Fix the uninitialized brainfuck tape (fib.bf going haywire after 89)

[g-w1]

What was wrong

hello.bf worked but fib.bf "slowly went haywire" after printing 89 — the blog post blamed "some weird miscompilation in the number-printing routine." It isn't a miscompilation; the emitted x86-64 is fine. The brainfuck tape was never zero-initialized — it was aliased on top of the ELF section-header table.

In main.zig:

• bss_o = shstrtab_o + shstrtab.len comes out exactly equal to sections_off (because data is empty), so the address handed to the program as cell 0 (r10 = base_point + bss_o) pointed straight at the section-header table in the file image.
• the single PT_LOAD had p_filesz == p_memsz, so there was no zero-fill region at all — the .bss section is SHT_NOBITS but the segment gave us nothing zeroed.

So cells 0–7 read as 0 only by luck (they land inside the all-zero NULL section header) and cell 8+ read leftover section-header bytes. hello.bf only ever touches cells 0–6, so it survived. fib.bf keeps one decimal digit per cell and walks rightward as the numbers grow — the instant 89 → 144 needs a third digit cell it reads a never-written cell, gets garbage, and the output falls apart.

(Bonus: read() was reading from fd 1 (stdout) instead of fd 0 (stdin) — fixed too.)

The fix

Put the tape past the end of the file image and bump p_memsz to cover it, so the loader zero-fills it for us. Minimal change, no modernizing — still builds with the same Zig 0.8.1 it was written against.

-    const bss_o = shstrtab_o + shstrtab.len;
     const sections_off = header_off + c.len;
     ...
     const filesize = sh_off;
+    const bss_o = filesize;                                  // past the file image
+    const memsize = filesize + @as(usize, opts.bss_len);
     ...
-        .p_memsz = cast(filesize),
+        .p_memsz = cast(memsize),

Proof, on a real x86-64 Linux box (new CI workflow)

.github/workflows/bf.yml installs Zig 0.8.1, runs zig build, compiles the sample programs to native ELF executables and checks their output:

• hello.bf → Hello World!
• fib.bf → 0 1 1 2 3 5 8 13 21 34 55 89 144 233 377 610 — past 89
• torture test: >×3000, write 'A', walk back, all reads correct (BAB) — the pre-fix compiler couldn't keep even one page of tape, let alone a zeroed one
• test2.bf → No

Known follow-ups (left out to keep this minimal)

• cells are 8 bytes apart (add r10, 8) but .bss is bss_len bytes (default 30000), so only ~3750 cells are usable; either narrow the stride or scale bss_len.
• -b is parsed into a ?u32 but never passed to genElfAndWriteToFs (ElfOpts.bss_len is also u16).

🤖 Generated with Claude Code