llvm/lib/CodeGen/ExpandLargeFpConvert.cpp

   1 //===--- ExpandLargeFpConvert.cpp - Expand large fp convert----------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9
  10 // This pass expands ‘fptoui .. to’, ‘fptosi .. to’, ‘uitofp .. to’,
  11 // ‘sitofp .. to’ instructions with a bitwidth above a threshold into
  12 // auto-generated functions. This is useful for targets like x86_64 that cannot
  13 // lower fp convertions with more than 128 bits.
  14 //
  15 //===----------------------------------------------------------------------===//
  16
  17 #include "llvm/ADT/SmallVector.h"
  18 #include "llvm/ADT/StringExtras.h"
  19 #include "llvm/Analysis/GlobalsModRef.h"
  20 #include "llvm/CodeGen/Passes.h"
  21 #include "llvm/CodeGen/TargetLowering.h"
  22 #include "llvm/CodeGen/TargetPassConfig.h"
  23 #include "llvm/CodeGen/TargetSubtargetInfo.h"
  24 #include "llvm/IR/IRBuilder.h"
  25 #include "llvm/IR/InstIterator.h"
  26 #include "llvm/IR/PassManager.h"
  27 #include "llvm/InitializePasses.h"
  28 #include "llvm/Pass.h"
  29 #include "llvm/Support/CommandLine.h"
  30 #include "llvm/Target/TargetMachine.h"
  31
  32 using namespace llvm;
  33
  34 static cl::opt<unsigned>
  35     ExpandFpConvertBits("expand-fp-convert-bits", cl::Hidden,
  36                      cl::init(llvm::IntegerType::MAX_INT_BITS),
  37                      cl::desc("fp convert instructions on integers with "
  38                               "more than <N> bits are expanded."));
  39
  40 /// Generate code to convert a fp number to integer, replacing FPToS(U)I with
  41 /// the generated code. This currently generates code similarly to compiler-rt's
  42 /// implementations.
  43 ///
  44 /// An example IR generated from compiler-rt/fixsfdi.c looks like below:
  45 /// define dso_local i64 @foo(float noundef %a) local_unnamed_addr #0 {
  46 /// entry:
  47 ///   %0 = bitcast float %a to i32
  48 ///   %conv.i = zext i32 %0 to i64
  49 ///   %tobool.not = icmp sgt i32 %0, -1
  50 ///   %conv = select i1 %tobool.not, i64 1, i64 -1
  51 ///   %and = lshr i64 %conv.i, 23
  52 ///   %shr = and i64 %and, 255
  53 ///   %and2 = and i64 %conv.i, 8388607
  54 ///   %or = or i64 %and2, 8388608
  55 ///   %cmp = icmp ult i64 %shr, 127
  56 ///   br i1 %cmp, label %cleanup, label %if.end
  57 ///
  58 /// if.end:                                           ; preds = %entry
  59 ///   %sub = add nuw nsw i64 %shr, 4294967169
  60 ///   %conv5 = and i64 %sub, 4294967232
  61 ///   %cmp6.not = icmp eq i64 %conv5, 0
  62 ///   br i1 %cmp6.not, label %if.end12, label %if.then8
  63 ///
  64 /// if.then8:                                         ; preds = %if.end
  65 ///   %cond11 = select i1 %tobool.not, i64 9223372036854775807, i64 -9223372036854775808
  66 ///   br label %cleanup
  67 ///
  68 /// if.end12:                                         ; preds = %if.end
  69 ///   %cmp13 = icmp ult i64 %shr, 150
  70 ///   br i1 %cmp13, label %if.then15, label %if.else
  71 ///
  72 /// if.then15:                                        ; preds = %if.end12
  73 ///   %sub16 = sub nuw nsw i64 150, %shr
  74 ///   %shr17 = lshr i64 %or, %sub16
  75 ///   %mul = mul nsw i64 %shr17, %conv
  76 ///   br label %cleanup
  77 ///
  78 /// if.else:                                          ; preds = %if.end12
  79 ///   %sub18 = add nsw i64 %shr, -150
  80 ///   %shl = shl i64 %or, %sub18
  81 ///   %mul19 = mul nsw i64 %shl, %conv
  82 ///   br label %cleanup
  83 ///
  84 /// cleanup:                                          ; preds = %entry, %if.else, %if.then15, %if.then8
  85 ///   %retval.0 = phi i64 [ %cond11, %if.then8 ], [ %mul, %if.then15 ], [ %mul19, %if.else ], [ 0, %entry ]
  86 ///   ret i64 %retval.0
  87 /// }
  88 ///
  89 /// Replace fp to integer with generated code.
