2001-12-03 17:28:42 +00:00
|
|
|
//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
|
2005-04-21 23:48:37 +00:00
|
|
|
//
|
2003-10-20 19:43:21 +00:00
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
2007-12-29 20:36:04 +00:00
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
2005-04-21 23:48:37 +00:00
|
|
|
//
|
2003-10-20 19:43:21 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
2001-12-03 17:28:42 +00:00
|
|
|
//
|
2004-04-02 20:24:31 +00:00
|
|
|
// This transformation analyzes and transforms the induction variables (and
|
|
|
|
// computations derived from them) into simpler forms suitable for subsequent
|
|
|
|
// analysis and transformation.
|
|
|
|
//
|
2006-08-18 09:01:07 +00:00
|
|
|
// This transformation makes the following changes to each loop with an
|
2004-04-02 20:24:31 +00:00
|
|
|
// identifiable induction variable:
|
|
|
|
// 1. All loops are transformed to have a SINGLE canonical induction variable
|
|
|
|
// which starts at zero and steps by one.
|
|
|
|
// 2. The canonical induction variable is guaranteed to be the first PHI node
|
|
|
|
// in the loop header block.
|
2009-06-14 22:38:41 +00:00
|
|
|
// 3. The canonical induction variable is guaranteed to be in a wide enough
|
|
|
|
// type so that IV expressions need not be (directly) zero-extended or
|
|
|
|
// sign-extended.
|
|
|
|
// 4. Any pointer arithmetic recurrences are raised to use array subscripts.
|
2004-04-02 20:24:31 +00:00
|
|
|
//
|
|
|
|
// If the trip count of a loop is computable, this pass also makes the following
|
|
|
|
// changes:
|
|
|
|
// 1. The exit condition for the loop is canonicalized to compare the
|
|
|
|
// induction value against the exit value. This turns loops like:
|
|
|
|
// 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
|
|
|
|
// 2. Any use outside of the loop of an expression derived from the indvar
|
|
|
|
// is changed to compute the derived value outside of the loop, eliminating
|
|
|
|
// the dependence on the exit value of the induction variable. If the only
|
|
|
|
// purpose of the loop is to compute the exit value of some derived
|
|
|
|
// expression, this transformation will make the loop dead.
|
|
|
|
//
|
|
|
|
// This transformation should be followed by strength reduction after all of the
|
2009-05-19 20:38:47 +00:00
|
|
|
// desired loop transformations have been performed.
|
2001-12-03 17:28:42 +00:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2006-12-19 21:40:18 +00:00
|
|
|
#define DEBUG_TYPE "indvars"
|
2002-05-07 20:03:00 +00:00
|
|
|
#include "llvm/Transforms/Scalar.h"
|
2004-04-02 20:24:31 +00:00
|
|
|
#include "llvm/BasicBlock.h"
|
Change the canonical induction variable that we insert.
Instead of producing code like this:
Loop:
X = phi 0, X2
...
X2 = X + 1
if (X != N-1) goto Loop
We now generate code that looks like this:
Loop:
X = phi 0, X2
...
X2 = X + 1
if (X2 != N) goto Loop
This has two big advantages:
1. The trip count of the loop is now explicit in the code, allowing
the direct implementation of Loop::getTripCount()
2. This reduces register pressure in the loop, and allows X and X2 to be
put into the same register.
As a consequence of the second point, the code we generate for loops went
from:
.LBB2: # no_exit.1
...
mov %EDI, %ESI
inc %EDI
cmp %ESI, 2
mov %ESI, %EDI
jne .LBB2 # PC rel: no_exit.1
To:
.LBB2: # no_exit.1
...
inc %ESI
cmp %ESI, 3
jne .LBB2 # PC rel: no_exit.1
... which has two fewer moves, and uses one less register.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12961 91177308-0d34-0410-b5e6-96231b3b80d8
2004-04-15 15:21:43 +00:00
|
|
|
#include "llvm/Constants.h"
|
2003-12-22 05:02:01 +00:00
|
|
|
#include "llvm/Instructions.h"
|
2010-03-15 22:23:03 +00:00
|
|
|
#include "llvm/IntrinsicInst.h"
|
2009-07-03 00:17:18 +00:00
|
|
|
#include "llvm/LLVMContext.h"
|
2004-04-02 20:24:31 +00:00
|
|
|
#include "llvm/Type.h"
|
2009-05-12 02:17:14 +00:00
|
|
|
#include "llvm/Analysis/Dominators.h"
|
|
|
|
#include "llvm/Analysis/IVUsers.h"
|
2005-07-30 00:12:19 +00:00
|
|
|
#include "llvm/Analysis/ScalarEvolutionExpander.h"
|
2003-12-18 17:19:19 +00:00
|
|
|
#include "llvm/Analysis/LoopInfo.h"
|
2007-03-07 06:39:01 +00:00
|
|
|
#include "llvm/Analysis/LoopPass.h"
|
2002-02-12 22:39:50 +00:00
|
|
|
#include "llvm/Support/CFG.h"
|
2011-06-28 03:01:46 +00:00
|
|
|
#include "llvm/Support/CommandLine.h"
|
2007-01-07 01:14:12 +00:00
|
|
|
#include "llvm/Support/Debug.h"
|
2009-08-23 04:37:46 +00:00
|
|
|
#include "llvm/Support/raw_ostream.h"
|
2003-12-18 17:19:19 +00:00
|
|
|
#include "llvm/Transforms/Utils/Local.h"
|
2009-05-12 02:17:14 +00:00
|
|
|
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
|
2011-05-04 02:10:13 +00:00
|
|
|
#include "llvm/Target/TargetData.h"
|
2011-07-06 20:50:43 +00:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
For PR1064:
Implement the arbitrary bit-width integer feature. The feature allows
integers of any bitwidth (up to 64) to be defined instead of just 1, 8,
16, 32, and 64 bit integers.
This change does several things:
1. Introduces a new Derived Type, IntegerType, to represent the number of
bits in an integer. The Type classes SubclassData field is used to
store the number of bits. This allows 2^23 bits in an integer type.
2. Removes the five integer Type::TypeID values for the 1, 8, 16, 32 and
64-bit integers. These are replaced with just IntegerType which is not
a primitive any more.
3. Adjust the rest of LLVM to account for this change.
Note that while this incremental change lays the foundation for arbitrary
bit-width integers, LLVM has not yet been converted to actually deal with
them in any significant way. Most optimization passes, for example, will
still only deal with the byte-width integer types. Future increments
will rectify this situation.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@33113 91177308-0d34-0410-b5e6-96231b3b80d8
2007-01-12 07:05:14 +00:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
2004-09-01 22:55:40 +00:00
|
|
|
#include "llvm/ADT/Statistic.h"
|
2009-05-12 02:17:14 +00:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2003-12-18 17:19:19 +00:00
|
|
|
using namespace llvm;
|
2003-11-11 22:41:34 +00:00
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
STATISTIC(NumRemoved , "Number of aux indvars removed");
|
|
|
|
STATISTIC(NumWidened , "Number of indvars widened");
|
|
|
|
STATISTIC(NumInserted , "Number of canonical indvars added");
|
|
|
|
STATISTIC(NumReplaced , "Number of exit values replaced");
|
|
|
|
STATISTIC(NumLFTR , "Number of loop exit tests replaced");
|
|
|
|
STATISTIC(NumElimIdentity, "Number of IV identities eliminated");
|
|
|
|
STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
|
|
|
|
STATISTIC(NumElimRem , "Number of IV remainder operations eliminated");
|
|
|
|
STATISTIC(NumElimCmp , "Number of IV comparisons eliminated");
|
2011-07-06 20:50:43 +00:00
|
|
|
STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2011-06-28 03:01:46 +00:00
|
|
|
static cl::opt<bool> DisableIVRewrite(
|
|
|
|
"disable-iv-rewrite", cl::Hidden,
|
|
|
|
cl::desc("Disable canonical induction variable rewriting"));
|
2011-05-04 02:10:13 +00:00
|
|
|
|
2006-12-19 21:40:18 +00:00
|
|
|
namespace {
|
2009-09-02 06:11:42 +00:00
|
|
|
class IndVarSimplify : public LoopPass {
|
2009-05-12 02:17:14 +00:00
|
|
|
IVUsers *IU;
|
2004-04-02 20:24:31 +00:00
|
|
|
LoopInfo *LI;
|
|
|
|
ScalarEvolution *SE;
|
2009-06-27 05:16:57 +00:00
|
|
|
DominatorTree *DT;
|
2011-05-04 02:10:13 +00:00
|
|
|
TargetData *TD;
|
2011-06-21 03:22:38 +00:00
|
|
|
|
2011-03-17 23:51:11 +00:00
|
|
|
SmallVector<WeakVH, 16> DeadInsts;
|
2003-12-23 07:47:09 +00:00
|
|
|
bool Changed;
|
2003-12-22 03:58:44 +00:00
|
|
|
public:
|
2007-05-01 21:15:47 +00:00
|
|
|
|
2009-07-15 01:26:32 +00:00
|
|
|
static char ID; // Pass identification, replacement for typeid
|
2011-06-21 03:22:38 +00:00
|
|
|
IndVarSimplify() : LoopPass(ID), IU(0), LI(0), SE(0), DT(0), TD(0),
|
2011-06-28 02:49:20 +00:00
|
|
|
Changed(false) {
|
2010-10-19 17:21:58 +00:00
|
|
|
initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
|
|
|
|
}
|
2007-05-01 21:15:47 +00:00
|
|
|
|
2009-07-15 01:26:32 +00:00
|
|
|
virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
|
2009-02-17 20:49:49 +00:00
|
|
|
|
2009-07-15 01:26:32 +00:00
|
|
|
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.addRequired<DominatorTree>();
|
|
|
|
AU.addRequired<LoopInfo>();
|
|
|
|
AU.addRequired<ScalarEvolution>();
|
|
|
|
AU.addRequiredID(LoopSimplifyID);
|
|
|
|
AU.addRequiredID(LCSSAID);
|
2011-06-28 03:01:46 +00:00
|
|
|
if (!DisableIVRewrite)
|
|
|
|
AU.addRequired<IVUsers>();
|
2009-07-15 01:26:32 +00:00
|
|
|
AU.addPreserved<ScalarEvolution>();
|
|
|
|
AU.addPreservedID(LoopSimplifyID);
|
|
|
|
AU.addPreservedID(LCSSAID);
|
2011-06-21 03:22:38 +00:00
|
|
|
if (!DisableIVRewrite)
|
|
|
|
AU.addPreserved<IVUsers>();
|
2009-07-15 01:26:32 +00:00
|
|
|
AU.setPreservesCFG();
|
|
|
|
}
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
private:
|
2011-07-06 20:50:43 +00:00
|
|
|
virtual void releaseMemory() {
|
|
|
|
DeadInsts.clear();
|
|
|
|
}
|
|
|
|
|
2011-03-17 23:51:11 +00:00
|
|
|
bool isValidRewrite(Value *FromVal, Value *ToVal);
|
2007-03-07 06:39:01 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
void HandleFloatingPointIV(Loop *L, PHINode *PH);
|
|
|
|
void RewriteNonIntegerIVs(Loop *L);
|
|
|
|
|
|
|
|
void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
|
|
|
|
|
2011-05-20 18:25:42 +00:00
|
|
|
void SimplifyIVUsers(SCEVExpander &Rewriter);
|
2011-06-21 03:22:38 +00:00
|
|
|
void SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter);
|
|
|
|
|
|
|
|
bool EliminateIVUser(Instruction *UseInst, Instruction *IVOperand);
|
2011-05-12 00:04:28 +00:00
|
|
|
void EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand);
|
|
|
|
void EliminateIVRemainder(BinaryOperator *Rem,
|
|
|
|
Value *IVOperand,
|
2011-06-21 15:43:52 +00:00
|
|
|
bool IsSigned);
|
2004-04-02 20:24:31 +00:00
|
|
|
|
2011-07-06 20:50:43 +00:00
|
|
|
void SimplifyCongruentIVs(Loop *L);
|
|
|
|
|
2010-02-22 04:11:59 +00:00
|
|
|
void RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter);
|
2008-09-09 21:41:07 +00:00
|
|
|
|
2011-07-16 01:18:53 +00:00
|
|
|
ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
|
2011-07-12 00:08:50 +00:00
|
|
|
PHINode *IndVar,
|
|
|
|
SCEVExpander &Rewriter);
|
2009-05-12 02:17:14 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
void SinkUnusedInvariants(Loop *L);
|
2003-12-22 03:58:44 +00:00
|
|
|
};
|
2002-09-10 05:24:05 +00:00
|
|
|
}
|
2001-12-04 04:32:29 +00:00
|
|
|
|
2008-05-13 00:00:25 +00:00
|
|
|
char IndVarSimplify::ID = 0;
|
2010-10-12 19:48:12 +00:00
|
|
|
INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
|
2011-05-04 02:10:13 +00:00
|
|
|
"Induction Variable Simplification", false, false)
|
2010-10-12 19:48:12 +00:00
|
|
|
INITIALIZE_PASS_DEPENDENCY(DominatorTree)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LoopInfo)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(LCSSA)
|
|
|
|
INITIALIZE_PASS_DEPENDENCY(IVUsers)
|
|
|
|
INITIALIZE_PASS_END(IndVarSimplify, "indvars",
|
2011-05-04 02:10:13 +00:00
|
|
|
"Induction Variable Simplification", false, false)
|
2008-05-13 00:00:25 +00:00
|
|
|
|
2008-10-22 23:32:42 +00:00
|
|
|
Pass *llvm::createIndVarSimplifyPass() {
|
2003-12-22 03:58:44 +00:00
|
|
|
return new IndVarSimplify();
|
2001-12-04 04:32:29 +00:00
|
|
|
}
|
|
|
|
|
2011-03-17 23:51:11 +00:00
|
|
|
/// isValidRewrite - Return true if the SCEV expansion generated by the
|
|
|
|
/// rewriter can replace the original value. SCEV guarantees that it
|
|
|
|
/// produces the same value, but the way it is produced may be illegal IR.
