LLVM  6.0.0svn
IntegerDivision.cpp
Go to the documentation of this file.
1 //===-- IntegerDivision.cpp - Expand integer division ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file contains an implementation of 32bit and 64bit scalar integer
11 // division for targets that don't have native support. It's largely derived
12 // from compiler-rt's implementations of __udivsi3 and __udivmoddi4,
13 // but hand-tuned for targets that prefer less control flow.
14 //
15 //===----------------------------------------------------------------------===//
16 
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/IRBuilder.h"
20 #include "llvm/IR/Instructions.h"
21 #include "llvm/IR/Intrinsics.h"
22 #include <utility>
23 
24 using namespace llvm;
25 
26 #define DEBUG_TYPE "integer-division"
27 
28 /// Generate code to compute the remainder of two signed integers. Returns the
29 /// remainder, which will have the sign of the dividend. Builder's insert point
30 /// should be pointing where the caller wants code generated, e.g. at the srem
31 /// instruction. This will generate a urem in the process, and Builder's insert
32 /// point will be pointing at the uren (if present, i.e. not folded), ready to
33 /// be expanded if the user wishes
34 static Value *generateSignedRemainderCode(Value *Dividend, Value *Divisor,
35  IRBuilder<> &Builder) {
36  unsigned BitWidth = Dividend->getType()->getIntegerBitWidth();
37  ConstantInt *Shift;
38 
39  if (BitWidth == 64) {
40  Shift = Builder.getInt64(63);
41  } else {
42  assert(BitWidth == 32 && "Unexpected bit width");
43  Shift = Builder.getInt32(31);
44  }
45 
46  // Following instructions are generated for both i32 (shift 31) and
47  // i64 (shift 63).
48 
49  // ; %dividend_sgn = ashr i32 %dividend, 31
50  // ; %divisor_sgn = ashr i32 %divisor, 31
51  // ; %dvd_xor = xor i32 %dividend, %dividend_sgn
52  // ; %dvs_xor = xor i32 %divisor, %divisor_sgn
53  // ; %u_dividend = sub i32 %dvd_xor, %dividend_sgn
54  // ; %u_divisor = sub i32 %dvs_xor, %divisor_sgn
55  // ; %urem = urem i32 %dividend, %divisor
56  // ; %xored = xor i32 %urem, %dividend_sgn
57  // ; %srem = sub i32 %xored, %dividend_sgn
58  Value *DividendSign = Builder.CreateAShr(Dividend, Shift);
59  Value *DivisorSign = Builder.CreateAShr(Divisor, Shift);
60  Value *DvdXor = Builder.CreateXor(Dividend, DividendSign);
61  Value *DvsXor = Builder.CreateXor(Divisor, DivisorSign);
62  Value *UDividend = Builder.CreateSub(DvdXor, DividendSign);
63  Value *UDivisor = Builder.CreateSub(DvsXor, DivisorSign);
64  Value *URem = Builder.CreateURem(UDividend, UDivisor);
65  Value *Xored = Builder.CreateXor(URem, DividendSign);
66  Value *SRem = Builder.CreateSub(Xored, DividendSign);
67 
68  if (Instruction *URemInst = dyn_cast<Instruction>(URem))
69  Builder.SetInsertPoint(URemInst);
70 
71  return SRem;
72 }
73 
74 
75 /// Generate code to compute the remainder of two unsigned integers. Returns the
76 /// remainder. Builder's insert point should be pointing where the caller wants
77 /// code generated, e.g. at the urem instruction. This will generate a udiv in
78 /// the process, and Builder's insert point will be pointing at the udiv (if
79 /// present, i.e. not folded), ready to be expanded if the user wishes
80 static Value *generatedUnsignedRemainderCode(Value *Dividend, Value *Divisor,
81  IRBuilder<> &Builder) {
82  // Remainder = Dividend - Quotient*Divisor
83 
84  // Following instructions are generated for both i32 and i64
85 
86  // ; %quotient = udiv i32 %dividend, %divisor
87  // ; %product = mul i32 %divisor, %quotient
88  // ; %remainder = sub i32 %dividend, %product
89  Value *Quotient = Builder.CreateUDiv(Dividend, Divisor);
90  Value *Product = Builder.CreateMul(Divisor, Quotient);
91  Value *Remainder = Builder.CreateSub(Dividend, Product);
92 
93  if (Instruction *UDiv = dyn_cast<Instruction>(Quotient))
94  Builder.SetInsertPoint(UDiv);
95 
96  return Remainder;
97 }
98 
99 /// Generate code to divide two signed integers. Returns the quotient, rounded
100 /// towards 0. Builder's insert point should be pointing where the caller wants
101 /// code generated, e.g. at the sdiv instruction. This will generate a udiv in
102 /// the process, and Builder's insert point will be pointing at the udiv (if
103 /// present, i.e. not folded), ready to be expanded if the user wishes.
