[Gmp-commit] /var/hg/gmp: Rewrite.

[mercurial at gmplib.org]

details:   /var/hg/gmp/rev/13b6df6a8028
changeset: 15872:13b6df6a8028
user:      Torbjorn Granlund <tege at gmplib.org>
date:      Mon Jul 15 21:36:51 2013 +0200
description:
Rewrite.

diffstat:

 mpn/x86_64/core2/aors_n.asm |  84 +++++++++++++++++++++-----------------------
 1 files changed, 41 insertions(+), 43 deletions(-)

diffs (145 lines):

diff -r 78b8732e93ae -r 13b6df6a8028 mpn/x86_64/core2/aors_n.asm
--- a/mpn/x86_64/core2/aors_n.asm	Mon Jul 15 13:08:07 2013 +0200
+++ b/mpn/x86_64/core2/aors_n.asm	Mon Jul 15 21:36:51 2013 +0200
@@ -1,6 +1,6 @@
-dnl  Intel P6-15 mpn_add_n/mpn_sub_n -- mpn add or subtract.
+dnl  Intel mpn_add_n/mpn_sub_n optimised for Conroe, Nehalem.
 
-dnl  Copyright 2006, 2007, 2011, 2012 Free Software Foundation, Inc.
+dnl  Copyright 2006, 2007, 2011, 2012, 2013 Free Software Foundation, Inc.
 
 dnl  This file is part of the GNU MP Library.
 
@@ -21,14 +21,14 @@
 
 
 C	     cycles/limb
-C AMD K8,K9	 2.25
+C AMD K8,K9	 2
 C AMD K10	 2
 C Intel P4	10
-C Intel core2	 2.05
-C Intel NHM	 2.3
-C Intel SBR	 1.9
-C Intel atom	 ?
-C VIA nano	 ?
+C Intel core2	 2
+C Intel NHM	 2
+C Intel SBR	 2
+C Intel atom	 9
+C VIA nano	 3
 
 C INPUT PARAMETERS
 define(`rp',	`%rdi')
@@ -54,12 +54,6 @@
 ASM_START()
 	TEXT
 	ALIGN(16)
-PROLOGUE(func_nc)
-	FUNC_ENTRY(4)
-IFDOS(`	mov	56(%rsp), %r8	')
-	jmp	L(start)
-EPILOGUE()
-
 PROLOGUE(func)
 	FUNC_ENTRY(4)
 	xor	%r8, %r8
@@ -67,10 +61,10 @@
 	mov	(up), %r10
 	mov	(vp), %r11
 
-	lea	-8(up,n,8), up
-	lea	-8(vp,n,8), vp
-	lea	-16(rp,n,8), rp
-	mov	R32(%rcx), R32(%rax)
+	lea	(up,n,8), up
+	lea	(vp,n,8), vp
+	lea	(rp,n,8), rp
+	mov	R32(n), R32(%rax)
 	neg	n
 	and	$3, R32(%rax)
 	je	L(b00)
@@ -79,54 +73,58 @@
 	jl	L(b01)
 	je	L(b10)
 
-L(b11):	shr	%r8			C set cy
+L(b11):	neg	%r8			C set cy
 	jmp	L(e11)
 
-L(b00):	shr	%r8			C set cy
+L(b00):	neg	%r8			C set cy
 	mov	%r10, %r8
 	mov	%r11, %r9
 	lea	4(n), n
 	jmp	L(e00)
 
-L(b01):	shr	%r8			C set cy
-	jmp	L(e01)
+	nop
+	nop
+	nop
+L(b01):	neg	%r8			C set cy
+	jmp	L(top)
 
-L(b10):	shr	%r8			C set cy
+L(b10):	neg	%r8			C set cy
 	mov	%r10, %r8
 	mov	%r11, %r9
 	jmp	L(e10)
 
 L(end):	ADCSBB	%r11, %r10
-	mov	%r10, 8(rp)
+	mov	%r10, -8(rp)
 	mov	R32(%rcx), R32(%rax)	C clear eax, ecx contains 0
 	adc	R32(%rax), R32(%rax)
 	FUNC_EXIT()
 	ret
 
 	ALIGN(16)
-L(top):
-	mov	-24(up,n,8), %r8
-	mov	-24(vp,n,8), %r9
+L(top):	jrcxz	L(end)
+	mov	(up,n,8), %r8
+	mov	(vp,n,8), %r9
+	lea	4(n), n
+	ADCSBB	%r11, %r10
+	mov	%r10, -40(rp,n,8)
+L(e00):	mov	-24(up,n,8), %r10
+	mov	-24(vp,n,8), %r11
+	ADCSBB	%r9, %r8
+	mov	%r8, -32(rp,n,8)
+L(e11):	mov	-16(up,n,8), %r8
+	mov	-16(vp,n,8), %r9
 	ADCSBB	%r11, %r10
 	mov	%r10, -24(rp,n,8)
-L(e00):
-	mov	-16(up,n,8), %r10
-	mov	-16(vp,n,8), %r11
+L(e10):	mov	-8(up,n,8), %r10
+	mov	-8(vp,n,8), %r11
 	ADCSBB	%r9, %r8
 	mov	%r8, -16(rp,n,8)
-L(e11):
-	mov	-8(up,n,8), %r8
-	mov	-8(vp,n,8), %r9
-	ADCSBB	%r11, %r10
-	mov	%r10, -8(rp,n,8)
-L(e10):
-	mov	(up,n,8), %r10
-	mov	(vp,n,8), %r11
-	ADCSBB	%r9, %r8
-	mov	%r8, (rp,n,8)
-L(e01):
-	jrcxz	L(end)
-	lea	4(n), n
 	jmp	L(top)
+EPILOGUE()
 
+PROLOGUE(func_nc)
+	FUNC_ENTRY(4)
+IFDOS(`	mov	56(%rsp), %r8	')
+	jmp	L(start)
 EPILOGUE()
+