|
|
|
|
/// Ideally, this function will only be called for verification.
|
|
|
|
bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
|
|
|
|
// If an SCEV expression subsumed multiple pointers, its expansion could
|
|
|
|
// reassociate the GEP changing the base pointer. This is illegal because the
|
|
|
|
// final address produced by a GEP chain must be inbounds relative to its
|
|
|
|
// underlying object. Otherwise basic alias analysis, among other things,
|
|
|
|
// could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
|
|
|
|
// producing an expression involving multiple pointers. Until then, we must
|
|
|
|
// bail out here.
|
|
|
|
//
|
|
|
|
// Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
|
|
|
|
// because it understands lcssa phis while SCEV does not.
|
|
|
|
Value *FromPtr = FromVal;
|
|
|
|
Value *ToPtr = ToVal;
|
|
|
|
if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
|
|
|
|
FromPtr = GEP->getPointerOperand();
|
|
|
|
}
|
|
|
|
if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
|
|
|
|
ToPtr = GEP->getPointerOperand();
|
|
|
|
}
|
|
|
|
if (FromPtr != FromVal || ToPtr != ToVal) {
|
|
|
|
// Quickly check the common case
|
|
|
|
if (FromPtr == ToPtr)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
// SCEV may have rewritten an expression that produces the GEP's pointer
|
|
|
|
// operand. That's ok as long as the pointer operand has the same base
|
|
|
|
// pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
|
|
|
|
// base of a recurrence. This handles the case in which SCEV expansion
|
|
|
|
// converts a pointer type recurrence into a nonrecurrent pointer base
|
|
|
|
// indexed by an integer recurrence.
|
|
|
|
const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
|
|
|
|
const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
|
|
|
|
if (FromBase == ToBase)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
|
|
|
|
<< *FromBase << " != " << *ToBase << "\n");
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// RewriteNonIntegerIVs and helpers. Prefer integer IVs.
|
|
|
|
//===----------------------------------------------------------------------===//
|
2011-05-03 22:24:10 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
/// ConvertToSInt - Convert APF to an integer, if possible.
|
|
|
|
static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
|
|
|
|
bool isExact = false;
|
|
|
|
if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
|
2011-05-03 22:24:10 +00:00
|
|
|
return false;
|
2011-07-12 00:08:50 +00:00
|
|
|
// See if we can convert this to an int64_t
|
|
|
|
uint64_t UIntVal;
|
|
|
|
if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
|
|
|
|
&isExact) != APFloat::opOK || !isExact)
|
2011-05-03 22:24:10 +00:00
|
|
|
return false;
|
2011-07-12 00:08:50 +00:00
|
|
|
IntVal = UIntVal;
|
2011-05-03 22:24:10 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
/// HandleFloatingPointIV - If the loop has floating induction variable
|
|
|
|
/// then insert corresponding integer induction variable if possible.
|
|
|
|
/// For example,
|
|
|
|
/// for(double i = 0; i < 10000; ++i)
|
|
|
|
/// bar(i)
|
|
|
|
/// is converted into
|
|
|
|
/// for(int i = 0; i < 10000; ++i)
|
|
|
|
/// bar((double)i);
|
2011-05-25 04:42:22 +00:00
|
|
|
///
|
2011-07-12 00:08:50 +00:00
|
|
|
void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
|
|
|
|
unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
|
|
|
|
unsigned BackEdge = IncomingEdge^1;
|
2011-05-25 04:42:22 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Check incoming value.
|
|
|
|
ConstantFP *InitValueVal =
|
|
|
|
dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
|
2011-05-25 04:42:22 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
int64_t InitValue;
|
|
|
|
if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
|
|
|
|
return;
|
2011-05-25 04:42:22 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Check IV increment. Reject this PN if increment operation is not
|
|
|
|
// an add or increment value can not be represented by an integer.
|
|
|
|
BinaryOperator *Incr =
|
|
|
|
dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
|
|
|
|
if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
|
2011-05-25 04:42:22 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// If this is not an add of the PHI with a constantfp, or if the constant fp
|
|
|
|
// is not an integer, bail out.
|
|
|
|
ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
|
|
|
|
int64_t IncValue;
|
|
|
|
if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
|
|
|
|
!ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
|
|
|
|
return;
|
2011-05-25 04:42:22 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Check Incr uses. One user is PN and the other user is an exit condition
|
|
|
|
// used by the conditional terminator.
|
|
|
|
Value::use_iterator IncrUse = Incr->use_begin();
|
|
|
|
Instruction *U1 = cast<Instruction>(*IncrUse++);
|
|
|
|
if (IncrUse == Incr->use_end()) return;
|
|
|
|
Instruction *U2 = cast<Instruction>(*IncrUse++);
|
|
|
|
if (IncrUse != Incr->use_end()) return;
|
2010-04-12 21:13:43 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
|
|
|
|
// only used by a branch, we can't transform it.
|
|
|
|
FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
|
|
|
|
if (!Compare)
|
|
|
|
Compare = dyn_cast<FCmpInst>(U2);
|
|
|
|
if (Compare == 0 || !Compare->hasOneUse() ||
|
|
|
|
!isa<BranchInst>(Compare->use_back()))
|
|
|
|
return;
|
2009-02-12 22:19:27 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
|
Change the canonical induction variable that we insert.
Instead of producing code like this:
Loop:
X = phi 0, X2
...
X2 = X + 1
if (X != N-1) goto Loop
We now generate code that looks like this:
Loop:
X = phi 0, X2
...
X2 = X + 1
if (X2 != N) goto Loop
This has two big advantages:
1. The trip count of the loop is now explicit in the code, allowing
the direct implementation of Loop::getTripCount()
2. This reduces register pressure in the loop, and allows X and X2 to be
put into the same register.
As a consequence of the second point, the code we generate for loops went
from:
.LBB2: # no_exit.1
...
mov %EDI, %ESI
inc %EDI
cmp %ESI, 2
mov %ESI, %EDI
jne .LBB2 # PC rel: no_exit.1
To:
.LBB2: # no_exit.1
...
inc %ESI
cmp %ESI, 3
jne .LBB2 # PC rel: no_exit.1
... which has two fewer moves, and uses one less register.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@12961 91177308-0d34-0410-b5e6-96231b3b80d8
2004-04-15 15:21:43 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// We need to verify that the branch actually controls the iteration count
|
|
|
|
// of the loop. If not, the new IV can overflow and no one will notice.
|
|
|
|
// The branch block must be in the loop and one of the successors must be out
|
|
|
|
// of the loop.
|
|
|
|
assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
|
|
|
|
if (!L->contains(TheBr->getParent()) ||
|
|
|
|
(L->contains(TheBr->getSuccessor(0)) &&
|
|
|
|
L->contains(TheBr->getSuccessor(1))))
|
|
|
|
return;
|
2004-04-02 20:24:31 +00:00
|
|
|
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// If it isn't a comparison with an integer-as-fp (the exit value), we can't
|
|
|
|
// transform it.
|
|
|
|
ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
|
|
|
|
int64_t ExitValue;
|
|
|
|
if (ExitValueVal == 0 ||
|
|
|
|
!ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
|
|
|
|
return;
|
2009-02-12 22:19:27 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Find new predicate for integer comparison.
|
|
|
|
CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
|
|
|
|
switch (Compare->getPredicate()) {
|
|
|
|
default: return; // Unknown comparison.
|
|
|
|
case CmpInst::FCMP_OEQ:
|
|
|
|
case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
|
|
|
|
case CmpInst::FCMP_ONE:
|
|
|
|
case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
|
|
|
|
case CmpInst::FCMP_OGT:
|
|
|
|
case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
|
|
|
|
case CmpInst::FCMP_OGE:
|
|
|
|
case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
|
|
|
|
case CmpInst::FCMP_OLT:
|
|
|
|
case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
|
|
|
|
case CmpInst::FCMP_OLE:
|
|
|
|
case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
|
|
|
|
}
|
2009-05-12 02:17:14 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// We convert the floating point induction variable to a signed i32 value if
|
|
|
|
// we can. This is only safe if the comparison will not overflow in a way
|
|
|
|
// that won't be trapped by the integer equivalent operations. Check for this
|
|
|
|
// now.
|
|
|
|
// TODO: We could use i64 if it is native and the range requires it.
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// The start/stride/exit values must all fit in signed i32.
|
|
|
|
if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
|
|
|
|
return;
|
2009-05-12 02:17:14 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// If not actually striding (add x, 0.0), avoid touching the code.
|
|
|
|
if (IncValue == 0)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Positive and negative strides have different safety conditions.
|
|
|
|
if (IncValue > 0) {
|
|
|
|
// If we have a positive stride, we require the init to be less than the
|
|
|
|
// exit value and an equality or less than comparison.
|
|
|
|
if (InitValue >= ExitValue ||
|
|
|
|
NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
|
|
|
|
return;
|
|
|
|
|
|
|
|
uint32_t Range = uint32_t(ExitValue-InitValue);
|
|
|
|
if (NewPred == CmpInst::ICMP_SLE) {
|
|
|
|
// Normalize SLE -> SLT, check for infinite loop.
|
|
|
|
if (++Range == 0) return; // Range overflows.
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned Leftover = Range % uint32_t(IncValue);
|
|
|
|
|
|
|
|
// If this is an equality comparison, we require that the strided value
|
|
|
|
// exactly land on the exit value, otherwise the IV condition will wrap
|
|
|
|
// around and do things the fp IV wouldn't.
|
|
|
|
if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
|
|
|
|
Leftover != 0)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// If the stride would wrap around the i32 before exiting, we can't
|
|
|
|
// transform the IV.
|
|
|
|
if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
|
|
|
|
return;
|
|
|
|
|
|
|
|
} else {
|
|
|
|
// If we have a negative stride, we require the init to be greater than the
|
|
|
|
// exit value and an equality or greater than comparison.
|
|
|
|
if (InitValue >= ExitValue ||
|
|
|
|
NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
|
|
|
|
return;
|
|
|
|
|
|
|
|
uint32_t Range = uint32_t(InitValue-ExitValue);
|
|
|
|
if (NewPred == CmpInst::ICMP_SGE) {
|
|
|
|
// Normalize SGE -> SGT, check for infinite loop.
|
|
|
|
if (++Range == 0) return; // Range overflows.
|
|
|
|
}
|
|
|
|
|
|
|
|
unsigned Leftover = Range % uint32_t(-IncValue);
|
|
|
|
|
|
|
|
// If this is an equality comparison, we require that the strided value
|
|
|
|
// exactly land on the exit value, otherwise the IV condition will wrap
|
|
|
|
// around and do things the fp IV wouldn't.
|
|
|
|
if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
|
|
|
|
Leftover != 0)
|
|
|
|
return;
|
|
|
|
|
|
|
|
// If the stride would wrap around the i32 before exiting, we can't
|
|
|
|
// transform the IV.
|
|
|
|
if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2011-07-18 04:54:35 +00:00
|
|
|
IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
|
2011-07-12 00:08:50 +00:00
|
|
|
|
|
|
|
// Insert new integer induction variable.
|
|
|
|
PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
|
|
|
|
NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
|
|
|
|
PN->getIncomingBlock(IncomingEdge));
|
|
|
|
|
|
|
|
Value *NewAdd =
|
|
|
|
BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
|
|
|
|
Incr->getName()+".int", Incr);
|
|
|
|
NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
|
|
|
|
|
|
|
|
ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
|
|
|
|
ConstantInt::get(Int32Ty, ExitValue),
|
|
|
|
Compare->getName());
|
|
|
|
|
|
|
|
// In the following deletions, PN may become dead and may be deleted.
|
|
|
|
// Use a WeakVH to observe whether this happens.
|
|
|
|
WeakVH WeakPH = PN;
|
|
|
|
|
|
|
|
// Delete the old floating point exit comparison. The branch starts using the
|
|
|
|
// new comparison.
|
|
|
|
NewCompare->takeName(Compare);
|
|
|
|
Compare->replaceAllUsesWith(NewCompare);
|
|
|
|
RecursivelyDeleteTriviallyDeadInstructions(Compare);
|
|
|
|
|
|
|
|
// Delete the old floating point increment.
|
|
|
|
Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
|
|
|
|
RecursivelyDeleteTriviallyDeadInstructions(Incr);
|
|
|
|
|
|
|
|
// If the FP induction variable still has uses, this is because something else
|
|
|
|
// in the loop uses its value. In order to canonicalize the induction
|
|
|
|
// variable, we chose to eliminate the IV and rewrite it in terms of an
|
|
|
|
// int->fp cast.
|
|
|
|
//
|
|
|
|
// We give preference to sitofp over uitofp because it is faster on most
|
|
|
|
// platforms.
|
|
|
|
if (WeakPH) {
|
|
|
|
Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
|
|
|
|
PN->getParent()->getFirstNonPHI());
|
|
|
|
PN->replaceAllUsesWith(Conv);
|
|
|
|
RecursivelyDeleteTriviallyDeadInstructions(PN);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Add a new IVUsers entry for the newly-created integer PHI.
|
|
|
|
if (IU)
|
|
|
|
IU->AddUsersIfInteresting(NewPHI);
|
|
|
|
}
|
|
|
|
|
|
|
|
void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
|
|
|
|
// First step. Check to see if there are any floating-point recurrences.
|
|
|
|
// If there are, change them into integer recurrences, permitting analysis by
|
|
|
|
// the SCEV routines.
