fixed mode suffix emitter
[libfirm] / ir / be / ia32 / ia32_optimize.c
index b5a5607..c464538 100644 (file)
@@ -21,6 +21,7 @@
 #include "irgmod.h"
 #include "irgwalk.h"
 #include "height.h"
+#include "irbitset.h"
 
 #include "../be_t.h"
 #include "../beabi.h"
 #include "gen_ia32_regalloc_if.h"     /* the generated interface (register type and class defenitions) */
 #include "ia32_transform.h"
 #include "ia32_dbg_stat.h"
+#include "ia32_util.h"
+
+typedef struct _ia32_place_env_t {
+       ia32_code_gen_t *cg;
+       bitset_t        *visited;
+} ia32_place_env_t;
 
 typedef enum {
        IA32_AM_CAND_NONE  = 0,
@@ -44,6 +51,7 @@ typedef enum {
 #define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
 
 typedef int is_op_func_t(const ir_node *n);
+typedef ir_node *load_func_t(dbg_info *db, ir_graph *irg, ir_node *block, ir_node *base, ir_node *index, ir_node *mem);
 
 /**
  * checks if a node represents the NOREG value
@@ -84,8 +92,6 @@ static ident *unique_id(const char *tag)
        return new_id_from_str(str);
 }
 
-
-
 /**
  * Transforms a SymConst.
  *
@@ -96,11 +102,11 @@ static ident *unique_id(const char *tag)
  * @return the created ia32 Const node
  */
 static ir_node *gen_SymConst(ia32_transform_env_t *env) {
-       ir_node  *cnst;
        dbg_info *dbg   = env->dbg;
        ir_mode  *mode  = env->mode;
        ir_graph *irg   = env->irg;
        ir_node  *block = env->block;
+       ir_node  *cnst;
 
        if (mode_is_float(mode)) {
                FP_USED(env->cg);
@@ -183,7 +189,7 @@ static entity *get_entity_for_tv(ia32_code_gen_t *cg, ir_node *cnst)
  * @return the created ia32 Const node
  */
 static ir_node *gen_Const(ia32_transform_env_t *env) {
-       ir_node *cnst;
+       ir_node *cnst, *load;
        symconst_symbol sym;
        ir_graph *irg   = env->irg;
        ir_node  *block = env->block;
@@ -203,9 +209,15 @@ static ir_node *gen_Const(ia32_transform_env_t *env) {
                }
                sym.entity_p = get_entity_for_tv(env->cg, node);
 
-               cnst = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent);
-               env->irn = cnst;
-               cnst = gen_SymConst(env);
+
+               cnst      = new_rd_SymConst(dbg, irg, block, sym, symconst_addr_ent);
+               load      = new_r_Load(irg, block, get_irg_no_mem(irg), cnst, mode);
+               load      = new_r_Proj(irg, block, load, mode, pn_Load_res);
+               env->irn  = cnst;
+               env->mode = mode_P;
+               cnst      = gen_SymConst(env);
+               set_Load_ptr(get_Proj_pred(load), cnst);
+               cnst      = load;
        }
        else {
                cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), get_irn_mode(node));
@@ -214,30 +226,74 @@ static ir_node *gen_Const(ia32_transform_env_t *env) {
        return cnst;
 }
 
-
-
 /**
  * Transforms (all) Const's into ia32_Const and places them in the
  * block where they are used (or in the cfg-pred Block in case of Phi's).
  * Additionally all reference nodes are changed into mode_Is nodes.
+ * NOTE: irn must be a firm constant!
  */
-void ia32_place_consts_set_modes(ir_node *irn, void *env) {
-       ia32_code_gen_t      *cg = env;
-       ia32_transform_env_t  tenv;
-       ir_mode              *mode;
-       ir_node              *pred, *cnst;
-       int                   i;
-       opcode                opc;
+static void ia32_transform_const(ir_node *irn, void *env) {
+       ia32_code_gen_t      *cg   = env;
+       ir_node              *cnst = NULL;
+       ia32_transform_env_t tenv;
+
+       tenv.cg   = cg;
+       tenv.irg  = cg->irg;
+       tenv.mode = get_irn_mode(irn);
+       tenv.dbg  = get_irn_dbg_info(irn);
+       tenv.irn  = irn;
+       DEBUG_ONLY(tenv.mod = cg->mod;)
+
+       /* place const either in the smallest dominator of all its users or the original block */
+       if (cg->opt & IA32_OPT_PLACECNST)
+               tenv.block = node_users_smallest_common_dominator(irn, 1);
+       else
+               tenv.block = get_nodes_block(irn);
 
