[PATCH, i386]: Improve rsqrt scehduling a bit

[Uros Bizjak]

Hello!

This patch changes NR fixup of rsqrt from:

rsqrt(a) = 0.5 * rsqrtss(a) * (3.0 - a * rsqrtss(a) * rsqrtss(a))

to:

rsqrt(a) = -0.5 * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) + (-3.0))

The benefit of this change is, that we can use commutative addss instead 
of subss, improving the sequence from:

        movaps  %xmm0, %xmm1
        movss   .LC0(%rip), %xmm0
        rsqrtss %xmm1, %xmm2
        mulss   %xmm2, %xmm1
        mulss   %xmm2, %xmm1
        mulss   .LC1(%rip), %xmm2
        subss   %xmm1, %xmm0
        mulss   %xmm2, %xmm0

to:

        rsqrtss %xmm0, %xmm1
        mulss   %xmm1, %xmm0
        mulss   %xmm1, %xmm0
        mulss   .LC1(%rip), %xmm1
        addss   .LC0(%rip), %xmm0
        mulss   %xmm1, %xmm0

Patch was bootstrapped and regression tested on x86_64-linux {,-m32}, 
committed to SVN.

2008-01-06  Uros Bizjak  <ubizjak@gmail.com>

        * config/i386/i386.c (ix86_emit_swsqrtsf): Use negative constants in
        order to use commutative addition instead of subtraction.

Uros.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: rsqrt.diff.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20080106/2b7973e1/attachment.txt>