lib/VMCore/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/AutoUpgrade.h"
  15 #include "llvm/Constants.h"
  16 #include "llvm/Function.h"
  17 #include "llvm/LLVMContext.h"
  18 #include "llvm/Module.h"
  19 #include "llvm/IntrinsicInst.h"
  20 #include "llvm/ADT/SmallVector.h"
  21 #include "llvm/Support/CallSite.h"
  22 #include "llvm/Support/ErrorHandling.h"
  23 #include "llvm/Support/IRBuilder.h"
  24 #include <cstring>
  25 using namespace llvm;
  26
  27
  28 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
  29   assert(F && "Illegal to upgrade a non-existent Function.");
  30
  31   // Get the Function's name.
  32   const std::string& Name = F->getName();
  33
  34   // Convenience
  35   const FunctionType *FTy = F->getFunctionType();
  36
  37   // Quickly eliminate it, if it's not a candidate.
  38   if (Name.length() <= 8 || Name[0] != 'l' || Name[1] != 'l' ||
  39       Name[2] != 'v' || Name[3] != 'm' || Name[4] != '.')
  40     return false;
  41
  42   Module *M = F->getParent();
  43   switch (Name[5]) {
  44   default: break;
  45   case 'a':
  46     // This upgrades the llvm.atomic.lcs, llvm.atomic.las, llvm.atomic.lss,
  47     // and atomics with default address spaces to their new names to their new
  48     // function name (e.g. llvm.atomic.add.i32 => llvm.atomic.add.i32.p0i32)
  49     if (Name.compare(5,7,"atomic.",7) == 0) {
  50       if (Name.compare(12,3,"lcs",3) == 0) {
  51         std::string::size_type delim = Name.find('.',12);
  52         F->setName("llvm.atomic.cmp.swap" + Name.substr(delim) +
  53                    ".p0" + Name.substr(delim+1));
  54         NewFn = F;
  55         return true;
  56       }
  57       else if (Name.compare(12,3,"las",3) == 0) {
  58         std::string::size_type delim = Name.find('.',12);
  59         F->setName("llvm.atomic.load.add"+Name.substr(delim)
  60                    + ".p0" + Name.substr(delim+1));
  61         NewFn = F;
  62         return true;
  63       }
  64       else if (Name.compare(12,3,"lss",3) == 0) {
  65         std::string::size_type delim = Name.find('.',12);
  66         F->setName("llvm.atomic.load.sub"+Name.substr(delim)
  67                    + ".p0" + Name.substr(delim+1));
  68         NewFn = F;
  69         return true;
  70       }
  71       else if (Name.rfind(".p") == std::string::npos) {
  72         // We don't have an address space qualifier so this has be upgraded
  73         // to the new name.  Copy the type name at the end of the intrinsic
  74         // and add to it
  75         std::string::size_type delim = Name.find_last_of('.');
  76         assert(delim != std::string::npos && "can not find type");
  77         F->setName(Name + ".p0" + Name.substr(delim+1));
  78         NewFn = F;
  79         return true;
  80       }
  81     } else if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
  82       if (((Name.compare(14, 5, "vmovl", 5) == 0 ||
  83             Name.compare(14, 5, "vaddl", 5) == 0 ||
  84             Name.compare(14, 5, "vsubl", 5) == 0 ||
  85             Name.compare(14, 5, "vaddw", 5) == 0 ||
  86             Name.compare(14, 5, "vsubw", 5) == 0 ||
  87             Name.compare(14, 5, "vmull", 5) == 0 ||
  88             Name.compare(14, 5, "vmlal", 5) == 0 ||
  89             Name.compare(14, 5, "vmlsl", 5) == 0 ||
  90             Name.compare(14, 5, "vabdl", 5) == 0 ||
  91             Name.compare(14, 5, "vabal", 5) == 0) &&
  92            (Name.compare(19, 2, "s.", 2) == 0 ||
  93             Name.compare(19, 2, "u.", 2) == 0)) ||
  94
  95           (Name.compare(14, 4, "vaba", 4) == 0 &&
  96            (Name.compare(18, 2, "s.", 2) == 0 ||
  97             Name.compare(18, 2, "u.", 2) == 0)) ||
  98
  99           (Name.compare(14, 6, "vmovn.", 6) == 0)) {
 100
 101         // Calls to these are transformed into IR without intrinsics.
 102         NewFn = 0;
 103         return true;
 104       }
 105       // Old versions of NEON ld/st intrinsics are missing alignment arguments.
 106       bool isVLd = (Name.compare(14, 3, "vld", 3) == 0);
 107       bool isVSt = (Name.compare(14, 3, "vst", 3) == 0);
 108       if (isVLd || isVSt) {
 109         unsigned NumVecs = Name.at(17) - '0';
 110         if (NumVecs == 0 || NumVecs > 4)
 111           return false;
 112         bool isLaneOp = (Name.compare(18, 5, "lane.", 5) == 0);
 113         if (!isLaneOp && Name.at(18) != '.')
 114           return false;
 115         unsigned ExpectedArgs = 2; // for the address and alignment
 116         if (isVSt || isLaneOp)
 117           ExpectedArgs += NumVecs;
 118         if (isLaneOp)
 119           ExpectedArgs += 1; // for the lane number
 120         unsigned NumP = FTy->getNumParams();
 121         if (NumP != ExpectedArgs - 1)
 122           return false;
 123
 124         // Change the name of the old (bad) intrinsic, because
 125         // its type is incorrect, but we cannot overload that name.
 126         F->setName("");
 127
 128         // One argument is missing: add the alignment argument.
 129         std::vector<const Type*> NewParams;
 130         for (unsigned p = 0; p < NumP; ++p)
 131           NewParams.push_back(FTy->getParamType(p));
 132         NewParams.push_back(Type::getInt32Ty(F->getContext()));
 133         FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(),
 134                                                  NewParams, false);
 135         NewFn = cast<Function>(M->getOrInsertFunction(Name, NewFTy));
 136         return true;
 137       }
 138     }
 139     break;
 140   case 'b':
 141     //  This upgrades the name of the llvm.bswap intrinsic function to only use
 142     //  a single type name for overloading. We only care about the old format
 143     //  'llvm.bswap.i*.i*', so check for 'bswap.' and then for there being
 144     //  a '.' after 'bswap.'
 145     if (Name.compare(5,6,"bswap.",6) == 0) {
 146       std::string::size_type delim = Name.find('.',11);
 147
 148       if (delim != std::string::npos) {
 149         //  Construct the new name as 'llvm.bswap' + '.i*'
 150         F->setName(Name.substr(0,10)+Name.substr(delim));
 151         NewFn = F;
 152         return true;
 153       }
 154     }
 155     break;
 156
 157   case 'c':
 158     //  We only want to fix the 'llvm.ct*' intrinsics which do not have the
 159     //  correct return type, so we check for the name, and then check if the
 160     //  return type does not match the parameter type.
 161     if ( (Name.compare(5,5,"ctpop",5) == 0 ||
 162           Name.compare(5,4,"ctlz",4) == 0 ||
 163           Name.compare(5,4,"cttz",4) == 0) &&
 164         FTy->getReturnType() != FTy->getParamType(0)) {
 165       //  We first need to change the name of the old (bad) intrinsic, because
 166       //  its type is incorrect, but we cannot overload that name. We
 167       //  arbitrarily unique it here allowing us to construct a correctly named
 168       //  and typed function below.
 169       F->setName("");
 170
 171       //  Now construct the new intrinsic with the correct name and type. We
 172       //  leave the old function around in order to query its type, whatever it
 173       //  may be, and correctly convert up to the new type.
 174       NewFn = cast<Function>(M->getOrInsertFunction(Name,
 175                                                     FTy->getParamType(0),
 176                                                     FTy->getParamType(0),
 177                                                     (Type *)0));
 178       return true;
 179     }
 180     break;
 181
 182   case 'e':
 183     //  The old llvm.eh.selector.i32 is equivalent to the new llvm.eh.selector.
