lib/IR/AutoUpgrade.cpp

   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/IR/AutoUpgrade.h"
  15 #include "llvm/IR/CFG.h"
  16 #include "llvm/IR/CallSite.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/Support/ErrorHandling.h"
  28 #include <cstring>
  29 using namespace llvm;
  30
  31 // Upgrade the declarations of the SSE4.1 functions whose arguments have
  32 // changed their type from v4f32 to v2i64.
  33 static bool UpgradeSSE41Function(Function* F, Intrinsic::ID IID,
  34                                  Function *&NewFn) {
  35   // Check whether this is an old version of the function, which received
  36   // v4f32 arguments.
  37   Type *Arg0Type = F->getFunctionType()->getParamType(0);
  38   if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
  39     return false;
  40
  41   // Yes, it's old, replace it with new version.
  42   F->setName(F->getName() + ".old");
  43   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
  44   return true;
  45 }
  46
  47 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
  48 // arguments have changed their type from i32 to i8.
  49 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
  50                                              Function *&NewFn) {
  51   // Check that the last argument is an i32.
  52   Type *LastArgType = F->getFunctionType()->getParamType(
  53      F->getFunctionType()->getNumParams() - 1);
  54   if (!LastArgType->isIntegerTy(32))
  55     return false;
  56
  57   // Move this function aside and map down.
  58   F->setName(F->getName() + ".old");
  59   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
  60   return true;
  61 }
  62
  63 // Upgrade the declarations of AVX-512 cmp intrinsic functions whose 8-bit
  64 // immediates have changed their type from i32 to i8.
  65 static bool UpgradeAVX512CmpIntrinsic(Function *F, Intrinsic::ID IID,
  66                                       Function *&NewFn) {
  67   // Check that the last argument is an i32.
  68   Type *LastArgType = F->getFunctionType()->getParamType(2);
  69   if (!LastArgType->isIntegerTy(32))
  70     return false;
  71
  72   // Move this function aside and map down.
  73   F->setName(F->getName() + ".old");
  74   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
  75   return true;
  76 }
  77
  78 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
  79   assert(F && "Illegal to upgrade a non-existent Function.");
  80
  81   // Quickly eliminate it, if it's not a candidate.
  82   StringRef Name = F->getName();
  83   if (Name.size() <= 8 || !Name.startswith("llvm."))
  84     return false;
  85   Name = Name.substr(5); // Strip off "llvm."
  86
  87   switch (Name[0]) {
  88   default: break;
  89   case 'a': {
  90     if (Name.startswith("arm.neon.vclz")) {
  91       Type* args[2] = {
  92         F->arg_begin()->getType(),
  93         Type::getInt1Ty(F->getContext())
  94       };
  95       // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
  96       // the end of the name. Change name from llvm.arm.neon.vclz.* to
  97       //  llvm.ctlz.*
  98       FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
  99       NewFn = Function::Create(fType, F->getLinkage(),
 100                                "llvm.ctlz." + Name.substr(14), F->getParent());
 101       return true;
 102     }
 103     if (Name.startswith("arm.neon.vcnt")) {
 104       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
 105                                         F->arg_begin()->getType());
 106       return true;
 107     }
 108     break;
 109   }
 110   case 'c': {
 111     if (Name.startswith("ctlz.") && F->arg_size() == 1) {
 112       F->setName(Name + ".old");
 113       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
 114                                         F->arg_begin()->getType());
 115       return true;
 116     }
 117     if (Name.startswith("cttz.") && F->arg_size() == 1) {
 118       F->setName(Name + ".old");
 119       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
 120                                         F->arg_begin()->getType());
 121       return true;
 122     }
 123     break;
 124   }
 125   case 'd': {
 126     if (Name.startswith("dbg.declare") && F->arg_size() == 2) {
 127       F->setName(Name + ".old");
 128       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_declare);
 129       return true;
 130     }
 131     if (Name.startswith("dbg.value") && F->arg_size() == 3) {
 132       F->setName(Name + ".old");
 133       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_value);
 134       return true;
 135     }
 136     break;
 137   }
 138
 139   case 'o':
 140     // We only need to change the name to match the mangling including the
 141     // address space.
