This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix execute/980205 on 32bit SPARC
- To: jakub at redhat dot com
- Subject: Re: [PATCH] Fix execute/980205 on 32bit SPARC
- From: "David S. Miller" <davem at redhat dot com>
- Date: Tue, 21 Mar 2000 09:34:57 -0800
- CC: rth at cygnus dot com, gcc-patches at gcc dot gnu dot org
- References: <20000321180459.R525@mff.cuni.cz>
Date: Tue, 21 Mar 2000 18:04:59 +0100
From: Jakub Jelinek <jakub@redhat.com>
If -O0, reload may happily store %fp + non-8-aligned const into
some stack slot, load it then again and mem_min_alignment will
assume it is 8 byte aligned which results in SIGBUS.
Ok, if this fixes that failure case, please apply this patch.
I've known for quite a while about this bug but never stumbled
upon a quick fix which cured it and didn't introduce new bugs. ;-)
I really wanted to fix this differently, but it's a lot more
work. The idea was to make the expanders force us into a particular
insn pattern based upon whether REGNO_POINTER_ALIGN was sufficient
or not for the ldd/std insns.
This way, we make the decision _before_ reload can run which is
when REGNO_POINTER_ALIGN information is still valid.
It really is a bug to do things the way we do now, it happens
to work, but it may not some time in the future. And this is
why we really should fix this properly as soon as we have the
chance.
Later,
David S. Miller
davem@redhat.com