|
|
|
|
//
|
|
|
|
BasicBlock *Header = L->getHeader();
|
|
|
|
|
|
|
|
SmallVector<WeakVH, 8> PHIs;
|
|
|
|
for (BasicBlock::iterator I = Header->begin();
|
|
|
|
PHINode *PN = dyn_cast<PHINode>(I); ++I)
|
|
|
|
PHIs.push_back(PN);
|
|
|
|
|
|
|
|
for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
|
|
|
|
if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
|
|
|
|
HandleFloatingPointIV(L, PN);
|
|
|
|
|
|
|
|
// If the loop previously had floating-point IV, ScalarEvolution
|
|
|
|
// may not have been able to compute a trip count. Now that we've done some
|
|
|
|
// re-writing, the trip count may be computable.
|
|
|
|
if (Changed)
|
|
|
|
SE->forgetLoop(L);
|
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// RewriteLoopExitValues - Optimize IV users outside the loop.
|
|
|
|
// As a side effect, reduces the amount of IV processing within the loop.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
/// RewriteLoopExitValues - Check to see if this loop has a computable
|
|
|
|
/// loop-invariant execution count. If so, this means that we can compute the
|
|
|
|
/// final value of any expressions that are recurrent in the loop, and
|
|
|
|
/// substitute the exit values from the loop into any instructions outside of
|
|
|
|
/// the loop that use the final values of the current expressions.
|
|
|
|
///
|
|
|
|
/// This is mostly redundant with the regular IndVarSimplify activities that
|
|
|
|
/// happen later, except that it's more powerful in some cases, because it's
|
|
|
|
/// able to brute-force evaluate arbitrary instructions as long as they have
|
|
|
|
/// constant operands at the beginning of the loop.
|
|
|
|
void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
|
|
|
|
// Verify the input to the pass in already in LCSSA form.
|
|
|
|
assert(L->isLCSSAForm(*DT));
|
|
|
|
|
|
|
|
SmallVector<BasicBlock*, 8> ExitBlocks;
|
2007-03-04 03:43:23 +00:00
|
|
|
L->getUniqueExitBlocks(ExitBlocks);
|
|
|
|
|
|
|
|
// Find all values that are computed inside the loop, but used outside of it.
|
|
|
|
// Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
|
|
|
|
// the exit blocks of the loop to find them.
|
|
|
|
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
|
|
|
|
BasicBlock *ExitBB = ExitBlocks[i];
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2007-03-04 03:43:23 +00:00
|
|
|
// If there are no PHI nodes in this exit block, then no values defined
|
|
|
|
// inside the loop are used on this path, skip it.
|
|
|
|
PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
|
|
|
|
if (!PN) continue;
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2007-03-04 03:43:23 +00:00
|
|
|
unsigned NumPreds = PN->getNumIncomingValues();
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2007-03-04 03:43:23 +00:00
|
|
|
// Iterate over all of the PHI nodes.
|
|
|
|
BasicBlock::iterator BBI = ExitBB->begin();
|
|
|
|
while ((PN = dyn_cast<PHINode>(BBI++))) {
|
2009-05-24 19:36:09 +00:00
|
|
|
if (PN->use_empty())
|
|
|
|
continue; // dead use, don't replace it
|
2010-02-18 21:34:02 +00:00
|
|
|
|
|
|
|
// SCEV only supports integer expressions for now.
|
|
|
|
if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
|
|
|
|
continue;
|
|
|
|
|
2010-02-19 07:14:22 +00:00
|
|
|
// It's necessary to tell ScalarEvolution about this explicitly so that
|
|
|
|
// it can walk the def-use list and forget all SCEVs, as it may not be
|
|
|
|
// watching the PHI itself. Once the new exit value is in place, there
|
|
|
|
// may not be a def-use connection between the loop and every instruction
|
|
|
|
// which got a SCEVAddRecExpr for that loop.
|
|
|
|
SE->forgetValue(PN);
|
|
|
|
|
2007-03-04 03:43:23 +00:00
|
|
|
// Iterate over all of the values in all the PHI nodes.
|
|
|
|
for (unsigned i = 0; i != NumPreds; ++i) {
|
|
|
|
// If the value being merged in is not integer or is not defined
|
|
|
|
// in the loop, skip it.
|
|
|
|
Value *InVal = PN->getIncomingValue(i);
|
2010-02-18 21:34:02 +00:00
|
|
|
if (!isa<Instruction>(InVal))
|
2007-03-04 03:43:23 +00:00
|
|
|
continue;
|
|
|
|
|
|
|
|
// If this pred is for a subloop, not L itself, skip it.
|
2009-02-17 19:13:57 +00:00
|
|
|
if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
|
2007-03-04 03:43:23 +00:00
|
|
|
continue; // The Block is in a subloop, skip it.
|
|
|
|
|
|
|
|
// Check that InVal is defined in the loop.
|
|
|
|
Instruction *Inst = cast<Instruction>(InVal);
|
2009-12-18 01:24:09 +00:00
|
|
|
if (!L->contains(Inst))
|
2007-03-04 03:43:23 +00:00
|
|
|
continue;
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2007-03-04 03:43:23 +00:00
|
|
|
// Okay, this instruction has a user outside of the current loop
|
|
|
|
// and varies predictably *inside* the loop. Evaluate the value it
|
|
|
|
// contains when the loop exits, if possible.
|
2009-07-07 17:06:11 +00:00
|
|
|
const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
|
2010-11-17 21:23:15 +00:00
|
|
|
if (!SE->isLoopInvariant(ExitValue, L))
|
2007-03-04 03:43:23 +00:00
|
|
|
continue;
|
|
|
|
|
2009-06-26 22:53:46 +00:00
|
|
|
Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2010-01-05 01:27:06 +00:00
|
|
|
DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
|
2009-08-23 04:37:46 +00:00
|
|
|
<< " LoopVal = " << *Inst << "\n");
|
2007-03-04 01:00:28 +00:00
|
|
|
|
2011-03-17 23:51:11 +00:00
|
|
|
if (!isValidRewrite(Inst, ExitVal)) {
|
|
|
|
DeadInsts.push_back(ExitVal);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
Changed = true;
|
|
|
|
++NumReplaced;
|
|
|
|
|
2007-03-04 03:43:23 +00:00
|
|
|
PN->setIncomingValue(i, ExitVal);
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2009-05-12 02:17:14 +00:00
|
|
|
// If this instruction is dead now, delete it.
|
|
|
|
RecursivelyDeleteTriviallyDeadInstructions(Inst);
|
2009-02-17 19:13:57 +00:00
|
|
|
|
2009-07-14 01:09:02 +00:00
|
|
|
if (NumPreds == 1) {
|
|
|
|
// Completely replace a single-pred PHI. This is safe, because the
|
|
|
|
// NewVal won't be variant in the loop, so we don't need an LCSSA phi
|
|
|
|
// node anymore.
|
2007-03-04 03:43:23 +00:00
|
|
|
PN->replaceAllUsesWith(ExitVal);
|
2009-05-12 02:17:14 +00:00
|
|
|
RecursivelyDeleteTriviallyDeadInstructions(PN);
|
2007-03-03 22:48:48 +00:00
|
|
|
}
|
2005-06-15 21:29:31 +00:00
|
|
|
}
|
2009-07-14 01:09:02 +00:00
|
|
|
if (NumPreds != 1) {
|
2009-06-26 22:53:46 +00:00
|
|
|
// Clone the PHI and delete the original one. This lets IVUsers and
|
|
|
|
// any other maps purge the original user from their records.
|
2009-10-27 22:16:29 +00:00
|
|
|
PHINode *NewPN = cast<PHINode>(PN->clone());
|
2009-06-26 22:53:46 +00:00
|
|
|
NewPN->takeName(PN);
|
|
|
|
NewPN->insertBefore(PN);
|
|
|
|
PN->replaceAllUsesWith(NewPN);
|
|
|
|
PN->eraseFromParent();
|
|
|
|
}
|
2007-03-03 22:48:48 +00:00
|
|
|
}
|
|
|
|
}
|
2010-03-20 03:53:53 +00:00
|
|
|
|
|
|
|
// The insertion point instruction may have been deleted; clear it out
|
|
|
|
// so that the rewriter doesn't trip over it later.
|
|
|
|
Rewriter.clearInsertPoint();
|
2004-04-02 20:24:31 +00:00
|
|
|
}
|
2003-12-23 07:47:09 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Rewrite IV users based on a canonical IV.
|
|
|
|
// To be replaced by -disable-iv-rewrite.
|
|
|
|
//===----------------------------------------------------------------------===//
|
2009-04-15 23:31:51 +00:00
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
/// SimplifyIVUsers - Iteratively perform simplification on IVUsers within this
|
|
|
|
/// loop. IVUsers is treated as a worklist. Each successive simplification may
|
|
|
|
/// push more users which may themselves be candidates for simplification.
|
|
|
|
///
|
|
|
|
/// This is the old approach to IV simplification to be replaced by
|
|
|
|
/// SimplifyIVUsersNoRewrite.
|
|
|
|
///
|
|
|
|
void IndVarSimplify::SimplifyIVUsers(SCEVExpander &Rewriter) {
|
|
|
|
// Each round of simplification involves a round of eliminating operations
|
|
|
|
// followed by a round of widening IVs. A single IVUsers worklist is used
|
|
|
|
// across all rounds. The inner loop advances the user. If widening exposes
|
|
|
|
// more uses, then another pass through the outer loop is triggered.
|
|
|
|
for (IVUsers::iterator I = IU->begin(); I != IU->end(); ++I) {
|
|
|
|
Instruction *UseInst = I->getUser();
|
|
|
|
Value *IVOperand = I->getOperandValToReplace();
|
|
|
|
|
|
|
|
if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
|
|
|
|
EliminateIVComparison(ICmp, IVOperand);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
|
|
|
|
bool IsSigned = Rem->getOpcode() == Instruction::SRem;
|
|
|
|
if (IsSigned || Rem->getOpcode() == Instruction::URem) {
|
2011-06-21 15:43:52 +00:00
|
|
|
EliminateIVRemainder(Rem, IVOperand, IsSigned);
|
2011-06-21 03:22:38 +00:00
|
|
|
continue;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// FIXME: It is an extremely bad idea to indvar substitute anything more
|
|
|
|
// complex than affine induction variables. Doing so will put expensive
|
|
|
|
// polynomial evaluations inside of the loop, and the str reduction pass
|
|
|
|
// currently can only reduce affine polynomials. For now just disable
|
|
|
|
// indvar subst on anything more complex than an affine addrec, unless
|
|
|
|
// it can be expanded to a trivial value.
|
|
|
|
static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
|
|
|
|
// Loop-invariant values are safe.
|
|
|
|
if (SE->isLoopInvariant(S, L)) return true;
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
|
|
|
|
// to transform them into efficient code.
|
|
|
|
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
|
|
|
|
return AR->isAffine();
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// An add is safe it all its operands are safe.
|
|
|
|
if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
|
|
|
|
for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
|
|
|
|
E = Commutative->op_end(); I != E; ++I)
|
|
|
|
if (!isSafe(*I, L, SE)) return false;
|
|
|
|
return true;
|
2011-05-20 18:25:42 +00:00
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// A cast is safe if its operand is.
|
|
|
|
if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
|
|
|
|
return isSafe(C->getOperand(), L, SE);
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// A udiv is safe if its operands are.
|
|
|
|
if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
|
|
|
|
return isSafe(UD->getLHS(), L, SE) &&
|
|
|
|
isSafe(UD->getRHS(), L, SE);
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// SCEVUnknown is always safe.
|
|
|
|
if (isa<SCEVUnknown>(S))
|
|
|
|
return true;
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Nothing else is safe.
|
|
|
|
return false;
|
|
|
|
}
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
|
|
|
|
// Rewrite all induction variable expressions in terms of the canonical
|
|
|
|
// induction variable.
|
|
|
|
//
|
|
|
|
// If there were induction variables of other sizes or offsets, manually
|
|
|
|
// add the offsets to the primary induction variable and cast, avoiding
|
|
|
|
// the need for the code evaluation methods to insert induction variables
|
|
|
|
// of different sizes.
|
|
|
|
for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
|
|
|
|
Value *Op = UI->getOperandValToReplace();
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *UseTy = Op->getType();
|
2011-07-12 00:08:50 +00:00
|
|
|
Instruction *User = UI->getUser();
|
|
|
|
|
|
|
|
// Compute the final addrec to expand into code.
|
|
|
|
const SCEV *AR = IU->getReplacementExpr(*UI);
|
|
|
|
|
|
|
|
// Evaluate the expression out of the loop, if possible.
|
|
|
|
if (!L->contains(UI->getUser())) {
|
|
|
|
const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
|
|
|
|
if (SE->isLoopInvariant(ExitVal, L))
|
|
|
|
AR = ExitVal;
|
|
|
|
}
|
|
|
|
|
|
|
|
// FIXME: It is an extremely bad idea to indvar substitute anything more
|
|
|
|
// complex than affine induction variables. Doing so will put expensive
|
|
|
|
// polynomial evaluations inside of the loop, and the str reduction pass
|
|
|
|
// currently can only reduce affine polynomials. For now just disable
|
|
|
|
// indvar subst on anything more complex than an affine addrec, unless
|
|
|
|
// it can be expanded to a trivial value.
|
|
|
|
if (!isSafe(AR, L, SE))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Determine the insertion point for this user. By default, insert
|
|
|
|
// immediately before the user. The SCEVExpander class will automatically
|
|
|
|
// hoist loop invariants out of the loop. For PHI nodes, there may be
|
|
|
|
// multiple uses, so compute the nearest common dominator for the
|
|
|
|
// incoming blocks.
