chacha/asm/chacha-armv8.pl: replace 3+1 code paths with 4+1.
authorAndy Polyakov <appro@openssl.org>
Wed, 17 Apr 2019 19:31:01 +0000 (21:31 +0200)
committerAndy Polyakov <appro@openssl.org>
Wed, 17 Apr 2019 19:31:01 +0000 (21:31 +0200)
The change is triggered by ThunderX2 where 3+1 was slower than scalar
code path, but it helps all processors [to handle <512 inputs].

Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8776)

crypto/chacha/asm/chacha-armv8.pl

index dc38cbd..1f51017 100755 (executable)
 #
 # ChaCha20 for ARMv8.
 #
+# April 2019
+#
+# Replace 3xNEON+1xIALU code path with 4+1. 4+1 is actually fastest
+# option on most(*), but not all, processors, yet 6+2 is retained.
+# This is because penalties are considered tolerable in comparison to
+# improvement on processors where 6+2 helps. Most notably +37% on
+# ThunderX2. It's server-oriented processor which will have to serve
+# as many requests as possible. While others are mostly clients, when
+# performance doesn't have to be absolute top-notch, just fast enough,
+# as majority of time is spent "entertaining" relatively slow human.
+#
 # Performance in cycles per byte out of large buffer.
 #
-#                      IALU/gcc-4.9    3xNEON+1xIALU   6xNEON+2xIALU
+#                      IALU/gcc-4.9    4xNEON+1xIALU   6xNEON+2xIALU
 #
-# Apple A7             5.50/+49%       3.33            1.70
-# Cortex-A53           8.40/+80%       4.72            4.72(*)
-# Cortex-A57           8.06/+43%       4.90            4.43(**)
-# Denver               4.50/+82%       2.63            2.67(*)
-# X-Gene               9.50/+46%       8.82            8.89(*)
-# Mongoose             8.00/+44%       3.64            3.25
-# Kryo                 8.17/+50%       4.83            4.65
-# ThunderX2            7.26/+48%       7.91            4.30
+# Apple A7             5.50/+49%       2.72            1.60
+# Cortex-A53           8.40/+80%       4.06            4.45(*)
+# Cortex-A57           8.06/+43%       4.15            4.40(*)
+# Denver               4.50/+82%       2.30            2.70(*)
+# X-Gene               9.50/+46%       8.20            8.90(*)
+# Mongoose             8.00/+44%       2.74            3.12(*)
+# Kryo                 8.17/+50%       4.47            4.65(*)
+# ThunderX2            7.22/+48%       5.64            4.10
 #
-# (*)  it's expected that doubling interleave factor doesn't help
-#      all processors, only those with higher NEON latency and
-#      higher instruction issue rate;
-# (**) expected improvement was actually higher;
+# (*)  slower than 4+1:-(
 
 $flavour=shift;
 $output=shift;
@@ -121,18 +129,21 @@ my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
 }
 
 $code.=<<___;
-#include "arm_arch.h"
+#ifndef        __KERNEL__
+# include "arm_arch.h"
+.extern        OPENSSL_armcap_P
+#endif
 
 .text
 
-.extern        OPENSSL_armcap_P
-
 .align 5
 .Lsigma:
 .quad  0x3320646e61707865,0x6b20657479622d32           // endian-neutral
 .Lone:
-.long  1,0,0,0
-.asciz "ChaCha20 for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
+.long  1,2,3,4
+.Lrot24:
+.long  0x02010003,0x06050407,0x0a09080b,0x0e0d0c0f
+.asciz "ChaCha20 for ARMv8, CRYPTOGAMS by \@dot-asm"
 
 .globl ChaCha20_ctr32
 .type  ChaCha20_ctr32,%function
@@ -142,10 +153,12 @@ ChaCha20_ctr32:
        cmp     $len,#192
        b.lo    .Lshort
 
+#ifndef        __KERNEL__
        adrp    x17,OPENSSL_armcap_P
        ldr     w17,[x17,#:lo12:OPENSSL_armcap_P]
        tst     w17,#ARMV7_NEON
        b.ne    .LChaCha20_neon
+#endif
 
 .Lshort:
        .inst   0xd503233f                      // paciasp
@@ -164,7 +177,7 @@ ChaCha20_ctr32:
        ldp     @d[2],@d[3],[$key]              // load key
        ldp     @d[4],@d[5],[$key,#16]
        ldp     @d[6],@d[7],[$ctr]              // load counter
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        ror     @d[2],@d[2],#32
        ror     @d[3],@d[3],#32
        ror     @d[4],@d[4],#32
@@ -233,7 +246,7 @@ $code.=<<___;
        add     @x[14],@x[14],@x[15],lsl#32
        ldp     @x[13],@x[15],[$inp,#48]
        add     $inp,$inp,#64
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        rev     @x[0],@x[0]
        rev     @x[2],@x[2]
        rev     @x[4],@x[4]
@@ -290,7 +303,7 @@ $code.=<<___;
        add     @x[10],@x[10],@x[11],lsl#32
        add     @x[12],@x[12],@x[13],lsl#32
        add     @x[14],@x[14],@x[15],lsl#32
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        rev     @x[0],@x[0]
        rev     @x[2],@x[2]
        rev     @x[4],@x[4]
@@ -331,43 +344,87 @@ $code.=<<___;
 ___
 
