xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/AggressiveInstCombine/TruncInstCombine.cpp (revision d5b0e70f7e04d971691517ce1304d86a1e367e2e)
1 //===- TruncInstCombine.cpp -----------------------------------------------===//
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 // TruncInstCombine - looks for expression dags post-dominated by TruncInst and
10 // for each eligible dag, it will create a reduced bit-width expression, replace
11 // the old expression with this new one and remove the old expression.
12 // Eligible expression dag is such that:
13 //   1. Contains only supported instructions.
14 //   2. Supported leaves: ZExtInst, SExtInst, TruncInst and Constant value.
15 //   3. Can be evaluated into type with reduced legal bit-width.
16 //   4. All instructions in the dag must not have users outside the dag.
17 //      The only exception is for {ZExt, SExt}Inst with operand type equal to
18 //      the new reduced type evaluated in (3).
19 //
20 // The motivation for this optimization is that evaluating and expression using
21 // smaller bit-width is preferable, especially for vectorization where we can
22 // fit more values in one vectorized instruction. In addition, this optimization
23 // may decrease the number of cast instructions, but will not increase it.
24 //
25 //===----------------------------------------------------------------------===//
26 
27 #include "AggressiveInstCombineInternal.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/TargetLibraryInfo.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instruction.h"
36 #include "llvm/Support/KnownBits.h"
37 
38 using namespace llvm;
39 
40 #define DEBUG_TYPE "aggressive-instcombine"
41 
42 STATISTIC(
43     NumDAGsReduced,
44     "Number of truncations eliminated by reducing bit width of expression DAG");
45 STATISTIC(NumInstrsReduced,
46           "Number of instructions whose bit width was reduced");
47 
48 /// Given an instruction and a container, it fills all the relevant operands of
49 /// that instruction, with respect to the Trunc expression dag optimizaton.
50 static void getRelevantOperands(Instruction *I, SmallVectorImpl<Value *> &Ops) {
51   unsigned Opc = I->getOpcode();
52   switch (Opc) {
53   case Instruction::Trunc:
54   case Instruction::ZExt:
55   case Instruction::SExt:
56     // These CastInst are considered leaves of the evaluated expression, thus,
57     // their operands are not relevent.
58     break;
59   case Instruction::Add:
60   case Instruction::Sub:
61   case Instruction::Mul:
62   case Instruction::And:
63   case Instruction::Or:
64   case Instruction::Xor:
65   case Instruction::Shl:
66   case Instruction::LShr:
67   case Instruction::AShr:
68   case Instruction::UDiv:
69   case Instruction::URem:
70   case Instruction::InsertElement:
71     Ops.push_back(I->getOperand(0));
72     Ops.push_back(I->getOperand(1));
73     break;
74   case Instruction::ExtractElement:
75     Ops.push_back(I->getOperand(0));
76     break;
77   case Instruction::Select:
78     Ops.push_back(I->getOperand(1));
79     Ops.push_back(I->getOperand(2));
80     break;
81   default:
82     llvm_unreachable("Unreachable!");
83   }
84 }
85 
86 bool TruncInstCombine::buildTruncExpressionDag() {
87   SmallVector<Value *, 8> Worklist;
88   SmallVector<Instruction *, 8> Stack;
89   // Clear old expression dag.
90   InstInfoMap.clear();
91 
92   Worklist.push_back(CurrentTruncInst->getOperand(0));
93 
94   while (!Worklist.empty()) {
95     Value *Curr = Worklist.back();
96 
97     if (isa<Constant>(Curr)) {
98       Worklist.pop_back();
99       continue;
100     }
101 
102     auto *I = dyn_cast<Instruction>(Curr);
103     if (!I)
104       return false;
105 
106     if (!Stack.empty() && Stack.back() == I) {
107       // Already handled all instruction operands, can remove it from both the
108       // Worklist and the Stack, and add it to the instruction info map.
109       Worklist.pop_back();
110       Stack.pop_back();
111       // Insert I to the Info map.
112       InstInfoMap.insert(std::make_pair(I, Info()));
113       continue;
114     }
115 
116     if (InstInfoMap.count(I)) {
117       Worklist.pop_back();
118       continue;
119     }
120 
121     // Add the instruction to the stack before start handling its operands.
