1
//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This transformation analyzes and transforms the induction variables (and
11
// computations derived from them) into simpler forms suitable for subsequent
12
// analysis and transformation.
14
// If the trip count of a loop is computable, this pass also makes the following
16
// 1. The exit condition for the loop is canonicalized to compare the
17
// induction value against the exit value. This turns loops like:
18
// 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
19
// 2. Any use outside of the loop of an expression derived from the indvar
20
// is changed to compute the derived value outside of the loop, eliminating
21
// the dependence on the exit value of the induction variable. If the only
22
// purpose of the loop is to compute the exit value of some derived
23
// expression, this transformation will make the loop dead.
25
//===----------------------------------------------------------------------===//
27
#include "llvm/Transforms/Scalar.h"
28
#include "llvm/ADT/DenseMap.h"
29
#include "llvm/ADT/SmallVector.h"
30
#include "llvm/ADT/Statistic.h"
31
#include "llvm/Analysis/LoopInfo.h"
32
#include "llvm/Analysis/LoopPass.h"
33
#include "llvm/Analysis/ScalarEvolutionExpander.h"
34
#include "llvm/Analysis/TargetLibraryInfo.h"
35
#include "llvm/Analysis/TargetTransformInfo.h"
36
#include "llvm/IR/BasicBlock.h"
37
#include "llvm/IR/CFG.h"
38
#include "llvm/IR/Constants.h"
39
#include "llvm/IR/DataLayout.h"
40
#include "llvm/IR/Dominators.h"
41
#include "llvm/IR/Instructions.h"
42
#include "llvm/IR/IntrinsicInst.h"
43
#include "llvm/IR/LLVMContext.h"
44
#include "llvm/IR/PatternMatch.h"
45
#include "llvm/IR/Type.h"
46
#include "llvm/Support/CommandLine.h"
47
#include "llvm/Support/Debug.h"
48
#include "llvm/Support/raw_ostream.h"
49
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
50
#include "llvm/Transforms/Utils/Local.h"
51
#include "llvm/Transforms/Utils/SimplifyIndVar.h"
54
#define DEBUG_TYPE "indvars"
56
STATISTIC(NumWidened , "Number of indvars widened");
57
STATISTIC(NumReplaced , "Number of exit values replaced");
58
STATISTIC(NumLFTR , "Number of loop exit tests replaced");
59
STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
60
STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
62
// Trip count verification can be enabled by default under NDEBUG if we
63
// implement a strong expression equivalence checker in SCEV. Until then, we
64
// use the verify-indvars flag, which may assert in some cases.
65
static cl::opt<bool> VerifyIndvars(
66
"verify-indvars", cl::Hidden,
67
cl::desc("Verify the ScalarEvolution result after running indvars"));
69
static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
70
cl::desc("Reduce live induction variables."));
72
enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl };
74
static cl::opt<ReplaceExitVal> ReplaceExitValue(
75
"replexitval", cl::Hidden, cl::init(OnlyCheapRepl),
76
cl::desc("Choose the strategy to replace exit value in IndVarSimplify"),
77
cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"),
78
clEnumValN(OnlyCheapRepl, "cheap",
79
"only replace exit value when the cost is cheap"),
80
clEnumValN(AlwaysRepl, "always",
81
"always replace exit value whenever possible"),
89
class IndVarSimplify : public LoopPass {
93
TargetLibraryInfo *TLI;
94
const TargetTransformInfo *TTI;
96
SmallVector<WeakVH, 16> DeadInsts;
100
static char ID; // Pass identification, replacement for typeid
102
: LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
103
initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
106
bool runOnLoop(Loop *L, LPPassManager &LPM) override;
108
void getAnalysisUsage(AnalysisUsage &AU) const override {
109
AU.addRequired<DominatorTreeWrapperPass>();
110
AU.addRequired<LoopInfoWrapperPass>();
111
AU.addRequired<ScalarEvolution>();
112
AU.addRequiredID(LoopSimplifyID);
113
AU.addRequiredID(LCSSAID);
114
AU.addPreserved<ScalarEvolution>();
115
AU.addPreservedID(LoopSimplifyID);
116
AU.addPreservedID(LCSSAID);
117
AU.setPreservesCFG();
121
void releaseMemory() override {
125
bool isValidRewrite(Value *FromVal, Value *ToVal);
127
void HandleFloatingPointIV(Loop *L, PHINode *PH);
128
void RewriteNonIntegerIVs(Loop *L);
130
void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
132
bool CanLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet);
133
void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
135
Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
136
PHINode *IndVar, SCEVExpander &Rewriter);
138
void SinkUnusedInvariants(Loop *L);
140
Value *ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L,
141
Instruction *InsertPt, Type *Ty,
142
bool &IsHighCostExpansion);
146
char IndVarSimplify::ID = 0;
147
INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
148
"Induction Variable Simplification", false, false)
149
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
150
INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
151
INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
152
INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
153
INITIALIZE_PASS_DEPENDENCY(LCSSA)
154
INITIALIZE_PASS_END(IndVarSimplify, "indvars",
155
"Induction Variable Simplification", false, false)
157
Pass *llvm::createIndVarSimplifyPass() {
158
return new IndVarSimplify();
161
/// isValidRewrite - Return true if the SCEV expansion generated by the
162
/// rewriter can replace the original value. SCEV guarantees that it
163
/// produces the same value, but the way it is produced may be illegal IR.
164
/// Ideally, this function will only be called for verification.
165
bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
166
// If an SCEV expression subsumed multiple pointers, its expansion could
167
// reassociate the GEP changing the base pointer. This is illegal because the
168
// final address produced by a GEP chain must be inbounds relative to its
169
// underlying object. Otherwise basic alias analysis, among other things,
170
// could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
171
// producing an expression involving multiple pointers. Until then, we must
174
// Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
175
// because it understands lcssa phis while SCEV does not.
176
Value *FromPtr = FromVal;
177
Value *ToPtr = ToVal;
178
if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
179
FromPtr = GEP->getPointerOperand();
181
if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
182
ToPtr = GEP->getPointerOperand();
184
if (FromPtr != FromVal || ToPtr != ToVal) {
185
// Quickly check the common case
186
if (FromPtr == ToPtr)
189
// SCEV may have rewritten an expression that produces the GEP's pointer
190
// operand. That's ok as long as the pointer operand has the same base
191
// pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
192
// base of a recurrence. This handles the case in which SCEV expansion
193
// converts a pointer type recurrence into a nonrecurrent pointer base
194
// indexed by an integer recurrence.
196
// If the GEP base pointer is a vector of pointers, abort.
197
if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
200
const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
201
const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
202
if (FromBase == ToBase)
205
DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
206
<< *FromBase << " != " << *ToBase << "\n");
213
/// Determine the insertion point for this user. By default, insert immediately
214
/// before the user. SCEVExpander or LICM will hoist loop invariants out of the
215
/// loop. For PHI nodes, there may be multiple uses, so compute the nearest
216
/// common dominator for the incoming blocks.
217
static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
219
PHINode *PHI = dyn_cast<PHINode>(User);
223
Instruction *InsertPt = nullptr;
224
for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
225
if (PHI->getIncomingValue(i) != Def)
228
BasicBlock *InsertBB = PHI->getIncomingBlock(i);
230
InsertPt = InsertBB->getTerminator();
233
InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
234
InsertPt = InsertBB->getTerminator();
236
assert(InsertPt && "Missing phi operand");
237
assert((!isa<Instruction>(Def) ||
238
DT->dominates(cast<Instruction>(Def), InsertPt)) &&
239
"def does not dominate all uses");
243
//===----------------------------------------------------------------------===//
244
// RewriteNonIntegerIVs and helpers. Prefer integer IVs.
245
//===----------------------------------------------------------------------===//
247
/// ConvertToSInt - Convert APF to an integer, if possible.
