Our implementation did the classic add-sub rounding trick `(y = x +/- C =+ C - x)`
with `C = 1 / eps(T) = 2^(mantissa - 1)`. This approach only works for values whose
magnitude is below the rounding capacity of the constant. For a 64-bit mantissa
(like f80 has), `C = 2^63` only rounds for `|x| < 2^63`. Before we allowed this to
be ran on `e < bias + 64` aka `|x| < 2^64`. And because it isn't large enough,
we lose a bit to rounding.
For reference, the musl implementation does the same thing, using `mantissa - 1`:
https://git.musl-libc.org/cgit/musl/tree/src/math/ceill.c#n18
where `LDBL_MANT_DIG` is 64 for `long double` on x86.
This commit also combines the floor and ceil implementations into one generic one.
Previously these functions made the assumption that
when performing a on the input digits,
there could be no collisions between the less
significant digits being larger than '9', and the
upper digits being small enough to get past the
checks.
Now we perform a correct check across all of the
digits to ensure they're in between '0'-'9', at
a minimal cost, since all digits are checked in
parallel.
I believe these tests may have been flaky as a result of the bug fixed
in the previous commit. A big hint is that they were all crashing with
SIGSEGV with no stack trace. I suspect that some lingering SIGIOs from
cancelations were being delivered to a thread after its `munmap` call,
which was happening because the test runner called `Io.Threaded.deinit`
to cause all of the (detached) worker threads to exit.
If this passes, I'll re-run the x86_64-linux CI jobs on this commit a
few times before merge to try and be sure there are no lingering
failures.
Resolves: https://codeberg.org/ziglang/zig/issues/30096
Resolves: https://codeberg.org/ziglang/zig/issues/30592
Resolves: https://codeberg.org/ziglang/zig/issues/30682
As the comment explains, if a signal were to arrive between a detached
thread's `munmap` and `exit` calls, the signal handler would immediately
trigger SIGSEGV due to the stack being unmapped. To solve this, we need
to block all signals before entering this logic. The musl implementation
which this logic was ported from does this exact thing; that logic was
just lost when porting.
Notably, this would lead to a crash with no stack trace, because the
SIGSEGV handler would itself crash due to the missing stack.
It doesn't make any sense for a task to be canceled while it's
panicking.
As a happy accident, this also solves some cases where safety panics in
`Io.Threaded` would cause stack traces not to print due to invalid
thread-local state: when cancelation is blocked, `Io.Threaded` doesn't
consult said thread-local state at all. For instance, try inserting a
panic just after a call to `Syscall.start()` in `Io.Threaded`, and then
call the `Io` function in question from a `concurrent` task. Before this
PR, the stack trace fails to print, because the panic handler sees the
thread-local cancelation state in an unexpected state, leading to a
recursive panic. After this PR, the stack trace prints fine.
Andrew Kelley