mirror of
https://github.com/ossrs/srs.git
synced 2025-03-09 15:49:59 +00:00
Upgrade openssl from 1.1.0e to 1.1.1b, with source code. 4.0.78
This commit is contained in:
winlin
parent
8f1c992379
commit
96dbd7bced
1476 changed files with 616554 additions and 4 deletions
1160
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-armv4.pl
vendored
Executable file
1160
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-armv4.pl
vendored
Executable file
File diff suppressed because it is too large
Load diff
1144
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-armv8.pl
vendored
Executable file
1144
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-armv8.pl
vendored
Executable file
File diff suppressed because it is too large
Load diff
926
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-c64xplus.pl
vendored
Executable file
926
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-c64xplus.pl
vendored
Executable file
|
|
@ -0,0 +1,926 @@
|
|||
#! /usr/bin/env perl
|
||||
# Copyright 2016 The OpenSSL Project Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the OpenSSL license (the "License"). You may not use
|
||||
# this file except in compliance with the License. You can obtain a copy
|
||||
# in the file LICENSE in the source distribution or at
|
||||
# https://www.openssl.org/source/license.html
|
||||
|
||||
#
|
||||
# ====================================================================
|
||||
# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
|
||||
# project. The module is, however, dual licensed under OpenSSL and
|
||||
# CRYPTOGAMS licenses depending on where you obtain it. For further
|
||||
# details see http://www.openssl.org/~appro/cryptogams/.
|
||||
# ====================================================================
|
||||
#
|
||||
# ChaCha20 for C64x+.
|
||||
#
|
||||
# October 2015
|
||||
#
|
||||
# Performance is 3.54 cycles per processed byte, which is ~4.3 times
|
||||
# faster than code generated by TI compiler. Compiler also disables
|
||||
# interrupts for some reason, thus making interrupt response time
|
||||
# dependent on input length. This module on the other hand is free
|
||||
# from such limitation.
|
||||
|
||||
$output=pop;
|
||||
open STDOUT,">$output";
|
||||
|
||||
($OUT,$INP,$LEN,$KEYB,$COUNTERA)=("A4","B4","A6","B6","A8");
|
||||
($KEYA,$COUNTERB,$STEP)=("A7","B7","A3");
|
||||
|
||||
@X= ("A16","B16","A17","B17","A18","B18","A19","B19",
|
||||
"A20","B20","A21","B21","A22","B22","A23","B23");
|
||||
@Y= ("A24","B24","A25","B25","A26","B26","A27","B27",
|
||||
"A28","B28","A29","B29","A30","B30","A31","B31");
|
||||
@DAT=("A6", "A7", "B6", "B7", "A8", "A9", "B8", "B9",
|
||||
"A10","A11","B10","B11","A12","A13","B12","B13");
|
||||
|
||||
# yes, overlaps with @DAT, used only in 2x interleave code path...
|
||||
@K2x=("A6", "B6", "A7", "B7", "A8", "B8", "A9", "B9",
|
||||
"A10","B10","A11","B11","A2", "B2", "A13","B13");
|
||||
|
||||
$code.=<<___;
|
||||
.text
|
||||
|
||||
.if .ASSEMBLER_VERSION<7000000
|
||||
.asg 0,__TI_EABI__
|
||||
.endif
|
||||
.if __TI_EABI__
|
||||
.asg ChaCha20_ctr32,_ChaCha20_ctr32
|
||||
.endif
|
||||
|
||||
.asg B3,RA
|
||||
.asg A15,FP
|
||||
.asg B15,SP
|
||||
|
||||
.global _ChaCha20_ctr32
|
||||
.align 32
|
||||
_ChaCha20_ctr32:
|
||||
.asmfunc stack_usage(40+64)
|
||||
MV $LEN,A0 ; reassign
|
||||
[!A0] BNOP RA ; no data
|
||||
|| [A0] STW FP,*SP--(40+64) ; save frame pointer and alloca(40+64)
|
||||
|| [A0] MV SP,FP
|
||||
[A0] STDW B13:B12,*SP[4+8] ; ABI says so
|
||||
|| [A0] MV $KEYB,$KEYA
|
||||
|| [A0] MV $COUNTERA,$COUNTERB
|
||||
[A0] STDW B11:B10,*SP[3+8]
|
||||
|| [A0] STDW A13:A12,*FP[-3]
|
||||
[A0] STDW A11:A10,*FP[-4]
|
||||
|| [A0] MVK 128,$STEP ; 2 * input block size
|
||||
|
||||
[A0] LDW *${KEYA}[0],@Y[4] ; load key
|
||||
|| [A0] LDW *${KEYB}[1],@Y[5]
|
||||
|| [A0] MVK 0x00007865,@Y[0] ; synthesize sigma
|
||||
|| [A0] MVK 0x0000646e,@Y[1]
|
||||
[A0] LDW *${KEYA}[2],@Y[6]
|
||||
|| [A0] LDW *${KEYB}[3],@Y[7]
|
||||
|| [A0] MVKH 0x61700000,@Y[0]
|
||||
|| [A0] MVKH 0x33200000,@Y[1]
|
||||
LDW *${KEYA}[4],@Y[8]
|
||||
|| LDW *${KEYB}[5],@Y[9]
|
||||
|| MVK 0x00002d32,@Y[2]
|
||||
|| MVK 0x00006574,@Y[3]
|
||||
LDW *${KEYA}[6],@Y[10]
|
||||
|| LDW *${KEYB}[7],@Y[11]
|
||||
|| MVKH 0x79620000,@Y[2]
|
||||
|| MVKH 0x6b200000,@Y[3]
|
||||
LDW *${COUNTERA}[0],@Y[12] ; load counter||nonce
|
||||
|| LDW *${COUNTERB}[1],@Y[13]
|
||||
|| CMPLTU A0,$STEP,A1 ; is length < 2*blocks?
|
||||
LDW *${COUNTERA}[2],@Y[14]
|
||||
|| LDW *${COUNTERB}[3],@Y[15]
|
||||
|| [A1] BNOP top1x?
|
||||
[A1] MVK 64,$STEP ; input block size
|
||||
|| MVK 10,B0 ; inner loop counter
|
||||
|
||||
DMV @Y[2],@Y[0],@X[2]:@X[0] ; copy block
|
||||
|| DMV @Y[3],@Y[1],@X[3]:@X[1]
|
||||
||[!A1] STDW @Y[2]:@Y[0],*FP[-12] ; offload key material to stack
|
||||
||[!A1] STDW @Y[3]:@Y[1],*SP[2]
|
||||
DMV @Y[6],@Y[4],@X[6]:@X[4]
|
||||
|| DMV @Y[7],@Y[5],@X[7]:@X[5]
|
||||
||[!A1] STDW @Y[6]:@Y[4],*FP[-10]
|
||||
||[!A1] STDW @Y[7]:@Y[5],*SP[4]
|
||||
DMV @Y[10],@Y[8],@X[10]:@X[8]
|
||||
|| DMV @Y[11],@Y[9],@X[11]:@X[9]
|
||||
||[!A1] STDW @Y[10]:@Y[8],*FP[-8]
|
||||
||[!A1] STDW @Y[11]:@Y[9],*SP[6]
|
||||
DMV @Y[14],@Y[12],@X[14]:@X[12]
|
||||
|| DMV @Y[15],@Y[13],@X[15]:@X[13]
|
||||
||[!A1] MV @Y[12],@K2x[12] ; counter
