1 //===-- AutoUpgrade.cpp - Implement auto-upgrade helper functions ---------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the auto-upgrade helper functions.
11 // This is where deprecated IR intrinsics and other IR features are updated to
12 // current specifications.
13 //
14 //===----------------------------------------------------------------------===//
15
16 #include "llvm/IR/AutoUpgrade.h"
17 #include "llvm/IR/CFG.h"
18 #include "llvm/IR/CallSite.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/DIBuilder.h"
21 #include "llvm/IR/DebugInfo.h"
22 #include "llvm/IR/DiagnosticInfo.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/Instruction.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/LLVMContext.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include <cstring>
31 using namespace llvm;
32
33 // Upgrade the declarations of the SSE4.1 functions whose arguments have
34 // changed their type from v4f32 to v2i64.
UpgradeSSE41Function(Function * F,Intrinsic::ID IID,Function * & NewFn)35 static bool UpgradeSSE41Function(Function* F, Intrinsic::ID IID,
36 Function *&NewFn) {
37 // Check whether this is an old version of the function, which received
38 // v4f32 arguments.
39 Type *Arg0Type = F->getFunctionType()->getParamType(0);
40 if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
41 return false;
42
43 // Yes, it's old, replace it with new version.
44 F->setName(F->getName() + ".old");
45 NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
46 return true;
47 }
48
49 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
50 // arguments have changed their type from i32 to i8.
UpgradeX86IntrinsicsWith8BitMask(Function * F,Intrinsic::ID IID,Function * & NewFn)51 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
52 Function *&NewFn) {
53 // Check that the last argument is an i32.
54 Type *LastArgType = F->getFunctionType()->getParamType(
55 F->getFunctionType()->getNumParams() - 1);
56 if (!LastArgType->isIntegerTy(32))
57 return false;
58
59 // Move this function aside and map down.
60 F->setName(F->getName() + ".old");
61 NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
62 return true;
63 }
64
65 // Upgrade the declarations of AVX-512 cmp intrinsic functions whose 8-bit
66 // immediates have changed their type from i32 to i8.
UpgradeAVX512CmpIntrinsic(Function * F,Intrinsic::ID IID,Function * & NewFn)67 static bool UpgradeAVX512CmpIntrinsic(Function *F, Intrinsic::ID IID,
68 Function *&NewFn) {
69 // Check that the last argument is an i32.
70 Type *LastArgType = F->getFunctionType()->getParamType(2);
71 if (!LastArgType->isIntegerTy(32))
72 return false;
73
74 // Move this function aside and map down.
75 F->setName(F->getName() + ".old");
76 NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
77 return true;
78 }
79
UpgradeIntrinsicFunction1(Function * F,Function * & NewFn)80 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
81 assert(F && "Illegal to upgrade a non-existent Function.");
82
83 // Quickly eliminate it, if it's not a candidate.
84 StringRef Name = F->getName();
85 if (Name.size() <= 8 || !Name.startswith("llvm."))
86 return false;
87 Name = Name.substr(5); // Strip off "llvm."
88
89 switch (Name[0]) {
90 default: break;
91 case 'a': {
92 if (Name.startswith("arm.neon.vclz")) {
93 Type* args[2] = {
94 F->arg_begin()->getType(),
95 Type::getInt1Ty(F->getContext())
96 };
97 // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
98 // the end of the name. Change name from llvm.arm.neon.vclz.* to
99 // llvm.ctlz.*
100 FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
101 NewFn = Function::Create(fType, F->getLinkage(),
102 "llvm.ctlz." + Name.substr(14), F->getParent());
103 return true;
104 }
105 if (Name.startswith("arm.neon.vcnt")) {
106 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
107 F->arg_begin()->getType());
108 return true;
109 }
110 break;
111 }
112 case 'c': {
113 if (Name.startswith("ctlz.") && F->arg_size() == 1) {
114 F->setName(Name + ".old");
115 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
116 F->arg_begin()->getType());
117 return true;
118 }
119 if (Name.startswith("cttz.") && F->arg_size() == 1) {
120 F->setName(Name + ".old");
121 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
122 F->arg_begin()->getType());
123 return true;
124 }
125 break;
126 }
127
128 case 'o':
129 // We only need to change the name to match the mangling including the
130 // address space.
