Skip to content

Releases

GFOpenAmigaGCC 16.2.0-open9 for x86_64-linux-gnu, glibc 2.35, RC1

Choose a tag to compare

@github-actions github-actions released this 19 Nov 09:38
80e8b7d

GFOpenAmigaGCC 16.2.0-open9 for x86_64-linux-gnu, glibc 2.35

A cross compiler on x86-64 Linux for x86-64 Linux programs: daletop, the CI runners and the AMD x86 appliance. It is GCC 16.2.0 and binutils 2.47 as GNU released them, with one fix from the m68k toolchains and the open9 name, over a sysroot of glibc 2.35 and Linux 5.15's headers: Ubuntu 22.04's, the oldest system AmiCradle's runtimes start on.

What is in it

  • GCC 16.2.0 and binutils 2.47, the same tarballs and SHA-256 as the m68k-aros target and the kernel compilers (scripts/userland.lock), with patches/linux/series (patches/gcc/0019, fold-mem-offsets).
  • C and C++: libgcc, libgcc_s, libstdc++, libatomic and libgomp. Programs are PIE by default, as Ubuntu's and Debian's GCC make them.
  • A sysroot inside the folder (x86_64-gfopen-linux-gnu/sysroot): glibc 2.35 built from GNU's tarball by this GCC, Linux 5.15's headers, and zlib 1.3.1 (a program that links -lz loads the system's libz.so.1). A program links against it and asks for no glibc symbol above 2.35, so it starts on Ubuntu 22.04, Debian 12 and every newer system. The sysroot's glibc is only linked against: the program loads the system's own glibc.
  • The compiler says GFOpenAmigaGCC 16.2.0-open9. Nothing of AmigaOS.

GCC 16's libstdc++ is newer than Ubuntu 22.04's, so a C++ program that must start there links it in with -static-libstdc++. GMP, MPFR and MPC are linked in statically: the compiler needs only the host's C library (glibc 2.39 or later, Ubuntu 24.04's), and the folder can go anywhere.

Use

tar -xJf gfopenamigagcc-16.2.0-open9-x86_64-linux-gnu-glibc2.35-x86_64-linux.tar.xz
CC=$PWD/gfopenamigagcc-16.2.0-open9-x86_64-linux-gnu-glibc2.35-x86_64-linux/bin/x86_64-linux-gnu-gcc

In AmiCradle it is the linux-x86_64-glibc-gcc16 stove (amicradle-stoves linux-glibc-gcc16/install.sh x86_64 STOVES), which gfosmotherboards' ci/build_runtimes.py and Cradle's deploy_dev.py build the native runtimes with.

Checks

tests/userland-check.sh PREFIX x86_64 on the built toolchain: the identity of gcc and g++, no GMP, MPFR or MPC shared library in gcc, g++, cc1, cc1plus, as and ld, the sysroot inside the folder holds glibc 2.35, and tests/userland/hello.c (threads, dlopen, libm, TLS) and hello.cpp (iostream, std::thread, exceptions, std::stoll, -static-libstdc++) are PIE, ask for nothing above GLIBC_2.35 and print x86_64 c ok and x86_64 c++ ok. tests/userland-recipe-test.sh checks the recipe before the build.

GFOpenAmigaGCC 16.2.0-open9 for aarch64-linux-gnu, glibc 2.35, RC1

Choose a tag to compare

@github-actions github-actions released this 19 Nov 09:38
80e8b7d

GFOpenAmigaGCC 16.2.0-open9 for aarch64-linux-gnu, glibc 2.35

A cross compiler on x86-64 Linux for 64-bit ARM Linux programs: the CM4 and Pi 5 cards. It is GCC 16.2.0 and binutils 2.47 as GNU released them, with one fix from the m68k toolchains and the open9 name, over a sysroot of glibc 2.35 and Linux 5.15's headers: Ubuntu 22.04's, the oldest system AmiCradle's runtimes start on.

