Stream: bugs

Topic: zig arena bug affects Roc

Steve Howell (Oct 26 2026 at 14:46):

See #show and tell > roc on a unikernel (venting about LLM policies) for a little more context, but the executive summary is that zig 0.17 has a double-allocation bug that can plausibly bloat Roc programs.

Here's a tiny Roc app:

app [make] { pf: platform "./platform/main.roc" }

make : U64 -> List(U8)
make = |n| {
        small = List.repeat(1, 100)
        big = List.repeat(2, n)
        big.concat(small)
}

The host's roc_alloc allocates from an ArenaAllocator, and I log what the arena asks of its own allocator. For make(1_000_000), Roc asks for 3,000,164 bytes in all:

stock Zig 0.17.0: arena node grows 1,500,120 -> 5,000,296; arena holds 5,000,272
patched 0.17.0: arena node grows 1,500,120 -> 3,500,152; arena holds 3,500,128

https://github.com/showell/patched-zig-compiler/commit/b2d40f4eff542f5a19c88107cb05b1b5872e57a4 is the fix. I'm not submitting upstream to zig for reasons that I mention on the other topic.

Steve Howell (Oct 26 2026 at 14:48):

platform ""
        requires {
                make : U64 -> List(U8)
        }
        exposes []
        packages {}
        provides { "roc_make": make_for_host }
        targets: {
                x64musl: {
                        inputs: [app],
                        output: Archive,
                },
        }

make_for_host : U64 -> List(U8)
make_for_host = make

notramo (Oct 26 2026 at 14:51):

Isn't it related somehow to hugepages?

notramo (Oct 26 2026 at 14:52):

Or maybe allocating a bit more upfront, to reduce fragmentation and speed up subsequent allocations?

Steve Howell (Oct 26 2026 at 15:03):

It's literally double allocating. Zig does allocate extra bytes optimistically, but the double allocation isn't intentional here. I saw the symptoms of it in a real-world program that's too big to post here.

Steve Howell (Oct 26 2026 at 15:14):

https://github.com/showell/patched-zig-compiler/commit/502e0e447597f0b3e73c354f816a8abb1083c19f has more details.

Steve Howell (Oct 26 2026 at 15:18):

(deleted message sent to wrong topic)

Last updated: Nov 04 2026 at 18:54 UTC