248
static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
249
bool isExact = false;
250
// See if we can convert this to an int64_t
252
if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
253
&isExact) != APFloat::opOK || !isExact)
259
/// HandleFloatingPointIV - If the loop has floating induction variable
260
/// then insert corresponding integer induction variable if possible.
262
/// for(double i = 0; i < 10000; ++i)
264
/// is converted into
265
/// for(int i = 0; i < 10000; ++i)
268
void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
269
unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
270
unsigned BackEdge = IncomingEdge^1;
272
// Check incoming value.
273
ConstantFP *InitValueVal =
274
dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
277
if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
280
// Check IV increment. Reject this PN if increment operation is not
281
// an add or increment value can not be represented by an integer.
282
BinaryOperator *Incr =
283
dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
284
if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
286
// If this is not an add of the PHI with a constantfp, or if the constant fp
287
// is not an integer, bail out.
288
ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
290
if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
291
!ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
294
// Check Incr uses. One user is PN and the other user is an exit condition
295
// used by the conditional terminator.
296
Value::user_iterator IncrUse = Incr->user_begin();
297
Instruction *U1 = cast<Instruction>(*IncrUse++);
298
if (IncrUse == Incr->user_end()) return;
299
Instruction *U2 = cast<Instruction>(*IncrUse++);
300
if (IncrUse != Incr->user_end()) return;
302
// Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
303
// only used by a branch, we can't transform it.
304
FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
306
Compare = dyn_cast<FCmpInst>(U2);
307
if (!Compare || !Compare->hasOneUse() ||
308
!isa<BranchInst>(Compare->user_back()))
311
BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
313
// We need to verify that the branch actually controls the iteration count
314
// of the loop. If not, the new IV can overflow and no one will notice.
315
// The branch block must be in the loop and one of the successors must be out
317
assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
318
if (!L->contains(TheBr->getParent()) ||
319
(L->contains(TheBr->getSuccessor(0)) &&
320
L->contains(TheBr->getSuccessor(1))))
324
// If it isn't a comparison with an integer-as-fp (the exit value), we can't
326
ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
328
if (ExitValueVal == nullptr ||
329
!ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
332
// Find new predicate for integer comparison.
333
CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
334
switch (Compare->getPredicate()) {
335
default: return; // Unknown comparison.
336
case CmpInst::FCMP_OEQ:
337
case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
338
case CmpInst::FCMP_ONE:
339
case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
340
case CmpInst::FCMP_OGT:
341
case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
342
case CmpInst::FCMP_OGE:
343
case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
344
case CmpInst::FCMP_OLT:
345
case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
346
case CmpInst::FCMP_OLE:
347
case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
350
// We convert the floating point induction variable to a signed i32 value if
351
// we can. This is only safe if the comparison will not overflow in a way
352
// that won't be trapped by the integer equivalent operations. Check for this
354
// TODO: We could use i64 if it is native and the range requires it.
356
// The start/stride/exit values must all fit in signed i32.
357
if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
360
// If not actually striding (add x, 0.0), avoid touching the code.
364
// Positive and negative strides have different safety conditions.
366
// If we have a positive stride, we require the init to be less than the
368
if (InitValue >= ExitValue)
371
uint32_t Range = uint32_t(ExitValue-InitValue);
372
// Check for infinite loop, either:
373
// while (i <= Exit) or until (i > Exit)
374
if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
375
if (++Range == 0) return; // Range overflows.
378
unsigned Leftover = Range % uint32_t(IncValue);
380
// If this is an equality comparison, we require that the strided value
381
// exactly land on the exit value, otherwise the IV condition will wrap
382
// around and do things the fp IV wouldn't.
383
if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
387
// If the stride would wrap around the i32 before exiting, we can't
389
if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
393
// If we have a negative stride, we require the init to be greater than the
395
if (InitValue <= ExitValue)
398
uint32_t Range = uint32_t(InitValue-ExitValue);
399
// Check for infinite loop, either:
400
// while (i >= Exit) or until (i < Exit)
401
if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
402
if (++Range == 0) return; // Range overflows.
405
unsigned Leftover = Range % uint32_t(-IncValue);
407
// If this is an equality comparison, we require that the strided value
408
// exactly land on the exit value, otherwise the IV condition will wrap
409
// around and do things the fp IV wouldn't.
410
if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
414
// If the stride would wrap around the i32 before exiting, we can't
416
if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
420
IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
422
// Insert new integer induction variable.
423
PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
424
NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
425
PN->getIncomingBlock(IncomingEdge));
428
BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
429
Incr->getName()+".int", Incr);
430
NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
432
ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
433
ConstantInt::get(Int32Ty, ExitValue),
436
// In the following deletions, PN may become dead and may be deleted.
437
// Use a WeakVH to observe whether this happens.
440
// Delete the old floating point exit comparison. The branch starts using the
442
NewCompare->takeName(Compare);
443
Compare->replaceAllUsesWith(NewCompare);
444
RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
446
// Delete the old floating point increment.
447
Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
448
RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
450
// If the FP induction variable still has uses, this is because something else
451
// in the loop uses its value. In order to canonicalize the induction
452
// variable, we chose to eliminate the IV and rewrite it in terms of an
455
// We give preference to sitofp over uitofp because it is faster on most
458
Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
459
PN->getParent()->getFirstInsertionPt());
460
PN->replaceAllUsesWith(Conv);
461
RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
466
void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
467
// First step. Check to see if there are any floating-point recurrences.
468
// If there are, change them into integer recurrences, permitting analysis by
469
// the SCEV routines.
471
BasicBlock *Header = L->getHeader();
473
SmallVector<WeakVH, 8> PHIs;
474
for (BasicBlock::iterator I = Header->begin();
475
PHINode *PN = dyn_cast<PHINode>(I); ++I)
478
for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
479
if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
480
HandleFloatingPointIV(L, PN);
482
// If the loop previously had floating-point IV, ScalarEvolution
483
// may not have been able to compute a trip count. Now that we've done some
484
// re-writing, the trip count may be computable.
490
// Collect information about PHI nodes which can be transformed in
491
// RewriteLoopExitValues.
494
unsigned Ith; // Ith incoming value.
495
Value *Val; // Exit value after expansion.
496
bool HighCost; // High Cost when expansion.
497
bool SafePhi; // LCSSASafePhiForRAUW.
499
RewritePhi(PHINode *P, unsigned I, Value *V, bool H, bool S)
500
: PN(P), Ith(I), Val(V), HighCost(H), SafePhi(S) {}
504
Value *IndVarSimplify::ExpandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S,
505
Loop *L, Instruction *InsertPt,
507
bool &IsHighCostExpansion) {
508
using namespace llvm::PatternMatch;
510
if (!Rewriter.isHighCostExpansion(S, L)) {
511
IsHighCostExpansion = false;
512
return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
515
// Before expanding S into an expensive LLVM expression, see if we can use an
516
// already existing value as the expansion for S. There is potential to make
517
// this significantly smarter, but this simple heuristic already gets some
518
// interesting cases.
520
SmallVector<BasicBlock *, 4> Latches;
521
L->getLoopLatches(Latches);
523
for (BasicBlock *BB : Latches) {
524
ICmpInst::Predicate Pred;
525
Instruction *LHS, *RHS;
526
BasicBlock *TrueBB, *FalseBB;
528
if (!match(BB->getTerminator(),
529
m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)),
533
if (SE->getSCEV(LHS) == S && DT->dominates(LHS, InsertPt)) {
534
IsHighCostExpansion = false;
538
if (SE->getSCEV(RHS) == S && DT->dominates(RHS, InsertPt)) {
539
IsHighCostExpansion = false;
544
// We didn't find anything, fall back to using SCEVExpander.
545
assert(Rewriter.isHighCostExpansion(S, L) && "this should not have changed!");
546
IsHighCostExpansion = true;
547
return Rewriter.expandCodeFor(S, ResultTy, InsertPt);
550
//===----------------------------------------------------------------------===//
551
// RewriteLoopExitValues - Optimize IV users outside the loop.
552
// As a side effect, reduces the amount of IV processing within the loop.
