[patch] PR33552, match_asm_constraints failures and miscompiles

[Richard Guenther]

On 9/27/07, Michael Matz <matz@suse.de> wrote:
> Hi,
>
> this patch fixes two bugs and changes another thing:
>
> 1) 4.3 can't compile glibc anymore on x86 with -O2 -fPIC , as
>    string-inlines.c get's reload failures because match_asm_constraints
>    generates a situation with one additional input operand, which reload
>    can't handle anymore
> 2) PR 33552 is about a miscompile of gmp, also caused by
>    match_asm_constraints, which doesn't check the validity of the
>    transform it does, in particular it overwrites an output operand
>    without checking that it isn't used also as input.
> 3) We now restrict ourself to only doing the transformation on REG_P
>    operands, based on an older discussion.  We could also restrict us to
>    !MEM_P or to REG_P and subregs of REG_P, but the situation which this
>    pass is trying to solve in practice only seems to be with REG_P.  This
>    let's us also not use copy_rtx.
>
> A former version of this patch was bootstrapped and regtested on
> ia64,ppc,x86_64 and i686.  I've restarted the process with this patch.
>
> There are two testcase, from Jakub's comment in PR33552, and one extracted
> from glibc.
>
>
> Ciao,
> Michael.
>         PR rtl-optimization/33552
>         * function.c (match_asm_constraints_1): Check for overlap in
>         inputs and replace all occurences.
>
>         dg.target/i386/pr33552.c: New runtime test.
>         dg.target/i386/strinline.c: New compile time test.

gcc.target/

This is ok if it passes the re-bootstrap & regtest.

Thanks,
Richard.