 {{{
-my ($A0,$B0,$C0,$D0,$A1,$B1,$C1,$D1,$A2,$B2,$C2,$D2,$T0,$T1,$T2,$T3) =
-    map("v$_.4s",(0..7,16..23));
-my (@K)=map("v$_.4s",(24..30));
-my $ONE="v31.4s";
+my @K = map("v$_.4s",(0..3));
+my ($xt0,$xt1,$xt2,$xt3, $CTR,$ROT24) = map("v$_.4s",(4..9));
+my @X = map("v$_.4s",(16,20,24,28, 17,21,25,29, 18,22,26,30, 19,23,27,31));
+my ($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
+    $xc0,$xc1,$xc2,$xc3, $xd0,$xd1,$xd2,$xd3) = @X;
 
-sub NEONROUND {
-my $odd = pop;
-my ($a,$b,$c,$d,$t)=@_;
+sub NEON_lane_ROUND {
+my ($a0,$b0,$c0,$d0)=@_;
+my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
+my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
+my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
+my @x=map("'$_'",@X);
 
        (
-       "&add           ('$a','$a','$b')",
-       "&eor           ('$d','$d','$a')",
-       "&rev32_16      ('$d','$d')",           # vrot ($d,16)
-
-       "&add           ('$c','$c','$d')",
-       "&eor           ('$t','$b','$c')",
-       "&ushr          ('$b','$t',20)",
-       "&sli           ('$b','$t',12)",
-
-       "&add           ('$a','$a','$b')",
-       "&eor           ('$t','$d','$a')",
-       "&ushr          ('$d','$t',24)",
-       "&sli           ('$d','$t',8)",
-
-       "&add           ('$c','$c','$d')",
-       "&eor           ('$t','$b','$c')",
-       "&ushr          ('$b','$t',25)",
-       "&sli           ('$b','$t',7)",
-
-       "&ext           ('$c','$c','$c',8)",
-       "&ext           ('$d','$d','$d',$odd?4:12)",
-       "&ext           ('$b','$b','$b',$odd?12:4)"
+       "&add           (@x[$a0],@x[$a0],@x[$b0])",     # Q1
+        "&add          (@x[$a1],@x[$a1],@x[$b1])",     # Q2
+         "&add         (@x[$a2],@x[$a2],@x[$b2])",     # Q3
+          "&add        (@x[$a3],@x[$a3],@x[$b3])",     # Q4
+       "&eor           (@x[$d0],@x[$d0],@x[$a0])",
+        "&eor          (@x[$d1],@x[$d1],@x[$a1])",
+         "&eor         (@x[$d2],@x[$d2],@x[$a2])",
+          "&eor        (@x[$d3],@x[$d3],@x[$a3])",
+       "&rev32_16      (@x[$d0],@x[$d0])",
+        "&rev32_16     (@x[$d1],@x[$d1])",
+         "&rev32_16    (@x[$d2],@x[$d2])",
+          "&rev32_16   (@x[$d3],@x[$d3])",
+
+       "&add           (@x[$c0],@x[$c0],@x[$d0])",
+        "&add          (@x[$c1],@x[$c1],@x[$d1])",
+         "&add         (@x[$c2],@x[$c2],@x[$d2])",
+          "&add        (@x[$c3],@x[$c3],@x[$d3])",
+       "&eor           ('$xt0',@x[$b0],@x[$c0])",
+        "&eor          ('$xt1',@x[$b1],@x[$c1])",
+         "&eor         ('$xt2',@x[$b2],@x[$c2])",
+          "&eor        ('$xt3',@x[$b3],@x[$c3])",
+       "&ushr          (@x[$b0],'$xt0',20)",
+        "&ushr         (@x[$b1],'$xt1',20)",
+         "&ushr        (@x[$b2],'$xt2',20)",
+          "&ushr       (@x[$b3],'$xt3',20)",
+       "&sli           (@x[$b0],'$xt0',12)",
+        "&sli          (@x[$b1],'$xt1',12)",
+         "&sli         (@x[$b2],'$xt2',12)",
+          "&sli        (@x[$b3],'$xt3',12)",
+
+       "&add           (@x[$a0],@x[$a0],@x[$b0])",
+        "&add          (@x[$a1],@x[$a1],@x[$b1])",
+         "&add         (@x[$a2],@x[$a2],@x[$b2])",
+          "&add        (@x[$a3],@x[$a3],@x[$b3])",
+       "&eor           ('$xt0',@x[$d0],@x[$a0])",
+        "&eor          ('$xt1',@x[$d1],@x[$a1])",
+         "&eor         ('$xt2',@x[$d2],@x[$a2])",
+          "&eor        ('$xt3',@x[$d3],@x[$a3])",
+       "&tbl           (@x[$d0],'{$xt0}','$ROT24')",
+        "&tbl          (@x[$d1],'{$xt1}','$ROT24')",
+         "&tbl         (@x[$d2],'{$xt2}','$ROT24')",
+          "&tbl        (@x[$d3],'{$xt3}','$ROT24')",
+
+       "&add           (@x[$c0],@x[$c0],@x[$d0])",
+        "&add          (@x[$c1],@x[$c1],@x[$d1])",
+         "&add         (@x[$c2],@x[$c2],@x[$d2])",
+          "&add        (@x[$c3],@x[$c3],@x[$d3])",
+       "&eor           ('$xt0',@x[$b0],@x[$c0])",
+        "&eor          ('$xt1',@x[$b1],@x[$c1])",
+         "&eor         ('$xt2',@x[$b2],@x[$c2])",
+          "&eor        ('$xt3',@x[$b3],@x[$c3])",
+       "&ushr          (@x[$b0],'$xt0',25)",
+        "&ushr         (@x[$b1],'$xt1',25)",
+         "&ushr        (@x[$b2],'$xt2',25)",
+          "&ushr       (@x[$b3],'$xt3',25)",
+       "&sli           (@x[$b0],'$xt0',7)",
+        "&sli          (@x[$b1],'$xt1',7)",
+         "&sli         (@x[$b2],'$xt2',7)",
+          "&sli        (@x[$b3],'$xt3',7)"
        );
 }
 
