Re: fold_builtin_memcpy cannot optimize <retval>?

[Jeffrey A Law <law at redhat dot com>]

On Fri, 2005-01-21 at 11:51 +0100, Jakub Jelinek wrote:
> On Fri, Jan 21, 2005 at 11:38:37AM +0100, Richard Guenther wrote:
> > The simple testcase
> > 
> > struct Foo {};
> > Foo bar(const Foo& a)
> > {
> >   Foo res;
> >   res = a;
> >   return res;
> > }
> > 
> > leaves us with
> > 
> > Foo bar(const Foo&) (a)
> > {
> >   struct Foo res;
> > 
> >   # BLOCK 0
> >   # PRED: ENTRY [100.0%]  (fallthru,exec)
> >   __builtin_memcpy (&<retval>, a_1, 0);
> >   return <retval>;
> >   # SUCC: EXIT [100.0%]
> > 
> > }
> > 
> > in the optimized tree-dump and obviously ugly asm code.
> > 
> > What's going on here?
> 
> The arguments to __builtin_memcpy fail validation (validate_arglist).
> The reason is that second argument in this case is not POINTER_TYPE
> as memcpy requires, but REFERENCE_TYPE.
> 
> Not sure if it is better to just make sure __builtin_memcpy will
> never see a REFERENCE_TYPE argument in this case, or make argument
> validation in builtins.c less strict and for POINTER_TYPE requirements
> allow REFERENCE_TYPEs as well.
I'm not an expert on the differences in the semantics of POINTER_TYPE
and REFERENCE_TYPE, but I do know that I've been treating them as
effectively the same.

Jeff