cargo update, profile, clippy, bloat misc

.github/workflows/artifact.yaml

@@ -1,7 +1,7 @@
 name: artifact
 on: push
 env:
-  RUST_TOOLCHAIN: "nightly-2024-04-15"
+  RUST_TOOLCHAIN: "nightly-2024-04-30"
   CARGO_UNSTABLE_SPARSE_REGISTRY: "true"
   UNSAFE_PYO3_SKIP_VERSION_CHECK: "1"
 jobs:
@@ -156,7 +156,7 @@ jobs:
         LDFLAGS: "-Wl,--as-needed"
         RUSTFLAGS: "-Z mir-opt-level=4 -Z threads=4 -D warnings -C target-feature=-crt-static"
       with:
-        rust-toolchain: nightly-2024-04-15
+        rust-toolchain: nightly-2024-04-30
         rustup-components: rust-src
         target: ${{ matrix.platform.target }}
         manylinux: musllinux_1_2
@@ -245,7 +245,7 @@ jobs:
         RUSTFLAGS: "${{ matrix.target.rustflags }}"
       with:
         target: ${{ matrix.target.target }}
-        rust-toolchain: nightly-2024-04-15
+        rust-toolchain: nightly-2024-04-30
         rustup-components: rust-src
         manylinux: auto
         args: --release --strip --out=dist --features=${{ matrix.target.features }} -i python${{ matrix.python.version }}
@@ -308,7 +308,7 @@ jobs:
 
     - uses: dtolnay/rust-toolchain@master
       with:
-        toolchain: "nightly-2024-04-15"
+        toolchain: "nightly-2024-04-30"
         targets: "aarch64-apple-darwin, x86_64-apple-darwin"
         components: "rust-src"
 
@@ -378,7 +378,7 @@ jobs:
 
     - uses: dtolnay/rust-toolchain@master
       with:
-        toolchain: "nightly-2024-04-15"
+        toolchain: "nightly-2024-04-30"
         targets: "aarch64-apple-darwin, x86_64-apple-darwin"
         components: "rust-src"
 

.github/workflows/debug.yaml

@@ -10,7 +10,7 @@ jobs:
         profile: [
           { rust: "1.72", features: "" },
           { rust: "1.72", features: "--features=yyjson" },
-          { rust: "nightly-2024-04-15", features: "--features=yyjson,unstable-simd" },
+          { rust: "nightly-2024-04-30", features: "--features=yyjson,unstable-simd"},
         ]
         python: [
           { version: '3.13' },

Cargo.toml

@@ -45,24 +45,24 @@ no-panic = [
 yyjson = []
 
 [dependencies]
-ahash = { version = "^0.8.9", default_features = false, features = ["compile-time-rng"] }
-arrayvec = { version = "0.7", default_features = false, features = ["std", "serde"] }
-associative-cache = { version = "2", default_features = false }
-beef = { version = "0.5", default_features = false, features = ["impl_serde"] }
-bytecount = { version = "^0.6.7", default_features = false, features = ["runtime-dispatch-simd"] }
-chrono = { version = "=0.4.34", default_features = false }
-compact_str = { version = "0.7", default_features = false, features = ["serde"] }
-encoding_rs = { version = "0.8", default_features = false }
-half = { version = "2", default_features = false, features = ["std"] }
-itoa = { version = "1", default_features = false }
+ahash = { version = "^0.8.9", default-features = false, features = ["compile-time-rng"] }
+arrayvec = { version = "0.7", default-features = false, features = ["std", "serde"] }
+associative-cache = { version = "2", default-features = false }
+beef = { version = "0.5", default-features = false, features = ["impl_serde"] }
+bytecount = { version = "^0.6.7", default-features = false, features = ["runtime-dispatch-simd"] }
+chrono = { version = "=0.4.34", default-features = false }
+compact_str = { version = "0.7", default-features = false, features = ["serde"] }
+encoding_rs = { version = "0.8", default-features = false }
+half = { version = "2", default-features = false, features = ["std"] }
+itoa = { version = "1", default-features = false }
 itoap = { version = "1", features = ["std", "simd"] }
-once_cell = { version = "1", default_features = false, features = ["race"] }
-pyo3-ffi = { version = "^0.21", default_features = false, features = ["extension-module"]}
-ryu = { version = "1", default_features = false }
-serde = { version = "1", default_features = false }
-serde_json = { version = "1", default_features = false, features = ["std", "float_roundtrip"] }
-simdutf8 = { version = "0.1", default_features = false, features = ["std", "aarch64_neon"] }
-smallvec = { version = "^1.11", default_features = false, features = ["union", "write"] }
+once_cell = { version = "1", default-features = false, features = ["race"] }
+pyo3-ffi = { version = "^0.21", default-features = false, features = ["extension-module"]}
+ryu = { version = "1", default-features = false }
+serde = { version = "1", default-features = false }
+serde_json = { version = "1", default-features = false, features = ["std", "float_roundtrip"] }
+simdutf8 = { version = "0.1", default-features = false, features = ["std", "aarch64_neon"] }
+smallvec = { version = "^1.11", default-features = false, features = ["union", "write"] }
 
 [build-dependencies]
 cc = { version = "1" }

README.md

@@ -1203,7 +1203,7 @@ It benefits from also having a C build environment to compile a faster
 deserialization backend. See this project's `manylinux_2_28` builds for an
 example using clang and LTO.
 
