Skip to content

Incorrect memoization of comptime function calls #17761

@ianprime0509

Description

@ianprime0509

Zig Version

0.12.0-dev.1324+5257643d3

Steps to Reproduce and Observed Behavior

Found while debugging a failing test in #17634.

Given the following program:

const std = @import("std");

fn GenericIntApplier(
    comptime Context: type,
    comptime applyFn: fn (context: Context, arg: u32) void,
) type {
    return struct {
        context: Context,

        const Self = @This();

        fn apply(self: Self, arg: u32) void {
            self.any().apply(arg);
        }

        inline fn any(self: *const Self) IntApplier {
            return .{
                .context = @ptrCast(&self.context),
                .applyFn = typeErasedApplyFn,
            };
        }

        fn typeErasedApplyFn(context: *const anyopaque, arg: u32) void {
            const ptr: *const Context = @alignCast(@ptrCast(context));
            applyFn(ptr.*, arg);
        }
    };
}

const IntApplier = struct {
    context: *const anyopaque,
    applyFn: *const fn (context: *const anyopaque, arg: u32) void,

    fn apply(ia: IntApplier, arg: u32) void {
        ia.applyFn(ia.context, arg);
    }
};

const Accumulator = struct {
    value: u32,

    const Applier = GenericIntApplier(*u32, add);

    fn applier(a: *Accumulator) Applier {
        return .{ .context = &a.value };
    }

    fn add(context: *u32, arg: u32) void {
        context.* += arg;
    }
};

fn s() u32 {
    var a: Accumulator = .{ .value = 0 };
    const applier = a.applier();
    applier.apply(1);
    applier.apply(1);
    return a.value;
}

test {
    try comptime std.testing.expect(s() == 2);
}

The test fails. Inserting a @compileLog before returning a.value reveals that the actual value is 1 rather than the expected value of 2.

Expected Behavior

The test should pass.

It seems that the calls to apply (or some part of them) are being incorrectly memoized: as long as applier.apply is called with a different argument each time, the code works as expected, but any subsequent calls with an argument already used will have no effect.

Through bisecting, I've found that this regressed in 405ba26.

This might be further reducible, as there are quite a few moving parts in this example. I'll update this issue later if I'm able to remove any of them.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugObserved behavior contradicts documented or intended behaviorfrontendTokenization, parsing, AstGen, Sema, and Liveness.

    Type

    No type

    Projects

    No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions