- implemented __builtin_return_address(0)
[libfirm] / ir / be / ia32 / bearch_ia32.c
index 9edb3cc..035d96f 100644 (file)
@@ -23,9 +23,7 @@
  * @author      Christian Wuerdig
  * @version     $Id$
  */
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif
 
 #include "lc_opts.h"
 #include "lc_opts_enum.h"
 #include "irprintf.h"
 #include "iredges_t.h"
 #include "ircons.h"
+#include "irflag.h"
 #include "irgmod.h"
 #include "irgopt.h"
 #include "irbitset.h"
 #include "irgopt.h"
-#include "irdump_grgen.h"
 #include "pdeq.h"
 #include "pset.h"
 #include "debug.h"
 #include "../begnuas.h"
 #include "../bestate.h"
 #include "../beflags.h"
+#include "../betranshlp.h"
 
 #include "bearch_ia32_t.h"
 
 #include "ia32_new_nodes.h"
 #include "gen_ia32_regalloc_if.h"
 #include "gen_ia32_machine.h"
+#include "ia32_common_transform.h"
 #include "ia32_transform.h"
 #include "ia32_emitter.h"
 #include "ia32_map_regs.h"
@@ -89,6 +89,8 @@
 
 #ifdef FIRM_GRGEN_BE
 #include "ia32_pbqp_transform.h"
+
+transformer_t be_transformer = TRANSFORMER_DEFAULT;
 #endif
 
 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
@@ -99,27 +101,30 @@ static set *cur_reg_set = NULL;
 ir_mode         *mode_fpcw       = NULL;
 ia32_code_gen_t *ia32_current_cg = NULL;
 
+/** The current omit-fp state */
+static unsigned ia32_curr_fp_ommitted  = 0;
+static ir_type *omit_fp_between_type   = NULL;
+static ir_type *between_type           = NULL;
+static ir_entity *old_bp_ent           = NULL;
+static ir_entity *ret_addr_ent         = NULL;
+static ir_entity *omit_fp_ret_addr_ent = NULL;
+
 /**
  * The environment for the intrinsic mapping.
  */
 static ia32_intrinsic_env_t intrinsic_env = {
        NULL,    /* the isa */
        NULL,    /* the irg, these entities belong to */
-       NULL,    /* entity for first div operand (move into FPU) */
-       NULL,    /* entity for second div operand (move into FPU) */
-       NULL,    /* entity for converts ll -> d */
-       NULL,    /* entity for converts d -> ll */
        NULL,    /* entity for __divdi3 library call */
        NULL,    /* entity for __moddi3 library call */
        NULL,    /* entity for __udivdi3 library call */
        NULL,    /* entity for __umoddi3 library call */
-       NULL,    /* bias value for conversion from float to unsigned 64 */
 };
 
 
-typedef ir_node *(*create_const_node_func) (dbg_info *dbg, ir_graph *irg, ir_node *block);
+typedef ir_node *(*create_const_node_func) (dbg_info *dbg, ir_node *block);
 
-static INLINE ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
+static inline ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
                                     create_const_node_func func,
                                     const arch_register_t* reg)
 {
@@ -129,8 +134,8 @@ static INLINE ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
                return *place;
 
        block = get_irg_start_block(cg->irg);
-       res = func(NULL, cg->irg, block);
-       arch_set_irn_register(cg->arch_env, res, reg);
+       res = func(NULL, block);
+       arch_set_irn_register(res, reg);
        *place = res;
 
        add_irn_dep(get_irg_end(cg->irg), res);
@@ -141,37 +146,37 @@ static INLINE ir_node *create_const(ia32_code_gen_t *cg, ir_node **place,
 
 /* Creates the unique per irg GP NoReg node. */
 ir_node *ia32_new_NoReg_gp(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->noreg_gp, new_rd_ia32_NoReg_GP,
+       return create_const(cg, &cg->noreg_gp, new_bd_ia32_NoReg_GP,
                            &ia32_gp_regs[REG_GP_NOREG]);
 }
 
 ir_node *ia32_new_NoReg_vfp(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->noreg_vfp, new_rd_ia32_NoReg_VFP,
+       return create_const(cg, &cg->noreg_vfp, new_bd_ia32_NoReg_VFP,
                            &ia32_vfp_regs[REG_VFP_NOREG]);
 }
 
 ir_node *ia32_new_NoReg_xmm(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->noreg_xmm, new_rd_ia32_NoReg_XMM,
+       return create_const(cg, &cg->noreg_xmm, new_bd_ia32_NoReg_XMM,
                            &ia32_xmm_regs[REG_XMM_NOREG]);
 }
 
 ir_node *ia32_new_Unknown_gp(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->unknown_gp, new_rd_ia32_Unknown_GP,
+       return create_const(cg, &cg->unknown_gp, new_bd_ia32_Unknown_GP,
                            &ia32_gp_regs[REG_GP_UKNWN]);
 }
 
 ir_node *ia32_new_Unknown_vfp(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->unknown_vfp, new_rd_ia32_Unknown_VFP,
+       return create_const(cg, &cg->unknown_vfp, new_bd_ia32_Unknown_VFP,
                            &ia32_vfp_regs[REG_VFP_UKNWN]);
 }
 
 ir_node *ia32_new_Unknown_xmm(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->unknown_xmm, new_rd_ia32_Unknown_XMM,
+       return create_const(cg, &cg->unknown_xmm, new_bd_ia32_Unknown_XMM,
                            &ia32_xmm_regs[REG_XMM_UKNWN]);
 }
 
 ir_node *ia32_new_Fpu_truncate(ia32_code_gen_t *cg) {
-       return create_const(cg, &cg->fpu_trunc_mode, new_rd_ia32_ChangeCW,
+       return create_const(cg, &cg->fpu_trunc_mode, new_bd_ia32_ChangeCW,
                         &ia32_fp_cw_regs[REG_FPCW]);
 }
 
