commit 55860e3e2b5ab753302ce9a17e7f4bb3421b8b5c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 3 19:13:54 2026 +0200

dev-util/hip: drop unreachable test plumbing in 10.0.0

Same unreachable-plumbing shape as the other ROCm 10.0 packages, except
this one also cost a download. RESTRICT="test" is unconditional, so
src_test can never run, yet SRC_URI fetched hip-tests.tar.gz under test?
and the Manifest carried its DIST line -- 1.7 MB nobody could reach.

Removed with it: TEST_S in both the live and release branches, the test?
unpack, the hip_test_wrapper helper, the test? BDEPEND on freeglut and
spirv-llvm-translator, and the src_prepare branch that patched the catch
test tree.

Why the phase is restricted -- many tests are broken, and the suite
exercises the installed runtime rather than the build tree -- now sits
in the comment above RESTRICT rather than reading as an unexplained
aside.

commit 552a3b4e1648ab06e8ea9d93cc83dadf6982da71
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 23:39:37 2026 +0200

dev-util/hip: drop 7.2.3

Superseded by 7.2.4, the last release of the 7.x series and the rollback
target for the ROCm 10.0 migration. Version retention keeps the last two
plus the last of each previous major series; 7.2.4 and 10.0.0 satisfy both,
so the third version is redundant.

commit 56c1e80b1fbed5242f56800a2b14f331b170d741
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 19:01:38 2026 +0200

dev-util/hip: assert the -Werror and asan-doc sed anchors

Two of the seven seds in src_prepare had no anchor assertion while the other
five did. Both are load-bearing:

s/ -Werror//g on hipamd/src/CMakeLists.txt -- -Werror is wrong for a
distribution build, where a warning from a compiler upstream did not test
becomes a hard failure.

/asan COMPONENT asan/d on hipamd/packaging/CMakeLists.txt -- this is the
silent one: the build still succeeds and a stray /usr/share/doc/$-asan
directory lands in the image.

Verified with `ebuild ... clean prepare`: src_prepare completes and neither
guard fires, so both anchors are present in therock-10.0.

commit b6d18d4bf864f4e2dc73278d9ffe335dd80115ce
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 14:50:22 2026 +0200

dev-util/hip: fix hipStreamCreate() segfaulting on AVX-512 hosts

hipStreamCreate() died with a general protection fault inside libamdhip64:

amd::roc::VirtualGPU::VirtualGPU(...) <-- SIGSEGV
amd::roc::Device::createVirtualDevice(amd::CommandQueue*)
amd::HostQueue::HostQueue(...)
hip::Stream::Stream(...)
hip::ihipStreamCreate(...)
hipStreamCreate

The faulting instruction is `vmovdqa64 %zmm0,-0x128(%rdi)`, a 64-byte-ALIGNED
AVX-512 store, to an address that is only 32-byte aligned -- which is why it is
a general protection fault and not a page fault.

VirtualGPU declares two alignas(64) members, so alignof(VirtualGPU) is 64. It
derives from amd::ReferenceCountedObject, which declares operator new(size_t).
A class-scope operator new HIDES the C++17 over-aligned overloads, so name
lookup stops there and `new VirtualGPU(...)` never reaches
::operator new(size_t, align_val_t): the object gets 16-byte-guaranteed memory
while the compiler emits stores that assume 64. Forward the aligned new/delete
from the shared base so over-aligned derived types get what they declare.

Invisible without AVX-512 -- the vectorizer then uses at most 32-byte stores,
which the returned alignment happened to satisfy -- so this only shows up on
znver5 and similar. It is also invisible to a build-check and to any test that
stays on the default stream: rocBLAS sgemm and the whole of rocFFT s runtime
kernel compilation pass without ever creating an explicit stream.

