[2/2] Fix bogus inner induction (PR 86725)

This patch is the second part of the fix for PR 86725.  The problem
in the original test is that for:

  outer1:
    x_1 = PHI <x_4(outer2), ...>;
    ...

  inner:
    x_2 = PHI <x_1(outer1), x_3(...)>;
    ...
    x_3 = ...;
    ...

  outer2:
    x_4 = PHI <x_3(inner)>;
    ...

there are corner cases in which it is possible to classify the
inner phi as an induction but not the outer phi.  The -4.c test
is a more direct example.

After failing to classify x_1 as an induction, we go on to
classify it as a double reduction (which is basically true).
But we still classified the inner phi as an induction rather
than as part of a reduction, leading to an ICE when trying
to vectorise the outer phi.

We analyse the phis for outer loops first, so the simplest
fix is not to classify the phi as an induction if outer loop
analysis said that it should be a reduction.

The -2.c test is from the original PR.  The -3.c test is a
version in which "wo" really is used a reduction; this was
already correctly rejected, but for the wrong reason ("inner-loop
induction only used outside of the outer vectorized loop").
The -4.c test is another way of tickling the original problem
without relying on the undefinedness of signed overflow.
The -5.c test shows an (uninteresting) example in which the
patch prevents a spurious failure to vectorise the outer loop.

Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf
and x86_64-linux-gnu.  OK for trunk and GCC 8?

Richard


2018-08-22  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	PR tree-optimization/86725
	* tree-vect-loop.c (vect_inner_phi_in_double_reduction_p): New
	function.
	(vect_analyze_scalar_cycles_1): Check it.

gcc/testsuite/
	PR tree-optimization/86725
	* gcc.dg/vect/no-scevccp-pr86725-2.c: New test.
	* gcc.dg/vect/no-scevccp-pr86725-3.c: Likewise.
	* gcc.dg/vect/no-scevccp-pr86725-4.c: Likewise.
	* gcc.dg/vect/no-scevccp-pr86725-5.c: Likewise.

On Wed, Aug 22, 2018 at 11:34 AM Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> This patch is the second part of the fix for PR 86725.  The problem
> in the original test is that for:
>
>   outer1:
>     x_1 = PHI <x_4(outer2), ...>;
>     ...
>
>   inner:
>     x_2 = PHI <x_1(outer1), x_3(...)>;
>     ...
>     x_3 = ...;
>     ...
>
>   outer2:
>     x_4 = PHI <x_3(inner)>;
>     ...
>
> there are corner cases in which it is possible to classify the
> inner phi as an induction but not the outer phi.  The -4.c test
> is a more direct example.
>
> After failing to classify x_1 as an induction, we go on to
> classify it as a double reduction (which is basically true).
> But we still classified the inner phi as an induction rather
> than as part of a reduction, leading to an ICE when trying
> to vectorise the outer phi.
>
> We analyse the phis for outer loops first, so the simplest
> fix is not to classify the phi as an induction if outer loop
> analysis said that it should be a reduction.
>
> The -2.c test is from the original PR.  The -3.c test is a
> version in which "wo" really is used a reduction; this was
> already correctly rejected, but for the wrong reason ("inner-loop
> induction only used outside of the outer vectorized loop").
> The -4.c test is another way of tickling the original problem
> without relying on the undefinedness of signed overflow.
> The -5.c test shows an (uninteresting) example in which the
> patch prevents a spurious failure to vectorise the outer loop.
>
> Tested on aarch64-linux-gnu (with and without SVE), aarch64_be-elf
> and x86_64-linux-gnu.  OK for trunk and GCC 8?

OK.

Thanks,
Richard.

