Simd array

[shawnl]

This is the first pull request that adds array access of vectors to zig, extracted from #2945. It implements

a[4] = 5;
i = a[4];

and that index also being a scalar integer. There will be another PR with @scatter and @gather and some vector array accesses that lower to these, and then another (not yet written) that implements pointers to vector elements the way gcc-9 does it.

[dimenus]

@shawnl I have a couple tests that currently fail.

these both fail to compile on vanilla zig, but segfault the compiler (ir.cpp:16111) in your pr:

test "vector literal" {
    const S = struct {
        fn doTheTest() void {
            var foo = @Vector(4, f32) { 1, 2, 3, 4 };
        }
    };
    S.doTheTest();
    //comptime S.doTheTest();
}

const Vec4Obj = struct {
    data: @Vector(4, f32),
};

test "vector in a struct" {
    const S = struct {
        fn doTheTest() void {
            var foo = Vec4Obj {
                .data = [_]f32 { 1, 2, 3, 4},
            };
        }
    };
    S.doTheTest();
    comptime S.doTheTest();
}

[shawnl]

@dimenus thanks. I fixed the first, and while I managed to make the second not crash the compiler, the assignment is not going anywhere because vectors do not use ResultLoc as they need to (I am working on this already for getting &a[4] to work).

[shawnl]

I guess actually supporting vector literals is a change in the language, but I don't see why we have to do it by casting from array (which was the original design and how I've always done it).

If we are going to prohibit vector literals, I think we should also prohibit array literals that use a array type (and only allow them with scalar types), so that this code no longer works:

test "array type literal" {
    const S = struct {
        const A = [4]f32;
        fn doTheTest() void {
            var foo = A{ 1, 2, 3, 4 };
        }
    };
    S.doTheTest();
    comptime S.doTheTest();
}