[flang-commits] [flang] [flang] [NFCI] Using getSource instead of getOriginalDef (PR #128984)

[Renaud Kauffmann via flang-commits]

================
@@ -621,38 +587,34 @@ AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v,
             if (mlir::isa<fir::PointerType>(boxTy.getEleTy()))
               attributes.set(Attribute::Pointer);
 
-            auto def = getOriginalDef(op.getMemref(), attributes,
-                                      isCapturedInInternalProcedure,
-                                      approximateSource);
-            if (auto addrOfOp = def.template getDefiningOp<fir::AddrOfOp>()) {
-              global = addrOfOp.getSymbol();
-
-              if (hasGlobalOpTargetAttr(def, addrOfOp))
-                attributes.set(Attribute::Target);
+            auto boxSrc = getSource(op.getMemref());
+            attributes |= boxSrc.attributes;
+            approximateSource |= boxSrc.approximateSource;
+            isCapturedInInternalProcedure |=
+                boxSrc.isCapturedInInternalProcedure;
 
+            global = llvm::dyn_cast<mlir::SymbolRefAttr>(boxSrc.origin.u);
+            if (global) {
               type = SourceKind::Global;
-            }
-            // TODO: Add support to fir.allocmem
-            else if (auto allocOp =
-                         def.template getDefiningOp<fir::AllocaOp>()) {
-              v = def;
-              defOp = v.getDefiningOp();
-              type = SourceKind::Allocate;
-            } else if (isDummyArgument(def)) {
-              defOp = nullptr;
-              v = def;
             } else {
-              type = SourceKind::Indirect;
+              auto def = llvm::cast<mlir::Value>(boxSrc.origin.u);
+              // TODO: Add support to fir.allocmem
+              if (auto allocOp = def.template getDefiningOp<fir::AllocaOp>()) {
+                v = def;
+                defOp = v.getDefiningOp();
+                type = SourceKind::Allocate;
----------------
Renaud-K wrote:

`Direct` was a misnomer because you never know what was stored at the address loaded. We can still use aliasing rules on these box indirections because we are comfortable relying on source level attributes (POINTER, TARGET,...). We assume that  FIR will not introduce an aliasing situation without these attributes.

FIR does not currently guarantee this for scalars. This distinction only became clear to me recently. So `Emboxed` would be more precise. If and when we can treat scalars like boxes, we can regroup `Emboxed` and `Indirect` and treat them the same.  





https://github.com/llvm/llvm-project/pull/128984