 184     if (Name.compare("llvm.eh.selector.i32") == 0) {
 185       F->setName("llvm.eh.selector");
 186       NewFn = F;
 187       return true;
 188     }
 189     //  The old llvm.eh.typeid.for.i32 is equivalent to llvm.eh.typeid.for.
 190     if (Name.compare("llvm.eh.typeid.for.i32") == 0) {
 191       F->setName("llvm.eh.typeid.for");
 192       NewFn = F;
 193       return true;
 194     }
 195     //  Convert the old llvm.eh.selector.i64 to a call to llvm.eh.selector.
 196     if (Name.compare("llvm.eh.selector.i64") == 0) {
 197       NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_selector);
 198       return true;
 199     }
 200     //  Convert the old llvm.eh.typeid.for.i64 to a call to llvm.eh.typeid.for.
 201     if (Name.compare("llvm.eh.typeid.for.i64") == 0) {
 202       NewFn = Intrinsic::getDeclaration(M, Intrinsic::eh_typeid_for);
 203       return true;
 204     }
 205     break;
 206
 207   case 'm': {
 208     // This upgrades the llvm.memcpy, llvm.memmove, and llvm.memset to the
 209     // new format that allows overloading the pointer for different address
 210     // space (e.g., llvm.memcpy.i16 => llvm.memcpy.p0i8.p0i8.i16)
 211     const char* NewFnName = NULL;
 212     if (Name.compare(5,8,"memcpy.i",8) == 0) {
 213       if (Name[13] == '8')
 214         NewFnName = "llvm.memcpy.p0i8.p0i8.i8";
 215       else if (Name.compare(13,2,"16") == 0)
 216         NewFnName = "llvm.memcpy.p0i8.p0i8.i16";
 217       else if (Name.compare(13,2,"32") == 0)
 218         NewFnName = "llvm.memcpy.p0i8.p0i8.i32";
 219       else if (Name.compare(13,2,"64") == 0)
 220         NewFnName = "llvm.memcpy.p0i8.p0i8.i64";
 221     } else if (Name.compare(5,9,"memmove.i",9) == 0) {
 222       if (Name[14] == '8')
 223         NewFnName = "llvm.memmove.p0i8.p0i8.i8";
 224       else if (Name.compare(14,2,"16") == 0)
 225         NewFnName = "llvm.memmove.p0i8.p0i8.i16";
 226       else if (Name.compare(14,2,"32") == 0)
 227         NewFnName = "llvm.memmove.p0i8.p0i8.i32";
 228       else if (Name.compare(14,2,"64") == 0)
 229         NewFnName = "llvm.memmove.p0i8.p0i8.i64";
 230     }
 231     else if (Name.compare(5,8,"memset.i",8) == 0) {
 232       if (Name[13] == '8')
 233         NewFnName = "llvm.memset.p0i8.i8";
 234       else if (Name.compare(13,2,"16") == 0)
 235         NewFnName = "llvm.memset.p0i8.i16";
 236       else if (Name.compare(13,2,"32") == 0)
 237         NewFnName = "llvm.memset.p0i8.i32";
 238       else if (Name.compare(13,2,"64") == 0)
 239         NewFnName = "llvm.memset.p0i8.i64";
 240     }
 241     if (NewFnName) {
 242       NewFn = cast<Function>(M->getOrInsertFunction(NewFnName,
 243                                             FTy->getReturnType(),
 244                                             FTy->getParamType(0),
 245                                             FTy->getParamType(1),
 246                                             FTy->getParamType(2),
 247                                             FTy->getParamType(3),
 248                                             Type::getInt1Ty(F->getContext()),
 249                                             (Type *)0));
 250       return true;
 251     }
 252     break;
 253   }
 254   case 'p':
 255     //  This upgrades the llvm.part.select overloaded intrinsic names to only
 256     //  use one type specifier in the name. We only care about the old format
 257     //  'llvm.part.select.i*.i*', and solve as above with bswap.
 258     if (Name.compare(5,12,"part.select.",12) == 0) {
 259       std::string::size_type delim = Name.find('.',17);
 260
 261       if (delim != std::string::npos) {
 262         //  Construct a new name as 'llvm.part.select' + '.i*'
 263         F->setName(Name.substr(0,16)+Name.substr(delim));
 264         NewFn = F;
 265         return true;
 266       }
 267       break;
 268     }
 269
 270     //  This upgrades the llvm.part.set intrinsics similarly as above, however
 271     //  we care about 'llvm.part.set.i*.i*.i*', but only the first two types
 272     //  must match. There is an additional type specifier after these two
 273     //  matching types that we must retain when upgrading.  Thus, we require
 274     //  finding 2 periods, not just one, after the intrinsic name.
 275     if (Name.compare(5,9,"part.set.",9) == 0) {
 276       std::string::size_type delim = Name.find('.',14);
 277
 278       if (delim != std::string::npos &&
 279           Name.find('.',delim+1) != std::string::npos) {
 280         //  Construct a new name as 'llvm.part.select' + '.i*.i*'
 281         F->setName(Name.substr(0,13)+Name.substr(delim));
 282         NewFn = F;
 283         return true;
 284       }
 285       break;
 286     }
 287
 288     break;
 289   case 'x':
 290     // This fixes all MMX shift intrinsic instructions to take a
 291     // x86_mmx instead of a v1i64, v2i32, v4i16, or v8i8.
 292     if (Name.compare(5, 8, "x86.mmx.", 8) == 0) {
 293       const Type *X86_MMXTy = VectorType::getX86_MMXTy(FTy->getContext());
 294
 295       if (Name.compare(13, 4, "padd", 4) == 0   ||
 296           Name.compare(13, 4, "psub", 4) == 0   ||
 297           Name.compare(13, 4, "pmul", 4) == 0   ||
 298           Name.compare(13, 5, "pmadd", 5) == 0  ||
 299           Name.compare(13, 4, "pand", 4) == 0   ||
 300           Name.compare(13, 3, "por", 3) == 0    ||
 301           Name.compare(13, 4, "pxor", 4) == 0   ||
 302           Name.compare(13, 4, "pavg", 4) == 0   ||
 303           Name.compare(13, 4, "pmax", 4) == 0   ||
 304           Name.compare(13, 4, "pmin", 4) == 0   ||
 305           Name.compare(13, 4, "psad", 4) == 0   ||
 306           Name.compare(13, 4, "psll", 4) == 0   ||
 307           Name.compare(13, 4, "psrl", 4) == 0   ||
 308           Name.compare(13, 4, "psra", 4) == 0   ||
 309           Name.compare(13, 4, "pack", 4) == 0   ||
 310           Name.compare(13, 6, "punpck", 6) == 0 ||
 311           Name.compare(13, 4, "pcmp", 4) == 0) {
 312         assert(FTy->getNumParams() == 2 && "MMX intrinsic takes 2 args!");
 313         const Type *SecondParamTy = X86_MMXTy;
 314
 315         if (Name.compare(13, 5, "pslli", 5) == 0 ||
 316             Name.compare(13, 5, "psrli", 5) == 0 ||
 317             Name.compare(13, 5, "psrai", 5) == 0)
 318           SecondParamTy = FTy->getParamType(1);
 319
 320         // Don't do anything if it has the correct types.
 321         if (FTy->getReturnType() == X86_MMXTy &&
 322             FTy->getParamType(0) == X86_MMXTy &&
 323             FTy->getParamType(1) == SecondParamTy)
 324           break;
 325
 326         // We first need to change the name of the old (bad) intrinsic, because
 327         // its type is incorrect, but we cannot overload that name. We
 328         // arbitrarily unique it here allowing us to construct a correctly named
 329         // and typed function below.
 330         F->setName("");
 331
 332         // Now construct the new intrinsic with the correct name and type. We
 333         // leave the old function around in order to query its type, whatever it
 334         // may be, and correctly convert up to the new type.