-       if (is_Block(irn))
+       switch (get_irn_opcode(irn)) {
+               case iro_Const:
+                       cnst = gen_Const(&tenv);
+                       break;
+               case iro_SymConst:
+                       cnst = gen_SymConst(&tenv);
+                       break;
+               default:
+                       assert(0 && "Wrong usage of ia32_transform_const!");
+       }
+
+       assert(cnst && "Could not create ia32 Const");
+
+       /* set the new ia32 const */
+       exchange(irn, cnst);
+}
+
+/**
+ * Transform all firm consts and assure, we visit each const only once.
+ */
+static void ia32_place_consts_walker(ir_node *irn, void *env) {
+       ia32_place_env_t *penv = env;
+       opcode           opc   = get_irn_opcode(irn);
+
+       /* transform only firm consts which are not already visited */
+       if ((opc != iro_Const && opc != iro_SymConst) || bitset_is_set(penv->visited, get_irn_idx(irn)))
                return;
 
-       mode = get_irn_mode(irn);
+       /* mark const visited */
+       bitset_set(penv->visited, get_irn_idx(irn));
+
+       ia32_transform_const(irn, penv->cg);
+}
 
-       /* transform all reference nodes into mode_Is nodes */
-       if (mode_is_reference(mode)) {
-               mode = mode_Is;
-               set_irn_mode(irn, mode);
+/**
+ * Replace reference modes with mode_Iu and preserve store value modes.
+ */
+static void ia32_set_modes(ir_node *irn, void *env) {
+       if (is_Block(irn))
+               return;
+
+       /* transform all reference nodes into mode_Iu nodes */
+       if (mode_is_reference(get_irn_mode(irn))) {
+               set_irn_mode(irn, mode_Iu);
        }
 
        /*
@@ -248,51 +304,33 @@ void ia32_place_consts_set_modes(ir_node *irn, void *env) {
        if (get_irn_opcode(irn) == iro_Store) {
                set_irn_link(irn, get_irn_mode(get_Store_value(irn)));
        }
+}
 
-       tenv.block    = get_nodes_block(irn);
-       tenv.cg       = cg;
-       tenv.irg      = cg->irg;
-       DEBUG_ONLY(tenv.mod      = cg->mod;)
-
-       /* Loop over all predecessors and check for Sym/Const nodes */
-       for (i = get_irn_arity(irn) - 1; i >= 0; --i) {
-               pred      = get_irn_n(irn, i);
-               cnst      = NULL;
-               opc       = get_irn_opcode(pred);
-               tenv.irn  = pred;
-               tenv.mode = get_irn_mode(pred);
-               tenv.dbg  = get_irn_dbg_info(pred);
-
-               /* If it's a Phi, then we need to create the */
-               /* new Const in it's predecessor block       */
-               if (is_Phi(irn)) {
-                       tenv.block = get_Block_cfgpred_block(get_nodes_block(irn), i);
-               }
-
-               /* put the const into the block where the original const was */
-               if (! (cg->opt & IA32_OPT_PLACECNST)) {
-                       tenv.block = get_nodes_block(pred);
-               }
-
-               switch (opc) {
-                       case iro_Const:
-                               cnst = gen_Const(&tenv);
-                               break;
-                       case iro_SymConst:
-                               cnst = gen_SymConst(&tenv);
-                               break;
-                       default:
-                               break;
-               }
+/**
+ * Walks over the graph, transforms all firm consts into ia32 consts
+ * and places them into the "best" block.
+ * @param cg  The ia32 codegenerator object
+ */
+static void ia32_transform_all_firm_consts(ia32_code_gen_t *cg) {
+       ia32_place_env_t penv;
 
-               /* if we found a const, then set it */
-               if (cnst) {
-                       set_irn_n(irn, i, cnst);
-               }
-       }
+       penv.cg      = cg;
+       penv.visited = bitset_irg_malloc(cg->irg);
+       irg_walk_graph(cg->irg, NULL, ia32_place_consts_walker, &penv);
+       bitset_free(penv.visited);
 }
 
