1 | //===--- ExpandFp.cpp - Expand fp instructions ----------------------------===// |
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 | // This pass expands certain floating point instructions at the IR level. |
9 | // |
10 | // It expands ‘fptoui .. to’, ‘fptosi .. to’, ‘uitofp .. to’, ‘sitofp |
11 | // .. to’ instructions with a bitwidth above a threshold. This is |
12 | // useful for targets like x86_64 that cannot lower fp convertions |
13 | // with more than 128 bits. |
14 | // |
15 | //===----------------------------------------------------------------------===// |
16 | |
17 | #include "llvm/CodeGen/ExpandFp.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(Val: llvm::IntegerType::MAX_INT_BITS), |
37 | cl::desc("fp convert instructions on integers with " |
38 | "more than <N> bits are expanded." )); |
39 | |
40 | // clang-format off: preserve formatting of the following example |
41 | |
42 | /// Generate code to convert a fp number to integer, replacing FPToS(U)I with |
43 | /// the generated code. This currently generates code similarly to compiler-rt's |
44 | /// implementations. |
45 | /// |
46 | /// An example IR generated from compiler-rt/fixsfdi.c looks like below: |
47 | /// define dso_local i64 @foo(float noundef %a) local_unnamed_addr #0 { |
48 | /// entry: |
49 | /// %0 = bitcast float %a to i32 |
50 | /// %conv.i = zext i32 %0 to i64 |
51 | /// %tobool.not = icmp sgt i32 %0, -1 |
52 | /// %conv = select i1 %tobool.not, i64 1, i64 -1 |
53 | /// %and = lshr i64 %conv.i, 23 |
54 | /// %shr = and i64 %and, 255 |
55 | /// %and2 = and i64 %conv.i, 8388607 |
56 | /// %or = or i64 %and2, 8388608 |
57 | /// %cmp = icmp ult i64 %shr, 127 |
58 | /// br i1 %cmp, label %cleanup, label %if.end |
59 | /// |
60 | /// if.end: ; preds = %entry |
61 | /// %sub = add nuw nsw i64 %shr, 4294967169 |
62 | /// %conv5 = and i64 %sub, 4294967232 |
63 | /// %cmp6.not = icmp eq i64 %conv5, 0 |
64 | /// br i1 %cmp6.not, label %if.end12, label %if.then8 |
65 | /// |
66 | /// if.then8: ; preds = %if.end |
67 | /// %cond11 = select i1 %tobool.not, i64 9223372036854775807, i64 -9223372036854775808 |
68 | /// br label %cleanup |
69 | /// |
70 | /// if.end12: ; preds = %if.end |
71 | /// %cmp13 = icmp ult i64 %shr, 150 |
72 | /// br i1 %cmp13, label %if.then15, label %if.else |
73 | /// |
74 | /// if.then15: ; preds = %if.end12 |
75 | /// %sub16 = sub nuw nsw i64 150, %shr |
76 | /// %shr17 = lshr i64 %or, %sub16 |
77 | /// %mul = mul nsw i64 %shr17, %conv |
78 | /// br label %cleanup |
79 | /// |
80 | /// if.else: ; preds = %if.end12 |
81 | /// %sub18 = add nsw i64 %shr, -150 |
82 | /// %shl = shl i64 %or, %sub18 |
83 | /// %mul19 = mul nsw i64 %shl, %conv |
84 | /// br label %cleanup |
85 | /// |
86 | /// cleanup: ; preds = %entry, %if.else, %if.then15, %if.then8 |
87 | /// %retval.0 = phi i64 [ %cond11, %if.then8 ], [ %mul, %if.then15 ], [ %mul19, %if.else ], [ 0, %entry ] |
88 | /// ret i64 %retval.0 |
89 | /// } |
90 | /// |
91 | /// Replace fp to integer with generated code. |
92 | static void expandFPToI(Instruction *FPToI) { |
93 | // clang-format on |
94 | IRBuilder<> Builder(FPToI); |
95 | auto *FloatVal = FPToI->getOperand(i: 0); |
96 | IntegerType *IntTy = cast<IntegerType>(Val: FPToI->getType()); |
97 | |
98 | unsigned BitWidth = FPToI->getType()->getIntegerBitWidth(); |
99 | unsigned FPMantissaWidth = FloatVal->getType()->getFPMantissaWidth() - 1; |
100 | |
101 | // FIXME: fp16's range is covered by i32. So `fptoi half` can convert |
102 | // to i32 first following a sext/zext to target integer type. |
103 | Value *A1 = nullptr; |
104 | if (FloatVal->getType()->isHalfTy()) { |
105 | if (FPToI->getOpcode() == Instruction::FPToUI) { |
106 | Value *A0 = Builder.CreateFPToUI(V: FloatVal, DestTy: Builder.getIntNTy(N: 32)); |
107 | A1 = Builder.CreateZExt(V: A0, DestTy: IntTy); |
108 | } else { // FPToSI |
109 | Value *A0 = Builder.CreateFPToSI(V: FloatVal, DestTy: Builder.getIntNTy(N: 32)); |
110 | A1 = Builder.CreateSExt(V: A0, DestTy: IntTy); |
111 | } |
112 | FPToI->replaceAllUsesWith(V: A1); |
