heavyweight dependencies:
devel/boost-python-libs
devel/libunwind
math/blas
math/cblas
math/lapack
math/metis
math/openblas
math/py-numpy
math/suitesparse
This is particularly painful on powerpc64, where openblas does not even
build (yet).
This patch turns on option KIG on x86 only.
The impact on arm is TBD.
PR: 238721
Approved by: tcberner
a symbol matches multiple clauses the last one takes precedence. If the
catch-all is last it captures everything. In the case of Qt5 libraries
this caused all symbols to have a Qt_5 label while some should have
Qt_5_PRIVATE_API. This only affects lld because GNU ld always gives the
catch-all lowest priority.
Older versions of Qt5Webengine exported some memory allocation symbols from
the bundled Chromium. Version 5.9 stopped exporting these [1] but the
symbols were kept as weak wrappers for the standard allocation functions to
maintain binary compatibility. [2][3] The problem is that the call to the
standard function in these weak wrappers is only resolved to the standard
function if there's a call to this standard function in other parts of
Qt5Webengine, because only then is there a non-weak symbol that takes
precedence over the weak one. If there's no such non-weak symbol the call
in the weak wrapper resolves to the weak wrapper itself creating an infinite
call loop that overflows the stack and causes a crash. Some of the
allocation functions are variants of C++ new and delete and it probably
depends on the compiler whether these variants are used in other parts of
Qt5Webengine.
Remove the weak wrappers (make them Linux specific). This isn't binary
compatible but we are already breaking that with the changes to the symbol
versions.
[1] 5c2cbfccf9
[2] 2ed5054e3a
[3] 009f5ebb4b
Bump all ports that depend on Qt5.
PR: 234070
Exp-run by: antoine
Approved by: kde (adridg)
