register slots are now automatically allocated together with the $ARCH_attribute
[libfirm] / ir / be / ia32 / ia32_optimize.c
index 6a88bda..769bebb 100644 (file)
 #include "bearch_ia32_t.h"
 #include "gen_ia32_regalloc_if.h"     /* the generated interface (register type and class defenitions) */
 #include "ia32_transform.h"
-
-/*----*/
-
-#include "irhooks.h"
-#include "dbginfo_t.h"
-#include "firmstat.h"
-
-/**
- * Merge the debug info due to a LEA creation.
- *
- * @param oldn  the node
- * @param n     the new constant holding the value
- */
-#define DBG_OPT_LEA(oldn, n)                                \
-  do {                                                         \
-         hook_merge_nodes(&n, 1, &oldn, 1, FS_BE_IA32_LEA);      \
-    __dbg_info_merge_pair(n, oldn, dbg_backend);                 \
-  } while(0)
-
+#include "ia32_dbg_stat.h"
 
 #undef is_NoMem
 #define is_NoMem(irn) (get_irn_op(irn) == op_NoMem)
@@ -102,15 +84,17 @@ static ir_node *gen_SymConst(ia32_transform_env_t *env) {
        ir_node  *block = env->block;
 
        if (mode_is_float(mode)) {
+               FP_USED(env->cg);
                if (USE_SSE2(env->cg))
-                       cnst = new_rd_ia32_fConst(dbg, irg, block, mode);
+                       cnst = new_rd_ia32_xConst(dbg, irg, block, get_irg_no_mem(irg), mode);
                else
-                       cnst = new_rd_ia32_vfConst(dbg, irg, block, mode);
-       }
-       else {
-               cnst = new_rd_ia32_Const(dbg, irg, block, mode);
+                       cnst = new_rd_ia32_vfConst(dbg, irg, block, get_irg_no_mem(irg), mode);
        }
+       else
+               cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), mode);
+
        set_ia32_Const_attr(cnst, env->irn);
+
        return cnst;
 }
 
@@ -189,6 +173,7 @@ static ir_node *gen_Const(ia32_transform_env_t *env) {
        ir_mode  *mode  = env->mode;
 
        if (mode_is_float(mode)) {
+               FP_USED(env->cg);
                if (! USE_SSE2(env->cg)) {
                        cnst_classify_t clss = classify_Const(node);
 
@@ -204,7 +189,7 @@ static ir_node *gen_Const(ia32_transform_env_t *env) {
                cnst = gen_SymConst(env);
        }
        else {
-               cnst = new_rd_ia32_Const(dbg, irg, block, get_irn_mode(node));
+               cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), get_irn_mode(node));
                set_ia32_Const_attr(cnst, node);
        }
        return cnst;
@@ -257,7 +242,7 @@ void ia32_place_consts_set_modes(ir_node *irn, void *env) {
                }
 
                /* put the const into the block where the original const was */
-               if (! cg->opt.placecnst) {
+               if (! (cg->opt & IA32_OPT_PLACECNST)) {
                        tenv.block = get_nodes_block(pred);
                }
 
@@ -388,7 +373,7 @@ static int is_CondJmp_replacement(ir_node *cand, ir_node *irn) {
        int same_args = 1;
 
        for (i = 0; i < n; i++) {
-               if (get_irn_n(cand, i) == get_irn_n(irn, i)) {
+               if (get_irn_n(cand, i) != get_irn_n(irn, i)) {
                        same_args = 0;
                        break;
                }
@@ -412,13 +397,125 @@ static void ia32_optimize_CondJmp(ir_node *irn, ia32_code_gen_t *cg) {
 
        if (replace) {
                DBG((cg->mod, LEVEL_1, "replacing %+F by ", irn));
+               DBG_OPT_CJMP(irn);
 
-               set_irn_op(irn, op_ia32_CJmp);
+               set_irn_op(irn, op_ia32_CJmpAM);
 