 $code.=<<___;
 
+#ifdef __KERNEL__
+.globl ChaCha20_neon
+#endif
 .type  ChaCha20_neon,%function
 .align 5
 ChaCha20_neon:
@@ -394,8 +451,9 @@ ChaCha20_neon:
        ld1     {@K[1],@K[2]},[$key]
        ldp     @d[6],@d[7],[$ctr]              // load counter
        ld1     {@K[3]},[$ctr]
-       ld1     {$ONE},[@x[0]]
-#ifdef __ARMEB__
+       stp     d8,d9,[sp]                      // meet ABI requirements
+       ld1     {$CTR,$ROT24},[@x[0]]
+#ifdef __AARCH64EB__
        rev64   @K[0],@K[0]
        ror     @d[2],@d[2],#32
        ror     @d[3],@d[3],#32
@@ -404,115 +462,129 @@ ChaCha20_neon:
        ror     @d[6],@d[6],#32
        ror     @d[7],@d[7],#32
 #endif
-       add     @K[3],@K[3],$ONE                // += 1
-       add     @K[4],@K[3],$ONE
-       add     @K[5],@K[4],$ONE
-       shl     $ONE,$ONE,#2                    // 1 -> 4
 
 .Loop_outer_neon:
-       mov.32  @x[0],@d[0]                     // unpack key block
-       lsr     @x[1],@d[0],#32
-        mov    $A0,@K[0]
-       mov.32  @x[2],@d[1]
-       lsr     @x[3],@d[1],#32
-        mov    $A1,@K[0]
-       mov.32  @x[4],@d[2]
-       lsr     @x[5],@d[2],#32
-        mov    $A2,@K[0]
-       mov.32  @x[6],@d[3]
-        mov    $B0,@K[1]
-       lsr     @x[7],@d[3],#32
-        mov    $B1,@K[1]
-       mov.32  @x[8],@d[4]
-        mov    $B2,@K[1]
-       lsr     @x[9],@d[4],#32
-        mov    $D0,@K[3]
-       mov.32  @x[10],@d[5]
-        mov    $D1,@K[4]
-       lsr     @x[11],@d[5],#32
-        mov    $D2,@K[5]
-       mov.32  @x[12],@d[6]
-        mov    $C0,@K[2]
-       lsr     @x[13],@d[6],#32
-        mov    $C1,@K[2]
-       mov.32  @x[14],@d[7]
-        mov    $C2,@K[2]
-       lsr     @x[15],@d[7],#32
+       dup     $xa0,@{K[0]}[0]                 // unpack key block
+        mov.32 @x[0],@d[0]
+       dup     $xa1,@{K[0]}[1]
+        lsr    @x[1],@d[0],#32
+       dup     $xa2,@{K[0]}[2]
+        mov.32 @x[2],@d[1]
+       dup     $xa3,@{K[0]}[3]
+        lsr    @x[3],@d[1],#32
+       dup     $xb0,@{K[1]}[0]
+        mov.32 @x[4],@d[2]
+       dup     $xb1,@{K[1]}[1]
+        lsr    @x[5],@d[2],#32
+       dup     $xb2,@{K[1]}[2]
+        mov.32 @x[6],@d[3]
+       dup     $xb3,@{K[1]}[3]
+        lsr    @x[7],@d[3],#32
+       dup     $xd0,@{K[3]}[0]
+        mov.32 @x[8],@d[4]
+       dup     $xd1,@{K[3]}[1]
+        lsr    @x[9],@d[4],#32
+       dup     $xd2,@{K[3]}[2]
+        mov.32 @x[10],@d[5]
+       dup     $xd3,@{K[3]}[3]
+        lsr    @x[11],@d[5],#32
+       add     $xd0,$xd0,$CTR
+        mov.32 @x[12],@d[6]
+       dup     $xc0,@{K[2]}[0]
+        lsr    @x[13],@d[6],#32
+       dup     $xc1,@{K[2]}[1]
+        mov.32 @x[14],@d[7]
+       dup     $xc2,@{K[2]}[2]
+        lsr    @x[15],@d[7],#32
+       dup     $xc3,@{K[2]}[3]
 
