The usage of pragmas within the macOS SDK requires LLVM Clang 8. This is
the version as our prebuilt Clang, however the minimum is worth noting here
as they may diverge and/or expert users might expect they could use an
earlier version.
If you compile using Clang 7 you'll see output like:
```bash
In file included from kernel/qcore_mac_objc.mm:44:
In file included from /bitcoin/depends/SDKs/Xcode-11.3.1-11C505-extracted-SDK-with-libcxx-headers/System/Library/Frameworks/AppKit.framework/Headers/NSText.h:9:
In file included from /bitcoin/depends/SDKs/Xcode-11.3.1-11C505-extracted-SDK-with-libcxx-headers/System/Library/Frameworks/AppKit.framework/Headers/NSView.h:19:
In file included from /bitcoin/depends/SDKs/Xcode-11.3.1-11C505-extracted-SDK-with-libcxx-headers/System/Library/Frameworks/AppKit.framework/Headers/NSResponder.h:10:
/bitcoin/depends/SDKs/Xcode-11.3.1-11C505-extracted-SDK-with-libcxx-headers/System/Library/Frameworks/AppKit.framework/Headers/NSEvent.h:19:1: error:
expected 'push' or 'pop' after '#pragma clang attribute'
/bitcoin/depends/SDKs/Xcode-11.3.1-11C505-extracted-SDK-with-libcxx-headers/usr/include/os/availability.h:104:273: note: expanded from macro
'API_UNAVAILABLE_BEGIN'
...__API_UNAVAILABLE_BEGIN5, __API_UNAVAILABLE_BEGIN4, __API_UNAVAILABLE_BEGIN3, __API_UNAVAILABLE_BEGIN2, __API_UNAVAILABLE_BEGIN1, 0)(__VA_A...
^
fatal error: too many errors emitted, stopping now [-ferror-limit=]
20 errors generated.
```
5962522fbc depends: bump native_cctools for fixed lto with external clang (Cory Fields)
00d1ba7aaa depends: enable lto support for Apple's ld64 (Cory Fields)
Pull request description:
This didn't work for a few reasons (various toolchain compatibility issues) the last time I tested it, but after the last round of bumps it works with no apparent issues.
Note that this does not _enable_ LTO by default in any way, only hooks up the machinery for ```-flto``` to work correctly when specified.
Lines were split for an easier rebase after #17919 is merged.
ACKs for top commit:
fanquake:
ACK 5962522fbc. The relevant option upstream is [here](https://github.com/tpoechtrager/cctools-port/blob/master/cctools/m4/llvm.m4#L4).
Tree-SHA512: df2775e74e7bc847e6cef94cb8457d503d6c9e2fdea861e51386fa6ed5a7ba688241db3685561ae1a32f66724c1b3801727252025f00c04b90a3bdc8a4f6f93b
2a701a1c42 build: pass -fcommon when building genisoimage (fanquake)
Pull request description:
Starting with the 10.1 release, GCC [defaults to -fno-common](https://gcc.gnu.org/gcc-10/porting_to.html). This causes
linking issues when building genisoimage:
```bash
[ 98%] Building C object genisoimage/CMakeFiles/genisoimage.dir/checksum.o
[100%] Linking C executable genisoimage
/usr/bin/ld: CMakeFiles/genisoimage.dir/apple.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
/usr/bin/ld: CMakeFiles/genisoimage.dir/boot.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
/usr/bin/ld: CMakeFiles/genisoimage.dir/desktop.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
/usr/bin/ld: CMakeFiles/genisoimage.dir/dvd_file.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
```
Rather than patching genisoimage further, just pass -fcommon to preserve
the legacy GCC behaviour.
Noticed [while testing #19530](https://github.com/bitcoin/bitcoin/pull/19530#issuecomment-659802512).
ACKs for top commit:
laanwj:
ACK 2a701a1c42
hebasto:
ACK 2a701a1c42, tested on Fedora 32 (x86_64, GCC 10.1.1):
Tree-SHA512: 0e6dd1b4c7c3f5b6997616da27a7ea54f757c81677766ee6813c776356527787a32b959f834e699fee9798553b13bcb142469f442c5e7a04f810f2c06e3a9505
Starting with the 10.1 release, GCC defaults to -fno-common. This causes
linking issues when building genisoimage:
```bash
[ 98%] Building C object genisoimage/CMakeFiles/genisoimage.dir/checksum.o
[100%] Linking C executable genisoimage
/usr/bin/ld: CMakeFiles/genisoimage.dir/apple.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
/usr/bin/ld: CMakeFiles/genisoimage.dir/boot.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
/usr/bin/ld: CMakeFiles/genisoimage.dir/desktop.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
/usr/bin/ld: CMakeFiles/genisoimage.dir/dvd_file.o:(.bss+0x0): multiple definition of `outfile'; CMakeFiles/genisoimage.dir/genisoimage.o:(.bss+0x0): first defined here
```
Rather than patching genisoimage further, pass -fcommon to preserve the
legacy GCC behaviour.