                DB((cg->mod, LEVEL_1, "%+F\n", irn));
        }
 }
 
+/**
+ * Creates a Push from Store(IncSP(gp_reg_size))
+ */
+static void ia32_create_Push(ir_node *irn, ia32_code_gen_t *cg) {
+       ir_node *sp  = get_irn_n(irn, 0);
+       ir_node *val, *next, *push, *bl, *proj_M, *proj_res, *old_proj_M;
+       const ir_edge_t *edge;
+
+       if (get_ia32_am_offs(irn) || !be_is_IncSP(sp))
+               return;
+
+       if (arch_get_irn_register(cg->arch_env, get_irn_n(irn, 1)) !=
+               &ia32_gp_regs[REG_GP_NOREG])
+               return;
+
+       val = get_irn_n(irn, 2);
+       if (mode_is_float(get_irn_mode(val)))
+               return;
+
+       if (be_get_IncSP_direction(sp) != be_stack_dir_expand ||
+               be_get_IncSP_offset(sp) != get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_gp].mode))
+               return;
+
+       /* ok, translate into Push */
+       edge = get_irn_out_edge_first(irn);
+       old_proj_M = get_edge_src_irn(edge);
+
+       next = sched_next(irn);
+       sched_remove(irn);
+       sched_remove(sp);
+
+       bl   = get_nodes_block(irn);
+       push = new_rd_ia32_Push(NULL, current_ir_graph, bl,
+               be_get_IncSP_pred(sp), val, be_get_IncSP_mem(sp));
+       proj_res = new_r_Proj(current_ir_graph, bl, push, get_irn_mode(sp), pn_ia32_Push_stack);
+       proj_M   = new_r_Proj(current_ir_graph, bl, push, mode_M, pn_ia32_Push_M);
+
+       /* copy a possible constant from the store */
+       set_ia32_id_cnst(push, get_ia32_id_cnst(irn));
+       set_ia32_immop_type(push, get_ia32_immop_type(irn));
+
+       /* the push must have SP out register */
+       arch_set_irn_register(cg->arch_env, push, arch_get_irn_register(cg->arch_env, sp));
+
+       exchange(old_proj_M, proj_M);
+       exchange(sp, proj_res);
+       sched_add_before(next, push);
+       sched_add_after(push, proj_res);
+}
+
+/**
+ * Creates a Pop from IncSP(Load(sp))
+ */
+static void ia32_create_Pop(ir_node *irn, ia32_code_gen_t *cg) {
+       ir_node *old_proj_M = be_get_IncSP_mem(irn);
+       ir_node *load = skip_Proj(old_proj_M);
+       ir_node *old_proj_res = NULL;
+       ir_node *bl, *pop, *next, *proj_res, *proj_sp, *proj_M;
+       const ir_edge_t *edge;
+       const arch_register_t *reg, *sp;
+
+       if (! is_ia32_Load(load) || get_ia32_am_offs(load))
+               return;
+
+       if (arch_get_irn_register(cg->arch_env, get_irn_n(load, 1)) !=
+               &ia32_gp_regs[REG_GP_NOREG])
+               return;
+       if (arch_get_irn_register(cg->arch_env, get_irn_n(load, 0)) != cg->isa->arch_isa.sp)
+               return;
+
+       /* ok, translate into pop */
+       foreach_out_edge(load, edge) {
+               ir_node *succ = get_edge_src_irn(edge);
+               if (succ != old_proj_M) {
+                       old_proj_res = succ;
+                       break;
+               }
+       }
+       if (! old_proj_res) {
+               assert(0);
+               return; /* should not happen */
+       }
+
+       bl = get_nodes_block(load);
+
+       /* IncSP is typically scheduled after the load, so remove it first */
+       sched_remove(irn);
+       next = sched_next(old_proj_res);
+       sched_remove(old_proj_res);
+       sched_remove(load);
+
+       reg = arch_get_irn_register(cg->arch_env, load);
+       sp  = arch_get_irn_register(cg->arch_env, irn);
+
+       pop      = new_rd_ia32_Pop(NULL, current_ir_graph, bl, get_irn_n(irn, 0), get_irn_n(load, 2));
+       proj_res = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(old_proj_res), pn_ia32_Pop_res);
+       proj_sp  = new_r_Proj(current_ir_graph, bl, pop, get_irn_mode(irn), pn_ia32_Pop_stack);
+       proj_M   = new_r_Proj(current_ir_graph, bl, pop, mode_M, pn_ia32_Pop_M);
+
+       exchange(old_proj_M, proj_M);
+       exchange(old_proj_res, proj_res);
+       exchange(irn, proj_sp);
+
+       arch_set_irn_register(cg->arch_env, proj_res, reg);
+       arch_set_irn_register(cg->arch_env, proj_sp, sp);
+
+       sched_add_before(next, proj_sp);
+       sched_add_before(proj_sp, proj_res);
+       sched_add_before(proj_res,pop);
+}
+
 /**
  * Tries to optimize two following IncSP.
  */
@@ -445,6 +542,9 @@ static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) {
                be_set_IncSP_offset(prev, 0);
                be_set_IncSP_offset(irn, (unsigned)new_ofs);
                be_set_IncSP_direction(irn, curr_dir);
+
+               /* Omit the optimized IncSP */
+               be_set_IncSP_pred(irn, be_get_IncSP_pred(prev));
        }
 }
 
