From e64256d499f2fa45a724269cb1a8043a7434ba8d Mon Sep 17 00:00:00 2001 From: Jörg Hohensohn Date: Thu, 18 Mar 2004 22:44:05 +0000 Subject: patch #919088: 17% faster bitswap(), by Jens "SH" Arnold git-svn-id: svn://svn.rockbox.org/rockbox/trunk@4407 a1c6a512-1295-4272-9138-f99709370657 --- firmware/bitswap.S | 160 +++++++++++++++++++++++++++++------------------------ 1 file changed, 89 insertions(+), 71 deletions(-) diff --git a/firmware/bitswap.S b/firmware/bitswap.S index 990ecb4d00..25d7a99ad2 100644 --- a/firmware/bitswap.S +++ b/firmware/bitswap.S @@ -7,7 +7,7 @@ * \/ \/ \/ \/ \/ * $Id$ * - * Copyright (C) 2002 by Magnus Holmgren + * Copyright (C) 2004 by Jens Arnold * * All files in this archive are subject to the GNU General Public License. * See the file COPYING in the source tree root for full license agreement. @@ -22,67 +22,87 @@ .global _bitswap .type _bitswap,@function -/* Registers used: +/* Flips the bits of all bytes in a memory area (required for mp3 data on + * the Archos). This version is optimized for speed and size. * - * r0 Temporary (required by some instructions) - * r1 Low byte - * r2 High byte / final result - * r4 &Data - * r5 Length - * r7 Flip table - */ - -/* The instruction order below is a bit strange, because: - * 1) Keeping load/stores on longword boundaries means the instruction fetch - * won't compete with the memory access (because instructions are fetched - * in pairs). - * 2) Using the result of a fetch in the next instruction causes a stall - * (except in certain circumstances). - * See the SH-1 programming manual for details. + * arguments: + * r4 - start address + * r5 - length + * + * return value: void + * + * register usage: + * r0 - temporary + * r1 - bit mask for rounding to long bound / low byte (after swap) + * r2 - high byte (after swap) / combined result + * r4 - data address - 4 + * r5 - end address - 4 + * r7 - flip table (addressing with signed offset) + * + * The instruction order below is devised in a way to utilize the pipelining + * of the SH1 to the max. */ _bitswap: - mov.l .fliptable,r7 - add #-2,r4 /* ptr is used shifted by 2 */ - add r4,r5 /* r5 = end_address - 2 */ - add #-1,r5 /* r5 = &last_byte - 2 */ + mova _fliptable,r0 + mov r0,r7 + add #-4,r4 /* address is shifted by 4 */ + add r4,r5 /* r5 = end_address - 4 */ + cmp/hi r4,r5 /* at least something to do? */ + bf .exit /* no, get out of here! */ + + add #-3,r5 /* end offset for flipping 4 bytes per pass */ mov r4,r0 tst #1,r0 /* even address? */ - bt .init /* yes */ + bt .start2_w /* yes, jump into main loop */ - add #1,r4 /* r4 now even */ - mov.b @(1,r4),r0 /* no, swap first byte */ - extu.b r0,r0 - mov.b @(r0,r7),r0 - mov.b r0,@(1,r4) + /* no, flip first byte */ + mov.b @(4,r4),r0 /* load byte, sign extension! */ + add #1,r4 /* early increment */ + mov.b @(r0,r7),r0 /* fliptable offset is signed */ + bra .start2_w /* jump into main loop */ + mov.b r0,@(3,r4) /* store byte */ -.init: - cmp/hi r4,r5 /* at least 2 bytes to swap? */ - bf .last /* no, skip main loop */ - -.loop: - mov.w @(2,r4),r0 /* data to flip */ - add #2,r4 /* early increment */ - swap.b r0,r2 /* get high byte */ - extu.b r0,r0 /* prepare low byte */ - mov.b @(r0,r7),r1 /* swap low byte */ - extu.b r2,r0 /* prepare high byte */ - mov.b @(r0,r7),r2 /* swap high byte */ - extu.b r1,r1 /* zero extend low byte */ - shll8 r2 /* shift high byte, low byte zeroed */ - or r1,r2 /* put low byte in result */ - mov.w r2,@r4 /* store result, ptr already incr'd */ - cmp/hi r4,r5 /* while &last_byte > data */ - bt .loop + /* main loop: flips 2 words per pass */ + .align 2 +.loop2_w: + mov.w @(6,r4),r0 /* load second word */ + add #4,r4 /* early increment */ + swap.b r0,r2 /* get high byte (2nd word) */ + exts.b