https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115845
Bug ID: 115845
Summary: 25% runtime regression of 527.cam4_r when enabling
--param vect-partial-vector-usage={1,2} ontop of
-Ofast --march=znver4
Product: gcc
Version: 14.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
There's a lot of cases like below
│7c40:┌─ lea (%rdi,%rax,1),%rbx
▒
│ │tau_w_f(1:ncol,1:pver,:) = tau_w_f(1:ncol,1:pver,:) +
twf(1:ncol,:,:) ▒
68 │ │ vmovupd (%rsi,%rax,1),%zmm14{%k1}
▒
│ │ add $0x40,%rax
▒
│ │ vmovupd (%rbx),%zmm17{%k1}
▒
11165 │ │ vaddpd %zmm17,%zmm14,%zmm20
▒
864 │ │ vmovupd %zmm20,(%rbx){%k1}
◆
│ │ mov %r11d,%ebx
▒
│ │ vpbroadcastw %r11d,%xmm20
▒
5 │ │ sub $0x8,%r11d
▒
│ │ add $0x8,%ebx
▒
│ │ vpcmpnleuw %xmm1,%xmm20,%k1
▒
1 │ │ cmp $0x8,%bx
▒
89 │ └──ja 7c40
resulting in
Samples: 1M of event 'cycles:u', Event count (approx.): 1356741812802 2
Overhead Samples Command Shared Object
Symbol
7.02% 79632 cam4_r_peak.gcc cam4_r_peak.gcc7-m64 [.]
__aer_rad_props_MOD_aer_rad_props_sw ◆
3.96% 43265 cam4_r_peak.gcc cam4_r_peak.gcc7-m64 [.]
__tracer_data_MOD_interpolate_trcdata.constprop.0 ▒
3.09% 34998 cam4_r_peak.gcc cam4_r_peak.gcc7-m64 [.]
__radsw_MOD_radcswmx ▒
2.94% 32597 cam4_r_base.gcc libm-2.31.so [.]
__ieee754_log_fma ▒
2.70% 30823 cam4_r_peak.gcc libm-2.31.so [.]
__ieee754_log_fma ▒
2.68% 29978 cam4_r_base.gcc cam4_r_base.gcc7-m64 [.]
__radsw_MOD_radcswmx ▒
2.28% 24998 cam4_r_base.gcc cam4_r_base.gcc7-m64 [.]
__tracer_data_MOD_interpolate_trcdata.constprop.0 ▒
2.21% 25215 cam4_r_peak.gcc cam4_r_peak.gcc7-m64 [.]
__radae_MOD_radabs ▒
2.07% 22878 cam4_r_base.gcc cam4_r_base.gcc7-m64 [.]
__radae_MOD_radabs ▒
1.77% 20098 cam4_r_peak.gcc cam4_r_peak.gcc7-m64 [.]
__zm_conv_MOD_ientropy.isra.0 ▒
1.62% 18145 cam4_r_base.gcc cam4_r_base.gcc7-m64 [.]
__aer_rad_props_MOD_aer_rad_props_sw
(topmost and bottom most entries are the same functions peak/base).
It almost feels like fault suppression kicking in, but on loads?!
