1 | //===-- IntegerDivision.cpp - Expand integer division ---------------------===// |
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 | // This file contains an implementation of 32bit and 64bit scalar integer |
10 | // division for targets that don't have native support. It's largely derived |
11 | // from compiler-rt's implementations of __udivsi3 and __udivmoddi4, |
12 | // but hand-tuned for targets that prefer less control flow. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "llvm/Transforms/Utils/IntegerDivision.h" |
17 | #include "llvm/IR/Function.h" |
18 | #include "llvm/IR/IRBuilder.h" |
19 | #include "llvm/IR/Instructions.h" |
20 | #include "llvm/IR/Intrinsics.h" |
21 | |
22 | using namespace llvm; |
23 | |
24 | #define DEBUG_TYPE "integer-division" |
25 | |
26 | /// Generate code to compute the remainder of two signed integers. Returns the |
27 | /// remainder, which will have the sign of the dividend. Builder's insert point |
28 | /// should be pointing where the caller wants code generated, e.g. at the srem |
29 | /// instruction. This will generate a urem in the process, and Builder's insert |
30 | /// point will be pointing at the uren (if present, i.e. not folded), ready to |
31 | /// be expanded if the user wishes |
32 | static Value *generateSignedRemainderCode(Value *Dividend, Value *Divisor, |
33 | IRBuilder<> &Builder) { |
34 | unsigned BitWidth = Dividend->getType()->getIntegerBitWidth(); |
35 | ConstantInt *Shift = Builder.getIntN(N: BitWidth, C: BitWidth - 1); |
36 | |
37 | // Following instructions are generated for both i32 (shift 31) and |
38 | // i64 (shift 63). |
39 | |
40 | // ; %dividend_sgn = ashr i32 %dividend, 31 |
41 | // ; %divisor_sgn = ashr i32 %divisor, 31 |
42 | // ; %dvd_xor = xor i32 %dividend, %dividend_sgn |
43 | // ; %dvs_xor = xor i32 %divisor, %divisor_sgn |
44 | // ; %u_dividend = sub i32 %dvd_xor, %dividend_sgn |
45 | // ; %u_divisor = sub i32 %dvs_xor, %divisor_sgn |
46 | // ; %urem = urem i32 %dividend, %divisor |
47 | // ; %xored = xor i32 %urem, %dividend_sgn |
48 | // ; %srem = sub i32 %xored, %dividend_sgn |
49 | Dividend = Builder.CreateFreeze(V: Dividend); |
50 | Divisor = Builder.CreateFreeze(V: Divisor); |
51 | Value *DividendSign = Builder.CreateAShr(LHS: Dividend, RHS: Shift); |
52 | Value *DivisorSign = Builder.CreateAShr(LHS: Divisor, RHS: Shift); |
53 | Value *DvdXor = Builder.CreateXor(LHS: Dividend, RHS: DividendSign); |
54 | Value *DvsXor = Builder.CreateXor(LHS: Divisor, RHS: DivisorSign); |
55 | Value *UDividend = Builder.CreateSub(LHS: DvdXor, RHS: DividendSign); |
56 | Value *UDivisor = Builder.CreateSub(LHS: DvsXor, RHS: DivisorSign); |
57 | Value *URem = Builder.CreateURem(LHS: UDividend, RHS: UDivisor); |
58 | Value *Xored = Builder.CreateXor(LHS: URem, RHS: DividendSign); |
59 | Value *SRem = Builder.CreateSub(LHS: Xored, RHS: DividendSign); |
60 | |
61 | if (Instruction *URemInst = dyn_cast<Instruction>(Val: URem)) |
62 | Builder.SetInsertPoint(URemInst); |
63 | |
64 | return SRem; |
65 | } |
66 | |
67 | |
68 | /// Generate code to compute the remainder of two unsigned integers. Returns the |
69 | /// remainder. Builder's insert point should be pointing where the caller wants |
70 | /// code generated, e.g. at the urem instruction. This will generate a udiv in |
71 | /// the process, and Builder's insert point will be pointing at the udiv (if |
72 | /// present, i.e. not folded), ready to be expanded if the user wishes |
73 | static Value *generatedUnsignedRemainderCode(Value *Dividend, Value *Divisor, |
74 | IRBuilder<> &Builder) { |
75 | // Remainder = Dividend - Quotient*Divisor |
76 | |
77 | // Following instructions are generated for both i32 and i64 |
78 | |
