xref: /llvm-project/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp (revision 735ab61ac828bd61398e6847d60e308fdf2b54ec) 
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/CodeGen/ExpandLargeFpConvert.h"
18 #include "llvm/ADT/SmallVector.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 =
119       PowerOf2Ceil(FloatVal->getType()->getScalarSizeInBits());
120   unsigned ExponentWidth = FloatWidth - FPMantissaWidth - 1;
121   unsigned ExponentBias = (1 << (ExponentWidth - 1)) - 1;
122   Value *ImplicitBit = Builder.CreateShl(
123       Builder.getIntN(BitWidth, 1), Builder.getIntN(BitWidth, FPMantissaWidth));
124   Value *SignificandMask =
125       Builder.CreateSub(ImplicitBit, Builder.getIntN(BitWidth, 1));
126   Value *NegOne = Builder.CreateSExt(
127       ConstantInt::getSigned(Builder.getInt32Ty(), -1), IntTy);
128   Value *NegInf =
129       Builder.CreateShl(ConstantInt::getSigned(IntTy, 1),
130                         ConstantInt::getSigned(IntTy, BitWidth - 1));
131 
132   BasicBlock *Entry = Builder.GetInsertBlock();
133   Function *F = Entry->getParent();
134   Entry->setName(Twine(Entry->getName(), "fp-to-i-entry"));
135   BasicBlock *End =
136       Entry->splitBasicBlock(Builder.GetInsertPoint(), "fp-to-i-cleanup");
137   BasicBlock *IfEnd =
138       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-end", F, End);
139   BasicBlock *IfThen5 =
140       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-then5", F, End);
141   BasicBlock *IfEnd9 =
142       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-end9", F, End);
143   BasicBlock *IfThen12 =
144       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-then12", F, End);
145   BasicBlock *IfElse =
146       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-else", F, End);
147 
148   Entry->getTerminator()->eraseFromParent();
149 
150   // entry:
151   Builder.SetInsertPoint(Entry);
152   Value *FloatVal0 = FloatVal;
153   // fp80 conversion is implemented by fpext to fp128 first then do the
154   // conversion.
155   if (FloatVal->getType()->isX86_FP80Ty())
156     FloatVal0 =
157         Builder.CreateFPExt(FloatVal, Type::getFP128Ty(Builder.getContext()));
158   Value *ARep0 =
159       Builder.CreateBitCast(FloatVal0, Builder.getIntNTy(FloatWidth));
160   Value *ARep = Builder.CreateZExt(ARep0, FPToI->getType());
161   Value *PosOrNeg = Builder.CreateICmpSGT(
162       ARep0, ConstantInt::getSigned(Builder.getIntNTy(FloatWidth), -1));
163   Value *Sign = Builder.CreateSelect(PosOrNeg, ConstantInt::getSigned(IntTy, 1),
164                                      ConstantInt::getSigned(IntTy, -1));
165   Value *And =
166       Builder.CreateLShr(ARep, Builder.getIntN(BitWidth, FPMantissaWidth));
167   Value *And2 = Builder.CreateAnd(
168       And, Builder.getIntN(BitWidth, (1 << ExponentWidth) - 1));
169   Value *Abs = Builder.CreateAnd(ARep, SignificandMask);
170   Value *Or = Builder.CreateOr(Abs, ImplicitBit);
171   Value *Cmp =
172       Builder.CreateICmpULT(And2, Builder.getIntN(BitWidth, ExponentBias));
173   Builder.CreateCondBr(Cmp, End, IfEnd);
174 
175   // if.end:
