[SPIR-V] Compilation failure with SPV_KHR_non_semantic_info

[DragonJoker]

Description
At work, we have a shader that compiles properly when SPV_KHR_non_semantic_info is not used, but fails to compile as soon as this extension is used.

Steps to Reproduce
The shader doesn't compile on Compiler Explorer, but I couldn't make sure the issue is the same.
Hence using the HLSL code from the Compiler Explorer link, with the provided command line options, the shader doesn't compile.

Actual Behavior
When compiling in CLI, using this command :
dxc.exe .\Shader.hlsl -E DigitBinningPass -T cs_6_0 -spirv -fspv-target-env="vulkan1.1" -Wall -Wno-unused-const-variable -D imBINDLESS=1 -fspv-extension=SPV_EXT_descriptor_indexing -fspv-extension=SPV_KHR_non_semantic_info -fspv-debug=vulkan-with-source
produces the following output :
`fatal error: generated SPIR-V is invalid: ID '2535[%2535]' has not been defined
%2534 = OpExtInst %void %2 DebugValue %805 %2495 %2194 %int_0 %2535 %int_0

note: please file a bug report on https://github.com/Microsoft/DirectXShaderCompiler/issues with source code if possible`

Environment

• DXC version 1.8.0.5082
• Host Operating System Windows 11 25H2

[s-perron]

This may have been fixed by KhronosGroup/SPIRV-Tools#6510. Try the latest dxc if possible.

[DragonJoker]

I've just tested with a local build of commit ff53861 on main branch, and the issue is still here:
fatal error: generated SPIR-V is invalid: ID '2536[%2536]' has not been defined %2535 = OpExtInst %void %2 DebugValue %805 %2496 %2192 %int_0 %2536 %int_0
Different ids, though.

[s-perron]

@SteveUrquhart Can you look into this? I can get compiler explorer to output the spir-v (https://godbolt.org/z/fnaodxaq3). The validator will fail because some ids are missing. However, the spir-v before optimizations are run are valid (https://godbolt.org/z/G618rn3eb).

The spirv-val command to use is: spirv-val b.spv --target-env vulkan1.3 --before-hlsl-legalization.

I'm guessing ADCE in spirv-opt again.

[SteveUrquhart]

@s-perron, You can assign this to me.

[SteveUrquhart]

There's also an error in compilation. Before invoking the optimizer, we see:

fatal error: generated SPIR-V is invalid: ID '461[%461]' has not been defined
%460 = OpExtInst %void %2 DebugTypeFunction %uint_3 %461

You can see this in the -fcgl Compiler Explorer just by dropping the -Vd. And, --before-hlsl-legalization doesn't seem to help this. Eventually, the operands get turned into ForwardReferences by fix-opextinst-opcodes, after which the other error message appears:

error: line 2539: ID '2487[%2487]' has not been defined
%2486 = OpExtInst %void %2 DebugValue %1400 %2434 %2415 %int_0 %2487 %int_0

I'm considering whether fix-opextinst-opcodes is missing DebugValue's?

Finally, I'm not sure why the indices reported by spirv-val are not matching what I see in the spvasm. For instance, the actual failing line of spvasm seems to be:

%31746 = OpExtInst %void %2 DebugValue %927 %31735 %31744 %int_0 %36712 %int_0

[SteveUrquhart]

I've tracked this down to a DebugValue placement problem in the scalar-replacement=0 pass. Here, we output
%31746 = OpExtInst %void %2 DebugValue %927 %31735 %31744 %int_0 %13721 %int_0
It's clearly being placed in the wrong place. It needs to go after:
%13721 = OpLoad %uint %13699

[SteveUrquhart]

I've naively modified /insert_before=/dbg_decl at https://github.com/KhronosGroup/SPIRV-Tools/blob/main/source/opt/scalar_replacement_pass.cpp#L189 and that does seem to work, However, a new problem is then uncovered where a DebugValue has an incorrect index:

NonSemantic.Shader.DebugInfo.100` DebugValue: expected operand Indexes is OpConstant, DebugGlobalVariable, or type is OpConstant with an integer scalar type
  %2894 = OpExtInst %void %2 DebugValue %804 %2495 %2191 %int_0 %2880 %int_0

I've attached
explorer.10.spvasm.txt which is the input to scalar-replacement=0.

I'll be out of the office next week, and the following week I'll be at the Shading Languages Symposium. It would be great to meet everyone.

[DragonJoker]

Hi, I have encountered another issue, still related to SPV_KHR_non_semantic_info.
The error is not the same, but I have found a workaround.
The failing shader is here: https://godbolt.org/z/51K9xK5jz
And the version that compiles is here: https://godbolt.org/z/Gd59Ejsbr
The only difference between the two shaders lies there:

	// Failing version
	bool IntersectAABBFine(AABB _oAABB)
	{
		bool bResult = true;
		uint i = 0;
		[unroll]
		while (bResult && i < 6u)
		{
			bResult = arrPlanes[i].Distance(GetPositiveVertex(_oAABB, arrPlanes[i].vNormal)) >= 0;
			++i;
		}
		return bResult;
	}

	// Compiling version
	bool IntersectAABBFine(AABB _oAABB)
	{
		bool bResult = true;
		uint i = 0;
		[unroll]
		while (i < 6u)
		{
			bResult = bResult && arrPlanes[i].Distance(GetPositiveVertex(_oAABB, arrPlanes[i].vNormal)) >= 0;
			++i;
		}
		return bResult;
	}

Should I create another issue for this one ?