[InstCombine] Signed saturation tests. NFC
[llvm-complete.git] / lib / IR / AutoUpgrade.cpp
blob79f580d0e14d964b841a45453aed16e7ba12db99
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.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/IR/AutoUpgrade.h"
16 #include "llvm/ADT/StringSwitch.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/DIBuilder.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/Instruction.h"
24 #include "llvm/IR/IntrinsicInst.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/IR/Verifier.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/Regex.h"
30 #include <cstring>
31 using namespace llvm;
33 static void rename(GlobalValue *GV) { GV->setName(GV->getName() + ".old"); }
35 // Upgrade the declarations of the SSE4.1 ptest intrinsics whose arguments have
36 // changed their type from v4f32 to v2i64.
37 static bool UpgradePTESTIntrinsic(Function* F, Intrinsic::ID IID,
38 Function *&NewFn) {
39 // Check whether this is an old version of the function, which received
40 // v4f32 arguments.
41 Type *Arg0Type = F->getFunctionType()->getParamType(0);
42 if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
43 return false;
45 // Yes, it's old, replace it with new version.
46 rename(F);
47 NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
48 return true;
51 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
52 // arguments have changed their type from i32 to i8.
53 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
54 Function *&NewFn) {
55 // Check that the last argument is an i32.
56 Type *LastArgType = F->getFunctionType()->getParamType(
57 F->getFunctionType()->getNumParams() - 1);
58 if (!LastArgType->isIntegerTy(32))
59 return false;
61 // Move this function aside and map down.
62 rename(F);
63 NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
64 return true;
67 static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
68 // All of the intrinsics matches below should be marked with which llvm
69 // version started autoupgrading them. At some point in the future we would
70 // like to use this information to remove upgrade code for some older
71 // intrinsics. It is currently undecided how we will determine that future
72 // point.
73 if (Name == "addcarryx.u32" || // Added in 8.0
74 Name == "addcarryx.u64" || // Added in 8.0
75 Name == "addcarry.u32" || // Added in 8.0
76 Name == "addcarry.u64" || // Added in 8.0
77 Name == "subborrow.u32" || // Added in 8.0
78 Name == "subborrow.u64" || // Added in 8.0
79 Name.startswith("sse2.padds.") || // Added in 8.0
80 Name.startswith("sse2.psubs.") || // Added in 8.0
81 Name.startswith("sse2.paddus.") || // Added in 8.0
82 Name.startswith("sse2.psubus.") || // Added in 8.0
83 Name.startswith("avx2.padds.") || // Added in 8.0
84 Name.startswith("avx2.psubs.") || // Added in 8.0
85 Name.startswith("avx2.paddus.") || // Added in 8.0
86 Name.startswith("avx2.psubus.") || // Added in 8.0
87 Name.startswith("avx512.padds.") || // Added in 8.0
88 Name.startswith("avx512.psubs.") || // Added in 8.0
89 Name.startswith("avx512.mask.padds.") || // Added in 8.0
90 Name.startswith("avx512.mask.psubs.") || // Added in 8.0
91 Name.startswith("avx512.mask.paddus.") || // Added in 8.0
92 Name.startswith("avx512.mask.psubus.") || // Added in 8.0
93 Name=="ssse3.pabs.b.128" || // Added in 6.0
94 Name=="ssse3.pabs.w.128" || // Added in 6.0
95 Name=="ssse3.pabs.d.128" || // Added in 6.0
96 Name.startswith("fma4.vfmadd.s") || // Added in 7.0
97 Name.startswith("fma.vfmadd.") || // Added in 7.0
98 Name.startswith("fma.vfmsub.") || // Added in 7.0
99 Name.startswith("fma.vfmaddsub.") || // Added in 7.0
100 Name.startswith("fma.vfmsubadd.") || // Added in 7.0
101 Name.startswith("fma.vfnmadd.") || // Added in 7.0
102 Name.startswith("fma.vfnmsub.") || // Added in 7.0
103 Name.startswith("avx512.mask.vfmadd.") || // Added in 7.0
104 Name.startswith("avx512.mask.vfnmadd.") || // Added in 7.0
105 Name.startswith("avx512.mask.vfnmsub.") || // Added in 7.0
106 Name.startswith("avx512.mask3.vfmadd.") || // Added in 7.0
107 Name.startswith("avx512.maskz.vfmadd.") || // Added in 7.0
108 Name.startswith("avx512.mask3.vfmsub.") || // Added in 7.0
109 Name.startswith("avx512.mask3.vfnmsub.") || // Added in 7.0
110 Name.startswith("avx512.mask.vfmaddsub.") || // Added in 7.0
111 Name.startswith("avx512.maskz.vfmaddsub.") || // Added in 7.0
112 Name.startswith("avx512.mask3.vfmaddsub.") || // Added in 7.0
113 Name.startswith("avx512.mask3.vfmsubadd.") || // Added in 7.0
114 Name.startswith("avx512.mask.shuf.i") || // Added in 6.0
115 Name.startswith("avx512.mask.shuf.f") || // Added in 6.0
116 Name.startswith("avx512.kunpck") || //added in 6.0
117 Name.startswith("avx2.pabs.") || // Added in 6.0
118 Name.startswith("avx512.mask.pabs.") || // Added in 6.0
119 Name.startswith("avx512.broadcastm") || // Added in 6.0
120 Name == "sse.sqrt.ss" || // Added in 7.0
121 Name == "sse2.sqrt.sd" || // Added in 7.0
122 Name.startswith("avx512.mask.sqrt.p") || // Added in 7.0
123 Name.startswith("avx.sqrt.p") || // Added in 7.0
124 Name.startswith("sse2.sqrt.p") || // Added in 7.0
125 Name.startswith("sse.sqrt.p") || // Added in 7.0
126 Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
127 Name.startswith("sse2.pcmpeq.") || // Added in 3.1
128 Name.startswith("sse2.pcmpgt.") || // Added in 3.1
129 Name.startswith("avx2.pcmpeq.") || // Added in 3.1
130 Name.startswith("avx2.pcmpgt.") || // Added in 3.1
131 Name.startswith("avx512.mask.pcmpeq.") || // Added in 3.9
132 Name.startswith("avx512.mask.pcmpgt.") || // Added in 3.9
133 Name.startswith("avx.vperm2f128.") || // Added in 6.0
134 Name == "avx2.vperm2i128" || // Added in 6.0
135 Name == "sse.add.ss" || // Added in 4.0
136 Name == "sse2.add.sd" || // Added in 4.0
137 Name == "sse.sub.ss" || // Added in 4.0
138 Name == "sse2.sub.sd" || // Added in 4.0
139 Name == "sse.mul.ss" || // Added in 4.0
140 Name == "sse2.mul.sd" || // Added in 4.0
141 Name == "sse.div.ss" || // Added in 4.0
142 Name == "sse2.div.sd" || // Added in 4.0
143 Name == "sse41.pmaxsb" || // Added in 3.9
144 Name == "sse2.pmaxs.w" || // Added in 3.9
145 Name == "sse41.pmaxsd" || // Added in 3.9
146 Name == "sse2.pmaxu.b" || // Added in 3.9
147 Name == "sse41.pmaxuw" || // Added in 3.9
148 Name == "sse41.pmaxud" || // Added in 3.9
149 Name == "sse41.pminsb" || // Added in 3.9
150 Name == "sse2.pmins.w" || // Added in 3.9
151 Name == "sse41.pminsd" || // Added in 3.9
152 Name == "sse2.pminu.b" || // Added in 3.9
153 Name == "sse41.pminuw" || // Added in 3.9
154 Name == "sse41.pminud" || // Added in 3.9
155 Name == "avx512.kand.w" || // Added in 7.0
156 Name == "avx512.kandn.w" || // Added in 7.0
157 Name == "avx512.knot.w" || // Added in 7.0
158 Name == "avx512.kor.w" || // Added in 7.0
159 Name == "avx512.kxor.w" || // Added in 7.0
160 Name == "avx512.kxnor.w" || // Added in 7.0
161 Name == "avx512.kortestc.w" || // Added in 7.0
162 Name == "avx512.kortestz.w" || // Added in 7.0
163 Name.startswith("avx512.mask.pshuf.b.") || // Added in 4.0
164 Name.startswith("avx2.pmax") || // Added in 3.9
165 Name.startswith("avx2.pmin") || // Added in 3.9
166 Name.startswith("avx512.mask.pmax") || // Added in 4.0
167 Name.startswith("avx512.mask.pmin") || // Added in 4.0
168 Name.startswith("avx2.vbroadcast") || // Added in 3.8
169 Name.startswith("avx2.pbroadcast") || // Added in 3.8
170 Name.startswith("avx.vpermil.") || // Added in 3.1
171 Name.startswith("sse2.pshuf") || // Added in 3.9
172 Name.startswith("avx512.pbroadcast") || // Added in 3.9
173 Name.startswith("avx512.mask.broadcast.s") || // Added in 3.9
174 Name.startswith("avx512.mask.movddup") || // Added in 3.9
175 Name.startswith("avx512.mask.movshdup") || // Added in 3.9
176 Name.startswith("avx512.mask.movsldup") || // Added in 3.9
177 Name.startswith("avx512.mask.pshuf.d.") || // Added in 3.9
178 Name.startswith("avx512.mask.pshufl.w.") || // Added in 3.9
179 Name.startswith("avx512.mask.pshufh.w.") || // Added in 3.9
180 Name.startswith("avx512.mask.shuf.p") || // Added in 4.0
181 Name.startswith("avx512.mask.vpermil.p") || // Added in 3.9
182 Name.startswith("avx512.mask.perm.df.") || // Added in 3.9
183 Name.startswith("avx512.mask.perm.di.") || // Added in 3.9
184 Name.startswith("avx512.mask.punpckl") || // Added in 3.9
185 Name.startswith("avx512.mask.punpckh") || // Added in 3.9
186 Name.startswith("avx512.mask.unpckl.") || // Added in 3.9
187 Name.startswith("avx512.mask.unpckh.") || // Added in 3.9
188 Name.startswith("avx512.mask.pand.") || // Added in 3.9
189 Name.startswith("avx512.mask.pandn.") || // Added in 3.9
190 Name.startswith("avx512.mask.por.") || // Added in 3.9
191 Name.startswith("avx512.mask.pxor.") || // Added in 3.9
192 Name.startswith("avx512.mask.and.") || // Added in 3.9
193 Name.startswith("avx512.mask.andn.") || // Added in 3.9
194 Name.startswith("avx512.mask.or.") || // Added in 3.9
195 Name.startswith("avx512.mask.xor.") || // Added in 3.9
196 Name.startswith("avx512.mask.padd.") || // Added in 4.0
197 Name.startswith("avx512.mask.psub.") || // Added in 4.0
198 Name.startswith("avx512.mask.pmull.") || // Added in 4.0
199 Name.startswith("avx512.mask.cvtdq2pd.") || // Added in 4.0
200 Name.startswith("avx512.mask.cvtudq2pd.") || // Added in 4.0
201 Name.startswith("avx512.mask.cvtudq2ps.") || // Added in 7.0 updated 9.0
202 Name.startswith("avx512.mask.cvtqq2pd.") || // Added in 7.0 updated 9.0
203 Name.startswith("avx512.mask.cvtuqq2pd.") || // Added in 7.0 updated 9.0
204 Name.startswith("avx512.mask.cvtdq2ps.") || // Added in 7.0 updated 9.0
205 Name == "avx512.mask.cvtqq2ps.256" || // Added in 9.0
206 Name == "avx512.mask.cvtqq2ps.512" || // Added in 9.0
207 Name == "avx512.mask.cvtuqq2ps.256" || // Added in 9.0
208 Name == "avx512.mask.cvtuqq2ps.512" || // Added in 9.0
209 Name == "avx512.mask.cvtpd2dq.256" || // Added in 7.0
210 Name == "avx512.mask.cvtpd2ps.256" || // Added in 7.0
211 Name == "avx512.mask.cvttpd2dq.256" || // Added in 7.0
212 Name == "avx512.mask.cvttps2dq.128" || // Added in 7.0
213 Name == "avx512.mask.cvttps2dq.256" || // Added in 7.0
214 Name == "avx512.mask.cvtps2pd.128" || // Added in 7.0
215 Name == "avx512.mask.cvtps2pd.256" || // Added in 7.0
216 Name == "avx512.cvtusi2sd" || // Added in 7.0
217 Name.startswith("avx512.mask.permvar.") || // Added in 7.0
218 Name == "sse2.pmulu.dq" || // Added in 7.0
219 Name == "sse41.pmuldq" || // Added in 7.0
220 Name == "avx2.pmulu.dq" || // Added in 7.0
221 Name == "avx2.pmul.dq" || // Added in 7.0
222 Name == "avx512.pmulu.dq.512" || // Added in 7.0
223 Name == "avx512.pmul.dq.512" || // Added in 7.0
224 Name.startswith("avx512.mask.pmul.dq.") || // Added in 4.0
225 Name.startswith("avx512.mask.pmulu.dq.") || // Added in 4.0
226 Name.startswith("avx512.mask.pmul.hr.sw.") || // Added in 7.0
227 Name.startswith("avx512.mask.pmulh.w.") || // Added in 7.0
228 Name.startswith("avx512.mask.pmulhu.w.") || // Added in 7.0
229 Name.startswith("avx512.mask.pmaddw.d.") || // Added in 7.0
230 Name.startswith("avx512.mask.pmaddubs.w.") || // Added in 7.0
231 Name.startswith("avx512.mask.packsswb.") || // Added in 5.0
232 Name.startswith("avx512.mask.packssdw.") || // Added in 5.0
233 Name.startswith("avx512.mask.packuswb.") || // Added in 5.0
234 Name.startswith("avx512.mask.packusdw.") || // Added in 5.0
235 Name.startswith("avx512.mask.cmp.b") || // Added in 5.0
236 Name.startswith("avx512.mask.cmp.d") || // Added in 5.0
237 Name.startswith("avx512.mask.cmp.q") || // Added in 5.0
238 Name.startswith("avx512.mask.cmp.w") || // Added in 5.0
239 Name.startswith("avx512.mask.cmp.p") || // Added in 7.0
240 Name.startswith("avx512.mask.ucmp.") || // Added in 5.0
241 Name.startswith("avx512.cvtb2mask.") || // Added in 7.0
242 Name.startswith("avx512.cvtw2mask.") || // Added in 7.0
243 Name.startswith("avx512.cvtd2mask.") || // Added in 7.0
244 Name.startswith("avx512.cvtq2mask.") || // Added in 7.0
245 Name.startswith("avx512.mask.vpermilvar.") || // Added in 4.0
246 Name.startswith("avx512.mask.psll.d") || // Added in 4.0
247 Name.startswith("avx512.mask.psll.q") || // Added in 4.0
248 Name.startswith("avx512.mask.psll.w") || // Added in 4.0
249 Name.startswith("avx512.mask.psra.d") || // Added in 4.0
250 Name.startswith("avx512.mask.psra.q") || // Added in 4.0
251 Name.startswith("avx512.mask.psra.w") || // Added in 4.0
252 Name.startswith("avx512.mask.psrl.d") || // Added in 4.0
253 Name.startswith("avx512.mask.psrl.q") || // Added in 4.0
254 Name.startswith("avx512.mask.psrl.w") || // Added in 4.0
255 Name.startswith("avx512.mask.pslli") || // Added in 4.0
256 Name.startswith("avx512.mask.psrai") || // Added in 4.0
257 Name.startswith("avx512.mask.psrli") || // Added in 4.0
258 Name.startswith("avx512.mask.psllv") || // Added in 4.0
259 Name.startswith("avx512.mask.psrav") || // Added in 4.0
260 Name.startswith("avx512.mask.psrlv") || // Added in 4.0
261 Name.startswith("sse41.pmovsx") || // Added in 3.8
262 Name.startswith("sse41.pmovzx") || // Added in 3.9
263 Name.startswith("avx2.pmovsx") || // Added in 3.9
264 Name.startswith("avx2.pmovzx") || // Added in 3.9
265 Name.startswith("avx512.mask.pmovsx") || // Added in 4.0
266 Name.startswith("avx512.mask.pmovzx") || // Added in 4.0
267 Name.startswith("avx512.mask.lzcnt.") || // Added in 5.0
268 Name.startswith("avx512.mask.pternlog.") || // Added in 7.0
269 Name.startswith("avx512.maskz.pternlog.") || // Added in 7.0
270 Name.startswith("avx512.mask.vpmadd52") || // Added in 7.0
271 Name.startswith("avx512.maskz.vpmadd52") || // Added in 7.0
272 Name.startswith("avx512.mask.vpermi2var.") || // Added in 7.0
273 Name.startswith("avx512.mask.vpermt2var.") || // Added in 7.0
274 Name.startswith("avx512.maskz.vpermt2var.") || // Added in 7.0
275 Name.startswith("avx512.mask.vpdpbusd.") || // Added in 7.0
276 Name.startswith("avx512.maskz.vpdpbusd.") || // Added in 7.0
277 Name.startswith("avx512.mask.vpdpbusds.") || // Added in 7.0
278 Name.startswith("avx512.maskz.vpdpbusds.") || // Added in 7.0
279 Name.startswith("avx512.mask.vpdpwssd.") || // Added in 7.0
280 Name.startswith("avx512.maskz.vpdpwssd.") || // Added in 7.0
281 Name.startswith("avx512.mask.vpdpwssds.") || // Added in 7.0
282 Name.startswith("avx512.maskz.vpdpwssds.") || // Added in 7.0
283 Name.startswith("avx512.mask.dbpsadbw.") || // Added in 7.0
284 Name.startswith("avx512.mask.vpshld.") || // Added in 7.0
285 Name.startswith("avx512.mask.vpshrd.") || // Added in 7.0
286 Name.startswith("avx512.mask.vpshldv.") || // Added in 8.0
287 Name.startswith("avx512.mask.vpshrdv.") || // Added in 8.0
288 Name.startswith("avx512.maskz.vpshldv.") || // Added in 8.0
289 Name.startswith("avx512.maskz.vpshrdv.") || // Added in 8.0
290 Name.startswith("avx512.vpshld.") || // Added in 8.0
291 Name.startswith("avx512.vpshrd.") || // Added in 8.0
292 Name.startswith("avx512.mask.add.p") || // Added in 7.0. 128/256 in 4.0
293 Name.startswith("avx512.mask.sub.p") || // Added in 7.0. 128/256 in 4.0
294 Name.startswith("avx512.mask.mul.p") || // Added in 7.0. 128/256 in 4.0
295 Name.startswith("avx512.mask.div.p") || // Added in 7.0. 128/256 in 4.0
296 Name.startswith("avx512.mask.max.p") || // Added in 7.0. 128/256 in 5.0
297 Name.startswith("avx512.mask.min.p") || // Added in 7.0. 128/256 in 5.0
298 Name.startswith("avx512.mask.fpclass.p") || // Added in 7.0
299 Name.startswith("avx512.mask.vpshufbitqmb.") || // Added in 8.0
300 Name.startswith("avx512.mask.pmultishift.qb.") || // Added in 8.0
301 Name.startswith("avx512.mask.conflict.") || // Added in 9.0
302 Name == "avx512.mask.pmov.qd.256" || // Added in 9.0
303 Name == "avx512.mask.pmov.qd.512" || // Added in 9.0
304 Name == "avx512.mask.pmov.wb.256" || // Added in 9.0
305 Name == "avx512.mask.pmov.wb.512" || // Added in 9.0
306 Name == "sse.cvtsi2ss" || // Added in 7.0
307 Name == "sse.cvtsi642ss" || // Added in 7.0
308 Name == "sse2.cvtsi2sd" || // Added in 7.0
309 Name == "sse2.cvtsi642sd" || // Added in 7.0
310 Name == "sse2.cvtss2sd" || // Added in 7.0
311 Name == "sse2.cvtdq2pd" || // Added in 3.9
312 Name == "sse2.cvtdq2ps" || // Added in 7.0
313 Name == "sse2.cvtps2pd" || // Added in 3.9
314 Name == "avx.cvtdq2.pd.256" || // Added in 3.9
315 Name == "avx.cvtdq2.ps.256" || // Added in 7.0
316 Name == "avx.cvt.ps2.pd.256" || // Added in 3.9
317 Name.startswith("avx.vinsertf128.") || // Added in 3.7
318 Name == "avx2.vinserti128" || // Added in 3.7
319 Name.startswith("avx512.mask.insert") || // Added in 4.0
320 Name.startswith("avx.vextractf128.") || // Added in 3.7
321 Name == "avx2.vextracti128" || // Added in 3.7
322 Name.startswith("avx512.mask.vextract") || // Added in 4.0
323 Name.startswith("sse4a.movnt.") || // Added in 3.9
324 Name.startswith("avx.movnt.") || // Added in 3.2
325 Name.startswith("avx512.storent.") || // Added in 3.9
326 Name == "sse41.movntdqa" || // Added in 5.0
327 Name == "avx2.movntdqa" || // Added in 5.0
328 Name == "avx512.movntdqa" || // Added in 5.0
329 Name == "sse2.storel.dq" || // Added in 3.9
330 Name.startswith("sse.storeu.") || // Added in 3.9
331 Name.startswith("sse2.storeu.") || // Added in 3.9
332 Name.startswith("avx.storeu.") || // Added in 3.9
333 Name.startswith("avx512.mask.storeu.") || // Added in 3.9
334 Name.startswith("avx512.mask.store.p") || // Added in 3.9
335 Name.startswith("avx512.mask.store.b.") || // Added in 3.9
336 Name.startswith("avx512.mask.store.w.") || // Added in 3.9
337 Name.startswith("avx512.mask.store.d.") || // Added in 3.9
338 Name.startswith("avx512.mask.store.q.") || // Added in 3.9
339 Name == "avx512.mask.store.ss" || // Added in 7.0
340 Name.startswith("avx512.mask.loadu.") || // Added in 3.9
341 Name.startswith("avx512.mask.load.") || // Added in 3.9
342 Name.startswith("avx512.mask.expand.load.") || // Added in 7.0
343 Name.startswith("avx512.mask.compress.store.") || // Added in 7.0
344 Name.startswith("avx512.mask.expand.b") || // Added in 9.0
345 Name.startswith("avx512.mask.expand.w") || // Added in 9.0
346 Name.startswith("avx512.mask.expand.d") || // Added in 9.0
347 Name.startswith("avx512.mask.expand.q") || // Added in 9.0
348 Name.startswith("avx512.mask.expand.p") || // Added in 9.0
349 Name.startswith("avx512.mask.compress.b") || // Added in 9.0
350 Name.startswith("avx512.mask.compress.w") || // Added in 9.0
351 Name.startswith("avx512.mask.compress.d") || // Added in 9.0
352 Name.startswith("avx512.mask.compress.q") || // Added in 9.0
353 Name.startswith("avx512.mask.compress.p") || // Added in 9.0
354 Name == "sse42.crc32.64.8" || // Added in 3.4
355 Name.startswith("avx.vbroadcast.s") || // Added in 3.5
356 Name.startswith("avx512.vbroadcast.s") || // Added in 7.0
357 Name.startswith("avx512.mask.palignr.") || // Added in 3.9
358 Name.startswith("avx512.mask.valign.") || // Added in 4.0
359 Name.startswith("sse2.psll.dq") || // Added in 3.7
360 Name.startswith("sse2.psrl.dq") || // Added in 3.7
361 Name.startswith("avx2.psll.dq") || // Added in 3.7
362 Name.startswith("avx2.psrl.dq") || // Added in 3.7
363 Name.startswith("avx512.psll.dq") || // Added in 3.9
364 Name.startswith("avx512.psrl.dq") || // Added in 3.9
365 Name == "sse41.pblendw" || // Added in 3.7
366 Name.startswith("sse41.blendp") || // Added in 3.7
367 Name.startswith("avx.blend.p") || // Added in 3.7
368 Name == "avx2.pblendw" || // Added in 3.7
369 Name.startswith("avx2.pblendd.") || // Added in 3.7
370 Name.startswith("avx.vbroadcastf128") || // Added in 4.0
371 Name == "avx2.vbroadcasti128" || // Added in 3.7
372 Name.startswith("avx512.mask.broadcastf") || // Added in 6.0
373 Name.startswith("avx512.mask.broadcasti") || // Added in 6.0
374 Name == "xop.vpcmov" || // Added in 3.8
375 Name == "xop.vpcmov.256" || // Added in 5.0
376 Name.startswith("avx512.mask.move.s") || // Added in 4.0
377 Name.startswith("avx512.cvtmask2") || // Added in 5.0
378 Name.startswith("xop.vpcom") || // Added in 3.2, Updated in 9.0
379 Name.startswith("xop.vprot") || // Added in 8.0
380 Name.startswith("avx512.prol") || // Added in 8.0
381 Name.startswith("avx512.pror") || // Added in 8.0
382 Name.startswith("avx512.mask.prorv.") || // Added in 8.0
383 Name.startswith("avx512.mask.pror.") || // Added in 8.0
384 Name.startswith("avx512.mask.prolv.") || // Added in 8.0
385 Name.startswith("avx512.mask.prol.") || // Added in 8.0
386 Name.startswith("avx512.ptestm") || //Added in 6.0
387 Name.startswith("avx512.ptestnm") || //Added in 6.0
388 Name.startswith("avx512.mask.pavg")) // Added in 6.0
389 return true;
391 return false;
394 static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
395 Function *&NewFn) {
396 // Only handle intrinsics that start with "x86.".
