[PATCH] [CodeGen] tighten objc ivar invariant.load attribution

[Saleem Abdulrasool]

On Fri, 15 Feb 2013, John McCall wrote:

> On Feb 14, 2013, at 3:14 PM, Saleem Abdulrasool <compnerd at compnerd.org> wrote:
> > On Thu, Feb 14, 2013 at 11:23 AM, John McCall <rjmccall at apple.com> wrote:
> >> On Feb 14, 2013, at 9:37 AM, Saleem Abdulrasool <compnerd at compnerd.org> wrote:
> >>> Add a few test cases and fix an inverted check.
> >>> 
> >>> Hi rjmccall,
> >>> 
> >>> http://llvm-reviews.chandlerc.com/D406
> >>> 
> >>> CHANGE SINCE LAST DIFF
> >>> http://llvm-reviews.chandlerc.com/D406?vs=976&id=982#toc
> >>> 
> >>> Files:
> >>> lib/CodeGen/CGObjCMac.cpp
> >>> test/CodeGenObjC/ivar-invariant-unimplemented.m
> >>> test/CodeGenObjC/ivar-invariant.m
> >>> 
> >>> Index: lib/CodeGen/CGObjCMac.cpp
> >>> ===================================================================
> >>> --- lib/CodeGen/CGObjCMac.cpp
> >>> +++ lib/CodeGen/CGObjCMac.cpp
> >>> @@ -6449,10 +6449,23 @@
> >>>                                               unsigned CVRQualifiers) {
> >>>  ObjCInterfaceDecl *ID = ObjectTy->getAs<ObjCObjectType>()->getInterface();
> >>>  llvm::Value *Offset = EmitIvarOffset(CGF, ID, Ivar);
> >>> +
> >>> +  // Annotate the load as an invariant load iff the object type is the type, or
> >>> +  // a derived type, of the class containing the ivar within an ObjC method.
> >>> +  // This check is needed because the ivar offset is a lazily initialised value
> >>> +  // that may depend on objc_msgSend to perform a fixup on the first message
> >>> +  // dispatch.
> >>> +  //
> >>> +  // An additional opportunity to mark the load as invariant arises when the
> >>> +  // base of the ivar access is a parameter to an Objective C method.  However,
> >>> +  // because the parameters are not available in the current interface, we
> >>> +  // cannot perform this check.
> >> 
> >> Go ahead and tweak the interface to propagate through an optional
> >> base expression and then implement your check based on that.
> > 
> > Do you mind if I were to do that as a follow on change?
> 
> No, that's fine.
>
> Do you need a commit?
>

Thanks; committed as r175386.

> John.