bn/asm/armv4-mont.pl: boost NEON performance.

Close difference gap on Cortex-A9, which resulted in further improvement
even on other processors.
Reviewed-by: NRichard Levitte <levitte@openssl.org>

crypto/bn/asm/armv4-mont.pl

@@ -38,6 +38,15 @@
 # for execution on all NEON-capable processors, because gain on
 # others outweighs the marginal loss on Cortex-A9.
 
+# September 2015
+#
+# Align Cortex-A9 performance with November 2013 improvements, i.e.
+# NEON code is now ~20-105% faster than integer-only one on this
+# processor. But this optimization further improved performance even
+# on other processors: NEON code path is ~45-180% faster than original
+# integer-only on Cortex-A8, ~10-210% on Cortex-A15, ~70-450% on
+# Snapdragon S4.
+
 $flavour = shift;
 if ($flavour=~/^\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
 else { while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {} }
@@ -272,19 +281,16 @@ bn_mul_mont:
 .size	bn_mul_mont,.-bn_mul_mont
 ___
 {
-sub Dlo()   { shift=~m|q([1]?[0-9])|?"d".($1*2):"";     }
-sub Dhi()   { shift=~m|q([1]?[0-9])|?"d".($1*2+1):"";   }
-
 my ($A0,$A1,$A2,$A3)=map("d$_",(0..3));
 my ($N0,$N1,$N2,$N3)=map("d$_",(4..7));
 my ($Z,$Temp)=("q4","q5");
-my ($A0xB,$A1xB,$A2xB,$A3xB,$A4xB,$A5xB,$A6xB,$A7xB)=map("q$_",(6..13));
+my @ACC=map("q$_",(6..13));
 my ($Bi,$Ni,$M0)=map("d$_",(28..31));
-my $zero=&Dlo($Z);
-my $temp=&Dlo($Temp);
+my $zero="$Z#lo";
+my $temp="$Temp#lo";
 
 my ($rptr,$aptr,$bptr,$nptr,$n0,$num)=map("r$_",(0..5));
-my ($tinptr,$toutptr,$inner,$outer)=map("r$_",(6..9));
+my ($tinptr,$toutptr,$inner,$outer,$bnptr)=map("r$_",(6..11));
 
 $code.=<<___;
 #if __ARM_MAX_ARCH__>=7
@@ -300,59 +306,58 @@ bn_mul8x_mont_neon:
 	ldmia	ip,{r4-r5}		@ load rest of parameter block
 	mov	ip,sp
 
-	sub		$toutptr,sp,#16
+	cmp	$num,#8
+	bhi	.LNEON_8n
+
+	@ special case for $num==8, everything is in register bank...
+
 	vld1.32		{${Bi}[0]}, [$bptr,:32]!
-	sub		$toutptr,$toutptr,$num,lsl#4
+	veor		$zero,$zero,$zero
+	sub		$toutptr,sp,$num,lsl#4
 	vld1.32		{$A0-$A3},  [$aptr]!		@ can't specify :32 :-(
 	and		$toutptr,$toutptr,#-64
 	vld1.32		{${M0}[0]}, [$n0,:32]
 	mov		sp,$toutptr			@ alloca
-	veor		$zero,$zero,$zero
-	subs		$inner,$num,#8
 	vzip.16		$Bi,$zero
 
-	vmull.u32	$A0xB,$Bi,${A0}[0]
-	vmull.u32	$A1xB,$Bi,${A0}[1]
-	vmull.u32	$A2xB,$Bi,${A1}[0]
-	vshl.i64	$temp,`&Dhi("$A0xB")`,#16
-	vmull.u32	$A3xB,$Bi,${A1}[1]
+	vmull.u32	@ACC[0],$Bi,${A0}[0]
+	vmull.u32	@ACC[1],$Bi,${A0}[1]
+	vmull.u32	@ACC[2],$Bi,${A1}[0]
+	vshl.i64	$Ni,@ACC[0]#hi,#16
+	vmull.u32	@ACC[3],$Bi,${A1}[1]
 
-	vadd.u64	$temp,$temp,`&Dlo("$A0xB")`
+	vadd.u64	$Ni,$Ni,@ACC[0]#lo
 	veor		$zero,$zero,$zero
-	vmul.u32	$Ni,$temp,$M0
+	vmul.u32	$Ni,$Ni,$M0
 
-	vmull.u32	$A4xB,$Bi,${A2}[0]
+	vmull.u32	@ACC[4],$Bi,${A2}[0]
 	 vld1.32	{$N0-$N3}, [$nptr]!
-	vmull.u32	$A5xB,$Bi,${A2}[1]
-	vmull.u32	$A6xB,$Bi,${A3}[0]
+	vmull.u32	@ACC[5],$Bi,${A2}[1]
+	vmull.u32	@ACC[6],$Bi,${A3}[0]
 	vzip.16		$Ni,$zero
-	vmull.u32	$A7xB,$Bi,${A3}[1]
-
-	bne	.LNEON_1st
+	vmull.u32	@ACC[7],$Bi,${A3}[1]
 
