Canonicalize input paths in source rewriter

[randomPoison]

Make RootDirectory and OutputDirectory canonical paths immediately after reading CLI args. #432 makes them absolute, but we need them to be fully canonical to reliably compare.

[fw-immunant]

We might prefer llvm::sys::fs::make_absolute just to follow the pattern of using llvm functions here and avoid the <filesystem> include.

[kkysen]

Isn't this already done in #432?

[randomPoison]

Oh weird you're right. I'm pretty sure I've synced and rebuilt IA2 since then, so I'm not sure why I was still seeing issues with relative paths. I'll dig into it on my end, maybe I just didn't have the latest version if IA2.

[randomPoison]

Seems like the logic added in #432 isn't correctly canonicalizing RootDirectory and OutputDirectory. If I run the rewriter with --root-directory=.. --output-directory=. and have the rewriter print out RootDirectory and OutputDirectory, I get:

RootDirectory: /home/legare/zlib/build/..
OutputDirectory: /home/legare/zlib/build/.

It looks like make_absolute doesn't fully canonicalize paths:

Path before make_absolute: ..
Path after make_absolute: /home/legare/zlib/build/..

[kkysen]

Does it need to fully canonicalized them or just make them absolute? If we need full canonicalization, can you fix #432 to use the canonicalization function instead? Either LLVM's if it has one or std's if it doesn't like you did here.

[ayrtonm]

Does it need to fully canonicalized them or just make them absolute?

I think it has to be canonical because we compare with paths we get from querying clang for things.

[kkysen]

Does it need to fully canonicalized them or just make them absolute?

I think it has to be canonical because we compare with paths we get from querying clang for things.

Ah, let's canonicalize then.

[randomPoison]

Yeah I'm seeing bad behavior from the rewriter (e.g. trying to copy files to the wrong directory) with the current absolute path logic. Things only seem to work with a fully canonicalized path.

I've updated the PR to use remove_dots after making the path absolute. I couldn't find a direct canonicalize path function, but there's an internal one in LLVM that just does make_absolute followed by remove_dots, and it works when I test with relative paths.