Re: [PATCH] Fix PR78189

[Richard Biener <rguenther at suse dot de>]

On Tue, 8 Nov 2016, Richard Biener wrote:

> On Mon, 7 Nov 2016, Christophe Lyon wrote:
> 
> > Hi Richard,
> > 
> > 
> > On 7 November 2016 at 09:01, Richard Biener <rguenther@suse.de> wrote:
> > >
> > > The following fixes an oversight when computing alignment in the
> > > vectorizer.
> > >
> > > Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
> > >
> > > Richard.
> > >
> > > 2016-11-07  Richard Biener  <rguenther@suse.de>
> > >
> > >         PR tree-optimization/78189
> > >         * tree-vect-data-refs.c (vect_compute_data_ref_alignment): Fix
> > >         alignment computation.
> > >
> > >         * g++.dg/torture/pr78189.C: New testcase.
> > >
> > > Index: gcc/testsuite/g++.dg/torture/pr78189.C
> > > ===================================================================
> > > --- gcc/testsuite/g++.dg/torture/pr78189.C      (revision 0)
> > > +++ gcc/testsuite/g++.dg/torture/pr78189.C      (working copy)
> > > @@ -0,0 +1,41 @@
> > > +/* { dg-do run } */
> > > +/* { dg-additional-options "-ftree-slp-vectorize -fno-vect-cost-model" } */
> > > +
> > > +#include <cstddef>
> > > +
> > > +struct A
> > > +{
> > > +  void * a;
> > > +  void * b;
> > > +};
> > > +
> > > +struct alignas(16) B
> > > +{
> > > +  void * pad;
> > > +  void * misaligned;
> > > +  void * pad2;
> > > +
> > > +  A a;
> > > +
> > > +  void Null();
> > > +};
> > > +
> > > +void B::Null()
> > > +{
> > > +  a.a = nullptr;
> > > +  a.b = nullptr;
> > > +}
> > > +
> > > +void __attribute__((noinline,noclone))
> > > +NullB(void * misalignedPtr)
> > > +{
> > > +  B* b = reinterpret_cast<B*>(reinterpret_cast<char *>(misalignedPtr) - offsetof(B, misaligned));
> > > +  b->Null();
> > > +}
> > > +
> > > +int main()
> > > +{
> > > +  B b;
> > > +  NullB(&b.misaligned);
> > > +  return 0;
> > > +}
> > > diff --git gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c
> > > index 9346cfe..b03cb1e 100644
> > > --- gcc/tree-vect-data-refs.c
> > > +++ gcc/tree-vect-data-refs.c
> > > @@ -773,10 +773,25 @@ vect_compute_data_ref_alignment (struct data_reference *dr)
> > >    base = ref;
> > >    while (handled_component_p (base))
> > >      base = TREE_OPERAND (base, 0);
> > > +  unsigned int base_alignment;
> > > +  unsigned HOST_WIDE_INT base_bitpos;
> > > +  get_object_alignment_1 (base, &base_alignment, &base_bitpos);
> > > +  /* As data-ref analysis strips the MEM_REF down to its base operand
> > > +     to form DR_BASE_ADDRESS and adds the offset to DR_INIT we have to
> > > +     adjust things to make base_alignment valid as the alignment of
> > > +     DR_BASE_ADDRESS.  */
> > >    if (TREE_CODE (base) == MEM_REF)
> > > -    base = build2 (MEM_REF, TREE_TYPE (base), base_addr,
> > > -                  build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)), 0));
> > > -  unsigned int base_alignment = get_object_alignment (base);
> > > +    {
> > > +      base_bitpos -= mem_ref_offset (base).to_short_addr () * BITS_PER_UNIT;
> > > +      base_bitpos &= (base_alignment - 1);
> > > +    }
> > > +  if (base_bitpos != 0)
> > > +    base_alignment = base_bitpos & -base_bitpos;
> > > +  /* Also look at the alignment of the base address DR analysis
> > > +     computed.  */
> > > +  unsigned int base_addr_alignment = get_pointer_alignment (base_addr);
> > > +  if (base_addr_alignment > base_alignment)
> > > +    base_alignment = base_addr_alignment;
> > >
> > >    if (base_alignment >= TYPE_ALIGN (TREE_TYPE (vectype)))
> > >      DR_VECT_AUX (dr)->base_element_aligned = true;
> > 
> > Since you committed this patch (r241892), I'm seeing execution failures:
> >   gcc.dg/vect/pr40074.c -flto -ffat-lto-objects execution test
> >   gcc.dg/vect/pr40074.c execution test
> > on armeb-none-linux-gnueabihf --with-mode=arm --with-cpu=cortex-a9
> > --with-fpu=neon-fp16
> > (using qemu as simulator)
> 
> The difference is that we now vectorize the testcase with versioning
> for alignment (but it should never execute the vectorized variant).
> I need arm peoples help to understand what is wrong.
> 
> At least the testcase shows there is (kind-of) a missed optimization
> that we no longer figure out versioning for alignment is useless.
> I'll look into that.

I'm testing the following which restores previous behavior.  You
might still want to figure out what went wrong.

Richard.

2016-11-08  Richard Biener  <rguenther@suse.de>

	* tree-vect-data-refs.c (vect_compute_data_ref_alignment):
	Look at the DR_BASE_ADDRESS object for forcing alignment.

diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c
index b03cb1e..48995f4 100644
--- a/gcc/tree-vect-data-refs.c
+++ b/gcc/tree-vect-data-refs.c
@@ -813,12 +813,7 @@ vect_compute_data_ref_alignment (struct 
data_reference *dr)
 
   if (base_alignment < TYPE_ALIGN (vectype))
     {
-      /* Strip an inner MEM_REF to a bare decl if possible.  */
-      if (TREE_CODE (base) == MEM_REF
-         && integer_zerop (TREE_OPERAND (base, 1))
-         && TREE_CODE (TREE_OPERAND (base, 0)) == ADDR_EXPR)
-       base = TREE_OPERAND (TREE_OPERAND (base, 0), 0);
-
+      base = TREE_OPERAND (base_addr, 0);
       if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype)))
        {
          if (dump_enabled_p ())