        mov     $ctr,#10
-       subs    $len,$len,#256
+       subs    $len,$len,#320
 .Loop_neon:
        sub     $ctr,$ctr,#1
 ___
-       my @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,0);
-       my @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,0);
-       my @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,0);
-       my @thread3=&ROUND(0,4,8,12);
-
-       foreach (@thread0) {
-               eval;                   eval(shift(@thread3));
-               eval(shift(@thread1));  eval(shift(@thread3));
-               eval(shift(@thread2));  eval(shift(@thread3));
-       }
+       my @plus_one=&ROUND(0,4,8,12);
+       foreach (&NEON_lane_ROUND(0,4,8,12))  { eval; eval(shift(@plus_one)); }
 
-       @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,1);
-       @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,1);
-       @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,1);
-       @thread3=&ROUND(0,5,10,15);
-
-       foreach (@thread0) {
-               eval;                   eval(shift(@thread3));
-               eval(shift(@thread1));  eval(shift(@thread3));
-               eval(shift(@thread2));  eval(shift(@thread3));
-       }
+       @plus_one=&ROUND(0,5,10,15);
+       foreach (&NEON_lane_ROUND(0,5,10,15)) { eval; eval(shift(@plus_one)); }
 $code.=<<___;
        cbnz    $ctr,.Loop_neon
 
-       add.32  @x[0],@x[0],@d[0]               // accumulate key block
-        add    $A0,$A0,@K[0]
-       add     @x[1],@x[1],@d[0],lsr#32
-        add    $A1,$A1,@K[0]
-       add.32  @x[2],@x[2],@d[1]
-        add    $A2,$A2,@K[0]
-       add     @x[3],@x[3],@d[1],lsr#32
-        add    $C0,$C0,@K[2]
-       add.32  @x[4],@x[4],@d[2]
-        add    $C1,$C1,@K[2]
-       add     @x[5],@x[5],@d[2],lsr#32
-        add    $C2,$C2,@K[2]
-       add.32  @x[6],@x[6],@d[3]
-        add    $D0,$D0,@K[3]
-       add     @x[7],@x[7],@d[3],lsr#32
-       add.32  @x[8],@x[8],@d[4]
-        add    $D1,$D1,@K[4]
-       add     @x[9],@x[9],@d[4],lsr#32
-       add.32  @x[10],@x[10],@d[5]
-        add    $D2,$D2,@K[5]
-       add     @x[11],@x[11],@d[5],lsr#32
-       add.32  @x[12],@x[12],@d[6]
-        add    $B0,$B0,@K[1]
-       add     @x[13],@x[13],@d[6],lsr#32
-       add.32  @x[14],@x[14],@d[7]
-        add    $B1,$B1,@K[1]
-       add     @x[15],@x[15],@d[7],lsr#32
-        add    $B2,$B2,@K[1]
+       add     $xd0,$xd0,$CTR
+
+       zip1    $xt0,$xa0,$xa1                  // transpose data
+       zip1    $xt1,$xa2,$xa3
+       zip2    $xt2,$xa0,$xa1
+       zip2    $xt3,$xa2,$xa3
+       zip1.64 $xa0,$xt0,$xt1
+       zip2.64 $xa1,$xt0,$xt1
+       zip1.64 $xa2,$xt2,$xt3
+       zip2.64 $xa3,$xt2,$xt3
+
+       zip1    $xt0,$xb0,$xb1
+       zip1    $xt1,$xb2,$xb3
+       zip2    $xt2,$xb0,$xb1
+       zip2    $xt3,$xb2,$xb3
+       zip1.64 $xb0,$xt0,$xt1
+       zip2.64 $xb1,$xt0,$xt1
+       zip1.64 $xb2,$xt2,$xt3
+       zip2.64 $xb3,$xt2,$xt3
+
+       zip1    $xt0,$xc0,$xc1
+        add.32 @x[0],@x[0],@d[0]               // accumulate key block
+       zip1    $xt1,$xc2,$xc3
+        add    @x[1],@x[1],@d[0],lsr#32
+       zip2    $xt2,$xc0,$xc1
+        add.32 @x[2],@x[2],@d[1]
+       zip2    $xt3,$xc2,$xc3
+        add    @x[3],@x[3],@d[1],lsr#32
+       zip1.64 $xc0,$xt0,$xt1
+        add.32 @x[4],@x[4],@d[2]
+       zip2.64 $xc1,$xt0,$xt1
+        add    @x[5],@x[5],@d[2],lsr#32
+       zip1.64 $xc2,$xt2,$xt3
+        add.32 @x[6],@x[6],@d[3]
+       zip2.64 $xc3,$xt2,$xt3
+        add    @x[7],@x[7],@d[3],lsr#32
+
+       zip1    $xt0,$xd0,$xd1
+        add.32 @x[8],@x[8],@d[4]
+       zip1    $xt1,$xd2,$xd3
+        add    @x[9],@x[9],@d[4],lsr#32
+       zip2    $xt2,$xd0,$xd1
+        add.32 @x[10],@x[10],@d[5]
+       zip2    $xt3,$xd2,$xd3
+        add    @x[11],@x[11],@d[5],lsr#32
+       zip1.64 $xd0,$xt0,$xt1
+        add.32 @x[12],@x[12],@d[6]
+       zip2.64 $xd1,$xt0,$xt1
+        add    @x[13],@x[13],@d[6],lsr#32
+       zip1.64 $xd2,$xt2,$xt3
+        add.32 @x[14],@x[14],@d[7]
+       zip2.64 $xd3,$xt2,$xt3
+        add    @x[15],@x[15],@d[7],lsr#32
 