-	@ special case for num=8, everything is in register bank...
-
-	vmlal.u32	$A0xB,$Ni,${N0}[0]
+	vmlal.u32	@ACC[0],$Ni,${N0}[0]
 	sub		$outer,$num,#1
-	vmlal.u32	$A1xB,$Ni,${N0}[1]
-	vmlal.u32	$A2xB,$Ni,${N1}[0]
-	vmlal.u32	$A3xB,$Ni,${N1}[1]
-
-	vmlal.u32	$A4xB,$Ni,${N2}[0]
-	vmov		$Temp,$A0xB
-	vmlal.u32	$A5xB,$Ni,${N2}[1]
-	vmov		$A0xB,$A1xB
-	vmlal.u32	$A6xB,$Ni,${N3}[0]
-	vmov		$A1xB,$A2xB
-	vmlal.u32	$A7xB,$Ni,${N3}[1]
-	vmov		$A2xB,$A3xB
-	vmov		$A3xB,$A4xB
+	vmlal.u32	@ACC[1],$Ni,${N0}[1]
+	vmlal.u32	@ACC[2],$Ni,${N1}[0]
+	vmlal.u32	@ACC[3],$Ni,${N1}[1]
+
+	vmlal.u32	@ACC[4],$Ni,${N2}[0]
+	vmov		$Temp,@ACC[0]
+	vmlal.u32	@ACC[5],$Ni,${N2}[1]
+	vmov		@ACC[0],@ACC[1]
+	vmlal.u32	@ACC[6],$Ni,${N3}[0]
+	vmov		@ACC[1],@ACC[2]
+	vmlal.u32	@ACC[7],$Ni,${N3}[1]
+	vmov		@ACC[2],@ACC[3]
+	vmov		@ACC[3],@ACC[4]
 	vshr.u64	$temp,$temp,#16
-	vmov		$A4xB,$A5xB
-	vmov		$A5xB,$A6xB
-	vadd.u64	$temp,$temp,`&Dhi("$Temp")`
-	vmov		$A6xB,$A7xB
-	veor		$A7xB,$A7xB
+	vmov		@ACC[4],@ACC[5]
+	vmov		@ACC[5],@ACC[6]
+	vadd.u64	$temp,$temp,$Temp#hi
+	vmov		@ACC[6],@ACC[7]
+	veor		@ACC[7],@ACC[7]
 	vshr.u64	$temp,$temp,#16
 
 	b	.LNEON_outer8
@@ -362,279 +367,302 @@ bn_mul8x_mont_neon:
 	vld1.32		{${Bi}[0]}, [$bptr,:32]!
 	veor		$zero,$zero,$zero
 	vzip.16		$Bi,$zero
-	vadd.u64	`&Dlo("$A0xB")`,`&Dlo("$A0xB")`,$temp
+	vadd.u64	@ACC[0]#lo,@ACC[0]#lo,$temp
 
-	vmlal.u32	$A0xB,$Bi,${A0}[0]
-	vmlal.u32	$A1xB,$Bi,${A0}[1]
-	vmlal.u32	$A2xB,$Bi,${A1}[0]
-	vshl.i64	$temp,`&Dhi("$A0xB")`,#16
-	vmlal.u32	$A3xB,$Bi,${A1}[1]
+	vmlal.u32	@ACC[0],$Bi,${A0}[0]
+	vmlal.u32	@ACC[1],$Bi,${A0}[1]
+	vmlal.u32	@ACC[2],$Bi,${A1}[0]
+	vshl.i64	$Ni,@ACC[0]#hi,#16
+	vmlal.u32	@ACC[3],$Bi,${A1}[1]
 
-	vadd.u64	$temp,$temp,`&Dlo("$A0xB")`
+	vadd.u64	$Ni,$Ni,@ACC[0]#lo
 	veor		$zero,$zero,$zero
 	subs		$outer,$outer,#1
-	vmul.u32	$Ni,$temp,$M0
+	vmul.u32	$Ni,$Ni,$M0
 
