2001-12-03 17:28:42 +00:00
|
|
|
//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
|
2003-10-20 19:43:21 +00:00
|
|
|
//
|
|
|
|
// 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.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
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.
|
|
|
|
//
|
|
|
|
// This transformation make the following changes to each loop with an
|
|
|
|
// 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.
|
|
|
|
// 3. Any pointer arithmetic recurrences are raised to use array subscripts.
|
|
|
|
//
|
|
|
|
// 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
|
|
|
|
// desired loop transformations have been performed. Additionally, on targets
|
|
|
|
// where it is profitable, the loop could be transformed to count down to zero
|
|
|
|
// (the "do loop" optimization).
|
2001-12-03 17:28:42 +00:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
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"
|
2004-04-02 20:24:31 +00:00
|
|
|
#include "llvm/Type.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/Analysis/ScalarEvolutionExpressions.h"
|
2003-12-18 17:19:19 +00:00
|
|
|
#include "llvm/Analysis/LoopInfo.h"
|
2002-02-12 22:39:50 +00:00
|
|
|
#include "llvm/Support/CFG.h"
|
2003-12-18 17:19:19 +00:00
|
|
|
#include "llvm/Transforms/Utils/Local.h"
|
2004-04-02 20:24:31 +00:00
|
|
|
#include "Support/CommandLine.h"
|
2002-10-01 22:38:41 +00:00
|
|
|
#include "Support/Statistic.h"
|
2003-12-18 17:19:19 +00:00
|
|
|
using namespace llvm;
|
2003-11-11 22:41:34 +00:00
|
|
|
|
2002-09-10 05:24:05 +00:00
|
|
|
namespace {
|
2002-10-01 22:38:41 +00:00
|
|
|
Statistic<> NumRemoved ("indvars", "Number of aux indvars removed");
|
2004-04-02 20:24:31 +00:00
|
|
|
Statistic<> NumPointer ("indvars", "Number of pointer indvars promoted");
|
2003-09-10 05:24:46 +00:00
|
|
|
Statistic<> NumInserted("indvars", "Number of canonical indvars added");
|
2004-04-02 20:24:31 +00:00
|
|
|
Statistic<> NumReplaced("indvars", "Number of exit values replaced");
|
|
|
|
Statistic<> NumLFTR ("indvars", "Number of loop exit tests replaced");
|
2003-12-22 03:58:44 +00:00
|
|
|
|
|
|
|
class IndVarSimplify : public FunctionPass {
|
2004-04-02 20:24:31 +00:00
|
|
|
LoopInfo *LI;
|
|
|
|
ScalarEvolution *SE;
|
2003-12-23 07:47:09 +00:00
|
|
|
bool Changed;
|
2003-12-22 03:58:44 +00:00
|
|
|
public:
|
|
|
|
virtual bool runOnFunction(Function &) {
|
2004-04-02 20:24:31 +00:00
|
|
|
LI = &getAnalysis<LoopInfo>();
|
|
|
|
SE = &getAnalysis<ScalarEvolution>();
|
2003-12-23 07:47:09 +00:00
|
|
|
Changed = false;
|
|
|
|
|
2003-12-22 03:58:44 +00:00
|
|
|
// Induction Variables live in the header nodes of loops
|
2004-04-02 20:24:31 +00:00
|
|
|
for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
|
2004-01-08 00:09:44 +00:00
|
|
|
runOnLoop(*I);
|
2003-12-22 03:58:44 +00:00
|
|
|
return Changed;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
|
|
|
|
AU.addRequiredID(LoopSimplifyID);
|
2004-04-02 20:24:31 +00:00
|
|
|
AU.addRequired<ScalarEvolution>();
|
|
|
|
AU.addRequired<LoopInfo>();
|
2003-12-22 03:58:44 +00:00
|
|
|
AU.addPreservedID(LoopSimplifyID);
|
|
|
|
AU.setPreservesCFG();
|
|
|
|
}
|
2004-04-02 20:24:31 +00:00
|
|
|
private:
|
|
|
|
void runOnLoop(Loop *L);
|
|
|
|
void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
|
|
|
|
std::set<Instruction*> &DeadInsts);
|
|
|
|
void LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
