sx/src/codegen.zig

const std = @import("std");
const ast = @import("ast.zig");
const Node = ast.Node;
const Span = ast.Span;
const llvm = @import("llvm_api.zig");
const c = llvm.c;
const types = @import("types.zig");
const Type = types.Type;
const Builtins = @import("builtins.zig").Builtins;
const Parser = @import("parser.zig").Parser;
const errors = @import("errors.zig");
const sema = @import("sema.zig");
const comptime_mod = @import("comptime.zig");

pub const TargetConfig = struct {
    /// Target triple (e.g. "aarch64-apple-darwin"). Null = host default.
    triple: ?[*:0]const u8 = null,
    /// CPU name (e.g. "generic", "apple-m1"). Null = "generic".
    cpu: ?[*:0]const u8 = null,
    /// CPU features string (e.g. "+avx2"). Null = "".
    features: ?[*:0]const u8 = null,
    /// Optimization level.
    opt_level: OptLevel = .default,
    /// Library search paths (-L flags).
    lib_paths: []const []const u8 = &.{},
    /// Output path override.
    output_path: ?[]const u8 = null,
    /// Linker command (null = "cc" on Unix, "link.exe" on Windows).
    linker: ?[]const u8 = null,
    /// Sysroot for cross-compilation (passed as --sysroot to linker).
    sysroot: ?[]const u8 = null,

    pub const OptLevel = enum {
        none,
        less,
        default,
        aggressive,

        pub fn toLLVM(self: OptLevel) c.LLVMCodeGenOptLevel {
            return switch (self) {
                .none => c.LLVMCodeGenLevelNone,
                .less => c.LLVMCodeGenLevelLess,
                .default => c.LLVMCodeGenLevelDefault,
                .aggressive => c.LLVMCodeGenLevelAggressive,
            };
        }
    };

    /// Check if target triple indicates aarch64/arm64 (runtime check, not comptime).
    pub fn isAarch64(self: TargetConfig) bool {
        return self.tripleHasPrefix("aarch64", "arm64");
    }

    /// Check if target triple indicates x86_64/x86-64.
    pub fn isX86_64(self: TargetConfig) bool {
        return self.tripleHasPrefix("x86_64", "x86-64");
    }

    /// Check if target triple indicates Windows (contains "windows" or "win32").
    pub fn isWindows(self: TargetConfig) bool {
        return self.tripleContains("windows") or self.tripleContains("win32");
    }

    fn tripleHasPrefix(self: TargetConfig, prefix1: []const u8, prefix2: []const u8) bool {
        if (self.triple) |t| {
            const span = std.mem.span(t);
            return std.mem.startsWith(u8, span, prefix1) or std.mem.startsWith(u8, span, prefix2);
        }
        const dt = c.LLVMGetDefaultTargetTriple();
        defer c.LLVMDisposeMessage(dt);
        const span = std.mem.span(dt);
        return std.mem.startsWith(u8, span, prefix1) or std.mem.startsWith(u8, span, prefix2);
    }

    fn tripleContains(self: TargetConfig, needle: []const u8) bool {
        if (self.triple) |t| {
            return std.mem.indexOf(u8, std.mem.span(t), needle) != null;
        }
        const dt = c.LLVMGetDefaultTargetTriple();
        defer c.LLVMDisposeMessage(dt);
        return std.mem.indexOf(u8, std.mem.span(dt), needle) != null;
    }

    pub fn getCpu(self: TargetConfig) [*:0]const u8 {
        return self.cpu orelse "generic";
    }

    pub fn getFeatures(self: TargetConfig) [*:0]const u8 {
        return self.features orelse "";
    }

    pub fn getLinker(self: TargetConfig) []const u8 {
        return self.linker orelse "cc";
    }
};

pub const CodeGen = struct {
    context: c.LLVMContextRef,
    module: c.LLVMModuleRef,
    builder: c.LLVMBuilderRef,
    allocator: std.mem.Allocator,

    // Symbol table: maps variable names to their alloca pointers
    named_values: std.StringHashMap(NamedValue),
    // Enum type registry: maps enum name to variant list
    enum_types: std.StringHashMap([]const []const u8),
    // Type alias registry: maps alias name to target type name
    type_aliases: std.StringHashMap([]const u8),
    // Struct type registry: maps struct name to field info + LLVM type
    struct_types: std.StringHashMap(StructInfo),
    // Union type registry: maps union name to variant info + LLVM type
    union_types: std.StringHashMap(UnionInfo),
    // Built-in functions (printf, etc.)
    builtins: ?Builtins,
    // Current function being generated (for alloca insertion)
    current_function: c.LLVMValueRef,
    // Return type of the current function being generated
    current_return_type: Type = .void_type,
    // Scope save stack: each entry records shadowed names to restore on scope exit
    scope_saves: std.ArrayList(std.ArrayList(ScopeEntry)),
    // Defer stack: parallel to scope_saves, each entry holds deferred expressions
    defer_stack: std.ArrayList(std.ArrayList(*Node)),
    // Compile-time globals: maps name to global variable info for #run results
    comptime_globals: std.StringHashMap(ComptimeGlobal),
    // Top-level #run expressions for side effects only
    comptime_side_effects: std.ArrayList(*Node),
    // Generic function templates: maps name to AST for deferred monomorphization
    generic_templates: std.StringHashMap(GenericTemplate),
    // Instantiated generic functions: maps mangled name to LLVM function
    generic_instances: std.StringHashMap(c.LLVMValueRef),
    // Active type parameter bindings during generic instantiation (null when not instantiating)
    type_param_bindings: ?std.StringHashMap(Type) = null,
    // Active value parameter bindings during generic struct instantiation
    value_param_bindings: ?std.StringHashMap(i64) = null,
    // Active comptime param AST nodes during generic function instantiation (for #insert substitution)
    comptime_param_nodes: ?std.StringHashMap(*Node) = null,
    // Generic struct templates: maps name to AST for deferred instantiation
    generic_struct_templates: std.StringHashMap(GenericStructTemplate),
    // Known namespace names (for import resolution)
    namespaces: std.StringHashMap(void),
    // Functions declared with #builtin (only available when imported)
    builtin_functions: std.StringHashMap(void),
    // Function signatures: maps function name to display signature (e.g. "test :: () -> s32")
    fn_signatures: std.StringHashMap([]const u8),
    // Active namespace during body generation of imported modules
    current_namespace: ?[]const u8 = null,
    // Diagnostics list (optional, for structured error reporting)
    diagnostics: ?*errors.DiagnosticList = null,
    // Current source span (set at genExpr/genStmt/genExprAsType entry)
    current_span: Span = .{ .start = 0, .end = 0 },
    // Loop context: break/continue target basic blocks (null when not in a loop)
    loop_break_bb: c.LLVMBasicBlockRef = null,
    loop_continue_bb: c.LLVMBasicBlockRef = null,
    // Sema result (optional, for type-aware comptime evaluation)
    sema_result: ?*const sema.SemaResult = null,
    // Root declarations from the AST (for VM on-demand function compilation)
    root_decls: []const *Node = &.{},
    // Cached LLVM struct type for string slices {ptr, i32}
    string_struct_type: c.LLVMTypeRef = null,
    // Cached LLVM struct type for Any {i32 tag, i64 value}
    any_struct_type: c.LLVMTypeRef = null,
    // Dynamic type ID assignment for Any tags (named types get unique IDs starting from 7)
    any_type_id_map: std.StringHashMap(u64),
    next_any_type_id: u64 = 7,
    // Cache of auto-generated to_string functions for complex types
    // Variadic function info: maps function name to variadic metadata
    variadic_functions: std.StringHashMap(VariadicInfo),
    // Maps function name to resolved sx parameter types (for accurate type conversion at call sites)
    fn_param_types: std.StringHashMap([]const Type),
    // Enriched Any type entries: maps type name to tag + category + sx type
    any_type_entries: std.StringHashMap(AnyTypeEntry),
    // Current match arm type entries (set during category match arm body generation)
    current_match_tags: ?[]const u64 = null,
    // Functions deferred to compile after all types are registered (e.g. any_to_string)
    deferred_fn_bodies: std.ArrayList(DeferredFn),
    // Libraries to link against (from #library directives)
    foreign_libraries: std.ArrayList([]const u8),
    // Set of foreign function names (for ABI lowering at call sites)
    foreign_fns: std.StringHashMap(void),
    // Global mutable variables (from top-level var_decl, e.g. function pointers loaded at runtime)
    global_mutable_vars: std.StringHashMap(NamedValue),
    // Declared return types for non-generic functions (preserves signedness lost by LLVM round-trip)
    function_return_types: std.StringHashMap(Type),
    // Target configuration (triple, cpu, opt level, lib paths, linker)
    target_config: TargetConfig = .{},

    const DeferredFn = struct {
        fd: ast.FnDecl,
        name: []const u8, // qualified name (may differ from fd.name for namespaced functions)
        namespace: ?[]const u8 = null,
    };

    const TypeCategory = enum {
        struct_cat,
        enum_cat,
        union_cat,
        vector_cat,
        array_cat,
        slice_cat,
        pointer_cat,
    };

    const AnyTypeEntry = struct {
        tag_id: u64,
        category: TypeCategory,
        sx_type: Type,
    };

    const VariadicInfo = struct {
        fixed_param_count: u32, // number of non-variadic params
        element_type_name: []const u8, // element type of the variadic slice (e.g. "s32")
    };

    const GenericTemplate = struct {
        fd: ast.FnDecl,
    };

    const GenericStructTemplate = struct {
        sd: ast.StructDecl,
    };

    const ComptimeGlobal = struct {
        global: c.LLVMValueRef, // LLVM global variable
        ty: Type, // sx type
        expr: *Node, // the inner expression to JIT-evaluate
        is_resolved: bool = false, // true if initializer already set (no JIT needed)
    };

    const StructInfo = struct {
        field_names: []const []const u8,
        field_types: []const Type,
        field_defaults: []const ?*Node,
        llvm_type: c.LLVMTypeRef,
        display_name: ?[]const u8 = null, // pretty name for generic instances
        type_param_names: []const []const u8 = &.{}, // original type param names (e.g. ["T"])
        type_param_types: []const Type = &.{}, // resolved types (e.g. [s32])
        template_name: ?[]const u8 = null, // original template name (e.g. "List")
    };

    const UnionInfo = struct {
        variant_names: []const []const u8,
        variant_types: []const Type, // void_type for void variants
        llvm_type: c.LLVMTypeRef, // { i32, [max_payload_size x i8] }
        max_payload_size: u64,
    };

    // Scope stack entry: records what a name mapped to before being shadowed
    const ScopeEntry = struct {
        name: []const u8,
        prev: ?NamedValue, // null = name didn't exist before this scope
    };

    const NamedValue = struct {
        ptr: c.LLVMValueRef, // alloca pointer
        ty: Type, // sx type
    };

    pub fn init(allocator: std.mem.Allocator, module_name: [*:0]const u8, target_config: TargetConfig) CodeGen {
        const ctx = c.LLVMContextCreate();
        const module = c.LLVMModuleCreateWithNameInContext(module_name, ctx);
        const builder = c.LLVMCreateBuilderInContext(ctx);

        // Set target triple on module so it appears in IR output
        if (target_config.triple) |t| {
            c.LLVMSetTarget(module, t);
        } else {
            const default_triple = c.LLVMGetDefaultTargetTriple();
            c.LLVMSetTarget(module, default_triple);
            c.LLVMDisposeMessage(default_triple);
        }
        return .{
            .context = ctx,
            .module = module,
            .builder = builder,
            .allocator = allocator,
            .named_values = std.StringHashMap(NamedValue).init(allocator),
            .enum_types = std.StringHashMap([]const []const u8).init(allocator),
            .type_aliases = std.StringHashMap([]const u8).init(allocator),
            .struct_types = std.StringHashMap(StructInfo).init(allocator),
            .union_types = std.StringHashMap(UnionInfo).init(allocator),
            .builtins = null,
            .current_function = null,
            .scope_saves = std.ArrayList(std.ArrayList(ScopeEntry)).empty,
            .defer_stack = std.ArrayList(std.ArrayList(*Node)).empty,
            .comptime_globals = std.StringHashMap(ComptimeGlobal).init(allocator),
            .comptime_side_effects = std.ArrayList(*Node).empty,
            .generic_templates = std.StringHashMap(GenericTemplate).init(allocator),
            .generic_instances = std.StringHashMap(c.LLVMValueRef).init(allocator),
            .generic_struct_templates = std.StringHashMap(GenericStructTemplate).init(allocator),
            .namespaces = std.StringHashMap(void).init(allocator),
            .builtin_functions = std.StringHashMap(void).init(allocator),
            .fn_signatures = std.StringHashMap([]const u8).init(allocator),
            .variadic_functions = std.StringHashMap(VariadicInfo).init(allocator),
            .fn_param_types = std.StringHashMap([]const Type).init(allocator),
            .any_type_id_map = std.StringHashMap(u64).init(allocator),
            .any_type_entries = std.StringHashMap(AnyTypeEntry).init(allocator),
            .deferred_fn_bodies = std.ArrayList(DeferredFn).empty,
            .foreign_libraries = std.ArrayList([]const u8).empty,
            .foreign_fns = std.StringHashMap(void).init(allocator),
            .global_mutable_vars = std.StringHashMap(NamedValue).init(allocator),
            .function_return_types = std.StringHashMap(Type).init(allocator),
            .target_config = target_config,
        };
    }

    pub fn deinit(self: *CodeGen) void {
        self.named_values.deinit();
        self.enum_types.deinit();
        self.type_aliases.deinit();
        self.struct_types.deinit();
        self.union_types.deinit();
        self.comptime_globals.deinit();
        self.generic_templates.deinit();
        self.generic_instances.deinit();
        self.generic_struct_templates.deinit();
        self.namespaces.deinit();
        self.builtin_functions.deinit();
        self.fn_signatures.deinit();
        self.variadic_functions.deinit();
        self.any_type_id_map.deinit();
        self.any_type_entries.deinit();
        self.deferred_fn_bodies.deinit(self.allocator);
        self.foreign_libraries.deinit(self.allocator);
        self.foreign_fns.deinit();
        c.LLVMDisposeBuilder(self.builder);
        c.LLVMDisposeModule(self.module);
        c.LLVMContextDispose(self.context);
    }

    fn emitError(self: *CodeGen, msg: []const u8) error{CodeGenError} {
        if (self.diagnostics) |diags| diags.add(.err, msg, self.current_span);
        return error.CodeGenError;
    }

    fn emitErrorFmt(self: *CodeGen, comptime fmt: []const u8, args: anytype) error{CodeGenError} {
        if (self.diagnostics) |diags| diags.addFmt(.err, self.current_span, fmt, args);
        return error.CodeGenError;
    }

    /// Build an alloca in the entry block of the current function so that
    /// stack space is reserved once, not on every loop iteration.
    fn buildEntryBlockAlloca(self: *CodeGen, ty: c.LLVMTypeRef, name: [*:0]const u8) c.LLVMValueRef {
        const entry_bb = c.LLVMGetEntryBasicBlock(self.current_function);
        const first_instr = c.LLVMGetFirstInstruction(entry_bb);
        const tmp_builder = c.LLVMCreateBuilderInContext(self.context);
        defer c.LLVMDisposeBuilder(tmp_builder);
        if (first_instr != null) {
            c.LLVMPositionBuilderBefore(tmp_builder, first_instr);
        } else {
            c.LLVMPositionBuilderAtEnd(tmp_builder, entry_bb);
        }
        return c.LLVMBuildAlloca(tmp_builder, ty, name);
    }

    pub fn typeToLLVM(self: *CodeGen, ty: Type) c.LLVMTypeRef {
        return switch (ty) {
            .signed => |w| c.LLVMIntTypeInContext(self.context, w),
            .unsigned => |w| c.LLVMIntTypeInContext(self.context, w),
            .f32 => c.LLVMFloatTypeInContext(self.context),
            .f64 => c.LLVMDoubleTypeInContext(self.context),
            .void_type => c.LLVMVoidTypeInContext(self.context),
            .boolean => c.LLVMInt1TypeInContext(self.context),
            .string_type, .slice_type => self.getStringStructType(), // slices use same {ptr, i32} layout
            .enum_type => c.LLVMInt64TypeInContext(self.context),
            .struct_type => |name| if (self.struct_types.get(name)) |info| info.llvm_type else unreachable,
            .union_type => |name| if (self.union_types.get(name)) |info| info.llvm_type else unreachable,
            .array_type => |info| {
                const elem_ty = Type.fromName(info.element_name) orelse unreachable;
                return c.LLVMArrayType2(self.typeToLLVM(elem_ty), info.length);
            },
            .vector_type => |info| {
                const elem_ty = Type.fromName(info.element_name) orelse unreachable;
                return c.LLVMVectorType(self.typeToLLVM(elem_ty), info.length);
            },
            .pointer_type, .many_pointer_type, .function_type => c.LLVMPointerTypeInContext(self.context, 0),
            .any_type => self.getAnyStructType(),
            .meta_type => c.LLVMPointerTypeInContext(self.context, 0),
        };
    }

    fn getAnyStructType(self: *CodeGen) c.LLVMTypeRef {
        if (self.any_struct_type) |t| return t;
        var field_types = [_]c.LLVMTypeRef{
            c.LLVMInt64TypeInContext(self.context), // type tag
            c.LLVMInt64TypeInContext(self.context), // value (fits all primitives)
        };
        self.any_struct_type = c.LLVMStructTypeInContext(self.context, &field_types, 2, 0);
        return self.any_struct_type.?;
    }

    /// Type tag constants for Any type (builtins: 0-6, Type: 10, named types: 7+ dynamic)
    const ANY_TAG_VOID: u64 = 0;
    const ANY_TAG_BOOL: u64 = 1;
    const ANY_TAG_S32: u64 = 2;
    const ANY_TAG_S64: u64 = 3;
    const ANY_TAG_F32: u64 = 4;
    const ANY_TAG_F64: u64 = 5;
    const ANY_TAG_STRING: u64 = 6;
    const ANY_TAG_TYPE: u64 = 10;

    /// Get or assign a unique type ID for a named type (struct, enum, union, vector, array).
    /// IDs start at 7 and are assigned dynamically per compilation.
    /// Also populates `any_type_entries` with category and type info.
    fn getAnyTypeId(self: *CodeGen, name: []const u8, sx_type: Type) !u64 {
        const gop = try self.any_type_id_map.getOrPut(name);
        if (!gop.found_existing) {
            gop.value_ptr.* = self.next_any_type_id;
            self.next_any_type_id += 1;
            // Skip over reserved slot 10 (ANY_TAG_TYPE)
            if (self.next_any_type_id == ANY_TAG_TYPE) self.next_any_type_id += 1;

            // Determine category from the sx type
            const category: TypeCategory = switch (sx_type) {
                .struct_type => .struct_cat,
                .enum_type => .enum_cat,
                .union_type => .union_cat,
                .vector_type => .vector_cat,
                .array_type => .array_cat,
                .slice_type => .slice_cat,
                .pointer_type, .many_pointer_type => .pointer_cat,
                else => .struct_cat, // fallback
            };
            try self.any_type_entries.put(name, .{
                .tag_id = gop.value_ptr.*,
                .category = category,
                .sx_type = sx_type,
            });
        }
        return gop.value_ptr.*;
    }

    /// Check if a function should have its body compilation deferred until after all types are registered.
    /// Functions with non-variadic Any parameters (like any_to_string) use type-based match expressions
    /// that need all types registered before compilation.
    fn shouldDeferFnBody(fd: ast.FnDecl) bool {
        for (fd.params) |param| {
            if (!param.is_variadic and param.type_expr.data == .type_expr and
                std.mem.eql(u8, param.type_expr.data.type_expr.name, "Any"))
            {
                return true;
            }
        }
        return false;
    }

    /// Pre-register a type in the Any type system so category matching (case slice:, case array:, etc.)
    /// works in any_to_string even before buildAnyValue is called for this type.
    fn preRegisterAnyType(self: *CodeGen, sx_type: Type) !void {
        switch (sx_type) {
            .struct_type => |name| {
                _ = try self.getAnyTypeId(name, sx_type);
                // Recursively register struct field types
                if (self.struct_types.get(name)) |info| {
                    for (info.field_types) |ft| {
                        try self.preRegisterAnyType(ft);
                    }
                }
            },
            .enum_type => |name| _ = try self.getAnyTypeId(name, sx_type),
            .union_type => |name| _ = try self.getAnyTypeId(name, sx_type),
            .vector_type => |info| _ = try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "vec[{d}]{s}", .{ info.length, info.element_name }), sx_type),
            .array_type => |info| _ = try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "[{d}]{s}", .{ info.length, info.element_name }), sx_type),
            .slice_type => |info| _ = try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "[]{s}", .{info.element_name}), sx_type),
            .pointer_type => |info| _ = try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "*{s}", .{info.pointee_name}), sx_type),
            .many_pointer_type => |info| _ = try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "[*]{s}", .{info.element_name}), sx_type),
            else => {},
        }
    }

    /// Build an Any value { tag: i32, value: i64 } from a typed LLVM value.
    /// Small values (ints, floats, bools, enums) are stored inline in the i64.
    /// Complex values (strings, structs, unions) are stored via pointer (alloca + ptr-to-int).
    fn buildAnyValue(self: *CodeGen, val: c.LLVMValueRef, in_ty: Type) !c.LLVMValueRef {
        const any_ty = self.getAnyStructType();
        const i64_ty = c.LLVMInt64TypeInContext(self.context);
        const undef = c.LLVMGetUndef(any_ty);

        // []u8 boxes as string (same repr, same Any tag)
        const ty: Type = if (in_ty.isSlice() and std.mem.eql(u8, in_ty.slice_type.element_name, "u8"))
            .string_type
        else
            in_ty;

        // Determine tag
        const tag: u64 = switch (ty) {
            .void_type => ANY_TAG_VOID,
            .boolean => ANY_TAG_BOOL,
            .signed => |w| if (w <= 32) ANY_TAG_S32 else ANY_TAG_S64,
            .unsigned => |w| if (w <= 32) ANY_TAG_S32 else ANY_TAG_S64,
            .f32 => ANY_TAG_F32,
            .f64 => ANY_TAG_F64,
            .string_type => ANY_TAG_STRING,
            .struct_type => |name| try self.getAnyTypeId(name, ty),
            .enum_type => |name| try self.getAnyTypeId(name, ty),
            .union_type => |name| try self.getAnyTypeId(name, ty),
            .vector_type => |info| try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "vec[{d}]{s}", .{ info.length, info.element_name }), ty),
            .array_type => |info| try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "[{d}]{s}", .{ info.length, info.element_name }), ty),
            .slice_type => |info| try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "[]{s}", .{info.element_name}), ty),
            .pointer_type => |info| try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "*{s}", .{info.pointee_name}), ty),
            .many_pointer_type => |info| try self.getAnyTypeId(try std.fmt.allocPrint(self.allocator, "[*]{s}", .{info.element_name}), ty),
            .meta_type => ANY_TAG_TYPE,
            else => ANY_TAG_S32,
        };
        const tag_val = c.LLVMConstInt(i64_ty, tag, 0);
        const with_tag = c.LLVMBuildInsertValue(self.builder, undef, tag_val, 0, "any_tag");

        // Convert value to i64
        const val_as_i64 = switch (ty) {
            .boolean => c.LLVMBuildZExt(self.builder, val, i64_ty, "any_bool"),
            .signed => |w| if (w <= 32)
                c.LLVMBuildSExt(self.builder, val, i64_ty, "any_int")
            else
                val,
            .unsigned => |w| if (w <= 32)
                c.LLVMBuildZExt(self.builder, val, i64_ty, "any_uint")
            else
                val,
            .f32 => blk: {
                // f32 -> f64 -> bitcast to i64
                const as_f64 = c.LLVMBuildFPExt(self.builder, val, c.LLVMDoubleTypeInContext(self.context), "f32_to_f64");
                break :blk c.LLVMBuildBitCast(self.builder, as_f64, i64_ty, "any_f32");
            },
            .f64 => c.LLVMBuildBitCast(self.builder, val, i64_ty, "any_f64"),
            .string_type => blk: {
                // String is {ptr, i32} — store to alloca, pass alloca as i64
                const str_alloca = self.buildEntryBlockAlloca(self.getStringStructType(), "any_str_tmp");
                _ = c.LLVMBuildStore(self.builder, val, str_alloca);
                break :blk c.LLVMBuildPtrToInt(self.builder, str_alloca, i64_ty, "any_str");
            },
            .struct_type => |sname| blk: {
                // Struct — store to alloca, pass pointer as i64
                const info = self.struct_types.get(sname) orelse
                    return c.LLVMGetUndef(any_ty);
                const alloca = self.buildEntryBlockAlloca(info.llvm_type, "any_struct_tmp");
                _ = c.LLVMBuildStore(self.builder, val, alloca);
                break :blk c.LLVMBuildPtrToInt(self.builder, alloca, i64_ty, "any_struct");
            },
            .enum_type => blk: {
                // Enum is i32 tag — extend to i64
                break :blk c.LLVMBuildZExt(self.builder, val, i64_ty, "any_enum");
            },
            .union_type => |uname| blk: {
                // Union — store to alloca, pass pointer as i64
                const info = self.union_types.get(uname) orelse
                    return c.LLVMGetUndef(any_ty);
                const alloca = self.buildEntryBlockAlloca(info.llvm_type, "any_union_tmp");
                _ = c.LLVMBuildStore(self.builder, val, alloca);
                break :blk c.LLVMBuildPtrToInt(self.builder, alloca, i64_ty, "any_union");
            },
            .vector_type, .array_type => blk: {
                // Vector/Array — store to alloca, pass pointer as i64
                const llvm_ty = self.typeToLLVM(ty);
                const alloca = self.buildEntryBlockAlloca(llvm_ty, "any_vec_tmp");
                _ = c.LLVMBuildStore(self.builder, val, alloca);
                break :blk c.LLVMBuildPtrToInt(self.builder, alloca, i64_ty, "any_vec");
            },
            .slice_type => blk: {
                // Slice {ptr, i32} — store to alloca, pass pointer as i64
                const alloca = self.buildEntryBlockAlloca(self.getStringStructType(), "any_slice_tmp");
                _ = c.LLVMBuildStore(self.builder, val, alloca);
                break :blk c.LLVMBuildPtrToInt(self.builder, alloca, i64_ty, "any_slice");
            },
            .pointer_type, .many_pointer_type => c.LLVMBuildPtrToInt(self.builder, val, i64_ty, "any_ptr"),
            .meta_type => |mt| blk: {
                // Meta type: wrap raw char ptr in string slice {ptr, len} for extraction
                const str_slice = self.buildStringSlice(val, mt.name.len);
                const alloca = self.buildEntryBlockAlloca(self.getStringStructType(), "any_type_tmp");
                _ = c.LLVMBuildStore(self.builder, str_slice, alloca);
                break :blk c.LLVMBuildPtrToInt(self.builder, alloca, i64_ty, "any_type");
            },
            else => c.LLVMBuildSExt(self.builder, val, i64_ty, "any_val"),
        };
        return c.LLVMBuildInsertValue(self.builder, with_tag, val_as_i64, 1, "any_value");
    }

    fn getStringStructType(self: *CodeGen) c.LLVMTypeRef {
        if (self.string_struct_type) |t| return t;
        var field_types = [_]c.LLVMTypeRef{
            c.LLVMPointerTypeInContext(self.context, 0), // ptr
            c.LLVMInt64TypeInContext(self.context), // len
        };
        self.string_struct_type = c.LLVMStructTypeInContext(self.context, &field_types, 2, 0);
        return self.string_struct_type.?;
    }

    /// Build a string slice {ptr, len} from a raw pointer and a constant length.
    fn buildStringSlice(self: *CodeGen, ptr: c.LLVMValueRef, len: u64) c.LLVMValueRef {
        const str_ty = self.getStringStructType();
        const undef = c.LLVMGetUndef(str_ty);
        const with_ptr = c.LLVMBuildInsertValue(self.builder, undef, ptr, 0, "str_ptr");
        const len_val = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), len, 0);
        return c.LLVMBuildInsertValue(self.builder, with_ptr, len_val, 1, "str_slice");
    }

    /// Build a string slice {ptr, len} from a raw pointer and a runtime length value.
    fn buildStringSliceRT(self: *CodeGen, ptr: c.LLVMValueRef, len_val: c.LLVMValueRef) c.LLVMValueRef {
        const str_ty = self.getStringStructType();
        const undef = c.LLVMGetUndef(str_ty);
        const with_ptr = c.LLVMBuildInsertValue(self.builder, undef, ptr, 0, "str_ptr");
        return c.LLVMBuildInsertValue(self.builder, with_ptr, len_val, 1, "str_slice");
    }

    fn pushScope(self: *CodeGen) !void {
        try self.scope_saves.append(self.allocator, std.ArrayList(ScopeEntry).empty);
        try self.defer_stack.append(self.allocator, std.ArrayList(*Node).empty);
    }

    fn popScope(self: *CodeGen) !void {
        // 1. Execute deferred expressions in LIFO order
        if (self.defer_stack.items.len > 0) {
            const defers = self.defer_stack.items[self.defer_stack.items.len - 1];
            var i: usize = defers.items.len;
            while (i > 0) {
                i -= 1;
                _ = try self.genExpr(defers.items[i]);
            }
            _ = self.defer_stack.pop();
        }

        // 2. Restore shadowed variables
        if (self.scope_saves.items.len > 0) {
            const saves = self.scope_saves.items[self.scope_saves.items.len - 1];
            // Restore in reverse order
            var i: usize = saves.items.len;
            while (i > 0) {
                i -= 1;
                const entry = saves.items[i];
                if (entry.prev) |prev| {
                    self.named_values.putAssumeCapacity(entry.name, prev);
                } else {
                    _ = self.named_values.remove(entry.name);
                }
            }
            _ = self.scope_saves.pop();
        }
    }

    /// Emit all pending deferred expressions from all active scopes (LIFO order,
    /// innermost scope first). Does NOT pop the stacks — used before `return`
    /// so that popScope() can still clean up the data structures later.
    fn emitAllDefers(self: *CodeGen) !void {
        var i: usize = self.defer_stack.items.len;
        while (i > 0) {
            i -= 1;
            const defers = self.defer_stack.items[i];
            var j: usize = defers.items.len;
            while (j > 0) {
                j -= 1;
                _ = try self.genExpr(defers.items[j]);
            }
        }
    }

    fn saveShadowed(self: *CodeGen, name: []const u8) !void {
        if (self.scope_saves.items.len == 0) return;
        const top = &self.scope_saves.items[self.scope_saves.items.len - 1];
        const prev = self.named_values.get(name);
        try top.append(self.allocator, .{ .name = name, .prev = prev });
    }

    pub fn generate(self: *CodeGen, root: *Node) !void {
        if (root.data != .root) return self.emitError("expected root node for code generation");
        // Store root decls for VM on-demand function compilation
        self.root_decls = root.data.root.decls;
        // Initialize built-in function declarations (printf, etc.)
        self.builtins = Builtins.init(self.module, self.context);

        // Pass 1: Register all declarations (signatures only, no bodies)
        for (root.data.root.decls) |decl| {
            switch (decl.data) {
                .fn_decl => |fd| {
                    if (fd.body.data == .builtin_expr) {
                        try self.builtin_functions.put(fd.name, {});
                    } else if (fd.body.data == .foreign_expr) {
                        // External C function — register LLVM declaration (no body)
                        try self.registerFnDecl(fd);
                    } else if (fd.type_params.len > 0) {
                        try self.generic_templates.put(fd.name, .{ .fd = fd });
                    } else {
                        try self.registerFnDecl(fd);
                    }
                    try self.fn_signatures.put(fd.name, self.buildFnSignature(fd));
                },
                .library_decl => |ld| {
                    try self.foreign_libraries.append(self.allocator, ld.lib_name);
                },
                .enum_decl => |ed| {
                    try self.enum_types.put(ed.name, ed.variants);
                    _ = try self.getAnyTypeId(ed.name, .{ .enum_type = ed.name });
                },
                .struct_decl => |sd| try self.registerStructType(sd),
                .union_decl => |ud| try self.registerUnionType(ud),
                .const_decl => |cd| {
                    if (cd.value.data == .builtin_expr) {
                        // #builtin constant — skip codegen
                    } else if (cd.value.data == .lambda) {
                        try self.registerLambdaAsFunction(cd.name, cd.value.data.lambda);
                    } else if (cd.value.data == .type_expr) {
                        try self.type_aliases.put(cd.name, cd.value.data.type_expr.name);
                    } else if (cd.value.data == .call) {
                        // Check if this is a generic struct or type function instantiation
                        const callee_name = if (cd.value.data.call.callee.data == .identifier)
                            cd.value.data.call.callee.data.identifier.name
                        else
                            null;
                        if (callee_name) |cn| {
                            if (self.generic_struct_templates.get(cn)) |tmpl| {
                                // Generic struct instantiation: Vec3 :: Vec(3, f32);
                                const result_ty = try self.instantiateGenericStruct(cn, tmpl, cd.value.data.call.args);
                                if (result_ty.isStruct()) {
                                    try self.type_aliases.put(cd.name, result_ty.struct_type);
                                }
                            } else if (self.generic_templates.get(cn)) |tmpl| {
                                // Type-returning function: Foo :: Complex(u32);
                                const result_ty = try self.instantiateTypeFunction(cd.name, cn, tmpl, cd.value.data.call.args);
                                if (result_ty.isStruct()) {
                                    try self.type_aliases.put(cd.name, result_ty.struct_type);
                                } else if (result_ty.isUnion()) {
                                    try self.type_aliases.put(cd.name, result_ty.union_type);
                                }
                            } else if (self.builtin_functions.contains(cn)) {
                                // Builtin type function (e.g., Vector(4, f32), Array(5, s32))
                                if (self.resolveBuiltinType(cn, cd.value.data.call.args)) |result_ty| {
                                    const display = try result_ty.displayName(self.allocator);
                                    try self.type_aliases.put(cd.name, display);
                                } else {
                                    try self.registerTopLevelConstant(cd);
                                }
                            } else {
                                try self.registerTopLevelConstant(cd);
                            }
                        } else {
                            try self.registerTopLevelConstant(cd);
                        }
                    } else if (cd.value.data == .comptime_expr) {
                        // Use explicit type annotation if available
                        const ct_type_override: ?Type = if (cd.type_annotation) |te| Type.fromTypeExpr(te) else null;
                        try self.registerComptimeGlobal(cd.name, cd.value.data.comptime_expr.expr, ct_type_override);
                    } else {
                        // Top-level value constant (e.g., SPECIAL_VALUE :u8: 42;)
                        try self.registerTopLevelConstant(cd);
                    }
                },
                .comptime_expr => |ct| {
                    try self.comptime_side_effects.append(self.allocator, ct.expr);
                },
                .namespace_decl => |ns| {
                    try self.registerNamespace(ns);
                },
                .var_decl => |vd| {
                    try self.registerGlobalVar(vd);
                },
                else => {},
            }
        }

        // Pre-register all known types for Any type ID assignment
        {
            var it = self.struct_types.iterator();
            while (it.next()) |entry| {
                _ = try self.getAnyTypeId(entry.key_ptr.*, .{ .struct_type = entry.key_ptr.* });
            }
        }
        {
            var it = self.enum_types.iterator();
            while (it.next()) |entry| {
                _ = try self.getAnyTypeId(entry.key_ptr.*, .{ .enum_type = entry.key_ptr.* });
            }
        }
        {
            var it = self.union_types.iterator();
            while (it.next()) |entry| {
                _ = try self.getAnyTypeId(entry.key_ptr.*, .{ .union_type = entry.key_ptr.* });
            }
        }

        // Pass 2: Generate all function bodies
        // Functions with Any parameters (like any_to_string) are deferred to Pass 3
        // so that all types are registered before their type-match expressions are compiled.
        for (root.data.root.decls) |decl| {
            switch (decl.data) {
                .fn_decl => |fd| {
                    if (fd.body.data == .builtin_expr or fd.body.data == .foreign_expr) {
                        // skip — no body to generate
                    } else if (fd.type_params.len == 0) {
                        if (shouldDeferFnBody(fd)) {
                            try self.deferred_fn_bodies.append(self.allocator, .{ .fd = fd, .name = fd.name });
                        } else {
                            try self.genFnBody(fd);
                        }
                    }
                },
                .const_decl => |cd| {
                    if (cd.value.data == .lambda) {
                        try self.genLambdaBody(cd.name, cd.value.data.lambda);
                    }
                },
                .namespace_decl => |ns| {
                    try self.genNamespaceBodies(ns);
                },
                else => {},
            }
        }

        // Pass 3: Compile deferred function bodies (after all types are registered)
        for (self.deferred_fn_bodies.items) |deferred| {
            const saved_ns = self.current_namespace;
            self.current_namespace = deferred.namespace;
            defer self.current_namespace = saved_ns;
            try self.genFnBodyAs(deferred.fd, deferred.name);
        }

        // Execute comptime side effects via bytecode VM (e.g., #run main();)
        for (self.comptime_side_effects.items) |expr| {
            _ = try self.comptimeEval(expr, .void_type);
        }
    }

    /// Evaluate a comptime expression using the bytecode VM.
    /// No LLVM state save/restore needed — the VM operates independently.
    fn comptimeEval(self: *CodeGen, expr: *Node, expected_type: Type) !comptime_mod.Value {
        _ = expected_type; // VM infers types from values; expected_type used by caller for LLVM conversion

        var compiler = comptime_mod.Compiler.init(self.allocator, if (self.sema_result) |sr| sr else null, self.root_decls, self);
        const chunk = compiler.compile(expr) catch |err| {
            return self.emitErrorFmt("comptime compilation failed: {s}", .{@errorName(err)});
        };

        var vm = comptime_mod.VM.init(self.allocator, if (self.sema_result) |sr| sr else null, self.root_decls, self);
        return vm.execute(&chunk) catch |err| {
            return self.emitErrorFmt("comptime execution failed: {s}", .{@errorName(err)});
        };
    }

    /// Substitute comptime param identifiers in an AST expression with their literal nodes.
    /// Used before comptimeEval in #insert to resolve comptime function params.
    fn substituteComptimeNodes(self: *CodeGen, node: *Node) !*Node {
        const cpn = self.comptime_param_nodes orelse return node;