104 static Value *generateSignedDivisionCode(Value *Dividend, Value *Divisor,
105  IRBuilder<> &Builder) {
106  // Implementation taken from compiler-rt's __divsi3 and __divdi3
107 
108  unsigned BitWidth = Dividend->getType()->getIntegerBitWidth();
109  ConstantInt *Shift;
110 
111  if (BitWidth == 64) {
112  Shift = Builder.getInt64(63);
113  } else {
114  assert(BitWidth == 32 && "Unexpected bit width");
115  Shift = Builder.getInt32(31);
116  }
117 
118  // Following instructions are generated for both i32 (shift 31) and
119  // i64 (shift 63).
120 
121  // ; %tmp = ashr i32 %dividend, 31
122  // ; %tmp1 = ashr i32 %divisor, 31
123  // ; %tmp2 = xor i32 %tmp, %dividend
124  // ; %u_dvnd = sub nsw i32 %tmp2, %tmp
125  // ; %tmp3 = xor i32 %tmp1, %divisor
126  // ; %u_dvsr = sub nsw i32 %tmp3, %tmp1
127  // ; %q_sgn = xor i32 %tmp1, %tmp
128  // ; %q_mag = udiv i32 %u_dvnd, %u_dvsr
129  // ; %tmp4 = xor i32 %q_mag, %q_sgn
130  // ; %q = sub i32 %tmp4, %q_sgn
131  Value *Tmp = Builder.CreateAShr(Dividend, Shift);
132  Value *Tmp1 = Builder.CreateAShr(Divisor, Shift);
133  Value *Tmp2 = Builder.CreateXor(Tmp, Dividend);
134  Value *U_Dvnd = Builder.CreateSub(Tmp2, Tmp);
135  Value *Tmp3 = Builder.CreateXor(Tmp1, Divisor);
136  Value *U_Dvsr = Builder.CreateSub(Tmp3, Tmp1);
137  Value *Q_Sgn = Builder.CreateXor(Tmp1, Tmp);
138  Value *Q_Mag = Builder.CreateUDiv(U_Dvnd, U_Dvsr);
139  Value *Tmp4 = Builder.CreateXor(Q_Mag, Q_Sgn);
140  Value *Q = Builder.CreateSub(Tmp4, Q_Sgn);
141 
142  if (Instruction *UDiv = dyn_cast<Instruction>(Q_Mag))
143  Builder.SetInsertPoint(UDiv);
144 
145  return Q;
146 }
147 
148 /// Generates code to divide two unsigned scalar 32-bit or 64-bit integers.
149 /// Returns the quotient, rounded towards 0. Builder's insert point should
150 /// point where the caller wants code generated, e.g. at the udiv instruction.
151 static Value *generateUnsignedDivisionCode(Value *Dividend, Value *Divisor,
152  IRBuilder<> &Builder) {
153  // The basic algorithm can be found in the compiler-rt project's
154  // implementation of __udivsi3.c. Here, we do a lower-level IR based approach
155  // that's been hand-tuned to lessen the amount of control flow involved.