|
|
|
|
Instruction *InsertPt = User;
|
|
|
|
if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
|
|
|
|
for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
|
|
|
|
if (PHI->getIncomingValue(i) == Op) {
|
|
|
|
if (InsertPt == User)
|
|
|
|
InsertPt = PHI->getIncomingBlock(i)->getTerminator();
|
|
|
|
else
|
|
|
|
InsertPt =
|
|
|
|
DT->findNearestCommonDominator(InsertPt->getParent(),
|
|
|
|
PHI->getIncomingBlock(i))
|
|
|
|
->getTerminator();
|
|
|
|
}
|
|
|
|
|
|
|
|
// Now expand it into actual Instructions and patch it into place.
|
|
|
|
Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
|
|
|
|
<< " into = " << *NewVal << "\n");
|
|
|
|
|
|
|
|
if (!isValidRewrite(Op, NewVal)) {
|
|
|
|
DeadInsts.push_back(NewVal);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
// Inform ScalarEvolution that this value is changing. The change doesn't
|
|
|
|
// affect its value, but it does potentially affect which use lists the
|
|
|
|
// value will be on after the replacement, which affects ScalarEvolution's
|
|
|
|
// ability to walk use lists and drop dangling pointers when a value is
|
|
|
|
// deleted.
|
|
|
|
SE->forgetValue(User);
|
|
|
|
|
|
|
|
// Patch the new value into place.
|
|
|
|
if (Op->hasName())
|
|
|
|
NewVal->takeName(Op);
|
|
|
|
if (Instruction *NewValI = dyn_cast<Instruction>(NewVal))
|
|
|
|
NewValI->setDebugLoc(User->getDebugLoc());
|
|
|
|
User->replaceUsesOfWith(Op, NewVal);
|
|
|
|
UI->setOperandValToReplace(NewVal);
|
|
|
|
|
|
|
|
++NumRemoved;
|
|
|
|
Changed = true;
|
|
|
|
|
|
|
|
// The old value may be dead now.
|
|
|
|
DeadInsts.push_back(Op);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// IV Widening - Extend the width of an IV to cover its widest uses.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
// Collect information about induction variables that are used by sign/zero
|
|
|
|
// extend operations. This information is recorded by CollectExtend and
|
|
|
|
// provides the input to WidenIV.
|
|
|
|
struct WideIVInfo {
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *WidestNativeType; // Widest integer type created [sz]ext
|
2011-07-12 00:08:50 +00:00
|
|
|
bool IsSigned; // Was an sext user seen before a zext?
|
|
|
|
|
|
|
|
WideIVInfo() : WidestNativeType(0), IsSigned(false) {}
|
|
|
|
};
|
|
|
|
}
|
|
|
|
|
|
|
|
/// CollectExtend - Update information about the induction variable that is
|
|
|
|
/// extended by this sign or zero extend operation. This is used to determine
|
|
|
|
/// the final width of the IV before actually widening it.
|
|
|
|
static void CollectExtend(CastInst *Cast, bool IsSigned, WideIVInfo &WI,
|
|
|
|
ScalarEvolution *SE, const TargetData *TD) {
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *Ty = Cast->getType();
|
2011-07-12 00:08:50 +00:00
|
|
|
uint64_t Width = SE->getTypeSizeInBits(Ty);
|
|
|
|
if (TD && !TD->isLegalInteger(Width))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!WI.WidestNativeType) {
|
|
|
|
WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
|
|
|
|
WI.IsSigned = IsSigned;
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
// We extend the IV to satisfy the sign of its first user, arbitrarily.
|
|
|
|
if (WI.IsSigned != IsSigned)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
|
|
|
|
WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
|
|
|
|
}
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
/// WidenIV - The goal of this transform is to remove sign and zero extends
|
|
|
|
/// without creating any new induction variables. To do this, it creates a new
|
|
|
|
/// phi of the wider type and redirects all users, either removing extends or
|
|
|
|
/// inserting truncs whenever we stop propagating the type.
|
|
|
|
///
|
|
|
|
class WidenIV {
|
|
|
|
// Parameters
|
|
|
|
PHINode *OrigPhi;
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *WideType;
|
2011-07-12 00:08:50 +00:00
|
|
|
bool IsSigned;
|
|
|
|
|
|
|
|
// Context
|
|
|
|
LoopInfo *LI;
|
|
|
|
Loop *L;
|
|
|
|
ScalarEvolution *SE;
|
|
|
|
DominatorTree *DT;
|
|
|
|
|
|
|
|
// Result
|
|
|
|
PHINode *WidePhi;
|
|
|
|
Instruction *WideInc;
|
|
|
|
const SCEV *WideIncExpr;
|
2011-06-21 03:22:38 +00:00
|
|
|
SmallVectorImpl<WeakVH> &DeadInsts;
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
SmallPtrSet<Instruction*,16> Widened;
|
2011-07-02 02:34:25 +00:00
|
|
|
SmallVector<std::pair<Use *, Instruction *>, 8> NarrowIVUsers;
|
2011-05-20 18:25:42 +00:00
|
|
|
|
|
|
|
public:
|
2011-06-21 03:22:38 +00:00
|
|
|
WidenIV(PHINode *PN, const WideIVInfo &WI, LoopInfo *LInfo,
|
|
|
|
ScalarEvolution *SEv, DominatorTree *DTree,
|
2011-05-26 00:46:11 +00:00
|
|
|
SmallVectorImpl<WeakVH> &DI) :
|
2011-05-20 18:25:42 +00:00
|
|
|
OrigPhi(PN),
|
2011-06-21 03:22:38 +00:00
|
|
|
WideType(WI.WidestNativeType),
|
|
|
|
IsSigned(WI.IsSigned),
|
2011-05-20 18:25:42 +00:00
|
|
|
LI(LInfo),
|
|
|
|
L(LI->getLoopFor(OrigPhi->getParent())),
|
|
|
|
SE(SEv),
|
2011-05-26 00:46:11 +00:00
|
|
|
DT(DTree),
|
2011-05-20 18:25:42 +00:00
|
|
|
WidePhi(0),
|
|
|
|
WideInc(0),
|
2011-06-21 03:22:38 +00:00
|
|
|
WideIncExpr(0),
|
|
|
|
DeadInsts(DI) {
|
2011-05-20 18:25:42 +00:00
|
|
|
assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
|
|
|
|
}
|
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
PHINode *CreateWideIV(SCEVExpander &Rewriter);
|
2011-05-20 18:25:42 +00:00
|
|
|
|
|
|
|
protected:
|
|
|
|
Instruction *CloneIVUser(Instruction *NarrowUse,
|
|
|
|
Instruction *NarrowDef,
|
|
|
|
Instruction *WideDef);
|
|
|
|
|
2011-07-05 18:19:39 +00:00
|
|
|
const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
|
|
|
|
|
2011-06-29 23:03:57 +00:00
|
|
|
Instruction *WidenIVUse(Use &NarrowDefUse, Instruction *NarrowDef,
|
2011-05-20 18:25:42 +00:00
|
|
|
Instruction *WideDef);
|
2011-07-02 02:34:25 +00:00
|
|
|
|
|
|
|
void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
|
2011-05-20 18:25:42 +00:00
|
|
|
};
|
|
|
|
} // anonymous namespace
|
|
|
|
|
2011-07-18 04:54:35 +00:00
|
|
|
static Value *getExtend( Value *NarrowOper, Type *WideType,
|
2011-05-25 04:42:22 +00:00
|
|
|
bool IsSigned, IRBuilder<> &Builder) {
|
|
|
|
return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
|
|
|
|
Builder.CreateZExt(NarrowOper, WideType);
|
2011-05-20 18:25:42 +00:00
|
|
|
}
|
2010-04-12 02:21:50 +00:00
|
|
|
|
2011-05-20 18:25:42 +00:00
|
|
|
/// CloneIVUser - Instantiate a wide operation to replace a narrow
|
|
|
|
/// operation. This only needs to handle operations that can evaluation to
|
|
|
|
/// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
|
|
|
|
Instruction *WidenIV::CloneIVUser(Instruction *NarrowUse,
|
|
|
|
Instruction *NarrowDef,
|
|
|
|
Instruction *WideDef) {
|
|
|
|
unsigned Opcode = NarrowUse->getOpcode();
|
|
|
|
switch (Opcode) {
|
|
|
|
default:
|
|
|
|
return 0;
|
|
|
|
case Instruction::Add:
|
|
|
|
case Instruction::Mul:
|
|
|
|
case Instruction::UDiv:
|
|
|
|
case Instruction::Sub:
|
|
|
|
case Instruction::And:
|
|
|
|
case Instruction::Or:
|
|
|
|
case Instruction::Xor:
|
|
|
|
case Instruction::Shl:
|
|
|
|
case Instruction::LShr:
|
|
|
|
case Instruction::AShr:
|
|
|
|
DEBUG(dbgs() << "Cloning IVUser: " << *NarrowUse << "\n");
|
|
|
|
|
2011-05-25 04:42:22 +00:00
|
|
|
IRBuilder<> Builder(NarrowUse);
|
|
|
|
|
|
|
|
// Replace NarrowDef operands with WideDef. Otherwise, we don't know
|
|
|
|
// anything about the narrow operand yet so must insert a [sz]ext. It is
|
|
|
|
// probably loop invariant and will be folded or hoisted. If it actually
|
|
|
|
// comes from a widened IV, it should be removed during a future call to
|
|
|
|
// WidenIVUse.
|
|
|
|
Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) ? WideDef :
|
|
|
|
getExtend(NarrowUse->getOperand(0), WideType, IsSigned, Builder);
|
|
|
|
Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) ? WideDef :
|
|
|
|
getExtend(NarrowUse->getOperand(1), WideType, IsSigned, Builder);
|
|
|
|
|
2011-05-20 18:25:42 +00:00
|
|
|
BinaryOperator *NarrowBO = cast<BinaryOperator>(NarrowUse);
|
|
|
|
BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
|
2011-05-25 04:42:22 +00:00
|
|
|
LHS, RHS,
|
2011-05-20 18:25:42 +00:00
|
|
|
NarrowBO->getName());
|
|
|
|
Builder.Insert(WideBO);
|
2011-06-30 19:02:17 +00:00
|
|
|
if (const OverflowingBinaryOperator *OBO =
|
|
|
|
dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
|
|
|
|
if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
|
|
|
|
if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
|
|
|
|
}
|
2011-05-25 04:42:22 +00:00
|
|
|
return WideBO;
|
2011-05-20 18:25:42 +00:00
|
|
|
}
|
|
|
|
llvm_unreachable(0);
|
|
|
|
}
|
|
|
|
|
2011-05-26 00:46:11 +00:00
|
|
|
/// HoistStep - Attempt to hoist an IV increment above a potential use.
|
|
|
|
///
|
|
|
|
/// To successfully hoist, two criteria must be met:
|
|
|
|
/// - IncV operands dominate InsertPos and
|
|
|
|
/// - InsertPos dominates IncV
|
|
|
|
///
|
|
|
|
/// Meeting the second condition means that we don't need to check all of IncV's
|
|
|
|
/// existing uses (it's moving up in the domtree).
|
|
|
|
///
|
|
|
|
/// This does not yet recursively hoist the operands, although that would
|
|
|
|
/// not be difficult.
|
|
|
|
static bool HoistStep(Instruction *IncV, Instruction *InsertPos,
|
|
|
|
const DominatorTree *DT)
|
|
|
|
{
|
|
|
|
if (DT->dominates(IncV, InsertPos))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
if (!DT->dominates(InsertPos->getParent(), IncV->getParent()))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (IncV->mayHaveSideEffects())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Attempt to hoist IncV
|
|
|
|
for (User::op_iterator OI = IncV->op_begin(), OE = IncV->op_end();
|
|
|
|
OI != OE; ++OI) {
|
|
|
|
Instruction *OInst = dyn_cast<Instruction>(OI);
|
|
|
|
if (OInst && !DT->dominates(OInst, InsertPos))
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
IncV->moveBefore(InsertPos);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2011-07-05 18:19:39 +00:00
|
|
|
// GetWideRecurrence - Is this instruction potentially interesting from IVUsers'
|
|
|
|
// perspective after widening it's type? In other words, can the extend be
|
|
|
|
// safely hoisted out of the loop with SCEV reducing the value to a recurrence
|
|
|
|
// on the same loop. If so, return the sign or zero extended
|
|
|
|
// recurrence. Otherwise return NULL.
|
|
|
|
const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
|
|
|
|
if (!SE->isSCEVable(NarrowUse->getType()))
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
|
|
|
|
if (SE->getTypeSizeInBits(NarrowExpr->getType())
|
|
|
|
>= SE->getTypeSizeInBits(WideType)) {
|
|
|
|
// NarrowUse implicitly widens its operand. e.g. a gep with a narrow
|
|
|
|
// index. So don't follow this use.
