Don't cache the whole base directory; instead independently cache:
- the build output directory
- the ccache directory
This allow completely avoiding the need for building base if a
cache (keyed on the relevant parts of the git tree for base) is
successfully restored, as well as reducing the amount of cached
data.
Additionally, ensure caches are isolated by branch name (but allow
falling back to a "master" cache for restore): we don't want PRs
using each others' caches.
It should be a path to a directory, not a file.
https://circleci.com/docs/2.0/configuration-reference/#store_test_results
> Path (absolute, or relative to your working_directory) to directory containing subdirectories of JUnit XML or Cucumber JSON test metadata files
Perhaps the file needs to be named results.xml. We'll see.
The difference for builds from scratch is negligible, but for rebuilds (i.e., for us hackers & the CI here in front) it should bring a speed improvement.
Automatic fallback to Make when Ninja is not available, or override with `USE_MAKE=1`.
You can (theoretically) choose your own generator with a sensible combination of `CMAKE_FLAGS` and `CMAKE_MAKE_PROGRAM`.
I also added `MAKE_FLAGS` and `NINJA_FLAGS`. That way you can easily pass `MAKE_FLAGS=-n` or `NINJA_FLAGS=-n` for a dry run, for example.
To switch you might have to run something like `make dist-clean USE_MAKE=1`.
Cf. https://github.com/koreader/koreader-base/pull/861, https://github.com/koreader/koreader-base/pull/862 and https://github.com/koreader/virdevenv/pull/34/files
* junit test results; unfortunately this seems to conflict with the verbose out
* fix deps cache: two files can change independently
* verbose print obsoleted by gtest in upsream busted
The debug output from Coveralls was rubbish. `"Build processing error"`
Besides some basics like checking if we were actually sending
valid JSON that gives us absolutely nothing to work with.
* Enable parallelism in .circleci/config.yml
* Add BUSTED_SPEC_FILE to Makefile testfront
* Use it in .ci/test.sh with xargs for test parallelization
NB This is the dumb method of improving test time.
Ideally we want a workflow fan-in/fan-out approach.
* shellcheck 0.4.5 fix `LC_ALL: en_US.UTF8` (can be removed for shellcheck 0.4.6)
* hush updated luacheck on `reader.lua`; we're not assigning any variables but `= nil` is redundant