 142     if (F->arg_size() == 2 && Name.startswith("objectsize.")) {
 143       Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
 144       if (F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
 145         F->setName(Name + ".old");
 146         NewFn = Intrinsic::getDeclaration(F->getParent(),
 147                                           Intrinsic::objectsize, Tys);
 148         return true;
 149       }
 150     }
 151     break;
 152
 153   case 'x': {
 154     if (Name.startswith("x86.sse2.pcmpeq.") ||
 155         Name.startswith("x86.sse2.pcmpgt.") ||
 156         Name.startswith("x86.avx2.pcmpeq.") ||
 157         Name.startswith("x86.avx2.pcmpgt.") ||
 158         Name.startswith("x86.avx.vpermil.") ||
 159         Name == "x86.avx.movnt.dq.256" ||
 160         Name == "x86.avx.movnt.pd.256" ||
 161         Name == "x86.avx.movnt.ps.256" ||
 162         Name == "x86.sse42.crc32.64.8" ||
 163         Name == "x86.avx.vbroadcast.ss" ||
 164         Name == "x86.avx.vbroadcast.ss.256" ||
 165         Name == "x86.avx.vbroadcast.sd.256" ||
 166         Name == "x86.sse2.psll.dq" ||
 167         Name == "x86.sse2.psrl.dq" ||
 168         Name == "x86.avx2.psll.dq" ||
 169         Name == "x86.avx2.psrl.dq" ||
 170         Name == "x86.sse2.psll.dq.bs" ||
 171         Name == "x86.sse2.psrl.dq.bs" ||
 172         Name == "x86.avx2.psll.dq.bs" ||
 173         Name == "x86.avx2.psrl.dq.bs" ||
 174         (Name.startswith("x86.xop.vpcom") && F->arg_size() == 2)) {
 175       NewFn = nullptr;
 176       return true;
 177     }
 178     // SSE4.1 ptest functions may have an old signature.
 179     if (Name.startswith("x86.sse41.ptest")) {
 180       if (Name == "x86.sse41.ptestc")
 181         return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestc, NewFn);
 182       if (Name == "x86.sse41.ptestz")
 183         return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestz, NewFn);
 184       if (Name == "x86.sse41.ptestnzc")
 185         return UpgradeSSE41Function(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
 186     }
 187     // Several blend and other instructions with maskes used the wrong number of
 188     // bits.
 189     if (Name == "x86.sse41.pblendw")
 190       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_pblendw,
 191                                               NewFn);
 192     if (Name == "x86.sse41.blendpd")
 193       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_blendpd,
 194                                               NewFn);
 195     if (Name == "x86.sse41.blendps")
 196       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_blendps,
 197                                               NewFn);
 198     if (Name == "x86.sse41.insertps")
 199       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
 200                                               NewFn);
 201     if (Name == "x86.sse41.dppd")
 202       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
 203                                               NewFn);
 204     if (Name == "x86.sse41.dpps")
 205       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
 206                                               NewFn);
 207     if (Name == "x86.sse41.mpsadbw")
 208       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
 209                                               NewFn);
 210     if (Name == "x86.avx.blend.pd.256")
 211       return UpgradeX86IntrinsicsWith8BitMask(
 212           F, Intrinsic::x86_avx_blend_pd_256, NewFn);
 213     if (Name == "x86.avx.blend.ps.256")
 214       return UpgradeX86IntrinsicsWith8BitMask(
 215           F, Intrinsic::x86_avx_blend_ps_256, NewFn);
 216     if (Name == "x86.avx.dp.ps.256")
 217       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
 218                                               NewFn);
 219     if (Name == "x86.avx2.pblendw")
 220       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_pblendw,
 221                                               NewFn);
 222     if (Name == "x86.avx2.pblendd.128")
 223       return UpgradeX86IntrinsicsWith8BitMask(
 224           F, Intrinsic::x86_avx2_pblendd_128, NewFn);
 225     if (Name == "x86.avx2.pblendd.256")
 226       return UpgradeX86IntrinsicsWith8BitMask(
 227           F, Intrinsic::x86_avx2_pblendd_256, NewFn);
 228     if (Name == "x86.avx2.mpsadbw")
 229       return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
 230                                               NewFn);
 231
 232     if (Name == "x86.avx512.mask.cmp.ps.512")
 233       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_ps_512,
 234                                        NewFn);
 235     if (Name == "x86.avx512.mask.cmp.pd.512")
 236       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_pd_512,
 237                                        NewFn);
 238
 239     if (Name == "x86.avx512.mask.cmp.b.512")
 240       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_b_512,
 241                                        NewFn);
 242     if (Name == "x86.avx512.mask.cmp.w.512")
 243       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_w_512,
 244                                        NewFn);
 245     if (Name == "x86.avx512.mask.cmp.d.512")
 246       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_d_512,
 247                                        NewFn);
 248     if (Name == "x86.avx512.mask.cmp.q.512")
 249       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_q_512,
 250                                        NewFn);
 251     if (Name == "x86.avx512.mask.ucmp.b.512")