        b.lo    .Ltail_neon
 
        add     @x[0],@x[0],@x[1],lsl#32        // pack
        add     @x[2],@x[2],@x[3],lsl#32
        ldp     @x[1],@x[3],[$inp,#0]           // load input
+        add    $xa0,$xa0,@K[0]                 // accumulate key block
        add     @x[4],@x[4],@x[5],lsl#32
        add     @x[6],@x[6],@x[7],lsl#32
        ldp     @x[5],@x[7],[$inp,#16]
+        add    $xb0,$xb0,@K[1]
        add     @x[8],@x[8],@x[9],lsl#32
        add     @x[10],@x[10],@x[11],lsl#32
        ldp     @x[9],@x[11],[$inp,#32]
+        add    $xc0,$xc0,@K[2]
        add     @x[12],@x[12],@x[13],lsl#32
        add     @x[14],@x[14],@x[15],lsl#32
        ldp     @x[13],@x[15],[$inp,#48]
+        add    $xd0,$xd0,@K[3]
        add     $inp,$inp,#64
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        rev     @x[0],@x[0]
        rev     @x[2],@x[2]
        rev     @x[4],@x[4]
@@ -522,48 +594,68 @@ $code.=<<___;
        rev     @x[12],@x[12]
        rev     @x[14],@x[14]
 #endif
-       ld1.8   {$T0-$T3},[$inp],#64
+       ld1.8   {$xt0-$xt3},[$inp],#64
        eor     @x[0],@x[0],@x[1]
+        add    $xa1,$xa1,@K[0]
        eor     @x[2],@x[2],@x[3]
+        add    $xb1,$xb1,@K[1]
        eor     @x[4],@x[4],@x[5]
+        add    $xc1,$xc1,@K[2]
        eor     @x[6],@x[6],@x[7]
+        add    $xd1,$xd1,@K[3]
        eor     @x[8],@x[8],@x[9]
-        eor    $A0,$A0,$T0
+        eor    $xa0,$xa0,$xt0
+        movi   $xt0,#5
        eor     @x[10],@x[10],@x[11]
-        eor    $B0,$B0,$T1
+        eor    $xb0,$xb0,$xt1
        eor     @x[12],@x[12],@x[13]
-        eor    $C0,$C0,$T2
+        eor    $xc0,$xc0,$xt2
        eor     @x[14],@x[14],@x[15]
-        eor    $D0,$D0,$T3
-        ld1.8  {$T0-$T3},[$inp],#64
+        eor    $xd0,$xd0,$xt3
+        add    $CTR,$CTR,$xt0                  // += 5
+        ld1.8  {$xt0-$xt3},[$inp],#64
 
        stp     @x[0],@x[2],[$out,#0]           // store output
-        add    @d[6],@d[6],#4                  // increment counter
+        add    @d[6],@d[6],#5                  // increment counter
        stp     @x[4],@x[6],[$out,#16]
-        add    @K[3],@K[3],$ONE                // += 4
        stp     @x[8],@x[10],[$out,#32]
-        add    @K[4],@K[4],$ONE
        stp     @x[12],@x[14],[$out,#48]
-        add    @K[5],@K[5],$ONE
        add     $out,$out,#64
 