176   Builder.SetInsertPoint(IfEnd);
177   Value *Add1 = Builder.CreateAdd(
178       And2, ConstantInt::getSigned(
179                 IntTy, -static_cast<int64_t>(ExponentBias + BitWidth)));
180   Value *Cmp3 = Builder.CreateICmpULT(
181       Add1, ConstantInt::getSigned(IntTy, -static_cast<int64_t>(BitWidth)));
182   Builder.CreateCondBr(Cmp3, IfThen5, IfEnd9);
183 
184   // if.then5:
185   Builder.SetInsertPoint(IfThen5);
186   Value *PosInf = Builder.CreateXor(NegOne, NegInf);
187   Value *Cond8 = Builder.CreateSelect(PosOrNeg, PosInf, NegInf);
188   Builder.CreateBr(End);
189 
190   // if.end9:
191   Builder.SetInsertPoint(IfEnd9);
192   Value *Cmp10 = Builder.CreateICmpULT(
193       And2, Builder.getIntN(BitWidth, ExponentBias + FPMantissaWidth));
194   Builder.CreateCondBr(Cmp10, IfThen12, IfElse);
195 
196   // if.then12:
197   Builder.SetInsertPoint(IfThen12);
198   Value *Sub13 = Builder.CreateSub(
199       Builder.getIntN(BitWidth, ExponentBias + FPMantissaWidth), And2);
200   Value *Shr14 = Builder.CreateLShr(Or, Sub13);
201   Value *Mul = Builder.CreateMul(Shr14, Sign);
202   Builder.CreateBr(End);
203 
204   // if.else:
205   Builder.SetInsertPoint(IfElse);
206   Value *Sub15 = Builder.CreateAdd(
207       And2, ConstantInt::getSigned(
208                 IntTy, -static_cast<int64_t>(ExponentBias + FPMantissaWidth)));
209   Value *Shl = Builder.CreateShl(Or, Sub15);
210   Value *Mul16 = Builder.CreateMul(Shl, Sign);
211   Builder.CreateBr(End);
212 
213   // cleanup:
214   Builder.SetInsertPoint(End, End->begin());
215   PHINode *Retval0 = Builder.CreatePHI(FPToI->getType(), 4);
216 
217   Retval0->addIncoming(Cond8, IfThen5);
218   Retval0->addIncoming(Mul, IfThen12);
219   Retval0->addIncoming(Mul16, IfElse);
220   Retval0->addIncoming(Builder.getIntN(BitWidth, 0), Entry);
221 
222   FPToI->replaceAllUsesWith(Retval0);
223   FPToI->dropAllReferences();
224   FPToI->eraseFromParent();
225 }
226 
227 /// Generate code to convert a fp number to integer, replacing S(U)IToFP with
228 /// the generated code. This currently generates code similarly to compiler-rt's
229 /// implementations. This implementation has an implicit assumption that integer
230 /// width is larger than fp.
231 ///
232 /// An example IR generated from compiler-rt/floatdisf.c looks like below:
233 /// define dso_local float @__floatdisf(i64 noundef %a) local_unnamed_addr #0 {
234 /// entry:
235 ///   %cmp = icmp eq i64 %a, 0
236 ///   br i1 %cmp, label %return, label %if.end
237 ///
238 /// if.end:                                           ; preds = %entry
239 ///   %shr = ashr i64 %a, 63
240 ///   %xor = xor i64 %shr, %a
241 ///   %sub = sub nsw i64 %xor, %shr
242 ///   %0 = tail call i64 @llvm.ctlz.i64(i64 %sub, i1 true), !range !5
243 ///   %cast = trunc i64 %0 to i32
244 ///   %sub1 = sub nuw nsw i32 64, %cast
245 ///   %sub2 = xor i32 %cast, 63
246 ///   %cmp3 = icmp ult i32 %cast, 40
247 ///   br i1 %cmp3, label %if.then4, label %if.else
248 ///
249 /// if.then4:                                         ; preds = %if.end
250 ///   switch i32 %sub1, label %sw.default [
251 ///     i32 25, label %sw.bb