Verified: hipStreamCreate/Destroy return 0; rocfft_plan_create + execute now
succeed (FFT of a delta gives a flat unit spectrum, error 0.000000); rocBLAS
sgemm still correct (2I x 3I = 6I, error 0.000000).

commit e28b8836be66e47f5dfd4552ded310e385fc1626
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 12:36:45 2026 +0200

dev-util/hip: fix the __noinline__ clash against GCC 16 libstdc++

HIP's non-HCC (host-only) branch of host_defines.h does `#define __noinline__`
with an empty body. Any standard-library header spelling the GNU attribute as
[[__gnu__::__noinline__]] then expands the macro inside the attribute and fails
to parse. GCC 16 libstdc++ does exactly that in <format>, <stacktrace> and
<bits/shared_ptr_base.h>:

/usr/lib/gcc/.../g++-v16/format:4550:30: error: expected identifier
4550 | [[__gnu__::__noinline__]]

Caught building sci-libs/rocFFT-10.0.0, whose rtc_cache.cpp reaches <format>
through <chrono> after a local header has already pulled in HIP.

Reproduces on BOTH llvm-core/clang:23 and llvm-core/rocm-llvm with a two-line
translation unit, so this is a HIP/libstdc++ problem, not a compiler-fork one.
Upstream's own guard in the HCC branch, __has_feature(cuda_noinline_keyword),
does not help: that feature is absent in host-only mode. Redefining the macro
cannot help either, since the clash is with the NAME -- so the fix is to parse
the affected headers while it still means the real attribute.

Supersedes hip-7.0.2-fix-libcxx-noinline.patch, which covered only libc++
(upstream PR ROCm/rocm-systems#1875). The 7.2.x ebuilds keep the older one.

commit b7fd30dd598ce8ea8c9c27cebcf54562c04cf0c4
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 30 09:55:01 2026 +0200

dev-util/hip: drop the last libc++ patch and the stale test-branch seds

clr-fix-libcxx is fully obsolete, not half-obsolete. An earlier commit kept a
regenerated version of its <__clang_hip_math.h> half on the strength of the
patch's stated rationale -- that <__clang_cuda_complex_builtins.h> refers to
::max. Re-reading the 10.0 header shows it now uses __builtin_fmax/
__builtin_fmaxf instead, and including the two CUDA headers in the original
order compiles cleanly against clang 23 without any help. Patch deleted.

Three of the four substitutions in the USE=test branch matched nothing at 10.0:
"-Werror " is gone from the catch CMakeLists, cmake_policy(SET CMP0037 OLD) is
gone, and /opt/rocm/bin is gone from hipTestMain/hip_test_context.cc. Only
"-Wall -Wextra " survives, and it now asserts its anchor. These went unnoticed
because RESTRICT="test" means the branch does not normally run.

Also guards the @HIP_INSTALLS_HIPCC@ substitution, where a silent miss would
leave the placeholder unexpanded in the installed cmake config.

commit c1539e6774713d0adc26104046bb8f742bdb80db
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Aug 29 20:16:37 2026 +0200

dev-util/hip: add 10.0.0

Retarget all three assets (clr, hip, hip-tests) to therock-<major.minor>, move
LLVM_COMPAT to 23, and fix the 9999 branch's hardcoded SLOT="0/7.2".

Patches: 3 apply unchanged, 1 dropped, 1 split.
- fix-libcxx-ranges dropped: it pre-included <__ranges/join_view.h> ahead of
HIP's `#define __local` because libc++'s <ranges> uses __local as an
identifier. ROCm 10.0 removed that #define from the tree entirely
(device_library_decls.h is now 108 lines and defines it nowhere), so the
workaround guards nothing and its hunk targets a line past EOF.
- 6.3.0-clr-fix-libcxx replaced by a 10.0.0 rewrite keeping only its live half.
The same upstream removal killed its __local save/restore, but its
<__clang_hip_math.h> include is still needed. Splitting them keeps the
working half from being discarded with the dead one.

Install a /usr/lib/libamdhip64.so compat symlink. clang 23's HIP toolchain
appends the runtime to every link line as a HARDCODED
"<hip-path>/lib/libamdhip64.so" -- a literal "lib", assuming a non-multilib
ROCm layout -- so HIP links fail on Gentoo with "cannot find
/usr/lib/libamdhip64.so" even though CMake passes the correct lib64 path via
hip::amdhip64. Neither --rocm-path nor --hip-path changes the "lib" component,
and clang 22 did not add the library at all, so this is new in 23 and cannot be
worked around per-consumer. Guarded to no-op where libdir is already "lib".