131 if (F->arg_size() == 2 && Name.startswith("objectsize.")) {
132 Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
133 if (F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
134 F->setName(Name + ".old");
135 NewFn = Intrinsic::getDeclaration(F->getParent(),
136 Intrinsic::objectsize, Tys);
137 return true;
138 }
139 }
140 break;
141
142 case 'x': {
143 if (Name.startswith("x86.sse2.pcmpeq.") ||
144 Name.startswith("x86.sse2.pcmpgt.") ||
145 Name.startswith("x86.avx2.pcmpeq.") ||
146 Name.startswith("x86.avx2.pcmpgt.") ||
147 Name.startswith("x86.avx.vpermil.") ||
148 Name == "x86.avx.vinsertf128.pd.256" ||
149 Name == "x86.avx.vinsertf128.ps.256" ||
150 Name == "x86.avx.vinsertf128.si.256" ||
151 Name == "x86.avx2.vinserti128" ||
152 Name == "x86.avx.vextractf128.pd.256" ||
153 Name == "x86.avx.vextractf128.ps.256" ||
154 Name == "x86.avx.vextractf128.si.256" ||
155 Name == "x86.avx2.vextracti128" ||
156 Name == "x86.avx.movnt.dq.256" ||
157 Name == "x86.avx.movnt.pd.256" ||
158 Name == "x86.avx.movnt.ps.256" ||
159 Name == "x86.sse42.crc32.64.8" ||
160 Name == "x86.avx.vbroadcast.ss" ||
161 Name == "x86.avx.vbroadcast.ss.256" ||
162 Name == "x86.avx.vbroadcast.sd.256" ||
163 Name == "x86.sse2.psll.dq" ||
164 Name == "x86.sse2.psrl.dq" ||
165 Name == "x86.avx2.psll.dq" ||
166 Name == "x86.avx2.psrl.dq" ||
167 Name == "x86.sse2.psll.dq.bs" ||
168 Name == "x86.sse2.psrl.dq.bs" ||
169 Name == "x86.avx2.psll.dq.bs" ||
170 Name == "x86.avx2.psrl.dq.bs" ||
171 Name == "x86.sse41.pblendw" ||
172 Name == "x86.sse41.blendpd" ||
173 Name == "x86.sse41.blendps" ||
174 Name == "x86.avx.blend.pd.256" ||
175 Name == "x86.avx.blend.ps.256" ||
176 Name == "x86.avx2.pblendw" ||
177 Name == "x86.avx2.pblendd.128" ||
178 Name == "x86.avx2.pblendd.256" ||
179 Name == "x86.avx2.vbroadcasti128" ||
180 (Name.startswith("x86.xop.vpcom") && F->arg_size() == 2)) {
181 NewFn = nullptr;
182 return true;
183 }
184 // SSE4.1 ptest functions may have an old signature.
185 if (Name.startswith("x86.sse41.ptest")) {
186 if (Name == "x86.sse41.ptestc")
187 return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestc, NewFn);
188 if (Name == "x86.sse41.ptestz")
189 return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestz, NewFn);
190 if (Name == "x86.sse41.ptestnzc")
191 return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
192 }
193 // Several blend and other instructions with masks used the wrong number of
194 // bits.
195 if (Name == "x86.sse41.insertps")
196 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
197 NewFn);
198 if (Name == "x86.sse41.dppd")
199 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
200 NewFn);
201 if (Name == "x86.sse41.dpps")
202 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
203 NewFn);
204 if (Name == "x86.sse41.mpsadbw")
205 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
206 NewFn);
207 if (Name == "x86.avx.dp.ps.256")
208 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
209 NewFn);
210 if (Name == "x86.avx2.mpsadbw")
211 return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
212 NewFn);
213
214 if (Name == "x86.avx512.mask.cmp.ps.512")
215 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_ps_512,
216 NewFn);
217 if (Name == "x86.avx512.mask.cmp.pd.512")
218 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_pd_512,
219 NewFn);
220
221 if (Name == "x86.avx512.mask.cmp.b.512")
222 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_b_512,
223 NewFn);
224 if (Name == "x86.avx512.mask.cmp.w.512")
225 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_w_512,
226 NewFn);
227 if (Name == "x86.avx512.mask.cmp.d.512")
228 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_d_512,
229 NewFn);
230 if (Name == "x86.avx512.mask.cmp.q.512")
231 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_q_512,
232 NewFn);
233 if (Name == "x86.avx512.mask.ucmp.b.512")
234 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_b_512,
235 NewFn);
236 if (Name == "x86.avx512.mask.ucmp.w.512")
237 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_w_512,
238 NewFn);
239 if (Name == "x86.avx512.mask.ucmp.d.512")
240 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_d_512,
241 NewFn);
242 if (Name == "x86.avx512.mask.ucmp.q.512")
243 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_q_512,
244 NewFn);
245
246 if (Name == "x86.avx512.mask.cmp.b.256")
247 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_b_256,
248 NewFn);
249 if (Name == "x86.avx512.mask.cmp.w.256")
250 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_w_256,
251 NewFn);
252 if (Name == "x86.avx512.mask.cmp.d.256")
253 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_d_256,
254 NewFn);