113 | FPToI->dropAllReferences(); |
114 | FPToI->eraseFromParent(); |
115 | return; |
116 | } |
117 | |
118 | // fp80 conversion is implemented by fpext to fp128 first then do the |
119 | // conversion. |
120 | FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth; |
121 | unsigned FloatWidth = |
122 | PowerOf2Ceil(A: FloatVal->getType()->getScalarSizeInBits()); |
123 | unsigned ExponentWidth = FloatWidth - FPMantissaWidth - 1; |
124 | unsigned ExponentBias = (1 << (ExponentWidth - 1)) - 1; |
125 | Value *ImplicitBit = Builder.CreateShl( |
126 | LHS: Builder.getIntN(N: BitWidth, C: 1), RHS: Builder.getIntN(N: BitWidth, C: FPMantissaWidth)); |
127 | Value *SignificandMask = |
128 | Builder.CreateSub(LHS: ImplicitBit, RHS: Builder.getIntN(N: BitWidth, C: 1)); |
129 | Value *NegOne = Builder.CreateSExt( |
130 | V: ConstantInt::getSigned(Ty: Builder.getInt32Ty(), V: -1), DestTy: IntTy); |
131 | Value *NegInf = |
132 | Builder.CreateShl(LHS: ConstantInt::getSigned(Ty: IntTy, V: 1), |
133 | RHS: ConstantInt::getSigned(Ty: IntTy, V: BitWidth - 1)); |
134 | |
135 | BasicBlock *Entry = Builder.GetInsertBlock(); |
136 | Function *F = Entry->getParent(); |
137 | Entry->setName(Twine(Entry->getName(), "fp-to-i-entry" )); |
138 | BasicBlock *End = |
139 | Entry->splitBasicBlock(I: Builder.GetInsertPoint(), BBName: "fp-to-i-cleanup" ); |
140 | BasicBlock *IfEnd = |
141 | BasicBlock::Create(Context&: Builder.getContext(), Name: "fp-to-i-if-end" , Parent: F, InsertBefore: End); |
142 | BasicBlock *IfThen5 = |
143 | BasicBlock::Create(Context&: Builder.getContext(), Name: "fp-to-i-if-then5" , Parent: F, InsertBefore: End); |
144 | BasicBlock *IfEnd9 = |
145 | BasicBlock::Create(Context&: Builder.getContext(), Name: "fp-to-i-if-end9" , Parent: F, InsertBefore: End); |
146 | BasicBlock *IfThen12 = |
147 | BasicBlock::Create(Context&: Builder.getContext(), Name: "fp-to-i-if-then12" , Parent: F, InsertBefore: End); |
148 | BasicBlock *IfElse = |
149 | BasicBlock::Create(Context&: Builder.getContext(), Name: "fp-to-i-if-else" , Parent: F, InsertBefore: End); |
150 | |
151 | Entry->getTerminator()->eraseFromParent(); |
152 | |
153 | // entry: |
154 | Builder.SetInsertPoint(Entry); |
155 | Value *FloatVal0 = FloatVal; |
156 | // fp80 conversion is implemented by fpext to fp128 first then do the |
157 | // conversion. |
158 | if (FloatVal->getType()->isX86_FP80Ty()) |
159 | FloatVal0 = |
160 | Builder.CreateFPExt(V: FloatVal, DestTy: Type::getFP128Ty(C&: Builder.getContext())); |
161 | Value *ARep0 = |
162 | Builder.CreateBitCast(V: FloatVal0, DestTy: Builder.getIntNTy(N: FloatWidth)); |
163 | Value *ARep = Builder.CreateZExt(V: ARep0, DestTy: FPToI->getType()); |
164 | Value *PosOrNeg = Builder.CreateICmpSGT( |
165 | LHS: ARep0, RHS: ConstantInt::getSigned(Ty: Builder.getIntNTy(N: FloatWidth), V: -1)); |
166 | Value *Sign = Builder.CreateSelect(C: PosOrNeg, True: ConstantInt::getSigned(Ty: IntTy, V: 1), |
167 | False: ConstantInt::getSigned(Ty: IntTy, V: -1)); |
168 | Value *And = |
169 | Builder.CreateLShr(LHS: ARep, RHS: Builder.getIntN(N: BitWidth, C: FPMantissaWidth)); |
170 | Value *And2 = Builder.CreateAnd( |
171 | LHS: And, RHS: Builder.getIntN(N: BitWidth, C: (1 << ExponentWidth) - 1)); |
172 | Value *Abs = Builder.CreateAnd(LHS: ARep, RHS: SignificandMask); |
173 | Value *Or = Builder.CreateOr(LHS: Abs, RHS: ImplicitBit); |
174 | Value *Cmp = |
175 | Builder.CreateICmpULT(LHS: And2, RHS: Builder.getIntN(N: BitWidth, C: ExponentBias)); |
176 | Builder.CreateCondBr(Cond: Cmp, True: End, False: IfEnd); |
177 | |
178 | // if.end: |
179 | Builder.SetInsertPoint(IfEnd); |
180 | Value *Add1 = Builder.CreateAdd( |
181 | LHS: And2, RHS: ConstantInt::getSigned( |
182 | Ty: IntTy, V: -static_cast<int64_t>(ExponentBias + BitWidth))); |
183 | Value *Cmp3 = Builder.CreateICmpULT( |
184 | LHS: Add1, RHS: ConstantInt::getSigned(Ty: IntTy, V: -static_cast<int64_t>(BitWidth))); |
185 | Builder.CreateCondBr(Cond: Cmp3, True: IfThen5, False: IfEnd9); |
186 | |
187 | // if.then5: |
188 | Builder.SetInsertPoint(IfThen5); |