252 ///     i32 26, label %sw.epilog
253 ///   ]
254 ///
255 /// sw.bb:                                            ; preds = %if.then4
256 ///   %shl = shl i64 %sub, 1
257 ///   br label %sw.epilog
258 ///
259 /// sw.default:                                       ; preds = %if.then4
260 ///   %sub5 = sub nsw i64 38, %0
261 ///   %sh_prom = and i64 %sub5, 4294967295
262 ///   %shr6 = lshr i64 %sub, %sh_prom
263 ///   %shr9 = lshr i64 274877906943, %0
264 ///   %and = and i64 %shr9, %sub
265 ///   %cmp10 = icmp ne i64 %and, 0
266 ///   %conv11 = zext i1 %cmp10 to i64
267 ///   %or = or i64 %shr6, %conv11
268 ///   br label %sw.epilog
269 ///
270 /// sw.epilog:                                        ; preds = %sw.default, %if.then4, %sw.bb
271 ///   %a.addr.0 = phi i64 [ %or, %sw.default ], [ %sub, %if.then4 ], [ %shl, %sw.bb ]
272 ///   %1 = lshr i64 %a.addr.0, 2
273 ///   %2 = and i64 %1, 1
274 ///   %or16 = or i64 %2, %a.addr.0
275 ///   %inc = add nsw i64 %or16, 1
276 ///   %3 = and i64 %inc, 67108864
277 ///   %tobool.not = icmp eq i64 %3, 0
278 ///   %spec.select.v = select i1 %tobool.not, i64 2, i64 3
279 ///   %spec.select = ashr i64 %inc, %spec.select.v
280 ///   %spec.select56 = select i1 %tobool.not, i32 %sub2, i32 %sub1
281 ///   br label %if.end26
282 ///
283 /// if.else:                                          ; preds = %if.end
284 ///   %sub23 = add nuw nsw i64 %0, 4294967256
285 ///   %sh_prom24 = and i64 %sub23, 4294967295
286 ///   %shl25 = shl i64 %sub, %sh_prom24
287 ///   br label %if.end26
288 ///
289 /// if.end26:                                         ; preds = %sw.epilog, %if.else
290 ///   %a.addr.1 = phi i64 [ %shl25, %if.else ], [ %spec.select, %sw.epilog ]
291 ///   %e.0 = phi i32 [ %sub2, %if.else ], [ %spec.select56, %sw.epilog ]
292 ///   %conv27 = trunc i64 %shr to i32
293 ///   %and28 = and i32 %conv27, -2147483648
294 ///   %add = shl nuw nsw i32 %e.0, 23
295 ///   %shl29 = add nuw nsw i32 %add, 1065353216
296 ///   %conv31 = trunc i64 %a.addr.1 to i32
297 ///   %and32 = and i32 %conv31, 8388607
298 ///   %or30 = or i32 %and32, %and28
299 ///   %or33 = or i32 %or30, %shl29
300 ///   %4 = bitcast i32 %or33 to float
301 ///   br label %return
302 ///
303 /// return:                                           ; preds = %entry, %if.end26
304 ///   %retval.0 = phi float [ %4, %if.end26 ], [ 0.000000e+00, %entry ]
305 ///   ret float %retval.0
306 /// }
307 ///
308 /// Replace integer to fp with generated code.
309 static void expandIToFP(Instruction *IToFP) {
310   IRBuilder<> Builder(IToFP);
311   auto *IntVal = IToFP->getOperand(0);
312   IntegerType *IntTy = cast<IntegerType>(IntVal->getType());
313 
314   unsigned BitWidth = IntVal->getType()->getIntegerBitWidth();
315   unsigned FPMantissaWidth = IToFP->getType()->getFPMantissaWidth() - 1;
316   // fp80 conversion is implemented by conversion tp fp128 first following
317   // a fptrunc to fp80.
318   FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth;
319   // FIXME: As there is no related builtins added in compliler-rt,
320   // here currently utilized the fp32 <-> fp16 lib calls to implement.