 335         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 336                                                       X86_MMXTy, X86_MMXTy,
 337                                                       SecondParamTy, (Type*)0));
 338         return true;
 339       }
 340
 341       if (Name.compare(13, 8, "maskmovq", 8) == 0) {
 342         // Don't do anything if it has the correct types.
 343         if (FTy->getParamType(0) == X86_MMXTy &&
 344             FTy->getParamType(1) == X86_MMXTy)
 345           break;
 346
 347         F->setName("");
 348         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 349                                                       FTy->getReturnType(),
 350                                                       X86_MMXTy,
 351                                                       X86_MMXTy,
 352                                                       FTy->getParamType(2),
 353                                                       (Type*)0));
 354         return true;
 355       }
 356
 357       if (Name.compare(13, 8, "pmovmskb", 8) == 0) {
 358         if (FTy->getParamType(0) == X86_MMXTy)
 359           break;
 360
 361         F->setName("");
 362         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 363                                                       FTy->getReturnType(),
 364                                                       X86_MMXTy,
 365                                                       (Type*)0));
 366         return true;
 367       }
 368
 369       if (Name.compare(13, 5, "movnt", 5) == 0) {
 370         if (FTy->getParamType(1) == X86_MMXTy)
 371           break;
 372
 373         F->setName("");
 374         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 375                                                       FTy->getReturnType(),
 376                                                       FTy->getParamType(0),
 377                                                       X86_MMXTy,
 378                                                       (Type*)0));
 379         return true;
 380       }
 381
 382       if (Name.compare(13, 7, "palignr", 7) == 0) {
 383         if (FTy->getReturnType() == X86_MMXTy &&
 384             FTy->getParamType(0) == X86_MMXTy &&
 385             FTy->getParamType(1) == X86_MMXTy)
 386           break;
 387
 388         F->setName("");
 389         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 390                                                       X86_MMXTy,
 391                                                       X86_MMXTy,
 392                                                       X86_MMXTy,
 393                                                       FTy->getParamType(2),
 394                                                       (Type*)0));
 395         return true;
 396       }
 397
 398       if (Name.compare(13, 5, "pextr", 5) == 0) {
 399         if (FTy->getParamType(0) == X86_MMXTy)
 400           break;
 401
 402         F->setName("");
 403         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 404                                                       FTy->getReturnType(),
 405                                                       X86_MMXTy,
 406                                                       FTy->getParamType(1),
 407                                                       (Type*)0));
 408         return true;
 409       }
 410
 411       if (Name.compare(13, 5, "pinsr", 5) == 0) {
 412         if (FTy->getReturnType() == X86_MMXTy &&
 413             FTy->getParamType(0) == X86_MMXTy)
 414           break;
 415
 416         F->setName("");
 417         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 418                                                       X86_MMXTy,
 419                                                       X86_MMXTy,
 420                                                       FTy->getParamType(1),
 421                                                       FTy->getParamType(2),
 422                                                       (Type*)0));
 423         return true;
 424       }
 425
 426       if (Name.compare(13, 12, "cvtsi32.si64", 12) == 0) {
 427         if (FTy->getReturnType() == X86_MMXTy)
 428           break;
 429
 430         F->setName("");
 431         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 432                                                       X86_MMXTy,
 433                                                       FTy->getParamType(0),
 434                                                       (Type*)0));
 435         return true;
 436       }
 437
 438       if (Name.compare(13, 12, "cvtsi64.si32", 12) == 0) {
 439         if (FTy->getParamType(0) == X86_MMXTy)
 440           break;
 441
 442         F->setName("");
 443         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 444                                                       FTy->getReturnType(),
 445                                                       X86_MMXTy,
 446                                                       (Type*)0));
 447         return true;
 448       }
 449
 450       if (Name.compare(13, 8, "vec.init", 8) == 0) {
 451         if (FTy->getReturnType() == X86_MMXTy)
 452           break;
 453
 454         F->setName("");
 455
 456         if (Name.compare(21, 2, ".b", 2) == 0)
 457           NewFn = cast<Function>(M->getOrInsertFunction(Name,
 458                                                         X86_MMXTy,
 459                                                         FTy->getParamType(0),
 460                                                         FTy->getParamType(1),
 461                                                         FTy->getParamType(2),
 462                                                         FTy->getParamType(3),
 463                                                         FTy->getParamType(4),
 464                                                         FTy->getParamType(5),
 465                                                         FTy->getParamType(6),
 466                                                         FTy->getParamType(7),
 467                                                         (Type*)0));
 468         else if (Name.compare(21, 2, ".w", 2) == 0)
 469           NewFn = cast<Function>(M->getOrInsertFunction(Name,
 470                                                         X86_MMXTy,
 471                                                         FTy->getParamType(0),
 472                                                         FTy->getParamType(1),
 473                                                         FTy->getParamType(2),
 474                                                         FTy->getParamType(3),
 475                                                         (Type*)0));
 476         else if (Name.compare(21, 2, ".d", 2) == 0)
 477           NewFn = cast<Function>(M->getOrInsertFunction(Name,
 478                                                         X86_MMXTy,
 479                                                         FTy->getParamType(0),
 480                                                         FTy->getParamType(1),
 481                                                         (Type*)0));
 482         return true;
 483       }
 484
 485
 486       if (Name.compare(13, 9, "vec.ext.d", 9) == 0) {
 487         if (FTy->getReturnType() == X86_MMXTy &&
 488             FTy->getParamType(0) == X86_MMXTy)
 489           break;
 490
 491         F->setName("");
 492         NewFn = cast<Function>(M->getOrInsertFunction(Name,
 493                                                       X86_MMXTy,
 494                                                       X86_MMXTy,
 495                                                       FTy->getParamType(1),
 496                                                       (Type*)0));
 497         return true;
 498       }
 499
 500       if (Name.compare(13, 9, "emms", 4) == 0 ||
 501           Name.compare(13, 9, "femms", 5) == 0) {
 502         NewFn = 0;
 503         break;
 504       }
 505
 506       // We really shouldn't get here ever.
 507       assert(0 && "Invalid MMX intrinsic!");
 508       break;
 509     } else if (Name.compare(5,17,"x86.sse2.loadh.pd",17) == 0 ||
 510                Name.compare(5,17,"x86.sse2.loadl.pd",17) == 0 ||
 511                Name.compare(5,16,"x86.sse2.movl.dq",16) == 0 ||
 512                Name.compare(5,15,"x86.sse2.movs.d",15) == 0 ||
 513                Name.compare(5,16,"x86.sse2.shuf.pd",16) == 0 ||
 514                Name.compare(5,18,"x86.sse2.unpckh.pd",18) == 0 ||
 515                Name.compare(5,18,"x86.sse2.unpckl.pd",18) == 0 ||
 516                Name.compare(5,20,"x86.sse2.punpckh.qdq",20) == 0 ||
 517                Name.compare(5,20,"x86.sse2.punpckl.qdq",20) == 0) {
 518       // Calls to these intrinsics are transformed into ShuffleVector's.
 519       NewFn = 0;
 520       return true;
 521     } else if (Name.compare(5, 16, "x86.sse41.pmulld", 16) == 0) {
 522       // Calls to these intrinsics are transformed into vector multiplies.
 523       NewFn = 0;
 524       return true;
 525     } else if (Name.compare(5, 18, "x86.ssse3.palign.r", 18) == 0 ||
 526                Name.compare(5, 22, "x86.ssse3.palign.r.128", 22) == 0) {
 527       // Calls to these intrinsics are transformed into vector shuffles, shifts,
 528       // or 0.
 529       NewFn = 0;
 530       return true;
 531     } else if (Name.compare(5, 17, "x86.ssse3.pshuf.w", 17) == 0) {
 532       // This is an SSE/MMX instruction.
 533       const Type *X86_MMXTy = VectorType::getX86_MMXTy(FTy->getContext());
 534       NewFn =
 535         cast<Function>(M->getOrInsertFunction("llvm.x86.sse.pshuf.w",
 536                                               X86_MMXTy,
 537                                               X86_MMXTy,
 538                                               Type::getInt8Ty(F->getContext()),
 539                                               (Type*)0));
 540       return true;
 541     }
 542
 543     break;
 544   }
 545
 546   //  This may not belong here. This function is effectively being overloaded
 547   //  to both detect an intrinsic which needs upgrading, and to provide the
 548   //  upgraded form of the intrinsic. We should perhaps have two separate
 549   //  functions for this.