6457361e90 qt: Fix QFileDialog for static builds (Hennadii Stepanov)
Pull request description:
This change partially reverts 248e22bbc0 (#16386) and makes `QFileDialog`s work again for static builds.
Fixes https://github.com/bitcoin-core/gui/issues/32.
ACKs for top commit:
fanquake:
ACK 6457361e90. Although it would be good to know exactly _why_ this fixes the issue. At this stage I also don't think this should be a blocker for 0.20.1.
theuni:
ACK 6457361e90
Tree-SHA512: 8ad27e0bcae6debd02f73b7c374743e37d4edd806922b103a2fe494cf2d9930fe9ef3107b5a6c61f3c466cf7462de2641171880398954e7f2c4f417f5bb820d7
Note that this does not _enable_ lto by default in any way, only hooks up the
machinery for -flto to work correctly.
enable-lto-support is explicitly used for pinned-clang because we know it
works. It is neither enabled nor disabled in the external clang case so that
it can be auto-detected.
For depends builds this was fixed by fbcfcf69, which deleted the conflicting
headers. When we no longer control the clang installation, we need to ensure
that the SDK's libc++ headers are used rather than the ones shipped with clang.
We can do that by turning off the default include path and hard-coding our own.
This hard-coded path is ok because we control (via SDK packaging) where these
headers end-up.
Side-note: Now that this path is hard-coded in depends, we can potentially
package the SDK differently, as the c++ folder can live wherever is most
convenient for us.
This should be caught by the differing clang --version outputs, but because we
haven't yet extracted our pinned clang, the system one is actually used for the
version check. That's not a problem because bumping our pinned clang will
cause a rebuild of everything anyway.
a8d39b8840 doc: explain why passing -mlinker-version is required (fanquake)
Pull request description:
I have been down a 🐇 hole. Closes#19359.
When Clang is compiled, [a check is run](https://github.com/llvm/llvm-project/blob/release/8.x/clang/CMakeLists.txt#L353) to define `HOST_LINK_VERSION` as the output of `$CMAKE_LINKER -v`. Note the this is the version of the linker being used to compile Clang itself.. and this check is only run when compiling Clang for macOS.
In the Clang driver, if `HOST_LINK_VERSION` has been defined, there is some additional runtime functionality. An `-mlinker-version` argument, with the value of `HOST_LINK_VERSION` [will be added to the linker arguments](89de0d8dfb/clang/lib/Driver/Driver.cpp (L382)), if `-mlinker-version` has not been passed in by the user.
This is a bit weird, as by default, you are setting `-mlinker-version` to the version of the linker that was used to build the Clang binary, not the linker which will be used when compiling. The commit which introduced the functionality, 628fcf4e3b, described it as a "hack", that should be replaced. However, that was 10 years ago, and the behaviour is still here.
In the Darwin driver, [a check is done](89de0d8dfb/clang/lib/Driver/ToolChains/Darwin.cpp (L208)) for the `-mlinker-version` argument. If there is no argument, the version will default to `0`. Given the above, this should never happen when using Clang for macOS. A series of comparisons are then performed, to check whether the linker version is modern enough to enable certain features, like [`-demangle`](89de0d8dfb/clang/lib/Driver/ToolChains/Darwin.cpp (L215)).
### What this means
#### macOS
A Clang compiled for macOS, i.e `clang+llvm-8.0.0-x86_64-apple-darwin`, will have `HOST_LINKER_VERSION` set to the version of the linker used to compile Clang itself.
At runtime, `-mlinker-version=HOST_LINKER_VERSION` will be added to the linker args, if `-mlinker-version` wasn't passed in. In the Darwin driver, additional arguments, like `-demangle`, will be added to the linker arguments, because `HOST_LINKER_VERSION` was likely some very modern version of `lld` or `ld64`.
