Remove #[inline] from large functions, add small feature to CI

Drop #[inline] from functions with substantial code (short, parse,
parse_text, fmt_float, format_base10, trim_zeros, uscale, unpack64,
fixed_width, mul_pow10) to reduce WASM code duplication. Keep it on
tiny helpers only. Run tests and clippy for both feature configs in CI.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: gfx

.github/workflows/ci.yml

@@ -17,6 +17,7 @@ jobs:
       - uses: actions/checkout@v6
       - uses: dtolnay/rust-toolchain@stable
       - run: cargo test --workspace
+      - run: cargo test --workspace --features small
 
   integrity:
     name: Integrity
@@ -28,6 +29,9 @@ jobs:
           components: clippy, rustfmt
           targets: wasm32-unknown-unknown
       - run: cargo check --workspace
+      - run: cargo check --workspace --features small
       - run: cargo check --target wasm32-unknown-unknown
+      - run: cargo check --target wasm32-unknown-unknown --features small
       - run: cargo clippy --workspace -- -D warnings
+      - run: cargo clippy --workspace --features small -- -D warnings
       - run: cargo fmt --all --check

AGENTS.md

@@ -9,8 +9,9 @@ The crate is `no_std` and `wasm32-unknown-unknown` compatible.
 ## Development
 
 ```sh
-cargo test
+cargo test && cargo test --features small
 cargo fmt
-cargo clippy
+cargo clippy && cargo clippy --features small
 cargo bench -p bench
+cargo bench -p bench --features small
 ```

README.md

@@ -55,8 +55,8 @@ Measured on Apple M3 Pro, macOS 15.7.3 (aarch64):
 
 | Task                      |  fpfmt | fpfmt `small` |    ryu | stdlib |
 | ------------------------- | -----: | ------------: | -----: | -----: |
-| **format** (f64 → string) |  65 ns |        102 ns | 240 ns | 528 ns |
-| **parse** (string → f64)  | 697 ns |        701 ns |      — | 658 ns |
+| **format** (f64 → string) |  63 ns |         99 ns | 239 ns | 529 ns |
+| **parse** (string → f64)  | 690 ns |        700 ns |      — | 675 ns |
 
 ```sh
 cargo bench -p bench
@@ -67,8 +67,8 @@ cargo bench -p bench --features small
 
 | Configuration |            Size |
 | ------------- | --------------: |
-| default       |    14,428 bytes |
-| `small`       | **4,224 bytes** |
+| default       |    14,379 bytes |
+| `small`       | **4,175 bytes** |
 
 ```sh
 cargo build --target wasm32-unknown-unknown --release -p wasm-size

src/lib.rs

@@ -119,7 +119,6 @@ const UINT64_POW10: [u64; 20] = [
 /// `unpack64` returns (m, e) such that `f = m * 2**e`.
 /// The caller is expected to have handled 0, NaN, and +/-Inf already.
 /// To unpack an `f32`, use `unpack64(f as f64)`.
-#[inline]
 #[allow(clippy::many_single_char_names)]
 fn unpack64(f: f64) -> (u64, i32) {
     const SHIFT: u32 = 64 - 53; // 11
@@ -165,7 +164,6 @@ fn prescale(e: i32, p: i32, lp: i32) -> Scaler {
 }
 
 #[cfg(feature = "small")]
-#[inline]
 #[allow(clippy::many_single_char_names)]
 fn mul_pow10(p: i32) -> PmHiLo {
     let q = p.div_euclid(K);
@@ -225,7 +223,6 @@ fn prescale(e: i32, p: i32, lp: i32) -> Scaler {
 /// `uscale` returns `unround(x * 2**e * 10**p)`.
 /// The caller should pass `c = prescale(e, p, log2_pow10(p))`
 /// and should have left-justified x so its high bit is set.
-#[inline]
 fn uscale(x: u64, c: Scaler) -> Unrounded {
     let r = u128::from(x) * u128::from(c.pm.hi);
     let mut hi = (r >> 64) as u64;
@@ -247,7 +244,6 @@ fn uscale(x: u64, c: Scaler) -> Unrounded {
 ///
 /// Panics if `n > 18`.
 #[must_use]
-#[inline]
 #[allow(clippy::many_single_char_names)]
 pub fn fixed_width(f: f64, n: i32) -> (u64, i32) {
     debug_assert!(n <= 18, "too many digits");
@@ -270,7 +266,6 @@ pub fn fixed_width(f: f64, n: i32) -> (u64, i32) {
 ///
 /// Panics if `d > 10_000_000_000_000_000_000` (more than 19 digits).
 #[must_use]
-#[inline]
 #[allow(clippy::many_single_char_names)]
 pub fn parse(d: u64, p: i32) -> f64 {
     debug_assert!(d <= 10_000_000_000_000_000_000, "too many digits");
@@ -294,7 +289,6 @@ pub fn parse(d: u64, p: i32) -> f64 {
 /// Parses a decimal string and returns the nearest f64.
 /// Returns `None` if the input is malformed.
 #[must_use]
-#[inline]
 pub fn parse_text(s: &[u8]) -> Option<f64> {
     fn is_digit(c: u8) -> bool {
         c.wrapping_sub(b'0') <= 9
@@ -362,7 +356,6 @@ pub fn parse_text(s: &[u8]) -> Option<f64> {
 /// using as few digits as possible that will still round trip
 /// back to the original f64.
 #[must_use]
-#[inline]
 #[allow(clippy::many_single_char_names)]
 pub fn short(f: f64) -> (u64, i32) {
     const MIN_EXP: i32 = -1085;
@@ -411,7 +404,6 @@ fn skewed(e: i32) -> i32 {
 /// Removes trailing zeros from `x * 10**p`.
 /// If x ends in k zeros, returns `(x/10**k, p+k)`.
 /// Assumes that x ends in at most 16 zeros.
-#[inline]
 #[allow(clippy::unreadable_literal)]
 fn trim_zeros(x: u64, p: i32) -> (u64, i32) {
     const INV5P8: u64 = 0xc767074b22e90e21; // inverse of 5**8
@@ -472,7 +464,6 @@ const I2A: &[u8] = b"\
 /// Formats the decimal representation of u into a.
 /// The caller is responsible for ensuring that a is big enough to hold u.
 /// If a is too big, leading zeros will be filled in as needed.
-#[inline]
 fn format_base10(a: &mut [u8], mut u: u64) {
     let mut nd = a.len();
     while nd >= 8 {
@@ -523,7 +514,6 @@ fn format_base10(a: &mut [u8], mut u: u64) {
 /// The caller must pass nd set to the number of digits in d.
 /// Returns the number of bytes written to s.
 #[must_use]
-#[inline]
 pub fn fmt_float(s: &mut [u8], d: u64, p: i32, nd: i32) -> usize {
     let nd = nd as usize;
     // Put digits into s, leaving room for decimal point.