189 | Value *PosInf = Builder.CreateXor(LHS: NegOne, RHS: NegInf); |
190 | Value *Cond8 = Builder.CreateSelect(C: PosOrNeg, True: PosInf, False: NegInf); |
191 | Builder.CreateBr(Dest: End); |
192 | |
193 | // if.end9: |
194 | Builder.SetInsertPoint(IfEnd9); |
195 | Value *Cmp10 = Builder.CreateICmpULT( |
196 | LHS: And2, RHS: Builder.getIntN(N: BitWidth, C: ExponentBias + FPMantissaWidth)); |
197 | Builder.CreateCondBr(Cond: Cmp10, True: IfThen12, False: IfElse); |
198 | |
199 | // if.then12: |
200 | Builder.SetInsertPoint(IfThen12); |
201 | Value *Sub13 = Builder.CreateSub( |
202 | LHS: Builder.getIntN(N: BitWidth, C: ExponentBias + FPMantissaWidth), RHS: And2); |
203 | Value *Shr14 = Builder.CreateLShr(LHS: Or, RHS: Sub13); |
204 | Value *Mul = Builder.CreateMul(LHS: Shr14, RHS: Sign); |
205 | Builder.CreateBr(Dest: End); |
206 | |
207 | // if.else: |
208 | Builder.SetInsertPoint(IfElse); |
209 | Value *Sub15 = Builder.CreateAdd( |
210 | LHS: And2, RHS: ConstantInt::getSigned( |
211 | Ty: IntTy, V: -static_cast<int64_t>(ExponentBias + FPMantissaWidth))); |
212 | Value *Shl = Builder.CreateShl(LHS: Or, RHS: Sub15); |
213 | Value *Mul16 = Builder.CreateMul(LHS: Shl, RHS: Sign); |
214 | Builder.CreateBr(Dest: End); |
215 | |
216 | // cleanup: |
217 | Builder.SetInsertPoint(TheBB: End, IP: End->begin()); |
218 | PHINode *Retval0 = Builder.CreatePHI(Ty: FPToI->getType(), NumReservedValues: 4); |
219 | |
220 | Retval0->addIncoming(V: Cond8, BB: IfThen5); |
221 | Retval0->addIncoming(V: Mul, BB: IfThen12); |
222 | Retval0->addIncoming(V: Mul16, BB: IfElse); |
223 | Retval0->addIncoming(V: Builder.getIntN(N: BitWidth, C: 0), BB: Entry); |
224 | |
225 | FPToI->replaceAllUsesWith(V: Retval0); |
226 | FPToI->dropAllReferences(); |
227 | FPToI->eraseFromParent(); |
228 | } |
229 | |
230 | // clang-format off: preserve formatting of the following example |
231 | |
232 | /// Generate code to convert a fp number to integer, replacing S(U)IToFP with |
233 | /// the generated code. This currently generates code similarly to compiler-rt's |
234 | /// implementations. This implementation has an implicit assumption that integer |
235 | /// width is larger than fp. |
236 | /// |
237 | /// An example IR generated from compiler-rt/floatdisf.c looks like below: |
238 | /// define dso_local float @__floatdisf(i64 noundef %a) local_unnamed_addr #0 { |
239 | /// entry: |
240 | /// %cmp = icmp eq i64 %a, 0 |
241 | /// br i1 %cmp, label %return, label %if.end |
242 | /// |
243 | /// if.end: ; preds = %entry |
244 | /// %shr = ashr i64 %a, 63 |
245 | /// %xor = xor i64 %shr, %a |
246 | /// %sub = sub nsw i64 %xor, %shr |
247 | /// %0 = tail call i64 @llvm.ctlz.i64(i64 %sub, i1 true), !range !5 |
248 | /// %cast = trunc i64 %0 to i32 |
249 | /// %sub1 = sub nuw nsw i32 64, %cast |
250 | /// %sub2 = xor i32 %cast, 63 |
251 | /// %cmp3 = icmp ult i32 %cast, 40 |
252 | /// br i1 %cmp3, label %if.then4, label %if.else |
253 | /// |
254 | /// if.then4: ; preds = %if.end |
255 | /// switch i32 %sub1, label %sw.default [ |
256 | /// i32 25, label %sw.bb |
257 | /// i32 26, label %sw.epilog |
258 | /// ] |
259 | /// |
260 | /// sw.bb: ; preds = %if.then4 |
261 | /// %shl = shl i64 %sub, 1 |
262 | /// br label %sw.epilog |
263 | /// |
264 | /// sw.default: ; preds = %if.then4 |
265 | /// %sub5 = sub nsw i64 38, %0 |
266 | /// %sh_prom = and i64 %sub5, 4294967295 |
267 | /// %shr6 = lshr i64 %sub, %sh_prom |
268 | /// %shr9 = lshr i64 274877906943, %0 |
269 | /// %and = and i64 %shr9, %sub |
270 | /// %cmp10 = icmp ne i64 %and, 0 |
271 | /// %conv11 = zext i1 %cmp10 to i64 |
272 | /// %or = or i64 %shr6, %conv11 |
273 | /// br label %sw.epilog |
274 | /// |
275 | /// sw.epilog: ; preds = %sw.default, %if.then4, %sw.bb |
276 | /// %a.addr.0 = phi i64 [ %or, %sw.default ], [ %sub, %if.then4 ], [ %shl, %sw.bb ] |
277 | /// %1 = lshr i64 %a.addr.0, 2 |
278 | /// %2 = and i64 %1, 1 |
279 | /// %or16 = or i64 %2, %a.addr.0 |
280 | /// %inc = add nsw i64 %or16, 1 |
281 | /// %3 = and i64 %inc, 67108864 |
282 | /// %tobool.not = icmp eq i64 %3, 0 |