79 | // ; %quotient = udiv i32 %dividend, %divisor |
80 | // ; %product = mul i32 %divisor, %quotient |
81 | // ; %remainder = sub i32 %dividend, %product |
82 | Dividend = Builder.CreateFreeze(V: Dividend); |
83 | Divisor = Builder.CreateFreeze(V: Divisor); |
84 | Value *Quotient = Builder.CreateUDiv(LHS: Dividend, RHS: Divisor); |
85 | Value *Product = Builder.CreateMul(LHS: Divisor, RHS: Quotient); |
86 | Value *Remainder = Builder.CreateSub(LHS: Dividend, RHS: Product); |
87 | |
88 | if (Instruction *UDiv = dyn_cast<Instruction>(Val: Quotient)) |
89 | Builder.SetInsertPoint(UDiv); |
90 | |
91 | return Remainder; |
92 | } |
93 | |
94 | /// Generate code to divide two signed integers. Returns the quotient, rounded |
95 | /// towards 0. Builder's insert point should be pointing where the caller wants |
96 | /// code generated, e.g. at the sdiv instruction. This will generate a udiv in |
97 | /// the process, and Builder's insert point will be pointing at the udiv (if |
98 | /// present, i.e. not folded), ready to be expanded if the user wishes. |
99 | static Value *generateSignedDivisionCode(Value *Dividend, Value *Divisor, |
100 | IRBuilder<> &Builder) { |
101 | // Implementation taken from compiler-rt's __divsi3 and __divdi3 |
102 | |
103 | unsigned BitWidth = Dividend->getType()->getIntegerBitWidth(); |
104 | ConstantInt *Shift = Builder.getIntN(N: BitWidth, C: BitWidth - 1); |
105 | |
106 | // Following instructions are generated for both i32 (shift 31) and |
107 | // i64 (shift 63). |
108 | |
109 | // ; %tmp = ashr i32 %dividend, 31 |
110 | // ; %tmp1 = ashr i32 %divisor, 31 |
111 | // ; %tmp2 = xor i32 %tmp, %dividend |
112 | // ; %u_dvnd = sub nsw i32 %tmp2, %tmp |
113 | // ; %tmp3 = xor i32 %tmp1, %divisor |
114 | // ; %u_dvsr = sub nsw i32 %tmp3, %tmp1 |
115 | // ; %q_sgn = xor i32 %tmp1, %tmp |
116 | // ; %q_mag = udiv i32 %u_dvnd, %u_dvsr |
117 | // ; %tmp4 = xor i32 %q_mag, %q_sgn |
118 | // ; %q = sub i32 %tmp4, %q_sgn |
119 | Dividend = Builder.CreateFreeze(V: Dividend); |
120 | Divisor = Builder.CreateFreeze(V: Divisor); |
121 | Value *Tmp = Builder.CreateAShr(LHS: Dividend, RHS: Shift); |
122 | Value *Tmp1 = Builder.CreateAShr(LHS: Divisor, RHS: Shift); |
123 | Value *Tmp2 = Builder.CreateXor(LHS: Tmp, RHS: Dividend); |
124 | Value *U_Dvnd = Builder.CreateSub(LHS: Tmp2, RHS: Tmp); |
125 | Value *Tmp3 = Builder.CreateXor(LHS: Tmp1, RHS: Divisor); |
126 | Value *U_Dvsr = Builder.CreateSub(LHS: Tmp3, RHS: Tmp1); |
127 | Value *Q_Sgn = Builder.CreateXor(LHS: Tmp1, RHS: Tmp); |
128 | Value *Q_Mag = Builder.CreateUDiv(LHS: U_Dvnd, RHS: U_Dvsr); |
129 | Value *Tmp4 = Builder.CreateXor(LHS: Q_Mag, RHS: Q_Sgn); |
130 | Value *Q = Builder.CreateSub(LHS: Tmp4, RHS: Q_Sgn); |
131 | |
132 | if (Instruction *UDiv = dyn_cast<Instruction>(Val: Q_Mag)) |
133 | Builder.SetInsertPoint(UDiv); |
134 | |
135 | return Q; |
136 | } |
137 | |
138 | /// Generates code to divide two unsigned scalar 32-bit or 64-bit integers. |
139 | /// Returns the quotient, rounded towards 0. Builder's insert point should |
140 | /// point where the caller wants code generated, e.g. at the udiv instruction. |
141 | static Value *generateUnsignedDivisionCode(Value *Dividend, Value *Divisor, |
142 | IRBuilder<> &Builder) { |
143 | // The basic algorithm can be found in the compiler-rt project's |
144 | // implementation of __udivsi3.c. Here, we do a lower-level IR based approach |
145 | // that's been hand-tuned to lessen the amount of control flow involved. |
146 | |
147 | // Some helper values |
148 | IntegerType *DivTy = cast<IntegerType>(Val: Dividend->getType()); |
149 | unsigned BitWidth = DivTy->getBitWidth(); |
150 | |
151 | ConstantInt *Zero = ConstantInt::get(Ty: DivTy, V: 0); |