  90 static void expandFPToI(Instruction *FPToI) {
  91   IRBuilder<> Builder(FPToI);
  92   auto *FloatVal = FPToI->getOperand(0);
  93   IntegerType *IntTy = cast<IntegerType>(FPToI->getType());
  94
  95   unsigned BitWidth = FPToI->getType()->getIntegerBitWidth();
  96   unsigned FPMantissaWidth = FloatVal->getType()->getFPMantissaWidth() - 1;
  97
  98   // FIXME: fp16's range is covered by i32. So `fptoi half` can convert
  99   // to i32 first following a sext/zext to target integer type.
 100   Value *A1 = nullptr;
 101   if (FloatVal->getType()->isHalfTy()) {
 102     if (FPToI->getOpcode() == Instruction::FPToUI) {
 103       Value *A0 = Builder.CreateFPToUI(FloatVal, Builder.getIntNTy(32));
 104       A1 = Builder.CreateZExt(A0, IntTy);
 105     } else { // FPToSI
 106       Value *A0 = Builder.CreateFPToSI(FloatVal, Builder.getIntNTy(32));
 107       A1 = Builder.CreateSExt(A0, IntTy);
 108     }
 109     FPToI->replaceAllUsesWith(A1);
 110     FPToI->dropAllReferences();
 111     FPToI->eraseFromParent();
 112     return;
 113   }
 114
 115   // fp80 conversion is implemented by fpext to fp128 first then do the
 116   // conversion.
 117   FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth;
 118   unsigned FloatWidth = PowerOf2Ceil(FPMantissaWidth);
 119   unsigned ExponentWidth = FloatWidth - FPMantissaWidth - 1;
 120   unsigned ExponentBias = (1 << (ExponentWidth - 1)) - 1;
 121   Value *ImplicitBit = Builder.CreateShl(
 122       Builder.getIntN(BitWidth, 1), Builder.getIntN(BitWidth, FPMantissaWidth));
 123   Value *SignificandMask =
 124       Builder.CreateSub(ImplicitBit, Builder.getIntN(BitWidth, 1));
 125   Value *NegOne = Builder.CreateSExt(
 126       ConstantInt::getSigned(Builder.getInt32Ty(), -1), IntTy);
 127   Value *NegInf =
 128       Builder.CreateShl(ConstantInt::getSigned(IntTy, 1),
 129                         ConstantInt::getSigned(IntTy, BitWidth - 1));
 130
 131   BasicBlock *Entry = Builder.GetInsertBlock();
 132   Function *F = Entry->getParent();
 133   Entry->setName(Twine(Entry->getName(), "fp-to-i-entry"));
 134   BasicBlock *End =
 135       Entry->splitBasicBlock(Builder.GetInsertPoint(), "fp-to-i-cleanup");
 136   BasicBlock *IfEnd =
 137       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-end", F, End);
 138   BasicBlock *IfThen5 =
 139       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-then5", F, End);
 140   BasicBlock *IfEnd9 =
 141       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-end9", F, End);
 142   BasicBlock *IfThen12 =
 143       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-then12", F, End);
 144   BasicBlock *IfElse =
 145       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-else", F, End);
 146
 147   Entry->getTerminator()->eraseFromParent();
 148
 149   // entry:
 150   Builder.SetInsertPoint(Entry);
 151   Value *FloatVal0 = FloatVal;
 152   // fp80 conversion is implemented by fpext to fp128 first then do the
 153   // conversion.
