Segfault importing pytrellis.so on Apple M1

[ccattuto]

Apple Macbook Pro M1
OSX Monterey 12.2.1
Native arm64 Python 3.9.10 via Miniforge
Xcode 13.2.1
Apple clang version 13.0.0 (clang-1300.0.29.30)
boost 1.78.0_1 (Homebrew)

The library and executables build cleanly. No warnings of notice except for a few deprecations:

/opt/homebrew/include/boost/format/group.hpp:59:12: warning: definition of implicit copy constructor for 'group1<boost::io::detail::group0>' is deprecated because it has a user-declared copy assignment operator [-Wdeprecated-copy]
   group1& operator=(const group1&);

pytrellis.so is very big (~8 Mb). On importing it from Python, it segfaults:

Python 3.9.10 | packaged by conda-forge | (main, Feb  1 2022, 21:27:43) 
[Clang 11.1.0 ] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import pytrellis.so
zsh: segmentation fault  python

Digging a bit deeper shows that this happens when calling take_gil() when importing pytrellis.so.

(lldb) bt
* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x10)
  * frame #0: 0x000000010489cad8 libpython3.9.dylib`take_gil + 84
    frame #1: 0x00000001048ffce0 libpython3.9.dylib`PyGILState_Ensure + 56
    frame #2: 0x0000000103ff57ec pytrellis.so`pybind11::detail::get_internals() + 64
    frame #3: 0x0000000103ff567c pytrellis.so`PyInit_pytrellis + 120

[qcabanes-hobby]

Same issue here.

OSX Monterey 12.2.1
Native arm64
Python 3.10.2 (installed via Homebrew)
Boost 1.78.0
AppleClang 13.0.0.13000029

[ccattuto]

There's also an issue with dynamic linking of the executables:

% ./ecppack 
dyld[87459]: Library not loaded: @rpath/libtrellis.dylib
  Referenced from: /Users/ciro/prjtrellis/libtrellis/ecppack
  Reason: tried: '/Users/ciro/prjtrellis/libtrellis/../lib/trellis/libtrellis.dylib' (no such file), '/Users/ciro/prjtrellis/libtrellis/../lib/trellis/libtrellis.dylib' (no such file), '/usr/local/lib/libtrellis.dylib' (no such file), '/usr/lib/libtrellis.dylib' (no such file)
zsh: abort      ./ecppack

Using otool -L I get:

ecppack:
	@rpath/libtrellis.dylib (compatibility version 0.0.0, current version 0.0.0)
	/opt/homebrew/opt/boost/lib/libboost_filesystem-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
	/opt/homebrew/opt/boost/lib/libboost_program_options-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
	/opt/homebrew/opt/boost/lib/libboost_system-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
	/opt/homebrew/opt/boost/lib/libboost_thread-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
	/opt/homebrew/opt/boost/lib/libboost_chrono-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
	/opt/homebrew/opt/boost/lib/libboost_atomic-mt.dylib (compatibility version 0.0.0, current version 0.0.0)
	@rpath/libpython3.9.dylib (compatibility version 3.9.0, current version 3.9.0)
	/usr/lib/libc++.1.dylib (compatibility version 1.0.0, current version 1200.3.0)
	/usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1311.0.0)

Hence this can be fixed using install_name_tool to specify the path of libpython3.9.dylib:

install_name_tool ./ecppack -add_rpath MINIFORGE_LIB_PATH

Yielding:

% ecppack 
Error: input file is mandatory.

Project Trellis - Open Source Tools for ECP5 FPGAs
Version 1.2.1
ecppack: ECP5 bitstream packer

Copyright (C) 2018 gatecat <[email protected]>

Usage: ecppack input.config [output.bit] [options]
Allowed options:
  -h [ --help ]         show help
  -v [ --verbose ]      verbose output
  --db arg              Trellis database folder location
  --usercode arg        USERCODE to set in bitstream
  --idcode arg          IDCODE to override in bitstream
  --freq arg            config frequency in MHz
  --svf arg             output SVF file
  --svf-rowsize arg     SVF row size in bits (default 8000)
  --compress            compress bitstream to reduce size
  --spimode arg         SPI Mode to use (fast-read, dual-spi, qspi)
  --background          enable background reconfiguration in bitstream
  --delta arg           create a delta partial bitstream given a reference 
                        config
  --bootaddr arg        set next BOOTADDR in bitstream and enable multi-boot
  --input arg           input textual configuration
  --bit arg             output bitstream file
  --version             show current version and exit

[ccattuto]

Update: manually removing libpython3.9.dylib from the linker script for pytrellis.so (prjtrellis/libtrellis/CMakeFiles/pytrellis.dir/link.txt) fixes the segfault.

[qcabanes]

Works like a charm! I could finally install nextpnr for ECP5

[JamesTimothyMeech]

Has anyone else tried the above solution without success? Could me using python3.11 instead of 3.9 prevent the fix from working?