283 | /// %spec.select.v = select i1 %tobool.not, i64 2, i64 3 |
284 | /// %spec.select = ashr i64 %inc, %spec.select.v |
285 | /// %spec.select56 = select i1 %tobool.not, i32 %sub2, i32 %sub1 |
286 | /// br label %if.end26 |
287 | /// |
288 | /// if.else: ; preds = %if.end |
289 | /// %sub23 = add nuw nsw i64 %0, 4294967256 |
290 | /// %sh_prom24 = and i64 %sub23, 4294967295 |
291 | /// %shl25 = shl i64 %sub, %sh_prom24 |
292 | /// br label %if.end26 |
293 | /// |
294 | /// if.end26: ; preds = %sw.epilog, %if.else |
295 | /// %a.addr.1 = phi i64 [ %shl25, %if.else ], [ %spec.select, %sw.epilog ] |
296 | /// %e.0 = phi i32 [ %sub2, %if.else ], [ %spec.select56, %sw.epilog ] |
297 | /// %conv27 = trunc i64 %shr to i32 |
298 | /// %and28 = and i32 %conv27, -2147483648 |
299 | /// %add = shl nuw nsw i32 %e.0, 23 |
300 | /// %shl29 = add nuw nsw i32 %add, 1065353216 |
301 | /// %conv31 = trunc i64 %a.addr.1 to i32 |
302 | /// %and32 = and i32 %conv31, 8388607 |
303 | /// %or30 = or i32 %and32, %and28 |
304 | /// %or33 = or i32 %or30, %shl29 |
305 | /// %4 = bitcast i32 %or33 to float |
306 | /// br label %return |
307 | /// |
308 | /// return: ; preds = %entry, %if.end26 |
309 | /// %retval.0 = phi float [ %4, %if.end26 ], [ 0.000000e+00, %entry ] |
310 | /// ret float %retval.0 |
311 | /// } |
312 | /// |
313 | /// Replace integer to fp with generated code. |
314 | static void expandIToFP(Instruction *IToFP) { |
315 | // clang-format on |
316 | IRBuilder<> Builder(IToFP); |
317 | auto *IntVal = IToFP->getOperand(i: 0); |
318 | IntegerType *IntTy = cast<IntegerType>(Val: IntVal->getType()); |
319 | |
320 | unsigned BitWidth = IntVal->getType()->getIntegerBitWidth(); |
321 | unsigned FPMantissaWidth = IToFP->getType()->getFPMantissaWidth() - 1; |
322 | // fp80 conversion is implemented by conversion tp fp128 first following |
323 | // a fptrunc to fp80. |
324 | FPMantissaWidth = FPMantissaWidth == 63 ? 112 : FPMantissaWidth; |
325 | // FIXME: As there is no related builtins added in compliler-rt, |
326 | // here currently utilized the fp32 <-> fp16 lib calls to implement. |
327 | FPMantissaWidth = FPMantissaWidth == 10 ? 23 : FPMantissaWidth; |
328 | FPMantissaWidth = FPMantissaWidth == 7 ? 23 : FPMantissaWidth; |
329 | unsigned FloatWidth = PowerOf2Ceil(A: FPMantissaWidth); |
330 | bool IsSigned = IToFP->getOpcode() == Instruction::SIToFP; |
331 | |
332 | assert(BitWidth > FloatWidth && "Unexpected conversion. expandIToFP() " |
333 | "assumes integer width is larger than fp." ); |
334 | |
335 | Value *Temp1 = |
336 | Builder.CreateShl(LHS: Builder.getIntN(N: BitWidth, C: 1), |
337 | RHS: Builder.getIntN(N: BitWidth, C: FPMantissaWidth + 3)); |
338 | |
339 | BasicBlock *Entry = Builder.GetInsertBlock(); |
340 | Function *F = Entry->getParent(); |
341 | Entry->setName(Twine(Entry->getName(), "itofp-entry" )); |
342 | BasicBlock *End = |
343 | Entry->splitBasicBlock(I: Builder.GetInsertPoint(), BBName: "itofp-return" ); |
344 | BasicBlock *IfEnd = |
345 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-if-end" , Parent: F, InsertBefore: End); |
346 | BasicBlock *IfThen4 = |
347 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-if-then4" , Parent: F, InsertBefore: End); |
348 | BasicBlock *SwBB = |
349 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-sw-bb" , Parent: F, InsertBefore: End); |
350 | BasicBlock *SwDefault = |
351 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-sw-default" , Parent: F, InsertBefore: End); |
352 | BasicBlock *SwEpilog = |
353 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-sw-epilog" , Parent: F, InsertBefore: End); |
354 | BasicBlock *IfThen20 = |
355 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-if-then20" , Parent: F, InsertBefore: End); |
356 | BasicBlock *IfElse = |
357 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-if-else" , Parent: F, InsertBefore: End); |
358 | BasicBlock *IfEnd26 = |
359 | BasicBlock::Create(Context&: Builder.getContext(), Name: "itofp-if-end26" , Parent: F, InsertBefore: End); |
360 | |
361 | Entry->getTerminator()->eraseFromParent(); |
362 | |
