Re: [PATCH] tree-optimization/106131 - wrong code with FRE rewriting

[Richard Sandiford <richard.sandiford@arm.com>]

Richard Biener <rguenther@suse.de> writes:
> On Fri, 1 Jul 2022, Richard Sandiford wrote:
>
>> Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
>> > The following makes sure to not use the original TBAA type for
>> > looking up a value across an aggregate copy when we had to offset
>> > the read.
>> >
>> > Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed to trunk.
>> >
>> > 2022-06-30  Richard Biener  <rguenther@suse.de>
>> >
>> > 	PR tree-optimization/106131
>> > 	* tree-ssa-sccvn.cc (vn_reference_lookup_3): Force alias-set
>> > 	zero when offsetting the read looking through an aggregate
>> > 	copy.
>> >
>> > 	* g++.dg/torture/pr106131.C: New testcase.
>> > ---
>> >  gcc/testsuite/g++.dg/torture/pr106131.C | 34 +++++++++++++++++++++++++
>> >  gcc/tree-ssa-sccvn.cc                   | 16 +++++++++---
>> >  2 files changed, 46 insertions(+), 4 deletions(-)
>> >  create mode 100644 gcc/testsuite/g++.dg/torture/pr106131.C
>> >
>> > diff --git a/gcc/testsuite/g++.dg/torture/pr106131.C b/gcc/testsuite/g++.dg/torture/pr106131.C
>> > new file mode 100644
>> > index 00000000000..e110f4a8fe6
>> > --- /dev/null
>> > +++ b/gcc/testsuite/g++.dg/torture/pr106131.C
>> > @@ -0,0 +1,34 @@
>> > +// { dg-do run { target c++11 } }
>> > +
>> > +struct Pair {
>> > +    int a, b;
>> > +    Pair(const Pair &) = default;
>> > +    Pair(int _a, int _b) : a(_a), b(_b) {}
>> > +    Pair &operator=(const Pair &z) {
>> > +	a = z.a;
>> > +	b = z.b;
>> > +	return *this;
>> > +    }
>> > +};
>> > +
>> > +const int &max(const int &a, const int &b)
>> > +{
>> > +  return a < b ? b : a;
>> > +}
>> > +
>> > +int foo(Pair x, Pair y)
>> > +{
>> > +  return max(x.b, y.b);
>> > +}
>> > +
>> > +int main()
>> > +{
>> > +  auto f = new Pair[3] {{0, -11}, {0, -8}, {0, 2}};
>> > +  for (int i = 0; i < 1; i++) {
>> > +      f[i] = f[0];
>> > +      if(i == 0)
>> > +	f[i] = f[2];
>> > +      if (foo(f[i], f[1]) != 2)
>> > +	__builtin_abort();
>> > +  }
>> > +}
>> > diff --git a/gcc/tree-ssa-sccvn.cc b/gcc/tree-ssa-sccvn.cc
>> > index 9deedeac378..76d92895a3a 100644
>> > --- a/gcc/tree-ssa-sccvn.cc
>> > +++ b/gcc/tree-ssa-sccvn.cc
>> > @@ -3243,12 +3243,12 @@ vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
>> >        poly_int64 extra_off = 0;
>> >        if (j == 0 && i >= 0
>> >  	  && lhs_ops[0].opcode == MEM_REF
>> > -	  && maybe_ne (lhs_ops[0].off, -1))
>> > +	  && known_ne (lhs_ops[0].off, -1))
>> >  	{
>> >  	  if (known_eq (lhs_ops[0].off, vr->operands[i].off))
>> >  	    i--, j--;
>> >  	  else if (vr->operands[i].opcode == MEM_REF
>> > -		   && maybe_ne (vr->operands[i].off, -1))
>> > +		   && known_ne (vr->operands[i].off, -1))
>> 
>> These changes don't look right.  If -1 is a special marker,
>> it should be tested with known_eq (positive) or maybe_ne (negative).
>> 
>> In other words, we should enter the if body if off is not the
>> compile-time constant -1.
>
> Hmm, to me 'known_ne' was visually more correct (only if 'off' is
> an actual offset we may treat it as such).