156 
157  // Some helper values
158  IntegerType *DivTy = cast<IntegerType>(Dividend->getType());
159  unsigned BitWidth = DivTy->getBitWidth();
160 
161  ConstantInt *Zero;
162  ConstantInt *One;
163  ConstantInt *NegOne;
164  ConstantInt *MSB;
165 
166  if (BitWidth == 64) {
167  Zero = Builder.getInt64(0);
168  One = Builder.getInt64(1);
169  NegOne = ConstantInt::getSigned(DivTy, -1);
170  MSB = Builder.getInt64(63);
171  } else {
172  assert(BitWidth == 32 && "Unexpected bit width");
173  Zero = Builder.getInt32(0);
174  One = Builder.getInt32(1);
175  NegOne = ConstantInt::getSigned(DivTy, -1);
176  MSB = Builder.getInt32(31);
177  }
178 
179  ConstantInt *True = Builder.getTrue();
180 
181  BasicBlock *IBB = Builder.GetInsertBlock();
182  Function *F = IBB->getParent();
183  Function *CTLZ = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
184  DivTy);
185 
186  // Our CFG is going to look like:
187  // +---------------------+
188  // | special-cases |
189  // | ... |
190  // +---------------------+
191  // | |
192  // | +----------+
193  // | | bb1 |
194  // | | ... |
195  // | +----------+
196  // | | |
197  // | | +------------+
198  // | | | preheader |
199  // | | | ... |
200  // | | +------------+
201  // | | |
202  // | | | +---+
203  // | | | | |
204  // | | +------------+ |
205  // | | | do-while | |
206  // | | | ... | |
207  // | | +------------+ |
208  // | | | | |
209  // | +-----------+ +---+
210  // | | loop-exit |
211  // | | ... |
212  // | +-----------+
213  // | |
214  // +-------+
215  // | ... |
216  // | end |
217  // +-------+
218  BasicBlock *SpecialCases = Builder.GetInsertBlock();
219  SpecialCases->setName(Twine(SpecialCases->getName(), "_udiv-special-cases"));
220  BasicBlock *End = SpecialCases->splitBasicBlock(Builder.GetInsertPoint(),
221  "udiv-end");
222  BasicBlock *LoopExit = BasicBlock::Create(Builder.getContext(),
223  "udiv-loop-exit", F, End);
224  BasicBlock *DoWhile = BasicBlock::Create(Builder.getContext(),
225  "udiv-do-while", F, End);
226  BasicBlock *Preheader = BasicBlock::Create(Builder.getContext(),
227  "udiv-preheader", F, End);
228  BasicBlock *BB1 = BasicBlock::Create(Builder.getContext(),
229  "udiv-bb1", F, End);
230 
231  // We'll be overwriting the terminator to insert our extra blocks
232  SpecialCases->getTerminator()->eraseFromParent();
233 
234  // Same instructions are generated for both i32 (msb 31) and i64 (msb 63).
235 
236  // First off, check for special cases: dividend or divisor is zero, divisor
237  // is greater than dividend, and divisor is 1.
238  // ; special-cases:
239  // ; %ret0_1 = icmp eq i32 %divisor, 0
240  // ; %ret0_2 = icmp eq i32 %dividend, 0
241  // ; %ret0_3 = or i1 %ret0_1, %ret0_2
242  // ; %tmp0 = tail call i32 @llvm.ctlz.i32(i32 %divisor, i1 true)
243  // ; %tmp1 = tail call i32 @llvm.ctlz.i32(i32 %dividend, i1 true)
244  // ; %sr = sub nsw i32 %tmp0, %tmp1
245  // ; %ret0_4 = icmp ugt i32 %sr, 31
246  // ; %ret0 = or i1 %ret0_3, %ret0_4
247  // ; %retDividend = icmp eq i32 %sr, 31
248  // ; %retVal = select i1 %ret0, i32 0, i32 %dividend
249  // ; %earlyRet = or i1 %ret0, %retDividend
250  // ; br i1 %earlyRet, label %end, label %bb1
251  Builder.SetInsertPoint(SpecialCases);
252  Value *Ret0_1 = Builder.CreateICmpEQ(Divisor, Zero);
253  Value *Ret0_2 = Builder.CreateICmpEQ(Dividend, Zero);
254  Value *Ret0_3 = Builder.CreateOr(Ret0_1, Ret0_2);
255  Value *Tmp0 = Builder.CreateCall(CTLZ, {Divisor, True});
256  Value *Tmp1 = Builder.CreateCall(CTLZ, {Dividend, True});
257  Value *SR = Builder.CreateSub(Tmp0, Tmp1);
258  Value *Ret0_4 = Builder.CreateICmpUGT(SR, MSB);
259  Value *Ret0 = Builder.CreateOr(Ret0_3, Ret0_4);
260  Value *RetDividend = Builder.CreateICmpEQ(SR, MSB);
261  Value *RetVal = Builder.CreateSelect(Ret0, Zero, Dividend);
262  Value *EarlyRet = Builder.CreateOr(Ret0, RetDividend);
263  Builder.CreateCondBr(EarlyRet, End, BB1);
264 
265  // ; bb1: ; preds = %special-cases
266  // ; %sr_1 = add i32 %sr, 1
267  // ; %tmp2 = sub i32 31, %sr
268  // ; %q = shl i32 %dividend, %tmp2
269  // ; %skipLoop = icmp eq i32 %sr_1, 0
270  // ; br i1 %skipLoop, label %loop-exit, label %preheader
271  Builder.SetInsertPoint(BB1);
272  Value *SR_1 = Builder.CreateAdd(SR, One);
273  Value *Tmp2 = Builder.CreateSub(MSB, SR);
274  Value *Q = Builder.CreateShl(Dividend, Tmp2);
275  Value *SkipLoop = Builder.CreateICmpEQ(SR_1, Zero);