-	vmlal.u32	$A4xB,$Bi,${A2}[0]
-	vmlal.u32	$A5xB,$Bi,${A2}[1]
-	vmlal.u32	$A6xB,$Bi,${A3}[0]
+	vmlal.u32	@ACC[4],$Bi,${A2}[0]
+	vmlal.u32	@ACC[5],$Bi,${A2}[1]
+	vmlal.u32	@ACC[6],$Bi,${A3}[0]
 	vzip.16		$Ni,$zero
-	vmlal.u32	$A7xB,$Bi,${A3}[1]
-
-	vmlal.u32	$A0xB,$Ni,${N0}[0]
-	vmlal.u32	$A1xB,$Ni,${N0}[1]
-	vmlal.u32	$A2xB,$Ni,${N1}[0]
-	vmlal.u32	$A3xB,$Ni,${N1}[1]
-
-	vmlal.u32	$A4xB,$Ni,${N2}[0]
-	vmov		$Temp,$A0xB
-	vmlal.u32	$A5xB,$Ni,${N2}[1]
-	vmov		$A0xB,$A1xB
-	vmlal.u32	$A6xB,$Ni,${N3}[0]
-	vmov		$A1xB,$A2xB
-	vmlal.u32	$A7xB,$Ni,${N3}[1]
-	vmov		$A2xB,$A3xB
-	vmov		$A3xB,$A4xB
+	vmlal.u32	@ACC[7],$Bi,${A3}[1]
+
+	vmlal.u32	@ACC[0],$Ni,${N0}[0]
+	vmlal.u32	@ACC[1],$Ni,${N0}[1]
+	vmlal.u32	@ACC[2],$Ni,${N1}[0]
+	vmlal.u32	@ACC[3],$Ni,${N1}[1]
+
+	vmlal.u32	@ACC[4],$Ni,${N2}[0]
+	vmov		$Temp,@ACC[0]
+	vmlal.u32	@ACC[5],$Ni,${N2}[1]
+	vmov		@ACC[0],@ACC[1]
+	vmlal.u32	@ACC[6],$Ni,${N3}[0]
+	vmov		@ACC[1],@ACC[2]
+	vmlal.u32	@ACC[7],$Ni,${N3}[1]
+	vmov		@ACC[2],@ACC[3]
+	vmov		@ACC[3],@ACC[4]
 	vshr.u64	$temp,$temp,#16
-	vmov		$A4xB,$A5xB
-	vmov		$A5xB,$A6xB
-	vadd.u64	$temp,$temp,`&Dhi("$Temp")`
-	vmov		$A6xB,$A7xB
-	veor		$A7xB,$A7xB
+	vmov		@ACC[4],@ACC[5]
+	vmov		@ACC[5],@ACC[6]
+	vadd.u64	$temp,$temp,$Temp#hi
+	vmov		@ACC[6],@ACC[7]
+	veor		@ACC[7],@ACC[7]
 	vshr.u64	$temp,$temp,#16
 
 	bne	.LNEON_outer8
 
-	vadd.u64	`&Dlo("$A0xB")`,`&Dlo("$A0xB")`,$temp
+	vadd.u64	@ACC[0]#lo,@ACC[0]#lo,$temp
 	mov		$toutptr,sp
-	vshr.u64	$temp,`&Dlo("$A0xB")`,#16
+	vshr.u64	$temp,@ACC[0]#lo,#16
 	mov		$inner,$num
-	vadd.u64	`&Dhi("$A0xB")`,`&Dhi("$A0xB")`,$temp
-	add		$tinptr,sp,#16
-	vshr.u64	$temp,`&Dhi("$A0xB")`,#16
-	vzip.16		`&Dlo("$A0xB")`,`&Dhi("$A0xB")`
+	vadd.u64	@ACC[0]#hi,@ACC[0]#hi,$temp
+	add		$tinptr,sp,#96
+	vshr.u64	$temp,@ACC[0]#hi,#16
+	vzip.16		@ACC[0]#lo,@ACC[0]#hi
 
-	b	.LNEON_tail2
+	b	.LNEON_tail_entry
 
 .align	4
-.LNEON_1st:
-	vmlal.u32	$A0xB,$Ni,${N0}[0]
-	 vld1.32	{$A0-$A3}, [$aptr]!
-	vmlal.u32	$A1xB,$Ni,${N0}[1]
-	subs		$inner,$inner,#8
-	vmlal.u32	$A2xB,$Ni,${N1}[0]
-	vmlal.u32	$A3xB,$Ni,${N1}[1]
-
-	vmlal.u32	$A4xB,$Ni,${N2}[0]
-	 vld1.32	{$N0-$N1}, [$nptr]!
-	vmlal.u32	$A5xB,$Ni,${N2}[1]
-	 vst1.64	{$A0xB-$A1xB}, [$toutptr,:256]!
-	vmlal.u32	$A6xB,$Ni,${N3}[0]
-	vmlal.u32	$A7xB,$Ni,${N3}[1]
-	 vst1.64	{$A2xB-$A3xB}, [$toutptr,:256]!
-
-	vmull.u32	$A0xB,$Bi,${A0}[0]
-	 vld1.32	{$N2-$N3}, [$nptr]!
-	vmull.u32	$A1xB,$Bi,${A0}[1]
-	 vst1.64	{$A4xB-$A5xB}, [$toutptr,:256]!
-	vmull.u32	$A2xB,$Bi,${A1}[0]
-	vmull.u32	$A3xB,$Bi,${A1}[1]
-	 vst1.64	{$A6xB-$A7xB}, [$toutptr,:256]!
-
-	vmull.u32	$A4xB,$Bi,${A2}[0]
-	vmull.u32	$A5xB,$Bi,${A2}[1]
-	vmull.u32	$A6xB,$Bi,${A3}[0]
-	vmull.u32	$A7xB,$Bi,${A3}[1]
-
-	bne	.LNEON_1st
-
-	vmlal.u32	$A0xB,$Ni,${N0}[0]
-	add		$tinptr,sp,#16
-	vmlal.u32	$A1xB,$Ni,${N0}[1]
-	sub		$aptr,$aptr,$num,lsl#2		@ rewind $aptr
-	vmlal.u32	$A2xB,$Ni,${N1}[0]
-	 vld1.64	{$Temp}, [sp,:128]
-	vmlal.u32	$A3xB,$Ni,${N1}[1]
-	sub		$outer,$num,#1
-
-	vmlal.u32	$A4xB,$Ni,${N2}[0]
-	vst1.64		{$A0xB-$A1xB}, [$toutptr,:256]!
-	vmlal.u32	$A5xB,$Ni,${N2}[1]
-	vshr.u64	$temp,$temp,#16
-	 vld1.64	{$A0xB},       [$tinptr, :128]!
-	vmlal.u32	$A6xB,$Ni,${N3}[0]
-	vst1.64		{$A2xB-$A3xB}, [$toutptr,:256]!
-	vmlal.u32	$A7xB,$Ni,${N3}[1]
-
-	vst1.64		{$A4xB-$A5xB}, [$toutptr,:256]!
-	vadd.u64	$temp,$temp,`&Dhi("$Temp")`
-	veor		$Z,$Z,$Z
-	vst1.64		{$A6xB-$A7xB}, [$toutptr,:256]!
-	 vld1.64	{$A1xB-$A2xB}, [$tinptr, :256]!
-	vst1.64		{$Z},          [$toutptr,:128]
-	vshr.u64	$temp,$temp,#16
+.LNEON_8n:
+	veor		@ACC[0],@ACC[0],@ACC[0]
+	 sub		$toutptr,sp,#128
+	veor		@ACC[1],@ACC[1],@ACC[1]
+	 sub		$toutptr,$toutptr,$num,lsl#4
+	veor		@ACC[2],@ACC[2],@ACC[2]
+	 and		$toutptr,$toutptr,#-64
+	veor		@ACC[3],@ACC[3],@ACC[3]
+	 mov		sp,$toutptr			@ alloca
+	veor		@ACC[4],@ACC[4],@ACC[4]
+	 add		$toutptr,$toutptr,#256
+	veor		@ACC[5],@ACC[5],@ACC[5]
+	 sub		$inner,$num,#8
+	veor		@ACC[6],@ACC[6],@ACC[6]
+	veor		@ACC[7],@ACC[7],@ACC[7]
 
