Re-enable scalar-fp8 (vk) test once spirv-tools supports fp8 constant folding

[expipiplus1]

Summary

PR #11744 added tests/hlsl-intrinsic/scalar-fp8.slang (vk) to tests/expected-failure-no-gpu.txt as a temporary workaround.

The fp8 scalar test aborts during SPIR-V optimization: an fp8 scalar float constant reaches spirv-tools' constant folder, whose GetWordsFromScalarFloatConstant (folding_rules.cpp) asserts the float width is 16/32/64 and aborts on width 8.

Action

Once the spirv-tools fix for fp8 constant folding is upstreamed and pulled into Slang's external/spirv-tools submodule, remove the tests/hlsl-intrinsic/scalar-fp8.slang (vk) entry from tests/expected-failure-no-gpu.txt and confirm the test passes.

Follow-up to #11744 (requested by @szihs).

[nv-slang-bot]

Triage — tracking issue (blocked on an upstream spirv-tools fix)

• Status: Triaged. This is a follow-up cleanup task, not an immediately-fixable Slang bug — no Slang code change is actionable yet. Verified against master HEAD eb6a6efd3 by source inspection.
• Link / confirmation: The assert is real: external/spirv-tools/source/opt/folding_rules.cpp:156 does assert(width == 16 || width == 32 || width == 64), so an fp8 (width-8) scalar float constant aborts during constant folding. The test tests/hlsl-intrinsic/scalar-fp8.slang exists; the suppression entry is added by PR Mark scalar-fp8 (vk) as an expected failure on no-GPU runners #11744, which is still open (so the entry is not on master yet).
• Verdict: enhancement / tracking · severity low · component SPIR-V emit + external/spirv-tools submodule + CI · priority P3. The root defect is upstream in spirv-tools, not in Slang. Companion bug-side tracker: scalar-fp8 (vk) test aborts during SPIR-V constant folding on no-GPU runners #11767 (complementary, not a duplicate).
• Next action (resumption trigger): (1) PR Mark scalar-fp8 (vk) as an expected failure on no-GPU runners #11744 merges to land the suppression this issue removes; (2) upstream spirv-opt: Fix assertion with fp8 KhronosGroup/SPIRV-Tools#6677 ("spirv-opt: Fix assertion with fp8", the in-progress fix for upstream issue Problem converting ConstantBuffer<> arrays from slang to wgsl #6533) merges; (3) bump external/spirv-tools past it; (4) delete the tests/hlsl-intrinsic/scalar-fp8.slang (vk) line from tests/expected-failure-no-gpu.txt and confirm the test passes.
• Blocker: Upstream spirv-opt: Fix assertion with fp8 KhronosGroup/SPIRV-Tools#6677 is not yet merged, and the prerequisite Slang PR Mark scalar-fp8 (vk) as an expected failure on no-GPU runners #11744 is not yet merged. No Slang-side work to schedule until those land. (A local cherry-pick of the upstream fix into the vendored external/spirv-tools is possible to re-enable sooner, but creates submodule drift — only worthwhile if upstream stalls.)

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

I am not sure what the problem is exactly.

1. is it an intermittent failure or 100% deterministic failure?
2. is the failure observed only gh free runners?
3. why "(vk)" test is attempted to run on "no-GPU" test machine?

Until I have a clear repro instructions, I am not sure if this can be resolved.
I suspect that the following PR will resolve the issue but we need a clear repro steps:

• Default slang-test compiler invocations to -O0 #11805

My understanding is that the failure is observed while running the spirv-opt.
#11805 will disable the spirv-opt entirely for slang-test, and it may resolve/avoid the issue.

@nv-slang-bot , what would be a proper repro steps for this issue; can you trace back to the initial problems?
@expipiplus1 , can you tell us how to reproduce the problem?

[nv-slang-bot]

@jkwak-work — traced it back to the initial problem. The short version: it's a deterministic, GPU-free compiler crash inside spirv-opt, and yes, #11805 avoids it (it defaults spirv-opt off for slang-test). Details and a minimal repro below.

Minimal repro (no GPU, no render-test needed)

The abort is a pure compile-time crash, so slangc alone reproduces it — verified locally at master (0065572bf), Debug build:

# Debug build (assertions enabled), default optimization → ABORTS
$ ./build/Debug/bin/slangc tests/hlsl-intrinsic/scalar-fp8.slang \
      -target spirv -stage compute -entry computeMain -o /tmp/out.spv
slangc: .../external/spirv-tools/source/opt/folding_rules.cpp:156:
  ...GetWordsFromScalarFloatConstant(...):
  Assertion `width == 16 || width == 32 || width == 64' failed.
Aborted (core dumped)                       # exit 134

# Same build, add -O0 → no abort, exits 0
$ ./build/Debug/bin/slangc ... -O0 -o /tmp/out.spv            # exit 0

FloatE4M3(3.0) (line 39) is a compile-time fp8 scalar constant. At the default optimization level Slang runs spirv-opt, whose constant folder calls GetWordsFromScalarFloatConstant, which asserts the float width is 16/32/64 and aborts on width 8.

