Version 0.3.21 isn't built on M1

[devcrocod]

os: macOS
arch: aarch64
C compiler: gcc 11.3.0_1
Fortran compiler: gfortran 11.3.0_1
OpenBLAS version: 0.3.21

make command:

make CC=gcc FC=gfortran HOSTCC=gcc BINARY=64 F_COMPILER=GFORTRAN FEXTRALIB=-lgfortran USE_OPENMP=0 NO_AVX512=1 DYNAMIC_ARCH=1 NUM_THREADS=64

On version 0.3.20, the assembly was successful. When updating the library to 0.3.21, I got the following error:

In file included from ../common.h:133,
                 from ../kernel/arm64/../arm/iamin.c:37:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
In file included from ../common.h:133,
                 from ../kernel/arm64/../arm/imax.c:38:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
In file included from ../common.h:133,
                 from ../kernel/arm64/../arm/min.c:37:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
In file included from ../common.h:133,
                 from ../kernel/arm64/iamax_thunderx2t99.c:28:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
In file included from ../common.h:133,
                 from ../kernel/arm64/../arm/max.c:37:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
In file included from ../common.h:133,
                 from ../kernel/arm64/../arm/amin.c:37:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
In file included from ../common.h:133,
                 from ../kernel/arm64/../arm/imin.c:36:
/opt/homebrew/Cellar/gcc/11.3.0_1/lib/gcc/11/gcc/aarch64-apple-darwin21/11/include-fixed/math.h:68:5: error: #error "Unsupported value of __FLT_EVAL_METHOD__."
   68 | #   error "Unsupported value of __FLT_EVAL_METHOD__."
      |     ^~~~~
make[5]: *** [isamin_k_NEOVERSEN2.o] Error 1
make[5]: *** Waiting for unfinished jobs....
make[5]: *** [smax_k_NEOVERSEN2.o] Error 1
make[5]: *** [isamax_k_NEOVERSEN2.o] Error 1
make[5]: *** [smin_k_NEOVERSEN2.o] Error 1
make[5]: *** [ismax_k_NEOVERSEN2.o] Error 1
make[5]: *** [samin_k_NEOVERSEN2.o] Error 1
make[5]: *** [ismin_k_NEOVERSEN2.o] Error 1

[martin-frbg]

hm, looks like 11.3 may not be recent enough to build Neoverse N2 kernels.
if you need DYNAMIC_ARCH and cannot upgrade gcc, it should be possible to work around this by adding DYNAMIC_LIST="ARMV8 VORTEX" (and whatever other cpus you need except the N2)

[martin-frbg]

Actually this looks like a GCC build/packaging error, something similar was reported against the macports build of gcc-11 several months ago here: https://trac.macports.org/ticket/63072 (predating the addition of Neoverse N2 in OpenBLAS), with comments claiming that __FLT_EVAL_METHOD__ is supposed to be defined by the compiler itself, to one of the values supported in math.h (which seem to be -1,0,1 or 2, there is/was a similar issue with LLVM discussed at https://www.mail-archive.com/[email protected]/msg267040.html)
(Unfortunately the macports ticket appears to have been pretty much abandoned in an unsolved state)

[martin-frbg]

Reproduced with the homebrew gcc@11 on M1 - gcc@12 (currently 12.1.0) does not have this program, but here the compile fails for the NeoverseN2 simatcopy_cn kernel with a spurious assembler error "instruction requires sve" despite the -march=armv8.5a+sve+sve2+bf16 . Assembler is from XCode/LLVM (and homebrew binutils does not contain an alternative).

[martin-frbg]

I can get it to build with gcc-11 as well if I drop the SVE capability from the -march features string. This should do for the time being.