kernel-fxtec-pro1x/arch/x86/include/asm/bitops.h

522 lines
13 KiB
C
Raw Normal View History

#ifndef _ASM_X86_BITOPS_H
#define _ASM_X86_BITOPS_H
/*
* Copyright 1992, Linus Torvalds.
*
* Note: inlines with more than a single statement should be marked
* __always_inline to avoid problems with older gcc's inlining heuristics.
*/
#ifndef _LINUX_BITOPS_H
#error only <linux/bitops.h> can be included directly
#endif
#include <linux/compiler.h>
#include <asm/alternative.h>
#define BIT_64(n) (U64_C(1) << (n))
/*
* These have to be done with inline assembly: that way the bit-setting
* is guaranteed to be atomic. All bit operations return 0 if the bit
* was cleared before the operation and != 0 if it was not.
*
* bit 0 is the LSB of addr; bit 32 is the LSB of (addr+1).
*/
#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 1)
/* Technically wrong, but this avoids compilation errors on some gcc
versions. */
x86, bitops: make constant-bit set/clear_bit ops faster On Wed, 18 Jun 2008, Linus Torvalds wrote: > > And yes, the "lock andl" should be noticeably faster than the xchgl. I dunno. Here's a untested (!!) patch that turns constant-bit set/clear_bit ops into byte mask ops (lock orb/andb). It's not exactly pretty. The reason for using the byte versions is that a locked op is serialized in the memory pipeline anyway, so there are no forwarding issues (that could slow down things when we access things with different sizes), and the byte ops are a lot smaller than 32-bit and particularly 64-bit ops (big constants, and the 64-bit ops need the REX prefix byte too). [ Side note: I wonder if we should turn the "test_bit()" C version into a "char *" version too.. It could actually help with alias analysis, since char pointers can alias anything. So it might be the RightThing(tm) to do for multiple reasons. I dunno. It's a separate issue. ] It does actually shrink the kernel image a bit (a couple of hundred bytes on the text segment for my everything-compiled-in image), and while it's totally untested the (admittedly few) code generation points I looked at seemed sane. And "lock orb" should be noticeably faster than "lock bts". If somebody wants to play with it, go wild. I didn't do "change_bit()", because nobody sane uses that thing anyway. I guarantee nothing. And if it breaks, nobody saw me do anything. You can't prove this email wasn't sent by somebody who is good at forging smtp. This does require a gcc that is recent enough for "__builtin_constant_p()" to work in an inline function, but I suspect our kernel requirements are already higher than that. And if you do have an old gcc that is supported, the worst that would happen is that the optimization doesn't trigger. Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-06-18 22:03:26 -06:00
#define BITOP_ADDR(x) "=m" (*(volatile long *) (x))
#else
x86, bitops: make constant-bit set/clear_bit ops faster On Wed, 18 Jun 2008, Linus Torvalds wrote: > > And yes, the "lock andl" should be noticeably faster than the xchgl. I dunno. Here's a untested (!!) patch that turns constant-bit set/clear_bit ops into byte mask ops (lock orb/andb). It's not exactly pretty. The reason for using the byte versions is that a locked op is serialized in the memory pipeline anyway, so there are no forwarding issues (that could slow down things when we access things with different sizes), and the byte ops are a lot smaller than 32-bit and particularly 64-bit ops (big constants, and the 64-bit ops need the REX prefix byte too). [ Side note: I wonder if we should turn the "test_bit()" C version into a "char *" version too.. It could actually help with alias analysis, since char pointers can alias anything. So it might be the RightThing(tm) to do for multiple reasons. I dunno. It's a separate issue. ] It does actually shrink the kernel image a bit (a couple of hundred bytes on the text segment for my everything-compiled-in image), and while it's totally untested the (admittedly few) code generation points I looked at seemed sane. And "lock orb" should be noticeably faster than "lock bts". If somebody wants to play with it, go wild. I didn't do "change_bit()", because nobody sane uses that thing anyway. I guarantee nothing. And if it breaks, nobody saw me do anything. You can't prove this email wasn't sent by somebody who is good at forging smtp. This does require a gcc that is recent enough for "__builtin_constant_p()" to work in an inline function, but I suspect our kernel requirements are already higher than that. And if you do have an old gcc that is supported, the worst that would happen is that the optimization doesn't trigger. Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-06-18 22:03:26 -06:00
