On Fri, 8 Oct 2021 at 21:19, Richard Sandiford
<richard.sandif...@arm.com> wrote:
>
> Thanks for looking at this.
>
> Prathamesh Kulkarni <prathamesh.kulka...@linaro.org> writes:
> > Hi,
> > As mentioned in PR, for the following test-case:
> >
> > typedef unsigned char uint8_t;
> >
> > static inline uint8_t
> > x264_clip_uint8(uint8_t x)
> > {
> >   uint8_t t = -x;
> >   uint8_t t1 = x & ~63;
> >   return (t1 != 0) ? t : x;
> > }
> >
> > void
> > mc_weight(uint8_t *restrict dst, uint8_t *restrict src, int n)
> > {
> >   for (int x = 0; x < n*16; x++)
> >     dst[x] = x264_clip_uint8(src[x]);
> > }
> >
> > -O3 -mcpu=generic+sve generates following code for the inner loop:
> >
> > .L3:
> >         ld1b    z0.b, p0/z, [x1, x2]
> >         movprfx z2, z0
> >         and     z2.b, z2.b, #0xc0
> >         movprfx z1, z0
> >         neg     z1.b, p1/m, z0.b
> >         cmpeq   p2.b, p1/z, z2.b, #0
> >         sel     z0.b, p2, z0.b, z1.b
> >         st1b    z0.b, p0, [x0, x2]
> >         add     x2, x2, x4
> >         whilelo p0.b, w2, w3
> >         b.any   .L3
> >
> > The sel is redundant since we could conditionally negate z0 based on
> > the predicate
> > comparing z2 with 0.
> >
> > As suggested in the PR, the attached patch, introduces a new
> > conditional internal function .COND_NEG, and in gimple-isel replaces
> > the following sequence:
> >    op2 = -op1
> >    op0 = A cmp B
> >    lhs = op0 ? op1 : op2
> >
> > with:
> >    op0 = A inverted_cmp B
> >    lhs = .COND_NEG (op0, op1, op1).
> >
> > lhs = .COD_NEG (op0, op1, op1)
> > implies
> > lhs = neg (op1) if cond is true OR fall back to op1 if cond is false.
> >
> > With patch, it generates the following code-gen:
> > .L3:
> >         ld1b    z0.b, p0/z, [x1, x2]
> >         movprfx z1, z0
> >         and     z1.b, z1.b, #0xc0
> >         cmpne   p1.b, p2/z, z1.b, #0
> >         neg     z0.b, p1/m, z0.b
> >         st1b    z0.b, p0, [x0, x2]
> >         add     x2, x2, x4
> >         whilelo p0.b, w2, w3
> >         b.any   .L3
> >
> > While it seems to work for this test-case, I am not entirely sure if
> > the patch is correct. Does it look in the right direction ?
>
> For binary ops we use match.pd rather than isel:
>
> (for uncond_op (UNCOND_BINARY)
>      cond_op (COND_BINARY)
>  (simplify
>   (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)
>         && is_truth_type_for (op_type, TREE_TYPE (@0)))
>     (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)
>         && is_truth_type_for (op_type, TREE_TYPE (@0)))
>     (view_convert (cond_op (bit_not @0) @2 @3 (view_convert:op_type @1)))))))
>
> I think it'd be good to do the same here, using new (UN)COND_UNARY
> iterators.  (The iterators will only have one value to start with,
> but other unary ops could get the same treatment in future.)
Thanks for the suggestions.
The attached patch adds a pattern to match.pd to replace:
cond = a cmp b
r = cond ? x : -x
with:
cond = a inverted_cmp b
r = cond ? -x : x

Code-gen with patch for inner loop:
.L3:
        ld1b    z0.b, p0/z, [x1, x2]
        movprfx z1, z0
        and     z1.b, z1.b, #0xc0
        cmpne   p1.b, p2/z, z1.b, #0
        neg     z0.b, p1/m, z0.b
        st1b    z0.b, p0, [x0, x2]
        add     x2, x2, x4
        whilelo p0.b, w2, w3
        b.any   .L3

Does it look OK ?
I didn't add it under (UN)COND_UNARY since it inverts the comparison,
which we might not want to do for other unary ops ?

Also, I am not sure, how to test if target supports conditional
internal function ?
I tried to use:
(for cmp (tcc_comparison)
     icmp (inverted_tcc_comparison)
 (simplify
  (vec_cond (cmp@2 @0 @1) @3 (negate @3))
   (with { auto op_type = TREE_TYPE (@2); }
    (if (vectorized_internal_fn_supported_p (IFN_COND_NEG, op_type)
         && is_truth_type_for (op_type, TREE_TYPE (@0)))
      (IFN_COND_NEG (icmp:op_type @0 @1) @3 @3)))))

but both the conditions seem to fail.