@@ -179,10 +184,10 @@ ir_node *ia32_new_Fpu_truncate(ia32_code_gen_t *cg) {
 /**
  * Returns the admissible noreg register node for input register pos of node irn.
  */
-ir_node *ia32_get_admissible_noreg(ia32_code_gen_t *cg, ir_node *irn, int pos) {
-       const arch_register_req_t *req;
+static ir_node *ia32_get_admissible_noreg(ia32_code_gen_t *cg, ir_node *irn, int pos)
+{
+       const arch_register_req_t *req = arch_get_register_req(irn, pos);
 
-       req = arch_get_register_req(cg->arch_env, irn, pos);
        assert(req != NULL && "Missing register requirements");
        if (req->cls == &ia32_reg_classes[CLASS_ia32_gp])
                return ia32_new_NoReg_gp(cg);
@@ -247,62 +252,11 @@ static const arch_register_req_t *ia32_get_irn_reg_req(const ir_node *node,
        }
 
        /* unknowns should be transformed already */
-       assert(!is_Unknown(node));
        return arch_no_register_req;
 }
 
-static void ia32_set_irn_reg(ir_node *irn, const arch_register_t *reg)
-{
-       int    pos = 0;
-
-       if (get_irn_mode(irn) == mode_X) {
-               return;
-       }
-
-       if (is_Proj(irn)) {
-               pos = get_Proj_proj(irn);
-               irn = skip_Proj(irn);
-       }
-
-       if (is_ia32_irn(irn)) {
-               const arch_register_t **slots;
-
-               slots      = get_ia32_slots(irn);
-               slots[pos] = reg;
-       } else {
-               ia32_set_firm_reg(irn, reg, cur_reg_set);
-       }
-}
-
-static const arch_register_t *ia32_get_irn_reg(const ir_node *irn)
-{
-       int pos = 0;
-       const arch_register_t *reg = NULL;
-
-       if (is_Proj(irn)) {
-
-               if (get_irn_mode(irn) == mode_X) {
-                       return NULL;
-               }
-
-               pos = get_Proj_proj(irn);
-               irn = skip_Proj_const(irn);
-       }
-
-       if (is_ia32_irn(irn)) {
-               const arch_register_t **slots;
-               slots = get_ia32_slots(irn);
-               assert(pos < get_ia32_n_res(irn));
-               reg   = slots[pos];
-       } else {
-               reg = ia32_get_firm_reg(irn, cur_reg_set);
-       }
-
-       return reg;
-}
-
 static arch_irn_class_t ia32_classify(const ir_node *irn) {
-       arch_irn_class_t classification = arch_irn_class_normal;
+       arch_irn_class_t classification = 0;
 
        irn = skip_Proj_const(irn);
 
@@ -310,41 +264,18 @@ static arch_irn_class_t ia32_classify(const ir_node *irn) {
                classification |= arch_irn_class_branch;
 
        if (! is_ia32_irn(irn))
-               return classification & ~arch_irn_class_normal;
+               return classification;
 
-       if (is_ia32_Ld(irn))
-               classification |= arch_irn_class_load;
-
-       if (is_ia32_St(irn))
-               classification |= arch_irn_class_store;
-
-       if (is_ia32_need_stackent(irn))
+       if (is_ia32_is_reload(irn))
                classification |= arch_irn_class_reload;
 
-       return classification;
-}
-
-static arch_irn_flags_t ia32_get_flags(const ir_node *irn) {
-       arch_irn_flags_t flags = arch_irn_flags_none;
-
-       if (is_Unknown(irn))
-               return arch_irn_flags_ignore;
-
-       if(is_Proj(irn) && mode_is_datab(get_irn_mode(irn))) {
-               ir_node *pred = get_Proj_pred(irn);
-
-               if(is_ia32_irn(pred)) {
-                       flags = get_ia32_out_flags(pred, get_Proj_proj(irn));
-               }
+       if (is_ia32_is_spill(irn))
+               classification |= arch_irn_class_spill;
 
-               irn = pred;
-       }
-
-       if (is_ia32_irn(irn)) {
-               flags |= get_ia32_flags(irn);
-       }
+       if (is_ia32_is_remat(irn))
+               classification |= arch_irn_class_remat;
 
-       return flags;
+       return classification;
 }
 
 /**
@@ -384,6 +315,9 @@ static void ia32_set_frame_offset(ir_node *irn, int bias)
 
 static int ia32_get_sp_bias(const ir_node *node)
 {
+       if (is_ia32_Call(node))
+               return -(int)get_ia32_call_attr_const(node)->pop;
+
        if (is_ia32_Push(node))
                return 4;
 
@@ -393,67 +327,57 @@ static int ia32_get_sp_bias(const ir_node *node)
        return 0;
 }
 