397 if (!Name.startswith("x86."))
398 return false;
399 // Remove "x86." prefix.
400 Name = Name.substr(4);
402 if (ShouldUpgradeX86Intrinsic(F, Name)) {
403 NewFn = nullptr;
404 return true;
407 if (Name == "rdtscp") { // Added in 8.0
408 // If this intrinsic has 0 operands, it's the new version.
409 if (F->getFunctionType()->getNumParams() == 0)
410 return false;
412 rename(F);
413 NewFn = Intrinsic::getDeclaration(F->getParent(),
414 Intrinsic::x86_rdtscp);
415 return true;
418 // SSE4.1 ptest functions may have an old signature.
419 if (Name.startswith("sse41.ptest")) { // Added in 3.2
420 if (Name.substr(11) == "c")
421 return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestc, NewFn);
422 if (Name.substr(11) == "z")
423 return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestz, NewFn);
424 if (Name.substr(11) == "nzc")
425 return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
427 // Several blend and other instructions with masks used the wrong number of
428 // bits.
429 if (Name == "sse41.insertps") // Added in 3.6
430 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
431 NewFn);
432 if (Name == "sse41.dppd") // Added in 3.6
433 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
434 NewFn);
435 if (Name == "sse41.dpps") // Added in 3.6
436 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
437 NewFn);
438 if (Name == "sse41.mpsadbw") // Added in 3.6
439 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
440 NewFn);
441 if (Name == "avx.dp.ps.256") // Added in 3.6
442 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
443 NewFn);
444 if (Name == "avx2.mpsadbw") // Added in 3.6
445 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
446 NewFn);
448 // frcz.ss/sd may need to have an argument dropped. Added in 3.2
449 if (Name.startswith("xop.vfrcz.ss") && F->arg_size() == 2) {
450 rename(F);
451 NewFn = Intrinsic::getDeclaration(F->getParent(),
452 Intrinsic::x86_xop_vfrcz_ss);
453 return true;
455 if (Name.startswith("xop.vfrcz.sd") && F->arg_size() == 2) {
456 rename(F);
457 NewFn = Intrinsic::getDeclaration(F->getParent(),
458 Intrinsic::x86_xop_vfrcz_sd);
459 return true;
461 // Upgrade any XOP PERMIL2 index operand still using a float/double vector.
462 if (Name.startswith("xop.vpermil2")) { // Added in 3.9
463 auto Idx = F->getFunctionType()->getParamType(2);
464 if (Idx->isFPOrFPVectorTy()) {
465 rename(F);
466 unsigned IdxSize = Idx->getPrimitiveSizeInBits();
467 unsigned EltSize = Idx->getScalarSizeInBits();
468 Intrinsic::ID Permil2ID;
469 if (EltSize == 64 && IdxSize == 128)
470 Permil2ID = Intrinsic::x86_xop_vpermil2pd;
471 else if (EltSize == 32 && IdxSize == 128)
472 Permil2ID = Intrinsic::x86_xop_vpermil2ps;
473 else if (EltSize == 64 && IdxSize == 256)
474 Permil2ID = Intrinsic::x86_xop_vpermil2pd_256;
475 else
476 Permil2ID = Intrinsic::x86_xop_vpermil2ps_256;
477 NewFn = Intrinsic::getDeclaration(F->getParent(), Permil2ID);
478 return true;
482 if (Name == "seh.recoverfp") {
483 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::eh_recoverfp);
484 return true;
487 return false;
490 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
491 assert(F && "Illegal to upgrade a non-existent Function.");
493 // Quickly eliminate it, if it's not a candidate.
494 StringRef Name = F->getName();
495 if (Name.size() <= 8 || !Name.startswith("llvm."))
496 return false;
497 Name = Name.substr(5); // Strip off "llvm."
499 switch (Name[0]) {
500 default: break;
501 case 'a': {
502 if (Name.startswith("arm.rbit") || Name.startswith("aarch64.rbit")) {
503 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
504 F->arg_begin()->getType());
505 return true;
507 if (Name.startswith("arm.neon.vclz")) {
508 Type* args[2] = {
509 F->arg_begin()->getType(),
510 Type::getInt1Ty(F->getContext())
512 // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
513 // the end of the name. Change name from llvm.arm.neon.vclz.* to
514 // llvm.ctlz.*
515 FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
516 NewFn = Function::Create(fType, F->getLinkage(), F->getAddressSpace(),
517 "llvm.ctlz." + Name.substr(14), F->getParent());
518 return true;
520 if (Name.startswith("arm.neon.vcnt")) {
521 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
522 F->arg_begin()->getType());
523 return true;
525 static const Regex vldRegex("^arm\\.neon\\.vld([1234]|[234]lane)\\.v[a-z0-9]*$");
526 if (vldRegex.match(Name)) {
527 auto fArgs = F->getFunctionType()->params();
528 SmallVector<Type *, 4> Tys(fArgs.begin(), fArgs.end());
529 // Can't use Intrinsic::getDeclaration here as the return types might
530 // then only be structurally equal.
531 FunctionType* fType = FunctionType::get(F->getReturnType(), Tys, false);
532 NewFn = Function::Create(fType, F->getLinkage(), F->getAddressSpace(),
533 "llvm." + Name + ".p0i8", F->getParent());
534 return true;
536 static const Regex vstRegex("^arm\\.neon\\.vst([1234]|[234]lane)\\.v[a-z0-9]*$");
537 if (vstRegex.match(Name)) {
538 static const Intrinsic::ID StoreInts[] = {Intrinsic::arm_neon_vst1,
539 Intrinsic::arm_neon_vst2,
540 Intrinsic::arm_neon_vst3,
541 Intrinsic::arm_neon_vst4};
543 static const Intrinsic::ID StoreLaneInts[] = {
544 Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
545 Intrinsic::arm_neon_vst4lane
548 auto fArgs = F->getFunctionType()->params();
549 Type *Tys[] = {fArgs[0], fArgs[1]};
550 if (Name.find("lane") == StringRef::npos)
551 NewFn = Intrinsic::getDeclaration(F->getParent(),
552 StoreInts[fArgs.size() - 3], Tys);
553 else
554 NewFn = Intrinsic::getDeclaration(F->getParent(),
555 StoreLaneInts[fArgs.size() - 5], Tys);
556 return true;
558 if (Name == "aarch64.thread.pointer" || Name == "arm.thread.pointer") {
559 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
560 return true;
562 if (Name.startswith("aarch64.neon.addp")) {
563 if (F->arg_size() != 2)
564 break; // Invalid IR.
565 auto fArgs = F->getFunctionType()->params();
566 VectorType *ArgTy = dyn_cast<VectorType>(fArgs[0]);
567 if (ArgTy && ArgTy->getElementType()->isFloatingPointTy()) {
568 NewFn = Intrinsic::getDeclaration(F->getParent(),
569 Intrinsic::aarch64_neon_faddp, fArgs);
570 return true;
573 break;
576 case 'c': {
577 if (Name.startswith("ctlz.") && F->arg_size() == 1) {
578 rename(F);
579 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
580 F->arg_begin()->getType());
581 return true;
583 if (Name.startswith("cttz.") && F->arg_size() == 1) {
584 rename(F);
585 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
586 F->arg_begin()->getType());
587 return true;
589 break;
591 case 'd': {
592 if (Name == "dbg.value" && F->arg_size() == 4) {
593 rename(F);
594 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_value);
595 return true;
597 break;
599 case 'e': {
600 SmallVector<StringRef, 2> Groups;
601 static const Regex R("^experimental.vector.reduce.([a-z]+)\\.[fi][0-9]+");
602 if (R.match(Name, &Groups)) {
603 Intrinsic::ID ID = Intrinsic::not_intrinsic;
604 if (Groups[1] == "fadd")
605 ID = Intrinsic::experimental_vector_reduce_v2_fadd;
606 if (Groups[1] == "fmul")
607 ID = Intrinsic::experimental_vector_reduce_v2_fmul;
609 if (ID != Intrinsic::not_intrinsic) {
610 rename(F);
611 auto Args = F->getFunctionType()->params();
612 Type *Tys[] = {F->getFunctionType()->getReturnType(), Args[1]};
613 NewFn = Intrinsic::getDeclaration(F->getParent(), ID, Tys);
614 return true;
617 break;
619 case 'i':
620 case 'l': {
621 bool IsLifetimeStart = Name.startswith("lifetime.start");
622 if (IsLifetimeStart || Name.startswith("invariant.start")) {
623 Intrinsic::ID ID = IsLifetimeStart ?
624 Intrinsic::lifetime_start : Intrinsic::invariant_start;
625 auto Args = F->getFunctionType()->params();
626 Type* ObjectPtr[1] = {Args[1]};
627 if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
628 rename(F);
629 NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
630 return true;
634 bool IsLifetimeEnd = Name.startswith("lifetime.end");
635 if (IsLifetimeEnd || Name.startswith("invariant.end")) {
636 Intrinsic::ID ID = IsLifetimeEnd ?
637 Intrinsic::lifetime_end : Intrinsic::invariant_end;
639 auto Args = F->getFunctionType()->params();
640 Type* ObjectPtr[1] = {Args[IsLifetimeEnd ? 1 : 2]};
641 if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
642 rename(F);
643 NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
644 return true;
647 if (Name.startswith("invariant.group.barrier")) {
648 // Rename invariant.group.barrier to launder.invariant.group
649 auto Args = F->getFunctionType()->params();
650 Type* ObjectPtr[1] = {Args[0]};
651 rename(F);
652 NewFn = Intrinsic::getDeclaration(F->getParent(),
653 Intrinsic::launder_invariant_group, ObjectPtr);
654 return true;
658 break;
660 case 'm': {
661 if (Name.startswith("masked.load.")) {
662 Type *Tys[] = { F->getReturnType(), F->arg_begin()->getType() };
663 if (F->getName() != Intrinsic::getName(Intrinsic::masked_load, Tys)) {
664 rename(F);
665 NewFn = Intrinsic::getDeclaration(F->getParent(),
666 Intrinsic::masked_load,
667 Tys);
668 return true;
671 if (Name.startswith("masked.store.")) {
672 auto Args = F->getFunctionType()->params();
673 Type *Tys[] = { Args[0], Args[1] };
674 if (F->getName() != Intrinsic::getName(Intrinsic::masked_store, Tys)) {
675 rename(F);
676 NewFn = Intrinsic::getDeclaration(F->getParent(),
677 Intrinsic::masked_store,
678 Tys);
679 return true;
682 // Renaming gather/scatter intrinsics with no address space overloading
683 // to the new overload which includes an address space
684 if (Name.startswith("masked.gather.")) {
685 Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
686 if (F->getName() != Intrinsic::getName(Intrinsic::masked_gather, Tys)) {
687 rename(F);
688 NewFn = Intrinsic::getDeclaration(F->getParent(),
689 Intrinsic::masked_gather, Tys);
690 return true;
693 if (Name.startswith("masked.scatter.")) {
694 auto Args = F->getFunctionType()->params();
695 Type *Tys[] = {Args[0], Args[1]};
696 if (F->getName() != Intrinsic::getName(Intrinsic::masked_scatter, Tys)) {
697 rename(F);
698 NewFn = Intrinsic::getDeclaration(F->getParent(),
699 Intrinsic::masked_scatter, Tys);
700 return true;
703 // Updating the memory intrinsics (memcpy/memmove/memset) that have an
704 // alignment parameter to embedding the alignment as an attribute of
705 // the pointer args.
706 if (Name.startswith("memcpy.") && F->arg_size() == 5) {
707 rename(F);
708 // Get the types of dest, src, and len
709 ArrayRef<Type *> ParamTypes = F->getFunctionType()->params().slice(0, 3);
710 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memcpy,
711 ParamTypes);
712 return true;
714 if (Name.startswith("memmove.") && F->arg_size() == 5) {
715 rename(F);
716 // Get the types of dest, src, and len
717 ArrayRef<Type *> ParamTypes = F->getFunctionType()->params().slice(0, 3);
718 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memmove,
719 ParamTypes);
720 return true;
722 if (Name.startswith("memset.") && F->arg_size() == 5) {
723 rename(F);
724 // Get the types of dest, and len
725 const auto *FT = F->getFunctionType();
726 Type *ParamTypes[2] = {
727 FT->getParamType(0), // Dest
728 FT->getParamType(2) // len
730 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::memset,
731 ParamTypes);
732 return true;
734 break;
736 case 'n': {
737 if (Name.startswith("nvvm.")) {
738 Name = Name.substr(5);
740 // The following nvvm intrinsics correspond exactly to an LLVM intrinsic.
741 Intrinsic::ID IID = StringSwitch<Intrinsic::ID>(Name)
742 .Cases("brev32", "brev64", Intrinsic::bitreverse)
743 .Case("clz.i", Intrinsic::ctlz)
744 .Case("popc.i", Intrinsic::ctpop)
745 .Default(Intrinsic::not_intrinsic);
746 if (IID != Intrinsic::not_intrinsic && F->arg_size() == 1) {
747 NewFn = Intrinsic::getDeclaration(F->getParent(), IID,
748 {F->getReturnType()});
749 return true;
752 // The following nvvm intrinsics correspond exactly to an LLVM idiom, but
753 // not to an intrinsic alone. We expand them in UpgradeIntrinsicCall.
755 // TODO: We could add lohi.i2d.
756 bool Expand = StringSwitch<bool>(Name)
757 .Cases("abs.i", "abs.ll", true)
758 .Cases("clz.ll", "popc.ll", "h2f", true)
759 .Cases("max.i", "max.ll", "max.ui", "max.ull", true)
760 .Cases("min.i", "min.ll", "min.ui", "min.ull", true)
761 .StartsWith("atomic.load.add.f32.p", true)
762 .StartsWith("atomic.load.add.f64.p", true)
763 .Default(false);
764 if (Expand) {
765 NewFn = nullptr;
766 return true;
769 break;
771 case 'o':
772 // We only need to change the name to match the mangling including the
773 // address space.
774 if (Name.startswith("objectsize.")) {
775 Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
776 if (F->arg_size() == 2 || F->arg_size() == 3 ||
777 F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
778 rename(F);
779 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::objectsize,
780 Tys);
781 return true;
784 break;
786 case 'p':
787 if (Name == "prefetch") {
788 // Handle address space overloading.
789 Type *Tys[] = {F->arg_begin()->getType()};
790 if (F->getName() != Intrinsic::getName(Intrinsic::prefetch, Tys)) {
791 rename(F);
792 NewFn =
793 Intrinsic::getDeclaration(F->getParent(), Intrinsic::prefetch, Tys);
794 return true;
797 break;
799 case 's':
800 if (Name == "stackprotectorcheck") {
801 NewFn = nullptr;
802 return true;
804 break;
806 case 'x':
807 if (UpgradeX86IntrinsicFunction(F, Name, NewFn))
808 return true;
810 // Remangle our intrinsic since we upgrade the mangling
811 auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F);
812 if (Result != None) {
813 NewFn = Result.getValue();
814 return true;
817 // This may not belong here. This function is effectively being overloaded
818 // to both detect an intrinsic which needs upgrading, and to provide the
819 // upgraded form of the intrinsic. We should perhaps have two separate
820 // functions for this.
821 return false;
824 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
825 NewFn = nullptr;
826 bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
827 assert(F != NewFn && "Intrinsic function upgraded to the same function");
829 // Upgrade intrinsic attributes. This does not change the function.
830 if (NewFn)
831 F = NewFn;
832 if (Intrinsic::ID id = F->getIntrinsicID())
833 F->setAttributes(Intrinsic::getAttributes(F->getContext(), id));
834 return Upgraded;
837 GlobalVariable *llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
838 if (!(GV->hasName() && (GV->getName() == "llvm.global_ctors" ||
839 GV->getName() == "llvm.global_dtors")) ||
840 !GV->hasInitializer())
841 return nullptr;
842 ArrayType *ATy = dyn_cast<ArrayType>(GV->getValueType());
843 if (!ATy)
844 return nullptr;
845 StructType *STy = dyn_cast<StructType>(ATy->getElementType());
846 if (!STy || STy->getNumElements() != 2)
847 return nullptr;
849 LLVMContext &C = GV->getContext();
850 IRBuilder<> IRB(C);
851 auto EltTy = StructType::get(STy->getElementType(0), STy->getElementType(1),
852 IRB.getInt8PtrTy());
853 Constant *Init = GV->getInitializer();
854 unsigned N = Init->getNumOperands();
855 std::vector<Constant *> NewCtors(N);
856 for (unsigned i = 0; i != N; ++i) {
857 auto Ctor = cast<Constant>(Init->getOperand(i));
858 NewCtors[i] = ConstantStruct::get(
859 EltTy, Ctor->getAggregateElement(0u), Ctor->getAggregateElement(1),
860 Constant::getNullValue(IRB.getInt8PtrTy()));
862 Constant *NewInit = ConstantArray::get(ArrayType::get(EltTy, N), NewCtors);
864 return new GlobalVariable(NewInit->getType(), false, GV->getLinkage(),
865 NewInit, GV->getName());
868 // Handles upgrading SSE2/AVX2/AVX512BW PSLLDQ intrinsics by converting them
869 // to byte shuffles.
870 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder,
871 Value *Op, unsigned Shift) {
872 Type *ResultTy = Op->getType();
873 unsigned NumElts = ResultTy->getVectorNumElements() * 8;
875 // Bitcast from a 64-bit element type to a byte element type.
876 Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
877 Op = Builder.CreateBitCast(Op, VecTy, "cast");
879 // We'll be shuffling in zeroes.
880 Value *Res = Constant::getNullValue(VecTy);
882 // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
883 // we'll just return the zero vector.
884 if (Shift < 16) {
885 uint32_t Idxs[64];
886 // 256/512-bit version is split into 2/4 16-byte lanes.
887 for (unsigned l = 0; l != NumElts; l += 16)
888 for (unsigned i = 0; i != 16; ++i) {
889 unsigned Idx = NumElts + i - Shift;
890 if (Idx < NumElts)
891 Idx -= NumElts - 16; // end of lane, switch operand.
892 Idxs[l + i] = Idx + l;
895 Res = Builder.CreateShuffleVector(Res, Op, makeArrayRef(Idxs, NumElts));
898 // Bitcast back to a 64-bit element type.
899 return Builder.CreateBitCast(Res, ResultTy, "cast");
902 // Handles upgrading SSE2/AVX2/AVX512BW PSRLDQ intrinsics by converting them
903 // to byte shuffles.
904 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, Value *Op,
905 unsigned Shift) {
906 Type *ResultTy = Op->getType();
907 unsigned NumElts = ResultTy->getVectorNumElements() * 8;
909 // Bitcast from a 64-bit element type to a byte element type.
910 Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
911 Op = Builder.CreateBitCast(Op, VecTy, "cast");
913 // We'll be shuffling in zeroes.
914 Value *Res = Constant::getNullValue(VecTy);
916 // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
917 // we'll just return the zero vector.
918 if (Shift < 16) {
919 uint32_t Idxs[64];
920 // 256/512-bit version is split into 2/4 16-byte lanes.
921 for (unsigned l = 0; l != NumElts; l += 16)
922 for (unsigned i = 0; i != 16; ++i) {
923 unsigned Idx = i + Shift;
924 if (Idx >= 16)
925 Idx += NumElts - 16; // end of lane, switch operand.
926 Idxs[l + i] = Idx + l;
929 Res = Builder.CreateShuffleVector(Op, Res, makeArrayRef(Idxs, NumElts));
932 // Bitcast back to a 64-bit element type.
933 return Builder.CreateBitCast(Res, ResultTy, "cast");
936 static Value *getX86MaskVec(IRBuilder<> &Builder, Value *Mask,
937 unsigned NumElts) {
938 llvm::VectorType *MaskTy = llvm::VectorType::get(Builder.getInt1Ty(),
939 cast<IntegerType>(Mask->getType())->getBitWidth());
940 Mask = Builder.CreateBitCast(Mask, MaskTy);
942 // If we have less than 8 elements, then the starting mask was an i8 and
943 // we need to extract down to the right number of elements.
944 if (NumElts < 8) {
945 uint32_t Indices[4];
946 for (unsigned i = 0; i != NumElts; ++i)
947 Indices[i] = i;
948 Mask = Builder.CreateShuffleVector(Mask, Mask,
949 makeArrayRef(Indices, NumElts),
950 "extract");
953 return Mask;
956 static Value *EmitX86Select(IRBuilder<> &Builder, Value *Mask,
957 Value *Op0, Value *Op1) {
958 // If the mask is all ones just emit the first operation.
959 if (const auto *C = dyn_cast<Constant>(Mask))
960 if (C->isAllOnesValue())
961 return Op0;
963 Mask = getX86MaskVec(Builder, Mask, Op0->getType()->getVectorNumElements());
964 return Builder.CreateSelect(Mask, Op0, Op1);
967 static Value *EmitX86ScalarSelect(IRBuilder<> &Builder, Value *Mask,
968 Value *Op0, Value *Op1) {
969 // If the mask is all ones just emit the first operation.
970 if (const auto *C = dyn_cast<Constant>(Mask))
971 if (C->isAllOnesValue())
972 return Op0;
974 llvm::VectorType *MaskTy =
975 llvm::VectorType::get(Builder.getInt1Ty(),
976 Mask->getType()->getIntegerBitWidth());
977 Mask = Builder.CreateBitCast(Mask, MaskTy);
978 Mask = Builder.CreateExtractElement(Mask, (uint64_t)0);
979 return Builder.CreateSelect(Mask, Op0, Op1);
982 // Handle autoupgrade for masked PALIGNR and VALIGND/Q intrinsics.
983 // PALIGNR handles large immediates by shifting while VALIGN masks the immediate
984 // so we need to handle both cases. VALIGN also doesn't have 128-bit lanes.
985 static Value *UpgradeX86ALIGNIntrinsics(IRBuilder<> &Builder, Value *Op0,
986 Value *Op1, Value *Shift,
987 Value *Passthru, Value *Mask,
988 bool IsVALIGN) {
989 unsigned ShiftVal = cast<llvm::ConstantInt>(Shift)->getZExtValue();
991 unsigned NumElts = Op0->getType()->getVectorNumElements();
992 assert((IsVALIGN || NumElts % 16 == 0) && "Illegal NumElts for PALIGNR!");
993 assert((!IsVALIGN || NumElts <= 16) && "NumElts too large for VALIGN!");
994 assert(isPowerOf2_32(NumElts) && "NumElts not a power of 2!");
996 // Mask the immediate for VALIGN.
997 if (IsVALIGN)
998 ShiftVal &= (NumElts - 1);
1000 // If palignr is shifting the pair of vectors more than the size of two
1001 // lanes, emit zero.
1002 if (ShiftVal >= 32)
1003 return llvm::Constant::getNullValue(Op0->getType());
1005 // If palignr is shifting the pair of input vectors more than one lane,
1006 // but less than two lanes, convert to shifting in zeroes.
1007 if (ShiftVal > 16) {
1008 ShiftVal -= 16;
1009 Op1 = Op0;
1010 Op0 = llvm::Constant::getNullValue(Op0->getType());
1013 uint32_t Indices[64];
1014 // 256-bit palignr operates on 128-bit lanes so we need to handle that
1015 for (unsigned l = 0; l < NumElts; l += 16) {
1016 for (unsigned i = 0; i != 16; ++i) {
1017 unsigned Idx = ShiftVal + i;
1018 if (!IsVALIGN && Idx >= 16) // Disable wrap for VALIGN.
