add a parameter to limit the maximum size of methods when inlining, otherwise some...
[libfirm] / ir / ir / iropt.c
index 0318352..c2cecb1 100644 (file)
@@ -722,10 +722,16 @@ static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops
 static ir_node *equivalent_node_Block(ir_node *n)
 {
        ir_node *oldn = n;
-       int n_preds   = get_Block_n_cfgpreds(n);
+       int     n_preds;
 
-       /* The Block constructor does not call optimize, but mature_immBlock
-       calls the optimization. */
+       /* don't optimize dead blocks */
+       if (is_Block_dead(n))
+               return n;
+
+       n_preds = get_Block_n_cfgpreds(n);
+
+       /* The Block constructor does not call optimize, but mature_immBlock()
+          calls the optimization. */
        assert(get_Block_matured(n));
 
        /* Straightening: a single entry Block following a single exit Block
@@ -971,7 +977,7 @@ static ir_node *equivalent_node_Sub(ir_node *n) {
 
 
 /**
- * Optimize an "idempotent unary op", ie op(op(n)) = n.
+ * Optimize an "self-inverse unary op", ie op(op(n)) = n.
  *
  * @todo
  *   -(-a) == a, but might overflow two times.
@@ -1232,7 +1238,7 @@ static ir_node *equivalent_node_Phi(ir_node *n) {
        int i, n_preds;
 
        ir_node *oldn = n;
-       ir_node *block = NULL;     /* to shutup gcc */
+       ir_node *block;
        ir_node *first_val = NULL; /* to shutup gcc */
 
        if (!get_opt_normalize()) return n;
@@ -1240,8 +1246,6 @@ static ir_node *equivalent_node_Phi(ir_node *n) {
        n_preds = get_Phi_n_preds(n);
 
        block = get_nodes_block(n);
-       /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
-       assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
        if ((is_Block_dead(block)) ||                  /* Control dead */
                (block == get_irg_start_block(current_ir_graph)))  /* There should be no Phi nodes */
                return new_Bad();                                    /* in the Start Block. */
@@ -1299,45 +1303,36 @@ static ir_node *equivalent_node_Phi(ir_node *n) {
  *   themselves.
  */
 static ir_node *equivalent_node_Sync(ir_node *n) {
-       int i, n_preds;
-
-       ir_node *oldn = n;
-       ir_node *first_val = NULL; /* to shutup gcc */
-
-       if (!get_opt_normalize()) return n;
+       int arity = get_Sync_n_preds(n);
+       int i;
 
-       n_preds = get_Sync_n_preds(n);
+       for (i = 0; i < arity;) {
+               ir_node *pred = get_Sync_pred(n, i);
+               int      j;
 
-       /* Find first non-self-referencing input */
-       for (i = 0; i < n_preds; ++i) {
-               first_val = get_Sync_pred(n, i);
-               if ((first_val != n)  /* not self pointer */ &&
-                   (! is_Bad(first_val))
-                  ) {                /* value not dead */
-                       break;            /* then found first value. */
+               /* Remove Bad predecessors */
+               if (is_Bad(pred)) {
+                       del_Sync_n(n, i);
+                       --arity;
+                       continue;
                }
-       }
 
-       if (i >= n_preds)
-               /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
-               return new_Bad();
-
-       /* search the rest of inputs, determine if any of these
-          are non-self-referencing */
-       while (++i < n_preds) {
-               ir_node *scnd_val = get_Sync_pred(n, i);
-               if ((scnd_val != n) &&
-                   (scnd_val != first_val) &&
-                   (! is_Bad(scnd_val))
-                  )
-                       break;
+               /* Remove duplicate predecessors */
+               for (j = 0;; ++j) {
+                       if (j >= i) {
+                               ++i;
+                               break;
+                       }
+                       if (get_Sync_pred(n, j) == pred) {
+                               del_Sync_n(n, i);
+                               --arity;
+                               break;
+                       }
+               }
        }
 
