2005-07-30 00:12:19 +00:00
|
|
|
//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
|
|
|
// This file was developed by the LLVM research group and is distributed under
|
|
|
|
// the University of Illinois Open Source License. See LICENSE.TXT for details.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file contains the implementation of the scalar evolution expander,
|
|
|
|
// which is used to generate the code corresponding to a given scalar evolution
|
|
|
|
// expression.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "llvm/Analysis/ScalarEvolutionExpander.h"
|
2006-12-07 01:30:32 +00:00
|
|
|
#include "llvm/Analysis/LoopInfo.h"
|
2005-07-30 00:12:19 +00:00
|
|
|
using namespace llvm;
|
|
|
|
|
2006-02-04 09:51:53 +00:00
|
|
|
/// InsertCastOfTo - Insert a cast of V to the specified type, doing what
|
|
|
|
/// we can to share the casts.
|
2006-12-13 08:06:42 +00:00
|
|
|
Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V,
|
|
|
|
const Type *Ty) {
|
2006-02-04 09:51:53 +00:00
|
|
|
// FIXME: keep track of the cast instruction.
|
|
|
|
if (Constant *C = dyn_cast<Constant>(V))
|
2006-12-12 23:36:14 +00:00
|
|
|
return ConstantExpr::getCast(opcode, C, Ty);
|
2006-02-04 09:51:53 +00:00
|
|
|
|
|
|
|
if (Argument *A = dyn_cast<Argument>(V)) {
|
|
|
|
// Check to see if there is already a cast!
|
|
|
|
for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
|
|
|
|
UI != E; ++UI) {
|
|
|
|
if ((*UI)->getType() == Ty)
|
|
|
|
if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) {
|
2006-11-27 01:05:10 +00:00
|
|
|
// If the cast isn't the first instruction of the function, move it.
|
2006-02-04 09:51:53 +00:00
|
|
|
if (BasicBlock::iterator(CI) !=
|
|
|
|
A->getParent()->getEntryBlock().begin()) {
|
|
|
|
CI->moveBefore(A->getParent()->getEntryBlock().begin());
|
|
|
|
}
|
|
|
|
return CI;
|
|
|
|
}
|
|
|
|
}
|
2006-12-12 23:36:14 +00:00
|
|
|
return CastInst::create(opcode, V, Ty, V->getName(),
|
|
|
|
A->getParent()->getEntryBlock().begin());
|
2006-02-04 09:51:53 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
Instruction *I = cast<Instruction>(V);
|
|
|
|
|
|
|
|
// Check to see if there is already a cast. If there is, use it.
|
|
|
|
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
|
|
|
|
UI != E; ++UI) {
|
|
|
|
if ((*UI)->getType() == Ty)
|
|
|
|
if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) {
|
|
|
|
BasicBlock::iterator It = I; ++It;
|
|
|
|
if (isa<InvokeInst>(I))
|
|
|
|
It = cast<InvokeInst>(I)->getNormalDest()->begin();
|
|
|
|
while (isa<PHINode>(It)) ++It;
|
|
|
|
if (It != BasicBlock::iterator(CI)) {
|
|
|
|
// Splice the cast immediately after the operand in question.
|
|
|
|
CI->moveBefore(It);
|
|
|
|
}
|
|
|
|
return CI;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
BasicBlock::iterator IP = I; ++IP;
|
|
|
|
if (InvokeInst *II = dyn_cast<InvokeInst>(I))
|
|
|
|
IP = II->getNormalDest()->begin();
|
|
|
|
while (isa<PHINode>(IP)) ++IP;
|
2006-12-12 23:36:14 +00:00
|
|
|
return CastInst::create(opcode, V, Ty, V->getName(), IP);
|
2006-02-04 09:51:53 +00:00
|
|
|
}
|
|
|
|
|
2007-04-13 05:04:18 +00:00
|
|
|
/// InsertBinop - Insert the specified binary operator, doing a small amount
|
|
|
|
/// of work to avoid inserting an obviously redundant operation.
