Re: [PATCH] Check the type of mask while generating cond_op in gimple simplication.

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

Hongtao Liu via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> On Wed, Sep 1, 2021 at 8:52 PM Richard Sandiford
> <richard.sandiford@arm.com> wrote:
>>
>> Richard Biener <richard.guenther@gmail.com> writes:
>> > On Wed, Sep 1, 2021 at 8:28 AM Hongtao Liu <crazylht@gmail.com> wrote:
>> >>
>> >> On Tue, Aug 31, 2021 at 7:56 PM Richard Biener
>> >> <richard.guenther@gmail.com> wrote:
>> >> >
>> >> > On Tue, Aug 31, 2021 at 12:18 PM Hongtao Liu <crazylht@gmail.com> wrote:
>> >> > >
>> >> > > On Mon, Aug 30, 2021 at 8:25 PM Richard Biener via Gcc-patches
>> >> > > <gcc-patches@gcc.gnu.org> wrote:
>> >> > > >
>> >> > > > On Fri, Aug 27, 2021 at 8:53 AM liuhongt <hongtao.liu@intel.com> wrote:
>> >> > > > >
>> >> > > > >   When gimple simplifcation try to combine op and vec_cond_expr to cond_op,
>> >> > > > > it doesn't check if mask type matches. It causes an ICE when expand cond_op
>> >> > > > > with mismatched mode.
>> >> > > > >   This patch add a function named cond_vectorized_internal_fn_supported_p
>> >> > > > >  to additionally check mask type than vectorized_internal_fn_supported_p.
>> >> > > > >
>> >> > > > >   Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
>> >> > > > >   Ok for trunk?
>> >> > > > >
>> >> > > > > gcc/ChangeLog:
>> >> > > > >
>> >> > > > >         PR middle-end/102080
>> >> > > > >         * internal-fn.c (cond_vectorized_internal_fn_supported_p): New functions.
>> >> > > > >         * internal-fn.h (cond_vectorized_internal_fn_supported_p): New declaration.
>> >> > > > >         * match.pd: Check the type of mask while generating cond_op in
>> >> > > > >         gimple simplication.
>> >> > > > >
>> >> > > > > gcc/testsuite/ChangeLog:
>> >> > > > >
>> >> > > > >         PR middle-end/102080
>> >> > > > >         * gcc.target/i386/pr102080.c: New test.
>> >> > > > > ---
>> >> > > > >  gcc/internal-fn.c                        | 22 ++++++++++++++++++++++
>> >> > > > >  gcc/internal-fn.h                        |  1 +
>> >> > > > >  gcc/match.pd                             | 24 ++++++++++++++++--------
>> >> > > > >  gcc/testsuite/gcc.target/i386/pr102080.c | 16 ++++++++++++++++
>> >> > > > >  4 files changed, 55 insertions(+), 8 deletions(-)
>> >> > > > >  create mode 100644 gcc/testsuite/gcc.target/i386/pr102080.c
>> >> > > > >
>> >> > > > > diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c
>> >> > > > > index 1360a00f0b9..8b2b65db1a7 100644
>> >> > > > > --- a/gcc/internal-fn.c
>> >> > > > > +++ b/gcc/internal-fn.c
>> >> > > > > @@ -4102,6 +4102,28 @@ expand_internal_call (gcall *stmt)
>> >> > > > >    expand_internal_call (gimple_call_internal_fn (stmt), stmt);
>> >> > > > >  }
>> >> > > > >
>> >> > > > > +/* Check cond_op for vector modes since vectorized_internal_fn_supported_p
>> >> > > > > +   doesn't check if mask type matches.  */
>> >> > > > > +bool
>> >> > > > > +cond_vectorized_internal_fn_supported_p (internal_fn ifn, tree type,
>> >> > > > > +                                        tree mask_type)
>> >> > > > > +{
>> >> > > > > +  if (!vectorized_internal_fn_supported_p (ifn, type))
>> >> > > > > +    return false;
>> >> > > > > +
>> >> > > > > +  machine_mode mask_mode;
>> >> > > > > +  machine_mode vmode = TYPE_MODE (type);
>> >> > > > > +  int size1, size2;
>> >> > > > > +  if (VECTOR_MODE_P (vmode)
>> >> > > > > +      && targetm.vectorize.get_mask_mode (vmode).exists(&mask_mode)
>> >> > > > > +      && GET_MODE_SIZE (mask_mode).is_constant (&size1)
>> >> > > > > +      && GET_MODE_SIZE (TYPE_MODE (mask_type)).is_constant (&size2)
>> >> > > > > +      && size1 != size2)
>> >> > > >
>> >> > > > Why do we check for equal size rather than just mode equality which
>> >> > > I originally thought  TYPE_MODE of vector(8) <signed-boolean:1> was
>> >> > > not QImode, Changed the patch to check mode equality.
>> >> > > Update patch.
>> >> >
>> >> > Looking at all this it seems the match.pd patterns should have not
>> >> > used vectorized_internal_fn_supported_p but direct_internal_fn_supported_p
>> >> > which is equivalent here because we're always working with vector modes?
>>
>> Yeah, looks like it.
>>
