {
2270       unsigned NumElts = CI->getType()->getScalarSizeInBits();
2271       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), NumElts);
2272       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), NumElts);
2273       int Indices[64];
2274       for (unsigned i = 0; i != NumElts; ++i)
2275         Indices[i] = i;
2276 
2277       // First extract half of each vector. This gives better codegen than
2278       // doing it in a single shuffle.
2279       LHS =
2280           Builder.CreateShuffleVector(LHS, LHS, ArrayRef(Indices, NumElts / 2));
2281       RHS =
2282           Builder.CreateShuffleVector(RHS, RHS, ArrayRef(Indices, NumElts / 2));
2283       // Concat the vectors.
2284       // NOTE: Operands have to be swapped to match intrinsic definition.
2285       Rep = Builder.CreateShuffleVector(RHS, LHS, ArrayRef(Indices, NumElts));
2286       Rep = Builder.CreateBitCast(Rep, CI->getType());
2287     } else if (IsX86 && Name == "avx512.kand.w") {
2288       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2289       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
2290       Rep = Builder.CreateAnd(LHS, RHS);
2291       Rep = Builder.CreateBitCast(Rep, CI->getType());
2292     } else if (IsX86 && Name == "avx512.kandn.w") {
2293       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2294       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
2295       LHS = Builder.CreateNot(LHS);
2296       Rep = Builder.CreateAnd(LHS, RHS);
2297       Rep = Builder.CreateBitCast(Rep, CI->getType());
2298     } else if (IsX86 && Name == "avx512.kor.w") {
2299       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2300       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
2301       Rep = Builder.CreateOr(LHS, RHS);
2302       Rep = Builder.CreateBitCast(Rep, CI->getType());
2303     } else if (IsX86 && Name == "avx512.kxor.w") {
2304       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2305       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
2306       Rep = Builder.CreateXor(LHS, RHS);
2307       Rep = Builder.CreateBitCast(Rep, CI->getType());
2308     } else if (IsX86 && Name == "avx512.kxnor.w") {
2309       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2310       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
2311       LHS = Builder.CreateNot(LHS);
2312       Rep = Builder.CreateXor(LHS, RHS);
2313       Rep = Builder.CreateBitCast(Rep, CI->getType());
2314     } else if (IsX86 && Name == "avx512.knot.w") {
2315       Rep = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2316       Rep = Builder.CreateNot(Rep);
2317       Rep = Builder.CreateBitCast(Rep, CI->getType());
2318     } else if (IsX86 &&
2319                (Name == "avx512.kortestz.w" || Name == "avx512.kortestc.w")) {
2320       Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
2321       Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
2322       Rep = Builder.CreateOr(LHS, RHS);
2323       Rep = Builder.CreateBitCast(Rep, Builder.getInt16Ty());
2324       Value *C;
2325       if (Name[14] == 'c')
2326         C = ConstantInt::getAllOnesValue(Builder.getInt16Ty());
2327       else
2328         C = ConstantInt::getNullValue(Builder.getInt16Ty());
2329       Rep = Builder.CreateICmpEQ(Rep, C);
2330       Rep = Builder.CreateZExt(Rep, Builder.getInt32Ty());
2331     } else if (IsX86 && (Name == "sse.add.ss" || Name == "sse2.add.sd" ||
2332                          Name == "sse.sub.ss" || Name == "sse2.sub.sd" ||
2333                          Name == "sse.mul.ss" || Name == "sse2.mul.sd" ||
2334                          Name == "sse.div.ss" || Name == "sse2.div.sd")) {
2335       Type *I32Ty = Type::getInt32Ty(C);
2336       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
2337                                                  ConstantInt::get(I32Ty, 0));
2338       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
2339                                                  ConstantInt::get(I32Ty, 0));
2340       Value *EltOp;
2341       if (Name.contains(".add."))
2342         EltOp = Builder.CreateFAdd(Elt0, Elt1);
2343       else if (Name.contains(".sub."))
2344         EltOp = Builder.CreateFSub(Elt0, Elt1);
2345       else if (Name.contains(".mul."))
2346         EltOp = Builder.CreateFMul(Elt0, Elt1);
2347       else
2348         EltOp = Builder.CreateFDiv(Elt0, Elt1);
2349       Rep = Builder.CreateInsertElement(CI->getArgOperand(0), EltOp,
2350                                         ConstantInt::get(I32Ty, 0));
2351     } else if (IsX86 && Name.startswith("avx512.mask.pcmp")) {
2352       // "avx512.mask.pcmpeq." or "avx512.mask.pcmpgt."
2353       bool CmpEq = Name[16] == 'e';
2354       Rep = upgradeMaskedCompare(Builder, *CI, CmpEq ? 0 : 6, true);
2355     } else if (IsX86 && Name.startswith("avx512.mask.vpshufbitqmb.")) {
2356       Type *OpTy = CI->getArgOperand(0)->getType();
2357       unsigned VecWidth = OpTy->getPrimitiveSizeInBits();
2358       Intrinsic::ID IID;
2359       switch (VecWidth) {
2360       default: llvm_unreachable("Unexpected intrinsic");
2361       case 128: IID = Intrinsic::x86_avx512_vpshufbitqmb_128; break;
2362       case 256: IID = Intrinsic::x86_avx512_vpshufbitqmb_256; break;
2363       case 512: IID = Intrinsic::x86_avx512_vpshufbitqmb_512; break;
2364       }
2365 
2366       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2367                                { CI->getOperand(0), CI->getArgOperand(1) });
2368       Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, CI->getArgOperand(2));
2369     } else if (IsX86 && Name.startswith("avx512.mask.fpclass.p")) {
2370       Type *OpTy = CI->getArgOperand(0)->getType();
2371       unsigned VecWidth = OpTy->getPrimitiveSizeInBits();
2372       unsigned EltWidth = OpTy->getScalarSizeInBits();
2373       Intrinsic::ID IID;
2374       if (VecWidth == 128 && EltWidth == 32)
2375         IID = Intrinsic::x86_avx512_fpclass_ps_128;
2376       else if (VecWidth == 256 && EltWidth == 32)
2377         IID = Intrinsic::x86_avx512_fpclass_ps_256;
2378       else if (VecWidth == 512 && EltWidth == 32)
2379         IID = Intrinsic::x86_avx512_fpclass_ps_512;
2380       else if (VecWidth == 128 && EltWidth == 64)
2381         IID = Intrinsic::x86_avx512_fpclass_pd_128;
2382       else if (VecWidth == 256 && EltWidth == 64)
2383         IID = Intrinsic::x86_avx512_fpclass_pd_256;
2384       else if (VecWidth == 512 && EltWidth == 64)
2385         IID = Intrinsic::x86_avx512_fpclass_pd_512;
2386       else
2387         llvm_unreachable("Unexpected intrinsic");
2388 
2389       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2390                                { CI->getOperand(0), CI->getArgOperand(1) });
2391       Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, CI->getArgOperand(2));
2392     } else if (IsX86 && Name.startswith("avx512.cmp.p")) {
2393       SmallVector<Value *, 4> Args(CI->args());
2394       Type *OpTy = Args[0]->getType();
2395       unsigned VecWidth = OpTy->getPrimitiveSizeInBits();
2396       unsigned EltWidth = OpTy->getScalarSizeInBits();
2397       Intrinsic::ID IID;
2398       if (VecWidth == 128 && EltWidth == 32)
2399         IID = Intrinsic::x86_avx512_mask_cmp_ps_128;
2400       else if (VecWidth == 256 && EltWidth == 32)
2401         IID = Intrinsic::x86_avx512_mask_cmp_ps_256;
2402       else if (VecWidth == 512 && EltWidth == 32)
2403         IID = Intrinsic::x86_avx512_mask_cmp_ps_512;
2404       else if (VecWidth == 128 && EltWidth == 64)
2405         IID = Intrinsic::x86_avx512_mask_cmp_pd_128;
2406       else if (VecWidth == 256 && EltWidth == 64)
2407         IID = Intrinsic::x86_avx512_mask_cmp_pd_256;
2408       else if (VecWidth == 512 && EltWidth == 64)
2409         IID = Intrinsic::x86_avx512_mask_cmp_pd_512;
2410       else
2411         llvm_unreachable("Unexpected intrinsic");
2412 
2413       Value *Mask = Constant::getAllOnesValue(CI->getType());
2414       if (VecWidth == 512)
2415         std::swap(Mask, Args.back());
2416       Args.push_back(Mask);
2417 
2418       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2419                                Args);
2420     } else if (IsX86 && Name.startswith("avx512.mask.cmp.")) {
2421       // Integer compare intrinsics.
2422       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2423       Rep = upgradeMaskedCompare(Builder, *CI, Imm, true);
2424     } else if (IsX86 && Name.startswith("avx512.mask.ucmp.")) {
2425       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2426       Rep = upgradeMaskedCompare(Builder, *CI, Imm, false);
2427     } else if (IsX86 && (Name.startswith("avx512.cvtb2mask.") ||
2428                          Name.startswith("avx512.cvtw2mask.") ||
2429                          Name.startswith("avx512.cvtd2mask.") ||
2430                          Name.startswith("avx512.cvtq2mask."))) {
2431       Value *Op = CI->getArgOperand(0);
2432       Value *Zero = llvm::Constant::getNullValue(Op->getType());
2433       Rep = Builder.CreateICmp(ICmpInst::ICMP_SLT, Op, Zero);
2434       Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, nullptr);
2435     } else if(IsX86 && (Name == "ssse3.pabs.b.128" ||
2436                         Name == "ssse3.pabs.w.128" ||
2437                         Name == "ssse3.pabs.d.128" ||
2438                         Name.startswith("avx2.pabs") ||
2439                         Name.startswith("avx512.mask.pabs"))) {
2440       Rep = upgradeAbs(Builder, *CI);
2441     } else if (IsX86 && (Name == "sse41.pmaxsb" ||
2442                          Name == "sse2.pmaxs.w" ||
2443                          Name == "sse41.pmaxsd" ||
2444                          Name.startswith("avx2.pmaxs") ||
2445                          Name.startswith("avx512.mask.pmaxs"))) {
2446       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::smax);
2447     } else if (IsX86 && (Name == "sse2.pmaxu.b" ||
2448                          Name == "sse41.pmaxuw" ||
2449                          Name == "sse41.pmaxud" ||
2450                          Name.startswith("avx2.pmaxu") ||
2451                          Name.startswith("avx512.mask.pmaxu"))) {
2452       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::umax);
2453     } else if (IsX86 && (Name == "sse41.pminsb" ||
2454                          Name == "sse2.pmins.w" ||
2455                          Name == "sse41.pminsd" ||
2456                          Name.startswith("avx2.pmins") ||
2457                          Name.startswith("avx512.mask.pmins"))) {
2458       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::smin);
2459     } else if (IsX86 && (Name == "sse2.pminu.b" ||
2460                          Name == "sse41.pminuw" ||
2461                          Name == "sse41.pminud" ||
2462                          Name.startswith("avx2.pminu") ||
2463                          Name.startswith("avx512.mask.pminu"))) {
2464       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::umin);
2465     } else if (IsX86 && (Name == "sse2.pmulu.dq" ||
2466                          Name == "avx2.pmulu.dq" ||
2467                          Name == "avx512.pmulu.dq.512" ||
2468                          Name.startswith("avx512.mask.pmulu.dq."))) {
2469       Rep = upgradePMULDQ(Builder, *CI, /*Signed*/false);
2470     } else if (IsX86 && (Name == "sse41.pmuldq" ||
2471                          Name == "avx2.pmul.dq" ||
2472                          Name == "avx512.pmul.dq.512" ||
2473                          Name.startswith("avx512.mask.pmul.dq."))) {
2474       Rep = upgradePMULDQ(Builder, *CI, /*Signed*/true);
2475     } else if (IsX86 && (Name == "sse.cvtsi2ss" ||
2476                          Name == "sse2.cvtsi2sd" ||
2477                          Name == "sse.cvtsi642ss" ||
2478                          Name == "sse2.cvtsi642sd")) {
2479       Rep = Builder.CreateSIToFP(
2480           CI->getArgOperand(1),
2481           cast<VectorType>(CI->getType())->getElementType());
2482       Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep, (uint64_t)0);
2483     } else if (IsX86 && Name == "avx512.cvtusi2sd") {
2484       Rep = Builder.CreateUIToFP(
2485           CI->getArgOperand(1),
2486           cast<VectorType>(CI->getType())->getElementType());
2487       Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep, (uint64_t)0);
2488     } else if (IsX86 && Name == "sse2.cvtss2sd") {
2489       Rep = Builder.CreateExtractElement(CI->getArgOperand(1), (uint64_t)0);
2490       Rep = Builder.CreateFPExt(
2491           Rep, cast<VectorType>(CI->getType())->getElementType());
2492       Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep, (uint64_t)0);
2493     } else if (IsX86 && (Name == "sse2.cvtdq2pd" ||
2494                          Name == "sse2.cvtdq2ps" ||
2495                          Name == "avx.cvtdq2.pd.256" ||
2496                          Name == "avx.cvtdq2.ps.256" ||
2497                          Name.startswith("avx512.mask.cvtdq2pd.") ||
2498                          Name.startswith("avx512.mask.cvtudq2pd.") ||
2499                          Name.startswith("avx512.mask.cvtdq2ps.") ||
2500                          Name.startswith("avx512.mask.cvtudq2ps.") ||
2501                          Name.startswith("avx512.mask.cvtqq2pd.") ||
2502                          Name.startswith("avx512.mask.cvtuqq2pd.") ||
2503                          Name == "avx512.mask.cvtqq2ps.256" ||
2504                          Name == "avx512.mask.cvtqq2ps.512" ||
2505                          Name == "avx512.mask.cvtuqq2ps.256" ||
2506                          Name == "avx512.mask.cvtuqq2ps.512" ||
2507                          Name == "sse2.cvtps2pd" ||
2508                          Name == "avx.cvt.ps2.pd.256" ||
2509                          Name == "avx512.mask.cvtps2pd.128" ||
2510                          Name == "avx512.mask.cvtps2pd.256")) {
2511       auto *DstTy = cast<FixedVectorType>(CI->getType());
2512       Rep = CI->getArgOperand(0);
2513       auto *SrcTy = cast<FixedVectorType>(Rep->getType());
2514 
2515       unsigned NumDstElts = DstTy->getNumElements();
2516       if (NumDstElts < SrcTy->getNumElements()) {
2517         assert(NumDstElts == 2 && "Unexpected vector size");
2518         Rep = Builder.CreateShuffleVector(Rep, Rep, ArrayRef<int>{0, 1});
2519       }
2520 
2521       bool IsPS2PD = SrcTy->getElementType()->isFloatTy();
2522       bool IsUnsigned = (StringRef::npos != Name.find("cvtu"));
2523       if (IsPS2PD)
2524         Rep = Builder.CreateFPExt(Rep, DstTy, "cvtps2pd");
2525       else if (CI->arg_size() == 4 &&
2526                (!isa<ConstantInt>(CI->getArgOperand(3)) ||
2527                 cast<ConstantInt>(CI->getArgOperand(3))->getZExtValue() != 4)) {
2528         Intrinsic::ID IID = IsUnsigned ? Intrinsic::x86_avx512_uitofp_round
2529                                        : Intrinsic::x86_avx512_sitofp_round;
2530         Function *F = Intrinsic::getDeclaration(CI->getModule(), IID,
2531                                                 { DstTy, SrcTy });
2532         Rep = Builder.CreateCall(F, { Rep, CI->getArgOperand(3) });
2533       } else {
2534         Rep = IsUnsigned ? Builder.CreateUIToFP(Rep, DstTy, "cvt")
2535                          : Builder.CreateSIToFP(Rep, DstTy, "cvt");
2536       }
2537 
2538       if (CI->arg_size() >= 3)
2539         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2540                             CI->getArgOperand(1));
2541     } else if (IsX86 && (Name.startswith("avx512.mask.vcvtph2ps.") ||
2542                          Name.startswith("vcvtph2ps."))) {
2543       auto *DstTy = cast<FixedVectorType>(CI->getType());
2544       Rep = CI->getArgOperand(0);
2545       auto *SrcTy = cast<FixedVectorType>(Rep->getType());
2546       unsigned NumDstElts = DstTy->getNumElements();
2547       if (NumDstElts != SrcTy->getNumElements()) {
2548         assert(NumDstElts == 4 && "Unexpected vector size");
2549         Rep = Builder.CreateShuffleVector(Rep, Rep, ArrayRef<int>{0, 1, 2, 3});
2550       }
2551       Rep = Builder.CreateBitCast(
2552           Rep, FixedVectorType::get(Type::getHalfTy(C), NumDstElts));
2553       Rep = Builder.CreateFPExt(Rep, DstTy, "cvtph2ps");
2554       if (CI->arg_size() >= 3)
2555         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2556                             CI->getArgOperand(1));
2557     } else if (IsX86 && Name.startswith("avx512.mask.load")) {
2558       // "avx512.mask.loadu." or "avx512.mask.load."