152 | ConstantInt *One = ConstantInt::get(Ty: DivTy, V: 1); |
153 | ConstantInt *NegOne = ConstantInt::getSigned(Ty: DivTy, V: -1); |
154 | ConstantInt *MSB = ConstantInt::get(Ty: DivTy, V: BitWidth - 1); |
155 | |
156 | ConstantInt *True = Builder.getTrue(); |
157 | |
158 | BasicBlock *IBB = Builder.GetInsertBlock(); |
159 | Function *F = IBB->getParent(); |
160 | Function *CTLZ = Intrinsic::getDeclaration(M: F->getParent(), id: Intrinsic::ctlz, |
161 | Tys: DivTy); |
162 | |
163 | // Our CFG is going to look like: |
164 | // +---------------------+ |
165 | // | special-cases | |
166 | // | ... | |
167 | // +---------------------+ |
168 | // | | |
169 | // | +----------+ |
170 | // | | bb1 | |
171 | // | | ... | |
172 | // | +----------+ |
173 | // | | | |
174 | // | | +------------+ |
175 | // | | | preheader | |
176 | // | | | ... | |
177 | // | | +------------+ |
178 | // | | | |
179 | // | | | +---+ |
180 | // | | | | | |
181 | // | | +------------+ | |
182 | // | | | do-while | | |
183 | // | | | ... | | |
184 | // | | +------------+ | |
185 | // | | | | | |
186 | // | +-----------+ +---+ |
187 | // | | loop-exit | |
188 | // | | ... | |
189 | // | +-----------+ |
190 | // | | |
191 | // +-------+ |
192 | // | ... | |
193 | // | end | |
194 | // +-------+ |
195 | BasicBlock *SpecialCases = Builder.GetInsertBlock(); |
196 | SpecialCases->setName(Twine(SpecialCases->getName(), "_udiv-special-cases" )); |
197 | BasicBlock *End = SpecialCases->splitBasicBlock(I: Builder.GetInsertPoint(), |
198 | BBName: "udiv-end" ); |
199 | BasicBlock *LoopExit = BasicBlock::Create(Context&: Builder.getContext(), |
200 | Name: "udiv-loop-exit" , Parent: F, InsertBefore: End); |
201 | BasicBlock *DoWhile = BasicBlock::Create(Context&: Builder.getContext(), |
202 | Name: "udiv-do-while" , Parent: F, InsertBefore: End); |
203 | BasicBlock * = BasicBlock::Create(Context&: Builder.getContext(), |
204 | Name: "udiv-preheader" , Parent: F, InsertBefore: End); |
205 | BasicBlock *BB1 = BasicBlock::Create(Context&: Builder.getContext(), |
206 | Name: "udiv-bb1" , Parent: F, InsertBefore: End); |
207 | |
208 | // We'll be overwriting the terminator to insert our extra blocks |
209 | SpecialCases->getTerminator()->eraseFromParent(); |
210 | |
211 | // Same instructions are generated for both i32 (msb 31) and i64 (msb 63). |
212 | |
213 | // First off, check for special cases: dividend or divisor is zero, divisor |
214 | // is greater than dividend, and divisor is 1. |
215 | // ; special-cases: |
216 | // ; %ret0_1 = icmp eq i32 %divisor, 0 |
217 | // ; %ret0_2 = icmp eq i32 %dividend, 0 |
218 | // ; %ret0_3 = or i1 %ret0_1, %ret0_2 |
219 | // ; %tmp0 = tail call i32 @llvm.ctlz.i32(i32 %divisor, i1 true) |
220 | // ; %tmp1 = tail call i32 @llvm.ctlz.i32(i32 %dividend, i1 true) |
221 | // ; %sr = sub nsw i32 %tmp0, %tmp1 |
222 | // ; %ret0_4 = icmp ugt i32 %sr, 31 |
223 | // ; %ret0 = select i1 %ret0_3, i1 true, i1 %ret0_4 |
224 | // ; %retDividend = icmp eq i32 %sr, 31 |
225 | // ; %retVal = select i1 %ret0, i32 0, i32 %dividend |
226 | // ; %earlyRet = select i1 %ret0, i1 true, %retDividend |
227 | // ; br i1 %earlyRet, label %end, label %bb1 |
228 | Builder.SetInsertPoint(SpecialCases); |
229 | Divisor = Builder.CreateFreeze(V: Divisor); |
230 | Dividend = Builder.CreateFreeze(V: Dividend); |
231 | Value *Ret0_1 = Builder.CreateICmpEQ(LHS: Divisor, RHS: Zero); |
232 | Value *Ret0_2 = Builder.CreateICmpEQ(LHS: Dividend, RHS: Zero); |
233 | Value *Ret0_3 = Builder.CreateOr(LHS: Ret0_1, RHS: Ret0_2); |
234 | Value *Tmp0 = Builder.CreateCall(Callee: CTLZ, Args: {Divisor, True}); |
235 | Value *Tmp1 = Builder.CreateCall(Callee: CTLZ, Args: {Dividend, True}); |
