Re: optimize lookups in snapshot [sub]xip arrays

On Tue, Aug 02, 2022 at 03:55:39PM -0700, Andres Freund wrote:
> FWIW, I'd split the introduction of the helper and the use of it in snapmgr
> into separate patches.

Will do.

> On 2022-08-02 15:13:01 -0700, Nathan Bossart wrote:
>> diff --git a/src/include/c.h b/src/include/c.h
>> index d35405f191..2c1a47bc28 100644
>> --- a/src/include/c.h
>> +++ b/src/include/c.h
>> @@ -371,6 +371,14 @@ typedef void (*pg_funcptr_t) (void);
>> #endif
>> #endif
>>
>> +/*
>> + * Are SSE2 intrinsics available?
>> + */
>> +#if (defined(__x86_64__) || defined(_M_AMD64))
>> +#include <emmintrin.h>
>> +#define USE_SSE2
>> +#endif
>> +
>
> It doesn't strike me as a good idea to include this in every single
> translation unit in pg. That header (+dependencies) isn't small.
>
> I'm on board with normalizing defines for SSE availability somewhere central
> though.

Yeah, this is just a temporary hack for now. It'll go away once the
defines for SSE2 availability are committed.

>> +/*
>> + * pg_linearsearch_uint32
>> + *
>> + * Returns the address of the first element in 'base' that equals 'key', or
>> + * NULL if no match is found.
>> + */
>> +#ifdef USE_SSE2
>> +pg_attribute_no_sanitize_alignment()
>> +#endif
>
> What's the deal with this annotation? Needs a comment.

Will do. c.h suggests that this should only be used for x86-specific code.

>> +static inline uint32 *
>> +pg_linearsearch_uint32(uint32 key, uint32 *base, uint32 nelem)
>
> Hm. I suspect this could be a bit faster if we didn't search for the offset,
> but just for presence in the array? Most users don't need the offset.

Just under half of the callers in 0002 require the offset, but I don't know
if any of those are worth optimizing in the first place. I'll change it
for now. It's easy enough to add it back in the future if required.

--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com