553
//===----------------------------------------------------------------------===//
555
/// RewriteLoopExitValues - Check to see if this loop has a computable
556
/// loop-invariant execution count. If so, this means that we can compute the
557
/// final value of any expressions that are recurrent in the loop, and
558
/// substitute the exit values from the loop into any instructions outside of
559
/// the loop that use the final values of the current expressions.
561
/// This is mostly redundant with the regular IndVarSimplify activities that
562
/// happen later, except that it's more powerful in some cases, because it's
563
/// able to brute-force evaluate arbitrary instructions as long as they have
564
/// constant operands at the beginning of the loop.
565
void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
566
// Verify the input to the pass in already in LCSSA form.
567
assert(L->isLCSSAForm(*DT));
569
SmallVector<BasicBlock*, 8> ExitBlocks;
570
L->getUniqueExitBlocks(ExitBlocks);
572
SmallVector<RewritePhi, 8> RewritePhiSet;
573
// Find all values that are computed inside the loop, but used outside of it.
574
// Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
575
// the exit blocks of the loop to find them.
576
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
577
BasicBlock *ExitBB = ExitBlocks[i];
579
// If there are no PHI nodes in this exit block, then no values defined
580
// inside the loop are used on this path, skip it.
581
PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
584
unsigned NumPreds = PN->getNumIncomingValues();
586
// We would like to be able to RAUW single-incoming value PHI nodes. We
587
// have to be certain this is safe even when this is an LCSSA PHI node.
588
// While the computed exit value is no longer varying in *this* loop, the
589
// exit block may be an exit block for an outer containing loop as well,
590
// the exit value may be varying in the outer loop, and thus it may still
591
// require an LCSSA PHI node. The safe case is when this is
592
// single-predecessor PHI node (LCSSA) and the exit block containing it is
593
// part of the enclosing loop, or this is the outer most loop of the nest.
594
// In either case the exit value could (at most) be varying in the same
595
// loop body as the phi node itself. Thus if it is in turn used outside of
596
// an enclosing loop it will only be via a separate LCSSA node.
597
bool LCSSASafePhiForRAUW =
599
(!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
601
// Iterate over all of the PHI nodes.
602
BasicBlock::iterator BBI = ExitBB->begin();
603
while ((PN = dyn_cast<PHINode>(BBI++))) {
605
continue; // dead use, don't replace it
607
// SCEV only supports integer expressions for now.
608
if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
611
// It's necessary to tell ScalarEvolution about this explicitly so that
612
// it can walk the def-use list and forget all SCEVs, as it may not be
613
// watching the PHI itself. Once the new exit value is in place, there
614
// may not be a def-use connection between the loop and every instruction
615
// which got a SCEVAddRecExpr for that loop.
618
// Iterate over all of the values in all the PHI nodes.
619
for (unsigned i = 0; i != NumPreds; ++i) {
620
// If the value being merged in is not integer or is not defined
621
// in the loop, skip it.
622
Value *InVal = PN->getIncomingValue(i);
623
if (!isa<Instruction>(InVal))
626
// If this pred is for a subloop, not L itself, skip it.
627
if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
628
continue; // The Block is in a subloop, skip it.
630
// Check that InVal is defined in the loop.
631
Instruction *Inst = cast<Instruction>(InVal);
632
if (!L->contains(Inst))
635
// Okay, this instruction has a user outside of the current loop
636
// and varies predictably *inside* the loop. Evaluate the value it
637
// contains when the loop exits, if possible.
638
const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
639
if (!SE->isLoopInvariant(ExitValue, L) ||
640
!isSafeToExpand(ExitValue, *SE))
643
// Computing the value outside of the loop brings no benefit if :
644
// - it is definitely used inside the loop in a way which can not be
646
// - no use outside of the loop can take advantage of hoisting the
647
// computation out of the loop
648
if (ExitValue->getSCEVType()>=scMulExpr) {
649
unsigned NumHardInternalUses = 0;
650
unsigned NumSoftExternalUses = 0;
651
unsigned NumUses = 0;
652
for (auto IB = Inst->user_begin(), IE = Inst->user_end();
653
IB != IE && NumUses <= 6; ++IB) {
654
Instruction *UseInstr = cast<Instruction>(*IB);
655
unsigned Opc = UseInstr->getOpcode();
657
if (L->contains(UseInstr)) {
658
if (Opc == Instruction::Call || Opc == Instruction::Ret)
659
NumHardInternalUses++;
661
if (Opc == Instruction::PHI) {
662
// Do not count the Phi as a use. LCSSA may have inserted
663
// plenty of trivial ones.
665
for (auto PB = UseInstr->user_begin(),
666
PE = UseInstr->user_end();
667
PB != PE && NumUses <= 6; ++PB, ++NumUses) {
668
unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
669
if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
670
NumSoftExternalUses++;
674
if (Opc != Instruction::Call && Opc != Instruction::Ret)
675
NumSoftExternalUses++;
678
if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
682
bool HighCost = false;
683
Value *ExitVal = ExpandSCEVIfNeeded(Rewriter, ExitValue, L, Inst,
684
PN->getType(), HighCost);
686
DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
687
<< " LoopVal = " << *Inst << "\n");
689
if (!isValidRewrite(Inst, ExitVal)) {
690
DeadInsts.push_back(ExitVal);
694
// Collect all the candidate PHINodes to be rewritten.
695
RewritePhiSet.push_back(
696
RewritePhi(PN, i, ExitVal, HighCost, LCSSASafePhiForRAUW));
701
bool LoopCanBeDel = CanLoopBeDeleted(L, RewritePhiSet);
704
for (const RewritePhi &Phi : RewritePhiSet) {
705
PHINode *PN = Phi.PN;
706
Value *ExitVal = Phi.Val;
708
// Only do the rewrite when the ExitValue can be expanded cheaply.
709
// If LoopCanBeDel is true, rewrite exit value aggressively.
710
if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) {
711
DeadInsts.push_back(ExitVal);
717
Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith));
718
PN->setIncomingValue(Phi.Ith, ExitVal);
720
// If this instruction is dead now, delete it. Don't do it now to avoid
721
// invalidating iterators.
722
if (isInstructionTriviallyDead(Inst, TLI))
723
DeadInsts.push_back(Inst);
725
// If we determined that this PHI is safe to replace even if an LCSSA
728
PN->replaceAllUsesWith(ExitVal);
729
PN->eraseFromParent();
733
// The insertion point instruction may have been deleted; clear it out
734
// so that the rewriter doesn't trip over it later.
735
Rewriter.clearInsertPoint();
738
/// CanLoopBeDeleted - Check whether it is possible to delete the loop after
739
/// rewriting exit value. If it is possible, ignore ReplaceExitValue and
740
/// do rewriting aggressively.
741
bool IndVarSimplify::CanLoopBeDeleted(
742
Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) {
744
BasicBlock *Preheader = L->getLoopPreheader();
745
// If there is no preheader, the loop will not be deleted.
749
// In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1.
750
// We obviate multiple ExitingBlocks case for simplicity.
751
// TODO: If we see testcase with multiple ExitingBlocks can be deleted
752
// after exit value rewriting, we can enhance the logic here.
753
SmallVector<BasicBlock *, 4> ExitingBlocks;
754
L->getExitingBlocks(ExitingBlocks);
755
SmallVector<BasicBlock *, 8> ExitBlocks;
756
L->getUniqueExitBlocks(ExitBlocks);
757
if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1)
760
BasicBlock *ExitBlock = ExitBlocks[0];
761
BasicBlock::iterator BI = ExitBlock->begin();
762
while (PHINode *P = dyn_cast<PHINode>(BI)) {
763
Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]);
765
// If the Incoming value of P is found in RewritePhiSet, we know it
766
// could be rewritten to use a loop invariant value in transformation
767
// phase later. Skip it in the loop invariant check below.