|
||||
||[!A1] MV @Y[13],@K2x[13]
|
||||
||[!A1] STW @Y[14],*FP[-6*2]
|
||||
||[!A1] STW @Y[15],*SP[8*2]
|
||||
___
|
||||
{ ################################################################
|
||||
# 2x interleave gives 50% performance improvement
|
||||
#
|
||||
my ($a0,$a1,$a2,$a3) = (0..3);
|
||||
my ($b0,$b1,$b2,$b3) = (4..7);
|
||||
my ($c0,$c1,$c2,$c3) = (8..11);
|
||||
my ($d0,$d1,$d2,$d3) = (12..15);
|
||||
|
||||
$code.=<<___;
|
||||
outer2x?:
|
||||
ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
|| ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
|| ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| DMV @Y[2],@Y[0],@K2x[2]:@K2x[0]
|
||||
|| DMV @Y[3],@Y[1],@K2x[3]:@K2x[1]
|
||||
XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
|| XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
|| XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| DMV @Y[6],@Y[4],@K2x[6]:@K2x[4]
|
||||
|| DMV @Y[7],@Y[5],@K2x[7]:@K2x[5]
|
||||
SWAP2 @X[$d1],@X[$d1] ; rotate by 16
|
||||
|| SWAP2 @X[$d2],@X[$d2]
|
||||
|| SWAP2 @X[$d0],@X[$d0]
|
||||
|| SWAP2 @X[$d3],@X[$d3]
|
||||
|
||||
ADD @X[$d1],@X[$c1],@X[$c1]
|
||||
|| ADD @X[$d2],@X[$c2],@X[$c2]
|
||||
|| ADD @X[$d0],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c3],@X[$c3]
|
||||
|| DMV @Y[10],@Y[8],@K2x[10]:@K2x[8]
|
||||
|| DMV @Y[11],@Y[9],@K2x[11]:@K2x[9]
|
||||
XOR @X[$c1],@X[$b1],@X[$b1]
|
||||
|| XOR @X[$c2],@X[$b2],@X[$b2]
|
||||
|| XOR @X[$c0],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b3],@X[$b3]
|
||||
|| ADD 1,@Y[12],@Y[12] ; adjust counter for 2nd block
|
||||
ROTL @X[$b1],12,@X[$b1]
|
||||
|| ROTL @X[$b2],12,@X[$b2]
|
||||
|| MV @Y[14],@K2x[14]
|
||||
|| MV @Y[15],@K2x[15]
|
||||
top2x?:
|
||||
ROTL @X[$b0],12,@X[$b0]
|
||||
|| ROTL @X[$b3],12,@X[$b3]
|
||||
|| ADD @Y[$b1],@Y[$a1],@Y[$a1]
|
||||
|| ADD @Y[$b2],@Y[$a2],@Y[$a2]
|
||||
ADD @Y[$b0],@Y[$a0],@Y[$a0]
|
||||
|| ADD @Y[$b3],@Y[$a3],@Y[$a3]
|
||||
|
||||
|| ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
|| ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
|| XOR @Y[$a1],@Y[$d1],@Y[$d1]
|
||||
|| XOR @Y[$a2],@Y[$d2],@Y[$d2]
|
||||
XOR @Y[$a0],@Y[$d0],@Y[$d0]
|
||||
|| XOR @Y[$a3],@Y[$d3],@Y[$d3]
|
||||
|| ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
|| XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| ROTL @X[$d1],8,@X[$d1]
|
||||
|| ROTL @X[$d2],8,@X[$d2]
|
||||
|| SWAP2 @Y[$d1],@Y[$d1] ; rotate by 16
|
||||
|| SWAP2 @Y[$d2],@Y[$d2]
|
||||
|| SWAP2 @Y[$d0],@Y[$d0]
|
||||
|| SWAP2 @Y[$d3],@Y[$d3]
|
||||
ROTL @X[$d0],8,@X[$d0]
|
||||
|| ROTL @X[$d3],8,@X[$d3]
|
||||
|| ADD @Y[$d1],@Y[$c1],@Y[$c1]
|
||||
|| ADD @Y[$d2],@Y[$c2],@Y[$c2]
|
||||
|| ADD @Y[$d0],@Y[$c0],@Y[$c0]
|
||||
|| ADD @Y[$d3],@Y[$c3],@Y[$c3]
|
||||
|| BNOP middle2x1? ; protect from interrupt
|
||||
|
||||
ADD @X[$d1],@X[$c1],@X[$c1]
|
||||
|| ADD @X[$d2],@X[$c2],@X[$c2]
|
||||
|| XOR @Y[$c1],@Y[$b1],@Y[$b1]
|
||||
|| XOR @Y[$c2],@Y[$b2],@Y[$b2]
|
||||
|| XOR @Y[$c0],@Y[$b0],@Y[$b0]
|
||||
|| XOR @Y[$c3],@Y[$b3],@Y[$b3]
|
||||
ADD @X[$d0],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c3],@X[$c3]
|
||||
|| XOR @X[$c1],@X[$b1],@X[$b1]
|
||||
|| XOR @X[$c2],@X[$b2],@X[$b2]
|
||||
|| ROTL @X[$d1],0,@X[$d2] ; moved to avoid cross-path stall
|
||||
|| ROTL @X[$d2],0,@X[$d3]
|
||||
XOR @X[$c0],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b3],@X[$b3]
|
||||
|| MV @X[$d0],@X[$d1]
|
||||
|| MV @X[$d3],@X[$d0]
|
||||
|| ROTL @Y[$b1],12,@Y[$b1]
|
||||
|| ROTL @Y[$b2],12,@Y[$b2]
|
||||
ROTL @X[$b1],7,@X[$b0] ; avoided cross-path stall
|
||||
|| ROTL @X[$b2],7,@X[$b1]
|
||||
ROTL @X[$b0],7,@X[$b3]
|
||||
|| ROTL @X[$b3],7,@X[$b2]
|
||||
middle2x1?:
|
||||
|
||||
ROTL @Y[$b0],12,@Y[$b0]
|
||||
|| ROTL @Y[$b3],12,@Y[$b3]
|
||||
|| ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|
||||
|| ADD @Y[$b1],@Y[$a1],@Y[$a1]
|
||||
|| ADD @Y[$b2],@Y[$a2],@Y[$a2]
|
||||
|| XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| ADD @Y[$b0],@Y[$a0],@Y[$a0]
|
||||
|| ADD @Y[$b3],@Y[$a3],@Y[$a3]
|
||||
|| XOR @Y[$a1],@Y[$d1],@Y[$d1]
|
||||
|| XOR @Y[$a2],@Y[$d2],@Y[$d2]
|
||||
XOR @Y[$a0],@Y[$d0],@Y[$d0]
|
||||
|| XOR @Y[$a3],@Y[$d3],@Y[$d3]
|
||||
|| ROTL @Y[$d1],8,@Y[$d1]
|
||||
|| ROTL @Y[$d2],8,@Y[$d2]
|
||||
|| SWAP2 @X[$d0],@X[$d0] ; rotate by 16
|
||||
|| SWAP2 @X[$d1],@X[$d1]
|
||||
|| SWAP2 @X[$d2],@X[$d2]
|
||||
|| SWAP2 @X[$d3],@X[$d3]
|
||||
ROTL @Y[$d0],8,@Y[$d0]
|
||||
|| ROTL @Y[$d3],8,@Y[$d3]
|
||||
|| ADD @X[$d0],@X[$c2],@X[$c2]
|
||||
|| ADD @X[$d1],@X[$c3],@X[$c3]
|
||||
|| ADD @X[$d2],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c1],@X[$c1]
|
||||
|| BNOP middle2x2? ; protect from interrupt
|
||||
|
||||
ADD @Y[$d1],@Y[$c1],@Y[$c1]
|
||||
|| ADD @Y[$d2],@Y[$c2],@Y[$c2]
|
||||
|| XOR @X[$c2],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b1],@X[$b1]
|
||||
|| XOR @X[$c0],@X[$b2],@X[$b2]
|
||||
|| XOR @X[$c1],@X[$b3],@X[$b3]
|
||||
ADD @Y[$d0],@Y[$c0],@Y[$c0]
|
||||
|| ADD @Y[$d3],@Y[$c3],@Y[$c3]
|
||||
|| XOR @Y[$c1],@Y[$b1],@Y[$b1]
|
||||
|| XOR @Y[$c2],@Y[$b2],@Y[$b2]
|
||||
|| ROTL @Y[$d1],0,@Y[$d2] ; moved to avoid cross-path stall
|
||||
|| ROTL @Y[$d2],0,@Y[$d3]
|
||||
XOR @Y[$c0],@Y[$b0],@Y[$b0]
|
||||
|| XOR @Y[$c3],@Y[$b3],@Y[$b3]
|
||||
|| MV @Y[$d0],@Y[$d1]
|
||||
|| MV @Y[$d3],@Y[$d0]
