Per-compile linkIR / linkAndOptimizeIR regression since the auto-diff refactor (auto-diff artifacts linked into every program)

[jvepsalainen-nv]

Issue Description

Since the auto-diff refactor (the overhaul between releases 2026.5 and 2026.7), every compile
pays an inflated linkIR / linkAndOptimizeIR cost — even an empty shader that never
differentiates. The IR linker now deep-clones the differentiable-interface witness tables and their
annotations into every program, then carries that dead auto-diff IR through specialization /
simplify / DCE before eliminating it. This is a per-compile floor: it affects all targets and all
shaders, independent of auto-diff use.

Reproducer Code

An empty / minimal shader (the compile-perf minimal workload):

RWStructuredBuffer<float> outBuf;

[shader("compute")]
[numthreads(1, 1, 1)]
void computeMain()
{
    outBuf[0] = 1.0;
}

Command:

slangc minimal.slang -target spirv -emit-spirv-directly -report-perf-benchmark

Expected Behavior

A non-differentiating program should not link auto-diff witness tables / annotations; linkIR and
linkAndOptimizeIR should be near the pre-refactor (release 2026.5) baseline.

Actual Behavior

linkIR rose ~10–13×, and linkAndOptimizeIR similarly: the minimal linkAndOptimizeIR is
~15 ms versus a pre-refactor baseline of ~1.9 ms.

Test Plan

• tools/compile-perf per-compile-floor workloads (e.g. minimal); the linkIR /
  linkAndOptimizeIR timers, compared against the pre-refactor baseline.

---

Fixed by #11779. Companion codegen-side half (simplifyIR): #11780. Related broad
perf-regression tracking: #11474.

[nv-slang-bot]

Triage — verified at master HEAD (b1bdd88d4).

• Status: Confirmed regression; mechanism reproduced on top-of-tree. Compiling the minimal non-differentiating shader from this report (-target spirv -emit-spirv-directly) still links saturated auto-diff IR into the program — a -dump-ir shows hundreds of IDifferentiable / Differential / fwd_diff·bwd_diff witness-table and annotation insts that the shader never uses. The useAutodiff link gating that Refactor auto-diff implementation. #9808 removed is genuinely absent at HEAD: shouldDeepCloneWitnessTable returns true unconditionally for the differentiable-interface family (source/slang/slang-ir-link.cpp:687), and cloneAnnotations clones every (differentiability-only) IRAnnotation regardless of use.
• Classification: Bug · regression (introduced by Refactor auto-diff implementation. #9808) · IR linker (linkIR / linkAndOptimizeIR, source/slang/slang-ir-link.cpp). Companion codegen-side half: simplifyIR regression: differentiable numeric builtins pull float's differential closure into non-auto-diff shaders #11780 (simplifyIR). Umbrella perf-tracking: Compiler performance regression in MDL bench #11474.
• Verdict: The fix already exists in open draft PR Restore auto-diff link gating to the final codegen link #11779 ("Restore auto-diff link gating to the final codegen link", Fixes #11781) — this issue does not need a fresh patch. Verified against the PR diff: it adds an isFinalCodegenLink flag (true only in linkIR) and gates both cloneAnnotations and the differentiable-family branch of shouldDeepCloneWitnessTable on isFinalCodegenLink && !useAutodiff, restoring the pre-Refactor auto-diff implementation. #9808 behavior while leaving prelink / precompilation (whose output module must stay complete and serializable) untouched. This is the linker-side twin of simplifyIR regression: differentiable numeric builtins pull float's differential closure into non-auto-diff shaders #11780, and the scope split is clean: Restore auto-diff link gating to the final codegen link #11779 fixes the per-compile floor (unreferenced tables / empty shaders), simplifyIR regression: differentiable numeric builtins pull float's differential closure into non-auto-diff shaders #11780 covers the entries a shader references via sin/sqrt.
• Next action: Land PR Restore auto-diff link gating to the final codegen link #11779 — no separate fix needed. It is currently a draft (REVIEW_REQUIRED, CI in flight) and will auto-close this issue on merge once promoted to ready and reviewed.
• Blocker: None for triage. The reported magnitude (~10–13× linkIR; ~15 ms vs ~1.9 ms linkAndOptimizeIR) is the author's tools/compile-perf measurement and was not independently re-measured here — only the mechanism (dead auto-diff IR in a non-diff program) and the missing gating were verified at HEAD.

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