openblas: add "locking" variant, updated blaspp (#21770)

BlasPP by ECP SLATE will fail to install by default
(`spack install blaspp`) because:
- the default BLAS installation in Spack is OpenBLAS
- BlasPP conflicts with `threads=none` for all recent OpenBLAS releases

OpenBLAS introduced a threadsafe compile option
with 0.3.7+ aka `USE_LOCKING`:
```
   61 # If you want to build a single-threaded OpenBLAS, but expect to call this
   62 # from several concurrent threads in some other program, comment this in for
   63 # thread safety. (This is done automatically for USE_THREAD=1 , and should not
   64 # be necessary when USE_OPENMP=1)
   65 # USE_LOCKING = 1
```

According to tests, with `spack install --test root blaspp`,
this exactly addresses the issues in BlasPP tests.

It also seems to be a good option to set by default for OpenBLAS and
users that do not need this safety net can always disable it.

Solve issues with newer OpenBLAS by requiring
`+locking` over none-default threading options.

var/spack/repos/builtin/packages/blaspp/package.py

@@ -29,8 +29,13 @@ class Blaspp(CMakePackage, CudaPackage):
     depends_on('cmake@3.15.0:', type='build')
     depends_on('blas')
 
+    # only supported with clingo solver: virtual dependency preferences
+    # depends_on('openblas threads=openmp', when='+openmp ^openblas')
+
     # BLASpp tests will fail when using openblas > 0.3.5 without multithreading support
-    conflicts('^openblas@0.3.6: threads=none', msg='BLASpp requires openblas multithreading support')
+    # locking is only supported in openblas 3.7+
+    conflicts('^openblas@0.3.6 threads=none', msg='BLASpp requires a threadsafe openblas')
+    conflicts('^openblas@0.3.7: ~locking', msg='BLASpp requires a threadsafe openblas')
 
     def cmake_args(self):
         spec = self.spec

var/spack/repos/builtin/packages/openblas/package.py

@@ -44,6 +44,7 @@ class Openblas(MakefilePackage):
     variant('shared', default=True, description='Build shared libraries')
     variant('consistent_fpcsr', default=False, description='Synchronize FP CSR between threads (x86/x86_64 only)')
 
+    variant('locking', default=True, description='Build with thread safety')
     variant(
         'threads', default='none',
         description='Multithreading support',
@@ -126,6 +127,7 @@ class Openblas(MakefilePackage):
     conflicts('+consistent_fpcsr', when='threads=none',
               msg='FPCSR consistency only applies to multithreading')
 
+    conflicts('threads=pthreads', when='~locking', msg='Pthread support requires +locking')
     conflicts('threads=openmp', when='%apple-clang', msg="Apple's clang does not support OpenMP")
     conflicts('threads=openmp @:0.2.19', when='%clang', msg='OpenBLAS @:0.2.19 does not support OpenMP with clang!')
 
@@ -255,6 +257,13 @@ def make_defs(self):
         if self.spec.satisfies('@0.2.16'):
             make_defs += ['BUILD_LAPACK_DEPRECATED=1']
 
+        # serial, but still thread-safe version
+        if self.spec.satisfies('@0.3.7:'):
+            if '+locking' in self.spec:
+                make_defs += ['USE_LOCKING=1']
+            else:
+                make_defs += ['USE_LOCKING=0']
+
         # Add support for multithreading
         if self.spec.satisfies('threads=openmp'):
             make_defs += ['USE_OPENMP=1', 'USE_THREAD=1']