#define BITOP_ADDR(x) "+m" (*(volatile long *) (x))
#endif
#define ADDR BITOP_ADDR(addr)
x86, bitops: make constant-bit set/clear_bit ops faster On Wed, 18 Jun 2008, Linus Torvalds wrote: > > And yes, the "lock andl" should be noticeably faster than the xchgl. I dunno. Here's a untested (!!) patch that turns constant-bit set/clear_bit ops into byte mask ops (lock orb/andb). It's not exactly pretty. The reason for using the byte versions is that a locked op is serialized in the memory pipeline anyway, so there are no forwarding issues (that could slow down things when we access things with different sizes), and the byte ops are a lot smaller than 32-bit and particularly 64-bit ops (big constants, and the 64-bit ops need the REX prefix byte too). [ Side note: I wonder if we should turn the "test_bit()" C version into a "char *" version too.. It could actually help with alias analysis, since char pointers can alias anything. So it might be the RightThing(tm) to do for multiple reasons. I dunno. It's a separate issue. ] It does actually shrink the kernel image a bit (a couple of hundred bytes on the text segment for my everything-compiled-in image), and while it's totally untested the (admittedly few) code generation points I looked at seemed sane. And "lock orb" should be noticeably faster than "lock bts". If somebody wants to play with it, go wild. I didn't do "change_bit()", because nobody sane uses that thing anyway. I guarantee nothing. And if it breaks, nobody saw me do anything. You can't prove this email wasn't sent by somebody who is good at forging smtp. This does require a gcc that is recent enough for "__builtin_constant_p()" to work in an inline function, but I suspect our kernel requirements are already higher than that. And if you do have an old gcc that is supported, the worst that would happen is that the optimization doesn't trigger. Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-06-18 22:03:26 -06:00
/*
* We do the locked ops that don't return the old value as
* a mask operation on a byte.
*/
#define IS_IMMEDIATE(nr) (__builtin_constant_p(nr))
#define CONST_MASK_ADDR(nr, addr) BITOP_ADDR((void *)(addr) + ((nr)>>3))
#define CONST_MASK(nr) (1 << ((nr) & 7))
x86, bitops: make constant-bit set/clear_bit ops faster On Wed, 18 Jun 2008, Linus Torvalds wrote: > > And yes, the "lock andl" should be noticeably faster than the xchgl. I dunno. Here's a untested (!!) patch that turns constant-bit set/clear_bit ops into byte mask ops (lock orb/andb). It's not exactly pretty. The reason for using the byte versions is that a locked op is serialized in the memory pipeline anyway, so there are no forwarding issues (that could slow down things when we access things with different sizes), and the byte ops are a lot smaller than 32-bit and particularly 64-bit ops (big constants, and the 64-bit ops need the REX prefix byte too). [ Side note: I wonder if we should turn the "test_bit()" C version into a "char *" version too.. It could actually help with alias analysis, since char pointers can alias anything. So it might be the RightThing(tm) to do for multiple reasons. I dunno. It's a separate issue. ] It does actually shrink the kernel image a bit (a couple of hundred bytes on the text segment for my everything-compiled-in image), and while it's totally untested the (admittedly few) code generation points I looked at seemed sane. And "lock orb" should be noticeably faster than "lock bts". If somebody wants to play with it, go wild. I didn't do "change_bit()", because nobody sane uses that thing anyway. I guarantee nothing. And if it breaks, nobody saw me do anything. You can't prove this email wasn't sent by somebody who is good at forging smtp. This does require a gcc that is recent enough for "__builtin_constant_p()" to work in an inline function, but I suspect our kernel requirements are already higher than that. And if you do have an old gcc that is supported, the worst that would happen is that the optimization doesn't trigger. Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-06-18 22:03:26 -06:00
/**
* set_bit - Atomically set a bit in memory
* @nr: the bit to set
* @addr: the address to start counting from
*
* This function is atomic and may not be reordered. See __set_bit()
* if you do not require the atomic guarantees.
*
* Note: there are no guarantees that this function will not be reordered
* on non x86 architectures, so if you are writing portable code,
* make sure not to rely on its reordering guarantees.
*
* Note that @nr may be almost arbitrarily large; this function is not
* restricted to acting on a single-word quantity.