-/**
- * Put all registers which are saved by the prologue/epilogue in a set.
- *
- * @param self  The callback object.
- * @param s     The result set.
- */
-static void ia32_abi_dont_save_regs(void *self, pset *s)
-{
-       ia32_abi_env_t *env = self;
-       if(env->flags.try_omit_fp)
-               pset_insert_ptr(s, env->aenv->bp);
-}
-
 /**
  * Generate the routine prologue.
  *
- * @param self    The callback object.
- * @param mem     A pointer to the mem node. Update this if you define new memory.
- * @param reg_map A map mapping all callee_save/ignore/parameter registers to their defining nodes.
+ * @param self       The callback object.
+ * @param mem        A pointer to the mem node. Update this if you define new memory.
+ * @param reg_map    A map mapping all callee_save/ignore/parameter registers to their defining nodes.
+ * @param stack_bias Points to the current stack bias, can be modified if needed.
  *
- * @return        The register which shall be used as a stack frame base.
+ * @return           The register which shall be used as a stack frame base.
  *
  * All nodes which define registers in @p reg_map must keep @p reg_map current.
  */
-static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map)
+static const arch_register_t *ia32_abi_prologue(void *self, ir_node **mem, pmap *reg_map, int *stack_bias)
 {
        ia32_abi_env_t   *env      = self;
        ia32_code_gen_t  *cg       = ia32_current_cg;
        const arch_env_t *arch_env = env->aenv;
 
+       ia32_curr_fp_ommitted = env->flags.try_omit_fp;
        if (! env->flags.try_omit_fp) {
-               ir_graph *irg     =env->irg;
+               ir_graph *irg     = env->irg;
                ir_node  *bl      = get_irg_start_block(irg);
                ir_node  *curr_sp = be_abi_reg_map_get(reg_map, arch_env->sp);
                ir_node  *curr_bp = be_abi_reg_map_get(reg_map, arch_env->bp);
-               ir_node  *noreg = ia32_new_NoReg_gp(cg);
+               ir_node  *noreg   = ia32_new_NoReg_gp(cg);
                ir_node  *push;
 
-               /* ALL nodes representing bp must be set to ignore. */
-               be_node_set_flags(get_Proj_pred(curr_bp), BE_OUT_POS(get_Proj_proj(curr_bp)), arch_irn_flags_ignore);
+               /* mark bp register as ignore */
+               be_set_constr_single_reg_out(get_Proj_pred(curr_bp),
+                               get_Proj_proj(curr_bp), arch_env->bp, arch_register_req_type_ignore);
 
                /* push ebp */
-               push    = new_rd_ia32_Push(NULL, irg, bl, noreg, noreg, *mem, curr_bp, curr_sp);
+               push    = new_bd_ia32_Push(NULL, bl, noreg, noreg, *mem, curr_bp, curr_sp);
                curr_sp = new_r_Proj(irg, bl, push, get_irn_mode(curr_sp), pn_ia32_Push_stack);
                *mem    = new_r_Proj(irg, bl, push, mode_M, pn_ia32_Push_M);
 
                /* the push must have SP out register */
-               arch_set_irn_register(arch_env, curr_sp, arch_env->sp);
-               set_ia32_flags(push, arch_irn_flags_ignore);
+               arch_set_irn_register(curr_sp, arch_env->sp);
+
+               /* this modifies the stack bias, because we pushed 32bit */
+               *stack_bias -= 4;
 
                /* move esp to ebp */
-               curr_bp  = be_new_Copy(arch_env->bp->reg_class, irg, bl, curr_sp);
-               be_set_constr_single_reg(curr_bp, BE_OUT_POS(0), arch_env->bp);
-               arch_set_irn_register(arch_env, curr_bp, arch_env->bp);
-               be_node_set_flags(curr_bp, BE_OUT_POS(0), arch_irn_flags_ignore);
+               curr_bp = be_new_Copy(arch_env->bp->reg_class, irg, bl, curr_sp);
+               be_set_constr_single_reg_out(curr_bp, 0, arch_env->bp,
+                                            arch_register_req_type_ignore);
 
                /* beware: the copy must be done before any other sp use */
                curr_sp = be_new_CopyKeep_single(arch_env->sp->reg_class, irg, bl, curr_sp, curr_bp, get_irn_mode(curr_sp));
-               be_set_constr_single_reg(curr_sp, BE_OUT_POS(0), arch_env->sp);
-               arch_set_irn_register(arch_env, curr_sp, arch_env->sp);
-               be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
+               be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
+                                                    arch_register_req_type_produces_sp);
 
                be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
                be_abi_reg_map_set(reg_map, arch_env->bp, curr_bp);
@@ -485,7 +409,6 @@ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_
        if (env->flags.try_omit_fp) {
                /* simply remove the stack frame here */
                curr_sp = be_new_IncSP(arch_env->sp, irg, bl, curr_sp, BE_STACK_FRAME_SIZE_SHRINK, 0);
-               add_irn_dep(curr_sp, *mem);
        } else {
                ir_mode *mode_bp = arch_env->bp->reg_class->mode;
 
@@ -493,8 +416,7 @@ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_
                        ir_node *leave;
 
                        /* leave */
-                       leave   = new_rd_ia32_Leave(NULL, irg, bl, curr_sp, curr_bp);
-                       set_ia32_flags(leave, arch_irn_flags_ignore);
+                       leave   = new_bd_ia32_Leave(NULL, bl, curr_bp);
                        curr_bp = new_r_Proj(irg, bl, leave, mode_bp, pn_ia32_Leave_frame);
                        curr_sp = new_r_Proj(irg, bl, leave, get_irn_mode(curr_sp), pn_ia32_Leave_stack);
                } else {
@@ -502,23 +424,23 @@ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_
 
