-march=<name of native> doesn't seem to set flags -march=native does?

Manuel López-Ibáñez lopezibanez@gmail.com
Thu Jul 14 23:25:00 GMT 2016

On 14/07/16 10:41, Marc Glisse wrote:
> On Thu, 14 Jul 2016, Jonathan Wakely wrote:
>> On 14 July 2016 at 01:05, Roger Pack wrote:
>>> Hello.
>>> As a note after a discussion [1]
>>> I ran into this oddness:
>>> $ gcc-6 -march=native -Q --help=target | grep march
>>>   -march=                               ivybridge
>>> $ gcc-6 -march=native -Q --help=target | grep sse3
>>>   -msse3                                [enabled]
>>>   -mssse3                               [enabled]
>>> $ gcc-6 -march=ivybridge -Q --help=target | grep sse3
>>>   -msse3                                [disabled]
>>>   -mssse3                               [disabled]
>>> It appears that specifying (in this case) "-march=ivybridge" which is
>>> the native, isn't enough/the same as specifying -march=native?
>>> In addition, the "cache sizes" don't seem to be called out with
>>> "-march=ivybridge" as they are with "-march=native"
>>> Anybody know what I'm missing here?
>> The named arch is a pre-configured set of supported instructions,
>> which should be true for all processors in that range.
>> Whereas "native" probes the CPU flags and enables precisely the set of
>> instructions it supports.
>> That they are different suggests either the preconfigured settings for
>> ivybridge are wrong, or not all processors in that family support all
>> the instructions your particular model supports.
>> Ivybridge should enable SSE3 and SSSE3 though, so I'm nto sure what's
>> happening there.
> Option reporting is not reliable. Try preprocessing a file that contains the
> macro __SSE3__ to check if sse3 is enabled.

-fverbose-asm should also be quite reliable: 

I think the problem is a misdesign/limitation of option reporting with --help, 
which only takes account the invokation of the compiler (see the differences in 
output when you add -### to your command-lines).

It may be worth opening a PR if there is not one already open: 


More information about the Gcc-help mailing list