*/
static __always_inline void
set_bit(unsigned int nr, volatile unsigned long *addr)
{
if (IS_IMMEDIATE(nr)) {
asm volatile(LOCK_PREFIX "orb %1,%0"
: CONST_MASK_ADDR(nr, addr)
: "iq" ((u8)CONST_MASK(nr))
: "memory");
} else {
asm volatile(LOCK_PREFIX "bts %1,%0"
: BITOP_ADDR(addr) : "Ir" (nr) : "memory");
}
}
/**
* __set_bit - Set a bit in memory
* @nr: the bit to set
* @addr: the address to start counting from
*
* Unlike set_bit(), this function is non-atomic and may be reordered.
* If it's called on the same region of memory simultaneously, the effect
* may be that only one operation succeeds.
*/
static inline void __set_bit(int nr, volatile unsigned long *addr)
{
asm volatile("bts %1,%0" : ADDR : "Ir" (nr) : "memory");
}
/**
* clear_bit - Clears a bit in memory
* @nr: Bit to clear
* @addr: Address to start counting from
*
* clear_bit() is atomic and may not be reordered. However, it does
* not contain a memory barrier, so if it is used for locking purposes,
* you should call smp_mb__before_clear_bit() and/or smp_mb__after_clear_bit()
* in order to ensure changes are visible on other processors.
*/
static __always_inline void
clear_bit(int nr, volatile unsigned long *addr)
{
if (IS_IMMEDIATE(nr)) {
asm volatile(LOCK_PREFIX "andb %1,%0"
: CONST_MASK_ADDR(nr, addr)
: "iq" ((u8)~CONST_MASK(nr)));
} else {
asm volatile(LOCK_PREFIX "btr %1,%0"
: BITOP_ADDR(addr)
: "Ir" (nr));
}
}
/*
* clear_bit_unlock - Clears a bit in memory
* @nr: Bit to clear
* @addr: Address to start counting from
*
* clear_bit() is atomic and implies release semantics before the memory
* operation. It can be used for an unlock.
*/
static inline void clear_bit_unlock(unsigned nr, volatile unsigned long *addr)
{
barrier();
clear_bit(nr, addr);
}
static inline void __clear_bit(int nr, volatile unsigned long *addr)
{
asm volatile("btr %1,%0" : ADDR : "Ir" (nr));
}
/*
* __clear_bit_unlock - Clears a bit in memory
* @nr: Bit to clear
* @addr: Address to start counting from
*
* __clear_bit() is non-atomic and implies release semantics before the memory
* operation. It can be used for an unlock if no other CPUs can concurrently
* modify other bits in the word.
*
* No memory barrier is required here, because x86 cannot reorder stores past
* older loads. Same principle as spin_unlock.
*/
static inline void __clear_bit_unlock(unsigned nr, volatile unsigned long *addr)
{
barrier();
__clear_bit(nr, addr);
}
#define smp_mb__before_clear_bit() barrier()
#define smp_mb__after_clear_bit() barrier()
/**
* __change_bit - Toggle a bit in memory
* @nr: the bit to change
* @addr: the address to start counting from
*
* Unlike change_bit(), this function is non-atomic and may be reordered.
* If it's called on the same region of memory simultaneously, the effect
* may be that only one operation succeeds.
*/
static inline void __change_bit(int nr, volatile unsigned long *addr)
{
asm volatile("btc %1,%0" : ADDR : "Ir" (nr));
}
/**
* change_bit - Toggle a bit in memory
* @nr: Bit to change
* @addr: Address to start counting from
*
* change_bit() is atomic and may not be reordered.
* Note that @nr may be almost arbitrarily large; this function is not
* restricted to acting on a single-word quantity.
*/
static inline void change_bit(int nr, volatile unsigned long *addr)
{
if (IS_IMMEDIATE(nr)) {
asm volatile(LOCK_PREFIX "xorb %1,%0"
: CONST_MASK_ADDR(nr, addr)
: "iq" ((u8)CONST_MASK(nr)));
} else {
asm volatile(LOCK_PREFIX "btc %1,%0"
: BITOP_ADDR(addr)
: "Ir" (nr));
}
}
/**
* test_and_set_bit - Set a bit and return its old value
* @nr: Bit to set
* @addr: Address to count from
*
* This operation is atomic and cannot be reordered.
* It also implies a memory barrier.
*/
static inline int test_and_set_bit(int nr, volatile unsigned long *addr)
{
int oldbit;
asm volatile(LOCK_PREFIX "bts %2,%1\n\t"
"sbb %0,%0" : "=r" (oldbit), ADDR : "Ir" (nr) : "memory");
return oldbit;
}
/**
* test_and_set_bit_lock - Set a bit and return its old value for lock
* @nr: Bit to set
* @addr: Address to count from
*
* This is the same as test_and_set_bit on x86.
