It curently uses some magic test to decide whether handcrafted code
works with or without -latomic. But it can claim that -latomic is not
needed for that case, while it is still needed for other cases.

> okay so append-atomic-flags does not work for me in this case
> noise-suppression-for-voice is doing `struct RnNoiseStats { uint32_t a, b, c, 
> d; }; std::atomic<RnNoiseStats> m_stats;`
> not just a single large integer

It is simplest to always add -latomic when an ebuild gets that deep
feeling that yeah, it would like some atomics please. The downsides to
listing a linker library are exactly:

- it might be unavailable
- it might be unneeded

And the former case is trivial to solve -- this function already does so
-- while the latter case has a sanctioned approach that is already used
for other intrinsic compiler libraries, but not for atomic "because the
build system would have a hard time if we had to build atomic early on"
which isn't a very good reason to break ebuilds which aren't building
sys-devel/gcc.

As a side benefit, we now handle -latomic such that a package which
requires it, but only for parts of the installed package, does not
overlink to libatomic in *all* binaries/libraries, even if the default
LDFLAGS are overridden and the global -Wl,--as-needed disappears.

Bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81358
Bug: https://bugs.gentoo.org/820101
Bug: https://bugs.gentoo.org/925672
Signed-off-by: Eli Schwartz <eschwart...@gmail.com>
---
 eclass/flag-o-matic.eclass | 80 +++++++++-----------------------------
 1 file changed, 19 insertions(+), 61 deletions(-)

diff --git a/eclass/flag-o-matic.eclass b/eclass/flag-o-matic.eclass
index 5ce7601fdde2..0e5271c7824f 100644
--- a/eclass/flag-o-matic.eclass
+++ b/eclass/flag-o-matic.eclass
@@ -1015,69 +1015,27 @@ test-compile() {
 }
 
 # @FUNCTION: append-atomic-flags
-# @USAGE: [bytes]
 # @DESCRIPTION:
-# Attempts to detect if appending -latomic is required to use
-# a specific-sized atomic intrinsic, and if so, appends it.  If the bytesize
-# is not specified, then check the four most common byte sizes (1, 2, 4, 8).
-# >=16-byte atomics are not included in this default set and must be explicitly
-# passed if required.  This may require you to add a macro definition like
-# -Duint128_t=__uint128_t to your CFLAGS.
+# Attempts to detect if appending -latomic works, and does so.
 append-atomic-flags() {
-       # this implementation is as described in bug #820101
-       local code
-
-       # first, ensure we can compile a trivial program
-       # this is because we can't distinguish if test-compile
-       # fails because -latomic is actually needed or if we have a
-       # broken toolchain (like due to bad FLAGS)
-       read -r -d '' code <<- EOF
-               int main(void)
-               {
-                       return 0;
-               }
-       EOF
-
-       # if toolchain is broken, just return silently.  it's better to
-       # let other pieces of the build fail later down the line than to
-       # make people think that something to do with atomic support is the
-       # cause of their problems.
-       test-compile "c+ld" "${code}" || return
-
-       local bytesizes
-       [[ "${#}" == "0" ]] && bytesizes=( "1" "2" "4" "8" ) || bytesizes="${@}"
-
-       for bytesize in ${bytesizes[@]}
-       do
-               # this sample program is informed by the great testing from the 
buildroot project:
-               # 
https://github.com/buildroot/buildroot/commit/6856e417da4f3aa77e2a814db2a89429af072f7d
-               read -r -d '' code <<- EOF
-                       #include <stdint.h>
-                       int main(void)
-                       {
-                               uint$((${bytesize} * 8))_t a = 0;
-                               __atomic_add_fetch(&a, 3, __ATOMIC_RELAXED);
-                               __atomic_compare_exchange_n(&a, &a, 2, 1, 
__ATOMIC_RELAXED, __ATOMIC_RELAXED);
-                               return 0;
-                       }
-               EOF
-
-               # do nothing if test program links fine
-               test-compile "c+ld" "${code}" && continue
-
-               # ensure that the toolchain supports -latomic
-               test-flags-CCLD "-latomic" &>/dev/null || die "-latomic is 
required but not supported by $(tc-getCC)"
-
-               append-libs "-latomic"
-
-               # verify that this did indeed fix the problem
-               test-compile "c+ld" "${code}" || \
-                       die "libatomic does not include an implementation of 
${bytesize}-byte atomics for this toolchain"
-
-               # if any of the required bytesizes require -latomic, no need to 
continue
-               # checking the others
-               return
-       done
+       # Make sure that the flag is actually valid. If it isn't, then maybe the
+       # library both doesn't exist and is redundant, or maybe the toolchain is
+       # broken, but let the build succeed or fail on its own.
+       test-flags-CCLD "-latomic" &>/dev/null || return
+
+       # We unconditionally append this flag. In the case that it's needed, the
+       # flag is, well, needed. In the case that it's not needed, it causes no
+       # harm, because we ensure that this specific library is definitely
+       # certainly linked with as-needed.
+       #
+       # Really, this should be implemented directly in the compiler, including
+       # the use of push/pop for as-needed. It's exactly what the gcc spec file
+       # does for e.g. -lgcc_s, but gcc is concerned about doing so due to 
build
+       # system internals and as a result all users have to deal with this mess
+       # instead.
+       #
+       # See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81358
+       append-libs "-Wl,--push-state,--as-needed,-latomic,--pop-state"
 }
 
 fi
-- 
2.43.2


Reply via email to