This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch] (4.1 stage 2 projects): vectorize reduction, part 3/n
- From: Paolo Bonzini <paolo dot bonzini at lu dot unisi dot ch>
- To: Richard Henderson <rth at redhat dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 05 Jul 2005 09:33:40 +0200
- Subject: Re: [patch] (4.1 stage 2 projects): vectorize reduction, part 3/n
- References: <20050619172501.GA20194@redhat.com> <OFA7F7C79E.BEFDF6B2-ONC2257026.002CCCDD-C2257026.005118FC@il.ibm.com> <20050620181455.GA22847@redhat.com> <d98dp9$3mj$1@sea.gmane.org> <20050627044507.GA27305@redhat.com>
Richard Henderson wrote:
On Tue, Jun 21, 2005 at 09:03:48AM +0200, Paolo Bonzini wrote:
I think min and max should be implemented in term of unsigned saturated
subtraction for unsigned values.
Hmm, yes.
uminv8hi:
movdqa a, t1
psubusw b, t1 /* t1[n] == 0 iff b[n] >= a[n] */
pxor t2, t2
pcmpeqw t1, t2 /* t2[n] = b[n] < a[n] ? -1 : 0 */
movdqa t2, t3
pand a, t3
pandn b, t2
por t3, t2
I had in mind something like
uminv8hi:
movdqa a, t1
movdqa t1, t2
psubusw b, t2 /* t2 = b[n] >= a[n] ? 0 : a[n] - b[n] */
psubw t1, t2 /* t1 = b[n] >= a[n] ? a[n] : b[n] */
For maximum, I could not come up with anything better than this
umaxv8hi:
movdqa a, t1
movdqa t1, t2
psubusw b, t2 /* t2 = b[n] >= a[n] ? 0 : a[n] - b[n] */
psubw t1, t2 /* t1 = b[n] >= a[n] ? a[n] : b[n] */
pxor t1, t2 /* t2 = a[n] ^ min(a[n], b[n]) */
pxor b, t2 /* t2 = a[n] ^ b[n] ^ min(a[n], b[n]) */
which has a lot of dependencies, but very low register pressure.
Paolo