122     Stack.push_back(I);
123 
124     unsigned Opc = I->getOpcode();
125     switch (Opc) {
126     case Instruction::Trunc:
127     case Instruction::ZExt:
128     case Instruction::SExt:
129       // trunc(trunc(x)) -> trunc(x)
130       // trunc(ext(x)) -> ext(x) if the source type is smaller than the new dest
131       // trunc(ext(x)) -> trunc(x) if the source type is larger than the new
132       // dest
133       break;
134     case Instruction::Add:
135     case Instruction::Sub:
136     case Instruction::Mul:
137     case Instruction::And:
138     case Instruction::Or:
139     case Instruction::Xor:
140     case Instruction::Shl:
141     case Instruction::LShr:
142     case Instruction::AShr:
143     case Instruction::UDiv:
144     case Instruction::URem:
145     case Instruction::InsertElement:
146     case Instruction::ExtractElement:
147     case Instruction::Select: {
148       SmallVector<Value *, 2> Operands;
149       getRelevantOperands(I, Operands);
150       append_range(Worklist, Operands);
151       break;
152     }
153     default:
154       // TODO: Can handle more cases here:
155       // 1. shufflevector
156       // 2. sdiv, srem
157       // 3. phi node(and loop handling)
158       // ...
159       return false;
160     }
161   }
162   return true;
163 }
164 
165 unsigned TruncInstCombine::getMinBitWidth() {
166   SmallVector<Value *, 8> Worklist;
167   SmallVector<Instruction *, 8> Stack;
168 
169   Value *Src = CurrentTruncInst->getOperand(0);
170   Type *DstTy = CurrentTruncInst->getType();
171   unsigned TruncBitWidth = DstTy->getScalarSizeInBits();
172   unsigned OrigBitWidth =
173       CurrentTruncInst->getOperand(0)->getType()->getScalarSizeInBits();
174 
175   if (isa<Constant>(Src))
176     return TruncBitWidth;
177 
178   Worklist.push_back(Src);
179   InstInfoMap[cast<Instruction>(Src)].ValidBitWidth = TruncBitWidth;
180 
181   while (!Worklist.empty()) {
182     Value *Curr = Worklist.back();
183 
184     if (isa<Constant>(Curr)) {
185       Worklist.pop_back();
186       continue;
187     }
188 
189     // Otherwise, it must be an instruction.
190     auto *I = cast<Instruction>(Curr);
191 
192     auto &Info = InstInfoMap[I];
193 
194     SmallVector<Value *, 2> Operands;
195     getRelevantOperands(I, Operands);
196 
197     if (!Stack.empty() && Stack.back() == I) {
198       // Already handled all instruction operands, can remove it from both, the
199       // Worklist and the Stack, and update MinBitWidth.
200       Worklist.pop_back();
201       Stack.pop_back();
202       for (auto *Operand : Operands)
203         if (auto *IOp = dyn_cast<Instruction>(Operand))
204           Info.MinBitWidth =
205               std::max(Info.MinBitWidth, InstInfoMap[IOp].MinBitWidth);
206       continue;
207     }
208 
209     // Add the instruction to the stack before start handling its operands.
210     Stack.push_back(I);
211     unsigned ValidBitWidth = Info.ValidBitWidth;
212 
213     // Update minimum bit-width before handling its operands. This is required
214     // when the instruction is part of a loop.
215     Info.MinBitWidth = std::max(Info.MinBitWidth, Info.ValidBitWidth);
216 
217     for (auto *Operand : Operands)
218       if (auto *IOp = dyn_cast<Instruction>(Operand)) {
219         // If we already calculated the minimum bit-width for this valid
220         // bit-width, or for a smaller valid bit-width, then just keep the
221         // answer we already calculated.
222         unsigned IOpBitwidth = InstInfoMap.lookup(IOp).ValidBitWidth;
223         if (IOpBitwidth >= ValidBitWidth)
224           continue;
225         InstInfoMap[IOp].ValidBitWidth = ValidBitWidth;
226         Worklist.push_back(IOp);
227       }
228   }
229   unsigned MinBitWidth = InstInfoMap.lookup(cast<Instruction>(Src)).MinBitWidth;
230   assert(MinBitWidth >= TruncBitWidth);
231 
232   if (MinBitWidth > TruncBitWidth) {
233     // In this case reducing expression with vector type might generate a new
234     // vector type, which is not preferable as it might result in generating
235     // sub-optimal code.
236     if (DstTy->isVectorTy())
237       return OrigBitWidth;
238     // Use the smallest integer type in the range [MinBitWidth, OrigBitWidth).