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
const SCEV *WideExpr = IsSigned ?
|
|
|
|
SE->getSignExtendExpr(NarrowExpr, WideType) :
|
|
|
|
SE->getZeroExtendExpr(NarrowExpr, WideType);
|
|
|
|
const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
|
|
|
|
if (!AddRec || AddRec->getLoop() != L)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
return AddRec;
|
|
|
|
}
|
|
|
|
|
2011-05-20 18:25:42 +00:00
|
|
|
/// WidenIVUse - Determine whether an individual user of the narrow IV can be
|
|
|
|
/// widened. If so, return the wide clone of the user.
|
2011-06-29 23:03:57 +00:00
|
|
|
Instruction *WidenIV::WidenIVUse(Use &NarrowDefUse, Instruction *NarrowDef,
|
2011-05-20 18:25:42 +00:00
|
|
|
Instruction *WideDef) {
|
2011-06-29 23:03:57 +00:00
|
|
|
Instruction *NarrowUse = cast<Instruction>(NarrowDefUse.getUser());
|
|
|
|
|
2011-07-02 02:34:25 +00:00
|
|
|
// Stop traversing the def-use chain at inner-loop phis or post-loop phis.
|
2011-05-20 18:25:42 +00:00
|
|
|
if (isa<PHINode>(NarrowUse) && LI->getLoopFor(NarrowUse->getParent()) != L)
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
// Our raison d'etre! Eliminate sign and zero extension.
|
|
|
|
if (IsSigned ? isa<SExtInst>(NarrowUse) : isa<ZExtInst>(NarrowUse)) {
|
2011-05-25 04:42:22 +00:00
|
|
|
Value *NewDef = WideDef;
|
|
|
|
if (NarrowUse->getType() != WideType) {
|
|
|
|
unsigned CastWidth = SE->getTypeSizeInBits(NarrowUse->getType());
|
|
|
|
unsigned IVWidth = SE->getTypeSizeInBits(WideType);
|
|
|
|
if (CastWidth < IVWidth) {
|
|
|
|
// The cast isn't as wide as the IV, so insert a Trunc.
|
2011-06-29 23:03:57 +00:00
|
|
|
IRBuilder<> Builder(NarrowDefUse);
|
2011-05-25 04:42:22 +00:00
|
|
|
NewDef = Builder.CreateTrunc(WideDef, NarrowUse->getType());
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
// A wider extend was hidden behind a narrower one. This may induce
|
|
|
|
// another round of IV widening in which the intermediate IV becomes
|
|
|
|
// dead. It should be very rare.
|
|
|
|
DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
|
|
|
|
<< " not wide enough to subsume " << *NarrowUse << "\n");
|
|
|
|
NarrowUse->replaceUsesOfWith(NarrowDef, WideDef);
|
|
|
|
NewDef = NarrowUse;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (NewDef != NarrowUse) {
|
|
|
|
DEBUG(dbgs() << "INDVARS: eliminating " << *NarrowUse
|
|
|
|
<< " replaced by " << *WideDef << "\n");
|
|
|
|
++NumElimExt;
|
|
|
|
NarrowUse->replaceAllUsesWith(NewDef);
|
|
|
|
DeadInsts.push_back(NarrowUse);
|
|
|
|
}
|
2011-06-21 03:22:38 +00:00
|
|
|
// Now that the extend is gone, we want to expose it's uses for potential
|
|
|
|
// further simplification. We don't need to directly inform SimplifyIVUsers
|
|
|
|
// of the new users, because their parent IV will be processed later as a
|
|
|
|
// new loop phi. If we preserved IVUsers analysis, we would also want to
|
|
|
|
// push the uses of WideDef here.
|
2011-05-20 18:25:42 +00:00
|
|
|
|
|
|
|
// No further widening is needed. The deceased [sz]ext had done it for us.
|
|
|
|
return 0;
|
|
|
|
}
|
2011-07-02 02:34:25 +00:00
|
|
|
|
|
|
|
// Does this user itself evaluate to a recurrence after widening?
|
2011-07-05 18:19:39 +00:00
|
|
|
const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(NarrowUse);
|
2011-05-20 18:25:42 +00:00
|
|
|
if (!WideAddRec) {
|
|
|
|
// This user does not evaluate to a recurence after widening, so don't
|
|
|
|
// follow it. Instead insert a Trunc to kill off the original use,
|
|
|
|
// eventually isolating the original narrow IV so it can be removed.
|
2011-06-29 23:03:57 +00:00
|
|
|
IRBuilder<> Builder(NarrowDefUse);
|
2011-05-25 04:42:22 +00:00
|
|
|
Value *Trunc = Builder.CreateTrunc(WideDef, NarrowDef->getType());
|
2011-05-20 18:25:42 +00:00
|
|
|
NarrowUse->replaceUsesOfWith(NarrowDef, Trunc);
|
|
|
|
return 0;
|
|
|
|
}
|
2011-07-02 02:34:25 +00:00
|
|
|
// We assume that block terminators are not SCEVable. We wouldn't want to
|
|
|
|
// insert a Trunc after a terminator if there happens to be a critical edge.
|
2011-06-29 23:03:57 +00:00
|
|
|
assert(NarrowUse != NarrowUse->getParent()->getTerminator() &&
|
2011-07-02 02:34:25 +00:00
|
|
|
"SCEV is not expected to evaluate a block terminator");
|
2011-06-29 23:03:57 +00:00
|
|
|
|
2011-05-26 00:46:11 +00:00
|
|
|
// Reuse the IV increment that SCEVExpander created as long as it dominates
|
|
|
|
// NarrowUse.
|
2011-05-20 18:25:42 +00:00
|
|
|
Instruction *WideUse = 0;
|
2011-05-26 00:46:11 +00:00
|
|
|
if (WideAddRec == WideIncExpr && HoistStep(WideInc, NarrowUse, DT)) {
|
2011-05-20 18:25:42 +00:00
|
|
|
WideUse = WideInc;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
WideUse = CloneIVUser(NarrowUse, NarrowDef, WideDef);
|
|
|
|
if (!WideUse)
|
|
|
|
return 0;
|
|
|
|
}
|
2011-07-02 02:34:25 +00:00
|
|
|
// Evaluation of WideAddRec ensured that the narrow expression could be
|
|
|
|
// extended outside the loop without overflow. This suggests that the wide use
|
2011-05-20 18:25:42 +00:00
|
|
|
// evaluates to the same expression as the extended narrow use, but doesn't
|
|
|
|
// absolutely guarantee it. Hence the following failsafe check. In rare cases
|
2011-06-21 03:22:38 +00:00
|
|
|
// where it fails, we simply throw away the newly created wide use.
|
2011-05-20 18:25:42 +00:00
|
|
|
if (WideAddRec != SE->getSCEV(WideUse)) {
|
|
|
|
DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
|
|
|
|
<< ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
|
|
|
|
DeadInsts.push_back(WideUse);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Returning WideUse pushes it on the worklist.
|
|
|
|
return WideUse;
|
|
|
|
}
|
|
|
|
|
2011-07-02 02:34:25 +00:00
|
|
|
/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
|
|
|
|
///
|
|
|
|
void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
|
|
|
|
for (Value::use_iterator UI = NarrowDef->use_begin(),
|
|
|
|
UE = NarrowDef->use_end(); UI != UE; ++UI) {
|
|
|
|
Use &U = UI.getUse();
|
|
|
|
|
|
|
|
// Handle data flow merges and bizarre phi cycles.
|
|
|
|
if (!Widened.insert(cast<Instruction>(U.getUser())))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
NarrowIVUsers.push_back(std::make_pair(&UI.getUse(), WideDef));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-05-20 18:25:42 +00:00
|
|
|
/// CreateWideIV - Process a single induction variable. First use the
|
|
|
|
/// SCEVExpander to create a wide induction variable that evaluates to the same
|
|
|
|
/// recurrence as the original narrow IV. Then use a worklist to forward
|
2011-06-21 03:22:38 +00:00
|
|
|
/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
|
2011-05-20 18:25:42 +00:00
|
|
|
/// interesting IV users, the narrow IV will be isolated for removal by
|
|
|
|
/// DeleteDeadPHIs.
|
|
|
|
///
|
|
|
|
/// It would be simpler to delete uses as they are processed, but we must avoid
|
|
|
|
/// invalidating SCEV expressions.
|
|
|
|
///
|
2011-06-21 03:22:38 +00:00
|
|
|
PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
|
2011-05-20 18:25:42 +00:00
|
|
|
// Is this phi an induction variable?
|
|
|
|
const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
|
|
|
|
if (!AddRec)
|
2011-06-21 03:22:38 +00:00
|
|
|
return NULL;
|
2011-05-20 18:25:42 +00:00
|
|
|
|
|
|
|
// Widen the induction variable expression.
|
|
|
|
const SCEV *WideIVExpr = IsSigned ?
|
|
|
|
SE->getSignExtendExpr(AddRec, WideType) :
|
|
|
|
SE->getZeroExtendExpr(AddRec, WideType);
|
|
|
|
|
|
|
|
assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
|
|
|
|
"Expect the new IV expression to preserve its type");
|
|
|
|
|
|
|
|
// Can the IV be extended outside the loop without overflow?
|
|
|
|
AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
|
|
|
|
if (!AddRec || AddRec->getLoop() != L)
|
2011-06-21 03:22:38 +00:00
|
|
|
return NULL;
|
2011-05-20 18:25:42 +00:00
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
// An AddRec must have loop-invariant operands. Since this AddRec is
|
2011-05-20 18:25:42 +00:00
|
|
|
// materialized by a loop header phi, the expression cannot have any post-loop
|
|
|
|
// operands, so they must dominate the loop header.
|
|
|
|
assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
|
|
|
|
SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
|
|
|
|
&& "Loop header phi recurrence inputs do not dominate the loop");
|
|
|
|
|
|
|
|
// The rewriter provides a value for the desired IV expression. This may
|
|
|
|
// either find an existing phi or materialize a new one. Either way, we
|
|
|
|
// expect a well-formed cyclic phi-with-increments. i.e. any operand not part
|
|
|
|
// of the phi-SCC dominates the loop entry.
|
|
|
|
Instruction *InsertPt = L->getHeader()->begin();
|
|
|
|
WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
|
|
|
|
|
|
|
|
// Remembering the WideIV increment generated by SCEVExpander allows
|
|
|
|
// WidenIVUse to reuse it when widening the narrow IV's increment. We don't
|
|
|
|
// employ a general reuse mechanism because the call above is the only call to
|
|
|
|
// SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
|
2011-05-26 00:46:11 +00:00
|
|
|
if (BasicBlock *LatchBlock = L->getLoopLatch()) {
|
|
|
|
WideInc =
|
|
|
|
cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
|
|
|
|
WideIncExpr = SE->getSCEV(WideInc);
|
|
|
|
}
|
2011-05-20 18:25:42 +00:00
|
|
|
|
|
|
|
DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
|
|
|
|
++NumWidened;
|
|
|
|
|
|
|
|
// Traverse the def-use chain using a worklist starting at the original IV.
|
2011-07-02 02:34:25 +00:00
|
|
|
assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
|
|
|
|
|
|
|
|
Widened.insert(OrigPhi);
|
|
|
|
pushNarrowIVUsers(OrigPhi, WidePhi);
|
2011-05-20 18:25:42 +00:00
|
|
|
|
|
|
|
while (!NarrowIVUsers.empty()) {
|
2011-06-29 23:03:57 +00:00
|
|
|
Use *UsePtr;
|
2011-05-26 00:46:11 +00:00
|
|
|
Instruction *WideDef;
|
2011-06-29 23:03:57 +00:00
|
|
|
tie(UsePtr, WideDef) = NarrowIVUsers.pop_back_val();
|
|
|
|
Use &NarrowDefUse = *UsePtr;
|
2011-05-26 00:46:11 +00:00
|
|
|
|
|
|
|
// Process a def-use edge. This may replace the use, so don't hold a
|
|
|
|
// use_iterator across it.
|
2011-06-29 23:03:57 +00:00
|
|
|
Instruction *NarrowDef = cast<Instruction>(NarrowDefUse.get());
|
|
|
|
Instruction *WideUse = WidenIVUse(NarrowDefUse, NarrowDef, WideDef);
|
2011-05-26 00:46:11 +00:00
|
|
|
|
|
|
|
// Follow all def-use edges from the previous narrow use.
|
2011-07-02 02:34:25 +00:00
|
|
|
if (WideUse)
|
|
|
|
pushNarrowIVUsers(cast<Instruction>(NarrowDefUse.getUser()), WideUse);
|
|
|
|
|
2011-05-26 00:46:11 +00:00
|
|
|
// WidenIVUse may have removed the def-use edge.
|
|
|
|
if (NarrowDef->use_empty())
|
|
|
|
DeadInsts.push_back(NarrowDef);
|
2010-04-12 02:21:50 +00:00
|
|
|
}
|
2011-06-21 03:22:38 +00:00
|
|
|
return WidePhi;
|
2010-04-12 02:21:50 +00:00
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Simplification of IV users based on SCEV evaluation.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2011-05-12 00:04:28 +00:00
|
|
|
void IndVarSimplify::EliminateIVComparison(ICmpInst *ICmp, Value *IVOperand) {
|
|
|
|
unsigned IVOperIdx = 0;
|
|
|
|
ICmpInst::Predicate Pred = ICmp->getPredicate();
|
|
|
|
if (IVOperand != ICmp->getOperand(0)) {
|
|
|
|
// Swapped
|
|
|
|
assert(IVOperand == ICmp->getOperand(1) && "Can't find IVOperand");
|
|
|
|
IVOperIdx = 1;
|
|
|
|
Pred = ICmpInst::getSwappedPredicate(Pred);
|
|
|
|
}
|
2010-04-13 01:46:36 +00:00
|
|
|
|
2011-05-12 00:04:28 +00:00
|
|
|
// Get the SCEVs for the ICmp operands.
|
|
|
|
const SCEV *S = SE->getSCEV(ICmp->getOperand(IVOperIdx));
|
|
|
|
const SCEV *X = SE->getSCEV(ICmp->getOperand(1 - IVOperIdx));
|
|
|
|
|
|
|
|
// Simplify unnecessary loops away.
|
|
|
|
const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
|
|
|
|
S = SE->getSCEVAtScope(S, ICmpLoop);
|
|
|
|
X = SE->getSCEVAtScope(X, ICmpLoop);
|
|
|
|
|
|
|
|
// If the condition is always true or always false, replace it with
|
|
|
|
// a constant value.