> Richard
>
>
> 2018-08-22  Richard Sandiford  <richard.sandiford@arm.com>
>
> gcc/
>         PR tree-optimization/86725
>         * tree-vect-loop.c (vect_inner_phi_in_double_reduction_p): New
>         function.
>         (vect_analyze_scalar_cycles_1): Check it.
>
> gcc/testsuite/
>         PR tree-optimization/86725
>         * gcc.dg/vect/no-scevccp-pr86725-2.c: New test.
>         * gcc.dg/vect/no-scevccp-pr86725-3.c: Likewise.
>         * gcc.dg/vect/no-scevccp-pr86725-4.c: Likewise.
>         * gcc.dg/vect/no-scevccp-pr86725-5.c: Likewise.
>
> Index: gcc/tree-vect-loop.c
> ===================================================================
> --- gcc/tree-vect-loop.c        2018-08-22 10:25:06.682099699 +0100
> +++ gcc/tree-vect-loop.c        2018-08-22 10:25:09.586074995 +0100
> @@ -462,6 +462,40 @@ vect_is_simple_iv_evolution (unsigned lo
>    return true;
>  }
>
> +/* Return true if PHI, described by STMT_INFO, is the inner PHI in
> +   what we are assuming is a double reduction.  For example, given
> +   a structure like this:
> +
> +      outer1:
> +       x_1 = PHI <x_4(outer2), ...>;
> +       ...
> +
> +      inner:
> +       x_2 = PHI <x_1(outer1), ...>;
> +       ...
> +       x_3 = ...;
> +       ...
> +
> +      outer2:
> +       x_4 = PHI <x_3(inner)>;
> +       ...
> +
> +   outer loop analysis would treat x_1 as a double reduction phi and
> +   this function would then return true for x_2.  */
> +
> +static bool
> +vect_inner_phi_in_double_reduction_p (stmt_vec_info stmt_info, gphi *phi)
> +{
> +  loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
> +  use_operand_p use_p;
> +  ssa_op_iter op_iter;
> +  FOR_EACH_PHI_ARG (use_p, phi, op_iter, SSA_OP_USE)
> +    if (stmt_vec_info def_info = loop_vinfo->lookup_def (USE_FROM_PTR (use_p)))
> +      if (STMT_VINFO_DEF_TYPE (def_info) == vect_double_reduction_def)
> +       return true;
> +  return false;
> +}
> +
>  /* Function vect_analyze_scalar_cycles_1.
>
>     Examine the cross iteration def-use cycles of scalar variables
> @@ -522,6 +556,7 @@ vect_analyze_scalar_cycles_1 (loop_vec_i
>         }
>
>        if (!access_fn
> +         || vect_inner_phi_in_double_reduction_p (stmt_vinfo, phi)
>           || !vect_is_simple_iv_evolution (loop->num, access_fn, &init, &step)
>           || (LOOP_VINFO_LOOP (loop_vinfo) != loop
>               && TREE_CODE (step) != INTEGER_CST))
> Index: gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-2.c
> ===================================================================
> --- /dev/null   2018-07-26 10:26:13.137955424 +0100
> +++ gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-2.c    2018-08-22 10:25:09.582075029 +0100
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O -w" } */
> +
> +int
> +nr (int xe)
> +{
> +  int oo, wo = 0;
> +
> +  for (oo = 0; oo < 4; ++oo)
> +    {
> +      int qq;
> +
> +      for (qq = 0; qq < 2; ++qq)
> +        {
> +          wo += 0x80000000;
> +          xe += wo;
> +        }
> +    }
> +  return xe;
> +}
> +
> +/* { dg-final { scan-tree-dump "reduction used in loop" "vect" { target vect_int } } } */
> +/* { dg-final { scan-tree-dump-not "OUTER LOOP VECTORIZED" "vect" } } */
> Index: gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-3.c
> ===================================================================
> --- /dev/null   2018-07-26 10:26:13.137955424 +0100
> +++ gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-3.c    2018-08-22 10:25:09.582075029 +0100
> @@ -0,0 +1,25 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O -w" } */
> +
> +int foo;
> +int
> +nr (int xe)
> +{
> +  int oo, wo = 0;
> +
> +  for (oo = 0; oo < 4; ++oo)
> +    {
> +      int qq;
> +
> +      for (qq = 0; qq < 2; ++qq)
> +        {
> +          wo += 0x80000000;
> +          xe += wo;
> +        }
> +    }
> +  foo = wo;
> +  return xe;
> +}
> +
> +/* { dg-final { scan-tree-dump "reduction used in loop" "vect" { target vect_int } } } */
> +/* { dg-final { scan-tree-dump-not "OUTER LOOP VECTORIZED" "vect" } } */
> Index: gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-4.c
> ===================================================================
> --- /dev/null   2018-07-26 10:26:13.137955424 +0100
> +++ gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-4.c    2018-08-22 10:25:09.582075029 +0100
> @@ -0,0 +1,23 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O -w" } */
> +
> +int
> +nr (unsigned int xe, unsigned int qqn)
> +{
> +  unsigned int oo, wo = 0;
> +
> +  for (oo = 0; oo < 4; ++oo)
> +    {
> +      unsigned int qq = qqn;
> +      do
> +        {
> +          wo += 1;
> +          xe += wo;
> +        }
> +      while (qq-- > 0);
> +    }
> +  return xe;
> +}
> +
> +/* { dg-final { scan-tree-dump "reduction used in loop" "vect" { target vect_int } } } */
> +/* { dg-final { scan-tree-dump-not "OUTER LOOP VECTORIZED" "vect" } } */
> Index: gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-5.c
> ===================================================================
> --- /dev/null   2018-07-26 10:26:13.137955424 +0100
> +++ gcc/testsuite/gcc.dg/vect/no-scevccp-pr86725-5.c    2018-08-22 10:25:09.582075029 +0100
> @@ -0,0 +1,24 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-O -w" } */
> +
> +unsigned int foo;
> +int
> +nr (unsigned int xe, unsigned int qqn)
> +{
> +  unsigned int oo, wo = 0;
> +
> +  for (oo = 0; oo < 4; ++oo)
> +    {
> +      unsigned int qq = qqn;
> +      do
> +        {
> +          wo += 1;
> +          xe += qq;
> +        }
> +      while (qq-- > 0);
> +    }
> +  foo = wo;
> +  return xe;
> +}
> +
> +/* { dg-final { scan-tree-dump "OUTER LOOP VECTORIZED" "vect" { target vect_int } } } */