Crystal Performance

Crystal Performance

Use this skill for measured optimization work only.

If the user explicitly wants benchmark design, throughput numbers, concurrency scaling data, or proof that a change helps instead of guesses, also use crystal-benchmarking.

Read the bundled references only as needed:

• references/profiling-tools.md
• references/performance-patterns.md
• references/alys.md

Rules

1. Identify the hot path before changing code.
2. Fix a broken benchmark harness before optimizing runtime code.
3. Capture a clear baseline and compare against it.
4. Treat each optimization as a separate experiment.
5. Keep failed or flat results; they are part of the deliverable.
6. Revert or discard changes without a measured win.
7. Re-run correctness gates after every kept change.

Workflow

1. Find the measurement entrypoint

Prefer the repo's own benchmark, profile, or performance plan. If the documented command is stale, repair the harness first.

2. Capture a baseline

Use release mode for timing unless you specifically need a debug-friendly call tree. Write down the exact command and the number you will compare against.

Run at least 3 times. Crystal benchmark noise is high — single-run results can appear ±20% from the mean. Discard the first (cold) run and average the rest.

If Crystal writes compiler cache or benchmark artifacts outside the workspace, redirect them before measuring. Prefer commands like:

mkdir -p bin .crystal-cache
CRYSTAL_CACHE_DIR=$PWD/.crystal-cache crystal build --release -o bin/bench ...

3. Name the hot path

Target one concrete cost such as allocation churn, duplicate scans, string building, closure creation, or expensive decoding.

4. Apply the smallest plausible change

Prefer less allocation, less copying, simpler data movement, and less repeated work. Avoid broad refactors while the signal is still unclear.

5. Re-measure and decide

Keep a change only if it measurably improves the targeted benchmark or allocation profile. Otherwise revert it and record the result.

6. Re-run correctness gates

Run focused specs first, then broader gates if the repo has them.

When optimizing multiple interchangeable implementations of the same API surface, add or extend a shared contract test before trusting benchmark wins. Benchmark speed is not evidence of semantic parity.

Benchmark Harness Checks

Repair the harness before drawing conclusions if any of these are true:

• The benchmark can be optimized away because the result is unused.
• The documented "MT" path only changes compile flags and does not spawn concurrent workers.
• Warmup and averaging are undocumented or absent.
• The benchmark writes outputs to directories that may not exist.
• The benchmark mixes setup cost into the timed region without meaning to.

For concurrency benchmarks specifically, separate these questions:

• Is the code compiled with MT/runtime flags?
• Is the workload actually concurrent?
• Is the benchmark measuring throughput, return latency, or background flush?

The first does not imply the second.

FFI Hot Paths (C binding libraries)

When optimizing Crystal code that wraps a C library, the dominant cost is almost always C function call overhead, not Crystal method dispatch or object allocation. LLVM inlines trivial Crystal methods in release mode.

Pattern: Redundant FFI bounds checks

Public methods often do bounds-checking FFI calls (e.g. child_count) that are redundant when called from a tight loop that already knows the bounds.

Fix: Inline the raw LibFoo.function(...) call directly in the iterator body, bypassing the public method's defensive FFI calls. Cache the bounds count once before the loop.

# BEFORE — 3 FFI calls per child (count + child + isnull)
def each_child(&)
  count = child_count
  i = 0
  while i < count
    yield child(i)   # child(i) calls child_count again!
    i += 1
  end
end

# AFTER — 2 FFI calls per child, count cached once
def each_child(&)
  unsafe = to_unsafe
  count = LibFoo.foo_node_child_count(unsafe)
  i = 0u32
  while i < count
    node = LibFoo.foo_node_child(unsafe, i)
    yield Node.new_unsafe(node) unless LibFoo.foo_node_is_null(node)
    i += 1
  end
end

Pattern: Redundant null checks after cursor move

Cursor-based iterators call current_node which does a ts_node_is_null FFI call. After a successful goto_next_sibling / goto_first_child, the cursor is guaranteed to be at a valid node.

Fix: Call ts_tree_cursor_current_node directly and skip the null check in iterator internals. Add a protected def unsafe_cursor_ptr on the cursor class for access.

When to stop: C-dominated paths

If a benchmark spends >95% of its time inside C library calls (e.g. query execution, parsing), optimizing the Crystal wrapper produces negligible gains. Identify these early by comparing total time against the per-iteration overhead. Cancel the experiment and record the finding.

Block methods vs Iterator classes

Crystal Iterator-based iterators require a heap-allocated class instance (typically 32–64B). Block-based methods (each_foo { |x| ... }) can be zero-allocation if they inline the FFI calls directly.

Provide both: the Iterator for chainable/lazy use, and the block method for hot paths. Document the tradeoff in comments.

StringPool for C strings

C APIs often return const char* without a length. Use StringPool.get(ptr, LibC.strlen(ptr)) instead of String.new(ptr) to deduplicate strings across repeated calls. This avoids allocating new String objects for the same C string returned repeatedly by the library.

Class with finalizer = unavoidable heap alloc

TreeCursor, QueryCursor, LookaheadIterator — any Crystal class with a finalize method to free C resources must be heap-allocated. There is no way to make these stack-allocated. If the allocation dominates a benchmark, the fix is to add a reuse pattern (pass an existing cursor as a parameter) or a block-based alternative that avoids the class entirely.

Deliverable

Report:

• the exact measurement command
• the named hot path
• the baseline used
• each experiment and what changed
• before/after numbers for kept changes
• flat or regressed numbers for discarded changes
• the correctness gates that ran

Notes

• Use alys only when allocation tracing is the actual question.
• If alys is not already present in shard.yml, ask before adding it.