From 6d1594e8918466be267c2d885bbf4d3c9c2846ff Mon Sep 17 00:00:00 2001
From: Taylor R Campbell <campbell+golang@mumble.net>
Date: Fri, 3 Dec 2021 12:10:15 +0000
Subject: [PATCH] runtime: Use rep stos for memclr -- no vectors or nontemporal
 stores.

---
 src/runtime/memclr_amd64.s | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)

diff --git a/src/runtime/memclr_amd64.s b/src/runtime/memclr_amd64.s
index 700bbd7b9b..8bffa7b192 100644
--- a/src/runtime/memclr_amd64.s
+++ b/src/runtime/memclr_amd64.s
@@ -17,6 +17,39 @@ TEXT runtime·memclrNoHeapPointers<ABIInternal>(SB), NOSPLIT, $0-16
 	MOVQ	AX, DI	// DI = ptr
 	XORQ	AX, AX
 
+#ifndef hack
+	CLD	// DF = 0 so STOSB moves forward
+	MOVQ	BX, CX
+	TESTQ	$0x1, CX
+	JNZ	byte
+	TESTQ	$0x3, CX
+	JNZ	word
+	TESTQ	$0x7, CX
+	JNZ	long
+	// DI = ptr
+	// CX = n
+	// AX = 0
+	// for (; CX != 0; CX--) *DI++ = AX;
+quad:
+	SHRQ	$3, CX
+	REP
+	STOSQ
+	RET
+long:
+	SHRQ	$2, CX
+	REP
+	STOSL
+	RET
+word:
+	SHRQ	$1, CX
+	REP
+	STOSW
+	RET
+byte:
+	REP
+	STOSB
+	RET
+#else
 	// MOVOU seems always faster than REP STOSQ.
 tail:
 	// BSR+branch table make almost all memmove/memclr benchmarks worse. Not worth doing.
@@ -177,3 +210,4 @@ _129through256:
 	MOVOU	X15, -32(DI)(BX*1)
 	MOVOU	X15, -16(DI)(BX*1)
 	RET
+#endif