2559       bool Aligned = Name[16] != 'u'; // "avx512.mask.loadu".
2560       Rep =
2561           UpgradeMaskedLoad(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
2562                             CI->getArgOperand(2), Aligned);
2563     } else if (IsX86 && Name.startswith("avx512.mask.expand.load.")) {
2564       auto *ResultTy = cast<FixedVectorType>(CI->getType());
2565       Type *PtrTy = ResultTy->getElementType();
2566 
2567       // Cast the pointer to element type.
2568       Value *Ptr = Builder.CreateBitCast(CI->getOperand(0),
2569                                          llvm::PointerType::getUnqual(PtrTy));
2570 
2571       Value *MaskVec = getX86MaskVec(Builder, CI->getArgOperand(2),
2572                                      ResultTy->getNumElements());
2573 
2574       Function *ELd = Intrinsic::getDeclaration(F->getParent(),
2575                                                 Intrinsic::masked_expandload,
2576                                                 ResultTy);
2577       Rep = Builder.CreateCall(ELd, { Ptr, MaskVec, CI->getOperand(1) });
2578     } else if (IsX86 && Name.startswith("avx512.mask.compress.store.")) {
2579       auto *ResultTy = cast<VectorType>(CI->getArgOperand(1)->getType());
2580       Type *PtrTy = ResultTy->getElementType();
2581 
2582       // Cast the pointer to element type.
2583       Value *Ptr = Builder.CreateBitCast(CI->getOperand(0),
2584                                          llvm::PointerType::getUnqual(PtrTy));
2585 
2586       Value *MaskVec =
2587           getX86MaskVec(Builder, CI->getArgOperand(2),
2588                         cast<FixedVectorType>(ResultTy)->getNumElements());
2589 
2590       Function *CSt = Intrinsic::getDeclaration(F->getParent(),
2591                                                 Intrinsic::masked_compressstore,
2592                                                 ResultTy);
2593       Rep = Builder.CreateCall(CSt, { CI->getArgOperand(1), Ptr, MaskVec });
2594     } else if (IsX86 && (Name.startswith("avx512.mask.compress.") ||
2595                          Name.startswith("avx512.mask.expand."))) {
2596       auto *ResultTy = cast<FixedVectorType>(CI->getType());
2597 
2598       Value *MaskVec = getX86MaskVec(Builder, CI->getArgOperand(2),
2599                                      ResultTy->getNumElements());
2600 
2601       bool IsCompress = Name[12] == 'c';
2602       Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress
2603                                      : Intrinsic::x86_avx512_mask_expand;
2604       Function *Intr = Intrinsic::getDeclaration(F->getParent(), IID, ResultTy);
2605       Rep = Builder.CreateCall(Intr, { CI->getOperand(0), CI->getOperand(1),
2606                                        MaskVec });
2607     } else if (IsX86 && Name.startswith("xop.vpcom")) {
2608       bool IsSigned;
2609       if (Name.endswith("ub") || Name.endswith("uw") || Name.endswith("ud") ||
2610           Name.endswith("uq"))
2611         IsSigned = false;
2612       else if (Name.endswith("b") || Name.endswith("w") || Name.endswith("d") ||
2613                Name.endswith("q"))
2614         IsSigned = true;
2615       else
2616         llvm_unreachable("Unknown suffix");
2617 
2618       unsigned Imm;
2619       if (CI->arg_size() == 3) {
2620         Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2621       } else {
2622         Name = Name.substr(9); // strip off "xop.vpcom"
2623         if (Name.startswith("lt"))
2624           Imm = 0;
2625         else if (Name.startswith("le"))
2626           Imm = 1;
2627         else if (Name.startswith("gt"))
2628           Imm = 2;
2629         else if (Name.startswith("ge"))
2630           Imm = 3;
2631         else if (Name.startswith("eq"))
2632           Imm = 4;
2633         else if (Name.startswith("ne"))
2634           Imm = 5;
2635         else if (Name.startswith("false"))
2636           Imm = 6;
2637         else if (Name.startswith("true"))
2638           Imm = 7;
2639         else
2640           llvm_unreachable("Unknown condition");
2641       }
2642 
2643       Rep = upgradeX86vpcom(Builder, *CI, Imm, IsSigned);
2644     } else if (IsX86 && Name.startswith("xop.vpcmov")) {
2645       Value *Sel = CI->getArgOperand(2);
2646       Value *NotSel = Builder.CreateNot(Sel);
2647       Value *Sel0 = Builder.CreateAnd(CI->getArgOperand(0), Sel);
2648       Value *Sel1 = Builder.CreateAnd(CI->getArgOperand(1), NotSel);
2649       Rep = Builder.CreateOr(Sel0, Sel1);
2650     } else if (IsX86 && (Name.startswith("xop.vprot") ||
2651                          Name.startswith("avx512.prol") ||
2652                          Name.startswith("avx512.mask.prol"))) {
2653       Rep = upgradeX86Rotate(Builder, *CI, false);
2654     } else if (IsX86 && (Name.startswith("avx512.pror") ||
2655                          Name.startswith("avx512.mask.pror"))) {
2656       Rep = upgradeX86Rotate(Builder, *CI, true);
2657     } else if (IsX86 && (Name.startswith("avx512.vpshld.") ||
2658                          Name.startswith("avx512.mask.vpshld") ||
2659                          Name.startswith("avx512.maskz.vpshld"))) {
2660       bool ZeroMask = Name[11] == 'z';
2661       Rep = upgradeX86ConcatShift(Builder, *CI, false, ZeroMask);
2662     } else if (IsX86 && (Name.startswith("avx512.vpshrd.") ||
2663                          Name.startswith("avx512.mask.vpshrd") ||
2664                          Name.startswith("avx512.maskz.vpshrd"))) {
2665       bool ZeroMask = Name[11] == 'z';
2666       Rep = upgradeX86ConcatShift(Builder, *CI, true, ZeroMask);
2667     } else if (IsX86 && Name == "sse42.crc32.64.8") {
2668       Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
2669                                                Intrinsic::x86_sse42_crc32_32_8);
2670       Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
2671       Rep = Builder.CreateCall(CRC32, {Trunc0, CI->getArgOperand(1)});
2672       Rep = Builder.CreateZExt(Rep, CI->getType(), "");
2673     } else if (IsX86 && (Name.startswith("avx.vbroadcast.s") ||
2674                          Name.startswith("avx512.vbroadcast.s"))) {
2675       // Replace broadcasts with a series of insertelements.
2676       auto *VecTy = cast<FixedVectorType>(CI->getType());
2677       Type *EltTy = VecTy->getElementType();
2678       unsigned EltNum = VecTy->getNumElements();
2679       Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
2680                                           EltTy->getPointerTo());
2681       Value *Load = Builder.CreateLoad(EltTy, Cast);
2682       Type *I32Ty = Type::getInt32Ty(C);
2683       Rep = PoisonValue::get(VecTy);
2684       for (unsigned I = 0; I < EltNum; ++I)
2685         Rep = Builder.CreateInsertElement(Rep, Load,
2686                                           ConstantInt::get(I32Ty, I));
2687     } else if (IsX86 && (Name.startswith("sse41.pmovsx") ||
2688                          Name.startswith("sse41.pmovzx") ||
2689                          Name.startswith("avx2.pmovsx") ||
2690                          Name.startswith("avx2.pmovzx") ||
2691                          Name.startswith("avx512.mask.pmovsx") ||
2692                          Name.startswith("avx512.mask.pmovzx"))) {
2693       auto *DstTy = cast<FixedVectorType>(CI->getType());
2694       unsigned NumDstElts = DstTy->getNumElements();
2695 
2696       // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
2697       SmallVector<int, 8> ShuffleMask(NumDstElts);
2698       for (unsigned i = 0; i != NumDstElts; ++i)
2699         ShuffleMask[i] = i;
2700 
2701       Value *SV =
2702           Builder.CreateShuffleVector(CI->getArgOperand(0), ShuffleMask);
2703 
2704       bool DoSext = (StringRef::npos != Name.find("pmovsx"));
2705       Rep = DoSext ? Builder.CreateSExt(SV, DstTy)
2706                    : Builder.CreateZExt(SV, DstTy);
2707       // If there are 3 arguments, it's a masked intrinsic so we need a select.
2708       if (CI->arg_size() == 3)
2709         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2710                             CI->getArgOperand(1));
2711     } else if (Name == "avx512.mask.pmov.qd.256" ||
2712                Name == "avx512.mask.pmov.qd.512" ||
2713                Name == "avx512.mask.pmov.wb.256" ||
2714                Name == "avx512.mask.pmov.wb.512") {
2715       Type *Ty = CI->getArgOperand(1)->getType();
2716       Rep = Builder.CreateTrunc(CI->getArgOperand(0), Ty);
2717       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2718                           CI->getArgOperand(1));
2719     } else if (IsX86 && (Name.startswith("avx.vbroadcastf128") ||
2720                          Name == "avx2.vbroadcasti128")) {
2721       // Replace vbroadcastf128/vbroadcasti128 with a vector load+shuffle.
2722       Type *EltTy = cast<VectorType>(CI->getType())->getElementType();
2723       unsigned NumSrcElts = 128 / EltTy->getPrimitiveSizeInBits();
2724       auto *VT = FixedVectorType::get(EltTy, NumSrcElts);
2725       Value *Op = Builder.CreatePointerCast(CI->getArgOperand(0),
2726                                             PointerType::getUnqual(VT));
2727       Value *Load = Builder.CreateAlignedLoad(VT, Op, Align(1));
2728       if (NumSrcElts == 2)
2729         Rep = Builder.CreateShuffleVector(Load, ArrayRef<int>{0, 1, 0, 1});
2730       else
2731         Rep = Builder.CreateShuffleVector(
2732             Load, ArrayRef<int>{0, 1, 2, 3, 0, 1, 2, 3});
2733     } else if (IsX86 && (Name.startswith("avx512.mask.shuf.i") ||
2734                          Name.startswith("avx512.mask.shuf.f"))) {
2735       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2736       Type *VT = CI->getType();
2737       unsigned NumLanes = VT->getPrimitiveSizeInBits() / 128;
2738       unsigned NumElementsInLane = 128 / VT->getScalarSizeInBits();
2739       unsigned ControlBitsMask = NumLanes - 1;
2740       unsigned NumControlBits = NumLanes / 2;
2741       SmallVector<int, 8> ShuffleMask(0);
2742 
2743       for (unsigned l = 0; l != NumLanes; ++l) {
2744         unsigned LaneMask = (Imm >> (l * NumControlBits)) & ControlBitsMask;
2745         // We actually need the other source.
2746         if (l >= NumLanes / 2)
2747           LaneMask += NumLanes;
2748         for (unsigned i = 0; i != NumElementsInLane; ++i)
2749           ShuffleMask.push_back(LaneMask * NumElementsInLane + i);
2750       }
2751       Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
2752                                         CI->getArgOperand(1), ShuffleMask);
2753       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
2754                           CI->getArgOperand(3));
2755     }else if (IsX86 && (Name.startswith("avx512.mask.broadcastf") ||
2756                          Name.startswith("avx512.mask.broadcasti"))) {
2757       unsigned NumSrcElts =
2758           cast<FixedVectorType>(CI->getArgOperand(0)->getType())
2759               ->getNumElements();
2760       unsigned NumDstElts =
2761           cast<FixedVectorType>(CI->getType())->getNumElements();
2762 
2763       SmallVector<int, 8> ShuffleMask(NumDstElts);
2764       for (unsigned i = 0; i != NumDstElts; ++i)
2765         ShuffleMask[i] = i % NumSrcElts;
2766 
2767       Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
2768                                         CI->getArgOperand(0),
2769                                         ShuffleMask);
2770       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2771                           CI->getArgOperand(1));
2772     } else if (IsX86 && (Name.startswith("avx2.pbroadcast") ||
2773                          Name.startswith("avx2.vbroadcast") ||
2774                          Name.startswith("avx512.pbroadcast") ||
2775                          Name.startswith("avx512.mask.broadcast.s"))) {
2776       // Replace vp?broadcasts with a vector shuffle.
2777       Value *Op = CI->getArgOperand(0);
2778       ElementCount EC = cast<VectorType>(CI->getType())->getElementCount();
2779       Type *MaskTy = VectorType::get(Type::getInt32Ty(C), EC);
2780       SmallVector<int, 8> M;
2781       ShuffleVectorInst::getShuffleMask(Constant::getNullValue(MaskTy), M);
2782       Rep = Builder.CreateShuffleVector(Op, M);
2783 
2784       if (CI->arg_size() == 3)
2785         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2786                             CI->getArgOperand(1));
2787     } else if (IsX86 && (Name.startswith("sse2.padds.") ||
2788                          Name.startswith("avx2.padds.") ||
2789                          Name.startswith("avx512.padds.") ||
2790                          Name.startswith("avx512.mask.padds."))) {
2791       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::sadd_sat);
2792     } else if (IsX86 && (Name.startswith("sse2.psubs.") ||
2793                          Name.startswith("avx2.psubs.") ||
2794                          Name.startswith("avx512.psubs.") ||
2795                          Name.startswith("avx512.mask.psubs."))) {
2796       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::ssub_sat);
2797     } else if (IsX86 && (Name.startswith("sse2.paddus.") ||
2798                          Name.startswith("avx2.paddus.") ||
2799                          Name.startswith("avx512.mask.paddus."))) {
2800       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::uadd_sat);
2801     } else if (IsX86 && (Name.startswith("sse2.psubus.") ||
2802                          Name.startswith("avx2.psubus.") ||
2803                          Name.startswith("avx512.mask.psubus."))) {
2804       Rep = UpgradeX86BinaryIntrinsics(Builder, *CI, Intrinsic::usub_sat);
2805     } else if (IsX86 && Name.startswith("avx512.mask.palignr.")) {
2806       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
2807                                       CI->getArgOperand(1),
2808                                       CI->getArgOperand(2),
2809                                       CI->getArgOperand(3),
2810                                       CI->getArgOperand(4),
2811                                       false);
2812     } else if (IsX86 && Name.startswith("avx512.mask.valign.")) {
2813       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
2814                                       CI->getArgOperand(1),
2815                                       CI->getArgOperand(2),
2816                                       CI->getArgOperand(3),
2817                                       CI->getArgOperand(4),
2818                                       true);
2819     } else if (IsX86 && (Name == "sse2.psll.dq" ||
2820                          Name == "avx2.psll.dq")) {
2821       // 128/256-bit shift left specified in bits.
2822       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2823       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0),
2824                                        Shift / 8); // Shift is in bits.
2825     } else if (IsX86 && (Name == "sse2.psrl.dq" ||
2826                          Name == "avx2.psrl.dq")) {
2827       // 128/256-bit shift right specified in bits.
2828       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2829       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0),
2830                                        Shift / 8); // Shift is in bits.
2831     } else if (IsX86 && (Name == "sse2.psll.dq.bs" ||
2832                          Name == "avx2.psll.dq.bs" ||
2833                          Name == "avx512.psll.dq.512")) {
2834       // 128/256/512-bit shift left specified in bytes.
2835       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2836       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
2837     } else if (IsX86 && (Name == "sse2.psrl.dq.bs" ||
2838                          Name == "avx2.psrl.dq.bs" ||
2839                          Name == "avx512.psrl.dq.512")) {
2840       // 128/256/512-bit shift right specified in bytes.
2841       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2842       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
2843     } else if (IsX86 && (Name == "sse41.pblendw" ||
2844                          Name.startswith("sse41.blendp") ||
2845                          Name.startswith("avx.blend.p") ||
2846                          Name == "avx2.pblendw" ||
2847                          Name.startswith("avx2.pblendd."))) {
2848       Value *Op0 = CI->getArgOperand(0);
2849       Value *Op1 = CI->getArgOperand(1);
2850       unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2851       auto *VecTy = cast<FixedVectorType>(CI->getType());
2852       unsigned NumElts = VecTy->getNumElements();
2853 
2854       SmallVector<int, 16> Idxs(NumElts);
2855       for (unsigned i = 0; i != NumElts; ++i)
2856         Idxs[i] = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
2857 
2858       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
2859     } else if (IsX86 && (Name.startswith("avx.vinsertf128.") ||
2860                          Name == "avx2.vinserti128" ||
2861                          Name.startswith("avx512.mask.insert"))) {
2862       Value *Op0 = CI->getArgOperand(0);
2863       Value *Op1 = CI->getArgOperand(1);
2864       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2865       unsigned DstNumElts =
2866           cast<FixedVectorType>(CI->getType())->getNumElements();
2867       unsigned SrcNumElts =
2868           cast<FixedVectorType>(Op1->getType())->getNumElements();
2869       unsigned Scale = DstNumElts / SrcNumElts;
2870 
2871       // Mask off the high bits of the immediate value; hardware ignores those.
2872       Imm = Imm % Scale;
2873 
2874       // Extend the second operand into a vector the size of the destination.
