summaryrefslogtreecommitdiff
path: root/lib/Transforms/Scalar
diff options
context:
space:
mode:
authorAndrew Trick <atrick@apple.com>2013-05-06 17:58:18 +0000
committerAndrew Trick <atrick@apple.com>2013-05-06 17:58:18 +0000
commitfcf79528da4da87996a936b9e86f669c1f0c1dc6 (patch)
tree23e21bd7e588fda564340acc57f34f00fd015d27 /lib/Transforms/Scalar
parent32c76107d029c1cad5935d08cdcde6139cf874bb (diff)
downloadllvm-fcf79528da4da87996a936b9e86f669c1f0c1dc6.tar.gz
llvm-fcf79528da4da87996a936b9e86f669c1f0c1dc6.tar.bz2
llvm-fcf79528da4da87996a936b9e86f669c1f0c1dc6.tar.xz
Rotate multi-exit loops even if the latch was simplified.
Test case by Michele Scandale! Fixes PR10293: Load not hoisted out of loop with multiple exits. There are few regressions with this patch, now tracked by rdar:13817079, and a roughly equal number of improvements. The regressions are almost certainly back luck because LoopRotate has very little idea of whether rotation is profitable. Doing better requires a more comprehensive solution. This checkin is a quick fix that lacks generality (PR10293 has a counter-example). But it trivially fixes the case in PR10293 without interfering with other cases, and it does satify the criteria that LoopRotate is a loop canonicalization pass that should avoid heuristics and special cases. I can think of two approaches that would probably be better in the long run. Ultimately they may both make sense. (1) LoopRotate should check that the current header would make a good loop guard, and that the loop does not already has a sufficient guard. The artifical SimplifiedLoopLatch check would be unnecessary, and the design would be more general and canonical. Two difficulties: - We need a strong guarantee that we won't endlessly rotate, so the analysis would need to be precise in order to avoid the SimplifiedLoopLatch precondition. - Analysis like this are usually based on SCEV, which we don't want to rely on. (2) Rotate on-demand in late loop passes. This could even be done by shoving the loop back on the queue after the optimization that needs it. This could work well when we find LICM opportunities in multi-branch loops. This requires some work, and it doesn't really solve the problem of SCEV wanting a loop guard before the analysis. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@181230 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Transforms/Scalar')
-rw-r--r--lib/Transforms/Scalar/LoopRotation.cpp43
1 files changed, 29 insertions, 14 deletions
diff --git a/lib/Transforms/Scalar/LoopRotation.cpp b/lib/Transforms/Scalar/LoopRotation.cpp
index e98ae953e5..14c5655f08 100644
--- a/lib/Transforms/Scalar/LoopRotation.cpp
+++ b/lib/Transforms/Scalar/LoopRotation.cpp
@@ -56,8 +56,8 @@ namespace {
}
bool runOnLoop(Loop *L, LPPassManager &LPM);
- void simplifyLoopLatch(Loop *L);
- bool rotateLoop(Loop *L);
+ bool simplifyLoopLatch(Loop *L);
+ bool rotateLoop(Loop *L, bool SimplifiedLatch);
private:
LoopInfo *LI;
@@ -84,13 +84,14 @@ bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) {
// Simplify the loop latch before attempting to rotate the header
// upward. Rotation may not be needed if the loop tail can be folded into the
// loop exit.
- simplifyLoopLatch(L);
+ bool SimplifiedLatch = simplifyLoopLatch(L);
// One loop can be rotated multiple times.
bool MadeChange = false;
- while (rotateLoop(L))
+ while (rotateLoop(L, SimplifiedLatch)) {
MadeChange = true;
-
+ SimplifiedLatch = false;
+ }
return MadeChange;
}
@@ -212,25 +213,25 @@ static bool shouldSpeculateInstrs(BasicBlock::iterator Begin,
/// canonical form so downstream passes can handle it.
///
/// I don't believe this invalidates SCEV.
-void LoopRotate::simplifyLoopLatch(Loop *L) {
+bool LoopRotate::simplifyLoopLatch(Loop *L) {
BasicBlock *Latch = L->getLoopLatch();
if (!Latch || Latch->hasAddressTaken())
- return;
+ return false;
BranchInst *Jmp = dyn_cast<BranchInst>(Latch->getTerminator());
if (!Jmp || !Jmp->isUnconditional())
- return;
+ return false;
BasicBlock *LastExit = Latch->getSinglePredecessor();
if (!LastExit || !L->isLoopExiting(LastExit))
- return;
+ return false;
BranchInst *BI = dyn_cast<BranchInst>(LastExit->getTerminator());
if (!BI)
- return;
+ return false;
if (!shouldSpeculateInstrs(Latch->begin(), Jmp))
- return;
+ return false;
DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into "
<< LastExit->getName() << "\n");
@@ -253,10 +254,20 @@ void LoopRotate::simplifyLoopLatch(Loop *L) {
if (DominatorTree *DT = getAnalysisIfAvailable<DominatorTree>())
DT->eraseNode(Latch);
Latch->eraseFromParent();
+ return true;
}
/// Rotate loop LP. Return true if the loop is rotated.
-bool LoopRotate::rotateLoop(Loop *L) {
+///
+/// \param SimplifiedLatch is true if the latch was just folded into the final
+/// loop exit. In this case we may want to rotate even though the new latch is
+/// now an exiting branch. This rotation would have happened had the latch not
+/// been simplified. However, if SimplifiedLatch is false, then we avoid
+/// rotating loops in which the latch exits to avoid excessive or endless
+/// rotation. LoopRotate should be repeatable and converge to a canonical
+/// form. This property is satisfied because simplifying the loop latch can only
+/// happen once across multiple invocations of the LoopRotate pass.
+bool LoopRotate::rotateLoop(Loop *L, bool SimplifiedLatch) {
// If the loop has only one block then there is not much to rotate.
if (L->getBlocks().size() == 1)
return false;
@@ -276,7 +287,12 @@ bool LoopRotate::rotateLoop(Loop *L) {
// If the loop latch already contains a branch that leaves the loop then the
// loop is already rotated.
- if (OrigLatch == 0 || L->isLoopExiting(OrigLatch))
+ if (OrigLatch == 0)
+ return false;
+
+ // Rotate if either the loop latch does *not* exit the loop, or if the loop
+ // latch was just simplified.
+ if (L->isLoopExiting(OrigLatch) && !SimplifiedLatch)
return false;
// Check size of original header and reject loop if it is very big or we can't
@@ -505,4 +521,3 @@ bool LoopRotate::rotateLoop(Loop *L) {
++NumRotated;
return true;
}
-