363 | Function *CTLZ = |
364 | Intrinsic::getOrInsertDeclaration(M: F->getParent(), id: Intrinsic::ctlz, Tys: IntTy); |
365 | ConstantInt *True = Builder.getTrue(); |
366 | |
367 | // entry: |
368 | Builder.SetInsertPoint(Entry); |
369 | Value *Cmp = Builder.CreateICmpEQ(LHS: IntVal, RHS: ConstantInt::getSigned(Ty: IntTy, V: 0)); |
370 | Builder.CreateCondBr(Cond: Cmp, True: End, False: IfEnd); |
371 | |
372 | // if.end: |
373 | Builder.SetInsertPoint(IfEnd); |
374 | Value *Shr = |
375 | Builder.CreateAShr(LHS: IntVal, RHS: Builder.getIntN(N: BitWidth, C: BitWidth - 1)); |
376 | Value *Xor = Builder.CreateXor(LHS: Shr, RHS: IntVal); |
377 | Value *Sub = Builder.CreateSub(LHS: Xor, RHS: Shr); |
378 | Value *Call = Builder.CreateCall(Callee: CTLZ, Args: {IsSigned ? Sub : IntVal, True}); |
379 | Value *Cast = Builder.CreateTrunc(V: Call, DestTy: Builder.getInt32Ty()); |
380 | int BitWidthNew = FloatWidth == 128 ? BitWidth : 32; |
381 | Value *Sub1 = Builder.CreateSub(LHS: Builder.getIntN(N: BitWidthNew, C: BitWidth), |
382 | RHS: FloatWidth == 128 ? Call : Cast); |
383 | Value *Sub2 = Builder.CreateSub(LHS: Builder.getIntN(N: BitWidthNew, C: BitWidth - 1), |
384 | RHS: FloatWidth == 128 ? Call : Cast); |
385 | Value *Cmp3 = Builder.CreateICmpSGT( |
386 | LHS: Sub1, RHS: Builder.getIntN(N: BitWidthNew, C: FPMantissaWidth + 1)); |
387 | Builder.CreateCondBr(Cond: Cmp3, True: IfThen4, False: IfElse); |
388 | |
389 | // if.then4: |
390 | Builder.SetInsertPoint(IfThen4); |
391 | llvm::SwitchInst *SI = Builder.CreateSwitch(V: Sub1, Dest: SwDefault); |
392 | SI->addCase(OnVal: Builder.getIntN(N: BitWidthNew, C: FPMantissaWidth + 2), Dest: SwBB); |
393 | SI->addCase(OnVal: Builder.getIntN(N: BitWidthNew, C: FPMantissaWidth + 3), Dest: SwEpilog); |
394 | |
395 | // sw.bb: |
396 | Builder.SetInsertPoint(SwBB); |
397 | Value *Shl = |
398 | Builder.CreateShl(LHS: IsSigned ? Sub : IntVal, RHS: Builder.getIntN(N: BitWidth, C: 1)); |
399 | Builder.CreateBr(Dest: SwEpilog); |
400 | |
401 | // sw.default: |
402 | Builder.SetInsertPoint(SwDefault); |
403 | Value *Sub5 = Builder.CreateSub( |
404 | LHS: Builder.getIntN(N: BitWidthNew, C: BitWidth - FPMantissaWidth - 3), |
405 | RHS: FloatWidth == 128 ? Call : Cast); |
406 | Value *ShProm = Builder.CreateZExt(V: Sub5, DestTy: IntTy); |
407 | Value *Shr6 = Builder.CreateLShr(LHS: IsSigned ? Sub : IntVal, |
408 | RHS: FloatWidth == 128 ? Sub5 : ShProm); |
409 | Value *Sub8 = |
410 | Builder.CreateAdd(LHS: FloatWidth == 128 ? Call : Cast, |
411 | RHS: Builder.getIntN(N: BitWidthNew, C: FPMantissaWidth + 3)); |
412 | Value *ShProm9 = Builder.CreateZExt(V: Sub8, DestTy: IntTy); |
413 | Value *Shr9 = Builder.CreateLShr(LHS: ConstantInt::getSigned(Ty: IntTy, V: -1), |
414 | RHS: FloatWidth == 128 ? Sub8 : ShProm9); |
415 | Value *And = Builder.CreateAnd(LHS: Shr9, RHS: IsSigned ? Sub : IntVal); |
416 | Value *Cmp10 = Builder.CreateICmpNE(LHS: And, RHS: Builder.getIntN(N: BitWidth, C: 0)); |
417 | Value *Conv11 = Builder.CreateZExt(V: Cmp10, DestTy: IntTy); |
418 | Value *Or = Builder.CreateOr(LHS: Shr6, RHS: Conv11); |
419 | Builder.CreateBr(Dest: SwEpilog); |
420 | |
421 | // sw.epilog: |
422 | Builder.SetInsertPoint(SwEpilog); |
423 | PHINode *AAddr0 = Builder.CreatePHI(Ty: IntTy, NumReservedValues: 3); |
424 | AAddr0->addIncoming(V: Or, BB: SwDefault); |
425 | AAddr0->addIncoming(V: IsSigned ? Sub : IntVal, BB: IfThen4); |
426 | AAddr0->addIncoming(V: Shl, BB: SwBB); |
427 | Value *A0 = Builder.CreateTrunc(V: AAddr0, DestTy: Builder.getInt32Ty()); |
428 | Value *A1 = Builder.CreateLShr(LHS: A0, RHS: Builder.getIntN(N: 32, C: 2)); |
429 | Value *A2 = Builder.CreateAnd(LHS: A1, RHS: Builder.getIntN(N: 32, C: 1)); |
430 | Value *Conv16 = Builder.CreateZExt(V: A2, DestTy: IntTy); |
431 | Value *Or17 = Builder.CreateOr(LHS: AAddr0, RHS: Conv16); |
432 | Value *Inc = Builder.CreateAdd(LHS: Or17, RHS: Builder.getIntN(N: BitWidth, C: 1)); |
433 | Value *Shr18 = nullptr; |
434 | if (IsSigned) |
435 | Shr18 = Builder.CreateAShr(LHS: Inc, RHS: Builder.getIntN(N: BitWidth, C: 2)); |
436 | else |