|
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
|
|
|
ScalarEvolutionRewriter &RW);
|
2004-04-02 20:24:31 +00:00
|
|
|
void RewriteLoopExitValues(Loop *L);
|
|
|
|
|
|
|
|
void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
|
2003-12-22 03:58:44 +00:00
|
|
|
};
|
|
|
|
RegisterOpt<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
|
2002-09-10 05:24:05 +00:00
|
|
|
}
|
2001-12-04 04:32:29 +00:00
|
|
|
|
2003-12-22 03:58:44 +00:00
|
|
|
Pass *llvm::createIndVarSimplifyPass() {
|
|
|
|
return new IndVarSimplify();
|
2001-12-04 04:32:29 +00:00
|
|
|
}
|
|
|
|
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
/// DeleteTriviallyDeadInstructions - If any of the instructions is the
|
|
|
|
/// specified set are trivially dead, delete them and see if this makes any of
|
|
|
|
/// their operands subsequently dead.
|
|
|
|
void IndVarSimplify::
|
|
|
|
DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
|
|
|
|
while (!Insts.empty()) {
|
|
|
|
Instruction *I = *Insts.begin();
|
|
|
|
Insts.erase(Insts.begin());
|
|
|
|
if (isInstructionTriviallyDead(I)) {
|
|
|
|
for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
|
|
|
|
if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
|
|
|
|
Insts.insert(U);
|
|
|
|
SE->deleteInstructionFromRecords(I);
|
|
|
|
I->getParent()->getInstList().erase(I);
|
|
|
|
Changed = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2001-12-03 17:28:42 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
/// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
|
|
|
|
/// recurrence. If so, change it into an integer recurrence, permitting
|
|
|
|
/// analysis by the SCEV routines.
|
|
|
|
void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
|
|
|
|
BasicBlock *Preheader,
|
|
|
|
std::set<Instruction*> &DeadInsts) {
|
|
|
|
assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
|
|
|
|
unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
|
|
|
|
unsigned BackedgeIdx = PreheaderIdx^1;
|
|
|
|
if (GetElementPtrInst *GEPI =
|
|
|
|
dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
|
|
|
|
if (GEPI->getOperand(0) == PN) {
|
|
|
|
assert(GEPI->getNumOperands() == 2 && "GEP types must mismatch!");
|
|
|
|
|
|
|
|
// Okay, we found a pointer recurrence. Transform this pointer
|
|
|
|
// recurrence into an integer recurrence. Compute the value that gets
|
|
|
|
// added to the pointer at every iteration.
|
|
|
|
Value *AddedVal = GEPI->getOperand(1);
|
|
|
|
|
|
|
|
// Insert a new integer PHI node into the top of the block.
|
|
|
|
PHINode *NewPhi = new PHINode(AddedVal->getType(),
|
|
|
|
PN->getName()+".rec", PN);
|
|
|
|
NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()),
|
|
|
|
Preheader);
|
|
|
|
// Create the new add instruction.
|
|
|
|
Value *NewAdd = BinaryOperator::create(Instruction::Add, NewPhi,
|
|
|
|
AddedVal,
|
|
|
|
GEPI->getName()+".rec", GEPI);
|
|
|
|
NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
|
|
|
|
|
|
|
|
// Update the existing GEP to use the recurrence.
|
|
|
|
GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
|
|
|
|
|
|
|
|
// Update the GEP to use the new recurrence we just inserted.
|
|
|
|
GEPI->setOperand(1, NewAdd);
|
|
|
|
|
|
|
|
// Finally, if there are any other users of the PHI node, we must
|
|
|
|
// insert a new GEP instruction that uses the pre-incremented version
|
|
|
|
// of the induction amount.
