On Wed, Jan 9, 2013 at 10:23 AM, Uros Bizjak <ubiz...@gmail.com> wrote:
>> No matter whether LRA (if it is a bug in there) is fixed or not,
>> *vec_concatv2df could handle for !avx sse3 x <- x, 1 alternative the same
>> as it handles x <- m, 1 alternative (using movddup).
>>
>> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>>
>> 2013-01-08 Jakub Jelinek <ja...@redhat.com>
>>
>> PR rtl-optimization/55829
>> * config/i386/sse.md (*vec_concatv2df): Add x <- x, 1 alternative
>> for sse3 but not avx.
>>
>> * gcc.target/i386/pr55829.c: New test.
>>
>> --- gcc/config/i386/sse.md.jj 2012-11-26 10:14:26.000000000 +0100
>> +++ gcc/config/i386/sse.md 2013-01-08 10:28:42.496819712 +0100
>> @@ -5183,10 +5183,10 @@ (define_insn "vec_dupv2df"
>> (set_attr "mode" "V2DF")])
>>
>> (define_insn "*vec_concatv2df"
>> - [(set (match_operand:V2DF 0 "register_operand" "=x,x,x,x,x,x,x,x")
>> + [(set (match_operand:V2DF 0 "register_operand" "=x,x,x, x,x,x,x,x")
>> (vec_concat:V2DF
>> - (match_operand:DF 1 "nonimmediate_operand" " 0,x,m,0,x,m,0,0")
>> - (match_operand:DF 2 "vector_move_operand" " x,x,1,m,m,C,x,m")))]
>> + (match_operand:DF 1 "nonimmediate_operand" " 0,x,xm,0,x,m,0,0")
>> + (match_operand:DF 2 "vector_move_operand" " x,x,1, m,m,C,x,m")))]
>
> This was done on purpose, since reload had some problems with similar
> pattern (please see PR 50875 [1] and [2]). If we are sure that LRA
> fixes this problem, then the patch is OK for mainline.
>
> Also, please revert "hack" that fixed PR 50875 in this case.
Looking into this problem a bit more: After Vladimir's LRA patch went
in, we generate for gcc.target/i386/pr55829.c:
movq p1(%rip), %r12 # 56 *movdi_internal_rex64/2 [length = 7]
movq %r12, (%rsp) # 57 *movdi_internal_rex64/4 [length = 4]
movddup (%rsp), %xmm1 # 23 *vec_concatv2df/3 [length = 5]
Combined with your proposed patch:
movq p1(%rip), %r12 # 60 *movdi_internal_rex64/2 [length = 7]
movq %r12, (%rsp) # 61 *movdi_internal_rex64/4 [length = 4]
movsd (%rsp), %xmm1 # 56 *movdf_internal_rex64/10
[length = 5]
unpcklpd %xmm1, %xmm1 # 23 *vec_concatv2df/1
[length = 4]
That is, one more move to use unpcklpd.
Based on this evidence, I think that the proposed patch should be
rejected, the generic LRA fix alone results in better code.
Thanks,
Uros.