276  Builder.CreateCondBr(SkipLoop, LoopExit, Preheader);
277 
278  // ; preheader: ; preds = %bb1
279  // ; %tmp3 = lshr i32 %dividend, %sr_1
280  // ; %tmp4 = add i32 %divisor, -1
281  // ; br label %do-while
282  Builder.SetInsertPoint(Preheader);
283  Value *Tmp3 = Builder.CreateLShr(Dividend, SR_1);
284  Value *Tmp4 = Builder.CreateAdd(Divisor, NegOne);
285  Builder.CreateBr(DoWhile);
286 
287  // ; do-while: ; preds = %do-while, %preheader
288  // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ]
289  // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ]
290  // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ]
291  // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ]
292  // ; %tmp5 = shl i32 %r_1, 1
293  // ; %tmp6 = lshr i32 %q_2, 31
294  // ; %tmp7 = or i32 %tmp5, %tmp6
295  // ; %tmp8 = shl i32 %q_2, 1
296  // ; %q_1 = or i32 %carry_1, %tmp8
297  // ; %tmp9 = sub i32 %tmp4, %tmp7
298  // ; %tmp10 = ashr i32 %tmp9, 31
299  // ; %carry = and i32 %tmp10, 1
300  // ; %tmp11 = and i32 %tmp10, %divisor
301  // ; %r = sub i32 %tmp7, %tmp11
302  // ; %sr_2 = add i32 %sr_3, -1
303  // ; %tmp12 = icmp eq i32 %sr_2, 0
304  // ; br i1 %tmp12, label %loop-exit, label %do-while
305  Builder.SetInsertPoint(DoWhile);
306  PHINode *Carry_1 = Builder.CreatePHI(DivTy, 2);
307  PHINode *SR_3 = Builder.CreatePHI(DivTy, 2);
308  PHINode *R_1 = Builder.CreatePHI(DivTy, 2);
309  PHINode *Q_2 = Builder.CreatePHI(DivTy, 2);
310  Value *Tmp5 = Builder.CreateShl(R_1, One);
311  Value *Tmp6 = Builder.CreateLShr(Q_2, MSB);
312  Value *Tmp7 = Builder.CreateOr(Tmp5, Tmp6);
313  Value *Tmp8 = Builder.CreateShl(Q_2, One);
314  Value *Q_1 = Builder.CreateOr(Carry_1, Tmp8);
315  Value *Tmp9 = Builder.CreateSub(Tmp4, Tmp7);
316  Value *Tmp10 = Builder.CreateAShr(Tmp9, MSB);
317  Value *Carry = Builder.CreateAnd(Tmp10, One);
318  Value *Tmp11 = Builder.CreateAnd(Tmp10, Divisor);
319  Value *R = Builder.CreateSub(Tmp7, Tmp11);
320  Value *SR_2 = Builder.CreateAdd(SR_3, NegOne);
321  Value *Tmp12 = Builder.CreateICmpEQ(SR_2, Zero);
322  Builder.CreateCondBr(Tmp12, LoopExit, DoWhile);
323 
324  // ; loop-exit: ; preds = %do-while, %bb1
325  // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ]
326  // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ]
327  // ; %tmp13 = shl i32 %q_3, 1
328  // ; %q_4 = or i32 %carry_2, %tmp13
329  // ; br label %end
330  Builder.SetInsertPoint(LoopExit);
331  PHINode *Carry_2 = Builder.CreatePHI(DivTy, 2);
332  PHINode *Q_3 = Builder.CreatePHI(DivTy, 2);
333  Value *Tmp13 = Builder.CreateShl(Q_3, One);
334  Value *Q_4 = Builder.CreateOr(Carry_2, Tmp13);
335  Builder.CreateBr(End);
336 
337  // ; end: ; preds = %loop-exit, %special-cases
338  // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ]
339  // ; ret i32 %q_5
340  Builder.SetInsertPoint(End, End->begin());
341  PHINode *Q_5 = Builder.CreatePHI(DivTy, 2);
342 
343  // Populate the Phis, since all values have now been created. Our Phis were:
344  // ; %carry_1 = phi i32 [ 0, %preheader ], [ %carry, %do-while ]
345  Carry_1->addIncoming(Zero, Preheader);
346  Carry_1->addIncoming(Carry, DoWhile);
347  // ; %sr_3 = phi i32 [ %sr_1, %preheader ], [ %sr_2, %do-while ]
348  SR_3->addIncoming(SR_1, Preheader);
349  SR_3->addIncoming(SR_2, DoWhile);
350  // ; %r_1 = phi i32 [ %tmp3, %preheader ], [ %r, %do-while ]
351  R_1->addIncoming(Tmp3, Preheader);
352  R_1->addIncoming(R, DoWhile);
353  // ; %q_2 = phi i32 [ %q, %preheader ], [ %q_1, %do-while ]
354  Q_2->addIncoming(Q, Preheader);
355  Q_2->addIncoming(Q_1, DoWhile);
356  // ; %carry_2 = phi i32 [ 0, %bb1 ], [ %carry, %do-while ]
357  Carry_2->addIncoming(Zero, BB1);
358  Carry_2->addIncoming(Carry, DoWhile);
359  // ; %q_3 = phi i32 [ %q, %bb1 ], [ %q_1, %do-while ]
360  Q_3->addIncoming(Q, BB1);
361  Q_3->addIncoming(Q_1, DoWhile);
362  // ; %q_5 = phi i32 [ %q_4, %loop-exit ], [ %retVal, %special-cases ]
363  Q_5->addIncoming(Q_4, LoopExit);
364  Q_5->addIncoming(RetVal, SpecialCases);
365 
366  return Q_5;
367 }
368 
369 /// Generate code to calculate the remainder of two integers, replacing Rem with
370 /// the generated code. This currently generates code using the udiv expansion,
371 /// but future work includes generating more specialized code, e.g. when more
372 /// information about the operands are known. Implements both 32bit and 64bit
373 /// scalar division.