@@ -458,8 +558,11 @@ void ia32_peephole_optimization(ir_node *irn, void *env) {
                ia32_optimize_TestJmp(irn, cg);
        else if (is_ia32_CondJmp(irn))
                ia32_optimize_CondJmp(irn, cg);
-       else if (be_is_IncSP(irn))
-               ia32_optimize_IncSP(irn, cg);
+       /* seems to be buggy when using Pushes */
+//     else if (be_is_IncSP(irn))
+//             ia32_optimize_IncSP(irn, cg);
+       else if (is_ia32_Store(irn))
+               ia32_create_Push(irn, cg);
 }
 
 
@@ -644,6 +747,112 @@ static int load_store_addr_is_equal(const ir_node *load, const ir_node *store,
        return is_equal;
 }
 
+typedef enum _ia32_take_lea_attr {
+       IA32_LEA_ATTR_NONE  = 0,
+       IA32_LEA_ATTR_BASE  = (1 << 0),
+       IA32_LEA_ATTR_INDEX = (1 << 1),
+       IA32_LEA_ATTR_OFFS  = (1 << 2),
+       IA32_LEA_ATTR_SCALE = (1 << 3),
+       IA32_LEA_ATTR_AMSC  = (1 << 4),
+       IA32_LEA_ATTR_FENT  = (1 << 5)
+} ia32_take_lea_attr;
+
+/**
+ * Decides if we have to keep the LEA operand or if we can assimilate it.
+ */
+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;
+       int      is_noreg_base  = be_is_NoReg(cg, base);
+       int      is_noreg_index = be_is_NoReg(cg, index);
+       ia32_am_flavour_t am_flav = get_ia32_am_flavour(lea);
+
+       /* If the Add and the LEA both have a different frame entity set: keep */
+       if (irn_ent && lea_ent && (irn_ent != lea_ent))
+               return IA32_LEA_ATTR_NONE;
+       else if (! irn_ent && lea_ent)
+               ret_val |= IA32_LEA_ATTR_FENT;
+
+       /* If the Add and the LEA both have already an address mode symconst: keep */
+       if (have_am_sc && get_ia32_am_sc(lea))
+               return IA32_LEA_ATTR_NONE;
+       else if (get_ia32_am_sc(lea))
+               ret_val |= IA32_LEA_ATTR_AMSC;
+
+       /* Check the different base-index combinations */
+
+       if (! is_noreg_base && ! is_noreg_index) {
+               /* Assimilate if base is the lea and the LEA is just a Base + Offset calculation */
+               if ((base == lea) && ! (am_flav & ia32_I ? 1 : 0)) {
+                       if (am_flav & ia32_O)
+                               ret_val |= IA32_LEA_ATTR_OFFS;
+
+                       ret_val |= IA32_LEA_ATTR_BASE;
+               }
+               else
+                       return IA32_LEA_ATTR_NONE;
+       }
+       else if (! is_noreg_base && is_noreg_index) {
+               /* Base is set but index not */
+               if (base == lea) {
+                       /* Base points to LEA: assimilate everything */
+                       if (am_flav & ia32_O)
+                               ret_val |= IA32_LEA_ATTR_OFFS;
+                       if (am_flav & ia32_S)
+                               ret_val |= IA32_LEA_ATTR_SCALE;
+                       if (am_flav & ia32_I)
+                               ret_val |= IA32_LEA_ATTR_INDEX;
+
+                       ret_val |= IA32_LEA_ATTR_BASE;
+               }
+               else if (am_flav & ia32_B ? 0 : 1) {
+                       /* Base is not the LEA but the LEA is an index only calculation: assimilate */
+                       if (am_flav & ia32_O)
+                               ret_val |= IA32_LEA_ATTR_OFFS;
+                       if (am_flav & ia32_S)
+                               ret_val |= IA32_LEA_ATTR_SCALE;
+
+                       ret_val |= IA32_LEA_ATTR_INDEX;
+               }
+               else
+                       return IA32_LEA_ATTR_NONE;
+       }
+       else if (is_noreg_base && ! is_noreg_index) {
+               /* Index is set but not base */
+               if (index == lea) {
+                       /* Index points to LEA: assimilate everything */
+                       if (am_flav & ia32_O)
+                               ret_val |= IA32_LEA_ATTR_OFFS;
+                       if (am_flav & ia32_S)
+                               ret_val |= IA32_LEA_ATTR_SCALE;
+                       if (am_flav & ia32_B)
+                               ret_val |= IA32_LEA_ATTR_BASE;
+
+                       ret_val |= IA32_LEA_ATTR_INDEX;
+               }
+               else if (am_flav & ia32_I ? 0 : 1) {
+                       /* Index is not the LEA but the LEA is a base only calculation: assimilate */
+                       if (am_flav & ia32_O)
+                               ret_val |= IA32_LEA_ATTR_OFFS;
+                       if (am_flav & ia32_S)
+                               ret_val |= IA32_LEA_ATTR_SCALE;
+
+                       ret_val |= IA32_LEA_ATTR_BASE;
+               }
+               else
+                       return IA32_LEA_ATTR_NONE;
+       }
+       else {
+               assert(0 && "There must have been set base or index");
+       }
+
+       return ret_val;
+}
 