321   FPMantissaWidth = FPMantissaWidth == 10 ? 23 : FPMantissaWidth;
322   FPMantissaWidth = FPMantissaWidth == 7 ? 23 : FPMantissaWidth;
323   unsigned FloatWidth = PowerOf2Ceil(FPMantissaWidth);
324   bool IsSigned = IToFP->getOpcode() == Instruction::SIToFP;
325 
326   assert(BitWidth > FloatWidth && "Unexpected conversion. expandIToFP() "
327                                   "assumes integer width is larger than fp.");
328 
329   Value *Temp1 =
330       Builder.CreateShl(Builder.getIntN(BitWidth, 1),
331                         Builder.getIntN(BitWidth, FPMantissaWidth + 3));
332 
333   BasicBlock *Entry = Builder.GetInsertBlock();
334   Function *F = Entry->getParent();
335   Entry->setName(Twine(Entry->getName(), "itofp-entry"));
336   BasicBlock *End =
337       Entry->splitBasicBlock(Builder.GetInsertPoint(), "itofp-return");
338   BasicBlock *IfEnd =
339       BasicBlock::Create(Builder.getContext(), "itofp-if-end", F, End);
340   BasicBlock *IfThen4 =
341       BasicBlock::Create(Builder.getContext(), "itofp-if-then4", F, End);
342   BasicBlock *SwBB =
343       BasicBlock::Create(Builder.getContext(), "itofp-sw-bb", F, End);
344   BasicBlock *SwDefault =
345       BasicBlock::Create(Builder.getContext(), "itofp-sw-default", F, End);
346   BasicBlock *SwEpilog =
347       BasicBlock::Create(Builder.getContext(), "itofp-sw-epilog", F, End);
348   BasicBlock *IfThen20 =
349       BasicBlock::Create(Builder.getContext(), "itofp-if-then20", F, End);
350   BasicBlock *IfElse =
351       BasicBlock::Create(Builder.getContext(), "itofp-if-else", F, End);
352   BasicBlock *IfEnd26 =
353       BasicBlock::Create(Builder.getContext(), "itofp-if-end26", F, End);
354 
355   Entry->getTerminator()->eraseFromParent();
356 
357   Function *CTLZ =
358       Intrinsic::getOrInsertDeclaration(F->getParent(), Intrinsic::ctlz, IntTy);
359   ConstantInt *True = Builder.getTrue();
360 
361   // entry:
362   Builder.SetInsertPoint(Entry);
363   Value *Cmp = Builder.CreateICmpEQ(IntVal, ConstantInt::getSigned(IntTy, 0));
364   Builder.CreateCondBr(Cmp, End, IfEnd);
365 
366   // if.end:
367   Builder.SetInsertPoint(IfEnd);
368   Value *Shr =
369       Builder.CreateAShr(IntVal, Builder.getIntN(BitWidth, BitWidth - 1));
370   Value *Xor = Builder.CreateXor(Shr, IntVal);
371   Value *Sub = Builder.CreateSub(Xor, Shr);
372   Value *Call = Builder.CreateCall(CTLZ, {IsSigned ? Sub : IntVal, True});
373   Value *Cast = Builder.CreateTrunc(Call, Builder.getInt32Ty());
374   int BitWidthNew = FloatWidth == 128 ? BitWidth : 32;
375   Value *Sub1 = Builder.CreateSub(Builder.getIntN(BitWidthNew, BitWidth),
376                                   FloatWidth == 128 ? Call : Cast);
377   Value *Sub2 = Builder.CreateSub(Builder.getIntN(BitWidthNew, BitWidth - 1),
378                                   FloatWidth == 128 ? Call : Cast);
379   Value *Cmp3 = Builder.CreateICmpSGT(
380       Sub1, Builder.getIntN(BitWidthNew, FPMantissaWidth + 1));
381   Builder.CreateCondBr(Cmp3, IfThen4, IfElse);
382 
383   // if.then4:
384   Builder.SetInsertPoint(IfThen4);
385   llvm::SwitchInst *SI = Builder.CreateSwitch(Sub1, SwDefault);
386   SI->addCase(Builder.getIntN(BitWidthNew, FPMantissaWidth + 2), SwBB);
387   SI->addCase(Builder.getIntN(BitWidthNew, FPMantissaWidth + 3), SwEpilog);
388 
389   // sw.bb:
390   Builder.SetInsertPoint(SwBB);
391   Value *Shl =
