X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fir%2Firopt.c;h=ca58b2ef4eeb06915ab19343d0987be77cbe1dc5;hb=a08e6f04aa3669cff094f94a9484c7c2bb1314d0;hp=dc2907678b653dc484a8ed23820cb7374a3afeef;hpb=630162b2084a7ef60e7593824d1552bc785e94eb;p=libfirm

diff --git a/ir/ir/iropt.c b/ir/ir/iropt.c
index dc2907678..ca58b2ef4 100644
--- a/ir/ir/iropt.c
+++ b/ir/ir/iropt.c
@@ -352,7 +352,7 @@ static ir_tarval *computed_value_Not(const ir_node *n)
 }  /* computed_value_Not */
 
 /**
- * Tests wether a shift shifts more bits than available in the mode
+ * Tests whether a shift shifts more bits than available in the mode
  */
 static bool is_oversize_shift(const ir_node *n)
 {
@@ -957,7 +957,7 @@ static ir_node *equivalent_node_Sub(ir_node *n)
 
 
 /**
- * Optimize an "self-inverse unary op", ie op(op(n)) = n.
+ * Optimize an "self-inverse unary op", i.e. op(op(n)) = n.
  *
  * @todo
  *   -(-a) == a, but might overflow two times.
@@ -2181,6 +2181,7 @@ restart:
 		}
 		DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_TO_ADD);
 		return n;
+#if 0
 	} else if (is_Mul(b)) { /* a - (b * C) -> a + (b * -C) */
 		ir_node *m_right = get_Mul_right(b);
 		if (is_Const(m_right)) {
@@ -2199,6 +2200,7 @@ restart:
 				return n;
 			}
 		}
+#endif
 	}
 
 	/* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
@@ -2663,7 +2665,7 @@ make_tuple:
 
 		/* skip a potential Pin */
 		mem = skip_Pin(mem);
-		turn_into_tuple(n, pn_Div_max);
+		turn_into_tuple(n, pn_Div_max+1);
 		set_Tuple_pred(n, pn_Div_M,         mem);
 		set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(blk));
 		set_Tuple_pred(n, pn_Div_X_except,  new_r_Bad(irg, mode_X));
@@ -2755,7 +2757,7 @@ make_tuple:
 
 		/* skip a potential Pin */
 		mem = skip_Pin(mem);
-		turn_into_tuple(n, pn_Mod_max);
+		turn_into_tuple(n, pn_Mod_max+1);
 		set_Tuple_pred(n, pn_Mod_M,         mem);
 		set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(blk));
 		set_Tuple_pred(n, pn_Mod_X_except,  new_r_Bad(irg, mode_X));
@@ -2789,7 +2791,7 @@ static ir_node *transform_node_Cond(ir_node *n)
 		   Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
 		ir_node *blk = get_nodes_block(n);
 		jmp = new_r_Jmp(blk);
-		turn_into_tuple(n, pn_Cond_max);
+		turn_into_tuple(n, pn_Cond_max+1);
 		if (ta == tarval_b_true) {
 			set_Tuple_pred(n, pn_Cond_false, new_r_Bad(irg, mode_X));
 			set_Tuple_pred(n, pn_Cond_true, jmp);
@@ -2930,15 +2932,109 @@ static bool is_shiftop(const ir_node *n)
 	return is_Shl(n) || is_Shr(n) || is_Shrs(n) || is_Rotl(n);
 }
 
