xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/ExpandLargeFpConvert.cpp (revision b1879975794772ee51f0b4865753364c7d7626c3) 
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/ADT/StringExtras.h"
20 #include "llvm/Analysis/GlobalsModRef.h"
21 #include "llvm/CodeGen/Passes.h"
22 #include "llvm/CodeGen/TargetLowering.h"
23 #include "llvm/CodeGen/TargetPassConfig.h"
24 #include "llvm/CodeGen/TargetSubtargetInfo.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/InstIterator.h"
27 #include "llvm/IR/PassManager.h"
28 #include "llvm/InitializePasses.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Target/TargetMachine.h"
32 
33 using namespace llvm;
34 
35 static cl::opt<unsigned>
36     ExpandFpConvertBits("expand-fp-convert-bits", cl::Hidden,
37                      cl::init(llvm::IntegerType::MAX_INT_BITS),
38                      cl::desc("fp convert instructions on integers with "
39                               "more than <N> bits are expanded."));
40 
41 /// Generate code to convert a fp number to integer, replacing FPToS(U)I with
42 /// the generated code. This currently generates code similarly to compiler-rt's
43 /// implementations.
44 ///
45 /// An example IR generated from compiler-rt/fixsfdi.c looks like below:
46 /// define dso_local i64 @foo(float noundef %a) local_unnamed_addr #0 {
47 /// entry:
48 ///   %0 = bitcast float %a to i32
49 ///   %conv.i = zext i32 %0 to i64
50 ///   %tobool.not = icmp sgt i32 %0, -1
51 ///   %conv = select i1 %tobool.not, i64 1, i64 -1
52 ///   %and = lshr i64 %conv.i, 23
53 ///   %shr = and i64 %and, 255
54 ///   %and2 = and i64 %conv.i, 8388607
55 ///   %or = or i64 %and2, 8388608
56 ///   %cmp = icmp ult i64 %shr, 127
57 ///   br i1 %cmp, label %cleanup, label %if.end
58 ///
59 /// if.end:                                           ; preds = %entry
60 ///   %sub = add nuw nsw i64 %shr, 4294967169
61 ///   %conv5 = and i64 %sub, 4294967232
62 ///   %cmp6.not = icmp eq i64 %conv5, 0
63 ///   br i1 %cmp6.not, label %if.end12, label %if.then8
64 ///
65 /// if.then8:                                         ; preds = %if.end
66 ///   %cond11 = select i1 %tobool.not, i64 9223372036854775807, i64 -9223372036854775808
67 ///   br label %cleanup
68 ///
69 /// if.end12:                                         ; preds = %if.end
70 ///   %cmp13 = icmp ult i64 %shr, 150
71 ///   br i1 %cmp13, label %if.then15, label %if.else
72 ///
73 /// if.then15:                                        ; preds = %if.end12
74 ///   %sub16 = sub nuw nsw i64 150, %shr
75 ///   %shr17 = lshr i64 %or, %sub16
76 ///   %mul = mul nsw i64 %shr17, %conv
77 ///   br label %cleanup
78 ///
79 /// if.else:                                          ; preds = %if.end12
80 ///   %sub18 = add nsw i64 %shr, -150
81 ///   %shl = shl i64 %or, %sub18
82 ///   %mul19 = mul nsw i64 %shl, %conv
83 ///   br label %cleanup
84 ///
85 /// cleanup:                                          ; preds = %entry, %if.else, %if.then15, %if.then8
86 ///   %retval.0 = phi i64 [ %cond11, %if.then8 ], [ %mul, %if.then15 ], [ %mul19, %if.else ], [ 0, %entry ]
87 ///   ret i64 %retval.0
88 /// }
89 ///
90 /// Replace fp to integer with generated code.
91 static void expandFPToI(Instruction *FPToI) {
92   IRBuilder<> Builder(FPToI);
93   auto *FloatVal = FPToI->getOperand(0);
94   IntegerType *IntTy = cast<IntegerType>(FPToI->getType());
95 
96   unsigned BitWidth = FPToI->getType()->getIntegerBitWidth();
97   unsigned FPMantissaWidth = FloatVal->getType()->getFPMantissaWidth() - 1;
98 
99   // FIXME: fp16's range is covered by i32. So `fptoi half` can convert
100   // to i32 first following a sext/zext to target integer type.
101   Value *A1 = nullptr;
102   if (FloatVal->getType()->isHalfTy()) {
103     if (FPToI->getOpcode() == Instruction::FPToUI) {
104       Value *A0 = Builder.CreateFPToUI(FloatVal, Builder.getIntNTy(32));
105       A1 = Builder.CreateZExt(A0, IntTy);
106     } else { // FPToSI
107       Value *A0 = Builder.CreateFPToSI(FloatVal, Builder.getIntNTy(32));
108       A1 = Builder.CreateSExt(A0, IntTy);
109     }
110     FPToI->replaceAllUsesWith(A1);
111     FPToI->dropAllReferences();
112     FPToI->eraseFromParent();
113     return;
114   }
115 
116   // fp80 conversion is implemented by fpext to fp128 first then do the
117   // conversion.