392       Builder.CreateShl(IsSigned ? Sub : IntVal, Builder.getIntN(BitWidth, 1));
393   Builder.CreateBr(SwEpilog);
394 
395   // sw.default:
396   Builder.SetInsertPoint(SwDefault);
397   Value *Sub5 = Builder.CreateSub(
398       Builder.getIntN(BitWidthNew, BitWidth - FPMantissaWidth - 3),
399       FloatWidth == 128 ? Call : Cast);
400   Value *ShProm = Builder.CreateZExt(Sub5, IntTy);
401   Value *Shr6 = Builder.CreateLShr(IsSigned ? Sub : IntVal,
402                                    FloatWidth == 128 ? Sub5 : ShProm);
403   Value *Sub8 =
404       Builder.CreateAdd(FloatWidth == 128 ? Call : Cast,
405                         Builder.getIntN(BitWidthNew, FPMantissaWidth + 3));
406   Value *ShProm9 = Builder.CreateZExt(Sub8, IntTy);
407   Value *Shr9 = Builder.CreateLShr(ConstantInt::getSigned(IntTy, -1),
408                                    FloatWidth == 128 ? Sub8 : ShProm9);
409   Value *And = Builder.CreateAnd(Shr9, IsSigned ? Sub : IntVal);
410   Value *Cmp10 = Builder.CreateICmpNE(And, Builder.getIntN(BitWidth, 0));
411   Value *Conv11 = Builder.CreateZExt(Cmp10, IntTy);
412   Value *Or = Builder.CreateOr(Shr6, Conv11);
413   Builder.CreateBr(SwEpilog);
414 
415   // sw.epilog:
416   Builder.SetInsertPoint(SwEpilog);
417   PHINode *AAddr0 = Builder.CreatePHI(IntTy, 3);
418   AAddr0->addIncoming(Or, SwDefault);
419   AAddr0->addIncoming(IsSigned ? Sub : IntVal, IfThen4);
420   AAddr0->addIncoming(Shl, SwBB);
421   Value *A0 = Builder.CreateTrunc(AAddr0, Builder.getInt32Ty());
422   Value *A1 = Builder.CreateLShr(A0, Builder.getIntN(32, 2));
423   Value *A2 = Builder.CreateAnd(A1, Builder.getIntN(32, 1));
424   Value *Conv16 = Builder.CreateZExt(A2, IntTy);
425   Value *Or17 = Builder.CreateOr(AAddr0, Conv16);
426   Value *Inc = Builder.CreateAdd(Or17, Builder.getIntN(BitWidth, 1));
427   Value *Shr18 = nullptr;
428   if (IsSigned)
429     Shr18 = Builder.CreateAShr(Inc, Builder.getIntN(BitWidth, 2));
430   else
431     Shr18 = Builder.CreateLShr(Inc, Builder.getIntN(BitWidth, 2));
432   Value *A3 = Builder.CreateAnd(Inc, Temp1, "a3");
433   Value *PosOrNeg = Builder.CreateICmpEQ(A3, Builder.getIntN(BitWidth, 0));
434   Value *ExtractT60 = Builder.CreateTrunc(Shr18, Builder.getIntNTy(FloatWidth));
435   Value *Extract63 = Builder.CreateLShr(Shr18, Builder.getIntN(BitWidth, 32));
436   Value *ExtractT64 = nullptr;
437   if (FloatWidth > 80)
438     ExtractT64 = Builder.CreateTrunc(Sub2, Builder.getInt64Ty());
439   else
440     ExtractT64 = Builder.CreateTrunc(Extract63, Builder.getInt32Ty());
441   Builder.CreateCondBr(PosOrNeg, IfEnd26, IfThen20);
442 
443   // if.then20
444   Builder.SetInsertPoint(IfThen20);
445   Value *Shr21 = nullptr;
446   if (IsSigned)
447     Shr21 = Builder.CreateAShr(Inc, Builder.getIntN(BitWidth, 3));
448   else
449     Shr21 = Builder.CreateLShr(Inc, Builder.getIntN(BitWidth, 3));
450   Value *ExtractT = Builder.CreateTrunc(Shr21, Builder.getIntNTy(FloatWidth));
451   Value *Extract = Builder.CreateLShr(Shr21, Builder.getIntN(BitWidth, 32));
452   Value *ExtractT62 = nullptr;
453   if (FloatWidth > 80)
454     ExtractT62 = Builder.CreateTrunc(Sub1, Builder.getIntNTy(64));
455   else
456     ExtractT62 = Builder.CreateTrunc(Extract, Builder.getIntNTy(32));
457   Builder.CreateBr(IfEnd26);
458 
459   // if.else:
460   Builder.SetInsertPoint(IfElse);
461   Value *Sub24 = Builder.CreateAdd(
462       FloatWidth == 128 ? Call : Cast,