The two seds anchored on COMMENT text in FindHIP.cmake still match at 10.0 and
now assert first; a silent miss there makes FindHIP fall back to path
heuristics instead of Gentoo's prefixes. The HIPCC_BIN_DIR sed is likewise
asserted -- a silent miss there is a file collision with dev-util/hipcc at
merge time rather than a build failure. hiprtc-config.cmake.in was dropped from
the cmake_minimum_required sed: it carries no such line at 10.0.

commit 7c384c8cac70105c06770e0ccb7af0fc295f2471
Author: Raukaan Cogbrother <cogbrother@raukaan.local>
Date: Sun May 31 03:29:35 2026 +0200

metadata: normalize tab → 2-space indent in 42 imported files

42 metadata.xml files inherited tab indentation from their origin
overlays (mostly the ROCm cluster, plus a handful of dev-python and
AMD-tooling imports). The overlay convention is 2-space; this brings
them in line with the other 498 files. Maintainer attributions
preserved verbatim.

commit 7b4ad321cdfc4cfce26bff126c0d49fd57a53c94
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri May 29 14:12:10 2026 +0200

dev-util/hip: add 7.2.4

rocm 7.2.4's clr/hip release-asset tarballs now carry their own top-level
directories; 7.2.3's unpacked flat, so src_unpack synthesised the wrapper dirs.
Unpack directly into WORKDIR to match the new monorepo layout.

commit d5828580a4aa18156ef11ecfad720cd1ea778bb4
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri May 8 22:05:57 2026 +0200

dev-util/hip: note inert NUMA cmake args

clr 7.2.3 dropped its find_package(NUMA), so the
CMAKE__FIND_PACKAGE_NUMA flags inherited from ::gentoo's
wiring are silently ignored at configure time. Comment notes the
condition so a future audit doesn't have to re-derive it; flags are
kept aligned with ::gentoo in case upstream restores detection.

commit 5ae35b51a5042dfedf3e9174d30c0f6081a7fa7e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri May 8 21:57:15 2026 +0200

dev-util/hip: restore numa USE flag

Lost when bumping from 7.1.0; metadata.xml still documents it. Mirrors
the upstream ::gentoo wiring: IUSE flag, conditional numactl dep,
CMAKE__FIND_PACKAGE_NUMA cmake args under the
amdgpu branch. Clears UnusedLocalUse pkgcheck warning.

commit 6ec9175a03f320f401a4dbdff9fd663ea1eddc8f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri May 8 20:15:32 2026 +0200

dev-util/hip: drop stale patches superseded by 7.2.3

commit 22830c5f18a1b919a5bf4462d1a6d28047d14b20
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed May 6 10:56:19 2026 +0200

dev-util/hip: 7.2.3 add src_unpack for flat upstream tarballs

The 7.2.3 clr.tar.gz, hip.tar.gz, and hip-tests.tar.gz from ROCm/rocm-systems all
unpack flat into ./ — extracting them directly into $ would collide. Add
src_unpack to extract each into its own clr/, hip/, hip-tests/ subdirectory,
restoring the layout the rest of the ebuild expects.

commit 4c68c249f91effde00aee82c87a4d07f9790ff2d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed May 6 01:54:25 2026 +0200

dev-util/hip: add 7.2.3

commit 8a7082f1ba65632398134de525d263330415ac96
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Dec 6 21:51:11 2025 +0100

*/*: drop rocm/hip as newer version is in main tree

Signed-off-by: Ivan S. Titov <iohann.s.titov@gmail.com>

commit cfa997ed467a25f0a98e30e69523013fb12b4136
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Oct 18 14:26:09 2025 +0200

added rocm and HIP v. 6.4.4 for now, as main tree lags