118   FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth;
119   unsigned FloatWidth =
120       PowerOf2Ceil(FloatVal->getType()->getScalarSizeInBits());
121   unsigned ExponentWidth = FloatWidth - FPMantissaWidth - 1;
122   unsigned ExponentBias = (1 << (ExponentWidth - 1)) - 1;
123   Value *ImplicitBit = Builder.CreateShl(
124       Builder.getIntN(BitWidth, 1), Builder.getIntN(BitWidth, FPMantissaWidth));
125   Value *SignificandMask =
126       Builder.CreateSub(ImplicitBit, Builder.getIntN(BitWidth, 1));
127   Value *NegOne = Builder.CreateSExt(
128       ConstantInt::getSigned(Builder.getInt32Ty(), -1), IntTy);
129   Value *NegInf =
130       Builder.CreateShl(ConstantInt::getSigned(IntTy, 1),
131                         ConstantInt::getSigned(IntTy, BitWidth - 1));
132 
133   BasicBlock *Entry = Builder.GetInsertBlock();
134   Function *F = Entry->getParent();
135   Entry->setName(Twine(Entry->getName(), "fp-to-i-entry"));
136   BasicBlock *End =
137       Entry->splitBasicBlock(Builder.GetInsertPoint(), "fp-to-i-cleanup");
138   BasicBlock *IfEnd =
139       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-end", F, End);
140   BasicBlock *IfThen5 =
141       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-then5", F, End);
142   BasicBlock *IfEnd9 =
143       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-end9", F, End);
144   BasicBlock *IfThen12 =
145       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-then12", F, End);
146   BasicBlock *IfElse =
147       BasicBlock::Create(Builder.getContext(), "fp-to-i-if-else", F, End);
148 
149   Entry->getTerminator()->eraseFromParent();
150 
151   // entry:
152   Builder.SetInsertPoint(Entry);
153   Value *FloatVal0 = FloatVal;
154   // fp80 conversion is implemented by fpext to fp128 first then do the
155   // conversion.
156   if (FloatVal->getType()->isX86_FP80Ty())
157     FloatVal0 =
158         Builder.CreateFPExt(FloatVal, Type::getFP128Ty(Builder.getContext()));
159   Value *ARep0 =
160       Builder.CreateBitCast(FloatVal0, Builder.getIntNTy(FloatWidth));
161   Value *ARep = Builder.CreateZExt(ARep0, FPToI->getType());
162   Value *PosOrNeg = Builder.CreateICmpSGT(
163       ARep0, ConstantInt::getSigned(Builder.getIntNTy(FloatWidth), -1));
164   Value *Sign = Builder.CreateSelect(PosOrNeg, ConstantInt::getSigned(IntTy, 1),
165                                      ConstantInt::getSigned(IntTy, -1));
166   Value *And =
167       Builder.CreateLShr(ARep, Builder.getIntN(BitWidth, FPMantissaWidth));
168   Value *And2 = Builder.CreateAnd(
169       And, Builder.getIntN(BitWidth, (1 << ExponentWidth) - 1));
170   Value *Abs = Builder.CreateAnd(ARep, SignificandMask);
171   Value *Or = Builder.CreateOr(Abs, ImplicitBit);
172   Value *Cmp =
173       Builder.CreateICmpULT(And2, Builder.getIntN(BitWidth, ExponentBias));
174   Builder.CreateCondBr(Cmp, End, IfEnd);
175 
176   // if.end:
177   Builder.SetInsertPoint(IfEnd);
178   Value *Add1 = Builder.CreateAdd(
179       And2, ConstantInt::getSigned(
180                 IntTy, -static_cast<int64_t>(ExponentBias + BitWidth)));
181   Value *Cmp3 = Builder.CreateICmpULT(
182       Add1, ConstantInt::getSigned(IntTy, -static_cast<int64_t>(BitWidth)));
183   Builder.CreateCondBr(Cmp3, IfThen5, IfEnd9);
184 
185   // if.then5:
186   Builder.SetInsertPoint(IfThen5);
187   Value *PosInf = Builder.CreateXor(NegOne, NegInf);
188   Value *Cond8 = Builder.CreateSelect(PosOrNeg, PosInf, NegInf);
189   Builder.CreateBr(End);
190 
191   // if.end9:
192   Builder.SetInsertPoint(IfEnd9);
193   Value *Cmp10 = Builder.CreateICmpULT(
194       And2, Builder.getIntN(BitWidth, ExponentBias + FPMantissaWidth));
195   Builder.CreateCondBr(Cmp10, IfThen12, IfElse);
196 
197   // if.then12:
198   Builder.SetInsertPoint(IfThen12);
