This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH, rs6000] Gimple folding for vec_madd()
- From: Richard Biener <richard dot guenther at gmail dot com>
- To: will_schmidt at vnet dot ibm dot com
- Cc: GCC Patches <gcc-patches at gcc dot gnu dot org>, Segher Boessenkool <segher at kernel dot crashing dot org>, Bill Schmidt <wschmidt at linux dot vnet dot ibm dot com>, David Edelsohn <dje dot gcc at gmail dot com>
- Date: Thu, 26 Oct 2017 11:05:53 +0200
- Subject: Re: [PATCH, rs6000] Gimple folding for vec_madd()
- Authentication-results: sourceware.org; auth=none
- References: <1508942330.26707.184.camel@brimstone.rchland.ibm.com>
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.
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;
>
>