 252       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_b_512,
 253                                        NewFn);
 254     if (Name == "x86.avx512.mask.ucmp.w.512")
 255       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_w_512,
 256                                        NewFn);
 257     if (Name == "x86.avx512.mask.ucmp.d.512")
 258       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_d_512,
 259                                        NewFn);
 260     if (Name == "x86.avx512.mask.ucmp.q.512")
 261       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_q_512,
 262                                        NewFn);
 263
 264     if (Name == "x86.avx512.mask.cmp.b.256")
 265       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_b_256,
 266                                        NewFn);
 267     if (Name == "x86.avx512.mask.cmp.w.256")
 268       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_w_256,
 269                                        NewFn);
 270     if (Name == "x86.avx512.mask.cmp.d.256")
 271       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_d_256,
 272                                        NewFn);
 273     if (Name == "x86.avx512.mask.cmp.q.256")
 274       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_q_256,
 275                                        NewFn);
 276     if (Name == "x86.avx512.mask.ucmp.b.256")
 277       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_b_256,
 278                                        NewFn);
 279     if (Name == "x86.avx512.mask.ucmp.w.256")
 280       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_w_256,
 281                                        NewFn);
 282     if (Name == "x86.avx512.mask.ucmp.d.256")
 283       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_d_256,
 284                                        NewFn);
 285     if (Name == "x86.avx512.mask.ucmp.q.256")
 286       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_q_256,
 287                                        NewFn);
 288
 289     if (Name == "x86.avx512.mask.cmp.b.128")
 290       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_b_128,
 291                                        NewFn);
 292     if (Name == "x86.avx512.mask.cmp.w.128")
 293       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_w_128,
 294                                        NewFn);
 295     if (Name == "x86.avx512.mask.cmp.d.128")
 296       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_d_128,
 297                                        NewFn);
 298     if (Name == "x86.avx512.mask.cmp.q.128")
 299       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_cmp_q_128,
 300                                        NewFn);
 301     if (Name == "x86.avx512.mask.ucmp.b.128")
 302       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_b_128,
 303                                        NewFn);
 304     if (Name == "x86.avx512.mask.ucmp.w.128")
 305       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_w_128,
 306                                        NewFn);
 307     if (Name == "x86.avx512.mask.ucmp.d.128")
 308       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_d_128,
 309                                        NewFn);
 310     if (Name == "x86.avx512.mask.ucmp.q.128")
 311       return UpgradeAVX512CmpIntrinsic(F, Intrinsic::x86_avx512_mask_ucmp_q_128,
 312                                        NewFn);
 313
 314     // frcz.ss/sd may need to have an argument dropped
 315     if (Name.startswith("x86.xop.vfrcz.ss") && F->arg_size() == 2) {
 316       F->setName(Name + ".old");
 317       NewFn = Intrinsic::getDeclaration(F->getParent(),
 318                                         Intrinsic::x86_xop_vfrcz_ss);
 319       return true;
 320     }
 321     if (Name.startswith("x86.xop.vfrcz.sd") && F->arg_size() == 2) {
 322       F->setName(Name + ".old");
 323       NewFn = Intrinsic::getDeclaration(F->getParent(),
 324                                         Intrinsic::x86_xop_vfrcz_sd);
 325       return true;
 326     }
 327     // Fix the FMA4 intrinsics to remove the 4
 328     if (Name.startswith("x86.fma4.")) {
 329       F->setName("llvm.x86.fma" + Name.substr(8));
 330       NewFn = F;
 331       return true;
 332     }
 333     break;
 334   }
 335   }
 336
 337   //  This may not belong here. This function is effectively being overloaded
 338   //  to both detect an intrinsic which needs upgrading, and to provide the
 339   //  upgraded form of the intrinsic. We should perhaps have two separate
 340   //  functions for this.
 341   return false;
 342 }
 343
 344 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
 345   NewFn = nullptr;
 346   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
 347
 348   // Upgrade intrinsic attributes.  This does not change the function.
 349   if (NewFn)
 350     F = NewFn;
 351   if (unsigned id = F->getIntrinsicID())
 352     F->setAttributes(Intrinsic::getAttributes(F->getContext(),
 353                                               (Intrinsic::ID)id));
 354   return Upgraded;
 355 }
 356
 357 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
 358   // Nothing to do yet.
 359   return false;
 360 }
 361
 362 static MDNode *getNodeField(const MDNode *DbgNode, unsigned Elt) {
 363   if (!DbgNode || Elt >= DbgNode->getNumOperands())
 364     return nullptr;
 365   return dyn_cast_or_null<MDNode>(DbgNode->getOperand(Elt));
 366 }
 367
 368 static MetadataAsValue *getExpression(Value *VarOperand, Function *F) {
 369   // Old-style DIVariables have an optional expression as the 8th element.
