------- Comment #7 from rguenth at gcc dot gnu dot org 2009-09-30 08:59 -------
Hm, on x86_64 with -O3 -funroll-loops I see
f:
.LFB0:
.cfi_startproc
xorl %eax, %eax
.p2align 4,,10
.p2align 3
.L2:
movdqu (%rsi,%rax), %xmm7
movdqu %xmm7, (%rdi,%rax)
movdqu 16(%rsi,%rax), %xmm6
movdqu %xmm6, 16(%rdi,%rax)
movdqu 32(%rsi,%rax), %xmm5
movdqu %xmm5, 32(%rdi,%rax)
movdqu 48(%rsi,%rax), %xmm4
movdqu %xmm4, 48(%rdi,%rax)
movdqu 64(%rsi,%rax), %xmm3
movdqu %xmm3, 64(%rdi,%rax)
movdqu 80(%rsi,%rax), %xmm2
movdqu %xmm2, 80(%rdi,%rax)
movdqu 96(%rsi,%rax), %xmm1
movdqu %xmm1, 96(%rdi,%rax)
movdqu 112(%rsi,%rax), %xmm0
movdqu %xmm0, 112(%rdi,%rax)
subq $-128, %rax
cmpq $4096, %rax
jne .L2
rep
ret
which looks pretty optimal to me ...
If you disable vectorization then restrict doesn't make a difference anymore
because of TARGET_MEM_REFs:
<bb 3>:
# i_15 = PHI <i_10(3), 0(2)>
D.2707_9 = MEM[base: s_7(D), index: i_15, step: 4];
MEM[base: t_4(D), index: i_15, step: 4] = D.2707_9;
i_10 = i_15 + 1;
if (i_10 != 1024)
goto <bb 3>;
The MEM_EXPRs we get from expansion do not have points-to information
anymore. It's possible to fix that, though it might be not without
losses elsewhere (TMRs suck).
Well, I'll try to have a look here during stage3.
--
rguenth at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|unassigned at gcc dot gnu |rguenth at gcc dot gnu dot
|dot org |org
Status|NEW |ASSIGNED
Last reconfirmed|2009-09-29 21:18:52 |2009-09-30 08:59:39
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22031
