src/simd: improve isa detection

src/simd/index_of.zig

@@ -1,5 +1,6 @@
 const std = @import("std");
 const builtin = @import("builtin");
+const isa = @import("isa.zig");
 const aarch64 = @import("aarch64.zig");
 
 // Note this is a reimplementation of std.mem.indexOfScalar. The Zig stdlib
@@ -91,7 +92,7 @@ fn testIndexOf(func: *const IndexOf) !void {
 }
 
 test "indexOf neon" {
-    // TODO: use ISA detection here
+    if (comptime !isa.possible(.neon)) return error.SkipZigTest;
-    if (comptime builtin.cpu.arch != .aarch64) return error.SkipZigTest;
+    const set = isa.detect();
-    try testIndexOf(&indexOfNeon);
+    if (set.contains(.neon)) try testIndexOf(&indexOfNeon);
 }

src/simd/isa.zig

@@ -1,26 +1,97 @@
 const std = @import("std");
 const builtin = @import("builtin");
+const assert = std.debug.assert;
+const x86_64 = @import("x86_64.zig");
 
-/// Possible instruction set architectures for SIMD operations. These are
+/// Raw comptime entry of poissible ISA. The arch is the arch that the
-/// coarse grained and are targeted specifically so we can detect exactly
+/// ISA is even possible on (e.g. neon is only possible on aarch64) but
-/// what is available to us in Ghostty.
+/// the actual ISA may not be available at runtime.
-pub const ISA = enum {
+const Entry = struct {
-    scalar,
+    name: [:0]const u8,
-    neon,
+    arch: []const std.Target.Cpu.Arch = &.{},
-    avx2,
+};
 
-    /// Detect the available ISA at runtime. This will use comptime information
+const entries: []const Entry = &.{
-    /// as well to minimize the number of runtime checks.
+    .{ .name = "scalar" },
-    pub fn detect() ISA {
+    .{ .name = "neon", .arch = &.{.aarch64} },
-        return switch (builtin.cpu.arch) {
+    .{ .name = "avx2", .arch = &.{ .x86, .x86_64 } },
-            // Neon is mandatory on aarch64. No runtime checks necessary.
+};
-            .aarch64 => .neon,
+
-            .x86_64 => detectX86(),
+/// Enum of possible ISAs for our SIMD operations. Note that these are
-            else => .scalar,
+/// coarse-grained because they match possible implementations rather than
-        };
+/// a fine-grained packed struct of available CPU features.
+pub const ISA = isa: {
+    const EnumField = std.builtin.Type.EnumField;
+    var fields: [entries.len]EnumField = undefined;
+    for (entries, 0..) |entry, i| {
+        fields[i] = .{ .name = entry.name, .value = i };
     }
 
-    fn detectX86() ISA {
+    break :isa @Type(.{ .Enum = .{
+        .tag_type = std.math.IntFittingRange(0, entries.len - 1),
+        .fields = &fields,
+        .decls = &.{},
+        .is_exhaustive = true,
+    } });
+};
+
+/// A set of ISAs.
+pub const Set = std.EnumSet(ISA);
+
+/// Check if the given ISA is possible on the current target. This is
+/// available at comptime to help prevent invalid architectures from
+/// being used.
+pub fn possible(comptime isa: ISA) bool {
+    inline for (entries) |entry| {
+        if (std.mem.eql(u8, entry.name, @tagName(isa))) {
+            for (entry.arch) |arch| {
+                if (arch == builtin.cpu.arch) return true;
+            }
+
+            // If we have no valid archs then its always valid.
+            return entry.arch.len == 0;
+        }
+    }
+
+    unreachable;
+}
+
+/// Detect all possible ISAs at runtime.
+pub fn detect() Set {
+    var set: Set = .{};
+    set.insert(.scalar);
+    switch (builtin.cpu.arch) {
+        // Neon is mandatory on aarch64. No runtime checks necessary.
+        .aarch64 => set.insert(.neon),
+        .x86_64 => detectX86(&set),
+        else => {},
+    }
+
+    return set;
+}
+
+/// Returns the preferred ISA to use that is available.
+pub fn preferred(set: Set) ISA {
+    const order: []const ISA = &.{ .avx2, .neon, .scalar };
+
+    // We should have all of our ISAs present in order
+    comptime {
+        for (@typeInfo(ISA).Enum.fields) |field| {
+            const v = @field(ISA, field.name);
+            assert(std.mem.indexOfScalar(ISA, order, v) != null);
+        }
+    }
+
+    inline for (order) |isa| {
+        if (comptime possible(isa)) {
+            if (set.contains(isa)) return isa;
+        }
+    }
+
+    return .scalar;
+}
+
+fn detectX86(set: *Set) void {
     // NOTE: this is just some boilerplate to detect AVX2. We
     // can probably support earlier forms of SIMD such as plain
     // SSE, and we can definitely take advtange of later forms. This
@@ -28,85 +99,72 @@ pub const ISA = enum {
 