|
|
|
|
if (SE->isKnownPredicate(Pred, S, X))
|
|
|
|
ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
|
|
|
|
else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
|
|
|
|
ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
|
|
|
|
else
|
|
|
|
return;
|
2010-04-13 01:46:36 +00:00
|
|
|
|
2011-05-12 00:04:28 +00:00
|
|
|
DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
|
2011-05-25 04:42:22 +00:00
|
|
|
++NumElimCmp;
|
2011-05-20 03:37:48 +00:00
|
|
|
Changed = true;
|
2011-05-12 00:04:28 +00:00
|
|
|
DeadInsts.push_back(ICmp);
|
|
|
|
}
|
2010-04-13 01:46:36 +00:00
|
|
|
|
2011-05-12 00:04:28 +00:00
|
|
|
void IndVarSimplify::EliminateIVRemainder(BinaryOperator *Rem,
|
|
|
|
Value *IVOperand,
|
2011-06-21 15:43:52 +00:00
|
|
|
bool IsSigned) {
|
2011-05-12 00:04:28 +00:00
|
|
|
// We're only interested in the case where we know something about
|
|
|
|
// the numerator.
|
|
|
|
if (IVOperand != Rem->getOperand(0))
|
|
|
|
return;
|
|
|
|
|
|
|
|
// Get the SCEVs for the ICmp operands.
|
|
|
|
const SCEV *S = SE->getSCEV(Rem->getOperand(0));
|
|
|
|
const SCEV *X = SE->getSCEV(Rem->getOperand(1));
|
|
|
|
|
|
|
|
// Simplify unnecessary loops away.
|
|
|
|
const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
|
|
|
|
S = SE->getSCEVAtScope(S, ICmpLoop);
|
|
|
|
X = SE->getSCEVAtScope(X, ICmpLoop);
|
|
|
|
|
|
|
|
// i % n --> i if i is in [0,n).
|
2011-05-20 03:37:48 +00:00
|
|
|
if ((!IsSigned || SE->isKnownNonNegative(S)) &&
|
|
|
|
SE->isKnownPredicate(IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
|
2011-05-12 00:04:28 +00:00
|
|
|
S, X))
|
|
|
|
Rem->replaceAllUsesWith(Rem->getOperand(0));
|
|
|
|
else {
|
|
|
|
// (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
|
|
|
|
const SCEV *LessOne =
|
|
|
|
SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
|
2011-05-20 03:37:48 +00:00
|
|
|
if (IsSigned && !SE->isKnownNonNegative(LessOne))
|
2011-05-12 00:04:28 +00:00
|
|
|
return;
|
|
|
|
|
2011-05-20 03:37:48 +00:00
|
|
|
if (!SE->isKnownPredicate(IsSigned ?
|
2011-05-12 00:04:28 +00:00
|
|
|
ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
|
|
|
|
LessOne, X))
|
|
|
|
return;
|
|
|
|
|
|
|
|
ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
|
|
|
|
Rem->getOperand(0), Rem->getOperand(1),
|
|
|
|
"tmp");
|
|
|
|
SelectInst *Sel =
|
|
|
|
SelectInst::Create(ICmp,
|
|
|
|
ConstantInt::get(Rem->getType(), 0),
|
|
|
|
Rem->getOperand(0), "tmp", Rem);
|
|
|
|
Rem->replaceAllUsesWith(Sel);
|
|
|
|
}
|
2010-04-13 01:46:36 +00:00
|
|
|
|
2011-05-12 00:04:28 +00:00
|
|
|
// Inform IVUsers about the new users.
|
2011-06-21 03:22:38 +00:00
|
|
|
if (IU) {
|
|
|
|
if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
|
2011-06-21 15:43:52 +00:00
|
|
|
IU->AddUsersIfInteresting(I);
|
2011-06-21 03:22:38 +00:00
|
|
|
}
|
2011-05-12 00:04:28 +00:00
|
|
|
DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
|
2011-05-25 04:42:22 +00:00
|
|
|
++NumElimRem;
|
2011-05-20 03:37:48 +00:00
|
|
|
Changed = true;
|
2011-05-12 00:04:28 +00:00
|
|
|
DeadInsts.push_back(Rem);
|
2010-04-13 01:46:36 +00:00
|
|
|
}
|
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
/// EliminateIVUser - Eliminate an operation that consumes a simple IV and has
|
|
|
|
/// no observable side-effect given the range of IV values.
|
|
|
|
bool IndVarSimplify::EliminateIVUser(Instruction *UseInst,
|
|
|
|
Instruction *IVOperand) {
|
|
|
|
if (ICmpInst *ICmp = dyn_cast<ICmpInst>(UseInst)) {
|
|
|
|
EliminateIVComparison(ICmp, IVOperand);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
if (BinaryOperator *Rem = dyn_cast<BinaryOperator>(UseInst)) {
|
|
|
|
bool IsSigned = Rem->getOpcode() == Instruction::SRem;
|
|
|
|
if (IsSigned || Rem->getOpcode() == Instruction::URem) {
|
2011-06-21 15:43:52 +00:00
|
|
|
EliminateIVRemainder(Rem, IVOperand, IsSigned);
|
2011-06-21 03:22:38 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Eliminate any operation that SCEV can prove is an identity function.
|
|
|
|
if (!SE->isSCEVable(UseInst->getType()) ||
|
2011-06-29 03:13:40 +00:00
|
|
|
(UseInst->getType() != IVOperand->getType()) ||
|
2011-06-21 03:22:38 +00:00
|
|
|
(SE->getSCEV(UseInst) != SE->getSCEV(IVOperand)))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "INDVARS: Eliminated identity: " << *UseInst << '\n');
|
2011-06-30 01:27:23 +00:00
|
|
|
|
|
|
|
UseInst->replaceAllUsesWith(IVOperand);
|
2011-06-21 03:22:38 +00:00
|
|
|
++NumElimIdentity;
|
|
|
|
Changed = true;
|
|
|
|
DeadInsts.push_back(UseInst);
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// pushIVUsers - Add all uses of Def to the current IV's worklist.
|
|
|
|
///
|
2011-06-28 02:49:20 +00:00
|
|
|
static void pushIVUsers(
|
|
|
|
Instruction *Def,
|
|
|
|
SmallPtrSet<Instruction*,16> &Simplified,
|
|
|
|
SmallVectorImpl< std::pair<Instruction*,Instruction*> > &SimpleIVUsers) {
|
2011-06-21 03:22:38 +00:00
|
|
|
|
|
|
|
for (Value::use_iterator UI = Def->use_begin(), E = Def->use_end();
|
|
|
|
UI != E; ++UI) {
|
|
|
|
Instruction *User = cast<Instruction>(*UI);
|
|
|
|
|
|
|
|
// Avoid infinite or exponential worklist processing.
|
|
|
|
// Also ensure unique worklist users.
|
2011-06-30 01:27:23 +00:00
|
|
|
// If Def is a LoopPhi, it may not be in the Simplified set, so check for
|
|
|
|
// self edges first.
|
|
|
|
if (User != Def && Simplified.insert(User))
|
2011-06-21 03:22:38 +00:00
|
|
|
SimpleIVUsers.push_back(std::make_pair(User, Def));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// isSimpleIVUser - Return true if this instruction generates a simple SCEV
|
|
|
|
/// expression in terms of that IV.
|
|
|
|
///
|
|
|
|
/// This is similar to IVUsers' isInsteresting() but processes each instruction
|
|
|
|
/// non-recursively when the operand is already known to be a simpleIVUser.
|
|
|
|
///
|
2011-07-12 00:08:50 +00:00
|
|
|
static bool isSimpleIVUser(Instruction *I, const Loop *L, ScalarEvolution *SE) {
|
2011-06-21 03:22:38 +00:00
|
|
|
if (!SE->isSCEVable(I->getType()))
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Get the symbolic expression for this instruction.
|
|
|
|
const SCEV *S = SE->getSCEV(I);
|
|
|
|
|
2011-06-29 23:03:57 +00:00
|
|
|
// We assume that terminators are not SCEVable.
|
|
|
|
assert((!S || I != I->getParent()->getTerminator()) &&
|
|
|
|
"can't fold terminators");
|
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
// Only consider affine recurrences.
|
|
|
|
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S);
|
|
|
|
if (AR && AR->getLoop() == L)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// SimplifyIVUsersNoRewrite - Iteratively perform simplification on a worklist
|
|
|
|
/// of IV users. Each successive simplification may push more users which may
|
|
|
|
/// themselves be candidates for simplification.
|
|
|
|
///
|
|
|
|
/// The "NoRewrite" algorithm does not require IVUsers analysis. Instead, it
|
|
|
|
/// simplifies instructions in-place during analysis. Rather than rewriting
|
|
|
|
/// induction variables bottom-up from their users, it transforms a chain of
|
|
|
|
/// IVUsers top-down, updating the IR only when it encouters a clear
|
|
|
|
/// optimization opportunitiy. A SCEVExpander "Rewriter" instance is still
|
|
|
|
/// needed, but only used to generate a new IV (phi) of wider type for sign/zero
|
|
|
|
/// extend elimination.
|
|
|
|
///
|
|
|
|
/// Once DisableIVRewrite is default, LSR will be the only client of IVUsers.
|
|
|
|
///
|
|
|
|
void IndVarSimplify::SimplifyIVUsersNoRewrite(Loop *L, SCEVExpander &Rewriter) {
|
2011-06-28 02:49:20 +00:00
|
|
|
std::map<PHINode *, WideIVInfo> WideIVMap;
|
|
|
|
|
2011-06-21 03:22:38 +00:00
|
|
|
SmallVector<PHINode*, 8> LoopPhis;
|
|
|
|
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
|
|
|
|
LoopPhis.push_back(cast<PHINode>(I));
|
|
|
|
}
|
2011-06-28 02:49:20 +00:00
|
|
|
// Each round of simplification iterates through the SimplifyIVUsers worklist
|
|
|
|
// for all current phis, then determines whether any IVs can be
|
|
|
|
// widened. Widening adds new phis to LoopPhis, inducing another round of
|
|
|
|
// simplification on the wide IVs.
|
2011-06-21 03:22:38 +00:00
|
|
|
while (!LoopPhis.empty()) {
|
2011-06-28 02:49:20 +00:00
|
|
|
// Evaluate as many IV expressions as possible before widening any IVs. This
|
2011-06-28 16:45:04 +00:00
|
|
|
// forces SCEV to set no-wrap flags before evaluating sign/zero
|
2011-06-28 02:49:20 +00:00
|
|
|
// extension. The first time SCEV attempts to normalize sign/zero extension,
|
|
|
|
// the result becomes final. So for the most predictable results, we delay
|
|
|
|
// evaluation of sign/zero extend evaluation until needed, and avoid running
|
|
|
|
// other SCEV based analysis prior to SimplifyIVUsersNoRewrite.
|
|
|
|
do {
|
|
|
|
PHINode *CurrIV = LoopPhis.pop_back_val();
|
2011-06-21 03:22:38 +00:00
|
|
|
|
2011-06-28 02:49:20 +00:00
|
|
|
// Information about sign/zero extensions of CurrIV.
|
|
|
|
WideIVInfo WI;
|
2011-06-21 03:22:38 +00:00
|
|
|
|
2011-06-28 02:49:20 +00:00
|
|
|
// Instructions processed by SimplifyIVUsers for CurrIV.
|
|
|
|
SmallPtrSet<Instruction*,16> Simplified;
|
2011-06-21 03:22:38 +00:00
|
|
|
|
2011-07-06 20:50:43 +00:00
|
|
|
// Use-def pairs if IV users waiting to be processed for CurrIV.
|
2011-06-28 02:49:20 +00:00
|
|
|
SmallVector<std::pair<Instruction*, Instruction*>, 8> SimpleIVUsers;
|
2011-06-21 03:22:38 +00:00
|
|
|
|
2011-06-30 01:27:23 +00:00
|
|
|
// Push users of the current LoopPhi. In rare cases, pushIVUsers may be
|
|
|
|
// called multiple times for the same LoopPhi. This is the proper thing to
|
|
|
|
// do for loop header phis that use each other.
|
2011-06-28 02:49:20 +00:00
|
|
|
pushIVUsers(CurrIV, Simplified, SimpleIVUsers);
|
|
|
|
|
|
|
|
while (!SimpleIVUsers.empty()) {
|
|
|
|
Instruction *UseInst, *Operand;
|
|
|
|
tie(UseInst, Operand) = SimpleIVUsers.pop_back_val();
|
2011-06-30 19:02:17 +00:00
|
|
|
// Bypass back edges to avoid extra work.
|
|
|
|
if (UseInst == CurrIV) continue;
|
2011-06-28 02:49:20 +00:00
|
|
|
|
|
|
|
if (EliminateIVUser(UseInst, Operand)) {
|
|
|
|
pushIVUsers(Operand, Simplified, SimpleIVUsers);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
if (CastInst *Cast = dyn_cast<CastInst>(UseInst)) {
|
|
|
|
bool IsSigned = Cast->getOpcode() == Instruction::SExt;
|
|
|
|
if (IsSigned || Cast->getOpcode() == Instruction::ZExt) {
|
|
|
|
CollectExtend(Cast, IsSigned, WI, SE, TD);
|
|
|
|
}
|
|
|
|
continue;
|
|
|
|
}
|
2011-07-12 00:08:50 +00:00
|
|
|
if (isSimpleIVUser(UseInst, L, SE)) {
|
2011-06-28 02:49:20 +00:00
|
|
|
pushIVUsers(UseInst, Simplified, SimpleIVUsers);
|
2011-06-21 03:22:38 +00:00
|
|
|
}
|
|
|
|
}
|
2011-06-28 02:49:20 +00:00
|
|
|
if (WI.WidestNativeType) {
|
|
|
|
WideIVMap[CurrIV] = WI;
|
2011-06-21 03:22:38 +00:00
|
|
|
}
|
2011-06-28 02:49:20 +00:00
|
|
|
} while(!LoopPhis.empty());
|
|
|
|
|
|
|
|
for (std::map<PHINode *, WideIVInfo>::const_iterator I = WideIVMap.begin(),
|
|
|
|
E = WideIVMap.end(); I != E; ++I) {
|
|
|
|
WidenIV Widener(I->first, I->second, LI, SE, DT, DeadInsts);
|
2011-06-21 03:22:38 +00:00
|
|
|
if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
|
|
|
|
Changed = true;
|
|
|
|
LoopPhis.push_back(WidePhi);
|
|
|
|
}
|
|
|
|
}
|
2011-06-28 02:49:20 +00:00
|
|
|
WideIVMap.clear();
|
2011-06-21 03:22:38 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-06 20:50:43 +00:00
|
|
|
/// SimplifyCongruentIVs - Check for congruent phis in this loop header and
|
|
|
|
/// populate ExprToIVMap for use later.
