[SYCL] Implement coverage instrumentation for device code

[0x12CC]

This PR extends Clang's source-based code coverage to work with SYCL device code. It includes the following changes:

1. The InstrProfilingLoweringPass was updated to lower profiling counters to SYCL device globals.
2. A new function was added to the compiler runtime to increment the host-side profiling counters.
3. The SYCL runtime was updated to send the contents of the device-side profiling counters to the new compiler runtime function when their device global map entries are being freed.

This feature may not work correctly for functions that differ between host and device due to the use of the __SYCL_DEVICE_ONLY__ macro. In such cases, it may not be possible to correlate the profiling counters from the device to the host. Resolves #7803.

[aelovikov-intel]

Do other projects enable profiling unconditionally in the build too?

[aelovikov-intel]

This is the only remaining question from me, the rest is 🔥 👍

[aelovikov-intel]

Would it make sense to add a high-level design document here as well (unless one exists already). Also, what about kernels using optional features? Are we merging info collected by different device images somehow? That needs to be explained either in some doc or at the PR title.

[0x12CC]

Would it make sense to add a high-level design document here as well (unless one exists already).

I think this makes sense. I'll add a document to sycl/doc/design.

Also, what about kernels using optional features?

I'm not sure I understand the question. If a kernel uses some features that prevent it from being submitted to a device, I would expect the corresponding profiling counters to contain zero values and not change the resulting coverage report.

Are we merging info collected by different device images somehow? That needs to be explained either in some doc or at the PR title.

The profiling counters collected from any device image are added to the host's profiling counters. If there are multiple device images that share device code, each one will increase the same profiling counters on the host. The assumption is that the host and device compilers instrument the same functions so it's possible to copy the counter values from the device to the host. The resulting profile output contains the values of the host profiling counters at the end of the program's execution. I'll make this clear in the design document.

[srividya-sundaram]

Driver changes LGTM

[maarquitos14]

LGTM, but ping me when the design doc is ready for another round :)

[0x12CC]

Thanks for the reviews. I've added sycl/doc/design/DeviceCodeCoverage.md in 8201ffc.

[bader]

How is the counter's size determined?
Can't we use one size (e.g. 8-byte integer) for all counters?

[0x12CC]

The counter variable is actually an array of integers, each of which is eight bytes. The number of elements in this array is equal to the number of regions in the function being instrumented. For the kernel in the coverage test case, there are two regions since there are two code branches. The resulting device global variable will have a size of (2 * sizeof(std::uint64_t)) bytes. I'll update the document to try to make this more clear.