Generalise new power6 addmul_1 to support also submul_1.

2 files changed, 53 insertions, 28 deletions

diff --git a/ChangeLog b/ChangeLog
index 1d6a44512..80e0f7a32 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,12 +1,12 @@
 2011-11-22  Torbjorn Granlund  <tege@gmplib.org>
 
+	* mpn/powerpc64/mode64/p6/aorsmul_1.asm: New file.
+
 	* configure.in: Don't fail fat builds under 64-bit DOS.
 
 	* mpn/powerpc64/mode64/aors_n.asm: Align loop for slightly better
 	power5 performance.
 
-	* mpn/powerpc64/mode64/p6/addmul_1.asm: New file.
-
 2011-11-21  Torbjorn Granlund  <tege@gmplib.org>
 
 	* gmp-h.in (__GNU_MP_RELEASE): Renamed from typo name.
diff --git a/mpn/powerpc64/mode64/p6/addmul_1.asm b/mpn/powerpc64/mode64/p6/aorsmul_1.asm
index bffa6f308..4bd508488 100644
--- a/mpn/powerpc64/mode64/p6/addmul_1.asm
+++ b/mpn/powerpc64/mode64/p6/aorsmul_1.asm
@@ -1,4 +1,4 @@
-dnl  PowerPC-64 mpn_addmul_1 optimised for power6.
+dnl  PowerPC-64 mpn_addmul_1 and mpn_submul_1 optimised for power6.
 
 dnl  Copyright 1999, 2000, 2001, 2003, 2004, 2005, 2006, 2008, 2010, 2011
 dnl  Free Software Foundation, Inc.
@@ -20,19 +20,20 @@ dnl  along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.
 
 include(`../config.m4')
 
-C		    cycles/limb
-C POWER3/PPC630		 ?
-C POWER4/PPC970		 ?
-C POWER5		 ?
-C POWER6		12.25
-C POWER7		 ?
+C               mpn_addmul_1    mpn_submul_1
+C               cycles/limb     cycles/limb
+C POWER3/PPC630     ?               ?
+C POWER4/PPC970     ?               ?
+C POWER5            ?               ?
+C POWER6           12.25           12.8
+C POWER7            ?               ?
 
 C TODO
 C  * Reduce register usage.
 C  * Schedule function entry code.
 C  * Unroll more.  8-way unrolling would bring us to 10 c/l, 16-way unrolling
 C    would bring us to 9 c/l.
-C  * Generalise to handle submul_1.
+C  * Handle n = 1 and perhaps n = 2 seperately, without saving any registers.
 
 C INPUT PARAMETERS
 define(`rp',  `r3')
@@ -40,8 +41,27 @@ define(`up',  `r4')
 define(`n',   `r5')
 define(`v0',  `r6')
 
+ifdef(`OPERATION_addmul_1',`
+  define(ADDSUBC,	adde)
+  define(ADDSUB,	addc)
+  define(func,		mpn_addmul_1)
+  define(func_nc,	mpn_addmul_1c)	C FIXME: not really supported
+  define(AM,		`$1')
+  define(SM,		`')
+  define(CLRRSC,	`addic	$1, r0, 0')
+')
+ifdef(`OPERATION_submul_1',`
+  define(ADDSUBC,	subfe)
+  define(ADDSUB,	subfc)
+  define(func,		mpn_submul_1)
+  define(func_nc,	mpn_submul_1c)	C FIXME: not really supported
+  define(AM,		`')
+  define(SM,		`$1')
+  define(CLRRSC,	`subfc	$1, r0, r0')
+')
+
 ASM_START()
-PROLOGUE(mpn_addmul_1)
+PROLOGUE(func)
 	std	r31, -8(r1)
 	std	r30, -16(r1)
 	std	r29, -24(r1)
@@ -57,11 +77,11 @@ PROLOGUE(mpn_addmul_1)
 	blt	cr6, L(b1)
 	beq	cr6, L(b2)
 
-L(b3):	addi	up, up, 16
+L(b3):	ld	r8, 0(up)
+	ld	r7, 8(up)
+	ld	r27, 16(up)
+	addi	up, up, 16
 	addi	rp, rp, 16
-	ld	r8, -16(up)
-	ld	r7, -8(up)
-	ld	r27, 0(up)
 	mulld	r5,  r8, v0
 	mulhdu	r8,  r8, v0
 	mulld	r9,  r7, v0
@@ -74,13 +94,13 @@ L(b3):	addi	up, up, 16
 	addc	r9, r9, r8
 	adde	r11, r11, r7
 	addze	r12, r27
-	addc	r5, r5, r29
+	ADDSUB	r5, r5, r29
 	b	L(l3)
 
-L(b2):	addi	up, up, 8
+L(b2):	ld	r7, 0(up)
+	ld	r27, 8(up)
+	addi	up, up, 8
 	addi	rp, rp, 8
-	ld	r7, -8(up)
-	ld	r27, 0(up)
 	mulld	r9,  r7, v0
 	mulhdu	r7,  r7, v0
 	mulld	r11, r27, v0
@@ -89,22 +109,25 @@ L(b2):	addi	up, up, 8
 	ld	r31, 0(rp)
 	addc	r11, r11, r7
 	addze	r12, r27
-	addc	r9, r9, r30
+	ADDSUB	r9, r9, r30
 	b	L(l2)
 
 L(b1):	ld	r27, 0(up)
 	ld	r31, 0(rp)
 	mulld	r11, r27, v0
 	mulhdu	r12, r27, v0
-	addc	r11, r11, r31
+	ADDSUB	r11, r11, r31
 	b	L(l1)
 
 L(b0):	addi	up, up, -8
 	addi	rp, rp, -8
-	addic	r12, r0, 0	C clear r12 and cy (use that r0 = 0)
+	CLRRSC(	r12)		C clear r12 and clr/set cy
 
 	ALIGN(32)
-L(top):	ld	r10, 8(up)
+L(top):
+SM(`	subfe	r11, r0, r0')	C complement...
+SM(`	addic	r11, r11, 1')	C ...carry flag
+	ld	r10, 8(up)
 	ld	r8, 16(up)
 	ld	r7, 24(up)
 	ld	r27, 32(up)
@@ -127,18 +150,20 @@ L(top):	ld	r10, 8(up)
 	adde	r9, r9, r8
 	adde	r11, r11, r7
 	addze	r12, r27
-	addc	r0, r0, r28
+	ADDSUB	r0, r0, r28
 	std	r0, -24(rp)
-	adde	r5, r5, r29
+	ADDSUBC	r5, r5, r29
 L(l3):	std	r5, -16(rp)
-	adde	r9, r9, r30
+	ADDSUBC	r9, r9, r30
 L(l2):	std	r9, -8(rp)
-	adde	r11, r11, r31
+	ADDSUBC	r11, r11, r31
 L(l1):	std	r11, 0(rp)
 	bdnz	L(top)
 
-	addze	r3, r12
+AM(`	addze	r3, r12')
+SM(`	subfe	r11, r0, r0')		C complement...
 	ld	r31, -8(r1)
+SM(`	subf	r3, r11, r12')
 	ld	r30, -16(r1)
 	ld	r29, -24(r1)
 	ld	r28, -32(r1)