ARM assembler pack: reschedule instructions for dual-issue pipeline.

Modest improvement coefficients mean that code already had some
parallelism and there was not very much room for improvement. Special
thanks to Ted Krovetz for benchmarking the code with such patience.

crypto/aes/asm/aes-armv4.pl

@@ -16,12 +16,17 @@
 # allows to merge logical or arithmetic operation with shift or rotate
 # in one instruction and emit combined result every cycle. The module
 # is endian-neutral. The performance is ~42 cycles/byte for 128-bit
-# key.
+# key [on single-issue Xscale PXA250 core].
 
 # May 2007.
 #
 # AES_set_[en|de]crypt_key is added.
 
+# July 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 12% improvement on
+# Cortex A8 core and ~25 cycles per byte processed with 128-bit key.
+
 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
@@ -167,24 +172,24 @@ AES_encrypt:
 	ldrb	$t2,[$rounds,#1]
 	ldrb	$t3,[$rounds,#0]
 	orr	$s0,$s0,$t1,lsl#8
-	orr	$s0,$s0,$t2,lsl#16
-	orr	$s0,$s0,$t3,lsl#24
 	ldrb	$s1,[$rounds,#7]
+	orr	$s0,$s0,$t2,lsl#16
 	ldrb	$t1,[$rounds,#6]
+	orr	$s0,$s0,$t3,lsl#24
 	ldrb	$t2,[$rounds,#5]
 	ldrb	$t3,[$rounds,#4]
 	orr	$s1,$s1,$t1,lsl#8
-	orr	$s1,$s1,$t2,lsl#16
-	orr	$s1,$s1,$t3,lsl#24
 	ldrb	$s2,[$rounds,#11]
+	orr	$s1,$s1,$t2,lsl#16
 	ldrb	$t1,[$rounds,#10]
+	orr	$s1,$s1,$t3,lsl#24
 	ldrb	$t2,[$rounds,#9]
 	ldrb	$t3,[$rounds,#8]
 	orr	$s2,$s2,$t1,lsl#8
-	orr	$s2,$s2,$t2,lsl#16
-	orr	$s2,$s2,$t3,lsl#24
 	ldrb	$s3,[$rounds,#15]
+	orr	$s2,$s2,$t2,lsl#16
 	ldrb	$t1,[$rounds,#14]
+	orr	$s2,$s2,$t3,lsl#24
 	ldrb	$t2,[$rounds,#13]
 	ldrb	$t3,[$rounds,#12]
 	orr	$s3,$s3,$t1,lsl#8
@@ -199,24 +204,24 @@ AES_encrypt:
 	mov	$t3,$s0,lsr#8
 	strb	$t1,[$rounds,#0]
 	strb	$t2,[$rounds,#1]
-	strb	$t3,[$rounds,#2]
-	strb	$s0,[$rounds,#3]
 	mov	$t1,$s1,lsr#24
+	strb	$t3,[$rounds,#2]
 	mov	$t2,$s1,lsr#16
+	strb	$s0,[$rounds,#3]
 	mov	$t3,$s1,lsr#8
 	strb	$t1,[$rounds,#4]
 	strb	$t2,[$rounds,#5]
-	strb	$t3,[$rounds,#6]
-	strb	$s1,[$rounds,#7]
 	mov	$t1,$s2,lsr#24
+	strb	$t3,[$rounds,#6]
 	mov	$t2,$s2,lsr#16
+	strb	$s1,[$rounds,#7]
 	mov	$t3,$s2,lsr#8
 	strb	$t1,[$rounds,#8]
 	strb	$t2,[$rounds,#9]
-	strb	$t3,[$rounds,#10]
-	strb	$s2,[$rounds,#11]
 	mov	$t1,$s3,lsr#24
+	strb	$t3,[$rounds,#10]
 	mov	$t2,$s3,lsr#16
+	strb	$s2,[$rounds,#11]
 	mov	$t3,$s3,lsr#8
 	strb	$t1,[$rounds,#12]
 	strb	$t2,[$rounds,#13]
@@ -233,141 +238,137 @@ AES_encrypt:
 .align	2
 _armv4_AES_encrypt:
 	str	lr,[sp,#-4]!		@ push lr
-	ldr	$t1,[$key],#16
-	ldr	$t2,[$key,#-12]
-	ldr	$t3,[$key,#-8]
-	ldr	$i1,[$key,#-4]
-	ldr	$rounds,[$key,#240-16]
+	ldmia	$key!,{$t1-$i1}
 	eor	$s0,$s0,$t1
+	ldr	$rounds,[$key,#240-16]
 	eor	$s1,$s1,$t2
 	eor	$s2,$s2,$t3
 	eor	$s3,$s3,$i1
 	sub	$rounds,$rounds,#1
 	mov	lr,#255
 
