Fix #11506: accept heap texel pointer as atomic destination

[nv-slang-bot]

Summary

InterlockedAdd (any atomic) on a texel of a DescriptorHandle<RWTexture2D<uint>> failed to compile with error[E41403]: invalid atomic destination — but only under -capability spvDescriptorHeapEXT. A plain RWTexture2D<uint> (or the same DescriptorHandle without the capability) did not hit E41403. This is a false-negative in the IR atomic-destination validator; the fix teaches it about the descriptor-heap-lowered texel pointer.

Diagnosis (root cause)

• Under spvDescriptorHeapEXT, processImageSubscript (source/slang/slang-ir-spirv-legalize.cpp:1357-1370) replaces the IRImageSubscript atomic destination with an IRSPIRVLoadTexelPointerFromHeap (a pointer in AddressSpace::Image). The plain-texture branch (:1352-1355) leaves the IRImageSubscript in place.
• Atomic-operation validation is deliberately deferred for SPIR-V (source/slang/slang-emit.cpp:1917-1921) and runs at the end of legalization (source/slang/slang-ir-spirv-legalize.cpp:2789) — i.e. after that rewrite.
• isValidAtomicDest (source/slang/slang-ir-validate.cpp:455-527) accepts IRImageSubscript (:463) and an address-space allowlist (:468-487), but has no case for IRSPIRVLoadTexelPointerFromHeap and no AddressSpace::Image case → falls through to return false (:526) → InvalidAtomicDestinationPointer / E41403.
• The op is genuinely a valid atomic target: it already emits as OpUntypedImageTexelPointerEXT (source/slang/slang-emit-spirv.cpp:4934-4959), atomicable just like OpImageTexelPointer. So emission was already correct — only the IR-level validator was out of date.

Approach

Accept IRSPIRVLoadTexelPointerFromHeap in isValidAtomicDest, immediately after the existing IRImageSubscript case — the post-lowering l-value form of the same construct the validator already accepts pre-lowering. One added check.

Alternatives considered and rejected:

• Broad case AddressSpace::Image: return true; — would bless any Image-space pointer as an atomic destination; less safe than recognizing the exact heap-texel-pointer opcode. (Independent review concurred: prefer the opcode check.)
• Reorder validation / change the heap rewrite — SPIR-V atomic validation is intentionally deferred until after address-space specialization so groupshared function-parameter atomics validate; reordering would regress that.

Files changed

• source/slang/slang-ir-validate.cpp — accept IRSPIRVLoadTexelPointerFromHeap as a valid atomic destination (+ a short rationale comment).
• tests/bugs/gh-11506-descriptor-handle-texture-atomic.slang — new compile-only regression test.

Tests

• Negative proof (without the fix): the repro fails with error[E41403]: invalid atomic destination.
• Positive proof (with the fix): compiles cleanly and emits OpUntypedImageTexelPointerEXT + OpAtomicIAdd.
• New test passes via slang-test.
• Regression: 36/36 atomic + descriptor-heap + image-atomic tests pass (tests/spirv/atomic*, tests/spirv/descriptor-heap*, tests/spirv/image-atomic*, tests/bugs/vk-image-atomics, etc.), 8 GPU-only tests ignored.
• Note: clang-format is not available in this environment, so extras/formatting.sh was not run locally; the change matches the surrounding style and CI's format check covers it.

⚠️ Known limitation — bindless image format (#11130)

This PR fixes only the spurious E41403 atomic-validator error. The front-end / IR path now compiles and emits the correct ops, but end-to-end binary SPIR-V still fails spirv-val for bindless texture atomics, because the bindless image type carries image format Unknown:

Expected the Image Format in Image to be R64i, R64ui, R32f, R32i, or R32ui
for Vulkan environment ... OpUntypedImageTexelPointerEXT

This missing-format defect is orthogonal to the atomic validator and is tracked separately by #11130. It reproduces for DescriptorHandle<RWTexture2D<uint>> atomics both with and without spvDescriptorHeapEXT (a plain RWTexture2D<uint> infers R32ui and validates; bindless handles get Unknown). It lives on the bindless image-format surface (see also #11499, #11503), not the atomic validator. Accordingly, the regression test uses -skip-spirv-validation and FileChecks the emitted opcodes (format-agnostic, so it stays green once #11130 lands).

Risk

Validation-only; the change widens what counts as a valid atomic destination by exactly one opcode that is produced solely on the texture-atomic heap-lowering path, so it cannot newly reject any existing program. No public header / ABI / language-surface change.

Fixes #11506.

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