255 if (Name == "x86.avx512.mask.cmp.q.256")
256 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_q_256,
257 NewFn);
258 if (Name == "x86.avx512.mask.ucmp.b.256")
259 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_b_256,
260 NewFn);
261 if (Name == "x86.avx512.mask.ucmp.w.256")
262 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_w_256,
263 NewFn);
264 if (Name == "x86.avx512.mask.ucmp.d.256")
265 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_d_256,
266 NewFn);
267 if (Name == "x86.avx512.mask.ucmp.q.256")
268 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_q_256,
269 NewFn);
270
271 if (Name == "x86.avx512.mask.cmp.b.128")
272 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_b_128,
273 NewFn);
274 if (Name == "x86.avx512.mask.cmp.w.128")
275 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_w_128,
276 NewFn);
277 if (Name == "x86.avx512.mask.cmp.d.128")
278 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_d_128,
279 NewFn);
280 if (Name == "x86.avx512.mask.cmp.q.128")
281 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_q_128,
282 NewFn);
283 if (Name == "x86.avx512.mask.ucmp.b.128")
284 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_b_128,
285 NewFn);
286 if (Name == "x86.avx512.mask.ucmp.w.128")
287 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_w_128,
288 NewFn);
289 if (Name == "x86.avx512.mask.ucmp.d.128")
290 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_d_128,
291 NewFn);
292 if (Name == "x86.avx512.mask.ucmp.q.128")
293 return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_q_128,
294 NewFn);
295
296 // frcz.ss/sd may need to have an argument dropped
297 if (Name.startswith("x86.xop.vfrcz.ss") && F->arg_size() == 2) {
298 F->setName(Name + ".old");
299 NewFn = Intrinsic::getDeclaration(F->getParent(),
300 Intrinsic::x86_xop_vfrcz_ss);
301 return true;
302 }
303 if (Name.startswith("x86.xop.vfrcz.sd") && F->arg_size() == 2) {
304 F->setName(Name + ".old");
305 NewFn = Intrinsic::getDeclaration(F->getParent(),
306 Intrinsic::x86_xop_vfrcz_sd);
307 return true;
308 }
309 // Fix the FMA4 intrinsics to remove the 4
310 if (Name.startswith("x86.fma4.")) {
311 F->setName("llvm.x86.fma" + Name.substr(8));
312 NewFn = F;
313 return true;
314 }
315 break;
316 }
317 }
318
319 // This may not belong here. This function is effectively being overloaded
320 // to both detect an intrinsic which needs upgrading, and to provide the
321 // upgraded form of the intrinsic. We should perhaps have two separate
322 // functions for this.
323 return false;
324 }
325
UpgradeIntrinsicFunction(Function * F,Function * & NewFn)326 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
327 NewFn = nullptr;
328 bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
329
330 // Upgrade intrinsic attributes. This does not change the function.
331 if (NewFn)
332 F = NewFn;
333 if (unsigned id = F->getIntrinsicID())
334 F->setAttributes(Intrinsic::getAttributes(F->getContext(),
335 (Intrinsic::ID)id));
336 return Upgraded;
337 }
338
UpgradeGlobalVariable(GlobalVariable * GV)339 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
340 // Nothing to do yet.
341 return false;
342 }
343
344 // Handles upgrading SSE2 and AVX2 PSLLDQ intrinsics by converting them
345 // to byte shuffles.
UpgradeX86PSLLDQIntrinsics(IRBuilder<> & Builder,LLVMContext & C,Value * Op,unsigned NumLanes,unsigned Shift)346 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder, LLVMContext &C,
347 Value *Op, unsigned NumLanes,
348 unsigned Shift) {
349 // Each lane is 16 bytes.
350 unsigned NumElts = NumLanes * 16;
351
352 // Bitcast from a 64-bit element type to a byte element type.
353 Op = Builder.CreateBitCast(Op,
354 VectorType::get(Type::getInt8Ty(C), NumElts),
355 "cast");
356 // We'll be shuffling in zeroes.
357 Value *Res = ConstantVector::getSplat(NumElts, Builder.getInt8(0));
358
359 // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
360 // we'll just return the zero vector.
361 if (Shift < 16) {
362 SmallVector<Constant*, 32> Idxs;
363 // 256-bit version is split into two 16-byte lanes.
364 for (unsigned l = 0; l != NumElts; l += 16)
365 for (unsigned i = 0; i != 16; ++i) {
366 unsigned Idx = NumElts + i - Shift;
367 if (Idx < NumElts)
368 Idx -= NumElts - 16; // end of lane, switch operand.
369 Idxs.push_back(Builder.getInt32(Idx + l));
370 }
371
372 Res = Builder.CreateShuffleVector(Res, Op, ConstantVector::get(Idxs));
373 }
374
375 // Bitcast back to a 64-bit element type.
376 return Builder.CreateBitCast(Res,
377 VectorType::get(Type::getInt64Ty(C), 2*NumLanes),
378 "cast");
379 }
380
381 // Handles upgrading SSE2 and AVX2 PSRLDQ intrinsics by converting them
382 // to byte shuffles.