-.Lenc_loop:
+	and	$i1,lr,$s0
 	and	$i2,lr,$s0,lsr#8
 	and	$i3,lr,$s0,lsr#16
-	and	$i1,lr,$s0
 	mov	$s0,$s0,lsr#24
+.Lenc_loop:
 	ldr	$t1,[$tbl,$i1,lsl#2]	@ Te3[s0>>0]
-	ldr	$s0,[$tbl,$s0,lsl#2]	@ Te0[s0>>24]
-	ldr	$t2,[$tbl,$i2,lsl#2]	@ Te2[s0>>8]
-	ldr	$t3,[$tbl,$i3,lsl#2]	@ Te1[s0>>16]
-
 	and	$i1,lr,$s1,lsr#16	@ i0
+	ldr	$t2,[$tbl,$i2,lsl#2]	@ Te2[s0>>8]
 	and	$i2,lr,$s1
+	ldr	$t3,[$tbl,$i3,lsl#2]	@ Te1[s0>>16]
 	and	$i3,lr,$s1,lsr#8
+	ldr	$s0,[$tbl,$s0,lsl#2]	@ Te0[s0>>24]
 	mov	$s1,$s1,lsr#24
+
 	ldr	$i1,[$tbl,$i1,lsl#2]	@ Te1[s1>>16]
-	ldr	$s1,[$tbl,$s1,lsl#2]	@ Te0[s1>>24]
 	ldr	$i2,[$tbl,$i2,lsl#2]	@ Te3[s1>>0]
 	ldr	$i3,[$tbl,$i3,lsl#2]	@ Te2[s1>>8]
 	eor	$s0,$s0,$i1,ror#8
-	eor	$s1,$s1,$t1,ror#24
-	eor	$t2,$t2,$i2,ror#8
-	eor	$t3,$t3,$i3,ror#8
-
+	ldr	$s1,[$tbl,$s1,lsl#2]	@ Te0[s1>>24]
 	and	$i1,lr,$s2,lsr#8	@ i0
+	eor	$t2,$t2,$i2,ror#8
 	and	$i2,lr,$s2,lsr#16	@ i1
+	eor	$t3,$t3,$i3,ror#8
 	and	$i3,lr,$s2
-	mov	$s2,$s2,lsr#24
+	eor	$s1,$s1,$t1,ror#24
 	ldr	$i1,[$tbl,$i1,lsl#2]	@ Te2[s2>>8]
+	mov	$s2,$s2,lsr#24
+
 	ldr	$i2,[$tbl,$i2,lsl#2]	@ Te1[s2>>16]
-	ldr	$s2,[$tbl,$s2,lsl#2]	@ Te0[s2>>24]
 	ldr	$i3,[$tbl,$i3,lsl#2]	@ Te3[s2>>0]
 	eor	$s0,$s0,$i1,ror#16
-	eor	$s1,$s1,$i2,ror#8
-	eor	$s2,$s2,$t2,ror#16
-	eor	$t3,$t3,$i3,ror#16
-
+	ldr	$s2,[$tbl,$s2,lsl#2]	@ Te0[s2>>24]
 	and	$i1,lr,$s3		@ i0
+	eor	$s1,$s1,$i2,ror#8
 	and	$i2,lr,$s3,lsr#8	@ i1
+	eor	$t3,$t3,$i3,ror#16
 	and	$i3,lr,$s3,lsr#16	@ i2
-	mov	$s3,$s3,lsr#24
+	eor	$s2,$s2,$t2,ror#16
 	ldr	$i1,[$tbl,$i1,lsl#2]	@ Te3[s3>>0]
+	mov	$s3,$s3,lsr#24
+
 	ldr	$i2,[$tbl,$i2,lsl#2]	@ Te2[s3>>8]
 	ldr	$i3,[$tbl,$i3,lsl#2]	@ Te1[s3>>16]
-	ldr	$s3,[$tbl,$s3,lsl#2]	@ Te0[s3>>24]
 	eor	$s0,$s0,$i1,ror#24
+	ldr	$s3,[$tbl,$s3,lsl#2]	@ Te0[s3>>24]
 	eor	$s1,$s1,$i2,ror#16
+	ldr	$i1,[$key],#16
 	eor	$s2,$s2,$i3,ror#8
+	ldr	$t1,[$key,#-12]
 	eor	$s3,$s3,$t3,ror#8
 
-	ldr	$t1,[$key],#16
-	ldr	$t2,[$key,#-12]
-	ldr	$t3,[$key,#-8]
-	ldr	$i1,[$key,#-4]
-	eor	$s0,$s0,$t1
-	eor	$s1,$s1,$t2
-	eor	$s2,$s2,$t3
-	eor	$s3,$s3,$i1
+	ldr	$t2,[$key,#-8]
+	eor	$s0,$s0,$i1
+	ldr	$t3,[$key,#-4]
+	and	$i1,lr,$s0
+	eor	$s1,$s1,$t1
+	and	$i2,lr,$s0,lsr#8
+	eor	$s2,$s2,$t2
+	and	$i3,lr,$s0,lsr#16
+	eor	$s3,$s3,$t3
+	mov	$s0,$s0,lsr#24
 
 	subs	$rounds,$rounds,#1
 	bne	.Lenc_loop
 
 	add	$tbl,$tbl,#2
 
-	and	$i1,lr,$s0
-	and	$i2,lr,$s0,lsr#8
-	and	$i3,lr,$s0,lsr#16
-	mov	$s0,$s0,lsr#24
 	ldrb	$t1,[$tbl,$i1,lsl#2]	@ Te4[s0>>0]
-	ldrb	$s0,[$tbl,$s0,lsl#2]	@ Te4[s0>>24]
-	ldrb	$t2,[$tbl,$i2,lsl#2]	@ Te4[s0>>8]
-	ldrb	$t3,[$tbl,$i3,lsl#2]	@ Te4[s0>>16]
-
 	and	$i1,lr,$s1,lsr#16	@ i0
+	ldrb	$t2,[$tbl,$i2,lsl#2]	@ Te4[s0>>8]
 	and	$i2,lr,$s1
+	ldrb	$t3,[$tbl,$i3,lsl#2]	@ Te4[s0>>16]
 	and	$i3,lr,$s1,lsr#8
+	ldrb	$s0,[$tbl,$s0,lsl#2]	@ Te4[s0>>24]
 	mov	$s1,$s1,lsr#24
+
 	ldrb	$i1,[$tbl,$i1,lsl#2]	@ Te4[s1>>16]
-	ldrb	$s1,[$tbl,$s1,lsl#2]	@ Te4[s1>>24]
 	ldrb	$i2,[$tbl,$i2,lsl#2]	@ Te4[s1>>0]
 	ldrb	$i3,[$tbl,$i3,lsl#2]	@ Te4[s1>>8]
 	eor	$s0,$i1,$s0,lsl#8
-	eor	$s1,$t1,$s1,lsl#24
-	eor	$t2,$i2,$t2,lsl#8
-	eor	$t3,$i3,$t3,lsl#8
-
+	ldrb	$s1,[$tbl,$s1,lsl#2]	@ Te4[s1>>24]
 	and	$i1,lr,$s2,lsr#8	@ i0
+	eor	$t2,$i2,$t2,lsl#8
 	and	$i2,lr,$s2,lsr#16	@ i1
+	eor	$t3,$i3,$t3,lsl#8
 	and	$i3,lr,$s2
-	mov	$s2,$s2,lsr#24
+	eor	$s1,$t1,$s1,lsl#24
 	ldrb	$i1,[$tbl,$i1,lsl#2]	@ Te4[s2>>8]
+	mov	$s2,$s2,lsr#24
+
 	ldrb	$i2,[$tbl,$i2,lsl#2]	@ Te4[s2>>16]
-	ldrb	$s2,[$tbl,$s2,lsl#2]	@ Te4[s2>>24]
 	ldrb	$i3,[$tbl,$i3,lsl#2]	@ Te4[s2>>0]
 	eor	$s0,$i1,$s0,lsl#8
-	eor	$s1,$s1,$i2,lsl#16
-	eor	$s2,$t2,$s2,lsl#24
-	eor	$t3,$i3,$t3,lsl#8
-
+	ldrb	$s2,[$tbl,$s2,lsl#2]	@ Te4[s2>>24]
 	and	$i1,lr,$s3		@ i0
+	eor	$s1,$s1,$i2,lsl#16
 	and	$i2,lr,$s3,lsr#8	@ i1
+	eor	$t3,$i3,$t3,lsl#8
 	and	$i3,lr,$s3,lsr#16	@ i2
-	mov	$s3,$s3,lsr#24
+	eor	$s2,$t2,$s2,lsl#24
 	ldrb	$i1,[$tbl,$i1,lsl#2]	@ Te4[s3>>0]
+	mov	$s3,$s3,lsr#24
+
 	ldrb	$i2,[$tbl,$i2,lsl#2]	@ Te4[s3>>8]
 	ldrb	$i3,[$tbl,$i3,lsl#2]	@ Te4[s3>>16]
-	ldrb	$s3,[$tbl,$s3,lsl#2]	@ Te4[s3>>24]
 	eor	$s0,$i1,$s0,lsl#8
+	ldrb	$s3,[$tbl,$s3,lsl#2]	@ Te4[s3>>24]
+	ldr	$i1,[$key,#0]
 	eor	$s1,$s1,$i2,lsl#8
+	ldr	$t1,[$key,#4]
 	eor	$s2,$s2,$i3,lsl#16
+	ldr	$t2,[$key,#8]
 	eor	$s3,$t3,$s3,lsl#24
+	ldr	$t3,[$key,#12]
 