 550   return false;
 551 }
 552
 553 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
 554   NewFn = 0;
 555   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
 556
 557   // Upgrade intrinsic attributes.  This does not change the function.
 558   if (NewFn)
 559     F = NewFn;
 560   if (unsigned id = F->getIntrinsicID())
 561     F->setAttributes(Intrinsic::getAttributes((Intrinsic::ID)id));
 562   return Upgraded;
 563 }
 564
 565 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
 566   StringRef Name(GV->getName());
 567
 568   // We are only upgrading one symbol here.
 569   if (Name == ".llvm.eh.catch.all.value") {
 570     GV->setName("llvm.eh.catch.all.value");
 571     return true;
 572   }
 573
 574   return false;
 575 }
 576
 577 /// ExtendNEONArgs - For NEON "long" and "wide" operations, where the results
 578 /// have vector elements twice as big as one or both source operands, do the
 579 /// sign- or zero-extension that used to be handled by intrinsics.  The
 580 /// extended values are returned via V0 and V1.
 581 static void ExtendNEONArgs(CallInst *CI, Value *Arg0, Value *Arg1,
 582                            Value *&V0, Value *&V1) {
 583   Function *F = CI->getCalledFunction();
 584   const std::string& Name = F->getName();
 585   bool isLong = (Name.at(18) == 'l');
 586   bool isSigned = (Name.at(19) == 's');
 587
 588   if (isSigned) {
 589     if (isLong)
 590       V0 = new SExtInst(Arg0, CI->getType(), "", CI);
 591     else
 592       V0 = Arg0;
 593     V1 = new SExtInst(Arg1, CI->getType(), "", CI);
 594   } else {
 595     if (isLong)
 596       V0 = new ZExtInst(Arg0, CI->getType(), "", CI);
 597     else
 598       V0 = Arg0;
 599     V1 = new ZExtInst(Arg1, CI->getType(), "", CI);
 600   }
 601 }
 602
 603 /// CallVABD - As part of expanding a call to one of the old NEON vabdl, vaba,
 604 /// or vabal intrinsics, construct a call to a vabd intrinsic.  Examine the
 605 /// name of the old intrinsic to determine whether to use a signed or unsigned
 606 /// vabd intrinsic.  Get the type from the old call instruction, adjusted for
 607 /// half-size vector elements if the old intrinsic was vabdl or vabal.
 608 static Instruction *CallVABD(CallInst *CI, Value *Arg0, Value *Arg1) {
 609   Function *F = CI->getCalledFunction();
 610   const std::string& Name = F->getName();
 611   bool isLong = (Name.at(18) == 'l');
 612   bool isSigned = (Name.at(isLong ? 19 : 18) == 's');
 613
 614   Intrinsic::ID intID;
 615   if (isSigned)
 616     intID = Intrinsic::arm_neon_vabds;
 617   else
 618     intID = Intrinsic::arm_neon_vabdu;
 619
 620   const Type *Ty = CI->getType();
 621   if (isLong)
 622     Ty = VectorType::getTruncatedElementVectorType(cast<const VectorType>(Ty));
 623
 624   Function *VABD = Intrinsic::getDeclaration(F->getParent(), intID, &Ty, 1);
 625   Value *Operands[2];
 626   Operands[0] = Arg0;
 627   Operands[1] = Arg1;
 628   return CallInst::Create(VABD, Operands, Operands+2,
 629                           "upgraded."+CI->getName(), CI);
 630 }
 631
 632 /// ConstructNewCallInst - Construct a new CallInst with the signature of NewFn.
 633 static void ConstructNewCallInst(Function *NewFn, CallInst *OldCI,
 634                                  Value **Operands, unsigned NumOps,
 635                                  bool AssignName = true) {
 636   // Construct a new CallInst.
 637   CallInst *NewCI =
 638     CallInst::Create(NewFn, Operands, Operands + NumOps,
 639                      AssignName ? "upgraded." + OldCI->getName() : "", OldCI);
 640
 641   NewCI->setTailCall(OldCI->isTailCall());
 642   NewCI->setCallingConv(OldCI->getCallingConv());
 643
 644   // Handle any uses of the old CallInst. If the type has changed, add a cast.
 645   if (!OldCI->use_empty()) {
 646     if (OldCI->getType() != NewCI->getType()) {
 647       Function *OldFn = OldCI->getCalledFunction();
 648       CastInst *RetCast =
 649         CastInst::Create(CastInst::getCastOpcode(NewCI, true,
 650                                                  OldFn->getReturnType(), true),
 651                          NewCI, OldFn->getReturnType(), NewCI->getName(),OldCI);
 652
 653       // Replace all uses of the old call with the new cast which has the
 654       // correct type.
 655       OldCI->replaceAllUsesWith(RetCast);
 656     } else {
 657       OldCI->replaceAllUsesWith(NewCI);
 658     }
 659   }
 660
 661   // Clean up the old call now that it has been completely upgraded.
 662   OldCI->eraseFromParent();
 663 }
 664
 665 // UpgradeIntrinsicCall - Upgrade a call to an old intrinsic to be a call the
 666 // upgraded intrinsic. All argument and return casting must be provided in
 667 // order to seamlessly integrate with existing context.
 668 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
 669   Function *F = CI->getCalledFunction();
 670   LLVMContext &C = CI->getContext();
 671   ImmutableCallSite CS(CI);
 672
 673   assert(F && "CallInst has no function associated with it.");
 674
 675   if (!NewFn) {
 676     // Get the Function's name.
 677     const std::string& Name = F->getName();
 678
 679     // Upgrade ARM NEON intrinsics.
 680     if (Name.compare(5, 9, "arm.neon.", 9) == 0) {
 681       Instruction *NewI;
 682       Value *V0, *V1;
 683       if (Name.compare(14, 7, "vmovls.", 7) == 0) {
 684         NewI = new SExtInst(CI->getArgOperand(0), CI->getType(),
 685                             "upgraded." + CI->getName(), CI);
 686       } else if (Name.compare(14, 7, "vmovlu.", 7) == 0) {
 687         NewI = new ZExtInst(CI->getArgOperand(0), CI->getType(),
 688                             "upgraded." + CI->getName(), CI);
 689       } else if (Name.compare(14, 4, "vadd", 4) == 0) {
 690         ExtendNEONArgs(CI, CI->getArgOperand(0), CI->getArgOperand(1), V0, V1);
 691         NewI = BinaryOperator::CreateAdd(V0, V1, "upgraded."+CI->getName(), CI);
 692       } else if (Name.compare(14, 4, "vsub", 4) == 0) {
 693         ExtendNEONArgs(CI, CI->getArgOperand(0), CI->getArgOperand(1), V0, V1);
 694         NewI = BinaryOperator::CreateSub(V0, V1,"upgraded."+CI->getName(),CI);
 695       } else if (Name.compare(14, 4, "vmul", 4) == 0) {
 696         ExtendNEONArgs(CI, CI->getArgOperand(0), CI->getArgOperand(1), V0, V1);
 697         NewI = BinaryOperator::CreateMul(V0, V1,"upgraded."+CI->getName(),CI);
 698       } else if (Name.compare(14, 4, "vmla", 4) == 0) {
 699         ExtendNEONArgs(CI, CI->getArgOperand(1), CI->getArgOperand(2), V0, V1);
 700         Instruction *MulI = BinaryOperator::CreateMul(V0, V1, "", CI);
 701         NewI = BinaryOperator::CreateAdd(CI->getArgOperand(0), MulI,
 702                                          "upgraded."+CI->getName(), CI);
 703       } else if (Name.compare(14, 4, "vmls", 4) == 0) {
 704         ExtendNEONArgs(CI, CI->getArgOperand(1), CI->getArgOperand(2), V0, V1);
 705         Instruction *MulI = BinaryOperator::CreateMul(V0, V1, "", CI);
 706         NewI = BinaryOperator::CreateSub(CI->getArgOperand(0), MulI,
 707                                          "upgraded."+CI->getName(), CI);
 708       } else if (Name.compare(14, 4, "vabd", 4) == 0) {
 709         NewI = CallVABD(CI, CI->getArgOperand(0), CI->getArgOperand(1));
 710         NewI = new ZExtInst(NewI, CI->getType(), "upgraded."+CI->getName(), CI);
 711       } else if (Name.compare(14, 4, "vaba", 4) == 0) {
 712         NewI = CallVABD(CI, CI->getArgOperand(1), CI->getArgOperand(2));
 713         if (Name.at(18) == 'l')
 714           NewI = new ZExtInst(NewI, CI->getType(), "", CI);
 715         NewI = BinaryOperator::CreateAdd(CI->getArgOperand(0), NewI,
 716                                          "upgraded."+CI->getName(), CI);
 717       } else if (Name.compare(14, 6, "vmovn.", 6) == 0) {
 718         NewI = new TruncInst(CI->getArgOperand(0), CI->getType(),
 719                              "upgraded." + CI->getName(), CI);
 720       } else {
 721         llvm_unreachable("Unknown arm.neon function for CallInst upgrade.");
 722       }
 723       // Replace any uses of the old CallInst.