|
|
|
|
Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS,
|
2007-04-17 23:43:50 +00:00
|
|
|
Value *RHS, Instruction *&InsertPt) {
|
2007-04-13 05:04:18 +00:00
|
|
|
// Do a quick scan to see if we have this binop nearby. If so, reuse it.
|
|
|
|
unsigned ScanLimit = 6;
|
|
|
|
for (BasicBlock::iterator IP = InsertPt, E = InsertPt->getParent()->begin();
|
|
|
|
ScanLimit; --IP, --ScanLimit) {
|
|
|
|
if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(IP))
|
|
|
|
if (BinOp->getOpcode() == Opcode && BinOp->getOperand(0) == LHS &&
|
2007-04-17 23:43:50 +00:00
|
|
|
BinOp->getOperand(1) == RHS) {
|
|
|
|
// If we found the instruction *at* the insert point, insert later
|
|
|
|
// instructions after it.
|
|
|
|
if (BinOp == InsertPt)
|
|
|
|
InsertPt = ++IP;
|
2007-04-13 05:04:18 +00:00
|
|
|
return BinOp;
|
2007-04-17 23:43:50 +00:00
|
|
|
}
|
2007-04-13 05:04:18 +00:00
|
|
|
if (IP == E) break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// If we don't have
|
|
|
|
return BinaryOperator::create(Opcode, LHS, RHS, "tmp.", InsertPt);
|
|
|
|
}
|
|
|
|
|
2005-07-30 00:12:19 +00:00
|
|
|
Value *SCEVExpander::visitMulExpr(SCEVMulExpr *S) {
|
|
|
|
const Type *Ty = S->getType();
|
|
|
|
int FirstOp = 0; // Set if we should emit a subtract.
|
|
|
|
if (SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
|
|
|
|
if (SC->getValue()->isAllOnesValue())
|
|
|
|
FirstOp = 1;
|
|
|
|
|
|
|
|
int i = S->getNumOperands()-2;
|
|
|
|
Value *V = expandInTy(S->getOperand(i+1), Ty);
|
|
|
|
|
|
|
|
// Emit a bunch of multiply instructions
|
|
|
|
for (; i >= FirstOp; --i)
|
2007-04-13 05:04:18 +00:00
|
|
|
V = InsertBinop(Instruction::Mul, V, expandInTy(S->getOperand(i), Ty),
|
|
|
|
InsertPt);
|
2005-07-30 00:12:19 +00:00
|
|
|
// -1 * ... ---> 0 - ...
|
|
|
|
if (FirstOp == 1)
|
2007-04-13 05:04:18 +00:00
|
|
|
V = InsertBinop(Instruction::Sub, Constant::getNullValue(V->getType()), V,
|
|
|
|
InsertPt);
|
2005-07-30 00:12:19 +00:00
|
|
|
return V;
|
|
|
|
}
|
|
|
|
|
|
|
|
Value *SCEVExpander::visitAddRecExpr(SCEVAddRecExpr *S) {
|
|
|
|
const Type *Ty = S->getType();
|
|
|
|
const Loop *L = S->getLoop();
|
|
|
|
// We cannot yet do fp recurrences, e.g. the xform of {X,+,F} --> X+{0,+,F}
|
2007-01-15 02:27:26 +00:00
|
|
|
assert(Ty->isInteger() && "Cannot expand fp recurrences yet!");
|
2005-07-30 00:12:19 +00:00
|
|
|
|
|
|
|
// {X,+,F} --> X + {0,+,F}
|
|
|
|
if (!isa<SCEVConstant>(S->getStart()) ||
|
2007-03-02 00:28:52 +00:00
|
|
|
!cast<SCEVConstant>(S->getStart())->getValue()->isZero()) {
|
2005-07-30 00:12:19 +00:00
|
|
|
Value *Start = expandInTy(S->getStart(), Ty);
|
|
|
|
std::vector<SCEVHandle> NewOps(S->op_begin(), S->op_end());
|
|
|
|
NewOps[0] = SCEVUnknown::getIntegerSCEV(0, Ty);
|
|
|
|
Value *Rest = expandInTy(SCEVAddRecExpr::get(NewOps, L), Ty);
|
|
|
|
|
|
|
|
// FIXME: look for an existing add to use.
