This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH][BZ #17801] Fix memcpy regression (five times slower on bulldozer.)
- From: "Carlos O'Donell" <carlos at redhat dot com>
- To: "H.J. Lu" <hjl dot tools at gmail dot com>, OndÅej BÃlka <neleai at seznam dot cz>, GNU C Library <libc-alpha at sourceware dot org>
- Date: Sat, 31 Jan 2015 01:50:32 -0500
- Subject: Re: [PATCH][BZ #17801] Fix memcpy regression (five times slower on bulldozer.)
- Authentication-results: sourceware.org; auth=none
- References: <20150106142939 dot GB5835 at domone> <CAMe9rOo4tmQc0bJ1Z=pjURvPBBMndwQ-ynbHc=Mpz3eD=eqjbg at mail dot gmail dot com> <20150130145656 dot GA26219 at gmail dot com> <CAMe9rOr9Vn93958O=1fFBTU8zTRJ_j-15EvrOq7J05SwTRejgw at mail dot gmail dot com>
On 01/30/2015 03:04 PM, H.J. Lu wrote:
>>>
>>> Please add a new feature bit, bit_Fast_AVX_Unaligned_Load, and turn it
>>> on together
>>> with bit_AVX2_Usable.
>>>
>>
>> I know we are in freeze. But I'd like to fix this regression in 2.21.
>> OK for master?
>
> Since this is a serious performance regression, I will check it in
> before the end of the day unless I am told otherwise.
In the future please TO: me so that I have high visibility on this change
as the release manager. I'm testing each of the changes to make sure things
are in good shape for the release.
Could you explain in detail why this is needed?
+#if index_AVX2_Usable != index_AVX_Fast_Unaligned_Load
+# error index_AVX2_Usable != index_AVX_Fast_Unaligned_Load
+#endif
Why do they have to be on the same index in the feature
array of bits? I don't see anywhere that checks them
both simultaneously. At the very least please add a detailed
comment why the error condition exists and how to fix it in
the future if another author needs to fix it.
Cheers,
Carlos.