-       if (i >= n_preds) {
-               /* Fold, if no multiple distinct non-self-referencing inputs */
-               n = first_val;
-               DBG_OPT_SYNC(oldn, n);
-       }
+       if (arity == 0) return new_Bad();
+       if (arity == 1) return get_Sync_pred(n, 0);
        return n;
 }  /* equivalent_node_Sync */
 
@@ -1445,41 +1440,54 @@ static ir_node *equivalent_node_Mux(ir_node *n)
        else if (is_Proj(sel) && !mode_honor_signed_zeros(get_irn_mode(n))) {
                ir_node *cmp = get_Proj_pred(sel);
                long proj_nr = get_Proj_proj(sel);
-               ir_node *b   = get_Mux_false(n);
-               ir_node *a   = get_Mux_true(n);
+               ir_node *f   = get_Mux_false(n);
+               ir_node *t   = get_Mux_true(n);
 
                /*
-                * Note: normalization puts the constant on the right site,
-                * so we check only one case.
-                *
                 * 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 first one.
                 */
-               if (is_Cmp(cmp) && get_Cmp_left(cmp) == a) {
-                       ir_node *cmp_r = get_Cmp_right(cmp);
-                       if (is_Const(cmp_r) && is_Const_null(cmp_r)) {
-                               /* Mux(a CMP 0, X, a) */
-                               if (is_Minus(b) && get_Minus_op(b) == a) {
-                                       /* Mux(a CMP 0, -a, a) */
-                                       if (proj_nr == pn_Cmp_Eq) {
-                                               /* Mux(a == 0, -a, a)  ==>  -a */
-                                               n = b;
+               if (is_Cmp(cmp)) {
+                       ir_node *const cmp_l = get_Cmp_left(cmp);
+                       ir_node *const cmp_r = get_Cmp_right(cmp);
+
+                       switch (proj_nr) {
+                               case pn_Cmp_Eq:
+                                       if ((cmp_l == t && cmp_r == f) || /* Psi(t == f, t, f) -> f */
+                                                       (cmp_l == f && cmp_r == t)) { /* Psi(f == t, t, f) -> f */
+                                               n = f;
                                                DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
-                                       } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
-                                               /* Mux(a != 0, -a, a)  ==> a */
-                                               n = a;
+                                               return n;
+                                       }
+                                       break;
+
+                               case pn_Cmp_Lg:
+                               case pn_Cmp_Ne:
+                                       if ((cmp_l == t && cmp_r == f) || /* Psi(t != f, t, f) -> t */
+                                                       (cmp_l == f && cmp_r == t)) { /* Psi(f != t, t, f) -> t */
+                                               n = t;
                                                DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
+                                               return n;
                                        }
-                               } else if (is_Const(b) && is_Const_null(b)) {
-                                       /* Mux(a CMP 0, 0, a) */
-                                       if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
-                                               /* Mux(a != 0, 0, a) ==> a */
-                                               n = a;
+                                       break;
+                       }
+
+                       /*
+                        * Note: normalization puts the constant on the right side,
+                        * so we check only one case.
+                        */
+                       if (cmp_l == t && is_Const(cmp_r) && is_Const_null(cmp_r)) {
+                               /* Mux(t CMP 0, X, t) */
+                               if (is_Minus(f) && get_Minus_op(f) == t) {
+                                       /* Mux(t CMP 0, -t, t) */
+                                       if (proj_nr == pn_Cmp_Eq) {
+                                               /* Mux(t == 0, -t, t)  ==>  -t */
+                                               n = f;
                                                DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
-                                       } else if (proj_nr == pn_Cmp_Eq) {
-                                               /* Mux(a == 0, 0, a) ==> 0 */
-                                               n = b;
+                                       } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
+                                               /* Mux(t != 0, -t, t)  ==> t */
+                                               n = t;
                                                DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
                                        }
                                }
@@ -1767,7 +1775,7 @@ static ir_node *apply_binop_on_2_phis(ir_node *a, ir_node *b, tarval *(*eval)(ta
        int      i, n;
 
        if (get_nodes_block(a) != get_nodes_block(b))
-      return NULL;
+               return NULL;
 
        n = get_irn_arity(a);
        NEW_ARR_A(void *, res, n);
@@ -1880,12 +1888,12 @@ static ir_node *transform_node_AddSub(ir_node *n) {
                unsigned ref_bits = get_mode_size_bits(mode);
 
                if (is_Conv(left)) {
-                       ir_mode *mode = get_irn_mode(left);
-                       unsigned bits = get_mode_size_bits(mode);
+                       ir_mode *lmode = get_irn_mode(left);
+                       unsigned bits = get_mode_size_bits(lmode);
 