236 | Value *SR = Builder.CreateSub(LHS: Tmp0, RHS: Tmp1); |
237 | Value *Ret0_4 = Builder.CreateICmpUGT(LHS: SR, RHS: MSB); |
238 | Value *Ret0 = Builder.CreateLogicalOr(Cond1: Ret0_3, Cond2: Ret0_4); |
239 | Value *RetDividend = Builder.CreateICmpEQ(LHS: SR, RHS: MSB); |
240 | Value *RetVal = Builder.CreateSelect(C: Ret0, True: Zero, False: Dividend); |
241 | Value *EarlyRet = Builder.CreateLogicalOr(Cond1: Ret0, Cond2: RetDividend); |
242 | Builder.CreateCondBr(Cond: EarlyRet, True: End, False: BB1); |
243 | |
244 | // ; bb1: ; preds = %special-cases |
245 | // ; %sr_1 = add i32 %sr, 1 |
246 | // ; %tmp2 = sub i32 31, %sr |
247 | // ; %q = shl i32 %dividend, %tmp2 |
248 | // ; %skipLoop = icmp eq i32 %sr_1, 0 |
249 | // ; br i1 %skipLoop, label %loop-exit, label %preheader |
250 | Builder.SetInsertPoint(BB1); |
251 | Value *SR_1 = Builder.CreateAdd(LHS: SR, RHS: One); |
252 | Value *Tmp2 = Builder.CreateSub(LHS: MSB, RHS: SR); |
253 | Value *Q = Builder.CreateShl(LHS: Dividend, RHS: Tmp2); |
254 | Value *SkipLoop = Builder.CreateICmpEQ(LHS: SR_1, RHS: Zero); |
255 | Builder.CreateCondBr(Cond: SkipLoop, True: LoopExit, False: Preheader); |
256 | |
257 | // ; preheader: ; preds = %bb1 |
258 | // ; %tmp3 = lshr i32 %dividend, %sr_1 |
259 | // ; %tmp4 = add i32 %divisor, -1 |
260 | // ; br label %do-while |
261 | Builder.SetInsertPoint(Preheader); |
262 | Value *Tmp3 = Builder.CreateLShr(LHS: Dividend, RHS: SR_1); |
263 | Value *Tmp4 = Builder.CreateAdd(LHS: Divisor, RHS: NegOne); |
264 | Builder.CreateBr(Dest: DoWhile); |
265 | |
266 | // ; do-while: ; preds = %do-while, %preheader |
267 | // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ] |
268 | // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ] |
269 | // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ] |
270 | // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ] |
271 | // ; %tmp5 = shl i32 %r_1, 1 |
272 | // ; %tmp6 = lshr i32 %q_2, 31 |
273 | // ; %tmp7 = or i32 %tmp5, %tmp6 |
274 | // ; %tmp8 = shl i32 %q_2, 1 |
275 | // ; %q_1 = or i32 %carry_1, %tmp8 |
276 | // ; %tmp9 = sub i32 %tmp4, %tmp7 |
277 | // ; %tmp10 = ashr i32 %tmp9, 31 |
278 | // ; %carry = and i32 %tmp10, 1 |
279 | // ; %tmp11 = and i32 %tmp10, %divisor |
280 | // ; %r = sub i32 %tmp7, %tmp11 |
281 | // ; %sr_2 = add i32 %sr_3, -1 |
282 | // ; %tmp12 = icmp eq i32 %sr_2, 0 |
283 | // ; br i1 %tmp12, label %loop-exit, label %do-while |
284 | Builder.SetInsertPoint(DoWhile); |
285 | PHINode *Carry_1 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
286 | PHINode *SR_3 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
287 | PHINode *R_1 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
288 | PHINode *Q_2 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
289 | Value *Tmp5 = Builder.CreateShl(LHS: R_1, RHS: One); |
290 | Value *Tmp6 = Builder.CreateLShr(LHS: Q_2, RHS: MSB); |
291 | Value *Tmp7 = Builder.CreateOr(LHS: Tmp5, RHS: Tmp6); |
292 | Value *Tmp8 = Builder.CreateShl(LHS: Q_2, RHS: One); |
293 | Value *Q_1 = Builder.CreateOr(LHS: Carry_1, RHS: Tmp8); |
294 | Value *Tmp9 = Builder.CreateSub(LHS: Tmp4, RHS: Tmp7); |
295 | Value *Tmp10 = Builder.CreateAShr(LHS: Tmp9, RHS: MSB); |
296 | Value *Carry = Builder.CreateAnd(LHS: Tmp10, RHS: One); |
297 | Value *Tmp11 = Builder.CreateAnd(LHS: Tmp10, RHS: Divisor); |
298 | Value *R = Builder.CreateSub(LHS: Tmp7, RHS: Tmp11); |
299 | Value *SR_2 = Builder.CreateAdd(LHS: SR_3, RHS: NegOne); |
300 | Value *Tmp12 = Builder.CreateICmpEQ(LHS: SR_2, RHS: Zero); |
301 | Builder.CreateCondBr(Cond: Tmp12, True: LoopExit, False: DoWhile); |
302 | |
303 | // ; loop-exit: ; preds = %do-while, %bb1 |
304 | // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ] |
