performance issue for std::span on wasm C calling convention

#include<span>
#include<cstddef>

void baz(::std::size_t *ptr,::std::size_t n) noexcept;

void foo(::std::span<::std::size_t> sp) noexcept
{
	baz(sp.data(),sp.size());
}

void bar(::std::size_t *ptr,::std::size_t n) noexcept
{
	baz(ptr,n);
}

_Z3fooNSt3__14spanImLm4294967295EEE:    # @_Z3fooNSt3__14spanImLm4294967295EEE
	.functype	_Z3fooNSt3__14spanImLm4294967295EEE (i32) -> ()
# %bb.0:
	local.get	0
	i32.load	0
	local.get	0
	i32.load	4
	call	_Z3bazPmm
                                        # fallthrough-return
	end_function
                                        # -- End function
	.section	.text._Z3barPmm,"",@
	.hidden	_Z3barPmm                       # -- Begin function _Z3barPmm
	.globl	_Z3barPmm
	.type	_Z3barPmm,@function
_Z3barPmm:                              # @_Z3barPmm
	.functype	_Z3barPmm (i32, i32) -> ()
# %bb.0:
	local.get	0
	local.get	1
	call	_Z3bazPmm
                                        # fallthrough-return
	end_function

std::span is passed by memory, not by registers on wasm. which is extremely slow.

This is going to encourage people to overuse pointers. Please change the calling convention here. Also, the return value has the same issue.

Start with the WebAssembly calling-convention conventions relevant to the C++ std::span example and compare the shown foo and bar assembly. Determine how span parameters and return values should be passed, then document the convention and validate it against generated wasm for both cases.