UpgradeX86PSRLDQIntrinsics(IRBuilder<> & Builder,LLVMContext & C,Value * Op,unsigned NumLanes,unsigned Shift)383 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, LLVMContext &C,
384 Value *Op, unsigned NumLanes,
385 unsigned Shift) {
386 // Each lane is 16 bytes.
387 unsigned NumElts = NumLanes * 16;
388
389 // Bitcast from a 64-bit element type to a byte element type.
390 Op = Builder.CreateBitCast(Op,
391 VectorType::get(Type::getInt8Ty(C), NumElts),
392 "cast");
393 // We'll be shuffling in zeroes.
394 Value *Res = ConstantVector::getSplat(NumElts, Builder.getInt8(0));
395
396 // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
397 // we'll just return the zero vector.
398 if (Shift < 16) {
399 SmallVector<Constant*, 32> Idxs;
400 // 256-bit version is split into two 16-byte lanes.
401 for (unsigned l = 0; l != NumElts; l += 16)
402 for (unsigned i = 0; i != 16; ++i) {
403 unsigned Idx = i + Shift;
404 if (Idx >= 16)
405 Idx += NumElts - 16; // end of lane, switch operand.
406 Idxs.push_back(Builder.getInt32(Idx + l));
407 }
408
409 Res = Builder.CreateShuffleVector(Op, Res, ConstantVector::get(Idxs));
410 }
411
412 // Bitcast back to a 64-bit element type.
413 return Builder.CreateBitCast(Res,
414 VectorType::get(Type::getInt64Ty(C), 2*NumLanes),
415 "cast");
416 }
417
418 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
419 // upgraded intrinsic. All argument and return casting must be provided in
420 // order to seamlessly integrate with existing context.
UpgradeIntrinsicCall(CallInst * CI,Function * NewFn)421 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
422 Function *F = CI->getCalledFunction();
423 LLVMContext &C = CI->getContext();
424 IRBuilder<> Builder(C);
425 Builder.SetInsertPoint(CI->getParent(), CI);
426
427 assert(F && "Intrinsic call is not direct?");
428
429 if (!NewFn) {
430 // Get the Function's name.
431 StringRef Name = F->getName();
432
433 Value *Rep;
434 // Upgrade packed integer vector compares intrinsics to compare instructions
435 if (Name.startswith("llvm.x86.sse2.pcmpeq.") ||
436 Name.startswith("llvm.x86.avx2.pcmpeq.")) {
437 Rep = Builder.CreateICmpEQ(CI->getArgOperand(0), CI->getArgOperand(1),
438 "pcmpeq");
439 // need to sign extend since icmp returns vector of i1
440 Rep = Builder.CreateSExt(Rep, CI->getType(), "");
441 } else if (Name.startswith("llvm.x86.sse2.pcmpgt.") ||
442 Name.startswith("llvm.x86.avx2.pcmpgt.")) {
443 Rep = Builder.CreateICmpSGT(CI->getArgOperand(0), CI->getArgOperand(1),
444 "pcmpgt");
445 // need to sign extend since icmp returns vector of i1
446 Rep = Builder.CreateSExt(Rep, CI->getType(), "");
447 } else if (Name == "llvm.x86.avx.movnt.dq.256" ||
448 Name == "llvm.x86.avx.movnt.ps.256" ||
449 Name == "llvm.x86.avx.movnt.pd.256") {
450 IRBuilder<> Builder(C);
451 Builder.SetInsertPoint(CI->getParent(), CI);
452
453 Module *M = F->getParent();
454 SmallVector<Metadata *, 1> Elts;
455 Elts.push_back(
456 ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
457 MDNode *Node = MDNode::get(C, Elts);
458
459 Value *Arg0 = CI->getArgOperand(0);
460 Value *Arg1 = CI->getArgOperand(1);
461
462 // Convert the type of the pointer to a pointer to the stored type.
463 Value *BC = Builder.CreateBitCast(Arg0,
464 PointerType::getUnqual(Arg1->getType()),
465 "cast");
466 StoreInst *SI = Builder.CreateStore(Arg1, BC);
467 SI->setMetadata(M->getMDKindID("nontemporal"), Node);
468 SI->setAlignment(16);
469
470 // Remove intrinsic.
