optimize `float` -> `f64` conversions on non-abi3

benches/bench_frompyobject.rs

@@ -2,7 +2,7 @@ use criterion::{black_box, criterion_group, criterion_main, Bencher, Criterion};
 
 use pyo3::{
     prelude::*,
-    types::{PyList, PyString},
+    types::{PyFloat, PyList, PyString},
 };
 
 #[derive(FromPyObject)]
@@ -79,6 +79,15 @@ fn not_a_list_via_extract_enum(b: &mut Bencher<'_>) {
     })
 }
 
+fn f64_from_pyobject(b: &mut Bencher<'_>) {
+    Python::with_gil(|py| {
+        let obj = PyFloat::new(py, 1.234);
+        b.iter(|| {
+            let _: f64 = obj.extract().unwrap();
+        });
+    })
+}
+
 fn criterion_benchmark(c: &mut Criterion) {
     c.bench_function("enum_from_pyobject", enum_from_pyobject);
     c.bench_function("list_via_downcast", list_via_downcast);
@@ -86,6 +95,7 @@ fn criterion_benchmark(c: &mut Criterion) {
     c.bench_function("not_a_list_via_downcast", not_a_list_via_downcast);
     c.bench_function("not_a_list_via_extract", not_a_list_via_extract);
     c.bench_function("not_a_list_via_extract_enum", not_a_list_via_extract_enum);
+    c.bench_function("f64_from_pyobject", f64_from_pyobject);
 }
 
 criterion_group!(benches, criterion_benchmark);

newsfragments/3345.changed.md

@@ -0,0 +1 @@
+Optimize conversion of `float` to `f64` (and `PyFloat::value`) on non-abi3 builds.

pyo3-ffi/src/cpython/floatobject.rs

@@ -0,0 +1,27 @@
+use crate::{PyFloat_Check, PyObject};
+use std::os::raw::c_double;
+
+#[repr(C)]
+pub struct PyFloatObject {
+    pub ob_base: PyObject,
+    pub ob_fval: c_double,
+}
+
+#[inline]
+pub unsafe fn _PyFloat_CAST(op: *mut PyObject) -> *mut PyFloatObject {
+    debug_assert_eq!(PyFloat_Check(op), 1);
+    op.cast()
+}
+
+#[inline]
+pub unsafe fn PyFloat_AS_DOUBLE(op: *mut PyObject) -> c_double {
+    (*_PyFloat_CAST(op)).ob_fval
+}
+
+// skipped PyFloat_Pack2
+// skipped PyFloat_Pack4
+// skipped PyFloat_Pack8
+
+// skipped PyFloat_Unpack2
+// skipped PyFloat_Unpack4
+// skipped PyFloat_Unpack8

pyo3-ffi/src/cpython/mod.rs

@@ -29,6 +29,7 @@ pub(crate) mod pymem;
 pub(crate) mod pystate;
 pub(crate) mod pythonrun;
 // skipped sysmodule.h
+pub(crate) mod floatobject;
 pub(crate) mod tupleobject;
 pub(crate) mod unicodeobject;
 pub(crate) mod weakrefobject;
@@ -42,6 +43,7 @@ pub use self::compile::*;
 pub use self::descrobject::*;
 #[cfg(not(PyPy))]
 pub use self::dictobject::*;
+pub use self::floatobject::*;
 pub use self::frameobject::*;
 pub use self::funcobject::*;
 pub use self::genobject::*;

pyo3-ffi/src/floatobject.rs

@@ -6,13 +6,6 @@ use std::ptr::addr_of_mut;
 // TODO: remove (see https://github.com/PyO3/pyo3/pull/1341#issuecomment-751515985)
 opaque_struct!(PyFloatObject);
 
-#[cfg(not(Py_LIMITED_API))]
-#[repr(C)]
-pub struct PyFloatObject {
-    pub ob_base: PyObject,
-    pub ob_fval: c_double,
-}
-
 #[cfg_attr(windows, link(name = "pythonXY"))]
 extern "C" {
     #[cfg_attr(PyPy, link_name = "PyPyFloat_Type")]
@@ -44,12 +37,6 @@ extern "C" {
     pub fn PyFloat_AsDouble(arg1: *mut PyObject) -> c_double;
 }
 
