std: fix sentinel handling in Allocator interface

[ifreund]

Currently the only function that handles sentinel terminated slices
properly is free. All other uses of mem.sliceAsBytes() in the allocator
interface lack proper handling of a possible sentinel.

This commit changes the Allocator interface to use @ptrCast() plus
the new mem.absorbSentinel() instead.

This also makes incorrectly passing a pointer to array to
Allocator.free() a compile error. The proper function to free a pointer
to an array is Allocator.destroy().

Reported-by: David Vanderson david.vanderson@gmail.com
References: #19984
References: #22706
References: #23020

[ifreund]

Ah, it looks like zig1.wasm hasn't been updated with the new @ptrCast() features yet.

[Fri3dNstuff]

Should std.mem.absorbSentinel keep the mutability of pointers? A returned mutable slice allows accessing the position of, and overriding the sentinel - which is probably unwanted...
Though, the code below works, and prints { 1, 1, 1, 4 }, so maybe it's not that big a deal? (or even a positive)

pub fn main() !void {
    var a: [3:0]u8 = @splat(1);
    a[3] = 4;
    std.debug.print("{any}\n", .{a[0..4]});
}

[castholm]

Is forbidding passing pointer-to-arrays to Allocator.free() (#23020 (comment)) really the correct resolution for #23019?

Language design-wise, a *[n]T is treated as a specialized []T with a comptime-known length and the language goes great lengths to make using *[n]T in place of []T painless via coercion and synthesizing the len field, which helps avoid issues where bytes[0..end] may return a pointer-to-array or a slice depending on the comptime-knownness of end.

Allocator.free() defines the parameter as memory: anytype because there's no other viable way of expressing "any slice with any qualifiers" with the current type system. Had it been typed as memory: []T (disregarding mismatching qualifiers) you would have been able to pass *[n]T to it without issue.

[ifreund]

Should std.mem.absorbSentinel keep the mutability of pointers? A returned mutable slice allows accessing the position of, and overriding the sentinel - which is probably unwanted...

On the contrary, this is intentional behavior.

Is forbidding passing pointer-to-arrays to Allocator.free() (#23020 (comment)) really the correct resolution for #23019?

In my mind, the answer of whether Allocator.destroy() or Allocator.free() should be used for a given allocation is extremely straightforward. If the allocation is created using Allocator.create(), use Allocator.destroy(). If the allocation is created with Allocator.alloc(), use Allocator.free(). This is well documented in the doc comments.

A pointer to an array can only be directly allocated with create(), therefore it is incorrect for free() to support freeing such a pointer. If, for example, the user casts the slice returned by alloc() into a pointer to an array, I don't think it is unreasonable for them to need to cast it back to the exact slice returned by alloc() in order to use free().