305 | // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ] |
306 | // ; %tmp13 = shl i32 %q_3, 1 |
307 | // ; %q_4 = or i32 %carry_2, %tmp13 |
308 | // ; br label %end |
309 | Builder.SetInsertPoint(LoopExit); |
310 | PHINode *Carry_2 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
311 | PHINode *Q_3 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
312 | Value *Tmp13 = Builder.CreateShl(LHS: Q_3, RHS: One); |
313 | Value *Q_4 = Builder.CreateOr(LHS: Carry_2, RHS: Tmp13); |
314 | Builder.CreateBr(Dest: End); |
315 | |
316 | // ; end: ; preds = %loop-exit, %special-cases |
317 | // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ] |
318 | // ; ret i32 %q_5 |
319 | Builder.SetInsertPoint(TheBB: End, IP: End->begin()); |
320 | PHINode *Q_5 = Builder.CreatePHI(Ty: DivTy, NumReservedValues: 2); |
321 | |
322 | // Populate the Phis, since all values have now been created. Our Phis were: |
323 | // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ] |
324 | Carry_1->addIncoming(V: Zero, BB: Preheader); |
325 | Carry_1->addIncoming(V: Carry, BB: DoWhile); |
326 | // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ] |
327 | SR_3->addIncoming(V: SR_1, BB: Preheader); |
328 | SR_3->addIncoming(V: SR_2, BB: DoWhile); |
329 | // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ] |
330 | R_1->addIncoming(V: Tmp3, BB: Preheader); |
331 | R_1->addIncoming(V: R, BB: DoWhile); |
332 | // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ] |
333 | Q_2->addIncoming(V: Q, BB: Preheader); |
334 | Q_2->addIncoming(V: Q_1, BB: DoWhile); |
335 | // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ] |
336 | Carry_2->addIncoming(V: Zero, BB: BB1); |
337 | Carry_2->addIncoming(V: Carry, BB: DoWhile); |
338 | // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ] |
339 | Q_3->addIncoming(V: Q, BB: BB1); |
340 | Q_3->addIncoming(V: Q_1, BB: DoWhile); |
341 | // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ] |
342 | Q_5->addIncoming(V: Q_4, BB: LoopExit); |
343 | Q_5->addIncoming(V: RetVal, BB: SpecialCases); |
344 | |
345 | return Q_5; |
346 | } |
347 | |
348 | /// Generate code to calculate the remainder of two integers, replacing Rem with |
349 | /// the generated code. This currently generates code using the udiv expansion, |
350 | /// but future work includes generating more specialized code, e.g. when more |
351 | /// information about the operands are known. |
352 | /// |
353 | /// Replace Rem with generated code. |
354 | bool llvm::expandRemainder(BinaryOperator *Rem) { |
355 | assert((Rem->getOpcode() == Instruction::SRem || |
356 | Rem->getOpcode() == Instruction::URem) && |
357 | "Trying to expand remainder from a non-remainder function" ); |
358 | |
359 | IRBuilder<> Builder(Rem); |
360 | |
361 | assert(!Rem->getType()->isVectorTy() && "Div over vectors not supported" ); |
362 | |
363 | // First prepare the sign if it's a signed remainder |
364 | if (Rem->getOpcode() == Instruction::SRem) { |
365 | Value *Remainder = generateSignedRemainderCode(Dividend: Rem->getOperand(i_nocapture: 0), |
366 | Divisor: Rem->getOperand(i_nocapture: 1), Builder); |
367 | |
368 | // Check whether this is the insert point while Rem is still valid. |
369 | bool IsInsertPoint = Rem->getIterator() == Builder.GetInsertPoint(); |
370 | Rem->replaceAllUsesWith(V: Remainder); |
371 | Rem->dropAllReferences(); |
372 | Rem->eraseFromParent(); |
373 | |
374 | // If we didn't actually generate an urem instruction, we're done |
375 | // This happens for example if the input were constant. In this case the |
376 | // Builder insertion point was unchanged |
377 | if (IsInsertPoint) |
378 | return true; |
379 | |
380 | BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: Builder.GetInsertPoint()); |