 370   DIExpression Expr(getNodeField(
 371       cast<MDNode>(cast<MetadataAsValue>(VarOperand)->getMetadata()), 8));
 372   if (!Expr) {
 373     DIBuilder DIB(*F->getParent(), /*AllowUnresolved*/ false);
 374     Expr = DIB.createExpression();
 375   }
 376   return MetadataAsValue::get(F->getContext(), Expr);
 377 }
 378
 379 // Handles upgrading SSE2 and AVX2 PSLLDQ intrinsics by converting them
 380 // to byte shuffles.
 381 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder, LLVMContext &C,
 382                                          Value *Op, unsigned NumLanes,
 383                                          unsigned Shift) {
 384   // Each lane is 16 bytes.
 385   unsigned NumElts = NumLanes * 16;
 386
 387   // Bitcast from a 64-bit element type to a byte element type.
 388   Op = Builder.CreateBitCast(Op,
 389                              VectorType::get(Type::getInt8Ty(C), NumElts),
 390                              "cast");
 391   // We'll be shuffling in zeroes.
 392   Value *Res = ConstantVector::getSplat(NumElts, Builder.getInt8(0));
 393
 394   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
 395   // we'll just return the zero vector.
 396   if (Shift < 16) {
 397     SmallVector<Constant*, 32> Idxs;
 398     // 256-bit version is split into two 16-byte lanes.
 399     for (unsigned l = 0; l != NumElts; l += 16)
 400       for (unsigned i = 0; i != 16; ++i) {
 401         unsigned Idx = NumElts + i - Shift;
 402         if (Idx < NumElts)
 403           Idx -= NumElts - 16; // end of lane, switch operand.
 404         Idxs.push_back(Builder.getInt32(Idx + l));
 405       }
 406
 407     Res = Builder.CreateShuffleVector(Res, Op, ConstantVector::get(Idxs));
 408   }
 409
 410   // Bitcast back to a 64-bit element type.
 411   return Builder.CreateBitCast(Res,
 412                                VectorType::get(Type::getInt64Ty(C), 2*NumLanes),
 413                                "cast");
 414 }
 415
 416 // Handles upgrading SSE2 and AVX2 PSRLDQ intrinsics by converting them
 417 // to byte shuffles.
 418 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, LLVMContext &C,
 419                                          Value *Op, unsigned NumLanes,
 420                                          unsigned Shift) {
 421   // Each lane is 16 bytes.
 422   unsigned NumElts = NumLanes * 16;
 423
 424   // Bitcast from a 64-bit element type to a byte element type.
 425   Op = Builder.CreateBitCast(Op,
 426                              VectorType::get(Type::getInt8Ty(C), NumElts),
 427                              "cast");
 428   // We'll be shuffling in zeroes.
 429   Value *Res = ConstantVector::getSplat(NumElts, Builder.getInt8(0));
 430
 431   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
 432   // we'll just return the zero vector.
 433   if (Shift < 16) {
 434     SmallVector<Constant*, 32> Idxs;
 435     // 256-bit version is split into two 16-byte lanes.
 436     for (unsigned l = 0; l != NumElts; l += 16)
 437       for (unsigned i = 0; i != 16; ++i) {
 438         unsigned Idx = i + Shift;
 439         if (Idx >= 16)
 440           Idx += NumElts - 16; // end of lane, switch operand.
 441         Idxs.push_back(Builder.getInt32(Idx + l));
 442       }
 443
 444     Res = Builder.CreateShuffleVector(Op, Res, ConstantVector::get(Idxs));
 445   }
 446
 447   // Bitcast back to a 64-bit element type.
 448   return Builder.CreateBitCast(Res,
 449                                VectorType::get(Type::getInt64Ty(C), 2*NumLanes),
 450                                "cast");
 451 }
 452
 453 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
 454 // upgraded intrinsic. All argument and return casting must be provided in
 455 // order to seamlessly integrate with existing context.
 456 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
 457   Function *F = CI->getCalledFunction();
 458   LLVMContext &C = CI->getContext();
 459   IRBuilder<> Builder(C);
 460   Builder.SetInsertPoint(CI->getParent(), CI);
 461
 462   assert(F && "Intrinsic call is not direct?");
 463
 464   if (!NewFn) {
 465     // Get the Function's name.
