https://issues.dlang.org/show_bug.cgi?id=24129
--- Comment #2 from Jeffrey H. Johnson <tr...@pobox.com> ---
(In reply to Walter Bright from comment #1)
> The problem stems from when a.c and b.c are compiled together, as in:
>
> dmd a.c b.c
>
> The code from both gets placed in one object file, a.obj. Hence there are
> two copies of __acrt_local_get_ctype_array value, and they conflict with
> each other.
>
> The workaround is to compile them separately.
Thanks Walter! Do you think this might be something that can be worked around?
The reason is performance, for example, on Linux:
$ for i in *.c; do dmd -m64 -fPIC -fPIE -O -release -check=off -boundscheck=off
"${i}" -c -of=objout/${i}.o; done
$ dmd -m64 -fPIC -fPIE -O -release -check=off -boundscheck=off objout/*.o
-of=test1
$ strip test1
$ ls -la test1
885 KB Wed Sep 6 19:29:54 2023 test1*
$ rm -rf objout/*
$ dmd -m64 -fPIC -fPIE -O -release -check=off -boundscheck=off *.c -of=test2
$ strip test2
$ ls -la test2
838 KB Wed Sep 6 19:30:27 2023 test2*
It's quite close, but I can show (with statistical confidence) that the test2
build is a bit faster than test1 (in addition to being smaller and easier to
compile). I assume that this is due to some linker optimizations? Does the
compiler optimize better, almsot a "poor mans LTO" and treating it as a single
translation unit, perhaps?
I believe that using -inline (when it works, I made bug #24131 for that) also
results in faster binaries when they are all built together, rather than
separately.
However, as a workaround, this seems acceptable for now, thank you so much.
And I assume this is still a bug, so I'll leave this open for now.
--
