GCC 16.2 for the Amiga, fixed properly.

Code the compiler got wrong

Bugs in GCC's 68k back end and middle end: the program compiled, then did the wrong thing.

• Float to integer on a 68040. GCC could save the FPU control register into the very register a store was using as its address, so the number landed in the wrong place. How we found it
• memmove copying the wrong way. An overlapping block move of a known size copied forwards and overwrote what it was about to read. It broke shift-up loops, __builtin_memmove and std::copy_backward. How we found it
• 64-bit arguments read from the wrong word. A 64-bit value pushed from the stack for a library call read the word next to it, so some 64-bit divides returned 0 or -1. How we found it
• A calloc that calls itself. GCC turned a hand-written calloc (malloc then memset) into a call to calloc, which recursed until the stack ran out. How we found it
• Base-relative C++ exceptions. In -fbaserel and -resident programs the exception tables could not reach their type information, so catching by type failed to link.

Defaults that suit an Amiga

GCC assumes things about memory that are not true on a 68k Amiga. We changed the defaults so ordinary code is safe.

• Address 0 is real memory. At -O2 GCC replaced provable NULL reads with trap #7, which AmigaOS shows as Software Failure 80000027. Our driver turns that off by default. How we found it
• Loop distribution back on. With the memmove fix in place, shift and insert loops can safely become block moves again, which made them up to 3.8 times faster on the AC090. How we found it
• C++ destructors registered properly. -fuse-cxa-atexit is now the default, so thread_local destructors run and static objects are destroyed in the right order. How we found it
• -resident32 links the right libraries. It used to pick the plain libraries instead of the base-relative ones. How we found it

Starting and stopping a program

Fixes in libnix, the small C library every program links with.

• Asking for a bigger stack. __stack did not give a program its stack, and the stack swap clobbered a register (Software Failure 80000004). How we found it
• Returning to the Shell. Plain -fbaserel32 programs returned their stack size instead of their exit code. How we found it
• 32-bit maths on a 68000. A program's own UtilityBase, still NULL, made the first multiply or divide jump to address 0. How we found it
• Doubles rounded to floats. On an Amiga with an FPU, OS 3.2's maths libraries left the FPU in single precision, so software doubles lost their accuracy. How we found it
• C++ start-up in base-relative programs. Start-up read its constructor and exception lists from the wrong place, so these programs failed to link or hung before main.

Threads

C++ threads on AmigaOS 3.x, made dependable enough for the Open family's own code.

• A thread's record moved under it. Starting a new thread could move the records of running ones, so thread_local and call_once crashed or hung about half the time. How we found it
• Detach, join and signals. pthread_detach was a stub, a finished thread's slot was never given back, and pthread_join could swallow a Ctrl-F the program was waiting for. Join now has a signal of its own.
• Numbers in several threads at once. C++ streams formatting numbers in parallel hung the program, and sometimes the whole machine, because setlocale() was not safe to call from two threads. How we found it
• Threads in base-relative programs. A new thread started with no data pointer of its own and hung on its first global variable.

The AmigaOS headers

Two fixes to headers that GCC 16 reads differently from older compilers. The build fixes them for you.

• A library call inside another. In Stricmp(FilePart(name), ...) the outer call ran with the inner library's base in A6. The bug was in the tool that makes the NDK's inline headers. How we found it
• Network calls refused. Roadshow's socket headers listed argument registers as clobbered, which GCC 15 and later reject, so 31 network calls did not compile. How we found it

ARM64 Linux

An open2 build for ARM64 Linux, so the same compiler runs on a Raspberry Pi.

PowerPC and AmigaOS 4.1

A GCC 16.2 for AmigaOS 4.1 built the same way: from GNU's own sources, with open patches, and no borrowed toolchains.