2875       SmallVector<int, 8> Idxs(DstNumElts);
2876       for (unsigned i = 0; i != SrcNumElts; ++i)
2877         Idxs[i] = i;
2878       for (unsigned i = SrcNumElts; i != DstNumElts; ++i)
2879         Idxs[i] = SrcNumElts;
2880       Rep = Builder.CreateShuffleVector(Op1, Idxs);
2881 
2882       // Insert the second operand into the first operand.
2883 
2884       // Note that there is no guarantee that instruction lowering will actually
2885       // produce a vinsertf128 instruction for the created shuffles. In
2886       // particular, the 0 immediate case involves no lane changes, so it can
2887       // be handled as a blend.
2888 
2889       // Example of shuffle mask for 32-bit elements:
2890       // Imm = 1  <i32 0, i32 1, i32 2,  i32 3,  i32 8, i32 9, i32 10, i32 11>
2891       // Imm = 0  <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6,  i32 7 >
2892 
2893       // First fill with identify mask.
2894       for (unsigned i = 0; i != DstNumElts; ++i)
2895         Idxs[i] = i;
2896       // Then replace the elements where we need to insert.
2897       for (unsigned i = 0; i != SrcNumElts; ++i)
2898         Idxs[i + Imm * SrcNumElts] = i + DstNumElts;
2899       Rep = Builder.CreateShuffleVector(Op0, Rep, Idxs);
2900 
2901       // If the intrinsic has a mask operand, handle that.
2902       if (CI->arg_size() == 5)
2903         Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
2904                             CI->getArgOperand(3));
2905     } else if (IsX86 && (Name.startswith("avx.vextractf128.") ||
2906                          Name == "avx2.vextracti128" ||
2907                          Name.startswith("avx512.mask.vextract"))) {
2908       Value *Op0 = CI->getArgOperand(0);
2909       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2910       unsigned DstNumElts =
2911           cast<FixedVectorType>(CI->getType())->getNumElements();
2912       unsigned SrcNumElts =
2913           cast<FixedVectorType>(Op0->getType())->getNumElements();
2914       unsigned Scale = SrcNumElts / DstNumElts;
2915 
2916       // Mask off the high bits of the immediate value; hardware ignores those.
2917       Imm = Imm % Scale;
2918 
2919       // Get indexes for the subvector of the input vector.
2920       SmallVector<int, 8> Idxs(DstNumElts);
2921       for (unsigned i = 0; i != DstNumElts; ++i) {
2922         Idxs[i] = i + (Imm * DstNumElts);
2923       }
2924       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2925 
2926       // If the intrinsic has a mask operand, handle that.
2927       if (CI->arg_size() == 4)
2928         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2929                             CI->getArgOperand(2));
2930     } else if (!IsX86 && Name == "stackprotectorcheck") {
2931       Rep = nullptr;
2932     } else if (IsX86 && (Name.startswith("avx512.mask.perm.df.") ||
2933                          Name.startswith("avx512.mask.perm.di."))) {
2934       Value *Op0 = CI->getArgOperand(0);
2935       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2936       auto *VecTy = cast<FixedVectorType>(CI->getType());
2937       unsigned NumElts = VecTy->getNumElements();
2938 
2939       SmallVector<int, 8> Idxs(NumElts);
2940       for (unsigned i = 0; i != NumElts; ++i)
2941         Idxs[i] = (i & ~0x3) + ((Imm >> (2 * (i & 0x3))) & 3);
2942 
2943       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2944 
2945       if (CI->arg_size() == 4)
2946         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2947                             CI->getArgOperand(2));
2948     } else if (IsX86 && (Name.startswith("avx.vperm2f128.") ||
2949                          Name == "avx2.vperm2i128")) {
2950       // The immediate permute control byte looks like this:
2951       //    [1:0] - select 128 bits from sources for low half of destination
2952       //    [2]   - ignore
2953       //    [3]   - zero low half of destination
2954       //    [5:4] - select 128 bits from sources for high half of destination
2955       //    [6]   - ignore
2956       //    [7]   - zero high half of destination
2957 
2958       uint8_t Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2959 
2960       unsigned NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
2961       unsigned HalfSize = NumElts / 2;
2962       SmallVector<int, 8> ShuffleMask(NumElts);
2963 
2964       // Determine which operand(s) are actually in use for this instruction.
2965       Value *V0 = (Imm & 0x02) ? CI->getArgOperand(1) : CI->getArgOperand(0);
2966       Value *V1 = (Imm & 0x20) ? CI->getArgOperand(1) : CI->getArgOperand(0);
2967 
2968       // If needed, replace operands based on zero mask.
2969       V0 = (Imm & 0x08) ? ConstantAggregateZero::get(CI->getType()) : V0;
2970       V1 = (Imm & 0x80) ? ConstantAggregateZero::get(CI->getType()) : V1;
2971 
2972       // Permute low half of result.
2973       unsigned StartIndex = (Imm & 0x01) ? HalfSize : 0;
2974       for (unsigned i = 0; i < HalfSize; ++i)
2975         ShuffleMask[i] = StartIndex + i;
2976 
2977       // Permute high half of result.
2978       StartIndex = (Imm & 0x10) ? HalfSize : 0;
2979       for (unsigned i = 0; i < HalfSize; ++i)
2980         ShuffleMask[i + HalfSize] = NumElts + StartIndex + i;
2981 
2982       Rep = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
2983 
2984     } else if (IsX86 && (Name.startswith("avx.vpermil.") ||
2985                          Name == "sse2.pshuf.d" ||
2986                          Name.startswith("avx512.mask.vpermil.p") ||
2987                          Name.startswith("avx512.mask.pshuf.d."))) {
2988       Value *Op0 = CI->getArgOperand(0);
2989       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2990       auto *VecTy = cast<FixedVectorType>(CI->getType());
2991       unsigned NumElts = VecTy->getNumElements();
2992       // Calculate the size of each index in the immediate.
2993       unsigned IdxSize = 64 / VecTy->getScalarSizeInBits();
2994       unsigned IdxMask = ((1 << IdxSize) - 1);
2995 
2996       SmallVector<int, 8> Idxs(NumElts);
2997       // Lookup the bits for this element, wrapping around the immediate every
2998       // 8-bits. Elements are grouped into sets of 2 or 4 elements so we need
2999       // to offset by the first index of each group.
3000       for (unsigned i = 0; i != NumElts; ++i)
3001         Idxs[i] = ((Imm >> ((i * IdxSize) % 8)) & IdxMask) | (i & ~IdxMask);
3002 
3003       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
3004 
3005       if (CI->arg_size() == 4)
3006         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3007                             CI->getArgOperand(2));
3008     } else if (IsX86 && (Name == "sse2.pshufl.w" ||
3009                          Name.startswith("avx512.mask.pshufl.w."))) {
3010       Value *Op0 = CI->getArgOperand(0);
3011       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
3012       unsigned NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3013 
3014       SmallVector<int, 16> Idxs(NumElts);
3015       for (unsigned l = 0; l != NumElts; l += 8) {
3016         for (unsigned i = 0; i != 4; ++i)
3017           Idxs[i + l] = ((Imm >> (2 * i)) & 0x3) + l;
3018         for (unsigned i = 4; i != 8; ++i)
3019           Idxs[i + l] = i + l;
3020       }
3021 
3022       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
3023 
3024       if (CI->arg_size() == 4)
3025         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3026                             CI->getArgOperand(2));
3027     } else if (IsX86 && (Name == "sse2.pshufh.w" ||
3028                          Name.startswith("avx512.mask.pshufh.w."))) {
3029       Value *Op0 = CI->getArgOperand(0);
3030       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
3031       unsigned NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3032 
3033       SmallVector<int, 16> Idxs(NumElts);
3034       for (unsigned l = 0; l != NumElts; l += 8) {
3035         for (unsigned i = 0; i != 4; ++i)
3036           Idxs[i + l] = i + l;
3037         for (unsigned i = 0; i != 4; ++i)
3038           Idxs[i + l + 4] = ((Imm >> (2 * i)) & 0x3) + 4 + l;
3039       }
3040 
3041       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
3042 
3043       if (CI->arg_size() == 4)
3044         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3045                             CI->getArgOperand(2));
3046     } else if (IsX86 && Name.startswith("avx512.mask.shuf.p")) {
3047       Value *Op0 = CI->getArgOperand(0);
3048       Value *Op1 = CI->getArgOperand(1);
3049       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
3050       unsigned NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3051 
3052       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
3053       unsigned HalfLaneElts = NumLaneElts / 2;
3054 
3055       SmallVector<int, 16> Idxs(NumElts);
3056       for (unsigned i = 0; i != NumElts; ++i) {
3057         // Base index is the starting element of the lane.
3058         Idxs[i] = i - (i % NumLaneElts);
3059         // If we are half way through the lane switch to the other source.
3060         if ((i % NumLaneElts) >= HalfLaneElts)
3061           Idxs[i] += NumElts;
3062         // Now select the specific element. By adding HalfLaneElts bits from
3063         // the immediate. Wrapping around the immediate every 8-bits.
3064         Idxs[i] += (Imm >> ((i * HalfLaneElts) % 8)) & ((1 << HalfLaneElts) - 1);
3065       }
3066 
3067       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
3068 
3069       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
3070                           CI->getArgOperand(3));
3071     } else if (IsX86 && (Name.startswith("avx512.mask.movddup") ||
3072                          Name.startswith("avx512.mask.movshdup") ||
3073                          Name.startswith("avx512.mask.movsldup"))) {
3074       Value *Op0 = CI->getArgOperand(0);
3075       unsigned NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3076       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
3077 
3078       unsigned Offset = 0;
3079       if (Name.startswith("avx512.mask.movshdup."))
3080         Offset = 1;
3081 
3082       SmallVector<int, 16> Idxs(NumElts);
3083       for (unsigned l = 0; l != NumElts; l += NumLaneElts)
3084         for (unsigned i = 0; i != NumLaneElts; i += 2) {
3085           Idxs[i + l + 0] = i + l + Offset;
3086           Idxs[i + l + 1] = i + l + Offset;
3087         }
3088 
3089       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
3090 
3091       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
3092                           CI->getArgOperand(1));
3093     } else if (IsX86 && (Name.startswith("avx512.mask.punpckl") ||
3094                          Name.startswith("avx512.mask.unpckl."))) {
3095       Value *Op0 = CI->getArgOperand(0);
3096       Value *Op1 = CI->getArgOperand(1);
3097       int NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3098       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
3099 
3100       SmallVector<int, 64> Idxs(NumElts);
3101       for (int l = 0; l != NumElts; l += NumLaneElts)
3102         for (int i = 0; i != NumLaneElts; ++i)
3103           Idxs[i + l] = l + (i / 2) + NumElts * (i % 2);
3104 
3105       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
3106 
3107       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3108                           CI->getArgOperand(2));
3109     } else if (IsX86 && (Name.startswith("avx512.mask.punpckh") ||
3110                          Name.startswith("avx512.mask.unpckh."))) {
3111       Value *Op0 = CI->getArgOperand(0);
3112       Value *Op1 = CI->getArgOperand(1);
3113       int NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3114       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
3115 
3116       SmallVector<int, 64> Idxs(NumElts);
3117       for (int l = 0; l != NumElts; l += NumLaneElts)
3118         for (int i = 0; i != NumLaneElts; ++i)
3119           Idxs[i + l] = (NumLaneElts / 2) + l + (i / 2) + NumElts * (i % 2);
3120 
3121       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
3122 
3123       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3124                           CI->getArgOperand(2));
3125     } else if (IsX86 && (Name.startswith("avx512.mask.and.") ||
3126                          Name.startswith("avx512.mask.pand."))) {
3127       VectorType *FTy = cast<VectorType>(CI->getType());
3128       VectorType *ITy = VectorType::getInteger(FTy);
3129       Rep = Builder.CreateAnd(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
3130                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
3131       Rep = Builder.CreateBitCast(Rep, FTy);
3132       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3133                           CI->getArgOperand(2));
3134     } else if (IsX86 && (Name.startswith("avx512.mask.andn.") ||
3135                          Name.startswith("avx512.mask.pandn."))) {
3136       VectorType *FTy = cast<VectorType>(CI->getType());
3137       VectorType *ITy = VectorType::getInteger(FTy);
3138       Rep = Builder.CreateNot(Builder.CreateBitCast(CI->getArgOperand(0), ITy));
3139       Rep = Builder.CreateAnd(Rep,
3140                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
3141       Rep = Builder.CreateBitCast(Rep, FTy);
3142       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3143                           CI->getArgOperand(2));
3144     } else if (IsX86 && (Name.startswith("avx512.mask.or.") ||
3145                          Name.startswith("avx512.mask.por."))) {
3146       VectorType *FTy = cast<VectorType>(CI->getType());
3147       VectorType *ITy = VectorType::getInteger(FTy);
3148       Rep = Builder.CreateOr(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
3149                              Builder.CreateBitCast(CI->getArgOperand(1), ITy));
3150       Rep = Builder.CreateBitCast(Rep, FTy);
3151       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3152                           CI->getArgOperand(2));
3153     } else if (IsX86 && (Name.startswith("avx512.mask.xor.") ||
3154                          Name.startswith("avx512.mask.pxor."))) {
3155       VectorType *FTy = cast<VectorType>(CI->getType());
3156       VectorType *ITy = VectorType::getInteger(FTy);
3157       Rep = Builder.CreateXor(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
3158                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
3159       Rep = Builder.CreateBitCast(Rep, FTy);
3160       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3161                           CI->getArgOperand(2));
3162     } else if (IsX86 && Name.startswith("avx512.mask.padd.")) {
3163       Rep = Builder.CreateAdd(CI->getArgOperand(0), CI->getArgOperand(1));
3164       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3165                           CI->getArgOperand(2));
3166     } else if (IsX86 && Name.startswith("avx512.mask.psub.")) {
3167       Rep = Builder.CreateSub(CI->getArgOperand(0), CI->getArgOperand(1));
3168       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3169                           CI->getArgOperand(2));
3170     } else if (IsX86 && Name.startswith("avx512.mask.pmull.")) {
3171       Rep = Builder.CreateMul(CI->getArgOperand(0), CI->getArgOperand(1));
3172       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3173                           CI->getArgOperand(2));
3174     } else if (IsX86 && Name.startswith("avx512.mask.add.p")) {
3175       if (Name.endswith(".512")) {
3176         Intrinsic::ID IID;
3177         if (Name[17] == 's')
3178           IID = Intrinsic::x86_avx512_add_ps_512;
3179         else
3180           IID = Intrinsic::x86_avx512_add_pd_512;
3181 
3182         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3183                                  { CI->getArgOperand(0), CI->getArgOperand(1),
3184                                    CI->getArgOperand(4) });
3185       } else {
3186         Rep = Builder.CreateFAdd(CI->getArgOperand(0), CI->getArgOperand(1));
3187       }
3188       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3189                           CI->getArgOperand(2));
3190     } else if (IsX86 && Name.startswith("avx512.mask.div.p")) {
3191       if (Name.endswith(".512")) {
3192         Intrinsic::ID IID;
3193         if (Name[17] == 's')
3194           IID = Intrinsic::x86_avx512_div_ps_512;
3195         else
3196           IID = Intrinsic::x86_avx512_div_pd_512;
3197 
3198         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3199                                  { CI->getArgOperand(0), CI->getArgOperand(1),
3200                                    CI->getArgOperand(4) });
3201       } else {
3202         Rep = Builder.CreateFDiv(CI->getArgOperand(0), CI->getArgOperand(1));
3203       }
3204       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3205                           CI->getArgOperand(2));
3206     } else if (IsX86 && Name.startswith("avx512.mask.mul.p")) {
3207       if (Name.endswith(".512")) {
3208         Intrinsic::ID IID;
3209         if (Name[17] == 's')
3210           IID = Intrinsic::x86_avx512_mul_ps_512;
3211         else
3212           IID = Intrinsic::x86_avx512_mul_pd_512;
3213 
3214         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3215                                  { CI->getArgOperand(0), CI->getArgOperand(1),
3216                                    CI->getArgOperand(4) });
3217       } else {
3218         Rep = Builder.CreateFMul(CI->getArgOperand(0), CI->getArgOperand(1));
3219       }
3220       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3221                           CI->getArgOperand(2));
3222     } else if (IsX86 && Name.startswith("avx512.mask.sub.p")) {
3223       if (Name.endswith(".512")) {
3224         Intrinsic::ID IID;
3225         if (Name[17] == 's')
3226           IID = Intrinsic::x86_avx512_sub_ps_512;
3227         else
3228           IID = Intrinsic::x86_avx512_sub_pd_512;
3229 
3230         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3231                                  { CI->getArgOperand(0), CI->getArgOperand(1),
3232                                    CI->getArgOperand(4) });
3233       } else {
3234         Rep = Builder.CreateFSub(CI->getArgOperand(0), CI->getArgOperand(1));
3235       }
3236       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3237                           CI->getArgOperand(2));
3238     } else if (IsX86 && (Name.startswith("avx512.mask.max.p") ||
3239                          Name.startswith("avx512.mask.min.p")) &&
3240                Name.drop_front(18) == ".512") {
3241       bool IsDouble = Name[17] == 'd';