374 ///
375 /// @brief Replace Rem with generated code.
377  assert((Rem->getOpcode() == Instruction::SRem ||
378  Rem->getOpcode() == Instruction::URem) &&
379  "Trying to expand remainder from a non-remainder function");
380 
381  IRBuilder<> Builder(Rem);
382 
383  assert(!Rem->getType()->isVectorTy() && "Div over vectors not supported");
384  assert((Rem->getType()->getIntegerBitWidth() == 32 ||
385  Rem->getType()->getIntegerBitWidth() == 64) &&
386  "Div of bitwidth other than 32 or 64 not supported");
387 
388  // First prepare the sign if it's a signed remainder
389  if (Rem->getOpcode() == Instruction::SRem) {
390  Value *Remainder = generateSignedRemainderCode(Rem->getOperand(0),
391  Rem->getOperand(1), Builder);
392 
393  // Check whether this is the insert point while Rem is still valid.
394  bool IsInsertPoint = Rem->getIterator() == Builder.GetInsertPoint();
395  Rem->replaceAllUsesWith(Remainder);
396  Rem->dropAllReferences();
397  Rem->eraseFromParent();
398 
399  // If we didn't actually generate an urem instruction, we're done
400  // This happens for example if the input were constant. In this case the
401  // Builder insertion point was unchanged
402  if (IsInsertPoint)
403  return true;
404 
406  Rem = BO;
407  }
408 
409  Value *Remainder = generatedUnsignedRemainderCode(Rem->getOperand(0),
410  Rem->getOperand(1),
411  Builder);
412 
413  Rem->replaceAllUsesWith(Remainder);
414  Rem->dropAllReferences();
415  Rem->eraseFromParent();
416 
417  // Expand the udiv
418  if (BinaryOperator *UDiv = dyn_cast<BinaryOperator>(Builder.GetInsertPoint())) {
419  assert(UDiv->getOpcode() == Instruction::UDiv && "Non-udiv in expansion?");
420  expandDivision(UDiv);
421  }
422 
423  return true;
424 }
425 
426 
427 /// Generate code to divide two integers, replacing Div with the generated
428 /// code. This currently generates code similarly to compiler-rt's
429 /// implementations, but future work includes generating more specialized code
430 /// when more information about the operands are known. Implements both
431 /// 32bit and 64bit scalar division.
432 ///
433 /// @brief Replace Div with generated code.
435  assert((Div->getOpcode() == Instruction::SDiv ||
436  Div->getOpcode() == Instruction::UDiv) &&
437  "Trying to expand division from a non-division function");
438 
439  IRBuilder<> Builder(Div);
440 
441  assert(!Div->getType()->isVectorTy() && "Div over vectors not supported");
442  assert((Div->getType()->getIntegerBitWidth() == 32 ||
443  Div->getType()->getIntegerBitWidth() == 64) &&
444  "Div of bitwidth other than 32 or 64 not supported");
445 
446  // First prepare the sign if it's a signed division
447  if (Div->getOpcode() == Instruction::SDiv) {
448  // Lower the code to unsigned division, and reset Div to point to the udiv.
449  Value *Quotient = generateSignedDivisionCode(Div->getOperand(0),
450  Div->getOperand(1), Builder);
451 
452  // Check whether this is the insert point while Div is still valid.
