[PATCH] Scoped NoAlias Metadata

[hfinkel at anl.gov]

  Thanks for looking at this! I'll update and rebase soon.


================
Comment at: lib/IR/IRBuilder.cpp:75
@@ -73,3 +74,3 @@
   
   // Set the TBAA info if present.
   if (TBAATag)
----------------
Renato Golin wrote:
> Update this comment
No, this is still specifically TBAA.

Unlike everywhere else, for IRBuilder, I used three separate function parameters for the (now three) different MD types. This gives enough interface stability that Clang need not be updated with this change (and I assume the same is true of other clients).

================
Comment at: lib/IR/IRBuilder.cpp:106
@@ -97,3 +105,3 @@
 
   // Set the TBAA Struct info if present.
   if (TBAAStructTag)
----------------
Renato Golin wrote:
> update this comment
No, also still TBAA.

================
Comment at: lib/IR/IRBuilder.cpp:133
@@ -117,3 +132,3 @@
   
   // Set the TBAA info if present.
   if (TBAATag)
----------------
Renato Golin wrote:
> update this comment
No, also still TBAA.

================
Comment at: lib/Transforms/Utils/InlineFunction.cpp:426
@@ +425,3 @@
+          !isa<AtomicCmpXchgInst>(NI) && !isa<AtomicRMWInst>(NI) &&
+          !isa<MemIntrinsic>(NI))
+        continue;
----------------
Renato Golin wrote:
> Couldn't this be an else on the if block below? It's probably going to evaluate isa<> on each one (at least) anyway.
Yes, good point.

================
Comment at: lib/Transforms/Utils/InlineFunction.cpp:447
@@ +446,3 @@
+
+      // If we found on pointers, then this instruction is not suitable for
+      // pairing with an instruction to receive aliasing metadata.
----------------
Renato Golin wrote:
> found "no" pointers?
Yes, thanks!


http://llvm-reviews.chandlerc.com/D2194