 724       if (!CI->use_empty())
 725         CI->replaceAllUsesWith(NewI);
 726       CI->eraseFromParent();
 727       return;
 728     }
 729
 730     bool isLoadH = false, isLoadL = false, isMovL = false;
 731     bool isMovSD = false, isShufPD = false;
 732     bool isUnpckhPD = false, isUnpcklPD = false;
 733     bool isPunpckhQPD = false, isPunpcklQPD = false;
 734     if (F->getName() == "llvm.x86.sse2.loadh.pd")
 735       isLoadH = true;
 736     else if (F->getName() == "llvm.x86.sse2.loadl.pd")
 737       isLoadL = true;
 738     else if (F->getName() == "llvm.x86.sse2.movl.dq")
 739       isMovL = true;
 740     else if (F->getName() == "llvm.x86.sse2.movs.d")
 741       isMovSD = true;
 742     else if (F->getName() == "llvm.x86.sse2.shuf.pd")
 743       isShufPD = true;
 744     else if (F->getName() == "llvm.x86.sse2.unpckh.pd")
 745       isUnpckhPD = true;
 746     else if (F->getName() == "llvm.x86.sse2.unpckl.pd")
 747       isUnpcklPD = true;
 748     else if (F->getName() ==  "llvm.x86.sse2.punpckh.qdq")
 749       isPunpckhQPD = true;
 750     else if (F->getName() ==  "llvm.x86.sse2.punpckl.qdq")
 751       isPunpcklQPD = true;
 752
 753     if (isLoadH || isLoadL || isMovL || isMovSD || isShufPD ||
 754         isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
 755       std::vector<Constant*> Idxs;
 756       Value *Op0 = CI->getArgOperand(0);
 757       ShuffleVectorInst *SI = NULL;
 758       if (isLoadH || isLoadL) {
 759         Value *Op1 = UndefValue::get(Op0->getType());
 760         Value *Addr = new BitCastInst(CI->getArgOperand(1),
 761                                   Type::getDoublePtrTy(C),
 762                                       "upgraded.", CI);
 763         Value *Load = new LoadInst(Addr, "upgraded.", false, 8, CI);
 764         Value *Idx = ConstantInt::get(Type::getInt32Ty(C), 0);
 765         Op1 = InsertElementInst::Create(Op1, Load, Idx, "upgraded.", CI);
 766
 767         if (isLoadH) {
 768           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
 769           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
 770         } else {
 771           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
 772           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
 773         }
 774         Value *Mask = ConstantVector::get(Idxs);
 775         SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
 776       } else if (isMovL) {
 777         Constant *Zero = ConstantInt::get(Type::getInt32Ty(C), 0);
 778         Idxs.push_back(Zero);
 779         Idxs.push_back(Zero);
 780         Idxs.push_back(Zero);
 781         Idxs.push_back(Zero);
 782         Value *ZeroV = ConstantVector::get(Idxs);
 783
 784         Idxs.clear();
 785         Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 4));
 786         Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 5));
 787         Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
 788         Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
 789         Value *Mask = ConstantVector::get(Idxs);
 790         SI = new ShuffleVectorInst(ZeroV, Op0, Mask, "upgraded.", CI);
 791       } else if (isMovSD ||
 792                  isUnpckhPD || isUnpcklPD || isPunpckhQPD || isPunpcklQPD) {
 793         Value *Op1 = CI->getArgOperand(1);
 794         if (isMovSD) {
 795           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
 796           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
 797         } else if (isUnpckhPD || isPunpckhQPD) {
 798           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
 799           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 3));
 800         } else {
 801           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 0));
 802           Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), 2));
 803         }
 804         Value *Mask = ConstantVector::get(Idxs);
 805         SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
 806       } else if (isShufPD) {
 807         Value *Op1 = CI->getArgOperand(1);
 808         unsigned MaskVal =
 809                         cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
 810         Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C), MaskVal & 1));
 811         Idxs.push_back(ConstantInt::get(Type::getInt32Ty(C),
 812                                                ((MaskVal >> 1) & 1)+2));
 813         Value *Mask = ConstantVector::get(Idxs);
 814         SI = new ShuffleVectorInst(Op0, Op1, Mask, "upgraded.", CI);
 815       }
 816
 817       assert(SI && "Unexpected!");
 818
 819       // Handle any uses of the old CallInst.
 820       if (!CI->use_empty())
 821         //  Replace all uses of the old call with the new cast which has the
 822         //  correct type.
 823         CI->replaceAllUsesWith(SI);
 824
 825       //  Clean up the old call now that it has been completely upgraded.
 826       CI->eraseFromParent();
 827     } else if (F->getName() == "llvm.x86.sse41.pmulld") {
 828       // Upgrade this set of intrinsics into vector multiplies.
 829       Instruction *Mul = BinaryOperator::CreateMul(CI->getArgOperand(0),
 830                                                    CI->getArgOperand(1),
 831                                                    CI->getName(),
 832                                                    CI);
 833       // Fix up all the uses with our new multiply.
 834       if (!CI->use_empty())
 835         CI->replaceAllUsesWith(Mul);
 836
 837       // Remove upgraded multiply.
 838       CI->eraseFromParent();
 839     } else if (F->getName() == "llvm.x86.ssse3.palign.r") {
 840       Value *Op1 = CI->getArgOperand(0);
 841       Value *Op2 = CI->getArgOperand(1);
 842       Value *Op3 = CI->getArgOperand(2);
 843       unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
 844       Value *Rep;
 845       IRBuilder<> Builder(C);
 846       Builder.SetInsertPoint(CI->getParent(), CI);
 847
 848       // If palignr is shifting the pair of input vectors less than 9 bytes,
 849       // emit a shuffle instruction.
 850       if (shiftVal <= 8) {
 851         const Type *IntTy = Type::getInt32Ty(C);
 852         const Type *EltTy = Type::getInt8Ty(C);
 853         const Type *VecTy = VectorType::get(EltTy, 8);
 854
 855         Op2 = Builder.CreateBitCast(Op2, VecTy);
 856         Op1 = Builder.CreateBitCast(Op1, VecTy);
 857
 858         llvm::SmallVector<llvm::Constant*, 8> Indices;
 859         for (unsigned i = 0; i != 8; ++i)
 860           Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
 861
 862         Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
 863         Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
 864         Rep = Builder.CreateBitCast(Rep, F->getReturnType());
 865       }
 866
 867       // If palignr is shifting the pair of input vectors more than 8 but less
 868       // than 16 bytes, emit a logical right shift of the destination.
 869       else if (shiftVal < 16) {
 870         // MMX has these as 1 x i64 vectors for some odd optimization reasons.