*/
static __always_inline int
test_and_set_bit_lock(int nr, volatile unsigned long *addr)
{
return test_and_set_bit(nr, addr);
}
/**
* __test_and_set_bit - Set a bit and return its old value
* @nr: Bit to set
* @addr: Address to count from
*
* This operation is non-atomic and can be reordered.
* If two examples of this operation race, one can appear to succeed
* but actually fail. You must protect multiple accesses with a lock.
*/
static inline int __test_and_set_bit(int nr, volatile unsigned long *addr)
{
int oldbit;
asm("bts %2,%1\n\t"
"sbb %0,%0"
: "=r" (oldbit), ADDR
: "Ir" (nr));
return oldbit;
}
/**
* test_and_clear_bit - Clear a bit and return its old value
* @nr: Bit to clear
* @addr: Address to count from
*
* This operation is atomic and cannot be reordered.
* It also implies a memory barrier.
*/
static inline int test_and_clear_bit(int nr, volatile unsigned long *addr)
{
int oldbit;
asm volatile(LOCK_PREFIX "btr %2,%1\n\t"
"sbb %0,%0"
: "=r" (oldbit), ADDR : "Ir" (nr) : "memory");
return oldbit;
}
/**
* __test_and_clear_bit - Clear a bit and return its old value
* @nr: Bit to clear
* @addr: Address to count from
*
* This operation is non-atomic and can be reordered.
* If two examples of this operation race, one can appear to succeed
* but actually fail. You must protect multiple accesses with a lock.
*
* Note: the operation is performed atomically with respect to
* the local CPU, but not other CPUs. Portable code should not
* rely on this behaviour.
* KVM relies on this behaviour on x86 for modifying memory that is also
* accessed from a hypervisor on the same CPU if running in a VM: don't change
* this without also updating arch/x86/kernel/kvm.c
*/
static inline int __test_and_clear_bit(int nr, volatile unsigned long *addr)
{
int oldbit;
asm volatile("btr %2,%1\n\t"
"sbb %0,%0"
: "=r" (oldbit), ADDR
: "Ir" (nr));
return oldbit;
}
/* WARNING: non atomic and it can be reordered! */
static inline int __test_and_change_bit(int nr, volatile unsigned long *addr)
{
int oldbit;
asm volatile("btc %2,%1\n\t"
"sbb %0,%0"
: "=r" (oldbit), ADDR
: "Ir" (nr) : "memory");
return oldbit;
}
/**
* test_and_change_bit - Change a bit and return its old value
* @nr: Bit to change
* @addr: Address to count from
*
* This operation is atomic and cannot be reordered.
* It also implies a memory barrier.
*/
static inline int test_and_change_bit(int nr, volatile unsigned long *addr)
{
int oldbit;
asm volatile(LOCK_PREFIX "btc %2,%1\n\t"
"sbb %0,%0"
: "=r" (oldbit), ADDR : "Ir" (nr) : "memory");
return oldbit;
}
static __always_inline int constant_test_bit(unsigned int nr, const volatile unsigned long *addr)
{
return ((1UL << (nr % BITS_PER_LONG)) &
(addr[nr / BITS_PER_LONG])) != 0;
}
static inline int variable_test_bit(int nr, volatile const unsigned long *addr)
{
int oldbit;
asm volatile("bt %2,%1\n\t"
"sbb %0,%0"
: "=r" (oldbit)
: "m" (*(unsigned long *)addr), "Ir" (nr));
return oldbit;
}
#if 0 /* Fool kernel-doc since it doesn't do macros yet */
/**
* test_bit - Determine whether a bit is set
* @nr: bit number to test
* @addr: Address to start counting from
*/
static int test_bit(int nr, const volatile unsigned long *addr);
#endif
#define test_bit(nr, addr) \
(__builtin_constant_p((nr)) \
? constant_test_bit((nr), (addr)) \
: variable_test_bit((nr), (addr)))
/**
* __ffs - find first set bit in word
* @word: The word to search
*
* Undefined if no bit exists, so code should check against 0 first.
*/
static inline unsigned long __ffs(unsigned long word)
{
asm("rep; bsf %1,%0"
: "=r" (word)
: "rm" (word));
return word;
}
/**
* ffz - find first zero bit in word
* @word: The word to search
*
* Undefined if no zero exists, so code should check against ~0UL first.