                        /* the old SP is not needed anymore (kill the proj) */
                        assert(is_Proj(curr_sp));
-                       be_kill_node(curr_sp);
+                       kill_node(curr_sp);
 
                        /* copy ebp to esp */
                        curr_sp = be_new_Copy(&ia32_reg_classes[CLASS_ia32_gp], irg, bl, curr_bp);
-                       arch_set_irn_register(arch_env, curr_sp, arch_env->sp);
-                       be_node_set_flags(curr_sp, BE_OUT_POS(0), arch_irn_flags_ignore);
+                       arch_set_irn_register(curr_sp, arch_env->sp);
+                       be_set_constr_single_reg_out(curr_sp, 0, arch_env->sp,
+                                                        arch_register_req_type_ignore);
 
                        /* pop ebp */
-                       pop     = new_rd_ia32_Pop(NULL, env->irg, bl, *mem, curr_sp);
-                       set_ia32_flags(pop, arch_irn_flags_ignore);
+                       pop     = new_bd_ia32_PopEbp(NULL, bl, *mem, curr_sp);
                        curr_bp = new_r_Proj(irg, bl, pop, mode_bp, pn_ia32_Pop_res);
                        curr_sp = new_r_Proj(irg, bl, pop, get_irn_mode(curr_sp), pn_ia32_Pop_stack);
 
                        *mem = new_r_Proj(irg, bl, pop, mode_M, pn_ia32_Pop_M);
                }
-               arch_set_irn_register(arch_env, curr_sp, arch_env->sp);
-               arch_set_irn_register(arch_env, curr_bp, arch_env->bp);
+               arch_set_irn_register(curr_sp, arch_env->sp);
+               arch_set_irn_register(curr_bp, arch_env->bp);
        }
 
        be_abi_reg_map_set(reg_map, arch_env->sp, curr_sp);
@@ -534,8 +456,8 @@ static void ia32_abi_epilogue(void *self, ir_node *bl, ir_node **mem, pmap *reg_
  */
 static void *ia32_abi_init(const be_abi_call_t *call, const arch_env_t *aenv, ir_graph *irg)
 {
-       ia32_abi_env_t *env    = xmalloc(sizeof(env[0]));
-       be_abi_call_flags_t fl = be_abi_call_get_flags(call);
+       ia32_abi_env_t      *env = XMALLOC(ia32_abi_env_t);
+       be_abi_call_flags_t  fl  = be_abi_call_get_flags(call);
        env->flags = fl.bits;
        env->irg   = irg;
        env->aenv  = aenv;
@@ -551,23 +473,11 @@ static void ia32_abi_done(void *self) {
 }
 
 /**
- * Produces the type which sits between the stack args and the locals on the stack.
- * it will contain the return address and space to store the old base pointer.
- * @return The Firm type modeling the ABI between type.
+ * Build the between type and entities if not already build.
  */
-static ir_type *ia32_abi_get_between_type(void *self)
-{
+static void ia32_build_between_type(void) {
 #define IDENT(s) new_id_from_chars(s, sizeof(s)-1)
-       static ir_type *omit_fp_between_type = NULL;
-       static ir_type *between_type         = NULL;
-
-       ia32_abi_env_t *env = self;
-
        if (! between_type) {
-               ir_entity *old_bp_ent;
-               ir_entity *ret_addr_ent;
-               ir_entity *omit_fp_ret_addr_ent;
-
                ir_type *old_bp_type   = new_type_primitive(IDENT("bp"), mode_Iu);
                ir_type *ret_addr_type = new_type_primitive(IDENT("return_addr"), mode_Iu);
 
@@ -587,9 +497,28 @@ static ir_type *ia32_abi_get_between_type(void *self)
                set_type_size_bytes(omit_fp_between_type, get_type_size_bytes(ret_addr_type));
                set_type_state(omit_fp_between_type, layout_fixed);
        }
+#undef IDENT
+}
 
+/**
+ * Produces the type which sits between the stack args and the locals on the stack.
+ * it will contain the return address and space to store the old base pointer.
+ * @return The Firm type modeling the ABI between type.
+ */
+static ir_type *ia32_abi_get_between_type(void *self)
+{
+       ia32_abi_env_t *env = self;
+
+       ia32_build_between_type();
        return env->flags.try_omit_fp ? omit_fp_between_type : between_type;
-#undef IDENT
+}
+
+/**
+ * Return the stack entity that contains the return address.
+ */
+ir_entity *ia32_get_return_address_entity(void) {
+       ia32_build_between_type();
+       return ia32_curr_fp_ommitted ? omit_fp_ret_addr_ent : ret_addr_ent;
 }
 