769
for (const RewritePhi &Phi : RewritePhiSet) {
770
unsigned i = Phi.Ith;
771
if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) {
778
if (!found && (I = dyn_cast<Instruction>(Incoming)))
779
if (!L->hasLoopInvariantOperands(I))
785
for (Loop::block_iterator LI = L->block_begin(), LE = L->block_end();
787
for (BasicBlock::iterator BI = (*LI)->begin(), BE = (*LI)->end(); BI != BE;
789
if (BI->mayHaveSideEffects())
797
//===----------------------------------------------------------------------===//
798
// IV Widening - Extend the width of an IV to cover its widest uses.
799
//===----------------------------------------------------------------------===//
802
// Collect information about induction variables that are used by sign/zero
803
// extend operations. This information is recorded by CollectExtend and
804
// provides the input to WidenIV.
807
Type *WidestNativeType; // Widest integer type created [sz]ext
808
bool IsSigned; // Was a sext user seen before a zext?
810
WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
815
/// visitCast - Update information about the induction variable that is
816
/// extended by this sign or zero extend operation. This is used to determine
817
/// the final width of the IV before actually widening it.
818
static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
819
const TargetTransformInfo *TTI) {
820
bool IsSigned = Cast->getOpcode() == Instruction::SExt;
821
if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
824
Type *Ty = Cast->getType();
825
uint64_t Width = SE->getTypeSizeInBits(Ty);
826
if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
829
// Cast is either an sext or zext up to this point.
830
// We should not widen an indvar if arithmetics on the wider indvar are more
831
// expensive than those on the narrower indvar. We check only the cost of ADD
832
// because at least an ADD is required to increment the induction variable. We
833
// could compute more comprehensively the cost of all instructions on the
834
// induction variable when necessary.
836
TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
837
TTI->getArithmeticInstrCost(Instruction::Add,
838
Cast->getOperand(0)->getType())) {
842
if (!WI.WidestNativeType) {
843
WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
844
WI.IsSigned = IsSigned;
848
// We extend the IV to satisfy the sign of its first user, arbitrarily.
849
if (WI.IsSigned != IsSigned)
852
if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
853
WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
858
/// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
859
/// WideIV that computes the same value as the Narrow IV def. This avoids
860
/// caching Use* pointers.
861
struct NarrowIVDefUse {
862
Instruction *NarrowDef;
863
Instruction *NarrowUse;
864
Instruction *WideDef;
866
NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
868
NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
869
NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
872
/// WidenIV - The goal of this transform is to remove sign and zero extends
873
/// without creating any new induction variables. To do this, it creates a new
874
/// phi of the wider type and redirects all users, either removing extends or
875
/// inserting truncs whenever we stop propagating the type.
891
Instruction *WideInc;
892
const SCEV *WideIncExpr;
893
SmallVectorImpl<WeakVH> &DeadInsts;
895
SmallPtrSet<Instruction*,16> Widened;
896
SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
899
WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
900
ScalarEvolution *SEv, DominatorTree *DTree,
901
SmallVectorImpl<WeakVH> &DI) :
902
OrigPhi(WI.NarrowIV),
903
WideType(WI.WidestNativeType),
904
IsSigned(WI.IsSigned),
906
L(LI->getLoopFor(OrigPhi->getParent())),
911
WideIncExpr(nullptr),
913
assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
916
PHINode *CreateWideIV(SCEVExpander &Rewriter);
919
Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
922
Instruction *CloneIVUser(NarrowIVDefUse DU);
924
const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
926
const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
928
const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
929
unsigned OpCode) const;
931
Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
933
bool WidenLoopCompare(NarrowIVDefUse DU);
935
void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
937
} // anonymous namespace
939
/// isLoopInvariant - Perform a quick domtree based check for loop invariance
940
/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
941
/// gratuitous for this purpose.
942
static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
943
Instruction *Inst = dyn_cast<Instruction>(V);
947
return DT->properlyDominates(Inst->getParent(), L->getHeader());
950
Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
952
// Set the debug location and conservative insertion point.
953
IRBuilder<> Builder(Use);
954
// Hoist the insertion point into loop preheaders as far as possible.
955
for (const Loop *L = LI->getLoopFor(Use->getParent());
956
L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
957
L = L->getParentLoop())
958
Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
960
return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
961
Builder.CreateZExt(NarrowOper, WideType);
964
/// CloneIVUser - Instantiate a wide operation to replace a narrow
965
/// operation. This only needs to handle operations that can evaluation to
966
/// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
967
Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
968
unsigned Opcode = DU.NarrowUse->getOpcode();
972
case Instruction::Add:
973
case Instruction::Mul:
974
case Instruction::UDiv:
975
case Instruction::Sub:
976
case Instruction::And:
977
case Instruction::Or:
978
case Instruction::Xor:
979
case Instruction::Shl:
980
case Instruction::LShr:
981
case Instruction::AShr:
982
DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
984
// Replace NarrowDef operands with WideDef. Otherwise, we don't know
985
// anything about the narrow operand yet so must insert a [sz]ext. It is
986
// probably loop invariant and will be folded or hoisted. If it actually
987
// comes from a widened IV, it should be removed during a future call to
989
Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
990
getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
991
Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
992
getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
994
BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
995
BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
997
NarrowBO->getName());
998
IRBuilder<> Builder(DU.NarrowUse);
999
Builder.Insert(WideBO);
1000
if (const OverflowingBinaryOperator *OBO =
1001
dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
1002
if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
1003
if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
1009
const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
1010
unsigned OpCode) const {
1011
if (OpCode == Instruction::Add)
1012
return SE->getAddExpr(LHS, RHS);
1013
if (OpCode == Instruction::Sub)
1014
return SE->getMinusSCEV(LHS, RHS);
1015
if (OpCode == Instruction::Mul)
1016
return SE->getMulExpr(LHS, RHS);
1018
llvm_unreachable("Unsupported opcode.");
1021
/// No-wrap operations can transfer sign extension of their result to their
1022
/// operands. Generate the SCEV value for the widened operation without
1023
/// actually modifying the IR yet. If the expression after extending the
1024
/// operands is an AddRec for this loop, return it.
1025
const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
1027
// Handle the common case of add<nsw/nuw>
1028
const unsigned OpCode = DU.NarrowUse->getOpcode();
1029
// Only Add/Sub/Mul instructions supported yet.
1030
if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
1031
OpCode != Instruction::Mul)
1034
// One operand (NarrowDef) has already been extended to WideDef. Now determine
1035
// if extending the other will lead to a recurrence.
1036
const unsigned ExtendOperIdx =
1037
DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
1038
assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
1040
const SCEV *ExtendOperExpr = nullptr;
1041
const OverflowingBinaryOperator *OBO =
1042
cast<OverflowingBinaryOperator>(DU.NarrowUse);
1043
if (IsSigned && OBO->hasNoSignedWrap())
1044
ExtendOperExpr = SE->getSignExtendExpr(
1045
SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1046
else if(!IsSigned && OBO->hasNoUnsignedWrap())
1047
ExtendOperExpr = SE->getZeroExtendExpr(
1048
SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
1052
// When creating this SCEV expr, don't apply the current operations NSW or NUW
1053
// flags. This instruction may be guarded by control flow that the no-wrap
1054
// behavior depends on. Non-control-equivalent instructions can be mapped to
1055
// the same SCEV expression, and it would be incorrect to transfer NSW/NUW
1056
// semantics to those operations.
1057
const SCEV *lhs = SE->getSCEV(DU.WideDef);
1058
const SCEV *rhs = ExtendOperExpr;
1060
// Let's swap operands to the initial order for the case of non-commutative
1061
// operations, like SUB. See PR21014.
1062
if (ExtendOperIdx == 0)
1063
std::swap(lhs, rhs);
1064
const SCEVAddRecExpr *AddRec =
1065
dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
1067
if (!AddRec || AddRec->getLoop() != L)
1072
/// GetWideRecurrence - Is this instruction potentially interesting for further
1073
/// simplification after widening it's type? In other words, can the
1074
/// extend be safely hoisted out of the loop with SCEV reducing the value to a
1075
/// recurrence on the same loop. If so, return the sign or zero extended
1076
/// recurrence. Otherwise return NULL.
