Allow loop-idiom to run on multiple BB loops, but still only scan the loop
header for now for memset/memcpy opportunities. It turns out that loop-rotate
is successfully rotating loops, but *DOESN'T MERGE THE BLOCKS*, turning "for
loops" into 2 basic block loops that loop-idiom was ignoring.
With this fix, we form many *many* more memcpy and memsets than before, including
on the "history" loops in the viterbi benchmark, which look like this:
for (j=0; j<MAX_history; ++j) {
history_new[i][j+1] = history[2*i][j];
}
Transforming these loops into memcpy's speeds up the viterbi benchmark from
11.98s to 3.55s on my machine. Woo.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122685 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Transforms/Scalar/LoopIdiomRecognize.cpp b/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
index e9394cd..84e33f0 100644
--- a/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
+++ b/lib/Transforms/Scalar/LoopIdiomRecognize.cpp
@@ -126,11 +126,6 @@
bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
CurLoop = L;
- // We only look at trivial single basic block loops.
- // TODO: eventually support more complex loops, scanning the header.
- if (L->getBlocks().size() != 1)
- return false;
-
// The trip count of the loop must be analyzable.
SE = &getAnalysis<ScalarEvolution>();
if (!SE->hasLoopInvariantBackedgeTakenCount(L))
@@ -142,6 +137,11 @@
TD = getAnalysisIfAvailable<TargetData>();
if (TD == 0) return false;
+ // TODO: We currently only scan the header of the loop, because it is the only
+ // part that is known to execute and we don't want to make a conditional store
+ // into an unconditional one in the preheader. However, there can be diamonds
+ // and other things in the loop that would make other blocks "always executed"
+ // we should get the full set and scan each block.
BasicBlock *BB = L->getHeader();
DEBUG(dbgs() << "loop-idiom Scanning: F[" << BB->getParent()->getName()
<< "] Loop %" << BB->getName() << "\n");