|
|
|
|
if (!PN->use_empty()) {
|
|
|
|
BasicBlock::iterator InsertPos = PN; ++InsertPos;
|
|
|
|
while (isa<PHINode>(InsertPos)) ++InsertPos;
|
|
|
|
std::string Name = PN->getName(); PN->setName("");
|
|
|
|
Value *PreInc =
|
|
|
|
new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
|
|
|
|
std::vector<Value*>(1, NewPhi), Name,
|
|
|
|
InsertPos);
|
|
|
|
PN->replaceAllUsesWith(PreInc);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Delete the old PHI for sure, and the GEP if its otherwise unused.
|
|
|
|
DeadInsts.insert(PN);
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
++NumPointer;
|
|
|
|
Changed = true;
|
|
|
|
}
|
|
|
|
}
|
2003-12-22 03:58:44 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
/// LinearFunctionTestReplace - This method rewrites the exit condition of the
|
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
|
|
|
/// 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.
|
2004-04-02 20:24:31 +00:00
|
|
|
void IndVarSimplify::LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
|
|
|
|
ScalarEvolutionRewriter &RW) {
|
|
|
|
// Find the exit block for the loop. We can currently only handle loops with
|
|
|
|
// a single exit.
|
2004-04-18 22:14:10 +00:00
|
|
|
std::vector<BasicBlock*> ExitBlocks;
|
|
|
|
L->getExitBlocks(ExitBlocks);
|
|
|
|
if (ExitBlocks.size() != 1) return;
|
|
|
|
BasicBlock *ExitBlock = ExitBlocks[0];
|
2004-04-02 20:24:31 +00:00
|
|
|
|
|
|
|
// Make sure there is only one predecessor block in the loop.
|
|
|
|
BasicBlock *ExitingBlock = 0;
|
|
|
|
for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
|
|
|
|
PI != PE; ++PI)
|
|
|
|
if (L->contains(*PI)) {
|
|
|
|
if (ExitingBlock == 0)
|
|
|
|
ExitingBlock = *PI;
|
|
|
|
else
|
|
|
|
return; // Multiple exits from loop to this block.
|
|
|
|
}
|
|
|
|
assert(ExitingBlock && "Loop info is broken");
|
|
|
|
|
|
|
|
if (!isa<BranchInst>(ExitingBlock->getTerminator()))
|
|
|
|
return; // Can't rewrite non-branch yet
|
|
|
|
BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
|
|
|
|
assert(BI->isConditional() && "Must be conditional to be part of loop!");
|
|
|
|
|
|
|
|
std::set<Instruction*> InstructionsToDelete;
|
|
|
|
if (Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()))
|
|
|
|
InstructionsToDelete.insert(Cond);
|
|
|
|
|
2004-04-15 20:26:22 +00:00
|
|
|
// 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.
|
|
|
|
BasicBlock *Header = L->getHeader();
|
|
|
|
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?");
|
|
|
|
|
|
|
|
SCEVHandle TripCount = IterationCount;
|
|
|
|
Value *IndVar;
|
|
|
|
if (*HPI == ExitingBlock) {
|
|
|
|
// The IterationCount expression contains the number of times that the
|
|
|
|
// backedge actually branches to the loop header. This is one less than the
|
|
|
|
// number of times the loop executes, so add one to it.
|
|
|
|
Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
|
|
|
|
TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
|
|
|
|
IndVar = L->getCanonicalInductionVariableIncrement();
|
|
|
|
} else {
|
|
|
|
// We have to use the preincremented value...
|
|
|
|
IndVar = L->getCanonicalInductionVariable();
|
|
|
|
}
|
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
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// Expand the code for the iteration count into the preheader of the loop.
|
|
|
|
BasicBlock *Preheader = L->getLoopPreheader();
|
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
|
|
|
Value *ExitCnt = RW.ExpandCodeFor(TripCount, Preheader->getTerminator(),
|
2004-04-02 20:24:31 +00:00
|
|
|
IndVar->getType());
|
|
|
|
|
|
|
|
// Insert a new setne or seteq instruction before the branch.