#### Linux (cross compilation in depends)
A Clang compiled for Linux, i.e `clang+llvm-8.0.0-x86_64-linux-gnu-ubuntu-14.04`, which we now use for macOS builds in depends, will behave differently. As it's built for Linux, `HOST_LINKER_VERSION` was not defined at compile time, and there will be no default behaviour of appending `-mlinker-version=HOST_LINKER_VERSION` to the linker args. Thus, unless you pass in `-mlinker-version` yourself, when the version checks are done in the Darwin driver, no modern linker features will be enabled, as the version will have defaulted to `0`.
Therefore, it's important that we continue to pass `-mlinker-version="our LD64 version"` as part of our compilation flags, if we want to have "modern" linker features enabled for our macOS builds.
#### Summary
[Clang 8](https://releases.llvm.org/download.html#8.0.0). Building a macOS binary. Link line with path arguments trimmed.
| | default behaviour | `-mlinker-version=100` (`-demangle threshold`) | `-mlinker-version=530` |
| - | --------------- | --------------------- | ---------------------- |
| macOS Clang | `-demangle -lto_library ../libLTO.dylib -no_deduplicate -dynamic -arch x86_64 -macosx_version_min 10.15.0 -o a.out ../test-b8b9b3.o -lc++ -lSystem ../libclang_rt.osx.a` | `-demangle -dynamic -arch x86_64 -macosx_version_min 10.15.0 -o a.out ../test-a66966.o -lc++ -lSystem ../libclang_rt.osx.a` | same as default |
| Linux Clang | `-dynamic -arch x86_64 -macosx_version_min 10.12.0 -o a.out ../test-bfce57.o -lc++ -lSystem` | `-demangle -dynamic -arch x86_64 -macosx_version_min 10.12.0 -o a.out ../test-a846a3.o -lc++ -lSystem` | `-demangle -lto_library ../libLTO.dylib -no_deduplicate -dynamic -arch x86_64 -macosx_version_min 10.12.0 -o a.out ../test-de0280.o -lc++ -lSystem` |
Note: Most links here are pointing to the 8.x branch of LLVM/Clang, as we are using that version in depends.
Note: To add a little more confusion, you wont see `-mlinker-version X` in your compile flags, you'll see [`-target-linker-version X`](431daedee4/clang/lib/Driver/ToolChains/Clang.cpp (L4777)).
ACKs for top commit:
laanwj:
ACK a8d39b8840
Tree-SHA512: 92f93079a5e59a0d561e74336b5cb03e3bf5a34437f5850283b9128c7624494b8285ec16290b1fa8103fe87f8789a53ce44b17902b8c1db5fde24d74b76fb168
This enables of the use of AI_* definitions in the Windows headers,
specifically AI_ADDRCONFIG, which fixes an issue with libevent and
ipv6 on Windows.
It also aligns with what we define in configure when building Core.
libevent uses getaddrinfo when available, and falls back to gethostbyname
Windows has both, but gethostbyname only supports IPv4
libevent fails to detect Windows's getaddrinfo due to not including the right headers
This patches libevent's configure script to check it correctly
8a26848c46 build: Fix m4 escaping (Hennadii Stepanov)
9123ec15db build: Remove extra tokens warning (Hennadii Stepanov)
fded4f48c3 build: Remove duplicated QT_STATICPLUGIN define (Hennadii Stepanov)
05a93d5d96 build: Fix indentation in bitcoin_qt.m4 (Hennadii Stepanov)
ddbb419310 build: Use pkg-config in BITCOIN_QT_CONFIGURE for all hosts (Hennadii Stepanov)
492971de35 build: Fix mingw pkgconfig file and dependency naming (Hennadii Stepanov)
Pull request description:
This PR makes `bitcoin_qt.m4` to use `pkg-config` for all hosts and removes non-pkg-config paths from it. This is a step towards the idea which was clear [stated](https://github.com/bitcoin/bitcoin/pull/8314#issue-76644643) by Cory Fields:
> I believe the consensus is to treat Windows like the others and require pkg-config across the board. We can drop all of the non-pkg-config paths, and simply AC_REQUIRE(PKG_PROG_PKG_CONFIG)
There are two unsolved problems with this PR. If depends is built with `DEBUG=1` the `configure` script fails to pickup Qt:
- for macOS host (similar to, but not the same as #16391)
- for Windows host (regression)
The fix is ~on its way~ submitted in #18298 (as a followup).
Also this PR picks some small improvements from #17820.
