I have occasional problems getting ghc-5.04.1 to compile a very large source file with -O switched on. The source file is generated by another tool, and is rather large at 25762 lines (2776340 bytes).
Without -O, the compilation goes normally, taking about 10 minutes or so.
With -O, the compiler uses >280Mb of memory, which causes my machine with only 256Mb to swap like crazy, but that is OK, because after about half an hour, ghc produces its output in a .hc file. The C compiler (gcc-2.95.3) then produces a .raw_s file:
$ wc /tmp/ghc* 759338 837040 18321347 /tmp/ghc3992.hc 1711905 3174489 32014101 /tmp/ghc3992.raw_s
eek! :-S
It's the next stage, when (I presume) the ghc-mangler converts the .raw_s file to a simpler .s file, that doesn't run to completion. It simply reports
Killed make: *** [targets/ix86-Linux/hat-lib-ghc] Error 2
It seems that if I re-run ghc on the intermediate .hc file that was left behind in the /tmp directory, then the mangler is ok, and assembly proceeds right up to the .o file.
So is there any (fixable) reason why the one-step invocation fails, but the two-step invocation succeeds?
I have no idea what might cause this. If you put the .hc up for download somewhere, I'll take a look. In the meantime, you might want to try using the native code generator instead. Just add -fasm to the command line. Cheers, Simon