463       ConstantInt::getSigned(Builder.getIntNTy(BitWidthNew),
464                              -(BitWidth - FPMantissaWidth - 1)));
465   Value *ShProm25 = Builder.CreateZExt(Sub24, IntTy);
466   Value *Shl26 = Builder.CreateShl(IsSigned ? Sub : IntVal,
467                                    FloatWidth == 128 ? Sub24 : ShProm25);
468   Value *ExtractT61 = Builder.CreateTrunc(Shl26, Builder.getIntNTy(FloatWidth));
469   Value *Extract65 = Builder.CreateLShr(Shl26, Builder.getIntN(BitWidth, 32));
470   Value *ExtractT66 = nullptr;
471   if (FloatWidth > 80)
472     ExtractT66 = Builder.CreateTrunc(Sub2, Builder.getIntNTy(64));
473   else
474     ExtractT66 = Builder.CreateTrunc(Extract65, Builder.getInt32Ty());
475   Builder.CreateBr(IfEnd26);
476 
477   // if.end26:
478   Builder.SetInsertPoint(IfEnd26);
479   PHINode *AAddr1Off0 = Builder.CreatePHI(Builder.getIntNTy(FloatWidth), 3);
480   AAddr1Off0->addIncoming(ExtractT, IfThen20);
481   AAddr1Off0->addIncoming(ExtractT60, SwEpilog);
482   AAddr1Off0->addIncoming(ExtractT61, IfElse);
483   PHINode *AAddr1Off32 = nullptr;
484   if (FloatWidth > 32) {
485     AAddr1Off32 =
486         Builder.CreatePHI(Builder.getIntNTy(FloatWidth > 80 ? 64 : 32), 3);
487     AAddr1Off32->addIncoming(ExtractT62, IfThen20);
488     AAddr1Off32->addIncoming(ExtractT64, SwEpilog);
489     AAddr1Off32->addIncoming(ExtractT66, IfElse);
490   }
491   PHINode *E0 = nullptr;
492   if (FloatWidth <= 80) {
493     E0 = Builder.CreatePHI(Builder.getIntNTy(BitWidthNew), 3);
494     E0->addIncoming(Sub1, IfThen20);
495     E0->addIncoming(Sub2, SwEpilog);
496     E0->addIncoming(Sub2, IfElse);
497   }
498   Value *And29 = nullptr;
499   if (FloatWidth > 80) {
500     Value *Temp2 = Builder.CreateShl(Builder.getIntN(BitWidth, 1),
501                                      Builder.getIntN(BitWidth, 63));
502     And29 = Builder.CreateAnd(Shr, Temp2, "and29");
503   } else {
504     Value *Conv28 = Builder.CreateTrunc(Shr, Builder.getIntNTy(32));
505     And29 = Builder.CreateAnd(
506         Conv28, ConstantInt::getSigned(Builder.getIntNTy(32), 0x80000000));
507   }
508   unsigned TempMod = FPMantissaWidth % 32;
509   Value *And34 = nullptr;
510   Value *Shl30 = nullptr;
511   if (FloatWidth > 80) {
512     TempMod += 32;
513     Value *Add = Builder.CreateShl(AAddr1Off32, Builder.getIntN(64, TempMod));
514     Shl30 = Builder.CreateAdd(
515         Add,
516         Builder.getIntN(64, ((1ull << (62ull - TempMod)) - 1ull) << TempMod));
517     And34 = Builder.CreateZExt(Shl30, Builder.getIntNTy(128));
518   } else {
519     Value *Add = Builder.CreateShl(E0, Builder.getIntN(32, TempMod));
520     Shl30 = Builder.CreateAdd(
521         Add, Builder.getIntN(32, ((1 << (30 - TempMod)) - 1) << TempMod));
522     And34 = Builder.CreateAnd(FloatWidth > 32 ? AAddr1Off32 : AAddr1Off0,
523                               Builder.getIntN(32, (1 << TempMod) - 1));
524   }
525   Value *Or35 = nullptr;
526   if (FloatWidth > 80) {
527     Value *And29Trunc = Builder.CreateTrunc(And29, Builder.getIntNTy(128));
528     Value *Or31 = Builder.CreateOr(And29Trunc, And34);
529     Value *Or34 = Builder.CreateShl(Or31, Builder.getIntN(128, 64));
530     Value *Temp3 = Builder.CreateShl(Builder.getIntN(128, 1),
531                                      Builder.getIntN(128, FPMantissaWidth));
532     Value *Temp4 = Builder.CreateSub(Temp3, Builder.getIntN(128, 1));
