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 //
12 //===----------------------------------------------------------------------===//
13
14 #include "llvm/AutoUpgrade.h"
15 #include "llvm/Constants.h"
16 #include "llvm/Function.h"
17 #include "llvm/IRBuilder.h"
18 #include "llvm/Instruction.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/LLVMContext.h"
21 #include "llvm/Module.h"
22 #include "llvm/Support/CFG.h"
23 #include "llvm/Support/CallSite.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include <cstring>
26 using namespace llvm;
27
28 // Upgrade the declarations of the SSE4.1 functions whose arguments have
29 // changed their type from v4f32 to v2i64.
UpgradeSSE41Function(Function * F,Intrinsic::ID IID,Function * & NewFn)30 static bool UpgradeSSE41Function(Function* F, Intrinsic::ID IID,
31 Function *&NewFn) {
32 // Check whether this is an old version of the function, which received
33 // v4f32 arguments.
34 Type *Arg0Type = F->getFunctionType()->getParamType(0);
35 if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
36 return false;
37
38 // Yes, it's old, replace it with new version.
39 F->setName(F->getName() + ".old");
40 NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
41 return true;
42 }
43
UpgradeIntrinsicFunction1(Function * F,Function * & NewFn)44 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
45 assert(F && "Illegal to upgrade a non-existent Function.");
46
47 // Quickly eliminate it, if it's not a candidate.
48 StringRef Name = F->getName();
49 if (Name.size() <= 8 || !Name.startswith("llvm."))
50 return false;
51 Name = Name.substr(5); // Strip off "llvm."
52
53 switch (Name[0]) {
54 default: break;
55 case 'a': {
56 if (Name.startswith("arm.neon.vclz")) {
57 Type* args[2] = {
58 F->arg_begin()->getType(),
59 Type::getInt1Ty(F->getContext())
60 };
61 // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
62 // the end of the name. Change name from llvm.arm.neon.vclz.* to
63 // llvm.ctlz.*
64 FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
65 NewFn = Function::Create(fType, F->getLinkage(),
66 "llvm.ctlz." + Name.substr(14), F->getParent());
67 return true;
68 }
69 if (Name.startswith("arm.neon.vcnt")) {
70 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
71 F->arg_begin()->getType());
72 return true;
73 }
74 break;
75 }
76 case 'c': {
77 if (Name.startswith("ctlz.") && F->arg_size() == 1) {
78 F->setName(Name + ".old");
79 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
80 F->arg_begin()->getType());
81 return true;
82 }
83 if (Name.startswith("cttz.") && F->arg_size() == 1) {
84 F->setName(Name + ".old");
85 NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
86 F->arg_begin()->getType());
87 return true;
88 }
89 break;
90 }
91 case 'x': {
92 if (Name.startswith("x86.sse2.pcmpeq.") ||
93 Name.startswith("x86.sse2.pcmpgt.") ||
94 Name.startswith("x86.avx2.pcmpeq.") ||
95 Name.startswith("x86.avx2.pcmpgt.") ||
96 Name.startswith("x86.avx.vpermil.") ||
97 Name == "x86.avx.movnt.dq.256" ||
98 Name == "x86.avx.movnt.pd.256" ||
99 Name == "x86.avx.movnt.ps.256" ||
100 (Name.startswith("x86.xop.vpcom") && F->arg_size() == 2)) {
101 NewFn = 0;
102 return true;
103 }
104 // SSE4.1 ptest functions may have an old signature.
105 if (Name.startswith("x86.sse41.ptest")) {
106 if (Name == "x86.sse41.ptestc")
107 return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestc, NewFn);
108 if (Name == "x86.sse41.ptestz")
109 return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestz, NewFn);
110 if (Name == "x86.sse41.ptestnzc")
111 return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
112 }
113 // frcz.ss/sd may need to have an argument dropped
114 if (Name.startswith("x86.xop.vfrcz.ss") && F->arg_size() == 2) {
115 F->setName(Name + ".old");
116 NewFn = Intrinsic::getDeclaration(F->getParent(),
117 Intrinsic::x86_xop_vfrcz_ss);
118 return true;
119 }
120 if (Name.startswith("x86.xop.vfrcz.sd") && F->arg_size() == 2) {
121 F->setName(Name + ".old");
122 NewFn = Intrinsic::getDeclaration(F->getParent(),
123 Intrinsic::x86_xop_vfrcz_sd);
124 return true;
125 }
126 // Fix the FMA4 intrinsics to remove the 4
127 if (Name.startswith("x86.fma4.")) {
128 F->setName("llvm.x86.fma" + Name.substr(8));
129 NewFn = F;
130 return true;
131 }
132 break;
133 }
134 }
135
136 // This may not belong here. This function is effectively being overloaded
137 // to both detect an intrinsic which needs upgrading, and to provide the
138 // upgraded form of the intrinsic. We should perhaps have two separate
139 // functions for this.
140 return false;
141 }
142
UpgradeIntrinsicFunction(Function * F,Function * & NewFn)143 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
144 NewFn = 0;
145 bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
146
147 // Upgrade intrinsic attributes. This does not change the function.