>> >> > And then shouldn't we look at the actual optab whether the mask mode matches
>> >> > the expectation rather than going around via the target hook which may not have
>> >> > enough context to decide which mask mode to use?
>> >> How about this?
>> >>
>> >> +/* Return true if target supports cond_op with data TYPE and
>> >> +   mask MASK_TYPE.  */
>> >> +bool
>> >> +cond_internal_fn_supported_p (internal_fn ifn, tree type,
>> >> +       tree mask_type)
>> >> +{
>> >> +  tree_pair types = tree_pair (type, type);
>> >> +  optab tmp = direct_internal_fn_optab (ifn, types);
>> >> +  machine_mode vmode = TYPE_MODE (type);
>> >> +  insn_code icode = direct_optab_handler (tmp, vmode);
>> >> +  if (icode == CODE_FOR_nothing)
>> >> +    return false;
>> >> +
>> >> +  machine_mode mask_mode = TYPE_MODE (mask_type);
>> >> +  /* Can't create rtx and use insn_operand_matches here.  */
>> >> +  return insn_data[icode].operand[0].mode == vmode
>> >> +    && insn_data[icode].operand[1].mode == mask_mode;
>> >> +}
>> >> +
>> >
>> > Yeah, sth like that, though the operand[0].mode test should be
>> > redudnant.  I think we should assert or have a whiltelist
>> > for the internal function we support to be queried this way.
>> > Not sure if we can directly access the 'cond_binary/cond_ternary'
>> > classification used in internal-fn.def, that would be best.
>> >
>> > Richard, what are your thoughts about all this?
>>
>> IMO using get_mask_mode was right.  The optab documentation says:
>>
>>   Operands 0, 2, 3 and 4 all have mode @var{m}.  Operand 1 is a scalar
>>   integer if @var{m} is scalar, otherwise it has the mode returned by
>>   @code{TARGET_VECTORIZE_GET_MASK_MODE}.
>>
>> Allowing targets to use optabs to enforce different mask modes for
>> different operations would open up a mess of combinations.
>>
>> In other words, I think cond_vectorized_internal_fn_supported_p
>> is really testing two things:
>>
>> (1) is the mask type/vector type combination well-formed?
>> (2) is the internal function supported for the vector type?
>>
>> where (1) is a gimple question and (2) is a target question.
>>
>> I guess there's an argument that (1) should be part of the match.pd
>> condition instead, alongside the element_precision check.  That would
>> add to the cut-&-paste though. :-(
>>
>> Alternatively, I guess we would add:
>>
>>   bool is_truth_type_for (tree type, tree truth_type);
>>
>> to return true if truth_type is equal to truth_type_for (type)
>> (but without having to call truth_type_for).  We could then use:
>>
>>   is_truth_type_for (op_type, TREE_TYPE (@0))
>>
>> instead of:
>>
>>   element_precision (type) == element_precision (op_type)
>>
>> since it should be a strictly stronger condition.
> Thanks for your advice, it sounds more reasonable.
> Here is the updated patch.
>>
>> Thanks,
>> Richard
>
>
>
> -- 
> BR,
> Hongtao
>
> From ae192d4a164b8d73adb06d6e28864f717741158c Mon Sep 17 00:00:00 2001
> From: liuhongt <hongtao.liu@intel.com>
> Date: Thu, 2 Sep 2021 13:05:54 +0800
> Subject: [PATCH v4] Check mask type when doing cond_op related gimple
>  simplification.
>
> gcc/ChangeLog:
>
> 	PR middle-end/102080
> 	* match.pd: Check mask type when doing cond_op related gimple
> 	simplification.
> 	* tree.c (is_truth_type_for): New function.
> 	* tree.h (is_truth_type_for): New declaration.
>
> gcc/testsuite/ChangeLog:
>
> 	PR middle-end/102080
> 	* gcc.target/i386/pr102080.c: New test.
> ---
>  gcc/match.pd                             |  8 ++++----
>  gcc/testsuite/gcc.target/i386/pr102080.c | 19 +++++++++++++++++++
>  gcc/tree.c                               |  7 +++++++
>  gcc/tree.h                               |  1 +
>  4 files changed, 31 insertions(+), 4 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.target/i386/pr102080.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index f421c74b62c..c9a27f46ed2 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -6988,13 +6988,13 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>    (vec_cond @0 (view_convert? (uncond_op@4 @1 @2)) @3)
>    (with { tree op_type = TREE_TYPE (@4); }
>     (if (vectorized_internal_fn_supported_p (as_internal_fn (cond_op), op_type)
> -	&& element_precision (type) == element_precision (op_type))
> +	&& is_truth_type_for (type, TREE_TYPE (@0)))