        // Direct identifier match
        if (node.data == .identifier) {
            if (cpn.get(node.data.identifier.name)) |replacement| {
                return replacement;
            }
        }

        // Recurse into call arguments
        if (node.data == .call) {
            var new_args = try self.allocator.alloc(*Node, node.data.call.args.len);
            var changed = false;
            for (node.data.call.args, 0..) |arg, i| {
                new_args[i] = try self.substituteComptimeNodes(arg);
                if (new_args[i] != arg) changed = true;
            }
            if (changed) {
                const new_node = try self.allocator.create(Node);
                new_node.* = .{
                    .span = node.span,
                    .data = .{ .call = .{
                        .callee = node.data.call.callee,
                        .args = new_args,
                    } },
                };
                return new_node;
            }
        }

        // Recurse into binary ops
        if (node.data == .binary_op) {
            const new_lhs = try self.substituteComptimeNodes(node.data.binary_op.lhs);
            const new_rhs = try self.substituteComptimeNodes(node.data.binary_op.rhs);
            if (new_lhs != node.data.binary_op.lhs or new_rhs != node.data.binary_op.rhs) {
                const new_node = try self.allocator.create(Node);
                new_node.* = .{
                    .span = node.span,
                    .data = .{ .binary_op = .{
                        .op = node.data.binary_op.op,
                        .lhs = new_lhs,
                        .rhs = new_rhs,
                    } },
                };
                return new_node;
            }
        }

        return node;
    }

    /// Convert a comptime VM Value to an LLVM constant value.
    fn comptimeValueToLLVM(self: *CodeGen, value: comptime_mod.Value, ty: Type) c.LLVMValueRef {
        return switch (value) {
            .int_val => |v| c.LLVMConstInt(self.typeToLLVM(ty), @bitCast(v), 0),
            .float_val => |v| c.LLVMConstReal(c.LLVMDoubleTypeInContext(self.context), v),
            .float32_val => |v| c.LLVMConstReal(c.LLVMFloatTypeInContext(self.context), @as(f64, v)),
            .bool_val => |v| c.LLVMConstInt(c.LLVMInt1TypeInContext(self.context), if (v) 1 else 0, 0),
            .string_val => |v| blk: {
                const z = self.allocator.dupeZ(u8, v) catch unreachable;
                const ptr = c.LLVMBuildGlobalStringPtr(self.builder, z.ptr, "comptime_str");
                break :blk self.buildStringSlice(ptr, @intCast(v.len));
            },
            .void_val => c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0),
            .pointer_val => c.LLVMConstNull(c.LLVMPointerTypeInContext(self.context, 0)),
            .null_val => c.LLVMConstNull(c.LLVMPointerTypeInContext(self.context, 0)),
            .struct_val, .array_val, .type_val, .function_val => unreachable,
        };
    }

    /// Lazily resolve a comptime global by evaluating its expression via bytecode VM.
    fn resolveComptimeGlobal(self: *CodeGen, ct: *ComptimeGlobal) !void {
        const result = try self.comptimeEval(ct.expr, ct.ty);
        const const_val = self.comptimeValueToLLVM(result, ct.ty);
        c.LLVMSetInitializer(ct.global, const_val);
        c.LLVMSetGlobalConstant(ct.global, 1);
        ct.is_resolved = true;
    }

    fn resolveType(self: *CodeGen, type_node: ?*Node) Type {
        if (type_node) |tn| {
            if (Type.fromTypeExpr(tn)) |t| return t;
            // Array type: [N]T
            if (tn.data == .array_type_expr) {
                const ate = tn.data.array_type_expr;
                const length: u32 = @intCast(ate.length.data.int_literal.value);
                const elem_type = self.resolveType(ate.element_type);
                const elem_name = elem_type.displayName(self.allocator) catch unreachable;
                return .{ .array_type = .{ .element_name = elem_name, .length = length } };
            }
            // Slice type: []T
            if (tn.data == .slice_type_expr) {
                const ste = tn.data.slice_type_expr;
                const elem_type = self.resolveType(ste.element_type);
                const elem_name = elem_type.displayName(self.allocator) catch unreachable;
                return .{ .slice_type = .{ .element_name = elem_name } };
            }
            // Pointer type: *T
            if (tn.data == .pointer_type_expr) {
                const pte = tn.data.pointer_type_expr;
                const pointee_type = self.resolveType(pte.pointee_type);
                const pointee_name = pointee_type.displayName(self.allocator) catch unreachable;
                return .{ .pointer_type = .{ .pointee_name = pointee_name } };
            }
            // Many-pointer type: [*]T
            if (tn.data == .many_pointer_type_expr) {
                const mpte = tn.data.many_pointer_type_expr;
                const elem_type = self.resolveType(mpte.element_type);
                const elem_name = elem_type.displayName(self.allocator) catch unreachable;
                return .{ .many_pointer_type = .{ .element_name = elem_name } };
            }
            // Function pointer type: (ParamTypes) -> ReturnType
            if (tn.data == .function_type_expr) {
                const fte = tn.data.function_type_expr;
                var param_types = std.ArrayList(Type).empty;
                for (fte.param_types) |pt| {
                    param_types.append(self.allocator, self.resolveType(pt)) catch return .void_type;
                }
                const ret_ty = if (fte.return_type) |rt| self.resolveType(rt) else Type.void_type;
                const ret_ptr = self.allocator.create(Type) catch return .void_type;
                ret_ptr.* = ret_ty;
                return .{ .function_type = .{
                    .param_types = param_types.toOwnedSlice(self.allocator) catch return .void_type,
                    .return_type = ret_ptr,
                } };
            }
            // Parameterized type: Vector(N, T) or generic struct instantiation
            if (tn.data == .parameterized_type_expr) {
                const pte = tn.data.parameterized_type_expr;
                // Direct lookup (unqualified names from flat imports)
                if (self.builtin_functions.contains(pte.name)) {
                    if (self.resolveBuiltinType(pte.name, pte.args)) |ty| return ty;
                }
                if (self.generic_struct_templates.get(pte.name)) |tmpl| {
                    return self.instantiateGenericStruct(pte.name, tmpl, pte.args) catch .void_type;
                }
                // Progressive namespace resolution for dotted names (e.g. "std.Vector")
                if (std.mem.indexOfScalar(u8, pte.name, '.')) |dot| {
                    const ns = pte.name[0..dot];
                    if (self.namespaces.contains(ns)) {
                        // Namespace verified — look up qualified name in registries
                        if (self.builtin_functions.contains(pte.name)) {
                            if (self.resolveBuiltinType(pte.name, pte.args)) |ty| return ty;
                        }
                        if (self.generic_struct_templates.get(pte.name)) |tmpl| {
                            return self.instantiateGenericStruct(pte.name, tmpl, pte.args) catch .void_type;
                        }
                    }
                }
                if (self.diagnostics) |diags| diags.addFmt(.err, tn.span, "unresolved type '{s}'", .{pte.name});
                return .void_type;
            }
            // Call expression as type: Vec(3, f32) → generic struct/type function instantiation
            if (tn.data == .call) {
                const name = self.calleeToQualifiedName(tn.data.call.callee);
                if (name) |n| {
                    if (self.builtin_functions.contains(n)) {
                        if (self.resolveBuiltinType(n, tn.data.call.args)) |ty| return ty;
                    }
                    if (self.generic_struct_templates.get(n)) |tmpl| {
                        return self.instantiateGenericStruct(n, tmpl, tn.data.call.args) catch .void_type;
                    }
                    if (self.generic_templates.get(n)) |tmpl| {
                        return self.instantiateTypeFunction(n, n, tmpl, tn.data.call.args) catch .void_type;
                    }
                }
                return .void_type;
            }
            // Check type parameter bindings (during generic instantiation)
            if (tn.data == .type_expr or tn.data == .identifier) {
                const name = if (tn.data == .type_expr) tn.data.type_expr.name else tn.data.identifier.name;
                // Try primitive type name first
                if (Type.fromName(name)) |t| return t;
                if (self.type_param_bindings) |bindings| {
                    if (bindings.get(name)) |t| return t;
                }
                // Check type aliases
                if (self.type_aliases.get(name)) |target| {
                    if (Type.fromName(target)) |t| return t;
                    if (self.struct_types.contains(target)) return .{ .struct_type = target };
                    if (self.union_types.contains(target)) return .{ .union_type = target };
                }
                // Check enum types
                if (self.enum_types.contains(name)) return .{ .enum_type = name };
                // Check struct types
                if (self.struct_types.contains(name)) return .{ .struct_type = name };
                // Check union types
                if (self.union_types.contains(name)) return .{ .union_type = name };
            }
            // Safety net: inline declarations that should have been hoisted
            if (tn.data == .struct_decl) {
                const sn = tn.data.struct_decl.name;
                if (self.struct_types.contains(sn)) return .{ .struct_type = sn };
            }
            if (tn.data == .union_decl) {
                const un = tn.data.union_decl.name;
                if (self.union_types.contains(un)) return .{ .union_type = un };
            }
            if (tn.data == .enum_decl) {
                const en = tn.data.enum_decl.name;
                if (self.enum_types.contains(en)) return .{ .enum_type = en };
            }
            return .void_type;
        }
        return .void_type;
    }

    /// Resolve a value argument to an integer — handles int_literal and identifier referencing value_param_bindings.
    fn resolveValueArg(self: *CodeGen, node: *Node) i64 {
        if (node.data == .int_literal) return node.data.int_literal.value;
        if (node.data == .identifier or node.data == .type_expr) {
            const name = if (node.data == .identifier) node.data.identifier.name else node.data.type_expr.name;
            if (self.value_param_bindings) |bindings| {
                if (bindings.get(name)) |val| return val;
            }
        }
        return 0;
    }

    /// Instantiate a generic struct template with concrete arguments.
    /// Returns the struct_type for the instantiated struct (possibly cached).
    fn instantiateGenericStruct(self: *CodeGen, template_name: []const u8, tmpl: GenericStructTemplate, args: []const *Node) !Type {
        const sd = tmpl.sd;

        // Build bindings from template params + args
        var type_bindings = std.StringHashMap(Type).init(self.allocator);
        var val_bindings = std.StringHashMap(i64).init(self.allocator);

        for (sd.type_params, 0..) |tp, i| {
            if (i >= args.len) return self.emitErrorFmt("generic struct '{s}' expects {d} type arguments, got {d}", .{ template_name, sd.type_params.len, args.len });
            const constraint_name = if (tp.constraint.data == .type_expr) tp.constraint.data.type_expr.name else "";
            if (std.mem.eql(u8, constraint_name, "Type")) {
                // Type parameter: resolve arg as type
                const resolved = self.resolveType(args[i]);
                try type_bindings.put(tp.name, resolved);
            } else {
                // Value parameter: resolve arg as integer
                const val = self.resolveValueArg(args[i]);
                try val_bindings.put(tp.name, val);
            }
        }

        const mangled_name = try self.mangleGenericName(template_name, sd.type_params, type_bindings, val_bindings, null);

        // Check if already instantiated
        if (self.struct_types.contains(mangled_name)) {
            return .{ .struct_type = mangled_name };
        }

        // Instantiate: resolve field types with bindings active
        const saved_type_bindings = self.type_param_bindings;
        const saved_value_bindings = self.value_param_bindings;
        self.type_param_bindings = type_bindings;
        self.value_param_bindings = val_bindings;
        defer {
            self.type_param_bindings = saved_type_bindings;
            self.value_param_bindings = saved_value_bindings;
        }

        const build = try self.buildStructFields(mangled_name, sd.field_types);
        const resolved_defaults = try self.allocator.dupe(?*Node, sd.field_defaults);

        // Build pretty display name: Vec(3,f32)
        var display_buf = std.ArrayList(u8).empty;
        try display_buf.appendSlice(self.allocator, template_name);
        try display_buf.append(self.allocator, '(');
        for (sd.type_params, 0..) |tp, i| {
            if (i > 0) try display_buf.appendSlice(self.allocator, ",");
            const constraint_name = if (tp.constraint.data == .type_expr) tp.constraint.data.type_expr.name else "";
            if (std.mem.eql(u8, constraint_name, "Type")) {
                if (type_bindings.get(tp.name)) |ty| {
                    const dn = ty.displayName(self.allocator) catch "?";
                    try display_buf.appendSlice(self.allocator, dn);
                }
            } else {
                if (val_bindings.get(tp.name)) |val| {
                    var tmp: [20]u8 = undefined;
                    const s = std.fmt.bufPrint(&tmp, "{d}", .{val}) catch "0";
                    try display_buf.appendSlice(self.allocator, s);
                }
            }
        }
        try display_buf.append(self.allocator, ')');
        const display_name = try display_buf.toOwnedSlice(self.allocator);

        // Collect type param names and resolved types for later extraction
        var tp_names = std.ArrayList([]const u8).empty;
        var tp_types = std.ArrayList(Type).empty;
        for (sd.type_params) |tp| {
            const constraint_name = if (tp.constraint.data == .type_expr) tp.constraint.data.type_expr.name else "";
            if (std.mem.eql(u8, constraint_name, "Type")) {
                if (type_bindings.get(tp.name)) |ty| {
                    try tp_names.append(self.allocator, tp.name);
                    try tp_types.append(self.allocator, ty);
                }
            }
        }

        try self.struct_types.put(mangled_name, .{
            .field_names = sd.field_names,
            .field_types = build.field_sx_types,
            .field_defaults = resolved_defaults,
            .llvm_type = build.llvm_type,
            .display_name = display_name,
            .type_param_names = try tp_names.toOwnedSlice(self.allocator),
            .type_param_types = try tp_types.toOwnedSlice(self.allocator),
            .template_name = template_name,
        });
        _ = try self.getAnyTypeId(mangled_name, .{ .struct_type = mangled_name });

        return .{ .struct_type = mangled_name };
    }

    /// Instantiate a type-returning function (e.g. Complex(u32)) by walking the body AST
    /// to find `return struct { ... }` or `return union { ... }` and registering with bindings active.
    fn instantiateTypeFunction(self: *CodeGen, alias_name: []const u8, template_name: []const u8, tmpl: GenericTemplate, args: []const *Node) !Type {
        const fd = tmpl.fd;

        // Build type bindings from params + args
        var type_bindings = std.StringHashMap(Type).init(self.allocator);
        for (fd.type_params, 0..) |tp, i| {
            if (i >= args.len) return self.emitErrorFmt("type function '{s}' expects {d} type arguments, got {d}", .{ template_name, fd.type_params.len, args.len });
            const resolved = self.resolveType(args[i]);
            try type_bindings.put(tp.name, resolved);
        }

        // Activate bindings
        const saved_type_bindings = self.type_param_bindings;
        self.type_param_bindings = type_bindings;
        defer self.type_param_bindings = saved_type_bindings;

        const mangled_name = try self.mangleGenericName(template_name, fd.type_params, type_bindings, null, null);

        // Try struct first
        if (self.findStructInBody(fd.body)) |struct_decl| {
            if (self.struct_types.contains(mangled_name)) {
                return .{ .struct_type = mangled_name };
            }
            return self.registerInstantiatedStruct(mangled_name, alias_name, struct_decl);
        }

        // Try union
        if (self.findUnionInBody(fd.body)) |union_decl| {
            if (self.union_types.contains(mangled_name)) {
                return .{ .union_type = mangled_name };
            }
            return self.registerInstantiatedUnion(mangled_name, union_decl);
        }

        return self.emitErrorFmt("type function '{s}' does not return a struct or union", .{template_name});
    }

    fn registerInstantiatedStruct(self: *CodeGen, mangled_name: []const u8, alias_name: []const u8, struct_decl: ast.StructDecl) !Type {
        const build = try self.buildStructFields(mangled_name, struct_decl.field_types);

        const resolved_defaults = try self.allocator.dupe(?*Node, struct_decl.field_defaults);
        const display_name = try self.allocator.dupe(u8, alias_name);

        try self.struct_types.put(mangled_name, .{
            .field_names = struct_decl.field_names,
            .field_types = build.field_sx_types,
            .field_defaults = resolved_defaults,
            .llvm_type = build.llvm_type,
            .display_name = display_name,
        });
        _ = try self.getAnyTypeId(mangled_name, .{ .struct_type = mangled_name });

        return .{ .struct_type = mangled_name };
    }

    fn registerInstantiatedUnion(self: *CodeGen, mangled_name: []const u8, union_decl: ast.UnionDecl) !Type {
        const build = try self.buildUnionFields(mangled_name, union_decl.variant_types);

        try self.union_types.put(mangled_name, .{
            .variant_names = union_decl.variant_names,
            .variant_types = build.variant_sx_types,
            .llvm_type = build.llvm_type,
            .max_payload_size = build.max_payload_size,
        });
        _ = try self.getAnyTypeId(mangled_name, .{ .union_type = mangled_name });

        return .{ .union_type = mangled_name };
    }

    /// Walk an AST body to find a struct declaration (from `return struct { ... }` or bare struct expr).
    fn findDeclInBody(comptime T: type, comptime tag: std.meta.FieldEnum(@TypeOf(@as(Node, undefined).data)), body: *Node) ?T {
        const extract = struct {
            fn get(node: *Node) ?T {
                return if (@field(node.data, @tagName(tag)) != @as(?T, null))
                    @field(node.data, @tagName(tag))
                else
                    null;
            }
        };
        _ = extract;
        if (body.data == tag) return @field(body.data, @tagName(tag));
        if (body.data == .block) {
            for (body.data.block.stmts) |stmt| {
                if (stmt.data == .return_stmt) {
                    if (stmt.data.return_stmt.value) |val| {
                        if (val.data == tag) return @field(val.data, @tagName(tag));
                    }
                }
                if (stmt.data == tag) return @field(stmt.data, @tagName(tag));
            }
        }
        return null;
    }

    fn findStructInBody(_: *CodeGen, body: *Node) ?ast.StructDecl {
        return findDeclInBody(ast.StructDecl, .struct_decl, body);
    }

    fn findUnionInBody(_: *CodeGen, body: *Node) ?ast.UnionDecl {
        return findDeclInBody(ast.UnionDecl, .union_decl, body);
    }

    fn buildFnType(self: *CodeGen, params: []const ast.Param, return_type: ?*Node, name: []const u8) !c.LLVMTypeRef {
        return self.buildFnTypeEx(params, return_type, name, false);
    }

    fn buildFnTypeEx(self: *CodeGen, params: []const ast.Param, return_type: ?*Node, name: []const u8, is_foreign: bool) !c.LLVMTypeRef {
        const ret_sx_type = self.resolveType(return_type);
        const is_main = std.mem.eql(u8, name, "main");
        const ret_llvm_type = if (is_main)
            c.LLVMInt32TypeInContext(self.context)
        else if (is_foreign and ret_sx_type.isStruct())
            self.getForeignReturnABIType(ret_sx_type)
        else
            self.typeToLLVM(ret_sx_type);

        var param_llvm_types = std.ArrayList(c.LLVMTypeRef).empty;
        for (params) |param| {
            if (param.is_comptime) continue;
            if (param.is_variadic) {
                // Variadic param becomes a slice {ptr, i32} in the LLVM signature
                try param_llvm_types.append(self.allocator, self.getStringStructType());
            } else {
                const sx_ty = self.resolveType(param.type_expr);
                if (sx_ty == .void_type) return self.emitErrorFmt("parameter '{s}' in function '{s}' has unresolved type", .{ param.name, name });
                // Foreign functions: apply C ABI lowering
                if (is_foreign and sx_ty == .string_type) {
                    try param_llvm_types.append(self.allocator, c.LLVMPointerTypeInContext(self.context, 0));
                } else if (is_foreign and sx_ty.isStruct()) {
                    try param_llvm_types.append(self.allocator, self.getForeignParamABIType(sx_ty));
                } else if (is_foreign and sx_ty.isArray()) {
                    // [N]T → pointer in C ABI (C arrays decay to pointers)
                    try param_llvm_types.append(self.allocator, c.LLVMPointerTypeInContext(self.context, 0));
                } else {
                    try param_llvm_types.append(self.allocator, self.typeToLLVM(sx_ty));
                }
            }
        }
        const params_slice = try param_llvm_types.toOwnedSlice(self.allocator);

        return c.LLVMFunctionType(
            ret_llvm_type,
            if (params_slice.len > 0) params_slice.ptr else null,
            @intCast(params_slice.len),
            0,
        );
    }

    /// For foreign (C ABI) functions, struct parameters must be lowered to their
    /// ABI-equivalent types. LLVM does NOT do this automatically on all targets.
    /// Dispatches to architecture-specific lowering based on target config.
    fn getForeignParamABIType(self: *CodeGen, sx_ty: Type) c.LLVMTypeRef {
        if (!sx_ty.isStruct()) return self.typeToLLVM(sx_ty);

        const sname = self.type_aliases.get(sx_ty.struct_type) orelse sx_ty.struct_type;
        const info = self.struct_types.get(sname) orelse return self.typeToLLVM(sx_ty);

        if (self.target_config.isAarch64()) {
            return self.aarch64ParamABI(info);
        } else if (self.target_config.isX86_64()) {
            if (self.target_config.isWindows()) {
                return self.win64ParamABI(info);
            }
            return self.x86_64SysVParamABI(info);
        }
        // Unknown architecture: pass struct type as-is (let LLVM backend handle it)
        return info.llvm_type;
    }

    /// AArch64 ABI: struct parameter lowering.
    /// - HFA (1-4 same float/double fields): [N x float/double]
    /// - Non-HFA ≤ 8 bytes: i64
    /// - Non-HFA 9-16 bytes: [2 x i64]
    /// - > 16 bytes: pass as-is (indirect, not yet fully handled)
    fn aarch64ParamABI(self: *CodeGen, info: StructInfo) c.LLVMTypeRef {
        // Check HFA: 1-4 fields all of the same float type
        const field_types = info.field_types;
        if (field_types.len >= 1 and field_types.len <= 4) {
            const first = field_types[0];
            if (first == .f32 or first == .f64) {
                var all_same = true;
                for (field_types[1..]) |ft| {
                    if (!ft.eql(first)) {
                        all_same = false;
                        break;
                    }
                }
                if (all_same) {
                    const elem_ty = if (first == .f32)
                        c.LLVMFloatTypeInContext(self.context)
                    else
                        c.LLVMDoubleTypeInContext(self.context);
                    return c.LLVMArrayType2(elem_ty, @intCast(field_types.len));
                }
            }
        }

        // Non-HFA: pack into integer registers
        const data_layout = c.LLVMGetModuleDataLayout(self.module);
        const size = c.LLVMStoreSizeOfType(data_layout, info.llvm_type);
        if (size <= 8) return c.LLVMInt64TypeInContext(self.context);
        if (size <= 16) return c.LLVMArrayType2(c.LLVMInt64TypeInContext(self.context), 2);
        return info.llvm_type;
    }

    /// x86-64 SysV ABI: struct parameter lowering.
    /// Each 8-byte "eightbyte" is classified as INTEGER or SSE:
    /// - If all fields in the eightbyte are float/double: SSE (passed in XMM register)
    /// - If any field is integer/pointer: INTEGER (passed in GPR)
    /// - Structs > 16 bytes: passed in memory (by pointer)
    fn x86_64SysVParamABI(self: *CodeGen, info: StructInfo) c.LLVMTypeRef {
        const data_layout = c.LLVMGetModuleDataLayout(self.module);
        const size = c.LLVMStoreSizeOfType(data_layout, info.llvm_type);

        // > 16 bytes: MEMORY class (passed by pointer, handled by LLVM backend)
        if (size > 16) return info.llvm_type;

        // Single eightbyte (≤ 8 bytes)
        if (size <= 8) {
            return self.classifyEightbyte(info.field_types, size);
        }

        // Two eightbytes (9-16 bytes): classify each half independently
        // Split fields into first eightbyte (offset < 8) and second eightbyte (offset >= 8)
        var first_eb_types = std.ArrayList(Type).empty;
        var second_eb_types = std.ArrayList(Type).empty;
        var second_eb_size: u64 = 0;

        const struct_ty = info.llvm_type;
        for (info.field_types, 0..) |ft, idx| {
            const offset = c.LLVMOffsetOfElement(data_layout, struct_ty, @intCast(idx));
            if (offset < 8) {
                first_eb_types.append(self.allocator, ft) catch return info.llvm_type;
            } else {
                second_eb_types.append(self.allocator, ft) catch return info.llvm_type;
                const field_llvm = self.typeToLLVM(ft);
                second_eb_size += c.LLVMStoreSizeOfType(data_layout, field_llvm);
            }
        }

        const eb1 = self.classifyEightbyte(first_eb_types.items, 8);
        const eb2 = self.classifyEightbyte(second_eb_types.items, if (second_eb_size > 0) second_eb_size else size - 8);

        // Compose the two eightbytes into a struct type
        var members: [2]c.LLVMTypeRef = .{ eb1, eb2 };
        return c.LLVMStructTypeInContext(self.context, &members, 2, 0);
    }

    /// Classify a single x86-64 eightbyte: if all fields are float, return SSE type;
    /// otherwise return an integer type matching the byte size.
    fn classifyEightbyte(self: *CodeGen, field_types_in_eb: []const Type, byte_size: u64) c.LLVMTypeRef {
        if (field_types_in_eb.len == 0) {
            // No fields in this chunk — use integer padding
            return c.LLVMIntTypeInContext(self.context, @intCast(byte_size * 8));
        }

        // Check if all fields are SSE (float/double)
        var all_sse = true;
        var float_count: u32 = 0;
        var double_count: u32 = 0;
        for (field_types_in_eb) |ft| {
            if (ft == .f32) {
                float_count += 1;
            } else if (ft == .f64) {
                double_count += 1;
            } else {
                all_sse = false;
                break;
            }
        }

        if (all_sse) {
            // SSE class: return appropriate float type
            if (double_count > 0 and float_count == 0) {
                if (double_count == 1) return c.LLVMDoubleTypeInContext(self.context);
                // Multiple doubles shouldn't fit in one eightbyte (double = 8 bytes)
                return c.LLVMDoubleTypeInContext(self.context);
            }
            if (float_count > 0 and double_count == 0) {
                if (float_count == 1) return c.LLVMFloatTypeInContext(self.context);
                // 2 floats = 8 bytes, fits in one eightbyte
                return c.LLVMArrayType2(c.LLVMFloatTypeInContext(self.context), @intCast(float_count));
            }
            // Mixed float/double in one eightbyte shouldn't happen (float=4, double=8)
            // but fall through to integer just in case
        }

        // INTEGER class: coerce to integer matching the byte size
        return c.LLVMIntTypeInContext(self.context, @intCast(byte_size * 8));
    }

    /// Windows x64 ABI: struct parameter lowering.
    /// Only structs of exactly 1, 2, 4, or 8 bytes are passed in a register.
    /// Everything else is passed by pointer (handled by LLVM backend).
    fn win64ParamABI(self: *CodeGen, info: StructInfo) c.LLVMTypeRef {
        const data_layout = c.LLVMGetModuleDataLayout(self.module);
        const size = c.LLVMStoreSizeOfType(data_layout, info.llvm_type);

        // Windows x64: only power-of-2 sizes ≤ 8 passed in register
        if (size == 1 or size == 2 or size == 4 or size == 8) {
            return c.LLVMIntTypeInContext(self.context, @intCast(size * 8));
        }
        // All other sizes: passed by pointer (LLVM handles byval)
        return info.llvm_type;
    }

    /// For foreign functions returning structs, apply the same ABI lowering as parameters.
    /// The rules for return values match parameter rules on both AArch64 and x86-64 SysV
    /// for small structs (≤ 16 bytes). Larger structs use sret (handled by LLVM).
    fn getForeignReturnABIType(self: *CodeGen, sx_ty: Type) c.LLVMTypeRef {
        // Reuse the same classification as parameters — the rules are identical
        // for small struct returns on both AArch64 and x86-64 SysV.
        return self.getForeignParamABIType(sx_ty);
    }

    /// Convert a struct value to its C ABI representation for a foreign call.
    /// Stores the struct to memory, then loads as the ABI type.
    fn convertStructToABI(self: *CodeGen, struct_val: c.LLVMValueRef, struct_ty: c.LLVMTypeRef, abi_ty: c.LLVMTypeRef) c.LLVMValueRef {
        const data_layout = c.LLVMGetModuleDataLayout(self.module);
        const struct_size = c.LLVMStoreSizeOfType(data_layout, struct_ty);
        const abi_size = c.LLVMStoreSizeOfType(data_layout, abi_ty);

        if (struct_size == abi_size) {
            // Same size (e.g. {float, float} → [2 x float]): store and reload
            const alloca = self.buildEntryBlockAlloca(struct_ty, "abi_tmp");
            _ = c.LLVMBuildStore(self.builder, struct_val, alloca);
            return c.LLVMBuildLoad2(self.builder, abi_ty, alloca, "abi_arg");
        } else {
            // Struct smaller than ABI type (e.g. {i8,i8,i8,i8} → i64): zero-init, then store struct
            const alloca = self.buildEntryBlockAlloca(abi_ty, "abi_tmp");
            _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(abi_ty), alloca);
            _ = c.LLVMBuildStore(self.builder, struct_val, alloca);
            return c.LLVMBuildLoad2(self.builder, abi_ty, alloca, "abi_arg");
        }
    }

    fn registerFnDecl(self: *CodeGen, fd: ast.FnDecl) !void {
        return self.registerFnDeclAs(fd, fd.name);
    }

    fn registerFnDeclAs(self: *CodeGen, fd: ast.FnDecl, llvm_name: []const u8) !void {
        const is_foreign = fd.body.data == .foreign_expr;
        const fn_type = try self.buildFnTypeEx(fd.params, fd.return_type, fd.name, is_foreign);
        const name_z = try self.allocator.dupeZ(u8, llvm_name);
        _ = c.LLVMAddFunction(self.module, name_z.ptr, fn_type);
        // Track foreign functions for ABI lowering at call sites
        if (is_foreign) try self.foreign_fns.put(llvm_name, {});
        // Track resolved parameter types for accurate call-site conversion
        var param_types = std.ArrayList(Type).empty;
        for (fd.params) |param| {
            if (param.is_comptime) continue;
            if (param.is_variadic) {
                const elem_name = if (param.type_expr.data == .type_expr) param.type_expr.data.type_expr.name else "s32";
                try param_types.append(self.allocator, .{ .slice_type = .{ .element_name = elem_name } });
            } else {
                try param_types.append(self.allocator, self.resolveType(param.type_expr));
            }
        }
        try self.fn_param_types.put(llvm_name, try param_types.toOwnedSlice(self.allocator));
        // Track declared return type (preserves signedness lost by LLVM round-trip)
        const ret_ty = if (fd.return_type) |rt| self.resolveType(rt) else Type.void_type;
        try self.function_return_types.put(llvm_name, ret_ty);
        // Track variadic function info for call site packing
        for (fd.params, 0..) |param, i| {
            if (param.is_variadic) {
                const elem_name = if (param.type_expr.data == .type_expr) param.type_expr.data.type_expr.name else "s32";
                try self.variadic_functions.put(llvm_name, .{
                    .fixed_param_count = @intCast(i),
                    .element_type_name = elem_name,
                });
                break;
            }
        }
    }

    fn registerNamespace(self: *CodeGen, ns: ast.NamespaceDecl) !void {
        try self.namespaces.put(ns.name, {});
        for (ns.decls) |decl| {
            switch (decl.data) {
                .fn_decl => |fd| {
                    if (fd.body.data == .builtin_expr) {
                        const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, fd.name });
                        try self.builtin_functions.put(qualified, {});
                        continue;
                    }
                    const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, fd.name });
                    if (fd.body.data == .foreign_expr) {
                        // External C function in namespace — register LLVM declaration with C name only
                        try self.registerFnDeclAs(fd, fd.name);
                        // Also track qualified name as foreign for ABI lowering at call sites
                        try self.foreign_fns.put(qualified, {});
                        // Store param types under qualified name so call-site type resolution works
                        var param_types = std.ArrayList(Type).empty;
                        for (fd.params) |param| {
                            if (param.is_comptime) continue;
                            try param_types.append(self.allocator, self.resolveType(param.type_expr));
                        }
                        try self.fn_param_types.put(qualified, try param_types.toOwnedSlice(self.allocator));
                    } else if (fd.type_params.len > 0) {
                        try self.generic_templates.put(qualified, .{ .fd = fd });
                    } else {
                        try self.registerFnDeclAs(fd, qualified);
                    }
                },
                .enum_decl => |ed| {
                    const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, ed.name });
                    try self.enum_types.put(qualified, ed.variants);
                },
                .struct_decl => |sd| {
                    try self.registerStructType(sd);
                    // Register qualified alias so rl.Color resolves to Color
                    const qualified_s = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, sd.name });
                    try self.type_aliases.put(qualified_s, sd.name);
                },
                .union_decl => |ud| {
                    try self.registerUnionType(ud);
                    const qualified_u = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, ud.name });
                    try self.type_aliases.put(qualified_u, ud.name);
                },
                .const_decl => |cd| {
                    if (cd.value.data == .builtin_expr) {
                        // #builtin constant in namespace — skip codegen
                    } else if (cd.value.data == .lambda) {
                        const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, cd.name });
                        try self.registerLambdaAsFunction(qualified, cd.value.data.lambda);
                    } else if (cd.value.data == .type_expr) {
                        const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, cd.name });
                        try self.type_aliases.put(qualified, cd.value.data.type_expr.name);
                    }
                },
                .library_decl => |ld| {
                    try self.foreign_libraries.append(self.allocator, ld.lib_name);
                },
                .var_decl => |vd| {
                    try self.registerGlobalVar(vd);
                },
                else => {},
            }
        }
    }

    fn genNamespaceBodies(self: *CodeGen, ns: ast.NamespaceDecl) !void {
        const saved_ns = self.current_namespace;
        self.current_namespace = ns.name;
        defer self.current_namespace = saved_ns;

        for (ns.decls) |decl| {
            switch (decl.data) {
                .fn_decl => |fd| {
                    if (fd.body.data == .builtin_expr or fd.body.data == .foreign_expr) {
                        // skip — no body to generate
                    } else if (fd.type_params.len == 0) {
                        const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, fd.name });
                        if (shouldDeferFnBody(fd)) {
                            try self.deferred_fn_bodies.append(self.allocator, .{ .fd = fd, .name = qualified, .namespace = ns.name });
                        } else {
                            try self.genFnBodyAs(fd, qualified);
                        }
                    }
                },
                .const_decl => |cd| {
                    if (cd.value.data == .lambda) {
                        const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns.name, cd.name });
                        try self.genLambdaBody(qualified, cd.value.data.lambda);
                    }
                },
                else => {},
            }
        }
    }

    fn inferComptimeReturnType(self: *CodeGen, expr: *Node) Type {
        // xx: see through to inner expression
        if (expr.data == .unary_op and expr.data.unary_op.op == .xx) {
            return self.inferComptimeReturnType(expr.data.unary_op.operand);
        }
        // For function calls, look up the registered function's return type
        if (expr.data == .call) {
            if (self.resolveCalleeName(expr.data.call)) |callee_name| {
                const callee_name_z = self.allocator.dupeZ(u8, callee_name) catch return Type.s(64);
                const callee_fn = c.LLVMGetNamedFunction(self.module, callee_name_z.ptr) orelse return Type.s(64);
                const fn_type = c.LLVMGlobalGetValueType(callee_fn);
                const ret_llvm = c.LLVMGetReturnType(fn_type);
                return self.llvmTypeToSxType(ret_llvm);
            }
        }
        return self.inferType(expr);
    }

    /// Map an LLVM type back to a sx Type
    fn llvmTypeToSxType(self: *CodeGen, llvm_ty: c.LLVMTypeRef) Type {
        if (llvm_ty == c.LLVMInt1TypeInContext(self.context)) return .boolean;
        if (llvm_ty == c.LLVMInt8TypeInContext(self.context)) return Type.s(8);
        if (llvm_ty == c.LLVMInt16TypeInContext(self.context)) return Type.s(16);
        if (llvm_ty == c.LLVMInt32TypeInContext(self.context)) return Type.s(32);
        if (llvm_ty == c.LLVMInt64TypeInContext(self.context)) return Type.s(64);
        if (llvm_ty == c.LLVMFloatTypeInContext(self.context)) return .f32;
        if (llvm_ty == c.LLVMDoubleTypeInContext(self.context)) return .f64;
        if (llvm_ty == c.LLVMVoidTypeInContext(self.context)) return .void_type;
        if (llvm_ty == self.getStringStructType()) return .string_type;
        if (self.any_struct_type != null and llvm_ty == self.any_struct_type.?) return .any_type;
        if (llvm_ty == c.LLVMPointerTypeInContext(self.context, 0)) return .string_type; // raw ptr fallback (meta_type)
        // Handle arbitrary-width integer types (e.g. i3, i7, i12)
        if (c.LLVMGetTypeKind(llvm_ty) == c.LLVMIntegerTypeKind) {
            const width = c.LLVMGetIntTypeWidth(llvm_ty);
            if (width > 0 and width <= 64) return Type.s(@intCast(width));
        }
        // Check for named struct types
        if (c.LLVMGetTypeKind(llvm_ty) == c.LLVMStructTypeKind) {
            const name_ptr = c.LLVMGetStructName(llvm_ty);
            if (name_ptr != null) {
                const name = std.mem.span(name_ptr);
                if (self.struct_types.contains(name)) return .{ .struct_type = name };
                if (self.union_types.contains(name)) return .{ .union_type = name };
            }
        }
        // Check for array types
        if (c.LLVMGetTypeKind(llvm_ty) == c.LLVMArrayTypeKind) {
            const elem_llvm = c.LLVMGetElementType(llvm_ty);
            const length: u32 = @intCast(c.LLVMGetArrayLength2(llvm_ty));
            const elem_ty = self.llvmTypeToSxType(elem_llvm);
            const elem_name = elem_ty.displayName(self.allocator) catch return Type.s(64);
            return .{ .array_type = .{ .element_name = elem_name, .length = length } };
        }
        // Check for vector types
        if (c.LLVMGetTypeKind(llvm_ty) == c.LLVMVectorTypeKind) {
            const elem_llvm = c.LLVMGetElementType(llvm_ty);
            const length = c.LLVMGetVectorSize(llvm_ty);
            const elem_ty = self.llvmTypeToSxType(elem_llvm);
            const elem_name = elem_ty.displayName(self.allocator) catch return Type.s(64);
            return .{ .vector_type = .{ .element_name = elem_name, .length = length } };
        }
        return Type.s(64);
    }

    fn registerComptimeGlobal(self: *CodeGen, name: []const u8, expr: *Node, type_override: ?Type) !void {
        const ty = type_override orelse self.inferComptimeReturnType(expr);
        if (ty == .void_type) return self.emitErrorFmt("cannot infer type for comptime global '{s}'", .{name});

        const llvm_ty = self.typeToLLVM(ty);
        const name_z = try self.allocator.dupeZ(u8, name);
        const global = c.LLVMAddGlobal(self.module, llvm_ty, name_z.ptr);
        c.LLVMSetInitializer(global, c.LLVMConstInt(llvm_ty, 0, 0));

        try self.comptime_globals.put(name, .{ .global = global, .ty = ty, .expr = expr });
    }