-	b		.LNEON_outer
+.LNEON_8n_init:
+	vst1.64		{@ACC[0]-@ACC[1]},[$toutptr,:256]!
+	subs		$inner,$inner,#8
+	vst1.64		{@ACC[2]-@ACC[3]},[$toutptr,:256]!
+	vst1.64		{@ACC[4]-@ACC[5]},[$toutptr,:256]!
+	vst1.64		{@ACC[6]-@ACC[7]},[$toutptr,:256]!
+	bne		.LNEON_8n_init
+
+	add		$tinptr,sp,#256
+	vld1.32		{$A0-$A3},[$aptr]!
+	add		$bnptr,sp,#8
+	vld1.32		{${M0}[0]},[$n0,:32]
+	mov		$outer,$num
+	b		.LNEON_8n_outer
 
 .align	4
-.LNEON_outer:
-	vld1.32		{${Bi}[0]}, [$bptr,:32]!
-	sub		$nptr,$nptr,$num,lsl#2		@ rewind $nptr
-	vld1.32		{$A0-$A3},  [$aptr]!
+.LNEON_8n_outer:
+	vld1.32		{${Bi}[0]},[$bptr,:32]!	@ *b++
 	veor		$zero,$zero,$zero
-	mov		$toutptr,sp
 	vzip.16		$Bi,$zero
-	sub		$inner,$num,#8
-	vadd.u64	`&Dlo("$A0xB")`,`&Dlo("$A0xB")`,$temp
+	add		$toutptr,sp,#128
+	vld1.32		{$N0-$N3},[$nptr]!
 
