Flang NVidia offload fix (#18965)

* Make release_90 preferred version.

* Be more explicity about CUDA dependencies.

* Remove duplicate CUDA dependency in Flang package and introduce nvptx variant.

* Fix nvptx variant message.

2 files changed, 23 insertions, 8 deletions

diff --git a/var/spack/repos/builtin/packages/flang/package.py b/var/spack/repos/builtin/packages/flang/package.py
index 1f7d1e5da0..aa45185d6c 100644
--- a/var/spack/repos/builtin/packages/flang/package.py
+++ b/var/spack/repos/builtin/packages/flang/package.py
@@ -8,7 +8,7 @@ from spack import *
 import os
 
 
-class Flang(CMakePackage, CudaPackage):
+class Flang(CMakePackage):
     """Flang is a Fortran compiler targeting LLVM."""
 
     homepage = "https://github.com/flang-compiler/flang"
@@ -24,17 +24,27 @@ class Flang(CMakePackage, CudaPackage):
     version('20180921', sha256='f33bd1f054e474f1e8a204bb6f78d42f8f6ecf7a894fdddc3999f7c272350784')
     version('20180612', sha256='6af858bea013548e091371a97726ac784edbd4ff876222575eaae48a3c2920ed')
 
+    # variants
+    msg = ('Target OpenMP offloading to NVIDIA GPUs '
+           '(experimental w/ no support). '
+           'Only works with GCC compilers compatible with CUDA 9.x. '
+           'Specify ^llvm-flang+cuda cuda_arch=<value> explictly.')
+
+    variant('nvptx',
+            default=False,
+            description=msg)
+
     # Patched only relevant for March 2019 release with OpenMP Offload support
     patch('https://github.com/flang-compiler/flang/commit/b342225a64692d2b9c3aff7658a8e4f94a8923eb.diff',
           sha256='3bd2c7453131eaaf11328785a3031fa2298bdd0c02cfd5e2b478e6e847d5da43',
-          when='@20190329 +cuda')
+          when='@20190329 +nvptx')
 
     # Build dependency
     depends_on('cmake@3.8:', type='build')
     depends_on('python@2.7:', type='build')
 
-    depends_on('llvm-flang@release_90', when='@master')
-    depends_on('llvm-flang@20190329', when='@20190329')
+    depends_on('llvm-flang@release_90', when='@master ~nvptx')
+    depends_on('llvm-flang@20190329', when='@20190329 ~nvptx')
     depends_on('llvm-flang@20181226_70', when='@20181226')
     depends_on('llvm-flang@20180921', when='@20180921')
     depends_on('llvm-flang@20180612', when='@20180612')
@@ -45,10 +55,11 @@ class Flang(CMakePackage, CudaPackage):
     depends_on('pgmath@20180921', when='@20180921')
     depends_on('pgmath@20180612', when='@20180612')
 
-    depends_on('llvm-flang +cuda', when='+cuda')
+    depends_on('llvm-flang+cuda@release_70', when='@master +nvptx')
+    depends_on('llvm-flang+cuda@20190329', when='@20190329 +nvptx')
 
     # conflicts
-    conflicts('+cuda', when='@:20181226',
+    conflicts('+nvptx', when='@:20181226',
               msg='OpenMP offload to NVidia GPUs available 20190329 or later')
 
     # Spurious problems running in parallel the Makefile
diff --git a/var/spack/repos/builtin/packages/llvm-flang/package.py b/var/spack/repos/builtin/packages/llvm-flang/package.py
index 04ab53d219..4b9358ce06 100644
--- a/var/spack/repos/builtin/packages/llvm-flang/package.py
+++ b/var/spack/repos/builtin/packages/llvm-flang/package.py
@@ -16,7 +16,7 @@ class LlvmFlang(CMakePackage, CudaPackage):
 
     maintainer = ['naromero77']
 
-    version('release_90', branch='release_90')
+    version('release_90', branch='release_90', preferred=True)
     version('release_80', branch='release_80')
     version('release_70', branch='release_70')
     version('release_60', branch='release_60')
@@ -46,7 +46,11 @@ class LlvmFlang(CMakePackage, CudaPackage):
     # libomptarget dependencies
     depends_on('libelf', when='+cuda')
     depends_on('libffi', when='+cuda')
-    depends_on('cuda@:9', when='+cuda')  # llvm 7 not compatible with newer version of cuda
+    # note that there may be other llvm-flang + cuda conflicts, but these
+    # are the only versions that are relevant for the flang+nvptx package
+    # llvm 7 not compatible with newer version of cuda
+    depends_on('cuda@:9', when='+cuda@release_70')
+    depends_on('cuda@:9', when='+cuda@20190329')
 
     # LLVM-Flang Componentes: Driver, OpenMP
     resource(name='flang-driver',