-       st1.8   {$A0-$D0},[$out],#64
-       ld1.8   {$A0-$D0},[$inp],#64
-
-       eor     $A1,$A1,$T0
-       eor     $B1,$B1,$T1
-       eor     $C1,$C1,$T2
-       eor     $D1,$D1,$T3
-       st1.8   {$A1-$D1},[$out],#64
-
-       eor     $A2,$A2,$A0
-       eor     $B2,$B2,$B0
-       eor     $C2,$C2,$C0
-       eor     $D2,$D2,$D0
-       st1.8   {$A2-$D2},[$out],#64
+       st1.8   {$xa0-$xd0},[$out],#64
+        add    $xa2,$xa2,@K[0]
+        add    $xb2,$xb2,@K[1]
+        add    $xc2,$xc2,@K[2]
+        add    $xd2,$xd2,@K[3]
+       ld1.8   {$xa0-$xd0},[$inp],#64
+
+       eor     $xa1,$xa1,$xt0
+       eor     $xb1,$xb1,$xt1
+       eor     $xc1,$xc1,$xt2
+       eor     $xd1,$xd1,$xt3
+       st1.8   {$xa1-$xd1},[$out],#64
+        add    $xa3,$xa3,@K[0]
+        add    $xb3,$xb3,@K[1]
+        add    $xc3,$xc3,@K[2]
+        add    $xd3,$xd3,@K[3]
+       ld1.8   {$xa1-$xd1},[$inp],#64
+
+       eor     $xa2,$xa2,$xa0
+       eor     $xb2,$xb2,$xb0
+       eor     $xc2,$xc2,$xc0
+       eor     $xd2,$xd2,$xd0
+       st1.8   {$xa2-$xd2},[$out],#64
+
+       eor     $xa3,$xa3,$xa1
+       eor     $xb3,$xb3,$xb1
+       eor     $xc3,$xc3,$xc1
+       eor     $xd3,$xd3,$xd1
+       st1.8   {$xa3-$xd3},[$out],#64
 
        b.hi    .Loop_outer_neon
 
+       ldp     d8,d9,[sp]                      // meet ABI requirements
+
        ldp     x19,x20,[x29,#16]
        add     sp,sp,#64
        ldp     x21,x22,[x29,#32]
@@ -574,8 +666,10 @@ $code.=<<___;
        .inst   0xd50323bf                      // autiasp
        ret
 
+.align 4
 .Ltail_neon:
-       add     $len,$len,#256
+       add     $len,$len,#320
+       ldp     d8,d9,[sp]                      // meet ABI requirements
        cmp     $len,#64
        b.lo    .Less_than_64
 
@@ -592,7 +686,7 @@ $code.=<<___;
        add     @x[14],@x[14],@x[15],lsl#32
        ldp     @x[13],@x[15],[$inp,#48]
        add     $inp,$inp,#64
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        rev     @x[0],@x[0]
        rev     @x[2],@x[2]
        rev     @x[4],@x[4]
@@ -612,48 +706,68 @@ $code.=<<___;
        eor     @x[14],@x[14],@x[15]
 
        stp     @x[0],@x[2],[$out,#0]           // store output
-        add    @d[6],@d[6],#4                  // increment counter
+        add    $xa0,$xa0,@K[0]                 // accumulate key block
        stp     @x[4],@x[6],[$out,#16]
+        add    $xb0,$xb0,@K[1]
        stp     @x[8],@x[10],[$out,#32]
+        add    $xc0,$xc0,@K[2]
        stp     @x[12],@x[14],[$out,#48]
+        add    $xd0,$xd0,@K[3]
        add     $out,$out,#64
        b.eq    .Ldone_neon
        sub     $len,$len,#64
        cmp     $len,#64
-       b.lo    .Less_than_128
-
-       ld1.8   {$T0-$T3},[$inp],#64
-       eor     $A0,$A0,$T0
-       eor     $B0,$B0,$T1
-       eor     $C0,$C0,$T2
-       eor     $D0,$D0,$T3
-       st1.8   {$A0-$D0},[$out],#64
+       b.lo    .Last_neon
+
+       ld1.8   {$xt0-$xt3},[$inp],#64
+       eor     $xa0,$xa0,$xt0
+       eor     $xb0,$xb0,$xt1
+       eor     $xc0,$xc0,$xt2
+       eor     $xd0,$xd0,$xt3
+       st1.8   {$xa0-$xd0},[$out],#64
        b.eq    .Ldone_neon
+
+       add     $xa0,$xa1,@K[0]
+       add     $xb0,$xb1,@K[1]
        sub     $len,$len,#64
+       add     $xc0,$xc1,@K[2]
        cmp     $len,#64
-       b.lo    .Less_than_192
-
-       ld1.8   {$T0-$T3},[$inp],#64
-       eor     $A1,$A1,$T0
-       eor     $B1,$B1,$T1
-       eor     $C1,$C1,$T2
-       eor     $D1,$D1,$T3
-       st1.8   {$A1-$D1},[$out],#64
+       add     $xd0,$xd1,@K[3]
+       b.lo    .Last_neon
+
+       ld1.8   {$xt0-$xt3},[$inp],#64
+       eor     $xa1,$xa0,$xt0
+       eor     $xb1,$xb0,$xt1
+       eor     $xc1,$xc0,$xt2
+       eor     $xd1,$xd0,$xt3
+       st1.8   {$xa1-$xd1},[$out],#64
        b.eq    .Ldone_neon
-       sub     $len,$len,#64
 