-	vmlal.u32	$A0xB,$Bi,${A0}[0]
-	 vld1.64	{$A3xB-$A4xB},[$tinptr,:256]!
-	vmlal.u32	$A1xB,$Bi,${A0}[1]
-	vmlal.u32	$A2xB,$Bi,${A1}[0]
-	 vld1.64	{$A5xB-$A6xB},[$tinptr,:256]!
-	vmlal.u32	$A3xB,$Bi,${A1}[1]
-
-	vshl.i64	$temp,`&Dhi("$A0xB")`,#16
+	vmlal.u32	@ACC[0],$Bi,${A0}[0]
+	vmlal.u32	@ACC[1],$Bi,${A0}[1]
 	 veor		$zero,$zero,$zero
-	vadd.u64	$temp,$temp,`&Dlo("$A0xB")`
-	 vld1.64	{$A7xB},[$tinptr,:128]!
-	vmul.u32	$Ni,$temp,$M0
-
-	vmlal.u32	$A4xB,$Bi,${A2}[0]
-	 vld1.32	{$N0-$N3}, [$nptr]!
-	vmlal.u32	$A5xB,$Bi,${A2}[1]
-	vmlal.u32	$A6xB,$Bi,${A3}[0]
+	vmlal.u32	@ACC[2],$Bi,${A1}[0]
+	 vshl.i64	$Ni,@ACC[0]#hi,#16
+	vmlal.u32	@ACC[3],$Bi,${A1}[1]
+	 vadd.u64	$Ni,$Ni,@ACC[0]#lo
+	vmlal.u32	@ACC[4],$Bi,${A2}[0]
+	 vmul.u32	$Ni,$Ni,$M0
+	vmlal.u32	@ACC[5],$Bi,${A2}[1]
+	vst1.32		{$Bi},[sp,:64]		@ put aside smashed b[8*i+0]
+	vmlal.u32	@ACC[6],$Bi,${A3}[0]
+	 vzip.16	$Ni,$zero
+	vmlal.u32	@ACC[7],$Bi,${A3}[1]
+___
+for ($i=0; $i<7;) {
+$code.=<<___;
+	vld1.32		{${Bi}[0]},[$bptr,:32]!	@ *b++
+	vmlal.u32	@ACC[0],$Ni,${N0}[0]
+	veor		$temp,$temp,$temp
+	vmlal.u32	@ACC[1],$Ni,${N0}[1]
+	vzip.16		$Bi,$temp
+	vmlal.u32	@ACC[2],$Ni,${N1}[0]
+	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
+	vmlal.u32	@ACC[3],$Ni,${N1}[1]
+	vmlal.u32	@ACC[4],$Ni,${N2}[0]
+	 vadd.u64	@ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
+	vmlal.u32	@ACC[5],$Ni,${N2}[1]
+	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
+	vmlal.u32	@ACC[6],$Ni,${N3}[0]
+	vmlal.u32	@ACC[7],$Ni,${N3}[1]
+	 vadd.u64	@ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
+	vst1.32		{$Ni},[$bnptr,:64]!	@ put aside smashed m[8*i+$i]
+___
+	push(@ACC,shift(@ACC));	$i++;
+$code.=<<___;
+	vmlal.u32	@ACC[0],$Bi,${A0}[0]
+	vld1.64		{@ACC[7]},[$tinptr,:128]!
+	vmlal.u32	@ACC[1],$Bi,${A0}[1]
+	 veor		$zero,$zero,$zero
+	vmlal.u32	@ACC[2],$Bi,${A1}[0]
+	 vshl.i64	$Ni,@ACC[0]#hi,#16
+	vmlal.u32	@ACC[3],$Bi,${A1}[1]
+	 vadd.u64	$Ni,$Ni,@ACC[0]#lo
+	vmlal.u32	@ACC[4],$Bi,${A2}[0]
+	 vmul.u32	$Ni,$Ni,$M0
+	vmlal.u32	@ACC[5],$Bi,${A2}[1]
+	vst1.32		{$Bi},[$bnptr,:64]!	@ put aside smashed b[8*i+$i]
+	vmlal.u32	@ACC[6],$Bi,${A3}[0]
 	 vzip.16	$Ni,$zero
-	vmlal.u32	$A7xB,$Bi,${A3}[1]
+	vmlal.u32	@ACC[7],$Bi,${A3}[1]
+___
+}
+$code.=<<___;
+	vld1.32		{$Bi},[sp,:64]		@ pull smashed b[8*i+0]
+	vmlal.u32	@ACC[0],$Ni,${N0}[0]
+	vld1.32		{$A0-$A3},[$aptr]!
+	vmlal.u32	@ACC[1],$Ni,${N0}[1]
+	vmlal.u32	@ACC[2],$Ni,${N1}[0]
+	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
+	vmlal.u32	@ACC[3],$Ni,${N1}[1]
+	vmlal.u32	@ACC[4],$Ni,${N2}[0]
+	 vadd.u64	@ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
+	vmlal.u32	@ACC[5],$Ni,${N2}[1]
+	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
+	vmlal.u32	@ACC[6],$Ni,${N3}[0]
+	vmlal.u32	@ACC[7],$Ni,${N3}[1]
+	 vadd.u64	@ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
+	vst1.32		{$Ni},[$bnptr,:64]	@ put aside smashed m[8*i+$i]
+	add		$bnptr,sp,#8		@ rewind
+___
+	push(@ACC,shift(@ACC));
+$code.=<<___;
+	sub		$inner,$num,#8
+	b		.LNEON_8n_inner
 
