Which -march option to use for a generic 64-bit cpu?

Which -march option to use for a generic 64-bit cpu?

[Erik Leunissen]

L.S.

I'd like to generate code that runs on a selection of common 64-bit 
cpu's, both Intel and AMD[*].

     [*] Aside: I'm not yet entirely sure yet about the final selection 
of architectures, but in any case I would like to support core2, atom, 
athlon64 (not yet sure about opteron and AMD multicore cpu's).

To that end, I've been looking in the list of "-march" options to gcc 
for a specific option that indicates such a common set of architectures 
(much like "-march=i686" for 32-bit architectures).

I remain confused about how to do this for 64-bit cpu's. I found several 
options indicating specific cpu's that I'd like to run my code on. 
However, I guess that I'm clueless about "backward compatibility" of 
(closely) related 64-bit cpu types (like i386 code being supported by 
i686 with 32-bit cpu's). Also, I'm unsure about the differences between 
AMD and Intel in this regard.

- Do options indicating such a generic 64-bit cpu exist?, or
- Can I specifiy multiple -march options in one compile command?, or
- Do I need to compile separately for each 64-bit architecture?, or
- ... ?

I'd appreciate very much any directions about how to proceed.


Thanks for any insight,

Erik Leunissen.

Re: Which -march option to use for a generic 64-bit cpu?

[Jonathan Wakely]

On 21/01/2012, Erik Leunissen wrote:
> L.S.
>
> I'd like to generate code that runs on a selection of common 64-bit
> cpu's, both Intel and AMD[*].
>
>      [*] Aside: I'm not yet entirely sure yet about the final selection
> of architectures, but in any case I would like to support core2, atom,
> athlon64 (not yet sure about opteron and AMD multicore cpu's).
>
> To that end, I've been looking in the list of "-march" options to gcc
> for a specific option that indicates such a common set of architectures
> (much like "-march=i686" for 32-bit architectures).
>
> I remain confused about how to do this for 64-bit cpu's. I found several
> options indicating specific cpu's that I'd like to run my code on.
> However, I guess that I'm clueless about "backward compatibility" of
> (closely) related 64-bit cpu types (like i386 code being supported by
> i686 with 32-bit cpu's). Also, I'm unsure about the differences between
> AMD and Intel in this regard.
>
> - Do options indicating such a generic 64-bit cpu exist?, or

Why do you think you need to use any -march option?
If do don't use one you'll get code that runs on any x86_64 machine.

> - Can I specifiy multiple -march options in one compile command?, or

No.

> - Do I need to compile separately for each 64-bit architecture?, or
> - ... ?
>
> I'd appreciate very much any directions about how to proceed.

Just compile without any -march, unless you're sure your code will
always run on chips that support more specific features. Each -march
option lists the features it enables, so you should be able to work
out what options will work for all the machines you want to use.  e.g.
don't use -march=core2 if you want to support chips without SSSE3.

You can always use -mtune=generic to generate code that is tuned for
common chips, but still runs on older ones.

Re: Which -march option to use for a generic 64-bit cpu?

[Jonathan Wakely]

On 21 January 2012 15:33, Jonathan Wakely wrote:
>> - Do options indicating such a generic 64-bit cpu exist?, or
>
> Why do you think you need to use any -march option?
> If do don't use one you'll get code that runs on any x86_64 machine.

That was meant to say "if you don't use..."

>> - Can I specifiy multiple -march options in one compile command?, or
>
> No.

Technically you can, but only the last one will have any effect.

Re: Which -march option to use for a generic 64-bit cpu?

[Erik Leunissen]

On 21/01/12 16:33, Jonathan Wakely wrote:
> On 21/01/2012, Erik Leunissen wrote:
>>  L.S.
>>
>>  I'd like to generate code that runs on a selection of common 64-bit
>>  cpu's, both Intel and AMD[*].
>>
>>       [*] Aside: I'm not yet entirely sure yet about the final selection
>>  of architectures, but in any case I would like to support core2, atom,
>>  athlon64 (not yet sure about opteron and AMD multicore cpu's).
>>
>>  To that end, I've been looking in the list of "-march" options to gcc
>>  for a specific option that indicates such a common set of architectures
>>  (much like "-march=i686" for 32-bit architectures).
>>
>>  I remain confused about how to do this for 64-bit cpu's. I found several
>>  options indicating specific cpu's that I'd like to run my code on.
>>  However, I guess that I'm clueless about "backward compatibility" of
>>  (closely) related 64-bit cpu types (like i386 code being supported by
>>  i686 with 32-bit cpu's). Also, I'm unsure about the differences between
>>  AMD and Intel in this regard.
>>
>>  - Do options indicating such a generic 64-bit cpu exist?, or
>
> Why do you think you need to use any -march option?

Well, for all I knew, absence of any option could mean that gcc 
generates code for the machine where I run the compile command, and not 
for others.

Following that logic, I expected to need a -march option.

Apparently this is just one part where I was confused/ignorant, as 
predicted.

> If do don't use one you'll get code that runs on any x86_64 machine.
>

I simply didn't know this. It's just what I need.