 154   if (FloatVal->getType()->isX86_FP80Ty())
 155     FloatVal0 =
 156         Builder.CreateFPExt(FloatVal, Type::getFP128Ty(Builder.getContext()));
 157   Value *ARep0 =
 158       Builder.CreateBitCast(FloatVal0, Builder.getIntNTy(FloatWidth));
 159   Value *ARep = Builder.CreateZExt(ARep0, FPToI->getType());
 160   Value *PosOrNeg = Builder.CreateICmpSGT(
 161       ARep0, ConstantInt::getSigned(Builder.getIntNTy(FloatWidth), -1));
 162   Value *Sign = Builder.CreateSelect(PosOrNeg, ConstantInt::getSigned(IntTy, 1),
 163                                      ConstantInt::getSigned(IntTy, -1));
 164   Value *And =
 165       Builder.CreateLShr(ARep, Builder.getIntN(BitWidth, FPMantissaWidth));
 166   Value *And2 = Builder.CreateAnd(
 167       And, Builder.getIntN(BitWidth, (1 << ExponentWidth) - 1));
 168   Value *Abs = Builder.CreateAnd(ARep, SignificandMask);
 169   Value *Or = Builder.CreateOr(Abs, ImplicitBit);
 170   Value *Cmp =
 171       Builder.CreateICmpULT(And2, Builder.getIntN(BitWidth, ExponentBias));
 172   Builder.CreateCondBr(Cmp, End, IfEnd);
 173
 174   // if.end:
 175   Builder.SetInsertPoint(IfEnd);
 176   Value *Add1 = Builder.CreateAdd(
 177       And2, ConstantInt::getSigned(IntTy, -int64_t(ExponentBias + BitWidth)));
 178   Value *Cmp3 =
 179       Builder.CreateICmpULT(Add1, ConstantInt::getSigned(IntTy, -BitWidth));
 180   Builder.CreateCondBr(Cmp3, IfThen5, IfEnd9);
 181
 182   // if.then5:
 183   Builder.SetInsertPoint(IfThen5);
 184   Value *PosInf = Builder.CreateXor(NegOne, NegInf);
 185   Value *Cond8 = Builder.CreateSelect(PosOrNeg, PosInf, NegInf);
 186   Builder.CreateBr(End);
 187
 188   // if.end9:
 189   Builder.SetInsertPoint(IfEnd9);
 190   Value *Cmp10 = Builder.CreateICmpULT(
 191       And2, Builder.getIntN(BitWidth, ExponentBias + FPMantissaWidth));
 192   Builder.CreateCondBr(Cmp10, IfThen12, IfElse);
 193
 194   // if.then12:
 195   Builder.SetInsertPoint(IfThen12);
 196   Value *Sub13 = Builder.CreateSub(
 197       Builder.getIntN(BitWidth, ExponentBias + FPMantissaWidth), And2);
 198   Value *Shr14 = Builder.CreateLShr(Or, Sub13);
 199   Value *Mul = Builder.CreateMul(Shr14, Sign);
 200   Builder.CreateBr(End);
 201
 202   // if.else:
 203   Builder.SetInsertPoint(IfElse);
 204   Value *Sub15 = Builder.CreateAdd(
 205       And2,
 206       ConstantInt::getSigned(IntTy, -(ExponentBias + FPMantissaWidth)));
 207   Value *Shl = Builder.CreateShl(Or, Sub15);
 208   Value *Mul16 = Builder.CreateMul(Shl, Sign);
 209   Builder.CreateBr(End);
 210
 211   // cleanup:
 212   Builder.SetInsertPoint(End, End->begin());
 213   PHINode *Retval0 = Builder.CreatePHI(FPToI->getType(), 4);
 214
 215   Retval0->addIncoming(Cond8, IfThen5);
 216   Retval0->addIncoming(Mul, IfThen12);
 217   Retval0->addIncoming(Mul16, IfElse);
 218   Retval0->addIncoming(Builder.getIntN(BitWidth, 0), Entry);
 219
 220   FPToI->replaceAllUsesWith(Retval0);
 221   FPToI->dropAllReferences();
 222   FPToI->eraseFromParent();
 223 }
 224
 225 /// Generate code to convert a fp number to integer, replacing S(U)IToFP with
 226 /// the generated code. This currently generates code similarly to compiler-rt's
 227 /// implementations. This implementation has an implicit assumption that integer
 228 /// width is larger than fp.
 229 ///
 230 /// An example IR generated from compiler-rt/floatdisf.c looks like below:
 231 /// define dso_local float @__floatdisf(i64 noundef %a) local_unnamed_addr #0 {
 232 /// entry:
 233 ///   %cmp = icmp eq i64 %a, 0
 234 ///   br i1 %cmp, label %return, label %if.end
 235 ///
 236 /// if.end:                                           ; preds = %entry
 237 ///   %shr = ashr i64 %a, 63
 238 ///   %xor = xor i64 %shr, %a