453  bool IsInsertPoint = Div->getIterator() == Builder.GetInsertPoint();
454  Div->replaceAllUsesWith(Quotient);
455  Div->dropAllReferences();
456  Div->eraseFromParent();
457 
458  // If we didn't actually generate an udiv instruction, we're done
459  // This happens for example if the input were constant. In this case the
460  // Builder insertion point was unchanged
461  if (IsInsertPoint)
462  return true;
463 
465  Div = BO;
466  }
467 
468  // Insert the unsigned division code
469  Value *Quotient = generateUnsignedDivisionCode(Div->getOperand(0),
470  Div->getOperand(1),
471  Builder);
472  Div->replaceAllUsesWith(Quotient);
473  Div->dropAllReferences();
474  Div->eraseFromParent();
475 
476  return true;
477 }
478 
479 /// Generate code to compute the remainder of two integers of bitwidth up to
480 /// 32 bits. Uses the above routines and extends the inputs/truncates the
481 /// outputs to operate in 32 bits; that is, these routines are good for targets
482 /// that have no or very little suppport for smaller than 32 bit integer
483 /// arithmetic.
484 ///
485 /// @brief Replace Rem with emulation code.
487  assert((Rem->getOpcode() == Instruction::SRem ||
488  Rem->getOpcode() == Instruction::URem) &&
489  "Trying to expand remainder from a non-remainder function");
490 
491  Type *RemTy = Rem->getType();
492  assert(!RemTy->isVectorTy() && "Div over vectors not supported");
493 
494  unsigned RemTyBitWidth = RemTy->getIntegerBitWidth();
495 
496  assert(RemTyBitWidth <= 32 &&
497  "Div of bitwidth greater than 32 not supported");
498 
499  if (RemTyBitWidth == 32)
500  return expandRemainder(Rem);
501 
502  // If bitwidth smaller than 32 extend inputs, extend output and proceed
503  // with 32 bit division.
504  IRBuilder<> Builder(Rem);
505 
506  Value *ExtDividend;
507  Value *ExtDivisor;
508  Value *ExtRem;
509  Value *Trunc;
510  Type *Int32Ty = Builder.getInt32Ty();
511 
512  if (Rem->getOpcode() == Instruction::SRem) {
513  ExtDividend = Builder.CreateSExt(Rem->getOperand(0), Int32Ty);
514  ExtDivisor = Builder.CreateSExt(Rem->getOperand(1), Int32Ty);
515  ExtRem = Builder.CreateSRem(ExtDividend, ExtDivisor);
516  } else {
517  ExtDividend = Builder.CreateZExt(Rem->getOperand(0), Int32Ty);
518  ExtDivisor = Builder.CreateZExt(Rem->getOperand(1), Int32Ty);
519  ExtRem = Builder.CreateURem(ExtDividend, ExtDivisor);
520  }
521  Trunc = Builder.CreateTrunc(ExtRem, RemTy);
522 
523  Rem->replaceAllUsesWith(Trunc);
524  Rem->dropAllReferences();
525  Rem->eraseFromParent();
526 
527  return expandRemainder(cast<BinaryOperator>(ExtRem));
528 }
529 
530 /// Generate code to compute the remainder of two integers of bitwidth up to
531 /// 64 bits. Uses the above routines and extends the inputs/truncates the
532 /// outputs to operate in 64 bits.
533 ///
534 /// @brief Replace Rem with emulation code.
536  assert((Rem->getOpcode() == Instruction::SRem ||
537  Rem->getOpcode() == Instruction::URem) &&
538  "Trying to expand remainder from a non-remainder function");
539 
540  Type *RemTy = Rem->getType();
541  assert(!RemTy->isVectorTy() && "Div over vectors not supported");
542 
543  unsigned RemTyBitWidth = RemTy->getIntegerBitWidth();
544 
545  assert(RemTyBitWidth <= 64 && "Div of bitwidth greater than 64 not supported");
546 
547  if (RemTyBitWidth == 64)
548  return expandRemainder(Rem);
549 
550  // If bitwidth smaller than 64 extend inputs, extend output and proceed
551  // with 64 bit division.