                        if (ref_bits == bits &&
-                           mode_is_int(mode) &&
-                           get_mode_arithmetic(mode) == irma_twos_complement) {
+                           mode_is_int(lmode) &&
+                           get_mode_arithmetic(lmode) == irma_twos_complement) {
                                ir_node *pre      = get_Conv_op(left);
                                ir_mode *pre_mode = get_irn_mode(pre);
 
@@ -1903,12 +1911,12 @@ static ir_node *transform_node_AddSub(ir_node *n) {
                }
 
                if (is_Conv(right)) {
-                       ir_mode *mode = get_irn_mode(right);
-                       unsigned bits = get_mode_size_bits(mode);
+                       ir_mode *rmode = get_irn_mode(right);
+                       unsigned bits = get_mode_size_bits(rmode);
 
                        if (ref_bits == bits &&
-                               mode_is_int(mode) &&
-                               get_mode_arithmetic(mode) == irma_twos_complement) {
+                           mode_is_int(rmode) &&
+                           get_mode_arithmetic(rmode) == irma_twos_complement) {
                                ir_node *pre      = get_Conv_op(right);
                                ir_mode *pre_mode = get_irn_mode(pre);
 
@@ -1924,6 +1932,19 @@ static ir_node *transform_node_AddSub(ir_node *n) {
                                }
                        }
                }
+
+               /* let address arithmetic use unsigned modes */
+               if (is_Const(right)) {
+                       ir_mode *rmode = get_irn_mode(right);
+
+                       if (mode_is_signed(rmode) && get_mode_arithmetic(rmode) == irma_twos_complement) {
+                               /* convert a AddP(P, *s) into AddP(P, *u) */
+                               ir_mode *nm = get_reference_mode_unsigned_eq(mode);
+
+                               ir_node *pre = new_r_Conv(current_ir_graph, get_nodes_block(n), right, nm);
+                               set_binop_right(n, pre);
+                       }
+               }
        }
        return n;
 }  /* transform_node_AddSub */
@@ -2829,7 +2850,8 @@ static ir_node *transform_node_Abs(ir_node *n) {
 
                /*
                 * We can replace the Abs by -x here.
-                * We even could add a new Confirm here.
+                * We even could add a new Confirm here
+                * (if not twos complement)
                 *
                 * Note that -x would create a new node, so we could
                 * not run it in the equivalent_node() context.
@@ -2880,7 +2902,8 @@ static ir_node *transform_node_Cond(ir_node *n) {
            (get_opt_unreachable_code())) {
                /* It's a boolean Cond, branching on a boolean constant.
                   Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
-               jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
+               ir_node *blk = get_nodes_block(n);
+               jmp = new_r_Jmp(current_ir_graph, blk);
                turn_into_tuple(n, pn_Cond_max);
                if (ta == tarval_b_true) {
                        set_Tuple_pred(n, pn_Cond_false, new_Bad());
@@ -2890,12 +2913,16 @@ static ir_node *transform_node_Cond(ir_node *n) {
                        set_Tuple_pred(n, pn_Cond_true, new_Bad());
                }
                /* We might generate an endless loop, so keep it alive. */
-               add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
+               add_End_keepalive(get_irg_end(current_ir_graph), blk);
        }
        return n;
 }  /* transform_node_Cond */
 
-typedef ir_node* (*recursive_transform) (ir_node *n);
+/**
+ * Prototype of a recursive transform function
+ * for bitwise distributive transformations.
+ */
+typedef ir_node* (*recursive_transform)(ir_node *n);
 