+/**
+ * normalisation: (x & c1) >> c2   to   (x >> c2) & (c1 >> c2)
+ *  (we can use:
+ *    - and, or, xor          instead of &
+ *    - Shl, Shr, Shrs, rotl  instead of >>
+ *    (with a special case for Or/Xor + Shrs)
+ *
+ * This normalisation is good for things like x-(x&y) esp. in 186.crafty.
+ */
+static ir_node *transform_node_shift_bitop(ir_node *n)
+{
+	ir_graph  *irg   = get_irn_irg(n);
+	ir_node   *right = get_binop_right(n);
+	ir_mode   *mode  = get_irn_mode(n);
+	ir_node   *left;
+	ir_node   *bitop_left;
+	ir_node   *bitop_right;
+	ir_op     *op_left;
+	ir_node   *block;
+	dbg_info  *dbgi;
+	ir_node   *new_shift;
+	ir_node   *new_bitop;
+	ir_node   *new_const;
+	ir_tarval *tv1;
+	ir_tarval *tv2;
+	ir_tarval *tv_shift;
+
+	if (is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
+		return n;
+
+	assert(is_Shrs(n) || is_Shr(n) || is_Shl(n) || is_Rotl(n));
+
+	if (!is_Const(right))
+		return n;
+
+	left    = get_binop_left(n);
+	op_left = get_irn_op(left);
+	if (op_left != op_And && op_left != op_Or && op_left != op_Eor)
+		return n;
+
+	/* doing it with Shrs is not legal if the Or/Eor affects the topmost bit */
+	if (is_Shrs(n) && (op_left == op_Or || op_left == op_Eor)) {
+		/* TODO: test if sign bit is affectes */
+		return n;
+	}
+
+	bitop_right = get_binop_right(left);
+	if (!is_Const(bitop_right))
+		return n;
+
+	bitop_left = get_binop_left(left);
+
+	block = get_nodes_block(n);
+	dbgi  = get_irn_dbg_info(n);
+	tv1   = get_Const_tarval(bitop_right);
+	tv2   = get_Const_tarval(right);
+
+	assert(get_tarval_mode(tv1) == mode);
+
+	if (is_Shl(n)) {
+		new_shift = new_rd_Shl(dbgi, block, bitop_left, right, mode);
+		tv_shift  = tarval_shl(tv1, tv2);
+	} else if (is_Shr(n)) {
+		new_shift = new_rd_Shr(dbgi, block, bitop_left, right, mode);
+		tv_shift  = tarval_shr(tv1, tv2);
+	} else if (is_Shrs(n)) {
+		new_shift = new_rd_Shrs(dbgi, block, bitop_left, right, mode);
+		tv_shift  = tarval_shrs(tv1, tv2);
+	} else {
+		assert(is_Rotl(n));
+		new_shift = new_rd_Rotl(dbgi, block, bitop_left, right, mode);
+		tv_shift  = tarval_rotl(tv1, tv2);
+	}
+
+	assert(get_tarval_mode(tv_shift) == mode);
+	irg       = get_irn_irg(n);
+	new_const = new_r_Const(irg, tv_shift);
+
+	if (op_left == op_And) {
+		new_bitop = new_rd_And(dbgi, block, new_shift, new_const, mode);
+	} else if (op_left == op_Or) {
+		new_bitop = new_rd_Or(dbgi, block, new_shift, new_const, mode);
+	} else {
+		assert(op_left == op_Eor);
+		new_bitop = new_rd_Eor(dbgi, block, new_shift, new_const, mode);
+	}
+
+	return new_bitop;
+}
+
 /**
  * normalisation: (x >> c1) & c2   to   (x & (c2<<c1)) >> c1
  *  (we can use:
  *    - and, or, xor          instead of &
  *    - Shl, Shr, Shrs, rotl  instead of >>
  *    (with a special case for Or/Xor + Shrs)
+ *
+ * This normalisation is usually good for the backend since << C can often be
+ * matched as address-mode.
  */
 static ir_node *transform_node_bitop_shift(ir_node *n)
 {
+	ir_graph  *irg   = get_irn_irg(n);
 	ir_node   *left  = get_binop_left(n);
 	ir_node   *right = get_binop_right(n);
 	ir_mode   *mode  = get_irn_mode(n);
@@ -2947,7 +3043,6 @@ static ir_node *transform_node_bitop_shift(ir_node *n)
 	ir_node   *block;
 	dbg_info  *dbg_bitop;
 	dbg_info  *dbg_shift;
-	ir_graph  *irg;
 	ir_node   *new_bitop;
 	ir_node   *new_shift;
 	ir_node   *new_const;
@@ -2955,6 +3050,9 @@ static ir_node *transform_node_bitop_shift(ir_node *n)
 	ir_tarval *tv2;
 	ir_tarval *tv_bitop;
 
+	if (!is_irg_state(irg, IR_GRAPH_STATE_NORMALISATION2))
+		return n;
+
 	assert(is_And(n) || is_Or(n) || is_Eor(n));
 	if (!is_Const(right) || !is_shiftop(left))
 		return n;
@@ -3766,6 +3864,7 @@ static ir_node *transform_node_Cmp(ir_node *n)
 				ir_graph *irg = get_irn_irg(n);
 				left    = op_left;
 				right   = new_r_Const(irg, new_tv);
+				mode    = get_irn_mode(left);
 				changed = true;
 				DBG_OPT_ALGSIM0(n, n, FS_OPT_CMP_CONV);
 			}
@@ -5048,7 +5147,7 @@ typedef ir_node*(*new_shift_func)(dbg_info *dbgi, ir_node *block,
  * then we can use that to minimize the value of Add(x, const) or
  * Sub(Const, x). In particular this often avoids 1 instruction in some
  * backends for the Shift(x, Sub(Const, y)) case because it can be replaced
- * by Shift(x, Minus(y)) which doesnt't need an explicit Const constructed.
+ * by Shift(x, Minus(y)) which does not need an explicit Const constructed.
  */
 static ir_node *transform_node_shift_modulo(ir_node *n,
                                             new_shift_func new_shift)
@@ -5135,6 +5234,8 @@ static ir_node *transform_node_Shr(ir_node *n)
 		n = transform_node_shift_modulo(n, new_rd_Shr);
 	if (is_Shr(n))
 		n = transform_node_shl_shr(n);
+	if (is_Shr(n))
+		n = transform_node_shift_bitop(n);
 