Thanks,
Prathamesh


>
> Richard
>
>
> >
> > Thanks,
> > Prathamesh
> >
> > diff --git a/gcc/gimple-isel.cc b/gcc/gimple-isel.cc
> > index 38e90933c3e..5b0dd3c1993 100644
> > --- a/gcc/gimple-isel.cc
> > +++ b/gcc/gimple-isel.cc
> > @@ -39,6 +39,8 @@ along with GCC; see the file COPYING3.  If not see
> >  #include "optabs.h"
> >  #include "gimple-fold.h"
> >  #include "internal-fn.h"
> > +#include "fold-const.h"
> > +#include "tree-pretty-print.h"
> >
> >  /* Expand all ARRAY_REF(VIEW_CONVERT_EXPR) gimple assignments into calls to
> >     internal function based on vector type of selected expansion.
> > @@ -203,6 +205,35 @@ gimple_expand_vec_cond_expr (gimple_stmt_iterator *gsi,
> >             return new_stmt;
> >           }
> >
> > +       /* Replace:
> > +          op2 = -op1
> > +          op0 = A cmp B
> > +          lhs = op0 ? op1 : op2
> > +
> > +          with:
> > +          op0 = A inverted_cmp B
> > +          lhs = .COND_NEG (op0, op1, op1).  */
> > +
> > +       gassign *op1_def = nullptr;
> > +       if (TREE_CODE (op1) == SSA_NAME)
> > +         op1_def = static_cast<gassign *> (SSA_NAME_DEF_STMT (op1));
> > +
> > +       gassign *op2_def = nullptr;
> > +       if (TREE_CODE (op2) == SSA_NAME)
> > +         op2_def = static_cast<gassign *> (SSA_NAME_DEF_STMT (op2));
> > +
> > +       if (can_compute_op0 && op1_def && op2_def
> > +           && gimple_assign_rhs_code (op2_def) == NEGATE_EXPR
> > +           && operand_equal_p (gimple_assign_rhs1 (op2_def), op1, 0))
> > +         {
> > +           auto inverted_code
> > +             = invert_tree_comparison (gimple_assign_rhs_code (def_stmt), 
> > true);
> > +           gimple_assign_set_rhs_code (def_stmt, inverted_code);
> > +           auto gsi2 = gsi_for_stmt (op2_def);
> > +           gsi_remove (&gsi2, true);
> > +           return gimple_build_call_internal (IFN_COND_NEG, 3, op0, op1, 
> > op1);
> > +         }
> > +
> >         if (can_compute_op0
> >             && used_vec_cond_exprs >= 2
> >             && (get_vcond_mask_icode (mode, TYPE_MODE (op0_type))
> > diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c
> > index 78db25bbac4..b57c7a4ed3e 100644
> > --- a/gcc/internal-fn.c
> > +++ b/gcc/internal-fn.c
> > @@ -3877,7 +3877,8 @@ static void (*const internal_fn_expanders[]) 
> > (internal_fn, gcall *) = {
> >    T (BIT_IOR_EXPR, IFN_COND_IOR) \
> >    T (BIT_XOR_EXPR, IFN_COND_XOR) \
> >    T (LSHIFT_EXPR, IFN_COND_SHL) \
> > -  T (RSHIFT_EXPR, IFN_COND_SHR)
> > +  T (RSHIFT_EXPR, IFN_COND_SHR) \
> > +  T (NEGATE_EXPR, IFN_COND_NEG)
> >
> >  /* Return a function that only performs CODE when a certain condition is 
> > met
> >     and that uses a given fallback value otherwise.  For example, if CODE is
> > diff --git a/gcc/internal-fn.def b/gcc/internal-fn.def
> > index 88169ef4656..db40d1bfd18 100644
> > --- a/gcc/internal-fn.def
> > +++ b/gcc/internal-fn.def
> > @@ -204,6 +204,8 @@ DEF_INTERNAL_OPTAB_FN (COND_FMS, ECF_CONST, cond_fms, 
> > cond_ternary)
> >  DEF_INTERNAL_OPTAB_FN (COND_FNMA, ECF_CONST, cond_fnma, cond_ternary)
> >  DEF_INTERNAL_OPTAB_FN (COND_FNMS, ECF_CONST, cond_fnms, cond_ternary)
> >
> > +DEF_INTERNAL_OPTAB_FN (COND_NEG, ECF_CONST | ECF_NOTHROW, cond_neg, 
> > cond_unary)
> > +
> >  DEF_INTERNAL_OPTAB_FN (RSQRT, ECF_CONST, rsqrt, unary)
> >
> >  DEF_INTERNAL_OPTAB_FN (REDUC_PLUS, ECF_CONST | ECF_NOTHROW,
> > diff --git a/gcc/optabs.def b/gcc/optabs.def
> > index 201b8aae1c0..fc65a3a7c23 100644
> > --- a/gcc/optabs.def
> > +++ b/gcc/optabs.def
> > @@ -245,6 +245,7 @@ OPTAB_D (cond_fma_optab, "cond_fma$a")
> >  OPTAB_D (cond_fms_optab, "cond_fms$a")
> >  OPTAB_D (cond_fnma_optab, "cond_fnma$a")
> >  OPTAB_D (cond_fnms_optab, "cond_fnms$a")
> > +OPTAB_D (cond_neg_optab, "cond_neg$a")
> >  OPTAB_D (cmov_optab, "cmov$a6")
> >  OPTAB_D (cstore_optab, "cstore$a4")
> >  OPTAB_D (ctrap_optab, "ctrap$a4")

Attachment: pr93183-2.diff
Description: Binary data

Reply via email to