 /**
  * makes use of distributive laws for and, or, eor
@@ -2988,7 +3015,7 @@ static ir_node *transform_bitwise_distributive(ir_node *n,
                                n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
                        } else {
                                n = exact_copy(a);
-                               set_irn_n(n, -1, blk);
+                               set_nodes_block(n, blk);
                                set_binop_left(n, new_n);
                                set_binop_right(n, c);
                                add_identities(current_ir_graph->value_table, n);
@@ -3331,13 +3358,15 @@ static ir_node *transform_node_Minus(ir_node *n) {
                ir_node *mul_r = get_Mul_right(a);
                if (is_Const(mul_r)) {
                        tarval   *tv    = tarval_neg(get_Const_tarval(mul_r));
-                       ir_node  *cnst  = new_Const(mode, tv);
-                       dbg_info *dbg   = get_irn_dbg_info(a);
-                       ir_graph *irg   = current_ir_graph;
-                       ir_node  *block = get_nodes_block(a);
-                       n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
-                       DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
-                       return n;
+                       if(tv != tarval_bad) {
+                               ir_node  *cnst  = new_Const(mode, tv);
+                               dbg_info *dbg   = get_irn_dbg_info(a);
+                               ir_graph *irg   = current_ir_graph;
+                               ir_node  *block = get_nodes_block(a);
+                               n = new_rd_Mul(dbg, irg, block, mul_l, cnst, mode);
+                               DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_MUL_C);
+                               return n;
+                       }
                }
        }
 
@@ -3624,11 +3653,9 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
                break;
        }
 
-       /* remove Casts */
-       if (is_Cast(left))
-               left = get_Cast_op(left);
-       if (is_Cast(right))
-               right = get_Cast_op(right);
+       /* remove Casts of both sides */
+       left  = skip_Cast(left);
+       right = skip_Cast(right);
 
        /* Remove unnecessary conversions */
        /* TODO handle constants */
@@ -3663,7 +3690,7 @@ static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
                }
        }
 
-       /* remove operation of both sides if possible */
+       /* remove operation on both sides if possible */
        if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
                /*
                 * The following operations are NOT safe for floating point operations, for instance
@@ -4841,28 +4868,35 @@ static ir_node *transform_node_Psi(ir_node *n) {
  * of the other sync to our own inputs
  */
 static ir_node *transform_node_Sync(ir_node *n) {
-       int i, arity;
+       int arity = get_Sync_n_preds(n);
+       int i;
+
+       for (i = 0; i < arity;) {
+               ir_node *pred = get_Sync_pred(n, i);
+               int      pred_arity;
+               int      j;
 
-       arity = get_irn_arity(n);
-       for(i = 0; i < get_irn_arity(n); /*empty*/) {
-               int i2, arity2;
-               ir_node *in = get_irn_n(n, i);
-               if(!is_Sync(in)) {
+               if (!is_Sync(pred)) {
                        ++i;
                        continue;
                }
 
-               /* set sync input 0 instead of the sync */
-               set_irn_n(n, i, get_irn_n(in, 0));
-               /* so we check this input again for syncs */
-
-               /* append all other inputs of the sync to our sync */
-               arity2 = get_irn_arity(in);
-               for(i2 = 1; i2 < arity2; ++i2) {
-                       ir_node *in_in = get_irn_n(in, i2);
-                       add_irn_n(n, in_in);
-                       /* increase arity so we also check the new inputs for syncs */
-                       arity++;
+               del_Sync_n(n, i);
+               --arity;
+
+               pred_arity = get_Sync_n_preds(pred);
+               for (j = 0; j < pred_arity; ++j) {
+                       ir_node *pred_pred = get_Sync_pred(pred, j);
+                       int      k;
+
+                       for (k = 0;; ++k) {
+                               if (k >= arity) {
+                                       add_irn_n(n, pred_pred);
+                                       ++arity;
+                                       break;
+                               }
+                               if (get_Sync_pred(n, k) == pred_pred) break;
+                       }
                }
        }
 
