[feat] Improve access to the namespaces of both parameter and variable built-ins

[jjotero]

Rather than customising the test attribute access using the __getattribute__ method, this PR moves this customisation into the __getattr__ method instead, which is called when the default __getattribute__ raises an AttributeError.

This PR also makes the variable and parameter spaces behave as proper namespaces, which makes operations such as testing if a variable has been declared and so on more intuitive:

my_var_name in self._rfm_var_space

instead of

my_var_name in self._rfm_var_space.vars

However, adding new items into the variable or parameter spaces is disallowed:

class Foo(rfm.RegressionTest):
    pass

Foo.param_space['P0'] = (1,2,) # This raises an error
Foo._rfm_var_space['V0'] = 'a' # and this too

The corresponding unit tests have been added.

Closes #1841.

[pep8speaks]

Hello @jjotero, Thank you for updating!

Cheers! There are no PEP8 issues in this Pull Request!Do see the ReFrame Coding Style Guide

Comment last updated at 2021-03-16 20:26:19 UTC

[codecov-io]

Codecov Report

Merging #1819 (e95be91) into master (a3d0b0c) will decrease coverage by 1.52%.
The diff coverage is 43.41%.

@@            Coverage Diff             @@
##           master    #1819      +/-   ##
==========================================
- Coverage   87.61%   86.08%   -1.53%     
==========================================
  Files          49       49              
  Lines        8136     8378     +242     
==========================================
+ Hits         7128     7212      +84     
- Misses       1008     1166     +158     

Impacted Files	Coverage Δ
reframe/core/variables.py	51.21% <38.37%> (-45.26%)	⬇️
reframe/core/meta.py	100.00% <100.00%> (ø)
reframe/core/namespaces.py	93.33% <100.00%> (+0.22%)	⬆️
reframe/core/parameters.py	98.50% <100.00%> (+0.04%)	⬆️
reframe/core/pipeline.py	91.86% <100.00%> (-0.05%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a3d0b0c...e95be91. Read the comment docs.

[jjotero]

Variables now behave like regular python class attributes and multiple modifications within the same class body are now allowed. For example:

class Foo:
    pass

class MyTest(rfm.RegressionTest):
    v = variable(Foo, value=Foo())
    v.var_attribute = 'my attribute'

    # Override a regular attribute with a variable
    a = 4
    a = variable(int, value=40)
    print(a+1) # Prints 41

    # Modify an existing var
    a *= 10 # variable a is now 400

    # Initialize a variable as a function of another variable
    b = variable(float, value=a/3.141592)
    

    def __init__(self):
        print(self.v.var_attribute) # This prints 'my attribute'

• Parameters and variables may override regular class attributes
• A parameter may only be defined once per class body and it's value may not be altered after it was declared.
• A variable's value may be modified multiple times within the same class body.
• Overriding a variable with a parameter or the other way around results in undefined behaviour.

[vkarak]

Also coverage is a bit unhappy with the fact that none of the special functions are tested. I would add a small test test exercising all operators.

[vkarak]

One problem that I see, which would help a lot in eliminating the need for a defining a constructor is that currently you can't use in an expression a built-in that was defined in a base class. Look at this?

class Base(rfm.RegressionTest):
    foo = variable(str, value='hello')


@rfm.simple_test
class HelloTest2(Base):
    descr = 'C Hello World test ' + foo
    valid_systems = ['*']
    valid_prog_environs = ['*']
    sourcepath = 'hello.c'
    tags = {'foo', 'bar'}
    sanity_patterns = sn.assert_found(r'Hello, World\!', 'foo.txt')
    maintainers = ['VK']

This will complain about foo in the HelloTest2 definition. I am wondering whether this ca be fixed or not?

[jjotero]

I've added a few more unit tests covering the variable operators and the other issues mentioned in the PR review.

[jjotero]

One problem that I see, which would help a lot in eliminating the need for a defining a constructor is that currently you can't use in an expression a built-in that was defined in a base class.

This is standard Python behaviour, isn't it? I have a few ideas to extend this behaviour. Let me do some "exploring" :)

[vkarak]

This is standard Python behaviour, isn't it? I have a few ideas to extend this behaviour. Let me do some "exploring" :)

Yes, looks like that. It seems that a NameError is thrown before even looking it up in the MetaNamespace. The only reason I want this is to be able to define sanity_patterns in the class body. If you give a filename and not stdout, it just works with the current machinery. Technically, also, if you need something from the test instance, like the self.stdout, you should use a pipeline hook, but sanity patterns with stdout are so common and it would be nice to be able to define them in the class body. I am experimenting with a solution just for this. I wouldn't go too far to avoid this NameError (there is enough magic in the metaclasses already), I'm fine with that, as soon as I can put the damn sanity_patterns inside the class body.

[vkarak]

Lgtm, I have only a couple of minor style comments.

[vkarak]

lgtm