What is in it

  • GCC 16.2.0 and binutils 2.47, the same tarballs and SHA-256 as the m68k-aros target and the kernel compilers (scripts/userland.lock), with patches/linux/series (patches/gcc/0019, fold-mem-offsets).
  • C and C++: libgcc, libgcc_s, libstdc++, libatomic and libgomp. Programs are PIE by default, as Ubuntu's and Debian's GCC make them.
  • A sysroot inside the folder (aarch64-linux-gnu/sysroot): glibc 2.35 built from GNU's tarball by this GCC, Linux 5.15's headers, and zlib 1.3.1 (a program that links -lz loads the system's libz.so.1). A program links against it and asks for no glibc symbol above 2.35, so it starts on Ubuntu 22.04, Debian 12 and every newer system. The sysroot's glibc is only linked against: the program loads the system's own glibc.
  • The compiler says GFOpenAmigaGCC 16.2.0-open9. Nothing of AmigaOS.

GCC 16's libstdc++ is newer than Ubuntu 22.04's, so a C++ program that must start there links it in with -static-libstdc++. GMP, MPFR and MPC are linked in statically: the compiler needs only the host's C library (glibc 2.39 or later, Ubuntu 24.04's), and the folder can go anywhere.

Use

tar -xJf gfopenamigagcc-16.2.0-open9-aarch64-linux-gnu-glibc2.35-x86_64-linux.tar.xz
CC=$PWD/gfopenamigagcc-16.2.0-open9-aarch64-linux-gnu-glibc2.35-x86_64-linux/bin/aarch64-linux-gnu-gcc

In AmiCradle it is the linux-aarch64-glibc-gcc16 stove (amicradle-stoves linux-glibc-gcc16/install.sh aarch64 STOVES), which gfosmotherboards' ci/build_runtimes.py and Cradle's deploy_dev.py build the native runtimes with.

Checks

tests/userland-check.sh PREFIX aarch64 on the built toolchain: the identity of gcc and g++, no GMP, MPFR or MPC shared library in gcc, g++, cc1, cc1plus, as and ld, the sysroot inside the folder holds glibc 2.35, and tests/userland/hello.c (threads, dlopen, libm, TLS) and hello.cpp (iostream, std::thread, exceptions, std::stoll, -static-libstdc++) are PIE, ask for nothing above GLIBC_2.35 and print aarch64 c ok and aarch64 c++ ok under qemu-aarch64 with the sysroot's glibc. tests/userland-recipe-test.sh checks the recipe before the build.

GFOpenAmigaGCC 16.2.0-open9 for x86_64-linux-gnu RC1

Choose a tag to compare

GFOpenAmigaGCC 16.2.0-open9 for x86_64-linux-gnu

A cross compiler on x86-64 Linux for x86-64 Linux kernels: the AMD x86 appliance (Debian 13 amd64). It is built as x86_64-gfopen-linux-gnu, so GCC does not take it for a native compiler, and every program is also there as x86_64-linux-gnu-NAME. It is GCC 16.2.0 and binutils 2.47 as GNU released them, with one fix from the m68k toolchains and the open9 name.

What is in it

  • GCC 16.2.0 and binutils 2.47, the same tarballs and SHA-256 as the m68k-aros target (scripts/linux.lock).
  • patches/gcc/0019 (fold-mem-offsets, from open8): the pass no longer takes an instruction with memory on both sides as an access through its source's base only. It is in GCC's target-independent code, so it goes in here too (patches/linux/series).
  • The compiler says GFOpenAmigaGCC 16.2.0-open9.
  • Nothing of AmigaOS: no driver defaults, libnix, pthreads (aros-stuff 0011) or libstdc++ change.

It is a kernel compiler, like the kernel.org cross compilers: C only, and libgcc without a C library. There is no glibc and no C++, so it does not build Linux programs. GMP, MPFR and MPC are linked in statically: the programs need only the host's C library (glibc 2.39 or later, Ubuntu 24.04's), and the folder can go anywhere.

Use

tar -xJf gfopenamigagcc-16.2.0-open9-x86_64-linux-gnu-x86_64-linux.tar.xz
make ARCH=x86_64 CROSS_COMPILE=$PWD/gfopenamigagcc-16.2.0-open9-x86_64-linux-gnu-x86_64-linux/bin/x86_64-linux-gnu- ...