1019 Idx += NumElts - 16; // End of lane, switch operand.
1020 Indices[l + i] = Idx + l;
1024 Value *Align = Builder.CreateShuffleVector(Op1, Op0,
1025 makeArrayRef(Indices, NumElts),
1026 "palignr");
1028 return EmitX86Select(Builder, Mask, Align, Passthru);
1031 static Value *UpgradeX86VPERMT2Intrinsics(IRBuilder<> &Builder, CallInst &CI,
1032 bool ZeroMask, bool IndexForm) {
1033 Type *Ty = CI.getType();
1034 unsigned VecWidth = Ty->getPrimitiveSizeInBits();
1035 unsigned EltWidth = Ty->getScalarSizeInBits();
1036 bool IsFloat = Ty->isFPOrFPVectorTy();
1037 Intrinsic::ID IID;
1038 if (VecWidth == 128 && EltWidth == 32 && IsFloat)
1039 IID = Intrinsic::x86_avx512_vpermi2var_ps_128;
1040 else if (VecWidth == 128 && EltWidth == 32 && !IsFloat)
1041 IID = Intrinsic::x86_avx512_vpermi2var_d_128;
1042 else if (VecWidth == 128 && EltWidth == 64 && IsFloat)
1043 IID = Intrinsic::x86_avx512_vpermi2var_pd_128;
1044 else if (VecWidth == 128 && EltWidth == 64 && !IsFloat)
1045 IID = Intrinsic::x86_avx512_vpermi2var_q_128;
1046 else if (VecWidth == 256 && EltWidth == 32 && IsFloat)
1047 IID = Intrinsic::x86_avx512_vpermi2var_ps_256;
1048 else if (VecWidth == 256 && EltWidth == 32 && !IsFloat)
1049 IID = Intrinsic::x86_avx512_vpermi2var_d_256;
1050 else if (VecWidth == 256 && EltWidth == 64 && IsFloat)
1051 IID = Intrinsic::x86_avx512_vpermi2var_pd_256;
1052 else if (VecWidth == 256 && EltWidth == 64 && !IsFloat)
1053 IID = Intrinsic::x86_avx512_vpermi2var_q_256;
1054 else if (VecWidth == 512 && EltWidth == 32 && IsFloat)
1055 IID = Intrinsic::x86_avx512_vpermi2var_ps_512;
1056 else if (VecWidth == 512 && EltWidth == 32 && !IsFloat)
1057 IID = Intrinsic::x86_avx512_vpermi2var_d_512;
1058 else if (VecWidth == 512 && EltWidth == 64 && IsFloat)
1059 IID = Intrinsic::x86_avx512_vpermi2var_pd_512;
1060 else if (VecWidth == 512 && EltWidth == 64 && !IsFloat)
1061 IID = Intrinsic::x86_avx512_vpermi2var_q_512;
1062 else if (VecWidth == 128 && EltWidth == 16)
1063 IID = Intrinsic::x86_avx512_vpermi2var_hi_128;
1064 else if (VecWidth == 256 && EltWidth == 16)
1065 IID = Intrinsic::x86_avx512_vpermi2var_hi_256;
1066 else if (VecWidth == 512 && EltWidth == 16)
1067 IID = Intrinsic::x86_avx512_vpermi2var_hi_512;
1068 else if (VecWidth == 128 && EltWidth == 8)
1069 IID = Intrinsic::x86_avx512_vpermi2var_qi_128;
1070 else if (VecWidth == 256 && EltWidth == 8)
1071 IID = Intrinsic::x86_avx512_vpermi2var_qi_256;
1072 else if (VecWidth == 512 && EltWidth == 8)
1073 IID = Intrinsic::x86_avx512_vpermi2var_qi_512;
1074 else
1075 llvm_unreachable("Unexpected intrinsic");
1077 Value *Args[] = { CI.getArgOperand(0) , CI.getArgOperand(1),
1078 CI.getArgOperand(2) };
1080 // If this isn't index form we need to swap operand 0 and 1.
1081 if (!IndexForm)
1082 std::swap(Args[0], Args[1]);
1084 Value *V = Builder.CreateCall(Intrinsic::getDeclaration(CI.getModule(), IID),
1085 Args);
1086 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty)
1087 : Builder.CreateBitCast(CI.getArgOperand(1),
1088 Ty);
1089 return EmitX86Select(Builder, CI.getArgOperand(3), V, PassThru);
1092 static Value *UpgradeX86AddSubSatIntrinsics(IRBuilder<> &Builder, CallInst &CI,
1093 bool IsSigned, bool IsAddition) {
1094 Type *Ty = CI.getType();
1095 Value *Op0 = CI.getOperand(0);
1096 Value *Op1 = CI.getOperand(1);
1098 Intrinsic::ID IID =
1099 IsSigned ? (IsAddition ? Intrinsic::sadd_sat : Intrinsic::ssub_sat)
1100 : (IsAddition ? Intrinsic::uadd_sat : Intrinsic::usub_sat);
1101 Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID, Ty);
1102 Value *Res = Builder.CreateCall(Intrin, {Op0, Op1});
1104 if (CI.getNumArgOperands() == 4) { // For masked intrinsics.
1105 Value *VecSrc = CI.getOperand(2);
1106 Value *Mask = CI.getOperand(3);
1107 Res = EmitX86Select(Builder, Mask, Res, VecSrc);
1109 return Res;
1112 static Value *upgradeX86Rotate(IRBuilder<> &Builder, CallInst &CI,
1113 bool IsRotateRight) {
1114 Type *Ty = CI.getType();
1115 Value *Src = CI.getArgOperand(0);
1116 Value *Amt = CI.getArgOperand(1);
1118 // Amount may be scalar immediate, in which case create a splat vector.
1119 // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
1120 // we only care about the lowest log2 bits anyway.
1121 if (Amt->getType() != Ty) {
1122 unsigned NumElts = Ty->getVectorNumElements();
1123 Amt = Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
1124 Amt = Builder.CreateVectorSplat(NumElts, Amt);
1127 Intrinsic::ID IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl;
1128 Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID, Ty);
1129 Value *Res = Builder.CreateCall(Intrin, {Src, Src, Amt});
1131 if (CI.getNumArgOperands() == 4) { // For masked intrinsics.
1132 Value *VecSrc = CI.getOperand(2);
1133 Value *Mask = CI.getOperand(3);
1134 Res = EmitX86Select(Builder, Mask, Res, VecSrc);
1136 return Res;
1139 static Value *upgradeX86vpcom(IRBuilder<> &Builder, CallInst &CI, unsigned Imm,
1140 bool IsSigned) {
1141 Type *Ty = CI.getType();
1142 Value *LHS = CI.getArgOperand(0);
1143 Value *RHS = CI.getArgOperand(1);
1145 CmpInst::Predicate Pred;
1146 switch (Imm) {
1147 case 0x0:
1148 Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT;
1149 break;
1150 case 0x1:
1151 Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE;
1152 break;
1153 case 0x2:
1154 Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
1155 break;
1156 case 0x3:
1157 Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE;
1158 break;
1159 case 0x4:
1160 Pred = ICmpInst::ICMP_EQ;
1161 break;
1162 case 0x5:
1163 Pred = ICmpInst::ICMP_NE;
1164 break;
1165 case 0x6:
1166 return Constant::getNullValue(Ty); // FALSE
1167 case 0x7:
1168 return Constant::getAllOnesValue(Ty); // TRUE
1169 default:
1170 llvm_unreachable("Unknown XOP vpcom/vpcomu predicate");
1173 Value *Cmp = Builder.CreateICmp(Pred, LHS, RHS);
1174 Value *Ext = Builder.CreateSExt(Cmp, Ty);
1175 return Ext;
1178 static Value *upgradeX86ConcatShift(IRBuilder<> &Builder, CallInst &CI,
1179 bool IsShiftRight, bool ZeroMask) {
1180 Type *Ty = CI.getType();
1181 Value *Op0 = CI.getArgOperand(0);
1182 Value *Op1 = CI.getArgOperand(1);
1183 Value *Amt = CI.getArgOperand(2);
1185 if (IsShiftRight)
1186 std::swap(Op0, Op1);
1188 // Amount may be scalar immediate, in which case create a splat vector.
1189 // Funnel shifts amounts are treated as modulo and types are all power-of-2 so
1190 // we only care about the lowest log2 bits anyway.
1191 if (Amt->getType() != Ty) {
1192 unsigned NumElts = Ty->getVectorNumElements();
1193 Amt = Builder.CreateIntCast(Amt, Ty->getScalarType(), false);
1194 Amt = Builder.CreateVectorSplat(NumElts, Amt);
1197 Intrinsic::ID IID = IsShiftRight ? Intrinsic::fshr : Intrinsic::fshl;
1198 Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID, Ty);
1199 Value *Res = Builder.CreateCall(Intrin, {Op0, Op1, Amt});
1201 unsigned NumArgs = CI.getNumArgOperands();
1202 if (NumArgs >= 4) { // For masked intrinsics.
1203 Value *VecSrc = NumArgs == 5 ? CI.getArgOperand(3) :
1204 ZeroMask ? ConstantAggregateZero::get(CI.getType()) :
1205 CI.getArgOperand(0);
1206 Value *Mask = CI.getOperand(NumArgs - 1);
1207 Res = EmitX86Select(Builder, Mask, Res, VecSrc);
1209 return Res;
1212 static Value *UpgradeMaskedStore(IRBuilder<> &Builder,
1213 Value *Ptr, Value *Data, Value *Mask,
1214 bool Aligned) {
1215 // Cast the pointer to the right type.
1216 Ptr = Builder.CreateBitCast(Ptr,
1217 llvm::PointerType::getUnqual(Data->getType()));
1218 unsigned Align =
1219 Aligned ? cast<VectorType>(Data->getType())->getBitWidth() / 8 : 1;
1221 // If the mask is all ones just emit a regular store.
1222 if (const auto *C = dyn_cast<Constant>(Mask))
1223 if (C->isAllOnesValue())
1224 return Builder.CreateAlignedStore(Data, Ptr, Align);
1226 // Convert the mask from an integer type to a vector of i1.
1227 unsigned NumElts = Data->getType()->getVectorNumElements();
1228 Mask = getX86MaskVec(Builder, Mask, NumElts);
1229 return Builder.CreateMaskedStore(Data, Ptr, Align, Mask);
1232 static Value *UpgradeMaskedLoad(IRBuilder<> &Builder,
1233 Value *Ptr, Value *Passthru, Value *Mask,
1234 bool Aligned) {
1235 Type *ValTy = Passthru->getType();
1236 // Cast the pointer to the right type.
1237 Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(ValTy));
1238 unsigned Align =
1239 Aligned ? cast<VectorType>(Passthru->getType())->getBitWidth() / 8 : 1;
1241 // If the mask is all ones just emit a regular store.
1242 if (const auto *C = dyn_cast<Constant>(Mask))
1243 if (C->isAllOnesValue())
1244 return Builder.CreateAlignedLoad(ValTy, Ptr, Align);
1246 // Convert the mask from an integer type to a vector of i1.
1247 unsigned NumElts = Passthru->getType()->getVectorNumElements();
1248 Mask = getX86MaskVec(Builder, Mask, NumElts);
1249 return Builder.CreateMaskedLoad(Ptr, Align, Mask, Passthru);
1252 static Value *upgradeAbs(IRBuilder<> &Builder, CallInst &CI) {
1253 Value *Op0 = CI.getArgOperand(0);
1254 llvm::Type *Ty = Op0->getType();
1255 Value *Zero = llvm::Constant::getNullValue(Ty);
1256 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_SGT, Op0, Zero);
1257 Value *Neg = Builder.CreateNeg(Op0);
1258 Value *Res = Builder.CreateSelect(Cmp, Op0, Neg);
1260 if (CI.getNumArgOperands() == 3)
1261 Res = EmitX86Select(Builder,CI.getArgOperand(2), Res, CI.getArgOperand(1));
1263 return Res;
1266 static Value *upgradeIntMinMax(IRBuilder<> &Builder, CallInst &CI,
1267 ICmpInst::Predicate Pred) {
1268 Value *Op0 = CI.getArgOperand(0);
1269 Value *Op1 = CI.getArgOperand(1);
1270 Value *Cmp = Builder.CreateICmp(Pred, Op0, Op1);
1271 Value *Res = Builder.CreateSelect(Cmp, Op0, Op1);
1273 if (CI.getNumArgOperands() == 4)
1274 Res = EmitX86Select(Builder, CI.getArgOperand(3), Res, CI.getArgOperand(2));
1276 return Res;
1279 static Value *upgradePMULDQ(IRBuilder<> &Builder, CallInst &CI, bool IsSigned) {
1280 Type *Ty = CI.getType();
1282 // Arguments have a vXi32 type so cast to vXi64.
1283 Value *LHS = Builder.CreateBitCast(CI.getArgOperand(0), Ty);
1284 Value *RHS = Builder.CreateBitCast(CI.getArgOperand(1), Ty);
1286 if (IsSigned) {
1287 // Shift left then arithmetic shift right.
1288 Constant *ShiftAmt = ConstantInt::get(Ty, 32);
1289 LHS = Builder.CreateShl(LHS, ShiftAmt);
1290 LHS = Builder.CreateAShr(LHS, ShiftAmt);
1291 RHS = Builder.CreateShl(RHS, ShiftAmt);
1292 RHS = Builder.CreateAShr(RHS, ShiftAmt);
1293 } else {
1294 // Clear the upper bits.
1295 Constant *Mask = ConstantInt::get(Ty, 0xffffffff);
1296 LHS = Builder.CreateAnd(LHS, Mask);
1297 RHS = Builder.CreateAnd(RHS, Mask);
1300 Value *Res = Builder.CreateMul(LHS, RHS);
1302 if (CI.getNumArgOperands() == 4)
1303 Res = EmitX86Select(Builder, CI.getArgOperand(3), Res, CI.getArgOperand(2));
1305 return Res;
1308 // Applying mask on vector of i1's and make sure result is at least 8 bits wide.
1309 static Value *ApplyX86MaskOn1BitsVec(IRBuilder<> &Builder, Value *Vec,
1310 Value *Mask) {
1311 unsigned NumElts = Vec->getType()->getVectorNumElements();
1312 if (Mask) {
1313 const auto *C = dyn_cast<Constant>(Mask);
1314 if (!C || !C->isAllOnesValue())
1315 Vec = Builder.CreateAnd(Vec, getX86MaskVec(Builder, Mask, NumElts));
1318 if (NumElts < 8) {
1319 uint32_t Indices[8];
1320 for (unsigned i = 0; i != NumElts; ++i)
1321 Indices[i] = i;
1322 for (unsigned i = NumElts; i != 8; ++i)
1323 Indices[i] = NumElts + i % NumElts;
1324 Vec = Builder.CreateShuffleVector(Vec,
1325 Constant::getNullValue(Vec->getType()),
1326 Indices);
1328 return Builder.CreateBitCast(Vec, Builder.getIntNTy(std::max(NumElts, 8U)));
1331 static Value *upgradeMaskedCompare(IRBuilder<> &Builder, CallInst &CI,
1332 unsigned CC, bool Signed) {
1333 Value *Op0 = CI.getArgOperand(0);
1334 unsigned NumElts = Op0->getType()->getVectorNumElements();
1336 Value *Cmp;
1337 if (CC == 3) {
1338 Cmp = Constant::getNullValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
1339 } else if (CC == 7) {
1340 Cmp = Constant::getAllOnesValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
1341 } else {
1342 ICmpInst::Predicate Pred;
1343 switch (CC) {
1344 default: llvm_unreachable("Unknown condition code");
1345 case 0: Pred = ICmpInst::ICMP_EQ; break;
1346 case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
1347 case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
1348 case 4: Pred = ICmpInst::ICMP_NE; break;
1349 case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
1350 case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
1352 Cmp = Builder.CreateICmp(Pred, Op0, CI.getArgOperand(1));
1355 Value *Mask = CI.getArgOperand(CI.getNumArgOperands() - 1);
1357 return ApplyX86MaskOn1BitsVec(Builder, Cmp, Mask);
1360 // Replace a masked intrinsic with an older unmasked intrinsic.
1361 static Value *UpgradeX86MaskedShift(IRBuilder<> &Builder, CallInst &CI,
1362 Intrinsic::ID IID) {
1363 Function *Intrin = Intrinsic::getDeclaration(CI.getModule(), IID);
1364 Value *Rep = Builder.CreateCall(Intrin,
1365 { CI.getArgOperand(0), CI.getArgOperand(1) });
1366 return EmitX86Select(Builder, CI.getArgOperand(3), Rep, CI.getArgOperand(2));
1369 static Value* upgradeMaskedMove(IRBuilder<> &Builder, CallInst &CI) {
1370 Value* A = CI.getArgOperand(0);
1371 Value* B = CI.getArgOperand(1);
1372 Value* Src = CI.getArgOperand(2);
1373 Value* Mask = CI.getArgOperand(3);
1375 Value* AndNode = Builder.CreateAnd(Mask, APInt(8, 1));
1376 Value* Cmp = Builder.CreateIsNotNull(AndNode);
1377 Value* Extract1 = Builder.CreateExtractElement(B, (uint64_t)0);
1378 Value* Extract2 = Builder.CreateExtractElement(Src, (uint64_t)0);
1379 Value* Select = Builder.CreateSelect(Cmp, Extract1, Extract2);
1380 return Builder.CreateInsertElement(A, Select, (uint64_t)0);
1384 static Value* UpgradeMaskToInt(IRBuilder<> &Builder, CallInst &CI) {
1385 Value* Op = CI.getArgOperand(0);
1386 Type* ReturnOp = CI.getType();
1387 unsigned NumElts = CI.getType()->getVectorNumElements();
1388 Value *Mask = getX86MaskVec(Builder, Op, NumElts);
1389 return Builder.CreateSExt(Mask, ReturnOp, "vpmovm2");
1392 // Replace intrinsic with unmasked version and a select.
1393 static bool upgradeAVX512MaskToSelect(StringRef Name, IRBuilder<> &Builder,
1394 CallInst &CI, Value *&Rep) {
1395 Name = Name.substr(12); // Remove avx512.mask.
1397 unsigned VecWidth = CI.getType()->getPrimitiveSizeInBits();
1398 unsigned EltWidth = CI.getType()->getScalarSizeInBits();
1399 Intrinsic::ID IID;
1400 if (Name.startswith("max.p")) {
1401 if (VecWidth == 128 && EltWidth == 32)
1402 IID = Intrinsic::x86_sse_max_ps;
1403 else if (VecWidth == 128 && EltWidth == 64)
1404 IID = Intrinsic::x86_sse2_max_pd;
1405 else if (VecWidth == 256 && EltWidth == 32)
1406 IID = Intrinsic::x86_avx_max_ps_256;
1407 else if (VecWidth == 256 && EltWidth == 64)
1408 IID = Intrinsic::x86_avx_max_pd_256;
1409 else
1410 llvm_unreachable("Unexpected intrinsic");
1411 } else if (Name.startswith("min.p")) {
1412 if (VecWidth == 128 && EltWidth == 32)
1413 IID = Intrinsic::x86_sse_min_ps;
1414 else if (VecWidth == 128 && EltWidth == 64)
1415 IID = Intrinsic::x86_sse2_min_pd;
1416 else if (VecWidth == 256 && EltWidth == 32)
1417 IID = Intrinsic::x86_avx_min_ps_256;
1418 else if (VecWidth == 256 && EltWidth == 64)
1419 IID = Intrinsic::x86_avx_min_pd_256;
1420 else
1421 llvm_unreachable("Unexpected intrinsic");
1422 } else if (Name.startswith("pshuf.b.")) {
1423 if (VecWidth == 128)
1424 IID = Intrinsic::x86_ssse3_pshuf_b_128;
1425 else if (VecWidth == 256)
1426 IID = Intrinsic::x86_avx2_pshuf_b;
1427 else if (VecWidth == 512)
1428 IID = Intrinsic::x86_avx512_pshuf_b_512;
1429 else
1430 llvm_unreachable("Unexpected intrinsic");
1431 } else if (Name.startswith("pmul.hr.sw.")) {
1432 if (VecWidth == 128)
1433 IID = Intrinsic::x86_ssse3_pmul_hr_sw_128;
1434 else if (VecWidth == 256)
1435 IID = Intrinsic::x86_avx2_pmul_hr_sw;
1436 else if (VecWidth == 512)
1437 IID = Intrinsic::x86_avx512_pmul_hr_sw_512;
1438 else
1439 llvm_unreachable("Unexpected intrinsic");
1440 } else if (Name.startswith("pmulh.w.")) {
1441 if (VecWidth == 128)
1442 IID = Intrinsic::x86_sse2_pmulh_w;
1443 else if (VecWidth == 256)
1444 IID = Intrinsic::x86_avx2_pmulh_w;
1445 else if (VecWidth == 512)
1446 IID = Intrinsic::x86_avx512_pmulh_w_512;
1447 else
1448 llvm_unreachable("Unexpected intrinsic");
1449 } else if (Name.startswith("pmulhu.w.")) {
1450 if (VecWidth == 128)
1451 IID = Intrinsic::x86_sse2_pmulhu_w;
1452 else if (VecWidth == 256)
1453 IID = Intrinsic::x86_avx2_pmulhu_w;
1454 else if (VecWidth == 512)
1455 IID = Intrinsic::x86_avx512_pmulhu_w_512;
1456 else
1457 llvm_unreachable("Unexpected intrinsic");
1458 } else if (Name.startswith("pmaddw.d.")) {
1459 if (VecWidth == 128)
1460 IID = Intrinsic::x86_sse2_pmadd_wd;
1461 else if (VecWidth == 256)
1462 IID = Intrinsic::x86_avx2_pmadd_wd;
1463 else if (VecWidth == 512)
1464 IID = Intrinsic::x86_avx512_pmaddw_d_512;
1465 else
1466 llvm_unreachable("Unexpected intrinsic");
1467 } else if (Name.startswith("pmaddubs.w.")) {
1468 if (VecWidth == 128)
1469 IID = Intrinsic::x86_ssse3_pmadd_ub_sw_128;
1470 else if (VecWidth == 256)
1471 IID = Intrinsic::x86_avx2_pmadd_ub_sw;
1472 else if (VecWidth == 512)
1473 IID = Intrinsic::x86_avx512_pmaddubs_w_512;
1474 else
1475 llvm_unreachable("Unexpected intrinsic");
1476 } else if (Name.startswith("packsswb.")) {
1477 if (VecWidth == 128)
1478 IID = Intrinsic::x86_sse2_packsswb_128;
1479 else if (VecWidth == 256)
1480 IID = Intrinsic::x86_avx2_packsswb;
1481 else if (VecWidth == 512)
1482 IID = Intrinsic::x86_avx512_packsswb_512;
1483 else
1484 llvm_unreachable("Unexpected intrinsic");
1485 } else if (Name.startswith("packssdw.")) {
1486 if (VecWidth == 128)
1487 IID = Intrinsic::x86_sse2_packssdw_128;
1488 else if (VecWidth == 256)
1489 IID = Intrinsic::x86_avx2_packssdw;
1490 else if (VecWidth == 512)
1491 IID = Intrinsic::x86_avx512_packssdw_512;
1492 else
1493 llvm_unreachable("Unexpected intrinsic");
1494 } else if (Name.startswith("packuswb.")) {
1495 if (VecWidth == 128)
1496 IID = Intrinsic::x86_sse2_packuswb_128;
1497 else if (VecWidth == 256)
1498 IID = Intrinsic::x86_avx2_packuswb;
1499 else if (VecWidth == 512)
1500 IID = Intrinsic::x86_avx512_packuswb_512;
1501 else
1502 llvm_unreachable("Unexpected intrinsic");
1503 } else if (Name.startswith("packusdw.")) {
1504 if (VecWidth == 128)
1505 IID = Intrinsic::x86_sse41_packusdw;
1506 else if (VecWidth == 256)
1507 IID = Intrinsic::x86_avx2_packusdw;
1508 else if (VecWidth == 512)
1509 IID = Intrinsic::x86_avx512_packusdw_512;
1510 else
1511 llvm_unreachable("Unexpected intrinsic");
1512 } else if (Name.startswith("vpermilvar.")) {
1513 if (VecWidth == 128 && EltWidth == 32)
1514 IID = Intrinsic::x86_avx_vpermilvar_ps;
1515 else if (VecWidth == 128 && EltWidth == 64)
1516 IID = Intrinsic::x86_avx_vpermilvar_pd;
1517 else if (VecWidth == 256 && EltWidth == 32)
1518 IID = Intrinsic::x86_avx_vpermilvar_ps_256;
1519 else if (VecWidth == 256 && EltWidth == 64)
1520 IID = Intrinsic::x86_avx_vpermilvar_pd_256;
1521 else if (VecWidth == 512 && EltWidth == 32)
1522 IID = Intrinsic::x86_avx512_vpermilvar_ps_512;
1523 else if (VecWidth == 512 && EltWidth == 64)
1524 IID = Intrinsic::x86_avx512_vpermilvar_pd_512;
1525 else
1526 llvm_unreachable("Unexpected intrinsic");
1527 } else if (Name == "cvtpd2dq.256") {
1528 IID = Intrinsic::x86_avx_cvt_pd2dq_256;
1529 } else if (Name == "cvtpd2ps.256") {
1530 IID = Intrinsic::x86_avx_cvt_pd2_ps_256;
1531 } else if (Name == "cvttpd2dq.256") {
1532 IID = Intrinsic::x86_avx_cvtt_pd2dq_256;
1533 } else if (Name == "cvttps2dq.128") {
1534 IID = Intrinsic::x86_sse2_cvttps2dq;
1535 } else if (Name == "cvttps2dq.256") {
1536 IID = Intrinsic::x86_avx_cvtt_ps2dq_256;
1537 } else if (Name.startswith("permvar.")) {
1538 bool IsFloat = CI.getType()->isFPOrFPVectorTy();
1539 if (VecWidth == 256 && EltWidth == 32 && IsFloat)
1540 IID = Intrinsic::x86_avx2_permps;
1541 else if (VecWidth == 256 && EltWidth == 32 && !IsFloat)
1542 IID = Intrinsic::x86_avx2_permd;
1543 else if (VecWidth == 256 && EltWidth == 64 && IsFloat)
1544 IID = Intrinsic::x86_avx512_permvar_df_256;
1545 else if (VecWidth == 256 && EltWidth == 64 && !IsFloat)
1546 IID = Intrinsic::x86_avx512_permvar_di_256;
1547 else if (VecWidth == 512 && EltWidth == 32 && IsFloat)
1548 IID = Intrinsic::x86_avx512_permvar_sf_512;
1549 else if (VecWidth == 512 && EltWidth == 32 && !IsFloat)
1550 IID = Intrinsic::x86_avx512_permvar_si_512;
1551 else if (VecWidth == 512 && EltWidth == 64 && IsFloat)
1552 IID = Intrinsic::x86_avx512_permvar_df_512;
1553 else if (VecWidth == 512 && EltWidth == 64 && !IsFloat)
1554 IID = Intrinsic::x86_avx512_permvar_di_512;
1555 else if (VecWidth == 128 && EltWidth == 16)
1556 IID = Intrinsic::x86_avx512_permvar_hi_128;
1557 else if (VecWidth == 256 && EltWidth == 16)
1558 IID = Intrinsic::x86_avx512_permvar_hi_256;
1559 else if (VecWidth == 512 && EltWidth == 16)
1560 IID = Intrinsic::x86_avx512_permvar_hi_512;
1561 else if (VecWidth == 128 && EltWidth == 8)
1562 IID = Intrinsic::x86_avx512_permvar_qi_128;
1563 else if (VecWidth == 256 && EltWidth == 8)
1564 IID = Intrinsic::x86_avx512_permvar_qi_256;
1565 else if (VecWidth == 512 && EltWidth == 8)
1566 IID = Intrinsic::x86_avx512_permvar_qi_512;
1567 else
1568 llvm_unreachable("Unexpected intrinsic");
1569 } else if (Name.startswith("dbpsadbw.")) {
1570 if (VecWidth == 128)
1571 IID = Intrinsic::x86_avx512_dbpsadbw_128;
1572 else if (VecWidth == 256)
1573 IID = Intrinsic::x86_avx512_dbpsadbw_256;
1574 else if (VecWidth == 512)
1575 IID = Intrinsic::x86_avx512_dbpsadbw_512;
1576 else
1577 llvm_unreachable("Unexpected intrinsic");
1578 } else if (Name.startswith("pmultishift.qb.")) {
1579 if (VecWidth == 128)
1580 IID = Intrinsic::x86_avx512_pmultishift_qb_128;
1581 else if (VecWidth == 256)
1582 IID = Intrinsic::x86_avx512_pmultishift_qb_256;
1583 else if (VecWidth == 512)
1584 IID = Intrinsic::x86_avx512_pmultishift_qb_512;
1585 else
1586 llvm_unreachable("Unexpected intrinsic");
1587 } else if (Name.startswith("conflict.")) {
1588 if (Name[9] == 'd' && VecWidth == 128)
1589 IID = Intrinsic::x86_avx512_conflict_d_128;
1590 else if (Name[9] == 'd' && VecWidth == 256)
1591 IID = Intrinsic::x86_avx512_conflict_d_256;
1592 else if (Name[9] == 'd' && VecWidth == 512)
1593 IID = Intrinsic::x86_avx512_conflict_d_512;
1594 else if (Name[9] == 'q' && VecWidth == 128)
1595 IID = Intrinsic::x86_avx512_conflict_q_128;
1596 else if (Name[9] == 'q' && VecWidth == 256)
1597 IID = Intrinsic::x86_avx512_conflict_q_256;
1598 else if (Name[9] == 'q' && VecWidth == 512)
1599 IID = Intrinsic::x86_avx512_conflict_q_512;
1600 else
1601 llvm_unreachable("Unexpected intrinsic");
1602 } else if (Name.startswith("pavg.")) {
1603 if (Name[5] == 'b' && VecWidth == 128)
1604 IID = Intrinsic::x86_sse2_pavg_b;
1605 else if (Name[5] == 'b' && VecWidth == 256)
1606 IID = Intrinsic::x86_avx2_pavg_b;
1607 else if (Name[5] == 'b' && VecWidth == 512)
1608 IID = Intrinsic::x86_avx512_pavg_b_512;
1609 else if (Name[5] == 'w' && VecWidth == 128)
1610 IID = Intrinsic::x86_sse2_pavg_w;
1611 else if (Name[5] == 'w' && VecWidth == 256)
1612 IID = Intrinsic::x86_avx2_pavg_w;
1613 else if (Name[5] == 'w' && VecWidth == 512)
1614 IID = Intrinsic::x86_avx512_pavg_w_512;
1615 else
1616 llvm_unreachable("Unexpected intrinsic");
1617 } else
1618 return false;
1620 SmallVector<Value *, 4> Args(CI.arg_operands().begin(),
1621 CI.arg_operands().end());
1622 Args.pop_back();
1623 Args.pop_back();
1624 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI.getModule(), IID),
1625 Args);
1626 unsigned NumArgs = CI.getNumArgOperands();
1627 Rep = EmitX86Select(Builder, CI.getArgOperand(NumArgs - 1), Rep,
1628 CI.getArgOperand(NumArgs - 2));
1629 return true;
1632 /// Upgrade comment in call to inline asm that represents an objc retain release
1633 /// marker.