 239 ///   %sub = sub nsw i64 %xor, %shr
 240 ///   %0 = tail call i64 @llvm.ctlz.i64(i64 %sub, i1 true), !range !5
 241 ///   %cast = trunc i64 %0 to i32
 242 ///   %sub1 = sub nuw nsw i32 64, %cast
 243 ///   %sub2 = xor i32 %cast, 63
 244 ///   %cmp3 = icmp ult i32 %cast, 40
 245 ///   br i1 %cmp3, label %if.then4, label %if.else
 246 ///
 247 /// if.then4:                                         ; preds = %if.end
 248 ///   switch i32 %sub1, label %sw.default [
 249 ///     i32 25, label %sw.bb
 250 ///     i32 26, label %sw.epilog
 251 ///   ]
 252 ///
 253 /// sw.bb:                                            ; preds = %if.then4
 254 ///   %shl = shl i64 %sub, 1
 255 ///   br label %sw.epilog
 256 ///
 257 /// sw.default:                                       ; preds = %if.then4
 258 ///   %sub5 = sub nsw i64 38, %0
 259 ///   %sh_prom = and i64 %sub5, 4294967295
 260 ///   %shr6 = lshr i64 %sub, %sh_prom
 261 ///   %shr9 = lshr i64 274877906943, %0
 262 ///   %and = and i64 %shr9, %sub
 263 ///   %cmp10 = icmp ne i64 %and, 0
 264 ///   %conv11 = zext i1 %cmp10 to i64
 265 ///   %or = or i64 %shr6, %conv11
 266 ///   br label %sw.epilog
 267 ///
 268 /// sw.epilog:                                        ; preds = %sw.default, %if.then4, %sw.bb
 269 ///   %a.addr.0 = phi i64 [ %or, %sw.default ], [ %sub, %if.then4 ], [ %shl, %sw.bb ]
 270 ///   %1 = lshr i64 %a.addr.0, 2
 271 ///   %2 = and i64 %1, 1
 272 ///   %or16 = or i64 %2, %a.addr.0
 273 ///   %inc = add nsw i64 %or16, 1
 274 ///   %3 = and i64 %inc, 67108864
 275 ///   %tobool.not = icmp eq i64 %3, 0
 276 ///   %spec.select.v = select i1 %tobool.not, i64 2, i64 3
 277 ///   %spec.select = ashr i64 %inc, %spec.select.v
 278 ///   %spec.select56 = select i1 %tobool.not, i32 %sub2, i32 %sub1
 279 ///   br label %if.end26
 280 ///
 281 /// if.else:                                          ; preds = %if.end
 282 ///   %sub23 = add nuw nsw i64 %0, 4294967256
 283 ///   %sh_prom24 = and i64 %sub23, 4294967295
 284 ///   %shl25 = shl i64 %sub, %sh_prom24
 285 ///   br label %if.end26
 286 ///
 287 /// if.end26:                                         ; preds = %sw.epilog, %if.else
 288 ///   %a.addr.1 = phi i64 [ %shl25, %if.else ], [ %spec.select, %sw.epilog ]
 289 ///   %e.0 = phi i32 [ %sub2, %if.else ], [ %spec.select56, %sw.epilog ]
 290 ///   %conv27 = trunc i64 %shr to i32
 291 ///   %and28 = and i32 %conv27, -2147483648
 292 ///   %add = shl nuw nsw i32 %e.0, 23
 293 ///   %shl29 = add nuw nsw i32 %add, 1065353216
 294 ///   %conv31 = trunc i64 %a.addr.1 to i32
 295 ///   %and32 = and i32 %conv31, 8388607
 296 ///   %or30 = or i32 %and32, %and28
 297 ///   %or33 = or i32 %or30, %shl29
 298 ///   %4 = bitcast i32 %or33 to float
 299 ///   br label %return
 300 ///
 301 /// return:                                           ; preds = %entry, %if.end26
 302 ///   %retval.0 = phi float [ %4, %if.end26 ], [ 0.000000e+00, %entry ]
 303 ///   ret float %retval.0
 304 /// }
 305 ///
 306 /// Replace integer to fp with generated code.
 307 static void expandIToFP(Instruction *IToFP) {
 308   IRBuilder<> Builder(IToFP);
 309   auto *IntVal = IToFP->getOperand(0);
 310   IntegerType *IntTy = cast<IntegerType>(IntVal->getType());
 311
 312   unsigned BitWidth = IntVal->getType()->getIntegerBitWidth();
 313   unsigned FPMantissaWidth = IToFP->getType()->getFPMantissaWidth() - 1;
 314   // fp80 conversion is implemented by conversion tp fp128 first following
 315   // a fptrunc to fp80.
 316   FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth;
 317   // FIXME: As there is no related builtins added in compliler-rt,
 318   // here currently utilized the fp32 <-> fp16 lib calls to implement.