-	ldr	lr,[sp],#4		@ pop lr
-	ldr	$t1,[$key,#0]
-	ldr	$t2,[$key,#4]
-	ldr	$t3,[$key,#8]
-	ldr	$i1,[$key,#12]
-	eor	$s0,$s0,$t1
-	eor	$s1,$s1,$t2
-	eor	$s2,$s2,$t3
-	eor	$s3,$s3,$i1
+	eor	$s0,$s0,$i1
+	eor	$s1,$s1,$t1
+	eor	$s2,$s2,$t2
+	eor	$s3,$s3,$t3
 
 	sub	$tbl,$tbl,#2
-	mov	pc,lr			@ return
+	ldr	pc,[sp],#4		@ pop and return
 .size	_armv4_AES_encrypt,.-_armv4_AES_encrypt
 
 .global AES_set_encrypt_key
@@ -402,31 +403,31 @@ AES_set_encrypt_key:
 	ldrb	$t2,[$rounds,#1]
 	ldrb	$t3,[$rounds,#0]
 	orr	$s0,$s0,$t1,lsl#8
-	orr	$s0,$s0,$t2,lsl#16
-	orr	$s0,$s0,$t3,lsl#24
 	ldrb	$s1,[$rounds,#7]
+	orr	$s0,$s0,$t2,lsl#16
 	ldrb	$t1,[$rounds,#6]
+	orr	$s0,$s0,$t3,lsl#24
 	ldrb	$t2,[$rounds,#5]
 	ldrb	$t3,[$rounds,#4]
 	orr	$s1,$s1,$t1,lsl#8
-	orr	$s1,$s1,$t2,lsl#16
-	orr	$s1,$s1,$t3,lsl#24
 	ldrb	$s2,[$rounds,#11]
+	orr	$s1,$s1,$t2,lsl#16
 	ldrb	$t1,[$rounds,#10]
+	orr	$s1,$s1,$t3,lsl#24
 	ldrb	$t2,[$rounds,#9]
 	ldrb	$t3,[$rounds,#8]
 	orr	$s2,$s2,$t1,lsl#8
-	orr	$s2,$s2,$t2,lsl#16
-	orr	$s2,$s2,$t3,lsl#24
 	ldrb	$s3,[$rounds,#15]
+	orr	$s2,$s2,$t2,lsl#16
 	ldrb	$t1,[$rounds,#14]
+	orr	$s2,$s2,$t3,lsl#24
 	ldrb	$t2,[$rounds,#13]
 	ldrb	$t3,[$rounds,#12]
 	orr	$s3,$s3,$t1,lsl#8
-	orr	$s3,$s3,$t2,lsl#16
-	orr	$s3,$s3,$t3,lsl#24
 	str	$s0,[$key],#16
+	orr	$s3,$s3,$t2,lsl#16
 	str	$s1,[$key,#-12]
+	orr	$s3,$s3,$t3,lsl#24
 	str	$s2,[$key,#-8]
 	str	$s3,[$key,#-4]
 
@@ -440,27 +441,26 @@ AES_set_encrypt_key:
 .L128_loop:
 	and	$t2,lr,$s3,lsr#24
 	and	$i1,lr,$s3,lsr#16
-	and	$i2,lr,$s3,lsr#8
-	and	$i3,lr,$s3
 	ldrb	$t2,[$tbl,$t2]
+	and	$i2,lr,$s3,lsr#8
 	ldrb	$i1,[$tbl,$i1]
+	and	$i3,lr,$s3
 	ldrb	$i2,[$tbl,$i2]
-	ldrb	$i3,[$tbl,$i3]
-	ldr	$t1,[$t3],#4			@ rcon[i++]
 	orr	$t2,$t2,$i1,lsl#24
+	ldrb	$i3,[$tbl,$i3]
 	orr	$t2,$t2,$i2,lsl#16
+	ldr	$t1,[$t3],#4			@ rcon[i++]
 	orr	$t2,$t2,$i3,lsl#8
 	eor	$t2,$t2,$t1
 	eor	$s0,$s0,$t2			@ rk[4]=rk[0]^...
 	eor	$s1,$s1,$s0			@ rk[5]=rk[1]^rk[4]
-	eor	$s2,$s2,$s1			@ rk[6]=rk[2]^rk[5]
-	eor	$s3,$s3,$s2			@ rk[7]=rk[3]^rk[6]
 	str	$s0,[$key],#16
+	eor	$s2,$s2,$s1			@ rk[6]=rk[2]^rk[5]
 	str	$s1,[$key,#-12]
+	eor	$s3,$s3,$s2			@ rk[7]=rk[3]^rk[6]
 	str	$s2,[$key,#-8]
-	str	$s3,[$key,#-4]
-
 	subs	$rounds,$rounds,#1
+	str	$s3,[$key,#-4]
 	bne	.L128_loop
 	sub	r2,$key,#176
 	b	.Ldone
@@ -471,16 +471,16 @@ AES_set_encrypt_key:
 	ldrb	$t2,[$rounds,#17]
 	ldrb	$t3,[$rounds,#16]
 	orr	$i2,$i2,$t1,lsl#8
-	orr	$i2,$i2,$t2,lsl#16
-	orr	$i2,$i2,$t3,lsl#24
 	ldrb	$i3,[$rounds,#23]
+	orr	$i2,$i2,$t2,lsl#16
 	ldrb	$t1,[$rounds,#22]
+	orr	$i2,$i2,$t3,lsl#24
 	ldrb	$t2,[$rounds,#21]
 	ldrb	$t3,[$rounds,#20]
 	orr	$i3,$i3,$t1,lsl#8
 	orr	$i3,$i3,$t2,lsl#16
-	orr	$i3,$i3,$t3,lsl#24
 	str	$i2,[$key],#8
+	orr	$i3,$i3,$t3,lsl#24
 	str	$i3,[$key,#-4]
 