 466     StringRef Name = F->getName();
 467
 468     Value *Rep;
 469     // Upgrade packed integer vector compares intrinsics to compare instructions
 470     if (Name.startswith("llvm.x86.sse2.pcmpeq.") ||
 471         Name.startswith("llvm.x86.avx2.pcmpeq.")) {
 472       Rep = Builder.CreateICmpEQ(CI->getArgOperand(0), CI->getArgOperand(1),
 473                                  "pcmpeq");
 474       // need to sign extend since icmp returns vector of i1
 475       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
 476     } else if (Name.startswith("llvm.x86.sse2.pcmpgt.") ||
 477                Name.startswith("llvm.x86.avx2.pcmpgt.")) {
 478       Rep = Builder.CreateICmpSGT(CI->getArgOperand(0), CI->getArgOperand(1),
 479                                   "pcmpgt");
 480       // need to sign extend since icmp returns vector of i1
 481       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
 482     } else if (Name == "llvm.x86.avx.movnt.dq.256" ||
 483                Name == "llvm.x86.avx.movnt.ps.256" ||
 484                Name == "llvm.x86.avx.movnt.pd.256") {
 485       IRBuilder<> Builder(C);
 486       Builder.SetInsertPoint(CI->getParent(), CI);
 487
 488       Module *M = F->getParent();
 489       SmallVector<Metadata *, 1> Elts;
 490       Elts.push_back(
 491           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
 492       MDNode *Node = MDNode::get(C, Elts);
 493
 494       Value *Arg0 = CI->getArgOperand(0);
 495       Value *Arg1 = CI->getArgOperand(1);
 496
 497       // Convert the type of the pointer to a pointer to the stored type.
 498       Value *BC = Builder.CreateBitCast(Arg0,
 499                                         PointerType::getUnqual(Arg1->getType()),
 500                                         "cast");
 501       StoreInst *SI = Builder.CreateStore(Arg1, BC);
 502       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
 503       SI->setAlignment(16);
 504
 505       // Remove intrinsic.
 506       CI->eraseFromParent();
 507       return;
 508     } else if (Name.startswith("llvm.x86.xop.vpcom")) {
 509       Intrinsic::ID intID;
 510       if (Name.endswith("ub"))
 511         intID = Intrinsic::x86_xop_vpcomub;
 512       else if (Name.endswith("uw"))
 513         intID = Intrinsic::x86_xop_vpcomuw;
 514       else if (Name.endswith("ud"))
 515         intID = Intrinsic::x86_xop_vpcomud;
 516       else if (Name.endswith("uq"))
 517         intID = Intrinsic::x86_xop_vpcomuq;
 518       else if (Name.endswith("b"))
 519         intID = Intrinsic::x86_xop_vpcomb;
 520       else if (Name.endswith("w"))
 521         intID = Intrinsic::x86_xop_vpcomw;
 522       else if (Name.endswith("d"))
 523         intID = Intrinsic::x86_xop_vpcomd;
 524       else if (Name.endswith("q"))
 525         intID = Intrinsic::x86_xop_vpcomq;
 526       else
 527         llvm_unreachable("Unknown suffix");
 528
 529       Name = Name.substr(18); // strip off "llvm.x86.xop.vpcom"
 530       unsigned Imm;
 531       if (Name.startswith("lt"))
 532         Imm = 0;
 533       else if (Name.startswith("le"))
 534         Imm = 1;
 535       else if (Name.startswith("gt"))
 536         Imm = 2;
 537       else if (Name.startswith("ge"))
 538         Imm = 3;
 539       else if (Name.startswith("eq"))
 540         Imm = 4;
 541       else if (Name.startswith("ne"))
 542         Imm = 5;
 543       else if (Name.startswith("false"))
 544         Imm = 6;
 545       else if (Name.startswith("true"))
 546         Imm = 7;
 547       else
 548         llvm_unreachable("Unknown condition");
 549
 550       Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
 551       Rep = Builder.CreateCall3(VPCOM, CI->getArgOperand(0),
 552                                 CI->getArgOperand(1), Builder.getInt8(Imm));
 553     } else if (Name == "llvm.x86.sse42.crc32.64.8") {
 554       Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
 555                                                Intrinsic::x86_sse42_crc32_32_8);
 556       Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
 557       Rep = Builder.CreateCall2(CRC32, Trunc0, CI->getArgOperand(1));
 558       Rep = Builder.CreateZExt(Rep, CI->getType(), "");
 559     } else if (Name.startswith("llvm.x86.avx.vbroadcast")) {
 560       // Replace broadcasts with a series of insertelements.
 561       Type *VecTy = CI->getType();
 562       Type *EltTy = VecTy->getVectorElementType();
 563       unsigned EltNum = VecTy->getVectorNumElements();
 564       Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
 565                                           EltTy->getPointerTo());
 566       Value *Load = Builder.CreateLoad(Cast);
 567       Type *I32Ty = Type::getInt32Ty(C);
 568       Rep = UndefValue::get(VecTy);
 569       for (unsigned I = 0; I < EltNum; ++I)
 570         Rep = Builder.CreateInsertElement(Rep, Load,
 571                                           ConstantInt::get(I32Ty, I));
 572     } else if (Name == "llvm.x86.sse2.psll.dq") {
 573       // 128-bit shift left specified in bits.
