amd64: Transform x87 cmp -> fucomi

ir/be/amd64/amd64_spec.pl

@@ -820,6 +820,16 @@ fchs => {
 	emit     => "fchs",
 },
 
+fucomi => {
+	irn_flags => [ "rematerializable" ],
+	in_reqs   => [ "x87", "x87" ],
+	out_reqs  => [ "flags" ],
+	ins       => [ "left", "right" ],
+	outs      => [ "flags" ],
+	attr_type => "amd64_x87_attr_t",
+	emit      => "fucom%FPi %F0",
+},
+
 fdup => {
 	in_reqs     => [ "x87" ],
 	out_reqs    => [ "x87" ],

ir/be/amd64/amd64_transform.c

@@ -1893,6 +1893,16 @@ static ir_node *gen_Proj_Proj(ir_node *const node)
 	panic("amd64: unexpected Proj(Proj) after %+F", pred_pred);
 }
 
+static ir_node *match_cmp_x87(ir_node *const node, ir_node *const op0,
+                              ir_node *op1)
+{
+	dbg_info *const dbgi      = get_irn_dbg_info(node);
+	ir_node  *const new_block = be_transform_node(get_nodes_block(node));
+	ir_node  *const new_op0   = be_transform_node(op0);
+	ir_node  *const new_op1   = be_transform_node(op1);
+	return new_bd_amd64_fucomi(dbgi, new_block, new_op0, new_op1);
+}
+
 static ir_node *gen_Cmp(ir_node *const node)
 {
 	ir_node  *const op1       = get_Cmp_left(node);
@@ -1905,6 +1915,8 @@ static ir_node *gen_Cmp(ir_node *const node)
 	ir_node     *new_node;
 	amd64_args_t args;
 	if (mode_is_float(cmp_mode)) {
+		if (cmp_mode == x86_mode_E)
+			return match_cmp_x87(node, op1, op2);
 		match_binop(&args, block, cmp_mode, op1, op2, match_am);
 		new_node = new_bd_amd64_ucomis(dbgi, new_block, args.arity, args.in, args.reqs, &args.attr);
 	} else {

ir/be/amd64/amd64_x87.c

@@ -75,6 +75,21 @@ static void sim_amd64_fdiv(x87_state *const state, ir_node *const node)
 	x86_sim_x87_binop(state, node, n_amd64_fdiv_left, n_amd64_fdiv_right, out);
 }
 
+static void sim_amd64_fucomi(x87_state *const state, ir_node *const node)
+{
+	ir_node *const op0 = get_irn_n(node, n_amd64_fucomi_left);
+	ir_node *const op1 = get_irn_n(node, n_amd64_fucomi_right);
+	unsigned const additional_pop = x86_sim_x87_fucom(state, node, op0, op1);
+	if (additional_pop != ~0u)
+		x86_x87_create_fpop(state, node, additional_pop);
+
+	/* TODO remove the following check once all emitters that test flags
+	 * check the predecessor nodes for reverse condition */
+	x87_attr_t const *const x87 = amd64_get_x87_attr(node);
+	if (x87->reverse)
+		panic("reverse not handled yet");
+}
+
 static void prepare_callbacks(void)
 {
 	x86_prepare_x87_callbacks();
@@ -89,6 +104,7 @@ static void prepare_callbacks(void)
 	x86_register_x87_sim(op_amd64_fst,    sim_amd64_fst);
 	x86_register_x87_sim(op_amd64_fstp,   sim_amd64_fstp);
 	x86_register_x87_sim(op_amd64_fsub,   sim_amd64_fsub);
+	x86_register_x87_sim(op_amd64_fucomi, sim_amd64_fucomi);
 	x86_register_x87_sim(op_amd64_ret,    x86_sim_x87_ret);
 }
 

ir/be/ia32/x86_x87.c

@@ -474,16 +474,7 @@ static ir_node *x87_dup_operand(x87_state *const state, ir_node *const n, unsign
 	return fdup;
 }
 
-/**
- * Create a fpop before node n.
- * This overwrites st(pos) with st(0) and pops st(0).
- *
- * @param state   the x87 state
- * @param n       the node after which to schedule the fpop
- * @param pos     the index of the entry to remove the register stack
- * @return the fpop node
- */
-static ir_node *x87_create_fpop(x87_state *const state, ir_node *const n,
+ir_node *x86_x87_create_fpop(x87_state *const state, ir_node *const n,
                              unsigned const pos)
 {
 	if (pos != 0) {
@@ -940,7 +931,7 @@ static void sim_FtstFnstsw(x87_state *state, ir_node *n)
 	move_to_tos(state, n, val);
 
 	if (!is_fp_live(val, live))
-		x87_create_fpop(state, n, 0);
+		x86_x87_create_fpop(state, n, 0);
 }
 
 unsigned x86_sim_x87_fucom(x87_state *const state, ir_node *const node,
@@ -1045,7 +1036,7 @@ static void sim_ia32_FucomFnstsw(x87_state *const state, ir_node *const node)
 			set_irn_op(node, op_ia32_FucomppFnstsw);
 			x87_pop(state);
 		} else {
-			x87_create_fpop(state, node, additional_pop);
+			x86_x87_create_fpop(state, node, additional_pop);
 		}
 	}
 }
@@ -1059,7 +1050,7 @@ static void sim_ia32_Fucomi(x87_state *const state, ir_node *const node)
 	ia32_x87_attr_t *const attr = get_ia32_x87_attr(node);
 	attr->attr.ins_permuted ^= attr->x87.reverse;
 	if (additional_pop != ~0u)
-		x87_create_fpop(state, node, additional_pop);
+		x86_x87_create_fpop(state, node, additional_pop);
 }
 
 /**
@@ -1261,7 +1252,7 @@ free_all:;
 				}
 			}
 		}
-		insert      = x87_create_fpop(state, insert, i);
+		insert      = x86_x87_create_fpop(state, insert, i);
 		depth      -= 1;
 		kill_mask >>= 1;
 	}

ir/be/ia32/x86_x87.h

@@ -73,6 +73,18 @@ unsigned x86_sim_x87_fucom(x87_state *state, ir_node *node,
 /** Push a value on the x87 stack. Intended to be used in sim functions. */
 void x86_x87_push(x87_state *state, ir_node *value);
 
+/**
+ * Create a fpop before node n.
+ * This overwrites st(pos) with st(0) and pops st(0).
+ *
+ * @param state   the x87 state
+ * @param n       the node after which to schedule the fpop
+ * @param pos     the index of the entry to remove the register stack
+ * @return the fpop node
+ */
+ir_node *x86_x87_create_fpop(x87_state *const state, ir_node *const n,
+                             unsigned const pos);
+
 /**
  * Register a simulator function.
  *