r0,r0 /* prepare low byte (2nd word) */ + mov.b @(r0,r7),r1 /* swap low byte (2nd word) */ + exts.b r2,r0 /* prepare high byte (2nd word) */ + mov.b @(r0,r7),r2 /* swap high byte (2nd word) */ + extu.b r1,r0 /* zero extend low byte (2nd word) */ + mov.w @r4,r1 /* load first word */ + shll8 r2 /* shift high byte (2nd word), low byte zeroed */ + or r2,r0 /* put low byte (2nd word) in result */ + swap.b r1,r2 /* get high byte (1st word) */ + mov.w r0,@(2,r4) /* store result (2nd word) */ + exts.b r1,r0 /* prepare low byte (1st word) */ + mov.b @(r0,r7),r1 /* swap low byte (1st word) */ + exts.b r2,r0 /* prepare high byte (1st word) */ + mov.b @(r0,r7),r2 /* swap high byte (1st word) */ + extu.b r1,r0 /* zero extend low byte (1st word) */ + shll8 r2 /* shift high byte (1st word), low byte zeroed */ + or r2,r0 /* put low byte (1st word) in result */ + mov.w r0,@r4 /* store result (1st word) */ +.start2_w: + cmp/hi r4,r5 /* runs r4 up to last long bound */ + bt .loop2_w -.last: - cmp/eq r4,r5 /* if behind (&last_byte - 2), exit */ - bf .exit + bra .start_b2 /* jump into trailing byte loop */ + add #3,r5 /* reset end offset */ - mov.b @(2,r4),r0 /* swap last byte */ - extu.b r0,r0 - mov.b @(r0,r7),r0 - mov.b r0,@(2,r4) + /* trailing byte loop: flips 0..3 bytes */ +.loop_b2: + mov.b @(4,r4),r0 /* loand byte, sign extension! */ + add #1,r4 /* early increment */ + mov.b @(r0,r7),r0 /* fliptable offset is signed */ + mov.b r0,@(3,r4) /* store byte */ +.start_b2: + cmp/hi r4,r5 /* runs r4 up to end address */ + bt .loop_b2 .exit: rts @@ -90,9 +110,22 @@ _bitswap: .align 2 -.fliptable: - .long _fliptable - + .byte 0x01, 0x81, 0x41, 0xc1, 0x21, 0xa1, 0x61, 0xe1 + .byte 0x11, 0x91, 0x51, 0xd1, 0x31, 0xb1, 0x71, 0xf1 + .byte 0x09, 0x89, 0x49, 0xc9, 0x29, 0xa9, 0x69, 0xe9 + .byte 0x19, 0x99, 0x59, 0xd9, 0x39, 0xb9, 0x79, 0xf9 + .byte 0x05, 0x85, 0x45, 0xc5, 0x25, 0xa5, 0x65, 0xe5 + .byte 0x15, 0x95, 0x55, 0xd5, 0x35, 0xb5, 0x75, 0xf5 + .byte 0x0d, 0x8d, 0x4d, 0xcd, 0x2d, 0xad, 0x6d, 0xed + .byte 0x1d, 0x9d, 0x5d, 0xdd, 0x3d, 0xbd, 0x7d, 0xfd + .byte 0x03, 0x83, 0x43, 0xc3, 0x23, 0xa3, 0x63, 0xe3 + .byte 0x13, 0x93, 0x53, 0xd3, 0x33, 0xb3, 0x73, 0xf3 + .byte 0x0b, 0x8b, 0x4b, 0xcb, 0x2b, 0xab, 0x6b, 0xeb + .byte 0x1b, 0x9b, 0x5b, 0xdb, 0x3b, 0xbb, 0x7b, 0xfb + .byte 0x07, 0x87, 0x47, 0xc7, 0x27, 0xa7, 0x67, 0xe7 + .byte 0x17, 0x97, 0x57, 0xd7, 0x37, 0xb7, 0x77, 0xf7 + .byte 0x0f, 0x8f, 0x4f, 0xcf, 0x2f, 0xaf, 0x6f, 0xef + .byte 0x1f, 0x9f, 0x5f, 0xdf, 0x3f, 0xbf, 0x7f, 0xff _fliptable: .byte 0x00, 0x80, 0x40, 0xc0, 0x20, 0xa0, 0x60, 0xe0 .byte 0x10, 0x90, 0x50, 0xd0, 0x30, 0xb0, 0x70, 0xf0 @@ -110,22 +143,7 @@ _fliptable: .byte 0x16, 0x96, 0x56, 0xd6, 0x36, 0xb6, 0x76, 0xf6 .byte 0x0e, 0x8e, 0x4e, 0xce, 0x2e, 0xae, 0x6e, 0xee .byte 0x1e, 0x9e, 0x5e, 0xde, 0x3e, 0xbe, 0x7e, 0xfe - .byte 0x01, 0x81, 0x41, 0xc1, 0x21, 0xa1, 0x61, 0xe1 - .byte 0x11, 0x91, 0x51, 0xd1, 0x31, 0xb1, 0x71, 0xf1 - .byte 0x09, 0x89, 0x49, 0xc9, 0x29, 0xa9, 0x69, 0xe9 - .byte 0x19, 0x99, 0x59, 0xd9, 0x39, 0xb9, 0x79, 0xf9 - .byte 0x05, 0x85, 0x45, 0xc5, 0x25, 0xa5, 0x65, 0xe5 - .byte 0x15, 0x95, 0x55, 0xd5, 0x35, 0xb5, 0x75, 0xf5 - .byte 0x0d, 0x8d, 0x4d, 0xcd, 0x2d, 0xad, 0x6d, 0xed - .byte 0x1d, 0x9d, 0x5d, 0xdd, 0x3d, 0xbd, 0x7d, 0xfd - .byte 0x03, 0x83, 0x43, 0xc3, 0x23, 0xa3, 0x63, 0xe3 - .byte 0x13, 0x93, 0x53, 0xd3, 0x33, 0xb3, 0x73, 0xf3 - .byte 0x0b, 0x8b, 0x4b, 0xcb, 0x2b, 0xab, 0x6b, 0xeb - .byte 0x1b, 0x9b, 0x5b, 0xdb, 0x3b, 0xbb, 0x7b, 0xfb - .byte 0x07, 0x87, 0x47, 0xc7, 0x27, 0xa7, 0x67, 0xe7 - .byte 0x17, 0x97, 0x57, 0xd7, 0x37, 0xb7, 0x77, 0xf7 - .byte 0x0f, 0x8f, 0x4f, 0xcf, 0x2f, 0xaf, 0x6f, 0xef - .byte 0x1f, 0x9f, 0x5f, 0xdf, 0x3f, 0xbf, 0x7f, 0xff + .end: .size _bitswap,.end-_bitswap -- cgit v1.2.3