-       st1.8   {$A2-$D2},[sp]
-       b       .Last_neon
+       add     $xa0,$xa2,@K[0]
+       add     $xb0,$xb2,@K[1]
+       sub     $len,$len,#64
+       add     $xc0,$xc2,@K[2]
+       cmp     $len,#64
+       add     $xd0,$xd2,@K[3]
+       b.lo    .Last_neon
+
+       ld1.8   {$xt0-$xt3},[$inp],#64
+       eor     $xa2,$xa0,$xt0
+       eor     $xb2,$xb0,$xt1
+       eor     $xc2,$xc0,$xt2
+       eor     $xd2,$xd0,$xt3
+       st1.8   {$xa2-$xd2},[$out],#64
+       b.eq    .Ldone_neon
 
-.Less_than_128:
-       st1.8   {$A0-$D0},[sp]
-       b       .Last_neon
-.Less_than_192:
-       st1.8   {$A1-$D1},[sp]
-       b       .Last_neon
+       add     $xa0,$xa3,@K[0]
+       add     $xb0,$xb3,@K[1]
+       add     $xc0,$xc3,@K[2]
+       add     $xd0,$xd3,@K[3]
+       sub     $len,$len,#64
 
-.align 4
 .Last_neon:
+       st1.8   {$xa0-$xd0},[sp]
+
        sub     $out,$out,#1
        add     $inp,$inp,$len
        add     $out,$out,$len
@@ -686,9 +800,41 @@ $code.=<<___;
 .size  ChaCha20_neon,.-ChaCha20_neon
 ___
 {
+my @K = map("v$_.4s",(0..6));
 my ($T0,$T1,$T2,$T3,$T4,$T5)=@K;
 my ($A0,$B0,$C0,$D0,$A1,$B1,$C1,$D1,$A2,$B2,$C2,$D2,
-    $A3,$B3,$C3,$D3,$A4,$B4,$C4,$D4,$A5,$B5,$C5,$D5) = map("v$_.4s",(0..23));
+    $A3,$B3,$C3,$D3,$A4,$B4,$C4,$D4,$A5,$B5,$C5,$D5) = map("v$_.4s",(8..31));
+my $rot24 = @K[6];
+my $ONE = "v7.4s";
+
+sub NEONROUND {
+my $odd = pop;
+my ($a,$b,$c,$d,$t)=@_;
+
+       (
+       "&add           ('$a','$a','$b')",
+       "&eor           ('$d','$d','$a')",
+       "&rev32_16      ('$d','$d')",           # vrot ($d,16)
+
+       "&add           ('$c','$c','$d')",
+       "&eor           ('$t','$b','$c')",
+       "&ushr          ('$b','$t',20)",
+       "&sli           ('$b','$t',12)",
+
+       "&add           ('$a','$a','$b')",
+       "&eor           ('$d','$d','$a')",
+       "&tbl           ('$d','{$d}','$rot24')",
+
+       "&add           ('$c','$c','$d')",
+       "&eor           ('$t','$b','$c')",
+       "&ushr          ('$b','$t',25)",
+       "&sli           ('$b','$t',7)",
+
+       "&ext           ('$c','$c','$c',8)",
+       "&ext           ('$d','$d','$d',$odd?4:12)",
+       "&ext           ('$b','$b','$b',$odd?12:4)"
+       );
+}
 
 $code.=<<___;
 .type  ChaCha20_512_neon,%function
@@ -708,6 +854,7 @@ ChaCha20_512_neon:
 .L512_or_more_neon:
        sub     sp,sp,#128+64
 
