Re: [Discuss-gnuradio] Volk build error using MacPorts GCC 4.7

[Michael Dickens]

On Jan 25, 2013, at 1:19 AM, Albert Chun-Chieh Huang <address@hidden> wrote:
> After digging into this issue, I found an answer on MacPorts list, and
> it's related to Apple's ancient assembler(version 1.38). In this post,
> they try to replace Apple /usr/bin/as by clang's assembler.
> 
> http://lists.macosforge.org/pipermail/macports-dev/2011-October/016335.html
> 
> I tried this approach to compile gnuradio with gcc 4.7, it's still not
> successful. But it seems to be an interesting and possible way -- to use
> clang's assembler to optimize for AVX on MacOSX. GNU binutil on MacOSX
> does not come with GNU assembler, so gcc 4.7 on MacOSX uses the built-in
> GNU assembler, which is very old and does not support AVX. Only clang's
> assembler on MacOSX provides support to AVX instruction set.

Hi Albert - Nice sleuthing!  I don't generally read the MacPorts developers' 
list emails unless they are addressed to me; hence, that thread slipped through 
my reading.  It seems like this issue has multiple parts:

1) Does the target compiler's assembler support the "xgetbv" instruction?

2) Does the target CPU support the "xgetbv" instruction?

Here's my "hack" attempt.  Add the following code to lib/volk/CMakeLists.txt:132

{{{
if(CPU_IS_x86)
   execute_process(COMMAND ${CMAKE_C_COMPILER} -o
         ${CMAKE_CURRENT_BINARY_DIR}/temp
          ${CMAKE_CURRENT_SOURCE_DIR}/test_xgetbv.c
      OUTPUT_QUIET ERROR_QUIET
      RESULT_VARIABLE avx_compile_result)
   if(NOT ${avx_compile_result} EQUAL 0)
      OVERRULE_ARCH(avx "Compiler missing xgetbv")
   else()
     message(STATUS "Executing temp")
      execute_process(COMMAND ${CMAKE_CURRENT_BINARY_DIR}/temp
          OUTPUT_QUIET ERROR_QUIET
          RESULT_VARIABLE avx_exe_result)
     if(NOT ${avx_exe_result} EQUAL 0)
       OVERRULE_ARCH(avx "CPU missing xgetbv")
     endif()
   endif()
   file(REMOVE ${CMAKE_CURRENT_BINARY_DIR}/temp)
endif()
}}}

and add the file lib/volk/test_xgetbv.c :

{{{
static inline unsigned long long _xgetbv(unsigned int index) {
  unsigned int eax, edx;
  __asm__ __volatile__("xgetbv" : "=a"(eax), "=d"(edx) : "c"(index));
  return ((unsigned long long)edx << 32) | eax;
}

int main (void) {
  (void) _xgetbv(0);
  return (0);
}
}}}

I tried to do this solely in cmake using "file(WRITE …)", but it never wrote 
the file (latest cmake, btw).  Hence, using a separate file for the test.

On my system, this test result in cmake output "Compiler missing xgetbv, 
Overruled arch avx" which is as desired.  If I switch the __asm__ line to 
instead read:
{{{
__asm__ (".byte 0x0f, 0x01, 0xd0" : "=a"(eax), "=d"(edx) : "c"(index));
}}}
then this test results in "CPU missing xgetbv, Overruled arch avx", which is 
again as desired.  It would also be good for someone to test this on a system 
which -does- support xgetbv to make sure it works successfully.

I'll let the VOLK-powers-that-be determine if this test makes any sense to 
include, or something like it, but IMHO it's wiser to actually -test- for a 
feature when reasonably possible rather than assume it based on standard 
information such as program version. :) - MLD