linux/log2.h: reduce instruction count for is_power_of_2()

Follow an observation that (n ^ (n - 1)) will only ever retain the most
significant bit set in the word operated on if that is the only bit set in
the first place, and use it to determine whether a number is a whole power
of 2, avoiding the need for an explicit check for nonzero.

This reduces the sequence produced to 3 instructions only across Alpha,
MIPS, and RISC-V targets, down from 4, 5, and 4 respectively, removing a
branch in the two latter cases.  And it's 5 instructions on POWER and
x86-64 vs 8 and 9 respectively.  There are no branches now emitted here
for targets that have a suitable conditional set operation, although an
inline expansion will often end with one, depending on what code a call to
this function is used in.

Credit goes to GCC authors for coming up with this optimisation used as
the fallback for (__builtin_popcountl(n) == 1), equivalent to this code,
for targets where the hardware population count operation is considered
expensive.

Link: https://lkml.kernel.org/r/[email protected]
Signed-off-by: Maciej W. Rozycki <[email protected]>
Cc: Jens Axboe <[email protected]>
Cc: John Garry <[email protected]>
Cc: "Martin K. Petersen" <[email protected]>
Cc: Su Hui <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>

Author: maciej-w-rozycki

include/linux/log2.h

@@ -44,7 +44,7 @@ int __ilog2_u64(u64 n)
 static __always_inline __attribute__((const))
 bool is_power_of_2(unsigned long n)
 {
-	return (n != 0 && ((n & (n - 1)) == 0));
+	return n - 1 < (n ^ (n - 1));
 }
 
 /**