SPIRV CI: 5 pre-existing debug-info lit failures on amd-staging (LLVM verifier/llc drift)

[lamb-j]

Summary

The SPIRV CI (spirv-ci.yml) Test SPIRV translator lit job (check-amd-llvm-spirv) has 5 pre-existing failing lit tests on amd-staging. They are not regressions — they fail identically on both PR head and the amd-staging baseline, so the "new failures" gate stays green and they are currently informational. This issue tracks fixing/triaging them.

Latest observed run: ROCm/SPIRV-LLVM-Translator run 29958131088 — check-amd-llvm-spirv: 1113 discovered, 998 passed, 5 failed, 6 XFAIL, 104 unsupported.

Failing tests

Test	Failure stage	Root cause
test/DebugInfo/LocalAddressSpace.ll	llvm-as verifier	invalid retained nodes
test/DebugInfo/NonSemantic/Shader200/DebugInfoStringType.ll	llvm-as verifier	invalid retained nodes
test/DebugInfo/X86/dimodule-external-fortran.ll	llvm-as verifier	invalid retained nodes
test/DebugInfo/X86/DW_AT_specification.ll	llvm-as verifier	invalid retained nodes
test/transcoding/dbginfo-bug-on-bool-converts.ll	llc / CHECK-LLC	codegen output drift

Root cause

These are LLVM version drift failures, not translator code bugs. All 5 test .ll files are byte-identical to KhronosGroup/SPIRV-LLVM-Translator main (0 diff vs upstream/main). The failures occur in LLVM tooling (llvm-as, llc) before/around the translator, driven by the newer LLVM that amd-staging builds against.

Bucket 1 — invalid retained nodes (4 tests): llvm-as fails with:

invalid retained nodes, expected DILocalVariable, DILabel, DIImportedEntity or DIType
!13 = !{!0}
!0 = !DIGlobalVariableExpression(var: !1, expr: !DIExpression())
warning: ignoring invalid debug info in <stdin>

The stale fixtures place a DIGlobalVariableExpression in a DISubprogram's retainedNodes. Current LLVM's verifier rejects this and drops the debug info, so the downstream CHECK-SPIRV / CHECK patterns then report "expected string not found in input."

Bucket 2 — llc output drift (1 test): transcoding/dbginfo-bug-on-bool-converts.ll fails at CHECK-LLC (llc DWARF output no longer matches the expected pattern). No retained-nodes error — this is separate codegen-output drift.

Upstream status

The test inputs match Khronos main exactly, so upstream Khronos CI (paired with a compatible LLVM) passes them. They only fail here because amd-staging's LLVM verifier is stricter than the LLVM upstream Khronos pins against. Fixing forward (updating the fixtures' debug-info metadata to the current schema, and refreshing the llc CHECK lines) belongs in this repo's amd-staging; the changes are candidates to send upstream as well.

Proposed next steps

• Update the 4 retained-nodes fixtures so retainedNodes holds only verifier-accepted node kinds (move the DIGlobalVariableExpression out, per current LLVM debug-info schema).
• Regenerate the expected CHECK-LLC lines for dbginfo-bug-on-bool-converts.ll against current llc.
• Consider upstreaming the fixture updates to KhronosGroup/SPIRV-LLVM-Translator.
• Once resolved, these can be promoted from informational back to fully clean.

Notes

• Not a PR regression; gate is "new failures vs baseline" only.
• The Test LLVM SPIRV codegen suite (check-llvm-codegen-spirv) is clean (1127/1130 pass, rest unsupported).

[maarquitos14]

These are all observable upstream too:

1. https://github.com/KhronosGroup/SPIRV-LLVM-Translator/actions/runs/29839453003/job/88664255501?pr=3900
2. https://github.com/KhronosGroup/SPIRV-LLVM-Translator/actions/runs/29839452604/job/88664254943?pr=3900