CPU device: passing a Python list to a float[N] param segfaults in array-parameter marshalling

[nv-slang-bot]

Summary

On DeviceType.cpu, calling a Slang function that takes a fixed-size array parameter (e.g. float x[3]) with a Python list segfaults (SIGSEGV) during array-parameter marshalling. This is a distinct, pre-existing CPU-backend defect that was unmasked by the #1136 dispatch-limits fix: previously the "Device reports zero compute dispatch groups in X" throw (#1136) fired first, so this code path was never reached on CPU. GPU backends (Vulkan / CUDA / D3D12 / Metal) handle the same call correctly.

Repro

Requires a build where the #1136 regression is fixed (PR #1137, or slang-rhi#854 + a submodule bump) — otherwise the #1136 throw fires first and this path is unreachable.

import slangpy as spy

dev = spy.create_device(type=spy.DeviceType.cpu)
mod = spy.Module.load_from_source(dev, "m", "float first(float x[3]) { return x[0]; }")
mod.first([3.0, 4.0, 5.0])   # SIGSEGV during array-parameter marshalling

This is the DeviceType.cpu case of the existing slangpy/tests/slangpy_tests/test_simple_function_call.py::test_pass_float_array (which passes on all GPU backends).

Context / provenance

• Found by an automated SlangPy coworker while validating the Every compute dispatch on the CPU device fails: "Device reports zero compute dispatch groups in X" (regression in 0.43.0) #1136 fix on the CPU backend (linux-gcc debug). The simple grid/scalar dispatch added in Treat zero compute-dispatch-group limit as unbounded (fixes #1136) #1137 (test_cpu_dispatch.py) works end-to-end; only this array-parameter path segfaults.
• Consistent with slang-rhi disabling the CPU backend in its own Linux test harness (external/slang-rhi/tests/testing.cpp: "Known issues with CPU backend on linux").
• Not addressed by Treat zero compute-dispatch-group limit as unbounded (fixes #1136) #1137 (which fixes only the dispatch-group-limit regression). Tracked here on its own — this is not the same bug as Every compute dispatch on the CPU device fails: "Device reports zero compute dispatch groups in X" (regression in 0.43.0) #1136.

Environment

• slangpy main with the Every compute dispatch on the CPU device fails: "Device reports zero compute dispatch groups in X" (regression in 0.43.0) #1136 fix applied (PR Treat zero compute-dispatch-group limit as unbounded (fixes #1136) #1137, branch dev/slangpy-fixer/slangpy-1136), linux-gcc debug; slang-rhi submodule @ 22239042.
• DeviceType.cpu. GPU backends unaffected.

Layer (to be pinned by triage)

Candidate layers for the fault: slangpy_ext array-parameter marshalling (src/slangpy_ext/), the slang-rhi CPU backend, or the slang compiler's CPU (host-callable) target. Triage to follow. Suggested priority: P2 (CPU backend was already entirely unusable before the #1136 fix, so this is not a fresh regression).

---

Related: #1136 (dispatch-limits regression; durable root fix in shader-slang/slang-rhi#854), #1137 (the fix that unmasked this). This issue is intentionally not linked as Fixes #1136.

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

[nv-slang-bot]

Triage — resolved & closed: not reproducible on current toolchain (stale-slang artifact)

• Status: ✅ Closed — not reproducible on current head. The crash was real but on a stale slang toolchain, and is verified fixed on the pinned toolchain.
• Root of the discrepancy: The original SIGSEGV was on an older checkout (slang-rhi ee078c7 → slang 2026.4.1), inside a full pytest test_simple_function_call.py --device-types cpu run (crash at test_pass_float_array:431), deterministic (2/2). "Array marshalling" was inferred from the crash site, not proven in isolation. After rebasing to slang-rhi 22239042 (→ slang 2026.12.2), that test was never re-run — which produced the "live defect" impression.
• Verification on current head (slang-rhi 22239042 / slang 2026.12.2), two independent builds: isolated mod.first([3.0,4.0,5.0]) → 3.0, rc=0; full pytest test_simple_function_call.py --device-types cpu → 37 passed, 2 skipped, rc=0, no segfault. Also clean on slang 2026.16.1, across float/double/int/uint/int64, list & numpy, and float[5]. Reflection layout correct.
• Conclusion: A CPU/host-callable crash in slang 2026.4.1, already fixed or absent by the 2026.4.1 → 2026.12.2 slang bump. Not a live defect and not a compiler escalation. Treat zero compute-dispatch-group limit as unbounded (fixes #1136) #1137 is unaffected (built/tested on 2026.12.2).
• Disposition: Closed as not-reproducible-on-current-toolchain — broken in slang 2026.4.1, verified fixed in 2026.12.2. If you hit this on an older toolchain, upgrade slang (≥ 2026.12.2) rather than pinning. Re-open with a backtrace if it recurs on a current build.

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