471 CI->eraseFromParent();
472 return;
473 } else if (Name.startswith("llvm.x86.xop.vpcom")) {
474 Intrinsic::ID intID;
475 if (Name.endswith("ub"))
476 intID = Intrinsic::x86_xop_vpcomub;
477 else if (Name.endswith("uw"))
478 intID = Intrinsic::x86_xop_vpcomuw;
479 else if (Name.endswith("ud"))
480 intID = Intrinsic::x86_xop_vpcomud;
481 else if (Name.endswith("uq"))
482 intID = Intrinsic::x86_xop_vpcomuq;
483 else if (Name.endswith("b"))
484 intID = Intrinsic::x86_xop_vpcomb;
485 else if (Name.endswith("w"))
486 intID = Intrinsic::x86_xop_vpcomw;
487 else if (Name.endswith("d"))
488 intID = Intrinsic::x86_xop_vpcomd;
489 else if (Name.endswith("q"))
490 intID = Intrinsic::x86_xop_vpcomq;
491 else
492 llvm_unreachable("Unknown suffix");
493
494 Name = Name.substr(18); // strip off "llvm.x86.xop.vpcom"
495 unsigned Imm;
496 if (Name.startswith("lt"))
497 Imm = 0;
498 else if (Name.startswith("le"))
499 Imm = 1;
500 else if (Name.startswith("gt"))
501 Imm = 2;
502 else if (Name.startswith("ge"))
503 Imm = 3;
504 else if (Name.startswith("eq"))
505 Imm = 4;
506 else if (Name.startswith("ne"))
507 Imm = 5;
508 else if (Name.startswith("false"))
509 Imm = 6;
510 else if (Name.startswith("true"))
511 Imm = 7;
512 else
513 llvm_unreachable("Unknown condition");
514
515 Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
516 Rep = Builder.CreateCall3(VPCOM, CI->getArgOperand(0),
517 CI->getArgOperand(1), Builder.getInt8(Imm));
518 } else if (Name == "llvm.x86.sse42.crc32.64.8") {
519 Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
520 Intrinsic::x86_sse42_crc32_32_8);
521 Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
522 Rep = Builder.CreateCall2(CRC32, Trunc0, CI->getArgOperand(1));
523 Rep = Builder.CreateZExt(Rep, CI->getType(), "");
524 } else if (Name.startswith("llvm.x86.avx.vbroadcast")) {
525 // Replace broadcasts with a series of insertelements.
526 Type *VecTy = CI->getType();
527 Type *EltTy = VecTy->getVectorElementType();
528 unsigned EltNum = VecTy->getVectorNumElements();
529 Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
530 EltTy->getPointerTo());
531 Value *Load = Builder.CreateLoad(Cast);
532 Type *I32Ty = Type::getInt32Ty(C);
533 Rep = UndefValue::get(VecTy);
534 for (unsigned I = 0; I < EltNum; ++I)
535 Rep = Builder.CreateInsertElement(Rep, Load,
536 ConstantInt::get(I32Ty, I));
537 } else if (Name == "llvm.x86.avx2.vbroadcasti128") {
538 // Replace vbroadcasts with a vector shuffle.
539 Value *Op = Builder.CreatePointerCast(
540 CI->getArgOperand(0),
541 PointerType::getUnqual(VectorType::get(Type::getInt64Ty(C), 2)));
542 Value *Load = Builder.CreateLoad(Op);
543 const int Idxs[4] = { 0, 1, 0, 1 };
544 Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
545 Idxs);
546 } else if (Name == "llvm.x86.sse2.psll.dq") {
547 // 128-bit shift left specified in bits.
548 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
549 Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
550 Shift / 8); // Shift is in bits.
551 } else if (Name == "llvm.x86.sse2.psrl.dq") {
552 // 128-bit shift right specified in bits.
553 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
554 Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
555 Shift / 8); // Shift is in bits.
556 } else if (Name == "llvm.x86.avx2.psll.dq") {
557 // 256-bit shift left specified in bits.
558 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
559 Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
560 Shift / 8); // Shift is in bits.
561 } else if (Name == "llvm.x86.avx2.psrl.dq") {
562 // 256-bit shift right specified in bits.
563 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
564 Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
565 Shift / 8); // Shift is in bits.
566 } else if (Name == "llvm.x86.sse2.psll.dq.bs") {
567 // 128-bit shift left specified in bytes.
568 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
569 Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
570 Shift);
571 } else if (Name == "llvm.x86.sse2.psrl.dq.bs") {
572 // 128-bit shift right specified in bytes.
573 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
574 Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
575 Shift);
576 } else if (Name == "llvm.x86.avx2.psll.dq.bs") {
577 // 256-bit shift left specified in bytes.
578 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
579 Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
580 Shift);
581 } else if (Name == "llvm.x86.avx2.psrl.dq.bs") {
582 // 256-bit shift right specified in bytes.