 
 /**
@@ -654,6 +863,9 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
        dbg_info   *dbg        = get_irn_dbg_info(irn);
        ir_node    *block      = get_nodes_block(irn);
        ir_node    *res        = irn;
+       ir_node    *shift      = NULL;
+       ir_node    *lea_o      = NULL;
+       ir_node    *lea        = NULL;
        char       *offs       = NULL;
        const char *offs_cnst  = NULL;
        char       *offs_lea   = NULL;
@@ -663,6 +875,7 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
        int         have_am_sc = 0;
        int         am_sc_sign = 0;
        ident      *am_sc      = NULL;
+       entity     *lea_ent    = NULL;
        ir_node    *left, *right, *temp;
        ir_node    *base, *index;
        ia32_am_flavour_t am_flav;
@@ -745,6 +958,7 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
                am_sc_sign = is_ia32_am_sc_sign(temp);
                have_am_sc = 1;
                dolea      = 1;
+               lea_o      = temp;
        }
 
        if (isadd) {
@@ -762,7 +976,8 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
 
                /* check for SHL 1,2,3 */
                if (pred_is_specific_node(temp, is_ia32_Shl)) {
-                       temp = get_Proj_pred(temp);
+                       temp  = get_Proj_pred(temp);
+                       shift = temp;
 
                        if (get_ia32_Immop_tarval(temp)) {
                                scale = get_tarval_long(get_ia32_Immop_tarval(temp));
@@ -772,6 +987,10 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
 
                                        DBG((mod, LEVEL_1, "\tgot scaled index %+F\n", index));
                                }
+                               else {
+                                       scale = 0;
+                                       shift = NULL;
+                               }
                        }
                }
 
