Re: [PATCH] x86-64: Use _dl_runtime_resolve_opt only with AVX512F

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

On Wed, Aug 2, 2017 at 8:12 AM, H.J. Lu <hongjiu.lu@intel.com> wrote:
> On AVX machines with XGETBV (ECX == 1) like Skylake processors,
>
> (gdb) disass _dl_runtime_resolve_avx_opt
> Dump of assembler code for function _dl_runtime_resolve_avx_opt:
>    0x0000000000015890 <+0>:     push   %rax
>    0x0000000000015891 <+1>:     push   %rcx
>    0x0000000000015892 <+2>:     push   %rdx
>    0x0000000000015893 <+3>:     mov    $0x1,%ecx
>    0x0000000000015898 <+8>:     xgetbv
>    0x000000000001589b <+11>:    mov    %eax,%r11d
>    0x000000000001589e <+14>:    pop    %rdx
>    0x000000000001589f <+15>:    pop    %rcx
>    0x00000000000158a0 <+16>:    pop    %rax
>    0x00000000000158a1 <+17>:    and    $0x4,%r11d
>    0x00000000000158a5 <+21>:    bnd je 0x16200 <_dl_runtime_resolve_sse_vex>
> End of assembler dump.
>
> is slower than:
>
> (gdb) disass _dl_runtime_resolve_avx_slow
> Dump of assembler code for function _dl_runtime_resolve_avx_slow:
>    0x0000000000015850 <+0>:     vorpd  %ymm0,%ymm1,%ymm8
>    0x0000000000015854 <+4>:     vorpd  %ymm2,%ymm3,%ymm9
>    0x0000000000015858 <+8>:     vorpd  %ymm4,%ymm5,%ymm10
>    0x000000000001585c <+12>:    vorpd  %ymm6,%ymm7,%ymm11
>    0x0000000000015860 <+16>:    vorpd  %ymm8,%ymm9,%ymm9
>    0x0000000000015865 <+21>:    vorpd  %ymm10,%ymm11,%ymm10
>    0x000000000001586a <+26>:    vpcmpeqd %xmm8,%xmm8,%xmm8
>    0x000000000001586f <+31>:    vorpd  %ymm9,%ymm10,%ymm10
>    0x0000000000015874 <+36>:    vptest %ymm10,%ymm8
>    0x0000000000015879 <+41>:    bnd jae 0x158b0 <_dl_runtime_resolve_avx>
>    0x000000000001587c <+44>:    vzeroupper
>    0x000000000001587f <+47>:    bnd jmpq 0x16200 <_dl_runtime_resolve_sse_vex>
> End of assembler dump.
> (gdb)
>
> since xgetbv takes much more cycles than single cycle operations like
> vpord/vvpcmpeq/ptest.  _dl_runtime_resolve_opt should be used only with
> AVX512 where AVX512 instructions lead to lower CPU frequency on Skylake
> server.
>
> Any comments or objections?
>
> H.J.
> ---
>         [BZ #21871]
>         * sysdeps/x86/cpu-features.c (init_cpu_features): Set
>         bit_arch_Use_dl_runtime_resolve_opt only with AVX512F.
> ---
>  sysdeps/x86/cpu-features.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
> index 1d087ea732..6f900840d4 100644
> --- a/sysdeps/x86/cpu-features.c
> +++ b/sysdeps/x86/cpu-features.c
> @@ -244,10 +244,13 @@ init_cpu_features (struct cpu_features *cpu_features)
>           |= bit_arch_Prefer_No_AVX512;
>
>        /* To avoid SSE transition penalty, use _dl_runtime_resolve_slow.
> -         If XGETBV suports ECX == 1, use _dl_runtime_resolve_opt.  */
> +         If XGETBV suports ECX == 1, use _dl_runtime_resolve_opt.
> +        Use _dl_runtime_resolve_opt only with AVX512F since it is
> +        slower than _dl_runtime_resolve_slow with AVX.  */
>        cpu_features->feature[index_arch_Use_dl_runtime_resolve_slow]
>         |= bit_arch_Use_dl_runtime_resolve_slow;
> -      if (cpu_features->max_cpuid >= 0xd)
> +      if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable)
> +         && cpu_features->max_cpuid >= 0xd)
>         {
>           unsigned int eax;
>
> --
> 2.13.3
>

I am checking it in today and will backport it to 2.26/2.25/2.24 branches
next week.

-- 
H.J.