Fix ICE on loop over constant vector at -O

[Richard Biener]

On Mon, Dec 10, 2012 at 10:55 AM, Eric Botcazou <ebotcazou@adacore.com> wrote:
>> Well ... I would have expected that we'd have folded the CONST_DECL to
>> its DECL_INITIAL.  At least that is what we do for all other CONST_DECLs.
>> The only way CONST_DECLs should appear in the IL (after some optimization
>> of course) is when you take their address.
>
> That's what (essentially) happens here.  The CONST_DECL is wrapped up in a
> VIEW_CONVERT_EXPR to array type and a component with non-constant index is
> extracted from it.

Ah, ok ...

>> So - if you can burn some extra cycles to figure out why it's not
>> replaced by its DECL_INITIAL ...?
>
> We specifically prevent that in gigi because it will need to be spilled to
> memory given the above access pattern, so I don't really see the problem.
> Here's the access in .optimized:
>
> _38 = MEM[symbol: pkg__zero_unit, index: ivtmp.38_27, step: 4, offset: -4B];

... so if would really be a pessimization doing that.  Of course handling
CONST_DECL in for_each_index is indeed obvious - I just was curios if
it was a missed-optimization opportunity as well.

Thanks for clarifying,
Richard.

> --
> Eric Botcazou