 	return n;
 }  /* transform_node_Shr */
@@ -5164,6 +5265,8 @@ static ir_node *transform_node_Shrs(ir_node *n)
 
 	if (is_Shrs(n))
 		n = transform_node_shift_modulo(n, new_rd_Shrs);
+	if (is_Shrs(n))
+		n = transform_node_shift_bitop(n);
 
 	return n;
 }  /* transform_node_Shrs */
@@ -5185,6 +5288,8 @@ static ir_node *transform_node_Shl(ir_node *n)
 		n = transform_node_shift_modulo(n, new_rd_Shl);
 	if (is_Shl(n))
 		n = transform_node_shl_shr(n);
+	if (is_Shl(n))
+		n = transform_node_shift_bitop(n);
 
 	return n;
 }  /* transform_node_Shl */
@@ -5202,6 +5307,9 @@ static ir_node *transform_node_Rotl(ir_node *n)
 	HANDLE_BINOP_PHI((eval_func) tarval_rotl, a, b, c, mode);
 	n = transform_node_shift(n);
 
+	if (is_Rotl(n))
+		n = transform_node_shift_bitop(n);
+
 	return n;
 }  /* transform_node_Rotl */
 
@@ -5276,7 +5384,7 @@ static ir_node *transform_node_End(ir_node *n)
 		/* no need to keep Bad */
 		if (is_Bad(ka))
 			continue;
-		/* dont keep unreachable code */
+		/* do not keep unreachable code */
 		block = is_Block(ka) ? ka : get_nodes_block(ka);
 		if (is_block_unreachable(block))
 			continue;
@@ -5306,16 +5414,115 @@ int ir_is_negated_value(const ir_node *a, const ir_node *b)
 	return false;
 }
 