 574       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 575       Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
 576                                        Shift / 8); // Shift is in bits.
 577     } else if (Name == "llvm.x86.sse2.psrl.dq") {
 578       // 128-bit shift right specified in bits.
 579       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 580       Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
 581                                        Shift / 8); // Shift is in bits.
 582     } else if (Name == "llvm.x86.avx2.psll.dq") {
 583       // 256-bit shift left specified in bits.
 584       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 585       Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
 586                                        Shift / 8); // Shift is in bits.
 587     } else if (Name == "llvm.x86.avx2.psrl.dq") {
 588       // 256-bit shift right specified in bits.
 589       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 590       Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
 591                                        Shift / 8); // Shift is in bits.
 592     } else if (Name == "llvm.x86.sse2.psll.dq.bs") {
 593       // 128-bit shift left specified in bytes.
 594       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 595       Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
 596                                        Shift);
 597     } else if (Name == "llvm.x86.sse2.psrl.dq.bs") {
 598       // 128-bit shift right specified in bytes.
 599       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 600       Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 1,
 601                                        Shift);
 602     } else if (Name == "llvm.x86.avx2.psll.dq.bs") {
 603       // 256-bit shift left specified in bytes.
 604       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 605       Rep = UpgradeX86PSLLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
 606                                        Shift);
 607     } else if (Name == "llvm.x86.avx2.psrl.dq.bs") {
 608       // 256-bit shift right specified in bytes.
 609       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 610       Rep = UpgradeX86PSRLDQIntrinsics(Builder, C, CI->getArgOperand(0), 2,
 611                                        Shift);
 612     } else {
 613       bool PD128 = false, PD256 = false, PS128 = false, PS256 = false;
 614       if (Name == "llvm.x86.avx.vpermil.pd.256")
 615         PD256 = true;
 616       else if (Name == "llvm.x86.avx.vpermil.pd")
 617         PD128 = true;
 618       else if (Name == "llvm.x86.avx.vpermil.ps.256")
 619         PS256 = true;
 620       else if (Name == "llvm.x86.avx.vpermil.ps")
 621         PS128 = true;
 622
 623       if (PD256 || PD128 || PS256 || PS128) {
 624         Value *Op0 = CI->getArgOperand(0);
 625         unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
 626         SmallVector<Constant*, 8> Idxs;
 627
 628         if (PD128)
 629           for (unsigned i = 0; i != 2; ++i)
 630             Idxs.push_back(Builder.getInt32((Imm >> i) & 0x1));
 631         else if (PD256)
 632           for (unsigned l = 0; l != 4; l+=2)
 633             for (unsigned i = 0; i != 2; ++i)
 634               Idxs.push_back(Builder.getInt32(((Imm >> (l+i)) & 0x1) + l));
 635         else if (PS128)
 636           for (unsigned i = 0; i != 4; ++i)
 637             Idxs.push_back(Builder.getInt32((Imm >> (2 * i)) & 0x3));
 638         else if (PS256)
 639           for (unsigned l = 0; l != 8; l+=4)
 640             for (unsigned i = 0; i != 4; ++i)
 641               Idxs.push_back(Builder.getInt32(((Imm >> (2 * i)) & 0x3) + l));
 642         else
 643           llvm_unreachable("Unexpected function");
 644
 645         Rep = Builder.CreateShuffleVector(Op0, Op0, ConstantVector::get(Idxs));
 646       } else {
 647         llvm_unreachable("Unknown function for CallInst upgrade.");
 648       }
 649     }
 650
 651     CI->replaceAllUsesWith(Rep);
 652     CI->eraseFromParent();
 653     return;
 654   }
 655
 656   std::string Name = CI->getName().str();
 657   if (!Name.empty())
 658     CI->setName(Name + ".old");
 659
 660   switch (NewFn->getIntrinsicID()) {
 661   default:
 662     llvm_unreachable("Unknown function for CallInst upgrade.");
 663
 664   // Upgrade debug intrinsics to use an additional DIExpression argument.
