tharre / redo Goto Github PK

Add test cases to ensure redo keeps on working like expected.

While depending on environmental variables is by no means difficult at the moment, it can certainly be cumbersome. An alternative would be to create a new command, redo-env, that would expect environment variables as arguments and hash over their contents. If the hash changes between subsequent redo runs the target would be considered out-of-date.

Example:

redo-env CC

$ redo
redo  all
(build.c:309): redo: failed to get realpath() of /usr/lib/clang/3.7.0/include/stddef.h: No such file or directory
(build.c:140): redo: invoked .do script ./default.do failed: 1

Clang was updated to 3.8.0 and thus the directory of the dependency disappeared. redo should just declare the dependency out-of-date in that case instead of failing.

This won't be trivial, as redo depends on a lot of POSIX features like the shell. But it's not impossible either.

Currently, mkdir() is called thousands of times because we call mkpath() unnecessarily. This is bad for performance.

redo-ifchange foo bar

should build foo and bar in parallel.

Additionally, redo should support the jobserver protocol like apenwarr's implementation does.

There should be test cases checking $3, specifically for subdirectories, as this has been bugged before.

Commit that fixed this bug: 0a9f087

redo should provide some mechanism to catch all output files of certain commands or "broken" compilers (like javac). A fork of apenwarr's redo has already implemented a interesting mechanic to do so (see here for an explanation).

Note that even while this issue remains unfixed you can still manage multiple outputs with redo, it's just not as good as it should be.

redo saves hashes about everything it builds already in .redo/, so we might as well provide a clean command to save people from writing their own (often outdated) clean.do.

• redo shouldn't abort if files in .redo/ have been corrupted if it can be avoided. Corrupted files should just be deleted.
• Error messages should always print out readable (relative to REDO_ROOT) paths.

Using rand() for generating the magic numbers is bad because it only guarantees 2^15 different values.

Possible solutions:

• ship our own random number generator (needless complexity, existing implementations?)
• don't use random numbers but just increment by 1 (additional state)