|
|
|
|
Instruction::BinaryOps Opcode;
|
|
|
|
if (L->contains(BI->getSuccessor(0)))
|
|
|
|
Opcode = Instruction::SetNE;
|
|
|
|
else
|
|
|
|
Opcode = Instruction::SetEQ;
|
|
|
|
|
|
|
|
Value *Cond = new SetCondInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
|
|
|
|
BI->setCondition(Cond);
|
|
|
|
++NumLFTR;
|
|
|
|
Changed = true;
|
|
|
|
|
|
|
|
DeleteTriviallyDeadInstructions(InstructionsToDelete);
|
|
|
|
}
|
2003-12-22 03:58:44 +00:00
|
|
|
|
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
/// 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.
|
|
|
|
void IndVarSimplify::RewriteLoopExitValues(Loop *L) {
|
|
|
|
BasicBlock *Preheader = L->getLoopPreheader();
|
|
|
|
|
|
|
|
// Scan all of the instructions in the loop, looking at those that have
|
|
|
|
// extra-loop users and which are recurrences.
|
|
|
|
ScalarEvolutionRewriter Rewriter(*SE, *LI);
|
|
|
|
|
|
|
|
// We insert the code into the preheader of the loop if the loop contains
|
|
|
|
// multiple exit blocks, or in the exit block if there is exactly one.
|
|
|
|
BasicBlock *BlockToInsertInto;
|
2004-04-18 22:14:10 +00:00
|
|
|
std::vector<BasicBlock*> ExitBlocks;
|
|
|
|
L->getExitBlocks(ExitBlocks);
|
|
|
|
if (ExitBlocks.size() == 1)
|
|
|
|
BlockToInsertInto = ExitBlocks[0];
|
2004-04-02 20:24:31 +00:00
|
|
|
else
|
|
|
|
BlockToInsertInto = Preheader;
|
|
|
|
BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
|
|
|
|
while (isa<PHINode>(InsertPt)) ++InsertPt;
|
|
|
|
|
2004-04-17 18:44:09 +00:00
|
|
|
bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
|
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
std::set<Instruction*> InstructionsToDelete;
|
|
|
|
|
|
|
|
for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
|
|
|
|
if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
|
|
|
|
BasicBlock *BB = L->getBlocks()[i];
|
|
|
|
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
|
|
|
|
if (I->getType()->isInteger()) { // Is an integer instruction
|
|
|
|
SCEVHandle SH = SE->getSCEV(I);
|
2004-04-17 18:44:09 +00:00
|
|
|
if (SH->hasComputableLoopEvolution(L) || // Varies predictably
|
|
|
|
HasConstantItCount) {
|
2004-04-02 20:24:31 +00:00
|
|
|
// Find out if this predictably varying value is actually used
|
|
|
|
// outside of the loop. "extra" as opposed to "intra".
|
|
|
|
std::vector<User*> ExtraLoopUsers;
|
|
|
|
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
|
|
|
|
UI != E; ++UI)
|
|
|
|
if (!L->contains(cast<Instruction>(*UI)->getParent()))
|
|
|
|
ExtraLoopUsers.push_back(*UI);
|
|
|
|
if (!ExtraLoopUsers.empty()) {
|
|
|
|
// Okay, this instruction has a user outside of the current loop
|
|
|
|
// and varies predictably in this loop. Evaluate the value it
|
|
|
|
// contains when the loop exits, and insert code for it.
|
2004-04-17 18:44:09 +00:00
|
|
|
SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
|
2004-04-02 20:24:31 +00:00
|
|
|
if (!isa<SCEVCouldNotCompute>(ExitValue)) {
|
|
|
|
Changed = true;
|
|
|
|
++NumReplaced;
|
|
|
|
Value *NewVal = Rewriter.ExpandCodeFor(ExitValue, InsertPt,
|
|
|
|
I->getType());
|
|
|
|
|
|
|
|
// Rewrite any users of the computed value outside of the loop
|
|
|
|
// with the newly computed value.
