• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- BypassSlowDivision.cpp - Bypass slow 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 optimization for div and rem on architectures that
10 // execute short instructions significantly faster than longer instructions.
11 // For example, on Intel Atom 32-bit divides are slow enough that during
12 // runtime it is profitable to check the value of the operands, and if they are
13 // positive and less than 256 use an unsigned 8-bit divide.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/None.h"
20 #include "llvm/ADT/Optional.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/Transforms/Utils/Local.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/Instruction.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/Type.h"
34 #include "llvm/IR/Value.h"
35 #include "llvm/Support/Casting.h"
36 #include "llvm/Support/KnownBits.h"
37 #include <cassert>
38 #include <cstdint>
39 
40 using namespace llvm;
41 
42 #define DEBUG_TYPE "bypass-slow-division"
43 
44 namespace {
45 
46   struct QuotRemPair {
47     Value *Quotient;
48     Value *Remainder;
49 
QuotRemPair__anon84524bd30111::QuotRemPair50     QuotRemPair(Value *InQuotient, Value *InRemainder)
51         : Quotient(InQuotient), Remainder(InRemainder) {}
52   };
53 
54   /// A quotient and remainder, plus a BB from which they logically "originate".
55   /// If you use Quotient or Remainder in a Phi node, you should use BB as its
56   /// corresponding predecessor.
57   struct QuotRemWithBB {
58     BasicBlock *BB = nullptr;
59     Value *Quotient = nullptr;
60     Value *Remainder = nullptr;
61   };
62 
63 using DivCacheTy = DenseMap<DivRemMapKey, QuotRemPair>;
64 using BypassWidthsTy = DenseMap<unsigned, unsigned>;
65 using VisitedSetTy = SmallPtrSet<Instruction *, 4>;
66 
67 enum ValueRange {
68   /// Operand definitely fits into BypassType. No runtime checks are needed.
69   VALRNG_KNOWN_SHORT,
70   /// A runtime check is required, as value range is unknown.
71   VALRNG_UNKNOWN,
72   /// Operand is unlikely to fit into BypassType. The bypassing should be
73   /// disabled.
74   VALRNG_LIKELY_LONG
75 };
76 
77 class FastDivInsertionTask {
78   bool IsValidTask = false;
79   Instruction *SlowDivOrRem = nullptr;
80   IntegerType *BypassType = nullptr;
81   BasicBlock *MainBB = nullptr;
82 
83   bool isHashLikeValue(Value *V, VisitedSetTy &Visited);
84   ValueRange getValueRange(Value *Op, VisitedSetTy &Visited);
85   QuotRemWithBB createSlowBB(BasicBlock *Successor);
86   QuotRemWithBB createFastBB(BasicBlock *Successor);
87   QuotRemPair createDivRemPhiNodes(QuotRemWithBB &LHS, QuotRemWithBB &RHS,
88                                    BasicBlock *PhiBB);
89   Value *insertOperandRuntimeCheck(Value *Op1, Value *Op2);
90   Optional<QuotRemPair> insertFastDivAndRem();
91 
isSignedOp()92   bool isSignedOp() {
93     return SlowDivOrRem->getOpcode() == Instruction::SDiv ||
94            SlowDivOrRem->getOpcode() == Instruction::SRem;
95   }
96 
isDivisionOp()97   bool isDivisionOp() {
98     return SlowDivOrRem->getOpcode() == Instruction::SDiv ||
99            SlowDivOrRem->getOpcode() == Instruction::UDiv;
100   }
101 
getSlowType()102   Type *getSlowType() { return SlowDivOrRem->getType(); }
103 
104 public:
105   FastDivInsertionTask(Instruction *I, const BypassWidthsTy &BypassWidths);
106 
107   Value *getReplacement(DivCacheTy &Cache);
108 };
109 
110 } // end anonymous namespace
111 
FastDivInsertionTask(Instruction * I,const BypassWidthsTy & BypassWidths)112 FastDivInsertionTask::FastDivInsertionTask(Instruction *I,
113                                            const BypassWidthsTy &BypassWidths) {
114   switch (I->getOpcode()) {
115   case Instruction::UDiv:
116   case Instruction::SDiv:
117   case Instruction::URem:
118   case Instruction::SRem:
119     SlowDivOrRem = I;
120     break;
121   default:
122     // I is not a div/rem operation.
123     return;
124   }
125 
126   // Skip division on vector types. Only optimize integer instructions.