148 if (NewFn)
149 F = NewFn;
150 if (unsigned id = F->getIntrinsicID())
151 F->setAttributes(Intrinsic::getAttributes((Intrinsic::ID)id));
152 return Upgraded;
153 }
154
UpgradeGlobalVariable(GlobalVariable * GV)155 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
156 // Nothing to do yet.
157 return false;
158 }
159
160 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
161 // upgraded intrinsic. All argument and return casting must be provided in
162 // order to seamlessly integrate with existing context.
UpgradeIntrinsicCall(CallInst * CI,Function * NewFn)163 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
164 Function *F = CI->getCalledFunction();
165 LLVMContext &C = CI->getContext();
166 IRBuilder<> Builder(C);
167 Builder.SetInsertPoint(CI->getParent(), CI);
168
169 assert(F && "Intrinsic call is not direct?");
170
171 if (!NewFn) {
172 // Get the Function's name.
173 StringRef Name = F->getName();
174
175 Value *Rep;
176 // Upgrade packed integer vector compares intrinsics to compare instructions
177 if (Name.startswith("llvm.x86.sse2.pcmpeq.") ||
178 Name.startswith("llvm.x86.avx2.pcmpeq.")) {
179 Rep = Builder.CreateICmpEQ(CI->getArgOperand(0), CI->getArgOperand(1),
180 "pcmpeq");
181 // need to sign extend since icmp returns vector of i1
182 Rep = Builder.CreateSExt(Rep, CI->getType(), "");
183 } else if (Name.startswith("llvm.x86.sse2.pcmpgt.") ||
184 Name.startswith("llvm.x86.avx2.pcmpgt.")) {
185 Rep = Builder.CreateICmpSGT(CI->getArgOperand(0), CI->getArgOperand(1),
186 "pcmpgt");
187 // need to sign extend since icmp returns vector of i1
188 Rep = Builder.CreateSExt(Rep, CI->getType(), "");
189 } else if (Name == "llvm.x86.avx.movnt.dq.256" ||
190 Name == "llvm.x86.avx.movnt.ps.256" ||
191 Name == "llvm.x86.avx.movnt.pd.256") {
192 IRBuilder<> Builder(C);
193 Builder.SetInsertPoint(CI->getParent(), CI);
194
195 Module *M = F->getParent();
196 SmallVector<Value *, 1> Elts;
197 Elts.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
198 MDNode *Node = MDNode::get(C, Elts);
199
200 Value *Arg0 = CI->getArgOperand(0);
201 Value *Arg1 = CI->getArgOperand(1);
202
203 // Convert the type of the pointer to a pointer to the stored type.
204 Value *BC = Builder.CreateBitCast(Arg0,
205 PointerType::getUnqual(Arg1->getType()),
206 "cast");
207 StoreInst *SI = Builder.CreateStore(Arg1, BC);
208 SI->setMetadata(M->getMDKindID("nontemporal"), Node);
209 SI->setAlignment(16);
210
211 // Remove intrinsic.
212 CI->eraseFromParent();
213 return;
214 } else if (Name.startswith("llvm.x86.xop.vpcom")) {
215 Intrinsic::ID intID;
216 if (Name.endswith("ub"))
217 intID = Intrinsic::x86_xop_vpcomub;
218 else if (Name.endswith("uw"))
219 intID = Intrinsic::x86_xop_vpcomuw;
220 else if (Name.endswith("ud"))
221 intID = Intrinsic::x86_xop_vpcomud;
222 else if (Name.endswith("uq"))
223 intID = Intrinsic::x86_xop_vpcomuq;
224 else if (Name.endswith("b"))
225 intID = Intrinsic::x86_xop_vpcomb;
226 else if (Name.endswith("w"))
227 intID = Intrinsic::x86_xop_vpcomw;
228 else if (Name.endswith("d"))
229 intID = Intrinsic::x86_xop_vpcomd;
230 else if (Name.endswith("q"))
231 intID = Intrinsic::x86_xop_vpcomq;
232 else
233 llvm_unreachable("Unknown suffix");
234
235 Name = Name.substr(18); // strip off "llvm.x86.xop.vpcom"
236 unsigned Imm;
237 if (Name.startswith("lt"))
238 Imm = 0;
239 else if (Name.startswith("le"))
240 Imm = 1;
241 else if (Name.startswith("gt"))
242 Imm = 2;
243 else if (Name.startswith("ge"))
244 Imm = 3;
245 else if (Name.startswith("eq"))
246 Imm = 4;
247 else if (Name.startswith("ne"))
248 Imm = 5;
249 else if (Name.startswith("true"))
250 Imm = 6;
251 else if (Name.startswith("false"))
252 Imm = 7;
253 else
254 llvm_unreachable("Unknown condition");
255
256 Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
257 Rep = Builder.CreateCall3(VPCOM, CI->getArgOperand(0),
258 CI->getArgOperand(1), Builder.getInt8(Imm));
259 } else {
260 bool PD128 = false, PD256 = false, PS128 = false, PS256 = false;
261 if (Name == "llvm.x86.avx.vpermil.pd.256")
262 PD256 = true;
263 else if (Name == "llvm.x86.avx.vpermil.pd")
264 PD128 = true;