Checks

tests/linux-check.sh PREFIX x86_64 on the built toolchain: the identity, no GMP, MPFR or MPC shared library in gcc, cc1, as and ld, the x86-64 kernel's code flags (-mcmodel=kernel, no SSE or red zone, retpolines, -mharden-sls=all) compile to elf64-x86-64, control-register instructions assemble, -m32 (the 32-bit vDSO) and -m16 (real-mode code) compile, libgcc is there, and a freestanding program (tests/linux/hello-x86_64.c) prints x86_64 ok, run natively, at -O2 and -Os. tests/linux-recipe-test.sh checks the recipe before the build.

GFOpenAmigaGCC 16.2.0-open9 RC1

Pre-release

Choose a tag to compare

@SacredTrees SacredTrees released this 18 Nov 16:15

GFOpenAmigaGCC 16.2.0-open9

Open9 is open8 with two fixes, one in libpthread and one in libstdc++, for AmigaOS 3.x (m68k-amigaos) and AROS 68k (m68k-aros). Nothing else rides along: the compiler, binutils and libnix are open8's.

Changes from Open8

  • pthreads (aros-stuff patch 0011, #52): every pthread_cond_signal wakes a thread that is still waiting. It used to signal the first waiter on the condition's list and leave it there until that waiter had woken and locked the mutex again, so a second signal before the first waiter ran, or a signal to a timed wait that had just run out, woke nobody and the next waiter slept on with its work queued. Signal and broadcast now take each waiter they wake off the list, a woken waiter takes itself off before it locks the mutex, and a timed wait that ran out but was signalled as well returns 0. pthread_cond_destroy straight after a broadcast is safe. Thread pools, work queues and timed waits are the code it reaches.
  • tests/pthreads/condsignal.c (two signals, timed out, broadcast, destroy): fails three checks on open8's libpthread.a, passes all six on open9's (lab on #52). CI compiles it with -Wall -Werror.
  • libstdc++ (#51): std::stoll, std::stoull and std::stold for std::string. Its configure left _GLIBCXX11_USE_C99_STDLIB off for libnix, though libnix has strtoll, strtoull and strtold, so only the wstring ones were there and ports carried their own (shiromino's amiga_stdstring.h). scripts/libstdcxx-c99-stdlib.sh turns it (and the C++98 one) on in every installed c++config.h; tests/cxx/stoll.cpp (stoll, stoull, stold, stof, llabs, lldiv, atoll and the throws) is compiled and linked by CI, so a release without them fails. A port's own shim can go once it builds on open9.
  • The compiler says GFOpenAmigaGCC 16.2.0-open9.
  • The build workflow can publish a release candidate: the rc input makes the same build a pre-release tagged v<release>-rcN, with the release's own file names.

After open9

Programs link libpthread statically, so a program gets the fix only when it is relinked against open9 (or a stove refreshed with runtime/stove-pthreads). Code that does not link -lpthread builds exactly as on open8.

GFOpenAmigaGCC 16.2.0-open9 for powerpc-linux-gnu RC1

Choose a tag to compare

GFOpenAmigaGCC 16.2.0-open9 for powerpc-linux-gnu

A cross compiler on x86-64 Linux for 32-bit big-endian PowerPC Linux kernels, such as the AC460's 460EX (the kernel's 44x platform picks -mcpu=440 itself). It is GCC 16.2.0 and binutils 2.47 as GNU released them, with one fix from the m68k toolchains and the open9 name.

What is in it

  • GCC 16.2.0 and binutils 2.47, the same tarballs and SHA-256 as the m68k-aros target (scripts/linux.lock).
  • patches/gcc/0019 (fold-mem-offsets, from open8): the pass no longer takes an instruction with memory on both sides as an access through its source's base only. It is in GCC's target-independent code, so it goes in here too (patches/linux/series).
  • The compiler says GFOpenAmigaGCC 16.2.0-open9.
  • Nothing of AmigaOS: no driver defaults, libnix, pthreads (aros-stuff 0011) or libstdc++ change.

It is a kernel compiler, like the kernel.org cross compilers: C only, and libgcc without a C library. There is no glibc and no C++, so it does not build Linux programs. GMP, MPFR and MPC are linked in statically: the programs need only the host's C library (glibc 2.39 or later, Ubuntu 24.04's), and the folder can go anywhere.

