From f1be262ec11c1c35e6485f432415b5b52adb505d Mon Sep 17 00:00:00 2001
From: Mateusz Guzik <mjg@FreeBSD.org>
Date: Sun, 31 Jan 2021 16:46:18 +0100
Subject: [PATCH] amd64: move memcmp checks upfront

This is a tradeoff which saves jumps for smaller sizes while making
the 8-16 range slower (roughly in line with the other cases).

Tested with glibc test suite.

For example size 3 (most common with vfs namecache) (ops/s):
before:	407086026
after:	461391995

The regressed range of 8-16 (with 8 as example):
before:	540850489
after:	461671032
---
 lib/libc/amd64/string/memcmp.S | 50 ++++++++++++++++++--------------
 sys/amd64/amd64/support.S      | 52 +++++++++++++++++++---------------
 2 files changed, 57 insertions(+), 45 deletions(-)

diff --git a/lib/libc/amd64/string/memcmp.S b/lib/libc/amd64/string/memcmp.S
index 67c7df280679..0c8121f9d885 100644
--- a/lib/libc/amd64/string/memcmp.S
+++ b/lib/libc/amd64/string/memcmp.S
@@ -45,9 +45,25 @@ ENTRY(memcmp)
 	cmpq	$16,%rdx
 	ja	101632f
 
-100816:
 	cmpb	$8,%dl
-	jl	100408f
+	jg	100816f
+
+	cmpb	$4,%dl
+	jg	100408f
+
+	cmpb	$2,%dl
+	jge	100204f
+
+	cmpb	$1,%dl
+	jl	100000f
+	movzbl	(%rdi),%eax
+	movzbl	(%rsi),%r8d
+	subl	%r8d,%eax
+100000:
+	ret
+
+	ALIGN_TEXT
+100816:
 	movq	(%rdi),%r8
 	movq	(%rsi),%r9
 	cmpq	%r8,%r9
@@ -57,9 +73,8 @@ ENTRY(memcmp)
 	cmpq	%r8,%r9
 	jne	10081608f
 	ret
+	ALIGN_TEXT
 100408:
-	cmpb	$4,%dl
-	jl	100204f
 	movl	(%rdi),%r8d
 	movl	(%rsi),%r9d
 	cmpl	%r8d,%r9d
@@ -69,9 +84,8 @@ ENTRY(memcmp)
 	cmpl	%r8d,%r9d
 	jne	10040804f
 	ret
+	ALIGN_TEXT
 100204:
-	cmpb	$2,%dl
-	jl	100001f
 	movzwl	(%rdi),%r8d
 	movzwl	(%rsi),%r9d
 	cmpl	%r8d,%r9d
@@ -81,15 +95,7 @@ ENTRY(memcmp)
 	cmpl	%r8d,%r9d
 	jne	1f
 	ret
-100001:
-	cmpb	$1,%dl
-	jl	100000f
-	movzbl	(%rdi),%eax
-	movzbl	(%rsi),%r8d
-	subl	%r8d,%eax
-100000:
-	ret
-ALIGN_TEXT
+	ALIGN_TEXT
 101632:
 	cmpq	$32,%rdx
 	ja	103200f
@@ -110,7 +116,7 @@ ALIGN_TEXT
 	cmpq	%r8,%r9
 	jne	10163224f
 	ret
-ALIGN_TEXT
+	ALIGN_TEXT
 103200:
 	movq	(%rdi),%r8
 	movq	8(%rdi),%r9
@@ -140,7 +146,7 @@ ALIGN_TEXT
  *
  * Before we compute it we narrow down the range (16 -> 8 -> 4 bytes).
  */
-ALIGN_TEXT
+	ALIGN_TEXT
 10320016:
 	leaq	16(%rdi),%rdi
 	leaq	16(%rsi),%rsi
@@ -152,29 +158,29 @@ ALIGN_TEXT
 	leaq	8(%rdi),%rdi
 	leaq	8(%rsi),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10081608:
 10163224:
 	leaq	-8(%rdi,%rdx),%rdi
 	leaq	-8(%rsi,%rdx),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10163216:
 	leaq	-16(%rdi,%rdx),%rdi
 	leaq	-16(%rsi,%rdx),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10163208:
 	leaq	8(%rdi),%rdi
 	leaq	8(%rsi),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10040804:
 	leaq	-4(%rdi,%rdx),%rdi
 	leaq	-4(%rsi,%rdx),%rsi
 	jmp	1f
 
