Re: [PATCH, rs6000] Gimple folding for vec_madd()

[Richard Biener <richard dot guenther at gmail dot com>]

On Thu, Oct 26, 2017 at 5:13 PM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Thu, Oct 26, 2017 at 4:30 PM, Will Schmidt <will_schmidt@vnet.ibm.com> wrote:
>> On Thu, 2017-10-26 at 11:05 +0200, Richard Biener wrote:
>>> On Wed, Oct 25, 2017 at 4:38 PM, Will Schmidt <will_schmidt@vnet.ibm.com> wrote:
>>> > Hi,
>>> >
>>> > Add support for gimple folding of the vec_madd() (vector multiply-add)
>>> > intrinsics.
>>> > Testcase coverage is provided by the existing tests
>>> >  gcc.target/powerpc/fold-vec-madd-*.c
>>> >
>>> > Sniff-tests appear clean.  A full regtest is currently running across assorted Power systems. (P6-P9).
>>> > OK for trunk (pending clean run results)?
>>>
>>> You can use FMA_EXPR on integer operands as well.  Otherwise you risk
>>> the FMA be not matched by combine later when part of the operation is
>>> CSEd.
>>
>> I had tried that initially, without success,..   I'll probably need
>> another hint.  :-)
>> Looking a bit closer, I think I see why the assert fired, but I'm not
>> sure what the proper fix would be.
>>
>> So attempting to the FMA_EXPR on the integer operands. (vector shorts in
>> this case), I end up triggering this error:
>>
>> /home/willschm/gcc/gcc-mainline-vec_fold_misc/gcc/testsuite/gcc.target/powerpc/fold-vec-madd-short.c:14:10: internal compiler error: in expand_expr_real_2, at expr.c:8712
>> 0x10813303 expand_expr_real_2(separate_ops*, rtx_def*, machine_mode, expand_modifier)
>>         /home/willschm/gcc/gcc-mainline-vec_fold_misc/gcc/expr.c:8712
>> 0x1081822f expand_expr_real_1(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool)
>>         /home/willschm/gcc/gcc-mainline-vec_fold_misc/gcc/expr.c:9787
>> 0x1080f7bb expand_expr_real(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool)
>>         /home/willschm/gcc/gcc-mainline-vec_fold_misc/gcc/expr.c:8084
>> ...
>>
>>
>> which when followed back, I tripped an assert here:  (gcc/expr.c:
>> expand_expr_real_2() ~ line 8710)
>>
>>     case FMA_EXPR:
>>       {
>>         optab opt = fma_optab;
>>         gimple *def0, *def2;
>>         if (optab_handler (fma_optab, mode) == CODE_FOR_nothing)
>>           {
>>             tree fn = mathfn_built_in (TREE_TYPE (treeop0), BUILT_IN_FMA);
>>             tree call_expr;
>>
>>             gcc_assert (fn != NULL_TREE);
>>
>> where gcc/builtins.c
>> mathfn_built_in()->mathfn_built_in_1->mathfn_built_in_2 looks to have
>> returned END_BUILTINS/NULL_TREE, due to falling through the if/else
>> tree:
>>
>>   if (TYPE_MAIN_VARIANT (type) == double_type_node)
>>     return fcode;
>>   else if (TYPE_MAIN_VARIANT (type) == float_type_node)
>>     return fcodef;
>>   else if (TYPE_MAIN_VARIANT (type) == long_double_type_node)
>>     return fcodel;
>>   else
>>     return END_BUILTINS;
>>
>> Looks like that is all double/float/long double contents.  First blush
>> attempt would be to add V8HI_type_node/integer_type_node to that if/else
>> tree, but that doesn't look like it would be near enough.
>
> Well - we of course expect to have an optab for the fma with vector
> short.  I thought
> you had one given you have the intrinsic.  If you don't have an optab
> you of course
> have to open-code it.
>
> Just thought you expected an actual machine instruction doing the integer FMA.

So you have

(define_insn "altivec_vmladduhm"
  [(set (match_operand:V8HI 0 "register_operand" "=v")
        (plus:V8HI (mult:V8HI (match_operand:V8HI 1 "register_operand" "v")
                              (match_operand:V8HI 2 "register_operand" "v"))
                   (match_operand:V8HI 3 "register_operand" "v")))]
  "TARGET_ALTIVEC"
  "vmladduhm %0,%1,%2,%3"
  [(set_attr "type" "veccomplex")])

but not

(define_expand "fmav8hi4"
...

or define_insn in case that's also a way to register an optab.

Richard.



> Richard.
>
>> Thanks
>> -Will
>>
>>>
>>> Richard.
>>>
>>> > Thanks,
>>> > -Will
>>> >
>>> > [gcc]
>>> >
>>> > 2017-10-25  Will Schmidt <will_schmidt@vnet.ibm.com>
>>> >
>>> >         * config/rs6000/rs6000.c: (rs6000_gimple_fold_builtin) Add support for
>>> >           gimple folding of vec_madd() intrinsics.
>>> >
>>> > diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
>>> > index 4837e14..04c2b15 100644
>>> > --- a/gcc/config/rs6000/rs6000.c
>>> > +++ b/gcc/config/rs6000/rs6000.c
>>> > @@ -16606,10 +16606,43 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator *gsi)
>>> >                                            build_int_cst (arg2_type, 0)), arg0);
>>> >          gimple_set_location (g, loc);
>>> >          gsi_replace (gsi, g, true);
>>> >          return true;
>>> >        }
>>> > +
>>> > +    /* vec_madd (Float) */
>>> > +    case ALTIVEC_BUILTIN_VMADDFP:
>>> > +    case VSX_BUILTIN_XVMADDDP:
>>> > +      {
>>> > +       arg0 = gimple_call_arg (stmt, 0);
>>> > +       arg1 = gimple_call_arg (stmt, 1);
>>> > +       tree arg2 = gimple_call_arg (stmt, 2);
>>> > +       lhs = gimple_call_lhs (stmt);
>>> > +       gimple *g = gimple_build_assign (lhs, FMA_EXPR , arg0, arg1, arg2);
>>> > +       gimple_set_location (g, gimple_location (stmt));
>>> > +       gsi_replace (gsi, g, true);
>>> > +       return true;
>>> > +      }
>>> > +    /* vec_madd (Integral) */
>>> > +    case ALTIVEC_BUILTIN_VMLADDUHM:
>>> > +      {
>>> > +       arg0 = gimple_call_arg (stmt, 0);
>>> > +       arg1 = gimple_call_arg (stmt, 1);
>>> > +       tree arg2 = gimple_call_arg (stmt, 2);
>>> > +       lhs = gimple_call_lhs (stmt);
>>> > +       tree lhs_type = TREE_TYPE (lhs);
>>> > +       location_t loc = gimple_location (stmt);
>>> > +       gimple_seq stmts = NULL;
>>> > +       tree mult_result = gimple_build (&stmts, loc, MULT_EXPR,
>>> > +                                      lhs_type, arg0, arg1);
>>> > +       tree plus_result = gimple_build (&stmts, loc, PLUS_EXPR,
>>> > +                                      lhs_type, mult_result, arg2);
>>> > +       gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
>>> > +       update_call_from_tree (gsi, plus_result);
>>> > +       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;
>>> >
>>> >
>>>
>>
>>