From 9fccf1629ea99ab2613601e2ecd48646acb1e219 Mon Sep 17 00:00:00 2001
From: cinap_lenrek <cinap_lenrek@felloff.net>
Date: Fri, 20 Nov 2015 06:25:56 +0100
Subject: libmp: add mpvecadd()/mpvecsub() assembly versions for arm

---
 sys/src/libmp/arm/mkfile           |  2 +-
 sys/src/libmp/arm/mpvecadd.s       | 32 ++++++++++++++++++++++++++++++++
 sys/src/libmp/arm/mpvecdigmuladd.s |  4 ++--
 sys/src/libmp/arm/mpvecdigmulsub.s |  4 ++--
 sys/src/libmp/arm/mpvecsub.s       | 31 +++++++++++++++++++++++++++++++
 5 files changed, 68 insertions(+), 5 deletions(-)
 create mode 100644 sys/src/libmp/arm/mpvecadd.s
 create mode 100644 sys/src/libmp/arm/mpvecsub.s

diff --git a/sys/src/libmp/arm/mkfile b/sys/src/libmp/arm/mkfile
index c939f5c41..501c5ecc4 100644
--- a/sys/src/libmp/arm/mkfile
+++ b/sys/src/libmp/arm/mkfile
@@ -2,7 +2,7 @@ objtype=arm
 </$objtype/mkfile
 
 LIB=/$objtype/lib/libmp.a
-SFILES=mpvecdigmuladd.s mpvecdigmulsub.s
+SFILES=mpvecdigmuladd.s mpvecdigmulsub.s mpvecadd.s mpvecsub.s
 
 HFILES=/$objtype/include/u.h /sys/include/mp.h ../port/dat.h
 
diff --git a/sys/src/libmp/arm/mpvecadd.s b/sys/src/libmp/arm/mpvecadd.s
new file mode 100644
index 000000000..931213407
--- /dev/null
+++ b/sys/src/libmp/arm/mpvecadd.s
@@ -0,0 +1,32 @@
+TEXT mpvecadd(SB),$0
+	MOVW	alen+4(FP), R4
+	MOVW	b+8(FP), R5
+	MOVW	blen+12(FP), R6
+	MOVW	sum+16(FP), R7
+	MOVW	$0, R8
+	MOVW	R8, R3
+	CMP	R8, R6
+	B.EQ	_add1
+	SUB	R6, R4, R4
+_addloop1:
+	MOVW.WP	4(R0), R1
+	MOVW.WP	4(R5), R2
+	CMP	$1, R3
+	ADC.S	R2, R1
+	ADC	R8, R8, R3
+	MOVW.WP	R1, 4(R7)
+	SUB.S	$1, R6
+	B.NE	_addloop1
+_add1:
+	CMP	R8, R4
+	B.EQ	_addend
+_addloop2:
+	MOVW.WP	4(R0), R1
+	ADD.S	R3, R1
+	ADC	R8, R8, R3
+	MOVW.WP	R1, 4(R7)
+	SUB.S	$1, R4
+	B.NE	_addloop2
+_addend:
+	MOVW	R3, (R7)
+	RET
diff --git a/sys/src/libmp/arm/mpvecdigmuladd.s b/sys/src/libmp/arm/mpvecdigmuladd.s
index a3833bee0..d006ff81b 100644
--- a/sys/src/libmp/arm/mpvecdigmuladd.s
+++ b/sys/src/libmp/arm/mpvecdigmuladd.s
@@ -5,12 +5,12 @@ TEXT mpvecdigmuladd(SB),$0
 	MOVW	$0, R2
 _muladdloop:
 	MOVW	$0, R1
-	MOVW.W.P 4(R0), R3
+	MOVW.WP	4(R0), R3
 	MULALU	R3, R5, (R1, R2)
  	MOVW	(R6), R7
 	ADD.S	R2, R7
 	ADC	$0, R1, R2
-	MOVW.W.P R7, 4(R6)
+	MOVW.WP	R7, 4(R6)
 	SUB.S	$1, R4
 	B.NE	_muladdloop
 	MOVW	(R6), R7
diff --git a/sys/src/libmp/arm/mpvecdigmulsub.s b/sys/src/libmp/arm/mpvecdigmulsub.s
index 7f877a64b..8340734b6 100644
--- a/sys/src/libmp/arm/mpvecdigmulsub.s
+++ b/sys/src/libmp/arm/mpvecdigmulsub.s
@@ -5,13 +5,13 @@ TEXT mpvecdigmulsub(SB),$0
 	MOVW	$0, R2
 _mulsubloop:
 	MOVW	$0, R1
-	MOVW.W.P 4(R0), R3
+	MOVW.WP	4(R0), R3
 	MULALU	R3, R5, (R1, R2)
  	MOVW	(R6), R7
 	SUB.S	R2, R7
 	ADD.CC	$1, R1
 	MOVW	R1, R2
-	MOVW.W.P R7, 4(R6)
+	MOVW.WP	R7, 4(R6)
 	SUB.S	$1, R4
 	B.NE	_mulsubloop
 	MOVW	(R6), R7
diff --git a/sys/src/libmp/arm/mpvecsub.s b/sys/src/libmp/arm/mpvecsub.s
new file mode 100644
index 000000000..bfcb124ae
--- /dev/null
+++ b/sys/src/libmp/arm/mpvecsub.s
@@ -0,0 +1,31 @@
+TEXT mpvecsub(SB),$0
+	MOVW	alen+4(FP), R4
+	MOVW	b+8(FP), R5
+	MOVW	blen+12(FP), R6
+	MOVW	diff+16(FP), R7
+	MOVW	$0, R8
+	MOVW	R8, R3
+	CMP	R8, R6
+	B.EQ	_sub1
+	SUB	R6, R4, R4
+_subloop1:
+	MOVW.WP	4(R0), R1
+	MOVW.WP	4(R5), R2
+	CMP	R3, R8
+	SBC.S	R2, R1
+	SBC	R8, R8, R3
+	MOVW.WP	R1, 4(R7)
+	SUB.S	$1, R6
+	B.NE	_subloop1
+_sub1:
+	CMP	R8, R4
+	RET.EQ
+_subloop2:
+	MOVW.WP	4(R0), R1
+	CMP	R3, R8
+	SBC.S	R8, R1
+	SBC	R8, R8, R3
+	MOVW.WP	R1, 4(R7)
+	SUB.S	$1, R4
+	B.NE	_subloop2
+	RET
-- 
cgit v1.2.3