 665   case Intrinsic::dbg_declare: {
 666     auto NewCI =
 667         Builder.CreateCall3(NewFn, CI->getArgOperand(0), CI->getArgOperand(1),
 668                             getExpression(CI->getArgOperand(1), F), Name);
 669     NewCI->setDebugLoc(CI->getDebugLoc());
 670     CI->replaceAllUsesWith(NewCI);
 671     CI->eraseFromParent();
 672     return;
 673   }
 674   case Intrinsic::dbg_value: {
 675     auto NewCI = Builder.CreateCall4(
 676         NewFn, CI->getArgOperand(0), CI->getArgOperand(1), CI->getArgOperand(2),
 677         getExpression(CI->getArgOperand(2), F), Name);
 678     NewCI->setDebugLoc(CI->getDebugLoc());
 679     CI->replaceAllUsesWith(NewCI);
 680     CI->eraseFromParent();
 681     return;
 682   }
 683   case Intrinsic::ctlz:
 684   case Intrinsic::cttz:
 685     assert(CI->getNumArgOperands() == 1 &&
 686            "Mismatch between function args and call args");
 687     CI->replaceAllUsesWith(Builder.CreateCall2(NewFn, CI->getArgOperand(0),
 688                                                Builder.getFalse(), Name));
 689     CI->eraseFromParent();
 690     return;
 691
 692   case Intrinsic::objectsize:
 693     CI->replaceAllUsesWith(Builder.CreateCall2(NewFn,
 694                                                CI->getArgOperand(0),
 695                                                CI->getArgOperand(1),
 696                                                Name));
 697     CI->eraseFromParent();
 698     return;
 699
 700   case Intrinsic::ctpop: {
 701     CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(0)));
 702     CI->eraseFromParent();
 703     return;
 704   }
 705
 706   case Intrinsic::x86_xop_vfrcz_ss:
 707   case Intrinsic::x86_xop_vfrcz_sd:
 708     CI->replaceAllUsesWith(Builder.CreateCall(NewFn, CI->getArgOperand(1),
 709                                               Name));
 710     CI->eraseFromParent();
 711     return;
 712
 713   case Intrinsic::x86_sse41_ptestc:
 714   case Intrinsic::x86_sse41_ptestz:
 715   case Intrinsic::x86_sse41_ptestnzc: {
 716     // The arguments for these intrinsics used to be v4f32, and changed
 717     // to v2i64. This is purely a nop, since those are bitwise intrinsics.
 718     // So, the only thing required is a bitcast for both arguments.
 719     // First, check the arguments have the old type.
 720     Value *Arg0 = CI->getArgOperand(0);
 721     if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
 722       return;
 723
 724     // Old intrinsic, add bitcasts
 725     Value *Arg1 = CI->getArgOperand(1);
 726
 727     Value *BC0 =
 728       Builder.CreateBitCast(Arg0,
 729                             VectorType::get(Type::getInt64Ty(C), 2),
 730                             "cast");
 731     Value *BC1 =
 732       Builder.CreateBitCast(Arg1,
 733                             VectorType::get(Type::getInt64Ty(C), 2),
 734                             "cast");
 735
 736     CallInst* NewCall = Builder.CreateCall2(NewFn, BC0, BC1, Name);
 737     CI->replaceAllUsesWith(NewCall);
 738     CI->eraseFromParent();
 739     return;
 740   }
 741
 742   case Intrinsic::x86_sse41_pblendw:
 743   case Intrinsic::x86_sse41_blendpd:
 744   case Intrinsic::x86_sse41_blendps:
 745   case Intrinsic::x86_sse41_insertps:
 746   case Intrinsic::x86_sse41_dppd:
 747   case Intrinsic::x86_sse41_dpps:
 748   case Intrinsic::x86_sse41_mpsadbw:
 749   case Intrinsic::x86_avx_blend_pd_256:
 750   case Intrinsic::x86_avx_blend_ps_256:
 751   case Intrinsic::x86_avx_dp_ps_256:
 752   case Intrinsic::x86_avx2_pblendw:
 753   case Intrinsic::x86_avx2_pblendd_128:
 754   case Intrinsic::x86_avx2_pblendd_256:
 755   case Intrinsic::x86_avx2_mpsadbw: {
 756     // Need to truncate the last argument from i32 to i8 -- this argument models
 757     // an inherently 8-bit immediate operand to these x86 instructions.
 758     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
 759                                  CI->arg_operands().end());
 760
 761     // Replace the last argument with a trunc.
 762     Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
 763
 764     CallInst *NewCall = Builder.CreateCall(NewFn, Args);
 765     CI->replaceAllUsesWith(NewCall);
 766     CI->eraseFromParent();
 767     return;
 768   }
 769   case Intrinsic::x86_avx512_mask_cmp_ps_512:
 770   case Intrinsic::x86_avx512_mask_cmp_pd_512: {
 771     // Need to truncate the last argument from i32 to i8 -- this argument models
 772     // an inherently 8-bit immediate operand to these x86 instructions.
 773     SmallVector<Value *, 5> Args(CI->arg_operands().begin(),
 774                                  CI->arg_operands().end());
 775
 776     // Replace the last argument with a trunc.
