From 06024f217d607369f0ee0071034ebb03071d5fb2 Mon Sep 17 00:00:00 2001 From: Alexandre Oliva Date: Mon, 31 Oct 2005 18:29:36 -0200 Subject: [PATCH] x86-64: bitops fix for -Os This fixes the x86-64 find_[first|next]_zero_bit() function for the end-of-range case. It didn't test for a zero size, and the "rep scas" would do entirely the wrong thing. Signed-off-by: Alexandre Oliva Signed-off-by: Linus Torvalds --- arch/x86_64/lib/bitops.c | 66 ++++++++++++++++++++++++++++++++++++------------ 1 file changed, 50 insertions(+), 16 deletions(-) diff --git a/arch/x86_64/lib/bitops.c b/arch/x86_64/lib/bitops.c index a29fb75b33a..95b6d9639fb 100644 --- a/arch/x86_64/lib/bitops.c +++ b/arch/x86_64/lib/bitops.c @@ -5,19 +5,23 @@ #undef find_first_bit #undef find_next_bit -/** - * find_first_zero_bit - find the first zero bit in a memory region - * @addr: The address to start the search at - * @size: The maximum size to search - * - * Returns the bit-number of the first zero bit, not the number of the byte - * containing a bit. - */ -inline long find_first_zero_bit(const unsigned long * addr, unsigned long size) +static inline long +__find_first_zero_bit(const unsigned long * addr, unsigned long size) { long d0, d1, d2; long res; + /* + * We must test the size in words, not in bits, because + * otherwise incoming sizes in the range -63..-1 will not run + * any scasq instructions, and then the flags used by the je + * instruction will have whatever random value was in place + * before. Nobody should call us like that, but + * find_next_zero_bit() does when offset and size are at the + * same word and it fails to find a zero itself. + */ + size += 63; + size >>= 6; if (!size) return 0; asm volatile( @@ -30,11 +34,29 @@ inline long find_first_zero_bit(const unsigned long * addr, unsigned long size) " shlq $3,%%rdi\n" " addq %%rdi,%%rdx" :"=d" (res), "=&c" (d0), "=&D" (d1), "=&a" (d2) - :"0" (0ULL), "1" ((size + 63) >> 6), "2" (addr), "3" (-1ULL), - [addr] "r" (addr) : "memory"); + :"0" (0ULL), "1" (size), "2" (addr), "3" (-1ULL), + [addr] "S" (addr) : "memory"); + /* + * Any register would do for [addr] above, but GCC tends to + * prefer rbx over rsi, even though rsi is readily available + * and doesn't have to be saved. + */ return res; } +/** + * find_first_zero_bit - find the first zero bit in a memory region + * @addr: The address to start the search at + * @size: The maximum size to search + * + * Returns the bit-number of the first zero bit, not the number of the byte + * containing a bit. + */ +long find_first_zero_bit(const unsigned long * addr, unsigned long size) +{ + return __find_first_zero_bit (addr, size); +} + /** * find_next_zero_bit - find the first zero bit in a memory region * @addr: The address to base the search on @@ -43,7 +65,7 @@ inline long find_first_zero_bit(const unsigned long * addr, unsigned long size) */ long find_next_zero_bit (const unsigned long * addr, long size, long offset) { - unsigned long * p = ((unsigned long *) addr) + (offset >> 6); + const unsigned long * p = addr + (offset >> 6); unsigned long set = 0; unsigned long res, bit = offset&63; @@ -63,8 +85,8 @@ long find_next_zero_bit (const unsigned long * addr, long size, long offset) /* * No zero yet, search remaining full words for a zero */ - res = find_first_zero_bit ((const unsigned long *)p, - size - 64 * (p - (unsigned long *) addr)); + res = __find_first_zero_bit (p, size - 64 * (p - addr)); + return (offset + set + res); } @@ -74,6 +96,19 @@ __find_first_bit(const unsigned long * addr, unsigned long size) long d0, d1; long res; + /* + * We must test the size in words, not in bits, because + * otherwise incoming sizes in the range -63..-1 will not run + * any scasq instructions, and then the flags used by the jz + * instruction will have whatever random value was in place + * before. Nobody should call us like that, but + * find_next_bit() does when offset and size are at the same + * word and it fails to find a one itself. + */ + size += 63; + size >>= 6; + if (!size) + return 0; asm volatile( " repe; scasq\n" " jz 1f\n" @@ -83,8 +118,7 @@ __find_first_bit(const unsigned long * addr, unsigned long size) " shlq $3,%%rdi\n" " addq %%rdi,%%rax" :"=a" (res), "=&c" (d0), "=&D" (d1) - :"0" (0ULL), - "1" ((size + 63) >> 6), "2" (addr), + :"0" (0ULL), "1" (size), "2" (addr), [addr] "r" (addr) : "memory"); return res; } -- cgit v1.2.3