Make bitpos() handle arguments with many set bits

[dougm]

Summary

When bitpos can't be implemented with an inline ffs* instruction, change the binary search so that it does not depend on a single bit only being set in the bitmask. Use bitpos more generally, and avoid some clearing of bits to accommodate its current behavior.

[kib]

So why not use _Generic ?

subr_blist.c

224 ↗

(On Diff #57283)

Is this case ever occur ? I believe daddr_t is always 64bit.
For this argument, you might actually add 'long'.

[dougm]

Apply _Generic instead of switch. Not previously acquainted with that construct.

[dougm]

subr_blist.c

224 ↗

(On Diff #57283)

Where I'm testing, long long and long have the same size and the switch led to a compilation error.

Alan has suggested that someday he hopes to use a 32-bit daddr_t for 32-bit machines, which is why I bother sticking the int case in there.

[kib]

subr_blist.c

224 ↗

(On Diff #57283)

Did you tested with switch, or with _Generic. I would expect that you can use long with generic, since all three types 'long long', 'long' and 'int' are different, even if some of them have identical representation.

[dougm]

Add a 'long' case to the _Generic for bitpos, which compiles successfully.

[dougm]

_Generic broke several builds. I'll prepare a patch to undo it, in case you don't have a better solution to offer.

[kib]

In D20232#436021, @dougm wrote:

_Generic broke several builds. I'll prepare a patch to undo it, in case you don't have a better solution to offer.

If you intend reverting the commit, you got the implicit approval.

Can you point me to the reports ?