[PATCH] aarch64: optimize the unaligned case of memcmp

[Sebastian Pop]

On 06/23/2017 04:28 PM, Wilco Dijkstra wrote:
> Sebastian Pop wrote:
>
>> If I remove all the alignment code, I get less performance on the hikey
>> A53 board.
>> With this patch:
> @@ -142,9 +143,23 @@ ENTRY(memcmp)
>
>           .p2align 6
>    .Lmisaligned8:
> +
> +       cmp     limit, #8
> +       b.lo    .LmisalignedLt8
> +
> +       .p2align 5
> +.Lloop_part_aligned:
> +       ldr     data1, [src1], #8
> +       ldr     data2, [src2], #8
> +       subs    limit_wd, limit_wd, #1
> +.Lstart_part_realigned:
> +       eor     diff, data1, data2      /* Non-zero if differences found. */
> +       cbnz    diff, .Lnot_limit
> +       b.ne    .Lloop_part_aligned
> +
> +.LmisalignedLt8:
>           sub     limit, limit, #1
>    1:
> -       /* Perhaps we can do better than this.  */
>           ldrb    data1w, [src1], #1
>           ldrb    data2w, [src2], #1
>           subs    limit, limit, #1
>
> Where is the setup of limit_wd and limit???

You are right, my patch was not quite correct: I was missing the 
initialization of limit_wd, like so:

lsr     limit_wd, limit, #3

limit is the number of bytes to be compared passed in as a parameter to 
memcmp.

With this extra statement I am still seeing the same low performance on 
A53 hikey:

Benchmark                              Time           CPU Iterations
--------------------------------------------------------------------
BM_string_memcmp_unaligned/8         345 ns        345 ns 2026483   
22.0879MB/s
BM_string_memcmp_unaligned/16        539 ns        539 ns 1298687   
28.3159MB/s
BM_string_memcmp_unaligned/20        613 ns        613 ns 1142076   
31.1222MB/s
BM_string_memcmp_unaligned/30        794 ns        794 ns 881596   
36.0357MB/s
BM_string_memcmp_unaligned/42        957 ns        957 ns 731746   
41.8753MB/s
BM_string_memcmp_unaligned/55       1208 ns       1207 ns 579591   
43.4525MB/s
BM_string_memcmp_unaligned/60       1231 ns       1231 ns 568372   
46.4756MB/s
BM_string_memcmp_unaligned/64       1312 ns       1312 ns 475862   
46.5316MB/s

The base is with no patch applied to memcmp.S: (byte by byte memcmp)

Benchmark                              Time           CPU Iterations
--------------------------------------------------------------------
BM_string_memcmp_unaligned/8         339 ns        339 ns 2066820   
22.5274MB/s
BM_string_memcmp_unaligned/16        536 ns        536 ns 1306265   
28.4901MB/s
BM_string_memcmp_unaligned/20        612 ns        612 ns 1146573   
31.1479MB/s
BM_string_memcmp_unaligned/30        789 ns        789 ns 886755   
36.2472MB/s
BM_string_memcmp_unaligned/42       1009 ns       1009 ns 693760      
39.7MB/s
BM_string_memcmp_unaligned/55       1233 ns       1233 ns 567719   
42.5469MB/s
BM_string_memcmp_unaligned/60       1322 ns       1322 ns 529511   
43.2804MB/s
BM_string_memcmp_unaligned/64       1392 ns       1392 ns 502817   
43.8426MB/s

And with the patch submitted for review without computing max and 
aligning on src1:

Benchmark                              Time           CPU Iterations
--------------------------------------------------------------------
BM_string_memcmp_unaligned/8         282 ns        282 ns 2482713    
27.061MB/s
BM_string_memcmp_unaligned/16        304 ns        304 ns 2300275   
50.1401MB/s
BM_string_memcmp_unaligned/20        322 ns        322 ns 2176437   
59.2469MB/s
BM_string_memcmp_unaligned/30        352 ns        352 ns 1988315    
81.328MB/s
BM_string_memcmp_unaligned/42        412 ns        412 ns 1699818    
97.317MB/s
BM_string_memcmp_unaligned/55        503 ns        503 ns 1393029   
104.382MB/s
BM_string_memcmp_unaligned/60        522 ns        522 ns 1340682   
109.619MB/s
BM_string_memcmp_unaligned/64        541 ns        541 ns 1297637   
112.891MB/s