583 unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
584 Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
585 Shift);
586 } else if (Name == "llvm.x86.sse41.pblendw" ||
587 Name == "llvm.x86.sse41.blendpd" ||
588 Name == "llvm.x86.sse41.blendps" ||
589 Name == "llvm.x86.avx.blend.pd.256" ||
590 Name == "llvm.x86.avx.blend.ps.256" ||
591 Name == "llvm.x86.avx2.pblendw" ||
592 Name == "llvm.x86.avx2.pblendd.128" ||
593 Name == "llvm.x86.avx2.pblendd.256") {
594 Value *Op0 = CI->getArgOperand(0);
595 Value *Op1 = CI->getArgOperand(1);
596 unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
597 VectorType *VecTy = cast<VectorType>(CI->getType());
598 unsigned NumElts = VecTy->getNumElements();
599
600 SmallVector<Constant*, 16> Idxs;
601 for (unsigned i = 0; i != NumElts; ++i) {
602 unsigned Idx = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
603 Idxs.push_back(Builder.getInt32(Idx));
604 }
605
606 Rep = Builder.CreateShuffleVector(Op0, Op1, ConstantVector::get(Idxs));
607 } else if (Name == "llvm.x86.avx.vinsertf128.pd.256" ||
608 Name == "llvm.x86.avx.vinsertf128.ps.256" ||
609 Name == "llvm.x86.avx.vinsertf128.si.256" ||
610 Name == "llvm.x86.avx2.vinserti128") {
611 Value *Op0 = CI->getArgOperand(0);
612 Value *Op1 = CI->getArgOperand(1);
613 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
614 VectorType *VecTy = cast<VectorType>(CI->getType());
615 unsigned NumElts = VecTy->getNumElements();
616
617 // Mask off the high bits of the immediate value; hardware ignores those.
618 Imm = Imm & 1;
619
620 // Extend the second operand into a vector that is twice as big.
621 Value *UndefV = UndefValue::get(Op1->getType());
622 SmallVector<Constant*, 8> Idxs;
623 for (unsigned i = 0; i != NumElts; ++i) {
624 Idxs.push_back(Builder.getInt32(i));
625 }
626 Rep = Builder.CreateShuffleVector(Op1, UndefV, ConstantVector::get(Idxs));
627
628 // Insert the second operand into the first operand.
629
630 // Note that there is no guarantee that instruction lowering will actually
631 // produce a vinsertf128 instruction for the created shuffles. In
632 // particular, the 0 immediate case involves no lane changes, so it can
633 // be handled as a blend.
634
635 // Example of shuffle mask for 32-bit elements:
636 // Imm = 1 <i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11>
637 // Imm = 0 <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6, i32 7 >
638
639 SmallVector<Constant*, 8> Idxs2;
640 // The low half of the result is either the low half of the 1st operand
641 // or the low half of the 2nd operand (the inserted vector).
642 for (unsigned i = 0; i != NumElts / 2; ++i) {
643 unsigned Idx = Imm ? i : (i + NumElts);
644 Idxs2.push_back(Builder.getInt32(Idx));
645 }
646 // The high half of the result is either the low half of the 2nd operand
647 // (the inserted vector) or the high half of the 1st operand.
648 for (unsigned i = NumElts / 2; i != NumElts; ++i) {
649 unsigned Idx = Imm ? (i + NumElts / 2) : i;
650 Idxs2.push_back(Builder.getInt32(Idx));
651 }
652 Rep = Builder.CreateShuffleVector(Op0, Rep, ConstantVector::get(Idxs2));
653 } else if (Name == "llvm.x86.avx.vextractf128.pd.256" ||
654 Name == "llvm.x86.avx.vextractf128.ps.256" ||
655 Name == "llvm.x86.avx.vextractf128.si.256" ||
656 Name == "llvm.x86.avx2.vextracti128") {
657 Value *Op0 = CI->getArgOperand(0);
658 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
659 VectorType *VecTy = cast<VectorType>(CI->getType());
660 unsigned NumElts = VecTy->getNumElements();
661
662 // Mask off the high bits of the immediate value; hardware ignores those.
663 Imm = Imm & 1;
664
665 // Get indexes for either the high half or low half of the input vector.