199   Value *Sub13 = Builder.CreateSub(
200       Builder.getIntN(BitWidth, ExponentBias + FPMantissaWidth), And2);
201   Value *Shr14 = Builder.CreateLShr(Or, Sub13);
202   Value *Mul = Builder.CreateMul(Shr14, Sign);
203   Builder.CreateBr(End);
204 
205   // if.else:
206   Builder.SetInsertPoint(IfElse);
207   Value *Sub15 = Builder.CreateAdd(
208       And2, ConstantInt::getSigned(
209                 IntTy, -static_cast<int64_t>(ExponentBias + FPMantissaWidth)));
210   Value *Shl = Builder.CreateShl(Or, Sub15);
211   Value *Mul16 = Builder.CreateMul(Shl, Sign);
212   Builder.CreateBr(End);
213 
214   // cleanup:
215   Builder.SetInsertPoint(End, End->begin());
216   PHINode *Retval0 = Builder.CreatePHI(FPToI->getType(), 4);
217 
218   Retval0->addIncoming(Cond8, IfThen5);
219   Retval0->addIncoming(Mul, IfThen12);
220   Retval0->addIncoming(Mul16, IfElse);
221   Retval0->addIncoming(Builder.getIntN(BitWidth, 0), Entry);
222 
223   FPToI->replaceAllUsesWith(Retval0);
224   FPToI->dropAllReferences();
225   FPToI->eraseFromParent();
226 }
227 
228 /// Generate code to convert a fp number to integer, replacing S(U)IToFP with
229 /// the generated code. This currently generates code similarly to compiler-rt's
230 /// implementations. This implementation has an implicit assumption that integer
231 /// width is larger than fp.
232 ///
233 /// An example IR generated from compiler-rt/floatdisf.c looks like below:
234 /// define dso_local float @__floatdisf(i64 noundef %a) local_unnamed_addr #0 {
235 /// entry:
236 ///   %cmp = icmp eq i64 %a, 0
237 ///   br i1 %cmp, label %return, label %if.end
238 ///
239 /// if.end:                                           ; preds = %entry
240 ///   %shr = ashr i64 %a, 63
241 ///   %xor = xor i64 %shr, %a
242 ///   %sub = sub nsw i64 %xor, %shr
243 ///   %0 = tail call i64 @llvm.ctlz.i64(i64 %sub, i1 true), !range !5
244 ///   %cast = trunc i64 %0 to i32
245 ///   %sub1 = sub nuw nsw i32 64, %cast
246 ///   %sub2 = xor i32 %cast, 63
247 ///   %cmp3 = icmp ult i32 %cast, 40
248 ///   br i1 %cmp3, label %if.then4, label %if.else
249 ///
250 /// if.then4:                                         ; preds = %if.end
251 ///   switch i32 %sub1, label %sw.default [
252 ///     i32 25, label %sw.bb
253 ///     i32 26, label %sw.epilog
254 ///   ]
255 ///
256 /// sw.bb:                                            ; preds = %if.then4
257 ///   %shl = shl i64 %sub, 1
258 ///   br label %sw.epilog
259 ///
260 /// sw.default:                                       ; preds = %if.then4
261 ///   %sub5 = sub nsw i64 38, %0
262 ///   %sh_prom = and i64 %sub5, 4294967295
263 ///   %shr6 = lshr i64 %sub, %sh_prom
264 ///   %shr9 = lshr i64 274877906943, %0
265 ///   %and = and i64 %shr9, %sub
266 ///   %cmp10 = icmp ne i64 %and, 0
267 ///   %conv11 = zext i1 %cmp10 to i64
268 ///   %or = or i64 %shr6, %conv11
269 ///   br label %sw.epilog
270 ///
271 /// sw.epilog:                                        ; preds = %sw.default, %if.then4, %sw.bb
272 ///   %a.addr.0 = phi i64 [ %or, %sw.default ], [ %sub, %if.then4 ], [ %shl, %sw.bb ]
273 ///   %1 = lshr i64 %a.addr.0, 2
274 ///   %2 = and i64 %1, 1
275 ///   %or16 = or i64 %2, %a.addr.0
276 ///   %inc = add nsw i64 %or16, 1
277 ///   %3 = and i64 %inc, 67108864
278 ///   %tobool.not = icmp eq i64 %3, 0
279 ///   %spec.select.v = select i1 %tobool.not, i64 2, i64 3
280 ///   %spec.select = ashr i64 %inc, %spec.select.v
281 ///   %spec.select56 = select i1 %tobool.not, i32 %sub2, i32 %sub1
282 ///   br label %if.end26
283 ///
284 /// if.else:                                          ; preds = %if.end
285 ///   %sub23 = add nuw nsw i64 %0, 4294967256
286 ///   %sh_prom24 = and i64 %sub23, 4294967295
287 ///   %shl25 = shl i64 %sub, %sh_prom24
288 ///   br label %if.end26
