Re: [PATCH, i386]: Macroize min/max patterns some more

[Uros Bizjak <ubizjak at gmail dot com>]

On Wed, Sep 21, 2011 at 9:45 AM, Jakub Jelinek <jakub@redhat.com> wrote:

>> ? ? ? * config/i386/i386.md (maxmin): New code iterator.
>> ? ? ? * config/i386/sse.md (<maxmin:code><mode>3): Macroize expander
>> ? ? ? from <umaxmin:code><mode>3 and <smaxmin:code><mode>3 using maxmin
>> ? ? ? code iterator.
>> ? ? ? (*avx2_<maxmin:code><mode>3): Macroize isn from
>> ? ? ? *avx2_<umaxmin:code><mode>3 and *avx2_<smaxmin:code><mode>3 using
>> ? ? ? maxmin code iterator.
>> ? ? ? (<smaxmin:code><VI124_128:mode>3): Merge with <smaxmin:code>v8hi3.
>> ? ? ? (<umaxmin:code><VI124_128:mode>3): Merge with umaxv4si3 and
>> ? ? ? <umaxmin:code>v16qi3.
>>
>> Tested on x86_64-pc-linux-gnu {,-m32}, committed to mainline SVN.
>
> This regressed gcc.dg/vect/vect-reduc-10.c code quality with -msse2,
> - ? ? ? psubusw (%rax), %xmm0
> - ? ? ? paddw ? (%rax), %xmm0
> + ? ? ? movdqa ?(%rax), %xmm1
> ? ? ? ?addq ? ?$16, %rax
> ? ? ? ?cmpq ? ?$aus+2048, %rax
> + ? ? ? psubusw %xmm1, %xmm0
> + ? ? ? paddw ? %xmm1, %xmm0
>
> the problem is that the two expanders force arguments into registers
> unconditionally, while previously in some cases they were
> nonimmediate_operand instead of register_operand predicated.
> This patch fixes that by always using nonimmediate_operand and adding
> force_reg where needed.
>
> 2011-09-21 ?Jakub Jelinek ?<jakub@redhat.com>
>
> ? ? ? ?* config/i386/sse.md (<code><mode>3 smaxmin:VI124_128 expander): Use
> ? ? ? ?nonimmediate_operand instead of register_operand predicate for operands
> ? ? ? ?1 and 2, force them into registers if expanding them as comparison.
> ? ? ? ?(<code><mode>3 umaxmin:VI124_128 expander): Similarly. ?For UMAX
> ? ? ? ?V8HImode force into register just operand 1.

OK.

Thanks,
Uros.