 871         const Type *EltTy = Type::getInt64Ty(C);
 872         const Type *VecTy = VectorType::get(EltTy, 1);
 873
 874         Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
 875         Op2 = ConstantInt::get(VecTy, (shiftVal-8) * 8);
 876
 877         // create i32 constant
 878         Function *I =
 879           Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_mmx_psrl_q);
 880         Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
 881       }
 882
 883       // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
 884       else {
 885         Rep = Constant::getNullValue(F->getReturnType());
 886       }
 887
 888       // Replace any uses with our new instruction.
 889       if (!CI->use_empty())
 890         CI->replaceAllUsesWith(Rep);
 891
 892       // Remove upgraded instruction.
 893       CI->eraseFromParent();
 894
 895     } else if (F->getName() == "llvm.x86.ssse3.palign.r.128") {
 896       Value *Op1 = CI->getArgOperand(0);
 897       Value *Op2 = CI->getArgOperand(1);
 898       Value *Op3 = CI->getArgOperand(2);
 899       unsigned shiftVal = cast<ConstantInt>(Op3)->getZExtValue();
 900       Value *Rep;
 901       IRBuilder<> Builder(C);
 902       Builder.SetInsertPoint(CI->getParent(), CI);
 903
 904       // If palignr is shifting the pair of input vectors less than 17 bytes,
 905       // emit a shuffle instruction.
 906       if (shiftVal <= 16) {
 907         const Type *IntTy = Type::getInt32Ty(C);
 908         const Type *EltTy = Type::getInt8Ty(C);
 909         const Type *VecTy = VectorType::get(EltTy, 16);
 910
 911         Op2 = Builder.CreateBitCast(Op2, VecTy);
 912         Op1 = Builder.CreateBitCast(Op1, VecTy);
 913
 914         llvm::SmallVector<llvm::Constant*, 16> Indices;
 915         for (unsigned i = 0; i != 16; ++i)
 916           Indices.push_back(ConstantInt::get(IntTy, shiftVal + i));
 917
 918         Value *SV = ConstantVector::get(Indices.begin(), Indices.size());
 919         Rep = Builder.CreateShuffleVector(Op2, Op1, SV, "palignr");
 920         Rep = Builder.CreateBitCast(Rep, F->getReturnType());
 921       }
 922
 923       // If palignr is shifting the pair of input vectors more than 16 but less
 924       // than 32 bytes, emit a logical right shift of the destination.
 925       else if (shiftVal < 32) {
 926         const Type *EltTy = Type::getInt64Ty(C);
 927         const Type *VecTy = VectorType::get(EltTy, 2);
 928         const Type *IntTy = Type::getInt32Ty(C);
 929
 930         Op1 = Builder.CreateBitCast(Op1, VecTy, "cast");
 931         Op2 = ConstantInt::get(IntTy, (shiftVal-16) * 8);
 932
 933         // create i32 constant
 934         Function *I =
 935           Intrinsic::getDeclaration(F->getParent(), Intrinsic::x86_sse2_psrl_dq);
 936         Rep = Builder.CreateCall2(I, Op1, Op2, "palignr");
 937       }
 938
 939       // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
 940       else {
 941         Rep = Constant::getNullValue(F->getReturnType());
 942       }
 943
 944       // Replace any uses with our new instruction.
 945       if (!CI->use_empty())
 946         CI->replaceAllUsesWith(Rep);
 947
 948       // Remove upgraded instruction.
 949       CI->eraseFromParent();
 950
 951     } else {
 952       llvm_unreachable("Unknown function for CallInst upgrade.");
 953     }
 954     return;
 955   }
 956
 957   switch (NewFn->getIntrinsicID()) {
 958   default: llvm_unreachable("Unknown function for CallInst upgrade.");
 959   case Intrinsic::arm_neon_vld1:
 960   case Intrinsic::arm_neon_vld2:
 961   case Intrinsic::arm_neon_vld3:
 962   case Intrinsic::arm_neon_vld4:
 963   case Intrinsic::arm_neon_vst1:
 964   case Intrinsic::arm_neon_vst2:
 965   case Intrinsic::arm_neon_vst3:
 966   case Intrinsic::arm_neon_vst4:
 967   case Intrinsic::arm_neon_vld2lane:
 968   case Intrinsic::arm_neon_vld3lane:
 969   case Intrinsic::arm_neon_vld4lane:
 970   case Intrinsic::arm_neon_vst2lane:
 971   case Intrinsic::arm_neon_vst3lane:
 972   case Intrinsic::arm_neon_vst4lane: {
 973     // Add a default alignment argument of 1.
 974     SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
 975     Operands.push_back(ConstantInt::get(Type::getInt32Ty(C), 1));
 976     CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
 977                                        CI->getName(), CI);
 978     NewCI->setTailCall(CI->isTailCall());
 979     NewCI->setCallingConv(CI->getCallingConv());
 980
 981     //  Handle any uses of the old CallInst.
 982     if (!CI->use_empty())
 983       //  Replace all uses of the old call with the new cast which has the
 984       //  correct type.
 985       CI->replaceAllUsesWith(NewCI);
 986
 987     //  Clean up the old call now that it has been completely upgraded.
 988     CI->eraseFromParent();
 989     break;
 990   }
 991
 992   case Intrinsic::x86_mmx_padd_b:
 993   case Intrinsic::x86_mmx_padd_w:
 994   case Intrinsic::x86_mmx_padd_d:
 995   case Intrinsic::x86_mmx_padd_q:
 996   case Intrinsic::x86_mmx_padds_b:
 997   case Intrinsic::x86_mmx_padds_w:
 998   case Intrinsic::x86_mmx_paddus_b:
 999   case Intrinsic::x86_mmx_paddus_w:
1000   case Intrinsic::x86_mmx_psub_b:
1001   case Intrinsic::x86_mmx_psub_w:
1002   case Intrinsic::x86_mmx_psub_d:
1003   case Intrinsic::x86_mmx_psub_q:
1004   case Intrinsic::x86_mmx_psubs_b:
1005   case Intrinsic::x86_mmx_psubs_w:
1006   case Intrinsic::x86_mmx_psubus_b:
1007   case Intrinsic::x86_mmx_psubus_w:
1008   case Intrinsic::x86_mmx_pmulh_w:
1009   case Intrinsic::x86_mmx_pmull_w:
1010   case Intrinsic::x86_mmx_pmulhu_w:
1011   case Intrinsic::x86_mmx_pmulu_dq:
1012   case Intrinsic::x86_mmx_pmadd_wd:
1013   case Intrinsic::x86_mmx_pand:
1014   case Intrinsic::x86_mmx_pandn:
1015   case Intrinsic::x86_mmx_por:
1016   case Intrinsic::x86_mmx_pxor:
1017   case Intrinsic::x86_mmx_pavg_b:
1018   case Intrinsic::x86_mmx_pavg_w:
1019   case Intrinsic::x86_mmx_pmaxu_b:
1020   case Intrinsic::x86_mmx_pmaxs_w:
1021   case Intrinsic::x86_mmx_pminu_b:
1022   case Intrinsic::x86_mmx_pmins_w:
1023   case Intrinsic::x86_mmx_psad_bw:
1024   case Intrinsic::x86_mmx_psll_w:
1025   case Intrinsic::x86_mmx_psll_d:
1026   case Intrinsic::x86_mmx_psll_q:
1027   case Intrinsic::x86_mmx_pslli_w:
1028   case Intrinsic::x86_mmx_pslli_d:
1029   case Intrinsic::x86_mmx_pslli_q:
1030   case Intrinsic::x86_mmx_psrl_w:
1031   case Intrinsic::x86_mmx_psrl_d:
1032   case Intrinsic::x86_mmx_psrl_q:
1033   case Intrinsic::x86_mmx_psrli_w:
1034   case Intrinsic::x86_mmx_psrli_d:
1035   case Intrinsic::x86_mmx_psrli_q:
1036   case Intrinsic::x86_mmx_psra_w:
1037   case Intrinsic::x86_mmx_psra_d:
1038   case Intrinsic::x86_mmx_psrai_w:
1039   case Intrinsic::x86_mmx_psrai_d:
1040   case Intrinsic::x86_mmx_packsswb:
1041   case Intrinsic::x86_mmx_packssdw:
1042   case Intrinsic::x86_mmx_packuswb:
1043   case Intrinsic::x86_mmx_punpckhbw:
1044   case Intrinsic::x86_mmx_punpckhwd:
1045   case Intrinsic::x86_mmx_punpckhdq:
1046   case Intrinsic::x86_mmx_punpcklbw:
1047   case Intrinsic::x86_mmx_punpcklwd:
1048   case Intrinsic::x86_mmx_punpckldq:
1049   case Intrinsic::x86_mmx_pcmpeq_b:
1050   case Intrinsic::x86_mmx_pcmpeq_w:
1051   case Intrinsic::x86_mmx_pcmpeq_d:
1052   case Intrinsic::x86_mmx_pcmpgt_b:
1053   case Intrinsic::x86_mmx_pcmpgt_w:
1054   case Intrinsic::x86_mmx_pcmpgt_d: {
1055     Value *Operands[2];
1056
1057     // Cast the operand to the X86 MMX type.
