[flang-commits] [flang] [flang] AliasAnalysis: distinguish addr of arg vs. addr in arg (PR #87723)

[Joel E. Denny via flang-commits]

================
@@ -399,16 +407,18 @@ AliasAnalysis::Source AliasAnalysis::getSource(mlir::Value v) {
   if (!defOp && type == SourceKind::Unknown)
     // Check if the memory source is coming through a dummy argument.
     if (isDummyArgument(v)) {
-      type = SourceKind::Argument;
       ty = v.getType();
       if (fir::valueHasFirAttribute(v, fir::getTargetAttrName()))
         attributes.set(Attribute::Target);
-
       if (Source::isPointerReference(ty))
         attributes.set(Attribute::Pointer);
+      if (followBoxAddr && attributes.test(Attribute::Pointer))
----------------
jdenny-ornl wrote:

For the AddrOfOp case: I noticed that flang/test/Analysis/AliasAnalysis/alias-analysis-8.fir expects the relationship between the address *in* an allocatable global and the address *of* that allocatable global to be NoAlias.

For the isDummyArg() case: I just checked the same relationship for allocatable arguments and discovered my patch wasn't handling it correctly: it has it as MustAlias.  The pointer check we're discussing is too strict.  I'll push a patch shortly that relaxes that to `fir::isa_ref_type(ty)` to turn it into NoAlias.  However, flang/test/Transforms/tbaa2.fir then fails because AddAliasTags cannot yet handle Direct for the allocatable arguments there.  My patch also attempts to extend AddAliasTags accordingly.

I also tried adding `fir::isa_ref_type(ty)` to the AddrOfOp case, and check-flang still behaved.  So, it doesn't improve existing test cases, and I haven't found a `!fir.box<!fir.array<...>>` case for AddrOfOp (see previous comment), so for now I've abandoned that addition as apparently useless.

https://github.com/llvm/llvm-project/pull/87723