1077
const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
1078
if (!SE->isSCEVable(NarrowUse->getType()))
1081
const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
1082
if (SE->getTypeSizeInBits(NarrowExpr->getType())
1083
>= SE->getTypeSizeInBits(WideType)) {
1084
// NarrowUse implicitly widens its operand. e.g. a gep with a narrow
1085
// index. So don't follow this use.
1089
const SCEV *WideExpr = IsSigned ?
1090
SE->getSignExtendExpr(NarrowExpr, WideType) :
1091
SE->getZeroExtendExpr(NarrowExpr, WideType);
1092
const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
1093
if (!AddRec || AddRec->getLoop() != L)
1098
/// This IV user cannot be widen. Replace this use of the original narrow IV
1099
/// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
1100
static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
1101
DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
1102
<< " for user " << *DU.NarrowUse << "\n");
1103
IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1104
Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
1105
DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
1108
/// If the narrow use is a compare instruction, then widen the compare
1109
// (and possibly the other operand). The extend operation is hoisted into the
1110
// loop preheader as far as possible.
1111
bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
1112
ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
1116
// Sign of IV user and compare must match.
1117
if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
1120
Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
1121
unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
1122
unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1123
assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
1125
// Widen the compare instruction.
1126
IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
1127
DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1129
// Widen the other operand of the compare, if necessary.
1130
if (CastWidth < IVWidth) {
1131
Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
1132
DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
1137
/// WidenIVUse - Determine whether an individual user of the narrow IV can be
1138
/// widened. If so, return the wide clone of the user.
1139
Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
1141
// Stop traversing the def-use chain at inner-loop phis or post-loop phis.
1142
if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
1143
if (LI->getLoopFor(UsePhi->getParent()) != L) {
1144
// For LCSSA phis, sink the truncate outside the loop.
1145
// After SimplifyCFG most loop exit targets have a single predecessor.
1146
// Otherwise fall back to a truncate within the loop.
1147
if (UsePhi->getNumOperands() != 1)
1148
truncateIVUse(DU, DT);
1151
PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
1153
WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
1154
IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
1155
Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
1156
UsePhi->replaceAllUsesWith(Trunc);
1157
DeadInsts.emplace_back(UsePhi);
1158
DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
1159
<< " to " << *WidePhi << "\n");
1164
// Our raison d'etre! Eliminate sign and zero extension.
1165
if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
1166
Value *NewDef = DU.WideDef;
1167
if (DU.NarrowUse->getType() != WideType) {
1168
unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
1169
unsigned IVWidth = SE->getTypeSizeInBits(WideType);
1170
if (CastWidth < IVWidth) {
1171
// The cast isn't as wide as the IV, so insert a Trunc.
1172
IRBuilder<> Builder(DU.NarrowUse);
1173
NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
1176
// A wider extend was hidden behind a narrower one. This may induce
1177
// another round of IV widening in which the intermediate IV becomes
1178
// dead. It should be very rare.
1179
DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
1180
<< " not wide enough to subsume " << *DU.NarrowUse << "\n");
1181
DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
1182
NewDef = DU.NarrowUse;
1185
if (NewDef != DU.NarrowUse) {
1186
DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
1187
<< " replaced by " << *DU.WideDef << "\n");
1189
DU.NarrowUse->replaceAllUsesWith(NewDef);
1190
DeadInsts.emplace_back(DU.NarrowUse);
1192
// Now that the extend is gone, we want to expose it's uses for potential
1193
// further simplification. We don't need to directly inform SimplifyIVUsers
1194
// of the new users, because their parent IV will be processed later as a
1195
// new loop phi. If we preserved IVUsers analysis, we would also want to
1196
// push the uses of WideDef here.
1198
// No further widening is needed. The deceased [sz]ext had done it for us.
1202
// Does this user itself evaluate to a recurrence after widening?
1203
const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
1205
WideAddRec = GetExtendedOperandRecurrence(DU);
1208
// If use is a loop condition, try to promote the condition instead of
1209
// truncating the IV first.
1210
if (WidenLoopCompare(DU))
1213
// This user does not evaluate to a recurence after widening, so don't
1214
// follow it. Instead insert a Trunc to kill off the original use,
1215
// eventually isolating the original narrow IV so it can be removed.
1216
truncateIVUse(DU, DT);
1219
// Assume block terminators cannot evaluate to a recurrence. We can't to
1220
// insert a Trunc after a terminator if there happens to be a critical edge.
1221
assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
1222
"SCEV is not expected to evaluate a block terminator");
1224
// Reuse the IV increment that SCEVExpander created as long as it dominates
1226
Instruction *WideUse = nullptr;
1227
if (WideAddRec == WideIncExpr
1228
&& Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
1231
WideUse = CloneIVUser(DU);
1235
// Evaluation of WideAddRec ensured that the narrow expression could be
1236
// extended outside the loop without overflow. This suggests that the wide use
1237
// evaluates to the same expression as the extended narrow use, but doesn't
1238
// absolutely guarantee it. Hence the following failsafe check. In rare cases
1239
// where it fails, we simply throw away the newly created wide use.
1240
if (WideAddRec != SE->getSCEV(WideUse)) {
1241
DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
1242
<< ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
1243
DeadInsts.emplace_back(WideUse);
1247
// Returning WideUse pushes it on the worklist.
1251
/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
1253
void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
1254
for (User *U : NarrowDef->users()) {
1255
Instruction *NarrowUser = cast<Instruction>(U);
1257
// Handle data flow merges and bizarre phi cycles.
1258
if (!Widened.insert(NarrowUser).second)
1261
NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
1265
/// CreateWideIV - Process a single induction variable. First use the
1266
/// SCEVExpander to create a wide induction variable that evaluates to the same
1267
/// recurrence as the original narrow IV. Then use a worklist to forward
1268
/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
1269
/// interesting IV users, the narrow IV will be isolated for removal by
1272
/// It would be simpler to delete uses as they are processed, but we must avoid
1273
/// invalidating SCEV expressions.
1275
PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
1276
// Is this phi an induction variable?
1277
const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
1281
// Widen the induction variable expression.
1282
const SCEV *WideIVExpr = IsSigned ?
1283
SE->getSignExtendExpr(AddRec, WideType) :
1284
SE->getZeroExtendExpr(AddRec, WideType);
1286
assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
1287
"Expect the new IV expression to preserve its type");
1289
// Can the IV be extended outside the loop without overflow?
1290
AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
1291
if (!AddRec || AddRec->getLoop() != L)
1294
// An AddRec must have loop-invariant operands. Since this AddRec is
1295
// materialized by a loop header phi, the expression cannot have any post-loop
1296
// operands, so they must dominate the loop header.
1297
assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
1298
SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
1299
&& "Loop header phi recurrence inputs do not dominate the loop");
1301
// The rewriter provides a value for the desired IV expression. This may
1302
// either find an existing phi or materialize a new one. Either way, we
1303
// expect a well-formed cyclic phi-with-increments. i.e. any operand not part
1304
// of the phi-SCC dominates the loop entry.
1305
Instruction *InsertPt = L->getHeader()->begin();
1306
WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
1308
// Remembering the WideIV increment generated by SCEVExpander allows
1309
// WidenIVUse to reuse it when widening the narrow IV's increment. We don't
1310
// employ a general reuse mechanism because the call above is the only call to
1311
// SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
1312
if (BasicBlock *LatchBlock = L->getLoopLatch()) {
1314
cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
1315
WideIncExpr = SE->getSCEV(WideInc);
1318
DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
1321
// Traverse the def-use chain using a worklist starting at the original IV.
1322
assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
1324
Widened.insert(OrigPhi);
1325
pushNarrowIVUsers(OrigPhi, WidePhi);
1327
while (!NarrowIVUsers.empty()) {
1328
NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
1330
// Process a def-use edge. This may replace the use, so don't hold a
1331
// use_iterator across it.