437 | Shr18 = Builder.CreateLShr(LHS: Inc, RHS: Builder.getIntN(N: BitWidth, C: 2)); |
438 | Value *A3 = Builder.CreateAnd(LHS: Inc, RHS: Temp1, Name: "a3" ); |
439 | Value *PosOrNeg = Builder.CreateICmpEQ(LHS: A3, RHS: Builder.getIntN(N: BitWidth, C: 0)); |
440 | Value * = Builder.CreateTrunc(V: Shr18, DestTy: Builder.getIntNTy(N: FloatWidth)); |
441 | Value * = Builder.CreateLShr(LHS: Shr18, RHS: Builder.getIntN(N: BitWidth, C: 32)); |
442 | Value * = nullptr; |
443 | if (FloatWidth > 80) |
444 | ExtractT64 = Builder.CreateTrunc(V: Sub2, DestTy: Builder.getInt64Ty()); |
445 | else |
446 | ExtractT64 = Builder.CreateTrunc(V: Extract63, DestTy: Builder.getInt32Ty()); |
447 | Builder.CreateCondBr(Cond: PosOrNeg, True: IfEnd26, False: IfThen20); |
448 | |
449 | // if.then20 |
450 | Builder.SetInsertPoint(IfThen20); |
451 | Value *Shr21 = nullptr; |
452 | if (IsSigned) |
453 | Shr21 = Builder.CreateAShr(LHS: Inc, RHS: Builder.getIntN(N: BitWidth, C: 3)); |
454 | else |
455 | Shr21 = Builder.CreateLShr(LHS: Inc, RHS: Builder.getIntN(N: BitWidth, C: 3)); |
456 | Value * = Builder.CreateTrunc(V: Shr21, DestTy: Builder.getIntNTy(N: FloatWidth)); |
457 | Value * = Builder.CreateLShr(LHS: Shr21, RHS: Builder.getIntN(N: BitWidth, C: 32)); |
458 | Value * = nullptr; |
459 | if (FloatWidth > 80) |
460 | ExtractT62 = Builder.CreateTrunc(V: Sub1, DestTy: Builder.getIntNTy(N: 64)); |
461 | else |
462 | ExtractT62 = Builder.CreateTrunc(V: Extract, DestTy: Builder.getIntNTy(N: 32)); |
463 | Builder.CreateBr(Dest: IfEnd26); |
464 | |
465 | // if.else: |
466 | Builder.SetInsertPoint(IfElse); |
467 | Value *Sub24 = Builder.CreateAdd( |
468 | LHS: FloatWidth == 128 ? Call : Cast, |
469 | RHS: ConstantInt::getSigned(Ty: Builder.getIntNTy(N: BitWidthNew), |
470 | V: -(BitWidth - FPMantissaWidth - 1))); |
471 | Value *ShProm25 = Builder.CreateZExt(V: Sub24, DestTy: IntTy); |
472 | Value *Shl26 = Builder.CreateShl(LHS: IsSigned ? Sub : IntVal, |
473 | RHS: FloatWidth == 128 ? Sub24 : ShProm25); |
474 | Value * = Builder.CreateTrunc(V: Shl26, DestTy: Builder.getIntNTy(N: FloatWidth)); |
475 | Value * = Builder.CreateLShr(LHS: Shl26, RHS: Builder.getIntN(N: BitWidth, C: 32)); |
476 | Value * = nullptr; |
477 | if (FloatWidth > 80) |
478 | ExtractT66 = Builder.CreateTrunc(V: Sub2, DestTy: Builder.getIntNTy(N: 64)); |
479 | else |
480 | ExtractT66 = Builder.CreateTrunc(V: Extract65, DestTy: Builder.getInt32Ty()); |
481 | Builder.CreateBr(Dest: IfEnd26); |
482 | |
483 | // if.end26: |
484 | Builder.SetInsertPoint(IfEnd26); |
485 | PHINode *AAddr1Off0 = Builder.CreatePHI(Ty: Builder.getIntNTy(N: FloatWidth), NumReservedValues: 3); |
486 | AAddr1Off0->addIncoming(V: ExtractT, BB: IfThen20); |
487 | AAddr1Off0->addIncoming(V: ExtractT60, BB: SwEpilog); |
488 | AAddr1Off0->addIncoming(V: ExtractT61, BB: IfElse); |
489 | PHINode *AAddr1Off32 = nullptr; |
490 | if (FloatWidth > 32) { |
491 | AAddr1Off32 = |
492 | Builder.CreatePHI(Ty: Builder.getIntNTy(N: FloatWidth > 80 ? 64 : 32), NumReservedValues: 3); |
493 | AAddr1Off32->addIncoming(V: ExtractT62, BB: IfThen20); |
494 | AAddr1Off32->addIncoming(V: ExtractT64, BB: SwEpilog); |
495 | AAddr1Off32->addIncoming(V: ExtractT66, BB: IfElse); |
496 | } |
497 | PHINode *E0 = nullptr; |
498 | if (FloatWidth <= 80) { |
499 | E0 = Builder.CreatePHI(Ty: Builder.getIntNTy(N: BitWidthNew), NumReservedValues: 3); |
500 | E0->addIncoming(V: Sub1, BB: IfThen20); |
501 | E0->addIncoming(V: Sub2, BB: SwEpilog); |
502 | E0->addIncoming(V: Sub2, BB: IfElse); |
503 | } |
504 | Value *And29 = nullptr; |
505 | if (FloatWidth > 80) { |
506 | Value *Temp2 = Builder.CreateShl(LHS: Builder.getIntN(N: BitWidth, C: 1), |
507 | RHS: Builder.getIntN(N: BitWidth, C: 63)); |
508 | And29 = Builder.CreateAnd(LHS: Shr, RHS: Temp2, Name: "and29" ); |
509 | } else { |
510 | Value *Conv28 = Builder.CreateTrunc(V: Shr, DestTy: Builder.getIntNTy(N: 32)); |
511 | And29 = Builder.CreateAnd( |
512 | LHS: Conv28, RHS: ConstantInt::getSigned(Ty: Builder.getIntNTy(N: 32), V: 0x80000000)); |
513 | } |
514 | unsigned TempMod = FPMantissaWidth % 32; |