381 | Rem = BO; |
382 | } |
383 | |
384 | Value *Remainder = generatedUnsignedRemainderCode(Dividend: Rem->getOperand(i_nocapture: 0), |
385 | Divisor: Rem->getOperand(i_nocapture: 1), |
386 | Builder); |
387 | |
388 | Rem->replaceAllUsesWith(V: Remainder); |
389 | Rem->dropAllReferences(); |
390 | Rem->eraseFromParent(); |
391 | |
392 | // Expand the udiv |
393 | if (BinaryOperator *UDiv = dyn_cast<BinaryOperator>(Val: Builder.GetInsertPoint())) { |
394 | assert(UDiv->getOpcode() == Instruction::UDiv && "Non-udiv in expansion?" ); |
395 | expandDivision(Div: UDiv); |
396 | } |
397 | |
398 | return true; |
399 | } |
400 | |
401 | /// Generate code to divide two integers, replacing Div with the generated |
402 | /// code. This currently generates code similarly to compiler-rt's |
403 | /// implementations, but future work includes generating more specialized code |
404 | /// when more information about the operands are known. |
405 | /// |
406 | /// Replace Div with generated code. |
407 | bool llvm::expandDivision(BinaryOperator *Div) { |
408 | assert((Div->getOpcode() == Instruction::SDiv || |
409 | Div->getOpcode() == Instruction::UDiv) && |
410 | "Trying to expand division from a non-division function" ); |
411 | |
412 | IRBuilder<> Builder(Div); |
413 | |
414 | assert(!Div->getType()->isVectorTy() && "Div over vectors not supported" ); |
415 | |
416 | // First prepare the sign if it's a signed division |
417 | if (Div->getOpcode() == Instruction::SDiv) { |
418 | // Lower the code to unsigned division, and reset Div to point to the udiv. |
419 | Value *Quotient = generateSignedDivisionCode(Dividend: Div->getOperand(i_nocapture: 0), |
420 | Divisor: Div->getOperand(i_nocapture: 1), Builder); |
421 | |
422 | // Check whether this is the insert point while Div is still valid. |
423 | bool IsInsertPoint = Div->getIterator() == Builder.GetInsertPoint(); |
424 | Div->replaceAllUsesWith(V: Quotient); |
425 | Div->dropAllReferences(); |
426 | Div->eraseFromParent(); |
427 | |
428 | // If we didn't actually generate an udiv instruction, we're done |
429 | // This happens for example if the input were constant. In this case the |
430 | // Builder insertion point was unchanged |
431 | if (IsInsertPoint) |
432 | return true; |
433 | |
434 | BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: Builder.GetInsertPoint()); |
435 | Div = BO; |
436 | } |
437 | |
438 | // Insert the unsigned division code |
439 | Value *Quotient = generateUnsignedDivisionCode(Dividend: Div->getOperand(i_nocapture: 0), |
440 | Divisor: Div->getOperand(i_nocapture: 1), |
441 | Builder); |
442 | Div->replaceAllUsesWith(V: Quotient); |
443 | Div->dropAllReferences(); |
444 | Div->eraseFromParent(); |
445 | |
446 | return true; |
447 | } |
448 | |
449 | /// Generate code to compute the remainder of two integers of bitwidth up to |
450 | /// 32 bits. Uses the above routines and extends the inputs/truncates the |
451 | /// outputs to operate in 32 bits; that is, these routines are good for targets |
452 | /// that have no or very little suppport for smaller than 32 bit integer |
453 | /// arithmetic. |
454 | /// |
455 | /// Replace Rem with emulation code. |
456 | bool llvm::expandRemainderUpTo32Bits(BinaryOperator *Rem) { |
457 | assert((Rem->getOpcode() == Instruction::SRem || |
458 | Rem->getOpcode() == Instruction::URem) && |
459 | "Trying to expand remainder from a non-remainder function" ); |
460 | |
461 | Type *RemTy = Rem->getType(); |
462 | assert(!RemTy->isVectorTy() && "Div over vectors not supported" ); |
463 | |
464 | unsigned RemTyBitWidth = RemTy->getIntegerBitWidth(); |
465 | |
466 | assert(RemTyBitWidth <= 32 && |