 /**
@@ -629,9 +558,10 @@ static int ia32_get_op_estimated_cost(const ir_node *irn)
                        (we assume they are in cache), other memory operations cost 20
                        cycles.
                */
-               if(is_ia32_use_frame(irn) ||
-                               (is_ia32_NoReg_GP(get_irn_n(irn, 0)) &&
-                        is_ia32_NoReg_GP(get_irn_n(irn, 1)))) {
+               if (is_ia32_use_frame(irn) || (
+                       is_ia32_NoReg_GP(get_irn_n(irn, n_ia32_base)) &&
+                       is_ia32_NoReg_GP(get_irn_n(irn, n_ia32_index))
+                   )) {
                        cost += 5;
                } else {
                        cost += 20;
@@ -651,7 +581,6 @@ static int ia32_get_op_estimated_cost(const ir_node *irn)
  * @return          The inverse operation or NULL if operation invertible
  */
 static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_t *inverse, struct obstack *obst) {
-       ir_graph *irg;
        ir_mode  *mode;
        ir_mode  *irn_mode;
        ir_node  *block, *noreg, *nomem;
@@ -674,12 +603,11 @@ static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_
                   irn);
        return NULL;
 
-       irg      = get_irn_irg(irn);
        block    = get_nodes_block(irn);
        mode     = get_irn_mode(irn);
        irn_mode = get_irn_mode(irn);
        noreg    = get_irn_n(irn, 0);
-       nomem    = new_r_NoMem(irg);
+       nomem    = new_NoMem();
        dbg      = get_irn_dbg_info(irn);
 
        /* initialize structure */
@@ -693,7 +621,7 @@ static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_
                        if (get_ia32_immop_type(irn) == ia32_ImmConst) {
                                /* we have an add with a const here */
                                /* invers == add with negated const */
-                               inverse->nodes[0] = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
+                               inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
                                inverse->costs   += 1;
                                copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
                                set_ia32_Immop_tarval(inverse->nodes[0], tarval_neg(get_ia32_Immop_tarval(irn)));
@@ -702,13 +630,13 @@ static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_
                        else if (get_ia32_immop_type(irn) == ia32_ImmSymConst) {
                                /* we have an add with a symconst here */
                                /* invers == sub with const */
-                               inverse->nodes[0] = new_rd_ia32_Sub(dbg, irg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
+                               inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
                                inverse->costs   += 2;
                                copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
                        }
                        else {
                                /* normal add: inverse == sub */
-                               inverse->nodes[0] = new_rd_ia32_Sub(dbg, irg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, i ^ 1));
+                               inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, i ^ 1));
                                inverse->costs   += 2;
                        }
 #endif
@@ -718,17 +646,17 @@ static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_
                        if (get_ia32_immop_type(irn) != ia32_ImmNone) {
                                /* we have a sub with a const/symconst here */
                                /* invers == add with this const */
-                               inverse->nodes[0] = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
+                               inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
                                inverse->costs   += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
                                copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
                        }
                        else {
                                /* normal sub */
                                if (i == n_ia32_binary_left) {
-                                       inverse->nodes[0] = new_rd_ia32_Add(dbg, irg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, 3));
+                                       inverse->nodes[0] = new_bd_ia32_Add(dbg, block, noreg, noreg, nomem, (ir_node*) irn, get_irn_n(irn, 3));
                                }
                                else {
-                                       inverse->nodes[0] = new_rd_ia32_Sub(dbg, irg, block, noreg, noreg, nomem, get_irn_n(irn, n_ia32_binary_left), (ir_node*) irn);
+                                       inverse->nodes[0] = new_bd_ia32_Sub(dbg, block, noreg, noreg, nomem, get_irn_n(irn, n_ia32_binary_left), (ir_node*) irn);
                                }
                                inverse->costs += 1;
                        }
@@ -738,24 +666,24 @@ static arch_inverse_t *ia32_get_inverse(const ir_node *irn, int i, arch_inverse_
 #if 0
                        if (get_ia32_immop_type(irn) != ia32_ImmNone) {
                                /* xor with const: inverse = xor */
-                               inverse->nodes[0] = new_rd_ia32_Xor(dbg, irg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
+                               inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, get_irn_n(irn, i), noreg);
                                inverse->costs   += (get_ia32_immop_type(irn) == ia32_ImmSymConst) ? 5 : 1;
                                copy_ia32_Immop_attr(inverse->nodes[0], (ir_node *)irn);
                        }
                        else {
                                /* normal xor */
-                               inverse->nodes[0] = new_rd_ia32_Xor(dbg, irg, block, noreg, noreg, nomem, (ir_node *) irn, get_irn_n(irn, i));
+                               inverse->nodes[0] = new_bd_ia32_Xor(dbg, block, noreg, noreg, nomem, (ir_node *) irn, get_irn_n(irn, i));
                                inverse->costs   += 1;
                        }
 #endif
                        break;
                case iro_ia32_Not: {
-                       inverse->nodes[0] = new_rd_ia32_Not(dbg, irg, block, (ir_node*) irn);
+                       inverse->nodes[0] = new_bd_ia32_Not(dbg, block, (ir_node*) irn);
                        inverse->costs   += 1;
                        break;
                }
                case iro_ia32_Neg: {
-                       inverse->nodes[0] = new_rd_ia32_Neg(dbg, irg, block, (ir_node*) irn);
+                       inverse->nodes[0] = new_bd_ia32_Neg(dbg, block, (ir_node*) irn);
                        inverse->costs   += 1;
                        break;
                }
@@ -790,82 +718,110 @@ static ir_mode *get_spill_mode(const ir_node *node)
  */
 static int ia32_is_spillmode_compatible(const ir_mode *mode, const ir_mode *spillmode)
 {
-       if(mode_is_float(mode)) {
-               return mode == spillmode;
-       } else {
-               return 1;
-       }
+       return !mode_is_float(mode) || mode == spillmode;
 }
 
