Optimize Integer.parse/2 by matching digit ranges by preciz · Pull Request #16037 · elixir-lang/elixir

preciz · 2026-10-29T17:21:01Z

Assisted-by: Antigravity:Gemini-4-Argon

Faster or same speed.
Lower or same memory usage.
.beam size same (slightly smaller)

Replacing the 62 unrolled character clauses with range guards and a
base-10 fast path eliminates a 75-entry indirect jump table across 36
branch targets, allowing the JIT to fuse bounds checks into is_in_range
and use tagless, overflow-free arithmetic with the 2..36 base type
bounds. Matching the first digit directly in count_digits/2 also avoids
bitstring position backtracking on unsigned inputs, while the
count == byte_size(binary) fast path skips :erlang.split_binary/2 and
its tuple/sub-binary allocation when the entire binary is parsed.

Bench:

Mix.install([:benchee])

path = "lib/elixir/lib/integer.ex"
baseline_ref = "c2f16efdc"

{baseline_source, 0} = System.cmd("git", ["show", "#{baseline_ref}:#{path}"], cd: __DIR__)

for {source, module} <- [{baseline_source, Old}, {File.read!(Path.join(__DIR__, path)), New}] do
  source
  |> String.replace("defmodule Integer do", "defmodule #{inspect(module)} do")
  |> Code.compile_string(path)
end

base10_short = "12345"
base10_signed = "-9876543210"
base10_medium = String.duplicate("1234567890", 2)
base10_long = String.duplicate("1234567890", 50)
base10_with_rem = "404 not found"
base10_long_with_rem = base10_long <> " not found"
base16_short_lower = "deadbeef"
base16_short_lower_with_rem = "deadbeef_rem"
base16_medium = "deadBEEF0123456789abcdef"
base16_long = String.duplicate("deadBEEF0123456789abcdef", 20)
base16_long_with_rem = base16_long <> "_rem"
base36_medium = "ElixirIsAwesome1234567890"

base36_long =
  String.duplicate("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz", 10)

base2_long = String.duplicate("10110011", 64)
invalid = "not_an_integer"
invalid_sign_only = "-not_an_integer"

inputs = %{
  "base 10 (short, 5 digits)" => {base10_short, 10},
  "base 10 (signed, 10 digits)" => {base10_signed, 10},
  "base 10 (medium, 20 digits)" => {base10_medium, 10},
  "base 10 (long, 500 digits)" => {base10_long, 10},
  "base 10 (with remainder)" => {base10_with_rem, 10},
  "base 10 (long, with remainder)" => {base10_long_with_rem, 10},
  "base 2 (long, 512 digits)" => {base2_long, 2},
  "base 16 (short lowercase, 8 digits)" => {base16_short_lower, 16},
  "base 16 (short lowercase, with remainder)" => {base16_short_lower_with_rem, 16},
  "base 16 (medium, 24 digits)" => {base16_medium, 16},
  "base 16 (long, 480 digits)" => {base16_long, 16},
  "base 16 (long, with remainder)" => {base16_long_with_rem, 16},
  "base 36 (medium, 25 digits)" => {base36_medium, 36},
  "base 36 (long, 620 digits)" => {base36_long, 36},
  "invalid (empty)" => {"", 10},
  "invalid (no digits)" => {invalid, 10},
  "invalid (base 16, no digits)" => {invalid, 16},
  "invalid (sign only)" => {invalid_sign_only, 10}
}

Benchee.run(
  %{
    "a#{baseline_ref} (baseline)" => fn {bin, base} -> Old.parse(bin, base) end,
    "working tree (optimized)" => fn {bin, base} -> New.parse(bin, base) end
  },
  inputs: inputs,
  pre_check: :all_same,
  warmup: 1,
  time: 3,
  memory_time: 1
)

Results:

Operating System: Linux
CPU Information: AMD Ryzen 7 8845HS w
Number of Available Cores: 16
Available memory: 54.72 GB
Elixir 1.21.0-dev
Erlang 29.1.2
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 1 s
time: 3 s
memory time: 1 s
reduction time: 0 ns
parallel: 1
inputs: base 10 (long, 500 digits), base 10 (long, with remainder), base 10 (medium, 20 digits), base 10 (short, 5 digits), base 10 (signed, 10 digits), base 10 (with remainder), base 16 (long, 480 digits), base 16 (long, with remainder), base 16 (medium, 24 digits), base 16 (short lowercase, 8 digits), base 16 (short lowercase, with remainder), base 2 (long, 512 digits), base 36 (long, 620 digits), base 36 (medium, 25 digits), invalid (base 16, no digits), invalid (empty), invalid (no digits), invalid (sign only)
Estimated total run time: 3 min
Excluding outliers: false