467 | "Div of bitwidth greater than 32 not supported" ); |
468 | |
469 | if (RemTyBitWidth == 32) |
470 | return expandRemainder(Rem); |
471 | |
472 | // If bitwidth smaller than 32 extend inputs, extend output and proceed |
473 | // with 32 bit division. |
474 | IRBuilder<> Builder(Rem); |
475 | |
476 | Value *ExtDividend; |
477 | Value *ExtDivisor; |
478 | Value *ExtRem; |
479 | Value *Trunc; |
480 | Type *Int32Ty = Builder.getInt32Ty(); |
481 | |
482 | if (Rem->getOpcode() == Instruction::SRem) { |
483 | ExtDividend = Builder.CreateSExt(V: Rem->getOperand(i_nocapture: 0), DestTy: Int32Ty); |
484 | ExtDivisor = Builder.CreateSExt(V: Rem->getOperand(i_nocapture: 1), DestTy: Int32Ty); |
485 | ExtRem = Builder.CreateSRem(LHS: ExtDividend, RHS: ExtDivisor); |
486 | } else { |
487 | ExtDividend = Builder.CreateZExt(V: Rem->getOperand(i_nocapture: 0), DestTy: Int32Ty); |
488 | ExtDivisor = Builder.CreateZExt(V: Rem->getOperand(i_nocapture: 1), DestTy: Int32Ty); |
489 | ExtRem = Builder.CreateURem(LHS: ExtDividend, RHS: ExtDivisor); |
490 | } |
491 | Trunc = Builder.CreateTrunc(V: ExtRem, DestTy: RemTy); |
492 | |
493 | Rem->replaceAllUsesWith(V: Trunc); |
494 | Rem->dropAllReferences(); |
495 | Rem->eraseFromParent(); |
496 | |
497 | return expandRemainder(Rem: cast<BinaryOperator>(Val: ExtRem)); |
498 | } |
499 | |
500 | /// Generate code to compute the remainder of two integers of bitwidth up to |
501 | /// 64 bits. Uses the above routines and extends the inputs/truncates the |
502 | /// outputs to operate in 64 bits. |
503 | /// |
504 | /// Replace Rem with emulation code. |
505 | bool llvm::expandRemainderUpTo64Bits(BinaryOperator *Rem) { |
506 | assert((Rem->getOpcode() == Instruction::SRem || |
507 | Rem->getOpcode() == Instruction::URem) && |
508 | "Trying to expand remainder from a non-remainder function" ); |
509 | |
510 | Type *RemTy = Rem->getType(); |
511 | assert(!RemTy->isVectorTy() && "Div over vectors not supported" ); |
512 | |
513 | unsigned RemTyBitWidth = RemTy->getIntegerBitWidth(); |
514 | |
515 | if (RemTyBitWidth >= 64) |
516 | return expandRemainder(Rem); |
517 | |
518 | // If bitwidth smaller than 64 extend inputs, extend output and proceed |
519 | // with 64 bit division. |
520 | IRBuilder<> Builder(Rem); |
521 | |
522 | Value *ExtDividend; |
523 | Value *ExtDivisor; |
524 | Value *ExtRem; |
525 | Value *Trunc; |
526 | Type *Int64Ty = Builder.getInt64Ty(); |
527 | |
528 | if (Rem->getOpcode() == Instruction::SRem) { |
529 | ExtDividend = Builder.CreateSExt(V: Rem->getOperand(i_nocapture: 0), DestTy: Int64Ty); |
530 | ExtDivisor = Builder.CreateSExt(V: Rem->getOperand(i_nocapture: 1), DestTy: Int64Ty); |
531 | ExtRem = Builder.CreateSRem(LHS: ExtDividend, RHS: ExtDivisor); |
532 | } else { |
533 | ExtDividend = Builder.CreateZExt(V: Rem->getOperand(i_nocapture: 0), DestTy: Int64Ty); |
534 | ExtDivisor = Builder.CreateZExt(V: Rem->getOperand(i_nocapture: 1), DestTy: Int64Ty); |
535 | ExtRem = Builder.CreateURem(LHS: ExtDividend, RHS: ExtDivisor); |
536 | } |
537 | Trunc = Builder.CreateTrunc(V: ExtRem, DestTy: RemTy); |
538 | |
539 | Rem->replaceAllUsesWith(V: Trunc); |
540 | Rem->dropAllReferences(); |
541 | Rem->eraseFromParent(); |
542 | |
543 | return expandRemainder(Rem: cast<BinaryOperator>(Val: ExtRem)); |
544 | } |
545 | |
546 | /// Generate code to divide two integers of bitwidth up to 32 bits. Uses the |
547 | /// above routines and extends the inputs/truncates the outputs to operate |
548 | /// in 32 bits; that is, these routines are good for targets that have no |
549 | /// or very little support for smaller than 32 bit integer arithmetic. |
550 | /// |
551 | /// Replace Div with emulation code. |
552 | bool llvm::expandDivisionUpTo32Bits(BinaryOperator *Div) { |
