simd.ccm File Reference

SIMD compressed pointers and heap-relative gathers. More...

#include <native/attributes.h>
#include <native/targets.h>
#include <array>
#include <cassert>
#include <bit>
#include <concepts>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <span>
#include <type_traits>
#include <utility>
import native;
import jam:heap;

Include dependency graph for simd.ccm:

Go to the source code of this file.

Classes
struct	jam::tracer< native::simd< ptr< T >, N, A > >
	Trace actual pointer lanes, including null fields, in register order. More...
struct	jam::tracer< native::wide< V, K > >
	Trace each register's fields in the enclosing allocation. More...

Functions
template<class T, std::size_t N, native::isa<> A>
void	jam::unsafe_assign (native::simd< ptr< T >, N, A > &target, native::simd< ptr< T >, N, A > const &source) noexcept
	Copy SIMD pointer bits without registering old-to-young slots.
template<class T, std::size_t N, native::isa<> A>
void	jam::assign (native::simd< ptr< T >, N, A > &target, native::simd< ptr< T >, N, A > const &source) noexcept
	Bulk-copy a register and register its old-to-young slots.
template<class T, std::size_t N, native::isa<> A, std::size_t K>
void	jam::unsafe_assign (native::wide< native::simd< ptr< T >, N, A >, K > &target, native::wide< native::simd< ptr< T >, N, A >, K > const &source) noexcept
	Copy a pointer register pack without registration; caller owns its barriers.
template<class T, std::size_t N, native::isa<> A, std::size_t K>
void	jam::assign (native::wide< native::simd< ptr< T >, N, A >, K > &target, native::wide< native::simd< ptr< T >, N, A >, K > const &source) noexcept
	Copy a pointer register pack with one bulk barrier per register.
template<class T, class U, class Owner, std::size_t N, native::isa<> A>
native::simd< U, N, A >	jam::gather (native::simd< ptr< T >, N, A > const &nodes, U Owner::*member, typename native::simd< ptr< T >, N, A >::mask_type active) noexcept
	Gather a member from selected lanes; inactive lanes return zero or null.
template<class T, class U, class Owner, std::size_t N, native::isa<> A>
native::simd< U, N, A >	jam::gather (native::simd< ptr< T >, N, A > const &nodes, U Owner::*member) noexcept
	Gather with the mask nodes != nullptr; inactive lanes return zero or null.
template<class T, class U, class Owner, std::size_t N, native::isa<> A, std::size_t K>
native::wide< native::simd< U, N, A >, K >	jam::gather (native::wide< native::simd< ptr< T >, N, A >, K > const &nodes, U Owner::*member, native::wide< typename native::simd< ptr< T >, N, A >::mask_type, K > const &active) noexcept
	Gather each register using the supplied masks, preserving lane and register order.
template<class T, class U, class Owner, std::size_t N, native::isa<> A, std::size_t K>
native::wide< native::simd< U, N, A >, K >	jam::gather (native::wide< native::simd< ptr< T >, N, A >, K > const &nodes, U Owner::*member) noexcept
	Compute all nonnull register masks, then gather the requested member.

Detailed Description

SIMD compressed pointers and heap-relative gathers.

Definition in file simd.ccm.

Function Documentation

gather() [1/4]

template<class T, class U, class Owner, std::size_t N, native::isa<> A>

native::simd< U, N, A > jam::gather ( native::simd< ptr< T >, N, A > const & nodes, U Owner::* member )

exportnoexcept

Gather with the mask nodes != nullptr; inactive lanes return zero or null.

Shares the explicit-mask overload's heap, member and result-shape requirements.

gather() [2/4]

template<class T, class U, class Owner, std::size_t N, native::isa<> A>

native::simd< U, N, A > jam::gather ( native::simd< ptr< T >, N, A > const & nodes, U Owner::* member, typename native::simd< ptr< T >, N, A >::mask_type active )

exportnoexcept

Gather a member from selected lanes; inactive lanes return zero or null.

Precondition

The current heap is bound. Active pointers name live T records in it.

member is nonnull; inheritance is nonvirtual and native supports the result shape.

On x86, Clang's flat data-member-pointer ABI supplies the displacement: ptrdiff_t on Unix, int32_t on Windows. Extended MS representations are rejected. The offset is folded into the gather base; no sample object or mask extraction is needed to recover it. Pointer results outside the heap expire at collection.

gather() [3/4]

template<class T, class U, class Owner, std::size_t N, native::isa<> A, std::size_t K>

native::wide< native::simd< U, N, A >, K > jam::gather ( native::wide< native::simd< ptr< T >, N, A >, K > const & nodes, U Owner::* member )

exportnoexcept

Compute all nonnull register masks, then gather the requested member.

Each register obeys the SIMD gather's requirements.

gather() [4/4]

template<class T, class U, class Owner, std::size_t N, native::isa<> A, std::size_t K>

native::wide< native::simd< U, N, A >, K > jam::gather ( native::wide< native::simd< ptr< T >, N, A >, K > const & nodes, U Owner::* member, native::wide< typename native::simd< ptr< T >, N, A >::mask_type, K > const & active )

exportnoexcept

Gather each register using the supplied masks, preserving lane and register order.

Each register obeys the SIMD gather's requirements.