Regression/Miscompilation: Self-hosted x86_64 backend disagrees on `@bitSizeOf` value of arrays with the LLVM backend

[sagehane]

Zig Version

0.16.0-dev.699+529aa9f27

Steps to Reproduce and Observed Behavior

Reporting on behalf of a discussion in the #zig:tchncs.de Matrix group. The issue seems to be a regression from 0.14, presumably since https://ziglang.org/devlog/2025/#2025-06-08 due to the new self-hosted x86 backend. The regression can be observed on 0.15.1 too, and is likely x86_64-specific.

$ uname -a
Linux nixos-laptop 6.12.51 #1-NixOS SMP PREEMPT_DYNAMIC Mon Oct  6 09:17:53 UTC 2025 x86_64 GNU/Linux

Sample code:

const std = @import("std");

fn sizeOfArray(N: comptime_int, comptime T: type) comptime_int {
    return @bitSizeOf([N]T);
}

comptime {
    @compileLog(sizeOfArray(2, u2));
}

Run the above code with -fno-llvm and -fllvm:

$ zig test test.zig -fno-llvm
test.zig:8:5: error: found compile log statement
    @compileLog(sizeOfArray(2, u2));
    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Compile Log Output:
@as(comptime_int, 4)

$ zig test test.zig -fllvm
test.zig:8:5: error: found compile log statement
    @compileLog(sizeOfArray(2, u2));
    ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Compile Log Output:
@as(comptime_int, 10)

Expected Behavior

In Zig 0.14 and prior, -fno-llvm and -fllvm would agree on the @bitSizeOf value. The status quo of LLVM seems to be something like:

const std = @import("std");

fn foo(comptime T: type, N: comptime_int) !void {
    const Array = [N]T;
    const len = @sizeOf(Array) - @alignOf(Array);
    try std.testing.expectEqual(8 * len + @bitSizeOf(T), @bitSizeOf(Array));
}

test "-fllvm" {
    try foo(u2, 2);
    try foo(u3, 2);
    try foo(u3, 3);
    try foo(u9, 17);
    try foo(u64, 17);
}

Quoting #19755:

[N]T, for runtime-allowed T: N * @bitSizeOf(T) bits

It seems like the current behaviour of the new self-hosted backend might be the desired behaviour, so it's debatable which backend is in need of fixing.

[Rexicon226]

This is an intentional change; #24064, #19755

[sagehane]

Oh, okay, then I guess this can be closed and the proposal can be the tracking issue to the implementation progress.