Use

tar -xJf gfopenamigagcc-16.2.0-open9-powerpc-linux-gnu-x86_64-linux.tar.xz
make ARCH=powerpc CROSS_COMPILE=$PWD/gfopenamigagcc-16.2.0-open9-powerpc-linux-gnu-x86_64-linux/bin/powerpc-linux-gnu- ...

Checks

tests/linux-check.sh PREFIX powerpc on the built toolchain: the identity, no GMP, MPFR or MPC shared library in gcc, cc1, as and ld, the 44x kernel's code flags (-mcpu=440 -msoft-float -ffixed-r2) compile to elf32-powerpc, the 440/460 MSR, SPR and TLB instructions assemble, libgcc is there, and a freestanding program (tests/linux/hello-powerpc.c) prints powerpc ok under qemu-ppc at -O2 and -Os. tests/linux-recipe-test.sh checks the recipe before the build.

GFOpenAmigaGCC 16.2.0-open9 for aarch64-linux-gnu RC1

Choose a tag to compare

GFOpenAmigaGCC 16.2.0-open9 for aarch64-linux-gnu

A cross compiler on x86-64 Linux for 64-bit ARM Linux kernels: the CM4 and Pi 5 card kernels. It is GCC 16.2.0 and binutils 2.47 as GNU released them, with one fix from the m68k toolchains and the open9 name.

What is in it

  • GCC 16.2.0 and binutils 2.47, the same tarballs and SHA-256 as the m68k-aros target (scripts/aarch64.lock).
  • patches/gcc/0019 (fold-mem-offsets, from open8): the pass no longer takes an instruction with memory on both sides as an access through its source's base only. It is in GCC's target-independent code, so it goes in here too (patches/aarch64/series).
  • The compiler says GFOpenAmigaGCC 16.2.0-open9.
  • Nothing of AmigaOS: no driver defaults, libnix, pthreads (aros-stuff 0011) or libstdc++ change.

It is a kernel compiler, like the kernel.org cross compilers: C only, and libgcc without a C library. There is no glibc and no C++, so it does not build Linux programs. GMP, MPFR and MPC are linked in statically: the programs need only the host's C library (glibc 2.39 or later, Ubuntu 24.04's), and the folder can go anywhere.

Use

tar -xJf gfopenamigagcc-16.2.0-open9-aarch64-linux-gnu-x86_64-linux.tar.xz
make ARCH=arm64 CROSS_COMPILE=$PWD/gfopenamigagcc-16.2.0-open9-aarch64-linux-gnu-x86_64-linux/bin/aarch64-linux-gnu- ...

Checks

tests/aarch64-check.sh on the built toolchain: the identity, no GMP, MPFR or MPC shared library in gcc, cc1, as and ld, the arm64 kernel's code flags compile to elf64-littleaarch64, system-register and pointer-authentication instructions assemble, libgcc is there, and a freestanding program (tests/aarch64/hello.c) prints aarch64 ok under qemu-aarch64 at -O2 and -Os. tests/aarch64-recipe-test.sh checks the recipe before the build.

GFOpenAmigaGCC 16.2.0-open8

Choose a tag to compare

@SacredTrees SacredTrees released this 14 Nov 19:01
5d3bea4

GFOpenAmigaGCC 16.2.0-open8

Open8 is open7 with one wrong-code fix, for AmigaOS 3.x (m68k-amigaos). Nothing else rides along. Built from main 5d3bea4 (#50).

Changes from Open7

  • GCC 0019: the fold-mem-offsets pass no longer takes a memory-to-memory move (move.b (a1),2(a0)) as a load or store it may change. It used to fold an address add into the stores beside such a move and delete the add, leaving the move's destination 8 bytes early at -O2 with -m68040 or -m68060 (GFOpenBench wrote a prefs chunk id as WB\0C, gfopenbench #216). docs/fold-mem-offsets.md
  • tests/repro/fmo.c, run by tests/repro/prove.sh at -O2 and -O3 with -m68040 and -m68060, at -Os, and at -O2 with -fno-late-combine-instructions: wrong on open6, right on open8.
  • The compiler says GFOpenAmigaGCC 16.2.0-open8.