 319   FPMantissaWidth = FPMantissaWidth == 10 ? 23 : FPMantissaWidth;
 320   unsigned FloatWidth = PowerOf2Ceil(FPMantissaWidth);
 321   bool IsSigned = IToFP->getOpcode() == Instruction::SIToFP;
 322
 323   assert(BitWidth > FloatWidth && "Unexpected conversion. expandIToFP() "
 324                                   "assumes integer width is larger than fp.");
 325
 326   Value *Temp1 =
 327       Builder.CreateShl(Builder.getIntN(BitWidth, 1),
 328                         Builder.getIntN(BitWidth, FPMantissaWidth + 3));
 329
 330   BasicBlock *Entry = Builder.GetInsertBlock();
 331   Function *F = Entry->getParent();
 332   Entry->setName(Twine(Entry->getName(), "itofp-entry"));
 333   BasicBlock *End =
 334       Entry->splitBasicBlock(Builder.GetInsertPoint(), "itofp-return");
 335   BasicBlock *IfEnd =
 336       BasicBlock::Create(Builder.getContext(), "itofp-if-end", F, End);
 337   BasicBlock *IfThen4 =
 338       BasicBlock::Create(Builder.getContext(), "itofp-if-then4", F, End);
 339   BasicBlock *SwBB =
 340       BasicBlock::Create(Builder.getContext(), "itofp-sw-bb", F, End);
 341   BasicBlock *SwDefault =
 342       BasicBlock::Create(Builder.getContext(), "itofp-sw-default", F, End);
 343   BasicBlock *SwEpilog =
 344       BasicBlock::Create(Builder.getContext(), "itofp-sw-epilog", F, End);
 345   BasicBlock *IfThen20 =
 346       BasicBlock::Create(Builder.getContext(), "itofp-if-then20", F, End);
 347   BasicBlock *IfElse =
 348       BasicBlock::Create(Builder.getContext(), "itofp-if-else", F, End);
 349   BasicBlock *IfEnd26 =
 350       BasicBlock::Create(Builder.getContext(), "itofp-if-end26", F, End);
 351
 352   Entry->getTerminator()->eraseFromParent();
 353
 354   Function *CTLZ =
 355       Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz, IntTy);
 356   ConstantInt *True = Builder.getTrue();
 357
 358   // entry:
 359   Builder.SetInsertPoint(Entry);
 360   Value *Cmp = Builder.CreateICmpEQ(IntVal, ConstantInt::getSigned(IntTy, 0));
 361   Builder.CreateCondBr(Cmp, End, IfEnd);
 362
 363   // if.end:
 364   Builder.SetInsertPoint(IfEnd);
 365   Value *Shr =
 366       Builder.CreateAShr(IntVal, Builder.getIntN(BitWidth, BitWidth - 1));
 367   Value *Xor = Builder.CreateXor(Shr, IntVal);
 368   Value *Sub = Builder.CreateSub(Xor, Shr);
 369   Value *Call = Builder.CreateCall(CTLZ, {IsSigned ? Sub : IntVal, True});
 370   Value *Cast = Builder.CreateTrunc(Call, Builder.getInt32Ty());
 371   int BitWidthNew = FloatWidth == 128 ? BitWidth : 32;
 372   Value *Sub1 = Builder.CreateSub(Builder.getIntN(BitWidthNew, BitWidth),
 373                                   FloatWidth == 128 ? Call : Cast);
 374   Value *Sub2 = Builder.CreateSub(Builder.getIntN(BitWidthNew, BitWidth - 1),
 375                                   FloatWidth == 128 ? Call : Cast);
 376   Value *Cmp3 = Builder.CreateICmpSGT(
 377       Sub2, Builder.getIntN(BitWidthNew, FPMantissaWidth + 1));
 378   Builder.CreateCondBr(Cmp3, IfThen4, IfElse);
 379
 380   // if.then4:
 381   Builder.SetInsertPoint(IfThen4);
 382   llvm::SwitchInst *SI = Builder.CreateSwitch(Sub1, SwDefault);
 383   SI->addCase(Builder.getIntN(BitWidthNew, FPMantissaWidth + 2), SwBB);
 384   SI->addCase(Builder.getIntN(BitWidthNew, FPMantissaWidth + 3), SwEpilog);
 385
 386   // sw.bb:
 387   Builder.SetInsertPoint(SwBB);
 388   Value *Shl =
 389       Builder.CreateShl(IsSigned ? Sub : IntVal, Builder.getIntN(BitWidth, 1));
 390   Builder.CreateBr(SwEpilog);
 391
 392   // sw.default:
 393   Builder.SetInsertPoint(SwDefault);
 394   Value *Sub5 = Builder.CreateSub(
 395       Builder.getIntN(BitWidthNew, BitWidth - FPMantissaWidth - 3),
 396       FloatWidth == 128 ? Call : Cast);
 397   Value *ShProm = Builder.CreateZExt(Sub5, IntTy);
 398   Value *Shr6 = Builder.CreateLShr(IsSigned ? Sub : IntVal,
