[Mlir-commits] [mlir] [MLIR][Mem2Reg] Improve performance by avoiding recomputations (PR #91444)

[Tobias Gysi]

================
@@ -605,20 +627,24 @@ void MemorySlotPromoter::promoteSlot() {
 
 LogicalResult mlir::tryToPromoteMemorySlots(
     ArrayRef<PromotableAllocationOpInterface> allocators, OpBuilder &builder,
-    const DataLayout &dataLayout, Mem2RegStatistics statistics) {
+    const DataLayout &dataLayout, DominanceInfo &dominance,
+    Mem2RegStatistics statistics) {
   bool promotedAny = false;
 
+  // Cache for block index maps. This is required to avoid expensive
+  // recomputations.
+  BlockIndexCache blockIndexCache;
+
   for (PromotableAllocationOpInterface allocator : allocators) {
----------------
gysit wrote:

This sounds like a reasonable follow up step to me. The advantage is that the block indices could be cached longer and the possible disadvantage is that `tryToPromoteMemorySlots` always promotes all allocas of the region (which today we do anyways)? Are there other trade-offs?

https://github.com/llvm/llvm-project/pull/91444