|
|
|
|
for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i)
|
|
|
|
ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
|
|
|
|
|
|
|
|
// If this instruction is dead now, schedule it to be removed.
|
|
|
|
if (I->use_empty())
|
|
|
|
InstructionsToDelete.insert(I);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2003-12-22 03:58:44 +00:00
|
|
|
}
|
2001-12-03 17:28:42 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
DeleteTriviallyDeadInstructions(InstructionsToDelete);
|
|
|
|
}
|
2003-12-23 07:47:09 +00:00
|
|
|
|
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
void IndVarSimplify::runOnLoop(Loop *L) {
|
|
|
|
// First step. Check to see if there are any trivial GEP pointer recurrences.
|
|
|
|
// If there are, change them into integer recurrences, permitting analysis by
|
|
|
|
// the SCEV routines.
|
2003-12-23 07:47:09 +00:00
|
|
|
//
|
2004-04-02 20:24:31 +00:00
|
|
|
BasicBlock *Header = L->getHeader();
|
|
|
|
BasicBlock *Preheader = L->getLoopPreheader();
|
|
|
|
|
|
|
|
std::set<Instruction*> DeadInsts;
|
|
|
|
for (BasicBlock::iterator I = Header->begin();
|
|
|
|
PHINode *PN = dyn_cast<PHINode>(I); ++I)
|
|
|
|
if (isa<PointerType>(PN->getType()))
|
|
|
|
EliminatePointerRecurrence(PN, Preheader, DeadInsts);
|
2003-12-23 07:47:09 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
if (!DeadInsts.empty())
|
|
|
|
DeleteTriviallyDeadInstructions(DeadInsts);
|
2001-12-04 04:32:29 +00:00
|
|
|
|
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// Next, transform all loops nesting inside of this loop.
|
|
|
|
for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
|
|
|
|
runOnLoop(*I);
|
2001-12-04 04:32:29 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// 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.
|
2001-12-04 04:32:29 +00:00
|
|
|
//
|
2004-04-02 20:24:31 +00:00
|
|
|
SCEVHandle IterationCount = SE->getIterationCount(L);
|
|
|
|
if (!isa<SCEVCouldNotCompute>(IterationCount))
|
|
|
|
RewriteLoopExitValues(L);
|
|
|
|
|
|
|
|
// Next, analyze all of the induction variables in the loop, canonicalizing
|
|
|
|
// auxillary induction variables.
|
|
|
|
std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
|
|
|
|
|
|
|
|
for (BasicBlock::iterator I = Header->begin();
|
|
|
|
PHINode *PN = dyn_cast<PHINode>(I); ++I)
|
|
|
|
if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
|
|
|
|
SCEVHandle SCEV = SE->getSCEV(PN);
|
|
|
|
if (SCEV->hasComputableLoopEvolution(L))
|
|
|
|
if (SE->shouldSubstituteIndVar(SCEV)) // HACK!
|
|
|
|
IndVars.push_back(std::make_pair(PN, SCEV));
|
|
|
|
}
|
2002-05-22 17:17:27 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// If there are no induction variables in the loop, there is nothing more to
|
|
|
|
// do.
|
2004-04-17 18:08:33 +00:00
|
|
|
if (IndVars.empty()) {
|
|
|
|
// Actually, if we know how many times the loop iterates, lets insert a
|
|
|
|
// canonical induction variable to help subsequent passes.
|
|
|
|
if (!isa<SCEVCouldNotCompute>(IterationCount)) {
|
|
|
|
ScalarEvolutionRewriter Rewriter(*SE, *LI);
|
|
|
|
Rewriter.GetOrInsertCanonicalInductionVariable(L,
|
|
|
|
IterationCount->getType());
|
|
|
|
LinearFunctionTestReplace(L, IterationCount, Rewriter);
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
2002-05-22 17:17:27 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// Compute the type of the largest recurrence expression.
