llvm

e2795a52 - [SYCL] Reduce sycl/khr/split_headers/accessor.hpp compile-time overhead (#22073)

[SYCL] Reduce sycl/khr/split_headers/accessor.hpp compile-time overhead (#22073) This change cuts the compile-time cost of including `sycl/khr/split_headers/accessor.hpp` by ~27% in a minimal TU (single `#include <sycl/accessor.hpp>`, -fsycl -fsycl-device-only -std=c++20): | Metric | Before | After | Delta | |---------------------|---------|---------|--------------| | Transitive headers | 402 | 322 | -80 (-20%) | | Total source-time | 887.9ms | 643.4ms | -244.5ms | Measured with clang -ftime-trace, attributing self-time per header. * Introduce `sycl/detail/fwd/buffer.hpp` so accessor.hpp can drop `#include <sycl/buffer.hpp>` in favor of a forward declaration. The primary `class buffer` template default arguments live in the new fwd header (with the corresponding defaults removed from buffer.hpp) to keep them in one place. Companion fwd-decls in handler.hpp, buffer_properties.hpp, and detail/backend_traits_opencl.hpp are retired in favor of the new header. * Drop dead/unused includes from accessor.hpp: - `<sycl/properties/buffer_properties.hpp>`: accessor.hpp itself references no symbol from this header; the SYCL spec also does not mandate that buffer-construction properties be visible from the accessor header. The include is moved into buffer.hpp (its natural home) and into detail/core.hpp so existing tests code that only includes accessor.hpp + uses `property::buffer::*` still works in the canonical aggregator path. * Drop `<algorithm>` from sycl/detail/property_list_base.hpp. The comment claimed the include was for `iter_swap`, but no algorithm call exists in the file. Removing it shaves ~30ms via the `<bits/ranges_*>` chain. * Move the body of `detail::cannot_be_called_on_host` from accessor.hpp to source/detail/accessor_impl.cpp and drop the unused `<iostream>` include from accessor.hpp. Adds one new exported symbol. The function was used in detail namespace and looks as debug infrastructure. * Remove the debug-only `friend std::ostream& operator<<` from accessor_iterator (no in-tree consumers) and drop its `<ostream>` include. * Split the inline `operator<<(std::ostream&, backend)` from sycl/backend_types.hpp into a new sycl/detail/backend_types_io.hpp so backend_types.hpp no longer pulls `<ostream>`. The full operator remains reachable through the canonical entry points (`<sycl/backend.hpp>` and `<sycl/detail/core.hpp>`), so user code using either of those continues to compile. The four ESIMD type-traits headers (bfloat16/tfloat32/bf8/hf8) were previously relying on a transitive `<ostream>` to define their own `operator<<`. They are now self-contained.