Proof

tests/repro/prove.sh on this build (the x86-64 toolchain below, under qemu-m68k), 14 Nov 2026 18:46 UTC: 48 of 48 checks right, none wrong. Every fmo line right: -O2/-O3 with -m68040 and -m68060, -Os -m68040, and -O2 -fno-late-combine-instructions with both CPUs.

After open8

Code built at -O2 for the 68040 or 68060 with open6 or open7 should be rebuilt. -fno-fold-mem-offsets is the workaround for a build that cannot move yet.

Files

  • gfopenamigagcc-16.2.0-open8-x86_64-linux.tar.xz: the toolchain for x86-64 Linux, without an NDK.
  • gfopenamigagcc-16.2.0-open8-src.tar.xz: amiga-gcc and its modules as built, with this repository.
  • SHA256SUMS.

GFOpenAmigaGCC 16.2.0-open7

Choose a tag to compare

@SacredTrees SacredTrees released this 07 Nov 23:28
0f88181

GFOpenAmigaGCC 16.2.0-open7

Open7 is the correctness, determinism and measured-performance follow-up to
Open6 for AmigaOS 3.x (m68k-amigaos) and AROS 68k (m68k-aros).

Changes from Open6

  • Runtime m68k cost overrides are accepted only through the explicit
    M68K_COSTS_FILE environment variable. Ambient files in the build or launch
    directory can no longer silently change generated code, and invalid or
    excessive values are rejected.
  • AROS release testing verifies the integer libgcc helpers in every advertised
    multilib and links the exact -m68040 C++ reproducer. The historical Open6
    note describing an unresolved __divsi3 could not be reproduced in either
    current Open6 or its archived predecessor.
  • The pinned fpuauto implementation dispatches through aligned data pointers
    instead of rewriting wrapper instructions, removing the instruction-cache
    and multicore race in the earlier prototype.
  • -fstack-protector, -fstack-protector-all,
    -fstack-protector-strong and -fstack-protector-explicit now link a target
    runtime. A distinct archive is installed for every AmigaOS multilib. A
    corrupted canary calls abort(), which prints libnix's Program aborted
    and exits with 127. The guard is mixed at startup from the time, a stack
    address and ExecBase; it stops ordinary overruns but is not a secret from
    a local attacker. m68k-aros has no stack-protector runtime: such links
    fail with __stack_chk_fail and __stack_chk_guard undefined.
  • Unsupported -fhardened stops in the driver with a clear recommendation
    instead of continuing into broken PIE/GOT assembly.
  • Opt-in UBSan checks can use trap mode without a sanitizer runtime. Modes
    requiring a runtime remain unsupported; the new cross-CPU execution matrix
    checks valid controls, deliberate UB and unsupported-mode diagnostics.
  • Installed provenance includes the complete patch set and pinned component
    commits used by the build.
  • A build takes an exclusive lock on its work folder instead of allowing two
    invocations to remove or replace each other's generated objects.
  • GFOpenNDK's libamiga.a builder uses exactly the CPU selected by each
    multilib and compiles its 1,303 objects with the requested job limit. It no
    longer sends a hardcoded -m68040 followed by a contradictory CPU option to
    GCC and GAS.
  • libnix4.library builds and installs again. Its source-list generator had
    misspelled the exclusion for the ordinary __vfwprintf_total_size object,
    so both that copy and the deliberate FULL_SPECIFIERS copy were exported.
    Open7 fixes the spelling and treats a failed or missing shared library as a
    build failure instead of silently continuing with only libnix4.a.
  • True 68040 AmigaOS targets have six dedicated target-library variants:
    plain, base-relative and base-relative-32, each with soft- and hard-float
    ABIs. Compatibility-range targets such as -m68020-40 continue to select
    68020 libraries. This removes the mixed-tuning C++ COMDAT warning and
    raises the AmigaOS multilib set from eleven to seventeen. Their libnix and
    libpthread folders hold the 68020 builds, with libnix's hard-float maths
    in the hard-float folders, so a -m68040 -m68881 link takes them, and the build checks
    that a hard-float sin returns its result in fp0.
  • Build checks drain compiler output instead of occasionally failing with
    SIGPIPE under pipefail. SSP Hunk archives use stable relative member names,
    and the Info index is rebuilt serially to preserve every installed manual.
  • fd2sfd's configure script is regenerated explicitly instead of selecting
    Autoconf versions through checkout timestamp order; configuration caches
    and backups are excluded from source archives.
  • AROS collect-aros honors TMPDIR for its intermediate linked image instead
    of unconditionally using /tmp, preventing large C++ links from failing on
    hosts with a separate or quota-limited temporary filesystem.

