slice of allowzero pointer is rejected in for

[LemonBoy]

Reduced from clashos, this may be a regression ¯\_(ツ)_/¯

export fn _start() void {
    var a = [4]u32{ 1, 2, 3, 4 };
    var b = @ptrCast([*]allowzero u32, &a);

    for (b[0..4]) |ch| {
        if (ch > 4) @panic("");
    }
}
const builtin = @import("builtin");
pub fn panic(msg: []const u8, error_return_trace: ?*builtin.StackTrace) noreturn {
    while (true) {}
}

[mikdusan]

This looks to be a regression by #4152. We actually had a reduction that I later marked as off-topic to the PR, but I think it was just side-stepped. This comment is relevant: #4152 (comment)

[fengb]

Forcing a pointer capture works around the compiler error: for (b[0..4]) |*ch| {

The for loop capture seems to be trying to coerce into a non-allowzero pointer:

error: expected type '*const u32', found '*allowzero u32'

[LemonBoy]

The problem lies in how the compiler lowers the loop. The ch payload is desugared as a var ch: <elem-type> whose base backing memory (IrInstSrcDeclVar->ptr) is the i-th element of the array/slice/pointer it's looping on. The index i is increased every loop and so every time ch is dereferenced in the loop body the correct value is used.

Of course this trick falls short once you take into account that a pointer may be allowzero.