|
||||
|| ROTL @X[$b0],12,@X[$b0]
|
||||
|| ROTL @X[$b1],12,@X[$b1]
|
||||
ROTL @Y[$b1],7,@Y[$b0] ; avoided cross-path stall
|
||||
|| ROTL @Y[$b2],7,@Y[$b1]
|
||||
ROTL @Y[$b0],7,@Y[$b3]
|
||||
|| ROTL @Y[$b3],7,@Y[$b2]
|
||||
middle2x2?:
|
||||
|
||||
ROTL @X[$b2],12,@X[$b2]
|
||||
|| ROTL @X[$b3],12,@X[$b3]
|
||||
|| ADD @Y[$b0],@Y[$a0],@Y[$a0]
|
||||
|| ADD @Y[$b1],@Y[$a1],@Y[$a1]
|
||||
ADD @Y[$b2],@Y[$a2],@Y[$a2]
|
||||
|| ADD @Y[$b3],@Y[$a3],@Y[$a3]
|
||||
|
||||
|| ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
|| XOR @Y[$a0],@Y[$d0],@Y[$d0]
|
||||
|| XOR @Y[$a1],@Y[$d1],@Y[$d1]
|
||||
XOR @Y[$a2],@Y[$d2],@Y[$d2]
|
||||
|| XOR @Y[$a3],@Y[$d3],@Y[$d3]
|
||||
|| ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| ROTL @X[$d0],8,@X[$d0]
|
||||
|| ROTL @X[$d1],8,@X[$d1]
|
||||
|| SWAP2 @Y[$d0],@Y[$d0] ; rotate by 16
|
||||
|| SWAP2 @Y[$d1],@Y[$d1]
|
||||
|| SWAP2 @Y[$d2],@Y[$d2]
|
||||
|| SWAP2 @Y[$d3],@Y[$d3]
|
||||
ROTL @X[$d2],8,@X[$d2]
|
||||
|| ROTL @X[$d3],8,@X[$d3]
|
||||
|| ADD @Y[$d0],@Y[$c2],@Y[$c2]
|
||||
|| ADD @Y[$d1],@Y[$c3],@Y[$c3]
|
||||
|| ADD @Y[$d2],@Y[$c0],@Y[$c0]
|
||||
|| ADD @Y[$d3],@Y[$c1],@Y[$c1]
|
||||
|| BNOP bottom2x1? ; protect from interrupt
|
||||
|
||||
ADD @X[$d0],@X[$c2],@X[$c2]
|
||||
|| ADD @X[$d1],@X[$c3],@X[$c3]
|
||||
|| XOR @Y[$c2],@Y[$b0],@Y[$b0]
|
||||
|| XOR @Y[$c3],@Y[$b1],@Y[$b1]
|
||||
|| XOR @Y[$c0],@Y[$b2],@Y[$b2]
|
||||
|| XOR @Y[$c1],@Y[$b3],@Y[$b3]
|
||||
ADD @X[$d2],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c1],@X[$c1]
|
||||
|| XOR @X[$c2],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b1],@X[$b1]
|
||||
|| ROTL @X[$d0],0,@X[$d3] ; moved to avoid cross-path stall
|
||||
|| ROTL @X[$d1],0,@X[$d0]
|
||||
XOR @X[$c0],@X[$b2],@X[$b2]
|
||||
|| XOR @X[$c1],@X[$b3],@X[$b3]
|
||||
|| MV @X[$d2],@X[$d1]
|
||||
|| MV @X[$d3],@X[$d2]
|
||||
|| ROTL @Y[$b0],12,@Y[$b0]
|
||||
|| ROTL @Y[$b1],12,@Y[$b1]
|
||||
ROTL @X[$b0],7,@X[$b1] ; avoided cross-path stall
|
||||
|| ROTL @X[$b1],7,@X[$b2]
|
||||
ROTL @X[$b2],7,@X[$b3]
|
||||
|| ROTL @X[$b3],7,@X[$b0]
|
||||
|| [B0] SUB B0,1,B0 ; decrement inner loop counter
|
||||
bottom2x1?:
|
||||
|
||||
ROTL @Y[$b2],12,@Y[$b2]
|
||||
|| ROTL @Y[$b3],12,@Y[$b3]
|
||||
|| [B0] ADD @X[$b1],@X[$a1],@X[$a1] ; modulo-scheduled
|
||||
|| [B0] ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
[B0] ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| [B0] ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|
||||
|| ADD @Y[$b0],@Y[$a0],@Y[$a0]
|
||||
|| ADD @Y[$b1],@Y[$a1],@Y[$a1]
|
||||
|| [B0] XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
|| [B0] XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
[B0] XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| [B0] XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| ADD @Y[$b2],@Y[$a2],@Y[$a2]
|
||||
|| ADD @Y[$b3],@Y[$a3],@Y[$a3]
|
||||
|| XOR @Y[$a0],@Y[$d0],@Y[$d0]
|
||||
|| XOR @Y[$a1],@Y[$d1],@Y[$d1]
|
||||
XOR @Y[$a2],@Y[$d2],@Y[$d2]
|
||||
|| XOR @Y[$a3],@Y[$d3],@Y[$d3]
|
||||
|| ROTL @Y[$d0],8,@Y[$d0]
|
||||
|| ROTL @Y[$d1],8,@Y[$d1]
|
||||
|| [B0] SWAP2 @X[$d1],@X[$d1] ; rotate by 16
|
||||
|| [B0] SWAP2 @X[$d2],@X[$d2]
|
||||
|| [B0] SWAP2 @X[$d0],@X[$d0]
|
||||
|| [B0] SWAP2 @X[$d3],@X[$d3]
|
||||
ROTL @Y[$d2],8,@Y[$d2]
|
||||
|| ROTL @Y[$d3],8,@Y[$d3]
|
||||
|| [B0] ADD @X[$d1],@X[$c1],@X[$c1]
|
||||
|| [B0] ADD @X[$d2],@X[$c2],@X[$c2]
|
||||
|| [B0] ADD @X[$d0],@X[$c0],@X[$c0]
|
||||
|| [B0] ADD @X[$d3],@X[$c3],@X[$c3]
|
||||
|| [B0] BNOP top2x? ; even protects from interrupt
|
||||
|
||||
ADD @Y[$d0],@Y[$c2],@Y[$c2]
|
||||
|| ADD @Y[$d1],@Y[$c3],@Y[$c3]
|
||||
|| [B0] XOR @X[$c1],@X[$b1],@X[$b1]
|
||||
|| [B0] XOR @X[$c2],@X[$b2],@X[$b2]
|
||||
|| [B0] XOR @X[$c0],@X[$b0],@X[$b0]
|
||||
|| [B0] XOR @X[$c3],@X[$b3],@X[$b3]
|
||||
ADD @Y[$d2],@Y[$c0],@Y[$c0]
|
||||
|| ADD @Y[$d3],@Y[$c1],@Y[$c1]
|
||||
|| XOR @Y[$c2],@Y[$b0],@Y[$b0]
|
||||
|| XOR @Y[$c3],@Y[$b1],@Y[$b1]
|
||||
|| ROTL @Y[$d0],0,@Y[$d3] ; moved to avoid cross-path stall
|
||||
|| ROTL @Y[$d1],0,@Y[$d0]
|
||||
XOR @Y[$c0],@Y[$b2],@Y[$b2]
|
||||
|| XOR @Y[$c1],@Y[$b3],@Y[$b3]
|
||||
|| MV @Y[$d2],@Y[$d1]
|
||||
|| MV @Y[$d3],@Y[$d2]
|
||||
|| [B0] ROTL @X[$b1],12,@X[$b1]
|
||||
|| [B0] ROTL @X[$b2],12,@X[$b2]
|
||||
ROTL @Y[$b0],7,@Y[$b1] ; avoided cross-path stall
|
||||
|| ROTL @Y[$b1],7,@Y[$b2]
|
||||
ROTL @Y[$b2],7,@Y[$b3]
|
||||
|| ROTL @Y[$b3],7,@Y[$b0]
|
||||
bottom2x2?:
|
||||
___
|
||||
}
|
||||
|
||||
$code.=<<___;
|
||||
ADD @K2x[0],@X[0],@X[0] ; accumulate key material
|
||||
|| ADD @K2x[1],@X[1],@X[1]
|
||||
|| ADD @K2x[2],@X[2],@X[2]
|
||||
|| ADD @K2x[3],@X[3],@X[3]
|
||||
ADD @K2x[0],@Y[0],@Y[0]
|
||||
|| ADD @K2x[1],@Y[1],@Y[1]
|
||||
|| ADD @K2x[2],@Y[2],@Y[2]
|
||||
|| ADD @K2x[3],@Y[3],@Y[3]
|
||||
|| LDNDW *${INP}++[8],@DAT[1]:@DAT[0]
|
||||
ADD @K2x[4],@X[4],@X[4]
|
||||
|| ADD @K2x[5],@X[5],@X[5]
|
||||
|| ADD @K2x[6],@X[6],@X[6]
|
||||
|| ADD @K2x[7],@X[7],@X[7]
|
||||
|| LDNDW *${INP}[-7],@DAT[3]:@DAT[2]
|
||||
ADD @K2x[4],@Y[4],@Y[4]
|
||||
|| ADD @K2x[5],@Y[5],@Y[5]
|
||||
|| ADD @K2x[6],@Y[6],@Y[6]
|
||||
|| ADD @K2x[7],@Y[7],@Y[7]
|
||||
|| LDNDW *${INP}[-6],@DAT[5]:@DAT[4]
|
||||
ADD @K2x[8],@X[8],@X[8]
|
||||
|| ADD @K2x[9],@X[9],@X[9]
|
||||
|| ADD @K2x[10],@X[10],@X[10]
|
||||
