Simplify code generated for for_each_method_def and for_each_proto_slot

[1tgr]

This change replaces slice1.iter().chain(slice2)....chain(sliceN).for_each(f) with f(slice1); f(slice2); ... f(sliceN);.

On a real-world project it significantly reduces number of lines of LLVM code generated by the compiler, with a similar improvement in compilation time.

I adapted the word_count example to add a #[pyclass]:

#[pyclass]
struct WordCount {
    contents: String,
}

#[pymethods]
impl WordCount {
    #[new]
    fn new(contents: String) -> Self {
        Self { contents }
    }

    fn search(&self, needle: &str) -> usize {
        search(&self.contents, needle)
    }

    fn search_sequential(&self, needle: &str) -> usize {
        search_sequential(&self.contents, needle)
    }

    fn search_sequential_allow_threads(&self, py: Python, needle: &str) -> usize {
        search_sequential_allow_threads(py, &self.contents, needle)
    }
}
...
m.add_class::<WordCount>()?;

According to cargo llvm-lines --release, before this change the top 10 functions in the word_count example, by number of LLVM lines, are:

  Lines         Copies       Function name
  -----         ------       -------------
  50562 (100%)  2257 (100%)  (TOTAL)
   2323 (4.6%)    16 (0.7%)  <core::iter::adapters::chain::Chain<A,B> as core::iter::traits::iterator::Iterator>::fold
   2125 (4.2%)    24 (1.1%)  core::iter::traits::iterator::Iterator::fold
   1255 (2.5%)    21 (0.9%)  alloc::alloc::box_free
   1026 (2.0%)    16 (0.7%)  core::result::Result<T,E>::map_err
    844 (1.7%)    12 (0.5%)  std::panicking::try
    789 (1.6%)     8 (0.4%)  pyo3::callback::handle_panic
    753 (1.5%)     3 (0.1%)  word_count::<impl pyo3::class::impl_::PyMethods<word_count::WordCount> for pyo3::class::impl_::PyClassImplCollector<word_count::WordCount>>::py_methods::METHODS::__wrap::{{closure}}
    665 (1.3%)    16 (0.7%)  core::iter::adapters::chain::Chain<A,B>::new
    648 (1.3%)    16 (0.7%)  core::iter::traits::iterator::Iterator::chain
    598 (1.2%)    30 (1.3%)  core::ptr::read

And after:

  Lines         Copies       Function name
  -----         ------       -------------
  45552 (100%)  2168 (100%)  (TOTAL)
   1255 (2.8%)    21 (1.0%)  alloc::alloc::box_free
   1026 (2.3%)    16 (0.7%)  core::result::Result<T,E>::map_err
    844 (1.9%)    12 (0.6%)  std::panicking::try
    789 (1.7%)     8 (0.4%)  pyo3::callback::handle_panic
    753 (1.7%)     3 (0.1%)  word_count::<impl pyo3::class::impl_::PyMethods<word_count::WordCount> for pyo3::class::impl_::PyClassImplCollector<word_count::WordCount>>::py_methods::METHODS::__wrap::{{closure}}
    657 (1.4%)     7 (0.3%)  core::iter::traits::iterator::Iterator::fold
    598 (1.3%)    30 (1.4%)  core::ptr::read
    570 (1.3%)    10 (0.5%)  alloc::raw_vec::RawVec<T,A>::current_memory
    558 (1.2%)     8 (0.4%)  alloc::boxed::Box<T,A>::into_unique
    546 (1.2%)     7 (0.3%)  std::thread::local::LocalKey<T>::try_with

[1tgr]

I expect the actual machine code to be the same before and after, since slice's iterator doesn't do anything special in its for_each, meaning it's equivalent to this for loop.

[davidhewitt]

Thanks, the slice approach looks good to me!