metal2vulkan 0.1.0

Translate Metal AIR / LLVM IR to Vulkan SPIR-V with a native Rust emitter
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164

use super::*;
mod gep_linearize;
mod phi_compare;
mod phi_merge;
mod phi_provenance;
mod raw_select;
mod select;
mod select_gep;

/// Whether `ty` is a plain scalar (the "part" side of the whole-vs-part upgrade): an integer, a
/// float of any width, or a bool. Composites (vector/array/struct/pointer) and unresolved aliases
/// are not scalars.
pub(super) fn is_scalar_pointee(ty: &LlType) -> bool {
    matches!(
        ty,
        LlType::Int(_) | LlType::Float | LlType::Half | LlType::BFloat | LlType::Bool
    )
}

/// Bit width of a scalar `LlType`, or `None` for a non-scalar / width-less type. Used by the
/// M-A2(a) reinterpret upgrade to require the carrier be the SAME width as the recorded pointee
/// (a legal bit-reinterpret). `Bool` has no defined storage width here, so it is excluded.
fn scalar_bit_width(ty: &LlType) -> Option<u32> {
    match ty {
        LlType::Float => Some(32),
        LlType::Half | LlType::BFloat => Some(16),
        LlType::Int(w) => Some(*w),
        _ => None,
    }
}

/// M-A2(b) whole-vs-part shape test: whether the (already-resolved) `carrier` is the WHOLE composite
/// `Vector(S,N)` / `[N x S]` whose element is exactly the recorded `scalar_pointee` `S` — a pure
/// part→whole granularity widening. Any other carrier shape (different element, struct, nested, scalar)
/// is out of the family.
fn whole_part_widens(carrier: &LlType, scalar_pointee: &LlType) -> bool {
    match carrier {
        LlType::Vector(elem, _) | LlType::Array(elem, _) => elem.as_ref() == scalar_pointee,
        _ => false,
    }
}

/// M-A2(a) reinterpret shape test: whether the (already-resolved) `carrier` is a DIFFERENT scalar of
/// the SAME bit width as the recorded `scalar_pointee` (a legal bit-reinterpret, e.g.
/// `Float(32)`↔`Int(32)`). NOTE this family is proven UNSOUND to prefer (dead-end #14) — the test only
/// gates the default-off diagnostic flag.
fn reinterp_compatible(carrier: &LlType, scalar_pointee: &LlType) -> bool {
    is_scalar_pointee(carrier)
        && carrier != scalar_pointee
        && match (scalar_bit_width(carrier), scalar_bit_width(scalar_pointee)) {
            (Some(cw), Some(pw)) => cw == pw,
            _ => false,
        }
}

fn pointer_arithmetic_access_chain_op_for_storage(
    storage: StorageClass,
    base_is_indexed_container: bool,
    pointee: &LlType,
    indices: &[TypedValue],
) -> Op {
    let scalar_pointee = !matches!(pointee, LlType::Array(_, _) | LlType::Struct(_));
    if scalar_pointee
        && ptr_access_chain_allowed_storage(storage)
        && !base_is_indexed_container
        && indices.first().and_then(|idx| const_index(Some(idx))) != Some(0)
    {
        Op::PtrAccessChain
    } else {
        Op::InBoundsAccessChain
    }
}

#[cfg(test)]
mod pointee_upgrade_tests {
    //! Coverage for the M-A2 pointee-upgrade shape predicates (`is_scalar_pointee`,
    //! `scalar_bit_width`, `whole_part_widens`, `reinterp_compatible`). These decide whether the
    //! default-off byte→real / whole-vs-part / reinterpret flags upgrade a pointer's pointee; getting
    //! the shape test wrong would silently mis-upgrade (or fail to upgrade) at the `pointer_pointee_for_value`
    //! seam. Pure functions on already-resolved `LlType`s, so no emitter state is needed.
    use super::{is_scalar_pointee, reinterp_compatible, scalar_bit_width, whole_part_widens};
    use crate::native::ir::LlType;

    fn vec(elem: LlType, n: u32) -> LlType {
        LlType::Vector(Box::new(elem), n)
    }
    fn arr(elem: LlType, n: u32) -> LlType {
        LlType::Array(Box::new(elem), n)
    }

    #[test]
    fn scalars_are_scalar_composites_are_not() {
        for s in [
            LlType::Float,
            LlType::Half,
            LlType::BFloat,
            LlType::Bool,
            LlType::Int(32),
            LlType::Int(16),
            LlType::Int(8),
        ] {
            assert!(is_scalar_pointee(&s), "{s:?} should be scalar");
        }
        for c in [
            vec(LlType::Float, 4),
            arr(LlType::Float, 8),
            LlType::Struct(vec![LlType::Half, LlType::Half]),
            LlType::Ptr(1),
            LlType::Void,
            LlType::Named("x".into()),
        ] {
            assert!(!is_scalar_pointee(&c), "{c:?} should not be scalar");
        }
    }

    #[test]
    fn bit_widths() {
        assert_eq!(scalar_bit_width(&LlType::Float), Some(32));
        assert_eq!(scalar_bit_width(&LlType::Int(32)), Some(32));
        assert_eq!(scalar_bit_width(&LlType::Half), Some(16));
        assert_eq!(scalar_bit_width(&LlType::BFloat), Some(16));
        assert_eq!(scalar_bit_width(&LlType::Int(16)), Some(16));
        assert_eq!(scalar_bit_width(&LlType::Int(8)), Some(8));
        // Bool has no defined storage width here; composites are width-less.
        assert_eq!(scalar_bit_width(&LlType::Bool), None);
        assert_eq!(scalar_bit_width(&vec(LlType::Float, 4)), None);
        assert_eq!(scalar_bit_width(&LlType::Ptr(1)), None);
    }

    #[test]
    fn whole_part_upgrades_only_matching_element_composite() {
        // Vector/array WHOSE ELEMENT is the recorded scalar → widen.
        assert!(whole_part_widens(&vec(LlType::Float, 4), &LlType::Float));
        assert!(whole_part_widens(&arr(LlType::Half, 8), &LlType::Half));
        assert!(whole_part_widens(
            &vec(LlType::Int(32), 2),
            &LlType::Int(32)
        ));
        // Element mismatch, scalar carrier, or struct → do NOT widen.
        assert!(!whole_part_widens(&vec(LlType::Int(32), 4), &LlType::Float));
        assert!(!whole_part_widens(&LlType::Float, &LlType::Float));
        assert!(!whole_part_widens(
            &LlType::Struct(vec![LlType::Float]),
            &LlType::Float
        ));
    }

    #[test]
    fn reinterp_requires_same_width_different_scalar() {
        // Same width, different kind → reinterpret-compatible.
        assert!(reinterp_compatible(&LlType::Int(32), &LlType::Float));
        assert!(reinterp_compatible(&LlType::Float, &LlType::Int(32)));
        assert!(reinterp_compatible(&LlType::Int(16), &LlType::Half));
        assert!(reinterp_compatible(&LlType::Half, &LlType::Int(16)));
        // Same type → nothing to reinterpret.
        assert!(!reinterp_compatible(&LlType::Float, &LlType::Float));
        // Different width → not a legal reinterpret.
        assert!(!reinterp_compatible(&LlType::Int(16), &LlType::Float));
        assert!(!reinterp_compatible(&LlType::Half, &LlType::Int(32)));
        // Composite carrier / Bool → excluded (that is whole-vs-part / width-less).
        assert!(!reinterp_compatible(&vec(LlType::Float, 4), &LlType::Float));
        assert!(!reinterp_compatible(&LlType::Bool, &LlType::Int(8)));
    }
}