GMP 6.0.0a on AMD A8-6600K

Marc Glisse marc.glisse at
Sun Apr 12 14:21:33 UTC 2015

On Sun, 12 Apr 2015, Marc Glisse wrote:

> On Sun, 12 Apr 2015, Marc Glisse wrote:
>> What is strange is that the configure test "double -> float" should have 
>> generated vcvtsd2ss and caused an error (assuming all AVX instructions are 
>> rejected), and that would have caused configure to try adding -mno-avx, 
>> but:
>> Test compile: double -> float conversion
>> configure:7588: gcc -O2 -pedantic -fomit-frame-pointer -m64 -mtune=bdver1 
>> -march=bdver1 conftest.c >&5
>> configure:7591: $? = 0
>> configure:7596: ./a.out || ./b.out || ./a.exe || ./a_out.exe || ./conftest
>> configure:7599: $? = 0
> Silly me, main never calls the function that has this instruction, so it will 
> not fail. Either we need a configure test that actually executes some 
> instructions involving float or double (not just compile them), or we should 
> be more forceful in adding -mno-avx for *noavx platforms instead of waiting 
> for some test to fail.

Sorry, misunderstanding on my part, optlist should work as is.

Interestingly, it works fine for ABI=32 (it adds -mno-avx), while for 
ABI=64 it ignores noavx (doesn't ever try -mno-avx).

It is easy to reproduce on other systems with
./configure --build=piledrivernoavx-unknown-linux-gnu

Marc Glisse

More information about the gmp-bugs mailing list