[SYCL] Improve mock PI plugin

[romanovvlad]

The patch adds more logic to default redefinitions of PI APIs:

1. Allocate, refcount and deallocate various handles
2. Handle sub-buffer creation

Also the patch adds support for adding PI functions to be called in addition (before or after) to the original function. This allows intercepting PI API calls for introspection while still allowing original function take care of handles. Or add some post processing of the returned values.

[aelovikov-intel]

I love the idea of what we are doing here, but I'm not sure about how. However, please consider the below just as an alternative idea and not a request to make any changes.

What if do more functional programming style here? Pseudocode:

// Maybe an extra overload in addition to pre-existing simpler functionality.
Mock.redefine([&](std::function OldEntry) -> std::function {
  return [&]() {
    // Code before
    OldEntry();
    // Code after.
   };
});

or

Mock.redefine_wrap([&](auto OldEntry) {
   // Code before
  OldEntry
  // Code after
});

This is likely slower in run time but I think that for testing infrastructure the tradeoff between easier maintainability/speed could be more biased towards the former.

[romanovvlad]

I love the idea of what we are doing here, but I'm not sure about how. However, please consider the below just as an alternative idea and not a request to make any changes.

What if do more functional programming style here? Pseudocode:

// Maybe an extra overload in addition to pre-existing simpler functionality.
Mock.redefine([&](std::function OldEntry) -> std::function {
  return [&]() {
    // Code before
    OldEntry();
    // Code after.
   };
});

or

Mock.redefine_wrap([&](auto OldEntry) {
   // Code before
  OldEntry
  // Code after
});

This is likely slower in run time but I think that for testing infrastructure the tradeoff between easier maintainability/speed could be more biased towards the former.

We could do something like that https://godbolt.org/z/T3oPbvKv8 .
Not sure why it's better though, current tests show that people tend to provide a function pointer rather than write a lambda. Also(not sure we will ever need this), but it's not clear how to undo such a wrapper.

[steffenlarsen]

This is super good! It will make it a lot easier to for example add specific info queries without having to redefine the entire info query function. Few comments.

[steffenlarsen]

I think we would need that in case gtest ever decides to run tests in parallel. It may be enough to make the arrays thread_local. That might have problems when a specific test is multi-threaded though.

[romanovvlad]

As an alternative we could:

1. Create a global map pi_platfrom to array of function pointers map
2. create a global map from every handle(void*) to pi_platform
3. update the map(from 2)) any time we create or remove a handle in the mock functions
4. in the proxy function, for a given handle find a corresponding pi_platform using the map from 2)
5. Using pi_platform find an array of function pointers from 1)

[steffenlarsen]

I think that might be a good solution to the problem. We could even simplify the map (2) a little, defining a space for which a given platform can use, e.g. the first platform can create handles 0-10000, the next can use 10000-20000, etc. Biggest problem is if someone overwrites the behavior of a function returning handles and return something not in the allocated range/in the map.

[steffenlarsen]

Good stuff!

[pvchupin]

@romanovvlad, please see post commit issue: https://github.com/intel/llvm/actions/runs/3322958425/jobs/5492696367

[pvchupin]

@romanovvlad, please see post commit issue: https://github.com/intel/llvm/actions/runs/3322958425/jobs/5492696367

There are actually few separate issues in different configurations. Probably it would be the best to revert at this point and reapply as soon as addressed.