Your three questions

1. Intermittent or deterministic? 100% deterministic — not a race/flake. Every run of an assertions-enabled build at an optimization level that runs spirv-opt hits the same assert (exit 134 every time, above).

2. Only on GH free runners? No — it's not runner-specific. It's a property of (a) an assertions-enabled build × (b) an optimization level that runs spirv-opt. assert() at folding_rules.cpp:156 is compiled out under NDEBUG, so release builds don't abort (width-8 falls through to the "<32-bit → zero-extend to 32" return) — this last point is inferred from the assert semantics; I could only run the Debug binary in my sandbox. In our CI the only assertions-enabled no-GPU leg is test-linux-debug-gcc-aarch64 (ubuntu-24.04-arm, debug); the release no-GPU legs don't hit it, and the cpu-only leg filters -vk out via -api cpu+llvm. (The ASan/coverage legs are also assertions-enabled and use the same no-gpu list.)

3. Why is a (vk) test compiled on a no-GPU machine? Two parts:

• Scheduling: the no-GPU legs run slang-test with -api all (only cpu-only uses -api cpu+llvm), so -vk variants are dispatched to render-test rather than filtered out.
• Where the abort is: render-test compiles the shader (the step that runs spirv-opt) as part of running the test, and compilation is target-codegen that needs no GPU. Note the order in render-test-main.cpp: it creates the device (:1965) and checks -render-features fp8 (:2014-2021) before compiling (:2048 → compileWithLayout). If the leg's runner had no fp8-capable Vulkan device the test would instead be Ignored (SLANG_E_NOT_AVAILABLE → TestResult::Ignored, slang-test-tool-util.cpp:21) and never compile. So the abort means render-test got past that gate — i.e. the runner exposes a Vulkan device (presumably Mesa's software lavapipe on the hosted runner) that advertises fp8 — and proceeded into compilation. The practical point: it's a compiler crash, not an execution/GPU issue, which is exactly why the slangc repro above needs no device.

On #11805 — confirmed, and it changes the resumption trigger

#11805 defaults render-test-backed compiles to -Xslang -O0, and at -O0 spirv-opt is skipped entirely — gated in two places: slang-emit.cpp:3204 (needsOptimization = optLevel != None) and slang-glslang.cpp:268 (glslang_optimizeSPIRV early-returns for NONE, before the optimizer is even constructed). scalar-fp8.slang's -vk directive has no -O and isn't in #11805's opt-in list, so it will compile at -O0 → the fp8 constant never reaches the folder → no abort (reproduced: -O0 exits 0).

So once #11805 merges, the tests/hlsl-intrinsic/scalar-fp8.slang (vk) entry in expected-failure-no-gpu.txt can be removed and the test confirmed passing — it isn't optimization-sensitive (its CHECK values are execution results). This issue therefore no longer has to wait on upstream KhronosGroup/SPIRV-Tools#6677 + a submodule bump; #11805 unblocks the suppression on its own.

One caveat so the two fixes don't get conflated: #11805 only stops slang-test from running spirv-opt on this shader — it does not fix the underlying spirv-tools defect. Anyone compiling fp8 with optimization (-O1+) still hits the assert (debug) / the width-8 fold path (release). The real fix for that general case remains upstream #6677 (issue #6533), tracked on the compiler side by the companion #11767. In short: #11805 → drop this test suppression; #6677 → the actual optimizer fix, separately.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

@nv-slang-bot , I think we can now re-enable the disabled tests because the following PR is merged to ToT:

• Default slang-test compiler invocations to -O0 #11805

Please make a new PR that re-enables those test.

[nv-slang-bot]

Opened #11922 (draft) to re-enable tests/hlsl-intrinsic/scalar-fp8.slang (vk) as requested.

Now that #11805 defaults slang-test render-test compiles to -Xslang -O0 (skipping spirv-opt), the fp8 scalar-constant abort no longer occurs on the slang-test path, so the expected-failure-no-gpu.txt entry is dead weight. Verified on an asserts-enabled Debug build at ToT:

• Standalone slangc … -target spirv on the test aborts (exit 134) at folding_rules.cpp:156 at default opt / -O2, but exits 0 with valid SPIR-V at -O0 — the level Default slang-test compiler invocations to -O0 #11805 uses.
• slang-test tests/hlsl-intrinsic/scalar-fp8.slang (no suppression list): -api all → 100% passed (1/1), 1 ignored; the cuda leg passes, the vk leg cleanly ignores (no fp8 render feature advertised). No abort in any run.

This was the only fp8/spirv-opt suppression in the file — no sibling entries to re-enable.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}