 	teq	lr,#192
@@ -494,27 +494,26 @@ AES_set_encrypt_key:
 .L192_loop:
 	and	$t2,lr,$i3,lsr#24
 	and	$i1,lr,$i3,lsr#16
-	and	$i2,lr,$i3,lsr#8
-	and	$i3,lr,$i3
 	ldrb	$t2,[$tbl,$t2]
+	and	$i2,lr,$i3,lsr#8
 	ldrb	$i1,[$tbl,$i1]
+	and	$i3,lr,$i3
 	ldrb	$i2,[$tbl,$i2]
-	ldrb	$i3,[$tbl,$i3]
-	ldr	$t1,[$t3],#4			@ rcon[i++]
 	orr	$t2,$t2,$i1,lsl#24
+	ldrb	$i3,[$tbl,$i3]
 	orr	$t2,$t2,$i2,lsl#16
+	ldr	$t1,[$t3],#4			@ rcon[i++]
 	orr	$t2,$t2,$i3,lsl#8
 	eor	$i3,$t2,$t1
 	eor	$s0,$s0,$i3			@ rk[6]=rk[0]^...
 	eor	$s1,$s1,$s0			@ rk[7]=rk[1]^rk[6]
-	eor	$s2,$s2,$s1			@ rk[8]=rk[2]^rk[7]
-	eor	$s3,$s3,$s2			@ rk[9]=rk[3]^rk[8]
 	str	$s0,[$key],#24
+	eor	$s2,$s2,$s1			@ rk[8]=rk[2]^rk[7]
 	str	$s1,[$key,#-20]
+	eor	$s3,$s3,$s2			@ rk[9]=rk[3]^rk[8]
 	str	$s2,[$key,#-16]
-	str	$s3,[$key,#-12]
-
 	subs	$rounds,$rounds,#1
+	str	$s3,[$key,#-12]
 	subeq	r2,$key,#216
 	beq	.Ldone
 
@@ -532,16 +531,16 @@ AES_set_encrypt_key:
 	ldrb	$t2,[$rounds,#25]
 	ldrb	$t3,[$rounds,#24]
 	orr	$i2,$i2,$t1,lsl#8
-	orr	$i2,$i2,$t2,lsl#16
-	orr	$i2,$i2,$t3,lsl#24
 	ldrb	$i3,[$rounds,#31]
+	orr	$i2,$i2,$t2,lsl#16
 	ldrb	$t1,[$rounds,#30]
+	orr	$i2,$i2,$t3,lsl#24
 	ldrb	$t2,[$rounds,#29]
 	ldrb	$t3,[$rounds,#28]
 	orr	$i3,$i3,$t1,lsl#8
 	orr	$i3,$i3,$t2,lsl#16
-	orr	$i3,$i3,$t3,lsl#24
 	str	$i2,[$key],#8
+	orr	$i3,$i3,$t3,lsl#24
 	str	$i3,[$key,#-4]
 
 	mov	$rounds,#14
@@ -553,52 +552,51 @@ AES_set_encrypt_key:
 .L256_loop:
 	and	$t2,lr,$i3,lsr#24
 	and	$i1,lr,$i3,lsr#16
-	and	$i2,lr,$i3,lsr#8
-	and	$i3,lr,$i3
 	ldrb	$t2,[$tbl,$t2]
+	and	$i2,lr,$i3,lsr#8
 	ldrb	$i1,[$tbl,$i1]
+	and	$i3,lr,$i3
 	ldrb	$i2,[$tbl,$i2]
-	ldrb	$i3,[$tbl,$i3]
-	ldr	$t1,[$t3],#4			@ rcon[i++]
 	orr	$t2,$t2,$i1,lsl#24
+	ldrb	$i3,[$tbl,$i3]
 	orr	$t2,$t2,$i2,lsl#16
+	ldr	$t1,[$t3],#4			@ rcon[i++]
 	orr	$t2,$t2,$i3,lsl#8
 	eor	$i3,$t2,$t1
 	eor	$s0,$s0,$i3			@ rk[8]=rk[0]^...
 	eor	$s1,$s1,$s0			@ rk[9]=rk[1]^rk[8]
-	eor	$s2,$s2,$s1			@ rk[10]=rk[2]^rk[9]
-	eor	$s3,$s3,$s2			@ rk[11]=rk[3]^rk[10]
 	str	$s0,[$key],#32
+	eor	$s2,$s2,$s1			@ rk[10]=rk[2]^rk[9]
 	str	$s1,[$key,#-28]
+	eor	$s3,$s3,$s2			@ rk[11]=rk[3]^rk[10]
 	str	$s2,[$key,#-24]
-	str	$s3,[$key,#-20]
-
 	subs	$rounds,$rounds,#1
+	str	$s3,[$key,#-20]
 	subeq	r2,$key,#256
 	beq	.Ldone
 
 	and	$t2,lr,$s3
 	and	$i1,lr,$s3,lsr#8
-	and	$i2,lr,$s3,lsr#16
-	and	$i3,lr,$s3,lsr#24
 	ldrb	$t2,[$tbl,$t2]
+	and	$i2,lr,$s3,lsr#16
 	ldrb	$i1,[$tbl,$i1]
+	and	$i3,lr,$s3,lsr#24
 	ldrb	$i2,[$tbl,$i2]
-	ldrb	$i3,[$tbl,$i3]
 	orr	$t2,$t2,$i1,lsl#8
+	ldrb	$i3,[$tbl,$i3]
 	orr	$t2,$t2,$i2,lsl#16
+	ldr	$t1,[$key,#-48]
 	orr	$t2,$t2,$i3,lsl#24
 