3242       bool IsMin = Name[13] == 'i';
3243       static const Intrinsic::ID MinMaxTbl[2][2] = {
3244         { Intrinsic::x86_avx512_max_ps_512, Intrinsic::x86_avx512_max_pd_512 },
3245         { Intrinsic::x86_avx512_min_ps_512, Intrinsic::x86_avx512_min_pd_512 }
3246       };
3247       Intrinsic::ID IID = MinMaxTbl[IsMin][IsDouble];
3248 
3249       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3250                                { CI->getArgOperand(0), CI->getArgOperand(1),
3251                                  CI->getArgOperand(4) });
3252       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
3253                           CI->getArgOperand(2));
3254     } else if (IsX86 && Name.startswith("avx512.mask.lzcnt.")) {
3255       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
3256                                                          Intrinsic::ctlz,
3257                                                          CI->getType()),
3258                                { CI->getArgOperand(0), Builder.getInt1(false) });
3259       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
3260                           CI->getArgOperand(1));
3261     } else if (IsX86 && Name.startswith("avx512.mask.psll")) {
3262       bool IsImmediate = Name[16] == 'i' ||
3263                          (Name.size() > 18 && Name[18] == 'i');
3264       bool IsVariable = Name[16] == 'v';
3265       char Size = Name[16] == '.' ? Name[17] :
3266                   Name[17] == '.' ? Name[18] :
3267                   Name[18] == '.' ? Name[19] :
3268                                     Name[20];
3269 
3270       Intrinsic::ID IID;
3271       if (IsVariable && Name[17] != '.') {
3272         if (Size == 'd' && Name[17] == '2') // avx512.mask.psllv2.di
3273           IID = Intrinsic::x86_avx2_psllv_q;
3274         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psllv4.di
3275           IID = Intrinsic::x86_avx2_psllv_q_256;
3276         else if (Size == 's' && Name[17] == '4') // avx512.mask.psllv4.si
3277           IID = Intrinsic::x86_avx2_psllv_d;
3278         else if (Size == 's' && Name[17] == '8') // avx512.mask.psllv8.si
3279           IID = Intrinsic::x86_avx2_psllv_d_256;
3280         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psllv8.hi
3281           IID = Intrinsic::x86_avx512_psllv_w_128;
3282         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psllv16.hi
3283           IID = Intrinsic::x86_avx512_psllv_w_256;
3284         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psllv32hi
3285           IID = Intrinsic::x86_avx512_psllv_w_512;
3286         else
3287           llvm_unreachable("Unexpected size");
3288       } else if (Name.endswith(".128")) {
3289         if (Size == 'd') // avx512.mask.psll.d.128, avx512.mask.psll.di.128
3290           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_d
3291                             : Intrinsic::x86_sse2_psll_d;
3292         else if (Size == 'q') // avx512.mask.psll.q.128, avx512.mask.psll.qi.128
3293           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_q
3294                             : Intrinsic::x86_sse2_psll_q;
3295         else if (Size == 'w') // avx512.mask.psll.w.128, avx512.mask.psll.wi.128
3296           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_w
3297                             : Intrinsic::x86_sse2_psll_w;
3298         else
3299           llvm_unreachable("Unexpected size");
3300       } else if (Name.endswith(".256")) {
3301         if (Size == 'd') // avx512.mask.psll.d.256, avx512.mask.psll.di.256
3302           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_d
3303                             : Intrinsic::x86_avx2_psll_d;
3304         else if (Size == 'q') // avx512.mask.psll.q.256, avx512.mask.psll.qi.256
3305           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_q
3306                             : Intrinsic::x86_avx2_psll_q;
3307         else if (Size == 'w') // avx512.mask.psll.w.256, avx512.mask.psll.wi.256
3308           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_w
3309                             : Intrinsic::x86_avx2_psll_w;
3310         else
3311           llvm_unreachable("Unexpected size");
3312       } else {
3313         if (Size == 'd') // psll.di.512, pslli.d, psll.d, psllv.d.512
3314           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_d_512 :
3315                 IsVariable  ? Intrinsic::x86_avx512_psllv_d_512 :
3316                               Intrinsic::x86_avx512_psll_d_512;
3317         else if (Size == 'q') // psll.qi.512, pslli.q, psll.q, psllv.q.512
3318           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_q_512 :
3319                 IsVariable  ? Intrinsic::x86_avx512_psllv_q_512 :
3320                               Intrinsic::x86_avx512_psll_q_512;
3321         else if (Size == 'w') // psll.wi.512, pslli.w, psll.w
3322           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_w_512
3323                             : Intrinsic::x86_avx512_psll_w_512;
3324         else
3325           llvm_unreachable("Unexpected size");
3326       }
3327 
3328       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
3329     } else if (IsX86 && Name.startswith("avx512.mask.psrl")) {
3330       bool IsImmediate = Name[16] == 'i' ||
3331                          (Name.size() > 18 && Name[18] == 'i');
3332       bool IsVariable = Name[16] == 'v';
3333       char Size = Name[16] == '.' ? Name[17] :
3334                   Name[17] == '.' ? Name[18] :
3335                   Name[18] == '.' ? Name[19] :
3336                                     Name[20];
3337 
3338       Intrinsic::ID IID;
3339       if (IsVariable && Name[17] != '.') {
3340         if (Size == 'd' && Name[17] == '2') // avx512.mask.psrlv2.di
3341           IID = Intrinsic::x86_avx2_psrlv_q;
3342         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psrlv4.di
3343           IID = Intrinsic::x86_avx2_psrlv_q_256;
3344         else if (Size == 's' && Name[17] == '4') // avx512.mask.psrlv4.si
3345           IID = Intrinsic::x86_avx2_psrlv_d;
3346         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrlv8.si
3347           IID = Intrinsic::x86_avx2_psrlv_d_256;
3348         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrlv8.hi
3349           IID = Intrinsic::x86_avx512_psrlv_w_128;
3350         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrlv16.hi
3351           IID = Intrinsic::x86_avx512_psrlv_w_256;
3352         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrlv32hi
3353           IID = Intrinsic::x86_avx512_psrlv_w_512;
3354         else
3355           llvm_unreachable("Unexpected size");
3356       } else if (Name.endswith(".128")) {
3357         if (Size == 'd') // avx512.mask.psrl.d.128, avx512.mask.psrl.di.128
3358           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_d
3359                             : Intrinsic::x86_sse2_psrl_d;
3360         else if (Size == 'q') // avx512.mask.psrl.q.128, avx512.mask.psrl.qi.128
3361           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_q
3362                             : Intrinsic::x86_sse2_psrl_q;
3363         else if (Size == 'w') // avx512.mask.psrl.w.128, avx512.mask.psrl.wi.128
3364           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_w
3365                             : Intrinsic::x86_sse2_psrl_w;
3366         else
3367           llvm_unreachable("Unexpected size");
3368       } else if (Name.endswith(".256")) {
3369         if (Size == 'd') // avx512.mask.psrl.d.256, avx512.mask.psrl.di.256
3370           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_d
3371                             : Intrinsic::x86_avx2_psrl_d;
3372         else if (Size == 'q') // avx512.mask.psrl.q.256, avx512.mask.psrl.qi.256
3373           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_q
3374                             : Intrinsic::x86_avx2_psrl_q;
3375         else if (Size == 'w') // avx512.mask.psrl.w.256, avx512.mask.psrl.wi.256
3376           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_w
3377                             : Intrinsic::x86_avx2_psrl_w;
3378         else
3379           llvm_unreachable("Unexpected size");
3380       } else {
3381         if (Size == 'd') // psrl.di.512, psrli.d, psrl.d, psrl.d.512
3382           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_d_512 :
3383                 IsVariable  ? Intrinsic::x86_avx512_psrlv_d_512 :
3384                               Intrinsic::x86_avx512_psrl_d_512;
3385         else if (Size == 'q') // psrl.qi.512, psrli.q, psrl.q, psrl.q.512
3386           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_q_512 :
3387                 IsVariable  ? Intrinsic::x86_avx512_psrlv_q_512 :
3388                               Intrinsic::x86_avx512_psrl_q_512;
3389         else if (Size == 'w') // psrl.wi.512, psrli.w, psrl.w)
3390           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_w_512
3391                             : Intrinsic::x86_avx512_psrl_w_512;
3392         else
3393           llvm_unreachable("Unexpected size");
3394       }
3395 
3396       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
3397     } else if (IsX86 && Name.startswith("avx512.mask.psra")) {
3398       bool IsImmediate = Name[16] == 'i' ||
3399                          (Name.size() > 18 && Name[18] == 'i');
3400       bool IsVariable = Name[16] == 'v';
3401       char Size = Name[16] == '.' ? Name[17] :
3402                   Name[17] == '.' ? Name[18] :
3403                   Name[18] == '.' ? Name[19] :
3404                                     Name[20];
3405 
3406       Intrinsic::ID IID;
3407       if (IsVariable && Name[17] != '.') {
3408         if (Size == 's' && Name[17] == '4') // avx512.mask.psrav4.si
3409           IID = Intrinsic::x86_avx2_psrav_d;
3410         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrav8.si
3411           IID = Intrinsic::x86_avx2_psrav_d_256;
3412         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrav8.hi
3413           IID = Intrinsic::x86_avx512_psrav_w_128;
3414         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrav16.hi
3415           IID = Intrinsic::x86_avx512_psrav_w_256;
3416         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrav32hi
3417           IID = Intrinsic::x86_avx512_psrav_w_512;
3418         else
3419           llvm_unreachable("Unexpected size");
3420       } else if (Name.endswith(".128")) {
3421         if (Size == 'd') // avx512.mask.psra.d.128, avx512.mask.psra.di.128
3422           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_d
3423                             : Intrinsic::x86_sse2_psra_d;
3424         else if (Size == 'q') // avx512.mask.psra.q.128, avx512.mask.psra.qi.128
3425           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_128 :
3426                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_128 :
3427                               Intrinsic::x86_avx512_psra_q_128;
3428         else if (Size == 'w') // avx512.mask.psra.w.128, avx512.mask.psra.wi.128
3429           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_w
3430                             : Intrinsic::x86_sse2_psra_w;
3431         else
3432           llvm_unreachable("Unexpected size");
3433       } else if (Name.endswith(".256")) {
3434         if (Size == 'd') // avx512.mask.psra.d.256, avx512.mask.psra.di.256
3435           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_d
3436                             : Intrinsic::x86_avx2_psra_d;
3437         else if (Size == 'q') // avx512.mask.psra.q.256, avx512.mask.psra.qi.256
3438           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_256 :
3439                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_256 :
3440                               Intrinsic::x86_avx512_psra_q_256;
3441         else if (Size == 'w') // avx512.mask.psra.w.256, avx512.mask.psra.wi.256
3442           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_w
3443                             : Intrinsic::x86_avx2_psra_w;
3444         else
3445           llvm_unreachable("Unexpected size");
3446       } else {
3447         if (Size == 'd') // psra.di.512, psrai.d, psra.d, psrav.d.512
3448           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_d_512 :
3449                 IsVariable  ? Intrinsic::x86_avx512_psrav_d_512 :
3450                               Intrinsic::x86_avx512_psra_d_512;
3451         else if (Size == 'q') // psra.qi.512, psrai.q, psra.q
3452           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_512 :
3453                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_512 :
3454                               Intrinsic::x86_avx512_psra_q_512;
3455         else if (Size == 'w') // psra.wi.512, psrai.w, psra.w
3456           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_w_512
3457                             : Intrinsic::x86_avx512_psra_w_512;
3458         else
3459           llvm_unreachable("Unexpected size");
3460       }
3461 
3462       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
3463     } else if (IsX86 && Name.startswith("avx512.mask.move.s")) {
3464       Rep = upgradeMaskedMove(Builder, *CI);
3465     } else if (IsX86 && Name.startswith("avx512.cvtmask2")) {
3466       Rep = UpgradeMaskToInt(Builder, *CI);
3467     } else if (IsX86 && Name.endswith(".movntdqa")) {
3468       Module *M = F->getParent();
3469       MDNode *Node = MDNode::get(
3470           C, ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
3471 
3472       Value *Ptr = CI->getArgOperand(0);
3473 
3474       // Convert the type of the pointer to a pointer to the stored type.
3475       Value *BC = Builder.CreateBitCast(
3476           Ptr, PointerType::getUnqual(CI->getType()), "cast");
3477       LoadInst *LI = Builder.CreateAlignedLoad(
3478           CI->getType(), BC,
3479           Align(CI->getType()->getPrimitiveSizeInBits().getFixedValue() / 8));
3480       LI->setMetadata(M->getMDKindID("nontemporal"), Node);
3481       Rep = LI;
3482     } else if (IsX86 && (Name.startswith("fma.vfmadd.") ||
3483                          Name.startswith("fma.vfmsub.") ||
3484                          Name.startswith("fma.vfnmadd.") ||
3485                          Name.startswith("fma.vfnmsub."))) {
3486       bool NegMul = Name[6] == 'n';
3487       bool NegAcc = NegMul ? Name[8] == 's' : Name[7] == 's';
3488       bool IsScalar = NegMul ? Name[12] == 's' : Name[11] == 's';
3489 
3490       Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3491                        CI->getArgOperand(2) };
3492 
3493       if (IsScalar) {
3494         Ops[0] = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
3495         Ops[1] = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
3496         Ops[2] = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
3497       }
3498 
3499       if (NegMul && !IsScalar)
3500         Ops[0] = Builder.CreateFNeg(Ops[0]);
3501       if (NegMul && IsScalar)
3502         Ops[1] = Builder.CreateFNeg(Ops[1]);
3503       if (NegAcc)
3504         Ops[2] = Builder.CreateFNeg(Ops[2]);
3505 
3506       Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(),
3507                                                          Intrinsic::fma,
3508                                                          Ops[0]->getType()),
3509                                Ops);
3510 
3511       if (IsScalar)
3512         Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep,
3513                                           (uint64_t)0);
3514     } else if (IsX86 && Name.startswith("fma4.vfmadd.s")) {
3515       Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3516                        CI->getArgOperand(2) };
3517 
3518       Ops[0] = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
3519       Ops[1] = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
3520       Ops[2] = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
3521 
3522       Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(),
3523                                                          Intrinsic::fma,
3524                                                          Ops[0]->getType()),
3525                                Ops);
3526 
3527       Rep = Builder.CreateInsertElement(Constant::getNullValue(CI->getType()),
3528                                         Rep, (uint64_t)0);
3529     } else if (IsX86 && (Name.startswith("avx512.mask.vfmadd.s") ||
3530                          Name.startswith("avx512.maskz.vfmadd.s") ||
3531                          Name.startswith("avx512.mask3.vfmadd.s") ||
3532                          Name.startswith("avx512.mask3.vfmsub.s") ||
3533                          Name.startswith("avx512.mask3.vfnmsub.s"))) {
3534       bool IsMask3 = Name[11] == '3';
3535       bool IsMaskZ = Name[11] == 'z';
3536       // Drop the "avx512.mask." to make it easier.
3537       Name = Name.drop_front(IsMask3 || IsMaskZ ? 13 : 12);
3538       bool NegMul = Name[2] == 'n';
3539       bool NegAcc = NegMul ? Name[4] == 's' : Name[3] == 's';
3540 
3541       Value *A = CI->getArgOperand(0);
3542       Value *B = CI->getArgOperand(1);
3543       Value *C = CI->getArgOperand(2);
3544 
3545       if (NegMul && (IsMask3 || IsMaskZ))
3546         A = Builder.CreateFNeg(A);
3547       if (NegMul && !(IsMask3 || IsMaskZ))
3548         B = Builder.CreateFNeg(B);
3549       if (NegAcc)
3550         C = Builder.CreateFNeg(C);
3551 
3552       A = Builder.CreateExtractElement(A, (uint64_t)0);
3553       B = Builder.CreateExtractElement(B, (uint64_t)0);
3554       C = Builder.CreateExtractElement(C, (uint64_t)0);
3555 
3556       if (!isa<ConstantInt>(CI->getArgOperand(4)) ||
3557           cast<ConstantInt>(CI->getArgOperand(4))->getZExtValue() != 4) {
3558         Value *Ops[] = { A, B, C, CI->getArgOperand(4) };
3559 
3560         Intrinsic::ID IID;
3561         if (Name.back() == 'd')
3562           IID = Intrinsic::x86_avx512_vfmadd_f64;
3563         else
3564           IID = Intrinsic::x86_avx512_vfmadd_f32;
3565         Function *FMA = Intrinsic::getDeclaration(CI->getModule(), IID);
3566         Rep = Builder.CreateCall(FMA, Ops);
3567       } else {
3568         Function *FMA = Intrinsic::getDeclaration(CI->getModule(),
3569                                                   Intrinsic::fma,
3570                                                   A->getType());
3571         Rep = Builder.CreateCall(FMA, { A, B, C });
3572       }
3573 
3574       Value *PassThru = IsMaskZ ? Constant::getNullValue(Rep->getType()) :
3575                         IsMask3 ? C : A;
3576 
3577       // For Mask3 with NegAcc, we need to create a new extractelement that
3578       // avoids the negation above.