1058     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1059                                   NewFn->getFunctionType()->getParamType(0),
1060                                   "upgraded.", CI);
1061
1062     switch (NewFn->getIntrinsicID()) {
1063     default:
1064       // Cast to the X86 MMX type.
1065       Operands[1] = new BitCastInst(CI->getArgOperand(1),
1066                                     NewFn->getFunctionType()->getParamType(1),
1067                                     "upgraded.", CI);
1068       break;
1069     case Intrinsic::x86_mmx_pslli_w:
1070     case Intrinsic::x86_mmx_pslli_d:
1071     case Intrinsic::x86_mmx_pslli_q:
1072     case Intrinsic::x86_mmx_psrli_w:
1073     case Intrinsic::x86_mmx_psrli_d:
1074     case Intrinsic::x86_mmx_psrli_q:
1075     case Intrinsic::x86_mmx_psrai_w:
1076     case Intrinsic::x86_mmx_psrai_d:
1077       // These take an i32 as their second parameter.
1078       Operands[1] = CI->getArgOperand(1);
1079       break;
1080     }
1081
1082     ConstructNewCallInst(NewFn, CI, Operands, 2);
1083     break;
1084   }
1085   case Intrinsic::x86_mmx_maskmovq: {
1086     Value *Operands[3];
1087
1088     // Cast the operands to the X86 MMX type.
1089     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1090                                   NewFn->getFunctionType()->getParamType(0),
1091                                   "upgraded.", CI);
1092     Operands[1] = new BitCastInst(CI->getArgOperand(1),
1093                                   NewFn->getFunctionType()->getParamType(1),
1094                                   "upgraded.", CI);
1095     Operands[2] = CI->getArgOperand(2);
1096
1097     ConstructNewCallInst(NewFn, CI, Operands, 3, false);
1098     break;
1099   }
1100   case Intrinsic::x86_mmx_pmovmskb: {
1101     Value *Operands[1];
1102
1103     // Cast the operand to the X86 MMX type.
1104     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1105                                   NewFn->getFunctionType()->getParamType(0),
1106                                   "upgraded.", CI);
1107
1108     ConstructNewCallInst(NewFn, CI, Operands, 1);
1109     break;
1110   }
1111   case Intrinsic::x86_mmx_movnt_dq: {
1112     Value *Operands[2];
1113
1114     Operands[0] = CI->getArgOperand(0);
1115
1116     // Cast the operand to the X86 MMX type.
1117     Operands[1] = new BitCastInst(CI->getArgOperand(1),
1118                                   NewFn->getFunctionType()->getParamType(1),
1119                                   "upgraded.", CI);
1120
1121     ConstructNewCallInst(NewFn, CI, Operands, 2, false);
1122     break;
1123   }
1124   case Intrinsic::x86_mmx_palignr_b: {
1125     Value *Operands[3];
1126
1127     // Cast the operands to the X86 MMX type.
1128     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1129                                   NewFn->getFunctionType()->getParamType(0),
1130                                   "upgraded.", CI);
1131     Operands[1] = new BitCastInst(CI->getArgOperand(1),
1132                                   NewFn->getFunctionType()->getParamType(1),
1133                                   "upgraded.", CI);
1134     Operands[2] = CI->getArgOperand(2);
1135
1136     ConstructNewCallInst(NewFn, CI, Operands, 3);
1137     break;
1138   }
1139   case Intrinsic::x86_mmx_pextr_w: {
1140     Value *Operands[2];
1141
1142     // Cast the operands to the X86 MMX type.
1143     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1144                                   NewFn->getFunctionType()->getParamType(0),
1145                                   "upgraded.", CI);
1146     Operands[1] = CI->getArgOperand(1);
1147
1148     ConstructNewCallInst(NewFn, CI, Operands, 2);
1149     break;
1150   }
1151   case Intrinsic::x86_mmx_pinsr_w: {
1152     Value *Operands[3];
1153
1154     // Cast the operands to the X86 MMX type.
1155     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1156                                   NewFn->getFunctionType()->getParamType(0),
1157                                   "upgraded.", CI);
1158     Operands[1] = CI->getArgOperand(1);
1159     Operands[2] = CI->getArgOperand(2);
1160
1161     ConstructNewCallInst(NewFn, CI, Operands, 3);
1162     break;
1163   }
1164   case Intrinsic::x86_sse_pshuf_w: {
1165     IRBuilder<> Builder(C);
1166     Builder.SetInsertPoint(CI->getParent(), CI);
1167
1168     // Cast the operand to the X86 MMX type.
1169     Value *Operands[2];
1170     Operands[0] =
1171       Builder.CreateBitCast(CI->getArgOperand(0),
1172                             NewFn->getFunctionType()->getParamType(0),
1173                             "upgraded.");
1174     Operands[1] =
1175       Builder.CreateTrunc(CI->getArgOperand(1),
1176                           Type::getInt8Ty(C),
1177                           "upgraded.");
1178
1179     ConstructNewCallInst(NewFn, CI, Operands, 2);
1180     break;
1181   }
1182
1183 #if 0
1184   case Intrinsic::x86_mmx_cvtsi32_si64: {
1185     // The return type needs to be changed.
1186     Value *Operands[1];
1187     Operands[0] = CI->getArgOperand(0);
1188     ConstructNewCallInst(NewFn, CI, Operands, 1);
1189     break;
1190   }
1191   case Intrinsic::x86_mmx_cvtsi64_si32: {
1192     Value *Operands[1];
1193
1194     // Cast the operand to the X86 MMX type.
1195     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1196                                   NewFn->getFunctionType()->getParamType(0),
1197                                   "upgraded.", CI);
1198
1199     ConstructNewCallInst(NewFn, CI, Operands, 1);
1200     break;
1201   }
1202   case Intrinsic::x86_mmx_vec_init_b:
1203   case Intrinsic::x86_mmx_vec_init_w:
1204   case Intrinsic::x86_mmx_vec_init_d: {
1205     // The return type needs to be changed.
