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

[trcrsired]

#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.

[trcrsired]

https://developercommunity.visualstudio.com/t/I-present-a-novel-calling-convention-nam/10433601?q=wincall

[dschuff]

What happens if you enable -mmulti-value? IIRC we use a different calling convention there, which passes structures in registers.

[trcrsired]

What happens if you enable -mmulti-value? IIRC we use a different calling convention there, which passes structures in registers.

Nothing changes. It is an ABI issue.

[tlively]

The full set of magic flags @dschuff is thinking of is -mmultivalue -Xclang -target-abi -Xclang experimental-mv.

[dschuff]

Ah thanks for that, I had forgotten that the feature enabler and the ABI were separate flags.