    /// Evaluate a simple constant expression to an LLVM constant value.
    /// Returns null for expressions that can't be constant-folded at registration time.
    fn evalConstant(self: *CodeGen, node: *Node, target_ty: Type) ?c.LLVMValueRef {
        const llvm_ty = self.typeToLLVM(target_ty);
        switch (node.data) {
            .int_literal => |lit| {
                return c.LLVMConstInt(llvm_ty, @bitCast(@as(i64, lit.value)), 0);
            },
            .float_literal => |lit| {
                return c.LLVMConstReal(llvm_ty, lit.value);
            },
            .bool_literal => |lit| {
                return c.LLVMConstInt(llvm_ty, if (lit.value) 1 else 0, 0);
            },
            else => return null,
        }
    }

    /// Register a top-level value constant (e.g., `SPECIAL_VALUE :u8: 42;`) as an LLVM global.
    fn registerTopLevelConstant(self: *CodeGen, cd: ast.ConstDecl) !void {
        const ta = cd.type_annotation orelse return; // need explicit type for top-level constants
        const sx_ty = self.resolveType(ta);
        if (sx_ty == .void_type) return;

        const const_val = self.evalConstant(cd.value, sx_ty) orelse return;

        const name_z = try self.allocator.dupeZ(u8, cd.name);
        const global = c.LLVMAddGlobal(self.module, self.typeToLLVM(sx_ty), name_z.ptr);
        c.LLVMSetInitializer(global, const_val);
        c.LLVMSetGlobalConstant(global, 1);

        try self.comptime_globals.put(cd.name, .{
            .global = global,
            .ty = sx_ty,
            .expr = cd.value,
            .is_resolved = true,
        });
    }

    fn registerGlobalVar(self: *CodeGen, vd: ast.VarDecl) !void {
        const ta = vd.type_annotation orelse return;
        const sx_ty = self.resolveType(ta);
        if (sx_ty == .void_type) return;

        const llvm_ty = self.typeToLLVM(sx_ty);
        const name_z = try self.allocator.dupeZ(u8, vd.name);
        const global = c.LLVMAddGlobal(self.module, llvm_ty, name_z.ptr);
        // Initialize with undef (will be set at runtime, e.g. by load_gl)
        c.LLVMSetInitializer(global, c.LLVMGetUndef(llvm_ty));
        // NOT constant — this is a mutable global
        c.LLVMSetGlobalConstant(global, 0);

        try self.global_mutable_vars.put(vd.name, .{ .ptr = global, .ty = sx_ty });
    }

    fn bindParam(self: *CodeGen, function: c.LLVMValueRef, name: []const u8, sx_ty: Type, param_idx: u32) !void {
        const llvm_ty = self.typeToLLVM(sx_ty);
        const param_name_z = try self.allocator.dupeZ(u8, name);
        const alloca = c.LLVMBuildAlloca(self.builder, llvm_ty, param_name_z.ptr);
        const param_val = c.LLVMGetParam(function, param_idx);
        _ = c.LLVMBuildStore(self.builder, param_val, alloca);
        try self.named_values.put(name, .{ .ptr = alloca, .ty = sx_ty });
    }

    fn genFnBody(self: *CodeGen, fd: ast.FnDecl) !void {
        return self.genFnBodyAs(fd, fd.name);
    }

    fn genFnBodyAs(self: *CodeGen, fd: ast.FnDecl, llvm_name: []const u8) !void {
        self.named_values.clearRetainingCapacity();

        const ret_sx_type = self.resolveType(fd.return_type);
        const is_main = std.mem.eql(u8, llvm_name, "main");
        const ret_llvm_type = if (is_main)
            c.LLVMInt32TypeInContext(self.context)
        else
            self.typeToLLVM(ret_sx_type);

        self.current_return_type = if (is_main) Type.s(32) else ret_sx_type;

        const name_z = try self.allocator.dupeZ(u8, llvm_name);
        const function = c.LLVMGetNamedFunction(self.module, name_z.ptr) orelse return self.emitErrorFmt("function '{s}' not found in LLVM module", .{llvm_name});
        self.current_function = function;

        const entry = c.LLVMAppendBasicBlockInContext(self.context, function, "entry");
        c.LLVMPositionBuilderAtEnd(self.builder, entry);

        // Create allocas for parameters and store incoming values
        for (fd.params, 0..) |param, i| {
            const sx_ty = if (param.is_variadic) blk: {
                const elem_name = if (param.type_expr.data == .type_expr) param.type_expr.data.type_expr.name else "s32";
                break :blk Type{ .slice_type = .{ .element_name = elem_name } };
            } else self.resolveType(param.type_expr);
            if (sx_ty == .void_type) return self.emitErrorFmt("parameter '{s}' has unresolved type", .{param.name});
            try self.bindParam(function, param.name, sx_ty, @intCast(i));
        }

        // Push function-level scope so that function-body defers are tracked
        try self.pushScope();

        // Generate body
        const body = fd.body;
        if (body.data != .block) return self.emitError("function body must be a block");

        var last_val: c.LLVMValueRef = null;
        for (body.data.block.stmts) |stmt| {
            last_val = try self.genStmt(stmt);
        }

        // Return — skip if current block already has a terminator (from explicit return)
        const current_bb = c.LLVMGetInsertBlock(self.builder);
        if (c.LLVMGetBasicBlockTerminator(current_bb) == null) {
            // Implicit return path: pop scope (executes defers) then return
            try self.popScope();
            // Check if last_val is void-typed (e.g. call to void-returning function)
            const effective_last_val: ?c.LLVMValueRef = if (last_val) |val|
                (if (c.LLVMTypeOf(val) == c.LLVMVoidTypeInContext(self.context)) null else val)
            else
                null;

            if (ret_sx_type == .void_type and !is_main) {
                _ = c.LLVMBuildRetVoid(self.builder);
            } else if (effective_last_val) |val| {
                if (ret_sx_type.isStruct()) {
                    // Struct implicit return: val is an alloca pointer, load the value
                    const sname = ret_sx_type.struct_type;
                    const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                    const loaded = c.LLVMBuildLoad2(self.builder, info.llvm_type, val, "retval");
                    _ = c.LLVMBuildRet(self.builder, loaded);
                } else {
                    const src_ty = self.llvmTypeToSxType(c.LLVMTypeOf(val));
                    const ret_val = self.convertValue(val, src_ty, self.current_return_type);
                    _ = c.LLVMBuildRet(self.builder, ret_val);
                }
            } else if (is_main) {
                _ = c.LLVMBuildRet(self.builder, c.LLVMConstInt(ret_llvm_type, 0, 0));
            } else if (ret_sx_type != .void_type) {
                _ = c.LLVMBuildUnreachable(self.builder);
            } else {
                _ = c.LLVMBuildRetVoid(self.builder);
            }
        } else {
            // Explicit return already emitted defers; just clean up scope stacks
            if (self.defer_stack.items.len > 0) _ = self.defer_stack.pop();
            if (self.scope_saves.items.len > 0) _ = self.scope_saves.pop();
        }
    }

    fn registerLambdaAsFunction(self: *CodeGen, name: []const u8, lambda: ast.Lambda) !void {
        const ret_sx_type = self.inferType(lambda.body);
        const ret_llvm_type = self.typeToLLVM(ret_sx_type);

        var param_llvm_types = std.ArrayList(c.LLVMTypeRef).empty;
        for (lambda.params) |param| {
            const sx_ty = self.resolveType(param.type_expr);
            try param_llvm_types.append(self.allocator, self.typeToLLVM(sx_ty));
        }
        const params_slice = try param_llvm_types.toOwnedSlice(self.allocator);

        const fn_type = c.LLVMFunctionType(
            ret_llvm_type,
            if (params_slice.len > 0) params_slice.ptr else null,
            @intCast(params_slice.len),
            0,
        );

        const name_z = try self.allocator.dupeZ(u8, name);
        _ = c.LLVMAddFunction(self.module, name_z.ptr, fn_type);
    }

    fn genLambdaBody(self: *CodeGen, name: []const u8, lambda: ast.Lambda) !void {
        self.named_values.clearRetainingCapacity();

        const ret_sx_type = self.inferType(lambda.body);
        self.current_return_type = ret_sx_type;

        const name_z = try self.allocator.dupeZ(u8, name);
        const function = c.LLVMGetNamedFunction(self.module, name_z.ptr) orelse return self.emitErrorFmt("lambda '{s}' not found in LLVM module", .{name});
        self.current_function = function;

        const entry = c.LLVMAppendBasicBlockInContext(self.context, function, "entry");
        c.LLVMPositionBuilderAtEnd(self.builder, entry);

        for (lambda.params, 0..) |param, i| {
            const sx_ty = self.resolveType(param.type_expr);
            try self.bindParam(function, param.name, sx_ty, @intCast(i));
        }

        const ret_val = try self.genExpr(lambda.body);
        if (ret_val) |val| {
            const src_ty = self.llvmTypeToSxType(c.LLVMTypeOf(val));
            const converted = self.convertValue(val, src_ty, ret_sx_type);
            _ = c.LLVMBuildRet(self.builder, converted);
        } else {
            _ = c.LLVMBuildRetVoid(self.builder);
        }
    }

    fn genStmt(self: *CodeGen, node: *Node) !c.LLVMValueRef {
        self.current_span = node.span;
        switch (node.data) {
            .var_decl => |vd| {
                return self.genVarDecl(vd);
            },
            .const_decl => |cd| {
                return self.genConstDecl(cd);
            },
            .fn_decl => |fd| {
                // Local declaration inside a function body
                if (fd.type_params.len > 0) {
                    // Generic template / type function: register for lazy instantiation
                    try self.generic_templates.put(fd.name, .{ .fd = fd });
                } else {
                    // Non-generic local function
                    // Save outer function state
                    const saved_fn = self.current_function;
                    const saved_bb = c.LLVMGetInsertBlock(self.builder);
                    const saved_ret = self.current_return_type;
                    const saved_named = self.named_values;
                    self.named_values = std.StringHashMap(NamedValue).init(self.allocator);

                    // Register with correct types (null return_type = void)
                    try self.registerFnDeclAs(fd, fd.name);

                    // Generate body inline
                    const ret_sx_type = self.resolveType(fd.return_type);
                    self.current_return_type = ret_sx_type;
                    const name_z = try self.allocator.dupeZ(u8, fd.name);
                    const function = c.LLVMGetNamedFunction(self.module, name_z.ptr) orelse
                        return self.emitErrorFmt("local function '{s}' not found", .{fd.name});
                    self.current_function = function;
                    const entry = c.LLVMAppendBasicBlockInContext(self.context, function, "entry");
                    c.LLVMPositionBuilderAtEnd(self.builder, entry);

                    for (fd.params, 0..) |param, i| {
                        const sx_ty = self.resolveType(param.type_expr);
                        try self.bindParam(function, param.name, sx_ty, @intCast(i));
                    }

                    var last_val: c.LLVMValueRef = null;
                    if (fd.body.data == .block) {
                        for (fd.body.data.block.stmts) |stmt| {
                            last_val = try self.genStmt(stmt);
                        }
                    } else {
                        last_val = try self.genExpr(fd.body);
                    }

                    const current_bb2 = c.LLVMGetInsertBlock(self.builder);
                    if (c.LLVMGetBasicBlockTerminator(current_bb2) == null) {
                        if (ret_sx_type == .void_type) {
                            _ = c.LLVMBuildRetVoid(self.builder);
                        } else if (last_val) |val| {
                            const src_ty = self.llvmTypeToSxType(c.LLVMTypeOf(val));
                            const converted = self.convertValue(val, src_ty, ret_sx_type);
                            _ = c.LLVMBuildRet(self.builder, converted);
                        } else {
                            _ = c.LLVMBuildRetVoid(self.builder);
                        }
                    }

                    // Restore outer function state
                    self.named_values = saved_named;
                    self.current_return_type = saved_ret;
                    self.current_function = saved_fn;
                    c.LLVMPositionBuilderAtEnd(self.builder, saved_bb);
                }
                return null;
            },
            .struct_decl => |sd| {
                try self.registerStructType(sd);
                return null;
            },
            .union_decl => |ud| {
                try self.registerUnionType(ud);
                return null;
            },
            .assignment => |asgn| {
                return self.genAssignment(asgn);
            },
            .return_stmt => |rs| {
                // Evaluate return value first, then emit all defers, then return
                if (rs.value) |val_node| {
                    const raw_val = try self.genExpr(val_node);
                    if (self.current_return_type.isStruct()) {
                        // Struct return: raw_val is an alloca pointer, load the value
                        const sname = self.current_return_type.struct_type;
                        const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                        const loaded = c.LLVMBuildLoad2(self.builder, info.llvm_type, raw_val, "retval");
                        try self.emitAllDefers();
                        _ = c.LLVMBuildRet(self.builder, loaded);
                    } else {
                        const src_ty = self.llvmTypeToSxType(c.LLVMTypeOf(raw_val));
                        const val = self.convertValue(raw_val, src_ty, self.current_return_type);
                        try self.emitAllDefers();
                        _ = c.LLVMBuildRet(self.builder, val);
                    }
                } else {
                    try self.emitAllDefers();
                    _ = c.LLVMBuildRetVoid(self.builder);
                }
                // Create a dead basic block for any subsequent instructions
                const dead_bb = c.LLVMAppendBasicBlockInContext(self.context, self.current_function, "after_ret");
                c.LLVMPositionBuilderAtEnd(self.builder, dead_bb);
                return null;
            },
            .defer_stmt => |ds| {
                // Don't generate now — push onto current defer list for later execution
                if (self.defer_stack.items.len > 0) {
                    const top = &self.defer_stack.items[self.defer_stack.items.len - 1];
                    try top.append(self.allocator, ds.expr);
                }
                return null;
            },
            .insert_expr => |ins| {
                // Substitute comptime param nodes before evaluation (e.g., replace $fmt identifier with literal)
                const expr = if (self.comptime_param_nodes != null)
                    try self.substituteComptimeNodes(ins.expr)
                else
                    ins.expr;
                // Evaluate the inner expression via bytecode VM to get a string, parse it, generate inline
                const result = try self.comptimeEval(expr, .string_type);
                const code_z = try self.allocator.dupeZ(u8, result.string_val);
                var parser = Parser.init(self.allocator, code_z);
                var last_val: c.LLVMValueRef = null;
                while (parser.current.tag != .eof) {
                    const stmt = try parser.parseStmt();
                    last_val = try self.genStmt(stmt);
                }
                return last_val;
            },
            else => {
                return self.genExpr(node);
            },
        }
    }

    fn genVarDecl(self: *CodeGen, vd: ast.VarDecl) !c.LLVMValueRef {
        // Meta type variable: x := f64 or x := Vec4 → runtime string holding the type name
        if (vd.value) |val| {
            const meta_name = self.asTypeName(val);
            if (meta_name) |raw_name| {
                const type_name = try self.allocator.dupeZ(u8, raw_name);
                const name_z = try self.allocator.dupeZ(u8, vd.name);
                const ptr_ty = c.LLVMPointerTypeInContext(self.context, 0);
                const alloca = self.buildEntryBlockAlloca(ptr_ty, name_z.ptr);
                const str_val = c.LLVMBuildGlobalStringPtr(self.builder, type_name.ptr, "type_name");
                _ = c.LLVMBuildStore(self.builder, str_val, alloca);
                try self.saveShadowed(vd.name);
                try self.named_values.put(vd.name, .{ .ptr = alloca, .ty = .{ .meta_type = .{ .name = raw_name } } });
                return null;
            }
        }

        var sx_ty: Type = Type.s(64);

        if (vd.type_annotation) |ta| {
            sx_ty = self.resolveType(ta);
        } else if (vd.value) |val| {
            // Infer type from value
            if (val.data == .struct_literal) {
                const sl = val.data.struct_literal;
                if (sl.struct_name) |name| {
                    sx_ty = .{ .struct_type = name };
                } else if (sl.type_expr) |te| {
                    sx_ty = self.resolveType(te);
                } else {
                    return self.emitError("cannot infer struct type from untyped struct literal");
                }
            } else if (val.data == .array_literal and val.data.array_literal.type_expr != null) {
                sx_ty = self.resolveType(val.data.array_literal.type_expr);
            } else {
                sx_ty = self.inferType(val);
            }
        } else {
            return self.emitErrorFmt("variable '{s}' has no type annotation and no initializer", .{vd.name});
        }

        // Struct-typed variable
        if (sx_ty.isStruct()) {
            // Resolve type aliases (e.g. Vec3 -> Vec__3_f32)
            const sname = self.type_aliases.get(sx_ty.struct_type) orelse sx_ty.struct_type;
            sx_ty = .{ .struct_type = sname };
            const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
            const name_z = try self.allocator.dupeZ(u8, vd.name);
            const alloca = self.buildEntryBlockAlloca(info.llvm_type, name_z.ptr);

            if (vd.value == null) {
                // Default-init: per-field defaults or zero
                try self.genStructDefaultInit(alloca, info);
            } else if (vd.value.?.data == .undef_literal) {
                // Undef-init: entire struct is undefined
                _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(info.llvm_type), alloca);
            } else if (vd.value.?.data == .struct_literal) {
                // Struct literal codegen returns an alloca — use it directly instead
                const lit_alloca = try self.genStructLiteral(vd.value.?.data.struct_literal, sname);
                try self.saveShadowed(vd.name);
                try self.named_values.put(vd.name, .{ .ptr = lit_alloca, .ty = sx_ty });
                return null;
            } else if (vd.value.?.data == .call) {
                // Function call returning a struct — result is a value, store to alloca
                const val = try self.genExpr(vd.value.?);
                _ = c.LLVMBuildStore(self.builder, val, alloca);
            } else {
                // General expression (xx cast, identifier, etc.) — evaluate as target type
                const val = try self.genExprAsType(vd.value.?, sx_ty);
                _ = c.LLVMBuildStore(self.builder, val, alloca);
            }

            try self.saveShadowed(vd.name);
            try self.named_values.put(vd.name, .{ .ptr = alloca, .ty = sx_ty });
            return null;
        }

        // Union-typed variable
        if (sx_ty.isUnion()) {
            const uname = self.type_aliases.get(sx_ty.union_type) orelse sx_ty.union_type;
            sx_ty = .{ .union_type = uname };
            const info = self.union_types.get(uname) orelse return self.emitErrorFmt("unknown union type '{s}'", .{uname});
            const name_z = try self.allocator.dupeZ(u8, vd.name);
            const alloca = self.buildEntryBlockAlloca(info.llvm_type, name_z.ptr);

            if (vd.value == null) {
                // Zero-init: tag=0, payload zeroed
                _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(info.llvm_type), alloca);
            } else if (vd.value.?.data == .undef_literal) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(info.llvm_type), alloca);
            } else if (vd.value.?.data == .union_literal) {
                const lit_alloca = try self.genUnionLiteral(vd.value.?.data.union_literal, uname);
                try self.saveShadowed(vd.name);
                try self.named_values.put(vd.name, .{ .ptr = lit_alloca, .ty = sx_ty });
                return null;
            } else if (vd.value.?.data == .enum_literal) {
                // Void variant: .none assigned to union variable
                const ul = ast.UnionLiteral{
                    .union_name = uname,
                    .variant_name = vd.value.?.data.enum_literal.name,
                    .payload = null,
                };
                const lit_alloca = try self.genUnionLiteral(ul, uname);
                try self.saveShadowed(vd.name);
                try self.named_values.put(vd.name, .{ .ptr = lit_alloca, .ty = sx_ty });
                return null;
            } else if (vd.value.?.data == .call) {
                // Call returning a union (e.g., Shape.circle(3.14)) — genExpr returns alloca
                const result_alloca = try self.genExpr(vd.value.?);
                const loaded = c.LLVMBuildLoad2(self.builder, info.llvm_type, result_alloca, "union_load");
                _ = c.LLVMBuildStore(self.builder, loaded, alloca);
            } else {
                // Other expression — try genExprAsType
                const result_alloca = try self.genExprAsType(vd.value.?, sx_ty);
                const loaded = c.LLVMBuildLoad2(self.builder, info.llvm_type, result_alloca, "union_load");
                _ = c.LLVMBuildStore(self.builder, loaded, alloca);
            }

            try self.saveShadowed(vd.name);
            try self.named_values.put(vd.name, .{ .ptr = alloca, .ty = sx_ty });
            return null;
        }

        // Array-typed variable
        if (sx_ty.isArray()) {
            const arr_info = sx_ty.array_type;
            const llvm_arr_ty = self.typeToLLVM(sx_ty);
            const arr_name_z = try self.allocator.dupeZ(u8, vd.name);
            const arr_alloca = self.buildEntryBlockAlloca(llvm_arr_ty, arr_name_z.ptr);

            if (vd.value == null) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(llvm_arr_ty), arr_alloca);
            } else if (vd.value.?.data == .undef_literal) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(llvm_arr_ty), arr_alloca);
            } else if (vd.value.?.data == .array_literal) {
                const al = vd.value.?.data.array_literal;
                const elem_sx_ty = Type.fromName(arr_info.element_name) orelse return self.emitErrorFmt("unknown array element type '{s}'", .{arr_info.element_name});
                const elem_llvm_ty = self.typeToLLVM(elem_sx_ty);
                const len = @min(al.elements.len, arr_info.length);
                for (0..len) |i| {
                    const val = try self.genExprAsType(al.elements[i], elem_sx_ty);
                    var indices = [_]c.LLVMValueRef{
                        c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0),
                        c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(i), 0),
                    };
                    const gep = c.LLVMBuildGEP2(self.builder, llvm_arr_ty, arr_alloca, &indices, 2, "arr_elem");
                    _ = c.LLVMBuildStore(self.builder, val, gep);
                }
                // Zero-init remaining elements
                for (len..arr_info.length) |i| {
                    var indices = [_]c.LLVMValueRef{
                        c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0),
                        c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(i), 0),
                    };
                    const gep = c.LLVMBuildGEP2(self.builder, llvm_arr_ty, arr_alloca, &indices, 2, "arr_elem");
                    _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(elem_llvm_ty), gep);
                }
            } else {
                return self.emitErrorFmt("unsupported initializer for array variable '{s}'", .{vd.name});
            }

            try self.saveShadowed(vd.name);
            try self.named_values.put(vd.name, .{ .ptr = arr_alloca, .ty = sx_ty });
            return null;
        }

        // Vector-typed variable
        if (sx_ty.isVector()) {
            const llvm_vec_ty = self.typeToLLVM(sx_ty);
            const vec_name_z = try self.allocator.dupeZ(u8, vd.name);
            const vec_alloca = self.buildEntryBlockAlloca(llvm_vec_ty, vec_name_z.ptr);

            if (vd.value == null) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(llvm_vec_ty), vec_alloca);
            } else if (vd.value.?.data == .undef_literal) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(llvm_vec_ty), vec_alloca);
            } else if (vd.value.?.data == .array_literal) {
                const vec_val = try self.genVectorLiteral(vd.value.?.data.array_literal, sx_ty);
                _ = c.LLVMBuildStore(self.builder, vec_val, vec_alloca);
            } else {
                // Expression (e.g. function call) returning a vector
                const val = try self.genExpr(vd.value.?);
                _ = c.LLVMBuildStore(self.builder, val, vec_alloca);
            }

            try self.saveShadowed(vd.name);
            try self.named_values.put(vd.name, .{ .ptr = vec_alloca, .ty = sx_ty });
            return null;
        }

        // Function pointer typed variable
        if (sx_ty.isFunctionType()) {
            const llvm_ty = c.LLVMPointerTypeInContext(self.context, 0);
            const name_z = try self.allocator.dupeZ(u8, vd.name);
            const alloca = self.buildEntryBlockAlloca(llvm_ty, name_z.ptr);

            if (vd.value == null) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(llvm_ty), alloca);
            } else if (vd.value.?.data == .undef_literal) {
                _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(llvm_ty), alloca);
            } else if (vd.value.?.data == .unary_op and vd.value.?.data.unary_op.op == .xx) {
                // xx cast: e.g. xx SDL_GL_GetProcAddress("glClear")
                const inner = vd.value.?.data.unary_op.operand;
                const val = try self.genExpr(inner);
                const src_ty = self.inferType(inner);
                const converted = self.convertValue(val, src_ty, sx_ty);
                _ = c.LLVMBuildStore(self.builder, converted, alloca);
            } else {
                // Direct assignment: identifier (function name) or other expression
                const val = try self.genExpr(vd.value.?);
                _ = c.LLVMBuildStore(self.builder, val, alloca);
            }

            try self.saveShadowed(vd.name);
            try self.named_values.put(vd.name, .{ .ptr = alloca, .ty = sx_ty });
            return null;
        }

        // Guard: void type cannot be allocated (would crash LLVM)
        if (sx_ty == .void_type) {
            return self.emitErrorFmt("cannot declare variable '{s}' with void type", .{vd.name});
        }

        // Non-struct types
        const llvm_ty = self.typeToLLVM(sx_ty);
        const name_z = try self.allocator.dupeZ(u8, vd.name);
        const alloca = self.buildEntryBlockAlloca(llvm_ty, name_z.ptr);

        if (vd.value == null) {
            // Default-init: zero
            _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(llvm_ty), alloca);
        } else if (vd.value.?.data == .undef_literal) {
            // Undef-init
            _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(llvm_ty), alloca);
        } else {
            const val = vd.value.?;
            const enum_name: ?[]const u8 = if (sx_ty.isEnum()) sx_ty.enum_type else null;
            const init_val = if (val.data == .enum_literal and enum_name != null)
                self.genEnumLiteral(val.data.enum_literal.name, enum_name.?)
            else if (vd.type_annotation != null)
                try self.genExprAsType(val, sx_ty)
            else
                try self.genExpr(val);
            _ = c.LLVMBuildStore(self.builder, init_val, alloca);
        }

        try self.saveShadowed(vd.name);
        try self.named_values.put(vd.name, .{ .ptr = alloca, .ty = sx_ty });
        return null;
    }

    fn genStructDefaultInit(self: *CodeGen, alloca: c.LLVMValueRef, info: StructInfo) !void {
        for (info.field_names, 0..) |_, fi| {
            const ft = info.field_types[fi];
            const ft_llvm = self.typeToLLVM(ft);
            const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, @intCast(fi), "dinit");

            if (info.field_defaults.len > fi and info.field_defaults[fi] != null) {
                const default_node = info.field_defaults[fi].?;
                if (default_node.data == .undef_literal) {
                    // Field default is --- → store undef
                    _ = c.LLVMBuildStore(self.builder, c.LLVMGetUndef(ft_llvm), gep);
                } else {
                    // Field has expression default → evaluate and convert
                    const val = try self.genExprAsType(default_node, ft);
                    if (ft.isStruct() or ft.isUnion()) {
                        // Aggregate types: genExprAsType returns an alloca, load the value first
                        const loaded = c.LLVMBuildLoad2(self.builder, ft_llvm, val, "dinit_load");
                        _ = c.LLVMBuildStore(self.builder, loaded, gep);
                    } else {
                        _ = c.LLVMBuildStore(self.builder, val, gep);
                    }
                }
            } else {
                // No default → zero
                _ = c.LLVMBuildStore(self.builder, c.LLVMConstNull(ft_llvm), gep);
            }
        }
    }

    fn genConstDecl(self: *CodeGen, cd: ast.ConstDecl) !c.LLVMValueRef {
        // Compile-time evaluation: register as comptime global for JIT
        if (cd.value.data == .comptime_expr) {
            const ct_type_override: ?Type = if (cd.type_annotation) |te| Type.fromTypeExpr(te) else null;
            try self.registerComptimeGlobal(cd.name, cd.value.data.comptime_expr.expr, ct_type_override);
            return null;
        }

        var sx_ty: Type = Type.s(64);

        if (cd.type_annotation) |ta| {
            sx_ty = self.resolveType(ta);
        } else {
            sx_ty = self.inferType(cd.value);
        }

        // Union-typed constant: delegate to genExprAsType which handles enum_literal + union_literal
        if (sx_ty.isUnion()) {
            const val = try self.genExprAsType(cd.value, sx_ty);
            try self.saveShadowed(cd.name);
            try self.named_values.put(cd.name, .{ .ptr = val, .ty = sx_ty });
            return null;
        }

        // Function pointer typed constant
        if (sx_ty.isFunctionType()) {
            const llvm_ty = c.LLVMPointerTypeInContext(self.context, 0);
            const name_z = try self.allocator.dupeZ(u8, cd.name);
            const alloca = self.buildEntryBlockAlloca(llvm_ty, name_z.ptr);
            if (cd.value.data == .unary_op and cd.value.data.unary_op.op == .xx) {
                const inner = cd.value.data.unary_op.operand;
                const val = try self.genExpr(inner);
                const src_inner_ty = self.inferType(inner);
                const converted = self.convertValue(val, src_inner_ty, sx_ty);
                _ = c.LLVMBuildStore(self.builder, converted, alloca);
            } else {
                const val = try self.genExpr(cd.value);
                _ = c.LLVMBuildStore(self.builder, val, alloca);
            }
            try self.saveShadowed(cd.name);
            try self.named_values.put(cd.name, .{ .ptr = alloca, .ty = sx_ty });
            return null;
        }

        const enum_name: ?[]const u8 = if (sx_ty.isEnum()) sx_ty.enum_type else null;
        const init_val = if (cd.value.data == .enum_literal and enum_name != null)
            self.genEnumLiteral(cd.value.data.enum_literal.name, enum_name.?)
        else
            try self.genExpr(cd.value);

        const llvm_ty = self.typeToLLVM(sx_ty);
        const name_z = try self.allocator.dupeZ(u8, cd.name);
        const alloca = self.buildEntryBlockAlloca(llvm_ty, name_z.ptr);
        _ = c.LLVMBuildStore(self.builder, init_val, alloca);
        try self.saveShadowed(cd.name);
        try self.named_values.put(cd.name, .{ .ptr = alloca, .ty = sx_ty });
        return null;
    }

    fn genCompoundOp(self: *CodeGen, op: ast.Assignment.Op, cur: c.LLVMValueRef, rhs: c.LLVMValueRef, ty: Type) c.LLVMValueRef {
        return switch (op) {
            .add_assign => if (ty.isFloat()) c.LLVMBuildFAdd(self.builder, cur, rhs, "addtmp") else c.LLVMBuildAdd(self.builder, cur, rhs, "addtmp"),
            .sub_assign => if (ty.isFloat()) c.LLVMBuildFSub(self.builder, cur, rhs, "subtmp") else c.LLVMBuildSub(self.builder, cur, rhs, "subtmp"),
            .mul_assign => if (ty.isFloat()) c.LLVMBuildFMul(self.builder, cur, rhs, "multmp") else c.LLVMBuildMul(self.builder, cur, rhs, "multmp"),
            .div_assign => if (ty.isFloat()) c.LLVMBuildFDiv(self.builder, cur, rhs, "divtmp") else if (ty.isUnsigned()) c.LLVMBuildUDiv(self.builder, cur, rhs, "divtmp") else c.LLVMBuildSDiv(self.builder, cur, rhs, "divtmp"),
            .mod_assign => if (ty.isFloat()) c.LLVMBuildFRem(self.builder, cur, rhs, "modtmp") else if (ty.isUnsigned()) c.LLVMBuildURem(self.builder, cur, rhs, "modtmp") else c.LLVMBuildSRem(self.builder, cur, rhs, "modtmp"),
            .assign => unreachable,
        };
    }

    fn genAssignment(self: *CodeGen, asgn: ast.Assignment) !c.LLVMValueRef {
        // Field assignment: expr.field = value;
        if (asgn.target.data == .field_access) {
            return self.genFieldAssignment(asgn);
        }

        // Index assignment: expr[i] = value;
        if (asgn.target.data == .index_expr) {
            return self.genIndexAssignment(asgn);
        }

        // Deref assignment: p.* = value;
        if (asgn.target.data == .deref_expr) {
            const de = asgn.target.data.deref_expr;
            const ptr_val = try self.genExpr(de.operand);
            const ptr_ty = self.inferType(de.operand);
            if (!ptr_ty.isPointer()) return self.emitError("dereference assignment requires a pointer");
            const pointee_ty = self.resolveTypeFromName(ptr_ty.pointer_type.pointee_name) orelse return self.emitError("unknown pointee type");
            const new_val = try self.genExprAsType(asgn.value, pointee_ty);
            _ = c.LLVMBuildStore(self.builder, new_val, ptr_val);
            return null;
        }

        // Target must be an identifier
        if (asgn.target.data != .identifier) return self.emitError("assignment target must be a variable");
        const name = asgn.target.data.identifier.name;
        const entry = self.named_values.get(name) orelse
            self.global_mutable_vars.get(name) orelse {
            return self.emitErrorFmt("undefined variable '{s}'", .{name});
        };

        // Meta type reassignment: x = Vec4, x = f64, x = test
        if (entry.ty == .meta_type and asgn.op == .assign) {
            const raw_name = self.asTypeName(asgn.value) orelse blk: {
                // Also accept function names as meta_type values (use signature)
                if (asgn.value.data == .identifier) {
                    const fn_name = asgn.value.data.identifier.name;
                    if (self.fn_signatures.get(fn_name)) |sig| break :blk sig;
                }
                break :blk @as(?[]const u8, null);
            };
            if (raw_name) |rn| {
                const type_name = try self.allocator.dupeZ(u8, rn);
                const str_val = c.LLVMBuildGlobalStringPtr(self.builder, type_name.ptr, "type_name");
                _ = c.LLVMBuildStore(self.builder, str_val, entry.ptr);
                if (self.named_values.getPtr(name)) |entry_ptr| {
                    entry_ptr.ty = .{ .meta_type = .{ .name = rn } };
                }
                return null;
            }
            return self.emitErrorFmt("cannot assign non-type value to Type variable '{s}'", .{name});
        }

        // Function pointer reassignment
        if (entry.ty.isFunctionType() and asgn.op == .assign) {
            if (asgn.value.data == .unary_op and asgn.value.data.unary_op.op == .xx) {
                const inner = asgn.value.data.unary_op.operand;
                const val = try self.genExpr(inner);
                const src_ty = self.inferType(inner);
                const converted = self.convertValue(val, src_ty, entry.ty);
                _ = c.LLVMBuildStore(self.builder, converted, entry.ptr);
            } else {
                const val = try self.genExpr(asgn.value);
                _ = c.LLVMBuildStore(self.builder, val, entry.ptr);
            }
            return null;
        }

        // Union reassignment: s = .circle(3.14) or s = .none
        if (entry.ty.isUnion() and asgn.op == .assign) {
            const new_alloca = try self.genExprAsType(asgn.value, entry.ty);
            // Copy from new alloca to existing alloca
            const info = self.union_types.get(entry.ty.union_type).?;
            const loaded = c.LLVMBuildLoad2(self.builder, info.llvm_type, new_alloca, "union_load");
            _ = c.LLVMBuildStore(self.builder, loaded, entry.ptr);
            return null;
        }

        const new_val = try self.genExpr(asgn.value);
        const llvm_ty = self.typeToLLVM(entry.ty);

        const store_val = if (asgn.op == .assign) new_val else blk: {
            const cur = c.LLVMBuildLoad2(self.builder, llvm_ty, entry.ptr, "cur");
            break :blk self.genCompoundOp(asgn.op, cur, new_val, entry.ty);
        };

        _ = c.LLVMBuildStore(self.builder, store_val, entry.ptr);
        return null;
    }

    fn genFieldAssignment(self: *CodeGen, asgn: ast.Assignment) !c.LLVMValueRef {
        const fa = asgn.target.data.field_access;

        // Object must be an identifier for now
        if (fa.object.data != .identifier) return self.emitError("field assignment target must be a variable");
        const obj_name = fa.object.data.identifier.name;
        const entry = self.named_values.get(obj_name) orelse return self.emitErrorFmt("undefined variable '{s}'", .{obj_name});

        // Pointer auto-deref: p.field = val
        if (entry.ty.isPointer()) {
            const pointee_ty = self.resolveTypeFromName(entry.ty.pointer_type.pointee_name) orelse
                return self.emitError("unknown pointee type for field assignment");
            if (pointee_ty.isStruct()) {
                const sname = pointee_ty.struct_type;
                const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                const fi = self.findFieldIndex(info, fa.field) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fa.field, sname });
                const field_ty = info.field_types[fi];
                const loaded_ptr = c.LLVMBuildLoad2(self.builder,
                    c.LLVMPointerTypeInContext(self.context, 0), entry.ptr, "ptr_load");
                const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, loaded_ptr, @intCast(fi), "pfield_ptr");
                const rhs = try self.genExprAsType(asgn.value, field_ty);
                if (asgn.op == .assign) {
                    _ = c.LLVMBuildStore(self.builder, rhs, gep);
                } else {
                    const field_llvm_ty = self.typeToLLVM(field_ty);
                    const cur = c.LLVMBuildLoad2(self.builder, field_llvm_ty, gep, "pcur");
                    _ = c.LLVMBuildStore(self.builder, self.genCompoundOp(asgn.op, cur, rhs, field_ty), gep);
                }
                return null;
            }
            return self.emitError("field assignment through pointer requires a struct pointee");
        }

        if (!entry.ty.isStruct()) return self.emitErrorFmt("field access on non-struct variable '{s}'", .{obj_name});

        const sname = entry.ty.struct_type;
        const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
        const fi = self.findFieldIndex(info, fa.field) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fa.field, sname });
        const field_ty = info.field_types[fi];

        const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, entry.ptr, @intCast(fi), "fassign");