515 | Value *And34 = nullptr; |
516 | Value *Shl30 = nullptr; |
517 | if (FloatWidth > 80) { |
518 | TempMod += 32; |
519 | Value *Add = Builder.CreateShl(LHS: AAddr1Off32, RHS: Builder.getIntN(N: 64, C: TempMod)); |
520 | Shl30 = Builder.CreateAdd( |
521 | LHS: Add, |
522 | RHS: Builder.getIntN(N: 64, C: ((1ull << (62ull - TempMod)) - 1ull) << TempMod)); |
523 | And34 = Builder.CreateZExt(V: Shl30, DestTy: Builder.getIntNTy(N: 128)); |
524 | } else { |
525 | Value *Add = Builder.CreateShl(LHS: E0, RHS: Builder.getIntN(N: 32, C: TempMod)); |
526 | Shl30 = Builder.CreateAdd( |
527 | LHS: Add, RHS: Builder.getIntN(N: 32, C: ((1 << (30 - TempMod)) - 1) << TempMod)); |
528 | And34 = Builder.CreateAnd(LHS: FloatWidth > 32 ? AAddr1Off32 : AAddr1Off0, |
529 | RHS: Builder.getIntN(N: 32, C: (1 << TempMod) - 1)); |
530 | } |
531 | Value *Or35 = nullptr; |
532 | if (FloatWidth > 80) { |
533 | Value *And29Trunc = Builder.CreateTrunc(V: And29, DestTy: Builder.getIntNTy(N: 128)); |
534 | Value *Or31 = Builder.CreateOr(LHS: And29Trunc, RHS: And34); |
535 | Value *Or34 = Builder.CreateShl(LHS: Or31, RHS: Builder.getIntN(N: 128, C: 64)); |
536 | Value *Temp3 = Builder.CreateShl(LHS: Builder.getIntN(N: 128, C: 1), |
537 | RHS: Builder.getIntN(N: 128, C: FPMantissaWidth)); |
538 | Value *Temp4 = Builder.CreateSub(LHS: Temp3, RHS: Builder.getIntN(N: 128, C: 1)); |
539 | Value *A6 = Builder.CreateAnd(LHS: AAddr1Off0, RHS: Temp4); |
540 | Or35 = Builder.CreateOr(LHS: Or34, RHS: A6); |
541 | } else { |
542 | Value *Or31 = Builder.CreateOr(LHS: And34, RHS: And29); |
543 | Or35 = Builder.CreateOr(LHS: IsSigned ? Or31 : And34, RHS: Shl30); |
544 | } |
545 | Value *A4 = nullptr; |
546 | if (IToFP->getType()->isDoubleTy()) { |
547 | Value *ZExt1 = Builder.CreateZExt(V: Or35, DestTy: Builder.getIntNTy(N: FloatWidth)); |
548 | Value *Shl1 = Builder.CreateShl(LHS: ZExt1, RHS: Builder.getIntN(N: FloatWidth, C: 32)); |
549 | Value *And1 = |
550 | Builder.CreateAnd(LHS: AAddr1Off0, RHS: Builder.getIntN(N: FloatWidth, C: 0xFFFFFFFF)); |
551 | Value *Or1 = Builder.CreateOr(LHS: Shl1, RHS: And1); |
552 | A4 = Builder.CreateBitCast(V: Or1, DestTy: IToFP->getType()); |
553 | } else if (IToFP->getType()->isX86_FP80Ty()) { |
554 | Value *A40 = |
555 | Builder.CreateBitCast(V: Or35, DestTy: Type::getFP128Ty(C&: Builder.getContext())); |
556 | A4 = Builder.CreateFPTrunc(V: A40, DestTy: IToFP->getType()); |
557 | } else if (IToFP->getType()->isHalfTy() || IToFP->getType()->isBFloatTy()) { |
558 | // Deal with "half" situation. This is a workaround since we don't have |
559 | // floattihf.c currently as referring. |
560 | Value *A40 = |
561 | Builder.CreateBitCast(V: Or35, DestTy: Type::getFloatTy(C&: Builder.getContext())); |
562 | A4 = Builder.CreateFPTrunc(V: A40, DestTy: IToFP->getType()); |
563 | } else // float type |
564 | A4 = Builder.CreateBitCast(V: Or35, DestTy: IToFP->getType()); |
565 | Builder.CreateBr(Dest: End); |
566 | |
567 | // return: |
568 | Builder.SetInsertPoint(TheBB: End, IP: End->begin()); |
569 | PHINode *Retval0 = Builder.CreatePHI(Ty: IToFP->getType(), NumReservedValues: 2); |
570 | Retval0->addIncoming(V: A4, BB: IfEnd26); |
571 | Retval0->addIncoming(V: ConstantFP::getZero(Ty: IToFP->getType(), Negative: false), BB: Entry); |
572 | |
573 | IToFP->replaceAllUsesWith(V: Retval0); |
574 | IToFP->dropAllReferences(); |
575 | IToFP->eraseFromParent(); |
576 | } |
577 | |
578 | static void scalarize(Instruction *I, SmallVectorImpl<Instruction *> &Replace) { |
579 | VectorType *VTy = cast<FixedVectorType>(Val: I->getType()); |
580 | |
581 | IRBuilder<> Builder(I); |
582 | |
583 | unsigned NumElements = VTy->getElementCount().getFixedValue(); |
584 | Value *Result = PoisonValue::get(T: VTy); |
585 | for (unsigned Idx = 0; Idx < NumElements; ++Idx) { |
586 | Value *Ext = Builder.CreateExtractElement(Vec: I->getOperand(i: 0), Idx); |
587 | Value *Cast = Builder.CreateCast(Op: cast<CastInst>(Val: I)->getOpcode(), V: Ext, |
588 | DestTy: I->getType()->getScalarType()); |
589 | Result = Builder.CreateInsertElement(Vec: Result, NewElt: Cast, Idx); |