1634 void llvm::UpgradeInlineAsmString(std::string *AsmStr) {
1635 size_t Pos;
1636 if (AsmStr->find("mov\tfp") == 0 &&
1637 AsmStr->find("objc_retainAutoreleaseReturnValue") != std::string::npos &&
1638 (Pos = AsmStr->find("# marker")) != std::string::npos) {
1639 AsmStr->replace(Pos, 1, ";");
1641 return;
1644 /// Upgrade a call to an old intrinsic. All argument and return casting must be
1645 /// provided to seamlessly integrate with existing context.
1646 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
1647 Function *F = CI->getCalledFunction();
1648 LLVMContext &C = CI->getContext();
1649 IRBuilder<> Builder(C);
1650 Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
1652 assert(F && "Intrinsic call is not direct?");
1654 if (!NewFn) {
1655 // Get the Function's name.
1656 StringRef Name = F->getName();
1658 assert(Name.startswith("llvm.") && "Intrinsic doesn't start with 'llvm.'");
1659 Name = Name.substr(5);
1661 bool IsX86 = Name.startswith("x86.");
1662 if (IsX86)
1663 Name = Name.substr(4);
1664 bool IsNVVM = Name.startswith("nvvm.");
1665 if (IsNVVM)
1666 Name = Name.substr(5);
1668 if (IsX86 && Name.startswith("sse4a.movnt.")) {
1669 Module *M = F->getParent();
1670 SmallVector<Metadata *, 1> Elts;
1671 Elts.push_back(
1672 ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
1673 MDNode *Node = MDNode::get(C, Elts);
1675 Value *Arg0 = CI->getArgOperand(0);
1676 Value *Arg1 = CI->getArgOperand(1);
1678 // Nontemporal (unaligned) store of the 0'th element of the float/double
1679 // vector.
1680 Type *SrcEltTy = cast<VectorType>(Arg1->getType())->getElementType();
1681 PointerType *EltPtrTy = PointerType::getUnqual(SrcEltTy);
1682 Value *Addr = Builder.CreateBitCast(Arg0, EltPtrTy, "cast");
1683 Value *Extract =
1684 Builder.CreateExtractElement(Arg1, (uint64_t)0, "extractelement");
1686 StoreInst *SI = Builder.CreateAlignedStore(Extract, Addr, 1);
1687 SI->setMetadata(M->getMDKindID("nontemporal"), Node);
1689 // Remove intrinsic.
1690 CI->eraseFromParent();
1691 return;
1694 if (IsX86 && (Name.startswith("avx.movnt.") ||
1695 Name.startswith("avx512.storent."))) {
1696 Module *M = F->getParent();
1697 SmallVector<Metadata *, 1> Elts;
1698 Elts.push_back(
1699 ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
1700 MDNode *Node = MDNode::get(C, Elts);
1702 Value *Arg0 = CI->getArgOperand(0);
1703 Value *Arg1 = CI->getArgOperand(1);
1705 // Convert the type of the pointer to a pointer to the stored type.
1706 Value *BC = Builder.CreateBitCast(Arg0,
1707 PointerType::getUnqual(Arg1->getType()),
1708 "cast");
1709 VectorType *VTy = cast<VectorType>(Arg1->getType());
1710 StoreInst *SI = Builder.CreateAlignedStore(Arg1, BC,
1711 VTy->getBitWidth() / 8);
1712 SI->setMetadata(M->getMDKindID("nontemporal"), Node);
1714 // Remove intrinsic.
1715 CI->eraseFromParent();
1716 return;
1719 if (IsX86 && Name == "sse2.storel.dq") {
1720 Value *Arg0 = CI->getArgOperand(0);
1721 Value *Arg1 = CI->getArgOperand(1);
1723 Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
1724 Value *BC0 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
1725 Value *Elt = Builder.CreateExtractElement(BC0, (uint64_t)0);
1726 Value *BC = Builder.CreateBitCast(Arg0,
1727 PointerType::getUnqual(Elt->getType()),
1728 "cast");
1729 Builder.CreateAlignedStore(Elt, BC, 1);
1731 // Remove intrinsic.
1732 CI->eraseFromParent();
1733 return;
1736 if (IsX86 && (Name.startswith("sse.storeu.") ||
1737 Name.startswith("sse2.storeu.") ||
1738 Name.startswith("avx.storeu."))) {
1739 Value *Arg0 = CI->getArgOperand(0);
1740 Value *Arg1 = CI->getArgOperand(1);
1742 Arg0 = Builder.CreateBitCast(Arg0,
1743 PointerType::getUnqual(Arg1->getType()),
1744 "cast");
1745 Builder.CreateAlignedStore(Arg1, Arg0, 1);
1747 // Remove intrinsic.
1748 CI->eraseFromParent();
1749 return;
1752 if (IsX86 && Name == "avx512.mask.store.ss") {
1753 Value *Mask = Builder.CreateAnd(CI->getArgOperand(2), Builder.getInt8(1));
1754 UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
1755 Mask, false);
1757 // Remove intrinsic.
1758 CI->eraseFromParent();
1759 return;
1762 if (IsX86 && (Name.startswith("avx512.mask.store"))) {
1763 // "avx512.mask.storeu." or "avx512.mask.store."
1764 bool Aligned = Name[17] != 'u'; // "avx512.mask.storeu".
1765 UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
1766 CI->getArgOperand(2), Aligned);
1768 // Remove intrinsic.
1769 CI->eraseFromParent();
1770 return;
1773 Value *Rep;
1774 // Upgrade packed integer vector compare intrinsics to compare instructions.
1775 if (IsX86 && (Name.startswith("sse2.pcmp") ||
1776 Name.startswith("avx2.pcmp"))) {
1777 // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt."
1778 bool CmpEq = Name[9] == 'e';
1779 Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
1780 CI->getArgOperand(0), CI->getArgOperand(1));
1781 Rep = Builder.CreateSExt(Rep, CI->getType(), "");
1782 } else if (IsX86 && (Name.startswith("avx512.broadcastm"))) {
1783 Type *ExtTy = Type::getInt32Ty(C);
1784 if (CI->getOperand(0)->getType()->isIntegerTy(8))
1785 ExtTy = Type::getInt64Ty(C);
1786 unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() /
1787 ExtTy->getPrimitiveSizeInBits();
1788 Rep = Builder.CreateZExt(CI->getArgOperand(0), ExtTy);
1789 Rep = Builder.CreateVectorSplat(NumElts, Rep);
1790 } else if (IsX86 && (Name == "sse.sqrt.ss" ||
1791 Name == "sse2.sqrt.sd")) {
1792 Value *Vec = CI->getArgOperand(0);
1793 Value *Elt0 = Builder.CreateExtractElement(Vec, (uint64_t)0);
1794 Function *Intr = Intrinsic::getDeclaration(F->getParent(),
1795 Intrinsic::sqrt, Elt0->getType());
1796 Elt0 = Builder.CreateCall(Intr, Elt0);
1797 Rep = Builder.CreateInsertElement(Vec, Elt0, (uint64_t)0);
1798 } else if (IsX86 && (Name.startswith("avx.sqrt.p") ||
1799 Name.startswith("sse2.sqrt.p") ||
1800 Name.startswith("sse.sqrt.p"))) {
1801 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
1802 Intrinsic::sqrt,
1803 CI->getType()),
1804 {CI->getArgOperand(0)});
1805 } else if (IsX86 && (Name.startswith("avx512.mask.sqrt.p"))) {
1806 if (CI->getNumArgOperands() == 4 &&
1807 (!isa<ConstantInt>(CI->getArgOperand(3)) ||
1808 cast<ConstantInt>(CI->getArgOperand(3))->getZExtValue() != 4)) {
1809 Intrinsic::ID IID = Name[18] == 's' ? Intrinsic::x86_avx512_sqrt_ps_512
1810 : Intrinsic::x86_avx512_sqrt_pd_512;
1812 Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(3) };
1813 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(),
1814 IID), Args);
1815 } else {
1816 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
1817 Intrinsic::sqrt,
1818 CI->getType()),
1819 {CI->getArgOperand(0)});
1821 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1822 CI->getArgOperand(1));
1823 } else if (IsX86 && (Name.startswith("avx512.ptestm") ||
1824 Name.startswith("avx512.ptestnm"))) {
1825 Value *Op0 = CI->getArgOperand(0);
1826 Value *Op1 = CI->getArgOperand(1);
1827 Value *Mask = CI->getArgOperand(2);
1828 Rep = Builder.CreateAnd(Op0, Op1);
1829 llvm::Type *Ty = Op0->getType();
1830 Value *Zero = llvm::Constant::getNullValue(Ty);
1831 ICmpInst::Predicate Pred =
1832 Name.startswith("avx512.ptestm") ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ;
1833 Rep = Builder.CreateICmp(Pred, Rep, Zero);
1834 Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, Mask);
1835 } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))){
1836 unsigned NumElts =
1837 CI->getArgOperand(1)->getType()->getVectorNumElements();
1838 Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
1839 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1840 CI->getArgOperand(1));
1841 } else if (IsX86 && (Name.startswith("avx512.kunpck"))) {
1842 unsigned NumElts = CI->getType()->getScalarSizeInBits();
1843 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), NumElts);
1844 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), NumElts);
1845 uint32_t Indices[64];
1846 for (unsigned i = 0; i != NumElts; ++i)
1847 Indices[i] = i;
1849 // First extract half of each vector. This gives better codegen than
1850 // doing it in a single shuffle.
1851 LHS = Builder.CreateShuffleVector(LHS, LHS,
1852 makeArrayRef(Indices, NumElts / 2));
1853 RHS = Builder.CreateShuffleVector(RHS, RHS,
1854 makeArrayRef(Indices, NumElts / 2));
1855 // Concat the vectors.
1856 // NOTE: Operands have to be swapped to match intrinsic definition.
1857 Rep = Builder.CreateShuffleVector(RHS, LHS,
1858 makeArrayRef(Indices, NumElts));
1859 Rep = Builder.CreateBitCast(Rep, CI->getType());
1860 } else if (IsX86 && Name == "avx512.kand.w") {
1861 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1862 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1863 Rep = Builder.CreateAnd(LHS, RHS);
1864 Rep = Builder.CreateBitCast(Rep, CI->getType());
1865 } else if (IsX86 && Name == "avx512.kandn.w") {
1866 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1867 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1868 LHS = Builder.CreateNot(LHS);
1869 Rep = Builder.CreateAnd(LHS, RHS);
1870 Rep = Builder.CreateBitCast(Rep, CI->getType());
1871 } else if (IsX86 && Name == "avx512.kor.w") {
1872 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1873 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1874 Rep = Builder.CreateOr(LHS, RHS);
1875 Rep = Builder.CreateBitCast(Rep, CI->getType());
1876 } else if (IsX86 && Name == "avx512.kxor.w") {
1877 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1878 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1879 Rep = Builder.CreateXor(LHS, RHS);
1880 Rep = Builder.CreateBitCast(Rep, CI->getType());
1881 } else if (IsX86 && Name == "avx512.kxnor.w") {
1882 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1883 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1884 LHS = Builder.CreateNot(LHS);
1885 Rep = Builder.CreateXor(LHS, RHS);
1886 Rep = Builder.CreateBitCast(Rep, CI->getType());
1887 } else if (IsX86 && Name == "avx512.knot.w") {
1888 Rep = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1889 Rep = Builder.CreateNot(Rep);
1890 Rep = Builder.CreateBitCast(Rep, CI->getType());
1891 } else if (IsX86 &&
1892 (Name == "avx512.kortestz.w" || Name == "avx512.kortestc.w")) {
1893 Value *LHS = getX86MaskVec(Builder, CI->getArgOperand(0), 16);
1894 Value *RHS = getX86MaskVec(Builder, CI->getArgOperand(1), 16);
1895 Rep = Builder.CreateOr(LHS, RHS);
1896 Rep = Builder.CreateBitCast(Rep, Builder.getInt16Ty());
1897 Value *C;
1898 if (Name[14] == 'c')
1899 C = ConstantInt::getAllOnesValue(Builder.getInt16Ty());
1900 else
1901 C = ConstantInt::getNullValue(Builder.getInt16Ty());
1902 Rep = Builder.CreateICmpEQ(Rep, C);
1903 Rep = Builder.CreateZExt(Rep, Builder.getInt32Ty());
1904 } else if (IsX86 && (Name == "sse.add.ss" || Name == "sse2.add.sd" ||
1905 Name == "sse.sub.ss" || Name == "sse2.sub.sd" ||
1906 Name == "sse.mul.ss" || Name == "sse2.mul.sd" ||
1907 Name == "sse.div.ss" || Name == "sse2.div.sd")) {
1908 Type *I32Ty = Type::getInt32Ty(C);
1909 Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1910 ConstantInt::get(I32Ty, 0));
1911 Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1912 ConstantInt::get(I32Ty, 0));
1913 Value *EltOp;
1914 if (Name.contains(".add."))
1915 EltOp = Builder.CreateFAdd(Elt0, Elt1);
1916 else if (Name.contains(".sub."))
1917 EltOp = Builder.CreateFSub(Elt0, Elt1);
1918 else if (Name.contains(".mul."))
1919 EltOp = Builder.CreateFMul(Elt0, Elt1);
1920 else
1921 EltOp = Builder.CreateFDiv(Elt0, Elt1);
1922 Rep = Builder.CreateInsertElement(CI->getArgOperand(0), EltOp,
1923 ConstantInt::get(I32Ty, 0));
1924 } else if (IsX86 && Name.startswith("avx512.mask.pcmp")) {
1925 // "avx512.mask.pcmpeq." or "avx512.mask.pcmpgt."
1926 bool CmpEq = Name[16] == 'e';
1927 Rep = upgradeMaskedCompare(Builder, *CI, CmpEq ? 0 : 6, true);
1928 } else if (IsX86 && Name.startswith("avx512.mask.vpshufbitqmb.")) {
1929 Type *OpTy = CI->getArgOperand(0)->getType();
1930 unsigned VecWidth = OpTy->getPrimitiveSizeInBits();
1931 Intrinsic::ID IID;
1932 switch (VecWidth) {
1933 default: llvm_unreachable("Unexpected intrinsic");
1934 case 128: IID = Intrinsic::x86_avx512_vpshufbitqmb_128; break;
1935 case 256: IID = Intrinsic::x86_avx512_vpshufbitqmb_256; break;
1936 case 512: IID = Intrinsic::x86_avx512_vpshufbitqmb_512; break;
1939 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1940 { CI->getOperand(0), CI->getArgOperand(1) });
1941 Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, CI->getArgOperand(2));
1942 } else if (IsX86 && Name.startswith("avx512.mask.fpclass.p")) {
1943 Type *OpTy = CI->getArgOperand(0)->getType();
1944 unsigned VecWidth = OpTy->getPrimitiveSizeInBits();
1945 unsigned EltWidth = OpTy->getScalarSizeInBits();
1946 Intrinsic::ID IID;
1947 if (VecWidth == 128 && EltWidth == 32)
1948 IID = Intrinsic::x86_avx512_fpclass_ps_128;
1949 else if (VecWidth == 256 && EltWidth == 32)
1950 IID = Intrinsic::x86_avx512_fpclass_ps_256;
1951 else if (VecWidth == 512 && EltWidth == 32)
1952 IID = Intrinsic::x86_avx512_fpclass_ps_512;
1953 else if (VecWidth == 128 && EltWidth == 64)
1954 IID = Intrinsic::x86_avx512_fpclass_pd_128;
1955 else if (VecWidth == 256 && EltWidth == 64)
1956 IID = Intrinsic::x86_avx512_fpclass_pd_256;
1957 else if (VecWidth == 512 && EltWidth == 64)
1958 IID = Intrinsic::x86_avx512_fpclass_pd_512;
1959 else
1960 llvm_unreachable("Unexpected intrinsic");
1962 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1963 { CI->getOperand(0), CI->getArgOperand(1) });
1964 Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, CI->getArgOperand(2));
1965 } else if (IsX86 && Name.startswith("avx512.mask.cmp.p")) {
1966 Type *OpTy = CI->getArgOperand(0)->getType();
1967 unsigned VecWidth = OpTy->getPrimitiveSizeInBits();
1968 unsigned EltWidth = OpTy->getScalarSizeInBits();
1969 Intrinsic::ID IID;
1970 if (VecWidth == 128 && EltWidth == 32)
1971 IID = Intrinsic::x86_avx512_cmp_ps_128;
1972 else if (VecWidth == 256 && EltWidth == 32)
1973 IID = Intrinsic::x86_avx512_cmp_ps_256;
1974 else if (VecWidth == 512 && EltWidth == 32)
1975 IID = Intrinsic::x86_avx512_cmp_ps_512;
1976 else if (VecWidth == 128 && EltWidth == 64)
1977 IID = Intrinsic::x86_avx512_cmp_pd_128;
1978 else if (VecWidth == 256 && EltWidth == 64)
1979 IID = Intrinsic::x86_avx512_cmp_pd_256;
1980 else if (VecWidth == 512 && EltWidth == 64)
1981 IID = Intrinsic::x86_avx512_cmp_pd_512;
1982 else
1983 llvm_unreachable("Unexpected intrinsic");
1985 SmallVector<Value *, 4> Args;
1986 Args.push_back(CI->getArgOperand(0));
1987 Args.push_back(CI->getArgOperand(1));
1988 Args.push_back(CI->getArgOperand(2));
1989 if (CI->getNumArgOperands() == 5)
1990 Args.push_back(CI->getArgOperand(4));
1992 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1993 Args);
1994 Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, CI->getArgOperand(3));
1995 } else if (IsX86 && Name.startswith("avx512.mask.cmp.") &&
1996 Name[16] != 'p') {
1997 // Integer compare intrinsics.
