openblas: the build phase builds and runs the test suite, defeating the run_tests guard (hard failure on macOS x86_64)

[cekees]

@mathomp4 @sethrj I don't think this is a critical issue with openblas/package.py, but I'm raising it for your awareness. Would likely make the openblas build faster and more robust.

The issue is that it appears build runs make -s with no target, so it defaults to all, which includes tests per the Makefile:

all :: libs netlib $(RELA) tests shared 

There's a wordy comment with a small change on PR #6000 that uses explicit build targets, which post-pones tests until spack's optional testing phase.

This was only occuring on an old macox x86_64 machine that successfully built openblas but had a crash during the tests, possibly due to being overloaded.

[mathomp4]

@cekees I mean, I have no objections, though I usually build openblas by CMake (for no real reason now other than I can sort of hit issues with the CMake build). I wonder if that's why I never saw this on my Macs? The CMake build doesn't seem to even have an option to build/run tests...

I'm going to ping @martin-frbg to see if he thinks that "libs", "netlib", "shared" is a good set to build since I don't know the GNU Make targets.

[martin-frbg]

Not in favor of telling people (distributors) to build without testing, but doing just libs-netlib-shared should work. (The alternative is to claim you're cross-compiling by setting CROSS=1, maybe I should make it a bit less of a hack. (ISTR there's a very old issue aaking for this, but it went a bit further with a testing framework independent of make or something like that. Will take another look at that when able)

[martin-frbg]

(Note that CMake comes with its own ctest harness for running tests separately from the build phase, which OpenBLAS builds already use)

[sethrj]

The behavior of the make in openblas and the spack recipe have changed a lot over the years: I changed it to do this in #2936 to avoid linking failures during the test phase, and @haampie changed it back in #4569 to avoid running ar multiple times.

[sethrj]

It's complicated by the fact that versions of openblas have the race condition, some versions have tests that fail on some platforms/toolchains, etc.

[mathomp4]

In re testing:

Not in favor of telling people (distributors) to build without testing, but doing just libs-netlib-shared should work. (The alternative is to claim you're cross-compiling by setting CROSS=1, maybe I should make it a bit less of a hack. (ISTR there's a very old issue aaking for this, but it went a bit further with a testing framework independent of make or something like that. Will take another look at that when able)

It looks like the --test=root functionality is there for GNU Make:

spack-packages/repos/spack_repo/builtin/packages/openblas/package.py

Lines 656 to 659 in d117b0a

	@run_after("build")
	@on_package_attributes(run_tests=True)
	def check_build(self):
	make("tests", *self.make_defs, parallel=False)

(Note that CMake comes with its own ctest harness for running tests separately from the build phase, which OpenBLAS builds already use)

We could add something similar for CMake builds a la:

 @run_after("build") 
 @on_package_attributes(run_tests=True) 
 def check_build(self): 
     ctest_args = ["-j", str(make_jobs)]
     with working_dir(self.build_directory):
         ctest(*ctest_args)