 777     Args[2] = Builder.CreateTrunc(Args[2], Type::getInt8Ty(C), "trunc");
 778
 779     CallInst *NewCall = Builder.CreateCall(NewFn, Args);
 780     CI->replaceAllUsesWith(NewCall);
 781     CI->eraseFromParent();
 782     return;
 783   }
 784   }
 785 }
 786
 787 // This tests each Function to determine if it needs upgrading. When we find
 788 // one we are interested in, we then upgrade all calls to reflect the new
 789 // function.
 790 void llvm::UpgradeCallsToIntrinsic(Function* F) {
 791   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
 792
 793   // Upgrade the function and check if it is a totaly new function.
 794   Function *NewFn;
 795   if (UpgradeIntrinsicFunction(F, NewFn)) {
 796     if (NewFn != F) {
 797       // Replace all uses to the old function with the new one if necessary.
 798       for (Value::user_iterator UI = F->user_begin(), UE = F->user_end();
 799            UI != UE; ) {
 800         if (CallInst *CI = dyn_cast<CallInst>(*UI++))
 801           UpgradeIntrinsicCall(CI, NewFn);
 802       }
 803       // Remove old function, no longer used, from the module.
 804       F->eraseFromParent();
 805     }
 806   }
 807 }
 808
 809 void llvm::UpgradeInstWithTBAATag(Instruction *I) {
 810   MDNode *MD = I->getMetadata(LLVMContext::MD_tbaa);
 811   assert(MD && "UpgradeInstWithTBAATag should have a TBAA tag");
 812   // Check if the tag uses struct-path aware TBAA format.
 813   if (isa<MDNode>(MD->getOperand(0)) && MD->getNumOperands() >= 3)
 814     return;
 815
 816   if (MD->getNumOperands() == 3) {
 817     Metadata *Elts[] = {MD->getOperand(0), MD->getOperand(1)};
 818     MDNode *ScalarType = MDNode::get(I->getContext(), Elts);
 819     // Create a MDNode <ScalarType, ScalarType, offset 0, const>
 820     Metadata *Elts2[] = {ScalarType, ScalarType,
 821                          ConstantAsMetadata::get(Constant::getNullValue(
 822                              Type::getInt64Ty(I->getContext()))),
 823                          MD->getOperand(2)};
 824     I->setMetadata(LLVMContext::MD_tbaa, MDNode::get(I->getContext(), Elts2));
 825   } else {
 826     // Create a MDNode <MD, MD, offset 0>
 827     Metadata *Elts[] = {MD, MD, ConstantAsMetadata::get(Constant::getNullValue(
 828                                     Type::getInt64Ty(I->getContext())))};
 829     I->setMetadata(LLVMContext::MD_tbaa, MDNode::get(I->getContext(), Elts));
 830   }
 831 }
 832
 833 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
 834                                       Instruction *&Temp) {
 835   if (Opc != Instruction::BitCast)
 836     return nullptr;
 837
 838   Temp = nullptr;
 839   Type *SrcTy = V->getType();
 840   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
 841       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
 842     LLVMContext &Context = V->getContext();
 843
 844     // We have no information about target data layout, so we assume that
 845     // the maximum pointer size is 64bit.
 846     Type *MidTy = Type::getInt64Ty(Context);
 847     Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
 848
 849     return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
 850   }
 851
 852   return nullptr;
 853 }
 854
 855 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
 856   if (Opc != Instruction::BitCast)
 857     return nullptr;
 858
 859   Type *SrcTy = C->getType();
 860   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
 861       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
 862     LLVMContext &Context = C->getContext();
 863
 864     // We have no information about target data layout, so we assume that
 865     // the maximum pointer size is 64bit.
 866     Type *MidTy = Type::getInt64Ty(Context);
 867
 868     return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
 869                                      DestTy);
 870   }
 871
 872   return nullptr;
 873 }
 874
 875 /// Check the debug info version number, if it is out-dated, drop the debug
 876 /// info. Return true if module is modified.
 877 bool llvm::UpgradeDebugInfo(Module &M) {
 878   unsigned Version = getDebugMetadataVersionFromModule(M);
 879   if (Version == DEBUG_METADATA_VERSION)
 880     return false;
 881
 882   bool RetCode = StripDebugInfo(M);
 883   if (RetCode) {
 884     DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
 885     M.getContext().diagnose(DiagVersion);
 886   }
 887   return RetCode;
 888 }
 889
 890 void llvm::UpgradeMDStringConstant(std::string &String) {
 891   const std::string OldPrefix = "llvm.vectorizer.";
 892   if (String == "llvm.vectorizer.unroll") {
 893     String = "llvm.loop.interleave.count";
 894   } else if (String.find(OldPrefix) == 0) {
 895     String.replace(0, OldPrefix.size(), "llvm.loop.vectorize.");
 896   }
 897 }