289 ///
290 /// if.end26:                                         ; preds = %sw.epilog, %if.else
291 ///   %a.addr.1 = phi i64 [ %shl25, %if.else ], [ %spec.select, %sw.epilog ]
292 ///   %e.0 = phi i32 [ %sub2, %if.else ], [ %spec.select56, %sw.epilog ]
293 ///   %conv27 = trunc i64 %shr to i32
294 ///   %and28 = and i32 %conv27, -2147483648
295 ///   %add = shl nuw nsw i32 %e.0, 23
296 ///   %shl29 = add nuw nsw i32 %add, 1065353216
297 ///   %conv31 = trunc i64 %a.addr.1 to i32
298 ///   %and32 = and i32 %conv31, 8388607
299 ///   %or30 = or i32 %and32, %and28
300 ///   %or33 = or i32 %or30, %shl29
301 ///   %4 = bitcast i32 %or33 to float
302 ///   br label %return
303 ///
304 /// return:                                           ; preds = %entry, %if.end26
305 ///   %retval.0 = phi float [ %4, %if.end26 ], [ 0.000000e+00, %entry ]
306 ///   ret float %retval.0
307 /// }
308 ///
309 /// Replace integer to fp with generated code.
310 static void expandIToFP(Instruction *IToFP) {
311   IRBuilder<> Builder(IToFP);
312   auto *IntVal = IToFP->getOperand(0);
313   IntegerType *IntTy = cast<IntegerType>(IntVal->getType());
314 
315   unsigned BitWidth = IntVal->getType()->getIntegerBitWidth();
316   unsigned FPMantissaWidth = IToFP->getType()->getFPMantissaWidth() - 1;
317   // fp80 conversion is implemented by conversion tp fp128 first following
318   // a fptrunc to fp80.
319   FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth;
320   // FIXME: As there is no related builtins added in compliler-rt,
321   // here currently utilized the fp32 <-> fp16 lib calls to implement.
322   FPMantissaWidth = FPMantissaWidth == 10 ? 23 : FPMantissaWidth;
323   FPMantissaWidth = FPMantissaWidth == 7 ? 23 : FPMantissaWidth;
324   unsigned FloatWidth = PowerOf2Ceil(FPMantissaWidth);
325   bool IsSigned = IToFP->getOpcode() == Instruction::SIToFP;
326 
327   assert(BitWidth > FloatWidth && "Unexpected conversion. expandIToFP() "
328                                   "assumes integer width is larger than fp.");
329 
330   Value *Temp1 =
331       Builder.CreateShl(Builder.getIntN(BitWidth, 1),
332                         Builder.getIntN(BitWidth, FPMantissaWidth + 3));
333 
334   BasicBlock *Entry = Builder.GetInsertBlock();
335   Function *F = Entry->getParent();
336   Entry->setName(Twine(Entry->getName(), "itofp-entry"));
337   BasicBlock *End =
338       Entry->splitBasicBlock(Builder.GetInsertPoint(), "itofp-return");
339   BasicBlock *IfEnd =
340       BasicBlock::Create(Builder.getContext(), "itofp-if-end", F, End);
341   BasicBlock *IfThen4 =
342       BasicBlock::Create(Builder.getContext(), "itofp-if-then4", F, End);
343   BasicBlock *SwBB =
344       BasicBlock::Create(Builder.getContext(), "itofp-sw-bb", F, End);
345   BasicBlock *SwDefault =
346       BasicBlock::Create(Builder.getContext(), "itofp-sw-default", F, End);
347   BasicBlock *SwEpilog =
348       BasicBlock::Create(Builder.getContext(), "itofp-sw-epilog", F, End);
349   BasicBlock *IfThen20 =
350       BasicBlock::Create(Builder.getContext(), "itofp-if-then20", F, End);
351   BasicBlock *IfElse =
352       BasicBlock::Create(Builder.getContext(), "itofp-if-else", F, End);
353   BasicBlock *IfEnd26 =
354       BasicBlock::Create(Builder.getContext(), "itofp-if-end26", F, End);
355 
356   Entry->getTerminator()->eraseFromParent();
357 
358   Function *CTLZ =
359       Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz, IntTy);
360   ConstantInt *True = Builder.getTrue();
361 
362   // entry:
363   Builder.SetInsertPoint(Entry);
364   Value *Cmp = Builder.CreateICmpEQ(IntVal, ConstantInt::getSigned(IntTy, 0));
365   Builder.CreateCondBr(Cmp, End, IfEnd);
366 
367   // if.end:
368   Builder.SetInsertPoint(IfEnd);
369   Value *Shr =
