Add DebugBuildIdentifier to the live worklist during initialization

[jkwak-work]

DebugBuildIdentifier was missing from the set of NonSemantic debug instructions added to the worklist in InitializeModuleScopeLiveInstructions.
Instead it was handled in ProcessGlobalValues (after the worklist is
exhausted) using live_insts_.Set() on its direct operands.

live_insts_.Set() marks an instruction live but does not enqueue it, so transitive operand dependencies are never visited. In shaders where the only use of a type (e.g. OpTypeInt) is through a constant that is itself only referenced by DebugBuildIdentifier's flags argument, the type instruction is not marked live and is incorrectly killed by ADCE.

The surviving constant then references a deleted type, producing invalid SPIR-V. Any subsequent pass that rebuilds the DefUseManager (e.g. CCP) will fail with "Definition is not registered."

Fix: include NonSemanticShaderDebugInfo100DebugBuildIdentifier in the worklist loop alongside the other always-live debug instructions. The worklist's transitive closure correctly marks all operand dependencies live before global-value cleanup runs. The existing special-case code in ProcessGlobalValues becomes unreachable for this opcode (IsLive returns true, so the early continue fires) and is now dead; it is left in place for safety.

[CLAassistant]

All committers have signed the CLA.

[jkwak-work]

I am not able to make a simple repro SPIRV for this issue.

As far as I can tell the problem is on the fact that the constant variable used for DebugBuildIdentifier is not recognized and it causes an assertion failure as following:

Assertion failed: def && "Definition is not registered.", file D:\sbf\git\SPIRV-Tools\source\opt\def_use_manager.cpp, line 56

This issue was observed only with an option, --eliminate-dead-code-aggressive.

If I can somehow make a simpler repro case later, I will share it to have it as a regression test.

[jkwak-work]

I found a small repro Spirv Code for the issue.

Fixes #6619

[jkwak-work]

@SteveUrquhart , can you give me a review for this?

[SteveUrquhart]

I'll take a look at it, thanks.

[SteveUrquhart]

I think the change itself is clearly the right thing to do. It would be best if your repro was part of the commit, you can add it as a case to https://github.com/KhronosGroup/SPIRV-Tools/blob/main/test/opt/aggressive_dead_code_elim_test.cpp. You can CHECK for the presence of DebugBuildIdentifier, as well as verify that the output is valid.

[jkwak-work]

I had to close this PR and reopen a new PR:

• Add DebugBuildIdentifier to the live worklist during initialization #6626