[kfox]

GeosBoot.h copies three PROGMEM arrays with a plain memcpy: the GEOS 128 boot ROML (245 bytes), its vectors (7) and the TR probe (12). GCC turns a memcpy of a whole const array it can see into a copy of that array in .rodata, which the Teensy linker script places in DTCM, so all 264 bytes sat in RAM1 in every main image whether or not GEOS was used. The three copies now go through GeosCopyFromFlash(), which hides the source pointer from the optimizer with an empty __asm__, so the bytes stay in flash only. (__asm__ rather than asm because the GEOS host harnesses compile this header as -std=c99.)

Verification

Main image, Docker linux/amd64, Teensyduino 1.61.0, main (ac24701) vs this branch:

	TR+ before	TR+ after	TR before	TR after
FLASH code	285884	285924	252940	253068
FLASH data	1455196	1455452	1442976	1443232
RAM1 variables	215044	215044	204932	204932
RAM1 code	203576	203576	190472	190472
RAM1 padding	25800	25800	6136	6136
RAM1 free for local variables	79868	79868	122748	122748
RAM2	unchanged		unchanged

The reported RAM1 figures don't move: .data keeps its size on both targets, because the freed bytes fall inside the padding before its next alignment step. That padding grows by 272 bytes on TR+ and 264 on TR, which is headroom before the next change costs a 4 KiB step of free-for-locals.

• Before, the three byte patterns are in .data on both targets and none is in .text.progmem. After, none is in .data, and nm shows Geos128BootRoml and Geos128BootVectors in flash.
• npm test: pass. Shortening the copy by one byte fails 5 GEOS boot tests, so the tests reach all three copies.
• TR+ on an NTSC C128, main image at this commit: GEOS 128 from the Utilities menu reaches the DeskTop with Special IO set to GeoRAM (2 of 2 launches) and with REU+USB Mouse/Disk and no REU preload file (3 of 3). Not run on the TR.

Reviewed at low (13 changed code lines), no findings.

[SensoriumEmbedded]

Thank you for chasing these bytes back into flash, nice find!

I reviewed this merged onto main (ac24701) and built both targets.

Memory: it matches your numbers. .data content drops 272 bytes on TR+ and 264 on TR, ITCM is unchanged, and flash grows by 256 bytes of progmem plus 40 (TR+) / 128 (TR) bytes of code.

Nothing to change from my side.

Hardware testing: Just to document. I'll hit these at some point, not a request or requirement! :)

Your GEOS 128 launches on the TR+ cover the boot ROM copies.

• A first launch from a fresh stock disk image, so the CONFIGURE probe gets patched
• One GEOS 128 launch on a plain TR

Thank you!