1998 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1999 Rep = upgradeMaskedCompare(Builder, *CI, Imm, true);
2000 } else if (IsX86 && Name.startswith("avx512.mask.ucmp.")) {
2001 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2002 Rep = upgradeMaskedCompare(Builder, *CI, Imm, false);
2003 } else if (IsX86 && (Name.startswith("avx512.cvtb2mask.") ||
2004 Name.startswith("avx512.cvtw2mask.") ||
2005 Name.startswith("avx512.cvtd2mask.") ||
2006 Name.startswith("avx512.cvtq2mask."))) {
2007 Value *Op = CI->getArgOperand(0);
2008 Value *Zero = llvm::Constant::getNullValue(Op->getType());
2009 Rep = Builder.CreateICmp(ICmpInst::ICMP_SLT, Op, Zero);
2010 Rep = ApplyX86MaskOn1BitsVec(Builder, Rep, nullptr);
2011 } else if(IsX86 && (Name == "ssse3.pabs.b.128" ||
2012 Name == "ssse3.pabs.w.128" ||
2013 Name == "ssse3.pabs.d.128" ||
2014 Name.startswith("avx2.pabs") ||
2015 Name.startswith("avx512.mask.pabs"))) {
2016 Rep = upgradeAbs(Builder, *CI);
2017 } else if (IsX86 && (Name == "sse41.pmaxsb" ||
2018 Name == "sse2.pmaxs.w" ||
2019 Name == "sse41.pmaxsd" ||
2020 Name.startswith("avx2.pmaxs") ||
2021 Name.startswith("avx512.mask.pmaxs"))) {
2022 Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SGT);
2023 } else if (IsX86 && (Name == "sse2.pmaxu.b" ||
2024 Name == "sse41.pmaxuw" ||
2025 Name == "sse41.pmaxud" ||
2026 Name.startswith("avx2.pmaxu") ||
2027 Name.startswith("avx512.mask.pmaxu"))) {
2028 Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_UGT);
2029 } else if (IsX86 && (Name == "sse41.pminsb" ||
2030 Name == "sse2.pmins.w" ||
2031 Name == "sse41.pminsd" ||
2032 Name.startswith("avx2.pmins") ||
2033 Name.startswith("avx512.mask.pmins"))) {
2034 Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SLT);
2035 } else if (IsX86 && (Name == "sse2.pminu.b" ||
2036 Name == "sse41.pminuw" ||
2037 Name == "sse41.pminud" ||
2038 Name.startswith("avx2.pminu") ||
2039 Name.startswith("avx512.mask.pminu"))) {
2040 Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_ULT);
2041 } else if (IsX86 && (Name == "sse2.pmulu.dq" ||
2042 Name == "avx2.pmulu.dq" ||
2043 Name == "avx512.pmulu.dq.512" ||
2044 Name.startswith("avx512.mask.pmulu.dq."))) {
2045 Rep = upgradePMULDQ(Builder, *CI, /*Signed*/false);
2046 } else if (IsX86 && (Name == "sse41.pmuldq" ||
2047 Name == "avx2.pmul.dq" ||
2048 Name == "avx512.pmul.dq.512" ||
2049 Name.startswith("avx512.mask.pmul.dq."))) {
2050 Rep = upgradePMULDQ(Builder, *CI, /*Signed*/true);
2051 } else if (IsX86 && (Name == "sse.cvtsi2ss" ||
2052 Name == "sse2.cvtsi2sd" ||
2053 Name == "sse.cvtsi642ss" ||
2054 Name == "sse2.cvtsi642sd")) {
2055 Rep = Builder.CreateSIToFP(CI->getArgOperand(1),
2056 CI->getType()->getVectorElementType());
2057 Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep, (uint64_t)0);
2058 } else if (IsX86 && Name == "avx512.cvtusi2sd") {
2059 Rep = Builder.CreateUIToFP(CI->getArgOperand(1),
2060 CI->getType()->getVectorElementType());
2061 Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep, (uint64_t)0);
2062 } else if (IsX86 && Name == "sse2.cvtss2sd") {
2063 Rep = Builder.CreateExtractElement(CI->getArgOperand(1), (uint64_t)0);
2064 Rep = Builder.CreateFPExt(Rep, CI->getType()->getVectorElementType());
2065 Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep, (uint64_t)0);
2066 } else if (IsX86 && (Name == "sse2.cvtdq2pd" ||
2067 Name == "sse2.cvtdq2ps" ||
2068 Name == "avx.cvtdq2.pd.256" ||
2069 Name == "avx.cvtdq2.ps.256" ||
2070 Name.startswith("avx512.mask.cvtdq2pd.") ||
2071 Name.startswith("avx512.mask.cvtudq2pd.") ||
2072 Name.startswith("avx512.mask.cvtdq2ps.") ||
2073 Name.startswith("avx512.mask.cvtudq2ps.") ||
2074 Name.startswith("avx512.mask.cvtqq2pd.") ||
2075 Name.startswith("avx512.mask.cvtuqq2pd.") ||
2076 Name == "avx512.mask.cvtqq2ps.256" ||
2077 Name == "avx512.mask.cvtqq2ps.512" ||
2078 Name == "avx512.mask.cvtuqq2ps.256" ||
2079 Name == "avx512.mask.cvtuqq2ps.512" ||
2080 Name == "sse2.cvtps2pd" ||
2081 Name == "avx.cvt.ps2.pd.256" ||
2082 Name == "avx512.mask.cvtps2pd.128" ||
2083 Name == "avx512.mask.cvtps2pd.256")) {
2084 Type *DstTy = CI->getType();
2085 Rep = CI->getArgOperand(0);
2086 Type *SrcTy = Rep->getType();
2088 unsigned NumDstElts = DstTy->getVectorNumElements();
2089 if (NumDstElts < SrcTy->getVectorNumElements()) {
2090 assert(NumDstElts == 2 && "Unexpected vector size");
2091 uint32_t ShuffleMask[2] = { 0, 1 };
2092 Rep = Builder.CreateShuffleVector(Rep, Rep, ShuffleMask);
2095 bool IsPS2PD = SrcTy->getVectorElementType()->isFloatTy();
2096 bool IsUnsigned = (StringRef::npos != Name.find("cvtu"));
2097 if (IsPS2PD)
2098 Rep = Builder.CreateFPExt(Rep, DstTy, "cvtps2pd");
2099 else if (CI->getNumArgOperands() == 4 &&
2100 (!isa<ConstantInt>(CI->getArgOperand(3)) ||
2101 cast<ConstantInt>(CI->getArgOperand(3))->getZExtValue() != 4)) {
2102 Intrinsic::ID IID = IsUnsigned ? Intrinsic::x86_avx512_uitofp_round
2103 : Intrinsic::x86_avx512_sitofp_round;
2104 Function *F = Intrinsic::getDeclaration(CI->getModule(), IID,
2105 { DstTy, SrcTy });
2106 Rep = Builder.CreateCall(F, { Rep, CI->getArgOperand(3) });
2107 } else {
2108 Rep = IsUnsigned ? Builder.CreateUIToFP(Rep, DstTy, "cvt")
2109 : Builder.CreateSIToFP(Rep, DstTy, "cvt");
2112 if (CI->getNumArgOperands() >= 3)
2113 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2114 CI->getArgOperand(1));
2115 } else if (IsX86 && (Name.startswith("avx512.mask.loadu."))) {
2116 Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
2117 CI->getArgOperand(1), CI->getArgOperand(2),
2118 /*Aligned*/false);
2119 } else if (IsX86 && (Name.startswith("avx512.mask.load."))) {
2120 Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
2121 CI->getArgOperand(1),CI->getArgOperand(2),
2122 /*Aligned*/true);
2123 } else if (IsX86 && Name.startswith("avx512.mask.expand.load.")) {
2124 Type *ResultTy = CI->getType();
2125 Type *PtrTy = ResultTy->getVectorElementType();
2127 // Cast the pointer to element type.
2128 Value *Ptr = Builder.CreateBitCast(CI->getOperand(0),
2129 llvm::PointerType::getUnqual(PtrTy));
2131 Value *MaskVec = getX86MaskVec(Builder, CI->getArgOperand(2),
2132 ResultTy->getVectorNumElements());
2134 Function *ELd = Intrinsic::getDeclaration(F->getParent(),
2135 Intrinsic::masked_expandload,
2136 ResultTy);
2137 Rep = Builder.CreateCall(ELd, { Ptr, MaskVec, CI->getOperand(1) });
2138 } else if (IsX86 && Name.startswith("avx512.mask.compress.store.")) {
2139 Type *ResultTy = CI->getArgOperand(1)->getType();
2140 Type *PtrTy = ResultTy->getVectorElementType();
2142 // Cast the pointer to element type.
2143 Value *Ptr = Builder.CreateBitCast(CI->getOperand(0),
2144 llvm::PointerType::getUnqual(PtrTy));
2146 Value *MaskVec = getX86MaskVec(Builder, CI->getArgOperand(2),
2147 ResultTy->getVectorNumElements());
2149 Function *CSt = Intrinsic::getDeclaration(F->getParent(),
2150 Intrinsic::masked_compressstore,
2151 ResultTy);
2152 Rep = Builder.CreateCall(CSt, { CI->getArgOperand(1), Ptr, MaskVec });
2153 } else if (IsX86 && (Name.startswith("avx512.mask.compress.") ||
2154 Name.startswith("avx512.mask.expand."))) {
2155 Type *ResultTy = CI->getType();
2157 Value *MaskVec = getX86MaskVec(Builder, CI->getArgOperand(2),
2158 ResultTy->getVectorNumElements());
2160 bool IsCompress = Name[12] == 'c';
2161 Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress
2162 : Intrinsic::x86_avx512_mask_expand;
2163 Function *Intr = Intrinsic::getDeclaration(F->getParent(), IID, ResultTy);
2164 Rep = Builder.CreateCall(Intr, { CI->getOperand(0), CI->getOperand(1),
2165 MaskVec });
2166 } else if (IsX86 && Name.startswith("xop.vpcom")) {
2167 bool IsSigned;
2168 if (Name.endswith("ub") || Name.endswith("uw") || Name.endswith("ud") ||
2169 Name.endswith("uq"))
2170 IsSigned = false;
2171 else if (Name.endswith("b") || Name.endswith("w") || Name.endswith("d") ||
2172 Name.endswith("q"))
2173 IsSigned = true;
2174 else
2175 llvm_unreachable("Unknown suffix");
2177 unsigned Imm;
2178 if (CI->getNumArgOperands() == 3) {
2179 Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2180 } else {
2181 Name = Name.substr(9); // strip off "xop.vpcom"
2182 if (Name.startswith("lt"))
2183 Imm = 0;
2184 else if (Name.startswith("le"))
2185 Imm = 1;
2186 else if (Name.startswith("gt"))
2187 Imm = 2;
2188 else if (Name.startswith("ge"))
2189 Imm = 3;
2190 else if (Name.startswith("eq"))
2191 Imm = 4;
2192 else if (Name.startswith("ne"))
2193 Imm = 5;
2194 else if (Name.startswith("false"))
2195 Imm = 6;
2196 else if (Name.startswith("true"))
2197 Imm = 7;
2198 else
2199 llvm_unreachable("Unknown condition");
2202 Rep = upgradeX86vpcom(Builder, *CI, Imm, IsSigned);
2203 } else if (IsX86 && Name.startswith("xop.vpcmov")) {
2204 Value *Sel = CI->getArgOperand(2);
2205 Value *NotSel = Builder.CreateNot(Sel);
2206 Value *Sel0 = Builder.CreateAnd(CI->getArgOperand(0), Sel);
2207 Value *Sel1 = Builder.CreateAnd(CI->getArgOperand(1), NotSel);
2208 Rep = Builder.CreateOr(Sel0, Sel1);
2209 } else if (IsX86 && (Name.startswith("xop.vprot") ||
2210 Name.startswith("avx512.prol") ||
2211 Name.startswith("avx512.mask.prol"))) {
2212 Rep = upgradeX86Rotate(Builder, *CI, false);
2213 } else if (IsX86 && (Name.startswith("avx512.pror") ||
2214 Name.startswith("avx512.mask.pror"))) {
2215 Rep = upgradeX86Rotate(Builder, *CI, true);
2216 } else if (IsX86 && (Name.startswith("avx512.vpshld.") ||
2217 Name.startswith("avx512.mask.vpshld") ||
2218 Name.startswith("avx512.maskz.vpshld"))) {
2219 bool ZeroMask = Name[11] == 'z';
2220 Rep = upgradeX86ConcatShift(Builder, *CI, false, ZeroMask);
2221 } else if (IsX86 && (Name.startswith("avx512.vpshrd.") ||
2222 Name.startswith("avx512.mask.vpshrd") ||
2223 Name.startswith("avx512.maskz.vpshrd"))) {
2224 bool ZeroMask = Name[11] == 'z';
2225 Rep = upgradeX86ConcatShift(Builder, *CI, true, ZeroMask);
2226 } else if (IsX86 && Name == "sse42.crc32.64.8") {
2227 Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
2228 Intrinsic::x86_sse42_crc32_32_8);
2229 Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
2230 Rep = Builder.CreateCall(CRC32, {Trunc0, CI->getArgOperand(1)});
2231 Rep = Builder.CreateZExt(Rep, CI->getType(), "");
2232 } else if (IsX86 && (Name.startswith("avx.vbroadcast.s") ||
2233 Name.startswith("avx512.vbroadcast.s"))) {
2234 // Replace broadcasts with a series of insertelements.
2235 Type *VecTy = CI->getType();
2236 Type *EltTy = VecTy->getVectorElementType();
2237 unsigned EltNum = VecTy->getVectorNumElements();
2238 Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
2239 EltTy->getPointerTo());
2240 Value *Load = Builder.CreateLoad(EltTy, Cast);
2241 Type *I32Ty = Type::getInt32Ty(C);
2242 Rep = UndefValue::get(VecTy);
2243 for (unsigned I = 0; I < EltNum; ++I)
2244 Rep = Builder.CreateInsertElement(Rep, Load,
2245 ConstantInt::get(I32Ty, I));
2246 } else if (IsX86 && (Name.startswith("sse41.pmovsx") ||
2247 Name.startswith("sse41.pmovzx") ||
2248 Name.startswith("avx2.pmovsx") ||
2249 Name.startswith("avx2.pmovzx") ||
2250 Name.startswith("avx512.mask.pmovsx") ||
2251 Name.startswith("avx512.mask.pmovzx"))) {
2252 VectorType *SrcTy = cast<VectorType>(CI->getArgOperand(0)->getType());
2253 VectorType *DstTy = cast<VectorType>(CI->getType());
2254 unsigned NumDstElts = DstTy->getNumElements();
2256 // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
2257 SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
2258 for (unsigned i = 0; i != NumDstElts; ++i)
2259 ShuffleMask[i] = i;
2261 Value *SV = Builder.CreateShuffleVector(
2262 CI->getArgOperand(0), UndefValue::get(SrcTy), ShuffleMask);
2264 bool DoSext = (StringRef::npos != Name.find("pmovsx"));
2265 Rep = DoSext ? Builder.CreateSExt(SV, DstTy)
2266 : Builder.CreateZExt(SV, DstTy);
2267 // If there are 3 arguments, it's a masked intrinsic so we need a select.
2268 if (CI->getNumArgOperands() == 3)
2269 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2270 CI->getArgOperand(1));
2271 } else if (Name == "avx512.mask.pmov.qd.256" ||
2272 Name == "avx512.mask.pmov.qd.512" ||
2273 Name == "avx512.mask.pmov.wb.256" ||
2274 Name == "avx512.mask.pmov.wb.512") {
2275 Type *Ty = CI->getArgOperand(1)->getType();
2276 Rep = Builder.CreateTrunc(CI->getArgOperand(0), Ty);
2277 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2278 CI->getArgOperand(1));
2279 } else if (IsX86 && (Name.startswith("avx.vbroadcastf128") ||
2280 Name == "avx2.vbroadcasti128")) {
2281 // Replace vbroadcastf128/vbroadcasti128 with a vector load+shuffle.
2282 Type *EltTy = CI->getType()->getVectorElementType();
2283 unsigned NumSrcElts = 128 / EltTy->getPrimitiveSizeInBits();
2284 Type *VT = VectorType::get(EltTy, NumSrcElts);
2285 Value *Op = Builder.CreatePointerCast(CI->getArgOperand(0),
2286 PointerType::getUnqual(VT));
2287 Value *Load = Builder.CreateAlignedLoad(VT, Op, 1);
2288 if (NumSrcElts == 2)
2289 Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
2290 { 0, 1, 0, 1 });
2291 else
2292 Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
2293 { 0, 1, 2, 3, 0, 1, 2, 3 });
2294 } else if (IsX86 && (Name.startswith("avx512.mask.shuf.i") ||
2295 Name.startswith("avx512.mask.shuf.f"))) {
2296 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2297 Type *VT = CI->getType();
2298 unsigned NumLanes = VT->getPrimitiveSizeInBits() / 128;
2299 unsigned NumElementsInLane = 128 / VT->getScalarSizeInBits();
2300 unsigned ControlBitsMask = NumLanes - 1;
2301 unsigned NumControlBits = NumLanes / 2;
2302 SmallVector<uint32_t, 8> ShuffleMask(0);
2304 for (unsigned l = 0; l != NumLanes; ++l) {
2305 unsigned LaneMask = (Imm >> (l * NumControlBits)) & ControlBitsMask;
2306 // We actually need the other source.
2307 if (l >= NumLanes / 2)
2308 LaneMask += NumLanes;
2309 for (unsigned i = 0; i != NumElementsInLane; ++i)
2310 ShuffleMask.push_back(LaneMask * NumElementsInLane + i);
2312 Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
2313 CI->getArgOperand(1), ShuffleMask);
2314 Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
2315 CI->getArgOperand(3));
2316 }else if (IsX86 && (Name.startswith("avx512.mask.broadcastf") ||
2317 Name.startswith("avx512.mask.broadcasti"))) {
2318 unsigned NumSrcElts =
2319 CI->getArgOperand(0)->getType()->getVectorNumElements();
2320 unsigned NumDstElts = CI->getType()->getVectorNumElements();
2322 SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
2323 for (unsigned i = 0; i != NumDstElts; ++i)
2324 ShuffleMask[i] = i % NumSrcElts;
2326 Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
2327 CI->getArgOperand(0),
2328 ShuffleMask);
2329 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2330 CI->getArgOperand(1));
2331 } else if (IsX86 && (Name.startswith("avx2.pbroadcast") ||
2332 Name.startswith("avx2.vbroadcast") ||
2333 Name.startswith("avx512.pbroadcast") ||
2334 Name.startswith("avx512.mask.broadcast.s"))) {
2335 // Replace vp?broadcasts with a vector shuffle.
2336 Value *Op = CI->getArgOperand(0);
2337 unsigned NumElts = CI->getType()->getVectorNumElements();
2338 Type *MaskTy = VectorType::get(Type::getInt32Ty(C), NumElts);
2339 Rep = Builder.CreateShuffleVector(Op, UndefValue::get(Op->getType()),
2340 Constant::getNullValue(MaskTy));
2342 if (CI->getNumArgOperands() == 3)
2343 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2344 CI->getArgOperand(1));
2345 } else if (IsX86 && (Name.startswith("sse2.padds.") ||
2346 Name.startswith("sse2.psubs.") ||
2347 Name.startswith("avx2.padds.") ||
2348 Name.startswith("avx2.psubs.") ||
2349 Name.startswith("avx512.padds.") ||
2350 Name.startswith("avx512.psubs.") ||
2351 Name.startswith("avx512.mask.padds.") ||
2352 Name.startswith("avx512.mask.psubs."))) {
2353 bool IsAdd = Name.contains(".padds");
2354 Rep = UpgradeX86AddSubSatIntrinsics(Builder, *CI, true, IsAdd);
2355 } else if (IsX86 && (Name.startswith("sse2.paddus.") ||
2356 Name.startswith("sse2.psubus.") ||
2357 Name.startswith("avx2.paddus.") ||
2358 Name.startswith("avx2.psubus.") ||
2359 Name.startswith("avx512.mask.paddus.") ||
2360 Name.startswith("avx512.mask.psubus."))) {
2361 bool IsAdd = Name.contains(".paddus");
2362 Rep = UpgradeX86AddSubSatIntrinsics(Builder, *CI, false, IsAdd);
2363 } else if (IsX86 && Name.startswith("avx512.mask.palignr.")) {
2364 Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
2365 CI->getArgOperand(1),
2366 CI->getArgOperand(2),
2367 CI->getArgOperand(3),
2368 CI->getArgOperand(4),
2369 false);
2370 } else if (IsX86 && Name.startswith("avx512.mask.valign.")) {
2371 Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
2372 CI->getArgOperand(1),
2373 CI->getArgOperand(2),
2374 CI->getArgOperand(3),
2375 CI->getArgOperand(4),
2376 true);
2377 } else if (IsX86 && (Name == "sse2.psll.dq" ||
2378 Name == "avx2.psll.dq")) {
2379 // 128/256-bit shift left specified in bits.
2380 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2381 Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0),
2382 Shift / 8); // Shift is in bits.
2383 } else if (IsX86 && (Name == "sse2.psrl.dq" ||
2384 Name == "avx2.psrl.dq")) {
2385 // 128/256-bit shift right specified in bits.
2386 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2387 Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0),
2388 Shift / 8); // Shift is in bits.
2389 } else if (IsX86 && (Name == "sse2.psll.dq.bs" ||
2390 Name == "avx2.psll.dq.bs" ||
2391 Name == "avx512.psll.dq.512")) {
2392 // 128/256/512-bit shift left specified in bytes.
2393 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2394 Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
2395 } else if (IsX86 && (Name == "sse2.psrl.dq.bs" ||
2396 Name == "avx2.psrl.dq.bs" ||
2397 Name == "avx512.psrl.dq.512")) {
2398 // 128/256/512-bit shift right specified in bytes.
2399 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2400 Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
2401 } else if (IsX86 && (Name == "sse41.pblendw" ||
2402 Name.startswith("sse41.blendp") ||
2403 Name.startswith("avx.blend.p") ||
2404 Name == "avx2.pblendw" ||
2405 Name.startswith("avx2.pblendd."))) {
2406 Value *Op0 = CI->getArgOperand(0);
2407 Value *Op1 = CI->getArgOperand(1);
2408 unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2409 VectorType *VecTy = cast<VectorType>(CI->getType());
2410 unsigned NumElts = VecTy->getNumElements();
2412 SmallVector<uint32_t, 16> Idxs(NumElts);
2413 for (unsigned i = 0; i != NumElts; ++i)
2414 Idxs[i] = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
2416 Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
2417 } else if (IsX86 && (Name.startswith("avx.vinsertf128.") ||
2418 Name == "avx2.vinserti128" ||
2419 Name.startswith("avx512.mask.insert"))) {
2420 Value *Op0 = CI->getArgOperand(0);
2421 Value *Op1 = CI->getArgOperand(1);
2422 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2423 unsigned DstNumElts = CI->getType()->getVectorNumElements();
2424 unsigned SrcNumElts = Op1->getType()->getVectorNumElements();
2425 unsigned Scale = DstNumElts / SrcNumElts;
2427 // Mask off the high bits of the immediate value; hardware ignores those.
2428 Imm = Imm % Scale;
2430 // Extend the second operand into a vector the size of the destination.
2431 Value *UndefV = UndefValue::get(Op1->getType());
2432 SmallVector<uint32_t, 8> Idxs(DstNumElts);
2433 for (unsigned i = 0; i != SrcNumElts; ++i)
2434 Idxs[i] = i;
2435 for (unsigned i = SrcNumElts; i != DstNumElts; ++i)
2436 Idxs[i] = SrcNumElts;
2437 Rep = Builder.CreateShuffleVector(Op1, UndefV, Idxs);
2439 // Insert the second operand into the first operand.
2441 // Note that there is no guarantee that instruction lowering will actually
2442 // produce a vinsertf128 instruction for the created shuffles. In
2443 // particular, the 0 immediate case involves no lane changes, so it can
2444 // be handled as a blend.
2446 // Example of shuffle mask for 32-bit elements:
2447 // Imm = 1 <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>
2448 // Imm = 0 <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7 >
2450 // First fill with identify mask.
2451 for (unsigned i = 0; i != DstNumElts; ++i)
2452 Idxs[i] = i;
2453 // Then replace the elements where we need to insert.
2454 for (unsigned i = 0; i != SrcNumElts; ++i)
2455 Idxs[i + Imm * SrcNumElts] = i + DstNumElts;
2456 Rep = Builder.CreateShuffleVector(Op0, Rep, Idxs);
2458 // If the intrinsic has a mask operand, handle that.
2459 if (CI->getNumArgOperands() == 5)
2460 Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
2461 CI->getArgOperand(3));
2462 } else if (IsX86 && (Name.startswith("avx.vextractf128.") ||
2463 Name == "avx2.vextracti128" ||
2464 Name.startswith("avx512.mask.vextract"))) {
2465 Value *Op0 = CI->getArgOperand(0);
2466 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2467 unsigned DstNumElts = CI->getType()->getVectorNumElements();
2468 unsigned SrcNumElts = Op0->getType()->getVectorNumElements();
2469 unsigned Scale = SrcNumElts / DstNumElts;
2471 // Mask off the high bits of the immediate value; hardware ignores those.
2472 Imm = Imm % Scale;
2474 // Get indexes for the subvector of the input vector.
2475 SmallVector<uint32_t, 8> Idxs(DstNumElts);
2476 for (unsigned i = 0; i != DstNumElts; ++i) {
2477 Idxs[i] = i + (Imm * DstNumElts);
2479 Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2481 // If the intrinsic has a mask operand, handle that.
2482 if (CI->getNumArgOperands() == 4)
2483 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2484 CI->getArgOperand(2));
2485 } else if (!IsX86 && Name == "stackprotectorcheck") {
2486 Rep = nullptr;
2487 } else if (IsX86 && (Name.startswith("avx512.mask.perm.df.") ||
2488 Name.startswith("avx512.mask.perm.di."))) {
2489 Value *Op0 = CI->getArgOperand(0);
2490 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2491 VectorType *VecTy = cast<VectorType>(CI->getType());
2492 unsigned NumElts = VecTy->getNumElements();
2494 SmallVector<uint32_t, 8> Idxs(NumElts);
2495 for (unsigned i = 0; i != NumElts; ++i)
2496 Idxs[i] = (i & ~0x3) + ((Imm >> (2 * (i & 0x3))) & 3);
2498 Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2500 if (CI->getNumArgOperands() == 4)
2501 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2502 CI->getArgOperand(2));
2503 } else if (IsX86 && (Name.startswith("avx.vperm2f128.") ||
2504 Name == "avx2.vperm2i128")) {
2505 // The immediate permute control byte looks like this:
2506 // [1:0] - select 128 bits from sources for low half of destination
2507 // [2] - ignore
2508 // [3] - zero low half of destination
2509 // [5:4] - select 128 bits from sources for high half of destination
2510 // [6] - ignore
2511 // [7] - zero high half of destination
2513 uint8_t Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2515 unsigned NumElts = CI->getType()->getVectorNumElements();
2516 unsigned HalfSize = NumElts / 2;
2517 SmallVector<uint32_t, 8> ShuffleMask(NumElts);
2519 // Determine which operand(s) are actually in use for this instruction.
2520 Value *V0 = (Imm & 0x02) ? CI->getArgOperand(1) : CI->getArgOperand(0);
2521 Value *V1 = (Imm & 0x20) ? CI->getArgOperand(1) : CI->getArgOperand(0);
2523 // If needed, replace operands based on zero mask.
2524 V0 = (Imm & 0x08) ? ConstantAggregateZero::get(CI->getType()) : V0;
2525 V1 = (Imm & 0x80) ? ConstantAggregateZero::get(CI->getType()) : V1;
2527 // Permute low half of result.
2528 unsigned StartIndex = (Imm & 0x01) ? HalfSize : 0;
2529 for (unsigned i = 0; i < HalfSize; ++i)
2530 ShuffleMask[i] = StartIndex + i;
2532 // Permute high half of result.
2533 StartIndex = (Imm & 0x10) ? HalfSize : 0;
2534 for (unsigned i = 0; i < HalfSize; ++i)
2535 ShuffleMask[i + HalfSize] = NumElts + StartIndex + i;
2537 Rep = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
2539 } else if (IsX86 && (Name.startswith("avx.vpermil.") ||
2540 Name == "sse2.pshuf.d" ||
2541 Name.startswith("avx512.mask.vpermil.p") ||
2542 Name.startswith("avx512.mask.pshuf.d."))) {
2543 Value *Op0 = CI->getArgOperand(0);
2544 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2545 VectorType *VecTy = cast<VectorType>(CI->getType());
2546 unsigned NumElts = VecTy->getNumElements();
2547 // Calculate the size of each index in the immediate.