>>  - Can I specifiy multiple -march options in one compile command?, or
>
> No.
>
>>  - Do I need to compile separately for each 64-bit architecture?, or
>>  - ... ?
>>
>>  I'd appreciate very much any directions about how to proceed.
>
> Just compile without any -march, unless you're sure your code will
> always run on chips that support more specific features. Each -march
> option lists the features it enables, so you should be able to work
> out what options will work for all the machines you want to use.  e.g.
> don't use -march=core2 if you want to support chips without SSSE3.
>

OK

> You can always use -mtune=generic to generate code that is tuned for
> common chips, but still runs on older ones.
>

OK, thanks for clearing up my confusion/ignorance.

Erik.

Re: Which -march option to use for a generic 64-bit cpu?

[Tim Prince]

On 1/21/2012 12:22 PM, Erik Leunissen wrote:
> On 21/01/12 16:33, Jonathan Wakely wrote:

>>
>> Why do you think you need to use any -march option?
>
> Well, for all I knew, absence of any option could mean that gcc
> generates code for the machine where I run the compile command, and not
> for others.
>
> Following that logic, I expected to need a -march option.
You're thinking of -march=native.  It's not set by default, but you 
might find it set in a makefile.



-- 
Tim Prince

Re: Which -march option to use for a generic 64-bit cpu?

[Jonathan Wakely]

On 21 January 2012 17:22, Erik Leunissen wrote:
> On 21/01/12 16:33, Jonathan Wakely wrote:
>>
>> Why do you think you need to use any -march option?
>
>
> Well, for all I knew, absence of any option could mean that gcc generates
> code for the machine where I run the compile command, and not for others.

No, the docs say:

"While picking a specific cpu-type will schedule things appropriately
for that particular chip, the compiler will not generate any code that
does not run on the i386 without the -march=cpu-type option being
used."

That applies to x86_64 too, if you don't use -march you will get code
that runs on any x86_64 processor, i.e. using just the basic x86_64
instruction set.

I've updated that text for the GCC 4.7 manual to say:
"While picking a specific cpu-type will schedule things appropriately
for that particular chip, the compiler will not generate any code that
does not run on the default machine type without the -march=cpu-type
option being used. For example, if GCC is configured for
i686-pc-linux-gnu then -mtune=pentium4 will generate code that is
tuned for Pentium4 but will still run on i686 machines."

i.e. if your compiler is configured for amdfam10-unknown-linux-gnu
then it will behave as though -march=amdfam10 is given, unless you
override it with an explicit -march option.  But if it's configured
for x86_64-unknown-linux-gnu (the default in most cases) then it
generates code that will run on any x86_64.

As Tim pointed out, the behaviour you describe is enabled by
-march-native, documented like so

"Using -march=native will enable all instruction subsets supported by
the local machine (hence the result might not run on different
machines)."

Re: Which -march option to use for a generic 64-bit cpu?

[Erik Leunissen]

On 21/01/12 19:37, Jonathan Wakely wrote:
> On 21 January 2012 17:22, Erik Leunissen wrote:
>>  On 21/01/12 16:33, Jonathan Wakely wrote:
>>>
>>>  Why do you think you need to use any -march option?
>>
>>
>>  Well, for all I knew, absence of any option could mean that gcc generates
>>  code for the machine where I run the compile command, and not for others.
>
> No, the docs say:
>

I now understand completely. Central to that understanding is your 
explanation of the meaning of "the default machine type". The example 
helps a great deal too, to be able to check my understanding.

I agree that that can be read in the docs.

Recognizing the value of your explanation leads me to suggest a small 
addition to the current text in the manual. Not under the option -mtune, 
but under the option -march. Maybe something like:

-march=cpu-type
Without any -march option, gcc generates code for the default machine 
type, i.e. the target-cpu for which your specific gcc program has been 
configured during build time. You may retrieve the default machine type 
with "gcc -dumpmachine". Adding an option -march makes gcc generate 
instructions for the machine type cpu-type. The choices for cpu-type are 
the same as for -mtune. Moreover, specifying -march=cpu-type implies 
-mtune=cpu-type.


Greetings,

Erik.


> "While picking a specific cpu-type will schedule things appropriately
> for that particular chip, the compiler will not generate any code that
> does not run on the i386 without the -march=cpu-type option being
> used."
>
> That applies to x86_64 too, if you don't use -march you will get code
> that runs on any x86_64 processor, i.e. using just the basic x86_64
> instruction set.
>
> I've updated that text for the GCC 4.7 manual to say:
> "While picking a specific cpu-type will schedule things appropriately
> for that particular chip, the compiler will not generate any code that
> does not run on the default machine type without the -march=cpu-type
> option being used. For example, if GCC is configured for
> i686-pc-linux-gnu then -mtune=pentium4 will generate code that is
> tuned for Pentium4 but will still run on i686 machines."
>
> i.e. if your compiler is configured for amdfam10-unknown-linux-gnu
> then it will behave as though -march=amdfam10 is given, unless you
> override it with an explicit -march option.  But if it's configured
> for x86_64-unknown-linux-gnu (the default in most cases) then it
> generates code that will run on any x86_64.
>
> As Tim pointed out, the behaviour you describe is enabled by
> -march-native, documented like so
>
> "Using -march=native will enable all instruction subsets supported by
> the local machine (hence the result might not run on different
> machines)."
>