3579       if (NegAcc && IsMask3)
3580         PassThru = Builder.CreateExtractElement(CI->getArgOperand(2),
3581                                                 (uint64_t)0);
3582 
3583       Rep = EmitX86ScalarSelect(Builder, CI->getArgOperand(3),
3584                                 Rep, PassThru);
3585       Rep = Builder.CreateInsertElement(CI->getArgOperand(IsMask3 ? 2 : 0),
3586                                         Rep, (uint64_t)0);
3587     } else if (IsX86 && (Name.startswith("avx512.mask.vfmadd.p") ||
3588                          Name.startswith("avx512.mask.vfnmadd.p") ||
3589                          Name.startswith("avx512.mask.vfnmsub.p") ||
3590                          Name.startswith("avx512.mask3.vfmadd.p") ||
3591                          Name.startswith("avx512.mask3.vfmsub.p") ||
3592                          Name.startswith("avx512.mask3.vfnmsub.p") ||
3593                          Name.startswith("avx512.maskz.vfmadd.p"))) {
3594       bool IsMask3 = Name[11] == '3';
3595       bool IsMaskZ = Name[11] == 'z';
3596       // Drop the "avx512.mask." to make it easier.
3597       Name = Name.drop_front(IsMask3 || IsMaskZ ? 13 : 12);
3598       bool NegMul = Name[2] == 'n';
3599       bool NegAcc = NegMul ? Name[4] == 's' : Name[3] == 's';
3600 
3601       Value *A = CI->getArgOperand(0);
3602       Value *B = CI->getArgOperand(1);
3603       Value *C = CI->getArgOperand(2);
3604 
3605       if (NegMul && (IsMask3 || IsMaskZ))
3606         A = Builder.CreateFNeg(A);
3607       if (NegMul && !(IsMask3 || IsMaskZ))
3608         B = Builder.CreateFNeg(B);
3609       if (NegAcc)
3610         C = Builder.CreateFNeg(C);
3611 
3612       if (CI->arg_size() == 5 &&
3613           (!isa<ConstantInt>(CI->getArgOperand(4)) ||
3614            cast<ConstantInt>(CI->getArgOperand(4))->getZExtValue() != 4)) {
3615         Intrinsic::ID IID;
3616         // Check the character before ".512" in string.
3617         if (Name[Name.size()-5] == 's')
3618           IID = Intrinsic::x86_avx512_vfmadd_ps_512;
3619         else
3620           IID = Intrinsic::x86_avx512_vfmadd_pd_512;
3621 
3622         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3623                                  { A, B, C, CI->getArgOperand(4) });
3624       } else {
3625         Function *FMA = Intrinsic::getDeclaration(CI->getModule(),
3626                                                   Intrinsic::fma,
3627                                                   A->getType());
3628         Rep = Builder.CreateCall(FMA, { A, B, C });
3629       }
3630 
3631       Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(CI->getType()) :
3632                         IsMask3 ? CI->getArgOperand(2) :
3633                                   CI->getArgOperand(0);
3634 
3635       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3636     } else if (IsX86 &&  Name.startswith("fma.vfmsubadd.p")) {
3637       unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3638       unsigned EltWidth = CI->getType()->getScalarSizeInBits();
3639       Intrinsic::ID IID;
3640       if (VecWidth == 128 && EltWidth == 32)
3641         IID = Intrinsic::x86_fma_vfmaddsub_ps;
3642       else if (VecWidth == 256 && EltWidth == 32)
3643         IID = Intrinsic::x86_fma_vfmaddsub_ps_256;
3644       else if (VecWidth == 128 && EltWidth == 64)
3645         IID = Intrinsic::x86_fma_vfmaddsub_pd;
3646       else if (VecWidth == 256 && EltWidth == 64)
3647         IID = Intrinsic::x86_fma_vfmaddsub_pd_256;
3648       else
3649         llvm_unreachable("Unexpected intrinsic");
3650 
3651       Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3652                        CI->getArgOperand(2) };
3653       Ops[2] = Builder.CreateFNeg(Ops[2]);
3654       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3655                                Ops);
3656     } else if (IsX86 && (Name.startswith("avx512.mask.vfmaddsub.p") ||
3657                          Name.startswith("avx512.mask3.vfmaddsub.p") ||
3658                          Name.startswith("avx512.maskz.vfmaddsub.p") ||
3659                          Name.startswith("avx512.mask3.vfmsubadd.p"))) {
3660       bool IsMask3 = Name[11] == '3';
3661       bool IsMaskZ = Name[11] == 'z';
3662       // Drop the "avx512.mask." to make it easier.
3663       Name = Name.drop_front(IsMask3 || IsMaskZ ? 13 : 12);
3664       bool IsSubAdd = Name[3] == 's';
3665       if (CI->arg_size() == 5) {
3666         Intrinsic::ID IID;
3667         // Check the character before ".512" in string.
3668         if (Name[Name.size()-5] == 's')
3669           IID = Intrinsic::x86_avx512_vfmaddsub_ps_512;
3670         else
3671           IID = Intrinsic::x86_avx512_vfmaddsub_pd_512;
3672 
3673         Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3674                          CI->getArgOperand(2), CI->getArgOperand(4) };
3675         if (IsSubAdd)
3676           Ops[2] = Builder.CreateFNeg(Ops[2]);
3677 
3678         Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3679                                  Ops);
3680       } else {
3681         int NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
3682 
3683         Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3684                          CI->getArgOperand(2) };
3685 
3686         Function *FMA = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::fma,
3687                                                   Ops[0]->getType());
3688         Value *Odd = Builder.CreateCall(FMA, Ops);
3689         Ops[2] = Builder.CreateFNeg(Ops[2]);
3690         Value *Even = Builder.CreateCall(FMA, Ops);
3691 
3692         if (IsSubAdd)
3693           std::swap(Even, Odd);
3694 
3695         SmallVector<int, 32> Idxs(NumElts);
3696         for (int i = 0; i != NumElts; ++i)
3697           Idxs[i] = i + (i % 2) * NumElts;
3698 
3699         Rep = Builder.CreateShuffleVector(Even, Odd, Idxs);
3700       }
3701 
3702       Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(CI->getType()) :
3703                         IsMask3 ? CI->getArgOperand(2) :
3704                                   CI->getArgOperand(0);
3705 
3706       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3707     } else if (IsX86 && (Name.startswith("avx512.mask.pternlog.") ||
3708                          Name.startswith("avx512.maskz.pternlog."))) {
3709       bool ZeroMask = Name[11] == 'z';
3710       unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3711       unsigned EltWidth = CI->getType()->getScalarSizeInBits();
3712       Intrinsic::ID IID;
3713       if (VecWidth == 128 && EltWidth == 32)
3714         IID = Intrinsic::x86_avx512_pternlog_d_128;
3715       else if (VecWidth == 256 && EltWidth == 32)
3716         IID = Intrinsic::x86_avx512_pternlog_d_256;
3717       else if (VecWidth == 512 && EltWidth == 32)
3718         IID = Intrinsic::x86_avx512_pternlog_d_512;
3719       else if (VecWidth == 128 && EltWidth == 64)
3720         IID = Intrinsic::x86_avx512_pternlog_q_128;
3721       else if (VecWidth == 256 && EltWidth == 64)
3722         IID = Intrinsic::x86_avx512_pternlog_q_256;
3723       else if (VecWidth == 512 && EltWidth == 64)
3724         IID = Intrinsic::x86_avx512_pternlog_q_512;
3725       else
3726         llvm_unreachable("Unexpected intrinsic");
3727 
3728       Value *Args[] = { CI->getArgOperand(0) , CI->getArgOperand(1),
3729                         CI->getArgOperand(2), CI->getArgOperand(3) };
3730       Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3731                                Args);
3732       Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3733                                  : CI->getArgOperand(0);
3734       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep, PassThru);
3735     } else if (IsX86 && (Name.startswith("avx512.mask.vpmadd52") ||
3736                          Name.startswith("avx512.maskz.vpmadd52"))) {
3737       bool ZeroMask = Name[11] == 'z';
3738       bool High = Name[20] == 'h' || Name[21] == 'h';
3739       unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3740       Intrinsic::ID IID;
3741       if (VecWidth == 128 && !High)
3742         IID = Intrinsic::x86_avx512_vpmadd52l_uq_128;
3743       else if (VecWidth == 256 && !High)
3744         IID = Intrinsic::x86_avx512_vpmadd52l_uq_256;
3745       else if (VecWidth == 512 && !High)
3746         IID = Intrinsic::x86_avx512_vpmadd52l_uq_512;
3747       else if (VecWidth == 128 && High)
3748         IID = Intrinsic::x86_avx512_vpmadd52h_uq_128;
3749       else if (VecWidth == 256 && High)
3750         IID = Intrinsic::x86_avx512_vpmadd52h_uq_256;
3751       else if (VecWidth == 512 && High)
3752         IID = Intrinsic::x86_avx512_vpmadd52h_uq_512;
3753       else
3754         llvm_unreachable("Unexpected intrinsic");
3755 
3756       Value *Args[] = { CI->getArgOperand(0) , CI->getArgOperand(1),
3757                         CI->getArgOperand(2) };
3758       Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3759                                Args);
3760       Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3761                                  : CI->getArgOperand(0);
3762       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3763     } else if (IsX86 && (Name.startswith("avx512.mask.vpermi2var.") ||
3764                          Name.startswith("avx512.mask.vpermt2var.") ||
3765                          Name.startswith("avx512.maskz.vpermt2var."))) {
3766       bool ZeroMask = Name[11] == 'z';
3767       bool IndexForm = Name[17] == 'i';
3768       Rep = UpgradeX86VPERMT2Intrinsics(Builder, *CI, ZeroMask, IndexForm);
3769     } else if (IsX86 && (Name.startswith("avx512.mask.vpdpbusd.") ||
3770                          Name.startswith("avx512.maskz.vpdpbusd.") ||
3771                          Name.startswith("avx512.mask.vpdpbusds.") ||
3772                          Name.startswith("avx512.maskz.vpdpbusds."))) {
3773       bool ZeroMask = Name[11] == 'z';
3774       bool IsSaturating = Name[ZeroMask ? 21 : 20] == 's';
3775       unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3776       Intrinsic::ID IID;
3777       if (VecWidth == 128 && !IsSaturating)
3778         IID = Intrinsic::x86_avx512_vpdpbusd_128;
3779       else if (VecWidth == 256 && !IsSaturating)
3780         IID = Intrinsic::x86_avx512_vpdpbusd_256;
3781       else if (VecWidth == 512 && !IsSaturating)
3782         IID = Intrinsic::x86_avx512_vpdpbusd_512;
3783       else if (VecWidth == 128 && IsSaturating)
3784         IID = Intrinsic::x86_avx512_vpdpbusds_128;
3785       else if (VecWidth == 256 && IsSaturating)
3786         IID = Intrinsic::x86_avx512_vpdpbusds_256;
3787       else if (VecWidth == 512 && IsSaturating)
3788         IID = Intrinsic::x86_avx512_vpdpbusds_512;
3789       else
3790         llvm_unreachable("Unexpected intrinsic");
3791 
3792       Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3793                         CI->getArgOperand(2)  };
3794       Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3795                                Args);
3796       Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3797                                  : CI->getArgOperand(0);
3798       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3799     } else if (IsX86 && (Name.startswith("avx512.mask.vpdpwssd.") ||
3800                          Name.startswith("avx512.maskz.vpdpwssd.") ||
3801                          Name.startswith("avx512.mask.vpdpwssds.") ||
3802                          Name.startswith("avx512.maskz.vpdpwssds."))) {
3803       bool ZeroMask = Name[11] == 'z';
3804       bool IsSaturating = Name[ZeroMask ? 21 : 20] == 's';
3805       unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3806       Intrinsic::ID IID;
3807       if (VecWidth == 128 && !IsSaturating)
3808         IID = Intrinsic::x86_avx512_vpdpwssd_128;
3809       else if (VecWidth == 256 && !IsSaturating)
3810         IID = Intrinsic::x86_avx512_vpdpwssd_256;
3811       else if (VecWidth == 512 && !IsSaturating)
3812         IID = Intrinsic::x86_avx512_vpdpwssd_512;
3813       else if (VecWidth == 128 && IsSaturating)
3814         IID = Intrinsic::x86_avx512_vpdpwssds_128;
3815       else if (VecWidth == 256 && IsSaturating)
3816         IID = Intrinsic::x86_avx512_vpdpwssds_256;
3817       else if (VecWidth == 512 && IsSaturating)
3818         IID = Intrinsic::x86_avx512_vpdpwssds_512;
3819       else
3820         llvm_unreachable("Unexpected intrinsic");
3821 
3822       Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3823                         CI->getArgOperand(2)  };
3824       Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3825                                Args);
3826       Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3827                                  : CI->getArgOperand(0);
3828       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3829     } else if (IsX86 && (Name == "addcarryx.u32" || Name == "addcarryx.u64" ||
3830                          Name == "addcarry.u32" || Name == "addcarry.u64" ||
3831                          Name == "subborrow.u32" || Name == "subborrow.u64")) {
3832       Intrinsic::ID IID;
3833       if (Name[0] == 'a' && Name.back() == '2')
3834         IID = Intrinsic::x86_addcarry_32;
3835       else if (Name[0] == 'a' && Name.back() == '4')
3836         IID = Intrinsic::x86_addcarry_64;
3837       else if (Name[0] == 's' && Name.back() == '2')
3838         IID = Intrinsic::x86_subborrow_32;
3839       else if (Name[0] == 's' && Name.back() == '4')
3840         IID = Intrinsic::x86_subborrow_64;
3841       else
3842         llvm_unreachable("Unexpected intrinsic");
3843 
3844       // Make a call with 3 operands.
3845       Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3846                         CI->getArgOperand(2)};
3847       Value *NewCall = Builder.CreateCall(
3848                                 Intrinsic::getDeclaration(CI->getModule(), IID),
3849                                 Args);
3850 
3851       // Extract the second result and store it.
3852       Value *Data = Builder.CreateExtractValue(NewCall, 1);
3853       // Cast the pointer to the right type.
3854       Value *Ptr = Builder.CreateBitCast(CI->getArgOperand(3),
3855                                  llvm::PointerType::getUnqual(Data->getType()));
3856       Builder.CreateAlignedStore(Data, Ptr, Align(1));
3857       // Replace the original call result with the first result of the new call.
3858       Value *CF = Builder.CreateExtractValue(NewCall, 0);
3859 
3860       CI->replaceAllUsesWith(CF);
3861       Rep = nullptr;
3862     } else if (IsX86 && Name.startswith("avx512.mask.") &&
3863                upgradeAVX512MaskToSelect(Name, Builder, *CI, Rep)) {
3864       // Rep will be updated by the call in the condition.
3865     } else if (IsNVVM && (Name == "abs.i" || Name == "abs.ll")) {
3866       Value *Arg = CI->getArgOperand(0);
3867       Value *Neg = Builder.CreateNeg(Arg, "neg");
3868       Value *Cmp = Builder.CreateICmpSGE(
3869           Arg, llvm::Constant::getNullValue(Arg->getType()), "abs.cond");
3870       Rep = Builder.CreateSelect(Cmp, Arg, Neg, "abs");
3871     } else if (IsNVVM && (Name.startswith("atomic.load.add.f32.p") ||
3872                           Name.startswith("atomic.load.add.f64.p"))) {
3873       Value *Ptr = CI->getArgOperand(0);
3874       Value *Val = CI->getArgOperand(1);
3875       Rep = Builder.CreateAtomicRMW(AtomicRMWInst::FAdd, Ptr, Val, MaybeAlign(),
3876                                     AtomicOrdering::SequentiallyConsistent);
3877     } else if (IsNVVM && (Name == "max.i" || Name == "max.ll" ||
3878                           Name == "max.ui" || Name == "max.ull")) {
3879       Value *Arg0 = CI->getArgOperand(0);
3880       Value *Arg1 = CI->getArgOperand(1);
3881       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
3882                        ? Builder.CreateICmpUGE(Arg0, Arg1, "max.cond")
3883                        : Builder.CreateICmpSGE(Arg0, Arg1, "max.cond");
3884       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "max");
3885     } else if (IsNVVM && (Name == "min.i" || Name == "min.ll" ||
3886                           Name == "min.ui" || Name == "min.ull")) {
3887       Value *Arg0 = CI->getArgOperand(0);
3888       Value *Arg1 = CI->getArgOperand(1);
3889       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
3890                        ? Builder.CreateICmpULE(Arg0, Arg1, "min.cond")
3891                        : Builder.CreateICmpSLE(Arg0, Arg1, "min.cond");
3892       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "min");
3893     } else if (IsNVVM && Name == "clz.ll") {
3894       // llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 and returns an i64.
