Re: [PATCH, rs6000] folding of vector stores in GIMPLE

[Richard Biener <richard.guenther@gmail.com>]

On Thu, Sep 21, 2017 at 10:56 PM, Will Schmidt
<will_schmidt@vnet.ibm.com> wrote:
> Hi,
>
> Folding of vector stores in GIMPLE.
>
> - Add code to handle gimple folding for the vec_st (vector store) builtins.
> - Remove the now obsoleted folding code for vec_st from rs6000-c-c.
>
> There are two spots that I could use some feedback on.
>
> First -
>  An early exit remains in place prevent folding of statements that do not
>  have a LHS.  To allow folding of the stores to get past the check, I have
>  added a helper function (rs6000_builtin_valid_without_lhs) that allows
>  those store intrinsics to proceed.  I'm not sure the approach (or the name I chose)
>  is the best choice, so I'll defer to recommendations on how to improve that. :-)
>
> Second -
>  This code (as-is) is subject to a TBAA related issue (similar to what was noticed
>  in the gimple folding code for loads.   As-is, with a testcase such as :
>
> void testst_struct1b (vector double vd1, long long ll1, struct S *p)
>     {
>         vec_st (vd1, ll1, (vector double *)p);
>     }
>
>  will generate gimple that looks like:
>     MEM[(struct S *)D.3218] = vd1;
>
> If I rework the code, setting arg2_type to be ptr_type_node, i.e.
> + tree arg2_type = TREE_TYPE (arg2);
> to:
> + tree arg2_type = ptr_type_node;
>
> the generated gimple then looks like
>   MEM[(void *)D.3218] = vd1;
>
> Which is probably OK, but I cannot say for certain.  The generated .s content is at least equivalent.

It looks safe.

The question I had is whether vec_st (vd1, ll1, (vector double *)p) is
equivalent
to *(vector double *)((char *)p + ll1) = vd1; from a TBAA perspective.  Thus
whether the type of the tird argument to vec_st defines the type of the access
(in C standards meaning).  If so then what we do now is pessimizing (but
as you say previously (long long *) and (long *) were aliased together and
you got wrong-code with aliasing with regular stores of the "wrong" same type).

A proper fix would be to transition this type as seen from the frontend to
GIMPLE, for example in a similar way we do for MEM_REFs by piggy-backing
that on an extra argument, a constant zero pointer of the alias
pointer type to use
(which would also serve as a type indicator of the store itself).  I'd use a
target specific internal function for this (not sure if we can have those target
specific, but I guess if it's folded away then that's fine) and get away with
the overload set.

Richard.