370       Builder.CreateAShr(IntVal, Builder.getIntN(BitWidth, BitWidth - 1));
371   Value *Xor = Builder.CreateXor(Shr, IntVal);
372   Value *Sub = Builder.CreateSub(Xor, Shr);
373   Value *Call = Builder.CreateCall(CTLZ, {IsSigned ? Sub : IntVal, True});
374   Value *Cast = Builder.CreateTrunc(Call, Builder.getInt32Ty());
375   int BitWidthNew = FloatWidth == 128 ? BitWidth : 32;
376   Value *Sub1 = Builder.CreateSub(Builder.getIntN(BitWidthNew, BitWidth),
377                                   FloatWidth == 128 ? Call : Cast);
378   Value *Sub2 = Builder.CreateSub(Builder.getIntN(BitWidthNew, BitWidth - 1),
379                                   FloatWidth == 128 ? Call : Cast);
380   Value *Cmp3 = Builder.CreateICmpSGT(
381       Sub1, Builder.getIntN(BitWidthNew, FPMantissaWidth + 1));
382   Builder.CreateCondBr(Cmp3, IfThen4, IfElse);
383 
384   // if.then4:
385   Builder.SetInsertPoint(IfThen4);
386   llvm::SwitchInst *SI = Builder.CreateSwitch(Sub1, SwDefault);
387   SI->addCase(Builder.getIntN(BitWidthNew, FPMantissaWidth + 2), SwBB);
388   SI->addCase(Builder.getIntN(BitWidthNew, FPMantissaWidth + 3), SwEpilog);
389 
390   // sw.bb:
391   Builder.SetInsertPoint(SwBB);
392   Value *Shl =
393       Builder.CreateShl(IsSigned ? Sub : IntVal, Builder.getIntN(BitWidth, 1));
394   Builder.CreateBr(SwEpilog);
395 
396   // sw.default:
397   Builder.SetInsertPoint(SwDefault);
398   Value *Sub5 = Builder.CreateSub(
399       Builder.getIntN(BitWidthNew, BitWidth - FPMantissaWidth - 3),
400       FloatWidth == 128 ? Call : Cast);
401   Value *ShProm = Builder.CreateZExt(Sub5, IntTy);
402   Value *Shr6 = Builder.CreateLShr(IsSigned ? Sub : IntVal,
403                                    FloatWidth == 128 ? Sub5 : ShProm);
404   Value *Sub8 =
405       Builder.CreateAdd(FloatWidth == 128 ? Call : Cast,
406                         Builder.getIntN(BitWidthNew, FPMantissaWidth + 3));
407   Value *ShProm9 = Builder.CreateZExt(Sub8, IntTy);
408   Value *Shr9 = Builder.CreateLShr(ConstantInt::getSigned(IntTy, -1),
409                                    FloatWidth == 128 ? Sub8 : ShProm9);
410   Value *And = Builder.CreateAnd(Shr9, IsSigned ? Sub : IntVal);
411   Value *Cmp10 = Builder.CreateICmpNE(And, Builder.getIntN(BitWidth, 0));
412   Value *Conv11 = Builder.CreateZExt(Cmp10, IntTy);
413   Value *Or = Builder.CreateOr(Shr6, Conv11);
414   Builder.CreateBr(SwEpilog);
415 
416   // sw.epilog:
417   Builder.SetInsertPoint(SwEpilog);
418   PHINode *AAddr0 = Builder.CreatePHI(IntTy, 3);
419   AAddr0->addIncoming(Or, SwDefault);
420   AAddr0->addIncoming(IsSigned ? Sub : IntVal, IfThen4);
421   AAddr0->addIncoming(Shl, SwBB);
422   Value *A0 = Builder.CreateTrunc(AAddr0, Builder.getInt32Ty());
423   Value *A1 = Builder.CreateLShr(A0, Builder.getIntN(32, 2));
424   Value *A2 = Builder.CreateAnd(A1, Builder.getIntN(32, 1));
425   Value *Conv16 = Builder.CreateZExt(A2, IntTy);
426   Value *Or17 = Builder.CreateOr(AAddr0, Conv16);
427   Value *Inc = Builder.CreateAdd(Or17, Builder.getIntN(BitWidth, 1));
428   Value *Shr18 = nullptr;
429   if (IsSigned)
430     Shr18 = Builder.CreateAShr(Inc, Builder.getIntN(BitWidth, 2));
431   else
432     Shr18 = Builder.CreateLShr(Inc, Builder.getIntN(BitWidth, 2));
433   Value *A3 = Builder.CreateAnd(Inc, Temp1, "a3");
434   Value *PosOrNeg = Builder.CreateICmpEQ(A3, Builder.getIntN(BitWidth, 0));
435   Value *ExtractT60 = Builder.CreateTrunc(Shr18, Builder.getIntNTy(FloatWidth));
436   Value *Extract63 = Builder.CreateLShr(Shr18, Builder.getIntN(BitWidth, 32));
437   Value *ExtractT64 = nullptr;
438   if (FloatWidth > 80)
439     ExtractT64 = Builder.CreateTrunc(Sub2, Builder.getInt64Ty());
440   else
441     ExtractT64 = Builder.CreateTrunc(Extract63, Builder.getInt32Ty());
442   Builder.CreateCondBr(PosOrNeg, IfEnd26, IfThen20);
443 
