sysdetect: fix bug where Makefile sets -ffree-form everywhere

[tgamblin]

To fix ARM compiler builds, #108 added these lines to the Makefile for sysdetect tests:

+ FFLAGS += -ffree-form
+
  ifeq ($(notdir $(F77)),gfortran)
      FFLAGS +=-ffree-form -ffree-line-length-none
  else ifeq ($(notdir $(F77)),flang)
      FFLAGS +=-ffree-form
  else ifneq ($(findstring $(notdir $(F77)), $(intel_compilers)),)
      FFLAGS +=-free
  else ifneq ($(findstring $(notdir $(F77)), $(cray_compilers)),)
      FFLAGS +=-ffree
  endif

but this adds -ffree-form to every compiler. This breaks cray compiler builds for PAPI 7.1.0:

     1307    /home/gitlab-runner-protected-3/builds/Cj9btwqn/0/spack/spack/lib/
             spack/env/cce/ftn -ffree-form -ffree -I../../.. -o query_device_si
             mple_f query_device_simple_f.F ../../../libpapi.a
     1308    ftn-78 ftn: ERROR in command line
     1309      The -f option has an invalid argument, "free-form".

You can see that CCE gets both -ffree-form and -ffree there.

This patch removes the global -ffree-form and makes the flang condition above a bit more permissive, so that both flang, armflang, amdflang, or really anything ending in flang adds -free-form:

       FFLAGS +=-ffree-form -ffree-line-length-none
-  else ifeq ($(notdir $(F77)),flang)
+  else ifeq ($(patsubst %flang,,$(notdir $(F77))),)  # compiler name ends with flang
       FFLAGS +=-ffree-form

This should fix the bug that #108 fixed and it should also work on other compilers.

Pull Request Description

Author Checklist

• Description
  Why this PR exists. Reference all relevant information, including background, issues, test failures, etc
• Commits
  Commits are self contained and only do one thing
  Commits have a header of the form: module: short description
  Commits have a body (whenever relevant) containing a detailed description of the addressed problem and its solution
• Tests
  The PR needs to pass all the tests

[tgamblin]

The pipeline on spack/spack#41886, confirms this fixes the build for CCE (successful build is here: https://gitlab.spack.io/spack/spack/-/jobs/9685218).

We don't currently test PAPI 7.1.0 with ARM compilers, so maybe @gcongiu can verify whether this fixes the issue he tried to solve in #108 -- I think it should.

[tgamblin]

I'm not sure what to do with the CI failure here. papi_spack is failing because this patch is already in Spack, and the patch fails to apply :). We don't currently have a good way to notice that. We've tried before:

• patches: make re-applied patches idempotent spack/spack#26784

But unfortunately making patch application idempotent can cause us to silently fail to apply patches (unix patch does not give you a way to differentiate between the two cases) so we had to revert:

• Revert "patches: make re-applied patches idempotent" spack/spack#27625

Any suggestions?

[gcongiu]

@tgamblin I would merge this PR anyway as the failure is not related to the patch (unless @jagode prefers otherwise). Maybe we could avoid this problem from happening again in the future by creating and testing PRs for the new PAPI versions in spack before releasing them? I am open to suggestions

[tgamblin]

Actually, looking at your CI, think I can fix this on the spack side if I just restrict the patch to 7.1.0. You’re building papi@master in CI, and this patch will be in 7.1.1, right?

[gcongiu]

@tgamblin it will definitely be in the next release, but that will not necessarily be versioned as 7.1.1.

[tgamblin]

I should ask better questions. If I restrict the patch to 7.1.0, does that cover all the PAPIs we might see with the issue?

[gcongiu]

@tgamblin if you restrict the patch to papi@7.1.0 I think that should do it. The sysdetect component was introduced with papi@7.0.0 and papi@7.0.1 was released soon after to fix some gcc related link time optimisation problem with more recent gcc versions (IIRC). However, none of the aforementioned versions ever worked with spack.

[tgamblin]

Ok, spack/spack#41892 restricts the patch to 7.1.0 so CI should pass.

[G-Ragghianti]

I've triggered a rerun of the papi-spack CI test.

[G-Ragghianti]

It is fixed! A holiday miracle!