3895       Value *Arg = CI->getArgOperand(0);
3896       Value *Ctlz = Builder.CreateCall(
3897           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
3898                                     {Arg->getType()}),
3899           {Arg, Builder.getFalse()}, "ctlz");
3900       Rep = Builder.CreateTrunc(Ctlz, Builder.getInt32Ty(), "ctlz.trunc");
3901     } else if (IsNVVM && Name == "popc.ll") {
3902       // llvm.nvvm.popc.ll returns an i32, but llvm.ctpop.i64 and returns an
3903       // i64.
3904       Value *Arg = CI->getArgOperand(0);
3905       Value *Popc = Builder.CreateCall(
3906           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
3907                                     {Arg->getType()}),
3908           Arg, "ctpop");
3909       Rep = Builder.CreateTrunc(Popc, Builder.getInt32Ty(), "ctpop.trunc");
3910     } else if (IsNVVM && Name == "h2f") {
3911       Rep = Builder.CreateCall(Intrinsic::getDeclaration(
3912                                    F->getParent(), Intrinsic::convert_from_fp16,
3913                                    {Builder.getFloatTy()}),
3914                                CI->getArgOperand(0), "h2f");
3915     } else if (IsARM) {
3916       Rep = UpgradeARMIntrinsicCall(Name, CI, F, Builder);
3917     } else {
3918       llvm_unreachable("Unknown function for CallBase upgrade.");
3919     }
3920 
3921     if (Rep)
3922       CI->replaceAllUsesWith(Rep);
3923     CI->eraseFromParent();
3924     return;
3925   }
3926 
3927   const auto &DefaultCase = [&]() -> void {
3928     if (CI->getFunctionType() == NewFn->getFunctionType()) {
3929       // Handle generic mangling change.
3930       assert(
3931           (CI->getCalledFunction()->getName() != NewFn->getName()) &&
3932           "Unknown function for CallBase upgrade and isn't just a name change");
3933       CI->setCalledFunction(NewFn);
3934       return;
3935     }
3936 
3937     // This must be an upgrade from a named to a literal struct.
3938     if (auto *OldST = dyn_cast<StructType>(CI->getType())) {
3939       assert(OldST != NewFn->getReturnType() &&
3940              "Return type must have changed");
3941       assert(OldST->getNumElements() ==
3942                  cast<StructType>(NewFn->getReturnType())->getNumElements() &&
3943              "Must have same number of elements");
3944 
3945       SmallVector<Value *> Args(CI->args());
3946       Value *NewCI = Builder.CreateCall(NewFn, Args);
3947       Value *Res = PoisonValue::get(OldST);
3948       for (unsigned Idx = 0; Idx < OldST->getNumElements(); ++Idx) {
3949         Value *Elem = Builder.CreateExtractValue(NewCI, Idx);
3950         Res = Builder.CreateInsertValue(Res, Elem, Idx);
3951       }
3952       CI->replaceAllUsesWith(Res);
3953       CI->eraseFromParent();
3954       return;
3955     }
3956 
3957     // We're probably about to produce something invalid. Let the verifier catch
3958     // it instead of dying here.
3959     CI->setCalledOperand(
3960         ConstantExpr::getPointerCast(NewFn, CI->getCalledOperand()->getType()));
3961     return;
3962   };
3963   CallInst *NewCall = nullptr;
3964   switch (NewFn->getIntrinsicID()) {
3965   default: {
3966     DefaultCase();
3967     return;
3968   }
3969   case Intrinsic::arm_neon_vst1:
3970   case Intrinsic::arm_neon_vst2:
3971   case Intrinsic::arm_neon_vst3:
3972   case Intrinsic::arm_neon_vst4:
3973   case Intrinsic::arm_neon_vst2lane:
3974   case Intrinsic::arm_neon_vst3lane:
3975   case Intrinsic::arm_neon_vst4lane: {
3976     SmallVector<Value *, 4> Args(CI->args());
3977     NewCall = Builder.CreateCall(NewFn, Args);
3978     break;
3979   }
3980   case Intrinsic::aarch64_sve_bfmlalb_lane_v2:
3981   case Intrinsic::aarch64_sve_bfmlalt_lane_v2:
3982   case Intrinsic::aarch64_sve_bfdot_lane_v2: {
3983     LLVMContext &Ctx = F->getParent()->getContext();
3984     SmallVector<Value *, 4> Args(CI->args());
3985     Args[3] = ConstantInt::get(Type::getInt32Ty(Ctx),
3986                                cast<ConstantInt>(Args[3])->getZExtValue());
3987     NewCall = Builder.CreateCall(NewFn, Args);
3988     break;
3989   }
3990   case Intrinsic::aarch64_sve_ld3_sret:
3991   case Intrinsic::aarch64_sve_ld4_sret:
3992   case Intrinsic::aarch64_sve_ld2_sret: {
3993     StringRef Name = F->getName();
3994     Name = Name.substr(5);
3995     unsigned N = StringSwitch<unsigned>(Name)
3996                      .StartsWith("aarch64.sve.ld2", 2)
3997                      .StartsWith("aarch64.sve.ld3", 3)
3998                      .StartsWith("aarch64.sve.ld4", 4)
3999                      .Default(0);
4000     ScalableVectorType *RetTy =
4001         dyn_cast<ScalableVectorType>(F->getReturnType());
4002     unsigned MinElts = RetTy->getMinNumElements() / N;
4003     SmallVector<Value *, 2> Args(CI->args());
4004     Value *NewLdCall = Builder.CreateCall(NewFn, Args);
4005     Value *Ret = llvm::PoisonValue::get(RetTy);
4006     for (unsigned I = 0; I < N; I++) {
4007       Value *Idx = ConstantInt::get(Type::getInt64Ty(C), I * MinElts);
4008       Value *SRet = Builder.CreateExtractValue(NewLdCall, I);
4009       Ret = Builder.CreateInsertVector(RetTy, Ret, SRet, Idx);
4010     }
4011     NewCall = dyn_cast<CallInst>(Ret);
4012     break;
4013   }
4014 
4015   case Intrinsic::vector_extract: {
4016     StringRef Name = F->getName();
4017     Name = Name.substr(5); // Strip llvm
4018     if (!Name.startswith("aarch64.sve.tuple.get")) {
4019       DefaultCase();
4020       return;
4021     }
4022     ScalableVectorType *RetTy =
4023         dyn_cast<ScalableVectorType>(F->getReturnType());
4024     unsigned MinElts = RetTy->getMinNumElements();
4025     unsigned I = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
4026     Value *NewIdx = ConstantInt::get(Type::getInt64Ty(C), I * MinElts);
4027     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0), NewIdx});
4028     break;
4029   }
4030 
4031   case Intrinsic::vector_insert: {
4032     StringRef Name = F->getName();
4033     Name = Name.substr(5);
4034     if (!Name.startswith("aarch64.sve.tuple")) {
4035       DefaultCase();
4036       return;
4037     }
4038     if (Name.startswith("aarch64.sve.tuple.set")) {
4039       unsigned I = dyn_cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
4040       ScalableVectorType *Ty =
4041           dyn_cast<ScalableVectorType>(CI->getArgOperand(2)->getType());
4042       Value *NewIdx =
4043           ConstantInt::get(Type::getInt64Ty(C), I * Ty->getMinNumElements());
4044       NewCall = Builder.CreateCall(
4045           NewFn, {CI->getArgOperand(0), CI->getArgOperand(2), NewIdx});
4046       break;
4047     }
4048     if (Name.startswith("aarch64.sve.tuple.create")) {
4049       unsigned N = StringSwitch<unsigned>(Name)
4050                        .StartsWith("aarch64.sve.tuple.create2", 2)
4051                        .StartsWith("aarch64.sve.tuple.create3", 3)
4052                        .StartsWith("aarch64.sve.tuple.create4", 4)
4053                        .Default(0);
4054       assert(N > 1 && "Create is expected to be between 2-4");
4055       ScalableVectorType *RetTy =
4056           dyn_cast<ScalableVectorType>(F->getReturnType());
4057       Value *Ret = llvm::PoisonValue::get(RetTy);
4058       unsigned MinElts = RetTy->getMinNumElements() / N;
4059       for (unsigned I = 0; I < N; I++) {
4060         Value *Idx = ConstantInt::get(Type::getInt64Ty(C), I * MinElts);
4061         Value *V = CI->getArgOperand(I);
4062         Ret = Builder.CreateInsertVector(RetTy, Ret, V, Idx);
4063       }
4064       NewCall = dyn_cast<CallInst>(Ret);
4065     }
4066     break;
4067   }
4068 
4069   case Intrinsic::arm_neon_bfdot:
4070   case Intrinsic::arm_neon_bfmmla:
4071   case Intrinsic::arm_neon_bfmlalb:
4072   case Intrinsic::arm_neon_bfmlalt:
4073   case Intrinsic::aarch64_neon_bfdot:
4074   case Intrinsic::aarch64_neon_bfmmla:
4075   case Intrinsic::aarch64_neon_bfmlalb:
4076   case Intrinsic::aarch64_neon_bfmlalt: {
4077     SmallVector<Value *, 3> Args;
4078     assert(CI->arg_size() == 3 &&
4079            "Mismatch between function args and call args");
4080     size_t OperandWidth =
4081         CI->getArgOperand(1)->getType()->getPrimitiveSizeInBits();
4082     assert((OperandWidth == 64 || OperandWidth == 128) &&
4083            "Unexpected operand width");
4084     Type *NewTy = FixedVectorType::get(Type::getBFloatTy(C), OperandWidth / 16);
4085     auto Iter = CI->args().begin();
4086     Args.push_back(*Iter++);
4087     Args.push_back(Builder.CreateBitCast(*Iter++, NewTy));
4088     Args.push_back(Builder.CreateBitCast(*Iter++, NewTy));
4089     NewCall = Builder.CreateCall(NewFn, Args);
4090     break;
4091   }
4092 
4093   case Intrinsic::bitreverse:
4094     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
4095     break;
4096 
4097   case Intrinsic::ctlz:
4098   case Intrinsic::cttz:
4099     assert(CI->arg_size() == 1 &&
4100            "Mismatch between function args and call args");
4101     NewCall =
4102         Builder.CreateCall(NewFn, {CI->getArgOperand(0), Builder.getFalse()});
4103     break;
4104 
4105   case Intrinsic::objectsize: {
4106     Value *NullIsUnknownSize =
4107         CI->arg_size() == 2 ? Builder.getFalse() : CI->getArgOperand(2);
4108     Value *Dynamic =
4109         CI->arg_size() < 4 ? Builder.getFalse() : CI->getArgOperand(3);
4110     NewCall = Builder.CreateCall(
4111         NewFn, {CI->getArgOperand(0), CI->getArgOperand(1), NullIsUnknownSize, Dynamic});
4112     break;
4113   }
4114 
4115   case Intrinsic::ctpop:
4116     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
4117     break;
4118 
4119   case Intrinsic::convert_from_fp16:
4120     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
4121     break;
4122 
4123   case Intrinsic::dbg_value:
4124     // Upgrade from the old version that had an extra offset argument.
4125     assert(CI->arg_size() == 4);
4126     // Drop nonzero offsets instead of attempting to upgrade them.
4127     if (auto *Offset = dyn_cast_or_null<Constant>(CI->getArgOperand(1)))
4128       if (Offset->isZeroValue()) {
4129         NewCall = Builder.CreateCall(
4130             NewFn,
4131             {CI->getArgOperand(0), CI->getArgOperand(2), CI->getArgOperand(3)});
4132         break;
4133       }
4134     CI->eraseFromParent();
4135     return;
4136 
4137   case Intrinsic::ptr_annotation:
4138     // Upgrade from versions that lacked the annotation attribute argument.
4139     if (CI->arg_size() != 4) {
4140       DefaultCase();
4141       return;
4142     }
4143 
4144     // Create a new call with an added null annotation attribute argument.
4145     NewCall = Builder.CreateCall(
4146         NewFn,
4147         {CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2),
4148          CI->getArgOperand(3), Constant::getNullValue(Builder.getInt8PtrTy())});
4149     NewCall->takeName(CI);
4150     CI->replaceAllUsesWith(NewCall);
4151     CI->eraseFromParent();
4152     return;
4153 
4154   case Intrinsic::var_annotation:
4155     // Upgrade from versions that lacked the annotation attribute argument.
4156     if (CI->arg_size() != 4) {
4157       DefaultCase();
4158       return;
4159     }
4160     // Create a new call with an added null annotation attribute argument.
4161     NewCall = Builder.CreateCall(
4162         NewFn,
4163         {CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2),
4164          CI->getArgOperand(3), Constant::getNullValue(Builder.getInt8PtrTy())});
4165     NewCall->takeName(CI);
4166     CI->replaceAllUsesWith(NewCall);
4167     CI->eraseFromParent();
4168     return;
4169 
4170   case Intrinsic::x86_xop_vfrcz_ss:
4171   case Intrinsic::x86_xop_vfrcz_sd:
4172     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(1)});
4173     break;
4174 
4175   case Intrinsic::x86_xop_vpermil2pd:
4176   case Intrinsic::x86_xop_vpermil2ps:
4177   case Intrinsic::x86_xop_vpermil2pd_256:
4178   case Intrinsic::x86_xop_vpermil2ps_256: {
4179     SmallVector<Value *, 4> Args(CI->args());
4180     VectorType *FltIdxTy = cast<VectorType>(Args[2]->getType());
4181     VectorType *IntIdxTy = VectorType::getInteger(FltIdxTy);
4182     Args[2] = Builder.CreateBitCast(Args[2], IntIdxTy);
4183     NewCall = Builder.CreateCall(NewFn, Args);
4184     break;
4185   }
4186 
4187   case Intrinsic::x86_sse41_ptestc:
4188   case Intrinsic::x86_sse41_ptestz:
4189   case Intrinsic::x86_sse41_ptestnzc: {
4190     // The arguments for these intrinsics used to be v4f32, and changed
4191     // to v2i64. This is purely a nop, since those are bitwise intrinsics.
4192     // So, the only thing required is a bitcast for both arguments.
4193     // First, check the arguments have the old type.
4194     Value *Arg0 = CI->getArgOperand(0);
4195     if (Arg0->getType() != FixedVectorType::get(Type::getFloatTy(C), 4))
4196       return;
4197 
4198     // Old intrinsic, add bitcasts
4199     Value *Arg1 = CI->getArgOperand(1);
4200 
4201     auto *NewVecTy = FixedVectorType::get(Type::getInt64Ty(C), 2);
4202 
4203     Value *BC0 = Builder.CreateBitCast(Arg0, NewVecTy, "cast");
4204     Value *BC1 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
4205 
4206     NewCall = Builder.CreateCall(NewFn, {BC0, BC1});
4207     break;
4208   }
4209 
4210   case Intrinsic::x86_rdtscp: {
4211     // This used to take 1 arguments. If we have no arguments, it is already
4212     // upgraded.
4213     if (CI->getNumOperands() == 0)
4214       return;
4215 
4216     NewCall = Builder.CreateCall(NewFn);
4217     // Extract the second result and store it.
4218     Value *Data = Builder.CreateExtractValue(NewCall, 1);
4219     // Cast the pointer to the right type.
4220     Value *Ptr = Builder.CreateBitCast(CI->getArgOperand(0),
4221                                  llvm::PointerType::getUnqual(Data->getType()));
4222     Builder.CreateAlignedStore(Data, Ptr, Align(1));
4223     // Replace the original call result with the first result of the new call.
4224     Value *TSC = Builder.CreateExtractValue(NewCall, 0);
4225 
4226     NewCall->takeName(CI);
4227     CI->replaceAllUsesWith(TSC);
4228     CI->eraseFromParent();
4229     return;
4230   }
4231 
4232   case Intrinsic::x86_sse41_insertps:
4233   case Intrinsic::x86_sse41_dppd:
4234   case Intrinsic::x86_sse41_dpps:
4235   case Intrinsic::x86_sse41_mpsadbw:
4236   case Intrinsic::x86_avx_dp_ps_256:
4237   case Intrinsic::x86_avx2_mpsadbw: {
4238     // Need to truncate the last argument from i32 to i8 -- this argument models
4239     // an inherently 8-bit immediate operand to these x86 instructions.
4240     SmallVector<Value *, 4> Args(CI->args());
4241 
4242     // Replace the last argument with a trunc.