-	ldr	$t1,[$key,#-48]
 	ldr	$i1,[$key,#-44]
 	ldr	$i2,[$key,#-40]
-	ldr	$i3,[$key,#-36]
 	eor	$t1,$t1,$t2			@ rk[12]=rk[4]^...
+	ldr	$i3,[$key,#-36]
 	eor	$i1,$i1,$t1			@ rk[13]=rk[5]^rk[12]
-	eor	$i2,$i2,$i1			@ rk[14]=rk[6]^rk[13]
-	eor	$i3,$i3,$i2			@ rk[15]=rk[7]^rk[14]
 	str	$t1,[$key,#-16]
+	eor	$i2,$i2,$i1			@ rk[14]=rk[6]^rk[13]
 	str	$i1,[$key,#-12]
+	eor	$i3,$i3,$i2			@ rk[15]=rk[7]^rk[14]
 	str	$i2,[$key,#-8]
 	str	$i3,[$key,#-4]
 	b	.L256_loop
@@ -819,24 +817,24 @@ AES_decrypt:
 	ldrb	$t2,[$rounds,#1]
 	ldrb	$t3,[$rounds,#0]
 	orr	$s0,$s0,$t1,lsl#8
-	orr	$s0,$s0,$t2,lsl#16
-	orr	$s0,$s0,$t3,lsl#24
 	ldrb	$s1,[$rounds,#7]
+	orr	$s0,$s0,$t2,lsl#16
 	ldrb	$t1,[$rounds,#6]
+	orr	$s0,$s0,$t3,lsl#24
 	ldrb	$t2,[$rounds,#5]
 	ldrb	$t3,[$rounds,#4]
 	orr	$s1,$s1,$t1,lsl#8
-	orr	$s1,$s1,$t2,lsl#16
-	orr	$s1,$s1,$t3,lsl#24
 	ldrb	$s2,[$rounds,#11]
+	orr	$s1,$s1,$t2,lsl#16
 	ldrb	$t1,[$rounds,#10]
+	orr	$s1,$s1,$t3,lsl#24
 	ldrb	$t2,[$rounds,#9]
 	ldrb	$t3,[$rounds,#8]
 	orr	$s2,$s2,$t1,lsl#8
-	orr	$s2,$s2,$t2,lsl#16
-	orr	$s2,$s2,$t3,lsl#24
 	ldrb	$s3,[$rounds,#15]
+	orr	$s2,$s2,$t2,lsl#16
 	ldrb	$t1,[$rounds,#14]
+	orr	$s2,$s2,$t3,lsl#24
 	ldrb	$t2,[$rounds,#13]
 	ldrb	$t3,[$rounds,#12]
 	orr	$s3,$s3,$t1,lsl#8
@@ -851,24 +849,24 @@ AES_decrypt:
 	mov	$t3,$s0,lsr#8
 	strb	$t1,[$rounds,#0]
 	strb	$t2,[$rounds,#1]
-	strb	$t3,[$rounds,#2]
-	strb	$s0,[$rounds,#3]
 	mov	$t1,$s1,lsr#24
+	strb	$t3,[$rounds,#2]
 	mov	$t2,$s1,lsr#16
+	strb	$s0,[$rounds,#3]
 	mov	$t3,$s1,lsr#8
 	strb	$t1,[$rounds,#4]
 	strb	$t2,[$rounds,#5]
-	strb	$t3,[$rounds,#6]
-	strb	$s1,[$rounds,#7]
 	mov	$t1,$s2,lsr#24
+	strb	$t3,[$rounds,#6]
 	mov	$t2,$s2,lsr#16
+	strb	$s1,[$rounds,#7]
 	mov	$t3,$s2,lsr#8
 	strb	$t1,[$rounds,#8]
 	strb	$t2,[$rounds,#9]
-	strb	$t3,[$rounds,#10]
-	strb	$s2,[$rounds,#11]
 	mov	$t1,$s3,lsr#24
+	strb	$t3,[$rounds,#10]
 	mov	$t2,$s3,lsr#16
+	strb	$s2,[$rounds,#11]
 	mov	$t3,$s3,lsr#8
 	strb	$t1,[$rounds,#12]
 	strb	$t2,[$rounds,#13]
@@ -885,146 +883,143 @@ AES_decrypt:
 .align	2
 _armv4_AES_decrypt:
 	str	lr,[sp,#-4]!		@ push lr
-	ldr	$t1,[$key],#16
-	ldr	$t2,[$key,#-12]
-	ldr	$t3,[$key,#-8]
-	ldr	$i1,[$key,#-4]
-	ldr	$rounds,[$key,#240-16]
+	ldmia	$key!,{$t1-$i1}
 	eor	$s0,$s0,$t1
+	ldr	$rounds,[$key,#240-16]
 	eor	$s1,$s1,$t2
 	eor	$s2,$s2,$t3
 	eor	$s3,$s3,$i1
 	sub	$rounds,$rounds,#1
 	mov	lr,#255
 
-.Ldec_loop:
 	and	$i1,lr,$s0,lsr#16
 	and	$i2,lr,$s0,lsr#8
 	and	$i3,lr,$s0
 	mov	$s0,$s0,lsr#24
+.Ldec_loop:
 	ldr	$t1,[$tbl,$i1,lsl#2]	@ Td1[s0>>16]
-	ldr	$s0,[$tbl,$s0,lsl#2]	@ Td0[s0>>24]
-	ldr	$t2,[$tbl,$i2,lsl#2]	@ Td2[s0>>8]
-	ldr	$t3,[$tbl,$i3,lsl#2]	@ Td3[s0>>0]
-
 	and	$i1,lr,$s1		@ i0
+	ldr	$t2,[$tbl,$i2,lsl#2]	@ Td2[s0>>8]
 	and	$i2,lr,$s1,lsr#16
+	ldr	$t3,[$tbl,$i3,lsl#2]	@ Td3[s0>>0]
 	and	$i3,lr,$s1,lsr#8
+	ldr	$s0,[$tbl,$s0,lsl#2]	@ Td0[s0>>24]
 	mov	$s1,$s1,lsr#24
+
 	ldr	$i1,[$tbl,$i1,lsl#2]	@ Td3[s1>>0]
-	ldr	$s1,[$tbl,$s1,lsl#2]	@ Td0[s1>>24]
 	ldr	$i2,[$tbl,$i2,lsl#2]	@ Td1[s1>>16]
 	ldr	$i3,[$tbl,$i3,lsl#2]	@ Td2[s1>>8]
 	eor	$s0,$s0,$i1,ror#24
-	eor	$s1,$s1,$t1,ror#8
-	eor	$t2,$i2,$t2,ror#8
-	eor	$t3,$i3,$t3,ror#8
-
+	ldr	$s1,[$tbl,$s1,lsl#2]	@ Td0[s1>>24]
 	and	$i1,lr,$s2,lsr#8	@ i0
+	eor	$t2,$i2,$t2,ror#8
 	and	$i2,lr,$s2		@ i1
+	eor	$t3,$i3,$t3,ror#8
 	and	$i3,lr,$s2,lsr#16
-	mov	$s2,$s2,lsr#24
+	eor	$s1,$s1,$t1,ror#8
 	ldr	$i1,[$tbl,$i1,lsl#2]	@ Td2[s2>>8]
+	mov	$s2,$s2,lsr#24
+
 	ldr	$i2,[$tbl,$i2,lsl#2]	@ Td3[s2>>0]
-	ldr	$s2,[$tbl,$s2,lsl#2]	@ Td0[s2>>24]
 	ldr	$i3,[$tbl,$i3,lsl#2]	@ Td1[s2>>16]
 	eor	$s0,$s0,$i1,ror#16
-	eor	$s1,$s1,$i2,ror#24
-	eor	$s2,$s2,$t2,ror#8
-	eor	$t3,$i3,$t3,ror#8
-
+	ldr	$s2,[$tbl,$s2,lsl#2]	@ Td0[s2>>24]
 	and	$i1,lr,$s3,lsr#16	@ i0
+	eor	$s1,$s1,$i2,ror#24
 	and	$i2,lr,$s3,lsr#8	@ i1
+	eor	$t3,$i3,$t3,ror#8
 	and	$i3,lr,$s3		@ i2
-	mov	$s3,$s3,lsr#24
+	eor	$s2,$s2,$t2,ror#8
 	ldr	$i1,[$tbl,$i1,lsl#2]	@ Td1[s3>>16]
+	mov	$s3,$s3,lsr#24
+
 	ldr	$i2,[$tbl,$i2,lsl#2]	@ Td2[s3>>8]
 	ldr	$i3,[$tbl,$i3,lsl#2]	@ Td3[s3>>0]
-	ldr	$s3,[$tbl,$s3,lsl#2]	@ Td0[s3>>24]
 	eor	$s0,$s0,$i1,ror#8
+	ldr	$s3,[$tbl,$s3,lsl#2]	@ Td0[s3>>24]
 	eor	$s1,$s1,$i2,ror#16
 	eor	$s2,$s2,$i3,ror#24
+	ldr	$i1,[$key],#16
 	eor	$s3,$s3,$t3,ror#8
 