Qualification gates

Each gate is passed with its evidence below, or says plainly what it does
not cover.

  • Clean full AmigaOS and AROS builds with the Open7 identity: AmigaOS
    in progress: the fresh confirmation pair of this tree (two complete
    socket-disabled builds from fresh pinned sources) is running
    (offline receipt). AROS: the
    clean full-SDK offline build on 2026-11-06
    (AROS receipt).
  • Existing Open6 compiler, libnix, C++, pthread, base-relative and
    autoincrement regression suites
    : passed in every offline build (qemu
    reproducers with and without loop distribution, autoincrement, libnix
    end-pointer, lean C++, base-relative exception data), and the pthread and
    C++ runtime tests compile in CI.
  • Open7 determinism, hardening, AROS libgcc and fpuauto checks: passed
    (determinism,
    hardening,
    AROS libgcc,
    fpuauto).
  • Differential optimisation tests across supported CPUs and -O0, -O2,
    -O3, -Os and LTO
    : passed
    (matrix,
    fuzzing); each offline
    build repeats them.
  • Repeatable AC090 runtime and size benchmarks before changing any
    optimisation default
    : Open7 changes no optimisation default (the release
    default stays -O2), so nothing waits on this gate. The size matrix is in
    the optimiser record; an AC090 elapsed-time matrix of O2, Os, O3 and
    O2-LTO-whole builds is still needed before any future default change.
  • Two clean builds compared after normalising release timestamps: in
    progress: the fresh confirmation pair compares installed prefixes and
    binary and source archives
    (offline receipt).
    The earlier r3 attempt stopped on a harness fault (its console pipe closed),
    not a build difference; the harness was fixed.
  • Runtime qualification on AmigaOS, AROS and the AmigaChrome AC090
    instance, including a deliberate canary-corruption test
    : passed on
    Kickstart/AmigaOS 3.2.3, on the AC090 instance (Instance-32 copy) and on
    AROS 68k (runtime record). The
    canary corruption aborts with Program aborted, rc 127, on 68000 and 68040
    builds; UBSan trap mode stops at trap #7 (AmigaOS error #80000027). The
    runs found and fixed the soft-float maths in the 68040 hard-float folders.
    On AROS, stack protector and fpuauto are not available (see above), and
    the only AROS 68k lab is an FPU-less 68020, so AROS was run with
    -m68020 -msoft-float programs.
  • Warning-free C++ links for the -m68040 -m68881 AC090 profile: passed.
    The runtime C++ program (exceptions, std::function, threads) compiles and
    links with no output, and links byte-identically with
    -Wl,--fatal-warnings.

Release artefacts must come from a build of this tree. The matched r2
archives recorded on 2026-11-06 (0bf5a4d5...) and the
os32-gcc16-16.2.0-open7-candidate stove predate the 68040 maths fix and
must not be published.

GFOpenAmigaGCC 16.2.0-open6

Choose a tag to compare

@github-actions github-actions released this 04 Nov 01:04
a8526ba

GFOpenAmigaGCC 16.2.0-open6

GCC 16.2 for AmigaOS 3.x (m68k-amigaos) and AROS 68k (m68k-aros). Everything in open4 remains. Open6 is the first published release since open4 (open5 was prepared on 2 November 2026 and went into AmiCradle's os32-gcc16 stove, but was never published), so these notes list open5's changes and open6's.

Open6 is one compiler for both 68k targets. The m68k-aros target joins: stock GCC 16.2.0 with AROS's own diffs and the m68k fixes this repository carries, so AROS code and GFOpenKick's exec (an AROS module) get the late-combine fix (PR123853, open4's patches/gcc/0013) that AmiCradle's aros68k-gcc16 stove lacked. The m68k-amigaos compiler is open5's but for its identity.

