On 8/22/2018 12:58 PM, Duy Nguyen wrote:
On Wed, Aug 22, 2018 at 6:49 PM Derrick Stolee <stolee@xxxxxxxxx> wrote:
On 8/22/2018 12:26 PM, Jeff King wrote:
On Wed, Aug 22, 2018 at 06:14:24PM +0200, Duy Nguyen wrote:
On Wed, Aug 22, 2018 at 6:08 PM Duy Nguyen <pclouds@xxxxxxxxx> wrote:
On Wed, Aug 22, 2018 at 6:03 PM Jeff King <peff@xxxxxxxx> wrote:
On Wed, Aug 22, 2018 at 07:14:42AM -0400, Derrick Stolee wrote:
The other thing I was going to recommend (and I'll try to test this out
myself later) is to see if 'the_hash_algo->rawsz' is being treated as a
volatile variable, since it is being referenced through a pointer. Perhaps
storing the value locally and then casing on it would help?
I tried various sprinkling of "const" around the declarations to make it
clear that the values wouldn't change once we saw them. But I couldn't
detect any difference. At most I think that would let us hoist the "if"
out of the loop, but gcc still seems unwilling to expand the memcmp when
there are other branches.
I think if that's the thing we want to have happen, we really do need to
just write it out on that branch rather than saying "memcmp".
This reminds me of an old discussion about memcpy() vs doing explicit
compare loop with lots of performance measurements..
Ah found it. Not sure if it is still relevant in light of multiple hash support
https://public-inbox.org/git/20110427225114.GA16765@xxxxxxx/
Yes, that was what I meant. We actually did switch to that hand-rolled
loop, but later we went back to memcmp in 0b006014c8 (hashcmp: use
memcmp instead of open-coded loop, 2017-08-09).
Looking at that commit, I'm surprised the old logic was just a for loop, instead of a word-based approach, such as the following:
Might work on x86 but it breaks on cpu architectures with stricter
alignment. I don't think we have a guarantee that object_id is always
8 byte aligned.
You (and Peff) are probably correct here, which is unfortunate. I'm not
familiar with alignment constraints, but assume that such a word-based
approach is best.
Thanks,
-Stolee
