firmware-size-diff

---

name: firmware-size-diff description: Build the PolyKybd firmware twice — at HEAD (baseline) and with the current working-tree changes — then diff the resulting binaries. Reports (a) size deltas: total .uf2 bytes plus ELF section sizes (.text/.data/.bss, ROM/RAM); and (b) code-identity: whether the .text section is byte-for-byte identical, with a disassembly diff when it isn't. Use for both feature work ("did this bloat flash?") and refactors ("did this produce identical machine code?").

Firmware byte/size diff

Two questions this skill answers:

1. Size delta — for new features / optimisations: did flash/RAM usage move, and by how much?
2. Code identity — for refactors that should be behaviour-preserving: is the generated .text byte-for-byte identical to baseline? If not, where did it diverge?

Repo root: /home/thomaspollak/Repos/qmk_firmware Build output: .build/polykybd_split72_default.{elf,uf2} Toolchain on PATH: arm-none-eabi-size, arm-none-eabi-objcopy, arm-none-eabi-objdump, arm-none-eabi-nm.

When to expect which result

Change kind	Expected size delta	Expected .text identity
Pure refactor (rename, extract function, reorder)	~0 B	Identical — if not, investigate
Inlining / static change	small	Likely differs
New feature / new code path	>0 B	Differs
Compiler-flag change	varies	Differs everywhere

Tell the user up-front which mode applies based on the nature of the change (or ask if ambiguous).

Procedure

1. Sanity check in qmk_firmware/:

   • git status --porcelain — if no tracked changes, stop with "nothing to diff".
   • Note current branch + HEAD SHA so you can confirm restoration at the end.

2. Stash the working changes so HEAD is the baseline:

   git stash push -u -m "firmware-size-diff-after"
   

   If the stash is empty, abort.

3. Clean + build baseline. A clean build is required for trustworthy .text comparison — incremental builds can leave stale objects from the "after" state in .build/:

   make polykybd/split72:default:clean
   make polykybd/split72:default -j$(nproc)
   cp .build/polykybd_split72_default.elf /tmp/fw-baseline.elf
   cp .build/polykybd_split72_default.uf2 /tmp/fw-baseline.uf2
   arm-none-eabi-objcopy -O binary --only-section=.text /tmp/fw-baseline.elf /tmp/fw-baseline.text.bin
   

   On failure: git stash pop and report.

4. Restore working changes and clean-build candidate:

   git stash pop
   make polykybd/split72:default:clean
   make polykybd/split72:default -j$(nproc)
   cp .build/polykybd_split72_default.elf /tmp/fw-after.elf
   cp .build/polykybd_split72_default.uf2 /tmp/fw-after.uf2
   arm-none-eabi-objcopy -O binary --only-section=.text /tmp/fw-after.elf /tmp/fw-after.text.bin
   

5. Identity check (.text) — for refactors this is the headline result:

   sha256sum /tmp/fw-baseline.text.bin /tmp/fw-after.text.bin
   cmp /tmp/fw-baseline.text.bin /tmp/fw-after.text.bin && echo IDENTICAL || echo DIFFERS
   

   • If identical: report it plainly — the refactor produced bit-exact machine code. Still print the size table from step 6 for completeness (will be all zeros for .text).
   • If differs: in addition to the size table, produce a disassembly diff scoped to the changed region:

     arm-none-eabi-objdump -d --no-show-raw-insn /tmp/fw-baseline.elf > /tmp/fw-baseline.asm
     arm-none-eabi-objdump -d --no-show-raw-insn /tmp/fw-after.elf    > /tmp/fw-after.asm
     diff -u /tmp/fw-baseline.asm /tmp/fw-after.asm | head -200
     

     Also list the symbols whose size changed:

     arm-none-eabi-nm --size-sort --print-size /tmp/fw-baseline.elf > /tmp/syms-baseline.txt
     arm-none-eabi-nm --size-sort --print-size /tmp/fw-after.elf    > /tmp/syms-after.txt
     diff /tmp/syms-baseline.txt /tmp/syms-after.txt | head -80
     

     Summarise the top 5 functions by |Δ size|.

6. Size delta (always report):

   • arm-none-eabi-size -A /tmp/fw-baseline.elf /tmp/fw-after.elf → table with absolute and delta for .text, .data, .bss, plus computed ROM (.text + .data) and RAM (.data + .bss).
   • wc -c /tmp/fw-baseline.uf2 /tmp/fw-after.uf2 → uf2 byte count + delta + percentage.
   • cmp -l /tmp/fw-baseline.uf2 /tmp/fw-after.uf2 | wc -l → count of differing bytes in the uf2.
   • Highlight any section that moved by >1 KB.

7. Restore state: working tree was popped in step 4. Confirm git status --porcelain matches what it was before the skill ran, and confirm HEAD is unchanged.

Output format

Lead with the verdict appropriate to the change kind:

Refactor mode:

.text:  IDENTICAL  (sha256 6e2a…)  — refactor is bit-exact ✓
ROM:    ±0 B       238144 B
RAM:    ±0 B       54912  B
UF2:    ±0 B       482304 B

Or, if not identical:

.text:  DIFFERS  (baseline 6e2a…  after 91c4…)
        Top symbol deltas:
          fill_overlay_decompress  +24 B
          poly_suspend             -8  B
ROM:    +16 B   238144 → 238160 B
RAM:    ±0  B    54912 → 54912  B
UF2:    +32 B   482304 → 482336 B   (12 differing bytes)

Feature/size mode:

ROM:  +312 B (.text +312, .data ±0)   238144 → 238456 B
RAM:  ±0   B (.data ±0, .bss ±0)       54912 → 54912  B
UF2:  +624 B (0.13%)                  482304 → 482928 B
Differing bytes in .uf2: 1184

Also add a last line with a percentage number of how much RAM of the total 264kB is already used.

Notes / pitfalls

• Spurious .text differences in refactor mode: macros like __FILE__, __DATE__, __TIME__, __LINE__ and build timestamps embedded by QMK can make .text differ even when behaviour is unchanged. Common culprits:
  • QMK build date string in version.h — rebuild within the same minute or check whether the only differing region is the version string.
  • __FILE__ paths inside assert() / log calls — if a refactor moves code between files, expect string-table churn.
  • If the disassembly diff is only inside .rodata-referencing loads with no instruction changes, the bytes differ in .rodata, not the code. Confirm with arm-none-eabi-size -A showing .text unchanged but .rodata moved.
• Variants: default to split72. Diff split42 too if the user asks, or if git stash show --name-only stash@{0} (run before the stash pop) lists files under keyboards/polykybd/split42/ or the shared keyboards/polykybd/poly_keymap.c (the latter affects both variants). (split42 was formerly corne42.)
• Codegen outputs: lang/lang_lut.c and base/fonts/generated/*.h come from cog/create_fonts.sh. If they're touched, mention that size changes may reflect regenerated data rather than hand-written code.
• Clean builds are required for both passes — see step 3 rationale. Don't try to optimise this away.
• Don't modify the user's branch, push, or amend. The skill only uses git stash.