This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [Committed] Only allow INTEGER_CSTs of integral/pointer types.


Hi Jason,

On Sat, 11 Nov 2006, Jason Merrill wrote:
> Yep, still no regressions on x86_64-pc-linux-gnu.  If it seems
> reasonable to you, I'll go ahead and check it in.

It looks good to me.  However my one hesitation (and presumably also
yours as you're a global maintainer) is that this affects the definition
of gimple.  Given none of the gimple architects have commented or
complained, and this resolves a P1 ICE-on-valid regression affecting
4.1, 4.2 and 4.3, I'd say go for it.  It doesn't make sense to
decompose/gimplify the components of a constant vector, even if
the elements aren't gimple constants.

The one potential issue we may need to look out for in future, is if
we ever enable an optimization to propagate these constants out of
the vector, we'll need to confirm they're gimple constants before
CCPing them into stmts.  This seems like a reasonable constraint.

Many congratulations again on fixing this.  I was banging my head
against it for several days without success.

Roger
--


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]