[amd-spirv-backend][Debug Info] Add additional support debug information in SPIRV backend

[lamb-j]

We have some support for debug info in the SPIRV translator, less in the SPIRV backend. We should:

• Reach parity with the SPIRV translator for debug info support
• Reach parity with the CPU for debug info support
• Both CPU debug information, and GPU debug information (heterogeneous-dwarf)

This will involve:

• Assessing difference between translator and backend debug info support
• Developing a static description of the source language
• Locations of variables

Example commands:
clang file.hip -gheterogeneous-debug --offload-arch=amdgcnspirv --use-spirv-backend --device-only
clang file.hip -gheterogeneous-debug --offload-arch=amdgcnspirv --no-use-spirv-backend --device-only

clang file.hip -target-triple=spir64 -o file.spv
clang file.hip -emit-llvm -o file.bc
llvm-spirv file.spv

[MrSidims]

Few insights from me.

The translator supports 3 DebugInfo instructions sets:

1. OpenCL.DebugInfo.100 - a legacy set, was good for it's time, but has several issues, like lacking support for full spectrum of DWARF operations that are used in debug expressions. Also it wasn't built considering non-semanticness of DebugInfo.
2. NonSemantic.Shader.DebugInfo.100 - it's build on as a replacement of OpenCL.DebugInfo.100. Now it's explicitly relies on SPV_KHR_non_semantic_info with few consequences for SPIR-V generation. There are other changes, but dependency on nonsemantic instruction set is the main change. It's still lacking support for all spectrum of DWARF operations.
3. NonSemantic.Shader.DebugInfo.200 is built on top of NonSemantic.Shader.DebugInfo.100. Now it brings support for all DWARF operations improving debuggability of C/C++ kernels and also it adds few new instructions to assist debugging of Fortran code offloaded to GPU through SPIR-V. It's not yet ratified by KhronosGroup (yeah, my technical debt as former Intel engineer), but it's fully implemented in the translator.

Outstanding issues for all 3 debug informations:
(there should be a good presentation from Diego Novillo, but here are my few cents):

• Currently DebugInfo specs are too strict about placing DebugInfoNone instructions, limiting number of cases when it's allowed. In reality this should be relaxed at least for a case, when a variable was mem2reg'ed and therefore Debug instruction can not refer it (right now we should be dropping debug info per spec, but actually debugging tools could have a good use of DebugDeclare.
• There is a lack of support for brand new instructions/types/concepts like cooperative matrices and vectors.

But overall, the translator can preserve debug information during LLVM IR -> SPIR-V -> LLVM IR quite nicely, especially with .200 instruction set enabled.

Future:

• Currently NonSemantic.Shader.DebugInfo.100 is being implemented in SPIR-V backend by contractors hired by Khronos Group (though I'm not following their work).
• NonSemantic.Shader.DebugInfo.200 should also be ratified and implemented in the backend. Though, not sure who should champion it as I've switched to team Red :)
• There (probably) will be yet another DebugInfo spec with rolling version, see https://gitlab.khronos.org/spirv/SPIR-V/-/issues/891 .

[mgcarrasco]

@MrSidims thanks!

Regarding NonSemantic.Shader.DebugInfo.100 in the BE, this might be the PR. I'm getting more familiar with it and also with debug info in general.

[MrSidims]

Frankly, I got lost track of incoming PRs, as I've originally reviewed llvm#165302 , which is now close without merge :)

[mgcarrasco]

It was merged in a different PR, but it had to be reverted due to some build errors. llvm#169643

[mgcarrasco]

It's still lacking support for all spectrum of DWARF operations

@MrSidims By any chance, would you know some examples of these limitations (NonSemantic.Shader.DebugInfo.100)?

[MrSidims]

By any chance, would you know some examples of these limitations (NonSemantic.Shader.DebugInfo.100)?

Sure, DWARF supports >100 of operations plus LLVM IR has few on top of them. Meanwhile OpenCL.DebugInfo.100 and therefore NonSemantic.Shader.DebugInfo.100 has only 10, see https://github.com/KhronosGroup/SPIRV-Registry/blob/main/nonsemantic/NonSemantic.Shader.DebugInfo.100.asciidoc#debug-operations . More operations are covered by .200 spec, see https://github.com/KhronosGroup/SPIRV-Registry/blob/d7f0ac3c7c2bda4d35117b0c5f8e58972ee1247c/nonsemantic/NonSemantic.Shader.DebugInfo.200.asciidoc#debug-operations (see more possible arithmetic operations on an address, logical operations etc).