##### With input base 10 (long, 500 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)      395.60 K        2.53 μs   ±188.43%        2.42 μs        3.92 μs
ac2f16efdc (baseline)         324.01 K        3.09 μs   ±170.28%        2.85 μs        5.47 μs

Comparison: 
working tree (optimized)      395.60 K
ac2f16efdc (baseline)         324.01 K - 1.22x slower +0.56 μs

Memory usage statistics:

Name                        Memory usage
working tree (optimized)         2.36 KB
ac2f16efdc (baseline)            2.44 KB - 1.03x memory usage +0.0781 KB

**All measurements for memory usage were the same**

##### With input base 10 (long, with remainder) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)      387.48 K        2.58 μs   ±262.06%        2.47 μs        3.89 μs
ac2f16efdc (baseline)         180.04 K        5.55 μs   ±116.65%        5.39 μs        8.00 μs

Comparison: 
working tree (optimized)      387.48 K
ac2f16efdc (baseline)         180.04 K - 2.15x slower +2.97 μs

Memory usage statistics:

Name                        Memory usage
working tree (optimized)         2.45 KB
ac2f16efdc (baseline)            2.45 KB - 1.00x memory usage +0 KB

**All measurements for memory usage were the same**

##### With input base 10 (medium, 20 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)        5.20 M      192.36 ns  ±2251.27%         170 ns         330 ns
ac2f16efdc (baseline)           3.85 M      259.49 ns  ±1675.17%         221 ns         381 ns

Comparison: 
working tree (optimized)        5.20 M
ac2f16efdc (baseline)           3.85 M - 1.35x slower +67.13 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)           400 B
ac2f16efdc (baseline)              480 B - 1.20x memory usage +80 B

**All measurements for memory usage were the same**

##### With input base 10 (short, 5 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       17.21 M       58.10 ns  ±8591.75%          41 ns          70 ns
ac2f16efdc (baseline)          13.59 M       73.57 ns  ±1168.22%          70 ns         101 ns

Comparison: 
working tree (optimized)       17.21 M
ac2f16efdc (baseline)          13.59 M - 1.27x slower +15.47 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            64 B
ac2f16efdc (baseline)              128 B - 2.00x memory usage +64 B

**All measurements for memory usage were the same**

##### With input base 10 (signed, 10 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       13.30 M       75.21 ns  ±8769.63%          60 ns          90 ns
ac2f16efdc (baseline)          10.84 M       92.25 ns  ±3179.26%          80 ns         130 ns

Comparison: 
working tree (optimized)       13.30 M
ac2f16efdc (baseline)          10.84 M - 1.23x slower +17.04 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            64 B
ac2f16efdc (baseline)              136 B - 2.13x memory usage +72 B

**All measurements for memory usage were the same**

##### With input base 10 (with remainder) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       13.01 M       76.85 ns  ±4666.94%          60 ns         100 ns
ac2f16efdc (baseline)          12.32 M       81.16 ns  ±4661.15%          70 ns         100 ns

Comparison: 
working tree (optimized)       13.01 M
ac2f16efdc (baseline)          12.32 M - 1.06x slower +4.31 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)           144 B
ac2f16efdc (baseline)              144 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

##### With input base 16 (long, 480 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)      320.55 K        3.12 μs   ±227.13%        2.98 μs        4.73 μs
ac2f16efdc (baseline)         208.64 K        4.79 μs   ±155.76%        5.51 μs        7.50 μs

Comparison: 
working tree (optimized)      320.55 K
ac2f16efdc (baseline)         208.64 K - 1.54x slower +1.67 μs

Memory usage statistics:

Name                        Memory usage
working tree (optimized)         2.45 KB
ac2f16efdc (baseline)            2.52 KB - 1.03x memory usage +0.0781 KB

**All measurements for memory usage were the same**

##### With input base 16 (long, with remainder) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)      313.99 K        3.18 μs   ±237.66%        3.06 μs        4.80 μs
ac2f16efdc (baseline)         237.07 K        4.22 μs   ±120.80%        3.44 μs        7.25 μs

Comparison: 
working tree (optimized)      313.99 K
ac2f16efdc (baseline)         237.07 K - 1.32x slower +1.03 μs

Memory usage statistics:

Name                        Memory usage
working tree (optimized)         2.53 KB
ac2f16efdc (baseline)            2.53 KB - 1.00x memory usage +0 KB

**All measurements for memory usage were the same**

##### With input base 16 (medium, 24 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)        4.36 M      229.25 ns  ±1658.61%         210 ns         371 ns
ac2f16efdc (baseline)           2.85 M      351.34 ns  ±1698.47%         311 ns         551 ns