|| ADD @K2x[11],@X[11],@X[11]
|
||||
|| LDNDW *${INP}[-5],@DAT[7]:@DAT[6]
|
||||
ADD @K2x[8],@Y[8],@Y[8]
|
||||
|| ADD @K2x[9],@Y[9],@Y[9]
|
||||
|| ADD @K2x[10],@Y[10],@Y[10]
|
||||
|| ADD @K2x[11],@Y[11],@Y[11]
|
||||
|| LDNDW *${INP}[-4],@DAT[9]:@DAT[8]
|
||||
ADD @K2x[12],@X[12],@X[12]
|
||||
|| ADD @K2x[13],@X[13],@X[13]
|
||||
|| ADD @K2x[14],@X[14],@X[14]
|
||||
|| ADD @K2x[15],@X[15],@X[15]
|
||||
|| LDNDW *${INP}[-3],@DAT[11]:@DAT[10]
|
||||
ADD @K2x[12],@Y[12],@Y[12]
|
||||
|| ADD @K2x[13],@Y[13],@Y[13]
|
||||
|| ADD @K2x[14],@Y[14],@Y[14]
|
||||
|| ADD @K2x[15],@Y[15],@Y[15]
|
||||
|| LDNDW *${INP}[-2],@DAT[13]:@DAT[12]
|
||||
ADD 1,@Y[12],@Y[12] ; adjust counter for 2nd block
|
||||
|| ADD 2,@K2x[12],@K2x[12] ; increment counter
|
||||
|| LDNDW *${INP}[-1],@DAT[15]:@DAT[14]
|
||||
|
||||
.if .BIG_ENDIAN
|
||||
SWAP2 @X[0],@X[0]
|
||||
|| SWAP2 @X[1],@X[1]
|
||||
|| SWAP2 @X[2],@X[2]
|
||||
|| SWAP2 @X[3],@X[3]
|
||||
SWAP2 @X[4],@X[4]
|
||||
|| SWAP2 @X[5],@X[5]
|
||||
|| SWAP2 @X[6],@X[6]
|
||||
|| SWAP2 @X[7],@X[7]
|
||||
SWAP2 @X[8],@X[8]
|
||||
|| SWAP2 @X[9],@X[9]
|
||||
|| SWAP4 @X[0],@X[1]
|
||||
|| SWAP4 @X[1],@X[0]
|
||||
SWAP2 @X[10],@X[10]
|
||||
|| SWAP2 @X[11],@X[11]
|
||||
|| SWAP4 @X[2],@X[3]
|
||||
|| SWAP4 @X[3],@X[2]
|
||||
SWAP2 @X[12],@X[12]
|
||||
|| SWAP2 @X[13],@X[13]
|
||||
|| SWAP4 @X[4],@X[5]
|
||||
|| SWAP4 @X[5],@X[4]
|
||||
SWAP2 @X[14],@X[14]
|
||||
|| SWAP2 @X[15],@X[15]
|
||||
|| SWAP4 @X[6],@X[7]
|
||||
|| SWAP4 @X[7],@X[6]
|
||||
SWAP4 @X[8],@X[9]
|
||||
|| SWAP4 @X[9],@X[8]
|
||||
|| SWAP2 @Y[0],@Y[0]
|
||||
|| SWAP2 @Y[1],@Y[1]
|
||||
SWAP4 @X[10],@X[11]
|
||||
|| SWAP4 @X[11],@X[10]
|
||||
|| SWAP2 @Y[2],@Y[2]
|
||||
|| SWAP2 @Y[3],@Y[3]
|
||||
SWAP4 @X[12],@X[13]
|
||||
|| SWAP4 @X[13],@X[12]
|
||||
|| SWAP2 @Y[4],@Y[4]
|
||||
|| SWAP2 @Y[5],@Y[5]
|
||||
SWAP4 @X[14],@X[15]
|
||||
|| SWAP4 @X[15],@X[14]
|
||||
|| SWAP2 @Y[6],@Y[6]
|
||||
|| SWAP2 @Y[7],@Y[7]
|
||||
SWAP2 @Y[8],@Y[8]
|
||||
|| SWAP2 @Y[9],@Y[9]
|
||||
|| SWAP4 @Y[0],@Y[1]
|
||||
|| SWAP4 @Y[1],@Y[0]
|
||||
SWAP2 @Y[10],@Y[10]
|
||||
|| SWAP2 @Y[11],@Y[11]
|
||||
|| SWAP4 @Y[2],@Y[3]
|
||||
|| SWAP4 @Y[3],@Y[2]
|
||||
SWAP2 @Y[12],@Y[12]
|
||||
|| SWAP2 @Y[13],@Y[13]
|
||||
|| SWAP4 @Y[4],@Y[5]
|
||||
|| SWAP4 @Y[5],@Y[4]
|
||||
SWAP2 @Y[14],@Y[14]
|
||||
|| SWAP2 @Y[15],@Y[15]
|
||||
|| SWAP4 @Y[6],@Y[7]
|
||||
|| SWAP4 @Y[7],@Y[6]
|
||||
SWAP4 @Y[8],@Y[9]
|
||||
|| SWAP4 @Y[9],@Y[8]
|
||||
SWAP4 @Y[10],@Y[11]
|
||||
|| SWAP4 @Y[11],@Y[10]
|
||||
SWAP4 @Y[12],@Y[13]
|
||||
|| SWAP4 @Y[13],@Y[12]
|
||||
SWAP4 @Y[14],@Y[15]
|
||||
|| SWAP4 @Y[15],@Y[14]
|
||||
.endif
|
||||
|
||||
XOR @DAT[0],@X[0],@X[0] ; xor 1st block
|
||||
|| XOR @DAT[3],@X[3],@X[3]
|
||||
|| XOR @DAT[2],@X[2],@X[1]
|
||||
|| XOR @DAT[1],@X[1],@X[2]
|
||||
|| LDNDW *${INP}++[8],@DAT[1]:@DAT[0]
|
||||
XOR @DAT[4],@X[4],@X[4]
|
||||
|| XOR @DAT[7],@X[7],@X[7]
|
||||
|| LDNDW *${INP}[-7],@DAT[3]:@DAT[2]
|
||||
XOR @DAT[6],@X[6],@X[5]
|
||||
|| XOR @DAT[5],@X[5],@X[6]
|
||||
|| LDNDW *${INP}[-6],@DAT[5]:@DAT[4]
|
||||
XOR @DAT[8],@X[8],@X[8]
|
||||
|| XOR @DAT[11],@X[11],@X[11]
|
||||
|| LDNDW *${INP}[-5],@DAT[7]:@DAT[6]
|
||||
XOR @DAT[10],@X[10],@X[9]
|
||||
|| XOR @DAT[9],@X[9],@X[10]
|
||||
|| LDNDW *${INP}[-4],@DAT[9]:@DAT[8]
|
||||
XOR @DAT[12],@X[12],@X[12]
|
||||
|| XOR @DAT[15],@X[15],@X[15]
|
||||
|| LDNDW *${INP}[-3],@DAT[11]:@DAT[10]
|
||||
XOR @DAT[14],@X[14],@X[13]
|
||||
|| XOR @DAT[13],@X[13],@X[14]
|
||||
|| LDNDW *${INP}[-2],@DAT[13]:@DAT[12]
|
||||
[A0] SUB A0,$STEP,A0 ; SUB A0,128,A0
|
||||
|| LDNDW *${INP}[-1],@DAT[15]:@DAT[14]
|
||||
|
||||
XOR @Y[0],@DAT[0],@DAT[0] ; xor 2nd block
|
||||
|| XOR @Y[1],@DAT[1],@DAT[1]
|
||||
|| STNDW @X[2]:@X[0],*${OUT}++[8]
|
||||
XOR @Y[2],@DAT[2],@DAT[2]
|
||||
|| XOR @Y[3],@DAT[3],@DAT[3]
|
||||
|| STNDW @X[3]:@X[1],*${OUT}[-7]
|
||||
XOR @Y[4],@DAT[4],@DAT[4]
|
||||
|| [A0] LDDW *FP[-12],@X[2]:@X[0] ; re-load key material from stack
|
||||
|| [A0] LDDW *SP[2], @X[3]:@X[1]
|
||||
XOR @Y[5],@DAT[5],@DAT[5]
|
||||
|| STNDW @X[6]:@X[4],*${OUT}[-6]
|
||||
XOR @Y[6],@DAT[6],@DAT[6]
|
||||
|| XOR @Y[7],@DAT[7],@DAT[7]
|
||||
|| STNDW @X[7]:@X[5],*${OUT}[-5]
|
||||
XOR @Y[8],@DAT[8],@DAT[8]
|
||||
|| [A0] LDDW *FP[-10],@X[6]:@X[4]
|
||||
|| [A0] LDDW *SP[4], @X[7]:@X[5]
|
||||
XOR @Y[9],@DAT[9],@DAT[9]
|
||||
|| STNDW @X[10]:@X[8],*${OUT}[-4]
|
||||
XOR @Y[10],@DAT[10],@DAT[10]
|
||||
|| XOR @Y[11],@DAT[11],@DAT[11]
|
||||
|| STNDW @X[11]:@X[9],*${OUT}[-3]
|
||||
XOR @Y[12],@DAT[12],@DAT[12]
|
||||
|| [A0] LDDW *FP[-8], @X[10]:@X[8]
|
||||
|| [A0] LDDW *SP[6], @X[11]:@X[9]
|
||||
XOR @Y[13],@DAT[13],@DAT[13]
|
||||
|| STNDW @X[14]:@X[12],*${OUT}[-2]
|
||||
XOR @Y[14],@DAT[14],@DAT[14]
|
||||
|| XOR @Y[15],@DAT[15],@DAT[15]
|
||||
|| STNDW @X[15]:@X[13],*${OUT}[-1]
|
||||
|
||||
[A0] MV @K2x[12],@X[12]
|
||||
|| [A0] MV @K2x[13],@X[13]
|
||||
|| [A0] LDW *FP[-6*2], @X[14]
|
||||
|| [A0] LDW *SP[8*2], @X[15]
|
||||
|
||||
[A0] DMV @X[2],@X[0],@Y[2]:@Y[0] ; duplicate key material
|
||||
|| STNDW @DAT[1]:@DAT[0],*${OUT}++[8]
|
||||
[A0] DMV @X[3],@X[1],@Y[3]:@Y[1]
|
||||
|| STNDW @DAT[3]:@DAT[2],*${OUT}[-7]
|
||||
[A0] DMV @X[6],@X[4],@Y[6]:@Y[4]
|
||||
|| STNDW @DAT[5]:@DAT[4],*${OUT}[-6]
|
||||
|| CMPLTU A0,$STEP,A1 ; is remaining length < 2*blocks?
