[cfe-dev] [OpenCL patch] Vector literals are rvalues *updated*

[Anton Lokhmotov]

> Well, for example, you can pass the address of a compound literal to a
> function which takes a struct by pointer.
Thanks for the example.  This can be useful for dealing with legacy code not
under your control.  In OpenCL, all program code is in one module (passed as
one or more C-strings).

> Anyway, C99 is quite clear about compound literals being l-values,
> and we're not going to unilaterally change that.
Agree.

> > On Sep 21, 2011, at 10:01 AM, John McCall wrote:
> >> Right.  Ideally, this should just apply to OpenCL vector literals,
> >> i.e. this should be an r-value:
> >>  (int4) (3,2,1,0)
> >> and this should be an l-value:
> >>  (int4) {3,2,1,0}
> >
> > On 22 September 2011 00:57, Tanya Lattner wrote:
> >> Agreed. If you can have it apply only for OpenCL vector literals,
> >> then it looks good to me.
> >
> > Unfortunately, we can't, as these two forms are indistinguishable in
> > the AST.  (Strictly speaking, only the first form is valid in OpenCL, 
> > but as we discussed [1] we should permit the second one as well.)
> 
> Two forms being indistinguishable in the AST is theoretically solvable.
Even if so, I am afraid this will come at the expense of extra code.  More
importantly, neither OpenCL (see page 164 here:
http://www.khronos.org/registry/cl/specs/opencl-1.1.pdf) nor AltiVec (see
page 26 here:
http://www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPIM.pdf) permit
using curly braces for initialising vector literals.  Do we need to worry
about preserving behaviour of (theoretically existing) code that's illegal
to start with?

Thanks,
Anton.