552  IRBuilder<> Builder(Rem);
553 
554  Value *ExtDividend;
555  Value *ExtDivisor;
556  Value *ExtRem;
557  Value *Trunc;
558  Type *Int64Ty = Builder.getInt64Ty();
559 
560  if (Rem->getOpcode() == Instruction::SRem) {
561  ExtDividend = Builder.CreateSExt(Rem->getOperand(0), Int64Ty);
562  ExtDivisor = Builder.CreateSExt(Rem->getOperand(1), Int64Ty);
563  ExtRem = Builder.CreateSRem(ExtDividend, ExtDivisor);
564  } else {
565  ExtDividend = Builder.CreateZExt(Rem->getOperand(0), Int64Ty);
566  ExtDivisor = Builder.CreateZExt(Rem->getOperand(1), Int64Ty);
567  ExtRem = Builder.CreateURem(ExtDividend, ExtDivisor);
568  }
569  Trunc = Builder.CreateTrunc(ExtRem, RemTy);
570 
571  Rem->replaceAllUsesWith(Trunc);
572  Rem->dropAllReferences();
573  Rem->eraseFromParent();
574 
575  return expandRemainder(cast<BinaryOperator>(ExtRem));
576 }
577 
578 /// Generate code to divide two integers of bitwidth up to 32 bits. Uses the
579 /// above routines and extends the inputs/truncates the outputs to operate
580 /// in 32 bits; that is, these routines are good for targets that have no
581 /// or very little support for smaller than 32 bit integer arithmetic.
582 ///
583 /// @brief Replace Div with emulation code.
585  assert((Div->getOpcode() == Instruction::SDiv ||
586  Div->getOpcode() == Instruction::UDiv) &&
587  "Trying to expand division from a non-division function");
588 
589  Type *DivTy = Div->getType();
590  assert(!DivTy->isVectorTy() && "Div over vectors not supported");
591 
592  unsigned DivTyBitWidth = DivTy->getIntegerBitWidth();
593 
594  assert(DivTyBitWidth <= 32 && "Div of bitwidth greater than 32 not supported");
595 
596  if (DivTyBitWidth == 32)
597  return expandDivision(Div);
598 
599  // If bitwidth smaller than 32 extend inputs, extend output and proceed
600  // with 32 bit division.
601  IRBuilder<> Builder(Div);
602 
603  Value *ExtDividend;
604  Value *ExtDivisor;
605  Value *ExtDiv;
606  Value *Trunc;
607  Type *Int32Ty = Builder.getInt32Ty();
608 
609  if (Div->getOpcode() == Instruction::SDiv) {
610  ExtDividend = Builder.CreateSExt(Div->getOperand(0), Int32Ty);
611  ExtDivisor = Builder.CreateSExt(Div->getOperand(1), Int32Ty);
612  ExtDiv = Builder.CreateSDiv(ExtDividend, ExtDivisor);
613  } else {
614  ExtDividend = Builder.CreateZExt(Div->getOperand(0), Int32Ty);
615  ExtDivisor = Builder.CreateZExt(Div->getOperand(1), Int32Ty);
616  ExtDiv = Builder.CreateUDiv(ExtDividend, ExtDivisor);
617  }
618  Trunc = Builder.CreateTrunc(ExtDiv, DivTy);
619 
620  Div->replaceAllUsesWith(Trunc);
621  Div->dropAllReferences();
622  Div->eraseFromParent();
623 
624  return expandDivision(cast<BinaryOperator>(ExtDiv));
625 }
626 
627 /// Generate code to divide two integers of bitwidth up to 64 bits. Uses the
628 /// above routines and extends the inputs/truncates the outputs to operate
629 /// in 64 bits.
630 ///
631 /// @brief Replace Div with emulation code.
633  assert((Div->getOpcode() == Instruction::SDiv ||
634  Div->getOpcode() == Instruction::UDiv) &&
635  "Trying to expand division from a non-division function");
636 
637  Type *DivTy = Div->getType();
638  assert(!DivTy->isVectorTy() && "Div over vectors not supported");
639 
640  unsigned DivTyBitWidth = DivTy->getIntegerBitWidth();
641 
642  assert(DivTyBitWidth <= 64 &&
643  "Div of bitwidth greater than 64 not supported");
644 
645  if (DivTyBitWidth == 64)
646  return expandDivision(Div);
647 
648  // If bitwidth smaller than 64 extend inputs, extend output and proceed
649  // with 64 bit division.
650  IRBuilder<> Builder(Div);
651 
652  Value *ExtDividend;
653  Value *ExtDivisor;
654  Value *ExtDiv;
655  Value *Trunc;
656  Type *Int64Ty = Builder.getInt64Ty();
657 
658  if (Div->getOpcode() == Instruction::SDiv) {
659  ExtDividend = Builder.CreateSExt(Div->getOperand(0), Int64Ty);
660  ExtDivisor = Builder.CreateSExt(Div->getOperand(1), Int64Ty);
661  ExtDiv = Builder.CreateSDiv(ExtDividend, ExtDivisor);
662  } else {
663  ExtDividend = Builder.CreateZExt(Div->getOperand(0), Int64Ty);
664  ExtDivisor = Builder.CreateZExt(Div->getOperand(1), Int64Ty);
665  ExtDiv = Builder.CreateUDiv(ExtDividend, ExtDivisor);
666  }
667  Trunc = Builder.CreateTrunc(ExtDiv, DivTy);
668 
669  Div->replaceAllUsesWith(Trunc);
670  Div->dropAllReferences();
671  Div->eraseFromParent();
672 
673  return expandDivision(cast<BinaryOperator>(ExtDiv));
674 }
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
BranchInst * CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, MDNode *BranchWeights=nullptr, MDNode *Unpredictable=nullptr)
Create a conditional &#39;br Cond, TrueDest, FalseDest&#39; instruction.
