[SYCL] Initial printf support for non-constant AS format strings

[AGindinson]

Allow generic address space strings into experimental::printf and, consequently,
into __spirv_ocl_printf declarations. To mitigate the lack of device BE support for
generic address space literals, we implement a pass to move the literal argument
into constant address space whenever possible.

This is a temporary solution; long-term, we should replace this with proper support
of generic address-spaced format strings at the level of SPIR-V translation and device
backends.

Tested in intel/llvm-test-suite#569.

Signed-off-by: Artem Gindinson artem.gindinson@intel.com

[AlexeySachkov]

Overall the pass looks almost good to me, waiting for tests

[AGindinson]

/verify with intel/llvm-test-suite#569

[AGindinson]

/verify with intel/llvm-test-suite#569

[AGindinson]

@AlexeySachkov @mlychkov @v-klochkov I'm thinking about moving the pass from SYCLLowerIR into Transforms/IPO to minimize code-ownership conflicts between DPC++ Tools and DPC++ ESIMD. I don't know if this would reflect the nature of the SYCL printf address space mutation as well, but since the pass is intended as temporary, short-term maintainability seems like a bigger focus. What are your thoughts on this?

[AlexeySachkov]

@AlexeySachkov @mlychkov @v-klochkov I'm thinking about moving the pass from SYCLLowerIR into Transforms/IPO to minimize code-ownership conflicts between DPC++ Tools and DPC++ ESIMD. I don't know if this would reflect the nature of the SYCL printf address space mutation as well, but since the pass is intended as temporary, short-term maintainability seems like a bigger focus. What are your thoughts on this?

I would actually prefer if this pass resides in SYCLLowerIR, because it is easier for us to distinguish it from upstream passes. I'm less concerned about mixing regular DPC++ and DPC++ ESIMD passes than about mixing our customizations with upstream code

[AGindinson]

/verify with intel/llvm-test-suite#569

[AGindinson]

@intel/llvm-reviewers-runtime, @v-klochkov, could you please approve changes to printf signatures in sycl/include/?
@kbobrovs, @kychendev, @sndmitriev, @v-klochkov, could someone please approve the addition of the pass files under llvm/**/SYCLLowerIR directories?
@elizabethandrews, @premanandrao, @smanna12, could someone please approve changes to Clang's pass scheduling in clang/lib/CodeGen/BackendUtil.cpp?
@bader, could you please approve the pass registration under llvm/?

(should #5016 come in earlier, I'll rebase the patch and rely on team-based code-owner approval)

[elizabethandrews]

FE change (add a pass) looks good to me.

[AGindinson]

For the joint Precommit of product + test patches, CUDA/Win failures in DeviceLib tests are known to be unrelated.