[SYCL][Driver] Enable PCH inclusion while performing host compilation with -fsycl

[srividya-sundaram]

This PR allows users to include a Pre-Compiled header (PCH) file while performing the host compilation step with -fsycl.

// Create PCH
clang++ -x c-header test.h

// Use PCH

clang++ -c -fsycl -include test.h test.cpp
clang++ -c -fsycl -include-pch test.h.gch test.cpp

[mdtoguchi]

Any thoughts on consideration for allowing the use of PCH files with -fsycl, but restricting to only the host side?

[bader]

Any thoughts on consideration for allowing the use of PCH files with -fsycl, but restricting to only the host side?

What problems do we need to solve to enable PCH for device compilers?

@romanovvlad, FYI.

[mdtoguchi]

Any thoughts on consideration for allowing the use of PCH files with -fsycl, but restricting to only the host side?

What problems do we need to solve to enable PCH for device compilers?

@romanovvlad, FYI.

If the generation of the PCH files is OK for device, the issue is consumption. We could take it in a step in the direction of 'fat PCH' files. If we encounter a fat PCH input file, we would recognize the targets, unbundle and pass along the unbundled PCH files to the respective compilations (host/device). For any device targets that are not found, we would ignore for those targets. Pass along the host variant to the host compile.

[romanovvlad]

Any thoughts on consideration for allowing the use of PCH files with -fsycl, but restricting to only the host side?

What problems do we need to solve to enable PCH for device compilers?
@romanovvlad, FYI.

If the generation of the PCH files is OK for device, the issue is consumption. We could take it in a step in the direction of 'fat PCH' files. If we encounter a fat PCH input file, we would recognize the targets, unbundle and pass along the unbundled PCH files to the respective compilations (host/device). For any device targets that are not found, we would ignore for those targets. Pass along the host variant to the host compile.

I noticed that there is some problem with deserialization of SYCL builtins , see fix here: https://github.com/romanovvlad/llvm/pull/7/files#diff-e2fdcbb3b7d1e3d5aa680d864906f9c49605080caa556ca2500c72b9c7a182c4R1815
This is not production quality fix though.

[bader]

Any thoughts on consideration for allowing the use of PCH files with -fsycl, but restricting to only the host side?

What problems do we need to solve to enable PCH for device compilers?
@romanovvlad, FYI.

If the generation of the PCH files is OK for device, the issue is consumption. We could take it in a step in the direction of 'fat PCH' files. If we encounter a fat PCH input file, we would recognize the targets, unbundle and pass along the unbundled PCH files to the respective compilations (host/device). For any device targets that are not found, we would ignore for those targets. Pass along the host variant to the host compile.

I noticed that there is some problem with deserialization of SYCL builtins , see fix here: https://github.com/romanovvlad/llvm/pull/7/files#diff-e2fdcbb3b7d1e3d5aa680d864906f9c49605080caa556ca2500c72b9c7a182c4R1815 This is not production quality fix though.

This won't resolve the problem Mike mentioned, but I think we need to make this fix anyway. @elizabethandrews, @premanandrao, FYI.

[srividya-sundaram]

@bader Can you please merge the changes in this PR? Thanks!

[bader]

@bader Can you please merge the changes in this PR? Thanks!

Yes, but it's better to ping @intel/llvm-gatekeepers team as described in the contribution guide.