444   // if.then20
445   Builder.SetInsertPoint(IfThen20);
446   Value *Shr21 = nullptr;
447   if (IsSigned)
448     Shr21 = Builder.CreateAShr(Inc, Builder.getIntN(BitWidth, 3));
449   else
450     Shr21 = Builder.CreateLShr(Inc, Builder.getIntN(BitWidth, 3));
451   Value *ExtractT = Builder.CreateTrunc(Shr21, Builder.getIntNTy(FloatWidth));
452   Value *Extract = Builder.CreateLShr(Shr21, Builder.getIntN(BitWidth, 32));
453   Value *ExtractT62 = nullptr;
454   if (FloatWidth > 80)
455     ExtractT62 = Builder.CreateTrunc(Sub1, Builder.getIntNTy(64));
456   else
457     ExtractT62 = Builder.CreateTrunc(Extract, Builder.getIntNTy(32));
458   Builder.CreateBr(IfEnd26);
459 
460   // if.else:
461   Builder.SetInsertPoint(IfElse);
462   Value *Sub24 = Builder.CreateAdd(
463       FloatWidth == 128 ? Call : Cast,
464       ConstantInt::getSigned(Builder.getIntNTy(BitWidthNew),
465                              -(BitWidth - FPMantissaWidth - 1)));
466   Value *ShProm25 = Builder.CreateZExt(Sub24, IntTy);
467   Value *Shl26 = Builder.CreateShl(IsSigned ? Sub : IntVal,
468                                    FloatWidth == 128 ? Sub24 : ShProm25);
469   Value *ExtractT61 = Builder.CreateTrunc(Shl26, Builder.getIntNTy(FloatWidth));
470   Value *Extract65 = Builder.CreateLShr(Shl26, Builder.getIntN(BitWidth, 32));
471   Value *ExtractT66 = nullptr;
472   if (FloatWidth > 80)
473     ExtractT66 = Builder.CreateTrunc(Sub2, Builder.getIntNTy(64));
474   else
475     ExtractT66 = Builder.CreateTrunc(Extract65, Builder.getInt32Ty());
476   Builder.CreateBr(IfEnd26);
477 
478   // if.end26:
479   Builder.SetInsertPoint(IfEnd26);
480   PHINode *AAddr1Off0 = Builder.CreatePHI(Builder.getIntNTy(FloatWidth), 3);
481   AAddr1Off0->addIncoming(ExtractT, IfThen20);
482   AAddr1Off0->addIncoming(ExtractT60, SwEpilog);
483   AAddr1Off0->addIncoming(ExtractT61, IfElse);
484   PHINode *AAddr1Off32 = nullptr;
485   if (FloatWidth > 32) {
486     AAddr1Off32 =
487         Builder.CreatePHI(Builder.getIntNTy(FloatWidth > 80 ? 64 : 32), 3);
488     AAddr1Off32->addIncoming(ExtractT62, IfThen20);
489     AAddr1Off32->addIncoming(ExtractT64, SwEpilog);
490     AAddr1Off32->addIncoming(ExtractT66, IfElse);
491   }
492   PHINode *E0 = nullptr;
493   if (FloatWidth <= 80) {
494     E0 = Builder.CreatePHI(Builder.getIntNTy(BitWidthNew), 3);
495     E0->addIncoming(Sub1, IfThen20);
496     E0->addIncoming(Sub2, SwEpilog);
497     E0->addIncoming(Sub2, IfElse);
498   }
499   Value *And29 = nullptr;
500   if (FloatWidth > 80) {
501     Value *Temp2 = Builder.CreateShl(Builder.getIntN(BitWidth, 1),
502                                      Builder.getIntN(BitWidth, 63));
503     And29 = Builder.CreateAnd(Shr, Temp2, "and29");
504   } else {
505     Value *Conv28 = Builder.CreateTrunc(Shr, Builder.getIntNTy(32));
506     And29 = Builder.CreateAnd(
507         Conv28, ConstantInt::getSigned(Builder.getIntNTy(32), 0x80000000));
508   }
509   unsigned TempMod = FPMantissaWidth % 32;
510   Value *And34 = nullptr;
511   Value *Shl30 = nullptr;
512   if (FloatWidth > 80) {
513     TempMod += 32;
514     Value *Add = Builder.CreateShl(AAddr1Off32, Builder.getIntN(64, TempMod));
515     Shl30 = Builder.CreateAdd(
516         Add,
517         Builder.getIntN(64, ((1ull << (62ull - TempMod)) - 1ull) << TempMod));
518     And34 = Builder.CreateZExt(Shl30, Builder.getIntNTy(128));
519   } else {
520     Value *Add = Builder.CreateShl(E0, Builder.getIntN(32, TempMod));
521     Shl30 = Builder.CreateAdd(
522         Add, Builder.getIntN(32, ((1 << (30 - TempMod)) - 1) << TempMod));
523     And34 = Builder.CreateAnd(FloatWidth > 32 ? AAddr1Off32 : AAddr1Off0,
524                               Builder.getIntN(32, (1 << TempMod) - 1));
525   }
526   Value *Or35 = nullptr;
527   if (FloatWidth > 80) {
