This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH] reload: X constraint doesn't match everything
> Uh, I think it fixes an ICE with some asm("xxx":"X"(rrr)) .
> See pr 20491.
Thanks for the pointer. My patch actually looks related to that issue.
The original failure in the pr appeared with a paradoxical subreg which
survived until reload and for which find_reloads would have set the
force_reload flag as in the example I had. In spite of the 'X' constraint
reload claimed that it couldn't find a valid reload. This is strange since
the programmer explicitly said that his inline assembly is able to deal with
everything by setting the 'X' constraint. So I would say reload should shut
up about it and give the programmer the nastiest rtx it is able to find somewhere.
According to the pr the bug can't be reproduced anymore except on hppa. Perhaps
someone with access to a hppa machine can give the patch a try?