533     Value *A6 = Builder.CreateAnd(AAddr1Off0, Temp4);
534     Or35 = Builder.CreateOr(Or34, A6);
535   } else {
536     Value *Or31 = Builder.CreateOr(And34, And29);
537     Or35 = Builder.CreateOr(IsSigned ? Or31 : And34, Shl30);
538   }
539   Value *A4 = nullptr;
540   if (IToFP->getType()->isDoubleTy()) {
541     Value *ZExt1 = Builder.CreateZExt(Or35, Builder.getIntNTy(FloatWidth));
542     Value *Shl1 = Builder.CreateShl(ZExt1, Builder.getIntN(FloatWidth, 32));
543     Value *And1 =
544         Builder.CreateAnd(AAddr1Off0, Builder.getIntN(FloatWidth, 0xFFFFFFFF));
545     Value *Or1 = Builder.CreateOr(Shl1, And1);
546     A4 = Builder.CreateBitCast(Or1, IToFP->getType());
547   } else if (IToFP->getType()->isX86_FP80Ty()) {
548     Value *A40 =
549         Builder.CreateBitCast(Or35, Type::getFP128Ty(Builder.getContext()));
550     A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
551   } else if (IToFP->getType()->isHalfTy() || IToFP->getType()->isBFloatTy()) {
552     // Deal with "half" situation. This is a workaround since we don't have
553     // floattihf.c currently as referring.
554     Value *A40 =
555         Builder.CreateBitCast(Or35, Type::getFloatTy(Builder.getContext()));
556     A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
557   } else // float type
558     A4 = Builder.CreateBitCast(Or35, IToFP->getType());
559   Builder.CreateBr(End);
560 
561   // return:
562   Builder.SetInsertPoint(End, End->begin());
563   PHINode *Retval0 = Builder.CreatePHI(IToFP->getType(), 2);
564   Retval0->addIncoming(A4, IfEnd26);
565   Retval0->addIncoming(ConstantFP::getZero(IToFP->getType(), false), Entry);
566 
567   IToFP->replaceAllUsesWith(Retval0);
568   IToFP->dropAllReferences();
569   IToFP->eraseFromParent();
570 }
571 
572 static void scalarize(Instruction *I, SmallVectorImpl<Instruction *> &Replace) {
573   VectorType *VTy = cast<FixedVectorType>(I->getType());
574 
575   IRBuilder<> Builder(I);
576 
577   unsigned NumElements = VTy->getElementCount().getFixedValue();
578   Value *Result = PoisonValue::get(VTy);
579   for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
580     Value *Ext = Builder.CreateExtractElement(I->getOperand(0), Idx);
581     Value *Cast = Builder.CreateCast(cast<CastInst>(I)->getOpcode(), Ext,
582                                      I->getType()->getScalarType());
583     Result = Builder.CreateInsertElement(Result, Cast, Idx);
584     if (isa<Instruction>(Cast))
585       Replace.push_back(cast<Instruction>(Cast));
586   }
587   I->replaceAllUsesWith(Result);
588   I->dropAllReferences();
589   I->eraseFromParent();
590 }
591 
592 static bool runImpl(Function &F, const TargetLowering &TLI) {
593   SmallVector<Instruction *, 4> Replace;
594   SmallVector<Instruction *, 4> ReplaceVector;
595   bool Modified = false;
596 
597   unsigned MaxLegalFpConvertBitWidth =
598       TLI.getMaxLargeFPConvertBitWidthSupported();
599   if (ExpandFpConvertBits != llvm::IntegerType::MAX_INT_BITS)
600     MaxLegalFpConvertBitWidth = ExpandFpConvertBits;
601 
602   if (MaxLegalFpConvertBitWidth >= llvm::IntegerType::MAX_INT_BITS)
603     return false;
604 
605   for (auto &I : instructions(F)) {
606     switch (I.getOpcode()) {
607     case Instruction::FPToUI:
608     case Instruction::FPToSI: {
609       // TODO: This pass doesn't handle scalable vectors.