|
2007-04-13 05:04:18 +00:00
|
|
|
return InsertBinop(Instruction::Add, Rest, Start, InsertPt);
|
2005-07-30 00:12:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// {0,+,1} --> Insert a canonical induction variable into the loop!
|
|
|
|
if (S->getNumOperands() == 2 &&
|
|
|
|
S->getOperand(1) == SCEVUnknown::getIntegerSCEV(1, Ty)) {
|
|
|
|
// Create and insert the PHI node for the induction variable in the
|
|
|
|
// specified loop.
|
|
|
|
BasicBlock *Header = L->getHeader();
|
|
|
|
PHINode *PN = new PHINode(Ty, "indvar", Header->begin());
|
|
|
|
PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
|
|
|
|
|
|
|
|
pred_iterator HPI = pred_begin(Header);
|
|
|
|
assert(HPI != pred_end(Header) && "Loop with zero preds???");
|
|
|
|
if (!L->contains(*HPI)) ++HPI;
|
|
|
|
assert(HPI != pred_end(Header) && L->contains(*HPI) &&
|
|
|
|
"No backedge in loop?");
|
|
|
|
|
|
|
|
// Insert a unit add instruction right before the terminator corresponding
|
|
|
|
// to the back-edge.
|
2007-01-21 00:29:26 +00:00
|
|
|
Constant *One = ConstantInt::get(Ty, 1);
|
2005-07-30 00:12:19 +00:00
|
|
|
Instruction *Add = BinaryOperator::createAdd(PN, One, "indvar.next",
|
|
|
|
(*HPI)->getTerminator());
|
|
|
|
|
|
|
|
pred_iterator PI = pred_begin(Header);
|
|
|
|
if (*PI == L->getLoopPreheader())
|
|
|
|
++PI;
|
|
|
|
PN->addIncoming(Add, *PI);
|
|
|
|
return PN;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Get the canonical induction variable I for this loop.
|
|
|
|
Value *I = getOrInsertCanonicalInductionVariable(L, Ty);
|
|
|
|
|
Fix a problem that Nate noticed with LSR:
When inserting code for an addrec expression with a non-unit stride, be
more careful where we insert the multiply. In particular, insert the multiply
in the outermost loop we can, instead of the requested insertion point.
This allows LSR to notice the mul in the right loop, reducing it when it gets
to it. This allows it to reduce the multiply, where before it missed it.
This happens quite a bit in the test suite, for example, eliminating 2
multiplies in art, 3 in ammp, 4 in apsi, reducing from 1050 multiplies to
910 muls in galgel (!), from 877 to 859 in applu, and 36 to 30 in bzip2.
This speeds up galgel from 16.45s to 16.01s, applu from 14.21 to 13.94s and
fourinarow from 66.67s to 63.48s.
This implements Transforms/LoopStrengthReduce/nested-reduce.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24102 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-30 06:24:33 +00:00
|
|
|
// If this is a simple linear addrec, emit it now as a special case.
|
2005-07-30 00:12:19 +00:00
|
|
|
if (S->getNumOperands() == 2) { // {0,+,F} --> i*F
|
|
|
|
Value *F = expandInTy(S->getOperand(1), Ty);
|
Fix a problem that Nate noticed with LSR:
When inserting code for an addrec expression with a non-unit stride, be
more careful where we insert the multiply. In particular, insert the multiply
in the outermost loop we can, instead of the requested insertion point.