        // Generate RHS and convert to field type
        const rhs = try self.genExprAsType(asgn.value, field_ty);

        if (asgn.op == .assign) {
            _ = c.LLVMBuildStore(self.builder, rhs, gep);
        } else {
            const field_llvm_ty = self.typeToLLVM(field_ty);
            const cur = c.LLVMBuildLoad2(self.builder, field_llvm_ty, gep, "fcur");
            _ = c.LLVMBuildStore(self.builder, self.genCompoundOp(asgn.op, cur, rhs, field_ty), gep);
        }
        return null;
    }

    fn genIndexAssignment(self: *CodeGen, asgn: ast.Assignment) !c.LLVMValueRef {
        const ie = asgn.target.data.index_expr;
        const obj_ty = self.inferType(ie.object);
        if (obj_ty == .string_type) {
            // String index assignment: s[i] = c
            const str_val = try self.genExpr(ie.object);
            const ptr = c.LLVMBuildExtractValue(self.builder, str_val, 0, "str_ptr");
            const idx = try self.genExpr(ie.index);
            const val = try self.genExpr(asgn.value);
            const i8_type = c.LLVMInt8TypeInContext(self.context);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const gep_ptr = c.LLVMBuildGEP2(self.builder, i8_type, ptr, &gep_indices, 1, "stridx");
            const byte_val = c.LLVMBuildTrunc(self.builder, val, i8_type, "trunc_byte");
            _ = c.LLVMBuildStore(self.builder, byte_val, gep_ptr);
            return null;
        }
        if (obj_ty.isArray()) {
            if (ie.object.data == .identifier) {
                if (self.named_values.get(ie.object.data.identifier.name)) |entry| {
                    const idx = try self.genExpr(ie.index);
                    const val = try self.genExpr(asgn.value);
                    const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
                    var indices = [_]c.LLVMValueRef{ zero, idx };
                    const gep_ptr = c.LLVMBuildGEP2(self.builder, self.typeToLLVM(obj_ty), entry.ptr, &indices, 2, "arridx");
                    _ = c.LLVMBuildStore(self.builder, val, gep_ptr);
                    return null;
                }
            }
            // struct.field[i] = val — GEP through struct to array field, then index
            if (ie.object.data == .field_access) {
                const field_ptr = try self.genAddressOf(ie.object);
                const idx = try self.genExpr(ie.index);
                const val = try self.genExpr(asgn.value);
                const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
                var indices = [_]c.LLVMValueRef{ zero, idx };
                const gep_ptr = c.LLVMBuildGEP2(self.builder, self.typeToLLVM(obj_ty), field_ptr, &indices, 2, "field_arridx");
                _ = c.LLVMBuildStore(self.builder, val, gep_ptr);
                return null;
            }
        }
        if (obj_ty.isSlice()) {
            const slice_info = obj_ty.slice_type;
            const elem_ty = Type.fromName(slice_info.element_name) orelse return self.emitError("unknown slice element type");
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            // Load slice value to get ptr
            const slice_val = blk: {
                if (ie.object.data == .identifier) {
                    if (self.named_values.get(ie.object.data.identifier.name)) |entry| {
                        break :blk c.LLVMBuildLoad2(self.builder, self.getStringStructType(), entry.ptr, "slice_load");
                    }
                }
                break :blk try self.genExpr(ie.object);
            };
            const ptr = c.LLVMBuildExtractValue(self.builder, slice_val, 0, "slice_ptr");
            const idx = try self.genExpr(ie.index);
            const val = try self.genExpr(asgn.value);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const gep_ptr = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, ptr, &gep_indices, 1, "sliceidx");
            _ = c.LLVMBuildStore(self.builder, val, gep_ptr);
            return null;
        }
        // Many-pointer index assignment: mp[i] = val
        if (obj_ty.isManyPointer()) {
            const elem_ty = self.resolveTypeFromName(obj_ty.many_pointer_type.element_name) orelse return self.emitError("unknown many-pointer element type");
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            const ptr_val = try self.genExpr(ie.object);
            const idx = try self.genExpr(ie.index);
            const val = try self.genExprAsType(asgn.value, elem_ty);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const gep_ptr = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, ptr_val, &gep_indices, 1, "mptridx");
            _ = c.LLVMBuildStore(self.builder, val, gep_ptr);
            return null;
        }
        return self.emitError("index assignment requires a string, array, slice, or [*] pointer target");
    }

    fn unescapeString(allocator: std.mem.Allocator, raw: []const u8) ![]u8 {
        var result = try allocator.alloc(u8, raw.len);
        var i: usize = 0;
        var j: usize = 0;
        while (i < raw.len) {
            if (raw[i] == '\\' and i + 1 < raw.len) {
                i += 1;
                switch (raw[i]) {
                    'n' => {
                        result[j] = '\n';
                    },
                    't' => {
                        result[j] = '\t';
                    },
                    'r' => {
                        result[j] = '\r';
                    },
                    '\\' => {
                        result[j] = '\\';
                    },
                    '"' => {
                        result[j] = '"';
                    },
                    '0' => {
                        result[j] = 0;
                    },
                    '`' => {
                        result[j] = '`';
                    },
                    else => {
                        result[j] = raw[i];
                    },
                }
                j += 1;
                i += 1;
            } else {
                result[j] = raw[i];
                j += 1;
                i += 1;
            }
        }
        return result[0..j];
    }

    fn genExpr(self: *CodeGen, node: *Node) anyerror!c.LLVMValueRef {
        self.current_span = node.span;
        switch (node.data) {
            .int_literal => |lit| {
                const i64_type = c.LLVMInt64TypeInContext(self.context);
                return c.LLVMConstInt(i64_type, @bitCast(@as(i64, lit.value)), 0);
            },
            .float_literal => |lit| {
                const f32_type = c.LLVMFloatTypeInContext(self.context);
                return c.LLVMConstReal(f32_type, lit.value);
            },
            .bool_literal => |lit| {
                const i1_type = c.LLVMInt1TypeInContext(self.context);
                return c.LLVMConstInt(i1_type, if (lit.value) 1 else 0, 0);
            },
            .string_literal => |lit| {
                const content = if (lit.is_raw) lit.raw else try unescapeString(self.allocator, lit.raw);
                const str_z = try self.allocator.dupeZ(u8, content);
                const ptr = c.LLVMBuildGlobalStringPtr(self.builder, str_z.ptr, "str");
                return self.buildStringSlice(ptr, @intCast(content.len));
            },
            .identifier => |ident| {
                if (self.named_values.get(ident.name)) |entry| {
                    const llvm_ty = self.typeToLLVM(entry.ty);
                    return c.LLVMBuildLoad2(self.builder, llvm_ty, entry.ptr, "loadtmp");
                }
                // Fall back to comptime globals (lazy resolution)
                if (self.comptime_globals.getPtr(ident.name)) |ct| {
                    if (!ct.is_resolved) {
                        try self.resolveComptimeGlobal(ct);
                    }
                    const llvm_ty = self.typeToLLVM(ct.ty);
                    return c.LLVMBuildLoad2(self.builder, llvm_ty, ct.global, "ct_load");
                }
                // Fall back to global mutable variables (e.g. function pointers from opengl.sx)
                if (self.global_mutable_vars.get(ident.name)) |entry| {
                    const llvm_ty = self.typeToLLVM(entry.ty);
                    return c.LLVMBuildLoad2(self.builder, llvm_ty, entry.ptr, "global_load");
                }
                // Fall back to function name → function pointer value
                {
                    const name_z = try self.allocator.dupeZ(u8, ident.name);
                    var fn_val = c.LLVMGetNamedFunction(self.module, name_z.ptr);
                    if (fn_val == null) {
                        // Try qualified name with current namespace
                        if (self.current_namespace) |ns| {
                            const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, ident.name });
                            const q_z = try self.allocator.dupeZ(u8, qualified);
                            fn_val = c.LLVMGetNamedFunction(self.module, q_z.ptr);
                        }
                    }
                    if (fn_val != null) return fn_val.?;
                }
                return self.emitErrorFmt("undefined identifier '{s}'", .{ident.name});
            },
            .binary_op => |binop| {
                if (binop.op == .and_op) return self.genAndOp(binop);
                if (binop.op == .or_op) return self.genOrOp(binop);
                const lhs_ty = self.inferType(binop.lhs);
                const rhs_ty = self.inferType(binop.rhs);
                const result_type = Type.widen(lhs_ty, rhs_ty);
                const lhs = try self.genExprAsType(binop.lhs, result_type);
                const rhs = try self.genExprAsType(binop.rhs, result_type);
                return self.genBinaryOp(binop.op, lhs, rhs, result_type);
            },
            .chained_comparison => |chain| {
                return self.genChainedComparison(chain);
            },
            .unary_op => |unop| {
                if (unop.op == .xx) {
                    // xx requires a target type context (assignment, declaration, argument, return)
                    return self.emitError("'xx' cast requires a target type context");
                }
                if (unop.op == .address_of) {
                    return self.genAddressOf(unop.operand);
                }
                const operand = try self.genExpr(unop.operand);
                return switch (unop.op) {
                    .negate => blk: {
                        const operand_ty = self.inferType(unop.operand);
                        if (operand_ty.isVector()) {
                            const elem_ty = operand_ty.vectorElementType();
                            break :blk if (elem_ty != null and elem_ty.?.isFloat())
                                c.LLVMBuildFNeg(self.builder, operand, "vnegtmp")
                            else
                                c.LLVMBuildNeg(self.builder, operand, "vnegtmp");
                        }
                        break :blk if (self.exprIsFloat(unop.operand))
                            c.LLVMBuildFNeg(self.builder, operand, "negtmp")
                        else
                            c.LLVMBuildNeg(self.builder, operand, "negtmp");
                    },
                    .not => c.LLVMBuildNot(self.builder, operand, "nottmp"),
                    .xx, .address_of => unreachable,
                };
            },
            .struct_literal => |sl| {
                const ctx_name: ?[]const u8 = if (self.current_return_type.isStruct()) self.current_return_type.struct_type else null;
                return self.genStructLiteral(sl, ctx_name);
            },
            .union_literal => |ul| {
                return self.genUnionLiteral(ul, null);
            },
            .array_literal => |al| {
                // Typed array/vector/slice literal: Type.[elems]
                if (al.type_expr) |te| {
                    const ty = self.resolveType(te);
                    if (ty.isVector()) return self.genVectorLiteral(al, ty);
                    if (ty.isArray()) return self.genArrayLiteral(al, ty);
                    if (ty.isSlice()) return self.genSliceLiteral(al, ty);
                }
                // If current return type is vector, build as vector SSA value
                if (self.current_return_type.isVector()) {
                    return self.genVectorLiteral(al, self.current_return_type);
                }
                return self.genArrayLiteral(al, null);
            },
            .field_access => |fa| {
                return self.genFieldAccess(fa);
            },
            .index_expr => |ie| {
                return self.genIndexExpr(ie);
            },
            .slice_expr => |se| {
                return self.genSliceExpr(se);
            },
            .call => |call_node| {
                return self.genCall(call_node);
            },
            .if_expr => |ie| {
                return self.genIfExpr(ie);
            },
            .match_expr => |me| {
                return self.genMatchExpr(me);
            },
            .while_expr => |we| {
                return self.genWhileExpr(we);
            },
            .for_expr => |fe| {
                return self.genForExpr(fe);
            },
            .break_expr => {
                if (self.loop_break_bb) |break_bb| {
                    _ = c.LLVMBuildBr(self.builder, break_bb);
                    const dead_bb = c.LLVMAppendBasicBlockInContext(self.context, self.current_function, "after_break");
                    c.LLVMPositionBuilderAtEnd(self.builder, dead_bb);
                    return null;
                }
                return self.emitError("'break' outside of loop");
            },
            .continue_expr => {
                if (self.loop_continue_bb) |continue_bb| {
                    _ = c.LLVMBuildBr(self.builder, continue_bb);
                    const dead_bb = c.LLVMAppendBasicBlockInContext(self.context, self.current_function, "after_continue");
                    c.LLVMPositionBuilderAtEnd(self.builder, dead_bb);
                    return null;
                }
                return self.emitError("'continue' outside of loop");
            },
            .block => |blk| {
                try self.pushScope();
                var last_val: c.LLVMValueRef = null;
                for (blk.stmts) |stmt| {
                    last_val = try self.genStmt(stmt);
                }
                try self.popScope();
                return last_val;
            },
            .var_decl => |vd| {
                return self.genVarDecl(vd);
            },
            .const_decl => |cd| {
                return self.genConstDecl(cd);
            },
            .assignment => |asgn| {
                return self.genAssignment(asgn);
            },
            .return_stmt => |rs| {
                if (rs.value) |val_node| {
                    const raw_val = try self.genExpr(val_node);
                    const src_ty = self.llvmTypeToSxType(c.LLVMTypeOf(raw_val));
                    const val = self.convertValue(raw_val, src_ty, self.current_return_type);
                    try self.emitAllDefers();
                    _ = c.LLVMBuildRet(self.builder, val);
                } else {
                    try self.emitAllDefers();
                    _ = c.LLVMBuildRetVoid(self.builder);
                }
                const dead_bb = c.LLVMAppendBasicBlockInContext(self.context, self.current_function, "after_ret");
                c.LLVMPositionBuilderAtEnd(self.builder, dead_bb);
                return null;
            },
            .deref_expr => |de| {
                const ptr_val = try self.genExpr(de.operand);
                const ptr_ty = self.inferType(de.operand);
                if (ptr_ty.isPointer()) {
                    const pointee_ty = self.resolveTypeFromName(ptr_ty.pointer_type.pointee_name) orelse return self.emitError("unknown pointee type");
                    return c.LLVMBuildLoad2(self.builder, self.typeToLLVM(pointee_ty), ptr_val, "deref");
                }
                return self.emitError("dereference requires a pointer type");
            },
            .null_literal => {
                return c.LLVMConstNull(c.LLVMPointerTypeInContext(self.context, 0));
            },
            .comptime_expr => |ct| {
                return self.genExpr(ct.expr);
            },
            else => return self.emitError("unsupported expression"),
        }
    }

    fn genAddressOf(self: *CodeGen, operand: *Node) !c.LLVMValueRef {
        // &x — return the alloca pointer of the variable
        if (operand.data == .identifier) {
            if (self.named_values.get(operand.data.identifier.name)) |entry| {
                return entry.ptr;
            }
            return self.emitErrorFmt("undefined variable '{s}'", .{operand.data.identifier.name});
        }
        // &expr[i] — return GEP pointer to the indexed element
        if (operand.data == .index_expr) {
            const ie = operand.data.index_expr;
            const obj_ty = self.inferType(ie.object);
            const idx = try self.genExpr(ie.index);

            if (obj_ty.isArray()) {
                if (ie.object.data == .identifier) {
                    if (self.named_values.get(ie.object.data.identifier.name)) |entry| {
                        const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
                        var indices = [_]c.LLVMValueRef{ zero, idx };
                        return c.LLVMBuildGEP2(self.builder, self.typeToLLVM(obj_ty), entry.ptr, &indices, 2, "addr_elem");
                    }
                }
            }
            if (obj_ty.isSlice() or obj_ty == .string_type) {
                const slice_val = try self.genExpr(ie.object);
                const ptr = c.LLVMBuildExtractValue(self.builder, slice_val, 0, "slice_ptr");
                const elem_ty = if (obj_ty.isSlice())
                    obj_ty.sliceElementType() orelse return self.emitError("unknown slice element type")
                else
                    Type.u(8);
                var gep_indices = [_]c.LLVMValueRef{idx};
                return c.LLVMBuildGEP2(self.builder, self.typeToLLVM(elem_ty), ptr, &gep_indices, 1, "addr_elem");
            }
        }
        // &s.field — return GEP pointer to the struct field
        if (operand.data == .field_access) {
            const fa = operand.data.field_access;
            if (fa.object.data == .identifier) {
                if (self.named_values.get(fa.object.data.identifier.name)) |entry| {
                    if (entry.ty.isStruct()) {
                        const sname = entry.ty.struct_type;
                        const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                        const idx = self.findFieldIndex(info, fa.field) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fa.field, sname });
                        return c.LLVMBuildStructGEP2(self.builder, info.llvm_type, entry.ptr, @intCast(idx), "addr_field");
                    }
                    // &p.field where p is *Struct — auto-deref through pointer
                    if (entry.ty.isPointer()) {
                        const pointee_name = entry.ty.pointer_type.pointee_name;
                        if (self.struct_types.get(pointee_name)) |info| {
                            const loaded_ptr = c.LLVMBuildLoad2(self.builder,
                                c.LLVMPointerTypeInContext(self.context, 0), entry.ptr, "ptr_load");
                            const idx = self.findFieldIndex(info, fa.field) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fa.field, pointee_name });
                            return c.LLVMBuildStructGEP2(self.builder, info.llvm_type, loaded_ptr, @intCast(idx), "addr_pfield");
                        }
                    }
                }
            }
        }
        return self.emitError("address-of requires a variable, index, or field expression");
    }

    const StructBuildResult = struct {
        field_sx_types: []const Type,
        llvm_type: c.LLVMTypeRef,
    };

    fn buildStructFields(self: *CodeGen, name: []const u8, field_type_nodes: []const *Node) !StructBuildResult {
        var field_sx_types = std.ArrayList(Type).empty;
        var field_llvm_types = std.ArrayList(c.LLVMTypeRef).empty;

        for (field_type_nodes) |ft| {
            const sx_ty = self.resolveType(ft);
            try field_sx_types.append(self.allocator, sx_ty);
            try field_llvm_types.append(self.allocator, self.typeToLLVM(sx_ty));
        }

        const llvm_types_slice = try field_llvm_types.toOwnedSlice(self.allocator);
        const name_z = try self.allocator.dupeZ(u8, name);
        const struct_ty = c.LLVMStructCreateNamed(self.context, name_z.ptr);
        c.LLVMStructSetBody(struct_ty, if (llvm_types_slice.len > 0) llvm_types_slice.ptr else null, @intCast(llvm_types_slice.len), 0);

        return .{
            .field_sx_types = try field_sx_types.toOwnedSlice(self.allocator),
            .llvm_type = struct_ty,
        };
    }

    const UnionBuildResult = struct {
        variant_sx_types: []const Type,
        llvm_type: c.LLVMTypeRef,
        max_payload_size: u64,
    };

    fn buildUnionFields(self: *CodeGen, name: []const u8, variant_type_nodes: []const ?*Node) !UnionBuildResult {
        var variant_sx_types = std.ArrayList(Type).empty;
        var max_payload_size: u64 = 0;
        const data_layout = c.LLVMGetModuleDataLayout(self.module);

        for (variant_type_nodes) |vt| {
            if (vt) |type_node| {
                const sx_ty = self.resolveType(type_node);
                try variant_sx_types.append(self.allocator, sx_ty);
                const llvm_ty = self.typeToLLVM(sx_ty);
                const size = c.LLVMStoreSizeOfType(data_layout, llvm_ty);
                if (size > max_payload_size) max_payload_size = size;
            } else {
                try variant_sx_types.append(self.allocator, .void_type);
            }
        }

        const name_z = try self.allocator.dupeZ(u8, name);
        const union_ty = c.LLVMStructCreateNamed(self.context, name_z.ptr);
        const i64_ty = c.LLVMInt64TypeInContext(self.context);
        const i8_ty = c.LLVMInt8TypeInContext(self.context);
        const payload_array_ty = c.LLVMArrayType2(i8_ty, max_payload_size);
        var union_fields = [2]c.LLVMTypeRef{ i64_ty, payload_array_ty };
        c.LLVMStructSetBody(union_ty, &union_fields, 2, 0);

        return .{
            .variant_sx_types = try variant_sx_types.toOwnedSlice(self.allocator),
            .llvm_type = union_ty,
            .max_payload_size = max_payload_size,
        };
    }

    fn hoistInlineTypeDecl(self: *CodeGen, parent_name: []const u8, child_name: []const u8, type_node: *Node) anyerror!void {
        const synthetic_name = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ parent_name, child_name });
        switch (type_node.data) {
            .struct_decl => |inline_sd| {
                var hoisted = inline_sd;
                hoisted.name = synthetic_name;
                try self.registerStructType(hoisted);
                type_node.data = .{ .type_expr = .{ .name = synthetic_name } };
            },
            .union_decl => |inline_ud| {
                var hoisted = inline_ud;
                hoisted.name = synthetic_name;
                try self.registerUnionType(hoisted);
                type_node.data = .{ .type_expr = .{ .name = synthetic_name } };
            },
            .enum_decl => |inline_ed| {
                try self.enum_types.put(synthetic_name, inline_ed.variants);
                _ = try self.getAnyTypeId(synthetic_name, .{ .enum_type = synthetic_name });
                type_node.data = .{ .type_expr = .{ .name = synthetic_name } };
            },
            else => {},
        }
    }

    fn registerStructType(self: *CodeGen, sd: ast.StructDecl) anyerror!void {
        // Generic struct: store as template instead of registering now
        if (sd.type_params.len > 0) {
            try self.generic_struct_templates.put(sd.name, .{ .sd = sd });
            return;
        }

        // Pre-pass: hoist inline type declarations from field types
        for (sd.field_types, 0..) |ft, i| {
            try self.hoistInlineTypeDecl(sd.name, sd.field_names[i], ft);
        }

        const build = try self.buildStructFields(sd.name, sd.field_types);

        // Process field defaults: replace #run expressions with comptime global references
        var resolved_defaults = try self.allocator.alloc(?*Node, sd.field_defaults.len);
        for (sd.field_defaults, 0..) |fd, i| {
            if (fd != null and fd.?.data == .comptime_expr) {
                const synthetic_name = try std.fmt.allocPrint(self.allocator, "__struct_{s}_field_{d}", .{ sd.name, i });
                const field_type_override: ?Type = if (i < build.field_sx_types.len) build.field_sx_types[i] else null;
                try self.registerComptimeGlobal(synthetic_name, fd.?.data.comptime_expr.expr, field_type_override);
                const id_node = try self.allocator.create(Node);
                id_node.* = .{ .span = .{ .start = 0, .end = 0 }, .data = .{ .identifier = .{ .name = synthetic_name } } };
                resolved_defaults[i] = id_node;
            } else {
                resolved_defaults[i] = fd;
            }
        }

        try self.struct_types.put(sd.name, .{
            .field_names = sd.field_names,
            .field_types = build.field_sx_types,
            .field_defaults = resolved_defaults,
            .llvm_type = build.llvm_type,
        });
        _ = try self.getAnyTypeId(sd.name, .{ .struct_type = sd.name });
    }

    fn registerUnionType(self: *CodeGen, ud: ast.UnionDecl) !void {
        // Pre-pass: hoist inline type declarations from variant types
        for (ud.variant_types, 0..) |vt_opt, i| {
            if (vt_opt) |vt| {
                try self.hoistInlineTypeDecl(ud.name, ud.variant_names[i], vt);
            }
        }

        const build = try self.buildUnionFields(ud.name, ud.variant_types);

        try self.union_types.put(ud.name, .{
            .variant_names = ud.variant_names,
            .variant_types = build.variant_sx_types,
            .llvm_type = build.llvm_type,
            .max_payload_size = build.max_payload_size,
        });
        _ = try self.getAnyTypeId(ud.name, .{ .union_type = ud.name });
    }

    fn genUnionLiteral(self: *CodeGen, ul: ast.UnionLiteral, expected_union_name: ?[]const u8) !c.LLVMValueRef {
        const uname = ul.union_name orelse expected_union_name orelse
            (if (self.current_return_type.isUnion()) self.current_return_type.union_type else null) orelse
            return self.emitError("cannot infer union type for literal");
        const resolved_name = self.type_aliases.get(uname) orelse uname;
        const info = self.union_types.get(resolved_name) orelse return self.emitErrorFmt("unknown union type '{s}'", .{resolved_name});

        // Find variant index
        var variant_idx: ?u32 = null;
        for (info.variant_names, 0..) |vn, i| {
            if (std.mem.eql(u8, vn, ul.variant_name)) {
                variant_idx = @intCast(i);
                break;
            }
        }
        const idx = variant_idx orelse return self.emitErrorFmt("no variant '{s}' in union '{s}'", .{ ul.variant_name, resolved_name });

        // Alloca union
        const alloca = self.buildEntryBlockAlloca(info.llvm_type, "union_tmp");
        const i64_ty = c.LLVMInt64TypeInContext(self.context);

        // Store tag (field 0)
        const tag_gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, 0, "tag");
        _ = c.LLVMBuildStore(self.builder, c.LLVMConstInt(i64_ty, idx, 0), tag_gep);

        // Store payload (field 1) if not void
        if (ul.payload) |payload_node| {
            const variant_ty = info.variant_types[idx];
            if (variant_ty != .void_type) {
                const payload_val = try self.genExprAsType(payload_node, variant_ty);
                const payload_gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, 1, "payload");
                const payload_llvm_ty = self.typeToLLVM(variant_ty);
                // Bitcast payload area to the variant's type pointer and store
                if (variant_ty.isStruct()) {
                    // Struct payload: load from alloca, store to payload area
                    const struct_val = c.LLVMBuildLoad2(self.builder, payload_llvm_ty, payload_val, "struct_load");
                    _ = c.LLVMBuildStore(self.builder, struct_val, payload_gep);
                } else {
                    _ = c.LLVMBuildStore(self.builder, payload_val, payload_gep);
                }
            }
        }

        return alloca;
    }

    fn genStructLiteral(self: *CodeGen, sl: ast.StructLiteral, expected_struct_name: ?[]const u8) anyerror!c.LLVMValueRef {
        const raw_name = sl.struct_name orelse blk: {
            if (sl.type_expr) |te| {
                const ty = self.resolveType(te);
                if (ty.isStruct()) break :blk ty.struct_type;
            }
            break :blk expected_struct_name orelse return self.emitError("cannot infer struct type for literal");
        };
        // Resolve type aliases (e.g. Vec3 -> Vec__3_f32)
        const sname = self.type_aliases.get(raw_name) orelse raw_name;
        const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});

        // Alloca the struct and default-init all fields (zero or declared defaults)
        const name_z = try self.allocator.dupeZ(u8, sname);
        const alloca = self.buildEntryBlockAlloca(info.llvm_type, name_z.ptr);
        try self.genStructDefaultInit(alloca, info);

        // Determine if this is named or positional mode
        var has_named = false;
        for (sl.field_inits) |fi| {
            if (fi.name != null) {
                has_named = true;
                break;
            }
        }

        if (has_named) {
            // Named/shorthand mode: map by field name
            for (sl.field_inits) |fi| {
                const fname = fi.name orelse {
                    // Positional field mixed with named — treat as identifier shorthand
                    if (fi.value.data == .identifier) {
                        const idx = self.findFieldIndex(info, fi.value.data.identifier.name) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fi.value.data.identifier.name, sname });
                        const val = try self.genExprAsType(fi.value, info.field_types[idx]);
                        const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, @intCast(idx), "field");
                        _ = c.LLVMBuildStore(self.builder, val, gep);
                        continue;
                    }
                    return self.emitError("mixed positional and named fields in struct literal");
                };
                const idx = self.findFieldIndex(info, fname) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fname, sname });
                const val = try self.genExprAsType(fi.value, info.field_types[idx]);
                const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, @intCast(idx), "field");
                _ = c.LLVMBuildStore(self.builder, val, gep);
            }
        } else {
            // Positional mode: assign in order
            for (sl.field_inits, 0..) |fi, i| {
                if (i >= info.field_names.len) return self.emitErrorFmt("too many fields in struct literal (expected {d})", .{info.field_names.len});
                const val = try self.genExprAsType(fi.value, info.field_types[i]);
                const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, @intCast(i), "field");
                _ = c.LLVMBuildStore(self.builder, val, gep);
            }
        }

        return alloca;
    }

    /// Generate an array literal as an alloca with elements stored via GEP.
    /// If target_ty is provided, elements are converted to the array's element type.
    /// Otherwise, element type is inferred from the first element.
    fn genArrayLiteral(self: *CodeGen, al: ast.ArrayLiteral, target_ty_opt: ?Type) !c.LLVMValueRef {
        const arr_ty: Type = target_ty_opt orelse blk: {
            // Infer from first element
            if (al.elements.len == 0) return self.emitError("cannot infer type of empty array literal");
            const elem_ty = self.inferType(al.elements[0]);
            const elem_name = try elem_ty.displayName(self.allocator);
            break :blk .{ .array_type = .{ .element_name = elem_name, .length = @intCast(al.elements.len) } };
        };
        const arr_info = arr_ty.array_type;
        const elem_sx_ty = Type.fromName(arr_info.element_name) orelse return self.emitErrorFmt("unknown array element type '{s}'", .{arr_info.element_name});
        const llvm_arr_ty = self.typeToLLVM(arr_ty);
        const alloca = self.buildEntryBlockAlloca(llvm_arr_ty, "arr");

        const len = @min(al.elements.len, arr_info.length);
        for (0..len) |i| {
            const val = try self.genExprAsType(al.elements[i], elem_sx_ty);
            var indices = [_]c.LLVMValueRef{
                c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0),
                c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(i), 0),
            };
            const gep = c.LLVMBuildGEP2(self.builder, llvm_arr_ty, alloca, &indices, 2, "arr_elem");
            _ = c.LLVMBuildStore(self.builder, val, gep);
        }
        return alloca;
    }

    fn genSliceLiteral(self: *CodeGen, al: ast.ArrayLiteral, slice_ty: Type) !c.LLVMValueRef {
        const elem_name = slice_ty.slice_type.element_name;
        const elem_sx_ty = Type.fromName(elem_name) orelse return self.emitErrorFmt("unknown slice element type '{s}'", .{elem_name});
        const n: u32 = @intCast(al.elements.len);

        // Create backing array [N]elem on the stack
        const arr_ty: Type = .{ .array_type = .{ .element_name = elem_name, .length = n } };
        const llvm_arr_ty = self.typeToLLVM(arr_ty);
        const arr_alloca = self.buildEntryBlockAlloca(llvm_arr_ty, "slice_backing");

        // Fill elements
        for (0..n) |i| {
            const val = try self.genExprAsType(al.elements[i], elem_sx_ty);
            var indices = [_]c.LLVMValueRef{
                c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0),
                c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(i), 0),
            };
            const gep = c.LLVMBuildGEP2(self.builder, llvm_arr_ty, arr_alloca, &indices, 2, "slice_elem");
            _ = c.LLVMBuildStore(self.builder, val, gep);
        }

        // Build slice {ptr, len}
        const i64_ty = c.LLVMInt64TypeInContext(self.context);
        const zero = c.LLVMConstInt(i64_ty, 0, 0);
        var gep_indices = [_]c.LLVMValueRef{ zero, zero };
        const elem_ptr = c.LLVMBuildGEP2(self.builder, llvm_arr_ty, arr_alloca, &gep_indices, 2, "slice_data");
        const slice_llvm_ty = self.getStringStructType();
        var slice_val = c.LLVMGetUndef(slice_llvm_ty);
        slice_val = c.LLVMBuildInsertValue(self.builder, slice_val, elem_ptr, 0, "slice_ptr");
        const len_val = c.LLVMConstInt(i64_ty, n, 0);
        slice_val = c.LLVMBuildInsertValue(self.builder, slice_val, len_val, 1, "slice_len");
        return slice_val;
    }

    fn genVectorLiteral(self: *CodeGen, al: ast.ArrayLiteral, vec_ty: Type) !c.LLVMValueRef {
        const vec_info = vec_ty.vector_type;
        const elem_sx_ty = Type.fromName(vec_info.element_name) orelse return self.emitErrorFmt("unknown vector element type '{s}'", .{vec_info.element_name});
        const llvm_vec_ty = self.typeToLLVM(vec_ty);
        var vec_val = c.LLVMGetUndef(llvm_vec_ty);

        const len = @min(al.elements.len, vec_info.length);
        for (0..len) |i| {
            const elem_val = try self.genExprAsType(al.elements[i], elem_sx_ty);
            const idx = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(i), 0);
            vec_val = c.LLVMBuildInsertElement(self.builder, vec_val, elem_val, idx, "vec_ins");
        }
        return vec_val;
    }

    fn broadcastScalar(self: *CodeGen, scalar: c.LLVMValueRef, vec_ty: Type) c.LLVMValueRef {
        const vec_info = vec_ty.vector_type;
        const llvm_vec_ty = self.typeToLLVM(vec_ty);
        // Insert scalar at index 0 of undef vector
        var vec = c.LLVMGetUndef(llvm_vec_ty);
        const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
        vec = c.LLVMBuildInsertElement(self.builder, vec, scalar, zero, "splat_ins");
        // Shuffle with zeroinitializer mask to broadcast element 0 to all lanes
        const mask_ty = c.LLVMVectorType(c.LLVMInt32TypeInContext(self.context), vec_info.length);
        const mask = c.LLVMConstNull(mask_ty);
        return c.LLVMBuildShuffleVector(self.builder, vec, c.LLVMGetUndef(llvm_vec_ty), mask, "splat");
    }

    fn genExprAsType(self: *CodeGen, node: *Node, target_ty: Type) !c.LLVMValueRef {
        self.current_span = node.span;
        // xx prefix: unwrap and convert freely (explicit cast)
        if (node.data == .unary_op and node.data.unary_op.op == .xx) {
            const inner = node.data.unary_op.operand;
            const val = try self.genExpr(inner);
            const src_ty = self.inferType(inner);
            return self.convertValue(val, src_ty, target_ty);
        }

        // Function pointer target: bypass narrowing check, just produce the pointer value
        if (target_ty.isFunctionType()) {
            return try self.genExpr(node);
        }

        // String literal → pointer context: produce raw pointer directly (no {ptr, len} wrapping)
        if (node.data == .string_literal and target_ty.isPointer()) {
            const lit = node.data.string_literal;
            const content = if (lit.is_raw) lit.raw else try unescapeString(self.allocator, lit.raw);
            const str_z = try self.allocator.dupeZ(u8, content);
            return c.LLVMBuildGlobalStringPtr(self.builder, str_z.ptr, "str");
        }

        // Enum literal assigned to union type: construct tag-only (void variant) union
        if (node.data == .enum_literal and target_ty.isUnion()) {
            const ul = ast.UnionLiteral{
                .union_name = null,
                .variant_name = node.data.enum_literal.name,
                .payload = null,
            };
            return self.genUnionLiteral(ul, target_ty.union_type);
        }

        // Union literal with target union type: pass context
        if (node.data == .union_literal and target_ty.isUnion()) {
            return self.genUnionLiteral(node.data.union_literal, target_ty.union_type);
        }

        // Struct literal targeting union type: .Variant.{fields} pattern
        // Parsed as struct_literal with type_expr = enum_literal("Variant")
        if (node.data == .struct_literal and target_ty.isUnion()) {
            const sl = node.data.struct_literal;
            if (sl.struct_name == null) {
                if (sl.type_expr) |te| {
                    if (te.data == .enum_literal) {
                        const variant_name = te.data.enum_literal.name;
                        const uname = self.type_aliases.get(target_ty.union_type) orelse target_ty.union_type;
                        const info = self.union_types.get(uname) orelse
                            return self.emitErrorFmt("unknown union type '{s}'", .{uname});

                        // Find variant index
                        var variant_idx: ?u32 = null;
                        for (info.variant_names, 0..) |vn, vi| {
                            if (std.mem.eql(u8, vn, variant_name)) {
                                variant_idx = @intCast(vi);
                                break;
                            }
                        }
                        const idx = variant_idx orelse
                            return self.emitErrorFmt("no variant '{s}' in union '{s}'", .{ variant_name, uname });

                        const variant_ty = info.variant_types[idx];

                        // Alloca union, store tag
                        const alloca = self.buildEntryBlockAlloca(info.llvm_type, "union_lit");
                        const i32_ty = c.LLVMInt32TypeInContext(self.context);
                        const tag_gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, 0, "tag");
                        _ = c.LLVMBuildStore(self.builder, c.LLVMConstInt(i32_ty, idx, 0), tag_gep);

                        // Store struct payload
                        if (variant_ty != .void_type) {
                            const payload_struct_name = if (variant_ty.isStruct()) variant_ty.struct_type else null;
                            const payload_alloca = try self.genStructLiteral(.{
                                .struct_name = payload_struct_name,
                                .type_expr = null,
                                .field_inits = sl.field_inits,
                            }, payload_struct_name);
                            const payload_gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, alloca, 1, "payload");
                            const payload_llvm_ty = self.typeToLLVM(variant_ty);
                            const struct_val = c.LLVMBuildLoad2(self.builder, payload_llvm_ty, payload_alloca, "struct_load");
                            _ = c.LLVMBuildStore(self.builder, struct_val, payload_gep);
                        }

                        return alloca;
                    }
                }
            }
        }

        // Struct literal targeting struct type: pass struct name context
        if (node.data == .struct_literal and target_ty.isStruct()) {
            const alloca = try self.genStructLiteral(node.data.struct_literal, target_ty.struct_type);
            // genStructLiteral returns an alloca pointer — load the value for by-value passing
            const sname = self.type_aliases.get(target_ty.struct_type) orelse target_ty.struct_type;
            if (self.struct_types.get(sname)) |si| {
                return c.LLVMBuildLoad2(self.builder, si.llvm_type, alloca, "struct_val");
            }
            return alloca;
        }

        // Array literal with target array type: generate with element conversion
        if (node.data == .array_literal and target_ty.isArray()) {
            return self.genArrayLiteral(node.data.array_literal, target_ty);
        }

        // Array literal with target vector type: build via undef + InsertElement
        if (node.data == .array_literal and target_ty.isVector()) {
            return self.genVectorLiteral(node.data.array_literal, target_ty);
        }

        // Array literal with target slice type: build stack-backed slice
        if (node.data == .array_literal and target_ty.isSlice()) {
            return self.genSliceLiteral(node.data.array_literal, target_ty);
        }