 /**
  * Check if irn can load its operand at position i from memory (source addressmode).
- * @param self   Pointer to irn ops itself
  * @param irn    The irn to be checked
  * @param i      The operands position
  * @return Non-Zero if operand can be loaded
  */
-static int ia32_possible_memory_operand(const ir_node *irn, unsigned int i) {
-       ir_node *op = get_irn_n(irn, i);
-       const ir_mode *mode = get_irn_mode(op);
+static int ia32_possible_memory_operand(const ir_node *irn, unsigned int i)
+{
+       ir_node       *op        = get_irn_n(irn, i);
+       const ir_mode *mode      = get_irn_mode(op);
        const ir_mode *spillmode = get_spill_mode(op);
 
-       if (
-               (i != n_ia32_binary_left && i != n_ia32_binary_right) || /* a "real" operand position must be requested */
-               ! is_ia32_irn(irn)                                    ||  /* must be an ia32 irn */
-               get_ia32_am_arity(irn) != ia32_am_binary              ||  /* must be a binary operation TODO is this necessary? */
-               get_ia32_op_type(irn) != ia32_Normal                  ||  /* must not already be a addressmode irn */
-               ! (get_ia32_am_support(irn) & ia32_am_Source)         ||  /* must be capable of source addressmode */
-               ! ia32_is_spillmode_compatible(mode, spillmode)       ||
-               is_ia32_use_frame(irn))                                  /* must not already use frame */
+       if (!is_ia32_irn(irn)                              ||  /* must be an ia32 irn */
+           get_ia32_op_type(irn) != ia32_Normal           ||  /* must not already be a addressmode irn */
+           !ia32_is_spillmode_compatible(mode, spillmode) ||
+           is_ia32_use_frame(irn))                            /* must not already use frame */
                return 0;
 
-       if (i == n_ia32_binary_left) {
-               const arch_register_req_t *req;
-               if(!is_ia32_commutative(irn))
+       switch (get_ia32_am_support(irn)) {
+               case ia32_am_none:
                        return 0;
-               /* we can't swap left/right for limited registers
-                * (As this (currently) breaks constraint handling copies)
-                */
-               req = get_ia32_in_req(irn, n_ia32_binary_left);
-               if (req->type & arch_register_req_type_limited) {
-                       return 0;
-               }
+
+               case ia32_am_unary:
+                       if (i != n_ia32_unary_op)
+                               return 0;
+                       break;
+
+               case ia32_am_binary:
+                       switch (i) {
+                               case n_ia32_binary_left: {
+                                       const arch_register_req_t *req;
+                                       if (!is_ia32_commutative(irn))
+                                               return 0;
+
+                                       /* we can't swap left/right for limited registers
+                                        * (As this (currently) breaks constraint handling copies)
+                                        */
+                                       req = get_ia32_in_req(irn, n_ia32_binary_left);
+                                       if (req->type & arch_register_req_type_limited)
+                                               return 0;
+                                       break;
+                               }
+
+                               case n_ia32_binary_right:
+                                       break;
+
+                               default:
+                                       return 0;
+                       }
+                       break;
+
+               default:
+                       panic("Unknown AM type");
        }
 
+       /* HACK: must not already use "real" memory.
+        * This can happen for Call and Div */
+       if (!is_NoMem(get_irn_n(irn, n_ia32_mem)))
+               return 0;
+
        return 1;
 }
 
 static void ia32_perform_memory_operand(ir_node *irn, ir_node *spill,
                                         unsigned int i)
 {
-       ia32_code_gen_t *cg = ia32_current_cg;
+       ir_mode *load_mode;
+       ir_mode *dest_op_mode;
 
        assert(ia32_possible_memory_operand(irn, i) && "Cannot perform memory operand change");
 
-       if (i == n_ia32_binary_left) {
-               ia32_swap_left_right(irn);
-       }
-
        set_ia32_op_type(irn, ia32_AddrModeS);
-       set_ia32_ls_mode(irn, get_irn_mode(get_irn_n(irn, i)));
+
+       load_mode    = get_irn_mode(get_irn_n(irn, i));
+       dest_op_mode = get_ia32_ls_mode(irn);
+       if (get_mode_size_bits(load_mode) <= get_mode_size_bits(dest_op_mode)) {
+               set_ia32_ls_mode(irn, load_mode);
+       }
        set_ia32_use_frame(irn);
        set_ia32_need_stackent(irn);
 
-       set_irn_n(irn, n_ia32_base, get_irg_frame(get_irn_irg(irn)));
-       set_irn_n(irn, n_ia32_binary_right, ia32_get_admissible_noreg(cg, irn, n_ia32_binary_right));
-       set_irn_n(irn, n_ia32_mem, spill);
-
-       /* immediates are only allowed on the right side */
-       if (i == n_ia32_binary_left && is_ia32_Immediate(get_irn_n(irn, n_ia32_binary_left))) {
+       if (i == n_ia32_binary_left                    &&
+           get_ia32_am_support(irn) == ia32_am_binary &&
+           /* immediates are only allowed on the right side */
+           !is_ia32_Immediate(get_irn_n(irn, n_ia32_binary_right))) {
                ia32_swap_left_right(irn);
+               i = n_ia32_binary_right;
        }
+
+       assert(is_NoMem(get_irn_n(irn, n_ia32_mem)));
+
+       set_irn_n(irn, n_ia32_base, get_irg_frame(get_irn_irg(irn)));
+       set_irn_n(irn, n_ia32_mem,  spill);
+       set_irn_n(irn, i,           ia32_get_admissible_noreg(ia32_current_cg, irn, i));
+       set_ia32_is_reload(irn);
 }
 
