Re: AVR __progmem__ variable reading

[David Brown <david dot brown at hesbynett dot no>]

On 23/02/2019 19:38, Łukasz Kostka wrote:
> > Wiadomość napisana przez David Brown <david.brown@hesbynett.no> w dniu 23.02.2019, o godz. 16:34:
> >
> > On 22/02/2019 23:34, Łukasz Kostka wrote:
> > > Hi
> > > I am using for a while now gcc and especially __progmem__ attribute. I’d like to report a feature request for gcc to handle reading from flash memory variables. Compiler has all the knowledge (target device, availability of LPM, ELPM instructions etc.) in order to properly interpret variable is accessed (eg. by array subscription).
> > > This would remove need for any assembly code written to do this.
> > > Make user code much cleaner.
> > > GCC having all this knowledge can optimize end assembly code.
> > > Simple attribute addition will switch from array in memory to array in flash.
> > > Can serve as future implementations for other platforms.
> > > What do you think ?
> >
> > You don't need to write assembly to read flash data with AVR gcc - you have never needed it.  To use the "progmem" attribute, include the <avr/pgmspace.h> header and use the macros and functions from there, such as "pgm_read_byte".
> >
> > <https://www.nongnu.org/avr-libc/user-manual/pgmspace_8h.html>
> That is my point. I still need to use "ppm_read_byte"

Using pgm_read-byte is not writing assembly code.  I agree that it is an 
inconvenience, but it is far less of an inconvenience than writing 
assembly would be.

> > Newer versions of AVR gcc have named address spaces, making the process simpler.  I must admit I haven't written any AVR code since these were added to the compiler, but I assume they work fine:
> >
> > <https://gcc.gnu.org/onlinedocs/gcc/Named-Address-Spaces.html#Named-Address-Spaces>
> Thx. I didn’t read this before. This works only for C. I use C++ and there are no named address spaces extensions :-(

Are you sure?  I don't have a recent AVR compiler handy, but I'd be a 
little surprised if the named address spaces are C only.  (The gcc 
documentation about extensions is not always very clear about whether 
extensions apply to C and C++, or just C.)

With a little effort, I am confident that you could make C++ template 
classes that cover the effects of the address spaces and hide the need 
for pgm_read_byte (and friends) from user code.

> My end goal is to use all variables in exactly same way as in regular code without <arv/pgmspace.h>

I think that will be a bit optimistic.  The big killer here is pointers. 
 It would be reasonable for the compiler to know about direct access to 
variables, and that is in fact what you get with the named address 
spaces.  But if you pass around a "char *" pointer, there is no way for 
the compiler to know that it should point to flash memory rather than 
data memory.  You either have to manually add "pgm_read_byte" and 
friends when using the pointer, or annotate the pointer's type with an 
address space qualifier.

There is a "__memx" address space that can be used to make pointers and 
accesses generic, but that turns pointers into 24-bits and makes access 
significantly bigger and slower - you do not want that as the default.

This sort of thing has been an issue for all sorts of small 
microcontrollers, and all their compilers, since their inception.  It is 
not solvable in an ideal way that gives maximal convenience to 
programmers and still results in efficient code.  The only good solution 
is to move away from such cpu designs - there are very few reasons for 
choosing a core such as the AVR rather than an ARM, MIPS or RISC-V 
alternative.  (You might choose the AVR device for its peripherals, or 
pin package, or power usage - but not for its core.)