@@ -4988,14 +5022,14 @@ static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
        const symconst_attr *pa = get_irn_symconst_attr(a);
        const symconst_attr *pb = get_irn_symconst_attr(b);
-       return (pa->num        != pb->num)
+       return (pa->kind       != pb->kind)
            || (pa->sym.type_p != pb->sym.type_p)
            || (pa->tp         != pb->tp);
 }  /* node_cmp_attr_SymConst */
 
 /** Compares the attributes of two Call nodes. */
 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
-       return (get_irn_call_attr(a) != get_irn_call_attr(b));
+       return get_irn_call_attr(a) != get_irn_call_attr(b);
 }  /* node_cmp_attr_Call */
 
 /** Compares the attributes of two Sel nodes. */
@@ -5015,8 +5049,8 @@ static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
        /* we can only enter this function if both nodes have the same number of inputs,
           hence it is enough to check if one of them is a Phi0 */
        if (is_Phi0(a)) {
-               /* check the Phi0 attribute */
-               return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
+               /* check the Phi0 pos attribute */
+               return get_irn_phi_attr(a)->u.pos != get_irn_phi_attr(b)->u.pos;
        }
        return 0;
 }  /* node_cmp_attr_Phi */
@@ -5055,6 +5089,49 @@ static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
                get_Store_volatility(b) == volatility_is_volatile);
 }  /* node_cmp_attr_Store */
 
+/** Compares two exception attributes */
+static int node_cmp_exception(ir_node *a, ir_node *b) {
+       const except_attr *ea = get_irn_except_attr(a);
+       const except_attr *eb = get_irn_except_attr(b);
+
+       return ea->pin_state != eb->pin_state;
+}
+
+#define node_cmp_attr_Bound  node_cmp_exception
+
+/** Compares the attributes of two Div nodes. */
+static int node_cmp_attr_Div(ir_node *a, ir_node *b) {
+       const divmod_attr *ma = get_irn_divmod_attr(a);
+       const divmod_attr *mb = get_irn_divmod_attr(b);
+       return ma->exc.pin_state != mb->exc.pin_state ||
+                  ma->res_mode      != mb->res_mode ||
+                  ma->no_remainder  != mb->no_remainder;
+}  /* node_cmp_attr_Div */
+
+/** Compares the attributes of two DivMod nodes. */
+static int node_cmp_attr_DivMod(ir_node *a, ir_node *b) {
+       const divmod_attr *ma = get_irn_divmod_attr(a);
+       const divmod_attr *mb = get_irn_divmod_attr(b);
+       return ma->exc.pin_state != mb->exc.pin_state ||
+                  ma->res_mode      != mb->res_mode;
+}  /* node_cmp_attr_DivMod */
+
+/** Compares the attributes of two Mod nodes. */
+static int node_cmp_attr_Mod(ir_node *a, ir_node *b) {
+       const divmod_attr *ma = get_irn_divmod_attr(a);
+       const divmod_attr *mb = get_irn_divmod_attr(b);
+       return ma->exc.pin_state != mb->exc.pin_state ||
+                  ma->res_mode      != mb->res_mode;
+}  /* node_cmp_attr_Mod */
+
+/** Compares the attributes of two Quot nodes. */
+static int node_cmp_attr_Quot(ir_node *a, ir_node *b) {
+       const divmod_attr *ma = get_irn_divmod_attr(a);
+       const divmod_attr *mb = get_irn_divmod_attr(b);
+       return ma->exc.pin_state != mb->exc.pin_state ||
+                  ma->res_mode      != mb->res_mode;
+}  /* node_cmp_attr_Quot */
+
 /** Compares the attributes of two Confirm nodes. */
 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
        return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
@@ -5138,6 +5215,12 @@ static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
        CASE(Store);
        CASE(Confirm);
        CASE(ASM);
+       CASE(Div);
+       CASE(DivMod);
+       CASE(Mod);
+       CASE(Quot);
+       CASE(Bound);
+       /* FIXME CopyB */
        default:
          /* leave NULL */;
        }