NO_FILESYSTEM targets can't link since fdeb32e9: main_dict lost `const`, putting the 389 KB dictionary in .data

[lambdamikel]

Summary

Since commit fdeb32e ("big endian support", 2026-05-05), platforms/pico cannot link. main_dict lost its const qualifier, which moves the 397,720-byte embedded dictionary out of .rodata (flash) and into .data (SRAM). The RP2040 only has 264 KB of SRAM, so the link fails unconditionally:

arm-none-eabi/bin/ld: DECtalkMini.elf.elf section `.data' will not fit in region `RAM'
arm-none-eabi/bin/ld: address 0x2006a6e8 of DECtalkMini.elf.elf section `.heap' is not within region `RAM'
arm-none-eabi/bin/ld: section .stack1_dummy VMA [20040000,200407ff] overlaps section .data VMA [200000c0,20067707]
arm-none-eabi/bin/ld: region `RAM' overflowed by 173800 bytes
collect2: error: ld returned 1 exit status

This is still present on dectalk-develop HEAD (7f8f8b9). It affects any NO_FILESYSTEM target; I've only tested the Pico, but platforms/gba compiles the same path and has even less RAM, so it is presumably affected too.

Reproduction

export PICO_SDK_PATH=... PICO_EXTRAS_PATH=...
cd platforms/pico
cmake -S . -B build -DPICO_BOARD=pico -DCMAKE_BUILD_TYPE=Release
cmake --build build

Environment: pico-sdk 2.3.0, pico-extras, arm-none-eabi-gcc 13.2.1, PICO_BOARD=pico.

Root cause

fdeb32e made two related changes:

1. src/maindict.c:2 — const unsigned char main_dict[] → unsigned char main_dict[]
2. src/loaddict.c — the extern declaration lost const, and an in-place byte-swap loop was added (now init_dictionary() at src/loaddict.c:386):

b       = (S32*)&main_dict[8];
entries = get_long_int(main_dict);

for(i = 0; i < entries; i++) {
        b[i] = SWAP_32_LITTLE(b[i]);
}

The underlying goal is legitimate: the dictionary index is stored little-endian, so a big-endian host needs it swapped. But doing it in place forces the whole 389 KB blob to be writable.

The unfortunate part is that on little-endian targets this achieves nothing. include/port.h:95-103:

#ifndef __BIG_ENDIAN__
...
#define SWAP_32_LITTLE(x) (x)

So on every little-endian embedded target the loop rewrites each word with its own value — at a cost of 389 KB of RAM that the device does not have.

Worth noting src/epsonapi.c:61 still declares it the original way, so the tree is currently inconsistent about the qualifier:

extern const unsigned char main_dict[];

Suggested fix

Restore const on the definition and both extern declarations, and make the swap conditional so it never runs (and never needs a writable dictionary) on little-endian:

#ifdef __BIG_ENDIAN__
        /* index is stored little-endian; needs swapping on a big-endian host */
        ...
#endif

That restores the embedded targets. It does leave big-endian NO_FILESYSTEM builds needing a different approach, since they can't mutate a flash-resident blob either — byte-swapping on read, or building a small RAM-resident copy of just the index (entries * sizeof(S32) bytes, rather than the entire dictionary), would work for both.

I'm happy to open a PR if that direction seems right.

Why this went unnoticed

.github/workflows/build.yml builds Windows 32-bit, Windows 64-bit, AppImage, and Android. All four load the dictionary from a file, so none of them compile the NO_FILESYSTEM branch in loaddict.c — and 389 KB of .data is harmless on a desktop anyway. There is no CI coverage for the Pico or GBA ports, so a change to shared code broke them silently about three months ago. A link-only CI job for platforms/pico would catch this class of regression cheaply.

[NishiOwO]

You can open a PR to fix this if you want.

[NishiOwO]

One way to solve this would be introducing big endian dictionary, which I might do later.

[NishiOwO]

This should be fixed now