[Bug tree-optimization/110026] New: [Bug] 5% performance drop on important benchmark after r260951.

[Bug tree-optimization/110026] New: [Bug] 5% performance drop on important benchmark after r260951.

[d_vampile at 163 dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110026

            Bug ID: 110026
           Summary: [Bug] 5% performance drop on important benchmark after
                    r260951.
           Product: gcc
           Version: 10.3.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: d_vampile at 163 dot com
  Target Milestone: ---

Created attachment 55184
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55184&action=edit
Open-source stream benchmark

After the patch is submitted on AArch64, the performance of copying subitems in
the stream benchmark decreases by 3%.

Alternatively, you can obtain it from
https://github.com/jeffhammond/stream/archive/master.zip.

Compiling & Running:
gcc -fopenmp -O -DSTREAM_ARRAY_SIZE=100000000 stream.c  -o stream
./stream

Before modification: (copy subitem)
ldr x2, [x3, x0, lsl #3]
str x2, [x4, x0, lsl #3]
add x0, x0, #0x1
cmp x1, x0
b.ne 400a00 <main._omp_fn.4+0x54>
ldr x19, [sp, #16]
ldp x29, x30, [sp], #32
ret

After the modification:
ldr d0, [x2, x0, lsl #3]
str d0, [x3, x0, lsl #3]
add x0, x0, #0x1
cmp x1, x0
b.ne 400a00 <main._omp_fn.4+0x54>
ldr x19, [sp, #16]
ldp x29, x30, [sp], #32
ret

It can be seen that the vector register (X0) is used before the modification,
and the common register (D0) is used after the modification.

[Bug tree-optimization/110026] [Bug] 5% performance drop on important benchmark after r260951.

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110026

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Note, any benchmarking for speed with -O rather than -O2/-O3 is intentionally
missing various optimizations which can greatly improve performance.

[Bug target/110026] [Bug] 5% performance drop on important benchmark after r260951.

[d_vampile at 163 dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110026

--- Comment #2 from d_vampile <d_vampile at 163 dot com> ---
(In reply to Jakub Jelinek from comment #1)
> Note, any benchmarking for speed with -O rather than -O2/-O3 is
> intentionally missing various optimizations which can greatly improve
> performance.

O0 does miss a lot of optimizations. However, for the problem I mentioned, the
GPRs used before and the FP registers after modification are used. When
vectorization is not applicable, the X0 register is faster than the D0
register. Is it appropriate to modify here?

[Bug target/110026] [Bug] 5% performance drop on important benchmark after r260951.

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110026

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ra

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to d_vampile from comment #2)
> O0 does miss a lot of optimizations. However, for the problem I mentioned,
> the GPRs used before and the FP registers after modification are used. When
> vectorization is not applicable, the X0 register is faster than the D0
> register. Is it appropriate to modify here?


Well the generic_tunings has:
  { 4, /* load_int.  */
    4, /* store_int.  */
    4, /* load_fp.  */
    4, /* store_fp.  */
    4, /* load_pred.  */
    4 /* store_pred.  */
  }, /* memmov_cost.  */


Which says the load/store of fp has the same cost as ints (gprs) (this is the
same as a53's tuning).

If anything that should be changed ....

Of you should use -mcpu=* where appliable.