590 | if (isa<Instruction>(Val: Cast)) |
591 | Replace.push_back(Elt: cast<Instruction>(Val: Cast)); |
592 | } |
593 | I->replaceAllUsesWith(V: Result); |
594 | I->dropAllReferences(); |
595 | I->eraseFromParent(); |
596 | } |
597 | |
598 | static bool runImpl(Function &F, const TargetLowering &TLI) { |
599 | SmallVector<Instruction *, 4> Replace; |
600 | SmallVector<Instruction *, 4> ReplaceVector; |
601 | bool Modified = false; |
602 | |
603 | unsigned MaxLegalFpConvertBitWidth = |
604 | TLI.getMaxLargeFPConvertBitWidthSupported(); |
605 | if (ExpandFpConvertBits != llvm::IntegerType::MAX_INT_BITS) |
606 | MaxLegalFpConvertBitWidth = ExpandFpConvertBits; |
607 | |
608 | if (MaxLegalFpConvertBitWidth >= llvm::IntegerType::MAX_INT_BITS) |
609 | return false; |
610 | |
611 | for (auto &I : instructions(F)) { |
612 | switch (I.getOpcode()) { |
613 | case Instruction::FPToUI: |
614 | case Instruction::FPToSI: { |
615 | // TODO: This pass doesn't handle scalable vectors. |
616 | if (I.getOperand(i: 0)->getType()->isScalableTy()) |
617 | continue; |
618 | |
619 | auto *IntTy = cast<IntegerType>(Val: I.getType()->getScalarType()); |
620 | if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth) |
621 | continue; |
622 | |
623 | if (I.getOperand(i: 0)->getType()->isVectorTy()) |
624 | ReplaceVector.push_back(Elt: &I); |
625 | else |
626 | Replace.push_back(Elt: &I); |
627 | Modified = true; |
628 | break; |
629 | } |
630 | case Instruction::UIToFP: |
631 | case Instruction::SIToFP: { |
632 | // TODO: This pass doesn't handle scalable vectors. |
633 | if (I.getOperand(i: 0)->getType()->isScalableTy()) |
634 | continue; |
635 | |
636 | auto *IntTy = |
637 | cast<IntegerType>(Val: I.getOperand(i: 0)->getType()->getScalarType()); |
638 | if (IntTy->getIntegerBitWidth() <= MaxLegalFpConvertBitWidth) |
639 | continue; |
640 | |
641 | if (I.getOperand(i: 0)->getType()->isVectorTy()) |
642 | ReplaceVector.push_back(Elt: &I); |
643 | else |
644 | Replace.push_back(Elt: &I); |
645 | Modified = true; |
646 | break; |
647 | } |
648 | default: |
649 | break; |
650 | } |
651 | } |
652 | |
653 | while (!ReplaceVector.empty()) { |
654 | Instruction *I = ReplaceVector.pop_back_val(); |
655 | scalarize(I, Replace); |
656 | } |
657 | |
658 | if (Replace.empty()) |
659 | return false; |
660 | |
661 | while (!Replace.empty()) { |
662 | Instruction *I = Replace.pop_back_val(); |
663 | if (I->getOpcode() == Instruction::FPToUI || |
664 | I->getOpcode() == Instruction::FPToSI) { |
665 | expandFPToI(FPToI: I); |
666 | } else { |
667 | expandIToFP(IToFP: I); |
668 | } |
669 | } |
670 | |
671 | return Modified; |
672 | } |
673 | |
674 | namespace { |
675 | class ExpandFpLegacyPass : public FunctionPass { |
676 | public: |
677 | static char ID; |
678 | |
679 | ExpandFpLegacyPass() : FunctionPass(ID) { |
680 | initializeExpandFpLegacyPassPass(*PassRegistry::getPassRegistry()); |
681 | } |
682 | |
683 | bool runOnFunction(Function &F) override { |
684 | auto *TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); |
685 | auto *TLI = TM->getSubtargetImpl(F)->getTargetLowering(); |
686 | return runImpl(F, TLI: *TLI); |
687 | } |
688 | |
689 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
690 | AU.addRequired<TargetPassConfig>(); |
691 | AU.addPreserved<AAResultsWrapperPass>(); |
692 | AU.addPreserved<GlobalsAAWrapperPass>(); |
693 | } |
694 | }; |
695 | } // namespace |
696 | |
697 | PreservedAnalyses ExpandFpPass::run(Function &F, FunctionAnalysisManager &FAM) { |
698 | const TargetSubtargetInfo *STI = TM->getSubtargetImpl(F); |
699 | return runImpl(F, TLI: *STI->getTargetLowering()) ? PreservedAnalyses::none() |
700 | : PreservedAnalyses::all(); |
701 | } |
702 | |
703 | char ExpandFpLegacyPass::ID = 0; |
704 | INITIALIZE_PASS_BEGIN(ExpandFpLegacyPass, "expand-fp" , |
705 | "Expand certain fp instructions" , false, false) |
706 | INITIALIZE_PASS_END(ExpandFpLegacyPass, "expand-fp" , "Expand fp" , false, false) |
707 | |
708 | FunctionPass *llvm::createExpandFpPass() { return new ExpandFpLegacyPass(); } |
709 | |