XPerf
Guides

Static-check source collection

Public sources for performance and memory shapes a linter can see. A row here is not a measurement on this tree. Numbers below are the source's numbers. A new slug still needs a local before/after, and the default stays hint.

Dispositions:

  • already-ruled — a slug in that kit already covers the shape.
  • measure-first — the source describes a shape a scanner can see, and this tree has no slug for it yet.
  • rejected — the source's own result is small, reversed, or not a stable rule.
  • no-compiler — this machine cannot run that language, so the shape stays unmeasured.

cperf

  • alloca inside a loop keeps each chunk until the function returns. Stack Overflow. already-ruled (cp-alloca-hot). Local stack cover was 1280000 bytes versus 64. The touch time was about 1×.
  • fgetc / fputc one byte at a time. The question reports about 10× against a 4096-byte fread. Stack Overflow. rejected. A local block read was faster, and a line matcher cannot tell a byte parser from a bulk read.
  • malloc in a loop, realloc by one, strcat, sprintf, fopen, gettimeofday, and a wide run of boolean fields are already-ruled.

cppperf

  • Constructing std::function in a loop heap-allocates once the capture misses the small buffer. The cost of captures. rejected. A local sink rerun made the hoisted std::function slower (about 0.6×).
  • Copying shared_ptr by value bumps an atomic count. GotW #91 and a pass-by-value writeup. rejected. A local copy was about 4.8×, and make_shared of a changing value cannot be hoisted.
  • const std::string / const std::vector by value was about 15–21× versus a const reference when the copy allocates. already-ruled (cx-const-by-value). A 4-character string was about 1× and the hint still fires.
  • Walking vector<bool> was about 12× a byte vector, and the packed container used fewer bytes. already-ruled (cx-vector-bool).
  • std::string append without reserve, endl, regex construction, clocks, and ifstream are already-ruled. String streams were removed after a fair rerun.

csharpperf

  • Building a string with + inside a loop allocates. Microsoft Learn. already-ruled (cs-string-plus-hot).
  • C# 10 interpolation no longer lowers to the old boxing string.Format path. Microsoft .NET blog. rejected as a "interpolation is slow" rule until a local .NET 8 run says otherwise.
  • File reads, Stopwatch.GetTimestamp, GetField, JSON, regex, and ToList are already-ruled.

goperf

  • s += part in a loop copies the whole string. A pre-grown strings.Builder does one allocation. String concatenation vs strings.Builder. already-ruled (gp-string-plus-hot).
  • bytes.Buffer vs strings.Builder on a short fragment list was about 1.5× on that machine. brandur.org. rejected.
  • fmt in a loop, json of one payload, os.Open, time.Now, regexp compile, and FieldByName are already-ruled.

jsperf

  • delete on an object moves it to dictionary properties. V8 fast properties. already-ruled (jp-delete-shape).
  • Deleting array elements, or reading a hole, flips element storage. web.dev V8 tips. already-ruled (jp-delete-array-hole).
  • Pre-allocating a huge array (the page cites more than 64K) is called out on that page. measure-first only for a literal huge new Array(n) a scanner can see without flagging normal arrays.
  • Regex construction, JSON.parse of one payload, readFileSync, and Date.now / performance.now are already-ruled.

pyperf

  • CPython can make += on str faster than ''.join for some shapes. String concatenation in Python. already-ruled for str (pp-str-plus-hot). Do not add a second join rule from the old quadratic story alone.
  • [None] * n versus append was about 1.3× in the posted timing. Stack Overflow. rejected.
  • bytes built with += versus b''.join was reported far apart. Fast bytes concatenation. measure-first.
  • open and json.loads of one payload are already-ruled. perf_counter_ns and getattr were measured locally and not added.

rubyperf

  • Repeated + allocates. A buffer wants <<. fast-ruby concatenation. already-ruled (rb-string-plus-hot).
  • # frozen_string_literal: true was about 1.2–1.3× on a tiny literal loop in that post. FastRuby.io. rejected until a local run on a real file shows more.
  • File.read, Regexp.new, and JSON.parse of one payload are already-ruled. Process.clock_gettime and send were measured locally and not added.

rustperf

  • collect allocates. Return the iterator, or extend an existing Vec. The Rust Performance Book. already-ruled in part (rs-collect-vec-hot).
  • iter().cloned().collect() on a slice should be to_vec(). Clippy iter_cloned_collect. measure-first. The published claim is "faster", not a ratio.
  • .clone() of a heap value inside a loop deep-copies. Clones and allocations. measure-first. Clippy's redundant_clone has historically skipped loop bodies (rust-clippy#17495).
  • File::open, Instant::now, and Regex::new are already-ruled. There is no JSON parser in the standard library.