127   IntegerType *SlowType = dyn_cast<IntegerType>(SlowDivOrRem->getType());
128   if (!SlowType)
129     return;
130 
131   // Skip if this bitwidth is not bypassed.
132   auto BI = BypassWidths.find(SlowType->getBitWidth());
133   if (BI == BypassWidths.end())
134     return;
135 
136   // Get type for div/rem instruction with bypass bitwidth.
137   IntegerType *BT = IntegerType::get(I->getContext(), BI->second);
138   BypassType = BT;
139 
140   // The original basic block.
141   MainBB = I->getParent();
142 
143   // The instruction is indeed a slow div or rem operation.
144   IsValidTask = true;
145 }
146 
147 /// Reuses previously-computed dividend or remainder from the current BB if
148 /// operands and operation are identical. Otherwise calls insertFastDivAndRem to
149 /// perform the optimization and caches the resulting dividend and remainder.
150 /// If no replacement can be generated, nullptr is returned.
getReplacement(DivCacheTy & Cache)151 Value *FastDivInsertionTask::getReplacement(DivCacheTy &Cache) {
152   // First, make sure that the task is valid.
153   if (!IsValidTask)
154     return nullptr;
155 
156   // Then, look for a value in Cache.
157   Value *Dividend = SlowDivOrRem->getOperand(0);
158   Value *Divisor = SlowDivOrRem->getOperand(1);
159   DivRemMapKey Key(isSignedOp(), Dividend, Divisor);
160   auto CacheI = Cache.find(Key);
161 
162   if (CacheI == Cache.end()) {
163     // If previous instance does not exist, try to insert fast div.
164     Optional<QuotRemPair> OptResult = insertFastDivAndRem();
165     // Bail out if insertFastDivAndRem has failed.
166     if (!OptResult)
167       return nullptr;
168     CacheI = Cache.insert({Key, *OptResult}).first;
169   }
170 
171   QuotRemPair &Value = CacheI->second;
172   return isDivisionOp() ? Value.Quotient : Value.Remainder;
173 }
174 
175 /// Check if a value looks like a hash.
176 ///
177 /// The routine is expected to detect values computed using the most common hash
178 /// algorithms. Typically, hash computations end with one of the following
179 /// instructions:
180 ///
181 /// 1) MUL with a constant wider than BypassType
182 /// 2) XOR instruction
183 ///
184 /// And even if we are wrong and the value is not a hash, it is still quite
185 /// unlikely that such values will fit into BypassType.
186 ///
187 /// To detect string hash algorithms like FNV we have to look through PHI-nodes.
188 /// It is implemented as a depth-first search for values that look neither long
189 /// nor hash-like.
isHashLikeValue(Value * V,VisitedSetTy & Visited)190 bool FastDivInsertionTask::isHashLikeValue(Value *V, VisitedSetTy &Visited) {
191   Instruction *I = dyn_cast<Instruction>(V);
192   if (!I)
193     return false;
194 
195   switch (I->getOpcode()) {
196   case Instruction::Xor:
197     return true;
198   case Instruction::Mul: {
199     // After Constant Hoisting pass, long constants may be represented as
200     // bitcast instructions. As a result, some constants may look like an
201     // instruction at first, and an additional check is necessary to find out if
202     // an operand is actually a constant.
203     Value *Op1 = I->getOperand(1);
204     ConstantInt *C = dyn_cast<ConstantInt>(Op1);
205     if (!C && isa<BitCastInst>(Op1))
206       C = dyn_cast<ConstantInt>(cast<BitCastInst>(Op1)->getOperand(0));
207     return C && C->getValue().getMinSignedBits() > BypassType->getBitWidth();
208   }
209   case Instruction::PHI:
210     // Stop IR traversal in case of a crazy input code. This limits recursion
211     // depth.
212     if (Visited.size() >= 16)
213       return false;
214     // Do not visit nodes that have been visited already. We return true because
215     // it means that we couldn't find any value that doesn't look hash-like.
216     if (Visited.find(I) != Visited.end())
217       return true;
218     Visited.insert(I);
219     return llvm::all_of(cast<PHINode>(I)->incoming_values(), [&](Value *V) {
220       // Ignore undef values as they probably don't affect the division
221       // operands.
