[PATCH, alpha]: Fix PR target/42113 some more

[Uros Bizjak <ubizjak at gmail dot com>]

Hello!

For certain corner cases, gcc can generate invalid pattern from 
mentioned splitters:

/var/tmp/portage/sys-devel/gcc-4.3.4/work/gcc-4.3.4/libgcc/../gcc/libgcc2.c:551: error: unrecognizable insn:
(insn 131 59 132 6/var/tmp/portage/sys-devel/gcc-4.3.4/work/gcc-4.3.4/libgcc/../gcc/libgcc2.c:549 (set (reg:SI 114)
        (ne:DI (reg:DI 97)
            (const_int 0 [0x0]))) -1 (nil))
/var/tmp/portage/sys-devel/gcc-4.3.4/work/gcc-4.3.4/libgcc/../gcc/libgcc2.c:551: internal compiler error: in extract_insn, at recog.c:2005
Please submit a full bug report,
with preprocessed source if appropriate.


Please note SImode mode of target register of the cmp-to-zero 
comparison. This represents invalid instruction, since the target 
register should be in DImode. The problem is then in a followup SImode 
multiplication: we have to use SImode subreg of the scratch register 
there, which is what the patch does. Additionally, scratch register has 
to be generated as DImode paradoxical subreg of output regno for 
overlapped regs when no pseudo can be created.

My previous 42113 patch reversed the modes (gcc was able to avoid 
generating invalid insn for some reason), but not in the case above.

Attached patch should generate insns that are always valid.

2010-03-24  Uros Bizjak <ubizjak@gmail.com>

    PR target/42113
    * config/alpha/alpha.md (*cmp_sadd_si): Change mode
    of scratch register to DImode.  Split to DImode comparison operator.
    Use SImode subreg of scratch register in the multiplication.
    (*cmp_sadd_sidi): Ditto.
    (*cmp_ssub_si): Ditto.
    (*cmp_ssub_sidi): Ditto.

Patch was bootstrapped and regression tested on alphaev68-pc-linux-gnu 
for 4.3, 4.4 and 4.5.

OK for mainline and branches?

Uros.

Attachment: a.diff.txt
Description: Text document