|
||||
||[!A0] BNOP epilogue?
|
||||
[A0] DMV @X[7],@X[5],@Y[7]:@Y[5]
|
||||
|| STNDW @DAT[7]:@DAT[6],*${OUT}[-5]
|
||||
||[!A1] BNOP outer2x?
|
||||
[A0] DMV @X[10],@X[8],@Y[10]:@Y[8]
|
||||
|| STNDW @DAT[9]:@DAT[8],*${OUT}[-4]
|
||||
[A0] DMV @X[11],@X[9],@Y[11]:@Y[9]
|
||||
|| STNDW @DAT[11]:@DAT[10],*${OUT}[-3]
|
||||
[A0] DMV @X[14],@X[12],@Y[14]:@Y[12]
|
||||
|| STNDW @DAT[13]:@DAT[12],*${OUT}[-2]
|
||||
[A0] DMV @X[15],@X[13],@Y[15]:@Y[13]
|
||||
|| STNDW @DAT[15]:@DAT[14],*${OUT}[-1]
|
||||
;;===== branch to epilogue? is taken here
|
||||
[A1] MVK 64,$STEP
|
||||
|| [A0] MVK 10,B0 ; inner loop counter
|
||||
;;===== branch to outer2x? is taken here
|
||||
___
|
||||
{
|
||||
my ($a0,$a1,$a2,$a3) = (0..3);
|
||||
my ($b0,$b1,$b2,$b3) = (4..7);
|
||||
my ($c0,$c1,$c2,$c3) = (8..11);
|
||||
my ($d0,$d1,$d2,$d3) = (12..15);
|
||||
|
||||
$code.=<<___;
|
||||
top1x?:
|
||||
ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
|| ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
|| XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| SWAP2 @X[$d1],@X[$d1] ; rotate by 16
|
||||
|| SWAP2 @X[$d2],@X[$d2]
|
||||
SWAP2 @X[$d0],@X[$d0]
|
||||
|| SWAP2 @X[$d3],@X[$d3]
|
||||
|
||||
|| ADD @X[$d1],@X[$c1],@X[$c1]
|
||||
|| ADD @X[$d2],@X[$c2],@X[$c2]
|
||||
ADD @X[$d0],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c3],@X[$c3]
|
||||
|| XOR @X[$c1],@X[$b1],@X[$b1]
|
||||
|| XOR @X[$c2],@X[$b2],@X[$b2]
|
||||
XOR @X[$c0],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b3],@X[$b3]
|
||||
|| ROTL @X[$b1],12,@X[$b1]
|
||||
|| ROTL @X[$b2],12,@X[$b2]
|
||||
ROTL @X[$b0],12,@X[$b0]
|
||||
|| ROTL @X[$b3],12,@X[$b3]
|
||||
|
||||
ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
|| ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
|| XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| ROTL @X[$d1],8,@X[$d1]
|
||||
|| ROTL @X[$d2],8,@X[$d2]
|
||||
ROTL @X[$d0],8,@X[$d0]
|
||||
|| ROTL @X[$d3],8,@X[$d3]
|
||||
|| BNOP middle1x? ; protect from interrupt
|
||||
|
||||
ADD @X[$d1],@X[$c1],@X[$c1]
|
||||
|| ADD @X[$d2],@X[$c2],@X[$c2]
|
||||
ADD @X[$d0],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c3],@X[$c3]
|
||||
|| XOR @X[$c1],@X[$b1],@X[$b1]
|
||||
|| XOR @X[$c2],@X[$b2],@X[$b2]
|
||||
|| ROTL @X[$d1],0,@X[$d2] ; moved to avoid cross-path stall
|
||||
|| ROTL @X[$d2],0,@X[$d3]
|
||||
XOR @X[$c0],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b3],@X[$b3]
|
||||
|| ROTL @X[$d0],0,@X[$d1]
|
||||
|| ROTL @X[$d3],0,@X[$d0]
|
||||
ROTL @X[$b1],7,@X[$b0] ; avoided cross-path stall
|
||||
|| ROTL @X[$b2],7,@X[$b1]
|
||||
ROTL @X[$b0],7,@X[$b3]
|
||||
|| ROTL @X[$b3],7,@X[$b2]
|
||||
middle1x?:
|
||||
|
||||
ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| SWAP2 @X[$d0],@X[$d0] ; rotate by 16
|
||||
|| SWAP2 @X[$d1],@X[$d1]
|
||||
SWAP2 @X[$d2],@X[$d2]
|
||||
|| SWAP2 @X[$d3],@X[$d3]
|
||||
|
||||
|| ADD @X[$d0],@X[$c2],@X[$c2]
|
||||
|| ADD @X[$d1],@X[$c3],@X[$c3]
|
||||
ADD @X[$d2],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c1],@X[$c1]
|
||||
|| XOR @X[$c2],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b1],@X[$b1]
|
||||
XOR @X[$c0],@X[$b2],@X[$b2]
|
||||
|| XOR @X[$c1],@X[$b3],@X[$b3]
|
||||
|| ROTL @X[$b0],12,@X[$b0]
|
||||
|| ROTL @X[$b1],12,@X[$b1]
|
||||
ROTL @X[$b2],12,@X[$b2]
|
||||
|| ROTL @X[$b3],12,@X[$b3]
|
||||
|
||||
ADD @X[$b0],@X[$a0],@X[$a0]
|
||||
|| ADD @X[$b1],@X[$a1],@X[$a1]
|
||||
|| [B0] SUB B0,1,B0 ; decrement inner loop counter
|
||||
ADD @X[$b2],@X[$a2],@X[$a2]
|
||||
|| ADD @X[$b3],@X[$a3],@X[$a3]
|
||||
|| XOR @X[$a0],@X[$d0],@X[$d0]
|
||||
|| XOR @X[$a1],@X[$d1],@X[$d1]
|
||||
XOR @X[$a2],@X[$d2],@X[$d2]
|
||||
|| XOR @X[$a3],@X[$d3],@X[$d3]
|
||||
|| ROTL @X[$d0],8,@X[$d0]
|
||||
|| ROTL @X[$d1],8,@X[$d1]
|
||||
ROTL @X[$d2],8,@X[$d2]
|
||||
|| ROTL @X[$d3],8,@X[$d3]
|
||||
|| [B0] BNOP top1x? ; even protects from interrupt
|
||||
|
||||
ADD @X[$d0],@X[$c2],@X[$c2]
|
||||
|| ADD @X[$d1],@X[$c3],@X[$c3]
|
||||
ADD @X[$d2],@X[$c0],@X[$c0]
|
||||
|| ADD @X[$d3],@X[$c1],@X[$c1]
|
||||
|| XOR @X[$c2],@X[$b0],@X[$b0]
|
||||
|| XOR @X[$c3],@X[$b1],@X[$b1]
|
||||
|| ROTL @X[$d0],0,@X[$d3] ; moved to avoid cross-path stall
|
||||
|| ROTL @X[$d1],0,@X[$d0]
|
||||
XOR @X[$c0],@X[$b2],@X[$b2]
|
||||
|| XOR @X[$c1],@X[$b3],@X[$b3]
|
||||
|| ROTL @X[$d2],0,@X[$d1]
|
||||
|| ROTL @X[$d3],0,@X[$d2]
|
||||
ROTL @X[$b0],7,@X[$b1] ; avoided cross-path stall
|
||||
|| ROTL @X[$b1],7,@X[$b2]
|
||||
ROTL @X[$b2],7,@X[$b3]
|
||||
|| ROTL @X[$b3],7,@X[$b0]
|
||||
||[!B0] CMPLTU A0,$STEP,A1 ; less than 64 bytes left?
|
||||
bottom1x?:
|
||||
___
|
||||
}
|
||||
|
||||
$code.=<<___;
|
||||
ADD @Y[0],@X[0],@X[0] ; accumulate key material
|
||||
|| ADD @Y[1],@X[1],@X[1]
|
||||
|| ADD @Y[2],@X[2],@X[2]
|
||||
|| ADD @Y[3],@X[3],@X[3]
|
||||
||[!A1] LDNDW *${INP}++[8],@DAT[1]:@DAT[0]
|
||||
|| [A1] BNOP tail?