This allows LSR to notice the mul in the right loop, reducing it when it gets
to it. This allows it to reduce the multiply, where before it missed it.
This happens quite a bit in the test suite, for example, eliminating 2
multiplies in art, 3 in ammp, 4 in apsi, reducing from 1050 multiplies to
910 muls in galgel (!), from 877 to 859 in applu, and 36 to 30 in bzip2.
This speeds up galgel from 16.45s to 16.01s, applu from 14.21 to 13.94s and
fourinarow from 66.67s to 63.48s.
This implements Transforms/LoopStrengthReduce/nested-reduce.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24102 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-30 06:24:33 +00:00
|
|
|
|
|
|
|
// IF the step is by one, just return the inserted IV.
|
2007-01-11 12:24:14 +00:00
|
|
|
if (ConstantInt *CI = dyn_cast<ConstantInt>(F))
|
2007-03-01 19:45:00 +00:00
|
|
|
if (CI->getValue() == 1)
|
Fix a problem that Nate noticed with LSR:
When inserting code for an addrec expression with a non-unit stride, be
more careful where we insert the multiply. In particular, insert the multiply
in the outermost loop we can, instead of the requested insertion point.
This allows LSR to notice the mul in the right loop, reducing it when it gets
to it. This allows it to reduce the multiply, where before it missed it.
This happens quite a bit in the test suite, for example, eliminating 2
multiplies in art, 3 in ammp, 4 in apsi, reducing from 1050 multiplies to
910 muls in galgel (!), from 877 to 859 in applu, and 36 to 30 in bzip2.
This speeds up galgel from 16.45s to 16.01s, applu from 14.21 to 13.94s and
fourinarow from 66.67s to 63.48s.
This implements Transforms/LoopStrengthReduce/nested-reduce.ll
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24102 91177308-0d34-0410-b5e6-96231b3b80d8
2005-10-30 06:24:33 +00:00
|
|
|
return I;
|
|
|
|
|
|
|
|
// If the insert point is directly inside of the loop, emit the multiply at
|
|
|
|
// the insert point. Otherwise, L is a loop that is a parent of the insert
|
|
|
|
// point loop. If we can, move the multiply to the outer most loop that it
|
|
|
|
// is safe to be in.
|
|
|
|
Instruction *MulInsertPt = InsertPt;
|
|
|
|
Loop *InsertPtLoop = LI.getLoopFor(MulInsertPt->getParent());
|
|
|
|
if (InsertPtLoop != L && InsertPtLoop &&
|
|
|
|
L->contains(InsertPtLoop->getHeader())) {
|
|
|
|
while (InsertPtLoop != L) {
|
|
|
|
// If we cannot hoist the multiply out of this loop, don't.
|
|
|
|
if (!InsertPtLoop->isLoopInvariant(F)) break;
|
|
|
|
|
|
|
|
// Otherwise, move the insert point to the preheader of the loop.
|
|
|
|
MulInsertPt = InsertPtLoop->getLoopPreheader()->getTerminator();
|
|
|
|
InsertPtLoop = InsertPtLoop->getParentLoop();
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-04-13 05:04:18 +00:00
|
|
|
return InsertBinop(Instruction::Mul, I, F, MulInsertPt);
|
2005-07-30 00:12:19 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// If this is a chain of recurrences, turn it into a closed form, using the
|
|
|
|
// folders, then expandCodeFor the closed form. This allows the folders to
|
|
|
|
// simplify the expression without having to build a bunch of special code
|
|
|
|
// into this folder.
|
|
|
|
SCEVHandle IH = SCEVUnknown::get(I); // Get I as a "symbolic" SCEV.
|
|
|
|
|
|
|
|
SCEVHandle V = S->evaluateAtIteration(IH);
|
2006-12-07 01:30:32 +00:00
|
|
|
//cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
|
2005-07-30 00:12:19 +00:00
|
|
|
|
|
|
|
return expandInTy(V, Ty);
|
|
|
|
}
|