 static const be_abi_callbacks_t ia32_abi_callbacks = {
        ia32_abi_init,
        ia32_abi_done,
        ia32_abi_get_between_type,
-       ia32_abi_dont_save_regs,
        ia32_abi_prologue,
        ia32_abi_epilogue
 };
@@ -874,10 +830,7 @@ static const be_abi_callbacks_t ia32_abi_callbacks = {
 
 static const arch_irn_ops_t ia32_irn_ops = {
        ia32_get_irn_reg_req,
-       ia32_set_irn_reg,
-       ia32_get_irn_reg,
        ia32_classify,
-       ia32_get_flags,
        ia32_get_frame_entity,
        ia32_set_frame_entity,
        ia32_set_frame_offset,
@@ -906,7 +859,7 @@ static ir_entity *mcount = NULL;
 static void ia32_before_abi(void *self) {
        lower_mode_b_config_t lower_mode_b_config = {
                mode_Iu,  /* lowered mode */
-               mode_Bu,  /* prefered mode for set */
+               mode_Bu,  /* preferred mode for set */
                0,        /* don't lower direct compares */
        };
        ia32_code_gen_t *cg = self;
@@ -930,11 +883,18 @@ static void ia32_before_abi(void *self) {
  * Transforms the standard firm graph into
  * an ia32 firm graph
  */
-static void ia32_prepare_graph(void *self) {
-       ia32_code_gen_t *cg = self;
+static void ia32_prepare_graph(void *self)
+{
+       ia32_code_gen_t *cg  = self;
+       ir_graph        *irg = cg->irg;
 
-       /* do local optimisations */
-       optimize_graph_df(cg->irg);
+       /* do local optimizations */
+       optimize_graph_df(irg);
+
+       /* we have to do cfopt+remove_critical_edges as we can't have Bad-blocks
+        * or critical edges in the backend */
+       optimize_cf(irg);
+       remove_critical_cf_edges(irg);
 
        /* TODO: we often have dead code reachable through out-edges here. So for
         * now we rebuild edges (as we need correct user count for code selection)
@@ -947,18 +907,25 @@ static void ia32_prepare_graph(void *self) {
        if (cg->dump)
                be_dump(cg->irg, "-pre_transform", dump_ir_block_graph_sched);
 
+       switch (be_transformer) {
+       case TRANSFORMER_DEFAULT:
+               /* transform remaining nodes into assembler instructions */
+               ia32_transform_graph(cg);
+               break;
+
 #ifdef FIRM_GRGEN_BE
-       /* transform nodes into assembler instructions by PBQP magic */
-       ia32_transform_graph_by_pbqp(cg);
+       case TRANSFORMER_PBQP:
+       case TRANSFORMER_RAND:
+               /* transform nodes into assembler instructions by PBQP magic */
+               ia32_transform_graph_by_pbqp(cg);
+               break;
 #endif
 
-       if (cg->dump)
-               be_dump(cg->irg, "-after_pbqp_transform", dump_ir_block_graph_sched);
-
-       /* transform remaining nodes into assembler instructions */
-       ia32_transform_graph(cg);
+       default:
+               panic("invalid transformer");
+       }
 
-       /* do local optimisations (mainly CSE) */
+       /* do local optimizations (mainly CSE) */
        optimize_graph_df(cg->irg);
 
        if (cg->dump)
@@ -967,9 +934,6 @@ static void ia32_prepare_graph(void *self) {
        /* optimize address mode */
        ia32_optimize_graph(cg);
 
-       if (cg->dump)
-               be_dump(cg->irg, "-am", dump_ir_block_graph_sched);
-
        /* do code placement, to optimize the position of constants */
        place_code(cg->irg);
 
@@ -977,14 +941,7 @@ static void ia32_prepare_graph(void *self) {
                be_dump(cg->irg, "-place", dump_ir_block_graph_sched);
 }
 
-/**
- * Dummy functions for hooks we don't need but which must be filled.
- */
-static void ia32_before_sched(void *self) {
-       (void) self;
-}
-
-static void turn_back_am(ir_node *node)
+ir_node *turn_back_am(ir_node *node)
 {
        ir_graph *irg   = current_ir_graph;
        dbg_info *dbgi  = get_irn_dbg_info(node);
@@ -992,40 +949,34 @@ static void turn_back_am(ir_node *node)
        ir_node  *base  = get_irn_n(node, n_ia32_base);
        ir_node  *index = get_irn_n(node, n_ia32_index);
        ir_node  *mem   = get_irn_n(node, n_ia32_mem);
-       ir_node  *noreg = ia32_new_NoReg_gp(ia32_current_cg);
-       ir_node  *load;
-       ir_node  *load_res;
-       ir_node  *mem_proj;
-       const ir_edge_t *edge;
+       ir_node  *noreg;
 
-       load     = new_rd_ia32_Load(dbgi, irg, block, base, index, mem);
-       load_res = new_rd_Proj(dbgi, irg, block, load, mode_Iu, pn_ia32_Load_res);
+       ir_node  *load     = new_bd_ia32_Load(dbgi, block, base, index, mem);
+       ir_node  *load_res = new_rd_Proj(dbgi, irg, block, load, mode_Iu, pn_ia32_Load_res);
 
        ia32_copy_am_attrs(load, node);
+       if (is_ia32_is_reload(node))
+               set_ia32_is_reload(load);
        set_irn_n(node, n_ia32_mem, new_NoMem());
 