666 SmallVector<Constant*, 4> Idxs(NumElts);
667 for (unsigned i = 0; i != NumElts; ++i) {
668 unsigned Idx = Imm ? (i + NumElts) : i;
669 Idxs[i] = Builder.getInt32(Idx);
670 }
671
672 Value *UndefV = UndefValue::get(Op0->getType());
673 Rep = Builder.CreateShuffleVector(Op0, UndefV, ConstantVector::get(Idxs));
674 } else {
675 bool PD128 = false, PD256 = false, PS128 = false, PS256 = false;
676 if (Name == "llvm.x86.avx.vpermil.pd.256")
677 PD256 = true;
678 else if (Name == "llvm.x86.avx.vpermil.pd")
679 PD128 = true;
680 else if (Name == "llvm.x86.avx.vpermil.ps.256")
681 PS256 = true;
682 else if (Name == "llvm.x86.avx.vpermil.ps")
683 PS128 = true;
684
685 if (PD256 || PD128 || PS256 || PS128) {
686 Value *Op0 = CI->getArgOperand(0);
687 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
688 SmallVector<Constant*, 8> Idxs;
689
690 if (PD128)
691 for (unsigned i = 0; i != 2; ++i)
692 Idxs.push_back(Builder.getInt32((Imm >> i) & 0x1));
693 else if (PD256)
694 for (unsigned l = 0; l != 4; l+=2)
695 for (unsigned i = 0; i != 2; ++i)
696 Idxs.push_back(Builder.getInt32(((Imm >> (l+i)) & 0x1) + l));
697 else if (PS128)
698 for (unsigned i = 0; i != 4; ++i)
699 Idxs.push_back(Builder.getInt32((Imm >> (2 * i)) & 0x3));
700 else if (PS256)
701 for (unsigned l = 0; l != 8; l+=4)
702 for (unsigned i = 0; i != 4; ++i)
703 Idxs.push_back(Builder.getInt32(((Imm >> (2 * i)) & 0x3) + l));
704 else
705 llvm_unreachable("Unexpected function");
706
707 Rep = Builder.CreateShuffleVector(Op0, Op0, ConstantVector::get(Idxs));
708 } else {
709 llvm_unreachable("Unknown function for CallInst upgrade.");
710 }
711 }
712
713 CI->replaceAllUsesWith(Rep);
714 CI->eraseFromParent();
715 return;
716 }
717
718 std::string Name = CI->getName();
719 if (!Name.empty())
720 CI->setName(Name + ".old");
721
722 switch (NewFn->getIntrinsicID()) {
723 default:
724 llvm_unreachable("Unknown function for CallInst upgrade.");
725
726 case Intrinsic::ctlz:
727 case Intrinsic::cttz:
728 assert(CI->getNumArgOperands() == 1 &&
729 "Mismatch between function args and call args");
730 CI->replaceAllUsesWith(Builder.CreateCall2(NewFn, CI->getArgOperand(0),
731 Builder.getFalse(), Name));
732 CI->eraseFromParent();
733 return;
734
735 case Intrinsic::objectsize:
736 CI->replaceAllUsesWith(Builder.CreateCall2(NewFn,
737 CI->getArgOperand(0),
738 CI->getArgOperand(1),
739 Name));
740 CI->eraseFromParent();
741 return;
742
743 case Intrinsic::ctpop: {
744 CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(0)));
745 CI->eraseFromParent();
746 return;
747 }
748
749 case Intrinsic::x86_xop_vfrcz_ss:
750 case Intrinsic::x86_xop_vfrcz_sd:
751 CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(1),
752 Name));
753 CI->eraseFromParent();
754 return;
755
756 case Intrinsic::x86_sse41_ptestc:
757 case Intrinsic::x86_sse41_ptestz:
758 case Intrinsic::x86_sse41_ptestnzc: {
759 // The arguments for these intrinsics used to be v4f32, and changed
760 // to v2i64. This is purely a nop, since those are bitwise intrinsics.
761 // So, the only thing required is a bitcast for both arguments.
762 // First, check the arguments have the old type.
763 Value *Arg0 = CI->getArgOperand(0);
764 if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
765 return;
766
767 // Old intrinsic, add bitcasts
768 Value *Arg1 = CI->getArgOperand(1);
769
770 Value *BC0 =
771 Builder.CreateBitCast(Arg0,
772 VectorType::get(Type::getInt64Ty(C), 2),
773 "cast");
774 Value *BC1 =
775 Builder.CreateBitCast(Arg1,
776 VectorType::get(Type::getInt64Ty(C), 2),
777 "cast");
778
779 CallInst* NewCall = Builder.CreateCall2(NewFn, BC0, BC1, Name);
780 CI->replaceAllUsesWith(NewCall);
781 CI->eraseFromParent();
782 return;
783 }
784
785 case Intrinsic::x86_sse41_insertps:
786 case Intrinsic::x86_sse41_dppd:
787 case Intrinsic::x86_sse41_dpps:
788 case Intrinsic::x86_sse41_mpsadbw:
789 case Intrinsic::x86_avx_dp_ps_256:
790 case Intrinsic::x86_avx2_mpsadbw: {
791 // Need to truncate the last argument from i32 to i8 -- this argument models
792 // an inherently 8-bit immediate operand to these x86 instructions.
793 SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
794 CI->arg_operands().end());
795
796 // Replace the last argument with a trunc.
797 Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
798
799 CallInst *NewCall = Builder.CreateCall(NewFn, Args);
800 CI->replaceAllUsesWith(NewCall);
801 CI->eraseFromParent();
802 return;
803 }
804 case Intrinsic::x86_avx512_mask_cmp_ps_512:
805 case Intrinsic::x86_avx512_mask_cmp_pd_512: {
806 // Need to truncate the last argument from i32 to i8 -- this argument models
807 // an inherently 8-bit immediate operand to these x86 instructions.
808 SmallVector<Value *, 5> Args(CI->arg_operands().begin(),
809 CI->arg_operands().end());
810
811 // Replace the last argument with a trunc.
