LLVM error with MMX types on i586

[alexcrichton]

Playground

Turns out if a function doesn't have the "mmx" feature enabled (like the entire i586-unknown-linux-gnu target) then LLVM will codegen an error if we use the mmx registers there. I... am not entirely sure what the best way is to fix this.

[alexcrichton]

Some text I wrote elsewhere relating to this issue:

---

The types available in the std::vendor module are intended to be generally
usable in any context. They'll only work well (in terms of being fast) in
appropriate contexts, however. The Rust compiler is expected though to at least
not provide unusual errors when using the types in the wrong context by
accident.

Unfortunately, though, this is not currently implemented for the MMX types!
Let's take a look at an example. For example this function works:

use std::vendor::*;
use std::mem;

pub fn foo() {
    let a: __m256i = unsafe { mem::zeroed() };
    let b = a;
}

Despite this function not enabling AVX nor the program enabling AVX globally
this function will operate as intended and won't actually produce any faults. In
other words, this is expected to produce a non-faulting program on all targets.

This is not true, however, for MMX types:

use std::vendor::*;
use std::mem;

pub fn foo() {
    let a: __m64 = unsafe { mem::zeroed() };
    let b = a;
}

When compiled for a target that doesn't support MMX by default (for example
i586-unknown-linux-gnu) this program will reveal a codegen error in LLVM. This
is largely a bug in the compiler due to how __m64 is represented,
unfortunately.

[workingjubilee]

Direct MMX support is no longer in scope for std::arch: #890