-
+/* Place all consts and change pointer arithmetics into unsigned integer arithmetics. */
+void ia32_pre_transform_phase(ia32_code_gen_t *cg) {
+       /*
+               We need to transform the consts twice:
+               - the psi condition tree transformer needs existing constants to be ia32 constants
+               - the psi condition tree transformer inserts new firm constants which need to be transformed
+       */
+       ia32_transform_all_firm_consts(cg);
+       irg_walk_graph(cg->irg, ia32_set_modes, ia32_transform_psi_cond_tree, cg);
+       ia32_transform_all_firm_consts(cg);
+}
 
 /********************************************************************************************************
  *  _____                _           _         ____        _   _           _          _   _
@@ -475,6 +513,7 @@ static void ia32_create_Push(ir_node *irn, ia32_code_gen_t *cg) {
        /* ok, translate into Push */
        edge       = get_irn_out_edge_first(irn);
        old_proj_M = get_edge_src_irn(edge);
+       bl         = get_nodes_block(irn);
 
        next = sched_next(irn);
        sched_remove(irn);
@@ -496,7 +535,6 @@ static void ia32_create_Push(ir_node *irn, ia32_code_gen_t *cg) {
                mem   = new_r_Sync(irg, bl, 2, in);
        }
 
-       bl   = get_nodes_block(irn);
        push = new_rd_ia32_Push(NULL, irg, bl, be_get_IncSP_pred(sp), val, mem);
        proj_res = new_r_Proj(irg, bl, push, get_irn_mode(sp), pn_ia32_Push_stack);
        proj_M   = new_r_Proj(irg, bl, push, mode_M, pn_ia32_Push_M);
@@ -613,13 +651,16 @@ static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) {
 void ia32_peephole_optimization(ir_node *irn, void *env) {
        ia32_code_gen_t *cg = env;
 
-       if (is_ia32_TestJmp(irn))
-               ia32_optimize_TestJmp(irn, cg);
-       else if (is_ia32_CondJmp(irn))
-               ia32_optimize_CondJmp(irn, cg);
+       /* AMD CPUs want explicit compare before conditional jump  */
+       if (! ARCH_AMD(cg->opt_arch)) {
+               if (is_ia32_TestJmp(irn))
+                       ia32_optimize_TestJmp(irn, cg);
+               else if (is_ia32_CondJmp(irn))
+                       ia32_optimize_CondJmp(irn, cg);
+       }
        /* seems to be buggy when using Pushes */
-//     else if (be_is_IncSP(irn))
-//             ia32_optimize_IncSP(irn, cg);
+       else if (be_is_IncSP(irn))
+               ia32_optimize_IncSP(irn, cg);
        else if (is_ia32_Store(irn))
                ia32_create_Push(irn, cg);
 }
@@ -656,48 +697,6 @@ static int ia32_get_irn_n_edges(const ir_node *irn) {
        return cnt;
 }
 
-/**
- * Returns the first mode_M Proj connected to irn.
- */
-static ir_node *get_mem_proj(const ir_node *irn) {
-       const ir_edge_t *edge;
-       ir_node         *src;
-
-       assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
-
-       foreach_out_edge(irn, edge) {
-               src = get_edge_src_irn(edge);
-
-               assert(is_Proj(src) && "Proj expected");
-
-               if (get_irn_mode(src) == mode_M)
-                       return src;
-       }
-
-       return NULL;
-}
-
-/**
- * Returns the first Proj with mode != mode_M connected to irn.
- */
-static ir_node *get_res_proj(const ir_node *irn) {
-       const ir_edge_t *edge;
-       ir_node         *src;
-
-       assert(get_irn_mode(irn) == mode_T && "expected mode_T node");
-
-       foreach_out_edge(irn, edge) {
-               src = get_edge_src_irn(edge);
-
-               assert(is_Proj(src) && "Proj expected");
-
-               if (get_irn_mode(src) != mode_M)
-                       return src;
-       }
-
-       return NULL;
-}
-
 /**
  * Determines if pred is a Proj and if is_op_func returns true for it's predecessor.
  *
@@ -790,7 +789,8 @@ static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const i
        ir_node *in, *load, *other, *left, *right;
        int      n, is_cand = 0, cand;
 