|
||||
ADD @Y[4],@X[4],@X[4]
|
||||
|| ADD @Y[5],@X[5],@X[5]
|
||||
|| ADD @Y[6],@X[6],@X[6]
|
||||
|| ADD @Y[7],@X[7],@X[7]
|
||||
||[!A1] LDNDW *${INP}[-7],@DAT[3]:@DAT[2]
|
||||
ADD @Y[8],@X[8],@X[8]
|
||||
|| ADD @Y[9],@X[9],@X[9]
|
||||
|| ADD @Y[10],@X[10],@X[10]
|
||||
|| ADD @Y[11],@X[11],@X[11]
|
||||
||[!A1] LDNDW *${INP}[-6],@DAT[5]:@DAT[4]
|
||||
ADD @Y[12],@X[12],@X[12]
|
||||
|| ADD @Y[13],@X[13],@X[13]
|
||||
|| ADD @Y[14],@X[14],@X[14]
|
||||
|| ADD @Y[15],@X[15],@X[15]
|
||||
||[!A1] LDNDW *${INP}[-5],@DAT[7]:@DAT[6]
|
||||
[!A1] LDNDW *${INP}[-4],@DAT[9]:@DAT[8]
|
||||
[!A1] LDNDW *${INP}[-3],@DAT[11]:@DAT[10]
|
||||
LDNDW *${INP}[-2],@DAT[13]:@DAT[12]
|
||||
LDNDW *${INP}[-1],@DAT[15]:@DAT[14]
|
||||
|
||||
.if .BIG_ENDIAN
|
||||
SWAP2 @X[0],@X[0]
|
||||
|| SWAP2 @X[1],@X[1]
|
||||
|| SWAP2 @X[2],@X[2]
|
||||
|| SWAP2 @X[3],@X[3]
|
||||
SWAP2 @X[4],@X[4]
|
||||
|| SWAP2 @X[5],@X[5]
|
||||
|| SWAP2 @X[6],@X[6]
|
||||
|| SWAP2 @X[7],@X[7]
|
||||
SWAP2 @X[8],@X[8]
|
||||
|| SWAP2 @X[9],@X[9]
|
||||
|| SWAP4 @X[0],@X[1]
|
||||
|| SWAP4 @X[1],@X[0]
|
||||
SWAP2 @X[10],@X[10]
|
||||
|| SWAP2 @X[11],@X[11]
|
||||
|| SWAP4 @X[2],@X[3]
|
||||
|| SWAP4 @X[3],@X[2]
|
||||
SWAP2 @X[12],@X[12]
|
||||
|| SWAP2 @X[13],@X[13]
|
||||
|| SWAP4 @X[4],@X[5]
|
||||
|| SWAP4 @X[5],@X[4]
|
||||
SWAP2 @X[14],@X[14]
|
||||
|| SWAP2 @X[15],@X[15]
|
||||
|| SWAP4 @X[6],@X[7]
|
||||
|| SWAP4 @X[7],@X[6]
|
||||
SWAP4 @X[8],@X[9]
|
||||
|| SWAP4 @X[9],@X[8]
|
||||
SWAP4 @X[10],@X[11]
|
||||
|| SWAP4 @X[11],@X[10]
|
||||
SWAP4 @X[12],@X[13]
|
||||
|| SWAP4 @X[13],@X[12]
|
||||
SWAP4 @X[14],@X[15]
|
||||
|| SWAP4 @X[15],@X[14]
|
||||
.else
|
||||
NOP 1
|
||||
.endif
|
||||
|
||||
XOR @X[0],@DAT[0],@DAT[0] ; xor with input
|
||||
|| XOR @X[1],@DAT[1],@DAT[1]
|
||||
|| XOR @X[2],@DAT[2],@DAT[2]
|
||||
|| XOR @X[3],@DAT[3],@DAT[3]
|
||||
|| [A0] SUB A0,$STEP,A0 ; SUB A0,64,A0
|
||||
XOR @X[4],@DAT[4],@DAT[4]
|
||||
|| XOR @X[5],@DAT[5],@DAT[5]
|
||||
|| XOR @X[6],@DAT[6],@DAT[6]
|
||||
|| XOR @X[7],@DAT[7],@DAT[7]
|
||||
|| STNDW @DAT[1]:@DAT[0],*${OUT}++[8]
|
||||
XOR @X[8],@DAT[8],@DAT[8]
|
||||
|| XOR @X[9],@DAT[9],@DAT[9]
|
||||
|| XOR @X[10],@DAT[10],@DAT[10]
|
||||
|| XOR @X[11],@DAT[11],@DAT[11]
|
||||
|| STNDW @DAT[3]:@DAT[2],*${OUT}[-7]
|
||||
XOR @X[12],@DAT[12],@DAT[12]
|
||||
|| XOR @X[13],@DAT[13],@DAT[13]
|
||||
|| XOR @X[14],@DAT[14],@DAT[14]
|
||||
|| XOR @X[15],@DAT[15],@DAT[15]
|
||||
|| STNDW @DAT[5]:@DAT[4],*${OUT}[-6]
|
||||
|| [A0] BNOP top1x?
|
||||
[A0] DMV @Y[2],@Y[0],@X[2]:@X[0] ; duplicate key material
|
||||
|| [A0] DMV @Y[3],@Y[1],@X[3]:@X[1]
|
||||
|| STNDW @DAT[7]:@DAT[6],*${OUT}[-5]
|
||||
[A0] DMV @Y[6],@Y[4],@X[6]:@X[4]
|
||||
|| [A0] DMV @Y[7],@Y[5],@X[7]:@X[5]
|
||||
|| STNDW @DAT[9]:@DAT[8],*${OUT}[-4]
|
||||
[A0] DMV @Y[10],@Y[8],@X[10]:@X[8]
|
||||
|| [A0] DMV @Y[11],@Y[9],@X[11]:@X[9]
|
||||
|| [A0] ADD 1,@Y[12],@Y[12] ; increment counter
|
||||
|| STNDW @DAT[11]:@DAT[10],*${OUT}[-3]
|
||||
[A0] DMV @Y[14],@Y[12],@X[14]:@X[12]
|
||||
|| [A0] DMV @Y[15],@Y[13],@X[15]:@X[13]
|
||||
|| STNDW @DAT[13]:@DAT[12],*${OUT}[-2]
|
||||
[A0] MVK 10,B0 ; inner loop counter
|
||||
|| STNDW @DAT[15]:@DAT[14],*${OUT}[-1]
|
||||
;;===== branch to top1x? is taken here
|
||||
|
||||
epilogue?:
|
||||
LDDW *FP[-4],A11:A10 ; ABI says so
|
||||
LDDW *FP[-3],A13:A12
|
||||
|| LDDW *SP[3+8],B11:B10
|
||||
LDDW *SP[4+8],B13:B12
|
||||
|| BNOP RA
|
||||
LDW *++SP(40+64),FP ; restore frame pointer
|
||||
NOP 4
|
||||
|
||||
tail?:
|
||||
LDBU *${INP}++[1],B24 ; load byte by byte
|
||||
|| SUB A0,1,A0
|
||||
|| SUB A0,1,B1
|
||||
[!B1] BNOP epilogue? ; interrupts are disabled for whole time
|
||||
|| [A0] LDBU *${INP}++[1],B24
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
[!B1] BNOP epilogue?
|
||||
|| [A0] LDBU *${INP}++[1],B24
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
[!B1] BNOP epilogue?
|
||||
|| ROTL @X[0],0,A24
|
||||
|| [A0] LDBU *${INP}++[1],B24
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
[!B1] BNOP epilogue?
|
||||
|| ROTL @X[0],24,A24
|
||||
|| [A0] LDBU *${INP}++[1],A24
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
[!B1] BNOP epilogue?
|
||||
|| ROTL @X[0],16,A24
|
||||
|| [A0] LDBU *${INP}++[1],A24
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
|| XOR A24,B24,B25
|
||||
STB B25,*${OUT}++[1] ; store byte by byte
|
||||
||[!B1] BNOP epilogue?
|
||||
|| ROTL @X[0],8,A24
|
||||
|| [A0] LDBU *${INP}++[1],A24
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
|| XOR A24,B24,B25
|
||||
STB B25,*${OUT}++[1]
|
||||
___
|
||||
sub TAIL_STEP {
|
||||
my $Xi= shift;
|
||||
my $T = ($Xi=~/^B/?"B24":"A24"); # match @X[i] to avoid cross path
|
||||
my $D = $T; $D=~tr/AB/BA/;
|
||||
my $O = $D; $O=~s/24/25/;
|
||||
|
||||
$code.=<<___;
|
||||
||[!B1] BNOP epilogue?
|
||||
|| ROTL $Xi,0,$T
|
||||
|| [A0] LDBU *${INP}++[1],$D
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
|| XOR A24,B24,$O
|
||||
STB $O,*${OUT}++[1]
|
||||
||[!B1] BNOP epilogue?
|
||||
|| ROTL $Xi,24,$T
|
||||
|| [A0] LDBU *${INP}++[1],$T
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
|| XOR A24,B24,$O
|
||||
STB $O,*${OUT}++[1]
|
||||
||[!B1] BNOP epilogue?
|
||||
|| ROTL $Xi,16,$T
|
||||
|| [A0] LDBU *${INP}++[1],$T
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
|| XOR A24,B24,$O
|
||||
STB $O,*${OUT}++[1]
|
||||
||[!B1] BNOP epilogue?