Why pass “type” rather than “op_type”?  “type” is the type of the
original expression, and if the original vec_cond is well-formed,
@0 should already have the right truth type for “type”.  Here we're
trying to convert the expression into a conditional operation on
“op_type”, so we need to test whether that's valid.

>      (view_convert (cond_op @0 @1 @2 (view_convert:op_type @3))))))
>   (simplify
>    (vec_cond @0 @1 (view_convert? (uncond_op@4 @2 @3)))
>    (with { tree op_type = TREE_TYPE (@4); }
>     (if (vectorized_internal_fn_supported_p (as_internal_fn (cond_op), op_type)
> -	&& element_precision (type) == element_precision (op_type))
> +	&& is_truth_type_for (type, TREE_TYPE (@0)))
>      (view_convert (cond_op (bit_not @0) @2 @3 (view_convert:op_type @1)))))))
>  
>  /* Same for ternary operations.  */
> @@ -7004,13 +7004,13 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>    (vec_cond @0 (view_convert? (uncond_op@5 @1 @2 @3)) @4)
>    (with { tree op_type = TREE_TYPE (@5); }
>     (if (vectorized_internal_fn_supported_p (as_internal_fn (cond_op), op_type)
> -	&& element_precision (type) == element_precision (op_type))
> +	&& is_truth_type_for (type, TREE_TYPE (@0)))
>      (view_convert (cond_op @0 @1 @2 @3 (view_convert:op_type @4))))))
>   (simplify
>    (vec_cond @0 @1 (view_convert? (uncond_op@5 @2 @3 @4)))
>    (with { tree op_type = TREE_TYPE (@5); }
>     (if (vectorized_internal_fn_supported_p (as_internal_fn (cond_op), op_type)
> -	&& element_precision (type) == element_precision (op_type))
> +	&& is_truth_type_for (type, TREE_TYPE (@0)))
>      (view_convert (cond_op (bit_not @0) @2 @3 @4
>  		  (view_convert:op_type @1)))))))
>  #endif
> diff --git a/gcc/testsuite/gcc.target/i386/pr102080.c b/gcc/testsuite/gcc.target/i386/pr102080.c
> new file mode 100644
> index 00000000000..4c5ee32ee63
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr102080.c
> @@ -0,0 +1,19 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */
> +
> +#include<immintrin.h>
> +typedef float __m256 __attribute__((__vector_size__(32)));
> +__m256 _mm256_blendv_ps___Y, _mm256_blendv_ps___M, _mm256_mul_ps___A,
> +  _mm256_mul_ps___B, IfThenElse___trans_tmp_9;
> +
> +void
> +__attribute__ ((target("avx")))
> +IfThenElse (__m256 no) {
> +  IfThenElse___trans_tmp_9 = _mm256_blendv_ps (no, _mm256_blendv_ps___Y, _mm256_blendv_ps___M);
> +}
> +void
> +__attribute__ ((target("avx512vl")))
> +EncodedFromDisplay() {
> +  __m256 __trans_tmp_11 = _mm256_mul_ps___A * _mm256_mul_ps___B;
> +  IfThenElse(__trans_tmp_11);
> +}
> diff --git a/gcc/tree.c b/gcc/tree.c
> index cba3bca41b3..88c2221eabb 100644
> --- a/gcc/tree.c
> +++ b/gcc/tree.c
> @@ -10723,6 +10723,13 @@ signed_type_for (tree type)
>    return signed_or_unsigned_type_for (0, type);
>  }
>  
> +bool
> +is_truth_type_for (tree type, tree truth_type)
> +{
> +  tree tmp = truth_type_for (type);
> +  return tmp == truth_type;
> +}

The idea was to try to avoid calling truth_type to create a type.
Instead we can use similar logic to truth_type to tell whether type
has the right form.  I think the rules are:

- For VECTOR_TYPEs:
  - The truth type must be a VECTOR_BOOLEAN_TYPE.
  - The number of elements must match (known_eq).
  - Also:
    - If !VECTOR_BOOLEAN_TYPE_P and targetm.vectorize.get_mask_mode
      exists, the truth type must have that mode.
    - Otherwise, the types must have the same size.
- Otherwise, the truth type must be a BOOLEAN_TYPE.

(Richi please correct me if I'm wrong.)

Thanks,
Richard


> +
>  /* If TYPE is a vector type, return a signed integer vector type with the
>     same width and number of subparts. Otherwise return boolean_type_node.  */
>  
> diff --git a/gcc/tree.h b/gcc/tree.h
> index 060a41f6991..c8542bfd476 100644
> --- a/gcc/tree.h
> +++ b/gcc/tree.h
> @@ -4556,6 +4556,7 @@ extern tree build_string_literal (unsigned, const char * = NULL,
>  extern tree signed_or_unsigned_type_for (int, tree);
>  extern tree signed_type_for (tree);
>  extern tree unsigned_type_for (tree);
> +bool is_truth_type_for (tree, tree);
>  extern tree truth_type_for (tree);
>  extern tree build_pointer_type_for_mode (tree, machine_mode, bool);
>  extern tree build_pointer_type (tree);