in case the rotation is 0, `v >> (32 - r)` would end up doing a 32 right
shift which is equal to the width of `v` and thus undefined behavior.
ref: https://blog.regehr.org/archives/1063
* icons-hash: take total probe count into account as well
* icons-hash: use a better PRNG
the older method was using a multiplicative congruential generator (MCG)
which doesn't work too well especially with just 32 bits of state.
change it to a PCG instead with 64 bits of state (and 32 bits of output)
which should give better results.
and since we should get better rng - the search iteration has been
halved as well to save some build time.
* icons-hash: use an xor-rotate hash function
* icons-hash: fix some compiler warnings
1. hover on entry created in cwd
2. check user input doesn't end with /
3. check user input len < PATH_MAX
4. support creation in ~ directory
5. handle filter mode after creation
6. ensure absolute path len < PATH_MAX
nnn source code already uses C11, which introduced `_Alignas` [^0] as a
standardized alternative to __attribute__((aligned)).
there are no other usage of `__attribute__` in the nnn source code, so
in theory this should make things more portable across compilers.
also removes an unnecessary usage of `alloca()` with an aligned constant
sized buffer instead.
[^0]: https://en.cppreference.com/w/c/language/_Alignas
Found memory deallocation edge case
Update and move chunk limit check
Generalize maximum size of input
Remove hard-coded values
Remove superfluous check before free
Let the kernel deal with extra data
Handle signals while reading
Conform to the manpage
Make CI happy
use `size_t` instead of `ssize_t`
`ssize_t` was used just so `--i` when `i` was zero would become -1
instead of SIZE_MAX. for looping through something in reverse order, the
"goes-to" operator (`-->`) can be used instead which doesn't require `i`
to be signed anymore.
remove useless blank line
use a normal loop
don't see any reason why freeing in reverse order would've been needed.
Co-authored-by: N-R-K <nrk@disroot.org>
- support listing maximum 16K files
- check if target directory exists before directory tree creation
in most of the cases many files will be under the same directory
- make frequently used function 'inline'
* Added icons for erlang, elixir and lockfile
* Lockfile icon updated
* Added colors for Erlang and elixir
* Added few more color variations and configured colors and file
extensions
* remove erlang and lock, use 1 color for elixir
Co-authored-by: Tharindu Abeydeera <tharindu.a@vizuamatix.com>
Co-authored-by: NRK <nrk@disroot.org>
a lot of the extension use the same icon. this can be exploited via
having an array with all the unique icons and then storing a single byte
index into the unique array.
when using `O_EMOJI` this results in around ~1.7KiB drop in the total
table size. `O_NERD` and `O_ICONS` get roughly ~0.5KiB savings.
This pretty much reworks the entire icon system. Some notable changes:
* The extensions are put into a statically generated hash-table instead
of a sorted array. We use Robin-Hood insertion to reduce the max probe
length. Currently we need to probe only 2 slots for `O_EMOJI` and only
3 for `O_NERD`/`O_ICONS`.
* I've opted not to use a perfect-hash since the perfect hashes
generated by [`gperf`](https://www.gnu.org/software/gperf) used some
huge lookup table. The hash function also wasn't as minimal as I'd
like.
* Colors are now using X-Macros. This should speed up startup since we
don't have to search `icons_ext` linearly to find unique colors.
* The hash-table generator outputs a more space optimized `struct
icon_pair` using a char array instead of char pointer. This brings
down the binary size from `145KiB` when using `O_NERD` down to
`137KiB`.
* Some unnecessary duplication and indirection has been reduced by using
the `ICON_STR()` macro.
reduces some unncessary code. and when initializing larger objects,
compilers (gcc and clang at least) typically tend to compile it down to
a memset anyways.
anything not explicitly initialized will be implicitly initialized to
zero. this makes things more robust since comments are not checked by
the compiler and can be incorrect.
Reasons:
1. `b` is not a special key and will be interpreted as a filter
2. with this change pressing the standard `-` inside a symlinked
bookmark takes back to the bookmarks directory. This deviates
from the regular bookmarks behaviour.
