From 51533b615e605d86154ec1b4e585c8ca1b0b15b7 Mon Sep 17 00:00:00 2001 From: Mikael Starvik Date: Wed, 27 Jul 2005 11:44:44 -0700 Subject: [PATCH] CRIS update: new subarchitecture v32 New CRIS sub architecture named v32. From: Dave Jones Fix swapped kmalloc args Signed-off-by: Mikael Starvik Signed-off-by: Dave Jones Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- arch/cris/arch-v32/lib/checksum.S | 111 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 111 insertions(+) create mode 100644 arch/cris/arch-v32/lib/checksum.S (limited to 'arch/cris/arch-v32/lib/checksum.S') diff --git a/arch/cris/arch-v32/lib/checksum.S b/arch/cris/arch-v32/lib/checksum.S new file mode 100644 index 00000000000..32e66181b82 --- /dev/null +++ b/arch/cris/arch-v32/lib/checksum.S @@ -0,0 +1,111 @@ +/* + * A fast checksum routine using movem + * Copyright (c) 1998-2001, 2003 Axis Communications AB + * + * csum_partial(const unsigned char * buff, int len, unsigned int sum) + */ + + .globl csum_partial +csum_partial: + + ;; r10 - src + ;; r11 - length + ;; r12 - checksum + + ;; check for breakeven length between movem and normal word looping versions + ;; we also do _NOT_ want to compute a checksum over more than the + ;; actual length when length < 40 + + cmpu.w 80,$r11 + blo _word_loop + nop + + ;; need to save the registers we use below in the movem loop + ;; this overhead is why we have a check above for breakeven length + ;; only r0 - r8 have to be saved, the other ones are clobber-able + ;; according to the ABI + + subq 9*4,$sp + subq 10*4,$r11 ; update length for the first loop + movem $r8,[$sp] + + ;; do a movem checksum + +_mloop: movem [$r10+],$r9 ; read 10 longwords + + ;; perform dword checksumming on the 10 longwords + + add.d $r0,$r12 + addc $r1,$r12 + addc $r2,$r12 + addc $r3,$r12 + addc $r4,$r12 + addc $r5,$r12 + addc $r6,$r12 + addc $r7,$r12 + addc $r8,$r12 + addc $r9,$r12 + + ;; fold the carry into the checksum, to avoid having to loop the carry + ;; back into the top + + addc 0,$r12 + addc 0,$r12 ; do it again, since we might have generated a carry + + subq 10*4,$r11 + bge _mloop + nop + + addq 10*4,$r11 ; compensate for last loop underflowing length + + movem [$sp+],$r8 ; restore regs + +_word_loop: + ;; only fold if there is anything to fold. + + cmpq 0,$r12 + beq _no_fold + + ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below. + ;; r9 and r13 can be used as temporaries. + + moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9 + lsrq 16,$r9 + + move.d $r12,$r13 + lsrq 16,$r13 ; r13 = checksum >> 16 + and.d $r9,$r12 ; checksum = checksum & 0xffff + add.d $r13,$r12 ; checksum += r13 + move.d $r12,$r13 ; do the same again, maybe we got a carry last add + lsrq 16,$r13 + and.d $r9,$r12 + add.d $r13,$r12 + +_no_fold: + cmpq 2,$r11 + blt _no_words + nop + + ;; checksum the rest of the words + + subq 2,$r11 + +_wloop: subq 2,$r11 + bge _wloop + addu.w [$r10+],$r12 + + addq 2,$r11 + +_no_words: + ;; see if we have one odd byte more + cmpq 1,$r11 + beq _do_byte + nop + ret + move.d $r12,$r10 + +_do_byte: + ;; copy and checksum the last byte + addu.b [$r10],$r12 + ret + move.d $r12,$r10 -- cgit v1.2.3