1332
Instruction *WideUse = WidenIVUse(DU, Rewriter);
1334
// Follow all def-use edges from the previous narrow use.
1336
pushNarrowIVUsers(DU.NarrowUse, WideUse);
1338
// WidenIVUse may have removed the def-use edge.
1339
if (DU.NarrowDef->use_empty())
1340
DeadInsts.emplace_back(DU.NarrowDef);
1345
//===----------------------------------------------------------------------===//
1346
// Live IV Reduction - Minimize IVs live across the loop.
1347
//===----------------------------------------------------------------------===//
1350
//===----------------------------------------------------------------------===//
1351
// Simplification of IV users based on SCEV evaluation.
1352
//===----------------------------------------------------------------------===//
1355
class IndVarSimplifyVisitor : public IVVisitor {
1356
ScalarEvolution *SE;
1357
const TargetTransformInfo *TTI;
1363
IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
1364
const TargetTransformInfo *TTI,
1365
const DominatorTree *DTree)
1366
: SE(SCEV), TTI(TTI), IVPhi(IV) {
1368
WI.NarrowIV = IVPhi;
1370
setSplitOverflowIntrinsics();
1373
// Implement the interface used by simplifyUsersOfIV.
1374
void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
1378
/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
1379
/// users. Each successive simplification may push more users which may
1380
/// themselves be candidates for simplification.
1382
/// Sign/Zero extend elimination is interleaved with IV simplification.
1384
void IndVarSimplify::SimplifyAndExtend(Loop *L,
1385
SCEVExpander &Rewriter,
1386
LPPassManager &LPM) {
1387
SmallVector<WideIVInfo, 8> WideIVs;
1389
SmallVector<PHINode*, 8> LoopPhis;
1390
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1391
LoopPhis.push_back(cast<PHINode>(I));
1393
// Each round of simplification iterates through the SimplifyIVUsers worklist
1394
// for all current phis, then determines whether any IVs can be
1395
// widened. Widening adds new phis to LoopPhis, inducing another round of
1396
// simplification on the wide IVs.
1397
while (!LoopPhis.empty()) {
1398
// Evaluate as many IV expressions as possible before widening any IVs. This
1399
// forces SCEV to set no-wrap flags before evaluating sign/zero
1400
// extension. The first time SCEV attempts to normalize sign/zero extension,
1401
// the result becomes final. So for the most predictable results, we delay
1402
// evaluation of sign/zero extend evaluation until needed, and avoid running
1403
// other SCEV based analysis prior to SimplifyAndExtend.
1405
PHINode *CurrIV = LoopPhis.pop_back_val();
1407
// Information about sign/zero extensions of CurrIV.
1408
IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
1410
Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
1412
if (Visitor.WI.WidestNativeType) {
1413
WideIVs.push_back(Visitor.WI);
1415
} while(!LoopPhis.empty());
1417
for (; !WideIVs.empty(); WideIVs.pop_back()) {
1418
WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
1419
if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
1421
LoopPhis.push_back(WidePhi);
1427
//===----------------------------------------------------------------------===//
1428
// LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
1429
//===----------------------------------------------------------------------===//
1431
/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
1432
/// count expression can be safely and cheaply expanded into an instruction
1433
/// sequence that can be used by LinearFunctionTestReplace.
1435
/// TODO: This fails for pointer-type loop counters with greater than one byte
1436
/// strides, consequently preventing LFTR from running. For the purpose of LFTR
1437
/// we could skip this check in the case that the LFTR loop counter (chosen by
1438
/// FindLoopCounter) is also pointer type. Instead, we could directly convert
1439
/// the loop test to an inequality test by checking the target data's alignment
1440
/// of element types (given that the initial pointer value originates from or is
1441
/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
1442
/// However, we don't yet have a strong motivation for converting loop tests
1443
/// into inequality tests.
1444
static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE,
1445
SCEVExpander &Rewriter) {
1446
const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
1447
if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
1448
BackedgeTakenCount->isZero())
1451
if (!L->getExitingBlock())
1454
// Can't rewrite non-branch yet.
1455
if (!isa<BranchInst>(L->getExitingBlock()->getTerminator()))
1458
if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L))
1464
/// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
1465
/// invariant value to the phi.
1466
static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
1467
Instruction *IncI = dyn_cast<Instruction>(IncV);
1471
switch (IncI->getOpcode()) {
1472
case Instruction::Add:
1473
case Instruction::Sub:
1475
case Instruction::GetElementPtr:
1476
// An IV counter must preserve its type.
1477
if (IncI->getNumOperands() == 2)
1483
PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
1484
if (Phi && Phi->getParent() == L->getHeader()) {
1485
if (isLoopInvariant(IncI->getOperand(1), L, DT))
1489
if (IncI->getOpcode() == Instruction::GetElementPtr)
1492
// Allow add/sub to be commuted.
1493
Phi = dyn_cast<PHINode>(IncI->getOperand(1));
1494
if (Phi && Phi->getParent() == L->getHeader()) {
1495
if (isLoopInvariant(IncI->getOperand(0), L, DT))
1501
/// Return the compare guarding the loop latch, or NULL for unrecognized tests.
1502
static ICmpInst *getLoopTest(Loop *L) {
1503
assert(L->getExitingBlock() && "expected loop exit");
1505
BasicBlock *LatchBlock = L->getLoopLatch();
1506
// Don't bother with LFTR if the loop is not properly simplified.
1510
BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
1511
assert(BI && "expected exit branch");
1513
return dyn_cast<ICmpInst>(BI->getCondition());
1516
/// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
1517
/// that the current exit test is already sufficiently canonical.
1518
static bool needsLFTR(Loop *L, DominatorTree *DT) {
1519
// Do LFTR to simplify the exit condition to an ICMP.
1520
ICmpInst *Cond = getLoopTest(L);
1524
// Do LFTR to simplify the exit ICMP to EQ/NE
1525
ICmpInst::Predicate Pred = Cond->getPredicate();
1526
if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
1529
// Look for a loop invariant RHS
1530
Value *LHS = Cond->getOperand(0);
1531
Value *RHS = Cond->getOperand(1);
1532
if (!isLoopInvariant(RHS, L, DT)) {
1533
if (!isLoopInvariant(LHS, L, DT))
1535
std::swap(LHS, RHS);
1537
// Look for a simple IV counter LHS
1538
PHINode *Phi = dyn_cast<PHINode>(LHS);
1540
Phi = getLoopPhiForCounter(LHS, L, DT);
1545
// Do LFTR if PHI node is defined in the loop, but is *not* a counter.
1546
int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
1550
// Do LFTR if the exit condition's IV is *not* a simple counter.
1551
Value *IncV = Phi->getIncomingValue(Idx);
1552
return Phi != getLoopPhiForCounter(IncV, L, DT);
1555
/// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
1556
/// down to checking that all operands are constant and listing instructions
1557
/// that may hide undef.
1558
static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
1560
if (isa<Constant>(V))
1561
return !isa<UndefValue>(V);
1566
// Conservatively handle non-constant non-instructions. For example, Arguments
1568
Instruction *I = dyn_cast<Instruction>(V);
1572
// Load and return values may be undef.
1573
if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
1576
// Optimistically handle other instructions.
1577
for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
1578
if (!Visited.insert(*OI).second)
1580
if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
1586
/// Return true if the given value is concrete. We must prove that undef can
1589
/// TODO: If we decide that this is a good approach to checking for undef, we
1590
/// may factor it into a common location.
1591
static bool hasConcreteDef(Value *V) {
1592
SmallPtrSet<Value*, 8> Visited;
1594
return hasConcreteDefImpl(V, Visited, 0);
1597
/// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
1598
/// be rewritten) loop exit test.
1599
static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
1600
int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1601
Value *IncV = Phi->getIncomingValue(LatchIdx);
1603
for (User *U : Phi->users())
1604
if (U != Cond && U != IncV) return false;
1606
for (User *U : IncV->users())
1607
if (U != Cond && U != Phi) return false;
1611
/// FindLoopCounter - Find an affine IV in canonical form.