239     Type *Ty = DL.getSmallestLegalIntType(DstTy->getContext(), MinBitWidth);
240     // Update minimum bit-width with the new destination type bit-width if
241     // succeeded to find such, otherwise, with original bit-width.
242     MinBitWidth = Ty ? Ty->getScalarSizeInBits() : OrigBitWidth;
243   } else { // MinBitWidth == TruncBitWidth
244     // In this case the expression can be evaluated with the trunc instruction
245     // destination type, and trunc instruction can be omitted. However, we
246     // should not perform the evaluation if the original type is a legal scalar
247     // type and the target type is illegal.
248     bool FromLegal = MinBitWidth == 1 || DL.isLegalInteger(OrigBitWidth);
249     bool ToLegal = MinBitWidth == 1 || DL.isLegalInteger(MinBitWidth);
250     if (!DstTy->isVectorTy() && FromLegal && !ToLegal)
251       return OrigBitWidth;
252   }
253   return MinBitWidth;
254 }
255 
256 Type *TruncInstCombine::getBestTruncatedType() {
257   if (!buildTruncExpressionDag())
258     return nullptr;
259 
260   // We don't want to duplicate instructions, which isn't profitable. Thus, we
261   // can't shrink something that has multiple users, unless all users are
262   // post-dominated by the trunc instruction, i.e., were visited during the
263   // expression evaluation.
264   unsigned DesiredBitWidth = 0;
265   for (auto Itr : InstInfoMap) {
266     Instruction *I = Itr.first;
267     if (I->hasOneUse())
268       continue;
269     bool IsExtInst = (isa<ZExtInst>(I) || isa<SExtInst>(I));
270     for (auto *U : I->users())
271       if (auto *UI = dyn_cast<Instruction>(U))
272         if (UI != CurrentTruncInst && !InstInfoMap.count(UI)) {
273           if (!IsExtInst)
274             return nullptr;
275           // If this is an extension from the dest type, we can eliminate it,
276           // even if it has multiple users. Thus, update the DesiredBitWidth and
277           // validate all extension instructions agrees on same DesiredBitWidth.
278           unsigned ExtInstBitWidth =
279               I->getOperand(0)->getType()->getScalarSizeInBits();
280           if (DesiredBitWidth && DesiredBitWidth != ExtInstBitWidth)
281             return nullptr;
282           DesiredBitWidth = ExtInstBitWidth;
283         }
284   }
285 
286   unsigned OrigBitWidth =
287       CurrentTruncInst->getOperand(0)->getType()->getScalarSizeInBits();
288 
289   // Initialize MinBitWidth for shift instructions with the minimum number
290   // that is greater than shift amount (i.e. shift amount + 1).
291   // For `lshr` adjust MinBitWidth so that all potentially truncated
292   // bits of the value-to-be-shifted are zeros.
293   // For `ashr` adjust MinBitWidth so that all potentially truncated
294   // bits of the value-to-be-shifted are sign bits (all zeros or ones)
295   // and even one (first) untruncated bit is sign bit.
296   // Exit early if MinBitWidth is not less than original bitwidth.
297   for (auto &Itr : InstInfoMap) {
298     Instruction *I = Itr.first;
299     if (I->isShift()) {
300       KnownBits KnownRHS = computeKnownBits(I->getOperand(1));
301       unsigned MinBitWidth = KnownRHS.getMaxValue()
302                                  .uadd_sat(APInt(OrigBitWidth, 1))
303                                  .getLimitedValue(OrigBitWidth);
304       if (MinBitWidth == OrigBitWidth)
305         return nullptr;
306       if (I->getOpcode() == Instruction::LShr) {
307         KnownBits KnownLHS = computeKnownBits(I->getOperand(0));
308         MinBitWidth =
309             std::max(MinBitWidth, KnownLHS.getMaxValue().getActiveBits());
310       }
311       if (I->getOpcode() == Instruction::AShr) {
312         unsigned NumSignBits = ComputeNumSignBits(I->getOperand(0));
313         MinBitWidth = std::max(MinBitWidth, OrigBitWidth - NumSignBits + 1);
314       }
315       if (MinBitWidth >= OrigBitWidth)
316         return nullptr;
317       Itr.second.MinBitWidth = MinBitWidth;
318     }
319     if (I->getOpcode() == Instruction::UDiv ||
320         I->getOpcode() == Instruction::URem) {
321       unsigned MinBitWidth = 0;
322       for (const auto &Op : I->operands()) {
323         KnownBits Known = computeKnownBits(Op);
324         MinBitWidth =
325             std::max(Known.getMaxValue().getActiveBits(), MinBitWidth);
326         if (MinBitWidth >= OrigBitWidth)
327           return nullptr;
328       }
329       Itr.second.MinBitWidth = MinBitWidth;
330     }
331   }
332 
333   // Calculate minimum allowed bit-width allowed for shrinking the currently
334   // visited truncate's operand.
