On Thu, May 28, 2026 at 12:12:09PM -0500, Simon Glass wrote: > Hi Tom, > > On Wed, 27 May 2026 at 21:43, Tom Rini <[email protected]> wrote: > > > > On Wed, May 27, 2026 at 08:36:45PM +0200, Heinrich Schuchardt wrote: > > > On 5/27/26 18:10, Simon Glass wrote: > > > > Sandbox needs to include system headers in some files, but also wants > > > > to use alist. Adjust the headers to permit this. > > > > > > > > Signed-off-by: Simon Glass <[email protected]> > > > > --- > > > > > > > > include/alist.h | 16 +++++++++++----- > > > > lib/alist.c | 1 + > > > > 2 files changed, 12 insertions(+), 5 deletions(-) > > > > > > > > diff --git a/include/alist.h b/include/alist.h > > > > index b00d9ea97d6..69d7cdb722f 100644 > > > > --- a/include/alist.h > > > > +++ b/include/alist.h > > > > @@ -10,8 +10,14 @@ > > > > #define __ALIST_H > > > > #include <stdbool.h> > > > > -#include <linux/bitops.h> > > > > + > > > > +#ifdef USE_HOSTCC > > > > +#include <sys/types.h> > > > > +#include <stdint.h> > > > > +#else > > > > #include <linux/types.h> > > > > +#endif > > > > +#define BIT(nr) (1UL << (nr)) > > > > > > BIT(nr) is already defined in include/linux/bitops.h. > > > > > > We must not assume that no module includes both alist.h and > > > linux/bitops.h. > > > E.g. include/expo.h and include/lmb.h include both alist.h and > > > linux/bitops.h. We should avoid redefinitions. > > > > > > The following is an abuse of the enum type: > > > > > > enum alist_flags { > > > ALISTF_FAIL = BIT(0), > > > }; > > > > > > As this is the only use of BIT(), we could simply use the value "1" here. > > > > > > #define ALISTF_FAIL (1) > > > > > > But as ALISTF_FAIL is the only bit in flags used, the best solution would > > > be > > > replacing flags by a boolean called fail. > > > > I see this as another reminder about how frustrating it was that alist > > was introduced under the guise of "must have this for an x86 feature" > > but instead is a solution in need of a problem to solve, and whenever > > it's used for something else, yet another problem with it is shown. > > The alist was motivated by its original commit. It allows easy > iteration, uses less memory (no next/prev pointers in every element), > uses fewer allocations than linked lists (one malloc() for the whole > list instead of one for each node) and has much better cache locality > as a result (i.e. improved performance). > > I asked my friendly AI to compare them: > > - Reach for alist when you want a growable, index-addressable, > cache-friendly > collection of small value objects you mostly append to and iterate (e.g. > bootflow/bootdev scan results, parsed tables). > - Reach for list_head when elements need O(1) removal/splice from arbitrary > positions, must keep stable identity while linked, or already exist as > independently-allocated objects (most driver-model / subsystem object > lists). > > One nuance: alist_for_each is pure pointer arithmetic over the buffer, > whereas > list_for_each_entry uses container_of to recover the enclosing struct from > the embedded node — a direct consequence of array-of-values vs. > intrusive-links.
This changes absolutely nothing about what I said the last time you wanted to use alist more widely, and in turn what I said about it a year ago, or whenever that was. I'm not going to go ask an unfriendly AI to say what's wrong with alist, I'm just going to say that I already said it. -- Tom
signature.asc
Description: PGP signature