2548 unsigned IdxSize = 64 / VecTy->getScalarSizeInBits();
2549 unsigned IdxMask = ((1 << IdxSize) - 1);
2551 SmallVector<uint32_t, 8> Idxs(NumElts);
2552 // Lookup the bits for this element, wrapping around the immediate every
2553 // 8-bits. Elements are grouped into sets of 2 or 4 elements so we need
2554 // to offset by the first index of each group.
2555 for (unsigned i = 0; i != NumElts; ++i)
2556 Idxs[i] = ((Imm >> ((i * IdxSize) % 8)) & IdxMask) | (i & ~IdxMask);
2558 Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2560 if (CI->getNumArgOperands() == 4)
2561 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2562 CI->getArgOperand(2));
2563 } else if (IsX86 && (Name == "sse2.pshufl.w" ||
2564 Name.startswith("avx512.mask.pshufl.w."))) {
2565 Value *Op0 = CI->getArgOperand(0);
2566 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2567 unsigned NumElts = CI->getType()->getVectorNumElements();
2569 SmallVector<uint32_t, 16> Idxs(NumElts);
2570 for (unsigned l = 0; l != NumElts; l += 8) {
2571 for (unsigned i = 0; i != 4; ++i)
2572 Idxs[i + l] = ((Imm >> (2 * i)) & 0x3) + l;
2573 for (unsigned i = 4; i != 8; ++i)
2574 Idxs[i + l] = i + l;
2577 Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2579 if (CI->getNumArgOperands() == 4)
2580 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2581 CI->getArgOperand(2));
2582 } else if (IsX86 && (Name == "sse2.pshufh.w" ||
2583 Name.startswith("avx512.mask.pshufh.w."))) {
2584 Value *Op0 = CI->getArgOperand(0);
2585 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
2586 unsigned NumElts = CI->getType()->getVectorNumElements();
2588 SmallVector<uint32_t, 16> Idxs(NumElts);
2589 for (unsigned l = 0; l != NumElts; l += 8) {
2590 for (unsigned i = 0; i != 4; ++i)
2591 Idxs[i + l] = i + l;
2592 for (unsigned i = 0; i != 4; ++i)
2593 Idxs[i + l + 4] = ((Imm >> (2 * i)) & 0x3) + 4 + l;
2596 Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2598 if (CI->getNumArgOperands() == 4)
2599 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2600 CI->getArgOperand(2));
2601 } else if (IsX86 && Name.startswith("avx512.mask.shuf.p")) {
2602 Value *Op0 = CI->getArgOperand(0);
2603 Value *Op1 = CI->getArgOperand(1);
2604 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
2605 unsigned NumElts = CI->getType()->getVectorNumElements();
2607 unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
2608 unsigned HalfLaneElts = NumLaneElts / 2;
2610 SmallVector<uint32_t, 16> Idxs(NumElts);
2611 for (unsigned i = 0; i != NumElts; ++i) {
2612 // Base index is the starting element of the lane.
2613 Idxs[i] = i - (i % NumLaneElts);
2614 // If we are half way through the lane switch to the other source.
2615 if ((i % NumLaneElts) >= HalfLaneElts)
2616 Idxs[i] += NumElts;
2617 // Now select the specific element. By adding HalfLaneElts bits from
2618 // the immediate. Wrapping around the immediate every 8-bits.
2619 Idxs[i] += (Imm >> ((i * HalfLaneElts) % 8)) & ((1 << HalfLaneElts) - 1);
2622 Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
2624 Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
2625 CI->getArgOperand(3));
2626 } else if (IsX86 && (Name.startswith("avx512.mask.movddup") ||
2627 Name.startswith("avx512.mask.movshdup") ||
2628 Name.startswith("avx512.mask.movsldup"))) {
2629 Value *Op0 = CI->getArgOperand(0);
2630 unsigned NumElts = CI->getType()->getVectorNumElements();
2631 unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
2633 unsigned Offset = 0;
2634 if (Name.startswith("avx512.mask.movshdup."))
2635 Offset = 1;
2637 SmallVector<uint32_t, 16> Idxs(NumElts);
2638 for (unsigned l = 0; l != NumElts; l += NumLaneElts)
2639 for (unsigned i = 0; i != NumLaneElts; i += 2) {
2640 Idxs[i + l + 0] = i + l + Offset;
2641 Idxs[i + l + 1] = i + l + Offset;
2644 Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
2646 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2647 CI->getArgOperand(1));
2648 } else if (IsX86 && (Name.startswith("avx512.mask.punpckl") ||
2649 Name.startswith("avx512.mask.unpckl."))) {
2650 Value *Op0 = CI->getArgOperand(0);
2651 Value *Op1 = CI->getArgOperand(1);
2652 int NumElts = CI->getType()->getVectorNumElements();
2653 int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
2655 SmallVector<uint32_t, 64> Idxs(NumElts);
2656 for (int l = 0; l != NumElts; l += NumLaneElts)
2657 for (int i = 0; i != NumLaneElts; ++i)
2658 Idxs[i + l] = l + (i / 2) + NumElts * (i % 2);
2660 Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
2662 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2663 CI->getArgOperand(2));
2664 } else if (IsX86 && (Name.startswith("avx512.mask.punpckh") ||
2665 Name.startswith("avx512.mask.unpckh."))) {
2666 Value *Op0 = CI->getArgOperand(0);
2667 Value *Op1 = CI->getArgOperand(1);
2668 int NumElts = CI->getType()->getVectorNumElements();
2669 int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
2671 SmallVector<uint32_t, 64> Idxs(NumElts);
2672 for (int l = 0; l != NumElts; l += NumLaneElts)
2673 for (int i = 0; i != NumLaneElts; ++i)
2674 Idxs[i + l] = (NumLaneElts / 2) + l + (i / 2) + NumElts * (i % 2);
2676 Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
2678 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2679 CI->getArgOperand(2));
2680 } else if (IsX86 && (Name.startswith("avx512.mask.and.") ||
2681 Name.startswith("avx512.mask.pand."))) {
2682 VectorType *FTy = cast<VectorType>(CI->getType());
2683 VectorType *ITy = VectorType::getInteger(FTy);
2684 Rep = Builder.CreateAnd(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
2685 Builder.CreateBitCast(CI->getArgOperand(1), ITy));
2686 Rep = Builder.CreateBitCast(Rep, FTy);
2687 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2688 CI->getArgOperand(2));
2689 } else if (IsX86 && (Name.startswith("avx512.mask.andn.") ||
2690 Name.startswith("avx512.mask.pandn."))) {
2691 VectorType *FTy = cast<VectorType>(CI->getType());
2692 VectorType *ITy = VectorType::getInteger(FTy);
2693 Rep = Builder.CreateNot(Builder.CreateBitCast(CI->getArgOperand(0), ITy));
2694 Rep = Builder.CreateAnd(Rep,
2695 Builder.CreateBitCast(CI->getArgOperand(1), ITy));
2696 Rep = Builder.CreateBitCast(Rep, FTy);
2697 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2698 CI->getArgOperand(2));
2699 } else if (IsX86 && (Name.startswith("avx512.mask.or.") ||
2700 Name.startswith("avx512.mask.por."))) {
2701 VectorType *FTy = cast<VectorType>(CI->getType());
2702 VectorType *ITy = VectorType::getInteger(FTy);
2703 Rep = Builder.CreateOr(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
2704 Builder.CreateBitCast(CI->getArgOperand(1), ITy));
2705 Rep = Builder.CreateBitCast(Rep, FTy);
2706 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2707 CI->getArgOperand(2));
2708 } else if (IsX86 && (Name.startswith("avx512.mask.xor.") ||
2709 Name.startswith("avx512.mask.pxor."))) {
2710 VectorType *FTy = cast<VectorType>(CI->getType());
2711 VectorType *ITy = VectorType::getInteger(FTy);
2712 Rep = Builder.CreateXor(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
2713 Builder.CreateBitCast(CI->getArgOperand(1), ITy));
2714 Rep = Builder.CreateBitCast(Rep, FTy);
2715 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2716 CI->getArgOperand(2));
2717 } else if (IsX86 && Name.startswith("avx512.mask.padd.")) {
2718 Rep = Builder.CreateAdd(CI->getArgOperand(0), CI->getArgOperand(1));
2719 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2720 CI->getArgOperand(2));
2721 } else if (IsX86 && Name.startswith("avx512.mask.psub.")) {
2722 Rep = Builder.CreateSub(CI->getArgOperand(0), CI->getArgOperand(1));
2723 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2724 CI->getArgOperand(2));
2725 } else if (IsX86 && Name.startswith("avx512.mask.pmull.")) {
2726 Rep = Builder.CreateMul(CI->getArgOperand(0), CI->getArgOperand(1));
2727 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2728 CI->getArgOperand(2));
2729 } else if (IsX86 && Name.startswith("avx512.mask.add.p")) {
2730 if (Name.endswith(".512")) {
2731 Intrinsic::ID IID;
2732 if (Name[17] == 's')
2733 IID = Intrinsic::x86_avx512_add_ps_512;
2734 else
2735 IID = Intrinsic::x86_avx512_add_pd_512;
2737 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2738 { CI->getArgOperand(0), CI->getArgOperand(1),
2739 CI->getArgOperand(4) });
2740 } else {
2741 Rep = Builder.CreateFAdd(CI->getArgOperand(0), CI->getArgOperand(1));
2743 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2744 CI->getArgOperand(2));
2745 } else if (IsX86 && Name.startswith("avx512.mask.div.p")) {
2746 if (Name.endswith(".512")) {
2747 Intrinsic::ID IID;
2748 if (Name[17] == 's')
2749 IID = Intrinsic::x86_avx512_div_ps_512;
2750 else
2751 IID = Intrinsic::x86_avx512_div_pd_512;
2753 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2754 { CI->getArgOperand(0), CI->getArgOperand(1),
2755 CI->getArgOperand(4) });
2756 } else {
2757 Rep = Builder.CreateFDiv(CI->getArgOperand(0), CI->getArgOperand(1));
2759 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2760 CI->getArgOperand(2));
2761 } else if (IsX86 && Name.startswith("avx512.mask.mul.p")) {
2762 if (Name.endswith(".512")) {
2763 Intrinsic::ID IID;
2764 if (Name[17] == 's')
2765 IID = Intrinsic::x86_avx512_mul_ps_512;
2766 else
2767 IID = Intrinsic::x86_avx512_mul_pd_512;
2769 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2770 { CI->getArgOperand(0), CI->getArgOperand(1),
2771 CI->getArgOperand(4) });
2772 } else {
2773 Rep = Builder.CreateFMul(CI->getArgOperand(0), CI->getArgOperand(1));
2775 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2776 CI->getArgOperand(2));
2777 } else if (IsX86 && Name.startswith("avx512.mask.sub.p")) {
2778 if (Name.endswith(".512")) {
2779 Intrinsic::ID IID;
2780 if (Name[17] == 's')
2781 IID = Intrinsic::x86_avx512_sub_ps_512;
2782 else
2783 IID = Intrinsic::x86_avx512_sub_pd_512;
2785 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2786 { CI->getArgOperand(0), CI->getArgOperand(1),
2787 CI->getArgOperand(4) });
2788 } else {
2789 Rep = Builder.CreateFSub(CI->getArgOperand(0), CI->getArgOperand(1));
2791 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2792 CI->getArgOperand(2));
2793 } else if (IsX86 && (Name.startswith("avx512.mask.max.p") ||
2794 Name.startswith("avx512.mask.min.p")) &&
2795 Name.drop_front(18) == ".512") {
2796 bool IsDouble = Name[17] == 'd';
2797 bool IsMin = Name[13] == 'i';
2798 static const Intrinsic::ID MinMaxTbl[2][2] = {
2799 { Intrinsic::x86_avx512_max_ps_512, Intrinsic::x86_avx512_max_pd_512 },
2800 { Intrinsic::x86_avx512_min_ps_512, Intrinsic::x86_avx512_min_pd_512 }
2802 Intrinsic::ID IID = MinMaxTbl[IsMin][IsDouble];
2804 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
2805 { CI->getArgOperand(0), CI->getArgOperand(1),
2806 CI->getArgOperand(4) });
2807 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2808 CI->getArgOperand(2));
2809 } else if (IsX86 && Name.startswith("avx512.mask.lzcnt.")) {
2810 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
2811 Intrinsic::ctlz,
2812 CI->getType()),
2813 { CI->getArgOperand(0), Builder.getInt1(false) });
2814 Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
2815 CI->getArgOperand(1));
2816 } else if (IsX86 && Name.startswith("avx512.mask.psll")) {
2817 bool IsImmediate = Name[16] == 'i' ||
2818 (Name.size() > 18 && Name[18] == 'i');
2819 bool IsVariable = Name[16] == 'v';
2820 char Size = Name[16] == '.' ? Name[17] :
2821 Name[17] == '.' ? Name[18] :
2822 Name[18] == '.' ? Name[19] :
2823 Name[20];
2825 Intrinsic::ID IID;
2826 if (IsVariable && Name[17] != '.') {
2827 if (Size == 'd' && Name[17] == '2') // avx512.mask.psllv2.di
2828 IID = Intrinsic::x86_avx2_psllv_q;
2829 else if (Size == 'd' && Name[17] == '4') // avx512.mask.psllv4.di
2830 IID = Intrinsic::x86_avx2_psllv_q_256;
2831 else if (Size == 's' && Name[17] == '4') // avx512.mask.psllv4.si
2832 IID = Intrinsic::x86_avx2_psllv_d;
2833 else if (Size == 's' && Name[17] == '8') // avx512.mask.psllv8.si
2834 IID = Intrinsic::x86_avx2_psllv_d_256;
2835 else if (Size == 'h' && Name[17] == '8') // avx512.mask.psllv8.hi
2836 IID = Intrinsic::x86_avx512_psllv_w_128;
2837 else if (Size == 'h' && Name[17] == '1') // avx512.mask.psllv16.hi
2838 IID = Intrinsic::x86_avx512_psllv_w_256;
2839 else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psllv32hi
2840 IID = Intrinsic::x86_avx512_psllv_w_512;
2841 else
2842 llvm_unreachable("Unexpected size");
2843 } else if (Name.endswith(".128")) {
2844 if (Size == 'd') // avx512.mask.psll.d.128, avx512.mask.psll.di.128
2845 IID = IsImmediate ? Intrinsic::x86_sse2_pslli_d
2846 : Intrinsic::x86_sse2_psll_d;
2847 else if (Size == 'q') // avx512.mask.psll.q.128, avx512.mask.psll.qi.128
2848 IID = IsImmediate ? Intrinsic::x86_sse2_pslli_q
2849 : Intrinsic::x86_sse2_psll_q;
2850 else if (Size == 'w') // avx512.mask.psll.w.128, avx512.mask.psll.wi.128
2851 IID = IsImmediate ? Intrinsic::x86_sse2_pslli_w
2852 : Intrinsic::x86_sse2_psll_w;
2853 else
2854 llvm_unreachable("Unexpected size");
2855 } else if (Name.endswith(".256")) {
2856 if (Size == 'd') // avx512.mask.psll.d.256, avx512.mask.psll.di.256
2857 IID = IsImmediate ? Intrinsic::x86_avx2_pslli_d
2858 : Intrinsic::x86_avx2_psll_d;
2859 else if (Size == 'q') // avx512.mask.psll.q.256, avx512.mask.psll.qi.256
2860 IID = IsImmediate ? Intrinsic::x86_avx2_pslli_q
2861 : Intrinsic::x86_avx2_psll_q;
2862 else if (Size == 'w') // avx512.mask.psll.w.256, avx512.mask.psll.wi.256
2863 IID = IsImmediate ? Intrinsic::x86_avx2_pslli_w
2864 : Intrinsic::x86_avx2_psll_w;
2865 else
2866 llvm_unreachable("Unexpected size");
2867 } else {
2868 if (Size == 'd') // psll.di.512, pslli.d, psll.d, psllv.d.512
2869 IID = IsImmediate ? Intrinsic::x86_avx512_pslli_d_512 :
2870 IsVariable ? Intrinsic::x86_avx512_psllv_d_512 :
2871 Intrinsic::x86_avx512_psll_d_512;
2872 else if (Size == 'q') // psll.qi.512, pslli.q, psll.q, psllv.q.512
2873 IID = IsImmediate ? Intrinsic::x86_avx512_pslli_q_512 :
2874 IsVariable ? Intrinsic::x86_avx512_psllv_q_512 :
2875 Intrinsic::x86_avx512_psll_q_512;
2876 else if (Size == 'w') // psll.wi.512, pslli.w, psll.w
2877 IID = IsImmediate ? Intrinsic::x86_avx512_pslli_w_512
2878 : Intrinsic::x86_avx512_psll_w_512;
2879 else
2880 llvm_unreachable("Unexpected size");
2883 Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2884 } else if (IsX86 && Name.startswith("avx512.mask.psrl")) {
2885 bool IsImmediate = Name[16] == 'i' ||
2886 (Name.size() > 18 && Name[18] == 'i');
2887 bool IsVariable = Name[16] == 'v';
2888 char Size = Name[16] == '.' ? Name[17] :
2889 Name[17] == '.' ? Name[18] :
2890 Name[18] == '.' ? Name[19] :
2891 Name[20];
2893 Intrinsic::ID IID;
2894 if (IsVariable && Name[17] != '.') {
2895 if (Size == 'd' && Name[17] == '2') // avx512.mask.psrlv2.di
2896 IID = Intrinsic::x86_avx2_psrlv_q;
2897 else if (Size == 'd' && Name[17] == '4') // avx512.mask.psrlv4.di
2898 IID = Intrinsic::x86_avx2_psrlv_q_256;
2899 else if (Size == 's' && Name[17] == '4') // avx512.mask.psrlv4.si
2900 IID = Intrinsic::x86_avx2_psrlv_d;
2901 else if (Size == 's' && Name[17] == '8') // avx512.mask.psrlv8.si
2902 IID = Intrinsic::x86_avx2_psrlv_d_256;
2903 else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrlv8.hi
2904 IID = Intrinsic::x86_avx512_psrlv_w_128;
2905 else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrlv16.hi
2906 IID = Intrinsic::x86_avx512_psrlv_w_256;
2907 else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrlv32hi
2908 IID = Intrinsic::x86_avx512_psrlv_w_512;
2909 else
2910 llvm_unreachable("Unexpected size");
2911 } else if (Name.endswith(".128")) {
2912 if (Size == 'd') // avx512.mask.psrl.d.128, avx512.mask.psrl.di.128
2913 IID = IsImmediate ? Intrinsic::x86_sse2_psrli_d
2914 : Intrinsic::x86_sse2_psrl_d;
2915 else if (Size == 'q') // avx512.mask.psrl.q.128, avx512.mask.psrl.qi.128
2916 IID = IsImmediate ? Intrinsic::x86_sse2_psrli_q
2917 : Intrinsic::x86_sse2_psrl_q;
2918 else if (Size == 'w') // avx512.mask.psrl.w.128, avx512.mask.psrl.wi.128
2919 IID = IsImmediate ? Intrinsic::x86_sse2_psrli_w
2920 : Intrinsic::x86_sse2_psrl_w;
2921 else
2922 llvm_unreachable("Unexpected size");
2923 } else if (Name.endswith(".256")) {
2924 if (Size == 'd') // avx512.mask.psrl.d.256, avx512.mask.psrl.di.256
2925 IID = IsImmediate ? Intrinsic::x86_avx2_psrli_d
2926 : Intrinsic::x86_avx2_psrl_d;
2927 else if (Size == 'q') // avx512.mask.psrl.q.256, avx512.mask.psrl.qi.256
2928 IID = IsImmediate ? Intrinsic::x86_avx2_psrli_q
2929 : Intrinsic::x86_avx2_psrl_q;
2930 else if (Size == 'w') // avx512.mask.psrl.w.256, avx512.mask.psrl.wi.256
2931 IID = IsImmediate ? Intrinsic::x86_avx2_psrli_w
2932 : Intrinsic::x86_avx2_psrl_w;
2933 else
2934 llvm_unreachable("Unexpected size");
2935 } else {
2936 if (Size == 'd') // psrl.di.512, psrli.d, psrl.d, psrl.d.512
2937 IID = IsImmediate ? Intrinsic::x86_avx512_psrli_d_512 :
2938 IsVariable ? Intrinsic::x86_avx512_psrlv_d_512 :
2939 Intrinsic::x86_avx512_psrl_d_512;
2940 else if (Size == 'q') // psrl.qi.512, psrli.q, psrl.q, psrl.q.512
2941 IID = IsImmediate ? Intrinsic::x86_avx512_psrli_q_512 :
2942 IsVariable ? Intrinsic::x86_avx512_psrlv_q_512 :
2943 Intrinsic::x86_avx512_psrl_q_512;
2944 else if (Size == 'w') // psrl.wi.512, psrli.w, psrl.w)
2945 IID = IsImmediate ? Intrinsic::x86_avx512_psrli_w_512
2946 : Intrinsic::x86_avx512_psrl_w_512;
2947 else
2948 llvm_unreachable("Unexpected size");
2951 Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2952 } else if (IsX86 && Name.startswith("avx512.mask.psra")) {
2953 bool IsImmediate = Name[16] == 'i' ||
2954 (Name.size() > 18 && Name[18] == 'i');
2955 bool IsVariable = Name[16] == 'v';
2956 char Size = Name[16] == '.' ? Name[17] :
2957 Name[17] == '.' ? Name[18] :
2958 Name[18] == '.' ? Name[19] :
2959 Name[20];
2961 Intrinsic::ID IID;
2962 if (IsVariable && Name[17] != '.') {
2963 if (Size == 's' && Name[17] == '4') // avx512.mask.psrav4.si
2964 IID = Intrinsic::x86_avx2_psrav_d;
2965 else if (Size == 's' && Name[17] == '8') // avx512.mask.psrav8.si
2966 IID = Intrinsic::x86_avx2_psrav_d_256;
2967 else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrav8.hi
2968 IID = Intrinsic::x86_avx512_psrav_w_128;
2969 else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrav16.hi
2970 IID = Intrinsic::x86_avx512_psrav_w_256;
2971 else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrav32hi
2972 IID = Intrinsic::x86_avx512_psrav_w_512;
2973 else
2974 llvm_unreachable("Unexpected size");
2975 } else if (Name.endswith(".128")) {
2976 if (Size == 'd') // avx512.mask.psra.d.128, avx512.mask.psra.di.128
2977 IID = IsImmediate ? Intrinsic::x86_sse2_psrai_d
2978 : Intrinsic::x86_sse2_psra_d;
2979 else if (Size == 'q') // avx512.mask.psra.q.128, avx512.mask.psra.qi.128
2980 IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_128 :
2981 IsVariable ? Intrinsic::x86_avx512_psrav_q_128 :
2982 Intrinsic::x86_avx512_psra_q_128;
2983 else if (Size == 'w') // avx512.mask.psra.w.128, avx512.mask.psra.wi.128
2984 IID = IsImmediate ? Intrinsic::x86_sse2_psrai_w
2985 : Intrinsic::x86_sse2_psra_w;
2986 else
2987 llvm_unreachable("Unexpected size");
2988 } else if (Name.endswith(".256")) {
2989 if (Size == 'd') // avx512.mask.psra.d.256, avx512.mask.psra.di.256
2990 IID = IsImmediate ? Intrinsic::x86_avx2_psrai_d
2991 : Intrinsic::x86_avx2_psra_d;
2992 else if (Size == 'q') // avx512.mask.psra.q.256, avx512.mask.psra.qi.256
2993 IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_256 :
2994 IsVariable ? Intrinsic::x86_avx512_psrav_q_256 :
2995 Intrinsic::x86_avx512_psra_q_256;
2996 else if (Size == 'w') // avx512.mask.psra.w.256, avx512.mask.psra.wi.256
2997 IID = IsImmediate ? Intrinsic::x86_avx2_psrai_w
2998 : Intrinsic::x86_avx2_psra_w;
2999 else
3000 llvm_unreachable("Unexpected size");
3001 } else {
3002 if (Size == 'd') // psra.di.512, psrai.d, psra.d, psrav.d.512
3003 IID = IsImmediate ? Intrinsic::x86_avx512_psrai_d_512 :
3004 IsVariable ? Intrinsic::x86_avx512_psrav_d_512 :
3005 Intrinsic::x86_avx512_psra_d_512;
3006 else if (Size == 'q') // psra.qi.512, psrai.q, psra.q
3007 IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_512 :
3008 IsVariable ? Intrinsic::x86_avx512_psrav_q_512 :
3009 Intrinsic::x86_avx512_psra_q_512;
3010 else if (Size == 'w') // psra.wi.512, psrai.w, psra.w
3011 IID = IsImmediate ? Intrinsic::x86_avx512_psrai_w_512
3012 : Intrinsic::x86_avx512_psra_w_512;
3013 else
3014 llvm_unreachable("Unexpected size");
3017 Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
3018 } else if (IsX86 && Name.startswith("avx512.mask.move.s")) {
3019 Rep = upgradeMaskedMove(Builder, *CI);
3020 } else if (IsX86 && Name.startswith("avx512.cvtmask2")) {
3021 Rep = UpgradeMaskToInt(Builder, *CI);
3022 } else if (IsX86 && Name.endswith(".movntdqa")) {
3023 Module *M = F->getParent();
3024 MDNode *Node = MDNode::get(
3025 C, ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
3027 Value *Ptr = CI->getArgOperand(0);
3028 VectorType *VTy = cast<VectorType>(CI->getType());
3030 // Convert the type of the pointer to a pointer to the stored type.
