[SPIRV] Emit DebugFunction/Definition for both wrapper and real functions

[SteveUrquhart]

#6758 unfortunately caused a regression for shaders compiled without -fcgl. This PR extends the original fix and corrects the regression. We now emit DebugFunction and DebugFunctionDefinition pairs for both the wrapper and the main functions, and a DebugEntrypoint that points at the wrapper function. This survives inlining.

[s-perron]

Lots of tests are failing. Try running the spir-v tests: ninja check-clang-codegenspirv.

[SteveUrquhart]

I'm sorry, I have one test left to fix.

[SteveUrquhart]

@s-perron, the current failure in CI seems to be coming from an optimizer pass. I've attached the spvasm if we use fcgl or O0, and we see there are now 2 DebugFunctions, one for the wrapper and one for src.main. debug_info_manager.cpp:117 in the optimizer asserts.
assert(
fn_id_to_dbg_fn_.find(fn_id) == fn_id_to_dbg_fn_.end() &&
"Register DebugFunction for a function that already has DebugFunction");
I appreciate your advice on this.
spirv.debug.opline.end.of.shader.O0.spvasm.txt
spirv.debug.opline.end.of.shader.fcgl.spvasm.txt

[s-perron]

@SteveUrquhart The DebugFunction instructions do not look correct. Both of them have %src_main for their declaration, which is incorrect for two reasons. First they represent different functions, so they should be different. Second, that value should be the result of a DebugFunctionDeclaration, and not the result of the OpFunction.

   %src_main = OpFunction %void None %15
         %10 = OpExtInst %void %1 DebugFunction %8 %5 %6 3 1 %7 %9 FlagIsProtected|FlagIsPrivate 3 %src_main
         %14 = OpExtInst %void %1 DebugFunction %13 %5 %6 3 1 %7 %9 FlagIsProtected|FlagIsPrivate 3 %src_main

[SteveUrquhart]

@s-perron, I apologize for stepping in it again and not seeing that FunctionId duplication. Regarding the DebugFunctionDeclaration, the test we were discussing is using the OpenCL debug info instead of the NonSemantic debug info. The signature for DebugFunction is different between the two types, and I was messing up the forward reference in the OpenCL info. We don't emit the optional DebugFunctionDeclaration for now, for either debug info type.
I've pushed what I think is correct for both debug info types. There are still a handful of failures in the rich.debig.info tests, but I believe these to all be false failures due to rearranging of debug info. I don't want to be hasty again so I'm looking through those now.

[SteveUrquhart]

Apologies again for the delay in getting to this point.

[SteveUrquhart]

I've found the NonSemantic DebugScope for the entry function is incorrectly referencing the DebugFunction for the wrapper. I will enhance the testing to catch this and push again.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[SteveUrquhart]

I've added testing for NonSemantic DebugScope, and cleaned up the names the debug user sees for the wrapper and the original function. While debugging src_main, the user will want to see "main" from the HLSL.

[s-perron]

This looks good to me, but I'm not setup to test it in renderdoc. @Goshido Does this fix #6939?

[SteveUrquhart]

I've found one more problem with the logic. The new convenience function emitDebugFunction() was the wrong place to pushCurrentLexicalScope(). I have a fix for this and am adding test logic for it now.

[Goshido]

I gonna check PR when you are ready.

[SteveUrquhart]

@Goshido, please check it when you are ready.

[Goshido]

I checked branch fix-ns-debugfunctiondef, commit e1a4719d7b53ef6fddffa3f18422b6526f65a79b. DXC version 1.8.2407.93. It's working as expected:

Awesome! 👍

P.S. Quick question. What last digits of DXC version mean?

            minor   ???
              |      |
              ▼      ▼
DXC version 1.8.2407.93
            ^     ^
            |     |
          major   |
                  |
             year and ???