|
|
|
|
///
|
|
|
|
void IndVarSimplify::SimplifyCongruentIVs(Loop *L) {
|
2011-07-16 01:06:48 +00:00
|
|
|
DenseMap<const SCEV *, PHINode *> ExprToIVMap;
|
2011-07-06 20:50:43 +00:00
|
|
|
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
|
|
|
|
PHINode *Phi = cast<PHINode>(I);
|
2011-07-12 00:08:50 +00:00
|
|
|
if (!SE->isSCEVable(Phi->getType()))
|
|
|
|
continue;
|
|
|
|
|
2011-07-06 20:50:43 +00:00
|
|
|
const SCEV *S = SE->getSCEV(Phi);
|
2011-07-16 01:06:48 +00:00
|
|
|
DenseMap<const SCEV *, PHINode *>::const_iterator Pos;
|
2011-07-06 20:50:43 +00:00
|
|
|
bool Inserted;
|
|
|
|
tie(Pos, Inserted) = ExprToIVMap.insert(std::make_pair(S, Phi));
|
|
|
|
if (Inserted)
|
|
|
|
continue;
|
|
|
|
PHINode *OrigPhi = Pos->second;
|
|
|
|
// Replacing the congruent phi is sufficient because acyclic redundancy
|
|
|
|
// elimination, CSE/GVN, should handle the rest. However, once SCEV proves
|
|
|
|
// that a phi is congruent, it's almost certain to be the head of an IV
|
|
|
|
// user cycle that is isomorphic with the original phi. So it's worth
|
|
|
|
// eagerly cleaning up the common case of a single IV increment.
|
|
|
|
if (BasicBlock *LatchBlock = L->getLoopLatch()) {
|
|
|
|
Instruction *OrigInc =
|
|
|
|
cast<Instruction>(OrigPhi->getIncomingValueForBlock(LatchBlock));
|
|
|
|
Instruction *IsomorphicInc =
|
|
|
|
cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock));
|
|
|
|
if (OrigInc != IsomorphicInc &&
|
|
|
|
SE->getSCEV(OrigInc) == SE->getSCEV(IsomorphicInc) &&
|
|
|
|
HoistStep(OrigInc, IsomorphicInc, DT)) {
|
|
|
|
DEBUG(dbgs() << "INDVARS: Eliminated congruent iv.inc: "
|
|
|
|
<< *IsomorphicInc << '\n');
|
|
|
|
IsomorphicInc->replaceAllUsesWith(OrigInc);
|
|
|
|
DeadInsts.push_back(IsomorphicInc);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
DEBUG(dbgs() << "INDVARS: Eliminated congruent iv: " << *Phi << '\n');
|
|
|
|
++NumElimIV;
|
|
|
|
Phi->replaceAllUsesWith(OrigPhi);
|
|
|
|
DeadInsts.push_back(Phi);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2011-07-18 18:21:35 +00:00
|
|
|
// Check for expressions that ScalarEvolution generates to compute
|
|
|
|
// BackedgeTakenInfo. If these expressions have not been reduced, then expanding
|
|
|
|
// them may incur additional cost (albeit in the loop preheader).
|
|
|
|
static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
|
|
|
|
ScalarEvolution *SE) {
|
|
|
|
// If the backedge-taken count is a UDiv, it's very likely a UDiv that
|
|
|
|
// ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
|
|
|
|
// precise expression, rather than a UDiv from the user's code. If we can't
|
|
|
|
// find a UDiv in the code with some simple searching, assume the former and
|
|
|
|
// forego rewriting the loop.
|
|
|
|
if (isa<SCEVUDivExpr>(S)) {
|
|
|
|
ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
|
|
|
|
if (!OrigCond) return true;
|
|
|
|
const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
|
|
|
|
R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
|
|
|
|
if (R != S) {
|
|
|
|
const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
|
|
|
|
L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
|
|
|
|
if (L != S)
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!DisableIVRewrite)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
// Recurse past add expressions, which commonly occur in the
|
|
|
|
// BackedgeTakenCount. They may already exist in program code, and if not,
|
|
|
|
// they are not too expensive rematerialize.
|
|
|
|
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
|
|
|
|
for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
|
|
|
|
I != E; ++I) {
|
|
|
|
if (isHighCostExpansion(*I, BI, SE))
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// HowManyLessThans uses a Max expression whenever the loop is not guarded by
|
|
|
|
// the exit condition.
|
|
|
|
if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
// If we haven't recognized an expensive SCEV patter, assume its an expression
|
|
|
|
// produced by program code.
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
|
|
|
|
/// count expression can be safely and cheaply expanded into an instruction
|
|
|
|
/// sequence that can be used by LinearFunctionTestReplace.
|
|
|
|
static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
|
|
|
|
const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
|
|
|
|
if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
|
|
|
|
BackedgeTakenCount->isZero())
|
2010-06-18 01:35:11 +00:00
|
|
|
return false;
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
if (!L->getExitingBlock())
|
|
|
|
return false;
|
2011-05-04 02:10:13 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Can't rewrite non-branch yet.
|
|
|
|
BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
|
|
|
|
if (!BI)
|
|
|
|
return false;
|
2009-02-17 20:49:49 +00:00
|
|
|
|
2011-07-18 18:21:35 +00:00
|
|
|
if (isHighCostExpansion(BackedgeTakenCount, BI, SE))
|
|
|
|
return false;
|
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
return true;
|
|
|
|
}
|
2009-02-17 20:49:49 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
/// getBackedgeIVType - Get the widest type used by the loop test after peeking
|
|
|
|
/// through Truncs.
|
|
|
|
///
|
|
|
|
/// TODO: Unnecessary if LFTR does not force a canonical IV.
|
2011-07-18 04:54:35 +00:00
|
|
|
static Type *getBackedgeIVType(Loop *L) {
|
2011-07-12 00:08:50 +00:00
|
|
|
if (!L->getExitingBlock())
|
|
|
|
return 0;
|
2007-03-04 01:00:28 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
// Can't rewrite non-branch yet.
|
|
|
|
BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
|
|
|
|
if (!BI)
|
|
|
|
return 0;
|
2011-06-27 23:17:44 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
ICmpInst *Cond = dyn_cast<ICmpInst>(BI->getCondition());
|
|
|
|
if (!Cond)
|
|
|
|
return 0;
|
2011-05-04 02:10:13 +00:00
|
|
|
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *Ty = 0;
|
2011-07-12 00:08:50 +00:00
|
|
|
for(User::op_iterator OI = Cond->op_begin(), OE = Cond->op_end();
|
|
|
|
OI != OE; ++OI) {
|
|
|
|
assert((!Ty || Ty == (*OI)->getType()) && "bad icmp operand types");
|
|
|
|
TruncInst *Trunc = dyn_cast<TruncInst>(*OI);
|
|
|
|
if (!Trunc)
|
|
|
|
continue;
|
2004-04-02 20:24:31 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
return Trunc->getSrcTy();
|
|
|
|
}
|
|
|
|
return Ty;
|
|
|
|
}
|
2010-04-13 01:46:36 +00:00
|
|
|
|
2011-07-12 00:08:50 +00:00
|
|
|
/// LinearFunctionTestReplace - This method rewrites the exit condition of the
|
|
|
|
/// loop to be a canonical != comparison against the incremented loop induction
|
|
|
|
/// variable. This pass is able to rewrite the exit tests of any loop where the
|
|
|
|
/// SCEV analysis can determine a loop-invariant trip count of the loop, which
|
|
|
|
/// is actually a much broader range than just linear tests.
|
|
|
|
ICmpInst *IndVarSimplify::
|
|
|
|
LinearFunctionTestReplace(Loop *L,
|
|
|
|
const SCEV *BackedgeTakenCount,
|
|
|
|
PHINode *IndVar,
|
|
|
|
SCEVExpander &Rewriter) {
|
|
|
|
assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
|
|
|
|
BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
|
|
|
|
|
|
|
|
// If the exiting block is not the same as the backedge block, we must compare
|
|
|
|
// against the preincremented value, otherwise we prefer to compare against
|
|
|
|
// the post-incremented value.
|
|
|
|
Value *CmpIndVar;
|
|
|
|
const SCEV *RHS = BackedgeTakenCount;
|
|
|
|
if (L->getExitingBlock() == L->getLoopLatch()) {
|
|
|
|
// Add one to the "backedge-taken" count to get the trip count.
|
|
|
|
// If this addition may overflow, we have to be more pessimistic and
|
|
|
|
// cast the induction variable before doing the add.
|
|
|
|
const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
|
|
|
|
const SCEV *N =
|
|
|
|
SE->getAddExpr(BackedgeTakenCount,
|
|
|
|
SE->getConstant(BackedgeTakenCount->getType(), 1));
|
|
|
|
if ((isa<SCEVConstant>(N) && !N->isZero()) ||
|
|
|
|
SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
|
|
|
|
// No overflow. Cast the sum.
|
|
|
|
RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
|
|
|
|
} else {
|
|
|
|
// Potential overflow. Cast before doing the add.
|
|
|
|
RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
|
|
|
|
IndVar->getType());
|
|
|
|
RHS = SE->getAddExpr(RHS,
|
|
|
|
SE->getConstant(IndVar->getType(), 1));
|
|
|
|
}
|
|
|
|
|
|
|
|
// The BackedgeTaken expression contains the number of times that the
|
|
|
|
// backedge branches to the loop header. This is one less than the
|
|
|
|
// number of times the loop executes, so use the incremented indvar.
|
|
|
|
CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
|
|
|
|
} else {
|
|
|
|
// We have to use the preincremented value...
|
|
|
|
RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
|
|
|
|
IndVar->getType());
|
|
|
|
CmpIndVar = IndVar;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Expand the code for the iteration count.
|
|
|
|
assert(SE->isLoopInvariant(RHS, L) &&
|
|
|
|
"Computed iteration count is not loop invariant!");
|
|
|
|
Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
|
|
|
|
|
|
|
|
// Insert a new icmp_ne or icmp_eq instruction before the branch.
|
|
|
|
ICmpInst::Predicate Opcode;
|
|
|
|
if (L->contains(BI->getSuccessor(0)))
|
|
|
|
Opcode = ICmpInst::ICMP_NE;
|
|
|
|
else
|
|
|
|
Opcode = ICmpInst::ICMP_EQ;
|
|
|
|
|
|
|
|
DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
|
|
|
|
<< " LHS:" << *CmpIndVar << '\n'
|
|
|
|
<< " op:\t"
|
|
|
|
<< (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
|
|
|
|
<< " RHS:\t" << *RHS << "\n");
|
|
|
|
|
|
|
|
ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
|
|
|
|
Cond->setDebugLoc(BI->getDebugLoc());
|
|
|
|
Value *OrigCond = BI->getCondition();
|
|
|
|
// It's tempting to use replaceAllUsesWith here to fully replace the old
|
|
|
|
// comparison, but that's not immediately safe, since users of the old
|
|
|
|
// comparison may not be dominated by the new comparison. Instead, just
|
|
|
|
// update the branch to use the new comparison; in the common case this
|
|
|
|
// will make old comparison dead.
|
|
|
|
BI->setCondition(Cond);
|
|
|
|
DeadInsts.push_back(OrigCond);
|
|
|
|
|
|
|
|
++NumLFTR;
|
|
|
|
Changed = true;
|
|
|
|
return Cond;
|
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
/// If there's a single exit block, sink any loop-invariant values that
|
|
|
|
/// were defined in the preheader but not used inside the loop into the
|
|
|
|
/// exit block to reduce register pressure in the loop.
|
|
|
|
void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
|
|
|
|
BasicBlock *ExitBlock = L->getExitBlock();
|
|
|
|
if (!ExitBlock) return;
|
|
|
|
|
|
|
|
BasicBlock *Preheader = L->getLoopPreheader();
|
|
|
|
if (!Preheader) return;
|
|
|
|
|
|
|
|
Instruction *InsertPt = ExitBlock->getFirstNonPHI();
|
|
|
|
BasicBlock::iterator I = Preheader->getTerminator();
|
|
|
|
while (I != Preheader->begin()) {
|
|
|
|
--I;
|
|
|
|
// New instructions were inserted at the end of the preheader.
|
|
|
|
if (isa<PHINode>(I))
|
|
|
|
break;
|
|
|
|
|
|
|
|
// Don't move instructions which might have side effects, since the side
|
|
|
|
// effects need to complete before instructions inside the loop. Also don't
|
|
|
|
// move instructions which might read memory, since the loop may modify
|
|
|
|
// memory. Note that it's okay if the instruction might have undefined
|
|
|
|
// behavior: LoopSimplify guarantees that the preheader dominates the exit
|
|
|
|
// block.
|
|
|
|
if (I->mayHaveSideEffects() || I->mayReadFromMemory())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Skip debug info intrinsics.
|
|
|
|
if (isa<DbgInfoIntrinsic>(I))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Don't sink static AllocaInsts out of the entry block, which would
|
|
|
|
// turn them into dynamic allocas!
|
|
|
|
if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
|
|
|
|
if (AI->isStaticAlloca())
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Determine if there is a use in or before the loop (direct or
|
|
|
|
// otherwise).
|
|
|
|
bool UsedInLoop = false;
|
|
|
|
for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
|
|
|
|
UI != UE; ++UI) {
|
|
|
|
User *U = *UI;
|
|
|
|
BasicBlock *UseBB = cast<Instruction>(U)->getParent();
|
|
|
|
if (PHINode *P = dyn_cast<PHINode>(U)) {
|
|
|
|
unsigned i =
|
|
|
|
PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
|
|
|
|
UseBB = P->getIncomingBlock(i);
|
|
|
|
}
|
|
|
|
if (UseBB == Preheader || L->contains(UseBB)) {
|
|
|
|
UsedInLoop = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// If there is, the def must remain in the preheader.
|
|
|
|
if (UsedInLoop)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
// Otherwise, sink it to the exit block.
|
|
|
|
Instruction *ToMove = I;
|
|
|
|
bool Done = false;
|
|
|
|
|
|
|
|
if (I != Preheader->begin()) {
|
|
|
|
// Skip debug info intrinsics.
|
|
|
|
do {
|
|
|
|
--I;
|
|
|
|
} while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
|
|
|
|
|
|
|
|
if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
|
|
|
|
Done = true;
|
|
|
|
} else {
|
|
|
|
Done = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
ToMove->moveBefore(InsertPt);
|
|
|
|
if (Done) break;
|
|
|
|
InsertPt = ToMove;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// IndVarSimplify driver. Manage several subpasses of IV simplification.