|
|
|
|
//
|
|
|
|
const Type *LargestType = IndVars[0].first->getType();
|
|
|
|
for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
|
|
|
|
const Type *Ty = IndVars[i].first->getType();
|
|
|
|
if (Ty->getPrimitiveSize() > LargestType->getPrimitiveSize())
|
|
|
|
LargestType = Ty;
|
2003-12-22 09:53:29 +00:00
|
|
|
}
|
2001-12-04 04:32:29 +00:00
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// Create a rewriter object which we'll use to transform the code with.
|
|
|
|
ScalarEvolutionRewriter Rewriter(*SE, *LI);
|
|
|
|
|
|
|
|
// Now that we know the largest of of the induction variables in this loop,
|
|
|
|
// insert a canonical induction variable of the largest size.
|
2004-04-16 06:03:17 +00:00
|
|
|
LargestType = LargestType->getUnsignedVersion();
|
2004-04-02 20:24:31 +00:00
|
|
|
Value *IndVar = Rewriter.GetOrInsertCanonicalInductionVariable(L,LargestType);
|
|
|
|
++NumInserted;
|
|
|
|
Changed = true;
|
|
|
|
|
|
|
|
if (!isa<SCEVCouldNotCompute>(IterationCount))
|
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
|
|
|
LinearFunctionTestReplace(L, IterationCount, Rewriter);
|
2004-04-02 20:24:31 +00:00
|
|
|
|
|
|
|
// Now that we have a canonical induction variable, we can rewrite any
|
|
|
|
// recurrences in terms of the induction variable. Start with the auxillary
|
|
|
|
// induction variables, and recursively rewrite any of their uses.
|
|
|
|
BasicBlock::iterator InsertPt = Header->begin();
|
|
|
|
while (isa<PHINode>(InsertPt)) ++InsertPt;
|
|
|
|
|
Implement a fixme. The helps loops that have induction variables of different
types in them. Instead of creating an induction variable for all types, it
creates a single induction variable and casts to the other sizes. This generates
this code:
no_exit: ; preds = %entry, %no_exit
%indvar = phi uint [ %indvar.next, %no_exit ], [ 0, %entry ] ; <uint> [#uses=4]
*** %j.0.0 = cast uint %indvar to short ; <short> [#uses=1]
%indvar = cast uint %indvar to int ; <int> [#uses=1]
%tmp.7 = getelementptr short* %P, uint %indvar ; <short*> [#uses=1]
store short %j.0.0, short* %tmp.7
%inc.0 = add int %indvar, 1 ; <int> [#uses=2]
%tmp.2 = setlt int %inc.0, %N ; <bool> [#uses=1]
%indvar.next = add uint %indvar, 1 ; <uint> [#uses=1]
br bool %tmp.2, label %no_exit, label %loopexit
instead of:
no_exit: ; preds = %entry, %no_exit
%indvar = phi ushort [ %indvar.next, %no_exit ], [ 0, %entry ] ; <ushort> [#uses=2]
*** %indvar = phi uint [ %indvar.next, %no_exit ], [ 0, %entry ] ; <uint> [#uses=3]
%indvar = cast uint %indvar to int ; <int> [#uses=1]
%indvar = cast ushort %indvar to short ; <short> [#uses=1]
%tmp.7 = getelementptr short* %P, uint %indvar ; <short*> [#uses=1]
store short %indvar, short* %tmp.7
%inc.0 = add int %indvar, 1 ; <int> [#uses=2]
%tmp.2 = setlt int %inc.0, %N ; <bool> [#uses=1]
%indvar.next = add uint %indvar, 1
*** %indvar.next = add ushort %indvar, 1
br bool %tmp.2, label %no_exit, label %loopexit
This is an improvement in register pressure, but probably doesn't happen that
often.