 399                                    FloatWidth == 128 ? Sub5 : ShProm);
 400   Value *Sub8 =
 401       Builder.CreateAdd(FloatWidth == 128 ? Call : Cast,
 402                         Builder.getIntN(BitWidthNew, FPMantissaWidth + 3));
 403   Value *ShProm9 = Builder.CreateZExt(Sub8, IntTy);
 404   Value *Shr9 = Builder.CreateLShr(ConstantInt::getSigned(IntTy, -1),
 405                                    FloatWidth == 128 ? Sub8 : ShProm9);
 406   Value *And = Builder.CreateAnd(Shr9, IsSigned ? Sub : IntVal);
 407   Value *Cmp10 = Builder.CreateICmpNE(And, Builder.getIntN(BitWidth, 0));
 408   Value *Conv11 = Builder.CreateZExt(Cmp10, IntTy);
 409   Value *Or = Builder.CreateOr(Shr6, Conv11);
 410   Builder.CreateBr(SwEpilog);
 411
 412   // sw.epilog:
 413   Builder.SetInsertPoint(SwEpilog);
 414   PHINode *AAddr0 = Builder.CreatePHI(IntTy, 3);
 415   AAddr0->addIncoming(Or, SwDefault);
 416   AAddr0->addIncoming(IsSigned ? Sub : IntVal, IfThen4);
 417   AAddr0->addIncoming(Shl, SwBB);
 418   Value *A0 = Builder.CreateTrunc(AAddr0, Builder.getInt32Ty());
 419   Value *A1 = Builder.CreateLShr(A0, Builder.getIntN(32, 2));
 420   Value *A2 = Builder.CreateAnd(A1, Builder.getIntN(32, 1));
 421   Value *Conv16 = Builder.CreateZExt(A2, IntTy);
 422   Value *Or17 = Builder.CreateOr(AAddr0, Conv16);
 423   Value *Inc = Builder.CreateAdd(Or17, Builder.getIntN(BitWidth, 1));
 424   Value *Shr18 = nullptr;
 425   if (IsSigned)
 426     Shr18 = Builder.CreateAShr(Inc, Builder.getIntN(BitWidth, 2));
 427   else
 428     Shr18 = Builder.CreateLShr(Inc, Builder.getIntN(BitWidth, 2));
 429   Value *A3 = Builder.CreateAnd(Inc, Temp1, "a3");
 430   Value *PosOrNeg = Builder.CreateICmpEQ(A3, Builder.getIntN(BitWidth, 0));
 431   Value *ExtractT60 = Builder.CreateTrunc(Shr18, Builder.getIntNTy(FloatWidth));
 432   Value *Extract63 = Builder.CreateLShr(Shr18, Builder.getIntN(BitWidth, 32));
 433   Value *ExtractT64 = nullptr;
 434   if (FloatWidth > 80)
 435     ExtractT64 = Builder.CreateTrunc(Sub2, Builder.getInt64Ty());
 436   else
 437     ExtractT64 = Builder.CreateTrunc(Extract63, Builder.getInt32Ty());
 438   Builder.CreateCondBr(PosOrNeg, IfEnd26, IfThen20);
 439
 440   // if.then20
 441   Builder.SetInsertPoint(IfThen20);
 442   Value *Shr21 = nullptr;
 443   if (IsSigned)
 444     Shr21 = Builder.CreateAShr(Inc, Builder.getIntN(BitWidth, 3));
 445   else
 446     Shr21 = Builder.CreateLShr(Inc, Builder.getIntN(BitWidth, 3));
 447   Value *ExtractT = Builder.CreateTrunc(Shr21, Builder.getIntNTy(FloatWidth));
 448   Value *Extract = Builder.CreateLShr(Shr21, Builder.getIntN(BitWidth, 32));
 449   Value *ExtractT62 = nullptr;
 450   if (FloatWidth > 80)
 451     ExtractT62 = Builder.CreateTrunc(Sub1, Builder.getIntNTy(64));
 452   else
 453     ExtractT62 = Builder.CreateTrunc(Extract, Builder.getIntNTy(32));
 454   Builder.CreateBr(IfEnd26);
 455
 456   // if.else:
 457   Builder.SetInsertPoint(IfElse);
 458   Value *Sub24 = Builder.CreateAdd(
 459       FloatWidth == 128 ? Call : Cast,
 460       ConstantInt::getSigned(Builder.getIntNTy(BitWidthNew),
 461                              -(BitWidth - FPMantissaWidth - 1)));
 462   Value *ShProm25 = Builder.CreateZExt(Sub24, IntTy);
 463   Value *Shl26 = Builder.CreateShl(IsSigned ? Sub : IntVal,
 464                                    FloatWidth == 128 ? Sub24 : ShProm25);
 465   Value *ExtractT61 = Builder.CreateTrunc(Shl26, Builder.getIntNTy(FloatWidth));
 466   Value *Extract65 = Builder.CreateLShr(Shl26, Builder.getIntN(BitWidth, 32));
 467   Value *ExtractT66 = nullptr;
 468   if (FloatWidth > 80)
 469     ExtractT66 = Builder.CreateTrunc(Sub2, Builder.getIntNTy(64));
 470   else
 471     ExtractT66 = Builder.CreateTrunc(Extract65, Builder.getInt32Ty());
 472   Builder.CreateBr(IfEnd26);
 473