610       if (I.getOperand(0)->getType()->isScalableTy())
611         continue;
612 
613       auto *IntTy = cast<IntegerType>(I.getType()->getScalarType());
614       if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth)
615         continue;
616 
617       if (I.getOperand(0)->getType()->isVectorTy())
618         ReplaceVector.push_back(&I);
619       else
620         Replace.push_back(&I);
621       Modified = true;
622       break;
623     }
624     case Instruction::UIToFP:
625     case Instruction::SIToFP: {
626       // TODO: This pass doesn't handle scalable vectors.
627       if (I.getOperand(0)->getType()->isScalableTy())
628         continue;
629 
630       auto *IntTy =
631           cast<IntegerType>(I.getOperand(0)->getType()->getScalarType());
632       if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth)
633         continue;
634 
635       if (I.getOperand(0)->getType()->isVectorTy())
636         ReplaceVector.push_back(&I);
637       else
638         Replace.push_back(&I);
639       Modified = true;
640       break;
641     }
642     default:
643       break;
644     }
645   }
646 
647   while (!ReplaceVector.empty()) {
648     Instruction *I = ReplaceVector.pop_back_val();
649     scalarize(I, Replace);
650   }
651 
652   if (Replace.empty())
653     return false;
654 
655   while (!Replace.empty()) {
656     Instruction *I = Replace.pop_back_val();
657     if (I->getOpcode() == Instruction::FPToUI ||
658         I->getOpcode() == Instruction::FPToSI) {
659       expandFPToI(I);
660     } else {
661       expandIToFP(I);
662     }
663   }
664 
665   return Modified;
666 }
667 
668 namespace {
669 class ExpandLargeFpConvertLegacyPass : public FunctionPass {
670 public:
671   static char ID;
672 
673   ExpandLargeFpConvertLegacyPass() : FunctionPass(ID) {
674     initializeExpandLargeFpConvertLegacyPassPass(
675         *PassRegistry::getPassRegistry());
676   }
677 
678   bool runOnFunction(Function &F) override {
679     auto *TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
680     auto *TLI = TM->getSubtargetImpl(F)->getTargetLowering();
681     return runImpl(F, *TLI);
682   }
683 
684   void getAnalysisUsage(AnalysisUsage &AU) const override {
685     AU.addRequired<TargetPassConfig>();
686     AU.addPreserved<AAResultsWrapperPass>();
687     AU.addPreserved<GlobalsAAWrapperPass>();
688   }
689 };
690 } // namespace
691 
692 PreservedAnalyses ExpandLargeFpConvertPass::run(Function &F,
693                                                 FunctionAnalysisManager &FAM) {
694   const TargetSubtargetInfo *STI = TM->getSubtargetImpl(F);
695   return runImpl(F, *STI->getTargetLowering()) ? PreservedAnalyses::none()
696                                                : PreservedAnalyses::all();
697 }
698 
699 char ExpandLargeFpConvertLegacyPass::ID = 0;
700 INITIALIZE_PASS_BEGIN(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
701                       "Expand large fp convert", false, false)
702 INITIALIZE_PASS_END(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
703                     "Expand large fp convert", false, false)
704 
705 FunctionPass *llvm::createExpandLargeFpConvertPass() {
706   return new ExpandLargeFpConvertLegacyPass();
707 }
708