-	ldr	$t1,[$key],#16
-	ldr	$t2,[$key,#-12]
-	ldr	$t3,[$key,#-8]
-	ldr	$i1,[$key,#-4]
-	eor	$s0,$s0,$t1
-	eor	$s1,$s1,$t2
-	eor	$s2,$s2,$t3
-	eor	$s3,$s3,$i1
+	ldr	$t1,[$key,#-12]
+	ldr	$t2,[$key,#-8]
+	eor	$s0,$s0,$i1
+	ldr	$t3,[$key,#-4]
+	and	$i1,lr,$s0,lsr#16
+	eor	$s1,$s1,$t1
+	and	$i2,lr,$s0,lsr#8
+	eor	$s2,$s2,$t2
+	and	$i3,lr,$s0
+	eor	$s3,$s3,$t3
+	mov	$s0,$s0,lsr#24
 
 	subs	$rounds,$rounds,#1
 	bne	.Ldec_loop
 
 	add	$tbl,$tbl,#1024
 
-	ldr	$t1,[$tbl,#0]		@ prefetch Td4
-	ldr	$t2,[$tbl,#32]
-	ldr	$t3,[$tbl,#64]
-	ldr	$i1,[$tbl,#96]
-	ldr	$i2,[$tbl,#128]
-	ldr	$i3,[$tbl,#160]
-	ldr	$t1,[$tbl,#192]
-	ldr	$t2,[$tbl,#224]
+	ldr	$t2,[$tbl,#0]		@ prefetch Td4
+	ldr	$t3,[$tbl,#32]
+	ldr	$t1,[$tbl,#64]
+	ldr	$t2,[$tbl,#96]
+	ldr	$t3,[$tbl,#128]
+	ldr	$t1,[$tbl,#160]
+	ldr	$t2,[$tbl,#192]
+	ldr	$t3,[$tbl,#224]
 
-	and	$i1,lr,$s0,lsr#16
-	and	$i2,lr,$s0,lsr#8
-	and	$i3,lr,$s0
-	ldrb	$s0,[$tbl,$s0,lsr#24]	@ Td4[s0>>24]
+	ldrb	$s0,[$tbl,$s0]		@ Td4[s0>>24]
 	ldrb	$t1,[$tbl,$i1]		@ Td4[s0>>16]
-	ldrb	$t2,[$tbl,$i2]		@ Td4[s0>>8]
-	ldrb	$t3,[$tbl,$i3]		@ Td4[s0>>0]
-
 	and	$i1,lr,$s1		@ i0
+	ldrb	$t2,[$tbl,$i2]		@ Td4[s0>>8]
 	and	$i2,lr,$s1,lsr#16
+	ldrb	$t3,[$tbl,$i3]		@ Td4[s0>>0]
 	and	$i3,lr,$s1,lsr#8
+
 	ldrb	$i1,[$tbl,$i1]		@ Td4[s1>>0]
 	ldrb	$s1,[$tbl,$s1,lsr#24]	@ Td4[s1>>24]
 	ldrb	$i2,[$tbl,$i2]		@ Td4[s1>>16]
-	ldrb	$i3,[$tbl,$i3]		@ Td4[s1>>8]
 	eor	$s0,$i1,$s0,lsl#24
+	ldrb	$i3,[$tbl,$i3]		@ Td4[s1>>8]
 	eor	$s1,$t1,$s1,lsl#8
-	eor	$t2,$t2,$i2,lsl#8
-	eor	$t3,$t3,$i3,lsl#8
-
 	and	$i1,lr,$s2,lsr#8	@ i0
+	eor	$t2,$t2,$i2,lsl#8
 	and	$i2,lr,$s2		@ i1
-	and	$i3,lr,$s2,lsr#16
+	eor	$t3,$t3,$i3,lsl#8
 	ldrb	$i1,[$tbl,$i1]		@ Td4[s2>>8]
+	and	$i3,lr,$s2,lsr#16
+
 	ldrb	$i2,[$tbl,$i2]		@ Td4[s2>>0]
 	ldrb	$s2,[$tbl,$s2,lsr#24]	@ Td4[s2>>24]
-	ldrb	$i3,[$tbl,$i3]		@ Td4[s2>>16]
 	eor	$s0,$s0,$i1,lsl#8
+	ldrb	$i3,[$tbl,$i3]		@ Td4[s2>>16]
 	eor	$s1,$i2,$s1,lsl#16
-	eor	$s2,$t2,$s2,lsl#16
-	eor	$t3,$t3,$i3,lsl#16
-
 	and	$i1,lr,$s3,lsr#16	@ i0
+	eor	$s2,$t2,$s2,lsl#16
 	and	$i2,lr,$s3,lsr#8	@ i1
-	and	$i3,lr,$s3		@ i2
+	eor	$t3,$t3,$i3,lsl#16
 	ldrb	$i1,[$tbl,$i1]		@ Td4[s3>>16]
+	and	$i3,lr,$s3		@ i2
+
 	ldrb	$i2,[$tbl,$i2]		@ Td4[s3>>8]
 	ldrb	$i3,[$tbl,$i3]		@ Td4[s3>>0]
 	ldrb	$s3,[$tbl,$s3,lsr#24]	@ Td4[s3>>24]
 	eor	$s0,$s0,$i1,lsl#16
+	ldr	$i1,[$key,#0]
 	eor	$s1,$s1,$i2,lsl#8
+	ldr	$t1,[$key,#4]
 	eor	$s2,$i3,$s2,lsl#8
+	ldr	$t2,[$key,#8]
 	eor	$s3,$t3,$s3,lsl#24
+	ldr	$t3,[$key,#12]
 
