__attribute__ ifunc examples?

__attribute__ ifunc examples?

[James Cloos]

Does anyone have a worked example in C of using attribute ifunc to
dispatch over the various x86 SSE/AVX flags, as well as arm neon?

Must one call cpuid manually, or are flags already available to check?

I haven't found any via goog (and the C++ example I found was
incompatible with the gcc-4.6 syntax for ifunc), and glibc's
support for it is too opaque for grep.

Thanks,

-JimC
-- 
James Cloos <cloos@jhcloos.com>         OpenPGP: 1024D/ED7DAEA6

Re: __attribute__ ifunc examples?

[Ian Lance Taylor]

James Cloos <cloos@jhcloos.com> writes:

> Does anyone have a worked example in C of using attribute ifunc to
> dispatch over the various x86 SSE/AVX flags, as well as arm neon?

Not using SSE/AVX flags, but the basic idea is not too hard.  Here is an
example that sets the implementation of foo based on a global variable.


#include <stdio.h>

int foo (void) __attribute__ ((ifunc ("foo_ifunc")));

static int global = 1;

static int
f1 (void)
{
  return 0;
}

static int
f2 (void)
{
  return 1;
}

void *
foo_ifunc (void)
{
  return global == 1 ? f1 : f2;
}

int
main ()
{
  printf ("%d\n", foo());
}


For your case you need to write foo_ifunc using appropriate asm
instructions to check cpuid or whatever.  Note that foo_ifunc runs as
the program is just starting, so it can't call library routines or check
command line arguments or anything like that.  Checking cpuid is fine.


> Must one call cpuid manually, or are flags already available to check?

At present you have to call it manually.  There is an outstanding patch
to add __builtin_cpuid on x86 systems, but I don't think it has been
approved.

Ian

Re: __attribute__ ifunc examples?

[James Cloos]

>>>>> "ILT" == Ian Lance Taylor <iant@google.com> writes:

JC>> Does anyone have a worked example in C of using attribute ifunc to
JC>> dispatch over the various x86 SSE/AVX flags, as well as arm neon?

ILT> For your case you need to write foo_ifunc using appropriate asm
ILT> instructions to check cpuid or whatever.

Thanks, but that much I already got out of the docs.  I was hoping for
an example of actual checks for the various cpu features for x86 & arm
in the ifunc context.

The existing cpu feature detectors out in the wild, plus the i+a+v docs,
should provide enough extra clue, of course.  Just a matter of coding,
after all. 

ILT> At present you have to call it manually.  There is an outstanding patch
ILT> to add __builtin_cpuid on x86 systems, but I don't think it has been
ILT> approved.

That would be a useful addition.

Thank you.

-JimC
-- 
James Cloos <cloos@jhcloos.com>         OpenPGP: 1024D/ED7DAEA6