528     Value *And29Trunc = Builder.CreateTrunc(And29, Builder.getIntNTy(128));
529     Value *Or31 = Builder.CreateOr(And29Trunc, And34);
530     Value *Or34 = Builder.CreateShl(Or31, Builder.getIntN(128, 64));
531     Value *Temp3 = Builder.CreateShl(Builder.getIntN(128, 1),
532                                      Builder.getIntN(128, FPMantissaWidth));
533     Value *Temp4 = Builder.CreateSub(Temp3, Builder.getIntN(128, 1));
534     Value *A6 = Builder.CreateAnd(AAddr1Off0, Temp4);
535     Or35 = Builder.CreateOr(Or34, A6);
536   } else {
537     Value *Or31 = Builder.CreateOr(And34, And29);
538     Or35 = Builder.CreateOr(IsSigned ? Or31 : And34, Shl30);
539   }
540   Value *A4 = nullptr;
541   if (IToFP->getType()->isDoubleTy()) {
542     Value *ZExt1 = Builder.CreateZExt(Or35, Builder.getIntNTy(FloatWidth));
543     Value *Shl1 = Builder.CreateShl(ZExt1, Builder.getIntN(FloatWidth, 32));
544     Value *And1 =
545         Builder.CreateAnd(AAddr1Off0, Builder.getIntN(FloatWidth, 0xFFFFFFFF));
546     Value *Or1 = Builder.CreateOr(Shl1, And1);
547     A4 = Builder.CreateBitCast(Or1, IToFP->getType());
548   } else if (IToFP->getType()->isX86_FP80Ty()) {
549     Value *A40 =
550         Builder.CreateBitCast(Or35, Type::getFP128Ty(Builder.getContext()));
551     A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
552   } else if (IToFP->getType()->isHalfTy() || IToFP->getType()->isBFloatTy()) {
553     // Deal with "half" situation. This is a workaround since we don't have
554     // floattihf.c currently as referring.
555     Value *A40 =
556         Builder.CreateBitCast(Or35, Type::getFloatTy(Builder.getContext()));
557     A4 = Builder.CreateFPTrunc(A40, IToFP->getType());
558   } else // float type
559     A4 = Builder.CreateBitCast(Or35, IToFP->getType());
560   Builder.CreateBr(End);
561 
562   // return:
563   Builder.SetInsertPoint(End, End->begin());
564   PHINode *Retval0 = Builder.CreatePHI(IToFP->getType(), 2);
565   Retval0->addIncoming(A4, IfEnd26);
566   Retval0->addIncoming(ConstantFP::getZero(IToFP->getType(), false), Entry);
567 
568   IToFP->replaceAllUsesWith(Retval0);
569   IToFP->dropAllReferences();
570   IToFP->eraseFromParent();
571 }
572 
573 static void scalarize(Instruction *I, SmallVectorImpl<Instruction *> &Replace) {
574   VectorType *VTy = cast<FixedVectorType>(I->getType());
575 
576   IRBuilder<> Builder(I);
577 
578   unsigned NumElements = VTy->getElementCount().getFixedValue();
579   Value *Result = PoisonValue::get(VTy);
580   for (unsigned Idx = 0; Idx < NumElements; ++Idx) {
581     Value *Ext = Builder.CreateExtractElement(I->getOperand(0), Idx);
582     Value *Cast = Builder.CreateCast(cast<CastInst>(I)->getOpcode(), Ext,
583                                      I->getType()->getScalarType());
584     Result = Builder.CreateInsertElement(Result, Cast, Idx);
585     if (isa<Instruction>(Cast))
586       Replace.push_back(cast<Instruction>(Cast));
587   }
588   I->replaceAllUsesWith(Result);
589   I->dropAllReferences();
590   I->eraseFromParent();
591 }
592 
593 static bool runImpl(Function &F, const TargetLowering &TLI) {
594   SmallVector<Instruction *, 4> Replace;
595   SmallVector<Instruction *, 4> ReplaceVector;
596   bool Modified = false;
597 
598   unsigned MaxLegalFpConvertBitWidth =
599       TLI.getMaxLargeFPConvertBitWidthSupported();
600   if (ExpandFpConvertBits != llvm::IntegerType::MAX_INT_BITS)
601     MaxLegalFpConvertBitWidth = ExpandFpConvertBits;
602 
603   if (MaxLegalFpConvertBitWidth >= llvm::IntegerType::MAX_INT_BITS)
604     return false;
605 
606   for (auto &I : instructions(F)) {
607     switch (I.getOpcode()) {
608     case Instruction::FPToUI:
609     case Instruction::FPToSI: {
610       // TODO: This pass doesn't handle scalable vectors.