222       return getValueRange(V, Visited) == VALRNG_LIKELY_LONG ||
223              isa<UndefValue>(V);
224     });
225   default:
226     return false;
227   }
228 }
229 
230 /// Check if an integer value fits into our bypass type.
getValueRange(Value * V,VisitedSetTy & Visited)231 ValueRange FastDivInsertionTask::getValueRange(Value *V,
232                                                VisitedSetTy &Visited) {
233   unsigned ShortLen = BypassType->getBitWidth();
234   unsigned LongLen = V->getType()->getIntegerBitWidth();
235 
236   assert(LongLen > ShortLen && "Value type must be wider than BypassType");
237   unsigned HiBits = LongLen - ShortLen;
238 
239   const DataLayout &DL = SlowDivOrRem->getModule()->getDataLayout();
240   KnownBits Known(LongLen);
241 
242   computeKnownBits(V, Known, DL);
243 
244   if (Known.countMinLeadingZeros() >= HiBits)
245     return VALRNG_KNOWN_SHORT;
246 
247   if (Known.countMaxLeadingZeros() < HiBits)
248     return VALRNG_LIKELY_LONG;
249 
250   // Long integer divisions are often used in hashtable implementations. It's
251   // not worth bypassing such divisions because hash values are extremely
252   // unlikely to have enough leading zeros. The call below tries to detect
253   // values that are unlikely to fit BypassType (including hashes).
254   if (isHashLikeValue(V, Visited))
255     return VALRNG_LIKELY_LONG;
256 
257   return VALRNG_UNKNOWN;
258 }
259 
260 /// Add new basic block for slow div and rem operations and put it before
261 /// SuccessorBB.
createSlowBB(BasicBlock * SuccessorBB)262 QuotRemWithBB FastDivInsertionTask::createSlowBB(BasicBlock *SuccessorBB) {
263   QuotRemWithBB DivRemPair;
264   DivRemPair.BB = BasicBlock::Create(MainBB->getParent()->getContext(), "",
265                                      MainBB->getParent(), SuccessorBB);
266   IRBuilder<> Builder(DivRemPair.BB, DivRemPair.BB->begin());
267 
268   Value *Dividend = SlowDivOrRem->getOperand(0);
269   Value *Divisor = SlowDivOrRem->getOperand(1);
270 
271   if (isSignedOp()) {
272     DivRemPair.Quotient = Builder.CreateSDiv(Dividend, Divisor);
273     DivRemPair.Remainder = Builder.CreateSRem(Dividend, Divisor);
274   } else {
275     DivRemPair.Quotient = Builder.CreateUDiv(Dividend, Divisor);
276     DivRemPair.Remainder = Builder.CreateURem(Dividend, Divisor);
277   }
278 
279   Builder.CreateBr(SuccessorBB);
280   return DivRemPair;
281 }
282 
283 /// Add new basic block for fast div and rem operations and put it before
284 /// SuccessorBB.
createFastBB(BasicBlock * SuccessorBB)285 QuotRemWithBB FastDivInsertionTask::createFastBB(BasicBlock *SuccessorBB) {
286   QuotRemWithBB DivRemPair;
287   DivRemPair.BB = BasicBlock::Create(MainBB->getParent()->getContext(), "",
288                                      MainBB->getParent(), SuccessorBB);
289   IRBuilder<> Builder(DivRemPair.BB, DivRemPair.BB->begin());
290 
291   Value *Dividend = SlowDivOrRem->getOperand(0);
292   Value *Divisor = SlowDivOrRem->getOperand(1);
293   Value *ShortDivisorV =
294       Builder.CreateCast(Instruction::Trunc, Divisor, BypassType);
295   Value *ShortDividendV =
296       Builder.CreateCast(Instruction::Trunc, Dividend, BypassType);
297 
298   // udiv/urem because this optimization only handles positive numbers.
299   Value *ShortQV = Builder.CreateUDiv(ShortDividendV, ShortDivisorV);
300   Value *ShortRV = Builder.CreateURem(ShortDividendV, ShortDivisorV);
301   DivRemPair.Quotient =
302       Builder.CreateCast(Instruction::ZExt, ShortQV, getSlowType());
303   DivRemPair.Remainder =
304       Builder.CreateCast(Instruction::ZExt, ShortRV, getSlowType());
305   Builder.CreateBr(SuccessorBB);
306 
307   return DivRemPair;
308 }
309 
310 /// Creates Phi nodes for result of Div and Rem.
