Compilation of #include <sycl/sycl.hpp> is slow and that's especially problematic for SYCL RTC (run-time compilation). One way to overcome this is fine-grained includes that are being pursued separately. Another way is to employ clang's precompiled headers support which this PR is doing. Those two approaches can be combined, and this PR adds test-e2e/PerformanceTests/KernelCompiler/auto-pch.cpp that gives some idea of the PCH impact. The test shows PCH benefits when compiling some of the fine-grained includes on top of absolute minimum required to compiled SYCL RTC's "Hello world". From one of the CI runs:

It misses sycl/sycl.hpp line because that currently crashes FE when reading the generated PCH, the crash is being investigated/fixed separately.

Implementation-wise I'm reusing existing upstream clang::PrecompiledPreamble with one minor modification. It seems that PrecompiledPreamble's main usage is for things like clangd so it ignores errors in the code. I've modified it so that those errors would break pch-generation the same way normal compilation would break. I'm also not sure if we'd want that long-term, because it seems that making such "auto-pch" persistent would deviate from the upstream version of PrecompiledPreamble even more. I can imagine that in some near future we'd need to "fork" it into a separate utility. Still, seems to be fine for the first step.

Driver modifications are for the --auto-pch option support that should only be present on the SYCL RTC path and not for the regular clang invocations from the command line. I'm relatively confident those will stay in future.