-	ldr	lr,[sp],#4		@ pop lr
-	ldr	$t1,[$key,#0]
-	ldr	$t2,[$key,#4]
-	ldr	$t3,[$key,#8]
-	ldr	$i1,[$key,#12]
-	eor	$s0,$s0,$t1
-	eor	$s1,$s1,$t2
-	eor	$s2,$s2,$t3
-	eor	$s3,$s3,$i1
+	eor	$s0,$s0,$i1
+	eor	$s1,$s1,$t1
+	eor	$s2,$s2,$t2
+	eor	$s3,$s3,$t3
 
 	sub	$tbl,$tbl,#1024
-	mov	pc,lr			@ return
+	ldr	pc,[sp],#4		@ pop and return
 .size	_armv4_AES_decrypt,.-_armv4_AES_decrypt
 .asciz	"AES for ARMv4, CRYPTOGAMS by <appro\@openssl.org>"
 .align	2

crypto/modes/asm/ghash-armv4.pl

@@ -19,6 +19,12 @@
 # loop, this assembler loop body was found to be ~3x smaller than
 # compiler-generated one...
 #
+# July 2010
+#
+# Rescheduling for dual-issue pipeline resulted in 8.5% improvement on
+# Cortex A8 core and ~25 cycles per processed byte (which was observed
+# to be ~3 times faster than gcc-generated code:-)
+#
 # Note about "528B" variant. In ARM case it makes lesser sense to
 # implement it for following reasons:
 #
@@ -123,12 +129,12 @@ gcm_ghash_4bit:
 
 	add	$Zhh,$Htbl,$nlo,lsl#4
 	ldmia	$Zhh,{$Zll-$Zhh}	@ load Htbl[nlo]
+	add	$Thh,$Htbl,$nhi
 	ldrb	$nlo,[$inp,#14]
 
-	add	$Thh,$Htbl,$nhi
 	and	$nhi,$Zll,#0xf		@ rem
 	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nhi]
-	mov	$nhi,$nhi,lsl#1
+	add	$nhi,$nhi,$nhi
 	eor	$Zll,$Tll,$Zll,lsr#4
 	ldrh	$Tll,[sp,$nhi]		@ rem_4bit[rem]
 	eor	$Zll,$Zll,$Zlh,lsl#28
@@ -139,15 +145,15 @@ gcm_ghash_4bit:
 	eor	$Zhl,$Zhl,$Zhh,lsl#28
 	eor	$Zhh,$Thh,$Zhh,lsr#4
 	eor	$nlo,$nlo,$nhi
-	eor	$Zhh,$Zhh,$Tll,lsl#16
 	and	$nhi,$nlo,#0xf0
 	and	$nlo,$nlo,#0x0f
+	eor	$Zhh,$Zhh,$Tll,lsl#16
 
 .Loop:
 	add	$Thh,$Htbl,$nlo,lsl#4
 	subs	$cnt,$cnt,#1
-	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nlo]
 	and	$nlo,$Zll,#0xf		@ rem
+	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nlo]
 	add	$nlo,$nlo,$nlo
 	eor	$Zll,$Tll,$Zll,lsr#4
 	ldrh	$Tll,[sp,$nlo]		@ rem_4bit[rem]
@@ -161,22 +167,22 @@ gcm_ghash_4bit:
 
 	add	$Thh,$Htbl,$nhi
 	eor	$Zhh,$Zhh,$Tll,lsl#16	@ ^= rem_4bit[rem]
-	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nhi]
 	and	$nhi,$Zll,#0xf		@ rem
+	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nhi]
 	add	$nhi,$nhi,$nhi
 	eor	$Zll,$Tll,$Zll,lsr#4
 	ldrh	$Tll,[sp,$nhi]		@ rem_4bit[rem]
 	eor	$Zll,$Zll,$Zlh,lsl#28
-	ldrplb	$nhi,[$Xi,$cnt]
 	eor	$Zlh,$Tlh,$Zlh,lsr#4
+	ldrplb	$nhi,[$Xi,$cnt]
 	eor	$Zlh,$Zlh,$Zhl,lsl#28
 	eor	$Zhl,$Thl,$Zhl,lsr#4
 	eor	$Zhl,$Zhl,$Zhh,lsl#28
-	eor	$Zhh,$Thh,$Zhh,lsr#4
 	eorpl	$nlo,$nlo,$nhi
-	eor	$Zhh,$Zhh,$Tll,lsl#16	@ ^= rem_4bit[rem]
+	eor	$Zhh,$Thh,$Zhh,lsr#4
 	andpl	$nhi,$nlo,#0xf0
 	andpl	$nlo,$nlo,#0x0f
+	eor	$Zhh,$Zhh,$Tll,lsl#16	@ ^= rem_4bit[rem]
 	bpl	.Loop
 
 	ldr	$len,[sp,#32]		@ re-load $len/end
@@ -212,7 +218,7 @@ gcm_gmult_4bit:
 	add	$Thh,$Htbl,$nhi
 	and	$nhi,$Zll,#0xf		@ rem
 	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nhi]
-	mov	$nhi,$nhi,lsl#1
+	add	$nhi,$nhi,$nhi
 	eor	$Zll,$Tll,$Zll,lsr#4
 	ldrh	$Tll,[$rem_4bit,$nhi]	@ rem_4bit[rem]
 	eor	$Zll,$Zll,$Zlh,lsl#28
@@ -228,8 +234,8 @@ gcm_gmult_4bit:
 .Loop2:
 	add	$Thh,$Htbl,$nlo,lsl#4
 	subs	$cnt,$cnt,#1
-	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nlo]
 	and	$nlo,$Zll,#0xf		@ rem
+	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nlo]
 	add	$nlo,$nlo,$nlo
 	eor	$Zll,$Tll,$Zll,lsr#4
 	ldrh	$Tll,[$rem_4bit,$nlo]	@ rem_4bit[rem]
@@ -243,8 +249,8 @@ gcm_gmult_4bit:
 
 	add	$Thh,$Htbl,$nhi
 	eor	$Zhh,$Zhh,$Tll,lsl#16	@ ^= rem_4bit[rem]
-	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nhi]
 	and	$nhi,$Zll,#0xf		@ rem
+	ldmia	$Thh,{$Tll-$Thh}	@ load Htbl[nhi]
 	add	$nhi,$nhi,$nhi
 	eor	$Zll,$Tll,$Zll,lsr#4
 	ldrh	$Tll,[$rem_4bit,$nhi]	@ rem_4bit[rem]
@@ -255,8 +261,8 @@ gcm_gmult_4bit:
 	eor	$Zhl,$Zhl,$Zhh,lsl#28
 	eor	$Zhh,$Thh,$Zhh,lsr#4
 	andpl	$nhi,$nlo,#0xf0
-	eor	$Zhh,$Zhh,$Tll,lsl#16	@ ^= rem_4bit[rem]
 	andpl	$nlo,$nlo,#0x0f
+	eor	$Zhh,$Zhh,$Tll,lsl#16	@ ^= rem_4bit[rem]
 	bpl	.Loop2
 ___
 	&Zsmash();

crypto/sha/asm/sha256-armv4.pl

@@ -11,7 +11,12 @@
 
 # Performance is ~2x better than gcc 3.4 generated code and in "abso-
 # lute" terms is ~2250 cycles per 64-byte block or ~35 cycles per
-# byte.
+# byte [on single-issue Xscale PXA250 core].
+
+# July 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 22% improvement on
+# Cortex A8 core and ~20 cycles per processed byte.
 