That's also the intention behind maybe_ne though.  The point is that
-1 isn't really a number on a scale [M, N] (M < -1, N > -1).  It's a
just a C way of representing "nothing" in an "X or nothing" (std::optional).

So whether an offset happens to be numerically equal to -1 at runtime
isn't relevant.  A runtime off is an X in the "X or nothing" and so
needs to be treated in the same was as other Xs.

To put it another way: known_ne is harder to prove than maybe_ne.
known_ne must return false if it can't prove the arguments are
different for all possible combinations of indeterminates.
maybe_ne is instead the direct opposite of known_eq.

Another analogy might be: suppose that we used a -1 INTEGER_CST
as a special marker.  (Not a good choice, but bear with me.)
If we wanted to test whether a tree was this special marker,
we'd use integer_minus_onep.  To test that it isn't the marker
we'd use !integer_minus_onep.  known_ne would instead be the
equivalent of:

   integer_zerop (fold_build2 (boolean_type_node, NE_EXPR,
                               x, ...-1 node...))

and so a general x would be treated in the same way as -1.

> Yes, -1 is a special
> marker but still.  Practically of course
> known_ne (poly, integer constant) == maybe_ne (poly, integer_constant)?

Not in general.  E.g. if VL is the SVE vector length in bytes:

  maybe_ne (VL, 16)

is true (VL can be [1,16]*16) but:

  known_ne (VL, 16)

is false (VL might be 16 but might not).

> That is, if we'd have a poly-int that might be -1 we should still
> not use that as 'off'?

known_ne might be safe in this particular instance if no runtime offset
would ever evaluate to -1.  But IMO it's harder to reason about and
less obviously safe.

Thanks,
Richard

> I can of course revert that hunk.
>
> Thanks,
> Richard.
>
>> Thanks,
>> Richard
>> 
>> >  	    {
>> >  	      extra_off = vr->operands[i].off - lhs_ops[0].off;
>> >  	      i--, j--;
>> > @@ -3275,6 +3275,7 @@ vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
>> >        copy_reference_ops_from_ref (rhs1, &rhs);
>> >  
>> >        /* Apply an extra offset to the inner MEM_REF of the RHS.  */
>> > +      bool force_no_tbaa = false;
>> >        if (maybe_ne (extra_off, 0))
>> >  	{
>> >  	  if (rhs.length () < 2)
>> > @@ -3287,6 +3288,10 @@ vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
>> >  	  rhs[ix].op0 = int_const_binop (PLUS_EXPR, rhs[ix].op0,
>> >  					 build_int_cst (TREE_TYPE (rhs[ix].op0),
>> >  							extra_off));
>> > +	  /* When we have offsetted the RHS, reading only parts of it,
>> > +	     we can no longer use the original TBAA type, force alias-set
>> > +	     zero.  */
>> > +	  force_no_tbaa = true;
>> >  	}
>> >  
>> >        /* Save the operands since we need to use the original ones for
>> > @@ -3339,8 +3344,11 @@ vn_reference_lookup_3 (ao_ref *ref, tree vuse, void *data_,
>> >        /* Adjust *ref from the new operands.  */
>> >        ao_ref rhs1_ref;
>> >        ao_ref_init (&rhs1_ref, rhs1);
>> > -      if (!ao_ref_init_from_vn_reference (&r, ao_ref_alias_set (&rhs1_ref),
>> > -					  ao_ref_base_alias_set (&rhs1_ref),
>> > +      if (!ao_ref_init_from_vn_reference (&r,
>> > +					  force_no_tbaa ? 0
>> > +					  : ao_ref_alias_set (&rhs1_ref),
>> > +					  force_no_tbaa ? 0
>> > +					  : ao_ref_base_alias_set (&rhs1_ref),
>> >  					  vr->type, vr->operands))
>> >  	return (void *)-1;
>> >        /* This can happen with bitfields.  */
>>