at first I thought this must be a bug, but this could also be a space-
saving optimization to unique the redundant $PWD prefix of all paths.
Is this a bug or a feature?
Based on extremely quick experiments, I think it's simple-minded and
not purposely factoring out common prefixes; directories are just PWD
and -Ifoo, and filenames are the rest of the path, whether it comes
from the command line or an #include directive.
Interestingly, I get different .debug_line sections depending on
compiling straight to object versus compiling to assembler and then
assembling. When we compile to object, all the directories get moved
to the directory table. Assembler doesn't seem to do that.
But even in my Mr. Pedantic mode, it's hard to say it's wrong to put
directories into the file table. There's no semantic distinction in
DWARF, the consumer is expected to paste the strings together.
Yeah, I Think this is intentional to provide paths relative to some root - at least Chromium builds use -fdebug-compilation-dir to avoid baking in the random directory on various distributed build machines, but the “filename” is still qualified relative to that directory so that the debugger can lookup the relative paths itself.
Apparently LLVM is smart enough to not emit this as “/var/empty/var/empty/…”, but I don’t understand why we wouldn’t strip the directory prefix in CGDebugInfo::getOrCreateFile() to save space, or at least not emit a redundant directory.
I’ll send out a patch to strip a common prefix. It should save a little bit of memory and go better with my sense of aesthetics
Let me know if someone has an idea why this would be useful to keep.