-.LNEON_inner:
-	vmlal.u32	$A0xB,$Ni,${N0}[0]
-	 vld1.32	{$A0-$A3}, [$aptr]!
-	vmlal.u32	$A1xB,$Ni,${N0}[1]
+.align	4
+.LNEON_8n_inner:
 	subs		$inner,$inner,#8
-	vmlal.u32	$A2xB,$Ni,${N1}[0]
-	vmlal.u32	$A3xB,$Ni,${N1}[1]
-	vst1.64		{$A0xB-$A1xB}, [$toutptr,:256]!
-
-	vmlal.u32	$A4xB,$Ni,${N2}[0]
-	 vld1.64	{$A0xB},       [$tinptr, :128]!
-	vmlal.u32	$A5xB,$Ni,${N2}[1]
-	vst1.64		{$A2xB-$A3xB}, [$toutptr,:256]!
-	vmlal.u32	$A6xB,$Ni,${N3}[0]
-	 vld1.64	{$A1xB-$A2xB}, [$tinptr, :256]!
-	vmlal.u32	$A7xB,$Ni,${N3}[1]
-	vst1.64		{$A4xB-$A5xB}, [$toutptr,:256]!
-
-	vmlal.u32	$A0xB,$Bi,${A0}[0]
-	 vld1.64	{$A3xB-$A4xB}, [$tinptr, :256]!
-	vmlal.u32	$A1xB,$Bi,${A0}[1]
-	vst1.64		{$A6xB-$A7xB}, [$toutptr,:256]!
-	vmlal.u32	$A2xB,$Bi,${A1}[0]
-	 vld1.64	{$A5xB-$A6xB}, [$tinptr, :256]!
-	vmlal.u32	$A3xB,$Bi,${A1}[1]
-	 vld1.32	{$N0-$N3}, [$nptr]!
-
-	vmlal.u32	$A4xB,$Bi,${A2}[0]
-	 vld1.64	{$A7xB},       [$tinptr, :128]!
-	vmlal.u32	$A5xB,$Bi,${A2}[1]
-	vmlal.u32	$A6xB,$Bi,${A3}[0]
-	vmlal.u32	$A7xB,$Bi,${A3}[1]
-
-	bne	.LNEON_inner
-
-	vmlal.u32	$A0xB,$Ni,${N0}[0]
-	add		$tinptr,sp,#16
-	vmlal.u32	$A1xB,$Ni,${N0}[1]
-	sub		$aptr,$aptr,$num,lsl#2		@ rewind $aptr
-	vmlal.u32	$A2xB,$Ni,${N1}[0]
-	 vld1.64	{$Temp}, [sp,:128]
-	vmlal.u32	$A3xB,$Ni,${N1}[1]
-	subs		$outer,$outer,#1
-
-	vmlal.u32	$A4xB,$Ni,${N2}[0]
-	vst1.64		{$A0xB-$A1xB}, [$toutptr,:256]!
-	vmlal.u32	$A5xB,$Ni,${N2}[1]
-	 vld1.64	{$A0xB},       [$tinptr, :128]!
-	vshr.u64	$temp,$temp,#16
-	vst1.64		{$A2xB-$A3xB}, [$toutptr,:256]!
-	vmlal.u32	$A6xB,$Ni,${N3}[0]
-	 vld1.64	{$A1xB-$A2xB}, [$tinptr, :256]!
-	vmlal.u32	$A7xB,$Ni,${N3}[1]
-
-	vst1.64		{$A4xB-$A5xB}, [$toutptr,:256]!
-	vadd.u64	$temp,$temp,`&Dhi("$Temp")`
-	vst1.64		{$A6xB-$A7xB}, [$toutptr,:256]!
-	vshr.u64	$temp,$temp,#16
+	vmlal.u32	@ACC[0],$Bi,${A0}[0]
+	vld1.64		{@ACC[7]},[$tinptr,:128]
+	vmlal.u32	@ACC[1],$Bi,${A0}[1]
+	vld1.32		{$Ni},[$bnptr,:64]!	@ pull smashed m[8*i+0]
+	vmlal.u32	@ACC[2],$Bi,${A1}[0]
+	vld1.32		{$N0-$N3},[$nptr]!
+	vmlal.u32	@ACC[3],$Bi,${A1}[1]
+	it		ne