335   unsigned MinBitWidth = getMinBitWidth();
336 
337   // Check that we can shrink to smaller bit-width than original one and that
338   // it is similar to the DesiredBitWidth is such exists.
339   if (MinBitWidth >= OrigBitWidth ||
340       (DesiredBitWidth && DesiredBitWidth != MinBitWidth))
341     return nullptr;
342 
343   return IntegerType::get(CurrentTruncInst->getContext(), MinBitWidth);
344 }
345 
346 /// Given a reduced scalar type \p Ty and a \p V value, return a reduced type
347 /// for \p V, according to its type, if it vector type, return the vector
348 /// version of \p Ty, otherwise return \p Ty.
349 static Type *getReducedType(Value *V, Type *Ty) {
350   assert(Ty && !Ty->isVectorTy() && "Expect Scalar Type");
351   if (auto *VTy = dyn_cast<VectorType>(V->getType()))
352     return VectorType::get(Ty, VTy->getElementCount());
353   return Ty;
354 }
355 
356 Value *TruncInstCombine::getReducedOperand(Value *V, Type *SclTy) {
357   Type *Ty = getReducedType(V, SclTy);
358   if (auto *C = dyn_cast<Constant>(V)) {
359     C = ConstantExpr::getIntegerCast(C, Ty, false);
360     // If we got a constantexpr back, try to simplify it with DL info.
361     return ConstantFoldConstant(C, DL, &TLI);
362   }
363 
364   auto *I = cast<Instruction>(V);
365   Info Entry = InstInfoMap.lookup(I);
366   assert(Entry.NewValue);
367   return Entry.NewValue;
368 }
369 
370 void TruncInstCombine::ReduceExpressionDag(Type *SclTy) {
371   NumInstrsReduced += InstInfoMap.size();
372   for (auto &Itr : InstInfoMap) { // Forward
373     Instruction *I = Itr.first;
374     TruncInstCombine::Info &NodeInfo = Itr.second;
375 
376     assert(!NodeInfo.NewValue && "Instruction has been evaluated");
377 
378     IRBuilder<> Builder(I);
379     Value *Res = nullptr;
380     unsigned Opc = I->getOpcode();
381     switch (Opc) {
382     case Instruction::Trunc:
383     case Instruction::ZExt:
384     case Instruction::SExt: {
385       Type *Ty = getReducedType(I, SclTy);
386       // If the source type of the cast is the type we're trying for then we can
387       // just return the source.  There's no need to insert it because it is not
388       // new.
389       if (I->getOperand(0)->getType() == Ty) {
390         assert(!isa<TruncInst>(I) && "Cannot reach here with TruncInst");
391         NodeInfo.NewValue = I->getOperand(0);
392         continue;
393       }
394       // Otherwise, must be the same type of cast, so just reinsert a new one.
395       // This also handles the case of zext(trunc(x)) -> zext(x).
396       Res = Builder.CreateIntCast(I->getOperand(0), Ty,
397                                   Opc == Instruction::SExt);
398 
399       // Update Worklist entries with new value if needed.
400       // There are three possible changes to the Worklist:
401       // 1. Update Old-TruncInst -> New-TruncInst.
402       // 2. Remove Old-TruncInst (if New node is not TruncInst).
403       // 3. Add New-TruncInst (if Old node was not TruncInst).