 474   // if.end26:
 475   Builder.SetInsertPoint(IfEnd26);
 476   PHINode *AAddr1Off0 = Builder.CreatePHI(Builder.getIntNTy(FloatWidth), 3);
 477   AAddr1Off0->addIncoming(ExtractT, IfThen20);
 478   AAddr1Off0->addIncoming(ExtractT60, SwEpilog);
 479   AAddr1Off0->addIncoming(ExtractT61, IfElse);
 480   PHINode *AAddr1Off32 = nullptr;
 481   if (FloatWidth > 32) {
 482     AAddr1Off32 =
 483         Builder.CreatePHI(Builder.getIntNTy(FloatWidth > 80 ? 64 : 32), 3);
 484     AAddr1Off32->addIncoming(ExtractT62, IfThen20);
 485     AAddr1Off32->addIncoming(ExtractT64, SwEpilog);
 486     AAddr1Off32->addIncoming(ExtractT66, IfElse);
 487   }
 488   PHINode *E0 = nullptr;
 489   if (FloatWidth <= 80) {
 490     E0 = Builder.CreatePHI(Builder.getIntNTy(BitWidthNew), 3);
 491     E0->addIncoming(Sub1, IfThen20);
 492     E0->addIncoming(Sub2, SwEpilog);
 493     E0->addIncoming(Sub2, IfElse);
 494   }
 495   Value *And29 = nullptr;
 496   if (FloatWidth > 80) {
 497     Value *Temp2 = Builder.CreateShl(Builder.getIntN(BitWidth, 1),
 498                                      Builder.getIntN(BitWidth, 63));
 499     And29 = Builder.CreateAnd(Shr, Temp2, "and29");
 500   } else {
 501     Value *Conv28 = Builder.CreateTrunc(Shr, Builder.getIntNTy(32));
 502     And29 = Builder.CreateAnd(
 503         Conv28, ConstantInt::getSigned(Builder.getIntNTy(32), 0x80000000));
 504   }
 505   unsigned TempMod = FPMantissaWidth % 32;
 506   Value *And34 = nullptr;
 507   Value *Shl30 = nullptr;
 508   if (FloatWidth > 80) {
 509     TempMod += 32;
 510     Value *Add = Builder.CreateShl(AAddr1Off32, Builder.getIntN(64, TempMod));
 511     Shl30 = Builder.CreateAdd(
 512         Add,
 513         Builder.getIntN(64, ((1ull << (62ull - TempMod)) - 1ull) << TempMod));
 514     And34 = Builder.CreateZExt(Shl30, Builder.getIntNTy(128));
 515   } else {
 516     Value *Add = Builder.CreateShl(E0, Builder.getIntN(32, TempMod));
 517     Shl30 = Builder.CreateAdd(
 518         Add, Builder.getIntN(32, ((1 << (30 - TempMod)) - 1) << TempMod));
 519     And34 = Builder.CreateAnd(FloatWidth > 32 ? AAddr1Off32 : AAddr1Off0,
 520                               Builder.getIntN(32, (1 << TempMod) - 1));
 521   }
 522   Value *Or35 = nullptr;
 523   if (FloatWidth > 80) {
 524     Value *And29Trunc = Builder.CreateTrunc(And29, Builder.getIntNTy(128));
 525     Value *Or31 = Builder.CreateOr(And29Trunc, And34);
 526     Value *Or34 = Builder.CreateShl(Or31, Builder.getIntN(128, 64));
 527     Value *Temp3 = Builder.CreateShl(Builder.getIntN(128, 1),
 528                                      Builder.getIntN(128, FPMantissaWidth));
 529     Value *Temp4 = Builder.CreateSub(Temp3, Builder.getIntN(128, 1));
 530     Value *A6 = Builder.CreateAnd(AAddr1Off0, Temp4);
 531     Or35 = Builder.CreateOr(Or34, A6);
 532   } else {
 533     Value *Or31 = Builder.CreateOr(And34, And29);
 534     Or35 = Builder.CreateOr(IsSigned ? Or31 : And34, Shl30);
 535   }
 536   Value *A4 = nullptr;
 537   if (IToFP->getType()->isDoubleTy()) {
 538     Value *ZExt1 = Builder.CreateZExt(Or35, Builder.getIntNTy(FloatWidth));
 539     Value *Shl1 = Builder.CreateShl(ZExt1, Builder.getIntN(FloatWidth, 32));
 540     Value *And1 =
 541         Builder.CreateAnd(AAddr1Off0, Builder.getIntN(FloatWidth, 0xFFFFFFFF));
 542     Value *Or1 = Builder.CreateOr(Shl1, And1);
 543     A4 = Builder.CreateBitCast(Or1, IToFP->getType());
 544   } else if (IToFP->getType()->isX86_FP80Ty()) {
 545     Value *A40 =
 546         Builder.CreateBitCast(Or35, Type::getFP128Ty(Builder.getContext()));
 547     A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
 548   } else if (IToFP->getType()->isHalfTy()) {
 549     // Deal with "half" situation. This is a workaround since we don't have
 550     // floattihf.c currently as referring.
