On Sun, Aug 14, 2011 at 7:24 PM, Uros Bizjak <ubiz...@gmail.com> wrote:
> Hello!
>
> We can use ROUNDSP/ROUNDSD in round(a) expansion. Currently, we expand
> round(a) as (-O2 -ffast-math):
>
> .LFB0:
> .cfi_startproc
> movsd .LC1(%rip), %xmm1
> movapd %xmm0, %xmm2
> movsd .LC0(%rip), %xmm3
> andpd %xmm1, %xmm2
> ucomisd %xmm2, %xmm3
> jbe .L2
> addsd .LC2(%rip), %xmm2
> andnpd %xmm0, %xmm1
> movapd %xmm1, %xmm0
> cvttsd2siq %xmm2, %rax
> cvtsi2sdq %rax, %xmm2
> orpd %xmm2, %xmm0
> .L2:
> rep
> ret
>
> Adding -msse4, we now generate branchless code using roundsd:
>
> .LFB0:
> .cfi_startproc
> movsd .LC0(%rip), %xmm2
> movapd %xmm0, %xmm1
> andpd %xmm2, %xmm1
> andnpd %xmm0, %xmm2
> addsd .LC1(%rip), %xmm1
> roundsd $1, %xmm1, %xmm1
> orpd %xmm2, %xmm1
> movapd %xmm1, %xmm0
> ret
Hm, why do we need the sign-copy? If I read the docs correctly
we can simply use roundsd directly, no?
> The patch also simplifies a couple of checks in related patterns.
>
> 2011-08-14 Uros Bizjak <ubiz...@gmail.com>
>
> * config/i386/i386.c (ix86_expand_round_sse4): New function.
> * config/i386/i386-protos.h (ix86_expand_round_sse4): New prototype.
> * config/i386/i386.md (round<mode>2): Use ix86_expand_round_sse4
> for TARGET_ROUND.
>
> (rint<mode>2): Simplify TARGET_ROUND check.
> (floor<mode>2): Ditto.
> (ceil<mode>2): Ditto.
> (btrunc<mode>2): Ditto.
>
> Bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32},
> will be committed to mainline soon.
>
> Uros.
>