+static const ir_node *skip_upconv(const ir_node *node)
+{
+	while (is_Conv(node)) {
+		ir_mode       *mode    = get_irn_mode(node);
+		const ir_node *op      = get_Conv_op(node);
+		ir_mode       *op_mode = get_irn_mode(op);
+		if (!smaller_mode(op_mode, mode))
+			break;
+		node = op;
+	}
+	return node;
+}
+
+int ir_mux_is_abs(const ir_node *sel, const ir_node *mux_true,
+                  const ir_node *mux_false)
+{
+	ir_node    *cmp_left;
+	ir_node    *cmp_right;
+	ir_mode    *mode;
+	ir_relation relation;
+
+	if (!is_Cmp(sel))
+		return 0;
+
+	/**
+	 * Note further that these optimization work even for floating point
+	 * with NaN's because -NaN == NaN.
+	 * However, if +0 and -0 is handled differently, we cannot use the Abs/-Abs
+	 * transformations.
+	 */
+	mode = get_irn_mode(mux_true);
+	if (mode_honor_signed_zeros(mode))
+		return 0;
+
+	/* must be <, <=, >=, > */
+	relation = get_Cmp_relation(sel);
+	if ((relation & ir_relation_less_greater) == 0)
+		return 0;
+
+	if (!ir_is_negated_value(mux_true, mux_false))
+		return 0;
+
+	mux_true  = skip_upconv(mux_true);
+	mux_false = skip_upconv(mux_false);
+
+	/* must be x cmp 0 */
+	cmp_right = get_Cmp_right(sel);
+	if (!is_Const(cmp_right) || !is_Const_null(cmp_right))
+		return 0;
+
+	cmp_left = get_Cmp_left(sel);
+	if (cmp_left == mux_false) {
+		if (relation & ir_relation_less) {
+			return 1;
+		} else {
+			assert(relation & ir_relation_greater);
+			return -1;
+		}
+	} else if (cmp_left == mux_true) {
+		if (relation & ir_relation_less) {
+			return -1;
+		} else {
+			assert(relation & ir_relation_greater);
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+ir_node *ir_get_abs_op(const ir_node *sel, ir_node *mux_true,
+                       ir_node *mux_false)
+{
+	ir_node *cmp_left = get_Cmp_left(sel);
+	return cmp_left == skip_upconv(mux_false) ? mux_false : mux_true;
+}
+
 /**
  * Optimize a Mux into some simpler cases.
  */
 static ir_node *transform_node_Mux(ir_node *n)
 {
-	ir_node *oldn = n, *sel = get_Mux_sel(n);
-	ir_mode *mode = get_irn_mode(n);
-	ir_node  *t   = get_Mux_true(n);
-	ir_node  *f   = get_Mux_false(n);
-	ir_graph *irg = get_irn_irg(n);
+	ir_node  *oldn = n;
+	ir_node  *sel  = get_Mux_sel(n);
+	ir_mode  *mode = get_irn_mode(n);
+	ir_node  *t    = get_Mux_true(n);
+	ir_node  *f    = get_Mux_false(n);
+	ir_graph *irg  = get_irn_irg(n);
+
+	/* implement integer abs: abs(x) = x^(x >>s 31) - (x >>s 31) */
+	if (get_mode_arithmetic(mode) == irma_twos_complement) {
+		int abs = ir_mux_is_abs(sel, t, f);
+		if (abs != 0) {
+			dbg_info *dbgi       = get_irn_dbg_info(n);
+			ir_node  *block      = get_nodes_block(n);
+			ir_node  *op         = ir_get_abs_op(sel, t, f);
+			int       bits       = get_mode_size_bits(mode);
+			ir_node  *shiftconst = new_r_Const_long(irg, mode_Iu, bits-1);
+			ir_node  *sext       = new_rd_Shrs(dbgi, block, op, shiftconst, mode);
+			ir_node  *xorn       = new_rd_Eor(dbgi, block, op, sext, mode);
+			ir_node  *res;
+			if (abs > 0) {
+				res = new_rd_Sub(dbgi, block, xorn, sext, mode);
+			} else {
+				res = new_rd_Sub(dbgi, block, sext, xorn, mode);
+			}
+			return res;
+		}
+	}
 
 	if (is_irg_state(irg, IR_GRAPH_STATE_KEEP_MUX))
 		return n;
@@ -5626,7 +5833,7 @@ static ir_node *transform_node_Load(ir_node *n)
 			ir_node  *bad   = new_r_Bad(irg, mode_X);
 			ir_mode  *mode  = get_Load_mode(n);
 			ir_node  *res   = new_r_Proj(pred_load, mode, pn_Load_res);
-			ir_node  *in[pn_Load_max] = { mem, jmp, bad, res };
+			ir_node  *in[pn_Load_max+1] = { mem, res, jmp, bad };
 			ir_node  *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
 			return tuple;
 		}
@@ -5646,7 +5853,7 @@ static ir_node *transform_node_Load(ir_node *n)
 			ir_graph *irg   = get_irn_irg(n);
 			ir_node  *bad   = new_r_Bad(irg, mode_X);
 			ir_node  *res   = value;
-			ir_node  *in[pn_Load_max] = { mem, jmp, bad, res };
+			ir_node  *in[pn_Load_max+1] = { mem, res, jmp, bad };
 			ir_node  *tuple = new_r_Tuple(block, ARRAY_SIZE(in), in);
 			return tuple;
 		}
@@ -6109,10 +6316,23 @@ int identities_cmp(const void *elt, const void *key)
 		/* for pinned nodes, the block inputs must be equal */
 		if (get_irn_n(a, -1) != get_irn_n(b, -1))
 			return 1;
-	} else if (! get_opt_global_cse()) {
-		/* for block-local CSE both nodes must be in the same Block */
-		if (get_nodes_block(a) != get_nodes_block(b))
-			return 1;
+	} else {
+		ir_node *block_a = get_nodes_block(a);
+		ir_node *block_b = get_nodes_block(b);
+		if (! get_opt_global_cse()) {
+			/* for block-local CSE both nodes must be in the same Block */
+			if (block_a != block_b)
+				return 1;
+		} else {
+			/* The optimistic approach would be to do nothing here.
+			 * However doing GCSE optimistically produces a lot of partially dead code which appears
+			 * to be worse in practice than the missed opportunities.
+			 * So we use a very conservative variant here and only CSE if 1 value dominates the
+			 * other. */
+			if (!block_dominates(block_a, block_b)
+			    && !block_dominates(block_b, block_a))
+			    return 1;
+		}
 	}
 
 	/* compare a->in[0..ins] with b->in[0..ins] */