Bug 684682 – net-libs/webkit-gtk-2.24.1 - .../work/webkitgtk-2.24.1/Source/JavaScriptCore/assembler/ARM64Assembler.h:3769:100: error: ‘class JSC::AssemblerBuffer’ has no member named ‘data’ on machines with Cortex-A53 CPU

• Product: Gentoo Linux
• Status: RESOLVED FIXED

[Description Andrius Štikonas 2019-04-28 23:35:46 UTC]

webkit-gtk fails to build on my aarch64 (rk3399) system.

Reproducible: Always




[ebuild     U ~] net-libs/webkit-gtk-2.24.1:4/37::gentoo [2.22.6:4/37::gentoo] USE="X egl geolocation gstreamer introspection jpeg2k%* libnotify opengl spell wayland webgl (-aqua) -coverage -doc -gles2 -gnome-keyring -nsplugin -test (-jit%*)" 0 KiB

[Comment 1 Andrius Štikonas 2019-04-28 23:36:45 UTC]

Created attachment 574582 [details]
build.log.xz

Build log

[Comment 2 Andrius Štikonas 2019-09-02 06:29:41 UTC]

This bug is specific to Cortex-A53 CPUs.

[Comment 3 Mart Raudsepp 2019-09-02 11:27:55 UTC]

Yes, this looks like very clearly broken upstream code for this code path.

Source/JavaScriptCore/assembler/AssemblerBuffer.h has this for the data member:

#if !CPU(ARM64) // If we were to define this on arm64e, we'd need a way to update the hash as we write directly into the buffer.
        void* data() const { return m_storage.buffer(); }
#endif

while Source/JavaScriptCore/assembler/ARM64Assembler.h has this:

#if CPU(ARM64_CORTEXA53)
        CHECK_DATASIZE();
        if (datasize == 64) {
            if (LIKELY(m_buffer.codeSize() >= sizeof(int32_t))) {
                // From ARMv8 Reference Manual, Section C4.1: the encoding of the
                // instructions in the Loads and stores instruction group is:
                // ---- 1-0- ---- ---- ---- ---- ---- ----
                if (UNLIKELY((*reinterpret_cast_ptr<int32_t*>(reinterpret_cast_ptr<char*>(m_buffer.data()) + m_buffer.codeSize() - sizeof(int32_t)) & 0x0a000000) == 0x08000000))
                    nop();
            }
        }
#endif

There you can see if tries to access m_buffer.data(), but it's not even declared in the header if CPU(ARM64).

Please find upstream bug report for this, perhaps check it isn't fixed in their subversion trunk already (I think they have git mirror too, or have converted it all to git by now), and if nothing, file a new upstream bug and report the link here.

[Comment 4 Mart Raudsepp 2019-09-02 11:30:18 UTC]

The build system detects this CPU type from /proc/cpuinfo in Source/cmake/OptionsCommon.cmake and sets up the preprocessor vars and also prepends "-mfix-cortex-a53-835769" to CFLAGS/CXXFLAGS. The latter you may want to research and do yourself too if not already doing (albeit I'm not sure all a53 need that errata toolchain workaround).

[Comment 5 Nia 2020-10-11 09:40:13 UTC]

upstream bug report https://bugs.webkit.org/show_bug.cgi?id=197192.

The fix would land in webkit-gtk-2.29.1.