        // Array to slice coercion: [N]T → []T
        if (target_ty.isSlice()) {
            const src_ty = self.inferType(node);
            if (src_ty.isArray()) {
                const arr_info = src_ty.array_type;
                // Get the alloca pointer for the array (not the loaded value)
                const arr_alloca = blk: {
                    if (node.data == .identifier) {
                        if (self.named_values.get(node.data.identifier.name)) |entry| {
                            break :blk entry.ptr;
                        }
                    }
                    if (node.data == .field_access) {
                        break :blk try self.genAddressOf(node);
                    }
                    // Fallback: generate the expression and hope it returns a pointer
                    break :blk try self.genExpr(node);
                };
                // GEP to get pointer to first element
                const i64_ty = c.LLVMInt64TypeInContext(self.context);
                const zero = c.LLVMConstInt(i64_ty, 0, 0);
                var indices = [_]c.LLVMValueRef{ zero, zero };
                const elem_ptr = c.LLVMBuildGEP2(self.builder, self.typeToLLVM(src_ty), arr_alloca, &indices, 2, "arr_data");
                // Build slice struct {ptr, len}
                const slice_ty = self.getStringStructType();
                var slice_val = c.LLVMGetUndef(slice_ty);
                slice_val = c.LLVMBuildInsertValue(self.builder, slice_val, elem_ptr, 0, "slice_ptr");
                const len_val = c.LLVMConstInt(i64_ty, arr_info.length, 0);
                slice_val = c.LLVMBuildInsertValue(self.builder, slice_val, len_val, 1, "slice_len");
                return slice_val;
            }
        }

        // Array to many-pointer coercion: [N]T → [*]T
        if (target_ty.isManyPointer()) {
            const src_ty = self.inferType(node);
            if (src_ty.isArray()) {
                const arr_info = src_ty.array_type;
                if (std.mem.eql(u8, arr_info.element_name, target_ty.many_pointer_type.element_name)) {
                    const arr_alloca = blk: {
                        if (node.data == .identifier) {
                            if (self.named_values.get(node.data.identifier.name)) |entry| {
                                break :blk entry.ptr;
                            }
                        }
                        if (node.data == .field_access) {
                            break :blk try self.genAddressOf(node);
                        }
                        break :blk try self.genExpr(node);
                    };
                    const i64_ty = c.LLVMInt64TypeInContext(self.context);
                    const zero = c.LLVMConstInt(i64_ty, 0, 0);
                    var indices = [_]c.LLVMValueRef{ zero, zero };
                    return c.LLVMBuildGEP2(self.builder, self.typeToLLVM(src_ty), arr_alloca, &indices, 2, "arr_decay");
                }
            }
        }

        // Slice to many-pointer coercion: []T → [*]T (extract .ptr from fat pointer)
        if (target_ty.isManyPointer()) {
            const src_ty = self.inferType(node);
            if (src_ty.isSlice()) {
                if (std.mem.eql(u8, src_ty.slice_type.element_name, target_ty.many_pointer_type.element_name)) {
                    const slice_val = try self.genExpr(node);
                    return c.LLVMBuildExtractValue(self.builder, slice_val, 0, "slice_decay");
                }
            }
        }

        // Implicit address-of: passing T where *T is expected → auto &
        if (target_ty.isPointer()) {
            const src_ty = self.inferType(node);
            const pointee_name = target_ty.pointer_type.pointee_name;
            const src_matches = if (src_ty.isStruct())
                std.mem.eql(u8, src_ty.struct_type, pointee_name) or
                    (if (self.type_aliases.get(src_ty.struct_type)) |alias| std.mem.eql(u8, alias, pointee_name) else false) or
                    (if (self.type_aliases.get(pointee_name)) |alias| std.mem.eql(u8, alias, src_ty.struct_type) else false)
            else if (Type.fromName(pointee_name)) |pointee_ty|
                src_ty.eql(pointee_ty)
            else
                false;
            if (src_matches) {
                if (node.data == .identifier) {
                    if (self.named_values.get(node.data.identifier.name)) |entry| {
                        return entry.ptr;
                    }
                }
            }
        }

        var val = try self.genExpr(node);
        const src_ty = self.inferType(node);

        // Struct literals return alloca pointers — load the value for by-value passing
        if (src_ty.isStruct() and target_ty.isStruct()) {
            if (c.LLVMGetTypeKind(c.LLVMTypeOf(val)) == c.LLVMPointerTypeKind) {
                const info = self.struct_types.get(src_ty.struct_type) orelse
                    self.struct_types.get(self.type_aliases.get(src_ty.struct_type) orelse src_ty.struct_type);
                if (info) |si| {
                    val = c.LLVMBuildLoad2(self.builder, si.llvm_type, val, "struct_load");
                }
            }
        }

        // Scalar to vector broadcast
        if (target_ty.isVector() and !src_ty.isVector()) {
            const elem_ty = target_ty.vectorElementType() orelse return self.emitError("cannot determine vector element type");
            const converted = self.convertValue(val, src_ty, elem_ty);
            return self.broadcastScalar(converted, target_ty);
        }

        // Literals are exempt from narrowing checks
        if (node.data == .int_literal or node.data == .float_literal) {
            return self.convertValue(val, src_ty, target_ty);
        }

        // Check for narrowing conversion
        if (!src_ty.isImplicitlyConvertibleTo(target_ty)) {
            // Narrowing without xx — compile error
            return self.emitErrorFmt("narrowing conversion from '{s}' to '{s}' requires explicit 'xx' cast", .{
                src_ty.displayName(self.allocator) catch "?",
                target_ty.displayName(self.allocator) catch "?",
            });
        }

        return self.convertValue(val, src_ty, target_ty);
    }

    /// Convert an LLVM value from src_ty to target_ty, emitting appropriate casts.
    fn convertValue(self: *CodeGen, val: c.LLVMValueRef, src_ty: Type, target_ty: Type) c.LLVMValueRef {
        // Same type → return as-is
        if (src_ty.eql(target_ty)) return val;

        // string <-> []u8: identical LLVM type {ptr, i64}, no conversion needed
        if ((src_ty == .string_type and target_ty.isSlice() and
            std.mem.eql(u8, target_ty.slice_type.element_name, "u8")) or
            (src_ty.isSlice() and std.mem.eql(u8, src_ty.slice_type.element_name, "u8") and
            target_ty == .string_type))
            return val;

        const target_llvm = self.typeToLLVM(target_ty);

        // Any → concrete type: extract the i64 value and convert
        if (src_ty.isAny()) {
            const i64_val = c.LLVMBuildExtractValue(self.builder, val, 1, "any_extract");
            if (target_ty.isInt()) {
                if (target_ty.bitWidth() < 64) {
                    return c.LLVMBuildTrunc(self.builder, i64_val, target_llvm, "any_to_int");
                }
                return i64_val;
            }
            if (target_ty == .boolean) {
                return c.LLVMBuildTrunc(self.builder, i64_val, c.LLVMInt1TypeInContext(self.context), "any_to_bool");
            }
            if (target_ty == .f64) {
                return c.LLVMBuildBitCast(self.builder, i64_val, c.LLVMDoubleTypeInContext(self.context), "any_to_f64");
            }
            if (target_ty == .f32) {
                const as_f64 = c.LLVMBuildBitCast(self.builder, i64_val, c.LLVMDoubleTypeInContext(self.context), "any_f64_tmp");
                return c.LLVMBuildFPTrunc(self.builder, as_f64, c.LLVMFloatTypeInContext(self.context), "any_to_f32");
            }
            if (target_ty == .string_type) {
                // i64 is a pointer to {ptr, i32} on the stack
                const ptr = c.LLVMBuildIntToPtr(self.builder, i64_val, c.LLVMPointerTypeInContext(self.context, 0), "any_str_ptr");
                return c.LLVMBuildLoad2(self.builder, self.getStringStructType(), ptr, "any_to_str");
            }
            if (target_ty.isStruct()) {
                const sname = target_ty.struct_type;
                if (self.struct_types.get(sname)) |info| {
                    const ptr = c.LLVMBuildIntToPtr(self.builder, i64_val, c.LLVMPointerTypeInContext(self.context, 0), "any_struct_ptr");
                    return c.LLVMBuildLoad2(self.builder, info.llvm_type, ptr, "any_to_struct");
                }
            }
            if (target_ty.isEnum()) {
                return i64_val;
            }
            if (target_ty.isUnion()) {
                const uname = target_ty.union_type;
                if (self.union_types.get(uname)) |info| {
                    const ptr = c.LLVMBuildIntToPtr(self.builder, i64_val, c.LLVMPointerTypeInContext(self.context, 0), "any_union_ptr");
                    return c.LLVMBuildLoad2(self.builder, info.llvm_type, ptr, "any_to_union");
                }
            }
            if (target_ty.isPointer() or target_ty.isManyPointer()) {
                return c.LLVMBuildIntToPtr(self.builder, i64_val, c.LLVMPointerTypeInContext(self.context, 0), "any_to_ptr");
            }
            return i64_val;
        }

        // Float → float conversions
        if (src_ty.isFloat() and target_ty.isFloat()) {
            if (target_ty.bitWidth() > src_ty.bitWidth()) {
                return c.LLVMBuildFPExt(self.builder, val, target_llvm, "fext");
            } else {
                return c.LLVMBuildFPTrunc(self.builder, val, target_llvm, "ftrunc");
            }
        }

        // Int → float conversions
        if (src_ty.isInt() and target_ty.isFloat()) {
            if (src_ty.isSigned()) {
                return c.LLVMBuildSIToFP(self.builder, val, target_llvm, "sitofp");
            } else {
                return c.LLVMBuildUIToFP(self.builder, val, target_llvm, "uitofp");
            }
        }

        // Float → int conversions
        if (src_ty.isFloat() and target_ty.isInt()) {
            if (target_ty.isSigned()) {
                return c.LLVMBuildFPToSI(self.builder, val, target_llvm, "fptosi");
            } else {
                return c.LLVMBuildFPToUI(self.builder, val, target_llvm, "fptoui");
            }
        }

        // Pointer → int: PtrToInt
        if ((src_ty.isPointer() or src_ty.isManyPointer()) and target_ty.isInt()) {
            const as_i64 = c.LLVMBuildPtrToInt(self.builder, val, c.LLVMInt64TypeInContext(self.context), "ptrtoint");
            if (target_ty.bitWidth() < 64) {
                return c.LLVMBuildTrunc(self.builder, as_i64, target_llvm, "ptr_trunc");
            }
            return as_i64;
        }

        // Union → int: extract the tag field (index 0)
        if (src_ty.isUnion() and target_ty.isInt()) {
            const uname = src_ty.union_type;
            if (self.union_types.get(uname)) |info| {
                const tmp = self.buildEntryBlockAlloca(info.llvm_type, "union_cast");
                _ = c.LLVMBuildStore(self.builder, val, tmp);
                const tag_ptr = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, tmp, 0, "tag_ptr");
                const tag_val = c.LLVMBuildLoad2(self.builder, c.LLVMInt32TypeInContext(self.context), tag_ptr, "tag_val");
                if (target_ty.bitWidth() == 32) return tag_val;
                if (target_ty.bitWidth() > 32) return c.LLVMBuildSExt(self.builder, tag_val, target_llvm, "tag_ext");
                return c.LLVMBuildTrunc(self.builder, tag_val, target_llvm, "tag_trunc");
            }
        }

        // Int → int conversions
        if (src_ty.isInt() and target_ty.isInt()) {
            const sw = src_ty.bitWidth();
            const tw = target_ty.bitWidth();
            if (tw > sw) {
                // Extend — use SExt if source is signed, ZExt if unsigned
                if (src_ty.isSigned()) {
                    return c.LLVMBuildSExt(self.builder, val, target_llvm, "sext");
                } else {
                    return c.LLVMBuildZExt(self.builder, val, target_llvm, "zext");
                }
            } else if (tw < sw) {
                // Truncate
                return c.LLVMBuildTrunc(self.builder, val, target_llvm, "trunc");
            }
            // Same width, different signedness — no-op (bit pattern is the same)
            return val;
        }

        // Int → pointer/function_type: IntToPtr (for xx cast from integer to pointer)
        if (src_ty.isInt() and (target_ty.isPointer() or target_ty.isManyPointer() or target_ty.isFunctionType())) {
            return c.LLVMBuildIntToPtr(self.builder, val, c.LLVMPointerTypeInContext(self.context, 0), "inttoptr");
        }

        // Slice/string → pointer: extract .ptr from fat pointer
        if ((src_ty.isSlice() or src_ty == .string_type) and (target_ty.isPointer() or target_ty.isManyPointer())) {
            return c.LLVMBuildExtractValue(self.builder, val, 0, "slice_to_ptr");
        }

        // *[N]T → [*]T: pointer to array decays to many-pointer (both opaque ptrs, no-op)
        if (src_ty.isPointer() and target_ty.isManyPointer()) {
            return val;
        }

        // Pointer → function_type or function_type → pointer: both are opaque pointers, no-op
        if ((src_ty.isPointer() or src_ty.isManyPointer()) and target_ty.isFunctionType()) {
            return val;
        }
        if (src_ty.isFunctionType() and (target_ty.isPointer() or target_ty.isManyPointer())) {
            return val;
        }

        return val;
    }

    fn findFieldIndex(_: *CodeGen, info: StructInfo, name: []const u8) ?usize {
        for (info.field_names, 0..) |fn_name, i| {
            if (std.mem.eql(u8, fn_name, name)) return i;
        }
        return null;
    }

    fn componentToIndex(ch: u8) ?u32 {
        return switch (ch) {
            'x', 'r', 'u' => 0,
            'y', 'g', 'v' => 1,
            'z', 'b' => 2,
            'w', 'a' => 3,
            else => null,
        };
    }

    fn genSqrt(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("sqrt expects exactly 1 argument");
        const arg_val = try self.genExpr(call_node.args[0]);
        const arg_ty = self.inferType(call_node.args[0]);

        // Pick the right LLVM intrinsic based on float type
        const intrinsic_name: [*c]const u8 = if (std.meta.eql(arg_ty, Type.f64)) "llvm.sqrt.f64" else "llvm.sqrt.f32";
        const llvm_float_ty = if (std.meta.eql(arg_ty, Type.f64))
            c.LLVMDoubleTypeInContext(self.context)
        else
            c.LLVMFloatTypeInContext(self.context);

        // Get or declare the intrinsic
        var intrinsic_fn = c.LLVMGetNamedFunction(self.module, intrinsic_name);
        if (intrinsic_fn == null) {
            var param_types = [_]c.LLVMTypeRef{llvm_float_ty};
            const fn_type = c.LLVMFunctionType(llvm_float_ty, &param_types, 1, 0);
            intrinsic_fn = c.LLVMAddFunction(self.module, intrinsic_name, fn_type);
        }

        var args = [_]c.LLVMValueRef{arg_val};
        return c.LLVMBuildCall2(
            self.builder,
            c.LLVMGlobalGetValueType(intrinsic_fn.?),
            intrinsic_fn.?,
            &args,
            1,
            "sqrt",
        );
    }

    fn genSin(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("sin expects exactly 1 argument");
        const arg_val = try self.genExpr(call_node.args[0]);
        const arg_ty = self.inferType(call_node.args[0]);

        const intrinsic_name: [*c]const u8 = if (std.meta.eql(arg_ty, Type.f64)) "llvm.sin.f64" else "llvm.sin.f32";
        const llvm_float_ty = if (std.meta.eql(arg_ty, Type.f64))
            c.LLVMDoubleTypeInContext(self.context)
        else
            c.LLVMFloatTypeInContext(self.context);

        var intrinsic_fn = c.LLVMGetNamedFunction(self.module, intrinsic_name);
        if (intrinsic_fn == null) {
            var param_types = [_]c.LLVMTypeRef{llvm_float_ty};
            const fn_type = c.LLVMFunctionType(llvm_float_ty, &param_types, 1, 0);
            intrinsic_fn = c.LLVMAddFunction(self.module, intrinsic_name, fn_type);
        }

        var args = [_]c.LLVMValueRef{arg_val};
        return c.LLVMBuildCall2(self.builder, c.LLVMGlobalGetValueType(intrinsic_fn.?), intrinsic_fn.?, &args, 1, "sin");
    }

    fn genCos(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("cos expects exactly 1 argument");
        const arg_val = try self.genExpr(call_node.args[0]);
        const arg_ty = self.inferType(call_node.args[0]);

        const intrinsic_name: [*c]const u8 = if (std.meta.eql(arg_ty, Type.f64)) "llvm.cos.f64" else "llvm.cos.f32";
        const llvm_float_ty = if (std.meta.eql(arg_ty, Type.f64))
            c.LLVMDoubleTypeInContext(self.context)
        else
            c.LLVMFloatTypeInContext(self.context);

        var intrinsic_fn = c.LLVMGetNamedFunction(self.module, intrinsic_name);
        if (intrinsic_fn == null) {
            var param_types = [_]c.LLVMTypeRef{llvm_float_ty};
            const fn_type = c.LLVMFunctionType(llvm_float_ty, &param_types, 1, 0);
            intrinsic_fn = c.LLVMAddFunction(self.module, intrinsic_name, fn_type);
        }

        var args = [_]c.LLVMValueRef{arg_val};
        return c.LLVMBuildCall2(self.builder, c.LLVMGlobalGetValueType(intrinsic_fn.?), intrinsic_fn.?, &args, 1, "cos");
    }

    fn genSizeOf(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("size_of expects exactly 1 argument");
        const ty = self.resolveType(call_node.args[0]);
        if (std.meta.eql(ty, Type.void_type)) {
            return c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), 0, 0);
        }
        const llvm_ty = self.typeToLLVM(ty);
        const data_layout = c.LLVMGetModuleDataLayout(self.module);
        const size = c.LLVMStoreSizeOfType(data_layout, llvm_ty);
        return c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), size, 0);
    }

    fn genTypeOf(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("type_of expects exactly 1 argument");
        const arg = call_node.args[0];
        const arg_ty = self.inferType(arg);
        const i64_ty = c.LLVMInt64TypeInContext(self.context);

        // For Any values: extract the runtime tag (field 0)
        if (arg_ty.isAny()) {
            const val = try self.genExpr(arg);
            return c.LLVMBuildExtractValue(self.builder, val, 0, "type_of");
        }

        // For known types: return the constant tag value
        const tag: u64 = switch (arg_ty) {
            .void_type => ANY_TAG_VOID,
            .boolean => ANY_TAG_BOOL,
            .signed => |w| if (w <= 32) ANY_TAG_S32 else ANY_TAG_S64,
            .unsigned => |w| if (w <= 32) ANY_TAG_S32 else ANY_TAG_S64,
            .f32 => ANY_TAG_F32,
            .f64 => ANY_TAG_F64,
            .string_type => ANY_TAG_STRING,
            .struct_type => |name| try self.getAnyTypeId(name, arg_ty),
            .enum_type => |name| try self.getAnyTypeId(name, arg_ty),
            .union_type => |name| try self.getAnyTypeId(name, arg_ty),
            .meta_type => ANY_TAG_TYPE,
            else => ANY_TAG_S32,
        };
        return c.LLVMConstInt(i64_ty, tag, 0);
    }

    fn genTypeName(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("type_name expects exactly 1 argument");
        const ty = self.resolveType(call_node.args[0]);
        const name = try ty.displayName(self.allocator);
        return self.buildConstStr(name);
    }

    fn genFieldCount(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 1) return self.emitError("field_count expects exactly 1 argument");
        const ty = self.resolveType(call_node.args[0]);
        const i64_ty = c.LLVMInt64TypeInContext(self.context);
        if (ty.isStruct()) {
            const info = self.struct_types.get(ty.struct_type) orelse
                return self.emitErrorFmt("unknown struct type '{s}'", .{ty.struct_type});
            return c.LLVMConstInt(i64_ty, info.field_names.len, 0);
        }
        if (ty.isEnum()) {
            const variants = self.enum_types.get(ty.enum_type) orelse
                return self.emitErrorFmt("unknown enum type '{s}'", .{ty.enum_type});
            return c.LLVMConstInt(i64_ty, variants.len, 0);
        }
        if (ty.isVector()) {
            return c.LLVMConstInt(i64_ty, ty.vector_type.length, 0);
        }
        if (ty.isUnion()) {
            const info = self.union_types.get(ty.union_type) orelse
                return self.emitErrorFmt("unknown union type '{s}'", .{ty.union_type});
            return c.LLVMConstInt(i64_ty, info.variant_names.len, 0);
        }
        if (ty.isArray()) {
            return c.LLVMConstInt(i64_ty, ty.array_type.length, 0);
        }
        return self.emitError("field_count requires a struct, enum, vector, union, or array type");
    }

    fn genFieldName(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 2) return self.emitError("field_name expects 2 arguments: field_name(T, idx)");
        const ty = self.resolveType(call_node.args[0]);

        // Get the name list and type key
        const names: []const []const u8, const type_key: []const u8 = if (ty.isStruct()) blk: {
            const info = self.struct_types.get(ty.struct_type) orelse
                return self.emitErrorFmt("unknown struct type '{s}'", .{ty.struct_type});
            break :blk .{ info.field_names, ty.struct_type };
        } else if (ty.isEnum()) blk: {
            const variants = self.enum_types.get(ty.enum_type) orelse
                return self.emitErrorFmt("unknown enum type '{s}'", .{ty.enum_type});
            break :blk .{ variants, ty.enum_type };
        } else if (ty.isUnion()) blk: {
            const info = self.union_types.get(ty.union_type) orelse
                return self.emitErrorFmt("unknown union type '{s}'", .{ty.union_type});
            break :blk .{ info.variant_names, ty.union_type };
        } else return self.emitError("field_name requires a struct, enum, or union type");

        // Build a global array of string slices
        const n = names.len;
        const str_ty = self.getStringStructType();
        const arr_ty = c.LLVMArrayType2(str_ty, n);

        const vals = try self.allocator.alloc(c.LLVMValueRef, n);
        for (names, 0..) |name, i| {
            vals[i] = self.buildConstStrGlobal(name);
        }
        const arr_init = c.LLVMConstArray2(str_ty, vals.ptr, @intCast(n));
        const global_name = try self.allocator.dupeZ(u8, try std.fmt.allocPrint(self.allocator, "field_names.{s}", .{type_key}));
        var global = c.LLVMGetNamedGlobal(self.module, global_name.ptr);
        if (global == null) {
            global = c.LLVMAddGlobal(self.module, arr_ty, global_name.ptr);
            c.LLVMSetInitializer(global, arr_init);
            c.LLVMSetGlobalConstant(global, 1);
            c.LLVMSetLinkage(global, c.LLVMPrivateLinkage);
        }

        // GEP into the array with runtime index
        const idx = try self.genExpr(call_node.args[1]);
        const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
        var indices = [_]c.LLVMValueRef{ zero, idx };
        const elem_ptr = c.LLVMBuildGEP2(self.builder, arr_ty, global, &indices, 2, "field_name_ptr");
        return c.LLVMBuildLoad2(self.builder, str_ty, elem_ptr, "field_name");
    }

    fn genFieldValue(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 2) return self.emitError("field_value expects 2 arguments: field_value(s, idx)");

        const val = try self.genExpr(call_node.args[0]);
        const val_ty = self.inferType(call_node.args[0]);

        // Vector: extractelement + box as Any
        if (val_ty.isVector()) {
            const info = val_ty.vector_type;
            const elem_ty = Type.fromName(info.element_name) orelse
                return self.emitErrorFmt("unknown vector element type '{s}'", .{info.element_name});
            const idx = try self.genExpr(call_node.args[1]);
            const elem = c.LLVMBuildExtractElement(self.builder, val, idx, "vec_elem");
            return self.buildAnyValue(elem, elem_ty);
        }

        // Union: switch over tag, extract payload with correct type
        if (val_ty.isUnion()) {
            const uinfo = self.union_types.get(val_ty.union_type) orelse
                return self.emitErrorFmt("unknown union type '{s}'", .{val_ty.union_type});

            const union_alloca = self.buildEntryBlockAlloca(uinfo.llvm_type, "fv_union");
            _ = c.LLVMBuildStore(self.builder, val, union_alloca);

            // Read tag (field 0)
            const tag_ptr = c.LLVMBuildStructGEP2(self.builder, uinfo.llvm_type, union_alloca, 0, "fv_tag_ptr");
            const tag_val = c.LLVMBuildLoad2(self.builder, c.LLVMInt64TypeInContext(self.context), tag_ptr, "fv_tag");
            const payload_ptr = c.LLVMBuildStructGEP2(self.builder, uinfo.llvm_type, union_alloca, 1, "fv_payload_ptr");

            const n = uinfo.variant_names.len;
            const function = self.current_function;
            const any_ty = self.getAnyStructType();
            const merge_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "fv_merge");
            const default_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "fv_default");
            const sw = c.LLVMBuildSwitch(self.builder, tag_val, default_bb, @intCast(n));

            var phi_vals = std.ArrayList(c.LLVMValueRef).empty;
            var phi_bbs = std.ArrayList(c.LLVMBasicBlockRef).empty;

            for (uinfo.variant_types, 0..) |vty, vi| {
                const case_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "fv_ucase");
                c.LLVMAddCase(sw, c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), @intCast(vi), 0), case_bb);
                c.LLVMPositionBuilderAtEnd(self.builder, case_bb);

                const any_val = if (vty == .void_type) blk: {
                    // Void variant: return Any with void tag
                    const undef = c.LLVMGetUndef(any_ty);
                    const void_tag = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), ANY_TAG_VOID, 0);
                    const with_tag = c.LLVMBuildInsertValue(self.builder, undef, void_tag, 0, "void_tag");
                    const zero_val = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), 0, 0);
                    break :blk c.LLVMBuildInsertValue(self.builder, with_tag, zero_val, 1, "void_any");
                } else blk: {
                    const payload = c.LLVMBuildLoad2(self.builder, self.typeToLLVM(vty), payload_ptr, "fv_payload");
                    break :blk try self.buildAnyValue(payload, vty);
                };
                try phi_vals.append(self.allocator, any_val);
                try phi_bbs.append(self.allocator, c.LLVMGetInsertBlock(self.builder));
                _ = c.LLVMBuildBr(self.builder, merge_bb);
            }

            // Default: undef
            c.LLVMPositionBuilderAtEnd(self.builder, default_bb);
            try phi_vals.append(self.allocator, c.LLVMGetUndef(any_ty));
            try phi_bbs.append(self.allocator, default_bb);
            _ = c.LLVMBuildBr(self.builder, merge_bb);

            c.LLVMPositionBuilderAtEnd(self.builder, merge_bb);
            const vals_slice = try phi_vals.toOwnedSlice(self.allocator);
            const bbs_slice = try phi_bbs.toOwnedSlice(self.allocator);
            const phi = c.LLVMBuildPhi(self.builder, any_ty, "fv_uresult");
            c.LLVMAddIncoming(phi, vals_slice.ptr, bbs_slice.ptr, @intCast(vals_slice.len));
            return phi;
        }

        // Slice: extract ptr, GEP to element, load, box as Any
        if (val_ty.isSlice()) {
            const sinfo = val_ty.slice_type;
            const elem_ty = Type.fromName(sinfo.element_name) orelse
                return self.emitErrorFmt("unknown slice element type '{s}'", .{sinfo.element_name});
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            // val is {ptr, i32} — extract ptr
            const data_ptr = c.LLVMBuildExtractValue(self.builder, val, 0, "fv_sdata");
            const idx = try self.genExpr(call_node.args[1]);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const elem_ptr = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, data_ptr, &gep_indices, 1, "fv_selem");
            const elem = c.LLVMBuildLoad2(self.builder, elem_llvm_ty, elem_ptr, "fv_seval");
            return self.buildAnyValue(elem, elem_ty);
        }

        // Array: GEP + load + box as Any
        if (val_ty.isArray()) {
            const ainfo = val_ty.array_type;
            const elem_ty = Type.fromName(ainfo.element_name) orelse
                return self.emitErrorFmt("unknown array element type '{s}'", .{ainfo.element_name});
            const arr_llvm_ty = self.typeToLLVM(val_ty);
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            const arr_alloca = self.buildEntryBlockAlloca(arr_llvm_ty, "fv_arr");
            _ = c.LLVMBuildStore(self.builder, val, arr_alloca);
            const idx = try self.genExpr(call_node.args[1]);
            var gep_indices = [_]c.LLVMValueRef{
                c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0),
                idx,
            };
            const elem_ptr = c.LLVMBuildGEP2(self.builder, arr_llvm_ty, arr_alloca, &gep_indices, 2, "fv_aelem");
            const elem = c.LLVMBuildLoad2(self.builder, elem_llvm_ty, elem_ptr, "fv_aeval");
            return self.buildAnyValue(elem, elem_ty);
        }

        // Struct: switch over field indices
        const struct_val = val;
        const struct_ty = val_ty;
        if (!struct_ty.isStruct()) return self.emitError("field_value requires a struct, vector, union, or array value");

        const info = self.struct_types.get(struct_ty.struct_type) orelse
            return self.emitErrorFmt("unknown struct type '{s}'", .{struct_ty.struct_type});

        const idx = try self.genExpr(call_node.args[1]);
        const n = info.field_names.len;

        // Store struct to alloca BEFORE the switch (switch is a terminator)
        const struct_alloca = self.buildEntryBlockAlloca(info.llvm_type, "fv_struct");
        _ = c.LLVMBuildStore(self.builder, struct_val, struct_alloca);

        // Generate switch on idx with N cases
        const function = self.current_function;
        const merge_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "fv_merge");
        const default_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "fv_default");
        const sw = c.LLVMBuildSwitch(self.builder, idx, default_bb, @intCast(n));

        const any_ty = self.getAnyStructType();
        var phi_vals = std.ArrayList(c.LLVMValueRef).empty;
        var phi_bbs = std.ArrayList(c.LLVMBasicBlockRef).empty;

        for (0..n) |i| {
            const case_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "fv_case");
            const case_val = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), i, 0);
            c.LLVMAddCase(sw, case_val, case_bb);

            c.LLVMPositionBuilderAtEnd(self.builder, case_bb);
            // Extract field i via GEP + load
            const field_ptr = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, struct_alloca, @intCast(i), "fv_field_ptr");
            const field_llvm_ty = c.LLVMStructGetTypeAtIndex(info.llvm_type, @intCast(i));
            const field_val = c.LLVMBuildLoad2(self.builder, field_llvm_ty, field_ptr, "fv_field");
            const any_val = try self.buildAnyValue(field_val, info.field_types[i]);
            try phi_vals.append(self.allocator, any_val);
            try phi_bbs.append(self.allocator, c.LLVMGetInsertBlock(self.builder));
            _ = c.LLVMBuildBr(self.builder, merge_bb);
        }

        // Default: return undef Any
        c.LLVMPositionBuilderAtEnd(self.builder, default_bb);
        try phi_vals.append(self.allocator, c.LLVMGetUndef(any_ty));
        try phi_bbs.append(self.allocator, default_bb);
        _ = c.LLVMBuildBr(self.builder, merge_bb);

        c.LLVMPositionBuilderAtEnd(self.builder, merge_bb);
        const vals_slice = try phi_vals.toOwnedSlice(self.allocator);
        const bbs_slice = try phi_bbs.toOwnedSlice(self.allocator);
        const phi = c.LLVMBuildPhi(self.builder, any_ty, "fv_result");
        c.LLVMAddIncoming(phi, vals_slice.ptr, bbs_slice.ptr, @intCast(vals_slice.len));
        return phi;
    }

    fn genCast(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.args.len != 2) return self.emitError("cast expects: cast(Type) expr");
        const target_ty = self.resolveType(call_node.args[0]);
        const src_ty = self.inferType(call_node.args[1]);
        const val = try self.genExpr(call_node.args[1]);
        return self.convertValue(val, src_ty, target_ty);
    }

    fn genAlloc(self: *CodeGen, args: []const *Node) !c.LLVMValueRef {
        if (args.len != 1) return self.emitError("alloc expects exactly 1 argument: alloc(size)");
        const builtins = self.builtins orelse return self.emitError("builtins not available (missing #builtin import)");
        const size_val = try self.genExpr(args[0]);
        const i64_type = c.LLVMInt64TypeInContext(self.context);
        // calloc(size + 1, 1) — extra byte for null terminator
        const one_i64 = c.LLVMConstInt(i64_type, 1, 0);
        const size_plus_one = c.LLVMBuildAdd(self.builder, size_val, one_i64, "szp1");
        const calloc_fn = builtins.calloc_fn;
        const calloc_ty = c.LLVMGlobalGetValueType(calloc_fn);
        var calloc_args = [_]c.LLVMValueRef{ size_plus_one, one_i64 };
        const ptr = c.LLVMBuildCall2(self.builder, calloc_ty, calloc_fn, &calloc_args, 2, "alloc_ptr");
        // Build string slice: {ptr, size}
        return self.buildStringSliceRT(ptr, size_val);
    }

    fn genMalloc(self: *CodeGen, args: []const *Node) !c.LLVMValueRef {
        if (args.len != 1) return self.emitError("malloc expects exactly 1 argument: malloc(size)");
        const builtins = self.builtins orelse return self.emitError("builtins not available");
        const size_val = try self.genExpr(args[0]);
        const fn_ty = c.LLVMGlobalGetValueType(builtins.malloc_fn);
        var call_args = [_]c.LLVMValueRef{size_val};
        return c.LLVMBuildCall2(self.builder, fn_ty, builtins.malloc_fn, &call_args, 1, "malloc_ptr");
    }

    fn genFree(self: *CodeGen, args: []const *Node) !c.LLVMValueRef {
        if (args.len != 1) return self.emitError("free expects exactly 1 argument: free(ptr)");
        const builtins = self.builtins orelse return self.emitError("builtins not available");
        const ptr_val = try self.genExpr(args[0]);
        const fn_ty = c.LLVMGlobalGetValueType(builtins.free_fn);
        var call_args = [_]c.LLVMValueRef{ptr_val};
        _ = c.LLVMBuildCall2(self.builder, fn_ty, builtins.free_fn, &call_args, 1, "");
        return null;
    }

    fn genMemcpy(self: *CodeGen, args: []const *Node) !c.LLVMValueRef {
        if (args.len != 3) return self.emitError("memcpy expects 3 arguments: memcpy(dst, src, size)");
        const builtins = self.builtins orelse return self.emitError("builtins not available");
        const dst = try self.genExpr(args[0]);
        const src = try self.genExpr(args[1]);
        const size_val = try self.genExpr(args[2]);
        const fn_ty = c.LLVMGlobalGetValueType(builtins.memcpy_fn);
        var call_args = [_]c.LLVMValueRef{ dst, src, size_val };
        _ = c.LLVMBuildCall2(self.builder, fn_ty, builtins.memcpy_fn, &call_args, 3, "");
        return null;
    }

    fn genVectorExtract(self: *CodeGen, vec_val: c.LLVMValueRef, field: []const u8) !c.LLVMValueRef {
        if (field.len == 1) {
            const idx_val = componentToIndex(field[0]) orelse return self.emitErrorFmt("invalid vector component '{c}'", .{field[0]});
            const idx = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), idx_val, 0);
            return c.LLVMBuildExtractElement(self.builder, vec_val, idx, "comp");
        }
        return self.emitErrorFmt("unsupported vector swizzle '{s}'", .{field});
    }

    fn genFieldAccess(self: *CodeGen, fa: ast.FieldAccess) !c.LLVMValueRef {
        // Check if the object is a struct or vector variable
        if (fa.object.data == .identifier) {
            if (self.named_values.get(fa.object.data.identifier.name)) |entry| {
                // Pointer auto-deref: p.field → p.*.field
                if (entry.ty.isPointer()) {
                    const pointee_ty = self.resolveTypeFromName(entry.ty.pointer_type.pointee_name) orelse
                        return self.emitError("unknown pointee type for auto-deref");
                    const loaded_ptr = c.LLVMBuildLoad2(self.builder,
                        c.LLVMPointerTypeInContext(self.context, 0), entry.ptr, "ptr_load");
                    if (pointee_ty.isStruct()) {
                        const sname = pointee_ty.struct_type;
                        const info = self.struct_types.get(sname) orelse
                            return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                        const idx = self.findFieldIndex(info, fa.field) orelse
                            return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fa.field, sname });
                        const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, loaded_ptr,
                            @intCast(idx), "pfield");
                        return c.LLVMBuildLoad2(self.builder, self.typeToLLVM(info.field_types[idx]), gep, "pfieldval");
                    }
                    if (pointee_ty.isSlice()) {
                        const slice_val = c.LLVMBuildLoad2(self.builder, self.getStringStructType(), loaded_ptr, "pslice_load");
                        if (std.mem.eql(u8, fa.field, "len")) {
                            return c.LLVMBuildExtractValue(self.builder, slice_val, 1, "pslice_len");
                        }
                        if (std.mem.eql(u8, fa.field, "ptr")) {
                            return c.LLVMBuildExtractValue(self.builder, slice_val, 0, "pslice_ptr");
                        }
                        return self.emitErrorFmt("no field '{s}' on *slice (available: .len, .ptr)", .{fa.field});
                    }
                    return self.emitErrorFmt("no field '{s}' on pointer", .{fa.field});
                }
                if (entry.ty.isStruct()) {
                    const sname = entry.ty.struct_type;
                    const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                    const idx = self.findFieldIndex(info, fa.field) orelse return self.emitErrorFmt("no field '{s}' in struct '{s}'", .{ fa.field, sname });
                    const gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, entry.ptr, @intCast(idx), "field");
                    return c.LLVMBuildLoad2(self.builder, self.typeToLLVM(info.field_types[idx]), gep, "fieldval");
                }
                if (entry.ty.isUnion()) {
                    const uname = entry.ty.union_type;
                    const info = self.union_types.get(uname) orelse return self.emitErrorFmt("unknown union type '{s}'", .{uname});
                    // Find variant by name to determine payload type
                    var vidx: ?usize = null;
                    for (info.variant_names, 0..) |vn, i| {
                        if (std.mem.eql(u8, vn, fa.field)) {
                            vidx = i;
                            break;
                        }
                    }
                    const idx = vidx orelse return self.emitErrorFmt("no variant '{s}' in union '{s}'", .{ fa.field, uname });
                    const variant_ty = info.variant_types[idx];
                    if (variant_ty == .void_type) return self.emitErrorFmt("cannot access payload of void variant '{s}'", .{fa.field});
                    // GEP to field 1 (payload area), load as variant type
                    const payload_gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, entry.ptr, 1, "payload");
                    return c.LLVMBuildLoad2(self.builder, self.typeToLLVM(variant_ty), payload_gep, "union_payload");
                }
                if (entry.ty.isVector()) {
                    const vec_val = c.LLVMBuildLoad2(self.builder, self.typeToLLVM(entry.ty), entry.ptr, "vec_load");
                    return self.genVectorExtract(vec_val, fa.field);
                }
                if (entry.ty == .string_type) {
                    const str_val = c.LLVMBuildLoad2(self.builder, self.getStringStructType(), entry.ptr, "str_load");
                    if (std.mem.eql(u8, fa.field, "len")) {
                        return c.LLVMBuildExtractValue(self.builder, str_val, 1, "str_len");
                    }
                    if (std.mem.eql(u8, fa.field, "ptr")) {
                        return c.LLVMBuildExtractValue(self.builder, str_val, 0, "str_ptr");
                    }
                    return self.emitErrorFmt("no field '{s}' on string (available: .len, .ptr)", .{fa.field});
                }
                if (entry.ty.isSlice()) {
                    const slice_val = c.LLVMBuildLoad2(self.builder, self.getStringStructType(), entry.ptr, "slice_load");
                    if (std.mem.eql(u8, fa.field, "len")) {
                        return c.LLVMBuildExtractValue(self.builder, slice_val, 1, "slice_len");
                    }
                    if (std.mem.eql(u8, fa.field, "ptr")) {
                        return c.LLVMBuildExtractValue(self.builder, slice_val, 0, "slice_ptr");
                    }
                    return self.emitErrorFmt("no field '{s}' on slice (available: .len, .ptr)", .{fa.field});
                }
                if (entry.ty.isArray()) {
                    if (std.mem.eql(u8, fa.field, "len")) {
                        return c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), entry.ty.array_type.length, 0);
                    }
                    return self.emitErrorFmt("no field '{s}' on array (available: .len)", .{fa.field});
                }
                if (entry.ty.isAny()) {
                    const any_val = c.LLVMBuildLoad2(self.builder, self.getAnyStructType(), entry.ptr, "any_load");
                    if (std.mem.eql(u8, fa.field, "tag")) {
                        return c.LLVMBuildExtractValue(self.builder, any_val, 0, "any_tag");
                    }
                    if (std.mem.eql(u8, fa.field, "value")) {
                        return c.LLVMBuildExtractValue(self.builder, any_val, 1, "any_value");
                    }
                    return self.emitErrorFmt("no field '{s}' on Any (available: .tag, .value)", .{fa.field});
                }
            }
        }
        // Non-identifier object: evaluate expression and check type
        const obj_val = try self.genExpr(fa.object);
        const obj_ty = self.inferType(fa.object);
        if (obj_ty.isVector()) {
            return self.genVectorExtract(obj_val, fa.field);
        }
        if (obj_ty == .string_type) {
            if (std.mem.eql(u8, fa.field, "len")) {
                return c.LLVMBuildExtractValue(self.builder, obj_val, 1, "str_len");
            }
            if (std.mem.eql(u8, fa.field, "ptr")) {
                return c.LLVMBuildExtractValue(self.builder, obj_val, 0, "str_ptr");
            }
            return self.emitErrorFmt("no field '{s}' on string (available: .len, .ptr)", .{fa.field});
        }
        return self.emitError("field access on non-struct/non-vector expression");
    }

