[davecheney]

Recent changes that let interp move startup work to compile time did
not show up in TestBinarySize, so their savings went unmeasured and a
regression would go unnoticed. These samples cover the RP2 targets,
ESP32-C3, slice growth and os. Their sizes match on Go 1.25 to 1.27.

Run against the commits before and after each change, these rows show
(code/rodata bytes, opt=z):

• machine: keep RP2 PLL calculations out of startup #5837 RP2 PLL: pico -1308/-548, pico2 -1128/-552 (opt=2 code -5 to
  -6 KB). Visible only with builder: match size tests to the CLI interpreter limit #5975.
• interp: handle llvm.ctlz #5963 llvm.ctlz: pico -818/-142, pico2 -754/-142, size-slice -163/-5.
• os: avoid startup work for unused temporary files #5973 os temp file seed: pico2 size-os -16 (opt=2 -32).

[jakebailey]

I sort of didn't include any of this stuff in the size tests on purpose, because then the results depend on the Go version used to compile the code, hence why my test was import-free.

Is this actually stable?

[davecheney]

I sort of didn't include any of this stuff in the size tests on purpose, because then the results depend on the Go version used to compile the code, hence why my test was import-free.

Is this actually stable?

good question, let me check

[davecheney]

@jakebailey Good catch. size-sha256 changed with every Go version, even between 1.27.1 and 1.27.3, so I removed it.
The remaining rows are the same on Go 1.25, 1.26 and 1.27. size-os imports os, but the code it measures is in TinyGo's src/os.

[davecheney]

@deadprogram @jakebailey PTAL. I've removed the sha256 test, it was unstable across all Go versions, the other size tests were tested across several Go versions and are stable so they shouldn't introduce flakes.

I'd like to land this soon to lock in the interp improvements that have landed in 0.44-dev

[jakebailey]

If it's stable, seems fine

[davecheney]

If it's stable, seems fine

🫡 i'll watch them like a hawk. Any funny business and they're out.