Fix benchmark, add more inlines

CHANGELOG.md

@@ -6,6 +6,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+## [v0.8.1] 2024-09-03
+- Restore performance for recent rustc versions, force more functions to be inlined [!135](https://github.com/rsadsb/dump1090_rs/pull/135)
+
 ## [v0.8.0] 2024-09-02
 - Update MSRV to `1.74` to [!130](https://github.com/rsadsb/dump1090_rs/pull/130)
 - Properly decode short ADSB messages, thanks [@gariac](https://github.com/gariac) for finding. [!130](https://github.com/rsadsb/dump1090_rs/pull/130)

README.md

@@ -104,9 +104,9 @@ Reading from a 512KB iq sample to ADS-B bytes takes ~3.0 ms, but feel free to ru
 
 ### Intel i7-7700K CPU @ 4.20GHz
 ```
-01                      time:   [2.9850 ms 2.9859 ms 2.9868 ms]
-02                      time:   [2.9260 ms 2.9279 ms 2.9308 ms]
-03                      time:   [2.8449 ms 2.8525 ms 2.8653 ms]
+01                      time:   [3.6691 ms 3.6950 ms 3.7264 ms]
+02                      time:   [3.5941 ms 3.5987 ms 3.6040 ms]
+03                      time:   [3.4930 ms 3.4961 ms 3.4994 ms]
 ```
 
 # Changes

benches/demod_benchmark.rs

@@ -1,56 +1,28 @@
-// third-party
-use assert_hex::assert_eq_hex;
+use std::hint::black_box;
+
 use criterion::{criterion_group, criterion_main, Criterion};
-use hexlit::hex;
-// crate
 use libdump1090_rs::{demod_2400::demodulate2400, icao_filter::icao_flush, utils};
 use num_complex::Complex;
 
-fn routine(data: [Complex<i16>; 0x20000], expected_data: &Vec<Vec<u8>>) {
+fn routine(data: [Complex<i16>; 0x20000]) {
     // make sure icao starts in a deterministic position
     icao_flush();
     let outbuf = utils::to_mag(&data);
-    let data = demodulate2400(&outbuf).unwrap();
-    for (a, b) in data.iter().zip(expected_data.iter()) {
-        assert_eq_hex!(a.buffer(), *b);
-    }
+    let _ = black_box(demodulate2400(&outbuf).unwrap());
 }
 
 fn criterion_benchmark(c: &mut Criterion) {
     let filename = "test_iq/test_1641427457780.iq";
     let data_01 = utils::read_test_data(filename);
-    let expected_data_01 = Vec::from([
-        hex!("8dad929358b9c6273f002169c02e").to_vec(),
-        hex!("8daa2bc4f82100020049b8db9449").to_vec(),
-        hex!("8daa2bc4f82100020049b8db9449").to_vec(),
-        hex!("02e1971ce17c84").to_vec(),
-        hex!("8da0aaa058bf163fcf860013e840").to_vec(),
-    ]);
 
     let filename = "test_iq/test_1641428165033.iq";
     let data_02 = utils::read_test_data(filename);
-    let expected_data_02 = Vec::from([
-        hex!("8da79de99909932f780c9e2f2f8f").to_vec(),
-        hex!("8dac04d358a7820a86ac3709e689").to_vec(),
-        hex!("8dac04d3ea4288669b5c082751d4").to_vec(),
-        hex!("8da79de958bdf59c85104874adad").to_vec(),
-        hex!("5dad92936265f5").to_vec(),
-        hex!("5dad92936265f525be017735997b").to_vec(),
-    ]);
 