New since open4

Item open4 open6
Name and identity OpenAmigaGCC 16.2.0-open4; openamigagcc-* tarballs GFOpenAmigaGCC 16.2.0-open6 for both targets (patches/amiga-gcc/0002-open6-pkgversion.patch, scripts/build-aros.sh); gfopenamigagcc-* tarballs, the source tarball holding this repository as gfopengcc/ (open5's renaming)
m68k-aros (AROS 68k) not here: AmiCradle's aros68k-gcc16 stove, stock GCC 16.2.0 and AROS's diff only, built by AROS's crosstools build scripts/fetch-aros.sh and scripts/build-aros.sh: stock GCC 16.2.0, binutils 2.47 and isl 0.27 (scripts/aros.lock), AROS's diffs as AROS has them, and patches/gcc 0003, 0005, 0013 and 0014 as they are and 0008 ported (patches/aros/series, docs)
exec's oe_AlertText (exec_alert.c before amicradle-guest #582) at -m68040 -O2, m68k-aros move.b (a1)+,(0,a1,d0.l): each byte one place late lea into a register of its own, then move.b (An)+,(Am)
libamiga.a (GFOpenNDK) in the 68000 multilibs built for the 68040 and copied into every multilib folder: a -m68000 program calling RangeRand, ArgArrayInit or another helper got 68020 instructions built once per multilib with its own flags; build.sh checks the 68000 folders for 68020 code (docs) (open5)
A program whose thread crashed (a CPU exception leaves its task in Wait(0)) never returned to the Shell: exit joined every thread and waited for ever exit leaves a crashed thread behind and returns: patches/aros-stuff/0007 (open5)
pthread_cond_timedwait with an absolute time already past waited for ever times out at once: patches/aros-stuff/0007 (open5)
__thread/thread_local and pthread keys in a task pthreads did not make (an SDL thread from CreateNewProc/NP_Entry, a Workbench helper) pthread_self() gave an id with no record and pthread_getspecific() read NULL->tlsvalues (Kraptor's SDL thread wrote to $02002F3C on GFOpenKick) the task gets a record and TLS block of its own on its first pthreads call, reclaimed (key destructors run) once it has ended; with the table full it is the main thread; never a NULL record: patches/aros-stuff/0008, 0009 (tests/pthreads/foreigntls.c) (open5)
A check that read a whole list with grep -q under pipefail (build.sh's GFOpenNDK libamiga.a and trap #7 checks, package.sh's) could fail, or pass, at random: ar/objdump killed by SIGPIPE once grep had its match grep reads the whole list
Checks tests/autoinc/check.sh scans both targets' code for PR123853 with amicradle-guest's scan_autoinc_overlap.py; tests/aros-series.sh keeps every GCC patch decided for m68k-aros; tests/release-names.sh keeps the names in step

Open1 to open4 stay as they were published: their tarballs keep the openamigagcc- names and their compilers say OpenAmigaGCC. Tools that find a toolchain by its tarball name or its --version line should accept both.

The m68k-amigaos target libraries open6 builds are open5's: on the Cradle machine every one compared equal to the open5 os32-gcc16 stove's except for debug information and assert file names, which name the build folder; the open ROM built with open6 for both targets is byte for byte the one built with open5 and aros68k-gcc16.

The m68k-aros compiler is not in the release tarball: it needs an AROS SDK (from an AROS build) as its sysroot, so it is built on the machine that has one (AmiCradle's aros68k-gcc16 stove from open6). Its complete source and recipe are in the source tarball.

Checked with

The release is built and published by this repository's build workflow (run by hand with release: 16.2.0-open6). The tarballs are published only if every check in that same run passes, the aros job's too:

