[i386 PATCH] PR26600: SSE2 related ICE in reload
Richard Henderson
rth@redhat.com
Fri May 12 20:45:00 GMT 2006
On Thu, May 11, 2006 at 10:37:59AM -0600, Roger Sayle wrote:
> The little white lie is that 32-bit x86 doesn't actually have
> TImode hard registers which makes reloads slightly tricky.
Uh... yes it does. What do you think all those SSE registers are?
Since "movti" exists, TImode exists, QED.
> At first glance looks like a SET pattern where the SRC doesn't have
> the same mode as the destination. Fortunately, this is fixed up
> behind the scenes, but there's no good reason for it to be this ugly.
Actually, there is. Emitting raw SUBREG within patterns is
to be considered a mistake. It's too easy to get nested
subregs and other such evil things.
> Ok for mainline? And the 4.1 branch after a few days?
Not ok.
r~
More information about the Gcc-patches
mailing list