     let filename = "test_iq/test_1641428106243.iq";
     let data_03 = utils::read_test_data(filename);
-    let expected_data_03 = Vec::from([
-        hex!("8da8aac8990c30b51808aa24e573").to_vec(),
-        hex!("02e19838bff1d9").to_vec(),
-        hex!("8dada6b9990cf61e4848af2a8656").to_vec(),
-        hex!("8da4ba025885462008fa0a4a6eb2").to_vec(),
-        hex!("8da4ba025885462008fa0a4a6eb2").to_vec(),
-        hex!("8da4ba0299115f301074a72db6ff").to_vec(),
-    ]);
-    c.bench_function("01", |b| b.iter(|| routine(data_01, &expected_data_01)));
-    c.bench_function("02", |b| b.iter(|| routine(data_02, &expected_data_02)));
-    c.bench_function("03", |b| b.iter(|| routine(data_03, &expected_data_03)));
+    c.bench_function("01", |b| b.iter(|| routine(data_01)));
+    c.bench_function("02", |b| b.iter(|| routine(data_02)));
+    c.bench_function("03", |b| b.iter(|| routine(data_03)));
 }
 
 criterion_group!(benches, criterion_benchmark);

src/crc.rs

@@ -259,6 +259,7 @@ pub const CRC_TABLE: [u32; 256] = [
     0x00fa_0480,
 ];
 
+#[inline]
 pub fn modes_checksum(message: &[u8], bits: usize) -> u32 {
     let mut rem: u32 = 0;
     let n = bits / 8;

src/demod_2400.rs

@@ -19,6 +19,7 @@ enum Phase {
 }
 
 impl From<usize> for Phase {
+    #[inline(always)]
     fn from(num: usize) -> Self {
         match num % 5 {
             0 => Self::Zero,
@@ -33,6 +34,7 @@ impl From<usize> for Phase {
 
 impl Phase {
     /// Increment from 0..4 for incrementing the starting phase
+    #[inline(always)]
     fn next_start(self) -> Self {
         match self {
             Self::Zero => Self::One,
@@ -44,6 +46,7 @@ impl Phase {
     }
 
     /// Increment by expected next phase transition for bit denoting
+    #[inline(always)]
     fn next(self) -> Self {
         match self {
             Self::Zero => Self::Two,
@@ -55,6 +58,7 @@ impl Phase {
     }
 
     /// Amount of mag indexs used, for adding to the next start index
+    #[inline(always)]
     fn increment_index(self, index: usize) -> usize {
         index
             + match self {
@@ -98,6 +102,7 @@ pub struct ModeSMessage {
 }
 
 impl ModeSMessage {
+    #[inline(always)]
     pub fn buffer(&self) -> &[u8] {
         match self.msglen {
             MsgLen::Short => &self.msg[..MODES_SHORT_MSG_BYTES],
@@ -106,6 +111,7 @@ impl ModeSMessage {
     }
 }
 
+#[inline(always)]
 pub fn demodulate2400(mag: &MagnitudeBuffer) -> Result<Vec<ModeSMessage>, &'static str> {
     let mut results = vec![];
 
@@ -205,6 +211,7 @@ pub fn demodulate2400(mag: &MagnitudeBuffer) -> Result<Vec<ModeSMessage>, &'stat
     Ok(results)
 }
 
+#[inline(always)]
 fn check_preamble(preamble: &[u16]) -> Option<(i32, u32, u32)> {
     // This gets rid of the 3 core::panicking::panic_bounds_check calls,
     // but doesn't look to improve performance

src/mode_s/mod.rs

@@ -10,6 +10,7 @@ use super::{crc::modes_checksum, icao_filter::icao_filter_test};
 
 // mode_s.c:215
 #[must_use]
+#[inline(always)]
 pub fn getbits(data: &[u8], firstbit_1idx: usize, lastbit_1idx: usize) -> usize {
     let mut ans: usize = 0;
 
@@ -33,7 +34,7 @@ pub fn getbits(data: &[u8], firstbit_1idx: usize, lastbit_1idx: usize) -> usize
 pub fn score_modes_message(msg: &[u8]) -> Option<(MsgLen, i32)> {
     let validbits = msg.len() * 8;
 
-    if validbits < 56 {
+    if validbits < MODES_SHORT_MSG_BYTES * 8 {
         return None;
     }