createDivRemPhiNodes(QuotRemWithBB & LHS,QuotRemWithBB & RHS,BasicBlock * PhiBB)311 QuotRemPair FastDivInsertionTask::createDivRemPhiNodes(QuotRemWithBB &LHS,
312                                                        QuotRemWithBB &RHS,
313                                                        BasicBlock *PhiBB) {
314   IRBuilder<> Builder(PhiBB, PhiBB->begin());
315   PHINode *QuoPhi = Builder.CreatePHI(getSlowType(), 2);
316   QuoPhi->addIncoming(LHS.Quotient, LHS.BB);
317   QuoPhi->addIncoming(RHS.Quotient, RHS.BB);
318   PHINode *RemPhi = Builder.CreatePHI(getSlowType(), 2);
319   RemPhi->addIncoming(LHS.Remainder, LHS.BB);
320   RemPhi->addIncoming(RHS.Remainder, RHS.BB);
321   return QuotRemPair(QuoPhi, RemPhi);
322 }
323 
324 /// Creates a runtime check to test whether both the divisor and dividend fit
325 /// into BypassType. The check is inserted at the end of MainBB. True return
326 /// value means that the operands fit. Either of the operands may be NULL if it
327 /// doesn't need a runtime check.
insertOperandRuntimeCheck(Value * Op1,Value * Op2)328 Value *FastDivInsertionTask::insertOperandRuntimeCheck(Value *Op1, Value *Op2) {
329   assert((Op1 || Op2) && "Nothing to check");
330   IRBuilder<> Builder(MainBB, MainBB->end());
331 
332   Value *OrV;
333   if (Op1 && Op2)
334     OrV = Builder.CreateOr(Op1, Op2);
335   else
336     OrV = Op1 ? Op1 : Op2;
337 
338   // BitMask is inverted to check if the operands are
339   // larger than the bypass type
340   uint64_t BitMask = ~BypassType->getBitMask();
341   Value *AndV = Builder.CreateAnd(OrV, BitMask);
342 
343   // Compare operand values
344   Value *ZeroV = ConstantInt::getSigned(getSlowType(), 0);
345   return Builder.CreateICmpEQ(AndV, ZeroV);
346 }
347 
348 /// Substitutes the div/rem instruction with code that checks the value of the
349 /// operands and uses a shorter-faster div/rem instruction when possible.
insertFastDivAndRem()350 Optional<QuotRemPair> FastDivInsertionTask::insertFastDivAndRem() {
351   Value *Dividend = SlowDivOrRem->getOperand(0);
352   Value *Divisor = SlowDivOrRem->getOperand(1);
353 
354   VisitedSetTy SetL;
355   ValueRange DividendRange = getValueRange(Dividend, SetL);
356   if (DividendRange == VALRNG_LIKELY_LONG)
357     return None;
358 
359   VisitedSetTy SetR;
360   ValueRange DivisorRange = getValueRange(Divisor, SetR);
361   if (DivisorRange == VALRNG_LIKELY_LONG)
362     return None;
363 
364   bool DividendShort = (DividendRange == VALRNG_KNOWN_SHORT);
365   bool DivisorShort = (DivisorRange == VALRNG_KNOWN_SHORT);
366 
367   if (DividendShort && DivisorShort) {
368     // If both operands are known to be short then just replace the long
369     // division with a short one in-place.  Since we're not introducing control
370     // flow in this case, narrowing the division is always a win, even if the
371     // divisor is a constant (and will later get replaced by a multiplication).
372 
373     IRBuilder<> Builder(SlowDivOrRem);
374     Value *TruncDividend = Builder.CreateTrunc(Dividend, BypassType);
375     Value *TruncDivisor = Builder.CreateTrunc(Divisor, BypassType);
376     Value *TruncDiv = Builder.CreateUDiv(TruncDividend, TruncDivisor);
377     Value *TruncRem = Builder.CreateURem(TruncDividend, TruncDivisor);
378     Value *ExtDiv = Builder.CreateZExt(TruncDiv, getSlowType());
379     Value *ExtRem = Builder.CreateZExt(TruncRem, getSlowType());
380     return QuotRemPair(ExtDiv, ExtRem);
381   }
382 
383   if (isa<ConstantInt>(Divisor)) {
384     // If the divisor is not a constant, DAGCombiner will convert it to a
385     // multiplication by a magic constant.  It isn't clear if it is worth
386     // introducing control flow to get a narrower multiply.
