Re: [gomp4] Some progress on #pragma omp simd

[Aldy Hernandez <aldyh at redhat dot com>]

On 06/17/13 12:23, Richard Henderson wrote:
> On 06/17/2013 10:13 AM, Aldy Hernandez wrote:
> > -	  data.simduid = tree_low_cst (gimple_call_arg (stmt, 0), 1);
> > +	  data.simduid = gimple_call_arg (stmt, 0);
>
> Doesn't this copy the ADDR_EXPR from the call into simduid?
>
> >   simduid_to_vf::hash (const value_type *p)
> >   {
> > -  return p->simduid;
> > +  return htab_hash_pointer (p->simduid);
>
> ... at which point this bit is meaningless since all ADDR_EXPRs must of course
> have different pointers.
>
> I think we should validate the DECL_P extracted from the call_arg, and store
> that.  The hash should use DECL_UID to minimize hash variation due to memory
> layout.

As discussed on IRC.  Attached are these changes you requested, plus 
changing OMP_CLAUSE__SIMDUID__UID to OMP_CLAUSE__SIMDUID__DECL.

I will tackle the dot named builtins in the next iteration.

BTW, this patch bootstraps with no regressions.  I also manually 
inspected the gimple generated by the test below, and made sure that 
inlining func() into both foo() and bar() have different temporaries. 
Without this patch, the same constant was used incorrectly as arguments 
to __builtin_GOMP.simd_vf and __builtin_GOMP.simd_lane.

How does this look?
Aldy

#define N 1000

static inline int func (int *p)
{
  int x = 0, i;
#pragma simd reduction (+:x)
  for (i = 0; i < 1000; i++)
    x += p[i];
  return x;
}

int array[55555];
int dork[5555];
foo()
{
  return func(array);
}

bar()
{
  return func(dork) + 666;
}

Attachment: curr
Description: Text document