@@ -792,44 +1011,36 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
 
        /* Try to assimilate a LEA as left operand */
        if (is_ia32_Lea(left) && (get_ia32_am_flavour(left) != ia32_am_O)) {
-               ir_node *assim_lea_idx, *assim_lea_base;
-
-               am_flav        = get_ia32_am_flavour(left);
-               assim_lea_base = get_irn_n(left, 0);
-               assim_lea_idx  = get_irn_n(left, 1);
-
-
-               /* If we have an Add with a real right operand (not NoReg) and  */
-               /* the LEA contains already an index calculation then we create */
-               /* a new LEA.                                                   */
-               /* If the LEA contains already a frame_entity then we also      */
-               /* create a new one  otherwise we would loose it.               */
-               if ((isadd && ! be_is_NoReg(cg, index) && (am_flav & ia32_I)) || /* no new LEA if index already set */
-                       get_ia32_frame_ent(left)                                  || /* no new LEA if stack access */
-                       (have_am_sc && get_ia32_am_sc(left))                      || /* no new LEA if AM symconst already present */
-                       /* at least on of the LEA operands must be NOREG */
-                       (!be_is_NoReg(cg, assim_lea_base) && !be_is_NoReg(cg, assim_lea_idx)))
-               {
+               /* check if we can assimilate the LEA */
+               int take_attr = do_new_lea(irn, base, index, left, have_am_sc, cg);
+
+               if (take_attr == IA32_LEA_ATTR_NONE) {
                        DBG((mod, LEVEL_1, "\tleave old LEA, creating new one\n"));
                }
                else {
                        DBG((mod, LEVEL_1, "\tgot LEA as left operand ... assimilating\n"));
-                       offs       = get_ia32_am_offs(left);
-                       am_sc      = have_am_sc ? am_sc : get_ia32_am_sc(left);
-                       have_am_sc = am_sc ? 1 : 0;
-                       am_sc_sign = is_ia32_am_sc_sign(left);
-                       scale      = get_ia32_am_scale(left);
-
-                       if (be_is_NoReg(cg, assim_lea_base) && ! be_is_NoReg(cg, assim_lea_idx)) {
-                               /* assimilate index */
-                               assert(be_is_NoReg(cg, index) && ! be_is_NoReg(cg, base) && "operand mismatch for LEA assimilation");
-                               index = assim_lea_idx;
-                       }
-                       else if (! be_is_NoReg(cg, assim_lea_base) && be_is_NoReg(cg, assim_lea_idx)) {
-                               /* assimilate base */
-                               assert(! be_is_NoReg(cg, index) && (base == left) && "operand mismatch for LEA assimilation");
-                               base = assim_lea_base;
+                       lea = left; /* for statistics */
+
+                       if (take_attr & IA32_LEA_ATTR_OFFS)
+                               offs = get_ia32_am_offs(left);
+
+                       if (take_attr & IA32_LEA_ATTR_AMSC) {
+                               am_sc      = get_ia32_am_sc(left);
+                               have_am_sc = 1;
+                               am_sc_sign = is_ia32_am_sc_sign(left);
                        }
+
+                       if (take_attr & IA32_LEA_ATTR_SCALE)
+                               scale = get_ia32_am_scale(left);
+
+                       if (take_attr & IA32_LEA_ATTR_BASE)
+                               base = get_irn_n(left, 0);
+
+                       if (take_attr & IA32_LEA_ATTR_INDEX)
+                               index = get_irn_n(left, 1);
+
+                       if (take_attr & IA32_LEA_ATTR_FENT)
+                               lea_ent = get_ia32_frame_ent(left);
                }
        }
 
@@ -874,9 +1085,12 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
 
                /* copy the frame entity (could be set in case of Add */
                /* which was a FrameAddr) */
-               set_ia32_frame_ent(res, get_ia32_frame_ent(irn));
+               if (lea_ent)
+                       set_ia32_frame_ent(res, lea_ent);
+               else
+                       set_ia32_frame_ent(res, get_ia32_frame_ent(irn));
 
-               if (is_ia32_use_frame(irn))
+               if (get_ia32_frame_ent(res))
                        set_ia32_use_frame(res);
 
                /* set scale */
@@ -905,7 +1119,22 @@ 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 */
-               DBG_OPT_LEA(irn, res);
+               if (shift && lea && lea_o)
+                       DBG_OPT_LEA4(irn, lea_o, lea, shift, res);
+               else if (shift && lea)
+                       DBG_OPT_LEA3(irn, lea, shift, res);
+               else if (shift && lea_o)
+                       DBG_OPT_LEA3(irn, lea_o, shift, res);
+               else if (lea && lea_o)
+                       DBG_OPT_LEA3(irn, lea_o, lea, res);
+               else if (shift)
+                       DBG_OPT_LEA2(irn, shift, res);
+               else if (lea)
+                       DBG_OPT_LEA2(irn, lea, res);
+               else if (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);
@@ -919,6 +1148,68 @@ static ir_node *fold_addr(ia32_code_gen_t *cg, ir_node *irn, ir_node *noreg) {
        return res;
 }
 