     // If we support less than 7 for the maximum leaf level then we
     // don't support any AVX instructions.
-        var leaf = X86.cpuid(0, 0);
+    var leaf = x86_64.cpuid(0, 0);
-        if (leaf.eax < 7) return .scalar;
+    if (leaf.eax < 7) return;
 
     // If we don't have xsave or avx, then we don't support anything.
-        leaf = X86.cpuid(1, 0);
+    leaf = x86_64.cpuid(1, 0);
     const has_xsave = hasBit(leaf.ecx, 27);
     const has_avx = hasBit(leaf.ecx, 28);
-        if (!has_xsave or !has_avx) return .scalar;
+    if (!has_xsave or !has_avx) return;
 
     // We require AVX save state in order to use AVX instructions.
-        const xcr0_eax = X86.getXCR0(); // requires xsave+avx
+    const xcr0_eax = x86_64.getXCR0(); // requires xsave+avx
-        const has_avx_save = hasMask(xcr0_eax, X86.XCR0_XMM | X86.XCR0_YMM);
+    const has_avx_save = hasMask(xcr0_eax, x86_64.XCR0_XMM | x86_64.XCR0_YMM);
-        if (!has_avx_save) return .scalar;
+    if (!has_avx_save) return;
 
     // Check for AVX2.
-        leaf = X86.cpuid(7, 0);
+    leaf = x86_64.cpuid(7, 0);
     const has_avx2 = hasBit(leaf.ebx, 5);
-        if (has_avx2) return .avx2;
+    if (has_avx2) set.insert(.avx2);
-
+}
-        return .scalar;
-    }
-};
-
-/// Constants and functions related to x86 and x86_64. Reference for this
-/// can be found in the Intel Architectures Software Developer's Manual,
-/// mostly around the cpuid instruction.
-const X86 = struct {
-    const XCR0_XMM = 0x02;
-    const XCR0_YMM = 0x04;
-    const XCR0_MASKREG = 0x20;
-    const XCR0_ZMM0_15 = 0x40;
-    const XCR0_ZMM16_31 = 0x80;
-
-    const CpuidLeaf = packed struct {
-        eax: u32,
-        ebx: u32,
-        ecx: u32,
-        edx: u32,
-    };
-
-    /// Wrapper around x86 and x86_64 `cpuid` in order to gather processor
-    /// and feature information. This is explicitly and specifically only
-    /// for x86 and x86_64.
-    fn cpuid(leaf_id: u32, subid: u32) CpuidLeaf {
-        var eax: u32 = undefined;
-        var ebx: u32 = undefined;
-        var ecx: u32 = undefined;
-        var edx: u32 = undefined;
-
-        asm volatile ("cpuid"
-            : [_] "={eax}" (eax),
-              [_] "={ebx}" (ebx),
-              [_] "={ecx}" (ecx),
-              [_] "={edx}" (edx),
-            : [_] "{eax}" (leaf_id),
-              [_] "{ecx}" (subid),
-        );
-
-        return .{ .eax = eax, .ebx = ebx, .ecx = ecx, .edx = edx };
-    }
-
-    // Read control register 0 (XCR0). Used to detect features such as AVX.
-    fn getXCR0() u32 {
-        return asm volatile (
-            \\ xor %%ecx, %%ecx
-            \\ xgetbv
-            : [_] "={eax}" (-> u32),
-            :
-            : "edx", "ecx"
-        );
-    }
-};
 
