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 <[e-mail]>
Link: http://lkml.kernel.org/r/[e-mail]
Cc: Linus Torvalds <[e-mail]>
Signed-off-by: H. Peter Anvin <[e-mail]>

arch/x86/include/asm/bitops.h

···
 static inline int ffs(int x)
 {
 	int r;
-#ifdef CONFIG_X86_CMOV
+
+#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.
+	 */
+	long tmp = -1;
+	asm("bsfl %1,%0"
+	    : "=r" (r)
+	    : "rm" (x), "0" (tmp));
+#elif defined(CONFIG_X86_CMOV)
 	asm("bsfl %1,%0\n\t"
 	    "cmovzl %2,%0"
-	    : "=r" (r) : "rm" (x), "r" (-1));
+	    : "=&r" (r) : "rm" (x), "r" (-1));
 #else
 	asm("bsfl %1,%0\n\t"
 	    "jnz 1f\n\t"
···
 static inline int fls(int x)
 {
 	int r;
-#ifdef CONFIG_X86_CMOV
+
+#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.
+	 */
+	long tmp = -1;
+	asm("bsrl %1,%0"
+	    : "=r" (r)
+	    : "rm" (x), "0" (tmp));
+#elif defined(CONFIG_X86_CMOV)
 	asm("bsrl %1,%0\n\t"
 	    "cmovzl %2,%0"
 	    : "=&r" (r) : "rm" (x), "rm" (-1));
···
 	return r + 1;
 }
 
+/**
+ * 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)
+{
+	long bitpos = -1;
+	/*
+	 * 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,%0"
+	    : "+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>
···
 #include <asm/arch_hweight.h>
 
 #include <asm-generic/bitops/const_hweight.h>
-
-#include <asm-generic/bitops/fls64.h>
 
 #include <asm-generic/bitops/le.h>