387     return None;
388   }
389 
390   // After Constant Hoisting pass, long constants may be represented as
391   // bitcast instructions. As a result, some constants may look like an
392   // instruction at first, and an additional check is necessary to find out if
393   // an operand is actually a constant.
394   if (auto *BCI = dyn_cast<BitCastInst>(Divisor))
395     if (BCI->getParent() == SlowDivOrRem->getParent() &&
396         isa<ConstantInt>(BCI->getOperand(0)))
397       return None;
398 
399   if (DividendShort && !isSignedOp()) {
400     // If the division is unsigned and Dividend is known to be short, then
401     // either
402     // 1) Divisor is less or equal to Dividend, and the result can be computed
403     //    with a short division.
404     // 2) Divisor is greater than Dividend. In this case, no division is needed
405     //    at all: The quotient is 0 and the remainder is equal to Dividend.
406     //
407     // So instead of checking at runtime whether Divisor fits into BypassType,
408     // we emit a runtime check to differentiate between these two cases. This
409     // lets us entirely avoid a long div.
410 
411     // Split the basic block before the div/rem.
412     BasicBlock *SuccessorBB = MainBB->splitBasicBlock(SlowDivOrRem);
413     // Remove the unconditional branch from MainBB to SuccessorBB.
414     MainBB->getInstList().back().eraseFromParent();
415     QuotRemWithBB Long;
416     Long.BB = MainBB;
417     Long.Quotient = ConstantInt::get(getSlowType(), 0);
418     Long.Remainder = Dividend;
419     QuotRemWithBB Fast = createFastBB(SuccessorBB);
420     QuotRemPair Result = createDivRemPhiNodes(Fast, Long, SuccessorBB);
421     IRBuilder<> Builder(MainBB, MainBB->end());
422     Value *CmpV = Builder.CreateICmpUGE(Dividend, Divisor);
423     Builder.CreateCondBr(CmpV, Fast.BB, SuccessorBB);
424     return Result;
425   } else {
426     // General case. Create both slow and fast div/rem pairs and choose one of
427     // them at runtime.
428 
429     // Split the basic block before the div/rem.
430     BasicBlock *SuccessorBB = MainBB->splitBasicBlock(SlowDivOrRem);
431     // Remove the unconditional branch from MainBB to SuccessorBB.
432     MainBB->getInstList().back().eraseFromParent();
433     QuotRemWithBB Fast = createFastBB(SuccessorBB);
434     QuotRemWithBB Slow = createSlowBB(SuccessorBB);
435     QuotRemPair Result = createDivRemPhiNodes(Fast, Slow, SuccessorBB);
436     Value *CmpV = insertOperandRuntimeCheck(DividendShort ? nullptr : Dividend,
437                                             DivisorShort ? nullptr : Divisor);
438     IRBuilder<> Builder(MainBB, MainBB->end());
439     Builder.CreateCondBr(CmpV, Fast.BB, Slow.BB);
440     return Result;
441   }
442 }
443 
444 /// This optimization identifies DIV/REM instructions in a BB that can be
445 /// profitably bypassed and carried out with a shorter, faster divide.
bypassSlowDivision(BasicBlock * BB,const BypassWidthsTy & BypassWidths)446 bool llvm::bypassSlowDivision(BasicBlock *BB,
447                               const BypassWidthsTy &BypassWidths) {
448   DivCacheTy PerBBDivCache;
449 
450   bool MadeChange = false;
451   Instruction *Next = &*BB->begin();
452   while (Next != nullptr) {
453     // We may add instructions immediately after I, but we want to skip over
454     // them.
455     Instruction *I = Next;
456     Next = Next->getNextNode();
457 
458     // Ignore dead code to save time and avoid bugs.
459     if (I->hasNUses(0))
460       continue;
461 
462     FastDivInsertionTask Task(I, BypassWidths);
463     if (Value *Replacement = Task.getReplacement(PerBBDivCache)) {
464       I->replaceAllUsesWith(Replacement);
465       I->eraseFromParent();
466       MadeChange = true;
467     }
468   }
469 
470   // Above we eagerly create divs and rems, as pairs, so that we can efficiently
471   // create divrem machine instructions.  Now erase any unused divs / rems so we
472   // don't leave extra instructions sitting around.
473   for (auto &KV : PerBBDivCache)
474     for (Value *V : {KV.second.Quotient, KV.second.Remainder})
475       RecursivelyDeleteTriviallyDeadInstructions(V);
476 
477   return MadeChange;
478 }
479