-       if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn))
+       if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn) ||
+               is_ia32_GetST0(irn) || is_ia32_SetST0(irn))
                return 0;
 
        left  = get_irn_n(irn, 2);
@@ -805,10 +805,16 @@ static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const i
                load  = get_Proj_pred(in);
                other = right;
 
+               /* 8bit Loads are not supported, they cannot be used with every register */
+               if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16)
+                       is_cand = 0;
+
                /* If there is a data dependency of other irn from load: cannot use AM */
-               if (get_nodes_block(other) == block) {
+               if (is_cand && get_nodes_block(other) == block) {
                        other   = skip_Proj(other);
                        is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
+                       /* this could happen in loops */
+                       is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
                }
        }
 
@@ -823,10 +829,16 @@ static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const i
                load  = get_Proj_pred(in);
                other = left;
 
+               /* 8bit Loads are not supported, they cannot be used with every register */
+               if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16)
+                       is_cand = 0;
+
                /* If there is a data dependency of other irn from load: cannot use load */
-               if (get_nodes_block(other) == block) {
+               if (is_cand && get_nodes_block(other) == block) {
                        other   = skip_Proj(other);
                        is_cand = heights_reachable_in_block(h, other, load) ? 0 : is_cand;
+                       /* this could happen in loops */
+                       is_cand = heights_reachable_in_block(h, load, irn) ? 0 : is_cand;
                }
        }
 
@@ -897,8 +909,6 @@ typedef enum _ia32_take_lea_attr {
 static int do_new_lea(ir_node *irn, ir_node *base, ir_node *index, ir_node *lea,
                int have_am_sc, ia32_code_gen_t *cg)
 {
-       ir_node *lea_base = get_irn_n(lea, 0);
-       ir_node *lea_idx  = get_irn_n(lea, 1);
        entity  *irn_ent  = get_ia32_frame_ent(irn);
        entity  *lea_ent  = get_ia32_frame_ent(lea);
        int      ret_val  = 0;
@@ -988,6 +998,34 @@ static int do_new_lea(ir_node *irn, ir_node *base, ir_node *index, ir_node *lea,
        return ret_val;
 }
 
+/**
+ * Adds res before irn into schedule if irn was scheduled.
+ * @param irn  The schedule point
+ * @param res  The node to be scheduled
+ */
+static INLINE void try_add_to_sched(ir_node *irn, ir_node *res) {
+       if (sched_is_scheduled(irn))
+               sched_add_before(irn, res);
+}
+
+/**
+ * Removes irn from schedule if it was scheduled. If irn is a mode_T node
+ * all it's Projs are removed as well.
+ * @param irn  The irn to be removed from schedule
+ */
+static INLINE void try_remove_from_sched(ir_node *irn) {
+       if (sched_is_scheduled(irn)) {
+               if (get_irn_mode(irn) == mode_T) {
+                       const ir_edge_t *edge;
+                       foreach_out_edge(irn, edge) {
+                               ir_node *proj = get_edge_src_irn(edge);
+                               if (sched_is_scheduled(proj))
+                                       sched_remove(proj);
+                       }
+               }
+               sched_remove(irn);
+       }
+}
 
 /**
  * Folds Add or Sub to LEA if possible
@@ -1091,6 +1129,8 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
 
                if (temp == base)
                        base = noreg;
+               else if (temp == right)
+                       right = noreg;
        }
 
        if (isadd) {
@@ -1251,25 +1291,46 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
                DBG((mod, LEVEL_1, "\tLEA [%+F + %+F * %d + %s]\n", base, index, scale, get_ia32_am_offs(res)));
 