ACKs for top commit:
theuni:
Code review ACK 8a26848c46
dongcarl:
Code Review ACK 8a26848c46
laanwj:
Code review ACK 8a26848c46
Tree-SHA512: 3b25990934b939121983df7707997b31d61063b1207d909f539d69494c7cb85212f353092956d09ecffebb9fef28b869914dd1216a596d102fcb9744bb5487f7
f0d7ed10b4 depends: Propagate only specific CLI variables to sub-makes (Carl Dong)
0a33803f1c depends: boost: Use clang toolset if clang in CXX (Carl Dong)
1ce74bcde3 depends: boost: Split target-os from toolset (Carl Dong)
2d4e480813 depends: boost: Specify toolset to bootstrap.sh (Carl Dong)
3d6603e340 depends: Propagate well-known vars into depends (Carl Dong)
Pull request description:
From: https://github.com/bitcoin/bitcoin/pull/18308#issuecomment-598301117
The following monstrosity is quite useful when invoked inside `depends`, and reviewers can use it to compare the behaviour of this change against master.
```bash
make print-{{,{host,{,{i686,x86_64,riscv64}_}linux}_}{CC,CXX},boost_{cc,cxx}}
```
It would also be helpful to make sure that setting `HOST`, `CC`, and `CXX` does the right thing. The 3 hosts I found offered good coverage were: `{x86_64,i686,riscv64}-linux-gnu`. As we special-case the `x86_64` and `i686` hosts in `depends/hosts/linux.mk`, and `riscv64` is a sanity check for a non-special-cased host.
ACKs for top commit:
hebasto:
ACK f0d7ed10b4, tested on Linux Mint 19.3 (x86_64):
practicalswift:
ACK f0d7ed10b4 -- patch looks correct
laanwj:
Code review and concept ACK f0d7ed10b4
ryanofsky:
Code review ACK f0d7ed10b4. Changes since last review: adding comment explaining check for predefined make variables, dropping freetype commit, adding commit whitelisting overrides for recursive makes
Tree-SHA512: b6b8e76f713c26a0add6cd685824e2f5639109236ee9f89338f7c79cb1b1f2c3897bfb62b80b023d6d1943b5a6eb282a2f827f1f499c5e556eca015d6635fa65
We want to supply well-known vars to ./configure scripts to do with as
they please. However, we do _not_ want to override these well-known vars
at make-time as certain build systems expect a self-mangled version of
these well-known vars.
For example, freetype and bdb will prepend `libtool --mode=compile' to
CC and CXX, which, if we override CC on the command line at make-time,
will break the build.
Previously, we specified the target-os in the toolset (and sometimes
used the wrong command line flags), now we have a clear separation,
which is favored by ./bootstrap.sh and ./b2.
This means that all supported OSes will specify the correct target-os=
and toolset= on the command line.
b2 will pickup our user-config.jam just fine, however, bootstrap.sh has
its own toolset autodetect mechanism, which doesn't GAF about our
user-config.jam
These haven't been updated since their addition, so this updates the list that
controls which qt base translations are bundled with the macOS binary, to all the
languages that are available with qt 5.9.8.
This could probably be improved in some way, however qt updates are infrequent,
and I didn't want to spend any more time looking at this. Also given that no-one
seems to have noticed and/or reported this it wouldn't seem high-priority.
Could be backported to 0.20.1.
Catalina SDK clang stopped automatically searching the SDK include paths when
invoked without --sysroot:
https://github.com/bitcoin/bitcoin/pull/16367#issuecomment-594600985https://github.com/Homebrew/homebrew-core/issues/45061
This hasn't been a problem for current native depends packages because are
passing their own --sysroot values, and hasn't been a problem for current host
packages because they use `darwin_` commands instead of `build_darwin_`
commands. But the current `build_darwin_CC` and `build_darwin_CXX` commands
are still unnecessarily fragile, and incompatible with new native depends
packages added in https://github.com/bitcoin/bitcoin/pull/18677.
Cory Fields <cory-nospam-@coryfields.com> suggested in
https://github.com/bitcoin/bitcoin/pull/16367#issuecomment-595393546 switching
compiler from SDK clang to native clang (from $PATH) to avoid this problem.
This is easy and makes a certain amount of sense for building native packages,
as opposed to host packages. But fanquake <fanquake@gmail.com> pointed out in
https://github.com/bitcoin/bitcoin/pull/18677#discussion_r409934309 that it
would be inconsistent use switch to non-SDK compilers while still using other
SDK tools like ranlib and install_name_tool. So simplest, minimal fix seems to
be just adding the missing --sysroot option.