812 Args[2] = Builder.CreateTrunc(Args[2], Type::getInt8Ty(C), "trunc");
813
814 CallInst *NewCall = Builder.CreateCall(NewFn, Args);
815 CI->replaceAllUsesWith(NewCall);
816 CI->eraseFromParent();
817 return;
818 }
819 }
820 }
821
822 // This tests each Function to determine if it needs upgrading. When we find
823 // one we are interested in, we then upgrade all calls to reflect the new
824 // function.
UpgradeCallsToIntrinsic(Function * F)825 void llvm::UpgradeCallsToIntrinsic(Function* F) {
826 assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
827
828 // Upgrade the function and check if it is a totaly new function.
829 Function *NewFn;
830 if (UpgradeIntrinsicFunction(F, NewFn)) {
831 if (NewFn != F) {
832 // Replace all uses to the old function with the new one if necessary.
833 for (Value::user_iterator UI = F->user_begin(), UE = F->user_end();
834 UI != UE; ) {
835 if (CallInst *CI = dyn_cast<CallInst>(*UI++))
836 UpgradeIntrinsicCall(CI, NewFn);
837 }
838 // Remove old function, no longer used, from the module.
839 F->eraseFromParent();
840 }
841 }
842 }
843
UpgradeInstWithTBAATag(Instruction * I)844 void llvm::UpgradeInstWithTBAATag(Instruction *I) {
845 MDNode *MD = I->getMetadata(LLVMContext::MD_tbaa);
846 assert(MD && "UpgradeInstWithTBAATag should have a TBAA tag");
847 // Check if the tag uses struct-path aware TBAA format.
848 if (isa<MDNode>(MD->getOperand(0)) && MD->getNumOperands() >= 3)
849 return;
850
851 if (MD->getNumOperands() == 3) {
852 Metadata *Elts[] = {MD->getOperand(0), MD->getOperand(1)};
853 MDNode *ScalarType = MDNode::get(I->getContext(), Elts);
854 // Create a MDNode <ScalarType, ScalarType, offset 0, const>
855 Metadata *Elts2[] = {ScalarType, ScalarType,
856 ConstantAsMetadata::get(Constant::getNullValue(
857 Type::getInt64Ty(I->getContext()))),
858 MD->getOperand(2)};
859 I->setMetadata(LLVMContext::MD_tbaa, MDNode::get(I->getContext(), Elts2));
860 } else {
861 // Create a MDNode <MD, MD, offset 0>
862 Metadata *Elts[] = {MD, MD, ConstantAsMetadata::get(Constant::getNullValue(
863 Type::getInt64Ty(I->getContext())))};
864 I->setMetadata(LLVMContext::MD_tbaa, MDNode::get(I->getContext(), Elts));
865 }
866 }
867
UpgradeBitCastInst(unsigned Opc,Value * V,Type * DestTy,Instruction * & Temp)868 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
869 Instruction *&Temp) {
870 if (Opc != Instruction::BitCast)
871 return nullptr;
872
873 Temp = nullptr;
874 Type *SrcTy = V->getType();
875 if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
876 SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
877 LLVMContext &Context = V->getContext();
878
879 // We have no information about target data layout, so we assume that
880 // the maximum pointer size is 64bit.
881 Type *MidTy = Type::getInt64Ty(Context);
882 Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
883
884 return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
885 }
886
887 return nullptr;
888 }
889
UpgradeBitCastExpr(unsigned Opc,Constant * C,Type * DestTy)890 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
891 if (Opc != Instruction::BitCast)
892 return nullptr;
893
894 Type *SrcTy = C->getType();
895 if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
896 SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
897 LLVMContext &Context = C->getContext();
898
899 // We have no information about target data layout, so we assume that
900 // the maximum pointer size is 64bit.
901 Type *MidTy = Type::getInt64Ty(Context);
902
903 return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
904 DestTy);
905 }
906
907 return nullptr;
908 }
909
910 /// Check the debug info version number, if it is out-dated, drop the debug
911 /// info. Return true if module is modified.
UpgradeDebugInfo(Module & M)912 bool llvm::UpgradeDebugInfo(Module &M) {
913 unsigned Version = getDebugMetadataVersionFromModule(M);
914 if (Version == DEBUG_METADATA_VERSION)
915 return false;
916
917 bool RetCode = StripDebugInfo(M);
918 if (RetCode) {
919 DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
920 M.getContext().diagnose(DiagVersion);
921 }
922 return RetCode;
923 }
924
UpgradeMDStringConstant(std::string & String)925 void llvm::UpgradeMDStringConstant(std::string &String) {
926 const std::string OldPrefix = "llvm.vectorizer.";
927 if (String == "llvm.vectorizer.unroll") {
928 String = "llvm.loop.interleave.count";
929 } else if (String.find(OldPrefix) == 0) {
930 String.replace(0, OldPrefix.size(), "llvm.loop.vectorize.");
931 }
932 }
933