                /* we will exchange it, report here before the Proj is created */
-               if (shift && lea && lea_o)
+               if (shift && lea && lea_o) {
+                       try_remove_from_sched(shift);
+                       try_remove_from_sched(lea);
+                       try_remove_from_sched(lea_o);
                        DBG_OPT_LEA4(irn, lea_o, lea, shift, res);
-               else if (shift && lea)
+               }
+               else if (shift && lea) {
+                       try_remove_from_sched(shift);
+                       try_remove_from_sched(lea);
                        DBG_OPT_LEA3(irn, lea, shift, res);
-               else if (shift && lea_o)
+               }
+               else if (shift && lea_o) {
+                       try_remove_from_sched(shift);
+                       try_remove_from_sched(lea_o);
                        DBG_OPT_LEA3(irn, lea_o, shift, res);
-               else if (lea && lea_o)
+               }
+               else if (lea && lea_o) {
+                       try_remove_from_sched(lea);
+                       try_remove_from_sched(lea_o);
                        DBG_OPT_LEA3(irn, lea_o, lea, res);
-               else if (shift)
+               }
+               else if (shift) {
+                       try_remove_from_sched(shift);
                        DBG_OPT_LEA2(irn, shift, res);
-               else if (lea)
+               }
+               else if (lea) {
+                       try_remove_from_sched(lea);
                        DBG_OPT_LEA2(irn, lea, res);
-               else if (lea_o)
+               }
+               else if (lea_o) {
+                       try_remove_from_sched(lea_o);
                        DBG_OPT_LEA2(irn, lea_o, res);
+               }
                else
                        DBG_OPT_LEA1(irn, res);
 
                /* get the result Proj of the Add/Sub */
-               irn = get_res_proj(irn);
+               try_add_to_sched(irn, res);
+               try_remove_from_sched(irn);
+               irn = ia32_get_res_proj(irn);
 
                assert(irn && "Couldn't find result proj");
 
@@ -1313,6 +1374,8 @@ static void merge_loadstore_lea(ir_node *irn, ir_node *lea) {
        set_irn_n(irn, 0, get_irn_n(lea, 0));
        set_irn_n(irn, 1, get_irn_n(lea, 1));
 
+       try_remove_from_sched(lea);
+
        /* clear remat flag */
        set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
 
@@ -1394,7 +1457,7 @@ static void optimize_lea(ir_node *irn, void *env) {
                        foreach_out_edge_safe(left, edge, ne) {
                                src = get_edge_src_irn(edge);
 
-                               if (src && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) {
+                               if (src && (get_edge_src_pos(edge) == 0) && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) {
                                        DBG((cg->mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
                                        if (! is_ia32_got_lea(src))
                                                merge_loadstore_lea(src, left);
@@ -1423,7 +1486,7 @@ static void optimize_am(ir_node *irn, void *env) {
        int               need_exchange_on_fail = 0;
        DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
 
-       if (! is_ia32_irn(irn))
+       if (! is_ia32_irn(irn) || is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn))
                return;
 
        block    = get_nodes_block(irn);
@@ -1467,31 +1530,31 @@ static void optimize_am(ir_node *irn, void *env) {
                }
 
                /* normalize commutative ops */
-               if (node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_LEFT)) {
+               if (node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_RIGHT)) {
 
-                       /* Assure that right operand is always a Load if there is one    */
-                       /* because non-commutative ops can only use Dest AM if the right */
-                       /* operand is a load, so we only need to check right operand.    */
+                       /* Assure that left operand is always a Load if there is one    */
+                       /* because non-commutative ops can only use Dest AM if the left */
+                       /* operand is a load, so we only need to check left operand.    */
 
                        exchange_left_right(irn, &left, &right, 3, 2);
                        need_exchange_on_fail = 1;
 
                        /* now: load is right */
-                       cand = IA32_AM_CAND_RIGHT;
+                       cand = IA32_AM_CAND_LEFT;
                }
 
                /* check for Store -> op -> Load */
 
                /* Store -> op -> Load optimization is only possible if supported by op */
                /* and if right operand is a Load                                       */
-               if ((get_ia32_am_support(irn) & ia32_am_Dest) && (cand & IA32_AM_CAND_RIGHT))
+               if ((get_ia32_am_support(irn) & ia32_am_Dest) && (cand & IA32_AM_CAND_LEFT))
                {
                        /* An address mode capable op always has a result Proj.                  */
                        /* If this Proj is used by more than one other node, we don't need to    */
                        /* check further, otherwise we check for Store and remember the address, */
                        /* the Store points to. */
 
-                       succ = get_res_proj(irn);
+                       succ = ia32_get_res_proj(irn);
                        assert(succ && "Couldn't find result proj");
 
