[flang-commits] [flang] [flang] Support declarations scoping in FIR AA. (PR #201216)

[via flang-commits]

================
@@ -286,8 +288,74 @@ static fir::AliasAnalysis::Source mergeRegionBranchPredecessorSources(
 
   mlir::Type mergedTy = allTypesSame ? sources[0].valueType : fallbackType;
 
-  return {mergedOrigin, mergedKind, mergedTy,      mergedAttrs,
-          mergedApprox, mergedPath, mergedCaptured};
+  // Intersect scopedOrigins across predecessors by (scope, declValue).
+  // For an entry that matches in every predecessor, take the bitwise
+  // union (|=) of its 'attributes' and 'approximateSource' bits and
+  // keep the first predecessor's path steps (the steps must be equal
+  // to match, so there is nothing to merge there). Drop the entry on a
+  // path-step or isData mismatch, because different control-flow shapes
+  // from the same declare to the merge point cannot be summarised
+  // safely.
+  llvm::SmallVector<fir::AliasAnalysis::Source::ScopedOrigin, 4>
+      mergedScopedOrigins;
+  if (!sources.empty()) {
+    // Seed with the first predecessor's snapshots, keyed by
+    // (scope, declValue) for intersection lookups.
+    llvm::DenseMap<std::pair<void *, void *>, unsigned> indexInMerged;
+    mergedScopedOrigins.assign(sources[0].scopedOrigins.begin(),
+                               sources[0].scopedOrigins.end());
+    for (unsigned i = 0; i < mergedScopedOrigins.size(); ++i) {
+      const auto &scopedOrigin = mergedScopedOrigins[i];
+      indexInMerged[{scopedOrigin.scope.getAsOpaquePointer(),
----------------
jeanPerier wrote:

Is it really needed to use the scope as a key given the scope can be deduced from the declValue (so I think you should never have two keys with a same declare but different scopes)?

https://github.com/llvm/llvm-project/pull/201216