The more important fix will be to get rid of the redundant add.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@13101 91177308-0d34-0410-b5e6-96231b3b80d8
2004-04-21 22:22:01 +00:00
|
|
|
// If there were induction variables of other sizes, cast the primary
|
|
|
|
// induction variable to the right size for them, avoiding the need for the
|
|
|
|
// code evaluation methods to insert induction variables of different sizes.
|
|
|
|
std::map<unsigned, Value*> InsertedSizes;
|
|
|
|
InsertedSizes[LargestType->getPrimitiveSize()] = IndVar;
|
2004-04-02 20:24:31 +00:00
|
|
|
while (!IndVars.empty()) {
|
|
|
|
PHINode *PN = IndVars.back().first;
|
Implement a fixme. The helps loops that have induction variables of different
types in them. Instead of creating an induction variable for all types, it
creates a single induction variable and casts to the other sizes. This generates
this code:
no_exit: ; preds = %entry, %no_exit
%indvar = phi uint [ %indvar.next, %no_exit ], [ 0, %entry ] ; <uint> [#uses=4]
*** %j.0.0 = cast uint %indvar to short ; <short> [#uses=1]
%indvar = cast uint %indvar to int ; <int> [#uses=1]
%tmp.7 = getelementptr short* %P, uint %indvar ; <short*> [#uses=1]
store short %j.0.0, short* %tmp.7
%inc.0 = add int %indvar, 1 ; <int> [#uses=2]
%tmp.2 = setlt int %inc.0, %N ; <bool> [#uses=1]
%indvar.next = add uint %indvar, 1 ; <uint> [#uses=1]
br bool %tmp.2, label %no_exit, label %loopexit
instead of:
no_exit: ; preds = %entry, %no_exit
%indvar = phi ushort [ %indvar.next, %no_exit ], [ 0, %entry ] ; <ushort> [#uses=2]
*** %indvar = phi uint [ %indvar.next, %no_exit ], [ 0, %entry ] ; <uint> [#uses=3]
%indvar = cast uint %indvar to int ; <int> [#uses=1]
%indvar = cast ushort %indvar to short ; <short> [#uses=1]
%tmp.7 = getelementptr short* %P, uint %indvar ; <short*> [#uses=1]
store short %indvar, short* %tmp.7
%inc.0 = add int %indvar, 1 ; <int> [#uses=2]
%tmp.2 = setlt int %inc.0, %N ; <bool> [#uses=1]
%indvar.next = add uint %indvar, 1
*** %indvar.next = add ushort %indvar, 1
br bool %tmp.2, label %no_exit, label %loopexit
This is an improvement in register pressure, but probably doesn't happen that
often.
The more important fix will be to get rid of the redundant add.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@13101 91177308-0d34-0410-b5e6-96231b3b80d8
2004-04-21 22:22:01 +00:00
|
|
|
|
|
|
|
const Type *Ty = PN->getType()->getUnsignedVersion();
|
|
|
|
Value *&IV = InsertedSizes[Ty->getPrimitiveSize()];
|
|
|
|
if (IV == 0) {
|
|
|
|
// Insert a new cast instruction, which will hold this recurrence.
|
|
|
|
std::string Name = PN->getName();
|
|
|
|
PN->setName("");
|
|
|
|
IV = new CastInst(IndVar, Ty, Name, InsertPt);
|
|
|
|
}
|
|
|
|
|
|
|
|
Value *V = IV;
|
|
|
|
if (PN->getType() != Ty)
|
|
|
|
V = new CastInst(V, PN->getType(), V->getName(), InsertPt);
|
|
|
|
|
2004-04-02 20:24:31 +00:00
|
|
|
// Replace the old PHI Node with the inserted computation.