swiftperf

  • reserveCapacity once on an empty array avoids geometric growth. Calling it every iteration with count + k copies repeatedly. Hacking with Swift. Missing reserve is already-ruled (sp-array-no-reserve). Reserve-inside-the-loop is measure-first.
  • DateFormatter, JSONDecoder, NSRegularExpression, contentsOfFile, Date(), and value(forKey:) are already-ruled.

objcperf

  • A long loop of convenience constructors ([NSMutableString string]) can pile up autoreleased objects. Apple documents an @autoreleasepool inside the loop to cut the peak. Autorelease pool blocks and Stack Overflow. measure-first for the missing pool. Creating the pool every iteration is cheap (mikeash.com), so the rule must not flag a pool that is already there.
  • stringByAppendingString, NSLog, date formatters, JSON, regex, file reads, NSDate, and valueForKey: are already-ruled.

zperf

  • Building a new ArrayList on every iteration allocates. Reusing one list and calling clearRetainingCapacity matched a stack buffer in that thread. ziglang/zig#16982. Heap allocation in a loop is already-ruled (zp-alloc-hot, zp-create-destroy-hot). Do not add a second slug for the same shape.
  • std.fmt.allocPrint allocates a fresh slice. lib/std/fmt.zig. already-ruled (zp-alloc-print-hot) when the format string and arguments do not change. Local ReleaseFast medians: about 50× on the general-purpose allocator, about 12× on an arena, about 14× for a constant string. bufPrint and byte copies stayed at or under about 7× and were not added.
  • File open, Io.Clock, and json.parseFromSlice of one payload are already-ruled. The 0.17 standard library has no regular-expression compiler.

jvmperf

  • StringBuilder instead of + in a loop is the classic shape. byroot's comparison restates it. already-ruled (jp-string-concat-hot).
  • SparseArray avoids boxing integer keys and is meant for a few hundred entries. Android SparseArray. already-ruled (jp-android-sparse-array).
  • Boxed streams, regex, format, and the other JVM slugs stay as they are. This machine has no JDK that can run a new bench, so new JVM shapes are no-compiler.

kotlinperf

  • A lambda passed to a non-inline function allocates a closure. inline removes that object. Kotlin inline functions. no-compiler (no kotlinc here). JVM already has jp-lambda-to-non-inline-hot.
  • SparseArray on Android is the same note as jvmperf. already-ruled on the JVM side, not re-measured for Kotlin.
  • String +, println, and List growth are already-ruled in this kit.

phpperf

  • PHP strings are mutable. echo of several arguments avoids a concatenate. Stack Overflow. . in a loop is already-ruled (ph-string-dot-hot).
  • implode of a large array can use more peak memory than appending, in that SitePoint note. SitePoint. no-compiler.

dartperf

  • A cross-language string-build repo includes Dart and names StringBuffer as the builder. langs_string_build_test. That repo is not a controlled bench for this tree. String + in a loop is already-ruled (dt-string-plus-hot). Further Dart shapes are no-compiler.

luaperf

  • Collecting lines and calling table.concat cut a 350 KB read from 40 seconds to 0.5 seconds in Programming in Lua. PIL 11.6. already-ruled (lu-string-concat-hot). This machine has no lua, so the book figure is not re-run. no-compiler for anything beyond that slug.

beamperf

  • Binary <> copies. An iolist is a list of chunks, and IO functions write it without building one binary. Elixir IO data and iolist notes. List concatenation in a loop is already-ruled (be-list-concat-reduce). This machine has no Erlang or Elixir, so a new iolist slug is no-compiler.

nimperf

  • The same cross-language repo times Nim string building and does not isolate setLen. langs_string_build_test. String & and seq growth are already-ruled. Further Nim shapes are no-compiler.

arktsperf

  • If either operand of + is a string, the spec says a new string object is created when the expression is not constant. ArkTS spec, additive expressions. There is no string-plus slug in this kit yet. no-compiler. Do not copy JavaScript timings onto ArkTS.
  • Allocation in a loop, delete, console, includes, and JSON are already-ruled.

Rule ID map (historical — numeric → slug)

Previous Page

XPerf — what else (same playbook)

Next Page

On this page

cperfcppperfcsharpperfgoperfjsperfpyperfrubyperfrustperfswiftperfobjcperfzperfjvmperfkotlinperfphpperfdartperfluaperfbeamperfnimperfarktsperf