265 else if (Name == "llvm.x86.avx.vpermil.ps.256")
266 PS256 = true;
267 else if (Name == "llvm.x86.avx.vpermil.ps")
268 PS128 = true;
269
270 if (PD256 || PD128 || PS256 || PS128) {
271 Value *Op0 = CI->getArgOperand(0);
272 unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
273 SmallVector<Constant*, 8> Idxs;
274
275 if (PD128)
276 for (unsigned i = 0; i != 2; ++i)
277 Idxs.push_back(Builder.getInt32((Imm >> i) & 0x1));
278 else if (PD256)
279 for (unsigned l = 0; l != 4; l+=2)
280 for (unsigned i = 0; i != 2; ++i)
281 Idxs.push_back(Builder.getInt32(((Imm >> (l+i)) & 0x1) + l));
282 else if (PS128)
283 for (unsigned i = 0; i != 4; ++i)
284 Idxs.push_back(Builder.getInt32((Imm >> (2 * i)) & 0x3));
285 else if (PS256)
286 for (unsigned l = 0; l != 8; l+=4)
287 for (unsigned i = 0; i != 4; ++i)
288 Idxs.push_back(Builder.getInt32(((Imm >> (2 * i)) & 0x3) + l));
289 else
290 llvm_unreachable("Unexpected function");
291
292 Rep = Builder.CreateShuffleVector(Op0, Op0, ConstantVector::get(Idxs));
293 } else {
294 llvm_unreachable("Unknown function for CallInst upgrade.");
295 }
296 }
297
298 CI->replaceAllUsesWith(Rep);
299 CI->eraseFromParent();
300 return;
301 }
302
303 std::string Name = CI->getName().str();
304 CI->setName(Name + ".old");
305
306 switch (NewFn->getIntrinsicID()) {
307 default:
308 llvm_unreachable("Unknown function for CallInst upgrade.");
309
310 case Intrinsic::ctlz:
311 case Intrinsic::cttz:
312 assert(CI->getNumArgOperands() == 1 &&
313 "Mismatch between function args and call args");
314 CI->replaceAllUsesWith(Builder.CreateCall2(NewFn, CI->getArgOperand(0),
315 Builder.getFalse(), Name));
316 CI->eraseFromParent();
317 return;
318
319 case Intrinsic::arm_neon_vclz: {
320 // Change name from llvm.arm.neon.vclz.* to llvm.ctlz.*
321 CI->replaceAllUsesWith(Builder.CreateCall2(NewFn, CI->getArgOperand(0),
322 Builder.getFalse(),
323 "llvm.ctlz." + Name.substr(14)));
324 CI->eraseFromParent();
325 return;
326 }
327 case Intrinsic::ctpop: {
328 CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(0)));
329 CI->eraseFromParent();
330 return;
331 }
332
333 case Intrinsic::x86_xop_vfrcz_ss:
334 case Intrinsic::x86_xop_vfrcz_sd:
335 CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(1),
336 Name));
337 CI->eraseFromParent();
338 return;
339
340 case Intrinsic::x86_sse41_ptestc:
341 case Intrinsic::x86_sse41_ptestz:
342 case Intrinsic::x86_sse41_ptestnzc: {
343 // The arguments for these intrinsics used to be v4f32, and changed
344 // to v2i64. This is purely a nop, since those are bitwise intrinsics.
345 // So, the only thing required is a bitcast for both arguments.
346 // First, check the arguments have the old type.
347 Value *Arg0 = CI->getArgOperand(0);
348 if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
349 return;
350
351 // Old intrinsic, add bitcasts
352 Value *Arg1 = CI->getArgOperand(1);
353
354 Value *BC0 =
355 Builder.CreateBitCast(Arg0,
356 VectorType::get(Type::getInt64Ty(C), 2),
357 "cast");
358 Value *BC1 =
359 Builder.CreateBitCast(Arg1,
360 VectorType::get(Type::getInt64Ty(C), 2),
361 "cast");
362
363 CallInst* NewCall = Builder.CreateCall2(NewFn, BC0, BC1, Name);
364 CI->replaceAllUsesWith(NewCall);
365 CI->eraseFromParent();
366 return;
367 }
368 }
369 }
370
371 // This tests each Function to determine if it needs upgrading. When we find
372 // one we are interested in, we then upgrade all calls to reflect the new
373 // function.
UpgradeCallsToIntrinsic(Function * F)374 void llvm::UpgradeCallsToIntrinsic(Function* F) {
375 assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
376
377 // Upgrade the function and check if it is a totaly new function.
378 Function *NewFn;
379 if (UpgradeIntrinsicFunction(F, NewFn)) {
380 if (NewFn != F) {
381 // Replace all uses to the old function with the new one if necessary.
382 for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
383 UI != UE; ) {
384 if (CallInst *CI = dyn_cast<CallInst>(*UI++))
385 UpgradeIntrinsicCall(CI, NewFn);
386 }
387 // Remove old function, no longer used, from the module.
388 F->eraseFromParent();
389 }
390 }
391 }
392
393