Cross crate tests on windows are broken

[brson]

The compiletest driver is not setting up the aux paths correctly.

[yichoi]

I've seen it before but I cannot reproduce it from incoming branch

[metajack]

Have either of you seen this since?

[crnobog]

I've seen this today. For me the compiletest driver is building the executable as i686-pc-mingw32\test\run-pass\anon-extern-mod-cross-crate-2.stage2-i686-pc-mingw32.exe and the aux crate in i686-pc-mingw32\test\run-pass\anon-extern-mod-cross-crate-2.libaux

But then the PATH compiletest uses to run the executable has i686-pc-mingw32\test\run-pass\anon-extern-mod-cross-crate-2.stage2-i686-pc-mingw32.libaux in it, so it can't find the appropriate dlls.

I've had a look at runtest.rs but I haven't unravelled how these paths are constructed yet.

[crnobog]

Tracked it down, the issue is that the aux directory is determined during the build as output_base_name(config, testfile).with_filetype("libaux") which replaces the stage2 identifier with .libaux instead of appending it.
The win32 runner then just chops off the ".exe" from the test executable name to get the aux directory.

I think the safer way to fix it is to make the aux directory contain the stage identifier? I can't imagine you'd want different stages under test to share aux crates at any point.

[pnkfelix]

This particular testing infrastructure issue is not a 1.0 blocker. Assigning P-low.

[brson]

This looks to be fixed.