    fn genVectorComparison(self: *CodeGen, op: ast.BinaryOp.Op, lhs: c.LLVMValueRef, rhs: c.LLVMValueRef, vec_ty: Type, elem_ty: Type) c.LLVMValueRef {
        const vec_info = vec_ty.vector_type;
        const cmp = if (elem_ty.isFloat())
            (if (op == .eq) c.LLVMBuildFCmp(self.builder, c.LLVMRealOEQ, lhs, rhs, "vcmp") else c.LLVMBuildFCmp(self.builder, c.LLVMRealONE, lhs, rhs, "vcmp"))
        else
            (if (op == .eq) c.LLVMBuildICmp(self.builder, c.LLVMIntEQ, lhs, rhs, "vcmp") else c.LLVMBuildICmp(self.builder, c.LLVMIntNE, lhs, rhs, "vcmp"));
        // Reduce: extract each i1 and AND (eq) or OR (neq)
        var result = c.LLVMBuildExtractElement(self.builder, cmp, c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0), "cmp0");
        for (1..vec_info.length) |i| {
            const elem = c.LLVMBuildExtractElement(self.builder, cmp, c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(i), 0), "cmpi");
            result = if (op == .eq)
                c.LLVMBuildAnd(self.builder, result, elem, "andcmp")
            else
                c.LLVMBuildOr(self.builder, result, elem, "orcmp");
        }
        return result;
    }

    fn genIndexExpr(self: *CodeGen, ie: ast.IndexExpr) !c.LLVMValueRef {
        const obj_ty = self.inferType(ie.object);
        if (obj_ty.isVector()) {
            const vec_val = try self.genExpr(ie.object);
            const idx = try self.genExpr(ie.index);
            return c.LLVMBuildExtractElement(self.builder, vec_val, idx, "vidx");
        }
        if (obj_ty.isArray()) {
            const arr_info = obj_ty.array_type;
            const elem_ty = Type.fromName(arr_info.element_name) orelse return self.emitErrorFmt("unknown array element type '{s}'", .{arr_info.element_name});
            // Array index via identifier: load from GEP
            if (ie.object.data == .identifier) {
                if (self.named_values.get(ie.object.data.identifier.name)) |entry| {
                    const idx = try self.genExpr(ie.index);
                    const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
                    var indices = [_]c.LLVMValueRef{ zero, idx };
                    const gep = c.LLVMBuildGEP2(self.builder, self.typeToLLVM(obj_ty), entry.ptr, &indices, 2, "arridx");
                    return c.LLVMBuildLoad2(self.builder, self.typeToLLVM(elem_ty), gep, "arrval");
                }
            }
            // Array index via field access: GEP through struct field
            if (ie.object.data == .field_access) {
                const field_ptr = try self.genAddressOf(ie.object);
                const idx = try self.genExpr(ie.index);
                const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
                var indices = [_]c.LLVMValueRef{ zero, idx };
                const gep = c.LLVMBuildGEP2(self.builder, self.typeToLLVM(obj_ty), field_ptr, &indices, 2, "field_arridx");
                return c.LLVMBuildLoad2(self.builder, self.typeToLLVM(elem_ty), gep, "field_arrval");
            }
        }
        if (obj_ty == .string_type) {
            // String indexing: extract ptr from slice, GEP + load u8
            const str_val = try self.genExpr(ie.object);
            const ptr = c.LLVMBuildExtractValue(self.builder, str_val, 0, "str_ptr");
            const idx = try self.genExpr(ie.index);
            const i8_type = c.LLVMInt8TypeInContext(self.context);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const gep = c.LLVMBuildGEP2(self.builder, i8_type, ptr, &gep_indices, 1, "stridx");
            return c.LLVMBuildLoad2(self.builder, i8_type, gep, "byte");
        }
        if (obj_ty.isSlice()) {
            // Slice indexing: extract ptr, GEP with element type, load
            const slice_info = obj_ty.slice_type;
            const elem_ty = Type.fromName(slice_info.element_name) orelse return self.emitErrorFmt("unknown slice element type '{s}'", .{slice_info.element_name});
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            // For identifier objects, load the slice from alloca
            if (ie.object.data == .identifier) {
                if (self.named_values.get(ie.object.data.identifier.name)) |entry| {
                    const slice_val = c.LLVMBuildLoad2(self.builder, self.getStringStructType(), entry.ptr, "slice_load");
                    const ptr = c.LLVMBuildExtractValue(self.builder, slice_val, 0, "slice_ptr");
                    const idx = try self.genExpr(ie.index);
                    var gep_indices = [_]c.LLVMValueRef{idx};
                    const gep = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, ptr, &gep_indices, 1, "sliceidx");
                    return c.LLVMBuildLoad2(self.builder, elem_llvm_ty, gep, "sliceval");
                }
            }
            // Fallback for non-identifier slice expressions
            const slice_val = try self.genExpr(ie.object);
            const ptr = c.LLVMBuildExtractValue(self.builder, slice_val, 0, "slice_ptr");
            const idx = try self.genExpr(ie.index);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const gep = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, ptr, &gep_indices, 1, "sliceidx");
            return c.LLVMBuildLoad2(self.builder, elem_llvm_ty, gep, "sliceval");
        }
        // Many-pointer indexing: [*]T — GEP + load
        if (obj_ty.isManyPointer()) {
            const elem_ty = self.resolveTypeFromName(obj_ty.many_pointer_type.element_name) orelse return self.emitError("unknown many-pointer element type");
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            const ptr_val = try self.genExpr(ie.object);
            const idx = try self.genExpr(ie.index);
            var gep_indices = [_]c.LLVMValueRef{idx};
            const gep = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, ptr_val, &gep_indices, 1, "mptridx");
            return c.LLVMBuildLoad2(self.builder, elem_llvm_ty, gep, "mptrval");
        }
        return self.emitError("index expression requires an array, vector, string, slice, or [*] pointer");
    }

    fn genSliceExpr(self: *CodeGen, se: ast.SliceExpr) !c.LLVMValueRef {
        const obj_ty = self.inferType(se.object);
        const i64_ty = c.LLVMInt64TypeInContext(self.context);
        const zero = c.LLVMConstInt(i64_ty, 0, 0);
        const slice_struct_ty = self.getStringStructType();

        // Resolve start (default: 0)
        const start_val = if (se.start) |s| try self.genExpr(s) else zero;

        if (obj_ty.isArray()) {
            const arr_info = obj_ty.array_type;
            // Resolve end (default: array length)
            const end_val = if (se.end) |e| try self.genExpr(e) else c.LLVMConstInt(i64_ty, arr_info.length, 0);
            // Get array alloca
            const arr_ptr = blk: {
                if (se.object.data == .identifier) {
                    if (self.named_values.get(se.object.data.identifier.name)) |entry| {
                        break :blk entry.ptr;
                    }
                }
                break :blk try self.genExpr(se.object);
            };
            // GEP to arr[start]
            var indices = [_]c.LLVMValueRef{ zero, start_val };
            const elem_ptr = c.LLVMBuildGEP2(self.builder, self.typeToLLVM(obj_ty), arr_ptr, &indices, 2, "slice_start");
            // len = end - start
            const len_val = c.LLVMBuildSub(self.builder, end_val, start_val, "slice_len");
            // Build {ptr, len}
            var result = c.LLVMGetUndef(slice_struct_ty);
            result = c.LLVMBuildInsertValue(self.builder, result, elem_ptr, 0, "slice_ptr");
            result = c.LLVMBuildInsertValue(self.builder, result, len_val, 1, "slice_len");
            return result;
        }

        if (obj_ty == .string_type or obj_ty.isSlice()) {
            // Load {ptr, len} from variable or expression
            const obj_val = blk: {
                if (se.object.data == .identifier) {
                    if (self.named_values.get(se.object.data.identifier.name)) |entry| {
                        break :blk c.LLVMBuildLoad2(self.builder, slice_struct_ty, entry.ptr, "sslice_load");
                    }
                }
                break :blk try self.genExpr(se.object);
            };
            const base_ptr = c.LLVMBuildExtractValue(self.builder, obj_val, 0, "sslice_ptr");
            const base_len = c.LLVMBuildExtractValue(self.builder, obj_val, 1, "sslice_len");
            // Resolve end (default: original length)
            const end_val = if (se.end) |e| try self.genExpr(e) else base_len;
            // GEP base_ptr + start
            const elem_llvm_ty = if (obj_ty == .string_type)
                c.LLVMInt8TypeInContext(self.context)
            else
                self.typeToLLVM(Type.fromName(obj_ty.slice_type.element_name) orelse return self.emitError("unknown slice element type"));
            var gep_indices = [_]c.LLVMValueRef{start_val};
            const new_ptr = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, base_ptr, &gep_indices, 1, "sslice_off");
            // len = end - start
            const len_val = c.LLVMBuildSub(self.builder, end_val, start_val, "sslice_len");
            // Build {ptr, len}
            var result = c.LLVMGetUndef(slice_struct_ty);
            result = c.LLVMBuildInsertValue(self.builder, result, new_ptr, 0, "sslice_ptr");
            result = c.LLVMBuildInsertValue(self.builder, result, len_val, 1, "sslice_len");
            return result;
        }

        return self.emitError("slice expression requires an array, string, or slice");
    }

    fn genBinaryOp(self: *CodeGen, op: ast.BinaryOp.Op, lhs: c.LLVMValueRef, rhs: c.LLVMValueRef, result_type: Type) c.LLVMValueRef {
        // For vectors, dispatch based on element type; LLVM handles element-wise ops automatically
        const effective_ty = if (result_type.isVector())
            result_type.vectorElementType() orelse return lhs
        else
            result_type;

        // Vector comparison needs special handling (returns vector of i1)
        if (result_type.isVector() and (op == .eq or op == .neq)) {
            return self.genVectorComparison(op, lhs, rhs, result_type, effective_ty);
        }

        const b = self.builder;
        const is_float = effective_ty.isFloat();
        const is_unsigned = effective_ty.isUnsigned();

        return switch (op) {
            .add => if (is_float) c.LLVMBuildFAdd(b, lhs, rhs, "addtmp") else c.LLVMBuildAdd(b, lhs, rhs, "addtmp"),
            .sub => if (is_float) c.LLVMBuildFSub(b, lhs, rhs, "subtmp") else c.LLVMBuildSub(b, lhs, rhs, "subtmp"),
            .mul => if (is_float) c.LLVMBuildFMul(b, lhs, rhs, "multmp") else c.LLVMBuildMul(b, lhs, rhs, "multmp"),
            .div => if (is_float) c.LLVMBuildFDiv(b, lhs, rhs, "divtmp") else if (is_unsigned) c.LLVMBuildUDiv(b, lhs, rhs, "divtmp") else c.LLVMBuildSDiv(b, lhs, rhs, "divtmp"),
            .mod => if (is_float) c.LLVMBuildFRem(b, lhs, rhs, "modtmp") else if (is_unsigned) c.LLVMBuildURem(b, lhs, rhs, "modtmp") else c.LLVMBuildSRem(b, lhs, rhs, "modtmp"),
            .eq => if (is_float) c.LLVMBuildFCmp(b, c.LLVMRealOEQ, lhs, rhs, "eqtmp") else c.LLVMBuildICmp(b, c.LLVMIntEQ, lhs, rhs, "eqtmp"),
            .neq => if (is_float) c.LLVMBuildFCmp(b, c.LLVMRealONE, lhs, rhs, "neqtmp") else c.LLVMBuildICmp(b, c.LLVMIntNE, lhs, rhs, "neqtmp"),
            .lt => if (is_float) c.LLVMBuildFCmp(b, c.LLVMRealOLT, lhs, rhs, "lttmp") else if (is_unsigned) c.LLVMBuildICmp(b, c.LLVMIntULT, lhs, rhs, "lttmp") else c.LLVMBuildICmp(b, c.LLVMIntSLT, lhs, rhs, "lttmp"),
            .lte => if (is_float) c.LLVMBuildFCmp(b, c.LLVMRealOLE, lhs, rhs, "letmp") else if (is_unsigned) c.LLVMBuildICmp(b, c.LLVMIntULE, lhs, rhs, "letmp") else c.LLVMBuildICmp(b, c.LLVMIntSLE, lhs, rhs, "letmp"),
            .gt => if (is_float) c.LLVMBuildFCmp(b, c.LLVMRealOGT, lhs, rhs, "gttmp") else if (is_unsigned) c.LLVMBuildICmp(b, c.LLVMIntUGT, lhs, rhs, "gttmp") else c.LLVMBuildICmp(b, c.LLVMIntSGT, lhs, rhs, "gttmp"),
            .gte => if (is_float) c.LLVMBuildFCmp(b, c.LLVMRealOGE, lhs, rhs, "getmp") else if (is_unsigned) c.LLVMBuildICmp(b, c.LLVMIntUGE, lhs, rhs, "getmp") else c.LLVMBuildICmp(b, c.LLVMIntSGE, lhs, rhs, "getmp"),
            .and_op, .or_op => unreachable,
        };
    }

    fn genAndOp(self: *CodeGen, binop: ast.BinaryOp) !c.LLVMValueRef {
        const function = self.current_function;
        const i1_type = c.LLVMInt1TypeInContext(self.context);

        var lhs_val = try self.genExpr(binop.lhs);
        if (c.LLVMTypeOf(lhs_val) != i1_type) {
            lhs_val = c.LLVMBuildICmp(self.builder, c.LLVMIntNE, lhs_val, c.LLVMConstInt(c.LLVMTypeOf(lhs_val), 0, 0), "tobool");
        }
        const lhs_bb = c.LLVMGetInsertBlock(self.builder);

        const rhs_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "and.rhs");
        const merge_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "and.merge");

        _ = c.LLVMBuildCondBr(self.builder, lhs_val, rhs_bb, merge_bb);

        c.LLVMPositionBuilderAtEnd(self.builder, rhs_bb);
        var rhs_val = try self.genExpr(binop.rhs);
        if (c.LLVMTypeOf(rhs_val) != i1_type) {
            rhs_val = c.LLVMBuildICmp(self.builder, c.LLVMIntNE, rhs_val, c.LLVMConstInt(c.LLVMTypeOf(rhs_val), 0, 0), "tobool");
        }
        const rhs_end_bb = c.LLVMGetInsertBlock(self.builder);
        _ = c.LLVMBuildBr(self.builder, merge_bb);

        c.LLVMPositionBuilderAtEnd(self.builder, merge_bb);
        const phi = c.LLVMBuildPhi(self.builder, i1_type, "and.result");
        var vals = [2]c.LLVMValueRef{ c.LLVMConstInt(i1_type, 0, 0), rhs_val };
        var blocks = [2]c.LLVMBasicBlockRef{ lhs_bb, rhs_end_bb };
        c.LLVMAddIncoming(phi, &vals, &blocks, 2);

        return phi;
    }

    fn genOrOp(self: *CodeGen, binop: ast.BinaryOp) !c.LLVMValueRef {
        const function = self.current_function;
        const i1_type = c.LLVMInt1TypeInContext(self.context);

        var lhs_val = try self.genExpr(binop.lhs);
        if (c.LLVMTypeOf(lhs_val) != i1_type) {
            lhs_val = c.LLVMBuildICmp(self.builder, c.LLVMIntNE, lhs_val, c.LLVMConstInt(c.LLVMTypeOf(lhs_val), 0, 0), "tobool");
        }
        const lhs_bb = c.LLVMGetInsertBlock(self.builder);

        const rhs_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "or.rhs");
        const merge_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "or.merge");

        _ = c.LLVMBuildCondBr(self.builder, lhs_val, merge_bb, rhs_bb);

        c.LLVMPositionBuilderAtEnd(self.builder, rhs_bb);
        var rhs_val = try self.genExpr(binop.rhs);
        if (c.LLVMTypeOf(rhs_val) != i1_type) {
            rhs_val = c.LLVMBuildICmp(self.builder, c.LLVMIntNE, rhs_val, c.LLVMConstInt(c.LLVMTypeOf(rhs_val), 0, 0), "tobool");
        }
        const rhs_end_bb = c.LLVMGetInsertBlock(self.builder);
        _ = c.LLVMBuildBr(self.builder, merge_bb);

        c.LLVMPositionBuilderAtEnd(self.builder, merge_bb);
        const phi = c.LLVMBuildPhi(self.builder, i1_type, "or.result");
        var vals = [2]c.LLVMValueRef{ c.LLVMConstInt(i1_type, 1, 0), rhs_val };
        var blocks = [2]c.LLVMBasicBlockRef{ lhs_bb, rhs_end_bb };
        c.LLVMAddIncoming(phi, &vals, &blocks, 2);

        return phi;
    }

    fn genChainedComparison(self: *CodeGen, chain: ast.ChainedComparison) !c.LLVMValueRef {
        // Evaluate all operands exactly once
        var operand_vals = std.ArrayList(c.LLVMValueRef).empty;
        for (chain.operands) |operand| {
            const val = try self.genExpr(operand);
            try operand_vals.append(self.allocator, val);
        }

        // Compare pairwise and AND results together
        var result: c.LLVMValueRef = undefined;
        for (chain.ops, 0..) |op, i| {
            const lhs_ty = self.inferType(chain.operands[i]);
            const rhs_ty = self.inferType(chain.operands[i + 1]);
            const cmp_type = Type.widen(lhs_ty, rhs_ty);

            const lhs_conv = self.convertValue(operand_vals.items[i], lhs_ty, cmp_type);
            const rhs_conv = self.convertValue(operand_vals.items[i + 1], rhs_ty, cmp_type);

            const cmp = self.genBinaryOp(op, lhs_conv, rhs_conv, cmp_type);

            if (i == 0) {
                result = cmp;
            } else {
                result = c.LLVMBuildAnd(self.builder, result, cmp, "chain.and");
            }
        }

        return result;
    }

    fn genCall(self: *CodeGen, call_node: ast.Call) !c.LLVMValueRef {
        if (call_node.callee.data == .field_access) {
            const fa = call_node.callee.data.field_access;

            // Union construction: Shape.variant(payload)
            const resolved_type: ?Type = blk: {
                if (fa.object.data == .identifier) {
                    const name = self.type_aliases.get(fa.object.data.identifier.name) orelse fa.object.data.identifier.name;
                    if (self.union_types.contains(name)) break :blk .{ .union_type = name };
                    if (self.struct_types.contains(name)) break :blk .{ .struct_type = name };
                } else {
                    const ty = self.resolveType(fa.object);
                    if (ty.isUnion() or ty.isStruct()) break :blk ty;
                }
                break :blk null;
            };
            if (resolved_type) |rty| {
                if (rty.isUnion()) {
                    const type_name = rty.union_type;
                    const payload_node: ?*Node = if (call_node.args.len > 0) call_node.args[0] else null;
                    return self.genUnionLiteral(.{
                        .union_name = type_name,
                        .variant_name = fa.field,
                        .payload = payload_node,
                    }, type_name);
                }
            }

            // Namespaced call: namespace.func(args)
            if (fa.object.data == .identifier) {
                const ns_name = fa.object.data.identifier.name;
                if (self.namespaces.contains(ns_name)) {
                    const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns_name, fa.field });
                    return self.genCallByName(qualified, call_node);
                }
            }

            // UFCS: obj.method(args...) → method(obj, args...)
            const method_name = fa.field;
            const method_z = self.allocator.dupeZ(u8, method_name) catch method_name;
            if (self.generic_templates.contains(method_name) or
                c.LLVMGetNamedFunction(self.module, method_z.ptr) != null)
            {
                var ufcs_args = try self.allocator.alloc(*Node, call_node.args.len + 1);
                ufcs_args[0] = fa.object;
                for (call_node.args, 0..) |arg, i| {
                    ufcs_args[i + 1] = arg;
                }
                return self.genCallByName(method_name, .{
                    .callee = call_node.callee,
                    .args = ufcs_args,
                });
            }
        }

        // Resolve callee — must be an identifier
        if (call_node.callee.data != .identifier) return self.emitError("callee must be an identifier");
        const callee_name = call_node.callee.data.identifier.name;
        return self.genCallByName(callee_name, call_node);
    }

    fn genCallByName(self: *CodeGen, callee_name: []const u8, call_node: ast.Call) !c.LLVMValueRef {
        // Check if this is a generic function call
        if (self.generic_templates.get(callee_name)) |template| {
            return self.genGenericCall(callee_name, template, call_node);
        }
        // Intra-namespace fallback for generic templates
        if (self.current_namespace) |ns| {
            const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name });
            if (self.generic_templates.get(qualified)) |template| {
                return self.genGenericCall(qualified, template, call_node);
            }
        }

        // Check for #builtin function (only available when imported)
        if (self.builtin_functions.contains(callee_name)) {
            return self.dispatchBuiltin(callee_name, call_node);
        }
        // Intra-namespace fallback for builtins
        if (self.current_namespace) |ns| {
            const qualified_builtin = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name });
            if (self.builtin_functions.contains(qualified_builtin)) {
                return self.dispatchBuiltin(qualified_builtin, call_node);
            }
        }

        // Compiler intrinsics (always available, no #builtin declaration needed)
        if (std.mem.eql(u8, callee_name, "sqrt")) {
            return self.genSqrt(call_node);
        }
        if (std.mem.eql(u8, callee_name, "sin")) {
            return self.genSin(call_node);
        }
        if (std.mem.eql(u8, callee_name, "cos")) {
            return self.genCos(call_node);
        }
        if (std.mem.eql(u8, callee_name, "cast")) {
            return self.genCast(call_node);
        }
        if (std.mem.eql(u8, callee_name, "malloc")) {
            return self.genMalloc(call_node.args);
        }
        if (std.mem.eql(u8, callee_name, "free")) {
            return self.genFree(call_node.args);
        }
        if (std.mem.eql(u8, callee_name, "memcpy")) {
            return self.genMemcpy(call_node.args);
        }

        const name_z = try self.allocator.dupeZ(u8, callee_name);
        var callee_fn = c.LLVMGetNamedFunction(self.module, name_z.ptr);
        // Foreign function fallback: qualified name "ns.Func" → try unqualified "Func" (the C symbol)
        if (callee_fn == null) {
            if (std.mem.lastIndexOfScalar(u8, callee_name, '.')) |dot_idx| {
                const base_name = callee_name[dot_idx + 1 ..];
                const base_z = try self.allocator.dupeZ(u8, base_name);
                callee_fn = c.LLVMGetNamedFunction(self.module, base_z.ptr);
            }
        }
        // Intra-namespace fallback: try qualified name
        if (callee_fn == null) {
            if (self.current_namespace) |ns| {
                const qualified2 = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name });
                const qualified_z = try self.allocator.dupeZ(u8, qualified2);
                callee_fn = c.LLVMGetNamedFunction(self.module, qualified_z.ptr);
            }
        }
        // Function pointer indirect call: callee is a variable with function_type
        if (callee_fn == null) {
            const fp_entry = if (self.named_values.get(callee_name)) |e| e
                else self.global_mutable_vars.get(callee_name);
            if (fp_entry) |entry| {
                if (entry.ty.isFunctionType()) {
                    return self.genIndirectCall(entry, call_node);
                }
            }
            return self.emitErrorFmt("undefined function '{s}'", .{callee_name});
        }

        // Get function type (opaque pointers: use LLVMGlobalGetValueType)
        const fn_type = c.LLVMGlobalGetValueType(callee_fn.?);

        // Check if this is a variadic function call
        const var_info = self.variadic_functions.get(callee_name) orelse blk: {
            // Try qualified name lookup
            if (self.current_namespace) |ns| {
                const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name });
                break :blk self.variadic_functions.get(qualified);
            }
            break :blk null;
        };

        // Generate arguments with type conversion to match parameter types
        const num_params = c.LLVMCountParamTypes(fn_type);
        var param_llvm_types: [64]c.LLVMTypeRef = undefined;
        if (num_params > 0) {
            c.LLVMGetParamTypes(fn_type, &param_llvm_types);
        }

        var arg_vals = std.ArrayList(c.LLVMValueRef).empty;

        if (var_info) |vi| {
            // Variadic call: generate fixed args, then pack remaining into slice
            const fixed_count = vi.fixed_param_count;
            // Generate fixed args
            for (0..fixed_count) |i| {
                if (i < call_node.args.len) {
                    const param_ty = self.llvmTypeToSxType(param_llvm_types[i]);
                    try arg_vals.append(self.allocator, try self.genExprAsType(call_node.args[i], param_ty));
                }
            }
            // Pack variadic args into a slice {ptr, len}
            const elem_ty = Type.fromName(vi.element_type_name) orelse Type.s(64);
            const elem_llvm_ty = self.typeToLLVM(elem_ty);
            const var_arg_count = if (call_node.args.len > fixed_count) call_node.args.len - fixed_count else 0;

            // Check for spread operator: fn(..array) — single spread arg
            if (var_arg_count == 1 and call_node.args[fixed_count].data == .spread_expr) {
                const spread_operand = call_node.args[fixed_count].data.spread_expr.operand;
                const spread_ty = self.inferType(spread_operand);
                if (spread_ty.isArray()) {
                    // Spread an array: construct slice from array pointer + known length
                    const arr_info = spread_ty.array_type;
                    if (spread_operand.data == .identifier) {
                        if (self.named_values.get(spread_operand.data.identifier.name)) |entry| {
                            const arr_llvm_ty = self.typeToLLVM(spread_ty);
                            const zero = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), 0, 0);
                            var ptr_indices = [_]c.LLVMValueRef{ zero, zero };
                            const arr_ptr = c.LLVMBuildGEP2(self.builder, arr_llvm_ty, entry.ptr, &ptr_indices, 2, "spread_ptr");
                            const len_val = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), arr_info.length, 0);
                            const slice_val = self.buildStringSliceRT(arr_ptr, len_val);
                            try arg_vals.append(self.allocator, slice_val);
                        } else {
                            return self.emitError("spread operand not found");
                        }
                    } else {
                        return self.emitError("spread operator requires a named variable");
                    }
                } else if (spread_ty.isSlice()) {
                    // Spread a slice: pass through as-is
                    const slice_val = try self.genExpr(spread_operand);
                    try arg_vals.append(self.allocator, slice_val);
                } else {
                    return self.emitError("spread operator requires an array or slice");
                }
            } else if (var_arg_count > 0) {
                // Allocate array on stack: [N x elem_type]
                const arr_ty = c.LLVMArrayType2(elem_llvm_ty, @intCast(var_arg_count));
                const arr_alloca = self.buildEntryBlockAlloca(arr_ty, "varargs_arr");
                // Store each variadic arg
                for (0..var_arg_count) |vi_idx| {
                    const arg_val = if (elem_ty.isAny()) blk: {
                        // ..Any: wrap each arg in Any{tag, value}
                        const raw_val = try self.genExpr(call_node.args[fixed_count + vi_idx]);
                        const arg_ty = self.inferType(call_node.args[fixed_count + vi_idx]);
                        break :blk try self.buildAnyValue(raw_val, arg_ty);
                    } else try self.genExprAsType(call_node.args[fixed_count + vi_idx], elem_ty);
                    const zero = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), 0, 0);
                    const idx_val = c.LLVMConstInt(c.LLVMInt32TypeInContext(self.context), @intCast(vi_idx), 0);
                    var indices = [_]c.LLVMValueRef{ zero, idx_val };
                    const gep = c.LLVMBuildGEP2(self.builder, arr_ty, arr_alloca, &indices, 2, "vararg_elem");
                    _ = c.LLVMBuildStore(self.builder, arg_val, gep);
                }
                // Build slice: {ptr, len}
                const zero = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), 0, 0);
                var ptr_indices = [_]c.LLVMValueRef{ zero, zero };
                const arr_ptr = c.LLVMBuildGEP2(self.builder, arr_ty, arr_alloca, &ptr_indices, 2, "varargs_ptr");
                const len_val = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), @intCast(var_arg_count), 0);
                const slice_val = self.buildStringSliceRT(arr_ptr, len_val);
                try arg_vals.append(self.allocator, slice_val);
            } else {
                // Zero variadic args: pass empty slice {null, 0}
                const null_ptr = c.LLVMConstNull(c.LLVMPointerTypeInContext(self.context, 0));
                const zero_len = c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), 0, 0);
                const slice_val = self.buildStringSliceRT(null_ptr, zero_len);
                try arg_vals.append(self.allocator, slice_val);
            }
        } else {
            // Normal (non-variadic) call — use stored sx param types when available
            const stored_param_types = self.fn_param_types.get(callee_name) orelse blk: {
                if (self.current_namespace) |ns| {
                    const qualified = try std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name });
                    break :blk self.fn_param_types.get(qualified);
                }
                break :blk null;
            };
            for (call_node.args, 0..) |arg, i| {
                if (i < num_params) {
                    const param_ty = if (stored_param_types != null and i < stored_param_types.?.len)
                        stored_param_types.?[i]
                    else
                        self.llvmTypeToSxType(param_llvm_types[i]);
                    // For #foreign functions, [:0]u8 params have LLVM type ptr
                    const arg_ty = self.inferType(arg);
                    const llvm_param_is_ptr = (i < num_params and
                        c.LLVMGetTypeKind(param_llvm_types[i]) == c.LLVMPointerTypeKind);
                    const ptr_ty = Type{ .pointer_type = .{ .pointee_name = "u8" } };

                    var val = if (llvm_param_is_ptr and arg.data == .string_literal) blk: {
                        // String literal → pointer: produce raw ptr directly (context-dependent)
                        break :blk try self.genExprAsType(arg, ptr_ty);
                    } else if (llvm_param_is_ptr and arg_ty == .string_type) blk: {
                        // String variable → pointer: extract .ptr from {ptr, len}
                        const str_val = try self.genExpr(arg);
                        break :blk c.LLVMBuildExtractValue(self.builder, str_val, 0, "str_ptr");
                    } else if ((param_ty.isPointer() or llvm_param_is_ptr) and arg_ty.isSlice() and
                        std.mem.eql(u8, arg_ty.slice_type.element_name, "u8"))
                    {
                        return self.emitError(
                            "cannot pass []u8 to *u8: slice may not be null-terminated; use a string literal or xx cast",
                        );
                    } else try self.genExprAsType(arg, param_ty);
                    // Foreign calls: convert struct values to C ABI representation
                    if (param_ty.isStruct() and self.foreign_fns.contains(callee_name)) {
                        const struct_llvm_ty = self.typeToLLVM(param_ty);
                        if (struct_llvm_ty != param_llvm_types[i]) {
                            val = self.convertStructToABI(val, struct_llvm_ty, param_llvm_types[i]);
                        }
                    }
                    try arg_vals.append(self.allocator, val);
                } else {
                    try arg_vals.append(self.allocator, try self.genExpr(arg));
                }
            }
        }
        const args_slice = try arg_vals.toOwnedSlice(self.allocator);

        const ret_ty = c.LLVMGetReturnType(fn_type);
        const call_name: [*c]const u8 = if (ret_ty == c.LLVMVoidTypeInContext(self.context)) "" else "calltmp";
        return c.LLVMBuildCall2(
            self.builder,
            fn_type,
            callee_fn.?,
            if (args_slice.len > 0) args_slice.ptr else null,
            @intCast(args_slice.len),
            call_name,
        );
    }

    fn genIndirectCall(self: *CodeGen, entry: NamedValue, call_node: ast.Call) !c.LLVMValueRef {
        const fti = entry.ty.function_type;

        // Load the function pointer from the alloca
        const ptr_ty = c.LLVMPointerTypeInContext(self.context, 0);
        const fn_ptr = c.LLVMBuildLoad2(self.builder, ptr_ty, entry.ptr, "fn_ptr");

        // Build LLVM function type from FunctionTypeInfo
        const ptr_ty_llvm = c.LLVMPointerTypeInContext(self.context, 0);
        var param_llvm_types: [64]c.LLVMTypeRef = undefined;
        for (fti.param_types, 0..) |pt, i| {
            // [N]T and [:0]T params are pointers at the ABI level
            param_llvm_types[i] = if (pt.isArray() or pt == .string_type) ptr_ty_llvm else self.typeToLLVM(pt);
        }
        const ret_llvm = self.typeToLLVM(fti.return_type.*);
        const fn_type = c.LLVMFunctionType(
            ret_llvm,
            if (fti.param_types.len > 0) &param_llvm_types else null,
            @intCast(fti.param_types.len),
            0,
        );

        // Generate arguments with type conversion
        var arg_vals = std.ArrayList(c.LLVMValueRef).empty;
        for (call_node.args, 0..) |arg, i| {
            if (i < fti.param_types.len) {
                const pt = fti.param_types[i];
                // [N]T params: pass pointer via array decay
                if (pt.isArray()) {
                    const decay_target: Type = .{ .many_pointer_type = .{ .element_name = pt.array_type.element_name } };
                    try arg_vals.append(self.allocator, try self.genExprAsType(arg, decay_target));
                } else if (pt == .string_type) {
                    // [:0]u8 params: extract .ptr from fat pointer
                    const val = try self.genExprAsType(arg, pt);
                    try arg_vals.append(self.allocator, c.LLVMBuildExtractValue(self.builder, val, 0, "str_ptr"));
                } else {
                    try arg_vals.append(self.allocator, try self.genExprAsType(arg, pt));
                }
            } else {
                try arg_vals.append(self.allocator, try self.genExpr(arg));
            }
        }
        const args_slice = try arg_vals.toOwnedSlice(self.allocator);

        const call_name: [*c]const u8 = if (ret_llvm == c.LLVMVoidTypeInContext(self.context)) "" else "calltmp";
        return c.LLVMBuildCall2(
            self.builder,
            fn_type,
            fn_ptr,
            if (args_slice.len > 0) args_slice.ptr else null,
            @intCast(args_slice.len),
            call_name,
        );
    }

    fn genGenericCall(self: *CodeGen, qualified_name: []const u8, template: GenericTemplate, call_node: ast.Call) !c.LLVMValueRef {
        const fd = template.fd;

        // Check for runtime type dispatch: cast(runtime_type_var, any_val) as argument
        if (self.current_match_tags) |match_tags| {
            if (match_tags.len > 0) {
                for (call_node.args) |arg| {
                    if (arg.data == .call) {
                        if (arg.data.call.callee.data == .identifier) {
                            const cast_name = arg.data.call.callee.data.identifier.name;
                            if (std.mem.eql(u8, cast_name, "cast") or std.mem.eql(u8, cast_name, "std.cast")) {
                                if (arg.data.call.args.len == 2) {
                                    const type_arg = arg.data.call.args[0];
                                    // Check if first arg of cast is a runtime variable (not a type expression)
                                    if (type_arg.data == .identifier) {
                                        const name = type_arg.data.identifier.name;
                                        // It's a runtime type if it's a named_value, not a type name
                                        if (self.named_values.contains(name) and
                                            Type.fromName(name) == null and
                                            !self.struct_types.contains(name) and
                                            !self.enum_types.contains(name) and
                                            !self.union_types.contains(name) and
                                            !self.type_aliases.contains(name))
                                        {
                                            return self.genGenericCallWithRuntimeDispatch(template, call_node, match_tags);
                                        }
                                    }
                                }
                            }
                        }
                    }
                }
            }
        }

        // Check for comptime value params
        var has_comptime_values = false;
        var comptime_nodes = std.StringHashMap(*Node).init(self.allocator);
        for (fd.type_params) |tp| {
            const constraint_name = if (tp.constraint.data == .type_expr) tp.constraint.data.type_expr.name else "";
            if (!std.mem.eql(u8, constraint_name, "Type")) {
                // Value param — extract comptime value from call arg
                has_comptime_values = true;
                for (fd.params, 0..) |param, pi| {
                    if (std.mem.eql(u8, param.name, tp.name)) {
                        if (pi < call_node.args.len) {
                            try comptime_nodes.put(tp.name, @constCast(call_node.args[pi]));
                        }
                        break;
                    }
                }
            }
        }

