[llvm] [LiveDebugVariables] Stop holding SlotIndexes for erased instructions (PR #224467)
Petar Jovanovic via llvm-commits
llvm-commits at lists.llvm.org
Sun Sep 27 18:25:04 PDT 2026
================
@@ -1549,6 +1567,120 @@ splitRegister(Register OldReg, ArrayRef<Register> NewRegs, LiveIntervals &LIS) {
PImpl->splitRegister(OldReg, NewRegs);
}
+//===----------------------------------------------------------------------===//
+// Stale Index Canonicalization
+//===----------------------------------------------------------------------===//
+
+void UserValue::canonicalizeIndexes(const SlotIndexes &SI) {
+ unsigned NumStale = 0;
+ for (LocMap::const_iterator I = locInts.begin(); I.valid(); ++I)
+ NumStale += SI.isStaleIndex(I.start()) + SI.isStaleIndex(I.stop());
+ for (SlotIndex Idx : trimmedDefs)
+ NumStale += SI.isStaleIndex(Idx);
+ NumStaleIndexes += NumStale;
+
+ if (NumStale) {
+ // trimmedDefs is looked up by interval start. Remapping it here is safe:
+ // trimmed starts are block slots, so the Stop < Start case below cannot
+ // reach them, and a merge drops a start that then matches nothing.
+ if (!trimmedDefs.empty()) {
+ SmallVector<SlotIndex, 2> Defs(trimmedDefs.begin(), trimmedDefs.end());
----------------
petar-jovanovic wrote:
> Potentially expensive if this allocates -- can they not be updated in place?
It should not allocate: SmallSet<T, 2> is backed by a SmallVector<T, 2>, so the temporary has the same inline capacity. In-place updates are not really an option either, since this is a set keyed by the value being rewritten and mutating elements would violate its invariants. It is also rarely reached: trimmedDefs is only populated for inlined variables trimmed to a lexical scope.
https://github.com/llvm/llvm-project/pull/224467
More information about the llvm-commits
mailing list