[SYCL][Build] Add --enable-all-llvm-targets switch to configure.py

[pvchupin]

Can be used to enable all supported llvm compiler targets in the compiler build,
currently:

• NVPTX
• AMDGPU

Compiler backends do not require any 3rd party software to build and
test, this can be useful in some local testing, for example when
bringing changes from upstream

Also minor fixes to docs based on currently supported switches.

This is NFC for all existing processes.

[bader]

The change looks good to me, but I think libclc impacts compilation as well. Is it okay to skip libclc build?

[pvchupin]

The change looks good to me, but I think libclc impacts compilation as well. Is it okay to skip libclc build?

As far as I remember libclc has some dependency on 3rd party low-level runtimes, so it stays out of scope of this option, and included only when specific target is fully enabled in build (through --hip or --cuda switches).

[pvchupin]

I was wrong. It doesn't have any dependency it seems. At least build and testing is successful.
There is currently 1 LIT fail on AMDGPU target, libclc presence/absence in the build is not affecting that.

I also realized that after adding libclc there is going to be significant (but not a complete) overlap with --ci-defaults option, which seems actually is trying to accomplish the same idea, build everything without deps. The remaining differences in --ci-defaults:

• warnings by default
• compiler-rt and clang-tools-extra presence
• no AMDGPU llvm target, that's a gap I think. I'll submit a patch to address that separately... but LIT issue above should be fixed first.

On the other hand, and actually to answer @bader question, it seems libclc not required for NVPTX and AMDGPU (build and testing). So probably we should limit this change/switch to only that, to keep things limited to compiler targets (as new option says).

Tagging @AerialMantis, @npmiller for opinion

[bader]

On the other hand, and actually to answer @bader question, it seems libclc not required for NVPTX and AMDGPU (build and testing). So probably we should limit this change/switch to only that, to keep things limited to compiler targets (as new option says).

Building libclc is not required, but if a test program calls built-in functions implemented in libclc, having pre-built libclc improves compiler testing coverage. So, it really depends on what is our goal here. I assume it's to test as much as possible w/o external SW and HW dependencies. Am I right?

[pvchupin]

I assume it's to test as much as possible w/o external SW and HW dependencies. Am I right?

@bader, I started with this as a primary goal, then realized that --ci-defaults is sort of doing the same thing already.
Now I'm thinking this patch as is makes more sense, as it is only enables all compiler targets and corresponding LLVM BE LIT tests.
LLVM BE LIT tests don't require libclc, right?

If no other suggestions I'm leaning towards landing it in current state.

[npmiller]

I'm not sure how useful it will be in practice, as compiler changes should probably be tested against SYCL lit tests as well, but if you think there's a use case for it, I'm happy with the flag as it is as it's clear it only enables the compiler targets.