404       auto *Entry = find(Worklist, I);
405       if (Entry != Worklist.end()) {
406         if (auto *NewCI = dyn_cast<TruncInst>(Res))
407           *Entry = NewCI;
408         else
409           Worklist.erase(Entry);
410       } else if (auto *NewCI = dyn_cast<TruncInst>(Res))
411           Worklist.push_back(NewCI);
412       break;
413     }
414     case Instruction::Add:
415     case Instruction::Sub:
416     case Instruction::Mul:
417     case Instruction::And:
418     case Instruction::Or:
419     case Instruction::Xor:
420     case Instruction::Shl:
421     case Instruction::LShr:
422     case Instruction::AShr:
423     case Instruction::UDiv:
424     case Instruction::URem: {
425       Value *LHS = getReducedOperand(I->getOperand(0), SclTy);
426       Value *RHS = getReducedOperand(I->getOperand(1), SclTy);
427       Res = Builder.CreateBinOp((Instruction::BinaryOps)Opc, LHS, RHS);
428       // Preserve `exact` flag since truncation doesn't change exactness
429       if (auto *PEO = dyn_cast<PossiblyExactOperator>(I))
430         if (auto *ResI = dyn_cast<Instruction>(Res))
431           ResI->setIsExact(PEO->isExact());
432       break;
433     }
434     case Instruction::ExtractElement: {
435       Value *Vec = getReducedOperand(I->getOperand(0), SclTy);
436       Value *Idx = I->getOperand(1);
437       Res = Builder.CreateExtractElement(Vec, Idx);
438       break;
439     }
440     case Instruction::InsertElement: {
441       Value *Vec = getReducedOperand(I->getOperand(0), SclTy);
442       Value *NewElt = getReducedOperand(I->getOperand(1), SclTy);
443       Value *Idx = I->getOperand(2);
444       Res = Builder.CreateInsertElement(Vec, NewElt, Idx);
445       break;
446     }
447     case Instruction::Select: {
448       Value *Op0 = I->getOperand(0);
449       Value *LHS = getReducedOperand(I->getOperand(1), SclTy);
450       Value *RHS = getReducedOperand(I->getOperand(2), SclTy);
451       Res = Builder.CreateSelect(Op0, LHS, RHS);
452       break;
453     }
454     default:
455       llvm_unreachable("Unhandled instruction");
456     }
457 
458     NodeInfo.NewValue = Res;
459     if (auto *ResI = dyn_cast<Instruction>(Res))
460       ResI->takeName(I);
461   }
462 
463   Value *Res = getReducedOperand(CurrentTruncInst->getOperand(0), SclTy);
464   Type *DstTy = CurrentTruncInst->getType();
465   if (Res->getType() != DstTy) {
466     IRBuilder<> Builder(CurrentTruncInst);
467     Res = Builder.CreateIntCast(Res, DstTy, false);
468     if (auto *ResI = dyn_cast<Instruction>(Res))
469       ResI->takeName(CurrentTruncInst);
470   }
471   CurrentTruncInst->replaceAllUsesWith(Res);
472 
473   // Erase old expression dag, which was replaced by the reduced expression dag.
474   // We iterate backward, which means we visit the instruction before we visit
475   // any of its operands, this way, when we get to the operand, we already
476   // removed the instructions (from the expression dag) that uses it.
477   CurrentTruncInst->eraseFromParent();
478   for (auto &I : llvm::reverse(InstInfoMap)) {
479     // We still need to check that the instruction has no users before we erase
480     // it, because {SExt, ZExt}Inst Instruction might have other users that was
481     // not reduced, in such case, we need to keep that instruction.
482     if (I.first->use_empty())
483       I.first->eraseFromParent();
484   }
485 }
486 
487 bool TruncInstCombine::run(Function &F) {
488   bool MadeIRChange = false;
489 
490   // Collect all TruncInst in the function into the Worklist for evaluating.
491   for (auto &BB : F) {
492     // Ignore unreachable basic block.
493     if (!DT.isReachableFromEntry(&BB))
494       continue;
495     for (auto &I : BB)
496       if (auto *CI = dyn_cast<TruncInst>(&I))
497         Worklist.push_back(CI);
498   }
499 
500   // Process all TruncInst in the Worklist, for each instruction:
501   //   1. Check if it dominates an eligible expression dag to be reduced.
502   //   2. Create a reduced expression dag and replace the old one with it.
503   while (!Worklist.empty()) {
504     CurrentTruncInst = Worklist.pop_back_val();
505 
506     if (Type *NewDstSclTy = getBestTruncatedType()) {
507       LLVM_DEBUG(
508           dbgs() << "ICE: TruncInstCombine reducing type of expression dag "
509                     "dominated by: "
510                  << CurrentTruncInst << '\n');
511       ReduceExpressionDag(NewDstSclTy);
512       ++NumDAGsReduced;
513       MadeIRChange = true;
514     }
515   }
516 
517   return MadeIRChange;
518 }
519