diff options
Diffstat (limited to 'arch/sparc64/lib/memset.S')
| -rw-r--r-- | arch/sparc64/lib/memset.S | 196 |
1 files changed, 196 insertions, 0 deletions
diff --git a/arch/sparc64/lib/memset.S b/arch/sparc64/lib/memset.S new file mode 100644 index 000000000..55de4ea9d --- /dev/null +++ b/arch/sparc64/lib/memset.S @@ -0,0 +1,196 @@ +/* linux/arch/sparc64/lib/memset.S: Sparc optimized memset, bzero and clear_user code + * Copyright (C) 1991,1996 Free Software Foundation + * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu) + * Copyright (C) 1996,1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz) + * + * Returns 0, if ok, and number of bytes not yet set if exception + * occurs and we were called as clear_user. + */ + +#include <asm/asi.h> +#include <asm/ptrace.h> + +#define EX(x,y,a,b,z) \ +98: x,y; \ + .section .fixup,z##alloc,z##execinstr; \ + .align 4; \ +99: ba,pt %xcc, 30f; \ + a, b, %o0; \ + .section __ex_table,z##alloc; \ + .align 4; \ + .word 98b, 99b; \ + .text; \ + .align 4 + +#define EXT(start,end,handler,z) \ + .section __ex_table,z##alloc; \ + .align 4; \ + .word start, 0, end, handler; \ + .text; \ + .align 4 + +/* Please don't change these macros, unless you change the logic + * in the .fixup section below as well. + * Store 64 bytes at (BASE + OFFSET) using value SOURCE. */ +#define ZERO_BIG_BLOCK(base, offset, source) \ + stxa source, [base + offset + 0x00] %asi; \ + stxa source, [base + offset + 0x08] %asi; \ + stxa source, [base + offset + 0x10] %asi; \ + stxa source, [base + offset + 0x18] %asi; \ + stxa source, [base + offset + 0x20] %asi; \ + stxa source, [base + offset + 0x28] %asi; \ + stxa source, [base + offset + 0x30] %asi; \ + stxa source, [base + offset + 0x38] %asi; + +#define ZERO_LAST_BLOCKS(base, offset, source) \ + stxa source, [base - offset - 0x38] %asi; \ + stxa source, [base - offset - 0x30] %asi; \ + stxa source, [base - offset - 0x28] %asi; \ + stxa source, [base - offset - 0x20] %asi; \ + stxa source, [base - offset - 0x18] %asi; \ + stxa source, [base - offset - 0x10] %asi; \ + stxa source, [base - offset - 0x08] %asi; \ + stxa source, [base - offset - 0x00] %asi; + + .text + .align 4 + + .globl __bzero, __memset, __bzero_noasi + .globl memset, __memset_start, __memset_end +__memset_start: +__memset: +memset: + and %o1, 0xff, %g3 + sll %g3, 8, %g2 + or %g3, %g2, %g3 + sll %g3, 16, %g2 + or %g3, %g2, %g3 + mov %o2, %o1 + wr %g0, ASI_P, %asi + sllx %g3, 32, %g2 + ba,pt %xcc, 1f + or %g3, %g2, %g3 +__bzero: + wr %g0, ASI_P, %asi +__bzero_noasi: + mov %g0, %g3 +1: + cmp %o1, 7 + bleu,pn %xcc, 7f + andcc %o0, 3, %o2 + + be,a,pt %icc, 4f + andcc %o0, 4, %g0 + + cmp %o2, 3 + be,pn %icc, 2f + EX(stba %g3, [%o0] %asi, sub %o1, 0,#) + + cmp %o2, 2 + be,pt %icc, 2f + EX(stba %g3, [%o0 + 0x01] %asi, sub %o1, 1,#) + + EX(stba %g3, [%o0 + 0x02] %asi, sub %o1, 2,#) +2: + sub %o2, 4, %o2 + sub %o0, %o2, %o0 + add %o1, %o2, %o1 + andcc %o0, 4, %g0 +4: + be,a,pt %icc, 2f + andncc %o1, 0x7f, %o3 + + EX(sta %g3, [%o0] %asi, sub %o1, 0,#) + sub %o1, 4, %o1 + add %o0, 4, %o0 + andncc %o1, 0x7f, %o3 ! Now everything is 8 aligned and o1 is len to run +2: + be,pn %xcc, 9f + andcc %o1, 0x78, %o2 +10: + ZERO_BIG_BLOCK(%o0, 0x00, %g3) + subcc %o3, 128, %o3 + ZERO_BIG_BLOCK(%o0, 0x40, %g3) +11: + EXT(10b, 11b, 20f,#) + bne,pt %xcc, 10b + add %o0, 128, %o0 + + tst %o2 +9: + be,pn %xcc, 13f + andcc %o1, 7, %o1 +14: + rd %pc, %o4 + srl %o2, 1, %o3 + sub %o4, %o3, %o4 + jmpl %o4 + (13f - 14b), %g0 + add %o0, %o2, %o0 +12: + ZERO_LAST_BLOCKS(%o0, 0x48, %g3) + ZERO_LAST_BLOCKS(%o0, 0x08, %g3) +13: + be,pn %icc, 8f + andcc %o1, 4, %g0 + + be,pn %icc, 1f + andcc %o1, 2, %g0 + + EX(sta %g3, [%o0] %asi, and %o1, 7,#) + add %o0, 4, %o0 +1: + be,pn %icc, 1f + andcc %o1, 1, %g0 + + EX(stha %g3, [%o0] %asi, and %o1, 3,#) + add %o0, 2, %o0 +1: + bne,a,pn %icc, 8f + EX(stba %g3, [%o0] %asi, and %o1, 1,#) +8: + retl + clr %o0 +7: + be,pn %icc, 13b + orcc %o1, 0, %g0 + + be,pn %icc, 0f +8: + add %o0, 1, %o0 + subcc %o1, 1, %o1 + bne,a,pt %icc, 8b + EX(stba %g3, [%o0 - 1] %asi, add %o1, 1,#) +0: + retl + clr %o0 +__memset_end: + + .section .fixup,#alloc,#execinstr + .align 4 +20: + cmp %g2, 8 + bleu,pn %xcc, 1f + and %o1, 0x7f, %o1 + sub %g2, 9, %g2 + add %o3, 64, %o3 +1: + sll %g2, 3, %g2 + add %o3, %o1, %o0 + ba,pt %xcc, 30f + sub %o0, %g2, %o0 +21: + mov 8, %o0 + and %o1, 7, %o1 + sub %o0, %g2, %o0 + sll %o0, 3, %o0 + ba,pt %xcc, 30f + add %o0, %o1, %o0 +30: +/* %o4 is faulting address, %o5 is %pc where fault occured */ + save %sp, -160, %sp + mov %i5, %o0 + mov %i7, %o1 + call lookup_fault + mov %i4, %o2 + ret + restore |