1206     Value *Operands[8];
1207     unsigned NumOps = 0;
1208
1209     switch (NewFn->getIntrinsicID()) {
1210     default: break;
1211     case Intrinsic::x86_mmx_vec_init_b: NumOps = 8; break;
1212     case Intrinsic::x86_mmx_vec_init_w: NumOps = 4; break;
1213     case Intrinsic::x86_mmx_vec_init_d: NumOps = 2; break;
1214     }
1215
1216     switch (NewFn->getIntrinsicID()) {
1217     default: break;
1218     case Intrinsic::x86_mmx_vec_init_b:
1219       Operands[7] = CI->getArgOperand(7);
1220       Operands[6] = CI->getArgOperand(6);
1221       Operands[5] = CI->getArgOperand(5);
1222       Operands[4] = CI->getArgOperand(4);
1223       // FALLTHRU
1224     case Intrinsic::x86_mmx_vec_init_w:
1225       Operands[3] = CI->getArgOperand(3);
1226       Operands[2] = CI->getArgOperand(2);
1227       // FALLTHRU
1228     case Intrinsic::x86_mmx_vec_init_d:
1229       Operands[1] = CI->getArgOperand(1);
1230       Operands[0] = CI->getArgOperand(0);
1231       break;
1232     }
1233
1234     ConstructNewCallInst(NewFn, CI, Operands, NumOps);
1235     break;
1236   }
1237   case Intrinsic::x86_mmx_vec_ext_d: {
1238     Value *Operands[2];
1239
1240     // Cast the operand to the X86 MMX type.
1241     Operands[0] = new BitCastInst(CI->getArgOperand(0),
1242                                   NewFn->getFunctionType()->getParamType(0),
1243                                   "upgraded.", CI);
1244     Operands[1] = CI->getArgOperand(1);
1245
1246     ConstructNewCallInst(NewFn, CI, Operands, 2);
1247     break;
1248   }
1249 #endif
1250
1251   case Intrinsic::ctlz:
1252   case Intrinsic::ctpop:
1253   case Intrinsic::cttz: {
1254     //  Build a small vector of the original arguments.
1255     SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
1256
1257     //  Construct a new CallInst
1258     CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
1259                                        "upgraded."+CI->getName(), CI);
1260     NewCI->setTailCall(CI->isTailCall());
1261     NewCI->setCallingConv(CI->getCallingConv());
1262
1263     //  Handle any uses of the old CallInst.
1264     if (!CI->use_empty()) {
1265       //  Check for sign extend parameter attributes on the return values.
1266       bool SrcSExt = NewFn->getAttributes().paramHasAttr(0, Attribute::SExt);
1267       bool DestSExt = F->getAttributes().paramHasAttr(0, Attribute::SExt);
1268
1269       //  Construct an appropriate cast from the new return type to the old.
1270       CastInst *RetCast = CastInst::Create(
1271                             CastInst::getCastOpcode(NewCI, SrcSExt,
1272                                                     F->getReturnType(),
1273                                                     DestSExt),
1274                             NewCI, F->getReturnType(),
1275                             NewCI->getName(), CI);
1276       NewCI->moveBefore(RetCast);
1277
1278       //  Replace all uses of the old call with the new cast which has the
1279       //  correct type.
1280       CI->replaceAllUsesWith(RetCast);
1281     }
1282
1283     //  Clean up the old call now that it has been completely upgraded.
1284     CI->eraseFromParent();
1285   }
1286   break;
1287   case Intrinsic::eh_selector:
1288   case Intrinsic::eh_typeid_for: {
1289     // Only the return type changed.
1290     SmallVector<Value*, 8> Operands(CS.arg_begin(), CS.arg_end());
1291     CallInst *NewCI = CallInst::Create(NewFn, Operands.begin(), Operands.end(),
1292                                        "upgraded." + CI->getName(), CI);
1293     NewCI->setTailCall(CI->isTailCall());
1294     NewCI->setCallingConv(CI->getCallingConv());
1295
1296     //  Handle any uses of the old CallInst.
1297     if (!CI->use_empty()) {
1298       //  Construct an appropriate cast from the new return type to the old.
1299       CastInst *RetCast =
1300         CastInst::Create(CastInst::getCastOpcode(NewCI, true,
1301                                                  F->getReturnType(), true),
1302                          NewCI, F->getReturnType(), NewCI->getName(), CI);
1303       CI->replaceAllUsesWith(RetCast);
1304     }
1305     CI->eraseFromParent();
1306   }
1307   break;
1308   case Intrinsic::memcpy:
1309   case Intrinsic::memmove:
1310   case Intrinsic::memset: {
1311     // Add isVolatile
1312     const llvm::Type *I1Ty = llvm::Type::getInt1Ty(CI->getContext());
1313     Value *Operands[5] = { CI->getArgOperand(0), CI->getArgOperand(1),
1314                            CI->getArgOperand(2), CI->getArgOperand(3),
1315                            llvm::ConstantInt::get(I1Ty, 0) };
1316     CallInst *NewCI = CallInst::Create(NewFn, Operands, Operands+5,
1317                                        CI->getName(), CI);
1318     NewCI->setTailCall(CI->isTailCall());
1319     NewCI->setCallingConv(CI->getCallingConv());
1320     //  Handle any uses of the old CallInst.
1321     if (!CI->use_empty())
1322       //  Replace all uses of the old call with the new cast which has the
1323       //  correct type.
1324       CI->replaceAllUsesWith(NewCI);
1325
1326     //  Clean up the old call now that it has been completely upgraded.
1327     CI->eraseFromParent();
1328     break;
1329   }
1330   }
1331 }
1332
1333 // This tests each Function to determine if it needs upgrading. When we find
1334 // one we are interested in, we then upgrade all calls to reflect the new
1335 // function.
1336 void llvm::UpgradeCallsToIntrinsic(Function* F) {
1337   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
1338
1339   // Upgrade the function and check if it is a totaly new function.
1340   Function* NewFn;
1341   if (UpgradeIntrinsicFunction(F, NewFn)) {
1342     if (NewFn != F) {
1343       // Replace all uses to the old function with the new one if necessary.
1344       for (Value::use_iterator UI = F->use_begin(), UE = F->use_end();
1345            UI != UE; ) {
1346         if (CallInst* CI = dyn_cast<CallInst>(*UI++))
1347           UpgradeIntrinsicCall(CI, NewFn);
1348       }
1349       // Remove old function, no longer used, from the module.
1350       F->eraseFromParent();
1351     }
1352   }
1353 }
1354
1355 /// This function strips all debug info intrinsics, except for llvm.dbg.declare.
1356 /// If an llvm.dbg.declare intrinsic is invalid, then this function simply
1357 /// strips that use.
1358 void llvm::CheckDebugInfoIntrinsics(Module *M) {
1359
1360
1361   if (Function *FuncStart = M->getFunction("llvm.dbg.func.start")) {
1362     while (!FuncStart->use_empty()) {
1363       CallInst *CI = cast<CallInst>(FuncStart->use_back());
1364       CI->eraseFromParent();
1365     }
1366     FuncStart->eraseFromParent();
1367   }
1368
1369   if (Function *StopPoint = M->getFunction("llvm.dbg.stoppoint")) {
1370     while (!StopPoint->use_empty()) {
1371       CallInst *CI = cast<CallInst>(StopPoint->use_back());
1372       CI->eraseFromParent();
1373     }
1374     StopPoint->eraseFromParent();
1375   }
1376
1377   if (Function *RegionStart = M->getFunction("llvm.dbg.region.start")) {
1378     while (!RegionStart->use_empty()) {
1379       CallInst *CI = cast<CallInst>(RegionStart->use_back());
1380       CI->eraseFromParent();
1381     }
1382     RegionStart->eraseFromParent();
1383   }
1384
1385   if (Function *RegionEnd = M->getFunction("llvm.dbg.region.end")) {
1386     while (!RegionEnd->use_empty()) {
1387       CallInst *CI = cast<CallInst>(RegionEnd->use_back());
1388       CI->eraseFromParent();
1389     }
1390     RegionEnd->eraseFromParent();
1391   }
1392
1393   if (Function *Declare = M->getFunction("llvm.dbg.declare")) {
1394     if (!Declare->use_empty()) {
1395       DbgDeclareInst *DDI = cast<DbgDeclareInst>(Declare->use_back());
1396       if (!isa<MDNode>(DDI->getArgOperand(0)) ||
1397           !isa<MDNode>(DDI->getArgOperand(1))) {
1398         while (!Declare->use_empty()) {
1399           CallInst *CI = cast<CallInst>(Declare->use_back());
1400           CI->eraseFromParent();
1401         }
1402         Declare->eraseFromParent();
1403       }
1404     }
1405   }
1406 }