Stupid Compiler

Today, I spent a few hours trying to figure out why my SMLNJ install kept failing. For the benefit of others who may have run into this problem, I’m going to record my solution (which I haven’t seen anywhere).

Issue 1:
I had a native Win32 install of SMLNJ at some point and my first failure looks like:

$ ./config/install.sh
...
//fs01/Home/dblazakis/projects/smlnj/bin/.link-sml: line 45: c:smlnj/bin/.arch-n-opsys: No such file or directory
...

Notice the native windows path — once I saw that I remembered I had SMLNJ_HOME set. So, from bash, ‘unset SMLNJ_HOME’.

Issue 2:
With that out of the way, I figured I was golden… but I was wrong. The next error:

$ ./config/install.sh
...
./config/install.sh: CM metadata directory name is ".cm"
exception raised during init phase: SysErr: No such file or directory [noent]
//fs01/Home/dblazakis/projects/smlnj/bin/.run/run.x86-cygwin: Fatal error -- Uncaught exception SysErr with <unknown> raised at <stat.c>
./config/install.sh !!! Boot code failed, no heap image (sml.x86-cygwin).

This one is more troublesome. First thought is that it is a path issue, so I add a few echos to the install script and everything looks correct to me. Next, I add a strace to the call in the install script which is failing (a call to generate the initial heap image containing the preloaded modules). This generates a ton of output even after masking to output only syscalls. Sifting through it, it appears the first open or stat call after reading the PIDMAP uses a single ‘/’ instead of the usual double ‘/’ to denote a network share.

...
340 27275136 [main] run.x86-cygwin 2956 normalize_posix_path: src /fs01/Home/dblazakis/projects/smlnj/sml.boot.x86-unix/smlnj/init
...

Poor cygwin — it tries so hard. This appears to be a bug in the SMLNJ SrcPath module (cm/paths/srcpath.sml), but this is the first time I’ve waded through the SMLNJ code.

Long story short, I decided to move my build to a local directory (instead of a network share) and it worked on the first try.