                        addr_b = NULL;
@@ -1512,12 +1575,13 @@ static void optimize_am(ir_node *irn, void *env) {
                        if (store) {
                                /* we found a Store as single user: Now check for Load */
 
+                               /* skip the Proj for easier access */
+                               load = is_Proj(right) ? (is_ia32_Load(get_Proj_pred(right)) ? get_Proj_pred(right) : NULL) : NULL;
+
                                /* Extra check for commutative ops with two Loads */
-                               /* -> put the interesting Load right              */
-                               if (node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_BOTH)) {
-                                       if ((addr_b == get_irn_n(get_Proj_pred(left), 0)) &&
-                                               (addr_i == get_irn_n(get_Proj_pred(left), 1)))
-                                       {
+                               /* -> put the interesting Load left               */
+                               if (load && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_BOTH)) {
+                                       if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
                                                /* We exchange left and right, so it's easier to kill     */
                                                /* the correct Load later and to handle unary operations. */
                                                exchange_left_right(irn, &left, &right, 3, 2);
@@ -1526,11 +1590,11 @@ static void optimize_am(ir_node *irn, void *env) {
                                }
 
                                /* skip the Proj for easier access */
-                               load = get_Proj_pred(right);
+                               load = get_Proj_pred(left);
 
                                /* Compare Load and Store address */
                                if (load_store_addr_is_equal(load, store, addr_b, addr_i)) {
-                                       /* Right Load is from same address, so we can */
+                                       /* Left Load is from same address, so we can */
                                        /* disconnect the Load and Store here        */
 
                                        /* set new base, index and attributes */
@@ -1554,7 +1618,7 @@ static void optimize_am(ir_node *irn, void *env) {
                                        if (get_irn_arity(irn) == 5) {
                                                /* binary AMop */
                                                set_irn_n(irn, 4, get_irn_n(load, 2));
-                                               set_irn_n(irn, 3, noreg_gp);
+                                               set_irn_n(irn, 2, noreg_gp);
                                        }
                                        else {
                                                /* unary AMop */
@@ -1563,13 +1627,15 @@ static void optimize_am(ir_node *irn, void *env) {
                                        }
 
                                        /* connect the memory Proj of the Store to the op */
-                                       mem_proj = get_mem_proj(store);
+                                       mem_proj = ia32_get_proj_for_mode(store, mode_M);
                                        set_Proj_pred(mem_proj, irn);
                                        set_Proj_proj(mem_proj, 1);
 
                                        /* clear remat flag */
                                        set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
 
+                                       try_remove_from_sched(load);
+                                       try_remove_from_sched(store);
                                        DBG_OPT_AM_D(load, store, irn);
 
                                        DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
@@ -1596,83 +1662,74 @@ static void optimize_am(ir_node *irn, void *env) {
                need_exchange_on_fail = 0;
 
                /* normalize commutative ops */
-               if (check_am_src && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_RIGHT)) {
+               if (check_am_src && node_is_ia32_comm(irn) && (cand == IA32_AM_CAND_LEFT)) {
 
-                       /* Assure that left operand is always a Load if there is one */
-                       /* because non-commutative ops can only use Source AM if the */
-                       /* left operand is a Load, so we only need to check the left */
-                       /* operand afterwards.                                       */
+                       /* Assure that right operand is always a Load if there is one  */
+                       /* because non-commutative ops can only use Source AM if the   */
+                       /* right operand is a Load, so we only need to check the right */
+                       /* operand afterwards.                                         */
 
                        exchange_left_right(irn, &left, &right, 3, 2);
                        need_exchange_on_fail = 1;
 
                        /* now: load is left */
-                       cand = IA32_AM_CAND_LEFT;
+                       cand = IA32_AM_CAND_RIGHT;
                }
 
-               /* optimize op -> Load iff Load is only used by this op   */
-               /* and left operand is a Load which only used by this irn */
-               if (check_am_src               &&
-                       (cand & IA32_AM_CAND_LEFT) &&
-                       (ia32_get_irn_n_edges(left) == 1))
+               /* optimize op -> Load iff Load is only used by this op    */
+               /* and right operand is a Load which only used by this irn */
+               if (check_am_src                &&
+                       (cand & IA32_AM_CAND_RIGHT) &&
+                       (get_irn_arity(irn) == 5)   &&
+                       (ia32_get_irn_n_edges(right) == 1))
                {
-                       left = get_Proj_pred(left);
+                       right = get_Proj_pred(right);
 