*/
static inline unsigned long ffz(unsigned long word)
{
asm("rep; bsf %1,%0"
: "=r" (word)
: "r" (~word));
return word;
}
/*
* __fls: find last set bit in word
* @word: The word to search
*
* Undefined if no set bit exists, so code should check against 0 first.
*/
static inline unsigned long __fls(unsigned long word)
{
asm("bsr %1,%0"
: "=r" (word)
: "rm" (word));
return word;
}
#undef ADDR
#ifdef __KERNEL__
/**
* ffs - find first set bit in word
* @x: the word to search
*
* This is defined the same way as the libc and compiler builtin ffs
* routines, therefore differs in spirit from the other bitops.
*
* ffs(value) returns 0 if value is 0 or the position of the first
* set bit if value is nonzero. The first (least significant) bit
* is at position 1.
*/
static inline int ffs(int x)
{
int r;
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
#ifdef CONFIG_X86_64
/*
* AMD64 says BSFL won't clobber the dest reg if x==0; Intel64 says the
* dest reg is undefined if x==0, but their CPU architect says its
* value is written to set it to the same as before, except that the
* top 32 bits will be cleared.
*
* We cannot do this on 32 bits because at the very least some
* 486 CPUs did not behave this way.
*/
asm("bsfl %1,%0"
: "=r" (r)
: "rm" (x), "0" (-1));
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
#elif defined(CONFIG_X86_CMOV)
asm("bsfl %1,%0\n\t"
"cmovzl %2,%0"
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
: "=&r" (r) : "rm" (x), "r" (-1));
#else
asm("bsfl %1,%0\n\t"
"jnz 1f\n\t"
"movl $-1,%0\n"
"1:" : "=r" (r) : "rm" (x));
#endif
return r + 1;
}
/**
* fls - find last set bit in word
* @x: the word to search
*
* This is defined in a similar way as the libc and compiler builtin
* ffs, but returns the position of the most significant set bit.
*
* fls(value) returns 0 if value is 0 or the position of the last
* set bit if value is nonzero. The last (most significant) bit is
* at position 32.
*/
static inline int fls(int x)
{
int r;
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
#ifdef CONFIG_X86_64
/*
* AMD64 says BSRL won't clobber the dest reg if x==0; Intel64 says the
* dest reg is undefined if x==0, but their CPU architect says its
* value is written to set it to the same as before, except that the
* top 32 bits will be cleared.
*
* We cannot do this on 32 bits because at the very least some
* 486 CPUs did not behave this way.
*/
asm("bsrl %1,%0"
: "=r" (r)
: "rm" (x), "0" (-1));
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
#elif defined(CONFIG_X86_CMOV)
asm("bsrl %1,%0\n\t"
"cmovzl %2,%0"
: "=&r" (r) : "rm" (x), "rm" (-1));
#else
asm("bsrl %1,%0\n\t"
"jnz 1f\n\t"
"movl $-1,%0\n"
"1:" : "=r" (r) : "rm" (x));
#endif
return r + 1;
}
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
/**
* fls64 - find last set bit in a 64-bit word
* @x: the word to search
*
* This is defined in a similar way as the libc and compiler builtin
* ffsll, but returns the position of the most significant set bit.
*
* fls64(value) returns 0 if value is 0 or the position of the last
* set bit if value is nonzero. The last (most significant) bit is
* at position 64.
*/
#ifdef CONFIG_X86_64
static __always_inline int fls64(__u64 x)
{
int bitpos = -1;
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
/*
* AMD64 says BSRQ won't clobber the dest reg if x==0; Intel64 says the
* dest reg is undefined if x==0, but their CPU architect says its
* value is written to set it to the same as before.