 /// Check if a bit is set at the given offset
-inline fn hasBit(input: u32, offset: u5) bool {
+pub inline fn hasBit(input: u32, offset: u5) bool {
     return (input >> offset) & 1 != 0;
 }
 
 /// Checks if a mask exactly matches the input
-inline fn hasMask(input: u32, mask: u32) bool {
+pub inline fn hasMask(input: u32, mask: u32) bool {
     return (input & mask) == mask;
 }
+
+test "detect" {
+    const testing = std.testing;
+    const set = detect();
+    try testing.expect(set.contains(.scalar));
+
+    switch (builtin.cpu.arch) {
+        .aarch64 => {
+            // Neon is always available on aarch64
+            try testing.expect(set.contains(.neon));
+            try testing.expect(!set.contains(.avx2));
+        },
+
+        else => {},
+    }
+}
+
+test "preferred" {
+    _ = preferred(detect());
+}
+
+test "possible" {
+    const testing = std.testing;
+    try testing.expect(possible(.scalar)); // always possible
+
+    // hardcode some other common realities
+    switch (builtin.cpu.arch) {
+        .aarch64 => {
+            try testing.expect(possible(.neon));
+            try testing.expect(!possible(.avx2));
+        },
+
+        .x86, .x86_64 => {
+            try testing.expect(!possible(.neon));
+            try testing.expect(possible(.avx2));
+        },
+
+        else => {},
+    }
+}

src/simd/main.zig

@@ -1,15 +1,17 @@
 const std = @import("std");
 
-const isa = @import("isa.zig");
+pub const isa = @import("isa.zig");
 const index_of = @import("index_of.zig");
-pub usingnamespace isa;
 pub usingnamespace index_of;
 
+// const utf8 = @import("utf8.zig");
+// pub usingnamespace utf8;
+
 pub fn main() !void {
     //std.log.warn("ISA={}", .{isa.ISA.detect()});
     const input = "1234567\x1b1234567\x1b";
     //const input = "1234567812345678";
-    _ = index_of.indexOf(input, 0x1B);
+    std.log.warn("result={any}", .{index_of.indexOf(input, 0x1B)});
 }
 
 test {

src/simd/x86_64.zig

@@ -0,0 +1,44 @@
+pub const XCR0_XMM = 0x02;
+pub const XCR0_YMM = 0x04;
+pub const XCR0_MASKREG = 0x20;
+pub const XCR0_ZMM0_15 = 0x40;
+pub const XCR0_ZMM16_31 = 0x80;
+
+pub const CpuidLeaf = packed struct {
+    eax: u32,
+    ebx: u32,
+    ecx: u32,
+    edx: u32,
+};
+
+/// Wrapper around x86 and x86_64 `cpuid` in order to gather processor
+/// and feature information. This is explicitly and specifically only
+/// for x86 and x86_64.
+pub fn cpuid(leaf_id: u32, subid: u32) CpuidLeaf {
+    var eax: u32 = undefined;
+    var ebx: u32 = undefined;
+    var ecx: u32 = undefined;
+    var edx: u32 = undefined;
+
+    asm volatile ("cpuid"
+        : [_] "={eax}" (eax),
+          [_] "={ebx}" (ebx),
+          [_] "={ecx}" (ecx),
+          [_] "={edx}" (edx),
+        : [_] "{eax}" (leaf_id),
+          [_] "{ecx}" (subid),
+    );
+
+    return .{ .eax = eax, .ebx = ebx, .ecx = ecx, .edx = edx };
+}
+
+// Read control register 0 (XCR0). Used to detect features such as AVX.
+pub fn getXCR0() u32 {
+    return asm volatile (
+        \\ xor %%ecx, %%ecx
+        \\ xgetbv
+        : [_] "={eax}" (-> u32),
+        :
+        : "edx", "ecx"
+    );
+}