+	addne		$tinptr,$tinptr,#16	@ don't advance in last iteration
+	vmlal.u32	@ACC[4],$Bi,${A2}[0]
+	vmlal.u32	@ACC[5],$Bi,${A2}[1]
+	vmlal.u32	@ACC[6],$Bi,${A3}[0]
+	vmlal.u32	@ACC[7],$Bi,${A3}[1]
+___
+for ($i=1; $i<8; $i++) {
+$code.=<<___;
+	vld1.32		{$Bi},[$bnptr,:64]!	@ pull smashed b[8*i+$i]
+	vmlal.u32	@ACC[0],$Ni,${N0}[0]
+	vmlal.u32	@ACC[1],$Ni,${N0}[1]
+	vmlal.u32	@ACC[2],$Ni,${N1}[0]
+	vmlal.u32	@ACC[3],$Ni,${N1}[1]
+	vmlal.u32	@ACC[4],$Ni,${N2}[0]
+	vmlal.u32	@ACC[5],$Ni,${N2}[1]
+	vmlal.u32	@ACC[6],$Ni,${N3}[0]
+	vmlal.u32	@ACC[7],$Ni,${N3}[1]
+	vst1.64		{@ACC[0]},[$toutptr,:128]!
+___
+	push(@ACC,shift(@ACC));
+$code.=<<___;
+	vmlal.u32	@ACC[0],$Bi,${A0}[0]
+	vld1.64		{@ACC[7]},[$tinptr,:128]
+	vmlal.u32	@ACC[1],$Bi,${A0}[1]
+	vld1.32		{$Ni},[$bnptr,:64]!	@ pull smashed m[8*i+$i]
+	vmlal.u32	@ACC[2],$Bi,${A1}[0]
+	it		ne
+	addne		$tinptr,$tinptr,#16	@ don't advance in last iteration
+	vmlal.u32	@ACC[3],$Bi,${A1}[1]
+	vmlal.u32	@ACC[4],$Bi,${A2}[0]
+	vmlal.u32	@ACC[5],$Bi,${A2}[1]
+	vmlal.u32	@ACC[6],$Bi,${A3}[0]
+	vmlal.u32	@ACC[7],$Bi,${A3}[1]
+___
+}
+$code.=<<___;
+	it		eq
+	subeq		$aptr,$aptr,$num,lsl#2	@ rewind
+	vmlal.u32	@ACC[0],$Ni,${N0}[0]
+	vld1.32		{$Bi},[sp,:64]		@ pull smashed b[8*i+0]
+	vmlal.u32	@ACC[1],$Ni,${N0}[1]
+	vld1.32		{$A0-$A3},[$aptr]!
+	vmlal.u32	@ACC[2],$Ni,${N1}[0]
+	add		$bnptr,sp,#8		@ rewind
+	vmlal.u32	@ACC[3],$Ni,${N1}[1]
+	vmlal.u32	@ACC[4],$Ni,${N2}[0]
+	vmlal.u32	@ACC[5],$Ni,${N2}[1]
+	vmlal.u32	@ACC[6],$Ni,${N3}[0]
+	vst1.64		{@ACC[0]},[$toutptr,:128]!
+	vmlal.u32	@ACC[7],$Ni,${N3}[1]
+
+	bne		.LNEON_8n_inner
+___
+	push(@ACC,shift(@ACC));
+$code.=<<___;
+	add		$tinptr,sp,#128
+	vst1.64		{@ACC[0]-@ACC[1]},[$toutptr,:256]!
+	veor		q2,q2,q2		@ $N0-$N1
+	vst1.64		{@ACC[2]-@ACC[3]},[$toutptr,:256]!
+	veor		q3,q3,q3		@ $N2-$N3
+	vst1.64		{@ACC[4]-@ACC[5]},[$toutptr,:256]!
+	vst1.64		{@ACC[6]},[$toutptr,:128]
+
+	subs		$outer,$outer,#8
+	vld1.64		{@ACC[0]-@ACC[1]},[$tinptr,:256]!
+	vld1.64		{@ACC[2]-@ACC[3]},[$tinptr,:256]!
+	vld1.64		{@ACC[4]-@ACC[5]},[$tinptr,:256]!
+	vld1.64		{@ACC[6]-@ACC[7]},[$tinptr,:256]!
 
