From: Andre Ramnitz <andre.ramnitz@mail.de>
Date: Thu, 17 Sep 2026 18:02:20 +0000 (+0200)
Subject: sys-devel/gcc: znver5 tune branch misprediction
X-Git-Url: https://git.ramnitz.eu/?a=commitdiff_plain;p=etc_portage_patches.git

sys-devel/gcc: znver5 tune branch misprediction
---

diff --git a/sys-devel/gcc/patch_x86_increase_znver4_5_tune_branch_misprediction_cost.patch b/sys-devel/gcc/patch_x86_increase_znver4_5_tune_branch_misprediction_cost.patch
new file mode 100644
index 0000000..fe1503e
--- /dev/null
+++ b/sys-devel/gcc/patch_x86_increase_znver4_5_tune_branch_misprediction_cost.patch
@@ -0,0 +1,47 @@
+Increase the branch misprediction scale for znver4/5 tuning from
+COSTS_N_INSNS (2) to COSTS_N_INSNS (2) + 3.
+
+This improves 544.nab_r by 12% (-O3 -march=native -flto) on Znver5
+and 9% on Znver4 with single-copy.
+
+gcc/ChangeLog:
+
+	* config/i386/x86-tune-costs.h (znver4_cost): Increase branch
+	mispredict scale from COSTS_N_INSNS (2) to COSTS_N_INSNS (2) + 3.
+	(znver5_cost): Likewise.
+---
+
+Hi Maintainers,
+
+Bootstrapped and tested on x86_64-linux, Ok for staging?
+
+regards,
+Venkat.
+
+ gcc/config/i386/x86-tune-costs.h | 4 ++--
+ 1 file changed, 2 insertions(+), 2 deletions(-)
+
+diff --git a/gcc/config/i386/x86-tune-costs.h b/gcc/config/i386/x86-tune-costs.h
+index bc3bb694349..8be780d73da 100644
+--- a/gcc/config/i386/x86-tune-costs.h
++++ b/gcc/config/i386/x86-tune-costs.h
+@@ -2231,7 +2231,7 @@ struct processor_costs znver4_cost = {
+   "16",					/* Func alignment.  */
+   4,					/* Small unroll limit.  */
+   2,					/* Small unroll factor.  */
+-  COSTS_N_INSNS (2),			/* Branch mispredict scale.  */
++  COSTS_N_INSNS (2) + 3,		/* Branch mispredict scale.  */
+ };
+ 
+ /* This table currently replicates znver4_cost table. */
+@@ -2399,7 +2399,7 @@ struct processor_costs znver5_cost = {
+   "16",					/* Func alignment.  */
+   4,					/* Small unroll limit.  */
+   2,					/* Small unroll factor.  */
+-  COSTS_N_INSNS (2),			/* Branch mispredict scale.  */
++  COSTS_N_INSNS (2) + 3,		/* Branch mispredict scale.  */
+ };
+ 
+ /* skylake_cost should produce code tuned for Skylake family of CPUs.  */
+-- 
+2.34.1