https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105275
Jan Hubicka <hubicka at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
Last reconfirmed| |2025-04-26
--- Comment #7 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
I built with -march=native -Ofast -flto -fno-ipa-cp + PGO (without -flto the
problem does not reproduce).
imagick has bad train run and the hottest loop is not trained at all. Compared
to GCC15 the cold loop is now much slower:
50.26% imagick_r_peak. imagick_r_peak.trunk-pgolto-Ofast-native-m64
[.] MorphologyApply.cold
15.85% imagick_r_peak. imagick_r_peak.gcc15-pgolto-Ofast-native-m64
[.] MorphologyApply.cold
9.93% imagick_r_peak. imagick_r_peak.gcc15-pgolto-Ofast-native-m64
[.] MeanShiftImage
9.24% imagick_r_peak. imagick_r_peak.trunk-pgolto-Ofast-native-m64
[.] MeanShiftImage
4.41% imagick_r_peak. imagick_r_peak.trunk-pgolto-Ofast-native-m64
[.] GetVirtualPixelsFromNexus
3.64% imagick_r_peak. imagick_r_peak.gcc15-pgolto-Ofast-native-m64
[.] GetVirtualPixelsFromNexus
1.51% imagick_r_peak. imagick_r_peak.gcc15-pgolto-Ofast-native-m64
[.] MagickRound.lto_priv.1.lto_priv.0
1.38% imagick_r_peak. imagick_r_peak.trunk-pgolto-Ofast-native-m64
[.] MagickRound.lto_priv.1.lto_priv.0
1.21% imagick_r_peak. imagick_r_peak.trunk-pgolto-Ofast-native-m64
[.] GetOneCacheViewVirtualPixel
1.18% imagick_r_peak. imagick_r_peak.gcc15-pgolto-Ofast-native-m64
[.] GetOneCacheViewVirtualPixel
Codegen of the internal loop is same for both versions:
0.00 │2afd:┌─ cmp %rdx,%rcx ▒
0.00 │ │ jge 312f ▒
│ │result.red += (*k)*k_pixels[u].red; ▒
│ │ imul $0xfffffffffffffff8,%rcx,%r11 ▒
17.91 │ │ vmovsd (%rax,%r11,1),%xmm5 ▒
0.00 │ │ movzwl 0x4(%rbx,%rcx,8),%r11d ▒
0.00 │ │ vcvtsi2sd %r11d,%xmm6,%xmm6 ▒
│ │result.green += (*k)*k_pixels[u].green; ▒
22.46 │ │ movzwl 0x2(%rbx,%rcx,8),%r11d ▒
│ │result.red += (*k)*k_pixels[u].red; ▒
0.00 │ │ vfmadd231sd %xmm6,%xmm5,%xmm4 ▒
│ │result.green += (*k)*k_pixels[u].green; ▒
0.04 │ │ vcvtsi2sd %r11d,%xmm6,%xmm6 ▒
│ │result.blue += (*k)*k_pixels[u].blue; ▒
17.09 │ │ movzwl (%rbx,%rcx,8),%r11d ▒
│ │result.green += (*k)*k_pixels[u].green; ▒
0.00 │ │ vfmadd231sd %xmm6,%xmm5,%xmm3 ▒
│ │result.blue += (*k)*k_pixels[u].blue; ▒
12.99 │ │ vcvtsi2sd %r11d,%xmm6,%xmm6 ▒
│ │result.opacity += (*k)*k_pixels[u].opacity; ▒
0.15 │ │ movzwl 0x6(%rbx,%rcx,8),%r11d ▒
│ │result.blue += (*k)*k_pixels[u].blue; ▒
0.00 │ │ vfmadd231sd %xmm6,%xmm5,%xmm2 ▒
│ │result.opacity += (*k)*k_pixels[u].opacity; ▒
19.48 │ │ vcvtsi2sd %r11d,%xmm6,%xmm6 ▒
0.00 │ │ vfmadd231sd %xmm6,%xmm5,%xmm1 ▒
│ │if ( image->colorspace == CMYKColorspace) ▒
9.61 │ │ cmp $0xc,%esi ◆
0.00 │ │ je 317c ▒
│ │for (u=0; u < (ssize_t) kernel->width; u++, k--) { ▒
│2b58:│ inc %rcx ▒
0.00 │ └──jmp 2afd ▒
0.02 │2b34:┌─ cmp %rdx,%rcx ▒
3.62 │ │ jge 3166 ▒
│ │result.red += (*k)*k_pixels[u].red; ▒
│ │ imul $0xfffffffffffffff8,%rcx,%r11 ▒
0.19 │ │ vmovsd (%rax,%r11,1),%xmm5 ▒
8.36 │ │ movzwl 0x4(%rbx,%rcx,8),%r11d ▒
0.00 │ │ vcvtsi2sd %r11d,%xmm12,%xmm6 ▒
│ │result.green += (*k)*k_pixels[u].green; ▒
7.13 │ │ movzwl 0x2(%rbx,%rcx,8),%r11d ▒
│ │result.red += (*k)*k_pixels[u].red; ▒
19.29 │ │ vfmadd231sd %xmm6,%xmm5,%xmm4 ▒
│ │result.green += (*k)*k_pixels[u].green; ▒
0.10 │ │ vcvtsi2sd %r11d,%xmm12,%xmm6 ▒
│ │result.blue += (*k)*k_pixels[u].blue; ▒
4.76 │ │ movzwl (%rbx,%rcx,8),%r11d ▒
│ │result.green += (*k)*k_pixels[u].green; ▒
13.90 │ │ vfmadd231sd %xmm6,%xmm5,%xmm3 ▒
│ │result.blue += (*k)*k_pixels[u].blue; ▒
0.03 │ │ vcvtsi2sd %r11d,%xmm12,%xmm6 ▒
│ │result.opacity += (*k)*k_pixels[u].opacity; ▒
7.87 │ │ movzwl 0x6(%rbx,%rcx,8),%r11d ▒
│ │result.blue += (*k)*k_pixels[u].blue; ▒
11.34 │ │ vfmadd231sd %xmm6,%xmm5,%xmm2 ▒
│ │result.opacity += (*k)*k_pixels[u].opacity; ▒
0.07 │ │ vcvtsi2sd %r11d,%xmm12,%xmm6 ▒
4.52 │ │ vfmadd231sd %xmm6,%xmm5,%xmm1 ▒
│ │if ( image->colorspace == CMYKColorspace) ◆
7.03 │ │ cmp $0xc,%esi ▒
0.01 │ │ je 31b3 ▒
│ │for (u=0; u < (ssize_t) kernel->width; u++, k--) { ▒
│2b8f:│ inc %rcx ▒
11.13 │ └──jmp 2b34 ▒