-The project's own CI tests against `nightly-2024-04-15` and stable 1.72. It
+The project's own CI tests against `nightly-2024-04-30` and stable 1.72. It
 is prudent to pin the nightly version because that channel can introduce
 breaking changes.
 

build.rs

@@ -29,7 +29,7 @@ fn main() {
     }
 
     #[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
-    if !env::var("ORJSON_DISABLE_SIMD").is_ok() {
+    if env::var("ORJSON_DISABLE_SIMD").is_err() {
         if let Some(true) = version_check::supports_feature("portable_simd") {
             println!("cargo:rustc-cfg=feature=\"unstable-simd\"");
         }

ci/azure-pipelines.yml

@@ -1,5 +1,5 @@
 variables:
-  toolchain: nightly-2024-04-15
+  toolchain: nightly-2024-04-30
 
 jobs:
 

script/pynumpy

@@ -35,7 +35,7 @@ elif kind == "float64":
     array = numpy.random.random(size=(50000, 100))
     assert array.dtype == numpy.float64
 elif kind == "bool":
-    dtype = numpy.bool
+    dtype = numpy.bool_
     array = numpy.random.choice((True, False), size=(100000, 200))
 elif kind == "int8":
     dtype = numpy.int8

script/valgrind

@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+valgrind pytest -v --ignore=test/test_memory.py test

src/deserialize/cache.rs

@@ -38,6 +38,7 @@ pub type KeyMap =
 
 pub static mut KEY_MAP: OnceCell<KeyMap> = OnceCell::new();
 
+#[inline(always)]
 pub fn cache_hash(key: &[u8]) -> u64 {
     // try to omit code for >64 path in ahash
     assume!(key.len() <= 64);

src/deserialize/mod.rs

@@ -12,7 +12,6 @@ mod json;
 #[cfg(feature = "yyjson")]
 mod yyjson;
 
-pub use cache::KeyMap;
-pub use cache::KEY_MAP;
+pub use cache::{KeyMap, KEY_MAP};
 pub use deserializer::deserialize;
 pub use error::DeserializeError;

src/deserialize/pyobject.rs

@@ -5,34 +5,30 @@ use crate::str::{hash_str, unicode_from_str};
 use crate::typeref::{FALSE, NONE, TRUE};
 use core::ptr::NonNull;
 
+#[inline(always)]
 pub fn get_unicode_key(key_str: &str) -> *mut pyo3_ffi::PyObject {
-    let pykey = if unlikely!(key_str.len() > 64) {
-        create_str_impl(key_str)
+    if unlikely!(key_str.len() > 64) {
+        let pyob = unicode_from_str(key_str);
+        hash_str(pyob);
+        pyob
     } else {
-        get_unicode_key_impl(key_str)
-    };
-    debug_assert!(ffi!(Py_REFCNT(pykey)) >= 1);
-    debug_assert!(unsafe { (*pykey.cast::<pyo3_ffi::PyASCIIObject>()).hash != -1 });
-    pykey
-}
-
-fn get_unicode_key_impl(key_str: &str) -> *mut pyo3_ffi::PyObject {
-    let hash = cache_hash(key_str.as_bytes());
-    let map = unsafe { KEY_MAP.get_mut().unwrap_or_else(|| unreachable!()) };
-    let entry = map.entry(&hash).or_insert_with(
-        || hash,
-        || {
-            let pyob = create_str_impl(key_str);
-            CachedKey::new(pyob)
-        },
-    );
-    entry.get()
-}
-
-fn create_str_impl(key_str: &str) -> *mut pyo3_ffi::PyObject {
-    let pyob = unicode_from_str(key_str);
-    hash_str(pyob);
-    pyob
+        let hash = cache_hash(key_str.as_bytes());
+        unsafe {
+            let entry = KEY_MAP
+                .get_mut()
+                .unwrap_or_else(|| unreachable!())
+                .entry(&hash)
+                .or_insert_with(
+                    || hash,
+                    || {
+                        let pyob = unicode_from_str(key_str);
+                        hash_str(pyob);
+                        CachedKey::new(pyob)
+                    },
+                );
+            entry.get()
+        }
+    }
 }
 
 #[allow(dead_code)]

src/serialize/per_type/dataclass.rs

@@ -83,6 +83,7 @@ impl DataclassFastSerializer {
 }
 
 impl Serialize for DataclassFastSerializer {
+    #[inline(never)]
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
         S: Serializer,
@@ -92,12 +93,13 @@ impl Serialize for DataclassFastSerializer {
             return ZeroDictSerializer::new().serialize(serializer);
         }
         let mut map = serializer.serialize_map(None).unwrap();
-        let mut next_key: *mut pyo3_ffi::PyObject = core::ptr::null_mut();
-        let mut next_value: *mut pyo3_ffi::PyObject = core::ptr::null_mut();
 
         let mut pos = 0;
+        let mut next_key: *mut pyo3_ffi::PyObject = core::ptr::null_mut();
+        let mut next_value: *mut pyo3_ffi::PyObject = core::ptr::null_mut();
 
         pydict_next!(self.ptr, &mut pos, &mut next_key, &mut next_value);
+
         for _ in 0..ffi!(Py_SIZE(self.ptr)) as usize {
             let key = next_key;
             let value = next_value;
@@ -144,6 +146,8 @@ impl DataclassFallbackSerializer {
 }
 
 impl Serialize for DataclassFallbackSerializer {
+    #[cold]
+    #[inline(never)]
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
         S: Serializer,
@@ -157,12 +161,12 @@ impl Serialize for DataclassFallbackSerializer {
         }
         let mut map = serializer.serialize_map(None).unwrap();
 
+        let mut pos = 0;
         let mut next_key: *mut pyo3_ffi::PyObject = core::ptr::null_mut();
         let mut next_value: *mut pyo3_ffi::PyObject = core::ptr::null_mut();
 
-        let mut pos = 0;
-
         pydict_next!(fields, &mut pos, &mut next_key, &mut next_value);
+
         for _ in 0..ffi!(Py_SIZE(fields)) as usize {
             let attr = next_key;
             let field = next_value;