        // Normal generic call: Infer type bindings from arguments, widening across all args for the same type param
        var bindings = std.StringHashMap(Type).init(self.allocator);
        // Track bindings derived from parameterized struct types — these are authoritative and should not be widened
        var firm_bindings = std.StringHashMap(void).init(self.allocator);
        for (fd.params, 0..) |param, i| {
            if (param.is_comptime) continue;
            // Direct type param: (a: $T) introduces/widens, (a: T) only binds if not yet bound
            if (param.type_expr.data == .type_expr) {
                const type_name = param.type_expr.data.type_expr.name;
                // Check if this type name is a type parameter
                for (fd.type_params) |tp| {
                    if (std.mem.eql(u8, tp.name, type_name)) {
                        if (i < call_node.args.len) {
                            // Skip widening if binding was derived from a parameterized struct
                            if (!firm_bindings.contains(type_name)) {
                                const arg_ty = self.inferType(call_node.args[i]);
                                if (bindings.get(type_name)) |existing| {
                                    try bindings.put(type_name, Type.widen(existing, arg_ty));
                                } else {
                                    try bindings.put(type_name, arg_ty);
                                }
                            }
                        }
                        break;
                    }
                }
            }
            // Pointer to parameterized type: (p: *Foo($T)) — extract T from concrete struct
            if (param.type_expr.data == .pointer_type_expr) {
                const pointee = param.type_expr.data.pointer_type_expr.pointee_type;
                if (pointee.data == .parameterized_type_expr) {
                    const pte = pointee.data.parameterized_type_expr;
                    if (i < call_node.args.len) {
                        const arg_ty = self.inferType(call_node.args[i]);
                        // arg should be *StructName — get the struct's stored type param bindings
                        const struct_name = if (arg_ty.isPointer())
                            arg_ty.pointer_type.pointee_name
                        else if (arg_ty.isStruct())
                            arg_ty.struct_type
                        else
                            "";
                        if (self.struct_types.get(struct_name)) |info| {
                            if (info.template_name) |tmpl_name| {
                                if (std.mem.eql(u8, tmpl_name, pte.name)) {
                                    // Match generic args against stored type param bindings
                                    for (pte.args, 0..) |arg, ai| {
                                        if (arg.data == .type_expr and arg.data.type_expr.is_generic) {
                                            const gen_name = arg.data.type_expr.name;
                                            if (ai < info.type_param_types.len) {
                                                try bindings.put(gen_name, info.type_param_types[ai]);
                                                try firm_bindings.put(gen_name, {});
                                            }
                                        }
                                    }
                                }
                            }
                        }
                    }
                }
            }
            // Slice type param: (items: []$T) — infer T from array or slice element type
            if (param.type_expr.data == .slice_type_expr) {
                const elem_node = param.type_expr.data.slice_type_expr.element_type;
                if (elem_node.data == .type_expr) {
                    const type_name = elem_node.data.type_expr.name;
                    for (fd.type_params) |tp| {
                        if (std.mem.eql(u8, tp.name, type_name)) {
                            if (i < call_node.args.len) {
                                const arg_ty = self.inferType(call_node.args[i]);
                                const elem_ty = if (arg_ty.isArray())
                                    Type.fromName(arg_ty.array_type.element_name) orelse arg_ty
                                else if (arg_ty.isSlice())
                                    Type.fromName(arg_ty.slice_type.element_name) orelse arg_ty
                                else
                                    arg_ty;
                                if (bindings.get(type_name)) |existing| {
                                    try bindings.put(type_name, Type.widen(existing, elem_ty));
                                } else {
                                    try bindings.put(type_name, elem_ty);
                                }
                            }
                            break;
                        }
                    }
                }
            }
        }

        if (has_comptime_values) {
            return self.genComptimeCall(qualified_name, fd, call_node, bindings, comptime_nodes);
        }

        // Generate mangled name
        const mangled = try self.mangleGenericName(fd.name, fd.type_params, bindings, null, null);

        // Check cache
        const callee_fn = if (self.generic_instances.get(mangled)) |cached|
            cached
        else
            try self.instantiateGeneric(fd, bindings, mangled);

        // Generate arguments with type conversion to match parameter types
        self.type_param_bindings = bindings;
        var arg_vals = std.ArrayList(c.LLVMValueRef).empty;
        for (call_node.args, 0..) |arg, i| {
            if (i < fd.params.len) {
                const param_ty = self.resolveType(fd.params[i].type_expr);
                try arg_vals.append(self.allocator, try self.genExprAsType(arg, param_ty));
            } else {
                try arg_vals.append(self.allocator, try self.genExpr(arg));
            }
        }
        self.type_param_bindings = null;
        const args_slice = try arg_vals.toOwnedSlice(self.allocator);

        const fn_type = c.LLVMGlobalGetValueType(callee_fn);
        const ret_ty = c.LLVMGetReturnType(fn_type);
        const call_name: [*c]const u8 = if (ret_ty == c.LLVMVoidTypeInContext(self.context)) "" else "calltmp";
        return c.LLVMBuildCall2(
            self.builder,
            fn_type,
            callee_fn,
            if (args_slice.len > 0) args_slice.ptr else null,
            @intCast(args_slice.len),
            call_name,
        );
    }

    /// Generate a call to a generic function with comptime value parameters.
    /// Instantiates the function with the specific comptime values, then delegates to genCallByName
    /// with the mangled name and adjusted args (comptime args removed).
    fn genComptimeCall(
        self: *CodeGen,
        qualified_name: []const u8,
        fd: ast.FnDecl,
        call_node: ast.Call,
        type_bindings: std.StringHashMap(Type),
        comptime_nodes: std.StringHashMap(*Node),
    ) !c.LLVMValueRef {
        const mangled = try self.mangleGenericName(qualified_name, fd.type_params, type_bindings, null, comptime_nodes);

        // Instantiate if not cached
        if (!self.generic_instances.contains(mangled)) {
            // Set comptime param nodes for #insert substitution
            const saved_comptime_nodes = self.comptime_param_nodes;
            self.comptime_param_nodes = comptime_nodes;
            defer self.comptime_param_nodes = saved_comptime_nodes;

            // Set namespace context if the qualified name is namespaced (e.g. "std.print")
            const saved_namespace = self.current_namespace;
            if (std.mem.indexOfScalar(u8, qualified_name, '.')) |dot_pos| {
                self.current_namespace = qualified_name[0..dot_pos];
            }
            defer self.current_namespace = saved_namespace;

            // Pre-register Any type IDs for variadic args before function instantiation,
            // so type category matching (case slice:, case array:, etc.) in any_to_string
            // can find registered types during compilation of the function body.
            for (fd.params, 0..) |param, pi| {
                if (param.is_variadic) {
                    const elem_name_raw = if (param.type_expr.data == .type_expr) param.type_expr.data.type_expr.name else "";
                    if (std.mem.eql(u8, elem_name_raw, "Any") or std.mem.eql(u8, elem_name_raw, "std.Any")) {
                        for (pi..call_node.args.len) |ai| {
                            const arg_ty = self.inferType(call_node.args[ai]);
                            try self.preRegisterAnyType(arg_ty);
                        }
                    }
                    break;
                }
            }

            _ = try self.instantiateGeneric(fd, type_bindings, mangled);

            // Register variadic info for the mangled function (adjusted for removed comptime params)
            var comptime_before_variadic: u32 = 0;
            for (fd.params) |param| {
                if (param.is_variadic) break;
                if (param.is_comptime) comptime_before_variadic += 1;
            }
            for (fd.params, 0..) |param, i| {
                if (param.is_variadic) {
                    const elem_name = if (param.type_expr.data == .type_expr) param.type_expr.data.type_expr.name else "s32";
                    try self.variadic_functions.put(mangled, .{
                        .fixed_param_count = @intCast(i - comptime_before_variadic),
                        .element_type_name = elem_name,
                    });
                    break;
                }
            }
        }

        // Build adjusted call args (skip comptime args)
        var adjusted_args = std.ArrayList(*Node).empty;
        for (call_node.args, 0..) |arg, i| {
            if (i < fd.params.len and fd.params[i].is_comptime) continue;
            try adjusted_args.append(self.allocator, @constCast(arg));
        }
        const adjusted_args_slice = try adjusted_args.toOwnedSlice(self.allocator);

        const adjusted_call = ast.Call{
            .callee = call_node.callee,
            .args = adjusted_args_slice,
        };

        // Call the instantiated function through normal path (handles variadic packing etc.)
        return self.genCallByName(mangled, adjusted_call);
    }

    /// Generate a generic function call with runtime type dispatch.
    /// For each type tag in match_tags, monomorphize the generic function and dispatch via switch.
    fn genGenericCallWithRuntimeDispatch(
        self: *CodeGen,
        template: GenericTemplate,
        call_node: ast.Call,
        match_tags: []const u64,
    ) !c.LLVMValueRef {
        const fd = template.fd;

        // Find the cast argument and extract the runtime type tag + any value source
        var cast_arg_idx: usize = 0;
        var type_tag_node: *Node = undefined;
        var any_val_node: *Node = undefined;
        for (call_node.args, 0..) |arg, i| {
            if (arg.data == .call and arg.data.call.callee.data == .identifier) {
                const name = arg.data.call.callee.data.identifier.name;
                if ((std.mem.eql(u8, name, "cast") or std.mem.eql(u8, name, "std.cast")) and arg.data.call.args.len == 2) {
                    cast_arg_idx = i;
                    type_tag_node = arg.data.call.args[0];
                    any_val_node = arg.data.call.args[1];
                    break;
                }
            }
        }

        // Generate the runtime type tag value and the Any value
        const type_tag_val = try self.genExpr(type_tag_node);
        const any_val = try self.genExpr(any_val_node);
        // Generate non-cast arguments (evaluated once, before the switch)
        var other_arg_vals = std.ArrayList(?c.LLVMValueRef).empty;
        for (call_node.args, 0..) |arg, i| {
            if (i == cast_arg_idx) {
                try other_arg_vals.append(self.allocator, null); // placeholder
            } else {
                try other_arg_vals.append(self.allocator, try self.genExpr(arg));
            }
        }

        // Extract Any value i64 BEFORE the switch (switch is a terminator, nothing can follow it in the same BB)
        const any_i64 = c.LLVMBuildExtractValue(self.builder, any_val, 1, "any_payload");

        // Build dispatch switch
        const function = self.current_function;
        const dispatch_merge = c.LLVMAppendBasicBlockInContext(self.context, function, "dispatch_merge");
        const dispatch_default = c.LLVMAppendBasicBlockInContext(self.context, function, "dispatch_default");
        const sw = c.LLVMBuildSwitch(self.builder, type_tag_val, dispatch_default, @intCast(match_tags.len));

        // Determine return type from function signature
        const ret_ty = if (fd.return_type) |rt| self.resolveType(rt) else Type.void_type;
        // We'll use the first monomorphized function's return to determine LLVM type
        var result_llvm_ty: c.LLVMTypeRef = null;

        var phi_vals = std.ArrayList(c.LLVMValueRef).empty;
        var phi_bbs = std.ArrayList(c.LLVMBasicBlockRef).empty;

        for (match_tags) |tag| {
            // Find the AnyTypeEntry for this tag
            var entry_type: ?Type = null;
            var it = self.any_type_entries.iterator();
            while (it.next()) |entry| {
                if (entry.value_ptr.tag_id == tag) {
                    entry_type = entry.value_ptr.sx_type;
                    break;
                }
            }
            const sx_type = entry_type orelse continue;

            // Create case BB
            const case_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "dispatch_case");
            c.LLVMAddCase(sw, c.LLVMConstInt(c.LLVMInt64TypeInContext(self.context), tag, 0), case_bb);
            c.LLVMPositionBuilderAtEnd(self.builder, case_bb);

            // Convert Any payload to the concrete type
            const concrete_val = try self.extractAnyToConcreteType(any_i64, sx_type);

            // Monomorphize the generic function with this type
            var bindings = std.StringHashMap(Type).init(self.allocator);
            // Bind the type parameter from the cast argument's position
            if (cast_arg_idx < fd.params.len) {
                if (fd.params[cast_arg_idx].type_expr.data == .type_expr) {
                    const tp_name = fd.params[cast_arg_idx].type_expr.data.type_expr.name;
                    for (fd.type_params) |tp| {
                        if (std.mem.eql(u8, tp.name, tp_name)) {
                            try bindings.put(tp.name, sx_type);
                            break;
                        }
                    }
                } else if (fd.params[cast_arg_idx].type_expr.data == .slice_type_expr) {
                    // Slice type param: (items: []$T) — extract element type from concrete slice
                    const elem_node = fd.params[cast_arg_idx].type_expr.data.slice_type_expr.element_type;
                    if (elem_node.data == .type_expr) {
                        const tp_name = elem_node.data.type_expr.name;
                        for (fd.type_params) |tp| {
                            if (std.mem.eql(u8, tp.name, tp_name)) {
                                // Extract element type from concrete slice type
                                const elem_ty = if (sx_type.isSlice())
                                    Type.fromName(sx_type.slice_type.element_name) orelse sx_type
                                else if (sx_type.isArray())
                                    Type.fromName(sx_type.array_type.element_name) orelse sx_type
                                else
                                    sx_type;
                                try bindings.put(tp.name, elem_ty);
                                break;
                            }
                        }
                    }
                }
            }

            const mangled = try self.mangleGenericName(fd.name, fd.type_params, bindings, null, null);
            const callee_fn = if (self.generic_instances.get(mangled)) |cached|
                cached
            else
                try self.instantiateGeneric(fd, bindings, mangled);

            // Build argument list
            self.type_param_bindings = bindings;
            var arg_vals_list = std.ArrayList(c.LLVMValueRef).empty;
            for (other_arg_vals.items, 0..) |maybe_val, ai| {
                if (ai == cast_arg_idx) {
                    // Use the converted concrete value
                    try arg_vals_list.append(self.allocator, concrete_val);
                } else if (maybe_val) |v| {
                    try arg_vals_list.append(self.allocator, v);
                }
            }
            self.type_param_bindings = null;

            const args_slice = try arg_vals_list.toOwnedSlice(self.allocator);
            const fn_type = c.LLVMGlobalGetValueType(callee_fn);
            const call_result = c.LLVMBuildCall2(
                self.builder,
                fn_type,
                callee_fn,
                if (args_slice.len > 0) args_slice.ptr else null,
                @intCast(args_slice.len),
                if (ret_ty != .void_type) "dispatch_result" else "",
            );

            if (result_llvm_ty == null and ret_ty != .void_type) {
                result_llvm_ty = c.LLVMTypeOf(call_result);
            }

            if (ret_ty != .void_type) {
                try phi_vals.append(self.allocator, call_result);
                try phi_bbs.append(self.allocator, c.LLVMGetInsertBlock(self.builder));
            }
            _ = c.LLVMBuildBr(self.builder, dispatch_merge);
        }

        // Default case: return undef (should not be reached)
        c.LLVMPositionBuilderAtEnd(self.builder, dispatch_default);
        if (ret_ty != .void_type and result_llvm_ty != null) {
            try phi_vals.append(self.allocator, c.LLVMGetUndef(result_llvm_ty.?));
            try phi_bbs.append(self.allocator, dispatch_default);
        }
        _ = c.LLVMBuildBr(self.builder, dispatch_merge);

        // Merge
        c.LLVMPositionBuilderAtEnd(self.builder, dispatch_merge);
        if (ret_ty != .void_type and result_llvm_ty != null) {
            const vals_slice = try phi_vals.toOwnedSlice(self.allocator);
            const bbs_slice = try phi_bbs.toOwnedSlice(self.allocator);
            const phi = c.LLVMBuildPhi(self.builder, result_llvm_ty.?, "dispatch_phi");
            c.LLVMAddIncoming(phi, vals_slice.ptr, bbs_slice.ptr, @intCast(vals_slice.len));
            return phi;
        }

        return null;
    }

    /// Extract a concrete typed value from an Any i64 payload.
    fn extractAnyToConcreteType(self: *CodeGen, any_i64: c.LLVMValueRef, sx_type: Type) !c.LLVMValueRef {
        return switch (sx_type) {
            .boolean => c.LLVMBuildTrunc(self.builder, any_i64, c.LLVMInt1TypeInContext(self.context), "any_to_bool"),
            .signed => |w| if (w <= 32)
                c.LLVMBuildTrunc(self.builder, any_i64, c.LLVMIntTypeInContext(self.context, w), "any_to_int")
            else
                any_i64,
            .unsigned => |w| if (w <= 32)
                c.LLVMBuildTrunc(self.builder, any_i64, c.LLVMIntTypeInContext(self.context, w), "any_to_uint")
            else
                any_i64,
            .f32 => blk: {
                const as_f64 = c.LLVMBuildBitCast(self.builder, any_i64, c.LLVMDoubleTypeInContext(self.context), "i64_to_f64");
                break :blk c.LLVMBuildFPTrunc(self.builder, as_f64, c.LLVMFloatTypeInContext(self.context), "any_to_f32");
            },
            .f64 => c.LLVMBuildBitCast(self.builder, any_i64, c.LLVMDoubleTypeInContext(self.context), "any_to_f64"),
            .string_type => blk: {
                const ptr = c.LLVMBuildIntToPtr(self.builder, any_i64, c.LLVMPointerTypeInContext(self.context, 0), "any_to_str_ptr");
                break :blk c.LLVMBuildLoad2(self.builder, self.getStringStructType(), ptr, "any_to_str");
            },
            .struct_type => |sname| blk: {
                const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct '{s}'", .{sname});
                const ptr = c.LLVMBuildIntToPtr(self.builder, any_i64, c.LLVMPointerTypeInContext(self.context, 0), "any_to_struct_ptr");
                break :blk c.LLVMBuildLoad2(self.builder, info.llvm_type, ptr, "any_to_struct");
            },
            .enum_type => any_i64,
            .union_type => |uname| blk: {
                const info = self.union_types.get(uname) orelse return self.emitErrorFmt("unknown union '{s}'", .{uname});
                const ptr = c.LLVMBuildIntToPtr(self.builder, any_i64, c.LLVMPointerTypeInContext(self.context, 0), "any_to_union_ptr");
                break :blk c.LLVMBuildLoad2(self.builder, info.llvm_type, ptr, "any_to_union");
            },
            .vector_type, .array_type => blk: {
                const llvm_ty = self.typeToLLVM(sx_type);
                const ptr = c.LLVMBuildIntToPtr(self.builder, any_i64, c.LLVMPointerTypeInContext(self.context, 0), "any_to_vec_ptr");
                break :blk c.LLVMBuildLoad2(self.builder, llvm_ty, ptr, "any_to_vec");
            },
            .slice_type => blk: {
                const ptr = c.LLVMBuildIntToPtr(self.builder, any_i64, c.LLVMPointerTypeInContext(self.context, 0), "any_to_slice_ptr");
                break :blk c.LLVMBuildLoad2(self.builder, self.getStringStructType(), ptr, "any_to_slice");
            },
            .pointer_type, .many_pointer_type => c.LLVMBuildIntToPtr(self.builder, any_i64, c.LLVMPointerTypeInContext(self.context, 0), "any_to_ptr"),
            else => any_i64,
        };
    }

    fn mangleGenericName(
        self: *CodeGen,
        base: []const u8,
        type_params: []const ast.StructTypeParam,
        type_bindings: std.StringHashMap(Type),
        val_bindings: ?std.StringHashMap(i64),
        comptime_nodes: ?std.StringHashMap(*Node),
    ) ![]const u8 {
        var buf = std.ArrayList(u8).empty;
        try buf.appendSlice(self.allocator, base);
        try buf.appendSlice(self.allocator, "__");
        for (type_params, 0..) |tp, i| {
            if (i > 0) try buf.append(self.allocator, '_');
            const constraint_name = if (tp.constraint.data == .type_expr) tp.constraint.data.type_expr.name else "";
            if (std.mem.eql(u8, constraint_name, "Type")) {
                if (type_bindings.get(tp.name)) |ty| {
                    const name = try ty.displayName(self.allocator);
                    try buf.appendSlice(self.allocator, name);
                }
            } else if (comptime_nodes != null) {
                if (comptime_nodes.?.get(tp.name)) |node| {
                    if (node.data == .string_literal) {
                        const hash = std.hash.Wyhash.hash(0, node.data.string_literal.raw);
                        var hash_buf: [16]u8 = undefined;
                        const hash_str = std.fmt.bufPrint(&hash_buf, "{x}", .{hash}) catch "0";
                        try buf.appendSlice(self.allocator, hash_str);
                    } else if (node.data == .int_literal) {
                        var int_buf: [20]u8 = undefined;
                        const int_str = std.fmt.bufPrint(&int_buf, "{d}", .{node.data.int_literal.value}) catch "0";
                        try buf.appendSlice(self.allocator, int_str);
                    }
                }
            } else if (val_bindings != null) {
                if (val_bindings.?.get(tp.name)) |val| {
                    var tmp: [20]u8 = undefined;
                    const s = std.fmt.bufPrint(&tmp, "{d}", .{val}) catch "0";
                    try buf.appendSlice(self.allocator, s);
                }
            }
        }
        return try buf.toOwnedSlice(self.allocator);
    }

    fn instantiateGeneric(self: *CodeGen, fd: ast.FnDecl, bindings: std.StringHashMap(Type), mangled: []const u8) !c.LLVMValueRef {
        // Save current codegen state
        const saved_function = self.current_function;
        const saved_return_type = self.current_return_type;
        const saved_insert_bb = c.LLVMGetInsertBlock(self.builder);

        // Save named_values
        var saved_named_values = std.StringHashMap(NamedValue).init(self.allocator);
        var nv_iter = self.named_values.iterator();
        while (nv_iter.next()) |entry| {
            try saved_named_values.put(entry.key_ptr.*, entry.value_ptr.*);
        }

        // Save scope_saves and defer_stack — generic body must not pollute caller's scope tracking
        const saved_scope_saves = self.scope_saves;
        const saved_defer_stack = self.defer_stack;
        self.scope_saves = std.ArrayList(std.ArrayList(ScopeEntry)).empty;
        self.defer_stack = std.ArrayList(std.ArrayList(*Node)).empty;

        // Set type param bindings
        self.type_param_bindings = bindings;
        defer self.type_param_bindings = null;

        // Build the specialized function type
        const fn_type = try self.buildFnType(fd.params, fd.return_type, mangled);
        const mangled_z = try self.allocator.dupeZ(u8, mangled);
        const function = c.LLVMAddFunction(self.module, mangled_z.ptr, fn_type);

        // Cache before generating body (in case of recursion)
        try self.generic_instances.put(mangled, function);

        // Generate body
        self.named_values.clearRetainingCapacity();
        self.current_function = function;

        const entry_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "entry");
        c.LLVMPositionBuilderAtEnd(self.builder, entry_bb);

        // Create allocas for parameters
        var llvm_param_idx: u32 = 0;
        for (fd.params) |param| {
            if (param.is_comptime) {
                // Comptime param: create a constant in named_values from the call-site value
                if (self.comptime_param_nodes) |cpn| {
                    if (cpn.get(param.name)) |node| {
                        if (node.data == .string_literal) {
                            const slit = node.data.string_literal;
                            const raw = slit.raw;
                            const inner = if (!slit.is_raw and raw.len >= 2 and raw[0] == '"' and raw[raw.len - 1] == '"')
                                raw[1 .. raw.len - 1]
                            else
                                raw;
                            const content = if (slit.is_raw) inner else try unescapeString(self.allocator, inner);
                            const str_val = self.buildConstStr(content);
                            const param_name_z = try self.allocator.dupeZ(u8, param.name);
                            const alloca = c.LLVMBuildAlloca(self.builder, self.getStringStructType(), param_name_z.ptr);
                            _ = c.LLVMBuildStore(self.builder, str_val, alloca);
                            try self.named_values.put(param.name, .{ .ptr = alloca, .ty = .string_type });
                        } else if (node.data == .int_literal) {
                            const ct_sx_ty = self.resolveType(param.type_expr);
                            const ct_llvm_ty = self.typeToLLVM(ct_sx_ty);
                            const const_val = c.LLVMConstInt(ct_llvm_ty, @bitCast(node.data.int_literal.value), 0);
                            const param_name_z = try self.allocator.dupeZ(u8, param.name);
                            const alloca = c.LLVMBuildAlloca(self.builder, ct_llvm_ty, param_name_z.ptr);
                            _ = c.LLVMBuildStore(self.builder, const_val, alloca);
                            try self.named_values.put(param.name, .{ .ptr = alloca, .ty = ct_sx_ty });
                        }
                    }
                }
                continue;
            }
            // Variadic params: use slice_type (same as genFnBodyAs)
            const sx_ty = if (param.is_variadic) blk: {
                const elem_name = if (param.type_expr.data == .type_expr) param.type_expr.data.type_expr.name else "s32";
                break :blk Type{ .slice_type = .{ .element_name = elem_name } };
            } else self.resolveType(param.type_expr);
            try self.bindParam(function, param.name, sx_ty, llvm_param_idx);
            llvm_param_idx += 1;
        }

        // Generate body statements
        const body = fd.body;
        if (body.data != .block) return self.emitError("generic function body must be a block");

        const ret_sx_type = self.resolveType(fd.return_type);
        self.current_return_type = ret_sx_type;

        var last_val: c.LLVMValueRef = null;
        for (body.data.block.stmts) |stmt| {
            last_val = try self.genStmt(stmt);
        }

        // Emit return if current block has no terminator
        const current_bb = c.LLVMGetInsertBlock(self.builder);
        if (c.LLVMGetBasicBlockTerminator(current_bb) == null) {
            if (ret_sx_type == .void_type) {
                _ = c.LLVMBuildRetVoid(self.builder);
            } else if (last_val) |val| {
                if (ret_sx_type.isStruct()) {
                    const sname = ret_sx_type.struct_type;
                    const info = self.struct_types.get(sname) orelse return self.emitErrorFmt("unknown struct type '{s}'", .{sname});
                    const loaded = c.LLVMBuildLoad2(self.builder, info.llvm_type, val, "retval");
                    _ = c.LLVMBuildRet(self.builder, loaded);
                } else {
                    const src_ty = self.llvmTypeToSxType(c.LLVMTypeOf(val));
                    const converted = self.convertValue(val, src_ty, ret_sx_type);
                    _ = c.LLVMBuildRet(self.builder, converted);
                }
            } else {
                _ = c.LLVMBuildUnreachable(self.builder);
            }
        }

        // Restore codegen state
        self.current_function = saved_function;
        self.current_return_type = saved_return_type;
        if (saved_insert_bb) |bb| {
            c.LLVMPositionBuilderAtEnd(self.builder, bb);
        }
        self.named_values.clearRetainingCapacity();
        var restore_iter = saved_named_values.iterator();
        while (restore_iter.next()) |entry| {
            try self.named_values.put(entry.key_ptr.*, entry.value_ptr.*);
        }
        saved_named_values.deinit();

        // Restore scope_saves and defer_stack
        self.scope_saves = saved_scope_saves;
        self.defer_stack = saved_defer_stack;

        return function;
    }

    fn genIfExpr(self: *CodeGen, if_expr: ast.IfExpr) !c.LLVMValueRef {
        // Generate condition
        var cond_val = try self.genExpr(if_expr.condition);

        // Ensure condition is i1 (bool)
        const cond_type = c.LLVMTypeOf(cond_val);
        const i1_type = c.LLVMInt1TypeInContext(self.context);
        if (cond_type != i1_type) {
            cond_val = c.LLVMBuildICmp(self.builder, c.LLVMIntNE, cond_val, c.LLVMConstInt(cond_type, 0, 0), "tobool");
        }

        const function = self.current_function;
        const has_else = if_expr.else_branch != null;

        var then_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "then");
        var else_bb: c.LLVMBasicBlockRef = if (has_else)
            c.LLVMAppendBasicBlockInContext(self.context, function, "else")
        else
            null;
        const merge_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "merge");

        const false_dest = if (has_else) else_bb else merge_bb;
        _ = c.LLVMBuildCondBr(self.builder, cond_val, then_bb, false_dest);

        // Then branch
        c.LLVMPositionBuilderAtEnd(self.builder, then_bb);
        const then_val = try self.genExpr(if_expr.then_branch);
        then_bb = c.LLVMGetInsertBlock(self.builder); // may have changed due to nested control flow
        _ = c.LLVMBuildBr(self.builder, merge_bb);

        // Else branch
        var else_val: c.LLVMValueRef = null;
        if (if_expr.else_branch) |else_branch| {
            c.LLVMPositionBuilderAtEnd(self.builder, else_bb);
            else_val = try self.genExpr(else_branch);
            else_bb = c.LLVMGetInsertBlock(self.builder);
            _ = c.LLVMBuildBr(self.builder, merge_bb);
        }

        // Merge block
        c.LLVMPositionBuilderAtEnd(self.builder, merge_bb);

        // PHI node if both branches produced values
        if (then_val != null and else_val != null) {
            const phi = c.LLVMBuildPhi(self.builder, c.LLVMTypeOf(then_val), "iftmp");
            var vals = [2]c.LLVMValueRef{ then_val, else_val };
            var blocks = [2]c.LLVMBasicBlockRef{ then_bb, else_bb };
            c.LLVMAddIncoming(phi, &vals, &blocks, 2);
            return phi;
        }

        return null;
    }

    fn genWhileExpr(self: *CodeGen, while_expr: ast.WhileExpr) !c.LLVMValueRef {
        const function = self.current_function;

        // Create basic blocks: condition, body, after
        const cond_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "while.cond");
        const body_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "while.body");
        const after_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "while.after");

        // Branch from current block to condition check
        _ = c.LLVMBuildBr(self.builder, cond_bb);

        // Condition block
        c.LLVMPositionBuilderAtEnd(self.builder, cond_bb);
        var cond_val = try self.genExpr(while_expr.condition);

        // Ensure condition is i1 (bool)
        const cond_type = c.LLVMTypeOf(cond_val);
        const i1_type = c.LLVMInt1TypeInContext(self.context);
        if (cond_type != i1_type) {
            cond_val = c.LLVMBuildICmp(self.builder, c.LLVMIntNE, cond_val, c.LLVMConstInt(cond_type, 0, 0), "tobool");
        }

        _ = c.LLVMBuildCondBr(self.builder, cond_val, body_bb, after_bb);

        // Body block — save and set loop context for break/continue
        c.LLVMPositionBuilderAtEnd(self.builder, body_bb);
        const saved_break_bb = self.loop_break_bb;
        const saved_continue_bb = self.loop_continue_bb;
        self.loop_break_bb = after_bb;
        self.loop_continue_bb = cond_bb;

        _ = try self.genExpr(while_expr.body);

        // Restore loop context
        self.loop_break_bb = saved_break_bb;
        self.loop_continue_bb = saved_continue_bb;

        // Branch back to condition (if not already terminated by break/return)
        const current_bb = c.LLVMGetInsertBlock(self.builder);
        if (c.LLVMGetBasicBlockTerminator(current_bb) == null) {
            _ = c.LLVMBuildBr(self.builder, cond_bb);
        }

        // Position at after block
        c.LLVMPositionBuilderAtEnd(self.builder, after_bb);

        return null;
    }

    fn genForExpr(self: *CodeGen, for_expr: ast.ForExpr) !c.LLVMValueRef {
        const function = self.current_function;
        const i64_type = c.LLVMInt64TypeInContext(self.context);

        // Determine iterable type and get length + element access info
        const iter_ty = self.inferType(for_expr.iterable);
        var len_val: c.LLVMValueRef = undefined;
        var elem_ty: Type = Type.s(64);
        var iter_ptr: c.LLVMValueRef = undefined; // pointer to data
        var is_slice = false;

        if (iter_ty.isSlice()) {
            is_slice = true;
            const info = iter_ty.slice_type;
            elem_ty = Type.fromName(info.element_name) orelse Type.s(64);
            // Load slice value from alloca
            if (for_expr.iterable.data == .identifier) {
                if (self.named_values.get(for_expr.iterable.data.identifier.name)) |entry| {
                    const slice_val = c.LLVMBuildLoad2(self.builder, self.getStringStructType(), entry.ptr, "for_slice");
                    iter_ptr = c.LLVMBuildExtractValue(self.builder, slice_val, 0, "for_ptr");
                    len_val = c.LLVMBuildExtractValue(self.builder, slice_val, 1, "for_len");
                } else return self.emitError("for: iterable not found");
            } else return self.emitError("for: slice iterable must be a variable");
        } else if (iter_ty.isArray()) {
            const info = iter_ty.array_type;
            elem_ty = Type.fromName(info.element_name) orelse Type.s(64);
            len_val = c.LLVMConstInt(i64_type, info.length, 0);
            // Get pointer to array
            if (for_expr.iterable.data == .identifier) {
                if (self.named_values.get(for_expr.iterable.data.identifier.name)) |entry| {
                    iter_ptr = entry.ptr;
                } else return self.emitError("for: iterable not found");
            } else return self.emitError("for: array iterable must be a variable");
        } else {
            return self.emitError("for loop requires a slice or array iterable");
        }

        const elem_llvm_ty = self.typeToLLVM(elem_ty);

        // Allocate it_index (s64) and it (element type)
        const idx_alloca = self.buildEntryBlockAlloca(i64_type, "it_index");
        _ = c.LLVMBuildStore(self.builder, c.LLVMConstInt(i64_type, 0, 0), idx_alloca);
        const it_alloca = self.buildEntryBlockAlloca(elem_llvm_ty, "it");

        // Push scope and bind it, it_index
        try self.pushScope();
        try self.named_values.put("it", .{ .ptr = it_alloca, .ty = elem_ty });
        try self.named_values.put("it_index", .{ .ptr = idx_alloca, .ty = Type.s(64) });

        // Create basic blocks
        const cond_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "for.cond");
        const body_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "for.body");
        const after_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "for.after");

        _ = c.LLVMBuildBr(self.builder, cond_bb);

        // Condition: it_index < len
        c.LLVMPositionBuilderAtEnd(self.builder, cond_bb);
        const cur_idx = c.LLVMBuildLoad2(self.builder, i64_type, idx_alloca, "cur_idx");
        const cond_val = c.LLVMBuildICmp(self.builder, c.LLVMIntSLT, cur_idx, len_val, "for_cond");
        _ = c.LLVMBuildCondBr(self.builder, cond_val, body_bb, after_bb);

        // Body: load it = iterable[it_index], then execute body
        c.LLVMPositionBuilderAtEnd(self.builder, body_bb);
        const body_idx = c.LLVMBuildLoad2(self.builder, i64_type, idx_alloca, "body_idx");

        if (is_slice) {
            // Slice: GEP through data pointer
            var gep_indices = [_]c.LLVMValueRef{body_idx};
            const gep = c.LLVMBuildGEP2(self.builder, elem_llvm_ty, iter_ptr, &gep_indices, 1, "for_elem");
            const elem_val = c.LLVMBuildLoad2(self.builder, elem_llvm_ty, gep, "it_val");
            _ = c.LLVMBuildStore(self.builder, elem_val, it_alloca);
        } else {
            // Array: GEP with [0, idx]
            const arr_llvm_ty = self.typeToLLVM(iter_ty);
            const zero = c.LLVMConstInt(i64_type, 0, 0);
            var indices = [_]c.LLVMValueRef{ zero, body_idx };
            const gep = c.LLVMBuildGEP2(self.builder, arr_llvm_ty, iter_ptr, &indices, 2, "for_elem");
            const elem_val = c.LLVMBuildLoad2(self.builder, elem_llvm_ty, gep, "it_val");
            _ = c.LLVMBuildStore(self.builder, elem_val, it_alloca);
        }

        // Save and set loop context for break/continue
        const saved_break_bb = self.loop_break_bb;
        const saved_continue_bb = self.loop_continue_bb;
        self.loop_break_bb = after_bb;
        self.loop_continue_bb = cond_bb;

        _ = try self.genExpr(for_expr.body);

        self.loop_break_bb = saved_break_bb;
        self.loop_continue_bb = saved_continue_bb;

        // Increment it_index
        const current_bb = c.LLVMGetInsertBlock(self.builder);
        if (c.LLVMGetBasicBlockTerminator(current_bb) == null) {
            const inc_idx = c.LLVMBuildLoad2(self.builder, i64_type, idx_alloca, "inc_idx");
            const next_idx = c.LLVMBuildAdd(self.builder, inc_idx, c.LLVMConstInt(i64_type, 1, 0), "next_idx");
            _ = c.LLVMBuildStore(self.builder, next_idx, idx_alloca);
            _ = c.LLVMBuildBr(self.builder, cond_bb);
        }

        c.LLVMPositionBuilderAtEnd(self.builder, after_bb);

        try self.popScope();

        return null;
    }

    fn genEnumLiteral(self: *CodeGen, variant_name: []const u8, enum_type_name: []const u8) c.LLVMValueRef {
        const i64_type = c.LLVMInt64TypeInContext(self.context);
        const variants = self.enum_types.get(enum_type_name) orelse return c.LLVMConstInt(i64_type, 0, 0);
        for (variants, 0..) |v, i| {
            if (std.mem.eql(u8, v, variant_name)) {
                return c.LLVMConstInt(i64_type, @intCast(i), 0);
            }
        }
        return c.LLVMConstInt(i64_type, 0, 0);
    }

    fn lookupVariantIndex(variants: ?[]const []const u8, name: []const u8) u64 {
        if (variants) |vs| {
            for (vs, 0..) |v, i| {
                if (std.mem.eql(u8, v, name)) return i;
            }
        }
        return 0;
    }

    fn genMatchExpr(self: *CodeGen, match: ast.MatchExpr) !c.LLVMValueRef {
        // Determine subject type for enum vs union dispatch
        var enum_name: ?[]const u8 = null;
        var union_name: ?[]const u8 = null;
        if (match.subject.data == .identifier) {
            if (self.named_values.get(match.subject.data.identifier.name)) |entry| {
                if (entry.ty.isEnum()) enum_name = entry.ty.enum_type;
                if (entry.ty.isUnion()) union_name = entry.ty.union_type;
            }
        }