|
||||
|| ROTL $Xi,8,$T
|
||||
|| [A0] LDBU *${INP}++[1],$T
|
||||
|| [A0] SUB A0,1,A0
|
||||
|| SUB B1,1,B1
|
||||
|| XOR A24,B24,$O
|
||||
STB $O,*${OUT}++[1]
|
||||
___
|
||||
}
|
||||
foreach (1..14) { TAIL_STEP(@X[$_]); }
|
||||
$code.=<<___;
|
||||
||[!B1] BNOP epilogue?
|
||||
|| ROTL @X[15],0,B24
|
||||
|| XOR A24,B24,A25
|
||||
STB A25,*${OUT}++[1]
|
||||
|| ROTL @X[15],24,B24
|
||||
|| XOR A24,B24,A25
|
||||
STB A25,*${OUT}++[1]
|
||||
|| ROTL @X[15],16,B24
|
||||
|| XOR A24,B24,A25
|
||||
STB A25,*${OUT}++[1]
|
||||
|| XOR A24,B24,A25
|
||||
STB A25,*${OUT}++[1]
|
||||
|| XOR A24,B24,B25
|
||||
STB B25,*${OUT}++[1]
|
||||
.endasmfunc
|
||||
|
||||
.sect .const
|
||||
.cstring "ChaCha20 for C64x+, CRYPTOGAMS by <appro\@openssl.org>"
|
||||
.align 4
|
||||
___
|
||||
|
||||
print $code;
|
||||
close STDOUT;
|
||||
1348
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-ppc.pl
vendored
Executable file
1348
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-ppc.pl
vendored
Executable file
File diff suppressed because it is too large
Load diff
326
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-s390x.pl
vendored
Executable file
326
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-s390x.pl
vendored
Executable file
|
|
@ -0,0 +1,326 @@
|
|||
#! /usr/bin/env perl
|
||||
# Copyright 2016 The OpenSSL Project Authors. All Rights Reserved.
|
||||
#
|
||||
# Licensed under the OpenSSL license (the "License"). You may not use
|
||||
# this file except in compliance with the License. You can obtain a copy
|
||||
# in the file LICENSE in the source distribution or at
|
||||
# https://www.openssl.org/source/license.html
|
||||
|
||||
#
|
||||
# ====================================================================
|
||||
# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
|
||||
# project. The module is, however, dual licensed under OpenSSL and
|
||||
# CRYPTOGAMS licenses depending on where you obtain it. For further
|
||||
# details see http://www.openssl.org/~appro/cryptogams/.
|
||||
# ====================================================================
|
||||
#
|
||||
# December 2015
|
||||
#
|
||||
# ChaCha20 for s390x.
|
||||
#
|
||||
# 3 times faster than compiler-generated code.
|
||||
|
||||
$flavour = shift;
|
||||
|
||||
if ($flavour =~ /3[12]/) {
|
||||
$SIZE_T=4;
|
||||
$g="";
|
||||
} else {
|
||||
$SIZE_T=8;
|
||||
$g="g";
|
||||
}
|
||||
|
||||
while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {}
|
||||
open STDOUT,">$output";
|
||||
|
||||
sub AUTOLOAD() # thunk [simplified] x86-style perlasm
|
||||
{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
|
||||
$code .= "\t$opcode\t".join(',',@_)."\n";
|
||||
}
|
||||
|
||||
my $sp="%r15";
|
||||
|
||||
my $stdframe=16*$SIZE_T+4*8;
|
||||
my $frame=$stdframe+4*20;
|
||||
|
||||
my ($out,$inp,$len,$key,$counter)=map("%r$_",(2..6));
|
||||
|
||||
my @x=map("%r$_",(0..7,"x","x","x","x",(10..13)));
|
||||
my @t=map("%r$_",(8,9));
|
||||
|
||||
sub 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 ($xc,$xc_)=map("\"$_\"",@t);
|
||||
my @x=map("\"$_\"",@x);
|
||||
|
||||
# Consider order in which variables are addressed by their
|
||||
# index:
|
||||
#
|
||||
# a b c d
|
||||
#
|
||||
# 0 4 8 12 < even round
|
||||
# 1 5 9 13
|
||||
# 2 6 10 14
|
||||
# 3 7 11 15
|
||||
# 0 5 10 15 < odd round
|
||||
# 1 6 11 12
|
||||
# 2 7 8 13
|
||||
# 3 4 9 14
|
||||
#
|
||||
# 'a', 'b' and 'd's are permanently allocated in registers,
|
||||
# @x[0..7,12..15], while 'c's are maintained in memory. If
|
||||
# you observe 'c' column, you'll notice that pair of 'c's is
|
||||
# invariant between rounds. This means that we have to reload
|
||||
# them once per round, in the middle. This is why you'll see
|
||||
# 'c' stores and loads in the middle, but none in the beginning
|
||||
# or end.
|
||||
|
||||
(
|
||||
"&alr (@x[$a0],@x[$b0])", # Q1
|
||||
"&alr (@x[$a1],@x[$b1])", # Q2
|
||||
"&xr (@x[$d0],@x[$a0])",
|
||||
"&xr (@x[$d1],@x[$a1])",
|
||||
"&rll (@x[$d0],@x[$d0],16)",
|
||||
"&rll (@x[$d1],@x[$d1],16)",
|
||||
|
||||
"&alr ($xc,@x[$d0])",
|
||||
"&alr ($xc_,@x[$d1])",
|
||||
"&xr (@x[$b0],$xc)",
|
||||
"&xr (@x[$b1],$xc_)",
|
||||
"&rll (@x[$b0],@x[$b0],12)",
|
||||
"&rll (@x[$b1],@x[$b1],12)",
|
||||
|
||||
"&alr (@x[$a0],@x[$b0])",
|
||||
"&alr (@x[$a1],@x[$b1])",
|
||||
"&xr (@x[$d0],@x[$a0])",
|
||||
"&xr (@x[$d1],@x[$a1])",
|
||||
"&rll (@x[$d0],@x[$d0],8)",
|
||||
"&rll (@x[$d1],@x[$d1],8)",
|
||||
|
||||
"&alr ($xc,@x[$d0])",
|
||||
"&alr ($xc_,@x[$d1])",
|
||||
"&xr (@x[$b0],$xc)",
|
||||
"&xr (@x[$b1],$xc_)",
|
||||
"&rll (@x[$b0],@x[$b0],7)",
|
||||
"&rll (@x[$b1],@x[$b1],7)",
|
||||
|
||||
"&stm ($xc,$xc_,'$stdframe+4*8+4*$c0($sp)')", # reload pair of 'c's
|
||||
"&lm ($xc,$xc_,'$stdframe+4*8+4*$c2($sp)')",
|
||||
|
||||
"&alr (@x[$a2],@x[$b2])", # Q3
|
||||
"&alr (@x[$a3],@x[$b3])", # Q4
|
||||
"&xr (@x[$d2],@x[$a2])",
|
||||
"&xr (@x[$d3],@x[$a3])",
|
||||
"&rll (@x[$d2],@x[$d2],16)",
|
||||
"&rll (@x[$d3],@x[$d3],16)",
|
||||
|
||||
"&alr ($xc,@x[$d2])",
|
||||
"&alr ($xc_,@x[$d3])",
|
||||
"&xr (@x[$b2],$xc)",
|
||||
"&xr (@x[$b3],$xc_)",
|
||||
"&rll (@x[$b2],@x[$b2],12)",
|
||||
"&rll (@x[$b3],@x[$b3],12)",
|
||||
|
||||
"&alr (@x[$a2],@x[$b2])",
|
||||
"&alr (@x[$a3],@x[$b3])",
|
||||
"&xr (@x[$d2],@x[$a2])",
|
||||
"&xr (@x[$d3],@x[$a3])",
|
||||
"&rll (@x[$d2],@x[$d2],8)",
|
||||
"&rll (@x[$d3],@x[$d3],8)",
|
||||
|
||||
"&alr ($xc,@x[$d2])",
|
||||
"&alr ($xc_,@x[$d3])",
|
||||
"&xr (@x[$b2],$xc)",
|
||||
"&xr (@x[$b3],$xc_)",
|
||||
"&rll (@x[$b2],@x[$b2],7)",
|
||||
"&rll (@x[$b3],@x[$b3],7)"
|
||||
);
|
||||
}
|
||||
|
||||
$code.=<<___;
|
||||
.text
|
||||
|
||||
.globl ChaCha20_ctr32
|
||||
.type ChaCha20_ctr32,\@function
|
||||
.align 32
|
||||
ChaCha20_ctr32:
|
||||
lt${g}r $len,$len # $len==0?