 551     Value *A40 =
 552         Builder.CreateBitCast(Or35, Type::getFloatTy(Builder.getContext()));
 553     A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
 554   } else // float type
 555     A4 = Builder.CreateBitCast(Or35, IToFP->getType());
 556   Builder.CreateBr(End);
 557
 558   // return:
 559   Builder.SetInsertPoint(End, End->begin());
 560   PHINode *Retval0 = Builder.CreatePHI(IToFP->getType(), 2);
 561   Retval0->addIncoming(A4, IfEnd26);
 562   Retval0->addIncoming(ConstantFP::getZero(IToFP->getType(), false), Entry);
 563
 564   IToFP->replaceAllUsesWith(Retval0);
 565   IToFP->dropAllReferences();
 566   IToFP->eraseFromParent();
 567 }
 568
 569 static bool runImpl(Function &F, const TargetLowering &TLI) {
 570   SmallVector<Instruction *, 4> Replace;
 571   bool Modified = false;
 572
 573   unsigned MaxLegalFpConvertBitWidth =
 574       TLI.getMaxLargeFPConvertBitWidthSupported();
 575   if (ExpandFpConvertBits != llvm::IntegerType::MAX_INT_BITS)
 576     MaxLegalFpConvertBitWidth = ExpandFpConvertBits;
 577
 578   if (MaxLegalFpConvertBitWidth >= llvm::IntegerType::MAX_INT_BITS)
 579     return false;
 580
 581   for (auto &I : instructions(F)) {
 582     switch (I.getOpcode()) {
 583     case Instruction::FPToUI:
 584     case Instruction::FPToSI: {
 585       // TODO: This pass doesn't handle vectors.
 586       if (I.getOperand(0)->getType()->isVectorTy())
 587         continue;
 588
 589       auto *IntTy = dyn_cast<IntegerType>(I.getType());
 590       if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth)
 591         continue;
 592
 593       Replace.push_back(&I);
 594       Modified = true;
 595       break;
 596     }
 597     case Instruction::UIToFP:
 598     case Instruction::SIToFP: {
 599       // TODO: This pass doesn't handle vectors.
 600       if (I.getOperand(0)->getType()->isVectorTy())
 601         continue;
 602
 603       auto *IntTy = dyn_cast<IntegerType>(I.getOperand(0)->getType());
 604       if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth)
 605         continue;
 606
 607       Replace.push_back(&I);
 608       Modified = true;
 609       break;
 610     }
 611     default:
 612       break;
 613     }
 614   }
 615
 616   if (Replace.empty())
 617     return false;
 618
 619   while (!Replace.empty()) {
 620     Instruction *I = Replace.pop_back_val();
 621     if (I->getOpcode() == Instruction::FPToUI ||
 622         I->getOpcode() == Instruction::FPToSI) {
 623       expandFPToI(I);
 624     } else {
 625       expandIToFP(I);
 626     }
 627   }
 628
 629   return Modified;
 630 }
 631
 632 namespace {
 633 class ExpandLargeFpConvertLegacyPass : public FunctionPass {
 634 public:
 635   static char ID;
 636
 637   ExpandLargeFpConvertLegacyPass() : FunctionPass(ID) {
 638     initializeExpandLargeFpConvertLegacyPassPass(
 639         *PassRegistry::getPassRegistry());
 640   }
 641
 642   bool runOnFunction(Function &F) override {
 643     auto *TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
 644     auto *TLI = TM->getSubtargetImpl(F)->getTargetLowering();
 645     return runImpl(F, *TLI);
 646   }
 647
 648   void getAnalysisUsage(AnalysisUsage &AU) const override {
 649     AU.addRequired<TargetPassConfig>();
 650     AU.addPreserved<AAResultsWrapperPass>();
 651     AU.addPreserved<GlobalsAAWrapperPass>();
 652   }
 653 };
 654 } // namespace
 655
 656 char ExpandLargeFpConvertLegacyPass::ID = 0;
 657 INITIALIZE_PASS_BEGIN(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
 658                       "Expand large fp convert", false, false)
 659 INITIALIZE_PASS_END(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
 660                     "Expand large fp convert", false, false)
 661
 662 FunctionPass *llvm::createExpandLargeFpConvertPass() {
 663   return new ExpandLargeFpConvertLegacyPass();
 664 }