> They don't use the ELPM instruction rather, incorrectly,
> the LPM instruction.

The very problem is that AVR with > 64k flash is a segmented architecture and it's not easy to handle them correctly.

For example, ELPM Z+ might change RAMPZ so that this instruction is not of much use.

The root problem is the ABI and that it does not disallow crossing section boundaries, i.e. 64k boundaries.  No data must cross such boundaries and if ELPM Z+ shall be used to get efficient code (users like small/fast/low-ISR-overhead code) the last byte of each segment must be kept free.

Any other approach will lead to very inefficent code and/or strage artifacts.

The alternative is to take away RAMPZ from the compiler altogether and let the user set it appropriately, but the fact that crossing section boundaries does not work smooth won't go away with it, either.

The xxx_PF() functions modify the RAMPZ register, but on XMegas with EBI, RAMPZ influences also LD/LDD/ST/STD, making code after using xxx_PF() to access unintended data memory "pages".


In the attached file, in macros.inc, macro LPM_R0_ZPLUS_INIT is modified and macro LPM_R0_ZPLUS_FINI added, storing and restoring RAMPZ, for targets with _AVR_HAVE_RAMPD_ defined (which IMHO is unique indication of presence of EBI). Only LPM_R0_ZPLUS_FINI needs to be added to xxx_PF() functions (and documentation modified accordingly), as shown in the attached memcpy_PF(). All the xxx_PF() functions are affected and should be modified accordingly after consensus happens.


As a cheaper alternative, RAMPZ could be simply zeroed as all "standard" RAM access in current avr-gcc is restricted to 16-bit addresses; whether this is a better option I leave to decide for the more experienced developers on this forum.


This resolves https://savannah.nongnu.org/bugs/?func=detailitem&item_id=25120 which thus could be closed as duplicate (the issue with RAMPZ mods being non-interrupt safe is resolved by saving RAMPZ in the standard interrupt prologue/epilogue as of avr-gcc version 4.3+).


(file #20859)

This is a good point. I've considered this at the time of writing the functions but then I've discarded it considering that the code for pgm_read_xxxx_far() macros in avr/pgmspace.h has also the same problems and EEPROM routines are also affected. RAMPZ modification is documented but I agree that the side effects could be explained more clearly.

Making these functions reentrant secured imposes a small penalty (just to save and restore RAMPZ in any of the free call-used registers, 2 instructions, 2 machine cycles) for each function, probably small enough to avoid having two different versions, one reentrant and the other not. pgm_read_xxxx_far() macros probably could live in two flavors, reentrant and not reentrant. These macros will suffer more (relative to their small size) overhead costs adding RAMPZ preserving code.

This is all referred to interrupt handlers only. Preemptive task execution will require to save and restore RAMPZ as part of the task state for each context swap unless interrupts were disabled in the whole function (absurd)

Maybe it's time to debate the creation of a reentrant avr-libc version...

> 3. In future add support of "far" pointer in avr-gcc.

This sounds a bit optimistic but... who knows?

This code is to intended to access the whole flash memory in devices with flash address space bigger than 64KB.

The macro GET_FAR_ADDRESS() is the best version I've got after testing several ones but is not always a direct replacement of the & operator for getting far pointers. It has not been tested on ATmega2560, not even in simulation.

The string and memory copy functions for far program memory xxxx-PF() are still in beta stage and need more tests (not known bugs but not extensively tested). They had been parked in my hard disk for a long time but the big ROM devices users claim for some standard way to access the data in avr-libc.

Comments are welcomed.

Carlos Lamas.