|
||||
bzr %r14
|
||||
a${g}hi $len,-64
|
||||
l${g}hi %r1,-$frame
|
||||
stm${g} %r6,%r15,`6*$SIZE_T`($sp)
|
||||
sl${g}r $out,$inp # difference
|
||||
la $len,0($inp,$len) # end of input minus 64
|
||||
larl %r7,.Lsigma
|
||||
lgr %r0,$sp
|
||||
la $sp,0(%r1,$sp)
|
||||
st${g} %r0,0($sp)
|
||||
|
||||
lmg %r8,%r11,0($key) # load key
|
||||
lmg %r12,%r13,0($counter) # load counter
|
||||
lmg %r6,%r7,0(%r7) # load sigma constant
|
||||
|
||||
la %r14,0($inp)
|
||||
st${g} $out,$frame+3*$SIZE_T($sp)
|
||||
st${g} $len,$frame+4*$SIZE_T($sp)
|
||||
stmg %r6,%r13,$stdframe($sp) # copy key schedule to stack
|
||||
srlg @x[12],%r12,32 # 32-bit counter value
|
||||
j .Loop_outer
|
||||
|
||||
.align 16
|
||||
.Loop_outer:
|
||||
lm @x[0],@x[7],$stdframe+4*0($sp) # load x[0]-x[7]
|
||||
lm @t[0],@t[1],$stdframe+4*10($sp) # load x[10]-x[11]
|
||||
lm @x[13],@x[15],$stdframe+4*13($sp) # load x[13]-x[15]
|
||||
stm @t[0],@t[1],$stdframe+4*8+4*10($sp) # offload x[10]-x[11]
|
||||
lm @t[0],@t[1],$stdframe+4*8($sp) # load x[8]-x[9]
|
||||
st @x[12],$stdframe+4*12($sp) # save counter
|
||||
st${g} %r14,$frame+2*$SIZE_T($sp) # save input pointer
|
||||
lhi %r14,10
|
||||
j .Loop
|
||||
|
||||
.align 4
|
||||
.Loop:
|
||||
___
|
||||
foreach (&ROUND(0, 4, 8,12)) { eval; }
|
||||
foreach (&ROUND(0, 5,10,15)) { eval; }
|
||||
$code.=<<___;
|
||||
brct %r14,.Loop
|
||||
|
||||
l${g} %r14,$frame+2*$SIZE_T($sp) # pull input pointer
|
||||
stm @t[0],@t[1],$stdframe+4*8+4*8($sp) # offload x[8]-x[9]
|
||||
lm${g} @t[0],@t[1],$frame+3*$SIZE_T($sp)
|
||||
|
||||
al @x[0],$stdframe+4*0($sp) # accumulate key schedule
|
||||
al @x[1],$stdframe+4*1($sp)
|
||||
al @x[2],$stdframe+4*2($sp)
|
||||
al @x[3],$stdframe+4*3($sp)
|
||||
al @x[4],$stdframe+4*4($sp)
|
||||
al @x[5],$stdframe+4*5($sp)
|
||||
al @x[6],$stdframe+4*6($sp)
|
||||
al @x[7],$stdframe+4*7($sp)
|
||||
lrvr @x[0],@x[0]
|
||||
lrvr @x[1],@x[1]
|
||||
lrvr @x[2],@x[2]
|
||||
lrvr @x[3],@x[3]
|
||||
lrvr @x[4],@x[4]
|
||||
lrvr @x[5],@x[5]
|
||||
lrvr @x[6],@x[6]
|
||||
lrvr @x[7],@x[7]
|
||||
al @x[12],$stdframe+4*12($sp)
|
||||
al @x[13],$stdframe+4*13($sp)
|
||||
al @x[14],$stdframe+4*14($sp)
|
||||
al @x[15],$stdframe+4*15($sp)
|
||||
lrvr @x[12],@x[12]
|
||||
lrvr @x[13],@x[13]
|
||||
lrvr @x[14],@x[14]
|
||||
lrvr @x[15],@x[15]
|
||||
|
||||
la @t[0],0(@t[0],%r14) # reconstruct output pointer
|
||||
cl${g}r %r14,@t[1]
|
||||
jh .Ltail
|
||||
|
||||
x @x[0],4*0(%r14) # xor with input
|
||||
x @x[1],4*1(%r14)
|
||||
st @x[0],4*0(@t[0]) # store output
|
||||
x @x[2],4*2(%r14)
|
||||
st @x[1],4*1(@t[0])
|
||||
x @x[3],4*3(%r14)
|
||||
st @x[2],4*2(@t[0])
|
||||
x @x[4],4*4(%r14)
|
||||
st @x[3],4*3(@t[0])
|
||||
lm @x[0],@x[3],$stdframe+4*8+4*8($sp) # load x[8]-x[11]
|
||||
x @x[5],4*5(%r14)
|
||||
st @x[4],4*4(@t[0])
|
||||
x @x[6],4*6(%r14)
|
||||
al @x[0],$stdframe+4*8($sp)
|
||||
st @x[5],4*5(@t[0])
|
||||
x @x[7],4*7(%r14)
|
||||
al @x[1],$stdframe+4*9($sp)
|
||||
st @x[6],4*6(@t[0])
|
||||
x @x[12],4*12(%r14)
|
||||
al @x[2],$stdframe+4*10($sp)
|
||||
st @x[7],4*7(@t[0])
|
||||
x @x[13],4*13(%r14)
|
||||
al @x[3],$stdframe+4*11($sp)
|
||||
st @x[12],4*12(@t[0])
|
||||
x @x[14],4*14(%r14)
|
||||
st @x[13],4*13(@t[0])
|
||||
x @x[15],4*15(%r14)
|
||||
st @x[14],4*14(@t[0])
|
||||
lrvr @x[0],@x[0]
|
||||
st @x[15],4*15(@t[0])
|
||||
lrvr @x[1],@x[1]
|
||||
lrvr @x[2],@x[2]
|
||||
lrvr @x[3],@x[3]
|
||||
lhi @x[12],1
|
||||
x @x[0],4*8(%r14)
|
||||
al @x[12],$stdframe+4*12($sp) # increment counter
|
||||
x @x[1],4*9(%r14)
|
||||
st @x[0],4*8(@t[0])
|
||||
x @x[2],4*10(%r14)
|
||||
st @x[1],4*9(@t[0])
|
||||
x @x[3],4*11(%r14)
|
||||
st @x[2],4*10(@t[0])
|
||||
st @x[3],4*11(@t[0])
|
||||
|
||||
cl${g}r %r14,@t[1] # done yet?
|
||||
la %r14,64(%r14)
|
||||
jl .Loop_outer
|
||||
|
||||
.Ldone:
|
||||
xgr %r0,%r0
|
||||
xgr %r1,%r1
|
||||
xgr %r2,%r2
|
||||
xgr %r3,%r3
|
||||
stmg %r0,%r3,$stdframe+4*4($sp) # wipe key copy
|
||||
stmg %r0,%r3,$stdframe+4*12($sp)
|
||||
|
||||
lm${g} %r6,%r15,`$frame+6*$SIZE_T`($sp)
|
||||
br %r14
|
||||
|
||||
.align 16
|
||||
.Ltail:
|
||||
la @t[1],64($t[1])
|
||||
stm @x[0],@x[7],$stdframe+4*0($sp)
|
||||
sl${g}r @t[1],%r14
|
||||
lm @x[0],@x[3],$stdframe+4*8+4*8($sp)
|
||||
l${g}hi @x[6],0
|
||||
stm @x[12],@x[15],$stdframe+4*12($sp)
|
||||
al @x[0],$stdframe+4*8($sp)
|
||||
al @x[1],$stdframe+4*9($sp)
|
||||
al @x[2],$stdframe+4*10($sp)
|
||||
al @x[3],$stdframe+4*11($sp)
|
||||
lrvr @x[0],@x[0]
|
||||
lrvr @x[1],@x[1]
|
||||
lrvr @x[2],@x[2]
|
||||
lrvr @x[3],@x[3]
|
||||
stm @x[0],@x[3],$stdframe+4*8($sp)
|
||||
|
||||
.Loop_tail:
|
||||
llgc @x[4],0(@x[6],%r14)
|
||||
llgc @x[5],$stdframe(@x[6],$sp)
|
||||
xr @x[5],@x[4]
|
||||
stc @x[5],0(@x[6],@t[0])
|
||||
la @x[6],1(@x[6])
|
||||
brct @t[1],.Loop_tail
|
||||
|
||||
j .Ldone
|
||||
.size ChaCha20_ctr32,.-ChaCha20_ctr32
|
||||
|
||||
.align 32
|
||||
.Lsigma:
|
||||
.long 0x61707865,0x3320646e,0x79622d32,0x6b206574 # endian-neutral
|
||||
.asciz "ChaCha20 for s390x, CRYPTOGAMS by <appro\@openssl.org>"
|
||||
.align 4
|
||||
___
|
||||
|
||||
foreach (split("\n",$code)) {
|
||||
s/\`([^\`]*)\`/eval $1/ge;
|
||||
|
||||
print $_,"\n";
|
||||
}
|
||||
close STDOUT;
|
||||
1155
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-x86.pl
vendored
Executable file
1155
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-x86.pl
vendored
Executable file
File diff suppressed because it is too large
Load diff
4005
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-x86_64.pl
vendored
Executable file
4005
trunk/3rdparty/openssl-1.1-fit/crypto/chacha/asm/chacha-x86_64.pl
vendored
Executable file
File diff suppressed because it is too large
Load diff
Loading…
Add table
Add a link
Reference in a new issue