4243     Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
4244     NewCall = Builder.CreateCall(NewFn, Args);
4245     break;
4246   }
4247 
4248   case Intrinsic::x86_avx512_mask_cmp_pd_128:
4249   case Intrinsic::x86_avx512_mask_cmp_pd_256:
4250   case Intrinsic::x86_avx512_mask_cmp_pd_512:
4251   case Intrinsic::x86_avx512_mask_cmp_ps_128:
4252   case Intrinsic::x86_avx512_mask_cmp_ps_256:
4253   case Intrinsic::x86_avx512_mask_cmp_ps_512: {
4254     SmallVector<Value *, 4> Args(CI->args());
4255     unsigned NumElts =
4256         cast<FixedVectorType>(Args[0]->getType())->getNumElements();
4257     Args[3] = getX86MaskVec(Builder, Args[3], NumElts);
4258 
4259     NewCall = Builder.CreateCall(NewFn, Args);
4260     Value *Res = ApplyX86MaskOn1BitsVec(Builder, NewCall, nullptr);
4261 
4262     NewCall->takeName(CI);
4263     CI->replaceAllUsesWith(Res);
4264     CI->eraseFromParent();
4265     return;
4266   }
4267 
4268   case Intrinsic::x86_avx512bf16_cvtne2ps2bf16_128:
4269   case Intrinsic::x86_avx512bf16_cvtne2ps2bf16_256:
4270   case Intrinsic::x86_avx512bf16_cvtne2ps2bf16_512:
4271   case Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128:
4272   case Intrinsic::x86_avx512bf16_cvtneps2bf16_256:
4273   case Intrinsic::x86_avx512bf16_cvtneps2bf16_512: {
4274     SmallVector<Value *, 4> Args(CI->args());
4275     unsigned NumElts = cast<FixedVectorType>(CI->getType())->getNumElements();
4276     if (NewFn->getIntrinsicID() ==
4277         Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128)
4278       Args[1] = Builder.CreateBitCast(
4279           Args[1], FixedVectorType::get(Builder.getBFloatTy(), NumElts));
4280 
4281     NewCall = Builder.CreateCall(NewFn, Args);
4282     Value *Res = Builder.CreateBitCast(
4283         NewCall, FixedVectorType::get(Builder.getInt16Ty(), NumElts));
4284 
4285     NewCall->takeName(CI);
4286     CI->replaceAllUsesWith(Res);
4287     CI->eraseFromParent();
4288     return;
4289   }
4290   case Intrinsic::x86_avx512bf16_dpbf16ps_128:
4291   case Intrinsic::x86_avx512bf16_dpbf16ps_256:
4292   case Intrinsic::x86_avx512bf16_dpbf16ps_512:{
4293     SmallVector<Value *, 4> Args(CI->args());
4294     unsigned NumElts =
4295         cast<FixedVectorType>(CI->getType())->getNumElements() * 2;
4296     Args[1] = Builder.CreateBitCast(
4297         Args[1], FixedVectorType::get(Builder.getBFloatTy(), NumElts));
4298     Args[2] = Builder.CreateBitCast(
4299         Args[2], FixedVectorType::get(Builder.getBFloatTy(), NumElts));
4300 
4301     NewCall = Builder.CreateCall(NewFn, Args);
4302     break;
4303   }
4304 
4305   case Intrinsic::thread_pointer: {
4306     NewCall = Builder.CreateCall(NewFn, {});
4307     break;
4308   }
4309 
4310   case Intrinsic::invariant_start:
4311   case Intrinsic::invariant_end: {
4312     SmallVector<Value *, 4> Args(CI->args());
4313     NewCall = Builder.CreateCall(NewFn, Args);
4314     break;
4315   }
4316   case Intrinsic::masked_load:
4317   case Intrinsic::masked_store:
4318   case Intrinsic::masked_gather:
4319   case Intrinsic::masked_scatter: {
4320     SmallVector<Value *, 4> Args(CI->args());
4321     NewCall = Builder.CreateCall(NewFn, Args);
4322     NewCall->copyMetadata(*CI);
4323     break;
4324   }
4325 
4326   case Intrinsic::memcpy:
4327   case Intrinsic::memmove:
4328   case Intrinsic::memset: {
4329     // We have to make sure that the call signature is what we're expecting.
4330     // We only want to change the old signatures by removing the alignment arg:
4331     //  @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i32, i1)
4332     //    -> @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i1)
4333     //  @llvm.memset...(i8*, i8, i[32|64], i32, i1)
4334     //    -> @llvm.memset...(i8*, i8, i[32|64], i1)
4335     // Note: i8*'s in the above can be any pointer type
4336     if (CI->arg_size() != 5) {
4337       DefaultCase();
4338       return;
4339     }
4340     // Remove alignment argument (3), and add alignment attributes to the
4341     // dest/src pointers.
4342     Value *Args[4] = {CI->getArgOperand(0), CI->getArgOperand(1),
4343                       CI->getArgOperand(2), CI->getArgOperand(4)};
4344     NewCall = Builder.CreateCall(NewFn, Args);
4345     AttributeList OldAttrs = CI->getAttributes();
4346     AttributeList NewAttrs = AttributeList::get(
4347         C, OldAttrs.getFnAttrs(), OldAttrs.getRetAttrs(),
4348         {OldAttrs.getParamAttrs(0), OldAttrs.getParamAttrs(1),
4349          OldAttrs.getParamAttrs(2), OldAttrs.getParamAttrs(4)});
4350     NewCall->setAttributes(NewAttrs);
4351     auto *MemCI = cast<MemIntrinsic>(NewCall);
4352     // All mem intrinsics support dest alignment.
4353     const ConstantInt *Align = cast<ConstantInt>(CI->getArgOperand(3));
4354     MemCI->setDestAlignment(Align->getMaybeAlignValue());
4355     // Memcpy/Memmove also support source alignment.
4356     if (auto *MTI = dyn_cast<MemTransferInst>(MemCI))
4357       MTI->setSourceAlignment(Align->getMaybeAlignValue());
4358     break;
4359   }
4360   }
4361   assert(NewCall && "Should have either set this variable or returned through "
4362                     "the default case");
4363   NewCall->takeName(CI);
4364   CI->replaceAllUsesWith(NewCall);
4365   CI->eraseFromParent();
4366 }
4367 
UpgradeCallsToIntrinsic(Function * F)4368 void llvm::UpgradeCallsToIntrinsic(Function *F) {
4369   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
4370 
4371   // Check if this function should be upgraded and get the replacement function
4372   // if there is one.
4373   Function *NewFn;
4374   if (UpgradeIntrinsicFunction(F, NewFn)) {
4375     // Replace all users of the old function with the new function or new
4376     // instructions. This is not a range loop because the call is deleted.
4377     for (User *U : make_early_inc_range(F->users()))
4378       if (CallBase *CB = dyn_cast<CallBase>(U))
4379         UpgradeIntrinsicCall(CB, NewFn);
4380 
4381     // Remove old function, no longer used, from the module.
4382     F->eraseFromParent();
4383   }
4384 }
4385 
UpgradeTBAANode(MDNode & MD)4386 MDNode *llvm::UpgradeTBAANode(MDNode &MD) {
4387   // Check if the tag uses struct-path aware TBAA format.
4388   if (isa<MDNode>(MD.getOperand(0)) && MD.getNumOperands() >= 3)
4389     return &MD;
4390 
4391   auto &Context = MD.getContext();
4392   if (MD.getNumOperands() == 3) {
4393     Metadata *Elts[] = {MD.getOperand(0), MD.getOperand(1)};
4394     MDNode *ScalarType = MDNode::get(Context, Elts);
4395     // Create a MDNode <ScalarType, ScalarType, offset 0, const>
4396     Metadata *Elts2[] = {ScalarType, ScalarType,
4397                          ConstantAsMetadata::get(
4398                              Constant::getNullValue(Type::getInt64Ty(Context))),
4399                          MD.getOperand(2)};
4400     return MDNode::get(Context, Elts2);
4401   }
4402   // Create a MDNode <MD, MD, offset 0>
4403   Metadata *Elts[] = {&MD, &MD, ConstantAsMetadata::get(Constant::getNullValue(
4404                                     Type::getInt64Ty(Context)))};
4405   return MDNode::get(Context, Elts);
4406 }
4407 
UpgradeBitCastInst(unsigned Opc,Value * V,Type * DestTy,Instruction * & Temp)4408 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
4409                                       Instruction *&Temp) {
4410   if (Opc != Instruction::BitCast)
4411     return nullptr;
4412 
4413   Temp = nullptr;
4414   Type *SrcTy = V->getType();
4415   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
4416       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
4417     LLVMContext &Context = V->getContext();
4418 
4419     // We have no information about target data layout, so we assume that
4420     // the maximum pointer size is 64bit.
4421     Type *MidTy = Type::getInt64Ty(Context);
4422     Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
4423 
4424     return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
4425   }
4426 
4427   return nullptr;
4428 }
4429 
UpgradeBitCastExpr(unsigned Opc,Constant * C,Type * DestTy)4430 Constant *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
4431   if (Opc != Instruction::BitCast)
4432     return nullptr;
4433 
4434   Type *SrcTy = C->getType();
4435   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
4436       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
4437     LLVMContext &Context = C->getContext();
4438 
4439     // We have no information about target data layout, so we assume that
4440     // the maximum pointer size is 64bit.
4441     Type *MidTy = Type::getInt64Ty(Context);
4442 
4443     return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
4444                                      DestTy);
4445   }
4446 
4447   return nullptr;
4448 }
4449 
4450 /// Check the debug info version number, if it is out-dated, drop the debug
4451 /// info. Return true if module is modified.
UpgradeDebugInfo(Module & M)4452 bool llvm::UpgradeDebugInfo(Module &M) {
4453   unsigned Version = getDebugMetadataVersionFromModule(M);
4454   if (Version == DEBUG_METADATA_VERSION) {
4455     bool BrokenDebugInfo = false;
4456     if (verifyModule(M, &llvm::errs(), &BrokenDebugInfo))
4457       report_fatal_error("Broken module found, compilation aborted!");
4458     if (!BrokenDebugInfo)
4459       // Everything is ok.
4460       return false;
4461     else {
4462       // Diagnose malformed debug info.
4463       DiagnosticInfoIgnoringInvalidDebugMetadata Diag(M);
4464       M.getContext().diagnose(Diag);
4465     }
4466   }
4467   bool Modified = StripDebugInfo(M);
4468   if (Modified && Version != DEBUG_METADATA_VERSION) {
4469     // Diagnose a version mismatch.
4470     DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
4471     M.getContext().diagnose(DiagVersion);
4472   }
4473   return Modified;
4474 }
4475 
4476 /// This checks for objc retain release marker which should be upgraded. It
4477 /// returns true if module is modified.
UpgradeRetainReleaseMarker(Module & M)4478 static bool UpgradeRetainReleaseMarker(Module &M) {
4479   bool Changed = false;
4480   const char *MarkerKey = "clang.arc.retainAutoreleasedReturnValueMarker";
4481   NamedMDNode *ModRetainReleaseMarker = M.getNamedMetadata(MarkerKey);
4482   if (ModRetainReleaseMarker) {
4483     MDNode *Op = ModRetainReleaseMarker->getOperand(0);
4484     if (Op) {
4485       MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(0));
4486       if (ID) {
4487         SmallVector<StringRef, 4> ValueComp;
4488         ID->getString().split(ValueComp, "#");
4489         if (ValueComp.size() == 2) {
4490           std::string NewValue = ValueComp[0].str() + ";" + ValueComp[1].str();
4491           ID = MDString::get(M.getContext(), NewValue);
4492         }
4493         M.addModuleFlag(Module::Error, MarkerKey, ID);
4494         M.eraseNamedMetadata(ModRetainReleaseMarker);
4495         Changed = true;
4496       }
4497     }
4498   }
4499   return Changed;
4500 }
4501 
UpgradeARCRuntime(Module & M)4502 void llvm::UpgradeARCRuntime(Module &M) {
4503   // This lambda converts normal function calls to ARC runtime functions to
4504   // intrinsic calls.
4505   auto UpgradeToIntrinsic = [&](const char *OldFunc,
4506                                 llvm::Intrinsic::ID IntrinsicFunc) {
4507     Function *Fn = M.getFunction(OldFunc);
4508 
4509     if (!Fn)
4510       return;
4511 
4512     Function *NewFn = llvm::Intrinsic::getDeclaration(&M, IntrinsicFunc);
4513 
4514     for (User *U : make_early_inc_range(Fn->users())) {
4515       CallInst *CI = dyn_cast<CallInst>(U);
4516       if (!CI || CI->getCalledFunction() != Fn)
4517         continue;
4518 
4519       IRBuilder<> Builder(CI->getParent(), CI->getIterator());
4520       FunctionType *NewFuncTy = NewFn->getFunctionType();
4521       SmallVector<Value *, 2> Args;
4522 
4523       // Don't upgrade the intrinsic if it's not valid to bitcast the return
4524       // value to the return type of the old function.
4525       if (NewFuncTy->getReturnType() != CI->getType() &&
4526           !CastInst::castIsValid(Instruction::BitCast, CI,
4527                                  NewFuncTy->getReturnType()))
4528         continue;
4529 
4530       bool InvalidCast = false;
4531 
4532       for (unsigned I = 0, E = CI->arg_size(); I != E; ++I) {
4533         Value *Arg = CI->getArgOperand(I);
4534 
4535         // Bitcast argument to the parameter type of the new function if it's
4536         // not a variadic argument.
4537         if (I < NewFuncTy->getNumParams()) {
4538           // Don't upgrade the intrinsic if it's not valid to bitcast the argument
4539           // to the parameter type of the new function.
4540           if (!CastInst::castIsValid(Instruction::BitCast, Arg,
4541                                      NewFuncTy->getParamType(I))) {
4542             InvalidCast = true;
4543             break;
4544           }
4545           Arg = Builder.CreateBitCast(Arg, NewFuncTy->getParamType(I));
4546         }
4547         Args.push_back(Arg);
4548       }
4549 
4550       if (InvalidCast)
4551         continue;
4552 
4553       // Create a call instruction that calls the new function.
4554       CallInst *NewCall = Builder.CreateCall(NewFuncTy, NewFn, Args);
4555       NewCall->setTailCallKind(cast<CallInst>(CI)->getTailCallKind());
4556       NewCall->takeName(CI);
4557 
4558       // Bitcast the return value back to the type of the old call.
4559       Value *NewRetVal = Builder.CreateBitCast(NewCall, CI->getType());
4560 
4561       if (!CI->use_empty())
4562         CI->replaceAllUsesWith(NewRetVal);
4563       CI->eraseFromParent();
4564     }
4565 
4566     if (Fn->use_empty())
4567       Fn->eraseFromParent();
4568   };
4569 
4570   // Unconditionally convert a call to "clang.arc.use" to a call to
4571   // "llvm.objc.clang.arc.use".
4572   UpgradeToIntrinsic("clang.arc.use", llvm::Intrinsic::objc_clang_arc_use);
4573 
4574   // Upgrade the retain release marker. If there is no need to upgrade
4575   // the marker, that means either the module is already new enough to contain
4576   // new intrinsics or it is not ARC. There is no need to upgrade runtime call.
