在 2023/12/23 上午10:26, chenglulu 写道:
在 2023/12/22 下午3:21, chenglulu 写道:
在 2023/12/22 下午3:09, Xi Ruoyao 写道:
On Fri, 2023-12-22 at 11:44 +0800, chenglulu wrote:
在 2023/12/21 下午8:00, chenglulu 写道:
Sorry, I've been busy with something else these two days. I don't
think there's anything wrong with the code,
but I need to test the spec.:-)
Hi, Ruoyao:
After applying this patch, spec2006 464.h264 ref will have a 6.4%
performance drop. So I'm going to retest it.
I think 6.4% is large enough not to be a random error.
Is there an example showing the code regression?
And I'm wondering if keeping the peephole besides the new
define_insn_and_split produces a better result instead of solely
relying
on define_insn_and_split?
I haven't debugged this yet, I'm retesting, if there is still such a
big performance gap,
I think I need to see the reason.
The performance drop has nothing to do with this patch. I found that
the h264 performance compiled
by r14-6787 compared to r14-6421 dropped by 6.4%.
But there is a problem. My regression test has the following two fail
items.(based on r14-6787)
+FAIL: gcc.dg/cpp/_Pragma3.c (test for excess errors)
+FAIL: gcc.dg/pr86617.c scan-rtl-dump-times final "mem/v" 6