3031 Value *BC =
3032 Builder.CreateBitCast(Ptr, PointerType::getUnqual(VTy), "cast");
3033 LoadInst *LI = Builder.CreateAlignedLoad(VTy, BC, VTy->getBitWidth() / 8);
3034 LI->setMetadata(M->getMDKindID("nontemporal"), Node);
3035 Rep = LI;
3036 } else if (IsX86 && (Name.startswith("fma.vfmadd.") ||
3037 Name.startswith("fma.vfmsub.") ||
3038 Name.startswith("fma.vfnmadd.") ||
3039 Name.startswith("fma.vfnmsub."))) {
3040 bool NegMul = Name[6] == 'n';
3041 bool NegAcc = NegMul ? Name[8] == 's' : Name[7] == 's';
3042 bool IsScalar = NegMul ? Name[12] == 's' : Name[11] == 's';
3044 Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3045 CI->getArgOperand(2) };
3047 if (IsScalar) {
3048 Ops[0] = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
3049 Ops[1] = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
3050 Ops[2] = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
3053 if (NegMul && !IsScalar)
3054 Ops[0] = Builder.CreateFNeg(Ops[0]);
3055 if (NegMul && IsScalar)
3056 Ops[1] = Builder.CreateFNeg(Ops[1]);
3057 if (NegAcc)
3058 Ops[2] = Builder.CreateFNeg(Ops[2]);
3060 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(),
3061 Intrinsic::fma,
3062 Ops[0]->getType()),
3063 Ops);
3065 if (IsScalar)
3066 Rep = Builder.CreateInsertElement(CI->getArgOperand(0), Rep,
3067 (uint64_t)0);
3068 } else if (IsX86 && Name.startswith("fma4.vfmadd.s")) {
3069 Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3070 CI->getArgOperand(2) };
3072 Ops[0] = Builder.CreateExtractElement(Ops[0], (uint64_t)0);
3073 Ops[1] = Builder.CreateExtractElement(Ops[1], (uint64_t)0);
3074 Ops[2] = Builder.CreateExtractElement(Ops[2], (uint64_t)0);
3076 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(),
3077 Intrinsic::fma,
3078 Ops[0]->getType()),
3079 Ops);
3081 Rep = Builder.CreateInsertElement(Constant::getNullValue(CI->getType()),
3082 Rep, (uint64_t)0);
3083 } else if (IsX86 && (Name.startswith("avx512.mask.vfmadd.s") ||
3084 Name.startswith("avx512.maskz.vfmadd.s") ||
3085 Name.startswith("avx512.mask3.vfmadd.s") ||
3086 Name.startswith("avx512.mask3.vfmsub.s") ||
3087 Name.startswith("avx512.mask3.vfnmsub.s"))) {
3088 bool IsMask3 = Name[11] == '3';
3089 bool IsMaskZ = Name[11] == 'z';
3090 // Drop the "avx512.mask." to make it easier.
3091 Name = Name.drop_front(IsMask3 || IsMaskZ ? 13 : 12);
3092 bool NegMul = Name[2] == 'n';
3093 bool NegAcc = NegMul ? Name[4] == 's' : Name[3] == 's';
3095 Value *A = CI->getArgOperand(0);
3096 Value *B = CI->getArgOperand(1);
3097 Value *C = CI->getArgOperand(2);
3099 if (NegMul && (IsMask3 || IsMaskZ))
3100 A = Builder.CreateFNeg(A);
3101 if (NegMul && !(IsMask3 || IsMaskZ))
3102 B = Builder.CreateFNeg(B);
3103 if (NegAcc)
3104 C = Builder.CreateFNeg(C);
3106 A = Builder.CreateExtractElement(A, (uint64_t)0);
3107 B = Builder.CreateExtractElement(B, (uint64_t)0);
3108 C = Builder.CreateExtractElement(C, (uint64_t)0);
3110 if (!isa<ConstantInt>(CI->getArgOperand(4)) ||
3111 cast<ConstantInt>(CI->getArgOperand(4))->getZExtValue() != 4) {
3112 Value *Ops[] = { A, B, C, CI->getArgOperand(4) };
3114 Intrinsic::ID IID;
3115 if (Name.back() == 'd')
3116 IID = Intrinsic::x86_avx512_vfmadd_f64;
3117 else
3118 IID = Intrinsic::x86_avx512_vfmadd_f32;
3119 Function *FMA = Intrinsic::getDeclaration(CI->getModule(), IID);
3120 Rep = Builder.CreateCall(FMA, Ops);
3121 } else {
3122 Function *FMA = Intrinsic::getDeclaration(CI->getModule(),
3123 Intrinsic::fma,
3124 A->getType());
3125 Rep = Builder.CreateCall(FMA, { A, B, C });
3128 Value *PassThru = IsMaskZ ? Constant::getNullValue(Rep->getType()) :
3129 IsMask3 ? C : A;
3131 // For Mask3 with NegAcc, we need to create a new extractelement that
3132 // avoids the negation above.
3133 if (NegAcc && IsMask3)
3134 PassThru = Builder.CreateExtractElement(CI->getArgOperand(2),
3135 (uint64_t)0);
3137 Rep = EmitX86ScalarSelect(Builder, CI->getArgOperand(3),
3138 Rep, PassThru);
3139 Rep = Builder.CreateInsertElement(CI->getArgOperand(IsMask3 ? 2 : 0),
3140 Rep, (uint64_t)0);
3141 } else if (IsX86 && (Name.startswith("avx512.mask.vfmadd.p") ||
3142 Name.startswith("avx512.mask.vfnmadd.p") ||
3143 Name.startswith("avx512.mask.vfnmsub.p") ||
3144 Name.startswith("avx512.mask3.vfmadd.p") ||
3145 Name.startswith("avx512.mask3.vfmsub.p") ||
3146 Name.startswith("avx512.mask3.vfnmsub.p") ||
3147 Name.startswith("avx512.maskz.vfmadd.p"))) {
3148 bool IsMask3 = Name[11] == '3';
3149 bool IsMaskZ = Name[11] == 'z';
3150 // Drop the "avx512.mask." to make it easier.
3151 Name = Name.drop_front(IsMask3 || IsMaskZ ? 13 : 12);
3152 bool NegMul = Name[2] == 'n';
3153 bool NegAcc = NegMul ? Name[4] == 's' : Name[3] == 's';
3155 Value *A = CI->getArgOperand(0);
3156 Value *B = CI->getArgOperand(1);
3157 Value *C = CI->getArgOperand(2);
3159 if (NegMul && (IsMask3 || IsMaskZ))
3160 A = Builder.CreateFNeg(A);
3161 if (NegMul && !(IsMask3 || IsMaskZ))
3162 B = Builder.CreateFNeg(B);
3163 if (NegAcc)
3164 C = Builder.CreateFNeg(C);
3166 if (CI->getNumArgOperands() == 5 &&
3167 (!isa<ConstantInt>(CI->getArgOperand(4)) ||
3168 cast<ConstantInt>(CI->getArgOperand(4))->getZExtValue() != 4)) {
3169 Intrinsic::ID IID;
3170 // Check the character before ".512" in string.
3171 if (Name[Name.size()-5] == 's')
3172 IID = Intrinsic::x86_avx512_vfmadd_ps_512;
3173 else
3174 IID = Intrinsic::x86_avx512_vfmadd_pd_512;
3176 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3177 { A, B, C, CI->getArgOperand(4) });
3178 } else {
3179 Function *FMA = Intrinsic::getDeclaration(CI->getModule(),
3180 Intrinsic::fma,
3181 A->getType());
3182 Rep = Builder.CreateCall(FMA, { A, B, C });
3185 Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(CI->getType()) :
3186 IsMask3 ? CI->getArgOperand(2) :
3187 CI->getArgOperand(0);
3189 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3190 } else if (IsX86 && (Name.startswith("fma.vfmaddsub.p") ||
3191 Name.startswith("fma.vfmsubadd.p"))) {
3192 bool IsSubAdd = Name[7] == 's';
3193 int NumElts = CI->getType()->getVectorNumElements();
3195 Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3196 CI->getArgOperand(2) };
3198 Function *FMA = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::fma,
3199 Ops[0]->getType());
3200 Value *Odd = Builder.CreateCall(FMA, Ops);
3201 Ops[2] = Builder.CreateFNeg(Ops[2]);
3202 Value *Even = Builder.CreateCall(FMA, Ops);
3204 if (IsSubAdd)
3205 std::swap(Even, Odd);
3207 SmallVector<uint32_t, 32> Idxs(NumElts);
3208 for (int i = 0; i != NumElts; ++i)
3209 Idxs[i] = i + (i % 2) * NumElts;
3211 Rep = Builder.CreateShuffleVector(Even, Odd, Idxs);
3212 } else if (IsX86 && (Name.startswith("avx512.mask.vfmaddsub.p") ||
3213 Name.startswith("avx512.mask3.vfmaddsub.p") ||
3214 Name.startswith("avx512.maskz.vfmaddsub.p") ||
3215 Name.startswith("avx512.mask3.vfmsubadd.p"))) {
3216 bool IsMask3 = Name[11] == '3';
3217 bool IsMaskZ = Name[11] == 'z';
3218 // Drop the "avx512.mask." to make it easier.
3219 Name = Name.drop_front(IsMask3 || IsMaskZ ? 13 : 12);
3220 bool IsSubAdd = Name[3] == 's';
3221 if (CI->getNumArgOperands() == 5 &&
3222 (!isa<ConstantInt>(CI->getArgOperand(4)) ||
3223 cast<ConstantInt>(CI->getArgOperand(4))->getZExtValue() != 4)) {
3224 Intrinsic::ID IID;
3225 // Check the character before ".512" in string.
3226 if (Name[Name.size()-5] == 's')
3227 IID = Intrinsic::x86_avx512_vfmaddsub_ps_512;
3228 else
3229 IID = Intrinsic::x86_avx512_vfmaddsub_pd_512;
3231 Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3232 CI->getArgOperand(2), CI->getArgOperand(4) };
3233 if (IsSubAdd)
3234 Ops[2] = Builder.CreateFNeg(Ops[2]);
3236 Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
3237 {CI->getArgOperand(0), CI->getArgOperand(1),
3238 CI->getArgOperand(2), CI->getArgOperand(4)});
3239 } else {
3240 int NumElts = CI->getType()->getVectorNumElements();
3242 Value *Ops[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3243 CI->getArgOperand(2) };
3245 Function *FMA = Intrinsic::getDeclaration(CI->getModule(), Intrinsic::fma,
3246 Ops[0]->getType());
3247 Value *Odd = Builder.CreateCall(FMA, Ops);
3248 Ops[2] = Builder.CreateFNeg(Ops[2]);
3249 Value *Even = Builder.CreateCall(FMA, Ops);
3251 if (IsSubAdd)
3252 std::swap(Even, Odd);
3254 SmallVector<uint32_t, 32> Idxs(NumElts);
3255 for (int i = 0; i != NumElts; ++i)
3256 Idxs[i] = i + (i % 2) * NumElts;
3258 Rep = Builder.CreateShuffleVector(Even, Odd, Idxs);
3261 Value *PassThru = IsMaskZ ? llvm::Constant::getNullValue(CI->getType()) :
3262 IsMask3 ? CI->getArgOperand(2) :
3263 CI->getArgOperand(0);
3265 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3266 } else if (IsX86 && (Name.startswith("avx512.mask.pternlog.") ||
3267 Name.startswith("avx512.maskz.pternlog."))) {
3268 bool ZeroMask = Name[11] == 'z';
3269 unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3270 unsigned EltWidth = CI->getType()->getScalarSizeInBits();
3271 Intrinsic::ID IID;
3272 if (VecWidth == 128 && EltWidth == 32)
3273 IID = Intrinsic::x86_avx512_pternlog_d_128;
3274 else if (VecWidth == 256 && EltWidth == 32)
3275 IID = Intrinsic::x86_avx512_pternlog_d_256;
3276 else if (VecWidth == 512 && EltWidth == 32)
3277 IID = Intrinsic::x86_avx512_pternlog_d_512;
3278 else if (VecWidth == 128 && EltWidth == 64)
3279 IID = Intrinsic::x86_avx512_pternlog_q_128;
3280 else if (VecWidth == 256 && EltWidth == 64)
3281 IID = Intrinsic::x86_avx512_pternlog_q_256;
3282 else if (VecWidth == 512 && EltWidth == 64)
3283 IID = Intrinsic::x86_avx512_pternlog_q_512;
3284 else
3285 llvm_unreachable("Unexpected intrinsic");
3287 Value *Args[] = { CI->getArgOperand(0) , CI->getArgOperand(1),
3288 CI->getArgOperand(2), CI->getArgOperand(3) };
3289 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3290 Args);
3291 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3292 : CI->getArgOperand(0);
3293 Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep, PassThru);
3294 } else if (IsX86 && (Name.startswith("avx512.mask.vpmadd52") ||
3295 Name.startswith("avx512.maskz.vpmadd52"))) {
3296 bool ZeroMask = Name[11] == 'z';
3297 bool High = Name[20] == 'h' || Name[21] == 'h';
3298 unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3299 Intrinsic::ID IID;
3300 if (VecWidth == 128 && !High)
3301 IID = Intrinsic::x86_avx512_vpmadd52l_uq_128;
3302 else if (VecWidth == 256 && !High)
3303 IID = Intrinsic::x86_avx512_vpmadd52l_uq_256;
3304 else if (VecWidth == 512 && !High)
3305 IID = Intrinsic::x86_avx512_vpmadd52l_uq_512;
3306 else if (VecWidth == 128 && High)
3307 IID = Intrinsic::x86_avx512_vpmadd52h_uq_128;
3308 else if (VecWidth == 256 && High)
3309 IID = Intrinsic::x86_avx512_vpmadd52h_uq_256;
3310 else if (VecWidth == 512 && High)
3311 IID = Intrinsic::x86_avx512_vpmadd52h_uq_512;
3312 else
3313 llvm_unreachable("Unexpected intrinsic");
3315 Value *Args[] = { CI->getArgOperand(0) , CI->getArgOperand(1),
3316 CI->getArgOperand(2) };
3317 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3318 Args);
3319 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3320 : CI->getArgOperand(0);
3321 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3322 } else if (IsX86 && (Name.startswith("avx512.mask.vpermi2var.") ||
3323 Name.startswith("avx512.mask.vpermt2var.") ||
3324 Name.startswith("avx512.maskz.vpermt2var."))) {
3325 bool ZeroMask = Name[11] == 'z';
3326 bool IndexForm = Name[17] == 'i';
3327 Rep = UpgradeX86VPERMT2Intrinsics(Builder, *CI, ZeroMask, IndexForm);
3328 } else if (IsX86 && (Name.startswith("avx512.mask.vpdpbusd.") ||
3329 Name.startswith("avx512.maskz.vpdpbusd.") ||
3330 Name.startswith("avx512.mask.vpdpbusds.") ||
3331 Name.startswith("avx512.maskz.vpdpbusds."))) {
3332 bool ZeroMask = Name[11] == 'z';
3333 bool IsSaturating = Name[ZeroMask ? 21 : 20] == 's';
3334 unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3335 Intrinsic::ID IID;
3336 if (VecWidth == 128 && !IsSaturating)
3337 IID = Intrinsic::x86_avx512_vpdpbusd_128;
3338 else if (VecWidth == 256 && !IsSaturating)
3339 IID = Intrinsic::x86_avx512_vpdpbusd_256;
3340 else if (VecWidth == 512 && !IsSaturating)
3341 IID = Intrinsic::x86_avx512_vpdpbusd_512;
3342 else if (VecWidth == 128 && IsSaturating)
3343 IID = Intrinsic::x86_avx512_vpdpbusds_128;
3344 else if (VecWidth == 256 && IsSaturating)
3345 IID = Intrinsic::x86_avx512_vpdpbusds_256;
3346 else if (VecWidth == 512 && IsSaturating)
3347 IID = Intrinsic::x86_avx512_vpdpbusds_512;
3348 else
3349 llvm_unreachable("Unexpected intrinsic");
3351 Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3352 CI->getArgOperand(2) };
3353 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3354 Args);
3355 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3356 : CI->getArgOperand(0);
3357 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3358 } else if (IsX86 && (Name.startswith("avx512.mask.vpdpwssd.") ||
3359 Name.startswith("avx512.maskz.vpdpwssd.") ||
3360 Name.startswith("avx512.mask.vpdpwssds.") ||
3361 Name.startswith("avx512.maskz.vpdpwssds."))) {
3362 bool ZeroMask = Name[11] == 'z';
3363 bool IsSaturating = Name[ZeroMask ? 21 : 20] == 's';
3364 unsigned VecWidth = CI->getType()->getPrimitiveSizeInBits();
3365 Intrinsic::ID IID;
3366 if (VecWidth == 128 && !IsSaturating)
3367 IID = Intrinsic::x86_avx512_vpdpwssd_128;
3368 else if (VecWidth == 256 && !IsSaturating)
3369 IID = Intrinsic::x86_avx512_vpdpwssd_256;
3370 else if (VecWidth == 512 && !IsSaturating)
3371 IID = Intrinsic::x86_avx512_vpdpwssd_512;
3372 else if (VecWidth == 128 && IsSaturating)
3373 IID = Intrinsic::x86_avx512_vpdpwssds_128;
3374 else if (VecWidth == 256 && IsSaturating)
3375 IID = Intrinsic::x86_avx512_vpdpwssds_256;
3376 else if (VecWidth == 512 && IsSaturating)
3377 IID = Intrinsic::x86_avx512_vpdpwssds_512;
3378 else
3379 llvm_unreachable("Unexpected intrinsic");
3381 Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3382 CI->getArgOperand(2) };
3383 Rep = Builder.CreateCall(Intrinsic::getDeclaration(CI->getModule(), IID),
3384 Args);
3385 Value *PassThru = ZeroMask ? ConstantAggregateZero::get(CI->getType())
3386 : CI->getArgOperand(0);
3387 Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep, PassThru);
3388 } else if (IsX86 && (Name == "addcarryx.u32" || Name == "addcarryx.u64" ||
3389 Name == "addcarry.u32" || Name == "addcarry.u64" ||
3390 Name == "subborrow.u32" || Name == "subborrow.u64")) {
3391 Intrinsic::ID IID;
3392 if (Name[0] == 'a' && Name.back() == '2')
3393 IID = Intrinsic::x86_addcarry_32;
3394 else if (Name[0] == 'a' && Name.back() == '4')
3395 IID = Intrinsic::x86_addcarry_64;
3396 else if (Name[0] == 's' && Name.back() == '2')
3397 IID = Intrinsic::x86_subborrow_32;
3398 else if (Name[0] == 's' && Name.back() == '4')
3399 IID = Intrinsic::x86_subborrow_64;
3400 else
3401 llvm_unreachable("Unexpected intrinsic");
3403 // Make a call with 3 operands.
3404 Value *Args[] = { CI->getArgOperand(0), CI->getArgOperand(1),
3405 CI->getArgOperand(2)};
3406 Value *NewCall = Builder.CreateCall(
3407 Intrinsic::getDeclaration(CI->getModule(), IID),
3408 Args);
3410 // Extract the second result and store it.
3411 Value *Data = Builder.CreateExtractValue(NewCall, 1);
3412 // Cast the pointer to the right type.
3413 Value *Ptr = Builder.CreateBitCast(CI->getArgOperand(3),
3414 llvm::PointerType::getUnqual(Data->getType()));
3415 Builder.CreateAlignedStore(Data, Ptr, 1);
3416 // Replace the original call result with the first result of the new call.
3417 Value *CF = Builder.CreateExtractValue(NewCall, 0);
3419 CI->replaceAllUsesWith(CF);
3420 Rep = nullptr;
3421 } else if (IsX86 && Name.startswith("avx512.mask.") &&
3422 upgradeAVX512MaskToSelect(Name, Builder, *CI, Rep)) {
3423 // Rep will be updated by the call in the condition.
3424 } else if (IsNVVM && (Name == "abs.i" || Name == "abs.ll")) {
3425 Value *Arg = CI->getArgOperand(0);
3426 Value *Neg = Builder.CreateNeg(Arg, "neg");
3427 Value *Cmp = Builder.CreateICmpSGE(
3428 Arg, llvm::Constant::getNullValue(Arg->getType()), "abs.cond");
3429 Rep = Builder.CreateSelect(Cmp, Arg, Neg, "abs");
3430 } else if (IsNVVM && (Name.startswith("atomic.load.add.f32.p") ||
3431 Name.startswith("atomic.load.add.f64.p"))) {
3432 Value *Ptr = CI->getArgOperand(0);
3433 Value *Val = CI->getArgOperand(1);
3434 Rep = Builder.CreateAtomicRMW(AtomicRMWInst::FAdd, Ptr, Val,
3435 AtomicOrdering::SequentiallyConsistent);
3436 } else if (IsNVVM && (Name == "max.i" || Name == "max.ll" ||
3437 Name == "max.ui" || Name == "max.ull")) {
3438 Value *Arg0 = CI->getArgOperand(0);
3439 Value *Arg1 = CI->getArgOperand(1);
3440 Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
3441 ? Builder.CreateICmpUGE(Arg0, Arg1, "max.cond")
3442 : Builder.CreateICmpSGE(Arg0, Arg1, "max.cond");
3443 Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "max");
3444 } else if (IsNVVM && (Name == "min.i" || Name == "min.ll" ||
3445 Name == "min.ui" || Name == "min.ull")) {
3446 Value *Arg0 = CI->getArgOperand(0);
3447 Value *Arg1 = CI->getArgOperand(1);
3448 Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
3449 ? Builder.CreateICmpULE(Arg0, Arg1, "min.cond")
3450 : Builder.CreateICmpSLE(Arg0, Arg1, "min.cond");
3451 Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "min");
3452 } else if (IsNVVM && Name == "clz.ll") {
3453 // llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 and returns an i64.
3454 Value *Arg = CI->getArgOperand(0);
3455 Value *Ctlz = Builder.CreateCall(
3456 Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
3457 {Arg->getType()}),
3458 {Arg, Builder.getFalse()}, "ctlz");
3459 Rep = Builder.CreateTrunc(Ctlz, Builder.getInt32Ty(), "ctlz.trunc");
3460 } else if (IsNVVM && Name == "popc.ll") {
3461 // llvm.nvvm.popc.ll returns an i32, but llvm.ctpop.i64 and returns an
3462 // i64.
3463 Value *Arg = CI->getArgOperand(0);
3464 Value *Popc = Builder.CreateCall(
3465 Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
3466 {Arg->getType()}),
3467 Arg, "ctpop");
3468 Rep = Builder.CreateTrunc(Popc, Builder.getInt32Ty(), "ctpop.trunc");
3469 } else if (IsNVVM && Name == "h2f") {
3470 Rep = Builder.CreateCall(Intrinsic::getDeclaration(
3471 F->getParent(), Intrinsic::convert_from_fp16,
3472 {Builder.getFloatTy()}),
3473 CI->getArgOperand(0), "h2f");
3474 } else {
3475 llvm_unreachable("Unknown function for CallInst upgrade.");
3478 if (Rep)
3479 CI->replaceAllUsesWith(Rep);
3480 CI->eraseFromParent();
3481 return;
3484 const auto &DefaultCase = [&NewFn, &CI]() -> void {
3485 // Handle generic mangling change, but nothing else
3486 assert(
3487 (CI->getCalledFunction()->getName() != NewFn->getName()) &&
3488 "Unknown function for CallInst upgrade and isn't just a name change");
3489 CI->setCalledFunction(NewFn);
3491 CallInst *NewCall = nullptr;
3492 switch (NewFn->getIntrinsicID()) {
3493 default: {
3494 DefaultCase();
3495 return;
3497 case Intrinsic::experimental_vector_reduce_v2_fmul: {
3498 SmallVector<Value *, 2> Args;
3499 if (CI->isFast())
3500 Args.push_back(ConstantFP::get(CI->getOperand(0)->getType(), 1.0));
3501 else
3502 Args.push_back(CI->getOperand(0));
3503 Args.push_back(CI->getOperand(1));
3504 NewCall = Builder.CreateCall(NewFn, Args);
3505 cast<Instruction>(NewCall)->copyFastMathFlags(CI);
3506 break;
3508 case Intrinsic::experimental_vector_reduce_v2_fadd: {
3509 SmallVector<Value *, 2> Args;
3510 if (CI->isFast())
3511 Args.push_back(Constant::getNullValue(CI->getOperand(0)->getType()));
3512 else
3513 Args.push_back(CI->getOperand(0));
3514 Args.push_back(CI->getOperand(1));
3515 NewCall = Builder.CreateCall(NewFn, Args);
3516 cast<Instruction>(NewCall)->copyFastMathFlags(CI);
3517 break;
3519 case Intrinsic::arm_neon_vld1:
3520 case Intrinsic::arm_neon_vld2:
3521 case Intrinsic::arm_neon_vld3:
3522 case Intrinsic::arm_neon_vld4:
3523 case Intrinsic::arm_neon_vld2lane:
3524 case Intrinsic::arm_neon_vld3lane:
3525 case Intrinsic::arm_neon_vld4lane:
3526 case Intrinsic::arm_neon_vst1:
3527 case Intrinsic::arm_neon_vst2:
3528 case Intrinsic::arm_neon_vst3:
3529 case Intrinsic::arm_neon_vst4:
3530 case Intrinsic::arm_neon_vst2lane:
3531 case Intrinsic::arm_neon_vst3lane:
3532 case Intrinsic::arm_neon_vst4lane: {
3533 SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
3534 CI->arg_operands().end());
3535 NewCall = Builder.CreateCall(NewFn, Args);
3536 break;
3539 case Intrinsic::bitreverse:
3540 NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
3541 break;
3543 case Intrinsic::ctlz:
3544 case Intrinsic::cttz:
3545 assert(CI->getNumArgOperands() == 1 &&
3546 "Mismatch between function args and call args");
3547 NewCall =
3548 Builder.CreateCall(NewFn, {CI->getArgOperand(0), Builder.getFalse()});
3549 break;
3551 case Intrinsic::objectsize: {
3552 Value *NullIsUnknownSize = CI->getNumArgOperands() == 2
3553 ? Builder.getFalse()
3554 : CI->getArgOperand(2);
3555 Value *Dynamic =
3556 CI->getNumArgOperands() < 4 ? Builder.getFalse() : CI->getArgOperand(3);
3557 NewCall = Builder.CreateCall(
3558 NewFn, {CI->getArgOperand(0), CI->getArgOperand(1), NullIsUnknownSize, Dynamic});
3559 break;
3562 case Intrinsic::ctpop:
3563 NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
3564 break;
3566 case Intrinsic::convert_from_fp16:
3567 NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
3568 break;
3570 case Intrinsic::dbg_value:
3571 // Upgrade from the old version that had an extra offset argument.
