Performance

Why it's fast

graph LR
    subgraph "whereat"
        A["Location::caller()"] -->|"static ref"| B["Push to InlineVec"]
        B -->|"4 inline slots"| C["Done"]
    end
    subgraph "backtrace"
        D["Walk stack frames"] -->|"per frame"| E["Resolve symbols"]
        E -->|"read DWARF"| F["Format addresses"]
    end

#[track_caller] bakes file:line:col into the binary as static data. Location::caller() returns a pointer to data that already exists. No stack walking, no symbol resolution, no debug info parsing. The only runtime cost is pushing that pointer into a small vector.

The default InlineVec<4> stores 4 location slots inline (no heap allocation). Overflow spills to the heap via Vec. Optional features (_tinyvec-128-bytes, _smallvec-128-bytes, etc.) increase inline capacity — see ADVANCED.md.

Benchmarks

                                 Error creation time (lower is better)

Ok path (no error)      █ <1ns            ← ZERO overhead on success
plain enum error        █ <1ns
whereat (1 frame)       ███ 18ns          ← file:line:col captured
whereat (3 frames)      ███ 19ns
whereat (10 frames)     ██████████ 67ns

With RUST_BACKTRACE=1:
anyhow                  █████████████████████████████████████████████████ 2,500ns
backtrace crate         ████████████████████████████████████████████████████████████████████████████████████████████████████ 6,300ns
panic + catch_unwind    ██████████████████████████ 1,300ns

Same-depth comparison (10 frames, 10k iterations):

whereat .at()           █ 1.2ms           ← 100x faster than backtrace
panic + catch_unwind    ██████████████████████ 27ms
backtrace crate         ████████████████████████████████████████████████████████████████████████████████████████████████████ 119ms

anyhow/panic only capture backtraces when RUST_BACKTRACE=1. whereat always captures location.

Linux x86_64 (WSL2), 2026-01-18. Run cargo bench --bench overhead and cargo bench --bench nested_loops "fair_10fr" to reproduce.

Memory layout

At<E> is sizeof(E) + 8 bytes — your error stored inline, plus a single pointer to the boxed trace.

Component	Size	Notes
At<E>	sizeof(E) + 8	Error inline, trace pointer null when empty
AtTrace (heap)	~40 bytes + locations + contexts	Only allocated on error
Per location	8 bytes (static pointer)	&'static Location<'static>
Per context	varies	Static strings are just a pointer; typed data is boxed

On the Ok path: zero heap allocation. On the error path: one Box<AtTrace> allocation, then location pushes into the inline slots (no allocation for ≤4 frames).

Per-operation costs

Operation	Cost	Notes
at(err) / at!(err)	~18-23ns	Dominated by Box<AtTrace> allocation
.at() on Result	~6.5ns	Push static pointer to InlineVec
.at_str("msg")	~6.5ns	Push static pointer (no copy)
.at_string(|| format!(...))	varies	Closure runs + String allocation
.map_err_at(|e| ...)	<1ns	Moves trace pointer, no allocation
Ok path (any method)	<1ns	Short-circuit, no work done