-                       addr_b = get_irn_n(left, 0);
-                       addr_i = get_irn_n(left, 1);
+                       addr_b = get_irn_n(right, 0);
+                       addr_i = get_irn_n(right, 1);
 
                        /* set new base, index and attributes */
                        set_irn_n(irn, 0, addr_b);
                        set_irn_n(irn, 1, addr_i);
-                       add_ia32_am_offs(irn, get_ia32_am_offs(left));
-                       set_ia32_am_scale(irn, get_ia32_am_scale(left));
-                       set_ia32_am_flavour(irn, get_ia32_am_flavour(left));
+                       add_ia32_am_offs(irn, get_ia32_am_offs(right));
+                       set_ia32_am_scale(irn, get_ia32_am_scale(right));
+                       set_ia32_am_flavour(irn, get_ia32_am_flavour(right));
                        set_ia32_op_type(irn, ia32_AddrModeS);
-                       set_ia32_frame_ent(irn, get_ia32_frame_ent(left));
-                       set_ia32_ls_mode(irn, get_ia32_ls_mode(left));
+                       set_ia32_frame_ent(irn, get_ia32_frame_ent(right));
+                       set_ia32_ls_mode(irn, get_ia32_ls_mode(right));
 
-                       set_ia32_am_sc(irn, get_ia32_am_sc(left));
-                       if (is_ia32_am_sc_sign(left))
+                       set_ia32_am_sc(irn, get_ia32_am_sc(right));
+                       if (is_ia32_am_sc_sign(right))
                                set_ia32_am_sc_sign(irn);
 
                        /* clear remat flag */
                        set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
 
-                       if (is_ia32_use_frame(left))
+                       if (is_ia32_use_frame(right))
                                set_ia32_use_frame(irn);
 
                        /* connect to Load memory */
-                       if (get_irn_arity(irn) == 5) {
-                               /* binary AMop */
-                               set_irn_n(irn, 4, get_irn_n(left, 2));
-
-                               /* this is only needed for Compares, but currently ALL nodes
-                                * have this attribute :-) */
-                               set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
-
-                               /* disconnect from Load */
-                               /* (make second op -> first, set second in to noreg) */
-                               set_irn_n(irn, 2, get_irn_n(irn, 3));
-                               set_irn_n(irn, 3, noreg_gp);
-                       }
-                       else {
-                               /* unary AMop */
-                               set_irn_n(irn, 3, get_irn_n(left, 2));
+                       set_irn_n(irn, 4, get_irn_n(right, 2));
 
-                               /* disconnect from Load */
-                               set_irn_n(irn, 2, noreg_gp);
-                       }
+                       /* this is only needed for Compares, but currently ALL nodes
+                        * have this attribute :-) */
+                       set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
 
-                       DBG_OPT_AM_S(left, irn);
+                       /* disconnect from Load */
+                       set_irn_n(irn, 3, noreg_gp);
+
+                       DBG_OPT_AM_S(right, irn);
 
                        /* If Load has a memory Proj, connect it to the op */
-                       mem_proj = get_mem_proj(left);
+                       mem_proj = ia32_get_proj_for_mode(right, mode_M);
                        if (mem_proj) {
                                set_Proj_pred(mem_proj, irn);
                                set_Proj_proj(mem_proj, 1);
                        }
 
-                       DB((mod, LEVEL_1, "merged with %+F into source AM\n", left));
+                       try_remove_from_sched(right);
+
+                       DB((mod, LEVEL_1, "merged with %+F into source AM\n", right));
                }
                else {
                        /* was exchanged but optimize failed: exchange back */
@@ -1704,6 +1761,9 @@ void ia32_optimize_addressmode(ia32_code_gen_t *cg) {
                irg_walk_blkwise_graph(cg->irg, NULL, optimize_lea, cg);
        }
 
+       if (cg->dump)
+               be_dump(cg->irg, "-lea", dump_ir_block_graph_sched);
+
        if (cg->opt & IA32_OPT_DOAM) {
                /* we need height information for am optimization */
                heights_t *h = heights_new(cg->irg);