From 571a3ddd938b742af8fc2b02f26b4b3296ea8a94 Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Wed, 2 Mar 2022 16:12:40 -0800
Subject: [PATCH] x86: Adding an upper bound for Enhanced REP MOVSB.

In the process of optimizing memcpy for AMD machines, we have found the
vector move operations are outperforming enhanced REP MOVSB for data
transfers above the L2 cache size on Zen3 architectures.
To handle this use case, we are adding an upper bound parameter on
enhanced REP MOVSB:'__x86_rep_movsb_stop_threshold'.
As per large-bench results, we are configuring this parameter to the
L2 cache size for AMD machines and applicable from Zen3 architecture
supporting the ERMS feature.
For architectures other than AMD, it is the computed value of
non-temporal threshold parameter.

Reviewed-by: Premachandra Mallappa <premachandra.mallappa@amd.com>

(cherry picked from commit 6e02b3e9327b7dbb063958d2b124b64fcb4bbe3f)
---
 sysdeps/x86/cacheinfo.h                            | 14 ++++++++++++++
 .../x86_64/multiarch/memmove-vec-unaligned-erms.S  |  7 +++++--
 2 files changed, 19 insertions(+), 2 deletions(-)

diff --git a/sysdeps/x86/cacheinfo.h b/sysdeps/x86/cacheinfo.h
index 02556961..b982982f 100644
--- a/sysdeps/x86/cacheinfo.h
+++ b/sysdeps/x86/cacheinfo.h
@@ -45,6 +45,9 @@ long int __x86_rep_movsb_threshold attribute_hidden = 2048;
 /* Threshold to use Enhanced REP STOSB.  */
 long int __x86_rep_stosb_threshold attribute_hidden = 2048;
 
+/* Threshold to stop using Enhanced REP MOVSB.  */
+long int __x86_rep_movsb_stop_threshold attribute_hidden;
+
 static void
 get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr,
 		       long int core)
@@ -352,6 +355,12 @@ init_cacheinfo (void)
 	      shared += core;
             }
 	}
+
+      /* ERMS feature is implemented from AMD Zen3 architecture and it is
+	 performing poorly for data above L2 cache size. Henceforth, adding
+	 an upper bound threshold parameter to limit the usage of Enhanced
+	 REP MOVSB operations and setting its value to L2 cache size.  */
+      __x86_rep_movsb_stop_threshold = core;
     }
 
   if (cpu_features->data_cache_size != 0)
@@ -421,6 +430,11 @@ init_cacheinfo (void)
   else
     __x86_rep_movsb_threshold = rep_movsb_threshold;
 
+  /* Setting the upper bound of ERMS to the computed value of
+     non-temporal threshold for architectures other than AMD.  */
+  if (cpu_features->basic.kind != arch_kind_amd)
+    __x86_rep_movsb_stop_threshold = __x86_shared_non_temporal_threshold;
+
 # if HAVE_TUNABLES
   __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold;
 # endif
diff --git a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
index 572cef04..620ce3a8 100644
--- a/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
+++ b/sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S
@@ -30,7 +30,10 @@
       load and aligned store.  Load the last 4 * VEC and first VEC
       before the loop and store them after the loop to support
       overlapping addresses.
-   6. If size >= __x86_shared_non_temporal_threshold and there is no
+   6. On machines with ERMS feature, if size greater than equal or to
+      __x86_rep_movsb_threshold and less than
+      __x86_rep_movsb_stop_threshold, then REP MOVSB will be used.
+   7. If size >= __x86_shared_non_temporal_threshold and there is no
       overlap between destination and source, use non-temporal store
       instead of aligned store copying from either 2 or 4 pages at
       once.
@@ -311,7 +314,7 @@ L(return):
 #endif
 
 L(movsb):
-	cmp	__x86_shared_non_temporal_threshold(%rip), %RDX_LP
+	cmp     __x86_rep_movsb_stop_threshold(%rip), %RDX_LP
 	jae	L(more_8x_vec)
 	cmpq	%rsi, %rdi
 	jb	1f
-- 
GitLab