 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
@@ -52,27 +57,27 @@ $code.=<<___ if ($i<16);
 ___
 $code.=<<___;
 	ldr	$t2,[$Ktbl],#4			@ *K256++
-	str	$T1,[sp,#`$i%16`*4]
 	mov	$t0,$e,ror#$Sigma1[0]
+	str	$T1,[sp,#`$i%16`*4]
 	eor	$t0,$t0,$e,ror#$Sigma1[1]
-	eor	$t0,$t0,$e,ror#$Sigma1[2]	@ Sigma1(e)
-	add	$T1,$T1,$t0
 	eor	$t1,$f,$g
+	eor	$t0,$t0,$e,ror#$Sigma1[2]	@ Sigma1(e)
 	and	$t1,$t1,$e
+	add	$T1,$T1,$t0
 	eor	$t1,$t1,$g			@ Ch(e,f,g)
-	add	$T1,$T1,$t1
 	add	$T1,$T1,$h
-	add	$T1,$T1,$t2
 	mov	$h,$a,ror#$Sigma0[0]
+	add	$T1,$T1,$t1
 	eor	$h,$h,$a,ror#$Sigma0[1]
+	add	$T1,$T1,$t2
 	eor	$h,$h,$a,ror#$Sigma0[2]		@ Sigma0(a)
 	orr	$t0,$a,$b
-	and	$t0,$t0,$c
 	and	$t1,$a,$b
+	and	$t0,$t0,$c
+	add	$h,$h,$T1
 	orr	$t0,$t0,$t1			@ Maj(a,b,c)
-	add	$h,$h,$t0
 	add	$d,$d,$T1
-	add	$h,$h,$T1
+	add	$h,$h,$t0
 ___
 }
 
@@ -80,19 +85,19 @@ sub BODY_16_XX {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 
 $code.=<<___;
-	ldr	$t1,[sp,#`($i+1)%16`*4]	@ $i
+	ldr	$t1,[sp,#`($i+1)%16`*4]		@ $i
 	ldr	$t2,[sp,#`($i+14)%16`*4]
 	ldr	$T1,[sp,#`($i+0)%16`*4]
-	ldr	$inp,[sp,#`($i+9)%16`*4]
 	mov	$t0,$t1,ror#$sigma0[0]
+	ldr	$inp,[sp,#`($i+9)%16`*4]
 	eor	$t0,$t0,$t1,ror#$sigma0[1]
 	eor	$t0,$t0,$t1,lsr#$sigma0[2]	@ sigma0(X[i+1])
 	mov	$t1,$t2,ror#$sigma1[0]
+	add	$T1,$T1,$t0
 	eor	$t1,$t1,$t2,ror#$sigma1[1]
+	add	$T1,$T1,$inp
 	eor	$t1,$t1,$t2,lsr#$sigma1[2]	@ sigma1(X[i+14])
-	add	$T1,$T1,$t0
 	add	$T1,$T1,$t1
-	add	$T1,$T1,$inp
 ___
 	&BODY_00_15(@_);
 }

crypto/sha/asm/sha512-armv4.pl

@@ -10,7 +10,13 @@
 # SHA512 block procedure for ARMv4. September 2007.
 
 # This code is ~4.5 (four and a half) times faster than code generated
-# by gcc 3.4 and it spends ~72 clock cycles per byte. 
+# by gcc 3.4 and it spends ~72 clock cycles per byte [on single-issue
+# Xscale PXA250 core].
+#
+# July 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 6% improvement on
+# Cortex A8 core and ~40 cycles per processed byte.
 
 # Byte order [in]dependence. =========================================
 #
@@ -73,33 +79,31 @@ $code.=<<___;
 	eor	$t0,$t0,$Elo,lsl#23
 	eor	$t1,$t1,$Ehi,lsl#23	@ Sigma1(e)
 	adds	$Tlo,$Tlo,$t0
-	adc	$Thi,$Thi,$t1		@ T += Sigma1(e)
-	adds	$Tlo,$Tlo,$t2
-	adc	$Thi,$Thi,$t3		@ T += h
-
 	ldr	$t0,[sp,#$Foff+0]	@ f.lo
+	adc	$Thi,$Thi,$t1		@ T += Sigma1(e)
 	ldr	$t1,[sp,#$Foff+4]	@ f.hi
+	adds	$Tlo,$Tlo,$t2
 	ldr	$t2,[sp,#$Goff+0]	@ g.lo
+	adc	$Thi,$Thi,$t3		@ T += h
 	ldr	$t3,[sp,#$Goff+4]	@ g.hi
-	str	$Elo,[sp,#$Eoff+0]
-	str	$Ehi,[sp,#$Eoff+4]
-	str	$Alo,[sp,#$Aoff+0]
-	str	$Ahi,[sp,#$Aoff+4]
 
 	eor	$t0,$t0,$t2
+	str	$Elo,[sp,#$Eoff+0]
 	eor	$t1,$t1,$t3
+	str	$Ehi,[sp,#$Eoff+4]
 	and	$t0,$t0,$Elo
+	str	$Alo,[sp,#$Aoff+0]
 	and	$t1,$t1,$Ehi
+	str	$Ahi,[sp,#$Aoff+4]
 	eor	$t0,$t0,$t2
-	eor	$t1,$t1,$t3		@ Ch(e,f,g)
-
 	ldr	$t2,[$Ktbl,#4]		@ K[i].lo
+	eor	$t1,$t1,$t3		@ Ch(e,f,g)
 	ldr	$t3,[$Ktbl,#0]		@ K[i].hi
-	ldr	$Elo,[sp,#$Doff+0]	@ d.lo
-	ldr	$Ehi,[sp,#$Doff+4]	@ d.hi
 
 	adds	$Tlo,$Tlo,$t0
+	ldr	$Elo,[sp,#$Doff+0]	@ d.lo
 	adc	$Thi,$Thi,$t1		@ T += Ch(e,f,g)
+	ldr	$Ehi,[sp,#$Doff+4]	@ d.hi
 	adds	$Tlo,$Tlo,$t2
 	adc	$Thi,$Thi,$t3		@ T += K[i]
 	adds	$Elo,$Elo,$Tlo