On Tue, May 8, 2018 at 11:11 AM, Uros Bizjak <ubiz...@gmail.com> wrote:
> On Mon, Apr 30, 2018 at 9:19 PM, Jakub Jelinek <ja...@redhat.com> wrote:
>> Hi!
>>
>> Before avx512vl we don't have a single instruction to do V2DImode and
>> V4DImode abs, but that isn't much different from say V4SImode before SSE3
>> where we also just emit a short sequence that is better than elementwise
>> expansion. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for
>> trunk?
>>
>> 2018-04-30 Jakub Jelinek <ja...@redhat.com>
>>
>> PR target/85572
>> * config/i386/i386.c (ix86_expand_sse2_abs): Handle E_V2DImode and
>> E_V4DImode.
>> * config/i386/sse.md (abs<mode>2): Use VI_AVX2 iterator instead of
>> VI1248_AVX512VL_AVX512BW. Handle V2DImode and V4DImode if not
>> TARGET_AVX512VL using ix86_expand_sse2_abs. Formatting fixes.
>>
>> * g++.dg/other/sse2-pr85572-1.C: New test.
>> * g++.dg/other/sse2-pr85572-2.C: New test.
>> * g++.dg/other/sse4-pr85572-1.C: New test.
>> * g++.dg/other/avx2-pr85572-1.C: New test.
>
> LGTM.
>
> Thanks,
> Uros.
>
>> --- gcc/config/i386/i386.c.jj 2018-04-25 15:09:29.895453703 +0200
>> +++ gcc/config/i386/i386.c 2018-04-30 18:31:56.027101932 +0200
>> @@ -49806,39 +49806,74 @@ ix86_expand_sse2_abs (rtx target, rtx in
>>
>> switch (mode)
>> {
>> + case E_V2DImode:
>> + case E_V4DImode:
>> + /* For 64-bit signed integer X, with SSE4.2 use
>> + pxor t0, t0; pcmpgtq X, t0; pxor t0, X; psubq t0, X.
>> + Otherwise handle it similarly to V4SImode, except use 64 as W
>> instead of
>> + 32 and use logical instead of arithmetic right shift (which is
>> + unimplemented) and subtract. */
>> + if (TARGET_SSE4_2)
>> + {
>> + tmp0 = gen_reg_rtx (mode);
>> + tmp1 = gen_reg_rtx (mode);
>> + emit_move_insn (tmp1, CONST0_RTX (mode));
>> + if (mode == E_V2DImode)
>> + emit_insn (gen_sse4_2_gtv2di3 (tmp0, tmp1, input));
>> + else
>> + emit_insn (gen_avx2_gtv4di3 (tmp0, tmp1, input));
}
else
{
>> + tmp0 = expand_simple_binop (mode, LSHIFTRT, input,
>> + GEN_INT (GET_MODE_UNIT_BITSIZE (mode) - 1),
>> + NULL, 0, OPTAB_DIRECT);
>> + tmp0 = expand_simple_unop (mode, NEG, tmp0, NULL, false);
}
>> + tmp1 = expand_simple_binop (mode, XOR, tmp0, input,
>> + NULL, 0, OPTAB_DIRECT);
>> + x = expand_simple_binop (mode, MINUS, tmp1, tmp0,
>> + target, 0, OPTAB_DIRECT);
>> + break;
You could merge parts of the above code.
Uros.
