http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56200
--- Comment #2 from Alexander Monakov <amonakov at gcc dot gnu.org> 2013-02-04
21:36:38 UTC ---
(In reply to comment #1)
> What happens if you also use -fno-ivopts ?
For me, -fno-ivopts gives a small improvement, but still slower than -O0. I
think the slowdown is related to code layout in the Icache and branch
predictors. There is a hot region which is composed of three consecutive
conditional branches (cmp-jg-cmp-jg-cmp-jg in optimized code and
mov-cmp-jl-mov-cmp-jl-mov-cmp-jl at -O0). If I align the first _and_ the second
to a 16-byte boundary, I get better performance then -O0, but aligning only one
of those is still slower than -O0:
--- o1.s 2013-02-05 00:04:44.405072150 +0400
+++ o1h.s 2013-02-05 01:17:43.648014420 +0400
@@ -119,9 +119,11 @@ find:
movq %rdx, %rbp
leal 1(%r14), %eax
movl %eax, 12(%rsp)
+ .p2align 4,,7
.L18:
cmpl file(%r12), %r14d
jg .L17
+ .p2align 4,,7
cmpl (%r15,%r12), %r14d
jg .L17
cmpl (%rbx), %r14d
