Replace FMA_EXPR with one internal fn per optab

There are four optabs for various forms of fused multiply-add:
fma, fms, fnma and fnms.  Of these, only fma had a direct gimple
representation.  For the other three we relied on special pattern-
matching during expand, although tree-ssa-math-opts.c did have
some code to try to second-guess what expand would do.

This patch removes the old FMA_EXPR representation of fma and
introduces four new internal functions, one for each optab.
IFN_FMA is tied to BUILT_IN_FMA* while the other three are
independent directly-mapped internal functions.  It's then
possible to do the pattern-matching in match.pd and
tree-ssa-math-opts.c (via folding) can select the exact
FMA-based operation.

The BRIG & HSA parts are a best guess, but seem relatively simple.

2018-05-18  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* doc/sourcebuild.texi (scalar_all_fma): Document.
	* tree.def (FMA_EXPR): Delete.
	* internal-fn.def (FMA, FMS, FNMA, FNMS): New internal functions.
	* internal-fn.c (ternary_direct): New macro.
	(expand_ternary_optab_fn): Likewise.
	(direct_ternary_optab_supported_p): Likewise.
	* Makefile.in (build/genmatch.o): Depend on case-fn-macros.h.
	* builtins.c (fold_builtin_fma): Delete.
	(fold_builtin_3): Don't call it.
	* cfgexpand.c (expand_debug_expr): Remove FMA_EXPR handling.
	* expr.c (expand_expr_real_2): Likewise.
	* fold-const.c (operand_equal_p): Likewise.
	(fold_ternary_loc): Likewise.
	* gimple-pretty-print.c (dump_ternary_rhs): Likewise.
	* gimple.c (DEFTREECODE): Likewise.
	* gimplify.c (gimplify_expr): Likewise.
	* optabs-tree.c (optab_for_tree_code): Likewise.
	* tree-cfg.c (verify_gimple_assign_ternary): Likewise.
	* tree-eh.c (operation_could_trap_p): Likewise.
	(stmt_could_throw_1_p): Likewise.
	* tree-inline.c (estimate_operator_cost): Likewise.
	* tree-pretty-print.c (dump_generic_node): Likewise.
	(op_code_prio): Likewise.
	* tree-ssa-loop-im.c (stmt_cost): Likewise.
	* tree-ssa-operands.c (get_expr_operands): Likewise.
	* tree.c (commutative_ternary_tree_code, add_expr): Likewise.
	* fold-const-call.h (fold_fma): Delete.
	* fold-const-call.c (fold_const_call_ssss): Handle CFN_FMS,
	CFN_FNMA and CFN_FNMS.
	(fold_fma): Delete.
	* genmatch.c (combined_fn): New enum.
	(commutative_ternary_tree_code): Remove FMA_EXPR handling.
	(commutative_op): New function.
	(commutate): Use it.  Handle more than 2 operands.
	(dt_operand::gen_gimple_expr): Use commutative_op.
	(parser::parse_expr): Allow :c to be used with non-binary
	operators if the commutative operand is known.
	* gimple-ssa-backprop.c (backprop::process_builtin_call_use): Handle
	CFN_FMS, CFN_FNMA and CFN_FNMS.
	(backprop::process_assign_use): Remove FMA_EXPR handling.
	* hsa-gen.c (gen_hsa_insns_for_operation_assignment): Likewise.
	(gen_hsa_fma): New function.
	(gen_hsa_insn_for_internal_fn_call): Use it for IFN_FMA, IFN_FMS,
	IFN_FNMA and IFN_FNMS.
	* match.pd: Add folds for IFN_FMS, IFN_FNMA and IFN_FNMS.
	* gimple-fold.h (follow_all_ssa_edges): Declare.
	* gimple-fold.c (follow_all_ssa_edges): New function.
	* tree-ssa-math-opts.c (convert_mult_to_fma_1): Use the
	gimple_build interface and use follow_all_ssa_edges to fold the result.
	(convert_mult_to_fma): Use direct_internal_fn_suppoerted_p
	instead of checking for optabs directly.
	* config/i386/i386.c (ix86_add_stmt_cost): Recognize FMAs as calls
	rather than FMA_EXPRs.
	* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Create a
	call to IFN_FMA instead of an FMA_EXPR.

gcc/brig/
	* brigfrontend/brig-function.cc
	(brig_function::get_builtin_for_hsa_opcode): Use BUILT_IN_FMA
	for BRIG_OPCODE_FMA.
	(brig_function::get_tree_code_for_hsa_opcode): Treat BUILT_IN_FMA
	as a call.

gcc/c/
	* gimple-parser.c (c_parser_gimple_postfix_expression): Remove
	__FMA_EXPR handlng.

gcc/cp/
	* constexpr.c (cxx_eval_constant_expression): Remove FMA_EXPR handling.
	(potential_constant_expression_1): Likewise.

gcc/testsuite/
	* lib/target-supports.exp (check_effective_target_scalar_all_fma):
	New proc.
	* gcc.dg/fma-1.c: New test.
	* gcc.dg/fma-2.c: Likewise.
	* gcc.dg/fma-3.c: Likewise.
	* gcc.dg/fma-4.c: Likewise.
	* gcc.dg/fma-5.c: Likewise.
	* gcc.dg/fma-6.c: Likewise.
	* gcc.dg/fma-7.c: Likewise.
	* gcc.dg/gimplefe-26.c: Use .FMA instead of __FMA and require
	scalar_all_fma.
	* gfortran.dg/reassoc_7.f: Pass -ffp-contract=off.
	* gfortran.dg/reassoc_8.f: Likewise.
	* gfortran.dg/reassoc_9.f: Likewise.
	* gfortran.dg/reassoc_10.f: Likewise.