> The resulting code is verified by testcases powerpc/fold-vec-st-*.c, which
> has been posted separately.
>
> regtest looks clean on power6 and newer.
>
> pending feedback, OK for trunk?
>
> Thanks,
> -Will
>
>     [gcc]
>
>     2017-09-21  Will Schmidt  <will_schmidt@vnet.ibm.com>
>
>         * config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add handling
>         for early folding of vector stores (ALTIVEC_BUILTIN_ST_*).
>         (rs6000_builtin_valid_without_lhs): helper function.
>         * config/rs6000/rs6000-c.c (altivec_resolve_overloaded_builtin):
>         Remove obsoleted code for handling ALTIVEC_BUILTIN_VEC_ST.
>
> diff --git a/gcc/config/rs6000/rs6000-c.c b/gcc/config/rs6000/rs6000-c.c
> index a49db97..4a363a1 100644
> --- a/gcc/config/rs6000/rs6000-c.c
> +++ b/gcc/config/rs6000/rs6000-c.c
> @@ -6470,82 +6470,10 @@ altivec_resolve_overloaded_builtin (location_t loc, tree fndecl,
>                      convert (TREE_TYPE (stmt), arg0));
>        stmt = build2 (COMPOUND_EXPR, arg1_type, stmt, decl);
>        return stmt;
>      }
>
> -  /* Expand vec_st into an expression that masks the address and
> -     performs the store.  We need to expand this early to allow
> -     the best aliasing, as by the time we get into RTL we no longer
> -     are able to honor __restrict__, for example.  We may want to
> -     consider this for all memory access built-ins.
> -
> -     When -maltivec=be is specified, or the wrong number of arguments
> -     is provided, simply punt to existing built-in processing.  */
> -
> -  if (fcode == ALTIVEC_BUILTIN_VEC_ST
> -      && (BYTES_BIG_ENDIAN || !VECTOR_ELT_ORDER_BIG)
> -      && nargs == 3)
> -    {
> -      tree arg0 = (*arglist)[0];
> -      tree arg1 = (*arglist)[1];
> -      tree arg2 = (*arglist)[2];
> -
> -      /* Construct the masked address.  Let existing error handling take
> -        over if we don't have a constant offset.  */
> -      arg1 = fold (arg1);
> -
> -      if (TREE_CODE (arg1) == INTEGER_CST)
> -       {
> -         if (!ptrofftype_p (TREE_TYPE (arg1)))
> -           arg1 = build1 (NOP_EXPR, sizetype, arg1);
> -
> -         tree arg2_type = TREE_TYPE (arg2);
> -         if (TREE_CODE (arg2_type) == ARRAY_TYPE && c_dialect_cxx ())
> -           {
> -             /* Force array-to-pointer decay for C++.  */
> -             arg2 = default_conversion (arg2);
> -             arg2_type = TREE_TYPE (arg2);
> -           }
> -
> -         /* Find the built-in to make sure a compatible one exists; if not
> -            we fall back to default handling to get the error message.  */
> -         for (desc = altivec_overloaded_builtins;
> -              desc->code && desc->code != fcode; desc++)
> -           continue;
> -
> -         for (; desc->code == fcode; desc++)
> -           if (rs6000_builtin_type_compatible (TREE_TYPE (arg0), desc->op1)
> -               && rs6000_builtin_type_compatible (TREE_TYPE (arg1), desc->op2)
> -               && rs6000_builtin_type_compatible (TREE_TYPE (arg2),
> -                                                  desc->op3))
> -             {
> -               tree addr = fold_build2_loc (loc, POINTER_PLUS_EXPR, arg2_type,
> -                                            arg2, arg1);
> -               tree aligned
> -                 = fold_build2_loc (loc, BIT_AND_EXPR, arg2_type,
> -                                    addr, build_int_cst (arg2_type, -16));
> -
> -               tree arg0_type = TREE_TYPE (arg0);
> -               if (TYPE_MODE (arg0_type) == V2DImode)
> -                 /* Type-based aliasing analysis thinks vector long
> -                    and vector long long are different and will put them
> -                    in distinct alias classes.  Force our address type
> -                    to be a may-alias type to avoid this.  */
> -                 arg0_type
> -                   = build_pointer_type_for_mode (arg0_type, Pmode,
> -                                                  true/*can_alias_all*/);
> -               else
> -                 arg0_type = build_pointer_type (arg0_type);
> -               aligned = build1 (NOP_EXPR, arg0_type, aligned);
> -               tree stg = build_indirect_ref (loc, aligned, RO_NULL);
> -               tree retval = build2 (MODIFY_EXPR, TREE_TYPE (stg), stg,
> -                                     convert (TREE_TYPE (stg), arg0));
> -               return retval;
