[flang-commits] [flang] [flang] Preserve the alignment attribute across stack/heap conversions (PR #234898)

[Eugene Epshteyn via flang-commits]

================
@@ -215,9 +215,15 @@ void AllocationPlacementPass::runOnOperation() {
             ? placementHook(info, basePolicy, stackBytesUsed)
             : fir::decideAllocationPlacement(info, basePolicy, stackBytesUsed);
 
-    // Account for the decision in the running stack budget.
-    if (endsUpOnStack(placement, info.isCurrentlyOnStack) && info.byteSize)
+    // Account for the decision in the running stack budget, including the
+    // worst-case padding a requested alignment may need.
+    if (endsUpOnStack(placement, info.isCurrentlyOnStack) && info.byteSize) {
       stackBytesUsed += static_cast<std::size_t>(*info.byteSize);
+      std::optional<std::int64_t> alignment =
+          alloca ? alloca.getAlignment() : allocmem.getAlignment();
+      if (alignment && *alignment > 1)
+        stackBytesUsed += static_cast<std::size_t>(*alignment - 1);
----------------
eugeneepshteyn wrote:

Indeed: as written the first aligned temporary always slipped in under the limit. Done: the requested alignment is now part of `PendingAllocationInfo`, and `stackBytesNeeded()` (size plus worst-case padding) is what `shouldAllocateOnStack` compares against the limit and what the pass charges afterwards, so the two agree.


https://github.com/llvm/llvm-project/pull/234898