Skip to content

scripts_macos_build_fixes.md

Latest commit

 

History

History
150 lines (117 loc) · 6.66 KB

scripts_macos_build_fixes.md

File metadata and controls

150 lines (117 loc) · 6.66 KB
Raw

Fixing rebuild_all_macos.sh on Apple Silicon

rebuild_all_macos.sh and rebuild_all.sh are functionally identical (the macOS variant only hardcodes CPU_COUNT instead of reading /proc/cpuinfo). The failure was never in these wrapper scripts — it was five separate issues in the bundled host toolchain (binutils 2.32 / GCC 6.5.0 / GMP 6.1.2, all vintage 2016–2019) that only surface when building on macOS, and specifically on Apple Silicon. Diagnosed and fixed by actually running the build on an M-series Mac (macOS 26, arm64) and iterating on each failure in turn.

1. Ancient bundled zlib collides with the macOS SDK's fdopen

Symptom: error: expected identifier or '(' deep in zlib/zutil.c while building host binutils and GCC.

Cause: binutils and GCC each vendor a fallback copy of zlib 1.1.4. Its zutil.h has:

#if defined(MACOS) || defined(TARGET_OS_MAC)
#      define fdopen(fd,mode) NULL /* No fdopen() */
#endif

TARGET_OS_MAC is an Apple SDK macro that's defined on every Apple platform, including modern macOS — but this code predates Mac OS X and meant classic (pre-OS X) Mac OS, where fdopen() genuinely didn't exist. On real macOS it fires anyway, #define-ing fdopen to NULL, which then mangles the real fdopen() prototype pulled in from <stdio.h>.

Fix: guard the branch with && !defined(__APPLE__) in both vendored copies (tools/crosstools/gnu/binutils-2.32-aros.diff, tools/crosstools/gnu/gcc-6.5.0-aros.diff).

2. Vendored config.guess predates Apple Silicon

Symptom: configure: error: Oops, mp_limb_t is 64 bits, but the assembler code in mpn/ won't work on this system-abi combination while configuring GMP.

Cause: binutils/GCC/GMP each carry their own copy of GNU's config.guess/config.sub, all from before Apple Silicon existed (Nov 2020). On an M-series Mac, uname -m reports arm64, which these old scripts don't recognize — they fall back to a bare arm (implying 32-bit), producing the triple arm-apple-darwinNN. GMP then detects a real 64-bit mp_limb_t at compile time, sees a 32-bit-implying host triple, and aborts.

Fix: stop relying on the vendored config.guess entirely for Darwin. tools/crosstools/gnu/mmakefile.src now passes an explicit, correct --build=$(AROS_HOST_CPU)-apple-darwin --host=$(AROS_HOST_CPU)-apple-darwin to every host-compiler sub-package (gmp, isl, mpfr, mpc, binutils, gcc) when $(AROS_HOST_ARCH) is darwin. This bypasses the broken auto-detection.

3. GMP 6.1.2 has no Apple Silicon assembly

Symptom: error: unknown AArch64 fixup kind! / ADR/ADRP relocations must be GOT relative compiling invert_limb.lo, once (2) let GMP correctly detect an aarch64 host.

Cause: GMP 6.1.2's hand-written AArch64 assembly assumes a Linux/ELF assembler. It was written years before Apple Silicon existed and doesn't know about Apple's Mach-O/PIE relocation model, which Apple's assembler enforces.

Fix: pass --disable-assembly to GMP on Darwin (same mmakefile.src conditional as above), forcing its portable C implementation. Only affects build speed of the host compiler itself, not the resulting m68k target code.

4. GCC's own fibonacci_heap.h has a latent bug

Symptom: error: reference to non-static member function must be called in gcc/fibonacci_heap.h:481 while compiling bb-reorder.o.

Cause: genuine bug in upstream GCC 6.5.0. fibonacci_heap<K,V> has a member function V *min(), but union_of() uses it as if it were a data member:

if (heapb->min->compare (heapa->min) < 0)   // should be heapb->min()->compare(heapa->min())

GCC's own (lenient) C++ frontend tolerates this; Apple Clang — a much stricter, standards-conformant compiler — correctly rejects it. This is why Linux (host-compiled by GCC) builds fine and macOS (host-compiled by Clang) doesn't.

Fix: call min() properly: heapb->min ()->compare (heapa->min ()) < 0 (patched into gcc-6.5.0-aros.diff).

5. GCC has no host-hooks object for Apple Silicon

Symptom: Undefined symbols for architecture arm64: "_host_hooks" linking cc1.

Cause: GCC's generic gcc/config/host-darwin.c only supplies PCH helper functions — the actual const struct host_hooks host_hooks = ...; definition is expected to come from a CPU-specific companion file (config/i386/host-i386-darwin.c for Intel, an rs6000 equivalent for PowerPC). gcc/config.host never gained an aarch64-*-darwin* / arm64-*-darwin* case, because that hardware didn't exist yet, so no companion object is built and host_hooks is left undefined.

Fix: added gcc/config/aarch64/host-aarch64-darwin.c and gcc/config/aarch64/x-darwin (mirroring the i386 versions exactly), and a matching case in gcc/config.host.

6. Apple Silicon's 16KB page size breaks GCC's PCH buffer

Symptom: internal compiler error: in darwin_gt_pch_use_address, at config/host-darwin.c:51 — every invocation of the freshly-built xgcc crashes, including the trivial "can this compiler compile anything" autoconf probe for target libgcc.

Cause: host-darwin.c reserves a static 1GB buffer for precompiled headers, aligned to a hardcoded 4096 bytes:

static char pch_address_space[1024*1024*1024] __attribute__((aligned (4096)));

then asserts the buffer's address is a multiple of the runtime page size. Every Mac this code was ever tested on used 4KB pages — but Apple Silicon uses 16KB pages. 4096-byte alignment doesn't guarantee 16384-byte alignment, so the assertion can (and did) fail.

Fix: bump the alignment to 16384 in gcc-6.5.0-aros.diff. 16384 is also a multiple of 4096, so this is safe on Intel Macs too.

7. mkisofs isn't installed

Symptom: /bin/sh: mkisofs: command not found building distfiles/aros-amiga-m68k.iso.

Cause: not a code bug — dependencies_macos.sh already lists cdrtools (which provides mkisofs), it just hadn't been installed on this machine.

Fix: brew install cdrtools.

Result

With all of the above, rebuild_all_macos.sh's underlying build steps run to completion on Apple Silicon: the host GCC 6.50 cross-compiler builds, the full AROS/ApolloOS tree builds (make -j), make distfiles produces distfiles/aros-amiga-m68k.iso, and ApolloROM.V4 assembles correctly.

Files changed:

  • tools/crosstools/gnu/binutils-2.32-aros.diff — zlib fdopen guard fix.
  • tools/crosstools/gnu/gcc-6.5.0-aros.diff — zlib guard, fibonacci_heap.h fix, aarch64 host-hooks files + config.host case, PCH alignment fix.
  • tools/crosstools/gnu/mmakefile.src — explicit --build/--host triple and --disable-assembly for Darwin host-compiler sub-builds.

Not committed (local machine setup, not a repo bug):

  • brew install cdrtools for mkisofs.