This is the mail archive of the
mailing list for the GCC project.
Re: Pb with clearing const_tiny_rtx in clear_const_double_mem
- To: Olivier Hainque <hainque at act-europe dot fr>
- Subject: Re: Pb with clearing const_tiny_rtx in clear_const_double_mem
- From: Richard Henderson <rth at redhat dot com>
- Date: Fri, 24 Aug 2001 11:39:12 -0700
- Cc: gcc at gcc dot gnu dot org
- References: <email@example.com>
On Fri, Aug 24, 2001 at 03:53:03PM +0200, Olivier Hainque wrote:
> float f1 = 1.0;
> void bar (void)
> float f2 = 1.0;
> I'm unclear as to what the initial problem was.
The nested function wouldn't use the canonical values, which means all of
the checks for CONST0_RTX etc fail, which causes instructions to not match
in the alpha backend, as well as lots o missed optimizations.
One way to solve this problem is to swap out these tiny_rtx's in the nested
context so that they are not shared, but we still match CONST0_RTX. Care
should be taken so that the original functions are restored by the time we
get to integrate.c.
A larger issue is that it is damned stupid to not share constants between
functions on machines that just dump constants in .rodata and don't have
in-function constant pools. That's a somewhat large cleanup though.