Compile bug: since 5615 build on PowerPC fails with ICE: `ggml-cpu-impl.h: internal compiler error: in assemble_alias, at varasm.cc:6435`

[barracuda156]

Git commit

2e89f76

Operating systems

Mac

GGML backends

BLAS

Problem description & steps to reproduce

Since 5615 the build fails on PowerPC with ICE.

@xctan @ggerganov @i0ntempest FYI

First Bad Commit

f470bc3

Compile command

/opt/local/bin/cmake -G "CodeBlocks - Unix Makefiles" -DCMAKE_BUILD_TYPE=MacPorts -DCMAKE_INSTALL_PREFIX="/opt/local" -DCMAKE_INSTALL_NAME_DIR="/opt/local/lib" -DCMAKE_SYSTEM_PREFIX_PATH="/opt/local;/usr" -DCMAKE_C_COMPILER="$CC" -DCMAKE_CXX_COMPILER="$CXX" -DCMAKE_OBJC_COMPILER="$CC" -DCMAKE_OBJCXX_COMPILER="$CXX" -DCMAKE_POLICY_DEFAULT_CMP0025=NEW -DCMAKE_POLICY_DEFAULT_CMP0060=NEW -DCMAKE_VERBOSE_MAKEFILE=ON -DCMAKE_COLOR_MAKEFILE=ON -DCMAKE_FIND_FRAMEWORK=LAST -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_MAKE_PROGRAM=/usr/bin/make -DCMAKE_MODULE_PATH="/opt/local/share/cmake/Modules" -DCMAKE_PREFIX_PATH="/opt/local/share/cmake/Modules" -DCMAKE_BUILD_WITH_INSTALL_RPATH:BOOL=ON -DCMAKE_INSTALL_RPATH="/opt/local/lib" -Wno-dev -DGGML_LTO=ON -DGGML_CCACHE=OFF -DGGML_OPENMP=ON -DLLAMA_CURL=ON -DGGML_METAL=OFF -DGGML_METAL_EMBED_LIBRARY=OFF -DGGML_BLAS=ON -DGGML_ACCELLERATE=ON -DGGML_BLAS_VENDOR=Apple -DCMAKE_OSX_ARCHITECTURES="ppc" -DCMAKE_OSX_DEPLOYMENT_TARGET="10.6" -DCMAKE_OSX_SYSROOT="/" /opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615

Relevant log output

In file included from /opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615/ggml/src/ggml-cpu/quants.c:4:
/opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615/ggml/src/ggml-cpu/ggml-cpu-impl.h:510:28: internal compiler error: in assemble_alias, at varasm.cc:6435
  510 | #define GGML_DO_PRAGMA_(x) _Pragma (#x)
      |                            ^~~~~~~
/opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615/ggml/src/ggml-cpu/ggml-cpu-impl.h:511:27: note: in expansion of macro 'GGML_DO_PRAGMA_'
  511 | #define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
      |                           ^~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615/ggml/src/ggml-cpu/ggml-cpu-impl.h:520:39: note: in expansion of macro 'GGML_DO_PRAGMA'
  520 | # define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
      |                                       ^~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615/ggml/src/ggml-cpu/ggml-cpu-impl.h:530:36: note: in expansion of macro 'GGML_WEAK_ALIAS'
  530 | #define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)
      |                                    ^~~~~~~~~~~~~~~
/opt/local/var/macports/build/_opt_local_ppcports_llm_llama.cpp/llama.cpp/work/llama.cpp-5615/ggml/src/ggml-cpu/quants.c:41:1: note: in expansion of macro 'GGML_CPU_NATIVE_IMPL'
   41 | GGML_CPU_NATIVE_IMPL(quantize_row_q8_0)
      | ^~~~~~~~~~~~~~~~~~~~
Please submit a full bug report, with preprocessed source (by using -freport-bug).
See <https://trac.macports.org/newticket> for instructions.
make[2]: *** [ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/quants.c.o] Error 1
make[2]: *** Waiting for unfinished jobs....

[xctan]

Could you provide your cmake logs? BTW this issue is related to the weak alias feature, which is missing on Apple targets. Could you additionally provide predefined compiler macros of your compiler (e.g. $CC -dM -E empty.c -o macro.i)?

[barracuda156]

@xctan Thank you for responding!

The log and macros (is this what is expected?):

macro.txt

main-5615.log

[xctan]

I've pushed a potential fix in #14146. Would you be able to try it and let me know if it resolves the issue for you?

[barracuda156]

@xctan Thank you, I will check today and update you.