|
|
08c3a6 |
commit 6e008c884dad5a25f91085c68d044bb5e2d63761
|
|
|
08c3a6 |
Author: Noah Goldstein <goldstein.w.n@gmail.com>
|
|
|
08c3a6 |
Date: Tue Jun 14 13:50:11 2022 -0700
|
|
|
08c3a6 |
|
|
|
08c3a6 |
x86: Fix misordered logic for setting `rep_movsb_stop_threshold`
|
|
|
08c3a6 |
|
|
|
08c3a6 |
Move the setting of `rep_movsb_stop_threshold` to after the tunables
|
|
|
08c3a6 |
have been collected so that the `rep_movsb_stop_threshold` (which
|
|
|
08c3a6 |
is used to redirect control flow to the non_temporal case) will
|
|
|
08c3a6 |
use any user value for `non_temporal_threshold` (set using
|
|
|
08c3a6 |
glibc.cpu.x86_non_temporal_threshold)
|
|
|
08c3a6 |
|
|
|
08c3a6 |
(cherry picked from commit 035591551400cfc810b07244a015c9411e8bff7c)
|
|
|
08c3a6 |
|
|
|
08c3a6 |
diff --git a/sysdeps/x86/dl-cacheinfo.h b/sysdeps/x86/dl-cacheinfo.h
|
|
|
08c3a6 |
index 2e43e67e4f4037d3..560bf260e8fbd7bf 100644
|
|
|
08c3a6 |
--- a/sysdeps/x86/dl-cacheinfo.h
|
|
|
08c3a6 |
+++ b/sysdeps/x86/dl-cacheinfo.h
|
|
|
08c3a6 |
@@ -898,18 +898,6 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
|
|
|
08c3a6 |
if (CPU_FEATURE_USABLE_P (cpu_features, FSRM))
|
|
|
08c3a6 |
rep_movsb_threshold = 2112;
|
|
|
08c3a6 |
|
|
|
08c3a6 |
- unsigned long int rep_movsb_stop_threshold;
|
|
|
08c3a6 |
- /* ERMS feature is implemented from AMD Zen3 architecture and it is
|
|
|
08c3a6 |
- performing poorly for data above L2 cache size. Henceforth, adding
|
|
|
08c3a6 |
- an upper bound threshold parameter to limit the usage of Enhanced
|
|
|
08c3a6 |
- REP MOVSB operations and setting its value to L2 cache size. */
|
|
|
08c3a6 |
- if (cpu_features->basic.kind == arch_kind_amd)
|
|
|
08c3a6 |
- rep_movsb_stop_threshold = core;
|
|
|
08c3a6 |
- /* Setting the upper bound of ERMS to the computed value of
|
|
|
08c3a6 |
- non-temporal threshold for architectures other than AMD. */
|
|
|
08c3a6 |
- else
|
|
|
08c3a6 |
- rep_movsb_stop_threshold = non_temporal_threshold;
|
|
|
08c3a6 |
-
|
|
|
08c3a6 |
/* The default threshold to use Enhanced REP STOSB. */
|
|
|
08c3a6 |
unsigned long int rep_stosb_threshold = 2048;
|
|
|
08c3a6 |
|
|
|
08c3a6 |
@@ -951,6 +939,18 @@ dl_init_cacheinfo (struct cpu_features *cpu_features)
|
|
|
08c3a6 |
SIZE_MAX);
|
|
|
08c3a6 |
#endif
|
|
|
08c3a6 |
|
|
|
08c3a6 |
+ unsigned long int rep_movsb_stop_threshold;
|
|
|
08c3a6 |
+ /* ERMS feature is implemented from AMD Zen3 architecture and it is
|
|
|
08c3a6 |
+ performing poorly for data above L2 cache size. Henceforth, adding
|
|
|
08c3a6 |
+ an upper bound threshold parameter to limit the usage of Enhanced
|
|
|
08c3a6 |
+ REP MOVSB operations and setting its value to L2 cache size. */
|
|
|
08c3a6 |
+ if (cpu_features->basic.kind == arch_kind_amd)
|
|
|
08c3a6 |
+ rep_movsb_stop_threshold = core;
|
|
|
08c3a6 |
+ /* Setting the upper bound of ERMS to the computed value of
|
|
|
08c3a6 |
+ non-temporal threshold for architectures other than AMD. */
|
|
|
08c3a6 |
+ else
|
|
|
08c3a6 |
+ rep_movsb_stop_threshold = non_temporal_threshold;
|
|
|
08c3a6 |
+
|
|
|
08c3a6 |
cpu_features->data_cache_size = data;
|
|
|
08c3a6 |
cpu_features->shared_cache_size = shared;
|
|
|
08c3a6 |
cpu_features->non_temporal_threshold = non_temporal_threshold;
|