4577   if (!UpgradeRetainReleaseMarker(M))
4578     return;
4579 
4580   std::pair<const char *, llvm::Intrinsic::ID> RuntimeFuncs[] = {
4581       {"objc_autorelease", llvm::Intrinsic::objc_autorelease},
4582       {"objc_autoreleasePoolPop", llvm::Intrinsic::objc_autoreleasePoolPop},
4583       {"objc_autoreleasePoolPush", llvm::Intrinsic::objc_autoreleasePoolPush},
4584       {"objc_autoreleaseReturnValue",
4585        llvm::Intrinsic::objc_autoreleaseReturnValue},
4586       {"objc_copyWeak", llvm::Intrinsic::objc_copyWeak},
4587       {"objc_destroyWeak", llvm::Intrinsic::objc_destroyWeak},
4588       {"objc_initWeak", llvm::Intrinsic::objc_initWeak},
4589       {"objc_loadWeak", llvm::Intrinsic::objc_loadWeak},
4590       {"objc_loadWeakRetained", llvm::Intrinsic::objc_loadWeakRetained},
4591       {"objc_moveWeak", llvm::Intrinsic::objc_moveWeak},
4592       {"objc_release", llvm::Intrinsic::objc_release},
4593       {"objc_retain", llvm::Intrinsic::objc_retain},
4594       {"objc_retainAutorelease", llvm::Intrinsic::objc_retainAutorelease},
4595       {"objc_retainAutoreleaseReturnValue",
4596        llvm::Intrinsic::objc_retainAutoreleaseReturnValue},
4597       {"objc_retainAutoreleasedReturnValue",
4598        llvm::Intrinsic::objc_retainAutoreleasedReturnValue},
4599       {"objc_retainBlock", llvm::Intrinsic::objc_retainBlock},
4600       {"objc_storeStrong", llvm::Intrinsic::objc_storeStrong},
4601       {"objc_storeWeak", llvm::Intrinsic::objc_storeWeak},
4602       {"objc_unsafeClaimAutoreleasedReturnValue",
4603        llvm::Intrinsic::objc_unsafeClaimAutoreleasedReturnValue},
4604       {"objc_retainedObject", llvm::Intrinsic::objc_retainedObject},
4605       {"objc_unretainedObject", llvm::Intrinsic::objc_unretainedObject},
4606       {"objc_unretainedPointer", llvm::Intrinsic::objc_unretainedPointer},
4607       {"objc_retain_autorelease", llvm::Intrinsic::objc_retain_autorelease},
4608       {"objc_sync_enter", llvm::Intrinsic::objc_sync_enter},
4609       {"objc_sync_exit", llvm::Intrinsic::objc_sync_exit},
4610       {"objc_arc_annotation_topdown_bbstart",
4611        llvm::Intrinsic::objc_arc_annotation_topdown_bbstart},
4612       {"objc_arc_annotation_topdown_bbend",
4613        llvm::Intrinsic::objc_arc_annotation_topdown_bbend},
4614       {"objc_arc_annotation_bottomup_bbstart",
4615        llvm::Intrinsic::objc_arc_annotation_bottomup_bbstart},
4616       {"objc_arc_annotation_bottomup_bbend",
4617        llvm::Intrinsic::objc_arc_annotation_bottomup_bbend}};
4618 
4619   for (auto &I : RuntimeFuncs)
4620     UpgradeToIntrinsic(I.first, I.second);
4621 }
4622 
UpgradeModuleFlags(Module & M)4623 bool llvm::UpgradeModuleFlags(Module &M) {
4624   NamedMDNode *ModFlags = M.getModuleFlagsMetadata();
4625   if (!ModFlags)
4626     return false;
4627 
4628   bool HasObjCFlag = false, HasClassProperties = false, Changed = false;
4629   bool HasSwiftVersionFlag = false;
4630   uint8_t SwiftMajorVersion, SwiftMinorVersion;
4631   uint32_t SwiftABIVersion;
4632   auto Int8Ty = Type::getInt8Ty(M.getContext());
4633   auto Int32Ty = Type::getInt32Ty(M.getContext());
4634 
4635   for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) {
4636     MDNode *Op = ModFlags->getOperand(I);
4637     if (Op->getNumOperands() != 3)
4638       continue;
4639     MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(1));
4640     if (!ID)
4641       continue;
4642     auto SetBehavior = [&](Module::ModFlagBehavior B) {
4643       Metadata *Ops[3] = {ConstantAsMetadata::get(ConstantInt::get(
4644                               Type::getInt32Ty(M.getContext()), B)),
4645                           MDString::get(M.getContext(), ID->getString()),
4646                           Op->getOperand(2)};
4647       ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
4648       Changed = true;
4649     };
4650 
4651     if (ID->getString() == "Objective-C Image Info Version")
4652       HasObjCFlag = true;
4653     if (ID->getString() == "Objective-C Class Properties")
4654       HasClassProperties = true;
4655     // Upgrade PIC from Error/Max to Min.
4656     if (ID->getString() == "PIC Level") {
4657       if (auto *Behavior =
4658               mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0))) {
4659         uint64_t V = Behavior->getLimitedValue();
4660         if (V == Module::Error || V == Module::Max)
4661           SetBehavior(Module::Min);
4662       }
4663     }
4664     // Upgrade "PIE Level" from Error to Max.
4665     if (ID->getString() == "PIE Level")
4666       if (auto *Behavior =
4667               mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0)))
4668         if (Behavior->getLimitedValue() == Module::Error)
4669           SetBehavior(Module::Max);
4670 
4671     // Upgrade branch protection and return address signing module flags. The
4672     // module flag behavior for these fields were Error and now they are Min.
4673     if (ID->getString() == "branch-target-enforcement" ||
4674         ID->getString().startswith("sign-return-address")) {
4675       if (auto *Behavior =
4676               mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0))) {
4677         if (Behavior->getLimitedValue() == Module::Error) {
4678           Type *Int32Ty = Type::getInt32Ty(M.getContext());
4679           Metadata *Ops[3] = {
4680               ConstantAsMetadata::get(ConstantInt::get(Int32Ty, Module::Min)),
4681               Op->getOperand(1), Op->getOperand(2)};
4682           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
4683           Changed = true;
4684         }
4685       }
4686     }
4687 
4688     // Upgrade Objective-C Image Info Section. Removed the whitespce in the
4689     // section name so that llvm-lto will not complain about mismatching
4690     // module flags that is functionally the same.
4691     if (ID->getString() == "Objective-C Image Info Section") {
4692       if (auto *Value = dyn_cast_or_null<MDString>(Op->getOperand(2))) {
4693         SmallVector<StringRef, 4> ValueComp;
4694         Value->getString().split(ValueComp, " ");
4695         if (ValueComp.size() != 1) {
4696           std::string NewValue;
4697           for (auto &S : ValueComp)
4698             NewValue += S.str();
4699           Metadata *Ops[3] = {Op->getOperand(0), Op->getOperand(1),
4700                               MDString::get(M.getContext(), NewValue)};
4701           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
4702           Changed = true;
4703         }
4704       }
4705     }
4706 
4707     // IRUpgrader turns a i32 type "Objective-C Garbage Collection" into i8 value.
4708     // If the higher bits are set, it adds new module flag for swift info.
4709     if (ID->getString() == "Objective-C Garbage Collection") {
4710       auto Md = dyn_cast<ConstantAsMetadata>(Op->getOperand(2));
4711       if (Md) {
4712         assert(Md->getValue() && "Expected non-empty metadata");
4713         auto Type = Md->getValue()->getType();
4714         if (Type == Int8Ty)
4715           continue;
4716         unsigned Val = Md->getValue()->getUniqueInteger().getZExtValue();
4717         if ((Val & 0xff) != Val) {
4718           HasSwiftVersionFlag = true;
4719           SwiftABIVersion = (Val & 0xff00) >> 8;
4720           SwiftMajorVersion = (Val & 0xff000000) >> 24;
4721           SwiftMinorVersion = (Val & 0xff0000) >> 16;
4722         }
4723         Metadata *Ops[3] = {
4724           ConstantAsMetadata::get(ConstantInt::get(Int32Ty,Module::Error)),
4725           Op->getOperand(1),
4726           ConstantAsMetadata::get(ConstantInt::get(Int8Ty,Val & 0xff))};
4727         ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
4728         Changed = true;
4729       }
4730     }
4731   }
4732 
4733   // "Objective-C Class Properties" is recently added for Objective-C. We
4734   // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module
4735   // flag of value 0, so we can correclty downgrade this flag when trying to
4736   // link an ObjC bitcode without this module flag with an ObjC bitcode with
4737   // this module flag.
4738   if (HasObjCFlag && !HasClassProperties) {
4739     M.addModuleFlag(llvm::Module::Override, "Objective-C Class Properties",
4740                     (uint32_t)0);
4741     Changed = true;
4742   }
4743 
4744   if (HasSwiftVersionFlag) {
4745     M.addModuleFlag(Module::Error, "Swift ABI Version",
4746                     SwiftABIVersion);
4747     M.addModuleFlag(Module::Error, "Swift Major Version",
4748                     ConstantInt::get(Int8Ty, SwiftMajorVersion));
4749     M.addModuleFlag(Module::Error, "Swift Minor Version",
4750                     ConstantInt::get(Int8Ty, SwiftMinorVersion));
4751     Changed = true;
4752   }
4753 
4754   return Changed;
4755 }
4756 
UpgradeSectionAttributes(Module & M)4757 void llvm::UpgradeSectionAttributes(Module &M) {
4758   auto TrimSpaces = [](StringRef Section) -> std::string {
4759     SmallVector<StringRef, 5> Components;
4760     Section.split(Components, ',');
4761 
4762     SmallString<32> Buffer;
4763     raw_svector_ostream OS(Buffer);
4764 
4765     for (auto Component : Components)
4766       OS << ',' << Component.trim();
4767 
4768     return std::string(OS.str().substr(1));
4769   };
4770 
4771   for (auto &GV : M.globals()) {
4772     if (!GV.hasSection())
4773       continue;
4774 
4775     StringRef Section = GV.getSection();
4776 
4777     if (!Section.startswith("__DATA, __objc_catlist"))
4778       continue;
4779 
4780     // __DATA, __objc_catlist, regular, no_dead_strip
4781     // __DATA,__objc_catlist,regular,no_dead_strip
4782     GV.setSection(TrimSpaces(Section));
4783   }
4784 }
4785 
4786 namespace {
4787 // Prior to LLVM 10.0, the strictfp attribute could be used on individual
4788 // callsites within a function that did not also have the strictfp attribute.
4789 // Since 10.0, if strict FP semantics are needed within a function, the
4790 // function must have the strictfp attribute and all calls within the function
4791 // must also have the strictfp attribute. This latter restriction is
4792 // necessary to prevent unwanted libcall simplification when a function is
4793 // being cloned (such as for inlining).
4794 //
4795 // The "dangling" strictfp attribute usage was only used to prevent constant
4796 // folding and other libcall simplification. The nobuiltin attribute on the
4797 // callsite has the same effect.
4798 struct StrictFPUpgradeVisitor : public InstVisitor<StrictFPUpgradeVisitor> {
4799   StrictFPUpgradeVisitor() = default;
4800 
visitCallBase__anon0722971b0511::StrictFPUpgradeVisitor4801   void visitCallBase(CallBase &Call) {
4802     if (!Call.isStrictFP())
4803       return;
4804     if (isa<ConstrainedFPIntrinsic>(&Call))
4805       return;
4806     // If we get here, the caller doesn't have the strictfp attribute
4807     // but this callsite does. Replace the strictfp attribute with nobuiltin.
4808     Call.removeFnAttr(Attribute::StrictFP);
4809     Call.addFnAttr(Attribute::NoBuiltin);
4810   }
4811 };
4812 } // namespace
4813 
UpgradeFunctionAttributes(Function & F)4814 void llvm::UpgradeFunctionAttributes(Function &F) {
4815   // If a function definition doesn't have the strictfp attribute,
4816   // convert any callsite strictfp attributes to nobuiltin.
4817   if (!F.isDeclaration() && !F.hasFnAttribute(Attribute::StrictFP)) {
4818     StrictFPUpgradeVisitor SFPV;
4819     SFPV.visit(F);
4820   }
4821 
4822   // Remove all incompatibile attributes from function.
4823   F.removeRetAttrs(AttributeFuncs::typeIncompatible(F.getReturnType()));
4824   for (auto &Arg : F.args())
4825     Arg.removeAttrs(AttributeFuncs::typeIncompatible(Arg.getType()));
4826 }
4827 
isOldLoopArgument(Metadata * MD)4828 static bool isOldLoopArgument(Metadata *MD) {
4829   auto *T = dyn_cast_or_null<MDTuple>(MD);
4830   if (!T)
4831     return false;
4832   if (T->getNumOperands() < 1)
4833     return false;
4834   auto *S = dyn_cast_or_null<MDString>(T->getOperand(0));
4835   if (!S)
4836     return false;
4837   return S->getString().startswith("llvm.vectorizer.");
4838 }
4839 
upgradeLoopTag(LLVMContext & C,StringRef OldTag)4840 static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) {
4841   StringRef OldPrefix = "llvm.vectorizer.";
4842   assert(OldTag.startswith(OldPrefix) && "Expected old prefix");
4843 
4844   if (OldTag == "llvm.vectorizer.unroll")
4845     return MDString::get(C, "llvm.loop.interleave.count");
4846 
4847   return MDString::get(
4848       C, (Twine("llvm.loop.vectorize.") + OldTag.drop_front(OldPrefix.size()))
4849              .str());
4850 }
4851 
upgradeLoopArgument(Metadata * MD)4852 static Metadata *upgradeLoopArgument(Metadata *MD) {
4853   auto *T = dyn_cast_or_null<MDTuple>(MD);
4854   if (!T)
4855     return MD;
4856   if (T->getNumOperands() < 1)
4857     return MD;
4858   auto *OldTag = dyn_cast_or_null<MDString>(T->getOperand(0));
4859   if (!OldTag)
4860     return MD;
4861   if (!OldTag->getString().startswith("llvm.vectorizer."))
4862     return MD;
4863 
4864   // This has an old tag.  Upgrade it.
4865   SmallVector<Metadata *, 8> Ops;
4866   Ops.reserve(T->getNumOperands());
4867   Ops.push_back(upgradeLoopTag(T->getContext(), OldTag->getString()));
4868   for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I)
4869     Ops.push_back(T->getOperand(I));
4870 
4871   return MDTuple::get(T->getContext(), Ops);
4872 }
4873 
upgradeInstructionLoopAttachment(MDNode & N)4874 MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) {
4875   auto *T = dyn_cast<MDTuple>(&N);
4876   if (!T)
4877     return &N;
4878 
4879   if (none_of(T->operands(), isOldLoopArgument))
4880     return &N;
4881 
4882   SmallVector<Metadata *, 8> Ops;
4883   Ops.reserve(T->getNumOperands());
4884   for (Metadata *MD : T->operands())
4885     Ops.push_back(upgradeLoopArgument(MD));
4886 
4887   return MDTuple::get(T->getContext(), Ops);
4888 }
4889 
UpgradeDataLayoutString(StringRef DL,StringRef TT)4890 std::string llvm::UpgradeDataLayoutString(StringRef DL, StringRef TT) {
4891   Triple T(TT);
4892   // For AMDGPU we uprgrade older DataLayouts to include the default globals
4893   // address space of 1.
4894   if (T.isAMDGPU() && !DL.contains("-G") && !DL.startswith("G")) {
4895     return DL.empty() ? std::string("G1") : (DL + "-G1").str();
4896   }
4897 
4898   if (T.isRISCV64()) {
4899     // Make i32 a native type for 64-bit RISC-V.
4900     auto I = DL.find("-n64-");
4901     if (I != StringRef::npos)
4902       return (DL.take_front(I) + "-n32:64-" + DL.drop_front(I + 5)).str();
4903     return DL.str();
4904   }
4905 
4906   std::string Res = DL.str();
4907   if (!T.isX86())
4908     return Res;
4909 
4910   // If the datalayout matches the expected format, add pointer size address
4911   // spaces to the datalayout.
4912   std::string AddrSpaces = "-p270:32:32-p271:32:32-p272:64:64";
4913   if (!DL.contains(AddrSpaces)) {
4914     SmallVector<StringRef, 4> Groups;
4915     Regex R("(e-m:[a-z](-p:32:32)?)(-[if]64:.*$)");
4916     if (R.match(DL, &Groups))
4917       Res = (Groups[1] + AddrSpaces + Groups[3]).str();
4918   }
4919 
4920   // For 32-bit MSVC targets, raise the alignment of f80 values to 16 bytes.
4921   // Raising the alignment is safe because Clang did not produce f80 values in
4922   // the MSVC environment before this upgrade was added.
4923   if (T.isWindowsMSVCEnvironment() && !T.isArch64Bit()) {
4924     StringRef Ref = Res;
4925     auto I = Ref.find("-f80:32-");
4926     if (I != StringRef::npos)
4927       Res = (Ref.take_front(I) + "-f80:128-" + Ref.drop_front(I + 8)).str();
4928   }
4929 
4930   return Res;
4931 }
4932 
UpgradeAttributes(AttrBuilder & B)4933 void llvm::UpgradeAttributes(AttrBuilder &B) {
4934   StringRef FramePointer;
4935   Attribute A = B.getAttribute("no-frame-pointer-elim");
4936   if (A.isValid()) {
4937     // The value can be "true" or "false".
4938     FramePointer = A.getValueAsString() == "true" ? "all" : "none";
4939     B.removeAttribute("no-frame-pointer-elim");
4940   }
4941   if (B.contains("no-frame-pointer-elim-non-leaf")) {
4942     // The value is ignored. "no-frame-pointer-elim"="true" takes priority.
4943     if (FramePointer != "all")
4944       FramePointer = "non-leaf";
4945     B.removeAttribute("no-frame-pointer-elim-non-leaf");
4946   }
4947   if (!FramePointer.empty())
4948     B.addAttribute("frame-pointer", FramePointer);
4949 
4950   A = B.getAttribute("null-pointer-is-valid");
4951   if (A.isValid()) {
4952     // The value can be "true" or "false".
4953     bool NullPointerIsValid = A.getValueAsString() == "true";
4954     B.removeAttribute("null-pointer-is-valid");
4955     if (NullPointerIsValid)
4956       B.addAttribute(Attribute::NullPointerIsValid);
4957   }
4958 }
4959 
UpgradeOperandBundles(std::vector<OperandBundleDef> & Bundles)4960 void llvm::UpgradeOperandBundles(std::vector<OperandBundleDef> &Bundles) {
4961 
4962   // clang.arc.attachedcall bundles are now required to have an operand.
4963   // If they don't, it's okay to drop them entirely: when there is an operand,
4964   // the "attachedcall" is meaningful and required, but without an operand,
4965   // it's just a marker NOP.  Dropping it merely prevents an optimization.
4966   erase_if(Bundles, [&](OperandBundleDef &OBD) {
4967     return OBD.getTag() == "clang.arc.attachedcall" &&
4968            OBD.inputs().empty();
4969   });
4970 }
4971