1613
/// BECount may be an i8* pointer type. The pointer difference is already
1614
/// valid count without scaling the address stride, so it remains a pointer
1615
/// expression as far as SCEV is concerned.
1617
/// Currently only valid for LFTR. See the comments on hasConcreteDef below.
1619
/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
1621
/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
1622
/// This is difficult in general for SCEV because of potential overflow. But we
1623
/// could at least handle constant BECounts.
1624
static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
1625
ScalarEvolution *SE, DominatorTree *DT) {
1626
uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
1629
cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
1631
// Loop over all of the PHI nodes, looking for a simple counter.
1632
PHINode *BestPhi = nullptr;
1633
const SCEV *BestInit = nullptr;
1634
BasicBlock *LatchBlock = L->getLoopLatch();
1635
assert(LatchBlock && "needsLFTR should guarantee a loop latch");
1637
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
1638
PHINode *Phi = cast<PHINode>(I);
1639
if (!SE->isSCEVable(Phi->getType()))
1642
// Avoid comparing an integer IV against a pointer Limit.
1643
if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
1646
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
1647
if (!AR || AR->getLoop() != L || !AR->isAffine())
1650
// AR may be a pointer type, while BECount is an integer type.
1651
// AR may be wider than BECount. With eq/ne tests overflow is immaterial.
1652
// AR may not be a narrower type, or we may never exit.
1653
uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
1654
if (PhiWidth < BCWidth ||
1655
!L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
1658
const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
1659
if (!Step || !Step->isOne())
1662
int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
1663
Value *IncV = Phi->getIncomingValue(LatchIdx);
1664
if (getLoopPhiForCounter(IncV, L, DT) != Phi)
1667
// Avoid reusing a potentially undef value to compute other values that may
1668
// have originally had a concrete definition.
1669
if (!hasConcreteDef(Phi)) {
1670
// We explicitly allow unknown phis as long as they are already used by
1671
// the loop test. In this case we assume that performing LFTR could not
1672
// increase the number of undef users.
1673
if (ICmpInst *Cond = getLoopTest(L)) {
1674
if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
1675
&& Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
1680
const SCEV *Init = AR->getStart();
1682
if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
1683
// Don't force a live loop counter if another IV can be used.
1684
if (AlmostDeadIV(Phi, LatchBlock, Cond))
1687
// Prefer to count-from-zero. This is a more "canonical" counter form. It
1688
// also prefers integer to pointer IVs.
1689
if (BestInit->isZero() != Init->isZero()) {
1690
if (BestInit->isZero())
1693
// If two IVs both count from zero or both count from nonzero then the
1694
// narrower is likely a dead phi that has been widened. Use the wider phi
1695
// to allow the other to be eliminated.
1696
else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
1705
/// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
1706
/// holds the RHS of the new loop test.
1707
static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
1708
SCEVExpander &Rewriter, ScalarEvolution *SE) {
1709
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1710
assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
1711
const SCEV *IVInit = AR->getStart();
1713
// IVInit may be a pointer while IVCount is an integer when FindLoopCounter
1714
// finds a valid pointer IV. Sign extend BECount in order to materialize a
1715
// GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
1716
// the existing GEPs whenever possible.
1717
if (IndVar->getType()->isPointerTy()
1718
&& !IVCount->getType()->isPointerTy()) {
1720
// IVOffset will be the new GEP offset that is interpreted by GEP as a
1721
// signed value. IVCount on the other hand represents the loop trip count,
1722
// which is an unsigned value. FindLoopCounter only allows induction
1723
// variables that have a positive unit stride of one. This means we don't
1724
// have to handle the case of negative offsets (yet) and just need to zero
1726
Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
1727
const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
1729
// Expand the code for the iteration count.
1730
assert(SE->isLoopInvariant(IVOffset, L) &&
1731
"Computed iteration count is not loop invariant!");
1732
BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1733
Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
1735
Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
1736
assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
1737
// We could handle pointer IVs other than i8*, but we need to compensate for
1738
// gep index scaling. See canExpandBackedgeTakenCount comments.
1739
assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
1740
cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
1741
&& "unit stride pointer IV must be i8*");
1743
IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
1744
return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
1747
// In any other case, convert both IVInit and IVCount to integers before
1748
// comparing. This may result in SCEV expension of pointers, but in practice
1749
// SCEV will fold the pointer arithmetic away as such:
1750
// BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
1752
// Valid Cases: (1) both integers is most common; (2) both may be pointers
1753
// for simple memset-style loops.
1755
// IVInit integer and IVCount pointer would only occur if a canonical IV
1756
// were generated on top of case #2, which is not expected.
1758
const SCEV *IVLimit = nullptr;
1759
// For unit stride, IVCount = Start + BECount with 2's complement overflow.
1760
// For non-zero Start, compute IVCount here.
1761
if (AR->getStart()->isZero())
1764
assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
1765
const SCEV *IVInit = AR->getStart();
1767
// For integer IVs, truncate the IV before computing IVInit + BECount.
1768
if (SE->getTypeSizeInBits(IVInit->getType())
1769
> SE->getTypeSizeInBits(IVCount->getType()))
1770
IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
1772
IVLimit = SE->getAddExpr(IVInit, IVCount);
1774
// Expand the code for the iteration count.
1775
BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1776
IRBuilder<> Builder(BI);
1777
assert(SE->isLoopInvariant(IVLimit, L) &&
1778
"Computed iteration count is not loop invariant!");
1779
// Ensure that we generate the same type as IndVar, or a smaller integer
1780
// type. In the presence of null pointer values, we have an integer type
1781
// SCEV expression (IVInit) for a pointer type IV value (IndVar).
1782
Type *LimitTy = IVCount->getType()->isPointerTy() ?
1783
IndVar->getType() : IVCount->getType();
1784
return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
1788
/// LinearFunctionTestReplace - This method rewrites the exit condition of the
1789
/// loop to be a canonical != comparison against the incremented loop induction
1790
/// variable. This pass is able to rewrite the exit tests of any loop where the
1791
/// SCEV analysis can determine a loop-invariant trip count of the loop, which
1792
/// is actually a much broader range than just linear tests.
1793
Value *IndVarSimplify::
1794
LinearFunctionTestReplace(Loop *L,
1795
const SCEV *BackedgeTakenCount,
1797
SCEVExpander &Rewriter) {
1798
assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition");
1800
// Initialize CmpIndVar and IVCount to their preincremented values.
1801
Value *CmpIndVar = IndVar;
1802
const SCEV *IVCount = BackedgeTakenCount;
1804
// If the exiting block is the same as the backedge block, we prefer to
1805
// compare against the post-incremented value, otherwise we must compare
1806
// against the preincremented value.
1807
if (L->getExitingBlock() == L->getLoopLatch()) {
1808
// Add one to the "backedge-taken" count to get the trip count.
1809
// This addition may overflow, which is valid as long as the comparison is
1810
// truncated to BackedgeTakenCount->getType().
1811
IVCount = SE->getAddExpr(BackedgeTakenCount,
1812
SE->getConstant(BackedgeTakenCount->getType(), 1));
1813
// The BackedgeTaken expression contains the number of times that the
1814
// backedge branches to the loop header. This is one less than the
1815
// number of times the loop executes, so use the incremented indvar.
1816
CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
1819
Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
1820
assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
1821
&& "genLoopLimit missed a cast");
1823
// Insert a new icmp_ne or icmp_eq instruction before the branch.
1824
BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
1825
ICmpInst::Predicate P;
1826
if (L->contains(BI->getSuccessor(0)))
1827
P = ICmpInst::ICMP_NE;
1829
P = ICmpInst::ICMP_EQ;
1831
DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
1832
<< " LHS:" << *CmpIndVar << '\n'
1834
<< (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
1835
<< " RHS:\t" << *ExitCnt << "\n"
1836
<< " IVCount:\t" << *IVCount << "\n");
1838
IRBuilder<> Builder(BI);
1840
// LFTR can ignore IV overflow and truncate to the width of
1841
// BECount. This avoids materializing the add(zext(add)) expression.