  • tests/release-names.sh 16.2.0-open6: the identity patch, build.sh's and build-aros.sh's identity, the tarball names and these notes agree, before anything is built.
  • A complete build from scripts/sources.lock with every patch, and build.sh's own checks: the open6 identity, no libm881 for soft-float -m68040, no trap #7 for a read of address 0, C and C++23 programs link with libnix, every multilib links its own GFOpenNDK libamiga.a, and the 68000 ones have no 68020 code.
  • tests/repro/prove.sh, plain and with -fno-tree-loop-distribute-patterns, under qemu-m68k, and tests/autoinc/check.sh --amigaos.
  • tests/libnix/endptr/run.sh, tests/cxx/lean-check.sh and tests/baserel/eh-check.sh.
  • A compile of the tests that need a real AmigaOS (tests/cxx, tests/pthreads, tests/libnix); running them stays a lab job.
  • The aros job: tests/aros-series.sh, m68k-aros-gcc built without an AROS SDK, and tests/autoinc/check.sh --aros: no move addresses its destination through the register its source steps, for every CPU at -O2, -O3 and -Os.
  • scripts/package.sh: no NDK file in either tarball, and every libamiga.a is GFOpenNDK's.

Files

File
gfopenamigagcc-16.2.0-open6-x86_64-linux.tar.xz the m68k-amigaos toolchain for x86-64 Linux
gfopenamigagcc-16.2.0-open6-src.tar.xz the complete source, patches and build scripts (m68k-aros's too)
SHA256SUMS checksums

The NDK, Roadshow headers and optional developer kits are not included. See the install guide.

Known

  • The toolchain needs glibc 2.38 or later.
  • The shared libnix4.library still does not link; normal static programs do not need it.
  • For ARM64 Linux, use 16.2.0-open1's toolchain until an ARM64 build of a later release is made.
  • A work folder that built open4 or open5 already has that release's identity patch in amiga-gcc's Makefile, so open6's does not apply there: build from a fresh scripts/fetch.sh work folder (CI always does).
  • m68k-aros: a C++ program linked with -m68040 is left with __divsi3 undefined (as with the aros68k-gcc16 stove); link C++ at the default CPU, with -Wl,-u,__pthread_Init_Func -lpthread -latomic as the stove's tests do.

OpenAmigaGCC 16.2.0-open4

Choose a tag to compare

@github-actions github-actions released this 30 Oct 18:05

OpenAmigaGCC 16.2.0-open4

GCC 16.2 for AmigaOS 3.x (m68k-amigaos). Everything in open3 remains. Open4 adds the wrong-code and libnix fixes merged since open3.

New since open3

Item open3 open4
move.b (a2)+,(a0,a2.l): a move whose source steps a register addressed its destination through it (GCC 15's late-combine, PR123853) each element landed one place late at -O2 never generated: patches/gcc/0013 (docs)
strcat into a local array, inlined at -O1/-O2 for 68020 and up wrote to an absolute address (virtual-stack-vars left after vregs) addresses the array: patches/gcc/0014 (docs)
strtol/strtoul on empty or whitespace-only input left endptr unset endptr is the input: patches/libnix/0008 (docs)
Compiler identity OpenAmigaGCC 16.2.0-open3 OpenAmigaGCC 16.2.0-open4

Checked with

The release is built and published by this repository's build workflow (run by hand with release: 16.2.0-open4). The tarballs are published only if every check in that same run passes:

  • A complete build from scripts/sources.lock with every patch, and build.sh's own checks: the open4 identity, no libm881 for soft-float -m68040, no trap #7 for a read of address 0, C and C++23 programs link with libnix.
  • tests/repro/prove.sh, plain and with -fno-tree-loop-distribute-patterns, under qemu-m68k, including the 0013 and 0014 reproducers.
  • tests/libnix/endptr/run.sh: the built libnix's strtol/strtoul end pointers under qemu-m68k.
  • tests/cxx/lean-check.sh and tests/baserel/eh-check.sh.
  • A compile of the tests that need a real AmigaOS (tests/cxx, tests/pthreads, tests/libnix); running them stays a lab job.
  • scripts/package.sh: no NDK file in either tarball, and every libamiga.a is OpenNDK's.

Files

File
openamigagcc-16.2.0-open4-x86_64-linux.tar.xz the toolchain for x86-64 Linux
openamigagcc-16.2.0-open4-src.tar.xz the complete source, patches and build scripts
SHA256SUMS checksums

The NDK, Roadshow headers and optional developer kits are not included. See the install guide.

Known

  • The toolchain needs glibc 2.38 or later.
  • The shared libnix4.library still does not link; normal static programs do not need it.
  • For ARM64 Linux, use 16.2.0-open1's toolchain until an ARM64 build of open4 is made.