611       if (I.getOperand(0)->getType()->isScalableTy())
612         continue;
613 
614       auto *IntTy = cast<IntegerType>(I.getType()->getScalarType());
615       if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth)
616         continue;
617 
618       if (I.getOperand(0)->getType()->isVectorTy())
619         ReplaceVector.push_back(&I);
620       else
621         Replace.push_back(&I);
622       Modified = true;
623       break;
624     }
625     case Instruction::UIToFP:
626     case Instruction::SIToFP: {
627       // TODO: This pass doesn't handle scalable vectors.
628       if (I.getOperand(0)->getType()->isScalableTy())
629         continue;
630 
631       auto *IntTy =
632           cast<IntegerType>(I.getOperand(0)->getType()->getScalarType());
633       if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth)
634         continue;
635 
636       if (I.getOperand(0)->getType()->isVectorTy())
637         ReplaceVector.push_back(&I);
638       else
639         Replace.push_back(&I);
640       Modified = true;
641       break;
642     }
643     default:
644       break;
645     }
646   }
647 
648   while (!ReplaceVector.empty()) {
649     Instruction *I = ReplaceVector.pop_back_val();
650     scalarize(I, Replace);
651   }
652 
653   if (Replace.empty())
654     return false;
655 
656   while (!Replace.empty()) {
657     Instruction *I = Replace.pop_back_val();
658     if (I->getOpcode() == Instruction::FPToUI ||
659         I->getOpcode() == Instruction::FPToSI) {
660       expandFPToI(I);
661     } else {
662       expandIToFP(I);
663     }
664   }
665 
666   return Modified;
667 }
668 
669 namespace {
670 class ExpandLargeFpConvertLegacyPass : public FunctionPass {
671 public:
672   static char ID;
673 
674   ExpandLargeFpConvertLegacyPass() : FunctionPass(ID) {
675     initializeExpandLargeFpConvertLegacyPassPass(
676         *PassRegistry::getPassRegistry());
677   }
678 
679   bool runOnFunction(Function &F) override {
680     auto *TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
681     auto *TLI = TM->getSubtargetImpl(F)->getTargetLowering();
682     return runImpl(F, *TLI);
683   }
684 
685   void getAnalysisUsage(AnalysisUsage &AU) const override {
686     AU.addRequired<TargetPassConfig>();
687     AU.addPreserved<AAResultsWrapperPass>();
688     AU.addPreserved<GlobalsAAWrapperPass>();
689   }
690 };
691 } // namespace
692 
693 PreservedAnalyses ExpandLargeFpConvertPass::run(Function &F,
694                                                 FunctionAnalysisManager &FAM) {
695   const TargetSubtargetInfo *STI = TM->getSubtargetImpl(F);
696   return runImpl(F, *STI->getTargetLowering()) ? PreservedAnalyses::none()
697                                                : PreservedAnalyses::all();
698 }
699 
700 char ExpandLargeFpConvertLegacyPass::ID = 0;
701 INITIALIZE_PASS_BEGIN(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
702                       "Expand large fp convert", false, false)
703 INITIALIZE_PASS_END(ExpandLargeFpConvertLegacyPass, "expand-large-fp-convert",
704                     "Expand large fp convert", false, false)
705 
706 FunctionPass *llvm::createExpandLargeFpConvertPass() {
707   return new ExpandLargeFpConvertLegacyPass();
708 }
709