================
@@ -2168,6 +2170,26 @@ void LoopInterchangeTransform::restructureLoops(
SE->forgetLoop(NewOuter);
}
+/// Update the metadata attached to the latch branches of the two loops. The
+/// loop metadata and the branch weights are dropped. Some attributes of the
+/// loop metadata may not hold after interchange (e.g.,
+/// llvm.loop.parallel_accesses of the inner loop doesn't imply that the loop
is
+/// parallel when it becomes the outer loop), and no followup attributes are
+/// defined for interchange.
+/// TODO: Preserve the branch weights by swapping them.
+void LoopInterchangeTransform::updateLatchMetadata() {
+ auto *OuterLoopLatchBI =
+ cast<CondBrInst>(OuterLoop->getLoopLatch()->getTerminator());
+ auto *InnerLoopLatchBI =
+ cast<CondBrInst>(InnerLoop->getLoopLatch()->getTerminator());
+
+ for (CondBrInst *BI : {OuterLoopLatchBI, InnerLoopLatchBI}) {
+ BI->setMetadata(LLVMContext::MD_loop, nullptr);
----------------
kasuga-fj wrote:
I agree that keeping the attributes known to be safe makes sense, but I'd
prefer to do it as a follow-up after #235896 and keep this patch as simple as
possible. In #235896, each latch ends up controlling the other loop's
iterations, so the attributes we keep would also need to be moved to the other
latch. If I understand correctly, interchange isn't triggered very often at the
moment, and interchanging loops with user hints (e.g., `unroll(disable)`)
should be even rarer, so I think we can ignore the impact for now. As for
`mustprogress`, I think dropping it has little impact, since the trip counts of
interchanged loops are always computable.
https://github.com/llvm/llvm-project/pull/235892
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits