Optimizer: eliminate unused Unchecked.defaultof<'T> bindings

[Copilot]

Unchecked.defaultof<T> unused bindings were not eliminated even with optimizations enabled, because the optimizer's effect analysis incorrectly classified the underlying EI_ilzero IL instruction as having side effects.

Root cause

Unchecked.defaultof<T> inlines to (# "ilzero !0" type ('T) : 'T #), which the optimizer represents as TOp.ILAsm([EI_ilzero ty], ...). The IlAssemblyCodeInstrHasEffect function in Optimizer.fs did not recognize EI_ilzero as effect-free, so it fell through to the conservative _ -> true default, preventing dead binding elimination.

Change

src/Compiler/Optimize/Optimizer.fs — add EI_ilzero _ to the no-effect branch of IlAssemblyCodeInstrHasEffect. The instruction only produces a zero/default value (mapping to ldnull, ldc.i4.0, or initobj depending on type), with no observable side effects.

// Before: initobj instructions emitted for all three unused bindings
let f (n: float32) =
    Console.WriteLine n
    let _ = Unchecked.defaultof<decimal>   // initobj Decimal
    let _ = Unchecked.defaultof<decimal>   // initobj Decimal
    let _ = Unchecked.defaultof<decimal>   // initobj Decimal
    let n' = n * 2.f
    Console.WriteLine n'

// After: all three bindings eliminated; IL only contains the two Console.WriteLine calls

This is particularly relevant for SRTP-based patterns (common in libraries like FSharpPlus) where Unchecked.defaultof<'T> is used as a witness/dummy argument to drive overload resolution and ends up as unused bindings at call sites.

[github-actions]

❗ Release notes required

You can open this PR in browser to add release notes: open in github.dev

---

✅ Found changes and release notes in following paths:

Change path	Release notes path	Description
src/Compiler	docs/release-notes/.FSharp.Compiler.Service/11.0.100.md

[T-Gro]

Review: Optimizer – treat EI_ilzero as effect-free

Overall: Solid fix. The root-cause analysis is correct, the code change is minimal, and both positive and negative test cases are included. Approve direction; a few observations below.

---

1. Breadth of the ExprHasEffect / OptimizeExprOpFallback type-var guard

The freeInTypes check is now applied to all Expr.Op nodes that are otherwise effect-free, not only TOp.ILAsm [EI_ilzero _].
This means dead-binding elimination is also blocked for unused tuples, union cases, struct records, etc. when their tyargs reference a free typar – even the function's own type parameter (which is safe to reference via the IL generic context).

Example that could be affected:

sharp let f<'T> () = let _ = Unchecked.defaultof<'T> // safe to eliminate, 'T is the function's own typar 42

In practice this is unlikely to matter (who writes unused defaultof with their own typar?), but it would be worth adding a comment explaining the trade-off: "we accept false positives for non-SRTP typars to keep the fix simple."

If you later want a tighter check, you could limit the guard to TOp.ILAsm containing EI_ilzero, since that's the only instruction that can appear with zero value-args in an expression whose sole purpose is providing an SRTP witness/dummy.

2. Duplicate logic – consider a helper

The same freeInTypes CollectTyparsNoCaching tyargs check appears verbatim at lines 1642 and 2663. A small helper like:

sharp let tyargsHaveFreeTypars tyargs = not (List.isEmpty tyargs) && not (Zset.isEmpty (freeInTypes CollectTyparsNoCaching tyargs).FreeTypars)

would reduce duplication and make intent clearer at both call sites.

3. Performance consideration

freeInTypes traverses the type list and allocates a FreeTyvars. It's now called on every Expr.Op with non-empty tyargs whose op+args are effect-free. In large optimized builds this could add up. Have you spot-checked compiler self-build time or BenchmarkDotNet traces to confirm it's negligible? The "NoCaching" variant means no memoization across calls.

4. Tests

Good coverage:

• Concrete type → eliminated ✓
• SRTP type var (including nested Wrapper<^T>) → preserved ✓

Optional addition: a test with a non-SRTP generic function demonstrating the conservative "keep" behavior (to document expected behavior and guard against future over-optimization):

sharp let f<'T> () = let _ = Unchecked.defaultof<'T> 42

5. Minor: release-notes wording

"for concrete types"

The optimization applies whenever tyargs have no free typars, so Unchecked.defaultof<list<int>> is also eliminated – "concrete" is correct but might be clearer as "fully-resolved types" or "types without free type parameters."

---

Overall: the fix is correct and safe. The conservatism is acceptable. Nice work.

[github-actions]

🔍 Tooling Safety Check — Affects-Compiler-Output
Affects-Compiler-Output: modifies Optimizer.fs (IL dead-code elimination for EI_ilzero)

Generated by PR Tooling Safety Check · opus46 7.5M · ◷

[abonie]

@T-Gro most recent commit reverted all changes

[T-Gro]

@abonie :

Its difficult to find a balance between doing this optimization and not breaking a special FSharpPlus usage which intentionally uses a typed defaultof (with specified generic argument) to influence SRTP resolution.

I need more time here, I would really love to get this in.