-	bne	.LNEON_outer
+	itt		ne
+	subne		$nptr,$nptr,$num,lsl#2	@ rewind
+	bne		.LNEON_8n_outer
+
+	add		$toutptr,sp,#128
+	vst1.64		{q2-q3}, [sp,:256]!	@ start wiping stack frame
+	vshr.u64	$temp,@ACC[0]#lo,#16
+	vst1.64		{q2-q3},[sp,:256]!
+	vadd.u64	@ACC[0]#hi,@ACC[0]#hi,$temp
+	vst1.64		{q2-q3}, [sp,:256]!
+	vshr.u64	$temp,@ACC[0]#hi,#16
+	vst1.64		{q2-q3}, [sp,:256]!
+	vzip.16		@ACC[0]#lo,@ACC[0]#hi
 
-	mov		$toutptr,sp
 	mov		$inner,$num
+	b		.LNEON_tail_entry
 
+.align	4
 .LNEON_tail:
-	vadd.u64	`&Dlo("$A0xB")`,`&Dlo("$A0xB")`,$temp
-	vld1.64		{$A3xB-$A4xB}, [$tinptr, :256]!
-	vshr.u64	$temp,`&Dlo("$A0xB")`,#16
-	vadd.u64	`&Dhi("$A0xB")`,`&Dhi("$A0xB")`,$temp
-	vld1.64		{$A5xB-$A6xB}, [$tinptr, :256]!
-	vshr.u64	$temp,`&Dhi("$A0xB")`,#16
-	vld1.64		{$A7xB},       [$tinptr, :128]!
-	vzip.16		`&Dlo("$A0xB")`,`&Dhi("$A0xB")`
-
-.LNEON_tail2:
-	vadd.u64	`&Dlo("$A1xB")`,`&Dlo("$A1xB")`,$temp
-	vst1.32		{`&Dlo("$A0xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A1xB")`,#16
-	vadd.u64	`&Dhi("$A1xB")`,`&Dhi("$A1xB")`,$temp
-	vshr.u64	$temp,`&Dhi("$A1xB")`,#16
-	vzip.16		`&Dlo("$A1xB")`,`&Dhi("$A1xB")`
-
-	vadd.u64	`&Dlo("$A2xB")`,`&Dlo("$A2xB")`,$temp
-	vst1.32		{`&Dlo("$A1xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A2xB")`,#16
-	vadd.u64	`&Dhi("$A2xB")`,`&Dhi("$A2xB")`,$temp
-	vshr.u64	$temp,`&Dhi("$A2xB")`,#16
-	vzip.16		`&Dlo("$A2xB")`,`&Dhi("$A2xB")`
-
-	vadd.u64	`&Dlo("$A3xB")`,`&Dlo("$A3xB")`,$temp
-	vst1.32		{`&Dlo("$A2xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A3xB")`,#16
-	vadd.u64	`&Dhi("$A3xB")`,`&Dhi("$A3xB")`,$temp
-	vshr.u64	$temp,`&Dhi("$A3xB")`,#16
-	vzip.16		`&Dlo("$A3xB")`,`&Dhi("$A3xB")`
-
-	vadd.u64	`&Dlo("$A4xB")`,`&Dlo("$A4xB")`,$temp
-	vst1.32		{`&Dlo("$A3xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A4xB")`,#16
-	vadd.u64	`&Dhi("$A4xB")`,`&Dhi("$A4xB")`,$temp
-	vshr.u64	$temp,`&Dhi("$A4xB")`,#16
-	vzip.16		`&Dlo("$A4xB")`,`&Dhi("$A4xB")`
-
-	vadd.u64	`&Dlo("$A5xB")`,`&Dlo("$A5xB")`,$temp
-	vst1.32		{`&Dlo("$A4xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A5xB")`,#16
-	vadd.u64	`&Dhi("$A5xB")`,`&Dhi("$A5xB")`,$temp
-	vshr.u64	$temp,`&Dhi("$A5xB")`,#16
-	vzip.16		`&Dlo("$A5xB")`,`&Dhi("$A5xB")`
-
-	vadd.u64	`&Dlo("$A6xB")`,`&Dlo("$A6xB")`,$temp
-	vst1.32		{`&Dlo("$A5xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A6xB")`,#16
-	vadd.u64	`&Dhi("$A6xB")`,`&Dhi("$A6xB")`,$temp
-	vld1.64		{$A0xB}, [$tinptr, :128]!
-	vshr.u64	$temp,`&Dhi("$A6xB")`,#16
-	vzip.16		`&Dlo("$A6xB")`,`&Dhi("$A6xB")`
-
-	vadd.u64	`&Dlo("$A7xB")`,`&Dlo("$A7xB")`,$temp
-	vst1.32		{`&Dlo("$A6xB")`[0]}, [$toutptr, :32]!
-	vshr.u64	$temp,`&Dlo("$A7xB")`,#16
-	vadd.u64	`&Dhi("$A7xB")`,`&Dhi("$A7xB")`,$temp
-	vld1.64		{$A1xB-$A2xB},	[$tinptr, :256]!
-	vshr.u64	$temp,`&Dhi("$A7xB")`,#16
-	vzip.16		`&Dlo("$A7xB")`,`&Dhi("$A7xB")`
+	vadd.u64	@ACC[0]#lo,@ACC[0]#lo,$temp
+	vshr.u64	$temp,@ACC[0]#lo,#16
+	vld1.64		{@ACC[2]-@ACC[3]}, [$tinptr, :256]!
+	vadd.u64	@ACC[0]#hi,@ACC[0]#hi,$temp
+	vld1.64		{@ACC[4]-@ACC[5]}, [$tinptr, :256]!
+	vshr.u64	$temp,@ACC[0]#hi,#16
+	vld1.64		{@ACC[6]-@ACC[7]}, [$tinptr, :256]!
+	vzip.16		@ACC[0]#lo,@ACC[0]#hi
+
+.LNEON_tail_entry:
+___
+for ($i=1; $i<8; $i++) {
+$code.=<<___;
+	vadd.u64	@ACC[1]#lo,@ACC[1]#lo,$temp
+	vst1.32		{@ACC[0]#lo[0]}, [$toutptr, :32]!
+	vshr.u64	$temp,@ACC[1]#lo,#16
+	vadd.u64	@ACC[1]#hi,@ACC[1]#hi,$temp
+	vshr.u64	$temp,@ACC[1]#hi,#16
+	vzip.16		@ACC[1]#lo,@ACC[1]#hi
+___
+	push(@ACC,shift(@ACC));
+}
+	push(@ACC,shift(@ACC));
+$code.=<<___;
+	vld1.64		{@ACC[0]-@ACC[1]}, [$tinptr, :256]!
 	subs		$inner,$inner,#8
-	vst1.32		{`&Dlo("$A7xB")`[0]}, [$toutptr, :32]!
-
+	vst1.32		{@ACC[7]#lo[0]},   [$toutptr, :32]!
 	bne	.LNEON_tail
 
 	vst1.32	{${temp}[0]}, [$toutptr, :32]		@ top-most bit
@@ -708,8 +736,14 @@ $code.=<<___;
 #endif
 ___
 
-$code =~ s/\`([^\`]*)\`/eval $1/gem;
-$code =~ s/\bbx\s+lr\b/.word\t0xe12fff1e/gm;	# make it possible to compile with -march=armv4
-$code =~ s/\bret\b/bx	lr/gm;
-print $code;
+foreach (split("\n",$code)) {
+	s/\`([^\`]*)\`/eval $1/ge;
+
+	s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/ge	or
+	s/\bret\b/bx    lr/g						or
+	s/\bbx\s+lr\b/.word\t0xe12fff1e/g;	# make it possible to compile with -march=armv4
+
+	print $_,"\n";
+}
+
 close STDOUT;