3572 assert(CI->getNumArgOperands() == 4);
3573 // Drop nonzero offsets instead of attempting to upgrade them.
3574 if (auto *Offset = dyn_cast_or_null<Constant>(CI->getArgOperand(1)))
3575 if (Offset->isZeroValue()) {
3576 NewCall = Builder.CreateCall(
3577 NewFn,
3578 {CI->getArgOperand(0), CI->getArgOperand(2), CI->getArgOperand(3)});
3579 break;
3581 CI->eraseFromParent();
3582 return;
3584 case Intrinsic::x86_xop_vfrcz_ss:
3585 case Intrinsic::x86_xop_vfrcz_sd:
3586 NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(1)});
3587 break;
3589 case Intrinsic::x86_xop_vpermil2pd:
3590 case Intrinsic::x86_xop_vpermil2ps:
3591 case Intrinsic::x86_xop_vpermil2pd_256:
3592 case Intrinsic::x86_xop_vpermil2ps_256: {
3593 SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
3594 CI->arg_operands().end());
3595 VectorType *FltIdxTy = cast<VectorType>(Args[2]->getType());
3596 VectorType *IntIdxTy = VectorType::getInteger(FltIdxTy);
3597 Args[2] = Builder.CreateBitCast(Args[2], IntIdxTy);
3598 NewCall = Builder.CreateCall(NewFn, Args);
3599 break;
3602 case Intrinsic::x86_sse41_ptestc:
3603 case Intrinsic::x86_sse41_ptestz:
3604 case Intrinsic::x86_sse41_ptestnzc: {
3605 // The arguments for these intrinsics used to be v4f32, and changed
3606 // to v2i64. This is purely a nop, since those are bitwise intrinsics.
3607 // So, the only thing required is a bitcast for both arguments.
3608 // First, check the arguments have the old type.
3609 Value *Arg0 = CI->getArgOperand(0);
3610 if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
3611 return;
3613 // Old intrinsic, add bitcasts
3614 Value *Arg1 = CI->getArgOperand(1);
3616 Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
3618 Value *BC0 = Builder.CreateBitCast(Arg0, NewVecTy, "cast");
3619 Value *BC1 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
3621 NewCall = Builder.CreateCall(NewFn, {BC0, BC1});
3622 break;
3625 case Intrinsic::x86_rdtscp: {
3626 // This used to take 1 arguments. If we have no arguments, it is already
3627 // upgraded.
3628 if (CI->getNumOperands() == 0)
3629 return;
3631 NewCall = Builder.CreateCall(NewFn);
3632 // Extract the second result and store it.
3633 Value *Data = Builder.CreateExtractValue(NewCall, 1);
3634 // Cast the pointer to the right type.
3635 Value *Ptr = Builder.CreateBitCast(CI->getArgOperand(0),
3636 llvm::PointerType::getUnqual(Data->getType()));
3637 Builder.CreateAlignedStore(Data, Ptr, 1);
3638 // Replace the original call result with the first result of the new call.
3639 Value *TSC = Builder.CreateExtractValue(NewCall, 0);
3641 std::string Name = CI->getName();
3642 if (!Name.empty()) {
3643 CI->setName(Name + ".old");
3644 NewCall->setName(Name);
3646 CI->replaceAllUsesWith(TSC);
3647 CI->eraseFromParent();
3648 return;
3651 case Intrinsic::x86_sse41_insertps:
3652 case Intrinsic::x86_sse41_dppd:
3653 case Intrinsic::x86_sse41_dpps:
3654 case Intrinsic::x86_sse41_mpsadbw:
3655 case Intrinsic::x86_avx_dp_ps_256:
3656 case Intrinsic::x86_avx2_mpsadbw: {
3657 // Need to truncate the last argument from i32 to i8 -- this argument models
3658 // an inherently 8-bit immediate operand to these x86 instructions.
3659 SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
3660 CI->arg_operands().end());
3662 // Replace the last argument with a trunc.
3663 Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
3664 NewCall = Builder.CreateCall(NewFn, Args);
3665 break;
3668 case Intrinsic::thread_pointer: {
3669 NewCall = Builder.CreateCall(NewFn, {});
3670 break;
3673 case Intrinsic::invariant_start:
3674 case Intrinsic::invariant_end:
3675 case Intrinsic::masked_load:
3676 case Intrinsic::masked_store:
3677 case Intrinsic::masked_gather:
3678 case Intrinsic::masked_scatter: {
3679 SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
3680 CI->arg_operands().end());
3681 NewCall = Builder.CreateCall(NewFn, Args);
3682 break;
3685 case Intrinsic::memcpy:
3686 case Intrinsic::memmove:
3687 case Intrinsic::memset: {
3688 // We have to make sure that the call signature is what we're expecting.
3689 // We only want to change the old signatures by removing the alignment arg:
3690 // @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i32, i1)
3691 // -> @llvm.mem[cpy|move]...(i8*, i8*, i[32|i64], i1)
3692 // @llvm.memset...(i8*, i8, i[32|64], i32, i1)
3693 // -> @llvm.memset...(i8*, i8, i[32|64], i1)
3694 // Note: i8*'s in the above can be any pointer type
3695 if (CI->getNumArgOperands() != 5) {
3696 DefaultCase();
3697 return;
3699 // Remove alignment argument (3), and add alignment attributes to the
3700 // dest/src pointers.
3701 Value *Args[4] = {CI->getArgOperand(0), CI->getArgOperand(1),
3702 CI->getArgOperand(2), CI->getArgOperand(4)};
3703 NewCall = Builder.CreateCall(NewFn, Args);
3704 auto *MemCI = cast<MemIntrinsic>(NewCall);
3705 // All mem intrinsics support dest alignment.
3706 const ConstantInt *Align = cast<ConstantInt>(CI->getArgOperand(3));
3707 MemCI->setDestAlignment(Align->getZExtValue());
3708 // Memcpy/Memmove also support source alignment.
3709 if (auto *MTI = dyn_cast<MemTransferInst>(MemCI))
3710 MTI->setSourceAlignment(Align->getZExtValue());
3711 break;
3714 assert(NewCall && "Should have either set this variable or returned through "
3715 "the default case");
3716 std::string Name = CI->getName();
3717 if (!Name.empty()) {
3718 CI->setName(Name + ".old");
3719 NewCall->setName(Name);
3721 CI->replaceAllUsesWith(NewCall);
3722 CI->eraseFromParent();
3725 void llvm::UpgradeCallsToIntrinsic(Function *F) {
3726 assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
3728 // Check if this function should be upgraded and get the replacement function
3729 // if there is one.
3730 Function *NewFn;
3731 if (UpgradeIntrinsicFunction(F, NewFn)) {
3732 // Replace all users of the old function with the new function or new
3733 // instructions. This is not a range loop because the call is deleted.
3734 for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; )
3735 if (CallInst *CI = dyn_cast<CallInst>(*UI++))
3736 UpgradeIntrinsicCall(CI, NewFn);
3738 // Remove old function, no longer used, from the module.
3739 F->eraseFromParent();
3743 MDNode *llvm::UpgradeTBAANode(MDNode &MD) {
3744 // Check if the tag uses struct-path aware TBAA format.
3745 if (isa<MDNode>(MD.getOperand(0)) && MD.getNumOperands() >= 3)
3746 return &MD;
3748 auto &Context = MD.getContext();
3749 if (MD.getNumOperands() == 3) {
3750 Metadata *Elts[] = {MD.getOperand(0), MD.getOperand(1)};
3751 MDNode *ScalarType = MDNode::get(Context, Elts);
3752 // Create a MDNode <ScalarType, ScalarType, offset 0, const>
3753 Metadata *Elts2[] = {ScalarType, ScalarType,
3754 ConstantAsMetadata::get(
3755 Constant::getNullValue(Type::getInt64Ty(Context))),
3756 MD.getOperand(2)};
3757 return MDNode::get(Context, Elts2);
3759 // Create a MDNode <MD, MD, offset 0>
3760 Metadata *Elts[] = {&MD, &MD, ConstantAsMetadata::get(Constant::getNullValue(
3761 Type::getInt64Ty(Context)))};
3762 return MDNode::get(Context, Elts);
3765 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
3766 Instruction *&Temp) {
3767 if (Opc != Instruction::BitCast)
3768 return nullptr;
3770 Temp = nullptr;
3771 Type *SrcTy = V->getType();
3772 if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
3773 SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
3774 LLVMContext &Context = V->getContext();
3776 // We have no information about target data layout, so we assume that
3777 // the maximum pointer size is 64bit.
3778 Type *MidTy = Type::getInt64Ty(Context);
3779 Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
3781 return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
3784 return nullptr;
3787 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
3788 if (Opc != Instruction::BitCast)
3789 return nullptr;
3791 Type *SrcTy = C->getType();
3792 if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
3793 SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
3794 LLVMContext &Context = C->getContext();
3796 // We have no information about target data layout, so we assume that
3797 // the maximum pointer size is 64bit.
3798 Type *MidTy = Type::getInt64Ty(Context);
3800 return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
3801 DestTy);
3804 return nullptr;
3807 /// Check the debug info version number, if it is out-dated, drop the debug
3808 /// info. Return true if module is modified.
3809 bool llvm::UpgradeDebugInfo(Module &M) {
3810 unsigned Version = getDebugMetadataVersionFromModule(M);
3811 if (Version == DEBUG_METADATA_VERSION) {
3812 bool BrokenDebugInfo = false;
3813 if (verifyModule(M, &llvm::errs(), &BrokenDebugInfo))
3814 report_fatal_error("Broken module found, compilation aborted!");
3815 if (!BrokenDebugInfo)
3816 // Everything is ok.
3817 return false;
3818 else {
3819 // Diagnose malformed debug info.
3820 DiagnosticInfoIgnoringInvalidDebugMetadata Diag(M);
3821 M.getContext().diagnose(Diag);
3824 bool Modified = StripDebugInfo(M);
3825 if (Modified && Version != DEBUG_METADATA_VERSION) {
3826 // Diagnose a version mismatch.
3827 DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
3828 M.getContext().diagnose(DiagVersion);
3830 return Modified;
3833 /// This checks for objc retain release marker which should be upgraded. It
3834 /// returns true if module is modified.
3835 static bool UpgradeRetainReleaseMarker(Module &M) {
3836 bool Changed = false;
3837 const char *MarkerKey = "clang.arc.retainAutoreleasedReturnValueMarker";
3838 NamedMDNode *ModRetainReleaseMarker = M.getNamedMetadata(MarkerKey);
3839 if (ModRetainReleaseMarker) {
3840 MDNode *Op = ModRetainReleaseMarker->getOperand(0);
3841 if (Op) {
3842 MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(0));
3843 if (ID) {
3844 SmallVector<StringRef, 4> ValueComp;
3845 ID->getString().split(ValueComp, "#");
3846 if (ValueComp.size() == 2) {
3847 std::string NewValue = ValueComp[0].str() + ";" + ValueComp[1].str();
3848 ID = MDString::get(M.getContext(), NewValue);
3850 M.addModuleFlag(Module::Error, MarkerKey, ID);
3851 M.eraseNamedMetadata(ModRetainReleaseMarker);
3852 Changed = true;
3856 return Changed;
3859 void llvm::UpgradeARCRuntime(Module &M) {
3860 // This lambda converts normal function calls to ARC runtime functions to
3861 // intrinsic calls.
3862 auto UpgradeToIntrinsic = [&](const char *OldFunc,
3863 llvm::Intrinsic::ID IntrinsicFunc) {
3864 Function *Fn = M.getFunction(OldFunc);
3866 if (!Fn)
3867 return;
3869 Function *NewFn = llvm::Intrinsic::getDeclaration(&M, IntrinsicFunc);
3871 for (auto I = Fn->user_begin(), E = Fn->user_end(); I != E;) {
3872 CallInst *CI = dyn_cast<CallInst>(*I++);
3873 if (!CI || CI->getCalledFunction() != Fn)
3874 continue;
3876 IRBuilder<> Builder(CI->getParent(), CI->getIterator());
3877 FunctionType *NewFuncTy = NewFn->getFunctionType();
3878 SmallVector<Value *, 2> Args;
3880 for (unsigned I = 0, E = CI->getNumArgOperands(); I != E; ++I) {
3881 Value *Arg = CI->getArgOperand(I);
3882 // Bitcast argument to the parameter type of the new function if it's
3883 // not a variadic argument.
3884 if (I < NewFuncTy->getNumParams())
3885 Arg = Builder.CreateBitCast(Arg, NewFuncTy->getParamType(I));
3886 Args.push_back(Arg);
3889 // Create a call instruction that calls the new function.
3890 CallInst *NewCall = Builder.CreateCall(NewFuncTy, NewFn, Args);
3891 NewCall->setTailCallKind(cast<CallInst>(CI)->getTailCallKind());
3892 NewCall->setName(CI->getName());
3894 // Bitcast the return value back to the type of the old call.
3895 Value *NewRetVal = Builder.CreateBitCast(NewCall, CI->getType());
3897 if (!CI->use_empty())
3898 CI->replaceAllUsesWith(NewRetVal);
3899 CI->eraseFromParent();
3902 if (Fn->use_empty())
3903 Fn->eraseFromParent();
3906 // Unconditionally convert a call to "clang.arc.use" to a call to
3907 // "llvm.objc.clang.arc.use".
3908 UpgradeToIntrinsic("clang.arc.use", llvm::Intrinsic::objc_clang_arc_use);
3910 // Upgrade the retain release marker. If there is no need to upgrade
3911 // the marker, that means either the module is already new enough to contain
3912 // new intrinsics or it is not ARC. There is no need to upgrade runtime call.
3913 if (!UpgradeRetainReleaseMarker(M))
3914 return;
3916 std::pair<const char *, llvm::Intrinsic::ID> RuntimeFuncs[] = {
3917 {"objc_autorelease", llvm::Intrinsic::objc_autorelease},
3918 {"objc_autoreleasePoolPop", llvm::Intrinsic::objc_autoreleasePoolPop},
3919 {"objc_autoreleasePoolPush", llvm::Intrinsic::objc_autoreleasePoolPush},
3920 {"objc_autoreleaseReturnValue",
3921 llvm::Intrinsic::objc_autoreleaseReturnValue},
3922 {"objc_copyWeak", llvm::Intrinsic::objc_copyWeak},
3923 {"objc_destroyWeak", llvm::Intrinsic::objc_destroyWeak},
3924 {"objc_initWeak", llvm::Intrinsic::objc_initWeak},
3925 {"objc_loadWeak", llvm::Intrinsic::objc_loadWeak},
3926 {"objc_loadWeakRetained", llvm::Intrinsic::objc_loadWeakRetained},
3927 {"objc_moveWeak", llvm::Intrinsic::objc_moveWeak},
3928 {"objc_release", llvm::Intrinsic::objc_release},
3929 {"objc_retain", llvm::Intrinsic::objc_retain},
3930 {"objc_retainAutorelease", llvm::Intrinsic::objc_retainAutorelease},
3931 {"objc_retainAutoreleaseReturnValue",
3932 llvm::Intrinsic::objc_retainAutoreleaseReturnValue},
3933 {"objc_retainAutoreleasedReturnValue",
3934 llvm::Intrinsic::objc_retainAutoreleasedReturnValue},
3935 {"objc_retainBlock", llvm::Intrinsic::objc_retainBlock},
3936 {"objc_storeStrong", llvm::Intrinsic::objc_storeStrong},
3937 {"objc_storeWeak", llvm::Intrinsic::objc_storeWeak},
3938 {"objc_unsafeClaimAutoreleasedReturnValue",
3939 llvm::Intrinsic::objc_unsafeClaimAutoreleasedReturnValue},
3940 {"objc_retainedObject", llvm::Intrinsic::objc_retainedObject},
3941 {"objc_unretainedObject", llvm::Intrinsic::objc_unretainedObject},
3942 {"objc_unretainedPointer", llvm::Intrinsic::objc_unretainedPointer},
3943 {"objc_retain_autorelease", llvm::Intrinsic::objc_retain_autorelease},
3944 {"objc_sync_enter", llvm::Intrinsic::objc_sync_enter},
3945 {"objc_sync_exit", llvm::Intrinsic::objc_sync_exit},
3946 {"objc_arc_annotation_topdown_bbstart",
3947 llvm::Intrinsic::objc_arc_annotation_topdown_bbstart},
3948 {"objc_arc_annotation_topdown_bbend",
3949 llvm::Intrinsic::objc_arc_annotation_topdown_bbend},
3950 {"objc_arc_annotation_bottomup_bbstart",
3951 llvm::Intrinsic::objc_arc_annotation_bottomup_bbstart},
3952 {"objc_arc_annotation_bottomup_bbend",
3953 llvm::Intrinsic::objc_arc_annotation_bottomup_bbend}};
3955 for (auto &I : RuntimeFuncs)
3956 UpgradeToIntrinsic(I.first, I.second);
3959 bool llvm::UpgradeModuleFlags(Module &M) {
3960 NamedMDNode *ModFlags = M.getModuleFlagsMetadata();
3961 if (!ModFlags)
3962 return false;
3964 bool HasObjCFlag = false, HasClassProperties = false, Changed = false;
3965 for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) {
3966 MDNode *Op = ModFlags->getOperand(I);
3967 if (Op->getNumOperands() != 3)
3968 continue;
3969 MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(1));
3970 if (!ID)
3971 continue;
3972 if (ID->getString() == "Objective-C Image Info Version")
3973 HasObjCFlag = true;
3974 if (ID->getString() == "Objective-C Class Properties")
3975 HasClassProperties = true;
3976 // Upgrade PIC/PIE Module Flags. The module flag behavior for these two
3977 // field was Error and now they are Max.
3978 if (ID->getString() == "PIC Level" || ID->getString() == "PIE Level") {
3979 if (auto *Behavior =
3980 mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0))) {
3981 if (Behavior->getLimitedValue() == Module::Error) {
3982 Type *Int32Ty = Type::getInt32Ty(M.getContext());
3983 Metadata *Ops[3] = {
3984 ConstantAsMetadata::get(ConstantInt::get(Int32Ty, Module::Max)),
3985 MDString::get(M.getContext(), ID->getString()),
3986 Op->getOperand(2)};
3987 ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
3988 Changed = true;
3992 // Upgrade Objective-C Image Info Section. Removed the whitespce in the
3993 // section name so that llvm-lto will not complain about mismatching
3994 // module flags that is functionally the same.
3995 if (ID->getString() == "Objective-C Image Info Section") {
3996 if (auto *Value = dyn_cast_or_null<MDString>(Op->getOperand(2))) {
3997 SmallVector<StringRef, 4> ValueComp;
3998 Value->getString().split(ValueComp, " ");
3999 if (ValueComp.size() != 1) {
4000 std::string NewValue;
4001 for (auto &S : ValueComp)
4002 NewValue += S.str();
4003 Metadata *Ops[3] = {Op->getOperand(0), Op->getOperand(1),
4004 MDString::get(M.getContext(), NewValue)};
4005 ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
4006 Changed = true;
4012 // "Objective-C Class Properties" is recently added for Objective-C. We
4013 // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module
4014 // flag of value 0, so we can correclty downgrade this flag when trying to
4015 // link an ObjC bitcode without this module flag with an ObjC bitcode with
4016 // this module flag.
4017 if (HasObjCFlag && !HasClassProperties) {
4018 M.addModuleFlag(llvm::Module::Override, "Objective-C Class Properties",
4019 (uint32_t)0);
4020 Changed = true;
4023 return Changed;
4026 void llvm::UpgradeSectionAttributes(Module &M) {
4027 auto TrimSpaces = [](StringRef Section) -> std::string {
4028 SmallVector<StringRef, 5> Components;
4029 Section.split(Components, ',');
4031 SmallString<32> Buffer;
4032 raw_svector_ostream OS(Buffer);
4034 for (auto Component : Components)
4035 OS << ',' << Component.trim();
4037 return OS.str().substr(1);
4040 for (auto &GV : M.globals()) {
4041 if (!GV.hasSection())
4042 continue;
4044 StringRef Section = GV.getSection();
4046 if (!Section.startswith("__DATA, __objc_catlist"))
4047 continue;
4049 // __DATA, __objc_catlist, regular, no_dead_strip
4050 // __DATA,__objc_catlist,regular,no_dead_strip
4051 GV.setSection(TrimSpaces(Section));
4055 static bool isOldLoopArgument(Metadata *MD) {
4056 auto *T = dyn_cast_or_null<MDTuple>(MD);
4057 if (!T)
4058 return false;
4059 if (T->getNumOperands() < 1)
4060 return false;
4061 auto *S = dyn_cast_or_null<MDString>(T->getOperand(0));
4062 if (!S)
4063 return false;
4064 return S->getString().startswith("llvm.vectorizer.");
4067 static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) {
4068 StringRef OldPrefix = "llvm.vectorizer.";
4069 assert(OldTag.startswith(OldPrefix) && "Expected old prefix");
4071 if (OldTag == "llvm.vectorizer.unroll")
4072 return MDString::get(C, "llvm.loop.interleave.count");
4074 return MDString::get(
4075 C, (Twine("llvm.loop.vectorize.") + OldTag.drop_front(OldPrefix.size()))
4076 .str());
4079 static Metadata *upgradeLoopArgument(Metadata *MD) {
4080 auto *T = dyn_cast_or_null<MDTuple>(MD);
4081 if (!T)
4082 return MD;
4083 if (T->getNumOperands() < 1)
4084 return MD;
4085 auto *OldTag = dyn_cast_or_null<MDString>(T->getOperand(0));
4086 if (!OldTag)
4087 return MD;
4088 if (!OldTag->getString().startswith("llvm.vectorizer."))
4089 return MD;
4091 // This has an old tag. Upgrade it.
4092 SmallVector<Metadata *, 8> Ops;
4093 Ops.reserve(T->getNumOperands());
4094 Ops.push_back(upgradeLoopTag(T->getContext(), OldTag->getString()));
4095 for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I)
4096 Ops.push_back(T->getOperand(I));
4098 return MDTuple::get(T->getContext(), Ops);
4101 MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) {
4102 auto *T = dyn_cast<MDTuple>(&N);
4103 if (!T)
4104 return &N;
4106 if (none_of(T->operands(), isOldLoopArgument))
4107 return &N;
4109 SmallVector<Metadata *, 8> Ops;
4110 Ops.reserve(T->getNumOperands());
4111 for (Metadata *MD : T->operands())
4112 Ops.push_back(upgradeLoopArgument(MD));
4114 return MDTuple::get(T->getContext(), Ops);
4117 std::string llvm::UpgradeDataLayoutString(StringRef DL, StringRef TT) {
4118 std::string AddrSpaces = "-p270:32:32-p271:32:32-p272:64:64";
4120 // If X86, and the datalayout matches the expected format, add pointer size
4121 // address spaces to the datalayout.
4122 Triple::ArchType Arch = Triple(TT).getArch();
4123 if ((Arch != llvm::Triple::x86 && Arch != llvm::Triple::x86_64) ||
4124 DL.contains(AddrSpaces))
4125 return DL;
4127 SmallVector<StringRef, 4> Groups;
4128 Regex R("(e-m:[a-z](-p:32:32)?)(-[if]64:.*$)");
4129 if (!R.match(DL, &Groups))
4130 return DL;
4132 SmallString<1024> Buf;
4133 std::string Res = (Groups[1] + AddrSpaces + Groups[3]).toStringRef(Buf).str();
4134 return Res;