> -             }
> -       }
> -    }
> -
>    for (n = 0;
>         !VOID_TYPE_P (TREE_VALUE (fnargs)) && n < nargs;
>         fnargs = TREE_CHAIN (fnargs), n++)
>      {
>        tree decl_type = TREE_VALUE (fnargs);
> diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
> index 1978634..ef41534 100644
> --- a/gcc/config/rs6000/rs6000.c
> +++ b/gcc/config/rs6000/rs6000.c
> @@ -16155,10 +16155,29 @@ rs6000_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED,
>  #else
>    return NULL_TREE;
>  #endif
>  }
>
> +/*  Helper function to sort out which built-ins may be valid without having
> +    a LHS.  */
> +bool
> +rs6000_builtin_valid_without_lhs (enum rs6000_builtins fn_code)
> +{
> +   switch (fn_code)
> +      {
> +       case ALTIVEC_BUILTIN_STVX_V16QI:
> +       case ALTIVEC_BUILTIN_STVX_V8HI:
> +       case ALTIVEC_BUILTIN_STVX_V4SI:
> +       case ALTIVEC_BUILTIN_STVX_V4SF:
> +       case ALTIVEC_BUILTIN_STVX_V2DI:
> +       case ALTIVEC_BUILTIN_STVX_V2DF:
> +         return true;
> +       default:
> +         return false;
> +      }
> +}
> +
>  /* Fold a machine-dependent built-in in GIMPLE.  (For folding into
>     a constant, use rs6000_fold_builtin.)  */
>
>  bool
>  rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
> @@ -16182,12 +16201,13 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
>                fn_code, fn_name1, fn_name2);
>
>    if (!rs6000_fold_gimple)
>      return false;
>
> -  /* Generic solution to prevent gimple folding of code without a LHS.  */
> -  if (!gimple_call_lhs (stmt))
> +  /* Prevent gimple folding for code that does not have a LHS, unless it is
> +   allowed per the rs6000_builtin_valid_without_lhs helper function.  */
> +  if (!gimple_call_lhs (stmt) && !rs6000_builtin_valid_without_lhs (fn_code))
>      return false;
>
>    switch (fn_code)
>      {
>      /* Flavors of vec_add.  We deliberately don't expand
> @@ -16585,11 +16605,48 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
>                                                 build_int_cst (arg1_type, 0)));
>          gimple_set_location (g, loc);
>          gsi_replace (gsi, g, true);
>          return true;
>        }
> -
> +    /* Vector stores.  */
> +    case ALTIVEC_BUILTIN_STVX_V16QI:
> +    case ALTIVEC_BUILTIN_STVX_V8HI:
> +    case ALTIVEC_BUILTIN_STVX_V4SI:
> +    case ALTIVEC_BUILTIN_STVX_V4SF:
> +    case ALTIVEC_BUILTIN_STVX_V2DI:
> +    case ALTIVEC_BUILTIN_STVX_V2DF:
> +      {
> +        /* Do not fold for -maltivec=be on LE targets.  */
> +        if (VECTOR_ELT_ORDER_BIG && !BYTES_BIG_ENDIAN)
> +           return false;
> +        arg0 = gimple_call_arg (stmt, 0); /* Value to be stored.  */
> +        arg1 = gimple_call_arg (stmt, 1); /* Offset.  */
> +        tree arg2 = gimple_call_arg (stmt, 2); /* Store-to address.  */
> +        location_t loc = gimple_location (stmt);
> +        tree arg0_type = TREE_TYPE (arg0);
> +        tree arg2_type = TREE_TYPE (arg2);
> +        /* POINTER_PLUS_EXPR wants the offset to be of type 'sizetype'.  Create
> +           the tree using the value from arg0.  The resulting type will match
> +           the type of arg2.  */
> +        gimple_seq stmts = NULL;
> +        tree temp_offset = gimple_convert (&stmts, loc, sizetype, arg1);
> +        tree temp_addr = gimple_build (&stmts, loc, POINTER_PLUS_EXPR,
> +                                      arg2_type, arg2, temp_offset);
> +        /* Mask off any lower bits from the address.  */
> +        tree aligned_addr = gimple_build (&stmts, loc, BIT_AND_EXPR,
> +                                         arg2_type, temp_addr,
> +                                         build_int_cst (arg2_type, -16));
> +        gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
> +       /* The desired gimple result should be similar to:
> +        MEM[(__vector floatD.1407 *)_1] = vf1D.2697;  */
> +        gimple *g;
> +        g = gimple_build_assign (build2 (MEM_REF, arg0_type, aligned_addr,
> +                                          build_int_cst (arg2_type, 0)), arg0);
> +        gimple_set_location (g, loc);
> +        gsi_replace (gsi, g, true);
> +        return true;
> +      }
>      default:
>         if (TARGET_DEBUG_BUILTIN)
>            fprintf (stderr, "gimple builtin intrinsic not matched:%d %s %s\n",
>                     fn_code, fn_name1, fn_name2);
>        break;
>
>