+       eor     $ONE,$ONE,$ONE
        ldp     @d[0],@d[1],[@x[0]]             // load sigma
        ld1     {@K[0]},[@x[0]],#16
        ldp     @d[2],@d[3],[$key]              // load key
@@ -715,8 +862,9 @@ ChaCha20_512_neon:
        ld1     {@K[1],@K[2]},[$key]
        ldp     @d[6],@d[7],[$ctr]              // load counter
        ld1     {@K[3]},[$ctr]
-       ld1     {$ONE},[@x[0]]
-#ifdef __ARMEB__
+       ld1     {$ONE}[0],[@x[0]]
+       add     $key,@x[0],#16                  // .Lrot24
+#ifdef __AARCH64EB__
        rev64   @K[0],@K[0]
        ror     @d[2],@d[2],#32
        ror     @d[3],@d[3],#32
@@ -783,9 +931,10 @@ ChaCha20_512_neon:
         mov    $C4,@K[2]
         stp    @K[3],@K[4],[sp,#48]            // off-load key block, variable part
         mov    $C5,@K[2]
-        str    @K[5],[sp,#80]
+        stp    @K[5],@K[6],[sp,#80]
 
        mov     $ctr,#5
+       ld1     {$rot24},[$key]
        subs    $len,$len,#512
 .Loop_upper_neon:
        sub     $ctr,$ctr,#1
@@ -858,7 +1007,7 @@ $code.=<<___;
        add     @x[14],@x[14],@x[15],lsl#32
        ldp     @x[13],@x[15],[$inp,#48]
        add     $inp,$inp,#64
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        rev     @x[0],@x[0]
        rev     @x[2],@x[2]
        rev     @x[4],@x[4]
@@ -947,6 +1096,7 @@ $code.=<<___;
        add.32  @x[2],@x[2],@d[1]
         ldp    @K[4],@K[5],[sp,#64]
        add     @x[3],@x[3],@d[1],lsr#32
+        ldr    @K[6],[sp,#96]
         add    $A0,$A0,@K[0]
        add.32  @x[4],@x[4],@d[2]
         add    $A1,$A1,@K[0]
@@ -999,7 +1149,7 @@ $code.=<<___;
        add     $inp,$inp,#64
         add    $B5,$B5,@K[1]
 
-#ifdef __ARMEB__
+#ifdef __AARCH64EB__
        rev     @x[0],@x[0]
        rev     @x[2],@x[2]
        rev     @x[4],@x[4]
@@ -1077,24 +1227,24 @@ $code.=<<___;
        b.hs    .Loop_outer_512_neon
 
        adds    $len,$len,#512
-       ushr    $A0,$ONE,#2                     // 4 -> 1
+       ushr    $ONE,$ONE,#1                    // 4 -> 2
 
        ldp     d8,d9,[sp,#128+0]               // meet ABI requirements
        ldp     d10,d11,[sp,#128+16]
        ldp     d12,d13,[sp,#128+32]
        ldp     d14,d15,[sp,#128+48]
 
-       stp     @K[0],$ONE,[sp,#0]              // wipe off-load area
-       stp     @K[0],$ONE,[sp,#32]
-       stp     @K[0],$ONE,[sp,#64]
+       stp     @K[0],@K[0],[sp,#0]             // wipe off-load area
+       stp     @K[0],@K[0],[sp,#32]
+       stp     @K[0],@K[0],[sp,#64]
 
        b.eq    .Ldone_512_neon
 
+       sub     $key,$key,#16                   // .Lone
        cmp     $len,#192
-       sub     @K[3],@K[3],$A0                 // -= 1
-       sub     @K[4],@K[4],$A0
-       sub     @K[5],@K[5],$A0
        add     sp,sp,#128
+       sub     @K[3],@K[3],$ONE                // -= 2
+       ld1     {$CTR,$ROT24},[$key]
        b.hs    .Loop_outer_neon
 
        eor     @K[1],@K[1],@K[1]
@@ -1124,9 +1274,11 @@ foreach (split("\n",$code)) {
        s/\`([^\`]*)\`/eval $1/geo;
 
        (s/\b([a-z]+)\.32\b/$1/ and (s/x([0-9]+)/w$1/g or 1))   or
-       (m/\b(eor|ext|mov)\b/ and (s/\.4s/\.16b/g or 1))        or
+       (m/\b(eor|ext|mov|tbl)\b/ and (s/\.4s/\.16b/g or 1))    or
        (s/\b((?:ld|st)1)\.8\b/$1/ and (s/\.4s/\.16b/g or 1))   or
        (m/\b(ld|st)[rp]\b/ and (s/v([0-9]+)\.4s/q$1/g or 1))   or
+       (m/\b(dup|ld1)\b/ and (s/\.4(s}?\[[0-3]\])/.$1/g or 1)) or
+       (s/\b(zip[12])\.64\b/$1/ and (s/\.4s/\.2d/g or 1))      or
        (s/\brev32\.16\b/rev32/ and (s/\.4s/\.8h/g or 1));
 
        #s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo;