I found it: It's the call to longjmp that doesn't reach setjmp again.
But how can "restart_after_exception" get corrupted?

I found that some (older) MinGW versions have a bug that setjmp puts garbage into ebp, but the workaround -fno-omit-frame-pointer doesn't help so it's probably another bug.
Can you make a write breakpoint in gdb on restart_after_exception and longjmp in data_abort? The contents of restart_after_exceptions have to be the same as after the initial call to setjmp.

Edit: Set up a Win7 VM and compiled it. With gdb it outputs the same emu.c:551, argc = 65535, argv=something weird, and it indeed segfaults at main + 2088: call sched_process_pending_events(). As call pushes eip onto the stack ($esp), but $esp is somehow 0x10, it segfaults. That also explains why argc and argv are weird, they're stored on the stack.

I know this is not clean but I think it is better than the "prefix" solution there's in your makefile because prefix is tipically used for other things in a makefile.

You did it almost like I'd have done it, I'd just have a common section for not linux-native Makefile lines and a section for both MinGW and windows native builds.
And I'm getting a message, that "_WIN32_WINNT" would be redefined in gui-windows.c, so you may want to surround it with #ifndef.

Also you could help me finding the bug that prevents me from installing ndles 3.6 on a mingw compiled nspire_emu.
Could you add
printf("Data abort: address=%08x status=%02x pc=%08x\n", mva, status, arm.reg[15]);at the top of data_abort in emu.c and launch ndless_installer?
For me it outputs Data abort: address=0101010d status=01 pc=102fd8e4 before "Segmentation fault".

As it seems there's more people interested in a Qt gui I will think about it but a lot of work must be done before that.

The Qt GUI for linux could then be used as-is for windows as well, so not more work than removing gui-windows.c

http://ndless.me

Just a quick update:
I'm currently busy on vacation and next week school starts again so I'm busy learning and other unnecessary duties.
Redstone is fully implemented, but some blocks are still missing, so it's not very useful yet.
However, I worked a bit on my ndless fork and got exceptions to work! It's a rather nice feature, to minimize possible crashes.
Also, it'll be using a new executable format which will load faster and maybe also make the game faster!
It'll take a while to implement it in both ndless 3.6 and 3.1 though, I don't want that it runs on 3.6 only.

Hmm, lua extensions do definitely work, as clearly visible in the luaext.c file.
It's used for ndless 3.6 uninstallation, and that seems to work.

Question: In the menu, how are we supposed to save? When choosing save I tried the Menu, dot, 5, esc and Enter keys and whichever worked doesn't appear to save properly. When I restart the game and  try to reload the map, it does nothing so I lose all my progress.

Just press "menu" then either the touchpad or clickpad "click" or the 5 key. It should freeze for a short time and then close the menu.
Or just press Esc to "Save & Exit" in-game.
If it doesn't work, could you post your current crafti.map.tns if you have one?

Bad news: With newer MinGW versions it compiles and runs sucessfully, but the 3.6 ndless_installer.tns segfaults the emulator.
Any ideas?

How exactly would you use std::vector in C?
Also, push_back has no return value, an exception (std::bad_alloc, AFAIK) is thrown on malloc failure.