553 | assert((Div->getOpcode() == Instruction::SDiv || |
554 | Div->getOpcode() == Instruction::UDiv) && |
555 | "Trying to expand division from a non-division function" ); |
556 | |
557 | Type *DivTy = Div->getType(); |
558 | assert(!DivTy->isVectorTy() && "Div over vectors not supported" ); |
559 | |
560 | unsigned DivTyBitWidth = DivTy->getIntegerBitWidth(); |
561 | |
562 | assert(DivTyBitWidth <= 32 && "Div of bitwidth greater than 32 not supported" ); |
563 | |
564 | if (DivTyBitWidth == 32) |
565 | return expandDivision(Div); |
566 | |
567 | // If bitwidth smaller than 32 extend inputs, extend output and proceed |
568 | // with 32 bit division. |
569 | IRBuilder<> Builder(Div); |
570 | |
571 | Value *ExtDividend; |
572 | Value *ExtDivisor; |
573 | Value *ExtDiv; |
574 | Value *Trunc; |
575 | Type *Int32Ty = Builder.getInt32Ty(); |
576 | |
577 | if (Div->getOpcode() == Instruction::SDiv) { |
578 | ExtDividend = Builder.CreateSExt(V: Div->getOperand(i_nocapture: 0), DestTy: Int32Ty); |
579 | ExtDivisor = Builder.CreateSExt(V: Div->getOperand(i_nocapture: 1), DestTy: Int32Ty); |
580 | ExtDiv = Builder.CreateSDiv(LHS: ExtDividend, RHS: ExtDivisor); |
581 | } else { |
582 | ExtDividend = Builder.CreateZExt(V: Div->getOperand(i_nocapture: 0), DestTy: Int32Ty); |
583 | ExtDivisor = Builder.CreateZExt(V: Div->getOperand(i_nocapture: 1), DestTy: Int32Ty); |
584 | ExtDiv = Builder.CreateUDiv(LHS: ExtDividend, RHS: ExtDivisor); |
585 | } |
586 | Trunc = Builder.CreateTrunc(V: ExtDiv, DestTy: DivTy); |
587 | |
588 | Div->replaceAllUsesWith(V: Trunc); |
589 | Div->dropAllReferences(); |
590 | Div->eraseFromParent(); |
591 | |
592 | return expandDivision(Div: cast<BinaryOperator>(Val: ExtDiv)); |
593 | } |
594 | |
595 | /// Generate code to divide two integers of bitwidth up to 64 bits. Uses the |
596 | /// above routines and extends the inputs/truncates the outputs to operate |
597 | /// in 64 bits. |
598 | /// |
599 | /// Replace Div with emulation code. |
600 | bool llvm::expandDivisionUpTo64Bits(BinaryOperator *Div) { |
601 | assert((Div->getOpcode() == Instruction::SDiv || |
602 | Div->getOpcode() == Instruction::UDiv) && |
603 | "Trying to expand division from a non-division function" ); |
604 | |
605 | Type *DivTy = Div->getType(); |
606 | assert(!DivTy->isVectorTy() && "Div over vectors not supported" ); |
607 | |
608 | unsigned DivTyBitWidth = DivTy->getIntegerBitWidth(); |
609 | |
610 | if (DivTyBitWidth >= 64) |
611 | return expandDivision(Div); |
612 | |
613 | // If bitwidth smaller than 64 extend inputs, extend output and proceed |
614 | // with 64 bit division. |
615 | IRBuilder<> Builder(Div); |
616 | |
617 | Value *ExtDividend; |
618 | Value *ExtDivisor; |
619 | Value *ExtDiv; |
620 | Value *Trunc; |
621 | Type *Int64Ty = Builder.getInt64Ty(); |
622 | |
623 | if (Div->getOpcode() == Instruction::SDiv) { |
624 | ExtDividend = Builder.CreateSExt(V: Div->getOperand(i_nocapture: 0), DestTy: Int64Ty); |
625 | ExtDivisor = Builder.CreateSExt(V: Div->getOperand(i_nocapture: 1), DestTy: Int64Ty); |
626 | ExtDiv = Builder.CreateSDiv(LHS: ExtDividend, RHS: ExtDivisor); |
627 | } else { |
628 | ExtDividend = Builder.CreateZExt(V: Div->getOperand(i_nocapture: 0), DestTy: Int64Ty); |
629 | ExtDivisor = Builder.CreateZExt(V: Div->getOperand(i_nocapture: 1), DestTy: Int64Ty); |
630 | ExtDiv = Builder.CreateUDiv(LHS: ExtDividend, RHS: ExtDivisor); |
631 | } |
632 | Trunc = Builder.CreateTrunc(V: ExtDiv, DestTy: DivTy); |
633 | |
634 | Div->replaceAllUsesWith(V: Trunc); |
635 | Div->dropAllReferences(); |
636 | Div->eraseFromParent(); |
637 | |
638 | return expandDivision(Div: cast<BinaryOperator>(Val: ExtDiv)); |
639 | } |
640 | |