1842
unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
1843
unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
1844
if (CmpIndVarSize > ExitCntSize) {
1845
const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
1846
const SCEV *ARStart = AR->getStart();
1847
const SCEV *ARStep = AR->getStepRecurrence(*SE);
1848
// For constant IVCount, avoid truncation.
1849
if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
1850
const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
1851
APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
1852
// Note that the post-inc value of BackedgeTakenCount may have overflowed
1853
// above such that IVCount is now zero.
1854
if (IVCount != BackedgeTakenCount && Count == 0) {
1855
Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
1859
Count = Count.zext(CmpIndVarSize);
1861
if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
1862
NewLimit = Start - Count;
1864
NewLimit = Start + Count;
1865
ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
1867
DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
1869
CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
1873
Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
1874
Value *OrigCond = BI->getCondition();
1875
// It's tempting to use replaceAllUsesWith here to fully replace the old
1876
// comparison, but that's not immediately safe, since users of the old
1877
// comparison may not be dominated by the new comparison. Instead, just
1878
// update the branch to use the new comparison; in the common case this
1879
// will make old comparison dead.
1880
BI->setCondition(Cond);
1881
DeadInsts.push_back(OrigCond);
1888
//===----------------------------------------------------------------------===//
1889
// SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
1890
//===----------------------------------------------------------------------===//
1892
/// If there's a single exit block, sink any loop-invariant values that
1893
/// were defined in the preheader but not used inside the loop into the
1894
/// exit block to reduce register pressure in the loop.
1895
void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
1896
BasicBlock *ExitBlock = L->getExitBlock();
1897
if (!ExitBlock) return;
1899
BasicBlock *Preheader = L->getLoopPreheader();
1900
if (!Preheader) return;
1902
Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
1903
BasicBlock::iterator I = Preheader->getTerminator();
1904
while (I != Preheader->begin()) {
1906
// New instructions were inserted at the end of the preheader.
1907
if (isa<PHINode>(I))
1910
// Don't move instructions which might have side effects, since the side
1911
// effects need to complete before instructions inside the loop. Also don't
1912
// move instructions which might read memory, since the loop may modify
1913
// memory. Note that it's okay if the instruction might have undefined
1914
// behavior: LoopSimplify guarantees that the preheader dominates the exit
1916
if (I->mayHaveSideEffects() || I->mayReadFromMemory())
1919
// Skip debug info intrinsics.
1920
if (isa<DbgInfoIntrinsic>(I))
1923
// Skip landingpad instructions.
1924
if (isa<LandingPadInst>(I))
1927
// Don't sink alloca: we never want to sink static alloca's out of the
1928
// entry block, and correctly sinking dynamic alloca's requires
1929
// checks for stacksave/stackrestore intrinsics.
1930
// FIXME: Refactor this check somehow?
1931
if (isa<AllocaInst>(I))
1934
// Determine if there is a use in or before the loop (direct or
1936
bool UsedInLoop = false;
1937
for (Use &U : I->uses()) {
1938
Instruction *User = cast<Instruction>(U.getUser());
1939
BasicBlock *UseBB = User->getParent();
1940
if (PHINode *P = dyn_cast<PHINode>(User)) {
1942
PHINode::getIncomingValueNumForOperand(U.getOperandNo());
1943
UseBB = P->getIncomingBlock(i);
1945
if (UseBB == Preheader || L->contains(UseBB)) {
1951
// If there is, the def must remain in the preheader.
1955
// Otherwise, sink it to the exit block.
1956
Instruction *ToMove = I;
1959
if (I != Preheader->begin()) {
1960
// Skip debug info intrinsics.
1963
} while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
1965
if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
1971
ToMove->moveBefore(InsertPt);
1977
//===----------------------------------------------------------------------===//
1978
// IndVarSimplify driver. Manage several subpasses of IV simplification.
1979
//===----------------------------------------------------------------------===//
1981
bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
1982
if (skipOptnoneFunction(L))
1985
// If LoopSimplify form is not available, stay out of trouble. Some notes:
1986
// - LSR currently only supports LoopSimplify-form loops. Indvars'
1987
// canonicalization can be a pessimization without LSR to "clean up"
1989
// - We depend on having a preheader; in particular,
1990
// Loop::getCanonicalInductionVariable only supports loops with preheaders,
1991
// and we're in trouble if we can't find the induction variable even when
1992
// we've manually inserted one.
1993
if (!L->isLoopSimplifyForm())
1996
LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1997
SE = &getAnalysis<ScalarEvolution>();
1998
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1999
auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
2000
TLI = TLIP ? &TLIP->getTLI() : nullptr;
2001
auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
2002
TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
2003
const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
2008
// If there are any floating-point recurrences, attempt to
2009
// transform them to use integer recurrences.
2010
RewriteNonIntegerIVs(L);
2012
const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
2014
// Create a rewriter object which we'll use to transform the code with.
2015
SCEVExpander Rewriter(*SE, DL, "indvars");
2017
Rewriter.setDebugType(DEBUG_TYPE);
2020
// Eliminate redundant IV users.
2022
// Simplification works best when run before other consumers of SCEV. We
2023
// attempt to avoid evaluating SCEVs for sign/zero extend operations until
2024
// other expressions involving loop IVs have been evaluated. This helps SCEV
2025
// set no-wrap flags before normalizing sign/zero extension.
2026
Rewriter.disableCanonicalMode();
2027
SimplifyAndExtend(L, Rewriter, LPM);
2029
// Check to see if this loop has a computable loop-invariant execution count.
2030
// If so, this means that we can compute the final value of any expressions
2031
// that are recurrent in the loop, and substitute the exit values from the
2032
// loop into any instructions outside of the loop that use the final values of
2033
// the current expressions.
2035
if (ReplaceExitValue != NeverRepl &&
2036
!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
2037
RewriteLoopExitValues(L, Rewriter);
2039
// Eliminate redundant IV cycles.
2040
NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
2042
// If we have a trip count expression, rewrite the loop's exit condition
2043
// using it. We can currently only handle loops with a single exit.
2044
if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) {
2045
PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
2047
// Check preconditions for proper SCEVExpander operation. SCEV does not
2048
// express SCEVExpander's dependencies, such as LoopSimplify. Instead any
2049
// pass that uses the SCEVExpander must do it. This does not work well for
2050
// loop passes because SCEVExpander makes assumptions about all loops,
2051
// while LoopPassManager only forces the current loop to be simplified.
2053
// FIXME: SCEV expansion has no way to bail out, so the caller must
2054
// explicitly check any assumptions made by SCEV. Brittle.
2055
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
2056
if (!AR || AR->getLoop()->getLoopPreheader())
2057
(void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
2061
// Clear the rewriter cache, because values that are in the rewriter's cache
2062
// can be deleted in the loop below, causing the AssertingVH in the cache to
2066
// Now that we're done iterating through lists, clean up any instructions
2067
// which are now dead.
2068
while (!DeadInsts.empty())
2069
if (Instruction *Inst =
2070
dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
2071
RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
2073
// The Rewriter may not be used from this point on.
2075
// Loop-invariant instructions in the preheader that aren't used in the
2076
// loop may be sunk below the loop to reduce register pressure.
2077
SinkUnusedInvariants(L);
2079
// Clean up dead instructions.
2080
Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
2081
// Check a post-condition.
2082
assert(L->isLCSSAForm(*DT) &&
2083
"Indvars did not leave the loop in lcssa form!");
2085
// Verify that LFTR, and any other change have not interfered with SCEV's
2086
// ability to compute trip count.
2088
if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
2090
const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
2091
if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
2092
SE->getTypeSizeInBits(NewBECount->getType()))
2093
NewBECount = SE->getTruncateOrNoop(NewBECount,
2094
BackedgeTakenCount->getType());
2096
BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
2097
NewBECount->getType());
2098
assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");