-ALIGN_TEXT
+	ALIGN_TEXT
 80:
 	movl	(%rdi),%r8d
 	movl	(%rsi),%r9d
diff --git a/sys/amd64/amd64/support.S b/sys/amd64/amd64/support.S
index 49baa50ac294..b623fba277db 100644
--- a/sys/amd64/amd64/support.S
+++ b/sys/amd64/amd64/support.S
@@ -93,9 +93,26 @@ ENTRY(memcmp)
 	cmpq	$16,%rdx
 	ja	101632f
 
-100816:
 	cmpb	$8,%dl
-	jl	100408f
+	jg	100816f
+
+	cmpb	$4,%dl
+	jg	100408f
+
+	cmpb	$2,%dl
+	jge	100204f
+
+	cmpb	$1,%dl
+	jl	100000f
+	movzbl	(%rdi),%eax
+	movzbl	(%rsi),%r8d
+	subl	%r8d,%eax
+100000:
+	POP_FRAME_POINTER
+	ret
+
+	ALIGN_TEXT
+100816:
 	movq	(%rdi),%r8
 	movq	(%rsi),%r9
 	cmpq	%r8,%r9
@@ -106,9 +123,8 @@ ENTRY(memcmp)
 	jne	10081608f
 	POP_FRAME_POINTER
 	ret
+	ALIGN_TEXT
 100408:
-	cmpb	$4,%dl
-	jl	100204f
 	movl	(%rdi),%r8d
 	movl	(%rsi),%r9d
 	cmpl	%r8d,%r9d
@@ -119,9 +135,8 @@ ENTRY(memcmp)
 	jne	10040804f
 	POP_FRAME_POINTER
 	ret
+	ALIGN_TEXT
 100204:
-	cmpb	$2,%dl
-	jl	100001f
 	movzwl	(%rdi),%r8d
 	movzwl	(%rsi),%r9d
 	cmpl	%r8d,%r9d
@@ -132,16 +147,7 @@ ENTRY(memcmp)
 	jne	1f
 	POP_FRAME_POINTER
 	ret
-100001:
-	cmpb	$1,%dl
-	jl	100000f
-	movzbl	(%rdi),%eax
-	movzbl	(%rsi),%r8d
-	subl	%r8d,%eax
-100000:
-	POP_FRAME_POINTER
-	ret
-ALIGN_TEXT
+	ALIGN_TEXT
 101632:
 	cmpq	$32,%rdx
 	ja	103200f
@@ -163,7 +169,7 @@ ALIGN_TEXT
 	jne	10163224f
 	POP_FRAME_POINTER
 	ret
-ALIGN_TEXT
+	ALIGN_TEXT
 103200:
 	movq	(%rdi),%r8
 	movq	8(%rdi),%r9
@@ -194,7 +200,7 @@ ALIGN_TEXT
  *
  * Before we compute it we narrow down the range (16 -> 8 -> 4 bytes).
  */
-ALIGN_TEXT
+	ALIGN_TEXT
 10320016:
 	leaq	16(%rdi),%rdi
 	leaq	16(%rsi),%rsi
@@ -206,29 +212,29 @@ ALIGN_TEXT
 	leaq	8(%rdi),%rdi
 	leaq	8(%rsi),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10081608:
 10163224:
 	leaq	-8(%rdi,%rdx),%rdi
 	leaq	-8(%rsi,%rdx),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10163216:
 	leaq	-16(%rdi,%rdx),%rdi
 	leaq	-16(%rsi,%rdx),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10163208:
 	leaq	8(%rdi),%rdi
 	leaq	8(%rsi),%rsi
 	jmp	80f
-ALIGN_TEXT
+	ALIGN_TEXT
 10040804:
 	leaq	-4(%rdi,%rdx),%rdi
 	leaq	-4(%rsi,%rdx),%rsi
 	jmp	1f
 
-ALIGN_TEXT
+	ALIGN_TEXT
 80:
 	movl	(%rdi),%r8d
 	movl	(%rsi),%r9d