        // Get the switch value: for unions, load the tag from field 0; for enums, use the value directly
        const subject_val: c.LLVMValueRef = if (union_name != null) blk: {
            // Union: load tag from field 0 of the alloca
            const entry = self.named_values.get(match.subject.data.identifier.name).?;
            const info = self.union_types.get(union_name.?).?;
            const tag_gep = c.LLVMBuildStructGEP2(self.builder, info.llvm_type, entry.ptr, 0, "tag");
            break :blk c.LLVMBuildLoad2(self.builder, c.LLVMInt64TypeInContext(self.context), tag_gep, "tag_val");
        } else try self.genExpr(match.subject);

        const variants: ?[]const []const u8 = if (union_name) |un|
            (if (self.union_types.get(un)) |info| info.variant_names else null)
        else if (enum_name) |en|
            self.enum_types.get(en)
        else
            null;

        const function = self.current_function;
        const i64_type = c.LLVMInt64TypeInContext(self.context);
        const merge_bb = c.LLVMAppendBasicBlockInContext(self.context, function, "match_end");

        // Create case basic blocks
        var case_bbs = std.ArrayList(c.LLVMBasicBlockRef).empty;
        for (match.arms) |_| {
            try case_bbs.append(self.allocator, c.LLVMAppendBasicBlockInContext(self.context, function, "case"));
        }

        // Find else arm (null pattern) — use its BB as the switch default
        var else_arm_idx: ?usize = null;
        for (match.arms, 0..) |arm, i| {
            if (arm.pattern == null) {
                else_arm_idx = i;
                break;
            }
        }
        const default_bb = if (else_arm_idx) |idx|
            case_bbs.items[idx]
        else
            c.LLVMAppendBasicBlockInContext(self.context, function, "match_default");

        // Build switch instruction
        const sw = c.LLVMBuildSwitch(self.builder, subject_val, default_bb, @intCast(match.arms.len));
        for (match.arms, 0..) |arm, i| {
            const pat = arm.pattern orelse continue; // skip else arm
            if (pat.data == .enum_literal) {
                const idx = lookupVariantIndex(variants, pat.data.enum_literal.name);
                const case_val = c.LLVMConstInt(i64_type, idx, 0);
                c.LLVMAddCase(sw, case_val, case_bbs.items[i]);
            } else if (pat.data == .type_expr) {
                // Type-match: resolve type name to Any tag value(s)
                const tag_values = try self.resolveTypeMatchTags(pat.data.type_expr.name);
                for (tag_values) |tag| {
                    c.LLVMAddCase(sw, c.LLVMConstInt(i64_type, tag, 0), case_bbs.items[i]);
                }
            } else if (pat.data == .identifier) {
                // Named type (struct/enum/union name) or category (int/float)
                const tag_values = try self.resolveTypeMatchTags(pat.data.identifier.name);
                for (tag_values) |tag| {
                    c.LLVMAddCase(sw, c.LLVMConstInt(i64_type, tag, 0), case_bbs.items[i]);
                }
            }
        }

        // Generate arm bodies and collect PHI info
        var phi_vals = std.ArrayList(c.LLVMValueRef).empty;
        var phi_bbs = std.ArrayList(c.LLVMBasicBlockRef).empty;
        var has_value = false;
        var value_type: c.LLVMTypeRef = null;

        // Pre-collect tag values for each arm (for runtime dispatch context)
        var arm_tag_values = std.ArrayList([]const u64).empty;
        for (match.arms) |arm| {
            const tag_values: []const u64 = if (arm.pattern) |pat| blk: {
                break :blk if (pat.data == .type_expr)
                    try self.resolveTypeMatchTags(pat.data.type_expr.name)
                else if (pat.data == .identifier)
                    try self.resolveTypeMatchTags(pat.data.identifier.name)
                else
                    &.{};
            } else &.{};
            try arm_tag_values.append(self.allocator, tag_values);
        }

        for (match.arms, 0..) |arm, i| {
            c.LLVMPositionBuilderAtEnd(self.builder, case_bbs.items[i]);
            if (arm.is_break) {
                _ = c.LLVMBuildBr(self.builder, merge_bb);
            } else if (arm.pattern != null and arm_tag_values.items[i].len == 0 and
                (arm.pattern.?.data == .identifier or arm.pattern.?.data == .type_expr))
            {
                // Category/type arm with no matching types — BB is unreachable, skip body
                _ = c.LLVMBuildBr(self.builder, merge_bb);
            } else {
                // Set match arm context for runtime type dispatch
                const saved_match_tags = self.current_match_tags;
                self.current_match_tags = arm_tag_values.items[i];
                const val = try self.genExpr(arm.body);
                self.current_match_tags = saved_match_tags;
                const bb = c.LLVMGetInsertBlock(self.builder);
                _ = c.LLVMBuildBr(self.builder, merge_bb);
                if (val != null and c.LLVMGetTypeKind(c.LLVMTypeOf(val)) != c.LLVMVoidTypeKind) {
                    has_value = true;
                    if (value_type == null) value_type = c.LLVMTypeOf(val);
                    try phi_vals.append(self.allocator, val);
                    try phi_bbs.append(self.allocator, bb);
                }
            }
        }

        // Default block branches to merge (only if no else arm — else arm's body already generated above)
        if (else_arm_idx == null) {
            c.LLVMPositionBuilderAtEnd(self.builder, default_bb);
            _ = c.LLVMBuildBr(self.builder, merge_bb);
        }

        // Merge block
        c.LLVMPositionBuilderAtEnd(self.builder, merge_bb);

        if (has_value and value_type != null) {
            const undef_val = c.LLVMGetUndef(value_type);
            // Add undef entries for break arms and default block
            for (match.arms, 0..) |arm, i| {
                if (arm.is_break) {
                    try phi_vals.append(self.allocator, undef_val);
                    try phi_bbs.append(self.allocator, case_bbs.items[i]);
                }
            }
            if (else_arm_idx == null) {
                try phi_vals.append(self.allocator, undef_val);
                try phi_bbs.append(self.allocator, default_bb);
            }

            const vals_slice = try phi_vals.toOwnedSlice(self.allocator);
            const bbs_slice = try phi_bbs.toOwnedSlice(self.allocator);
            const phi = c.LLVMBuildPhi(self.builder, value_type, "matchtmp");
            c.LLVMAddIncoming(phi, vals_slice.ptr, bbs_slice.ptr, @intCast(vals_slice.len));
            return phi;
        }

        return null;
    }

    /// Resolve a type name to one or more Any tag values for type-switch matching.
    /// Categories: "int" matches s32+s64, "float" matches f32+f64.
    /// Specific types: "s32", "f64", "string", "bool", "Type".
    /// Named types: struct/enum/union names get dynamic IDs.
    fn resolveTypeMatchTags(self: *CodeGen, name: []const u8) ![]const u64 {
        // Category aliases
        if (std.mem.eql(u8, name, "int")) {
            const tags = try self.allocator.alloc(u64, 2);
            tags[0] = ANY_TAG_S32;
            tags[1] = ANY_TAG_S64;
            return tags;
        }
        if (std.mem.eql(u8, name, "float")) {
            const tags = try self.allocator.alloc(u64, 2);
            tags[0] = ANY_TAG_F32;
            tags[1] = ANY_TAG_F64;
            return tags;
        }
        // Type category aliases: "struct", "enum", "union", "vector", "array", "slice"
        const category: ?TypeCategory = if (std.mem.eql(u8, name, "struct"))
            .struct_cat
        else if (std.mem.eql(u8, name, "enum"))
            .enum_cat
        else if (std.mem.eql(u8, name, "union"))
            .union_cat
        else if (std.mem.eql(u8, name, "vector"))
            .vector_cat
        else if (std.mem.eql(u8, name, "array"))
            .array_cat
        else if (std.mem.eql(u8, name, "slice"))
            .slice_cat
        else if (std.mem.eql(u8, name, "pointer"))
            .pointer_cat
        else
            null;
        if (category) |cat| {
            var tag_list = std.ArrayList(u64).empty;
            var it = self.any_type_entries.iterator();
            while (it.next()) |entry| {
                if (entry.value_ptr.category == cat) {
                    try tag_list.append(self.allocator, entry.value_ptr.tag_id);
                }
            }
            if (tag_list.items.len > 0) {
                return try tag_list.toOwnedSlice(self.allocator);
            }
            // No types registered for this category — return empty slice
            return &.{};
        }
        // Specific builtin types
        const single_tag: ?u64 = if (std.mem.eql(u8, name, "bool"))
            ANY_TAG_BOOL
        else if (std.mem.eql(u8, name, "s32"))
            ANY_TAG_S32
        else if (std.mem.eql(u8, name, "s64"))
            ANY_TAG_S64
        else if (std.mem.eql(u8, name, "f32"))
            ANY_TAG_F32
        else if (std.mem.eql(u8, name, "f64"))
            ANY_TAG_F64
        else if (std.mem.eql(u8, name, "string"))
            ANY_TAG_STRING
        else if (std.mem.eql(u8, name, "Type") or std.mem.eql(u8, name, "type"))
            ANY_TAG_TYPE
        else if (std.mem.eql(u8, name, "void"))
            ANY_TAG_VOID
        else
            null;
        if (single_tag) |t| {
            const tags = try self.allocator.alloc(u64, 1);
            tags[0] = t;
            return tags;
        }
        // Named type (struct/enum/union) — get dynamic ID
        const sx_type: Type = if (self.struct_types.contains(name))
            .{ .struct_type = name }
        else if (self.enum_types.contains(name))
            .{ .enum_type = name }
        else if (self.union_types.contains(name))
            .{ .union_type = name }
        else
            .{ .struct_type = name }; // fallback
        const id = try self.getAnyTypeId(name, sx_type);
        const tags = try self.allocator.alloc(u64, 1);
        tags[0] = id;
        return tags;
    }

    /// Resolve a callee node to a function name string for type inference.
    /// Handles identifiers, namespaced calls, and intra-namespace fallback.
    fn resolveCalleeName(self: *CodeGen, call_node: ast.Call) ?[]const u8 {
        if (call_node.callee.data == .identifier) {
            return call_node.callee.data.identifier.name;
        }
        if (call_node.callee.data == .field_access) {
            const fa = call_node.callee.data.field_access;
            if (fa.object.data == .identifier) {
                if (self.namespaces.contains(fa.object.data.identifier.name)) {
                    return std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ fa.object.data.identifier.name, fa.field }) catch return null;
                }
            }
        }
        return null;
    }

    /// Resolve a builtin parameterized type (e.g. Vector(3, f32)).
    /// Strips namespace prefix to get base name, then dispatches.
    fn resolveBuiltinType(self: *CodeGen, name: []const u8, args: []const *Node) ?Type {
        const base = if (std.mem.lastIndexOfScalar(u8, name, '.')) |idx| name[idx + 1 ..] else name;
        if (std.mem.eql(u8, base, "Vector")) {
            if (args.len >= 2) {
                const n: u32 = @intCast(self.resolveValueArg(args[0]));
                const elem = self.resolveType(args[1]);
                const elem_name = elem.displayName(self.allocator) catch return null;
                const ty: Type = .{ .vector_type = .{ .element_name = elem_name, .length = n } };
                // Pre-register in any_type_entries so runtime dispatch knows about this type
                const any_name = std.fmt.allocPrint(self.allocator, "vec[{d}]{s}", .{ n, elem_name }) catch return null;
                _ = self.getAnyTypeId(any_name, ty) catch return null;
                return ty;
            }
        }
        if (std.mem.eql(u8, base, "Array")) {
            if (args.len >= 2) {
                const n: u32 = @intCast(self.resolveValueArg(args[0]));
                const elem = self.resolveType(args[1]);
                const elem_name = elem.displayName(self.allocator) catch return null;
                const ty: Type = .{ .array_type = .{ .element_name = elem_name, .length = n } };
                const any_name = std.fmt.allocPrint(self.allocator, "[{d}]{s}", .{ n, elem_name }) catch return null;
                _ = self.getAnyTypeId(any_name, ty) catch return null;
                return ty;
            }
        }
        return null;
    }

    fn dispatchBuiltin(self: *CodeGen, name: []const u8, call_node: ast.Call) !c.LLVMValueRef {
        // Extract base name (strip namespace prefix)
        const base = if (std.mem.lastIndexOfScalar(u8, name, '.')) |idx| name[idx + 1 ..] else name;
        if (std.mem.eql(u8, base, "write")) return self.genWriteCall(call_node.args);
        if (std.mem.eql(u8, base, "sqrt")) return self.genSqrt(call_node);
        if (std.mem.eql(u8, base, "sin")) return self.genSin(call_node);
        if (std.mem.eql(u8, base, "cos")) return self.genCos(call_node);
        if (std.mem.eql(u8, base, "size_of")) return self.genSizeOf(call_node);
        if (std.mem.eql(u8, base, "cast")) return self.genCast(call_node);
        if (std.mem.eql(u8, base, "alloc")) return self.genAlloc(call_node.args);
        if (std.mem.eql(u8, base, "malloc")) return self.genMalloc(call_node.args);
        if (std.mem.eql(u8, base, "free")) return self.genFree(call_node.args);
        if (std.mem.eql(u8, base, "memcpy")) return self.genMemcpy(call_node.args);
        if (std.mem.eql(u8, base, "type_of")) return self.genTypeOf(call_node);
        if (std.mem.eql(u8, base, "type_name")) return self.genTypeName(call_node);
        if (std.mem.eql(u8, base, "field_count")) return self.genFieldCount(call_node);
        if (std.mem.eql(u8, base, "field_name")) return self.genFieldName(call_node);
        if (std.mem.eql(u8, base, "field_value")) return self.genFieldValue(call_node);
        return self.emitErrorFmt("unknown builtin function '{s}'", .{name});
    }

    fn genWriteCall(self: *CodeGen, args: []const *Node) !c.LLVMValueRef {
        if (args.len != 1) return self.emitError("write expects exactly 1 argument");
        const builtins = self.builtins orelse return self.emitError("builtins not available (missing #builtin import)");
        const val = try self.genExpr(args[0]);
        // Extract ptr and len from string slice
        const ptr = c.LLVMBuildExtractValue(self.builder, val, 0, "str_ptr");
        const len_i64 = c.LLVMBuildExtractValue(self.builder, val, 1, "str_len");
        // printf %.*s precision is C int (i32) — truncate from i64
        const len = c.LLVMBuildTrunc(self.builder, len_i64, c.LLVMInt32TypeInContext(self.context), "len_trunc");
        const fmt = c.LLVMBuildGlobalStringPtr(self.builder, "%.*s", "write_fmt");
        const printf_fn = builtins.printf_fn;
        const fn_type = c.LLVMGlobalGetValueType(printf_fn);
        var call_args = [_]c.LLVMValueRef{ fmt, len, ptr };
        _ = c.LLVMBuildCall2(self.builder, fn_type, printf_fn, &call_args, 3, "");
        return null;
    }

    /// Helper: build a constant string slice in the current function
    fn buildConstStr(self: *CodeGen, s: []const u8) c.LLVMValueRef {
        const sz = self.allocator.dupeZ(u8, s) catch unreachable;
        const ptr = c.LLVMBuildGlobalStringPtr(self.builder, sz.ptr, "cstr");
        return self.buildStringSlice(ptr, @intCast(s.len));
    }

    /// Helper: build a constant string slice as a global constant (no builder needed).
    fn buildConstStrGlobal(self: *CodeGen, s: []const u8) c.LLVMValueRef {
        const sz = self.allocator.dupeZ(u8, s) catch unreachable;
        const i64_ty = c.LLVMInt64TypeInContext(self.context);
        const i8_ty = c.LLVMInt8TypeInContext(self.context);
        // Create a global string constant
        const str_const = c.LLVMConstStringInContext(self.context, sz.ptr, @intCast(s.len), 0);
        const global_name = (self.allocator.dupeZ(u8, std.fmt.allocPrint(self.allocator, ".str.{s}", .{s}) catch unreachable)) catch unreachable;
        var global = c.LLVMGetNamedGlobal(self.module, global_name.ptr);
        if (global == null) {
            const arr_ty = c.LLVMArrayType2(i8_ty, s.len + 1);
            global = c.LLVMAddGlobal(self.module, arr_ty, global_name.ptr);
            c.LLVMSetInitializer(global, str_const);
            c.LLVMSetGlobalConstant(global, 1);
            c.LLVMSetLinkage(global, c.LLVMPrivateLinkage);
        }
        // Build constant struct {ptr, i64}
        var fields = [_]c.LLVMValueRef{
            c.LLVMConstBitCast(global.?, c.LLVMPointerTypeInContext(self.context, 0)),
            c.LLVMConstInt(i64_ty, s.len, 0),
        };
        return c.LLVMConstStructInContext(self.context, &fields, 2, 0);
    }

    /// Check if a node refers to a type name. Returns the raw name or null.
    fn asTypeName(self: *CodeGen, node: *const Node) ?[]const u8 {
        if (node.data == .type_expr) return node.data.type_expr.name;
        if (node.data == .identifier) {
            const id = node.data.identifier.name;
            if (self.resolveTypeName(id) != null) return id;
        }
        return null;
    }

    /// Resolve a type name to its display string (null-terminated) for runtime use.
    fn resolveDisplayName(self: *CodeGen, name: []const u8) [:0]const u8 {
        const display = self.resolveTypeName(name) orelse name;
        return self.allocator.dupeZ(u8, display) catch unreachable;
    }

    /// Convert a type expression AST node to its source string representation.
    fn typeNodeToString(self: *CodeGen, node: *const Node) []const u8 {
        return switch (node.data) {
            .type_expr => |te| te.name,
            .identifier => |id| id.name,
            .pointer_type_expr => |pte| std.fmt.allocPrint(self.allocator, "*{s}", .{self.typeNodeToString(pte.pointee_type)}) catch "?",
            .many_pointer_type_expr => |mpte| std.fmt.allocPrint(self.allocator, "[*]{s}", .{self.typeNodeToString(mpte.element_type)}) catch "?",
            .slice_type_expr => |ste| std.fmt.allocPrint(self.allocator, "[]{s}", .{self.typeNodeToString(ste.element_type)}) catch "?",
            .array_type_expr => |ate| blk: {
                const elem = self.typeNodeToString(ate.element_type);
                if (ate.length.data == .int_literal) {
                    break :blk std.fmt.allocPrint(self.allocator, "[{d}]{s}", .{ ate.length.data.int_literal.value, elem }) catch "?";
                }
                break :blk std.fmt.allocPrint(self.allocator, "[?]{s}", .{elem}) catch "?";
            },
            else => "?",
        };
    }

    /// Build a function type string like "() -> s32" from an fn_decl.
    fn buildFnSignature(self: *CodeGen, fd: ast.FnDecl) []const u8 {
        var buf = std.ArrayList(u8).empty;
        buf.appendSlice(self.allocator, "(") catch return "?";
        for (fd.params, 0..) |param, i| {
            if (i > 0) buf.appendSlice(self.allocator, ", ") catch {};
            if (param.is_variadic) buf.appendSlice(self.allocator, "..") catch {};
            const ty_str = self.typeNodeToString(param.type_expr);
            buf.appendSlice(self.allocator, ty_str) catch {};
        }
        buf.appendSlice(self.allocator, ")") catch {};
        if (fd.return_type) |rt| {
            buf.appendSlice(self.allocator, " -> ") catch {};
            buf.appendSlice(self.allocator, self.typeNodeToString(rt)) catch {};
        }
        return buf.toOwnedSlice(self.allocator) catch "?";
    }

    /// Extract a qualified name from a callee expression (identifier or field_access chain).
    fn calleeToQualifiedName(self: *CodeGen, callee: *Node) ?[]const u8 {
        if (callee.data == .identifier) return callee.data.identifier.name;
        if (callee.data == .field_access) {
            const obj_name = self.calleeToQualifiedName(callee.data.field_access.object) orelse return null;
            return std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ obj_name, callee.data.field_access.field }) catch null;
        }
        return null;
    }

    /// Resolve a name to a type display string, or null if not a type.
    fn resolveTypeName(self: *CodeGen, name: []const u8) ?[]const u8 {
        const resolved = self.type_aliases.get(name) orelse name;
        if (self.struct_types.get(resolved)) |info| return info.display_name orelse resolved;
        if (self.resolveTypeFromName(name) != null) return resolved;
        return null;
    }

    /// Resolve a type name to a Type, checking primitives + registered structs/unions/enums.
    /// Unlike Type.fromName which only handles primitives.
    fn resolveTypeFromName(self: *CodeGen, name: []const u8) ?Type {
        // Primitives
        if (Type.fromName(name)) |t| return t;
        // Structs
        if (self.struct_types.contains(name)) return .{ .struct_type = name };
        // Unions
        if (self.union_types.contains(name)) return .{ .union_type = name };
        // Enums
        if (self.enum_types.contains(name)) return .{ .enum_type = name };
        // Vector display name: "Vector(N,T)"
        if (name.len > 8 and std.mem.startsWith(u8, name, "Vector(") and name[name.len - 1] == ')') {
            const inner = name[7 .. name.len - 1]; // "N,T"
            if (std.mem.indexOfScalar(u8, inner, ',')) |comma| {
                const n_str = inner[0..comma];
                const elem = inner[comma + 1 ..];
                const length = std.fmt.parseInt(u32, n_str, 10) catch return null;
                return .{ .vector_type = .{ .element_name = elem, .length = length } };
            }
        }
        // Type aliases
        if (self.type_aliases.get(name)) |target| return self.resolveTypeFromName(target);
        return null;
    }

    fn inferType(self: *CodeGen, node: *Node) Type {
        return switch (node.data) {
            .int_literal => Type.s(64),
            .float_literal => .f32,
            .bool_literal => .boolean,
            .string_literal => .string_type,
            .insert_expr => .void_type,
            .comptime_expr => |ct| self.inferType(ct.expr),
            .binary_op => |binop| {
                switch (binop.op) {
                    .eq, .neq, .lt, .lte, .gt, .gte, .and_op, .or_op => return .boolean,
                    else => {
                        const lhs_ty = self.inferType(binop.lhs);
                        const rhs_ty = self.inferType(binop.rhs);
                        return Type.widen(lhs_ty, rhs_ty);
                    },
                }
            },
            .chained_comparison => return .boolean,
            .identifier => |ident| {
                if (self.named_values.get(ident.name)) |entry| return entry.ty;
                if (self.comptime_globals.get(ident.name)) |ct| return ct.ty;
                if (self.global_mutable_vars.get(ident.name)) |entry| return entry.ty;
                return Type.s(64);
            },
            .if_expr => |ie| {
                return self.inferType(ie.then_branch);
            },
            .block => |blk| {
                if (blk.stmts.len > 0) {
                    return self.inferType(blk.stmts[blk.stmts.len - 1]);
                }
                return .void_type;
            },
            .union_literal => |ul| {
                if (ul.union_name) |uname| return .{ .union_type = uname };
                if (self.current_return_type.isUnion()) return self.current_return_type;
                return .void_type;
            },
            .enum_literal => {
                if (self.current_return_type.isEnum()) return self.current_return_type;
                if (self.current_return_type.isUnion()) return self.current_return_type;
                return .{ .enum_type = "" };
            },
            .match_expr => |me| {
                for (me.arms) |arm| {
                    if (!arm.is_break) return self.inferType(arm.body);
                }
                return .void_type;
            },
            .call => |call_node| {
                // Check for union literal pattern: Type.variant(payload)
                if (call_node.callee.data == .field_access) {
                    const fa = call_node.callee.data.field_access;
                    const obj_ty = blk: {
                        if (fa.object.data == .identifier) {
                            const name = self.type_aliases.get(fa.object.data.identifier.name) orelse fa.object.data.identifier.name;
                            if (self.union_types.contains(name)) break :blk Type{ .union_type = name };
                        }
                        const ty = self.resolveType(fa.object);
                        if (ty.isUnion()) break :blk ty;
                        break :blk @as(?Type, null);
                    };
                    if (obj_ty) |uty| return uty;
                }
                const callee_name = self.resolveCalleeName(call_node) orelse return Type.s(64);
                const base_name = if (std.mem.lastIndexOfScalar(u8, callee_name, '.')) |idx| callee_name[idx + 1 ..] else callee_name;
                // Built-in: sqrt/sin/cos returns same type as argument
                if (std.mem.eql(u8, base_name, "sqrt") or
                    std.mem.eql(u8, base_name, "sin") or
                    std.mem.eql(u8, base_name, "cos"))
                {
                    if (call_node.args.len > 0) return self.inferType(call_node.args[0]);
                    return .f32;
                }
                // Built-in: size_of returns s64
                if (std.mem.eql(u8, base_name, "size_of")) return Type.s(64);
                // Built-in: type_of returns s64 (type tag)
                if (std.mem.eql(u8, base_name, "type_of")) return Type.s(64);
                // Built-in: type_name returns string
                if (std.mem.eql(u8, base_name, "type_name")) return .string_type;
                // Built-in: field_count returns s64
                if (std.mem.eql(u8, base_name, "field_count")) return Type.s(64);
                // Built-in: field_name returns string
                if (std.mem.eql(u8, base_name, "field_name")) return .string_type;
                // Built-in: field_value returns Any
                if (std.mem.eql(u8, base_name, "field_value")) return .{ .any_type = {} };
                // Built-in: cast returns the target type (first arg)
                if (std.mem.eql(u8, base_name, "cast")) {
                    if (call_node.args.len > 0) return self.resolveType(call_node.args[0]);
                    return Type.s(64);
                }
                // Built-in: alloc returns string
                if (std.mem.eql(u8, base_name, "alloc")) return .string_type;
                if (std.mem.eql(u8, base_name, "malloc")) return .{ .pointer_type = .{ .pointee_name = "void" } };
                if (std.mem.eql(u8, base_name, "free")) return .void_type;
                if (std.mem.eql(u8, base_name, "memcpy")) return .void_type;
                // Check generic templates — infer return type from widened bindings
                const template = self.generic_templates.get(callee_name) orelse blk: {
                    // Intra-namespace fallback
                    if (self.current_namespace) |ns| {
                        const qualified = std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name }) catch return Type.s(64);
                        break :blk self.generic_templates.get(qualified);
                    }
                    break :blk null;
                };
                if (template) |tmpl| {
                    const gfd = tmpl.fd;
                    // Build widened type bindings from all call args
                    var inferred_bindings = std.StringHashMap(Type).init(self.allocator);
                    for (gfd.params, 0..) |param, pi| {
                        if (param.type_expr.data == .type_expr) {
                            for (gfd.type_params) |tp| {
                                if (std.mem.eql(u8, tp.name, param.type_expr.data.type_expr.name)) {
                                    if (pi < call_node.args.len) {
                                        const arg_ty = self.inferType(call_node.args[pi]);
                                        if (inferred_bindings.get(tp.name)) |existing| {
                                            inferred_bindings.put(tp.name, Type.widen(existing, arg_ty)) catch {};
                                        } else {
                                            inferred_bindings.put(tp.name, arg_ty) catch {};
                                        }
                                    }
                                    break;
                                }
                            }
                        }
                    }
                    // Resolve return type from bindings
                    if (gfd.return_type) |rt| {
                        if (rt.data == .type_expr) {
                            if (inferred_bindings.get(rt.data.type_expr.name)) |bound_ty| {
                                return bound_ty;
                            }
                        }
                        // Try resolving as a concrete type (e.g. -> string, -> s32)
                        const resolved = self.resolveType(rt);
                        if (!std.meta.eql(resolved, Type.void_type)) return resolved;
                    }
                    return Type.s(64);
                }
                // Check declared return types (preserves signedness)
                if (self.function_return_types.get(callee_name)) |ret_ty| return ret_ty;
                if (self.current_namespace) |ns| {
                    const qualified = std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns, callee_name }) catch return Type.s(64);
                    if (self.function_return_types.get(qualified)) |ret_ty| return ret_ty;
                }
                // Fallback: check non-generic LLVM functions
                const callee_name_z = self.allocator.dupeZ(u8, callee_name) catch return Type.s(64);
                var callee_fn_opt = c.LLVMGetNamedFunction(self.module, callee_name_z.ptr);
                // Intra-namespace fallback
                if (callee_fn_opt == null) {
                    if (self.current_namespace) |ns2| {
                        const q = std.fmt.allocPrint(self.allocator, "{s}.{s}", .{ ns2, callee_name }) catch return Type.s(64);
                        const qz = self.allocator.dupeZ(u8, q) catch return Type.s(64);
                        callee_fn_opt = c.LLVMGetNamedFunction(self.module, qz.ptr);
                    }
                }
                if (callee_fn_opt) |callee_fn| {
                    const fn_type = c.LLVMGlobalGetValueType(callee_fn);
                    const ret_llvm = c.LLVMGetReturnType(fn_type);
                    return self.llvmTypeToSxType(ret_llvm);
                }
                // Check if callee is a variable with function pointer type
                {
                    const fp_entry = if (self.named_values.get(callee_name)) |e| e
                        else self.global_mutable_vars.get(callee_name);
                    if (fp_entry) |entry| {
                        if (entry.ty.isFunctionType()) {
                            return entry.ty.function_type.return_type.*;
                        }
                    }
                }
                return Type.s(64);
            },
            .unary_op => |unop| {
                if (unop.op == .address_of) {
                    const operand_ty = self.inferType(unop.operand);
                    const name = operand_ty.displayName(self.allocator) catch return Type.s(64);
                    return .{ .pointer_type = .{ .pointee_name = name } };
                }
                return self.inferType(unop.operand);
            },
            .deref_expr => |de| {
                const ptr_ty = self.inferType(de.operand);
                if (ptr_ty.isPointer()) return self.resolveTypeFromName(ptr_ty.pointer_type.pointee_name) orelse Type.s(64);
                return Type.s(64);
            },
            .null_literal => return .{ .pointer_type = .{ .pointee_name = "void" } },
            .field_access => |fa| {
                var obj_ty = self.inferType(fa.object);
                // Auto-deref: if pointer, unwrap to pointee
                if (obj_ty.isPointer()) {
                    obj_ty = self.resolveTypeFromName(obj_ty.pointer_type.pointee_name) orelse Type.s(64);
                }
                if (obj_ty == .string_type) {
                    if (std.mem.eql(u8, fa.field, "len")) return Type.s(64);
                    if (std.mem.eql(u8, fa.field, "ptr")) return .{ .pointer_type = .{ .pointee_name = "u8" } };
                }
                if (obj_ty.isSlice()) {
                    if (std.mem.eql(u8, fa.field, "len")) return Type.s(64);
                }
                if (obj_ty.isArray()) {
                    if (std.mem.eql(u8, fa.field, "len")) return Type.s(64);
                }
                if (obj_ty.isAny()) {
                    if (std.mem.eql(u8, fa.field, "tag")) return Type.s(64);
                    if (std.mem.eql(u8, fa.field, "value")) return Type.s(64);
                }
                if (obj_ty.isVector()) {
                    return obj_ty.vectorElementType() orelse Type.s(64);
                }
                if (obj_ty.isStruct()) {
                    if (self.struct_types.get(obj_ty.struct_type)) |info| {
                        if (self.findFieldIndex(info, fa.field)) |idx| {
                            return info.field_types[idx];
                        }
                    }
                }
                if (obj_ty.isUnion()) {
                    if (self.union_types.get(obj_ty.union_type)) |info| {
                        for (info.variant_names, 0..) |vn, i| {
                            if (std.mem.eql(u8, vn, fa.field)) {
                                return info.variant_types[i];
                            }
                        }
                    }
                }
                return Type.s(64);
            },
            .index_expr => |ie| {
                const obj_ty = self.inferType(ie.object);
                if (obj_ty == .string_type) return Type.u(8);
                if (obj_ty.isVector()) {
                    return obj_ty.vectorElementType() orelse Type.s(64);
                }
                if (obj_ty.isArray()) {
                    return Type.fromName(obj_ty.array_type.element_name) orelse Type.s(64);
                }
                if (obj_ty.isSlice()) {
                    return obj_ty.sliceElementType() orelse Type.s(64);
                }
                if (obj_ty.isManyPointer()) {
                    return self.resolveTypeFromName(obj_ty.many_pointer_type.element_name) orelse Type.s(64);
                }
                return Type.s(64);
            },
            .slice_expr => |se| {
                const obj_ty = self.inferType(se.object);
                if (obj_ty == .string_type) return .{ .slice_type = .{ .element_name = "u8" } };
                if (obj_ty.isArray()) return .{ .slice_type = .{ .element_name = obj_ty.array_type.element_name } };
                if (obj_ty.isSlice()) return obj_ty;
                return .void_type;
            },
            .array_literal => |al| {
                if (al.elements.len == 0) return .void_type;
                const elem_ty = self.inferType(al.elements[0]);
                const elem_name = elem_ty.displayName(self.allocator) catch return Type.s(64);
                return .{ .array_type = .{ .element_name = elem_name, .length = @intCast(al.elements.len) } };
            },
            .while_expr, .for_expr, .break_expr, .continue_expr => .void_type,
            else => Type.s(64),
        };
    }

    fn exprIsFloat(self: *CodeGen, node: *Node) bool {
        return self.inferType(node).isFloat();
    }

    pub fn verify(self: *CodeGen) !void {
        var err_msg: [*c]u8 = null;
        if (c.LLVMVerifyModule(self.module, c.LLVMReturnStatusAction, &err_msg) != 0) {
            defer c.LLVMDisposeMessage(err_msg);
            const msg = std.mem.span(err_msg);
            return self.emitErrorFmt("LLVM verification failed: {s}", .{msg});
        }
    }

    pub fn printIR(self: *CodeGen) void {
        const ir = c.LLVMPrintModuleToString(self.module);
        defer c.LLVMDisposeMessage(ir);
        const len = std.mem.len(ir);
        std.debug.print("{s}\n", .{ir[0..len]});
    }

    fn emitToFile(self: *CodeGen, output_path: [*:0]const u8, file_type: c.LLVMCodeGenFileType) !void {
        llvm.initAllTargets();

        const cfg = self.target_config;
        const triple_owned = cfg.triple == null;
        const triple = cfg.triple orelse c.LLVMGetDefaultTargetTriple();
        defer if (triple_owned) c.LLVMDisposeMessage(@constCast(triple));

        var target: c.LLVMTargetRef = null;
        var err_msg: [*c]u8 = null;

        if (c.LLVMGetTargetFromTriple(triple, &target, &err_msg) != 0) {
            defer c.LLVMDisposeMessage(err_msg);
            const msg = std.mem.span(err_msg);
            return self.emitErrorFmt("failed to get target: {s}", .{msg});
        }

        const tm = c.LLVMCreateTargetMachine(
            target,
            triple,
            cfg.getCpu(),
            cfg.getFeatures(),
            cfg.opt_level.toLLVM(),
            c.LLVMRelocPIC,
            c.LLVMCodeModelDefault,
        );
        defer c.LLVMDisposeTargetMachine(tm);

        c.LLVMSetTarget(self.module, triple);

        var err_msg2: [*c]u8 = null;
        if (c.LLVMTargetMachineEmitToFile(tm, self.module, output_path, file_type, &err_msg2) != 0) {
            defer c.LLVMDisposeMessage(err_msg2);
            const msg = std.mem.span(err_msg2);
            return self.emitErrorFmt("failed to emit file: {s}", .{msg});
        }
    }

    pub fn emitObject(self: *CodeGen, output_path: [*:0]const u8) !void {
        return self.emitToFile(output_path, c.LLVMObjectFile);
    }

    pub fn emitAssembly(self: *CodeGen, output_path: [*:0]const u8) !void {
        return self.emitToFile(output_path, c.LLVMAssemblyFile);
    }

    pub fn link(allocator: std.mem.Allocator, io: std.Io, output_obj: []const u8, output_bin: []const u8, libraries: []const []const u8, target_config: TargetConfig) !void {
        var argv = std.ArrayList([]const u8).empty;

        if (target_config.isWindows()) {
            // Windows: MSVC-style linker flags
            const linker = target_config.linker orelse "link.exe";
            try argv.appendSlice(allocator, &.{ linker, output_obj });
            try argv.append(allocator, try std.fmt.allocPrint(allocator, "/OUT:{s}", .{output_bin}));

            for (target_config.lib_paths) |lp| {
                try argv.append(allocator, try std.fmt.allocPrint(allocator, "/LIBPATH:{s}", .{lp}));
            }
            for (libraries) |lib| {
                try argv.append(allocator, try std.fmt.allocPrint(allocator, "{s}.lib", .{lib}));
            }
        } else {
            // Unix: cc-style linker flags
            try argv.appendSlice(allocator, &.{ target_config.getLinker(), output_obj, "-o", output_bin });

            if (target_config.sysroot) |sr| {
                try argv.append(allocator, try std.fmt.allocPrint(allocator, "--sysroot={s}", .{sr}));
            }

            // User-supplied library paths first
            for (target_config.lib_paths) |lp| {
                try argv.append(allocator, try std.fmt.allocPrint(allocator, "-L{s}", .{lp}));
            }

            // Auto-detect host OS library paths when linking foreign libraries
            if (libraries.len > 0 and target_config.triple == null) {
                for (host_lib_paths) |path| {
                    try argv.append(allocator, try std.fmt.allocPrint(allocator, "-L{s}", .{path}));
                }
            }

            for (libraries) |lib| {
                try argv.append(allocator, try std.fmt.allocPrint(allocator, "-l{s}", .{lib}));
            }
        }

        const argv_slice = try argv.toOwnedSlice(allocator);
        var child = std.process.spawn(io, .{
            .argv = argv_slice,
        }) catch return error.LinkError;
        const result = child.wait(io) catch return error.LinkError;
        if (result != .exited) return error.LinkError;
        if (result.exited != 0) return error.LinkError;
    }

    /// Common library paths for the host OS, computed at comptime.
    const host_lib_paths = blk: {
        const builtin = @import("builtin");
        var paths: []const []const u8 = &.{};
        if (builtin.os.tag == .macos) {
            if (builtin.cpu.arch == .aarch64) {
                // Apple Silicon Homebrew
                paths = &.{ "/opt/homebrew/lib", "/usr/local/lib" };
            } else {
                // Intel Mac Homebrew
                paths = &.{"/usr/local/lib"};
            }
        } else if (builtin.os.tag == .linux) {
            paths = &.{ "/usr/local/lib", "/usr/lib" };
        }
        break :blk paths;
    };
};