Re: [PATCH] stdlib: reorganize stdlib Makefile routines by functionality

["H.J. Lu" <hjl.tools@gmail.com>]

On Mon, Apr 1, 2024 at 7:28 PM abush wang <abushwangs@gmail.com> wrote:
>
> Actually, not just d275970ab
> I found after a91bf4e0ff, there is also performance degradation on x86-64,
> even if this commit has nothing  to do with lrand48.
> This is my test data:
> before  a91bf4e0ff:
> Average time for lrand48: 1940 cycles
>
> after:
> Average time for lrand48: 5626 cycles

Please compare alignments of 2 versions of lrand48.

> It seems like there is a gradual performance degradation for lrand48.
>
>
> On Tue, Apr 2, 2024 at 10:17 AM abush wang <abushwangs@gmail.com> wrote:
>>
>> Yes，on x86-64.
>> I just compare the disassemble between d275970ab and before commit by objdump.
>> And __drand48_iterate will be more long distance after d275970ab, so I revert this
>> commit and found the performance will recover a little.
>>
>> Thanks,
>> abush
>>
>>
>> On Mon, Apr 1, 2024 at 9:12 PM Florian Weimer <fweimer@redhat.com> wrote:
>>>
>>> * abush wang:
>>>
>>> > This is test:
>>> > ```
>>> > uint64_t getnsecs() {
>>> >     uint32_t lo, hi;
>>> >     __asm__ __volatile__ (
>>> >         "rdtsc" : "=a"(lo), "=d"(hi)
>>> >     );
>>> >     return ((uint64_t)hi << 32) | lo;
>>> > }
>>> >
>>> > int main() {
>>> >     const int num_iterations = 1;
>>> >     uint64_t start, end, total_time = 0;
>>> >
>>> >     start = getnsecs();
>>> >     for (int i = 0; i < num_iterations; i++) {
>>> >         (void) lrand48();
>>> >     }
>>> >     end = getnsecs();
>>> >     total_time += (end - start);
>>> >
>>> >     printf("Average time for lrand48: %lu cycles\n", total_time / num_iterations);
>>> >     return 0;
>>> > }
>>> > ```
>>> > before:
>>> > Average time for lrand48: 21418 cycles
>>> >
>>> > after:
>>> > Average time for lrand48: 9892 cycles
>>>
>>> Do you see this on x86-64?  So this isn't a displacement range issue?
>>>
>>> It could be that this is a random performance change due to code
>>> alignment, and not actually caused by the direct call distance.
>>>
>>> Thanks,
>>> Florian
>>>


-- 
H.J.