|
Implement a fixme. The helps loops that have induction variables of different
types in them. Instead of creating an induction variable for all types, it
creates a single induction variable and casts to the other sizes. This generates
this code:
no_exit: ; preds = %entry, %no_exit
%indvar = phi uint [ %indvar.next, %no_exit ], [ 0, %entry ] ; <uint> [#uses=4]
*** %j.0.0 = cast uint %indvar to short ; <short> [#uses=1]
%indvar = cast uint %indvar to int ; <int> [#uses=1]
%tmp.7 = getelementptr short* %P, uint %indvar ; <short*> [#uses=1]
store short %j.0.0, short* %tmp.7
%inc.0 = add int %indvar, 1 ; <int> [#uses=2]
%tmp.2 = setlt int %inc.0, %N ; <bool> [#uses=1]
%indvar.next = add uint %indvar, 1 ; <uint> [#uses=1]
br bool %tmp.2, label %no_exit, label %loopexit
instead of:
no_exit: ; preds = %entry, %no_exit
%indvar = phi ushort [ %indvar.next, %no_exit ], [ 0, %entry ] ; <ushort> [#uses=2]
*** %indvar = phi uint [ %indvar.next, %no_exit ], [ 0, %entry ] ; <uint> [#uses=3]
%indvar = cast uint %indvar to int ; <int> [#uses=1]
%indvar = cast ushort %indvar to short ; <short> [#uses=1]
%tmp.7 = getelementptr short* %P, uint %indvar ; <short*> [#uses=1]
store short %indvar, short* %tmp.7
%inc.0 = add int %indvar, 1 ; <int> [#uses=2]
%tmp.2 = setlt int %inc.0, %N ; <bool> [#uses=1]
%indvar.next = add uint %indvar, 1
*** %indvar.next = add ushort %indvar, 1
br bool %tmp.2, label %no_exit, label %loopexit
This is an improvement in register pressure, but probably doesn't happen that
often.
The more important fix will be to get rid of the redundant add.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@13101 91177308-0d34-0410-b5e6-96231b3b80d8
2004-04-21 22:22:01 +00:00
|
|
|
PN->replaceAllUsesWith(V);
|
2004-04-02 20:24:31 +00:00
|
|
|
DeadInsts.insert(PN);
|
|
|
|
IndVars.pop_back();
|
|
|
|
++NumRemoved;
|
|
|
|
Changed = true;
|
2003-12-22 09:53:29 +00:00
|
|
|
}
|
2003-12-10 18:06:47 +00:00
|
|
|
|
2004-04-21 23:36:08 +00:00
|
|
|
// Now replace all derived expressions in the loop body with simpler
|
|
|
|
// expressions.
|
2004-04-02 20:24:31 +00:00
|
|
|
for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
|
|
|
|
if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
|
|
|
|
BasicBlock *BB = L->getBlocks()[i];
|
|
|
|
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
|
|
|
|
if (I->getType()->isInteger() && // Is an integer instruction
|
2004-04-21 23:36:08 +00:00
|
|
|
!I->use_empty() &&
|
2004-04-02 20:24:31 +00:00
|
|
|
!Rewriter.isInsertedInstruction(I)) {
|
|
|
|
SCEVHandle SH = SE->getSCEV(I);
|
2004-04-21 23:36:08 +00:00
|
|
|
Value *V = Rewriter.ExpandCodeFor(SH, I, I->getType());
|
|
|
|
if (V != I) {
|
|
|
|
if (isa<Instruction>(V)) {
|
|
|
|
std::string Name = I->getName();
|
|
|
|
I->setName("");
|
|
|
|
V->setName(Name);
|
|
|
|
}
|
|
|
|
I->replaceAllUsesWith(V);
|
|
|
|
DeadInsts.insert(I);
|
|
|
|
++NumRemoved;
|
|
|
|
Changed = true;
|
|
|
|
}
|
2004-04-02 20:24:31 +00:00
|
|
|
}
|
2001-12-04 04:32:29 +00:00
|
|
|
}
|
2004-04-21 23:36:08 +00:00
|
|
|
|
|
|
|
|
|
|
|
DeleteTriviallyDeadInstructions(DeadInsts);
|
2001-12-03 17:28:42 +00:00
|
|
|
}
|