Comparison: 
working tree (optimized)        4.36 M
ac2f16efdc (baseline)           2.85 M - 1.53x slower +122.09 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)           400 B
ac2f16efdc (baseline)              480 B - 1.20x memory usage +80 B

**All measurements for memory usage were the same**

##### With input base 16 (short lowercase, 8 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       13.69 M       73.06 ns  ±8229.39%          60 ns         100 ns
ac2f16efdc (baseline)          11.72 M       85.31 ns  ±1519.53%          80 ns         121 ns

Comparison: 
working tree (optimized)       13.69 M
ac2f16efdc (baseline)          11.72 M - 1.17x slower +12.25 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            64 B
ac2f16efdc (baseline)              128 B - 2.00x memory usage +64 B

**All measurements for memory usage were the same**

##### With input base 16 (short lowercase, with remainder) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       11.17 M       89.50 ns  ±2707.31%          80 ns         120 ns
ac2f16efdc (baseline)          10.85 M       92.12 ns  ±3120.87%          80 ns         120 ns

Comparison: 
working tree (optimized)       11.17 M
ac2f16efdc (baseline)          10.85 M - 1.03x slower +2.62 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)           136 B
ac2f16efdc (baseline)              136 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

##### With input base 2 (long, 512 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)      521.40 K        1.92 μs   ±345.64%        1.83 μs        2.59 μs
ac2f16efdc (baseline)         306.43 K        3.26 μs   ±175.08%        3.37 μs        4.60 μs

Comparison: 
working tree (optimized)      521.40 K
ac2f16efdc (baseline)         306.43 K - 1.70x slower +1.35 μs

Memory usage statistics:

Name                        Memory usage
working tree (optimized)         1.23 KB
ac2f16efdc (baseline)            1.30 KB - 1.06x memory usage +0.0781 KB

**All measurements for memory usage were the same**

##### With input base 36 (long, 620 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)      183.52 K        5.45 μs   ±108.83%        5.24 μs        8.19 μs
ac2f16efdc (baseline)         151.36 K        6.61 μs    ±44.12%        6.98 μs        9.56 μs

Comparison: 
working tree (optimized)      183.52 K
ac2f16efdc (baseline)         151.36 K - 1.21x slower +1.16 μs

Memory usage statistics:

Name                        Memory usage
working tree (optimized)         3.82 KB
ac2f16efdc (baseline)            3.90 KB - 1.02x memory usage +0.0781 KB

**All measurements for memory usage were the same**

##### With input base 36 (medium, 25 digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)        3.46 M      288.73 ns  ±1622.65%         260 ns         441 ns
ac2f16efdc (baseline)           2.79 M      358.41 ns    ±80.91%         360 ns         491 ns

Comparison: 
working tree (optimized)        3.46 M
ac2f16efdc (baseline)           2.79 M - 1.24x slower +69.68 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)           472 B
ac2f16efdc (baseline)              560 B - 1.19x memory usage +88 B

**All measurements for memory usage were the same**

##### With input invalid (base 16, no digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       28.72 M       34.82 ns  ±1562.68%          30 ns          50 ns
ac2f16efdc (baseline)          26.88 M       37.21 ns  ±1521.73%          30 ns          50 ns

Comparison: 
working tree (optimized)       28.72 M
ac2f16efdc (baseline)          26.88 M - 1.07x slower +2.38 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            40 B
ac2f16efdc (baseline)               40 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

##### With input invalid (empty) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       30.09 M       33.23 ns  ±1675.37%          30 ns          40 ns
ac2f16efdc (baseline)          29.25 M       34.19 ns  ±1849.49%          30 ns          50 ns

Comparison: 
working tree (optimized)       30.09 M
ac2f16efdc (baseline)          29.25 M - 1.03x slower +0.95 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            40 B
ac2f16efdc (baseline)               40 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

##### With input invalid (no digits) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       28.99 M       34.49 ns  ±1964.59%          30 ns          50 ns
ac2f16efdc (baseline)          25.86 M       38.67 ns  ±1632.46%          40 ns          50 ns

Comparison: 
working tree (optimized)       28.99 M
ac2f16efdc (baseline)          25.86 M - 1.12x slower +4.18 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            40 B
ac2f16efdc (baseline)               40 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

##### With input invalid (sign only) #####
Name                               ips        average  deviation         median         99th %
working tree (optimized)       27.84 M       35.92 ns  ±1562.84%          30 ns          50 ns
ac2f16efdc (baseline)          26.04 M       38.40 ns  ±1495.80%          31 ns          50 ns

Comparison: 
working tree (optimized)       27.84 M
ac2f16efdc (baseline)          26.04 M - 1.07x slower +2.47 ns

Memory usage statistics:

Name                        Memory usage
working tree (optimized)            40 B
ac2f16efdc (baseline)               40 B - 1.00x memory usage +0 B

**All measurements for memory usage were the same**

Replacing the 62 unrolled character clauses with range guards and a base-10 fast path eliminates a 75-entry indirect jump table across 36 branch targets, allowing the JIT to fuse bounds checks into is_in_range and use tagless, overflow-free arithmetic with the 2..36 base type bounds. Matching the first digit directly in count_digits/2 also avoids bitstring position backtracking on unsigned inputs, while the count == byte_size(binary) fast path skips :erlang.split_binary/2 and its tuple/sub-binary allocation when the entire binary is parsed. Assisted-by: Antigravity:Gemini-4-Argon

rdlu · 2026-10-29T18:38:10Z

Assisted-by: Claude Code 2.1.311:claude-opus-5-5

Independent verification of this PR on a different machine (AMD Ryzen 7 7735HS, Elixir 1.20.4, OTP 29.1.1, JIT).

Correctness: fuzzed Old.parse/2 (c2f16ef) vs New.parse/2 (3fa127d) on 300,840 inputs — random strings over 0-9a-zA-Z+-_ ., control and high bytes, all bases 2..36, plus edge cases (empty, lone sign, +-1, the chars adjacent to each range like / : @ [ ` {, unicode digits). 0 mismatches. Error behaviour is identical too (ArgumentError for bases 1, 37, 10.0, :a; FunctionClauseError for non-binaries).

Benchmark: the script from the description plus a few extra inputs, run twice with longer settings (2s warmup / 5s time, then 3s warmup / 15s time). Speedup of the PR over the baseline:

Input	PR description	5s run	15s run
base 10, 1 digit (extra)	—	1.67x	1.63x
base 10, 5 digits	1.27x	1.68x	1.62x
base 10, signed 10 digits	1.23x	1.48x	1.44x
base 10, 20 digits	1.35x	1.27x	1.25x
base 16, 8 digits lowercase	1.17x	1.33x	1.36x
base 16, 8 digits uppercase (extra)	—	1.31x	1.32x
base 2, 8 digits (extra)	—	1.49x	1.42x
base 10, 500 digits	1.22x	2.32x	1.57x
base 10, 500 digits + remainder	2.15x	2.20x	1.61x
base 16, 480 digits	1.54x	1.60x	1.33x
base 2, 512 digits	1.70x	2.01x	2.39x
base 36, 620 digits	1.21x	1.68x	1.68x
short with remainder (`"404 not found"`, `"-42px"`, `"deadbeef_rem"`)	1.03–1.06x	1.00–1.06x	1.00–1.12x
stops at first out-of-base char (`"7777777781"` base 8, `"ffffffffg"` base 16) (extra)	—	1.00–1.03x	1.01x
invalid inputs	1.03–1.12x	1.05–1.07x	1.01–1.11x

Faster or equal on every input in both runs. Short all-digit inputs are consistently 1.25–1.65x; long inputs vary run to run (bignum allocation/GC) but were always ≥1.2x. Short inputs with a remainder are roughly even, as expected since the count == byte_size(binary) fast path doesn't apply. Memory numbers match the description exactly (e.g. 128 B → 64 B for short full-digit inputs). .beam size: 19,528 → 19,292 bytes.

Suggestion: the existing parse/2 test doesn't cover the range-boundary cases the new guards depend on. Something like this could go into integer_test.exs (passes on both baseline and this PR):

assert Integer.parse("78", 8) === {7, "8"}
assert Integer.parse("fg", 16) === {15, "g"}
assert Integer.parse("FF", 16) === {255, ""}
assert Integer.parse("zZ", 36) === {1295, ""}
assert Integer.parse("9", 9) === :error
assert Integer.parse("-g", 16) === :error
assert Integer.parse("Zz", 35) === :error

Minor: the base in 2..36 added to the clause guard is redundant at runtime (the clause above already raises), so a short comment noting it's there for the compiler/JIT might help future readers.

josevalim merged commit 669f848 into elixir-lang:main Oct 30, 2026
15 checks passed

Navigation Menu

Repository navigation

Optimize Integer.parse/2 by matching digit ranges - #16037

Optimize Integer.parse/2 by matching digit ranges#16037
josevalim merged 1 commit into
elixir-lang:mainfrom
preciz:perf/integer-parse

preciz commented Oct 29, 2026 •

edited

Loading

Uh oh!

rdlu commented Oct 29, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Milestone

Development

Uh oh!

3 participants

Conversation

preciz commented Oct 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rdlu commented Oct 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Milestone

Development

Uh oh!

3 participants

preciz commented Oct 29, 2026 •

edited

Loading

rdlu commented Oct 29, 2026 •

edited

Loading