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
|
|
|
|
// If LoopSimplify form is not available, stay out of trouble. Some notes:
|
|
|
|
// - LSR currently only supports LoopSimplify-form loops. Indvars'
|
|
|
|
// canonicalization can be a pessimization without LSR to "clean up"
|
|
|
|
// afterwards.
|
|
|
|
// - We depend on having a preheader; in particular,
|
|
|
|
// Loop::getCanonicalInductionVariable only supports loops with preheaders,
|
|
|
|
// and we're in trouble if we can't find the induction variable even when
|
|
|
|
// we've manually inserted one.
|
|
|
|
if (!L->isLoopSimplifyForm())
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (!DisableIVRewrite)
|
|
|
|
IU = &getAnalysis<IVUsers>();
|
|
|
|
LI = &getAnalysis<LoopInfo>();
|
|
|
|
SE = &getAnalysis<ScalarEvolution>();
|
|
|
|
DT = &getAnalysis<DominatorTree>();
|
|
|
|
TD = getAnalysisIfAvailable<TargetData>();
|
|
|
|
|
|
|
|
DeadInsts.clear();
|
|
|
|
Changed = false;
|
|
|
|
|
|
|
|
// If there are any floating-point recurrences, attempt to
|
|
|
|
// transform them to use integer recurrences.
|
|
|
|
RewriteNonIntegerIVs(L);
|
|
|
|
|
|
|
|
const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
|
|
|
|
|
|
|
|
// Create a rewriter object which we'll use to transform the code with.
|
|
|
|
SCEVExpander Rewriter(*SE, "indvars");
|
|
|
|
|
|
|
|
// Eliminate redundant IV users.
|
|
|
|
//
|
|
|
|
// Simplification works best when run before other consumers of SCEV. We
|
|
|
|
// attempt to avoid evaluating SCEVs for sign/zero extend operations until
|
|
|
|
// other expressions involving loop IVs have been evaluated. This helps SCEV
|
|
|
|
// set no-wrap flags before normalizing sign/zero extension.
|
|
|
|
if (DisableIVRewrite) {
|
|
|
|
Rewriter.disableCanonicalMode();
|
|
|
|
SimplifyIVUsersNoRewrite(L, Rewriter);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check to see if this loop has a computable loop-invariant execution count.
|
|
|
|
// If so, this means that we can compute the final value of any expressions
|
|
|
|
// that are recurrent in the loop, and substitute the exit values from the
|
|
|
|
// loop into any instructions outside of the loop that use the final values of
|
|
|
|
// the current expressions.
|
|
|
|
//
|
|
|
|
if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
|
|
|
|
RewriteLoopExitValues(L, Rewriter);
|
|
|
|
|
|
|
|
// Eliminate redundant IV users.
|
|
|
|
if (!DisableIVRewrite)
|
|
|
|
SimplifyIVUsers(Rewriter);
|
|
|
|
|
2011-07-16 01:06:48 +00:00
|
|
|
// Eliminate redundant IV cycles.
|
2011-07-06 20:50:43 +00:00
|
|
|
if (DisableIVRewrite)
|
|
|
|
SimplifyCongruentIVs(L);
|
|
|
|
|
2009-05-12 02:17:14 +00:00
|
|
|
// Compute the type of the largest recurrence expression, and decide whether
|
|
|
|
// a canonical induction variable should be inserted.
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *LargestType = 0;
|
2009-05-12 02:17:14 +00:00
|
|
|
bool NeedCannIV = false;
|
2011-05-25 04:42:22 +00:00
|
|
|
bool ExpandBECount = canExpandBackedgeTakenCount(L, SE);
|
2011-05-03 22:24:10 +00:00
|
|
|
if (ExpandBECount) {
|
2009-05-12 02:17:14 +00:00
|
|
|
// If we have a known trip count and a single exit block, we'll be
|
|
|
|
// rewriting the loop exit test condition below, which requires a
|
|
|
|
// canonical induction variable.
|
2011-05-03 22:24:10 +00:00
|
|
|
NeedCannIV = true;
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *Ty = BackedgeTakenCount->getType();
|
2011-05-25 04:42:22 +00:00
|
|
|
if (DisableIVRewrite) {
|
|
|
|
// In this mode, SimplifyIVUsers may have already widened the IV used by
|
|
|
|
// the backedge test and inserted a Trunc on the compare's operand. Get
|
|
|
|
// the wider type to avoid creating a redundant narrow IV only used by the
|
|
|
|
// loop test.
|
|
|
|
LargestType = getBackedgeIVType(L);
|
|
|
|
}
|
2011-05-03 22:24:10 +00:00
|
|
|
if (!LargestType ||
|
|
|
|
SE->getTypeSizeInBits(Ty) >
|
|
|
|
SE->getTypeSizeInBits(LargestType))
|
|
|
|
LargestType = SE->getEffectiveSCEVType(Ty);
|
2004-04-17 18:08:33 +00:00
|
|
|
}
|
2011-05-04 02:10:13 +00:00
|
|
|
if (!DisableIVRewrite) {
|
|
|
|
for (IVUsers::const_iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
|
|
|
|
NeedCannIV = true;
|
2011-07-18 04:54:35 +00:00
|
|
|
Type *Ty =
|
2011-05-04 02:10:13 +00:00
|
|
|
SE->getEffectiveSCEVType(I->getOperandValToReplace()->getType());
|
|
|
|
if (!LargestType ||
|
|
|
|
SE->getTypeSizeInBits(Ty) >
|
2009-04-21 01:07:12 +00:00
|
|
|
SE->getTypeSizeInBits(LargestType))
|
2011-05-04 02:10:13 +00:00
|
|
|
LargestType = Ty;
|
|
|
|
}
|
2003-12-22 09:53:29 +00:00
|
|
|
}
|
2001-12-04 04:32:29 +00:00
|
|
|
|
2010-02-10 16:03:48 +00:00
|
|
|
// Now that we know the largest of the induction variable expressions
|
2009-05-12 02:17:14 +00:00
|
|
|
// in this loop, insert a canonical induction variable of the largest size.
|
2010-07-20 17:18:52 +00:00
|
|
|
PHINode *IndVar = 0;
|
2009-05-12 02:17:14 +00:00
|
|
|
if (NeedCannIV) {
|
2010-02-25 06:57:05 +00:00
|
|
|
// Check to see if the loop already has any canonical-looking induction
|
|
|
|
// variables. If any are present and wider than the planned canonical
|
|
|
|
// induction variable, temporarily remove them, so that the Rewriter
|
|
|
|
// doesn't attempt to reuse them.
|
|
|
|
SmallVector<PHINode *, 2> OldCannIVs;
|
|
|
|
while (PHINode *OldCannIV = L->getCanonicalInductionVariable()) {
|
2009-06-13 16:25:49 +00:00
|
|
|
if (SE->getTypeSizeInBits(OldCannIV->getType()) >
|
|
|
|
SE->getTypeSizeInBits(LargestType))
|
|
|
|
OldCannIV->removeFromParent();
|
|
|
|
else
|
2010-02-25 06:57:05 +00:00
|
|
|
break;
|
|
|
|
OldCannIVs.push_back(OldCannIV);
|
2009-06-13 16:25:49 +00:00
|
|
|
}
|
|
|
|
|
2009-06-26 22:53:46 +00:00
|
|
|
IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
|
2009-06-13 16:25:49 +00:00
|
|
|
|
2009-02-12 22:19:27 +00:00
|
|
|
++NumInserted;
|
|
|
|
Changed = true;
|
2010-01-05 01:27:06 +00:00
|
|
|
DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
|
2009-06-13 16:25:49 +00:00
|
|
|
|
|
|
|
// Now that the official induction variable is established, reinsert
|
2010-02-25 06:57:05 +00:00
|
|
|
// any old canonical-looking variables after it so that the IR remains
|
|
|
|
// consistent. They will be deleted as part of the dead-PHI deletion at
|
2009-06-13 16:25:49 +00:00
|
|
|
// the end of the pass.
|
2010-02-25 06:57:05 +00:00
|
|
|
while (!OldCannIVs.empty()) {
|
|
|
|
PHINode *OldCannIV = OldCannIVs.pop_back_val();
|
|
|
|
OldCannIV->insertBefore(L->getHeader()->getFirstNonPHI());
|
|
|
|
}
|
2007-06-15 14:38:12 +00:00
|
|
|
}
|
2004-04-02 20:24:31 +00:00
|
|
|
|
2009-02-12 22:19:27 +00:00
|
|
|
// If we have a trip count expression, rewrite the loop's exit condition
|
|
|
|
// using it. We can currently only handle loops with a single exit.
|
2009-05-12 02:17:14 +00:00
|
|
|
ICmpInst *NewICmp = 0;
|
2011-05-03 22:24:10 +00:00
|
|
|
if (ExpandBECount) {
|
2011-05-25 04:42:22 +00:00
|
|
|
assert(canExpandBackedgeTakenCount(L, SE) &&
|
2011-05-03 22:24:10 +00:00
|
|
|
"canonical IV disrupted BackedgeTaken expansion");
|
2009-05-12 02:17:14 +00:00
|
|
|
assert(NeedCannIV &&
|
|
|
|
"LinearFunctionTestReplace requires a canonical induction variable");
|
2011-07-16 01:18:53 +00:00
|
|
|
// Check preconditions for proper SCEVExpander operation. SCEV does not
|
|
|
|
// express SCEVExpander's dependencies, such as LoopSimplify. Instead any
|
|
|
|
// pass that uses the SCEVExpander must do it. This does not work well for
|
|
|
|
// loop passes because SCEVExpander makes assumptions about all loops, while
|
|
|
|
// LoopPassManager only forces the current loop to be simplified.
|
|
|
|
//
|
|
|
|
// FIXME: SCEV expansion has no way to bail out, so the caller must
|
|
|
|
// explicitly check any assumptions made by SCEV. Brittle.
|
|
|
|
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
|
|
|
|
if (!AR || AR->getLoop()->getLoopPreheader())
|
|
|
|
NewICmp =
|
|
|
|
LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, Rewriter);
|
2009-05-12 02:17:14 +00:00
|
|
|
}
|
2011-03-17 23:51:11 +00:00
|
|
|
// Rewrite IV-derived expressions.
|
2011-05-04 02:10:13 +00:00
|
|
|
if (!DisableIVRewrite)
|
|
|
|
RewriteIVExpressions(L, Rewriter);
|
2004-04-22 14:59:40 +00:00
|
|
|
|
2011-03-17 23:51:11 +00:00
|
|
|
// Clear the rewriter cache, because values that are in the rewriter's cache
|
|
|
|
// can be deleted in the loop below, causing the AssertingVH in the cache to
|
|
|
|
// trigger.
|
|
|
|
Rewriter.clear();
|
|
|
|
|
|
|
|
// Now that we're done iterating through lists, clean up any instructions
|
|
|
|
// which are now dead.
|
|
|
|
while (!DeadInsts.empty())
|
|
|
|
if (Instruction *Inst =
|
|
|
|
dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
|
|
|
|
RecursivelyDeleteTriviallyDeadInstructions(Inst);
|
|
|
|
|
2009-06-26 22:53:46 +00:00
|
|
|
// The Rewriter may not be used from this point on.
|
2009-05-24 20:08:21 +00:00
|
|
|
|
2009-05-12 02:17:14 +00:00
|
|
|
// Loop-invariant instructions in the preheader that aren't used in the
|
|
|
|
// loop may be sunk below the loop to reduce register pressure.
|
2009-06-26 22:53:46 +00:00
|
|
|
SinkUnusedInvariants(L);
|
2009-05-12 02:17:14 +00:00
|
|
|
|
|
|
|
// For completeness, inform IVUsers of the IV use in the newly-created
|
|
|
|
// loop exit test instruction.
|
2011-06-21 03:22:38 +00:00
|
|
|
if (NewICmp && IU)
|
2011-06-21 15:43:52 +00:00
|
|
|
IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
|
2009-05-12 02:17:14 +00:00
|
|
|
|
|
|
|
// Clean up dead instructions.
|
2010-01-05 16:31:45 +00:00
|
|
|
Changed |= DeleteDeadPHIs(L->getHeader());
|
2009-05-12 02:17:14 +00:00
|
|
|
// Check a post-condition.
|
2010-03-10 19:38:49 +00:00
|
|
|
assert(L->isLCSSAForm(*DT) && "Indvars did not leave the loop in lcssa form!");
|
2009-05-12 02:17:14 +00:00
|
|
|
return Changed;
|
|
|
|
}
|