------- Comment #4 from matz at gcc dot gnu dot org 2009-01-28 12:22 ------- I'm fairly sure it's the same underlying problem. The emitted file numbers (in the .s file for debug info) are in garbage collected memory. Hence the directives emitted when the PCH is not yet loaded and those emitted after it's loaded might use the same file number. In this case you use -g3, meaning to emit directives for defined macros. These are emitted (obviously) as soon as the include directive is seen. This will allocate numbers 1 and 2 for main.cpp and main.h. Then the PCH file is loaded, overwriting the internally (in dwarf2out.c:struct dwarf_file_data) stored file numbers, so we start counting from 2 again (we overwrite only data from after the main file), which is given to main.cpp again.
Hence the assembler gives the error message. Later compilers simply segfault already earlier in the process, because location info is also overwritten but still referenced, as detailed in the https://bugzilla.novell.com/show_bug.cgi?id=444153 . Also, as detailed there, it seems to be a fundamental design choice that the PCH has to be included from the main file, as the PCH machinery uses the garbage collection engine to actually retrieve the memory state like it was when precompiling the file. This obviously can only work if no stale references exist to then overwritten information and this in turn can only be ensured when there isn't "too much" information before loading the PCH file at all. No idea how easy this would be to fix, from a cursory look at the whole machinery it looks quite difficult (at least if one wants to really fix this, and not just work-around the cases that people happen to hit) -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38987