+
+/**
+ * Merges a Load/Store node with a LEA.
+ * @param irn The Load/Store node
+ * @param lea The LEA
+ */
+static void merge_loadstore_lea(ir_node *irn, ir_node *lea) {
+       entity *irn_ent = get_ia32_frame_ent(irn);
+       entity *lea_ent = get_ia32_frame_ent(lea);
+
+       /* If the irn and the LEA both have a different frame entity set: do not merge */
+       if (irn_ent && lea_ent && (irn_ent != lea_ent))
+               return;
+       else if (! irn_ent && lea_ent) {
+               set_ia32_frame_ent(irn, lea_ent);
+               set_ia32_use_frame(irn);
+       }
+
+       /* get the AM attributes from the LEA */
+       add_ia32_am_offs(irn, get_ia32_am_offs(lea));
+       set_ia32_am_scale(irn, get_ia32_am_scale(lea));
+       set_ia32_am_flavour(irn, get_ia32_am_flavour(lea));
+
+       set_ia32_am_sc(irn, get_ia32_am_sc(lea));
+       if (is_ia32_am_sc_sign(lea))
+               set_ia32_am_sc_sign(irn);
+
+       set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
+
+       /* set base and index */
+       set_irn_n(irn, 0, get_irn_n(lea, 0));
+       set_irn_n(irn, 1, get_irn_n(lea, 1));
+
+       /* clear remat flag */
+       set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
+
+       if (is_ia32_Ld(irn))
+               DBG_OPT_LOAD_LEA(lea, irn);
+       else
+               DBG_OPT_STORE_LEA(lea, irn);
+
+}
+
+/**
+ * Sets new_right index of irn to right and new_left index to left.
+ * Also exchange left and right
+ */
+static void exchange_left_right(ir_node *irn, ir_node **left, ir_node **right, int new_left, int new_right) {
+       ir_node *temp;
+
+       set_irn_n(irn, new_right, *right);
+       set_irn_n(irn, new_left, *left);
+
+       temp   = *left;
+       *left  = *right;
+       *right = temp;
+
+       /* this is only needed for Compares, but currently ALL nodes
+        * have this attribute :-) */
+       set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
+}
+
 /**
  * Optimizes a pattern around irn to address mode if possible.
  */
@@ -931,7 +1222,8 @@ void ia32_optimize_am(ir_node *irn, void *env) {
        ir_node           *left, *right, *temp;
        ir_node           *store, *load, *mem_proj;
        ir_node           *succ, *addr_b, *addr_i;
-       int                check_am_src = 0;
+       int               check_am_src = 0;
+       int               need_exchange_on_fail = 0;
        DEBUG_ONLY(firm_dbg_module_t *mod = cg->mod;)
 
        if (! is_ia32_irn(irn))
@@ -995,25 +1287,18 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                        left = get_irn_n(irn, 0);
 
                        if (is_ia32_Lea(left)) {
-                               DBG((mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
+                               const ir_edge_t *edge, *ne;
+                               ir_node *src;
 
-                               /* get the AM attributes from the LEA */
-                               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));
+                               /* merge all Loads/Stores connected to this LEA with the LEA */
+                               foreach_out_edge_safe(left, edge, ne) {
+                                       src = get_edge_src_irn(edge);
 