Definition: IRBuilder.h:779
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
LLVMContext & getContext() const
Definition: IRBuilder.h:124
void dropAllReferences()
Drop all references to operands.
Definition: User.h:279
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BinaryOps getOpcode() const
Definition: InstrTypes.h:523
Value * CreateXor(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1099
Value * CreateSExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1390
F(f)
bool expandRemainderUpTo64Bits(BinaryOperator *Rem)
Generate code to calculate the remainder of two integers, replacing Rem with the generated code...
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
IntegerType * getInt32Ty()
Fetch the type representing a 32-bit integer.
Definition: IRBuilder.h:348
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
IntegerType * getInt64Ty()
Fetch the type representing a 64-bit integer.
Definition: IRBuilder.h:353
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:668
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:893
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:286
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
Value * CreateICmpUGT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1556
BasicBlock * GetInsertBlock() const
Definition: IRBuilder.h:122
bool expandRemainderUpTo32Bits(BinaryOperator *Rem)
Generate code to calculate the remainder of two integers, replacing Rem with the generated code...
Value * CreateSub(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:915
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:430
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1387
bool expandRemainder(BinaryOperator *Rem)
Generate code to calculate the remainder of two integers, replacing Rem with the generated code...
unsigned getBitWidth() const
Get the number of bits in this IntegerType.
Definition: DerivedTypes.h:66
Function * getDeclaration(Module *M, ID id, ArrayRef< Type *> Tys=None)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:980
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block...
Definition: IRBuilder.h:128
Value * getOperand(unsigned i) const
Definition: User.h:154
Value * CreateOr(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1083
static Value * generateSignedDivisionCode(Value *Dividend, Value *Divisor, IRBuilder<> &Builder)
Generate code to divide two signed integers.
bool expandDivisionUpTo32Bits(BinaryOperator *Div)
Generate code to divide two integers, replacing Div with the generated code.
Value * CreateAShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1049
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
ConstantInt * getTrue()
Get the constant value for i1 true.
Definition: IRBuilder.h:288
bool expandDivision(BinaryOperator *Div)
Generate code to divide two integers, replacing Div with the generated code.
Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)
Definition: IRBuilder.h:1693
bool expandDivisionUpTo64Bits(BinaryOperator *Div)
Generate code to divide two integers, replacing Div with the generated code.
static const unsigned End
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
Definition: BasicBlock.h:101
Value * CreateICmpEQ(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1550
self_iterator getIterator()
Definition: ilist_node.h:82
Class to represent integer types.
Definition: DerivedTypes.h:40
ConstantInt * getInt64(uint64_t C)
Get a constant 64-bit value.
Definition: IRBuilder.h:313
Value * CreateMul(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:937
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1384
Value * CreateSRem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:997
static Value * generatedUnsignedRemainderCode(Value *Dividend, Value *Divisor, IRBuilder<> &Builder)
Generate code to compute the remainder of two unsigned integers.
PHINode * CreatePHI(Type *Ty, unsigned NumReservedValues, const Twine &Name="")
Definition: IRBuilder.h:1658
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Definition: IRBuilder.h:308
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
Definition: Constants.cpp:574
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:991
Value * CreateSDiv(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:971
Value * CreateUDiv(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:959
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1012
unsigned getIntegerBitWidth() const
Definition: DerivedTypes.h:97
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
static Value * generateUnsignedDivisionCode(Value *Dividend, Value *Divisor, IRBuilder<> &Builder)
Generates code to divide two unsigned scalar 32-bit or 64-bit integers.
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1067
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:556
LLVM Value Representation.
Definition: Value.h:73
BranchInst * CreateBr(BasicBlock *Dest)
Create an unconditional &#39;br label X&#39; instruction.
Definition: IRBuilder.h:773
static Value * generateSignedRemainderCode(Value *Dividend, Value *Divisor, IRBuilder<> &Builder)
Generate code to compute the remainder of two signed integers.
BasicBlock::iterator GetInsertPoint() const
Definition: IRBuilder.h:123
Value * CreateLShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1031
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
IntegerType * Int32Ty
CallInst * CreateCall(Value *Callee, ArrayRef< Value *> Args=None, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1663