-#[cfg(not(Py_LIMITED_API))]
-#[inline]
-pub unsafe fn PyFloat_AS_DOUBLE(op: *mut PyObject) -> c_double {
-    (*(op as *mut PyFloatObject)).ob_fval
-}
-
 // skipped non-limited _PyFloat_Pack2
 // skipped non-limited _PyFloat_Pack4
 // skipped non-limited _PyFloat_Pack8

src/types/floatob.rs

@@ -1,7 +1,8 @@
 #[cfg(feature = "experimental-inspect")]
 use crate::inspect::types::TypeInfo;
 use crate::{
-    ffi, AsPyPointer, FromPyObject, IntoPy, PyAny, PyErr, PyObject, PyResult, Python, ToPyObject,
+    ffi, AsPyPointer, FromPyObject, IntoPy, PyAny, PyErr, PyNativeType, PyObject, PyResult, Python,
+    ToPyObject,
 };
 use std::os::raw::c_double;
 
@@ -29,7 +30,16 @@ impl PyFloat {
 
     /// Gets the value of this float.
     pub fn value(&self) -> c_double {
-        unsafe { ffi::PyFloat_AsDouble(self.as_ptr()) }
+        #[cfg(not(Py_LIMITED_API))]
+        unsafe {
+            // Safety: self is PyFloat object
+            ffi::PyFloat_AS_DOUBLE(self.as_ptr())
+        }
+
+        #[cfg(Py_LIMITED_API)]
+        unsafe {
+            ffi::PyFloat_AsDouble(self.as_ptr())
+        }
     }
 }
 
@@ -54,6 +64,15 @@ impl<'source> FromPyObject<'source> for f64 {
     // PyFloat_AsDouble returns -1.0 upon failure
     #![allow(clippy::float_cmp)]
     fn extract(obj: &'source PyAny) -> PyResult<Self> {
+        // On non-limited API, .value() uses PyFloat_AS_DOUBLE which
+        // allows us to have an optimized fast path for the case when
+        // we have exactly a `float` object (it's not worth going through
+        // `isinstance` machinery for subclasses).
+        #[cfg(not(Py_LIMITED_API))]
+        if let Ok(float) = obj.downcast_exact::<PyFloat>() {
+            return Ok(float.value());
+        }
+
         let v = unsafe { ffi::PyFloat_AsDouble(obj.as_ptr()) };
 
         if v == -1.0 {
@@ -101,11 +120,7 @@ impl<'source> FromPyObject<'source> for f32 {
 
 #[cfg(test)]
 mod tests {
-    #[cfg(not(Py_LIMITED_API))]
+    use crate::{types::PyFloat, Python, ToPyObject};
-    use crate::ffi::PyFloat_AS_DOUBLE;
-    #[cfg(not(Py_LIMITED_API))]
-    use crate::AsPyPointer;
-    use crate::{Python, ToPyObject};
 
     macro_rules! num_to_py_object_and_back (
         ($func_name:ident, $t1:ty, $t2:ty) => (
@@ -127,15 +142,14 @@ mod tests {
     num_to_py_object_and_back!(to_from_f32, f32, f32);
     num_to_py_object_and_back!(int_to_float, i32, f64);
 
-    #[cfg(not(Py_LIMITED_API))]
     #[test]
-    fn test_as_double_macro() {
+    fn test_float_value() {
         use assert_approx_eq::assert_approx_eq;
 
         Python::with_gil(|py| {
             let v = 1.23f64;
-            let obj = v.to_object(py);
+            let obj = PyFloat::new(py, 1.23);
-            assert_approx_eq!(v, unsafe { PyFloat_AS_DOUBLE(obj.as_ptr()) });
+            assert_approx_eq!(v, obj.value());
         });
     }
 }