-                               set_ia32_am_sc(irn, get_ia32_am_sc(left));
-                               if (is_ia32_am_sc_sign(left))
-                                       set_ia32_am_sc_sign(irn);
-
-                               set_ia32_op_type(irn, is_ia32_Ld(irn) ? ia32_AddrModeS : ia32_AddrModeD);
-
-                               /* set base and index */
-                               set_irn_n(irn, 0, get_irn_n(left, 0));
-                               set_irn_n(irn, 1, get_irn_n(left, 1));
-
-                               /* clear remat flag */
-                               set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
+                                       if (src && (is_ia32_Ld(src) || is_ia32_St(src) || is_ia32_Store8Bit(src))) {
+                                               DBG((mod, LEVEL_1, "\nmerging %+F into %+F\n", left, irn));
+                                               merge_loadstore_lea(src, left);
+                                       }
+                               }
                        }
                }
                /* check if the node is an address mode candidate */
@@ -1036,12 +1321,8 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                                /* operand is a load, so we only need to check right operand.    */
                                if (pred_is_specific_nodeblock(block, left, is_ia32_Ld))
                                {
-                                       set_irn_n(irn, 2, right);
-                                       set_irn_n(irn, 3, left);
-
-                                       temp  = left;
-                                       left  = right;
-                                       right = temp;
+                                       exchange_left_right(irn, &left, &right, 3, 2);
+                                       need_exchange_on_fail = 1;
                                }
                        }
 
@@ -1069,20 +1350,10 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                                if (ia32_get_irn_n_edges(succ) == 1) {
                                        succ = get_edge_src_irn(get_irn_out_edge_first(succ));
 
-                                       if (is_ia32_fStore(succ) || is_ia32_Store(succ)) {
+                                       if (is_ia32_xStore(succ) || is_ia32_Store(succ)) {
                                                store  = succ;
                                                addr_b = get_irn_n(store, 0);
-
-                                               /* Could be that the Store is connected to the address    */
-                                               /* calculating LEA while the Load is already transformed. */
-                                               if (is_ia32_Lea(addr_b)) {
-                                                       succ   = addr_b;
-                                                       addr_b = get_irn_n(succ, 0);
-                                                       addr_i = get_irn_n(succ, 1);
-                                               }
-                                               else {
-                                                       addr_i = noreg_gp;
-                                               }
+                                               addr_i = get_irn_n(store, 1);
                                        }
                                }
 
@@ -1105,6 +1376,10 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                                                        temp  = left;
                                                        left  = right;
                                                        right = temp;
+
+                                                       /* this is only needed for Compares, but currently ALL nodes
+                                                        * have this attribute :-) */
+                                                       set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
                                                }
                                        }
 
@@ -1153,6 +1428,8 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                                                /* clear remat flag */
                                                set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
 
+                                               DBG_OPT_AM_D(load, store, irn);
+
                                                DB((mod, LEVEL_1, "merged with %+F and %+F into dest AM\n", load, store));
                                        }
                                } /* if (store) */
@@ -1166,19 +1443,21 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                                check_am_src = 1;
                        }
 
+                       /* was exchanged but optimize failed: exchange back */
+                       if (check_am_src && need_exchange_on_fail)
+                               exchange_left_right(irn, &left, &right, 3, 2);
+
+                       need_exchange_on_fail = 0;
+
                        /* normalize commutative ops */
-                       if (node_is_ia32_comm(irn)) {
+                       if (check_am_src && node_is_ia32_comm(irn)) {
                                /* 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.                                       */
                                if (pred_is_specific_nodeblock(block, right, is_ia32_Ld))       {
-                                       set_irn_n(irn, 2, right);
-                                       set_irn_n(irn, 3, left);
-
-                                       temp  = left;
-                                       left  = right;
-                                       right = temp;
+                                       exchange_left_right(irn, &left, &right, 3, 2);
+                                       need_exchange_on_fail = 1;
                                }
                        }
 
@@ -1217,14 +1496,25 @@ void ia32_optimize_am(ir_node *irn, void *env) {
                                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));
+
+                                       /* disconnect from Load */
+                                       set_irn_n(irn, 2, noreg_gp);
                                }
 
-                               /* disconnect from Load */
-                               set_irn_n(irn, 2, noreg_gp);
+                               DBG_OPT_AM_S(left, irn);
 
                                /* If Load has a memory Proj, connect it to the op */
                                mem_proj = get_mem_proj(left);
@@ -1235,6 +1525,11 @@ void ia32_optimize_am(ir_node *irn, void *env) {
 
                                DB((mod, LEVEL_1, "merged with %+F into source AM\n", left));
                        }
+                       else {
+                               /* was exchanged but optimize failed: exchange back */
+                               if (need_exchange_on_fail)
+                                       exchange_left_right(irn, &left, &right, 3, 2);
+                       }
                }
        }
 }