*/
asm("bsrq %1,%q0"
x86_64, asm: Optimise fls(), ffs() and fls64() fls(N), ffs(N) and fls64(N) can be optimised on x86_64. Currently they use a CMOV instruction after the BSR/BSF to set the destination register to -1 if the value to be scanned was 0 (in which case BSR/BSF set the Z flag). Instead, according to the AMD64 specification, we can make use of the fact that BSR/BSF doesn't modify its output register if its input is 0. By preloading the output with -1 and incrementing the result, we achieve the desired result without the need for a conditional check. The Intel x86_64 specification, however, says that the result of BSR/BSF in such a case is undefined. That said, when queried, one of the Intel CPU architects said that the behaviour on all Intel CPUs is that: (1) with BSRQ/BSFQ, the 64-bit destination register is written with its original value if the source is 0, thus, in essence, giving the effect we want. And, (2) with BSRL/BSFL, the lower half of the 64-bit destination register is written with its original value if the source is 0, and the upper half is cleared, thus giving us the effect we want (we return a 4-byte int). Further, it was indicated that they (Intel) are unlikely to get away with changing the behaviour. It might be possible to optimise the 32-bit versions of these functions, but there's a lot more variation, and so the effective non-destructive property of BSRL/BSRF cannot be relied on. [ hpa: specifically, some 486 chips are known to NOT have this property. ] I have benchmarked these functions on my Core2 Duo test machine using the following program: #include <stdlib.h> #include <stdio.h> #ifndef __x86_64__ #error #endif #define PAGE_SHIFT 12 typedef unsigned long long __u64, u64; typedef unsigned int __u32, u32; #define noinline __attribute__((noinline)) static __always_inline int fls64(__u64 x) { long bitpos = -1; asm("bsrq %1,%0" : "+r" (bitpos) : "rm" (x)); return bitpos + 1; } static inline unsigned long __fls(unsigned long word) { asm("bsr %1,%0" : "=r" (word) : "rm" (word)); return word; } static __always_inline int old_fls64(__u64 x) { if (x == 0) return 0; return __fls(x) + 1; } static noinline // __attribute__((const)) int old_get_order(unsigned long size) { int order; size = (size - 1) >> (PAGE_SHIFT - 1); order = -1; do { size >>= 1; order++; } while (size); return order; } static inline __attribute__((const)) int get_order_old_fls64(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = old_fls64(size); return order; } static inline __attribute__((const)) int get_order(unsigned long size) { int order; size--; size >>= PAGE_SHIFT; order = fls64(size); return order; } unsigned long prevent_optimise_out; static noinline unsigned long test_old_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += old_get_order(n); } } return total; } static noinline unsigned long test_get_order_old_fls64(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order_old_fls64(n); } } return total; } static noinline unsigned long test_get_order(void) { unsigned long n, total = 0; long rep, loop; for (rep = 1000000; rep > 0; rep--) { for (loop = 0; loop <= 16384; loop += 4) { n = 1UL << loop; total += get_order(n); } } return total; } int main(int argc, char **argv) { unsigned long total; switch (argc) { case 1: total = test_old_get_order(); break; case 2: total = test_get_order_old_fls64(); break; default: total = test_get_order(); break; } prevent_optimise_out = total; return 0; } This allows me to test the use of the old fls64() implementation and the new fls64() implementation and also to contrast these to the out-of-line loop-based implementation of get_order(). The results were: warthog>time ./get_order real 1m37.191s user 1m36.313s sys 0m0.861s warthog>time ./get_order x real 0m16.892s user 0m16.586s sys 0m0.287s warthog>time ./get_order x x real 0m7.731s user 0m7.727s sys 0m0.002s Using the current upstream fls64() as a basis for an inlined get_order() [the second result above] is much faster than using the current out-of-line loop-based get_order() [the first result above]. Using my optimised inline fls64()-based get_order() [the third result above] is even faster still. [ hpa: changed the selection of 32 vs 64 bits to use CONFIG_X86_64 instead of comparing BITS_PER_LONG, updated comments, rebased manually on top of 83d99df7c4bf x86, bitops: Move fls64.h inside __KERNEL__ ] Signed-off-by: David Howells <dhowells@redhat.com> Link: http://lkml.kernel.org/r/20111213145654.14362.39868.stgit@warthog.procyon.org.uk Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2011-12-13 07:56:54 -07:00
: "+r" (bitpos)
: "rm" (x));
return bitpos + 1;
}
#else
#include <asm-generic/bitops/fls64.h>
#endif
#include <asm-generic/bitops/find.h>
#include <asm-generic/bitops/sched.h>
#define ARCH_HAS_FAST_MULTIPLIER 1
#include <asm/arch_hweight.h>
#include <asm-generic/bitops/const_hweight.h>
#include <asm-generic/bitops/le.h>
#include <asm-generic/bitops/ext2-atomic-setbit.h>
#endif /* __KERNEL__ */
#endif /* _ASM_X86_BITOPS_H */