RP6502 build problem with RIA and xram overflow.

[markrvmurray]

Hi *,

I'm having a build problem on macOS Sonoma with the arm-none-eabi GCC from homebrew:

% mkdir build && cd build && cmake ..
% make -j4
[  0%] Performing build step for 'ELF2UF2Build'
:
[ 78%] Built target rp6502_vga
/Applications/ArmGNUToolchain/13.2.Rel1/arm-none-eabi/bin/../lib/gcc/arm-none-eabi/13.2.1/../../../../arm-none-eabi/bin/ld: rp6502_ria.elf section `.bss' will not fit in region `RAM'
/Applications/ArmGNUToolchain/13.2.Rel1/arm-none-eabi/bin/../lib/gcc/arm-none-eabi/13.2.1/../../../../arm-none-eabi/bin/ld: region RAM overflowed
/Applications/ArmGNUToolchain/13.2.Rel1/arm-none-eabi/bin/../lib/gcc/arm-none-eabi/13.2.1/../../../../arm-none-eabi/bin/ld: section .stack1_dummy VMA [20040000,200407ff] overlaps section .bss VMA [2002f718,200459b7]
/Applications/ArmGNUToolchain/13.2.Rel1/arm-none-eabi/bin/../lib/gcc/arm-none-eabi/13.2.1/../../../../arm-none-eabi/bin/ld: region `RAM' overflowed by 22968 bytes
collect2: error: ld returned 1 exit status
make[2]: *** [src/rp6502_ria.elf] Error 1
make[1]: *** [src/CMakeFiles/rp6502_ria.dir/all] Error 2
make: *** [all] Error 2

If I reduce the size of the xram array in src/ria/sys/mem.[ch] to e.g. 0x1000, then the above error goes away. No point in trying to run it, though!

Could some kind soul with a working build please make your rp6502_ria.elf.map file from the build available?

Thanks!

M

[rumbledethumps]

build.zip

[markrvmurray]

Thanks for the logs!

The problem is that with my later GCC-13 (you have GCC-11), the .text segment was too big. By starting the build with cmake -DCMAKE_BUILD_TYPE=MinSizeRel -GNinja .., I got a smaller .text segment which doesn't annoy the linker.

[markrvmurray]

Diffing the logs was a bit of a chore, but I bodged up sed horror that changed most of the gratuitous differences like paths, long numbers beginning with 0x00000000 etc to something that diff could ignore. After that, seeing that it was a simple increase in size of a lot of library parts. I didn't deep-dive into this, but I got the impression that the floating-point stuff was a lot bigger, particularly when it came to I/O and string conversions.

[markrvmurray]

I installed an Ubuntu 22.* VM for testing, and its GCC-12 built with no issues and no need to force minimum size. Seems GCC-13 is possibly doing overly aggressive optimisations such as loop unrolling.

[rumbledethumps]

`RAM' overflowed by 22968 bytes from inlining or unrolling? A few kb sure, but a bump in compiler versions would not have been my first guess.

[markrvmurray]

Looks like it could well be a combination of compiler+optimizations. I can do MinSizerel builds with no problem. This build optimises with -Os -DNDEBUG. Builds of Debug/-Og -g, Release/-O3 -DNDEBUG and RelWithDebugInfo/-O2 -g -DNDEBUG all fail with varying amounts of .text segment overflow, in the range of about 9k to 90k.

I'm attaching my diff, which is the difference between your and my Release rp6502_ria.elf.map logs, with a klugy sed script used to clean up gratuitous path and file name differences. Note that my build fails, so look for the string RAM, and disregard the obvious extra stuff from a few paragraphs below that onwards.

It's a big diff, so figuring out the problem is tedious, but I still reckon it's down to optimization, where none or too much both result in larger binaries overall.

yours_mine.patch

[markrvmurray]

The GCC optimisations are documented here: https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html. These docs specifically say that "-Os enables all -O2 optimizations except those that often increase code size:".