Error building with MSVC as the C compiler and using Cygwin as the shell
AlaskanEmily opened this issue · 6 comments
Trying to build with --with-cc=cl from the Cygwin shell, there are some issues with the ml script. It appears that when --no-mercury-stdlib-dir is specified, then the $LIBDIR variable is unset, which results in an error because merc_libdir_opts is set to -LIBPATH:$LIBPATH/$GRADE -LIBPATH:$LIBPATH.
The first option is clearly incorrect (it comes out as C:\Cygwin64$GRADE which is wrong), but the second empty -LIBPATH: option specifically causes a fatal linker error because the missing argument.
I remember this happening before, in like 2018 or 2019, but I don't remember if or how it was fixed (I might just have copied an older set of files from the scripts directory to fix it locally). I took a look at a few quick fixes just to test, like just not setting merc_libdir_opts when $LIBDIR is empty, but doing that caused its own new set of linking errors.
Is this bug report for the current ROTD, or for the beta of the 22.01 release?
This is with ROTD 2022-02-16, I'll test again with the newest ROTD to be sure
The problem also occurs when building with MSVC under MSYS2. It appears that the ml script has been broken at some point.
@AlaskanEmily the other step you (probably) need to take to avoid linker errors is set the definition of the variable LINKER_POST_FLAGS in the ml script to empty. I'll take further look into what has happened.
The problem here is that the code that forces static linkage for MSVC in the ml script, does so regardless of whether the option --no-mercury-stdlib-dir is given or not. If option processing determines that the value of mercury_libs is none, it should stay that way; in particular just using MSVC should not set it to static. I'm just testing a fix for this.
I'm testing the change, seems to work so far
Yep, doing a full rebuild this change fixes the issue. Thanks!