git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@260348 138bc75d-0d04-0410-961f-82ee72b054a4

1 files changed, 67 insertions, 17 deletions

diff --git a/gcc/hsa-gen.c b/gcc/hsa-gen.c
index 7974fffe360..74d5f07b3cb 100644
--- a/gcc/hsa-gen.c
+++ b/gcc/hsa-gen.c
@@ -3178,23 +3178,6 @@ gen_hsa_insns_for_operation_assignment (gimple *assign, hsa_bb *hbb)
     case NEGATE_EXPR:
       opcode = BRIG_OPCODE_NEG;
       break;
-    case FMA_EXPR:
-      /* There is a native HSA instruction for scalar FMAs but not for vector
-	 ones.  */
-      if (TREE_CODE (TREE_TYPE (lhs)) == VECTOR_TYPE)
-	{
-	  hsa_op_reg *dest
-	    = hsa_cfun->reg_for_gimple_ssa (gimple_assign_lhs (assign));
-	  hsa_op_with_type *op1 = hsa_reg_or_immed_for_gimple_op (rhs1, hbb);
-	  hsa_op_with_type *op2 = hsa_reg_or_immed_for_gimple_op (rhs2, hbb);
-	  hsa_op_with_type *op3 = hsa_reg_or_immed_for_gimple_op (rhs3, hbb);
-	  hsa_op_reg *tmp = new hsa_op_reg (dest->m_type);
-	  gen_hsa_binary_operation (BRIG_OPCODE_MUL, tmp, op1, op2, hbb);
-	  gen_hsa_binary_operation (BRIG_OPCODE_ADD, dest, tmp, op3, hbb);
-	  return;
-	}
-      opcode = BRIG_OPCODE_MAD;
-      break;
     case MIN_EXPR:
       opcode = BRIG_OPCODE_MIN;
       break;
@@ -4490,6 +4473,57 @@ gen_hsa_divmod (gcall *call, hsa_bb *hbb)
   insn->set_output_in_type (dest, 0, hbb);
 }
 
+/* Emit instructions that implement FMA, FMS, FNMA or FNMS call STMT.
+   Instructions are appended to basic block HBB.  NEGATE1 is true for
+   FNMA and FNMS.  NEGATE3 is true for FMS and FNMS.  */
+
+static void
+gen_hsa_fma (gcall *call, hsa_bb *hbb, bool negate1, bool negate3)
+{
+  tree lhs = gimple_call_lhs (call);
+  if (lhs == NULL_TREE)
+    return;
+
+  tree rhs1 = gimple_call_arg (call, 0);
+  tree rhs2 = gimple_call_arg (call, 1);
+  tree rhs3 = gimple_call_arg (call, 2);
+
+  hsa_op_reg *dest = hsa_cfun->reg_for_gimple_ssa (lhs);
+  hsa_op_with_type *op1 = hsa_reg_or_immed_for_gimple_op (rhs1, hbb);
+  hsa_op_with_type *op2 = hsa_reg_or_immed_for_gimple_op (rhs2, hbb);
+  hsa_op_with_type *op3 = hsa_reg_or_immed_for_gimple_op (rhs3, hbb);
+
+  if (negate1)
+    {
+      hsa_op_reg *tmp = new hsa_op_reg (dest->m_type);
+      gen_hsa_unary_operation (BRIG_OPCODE_NEG, tmp, op1, hbb);
+      op1 = tmp;
+    }
+
+  /* There is a native HSA instruction for scalar FMAs but not for vector
+     ones.  */
+  if (TREE_CODE (TREE_TYPE (lhs)) == VECTOR_TYPE)
+    {
+      hsa_op_reg *tmp = new hsa_op_reg (dest->m_type);
+      gen_hsa_binary_operation (BRIG_OPCODE_MUL, tmp, op1, op2, hbb);
+      gen_hsa_binary_operation (negate3 ? BRIG_OPCODE_SUB : BRIG_OPCODE_ADD,
+				dest, tmp, op3, hbb);
+    }
+  else
+    {
+      if (negate3)
+	{
+	  hsa_op_reg *tmp = new hsa_op_reg (dest->m_type);
+	  gen_hsa_unary_operation (BRIG_OPCODE_NEG, tmp, op3, hbb);
+	  op3 = tmp;
+	}
+      hsa_insn_basic *insn = new hsa_insn_basic (4, BRIG_OPCODE_MAD,
+						 dest->m_type, dest,
+						 op1, op2, op3);
+      hbb->append_insn (insn);
+    }
+}
+
 /* Set VALUE to a shadow kernel debug argument and append a new instruction
    to HBB basic block.  */
 
@@ -5224,6 +5258,22 @@ gen_hsa_insn_for_internal_fn_call (gcall *stmt, hsa_bb *hbb)
       gen_hsa_insns_for_call_of_internal_fn (stmt, hbb);
       break;
 
+    case IFN_FMA:
+      gen_hsa_fma (stmt, hbb, false, false);
+      break;
+
+    case IFN_FMS:
+      gen_hsa_fma (stmt, hbb, false, true);
+      break;
+
+    case IFN_FNMA:
+      gen_hsa_fma (stmt, hbb, true, false);
+      break;
+
+    case IFN_FNMS:
+      gen_hsa_fma (stmt, hbb, true, true);
+      break;
+
     default:
       HSA_SORRY_ATV (gimple_location (stmt),
 		     "support for HSA does not implement internal function: %s",