-       switch (get_ia32_am_arity(node)) {
+       switch (get_ia32_am_support(node)) {
                case ia32_am_unary:
                        set_irn_n(node, n_ia32_unary_op, load_res);
                        break;
 
                case ia32_am_binary:
-                       if (is_ia32_Immediate(get_irn_n(node, n_ia32_Cmp_right))) {
-                               assert(is_ia32_Cmp(node)  || is_ia32_Cmp8Bit(node) ||
-                                      is_ia32_Test(node) || is_ia32_Test8Bit(node));
+                       if (is_ia32_Immediate(get_irn_n(node, n_ia32_binary_right))) {
                                set_irn_n(node, n_ia32_binary_left, load_res);
                        } else {
                                set_irn_n(node, n_ia32_binary_right, load_res);
                        }
                        break;
 
-               case ia32_am_ternary:
-                       set_irn_n(node, n_ia32_binary_right, load_res);
-                       break;
-
-               default: break;
+               default:
+                       panic("Unknown AM type");
        }
-       set_irn_n(node, n_ia32_base, noreg);
+       noreg = ia32_new_NoReg_gp(ia32_current_cg);
+       set_irn_n(node, n_ia32_base,  noreg);
        set_irn_n(node, n_ia32_index, noreg);
        set_ia32_am_offs_int(node, 0);
        set_ia32_am_sc(node, NULL);
@@ -1034,24 +985,22 @@ static void turn_back_am(ir_node *node)
 
        /* rewire mem-proj */
        if (get_irn_mode(node) == mode_T) {
-               mem_proj = NULL;
+               const ir_edge_t *edge;
                foreach_out_edge(node, edge) {
                        ir_node *out = get_edge_src_irn(edge);
-                       if(get_irn_mode(out) == mode_M) {
-                               assert(mem_proj == NULL);
-                               mem_proj = out;
+                       if (get_irn_mode(out) == mode_M) {
+                               set_Proj_pred(out, load);
+                               set_Proj_proj(out, pn_ia32_Load_M);
+                               break;
                        }
                }
-
-               if(mem_proj != NULL) {
-                       set_Proj_pred(mem_proj, load);
-                       set_Proj_proj(mem_proj, pn_ia32_Load_M);
-               }
        }
 
        set_ia32_op_type(node, ia32_Normal);
        if (sched_is_scheduled(node))
                sched_add_before(node, load);
+
+       return load_res;
 }
 
 static ir_node *flags_remat(ir_node *node, ir_node *after)
@@ -1090,8 +1039,6 @@ static ir_node *flags_remat(ir_node *node, ir_node *after)
 
 /**
  * Called before the register allocator.
- * Calculate a block schedule here. We need it for the x87
- * simulator and the emitter.
  */
 static void ia32_before_ra(void *self) {
        ia32_code_gen_t *cg = self;
@@ -1130,21 +1077,22 @@ static void transform_to_Load(ia32_code_gen_t *cg, ir_node *node) {
 
        if (mode_is_float(spillmode)) {
                if (ia32_cg_config.use_sse2)
-                       new_op = new_rd_ia32_xLoad(dbg, irg, block, ptr, noreg, mem, spillmode);
+                       new_op = new_bd_ia32_xLoad(dbg, block, ptr, noreg, mem, spillmode);
                else
-                       new_op = new_rd_ia32_vfld(dbg, irg, block, ptr, noreg, mem, spillmode);
+                       new_op = new_bd_ia32_vfld(dbg, block, ptr, noreg, mem, spillmode);
        }
        else if (get_mode_size_bits(spillmode) == 128) {
                /* Reload 128 bit SSE registers */
-               new_op = new_rd_ia32_xxLoad(dbg, irg, block, ptr, noreg, mem);
+               new_op = new_bd_ia32_xxLoad(dbg, block, ptr, noreg, mem);
        }
        else
-               new_op = new_rd_ia32_Load(dbg, irg, block, ptr, noreg, mem);
+               new_op = new_bd_ia32_Load(dbg, block, ptr, noreg, mem);
 
        set_ia32_op_type(new_op, ia32_AddrModeS);
        set_ia32_ls_mode(new_op, spillmode);
        set_ia32_frame_ent(new_op, ent);
        set_ia32_use_frame(new_op);
+       set_ia32_is_reload(new_op);
 
        DBG_OPT_RELOAD2LD(node, new_op);
 
@@ -1156,10 +1104,10 @@ static void transform_to_Load(ia32_code_gen_t *cg, ir_node *node) {
        }
 
        /* copy the register from the old node to the new Load */
-       reg = arch_get_irn_register(cg->arch_env, node);
-       arch_set_irn_register(cg->arch_env, new_op, reg);
+       reg = arch_get_irn_register(node);
+       arch_set_irn_register(proj, reg);
 
-       SET_IA32_ORIG_NODE(new_op, ia32_get_old_node_name(cg, node));
+       SET_IA32_ORIG_NODE(new_op, node);
 
        exchange(node, proj);
 }
@@ -1175,7 +1123,7 @@ static void transform_to_Store(ia32_code_gen_t *cg, ir_node *node) {
        const ir_node *spillval = get_irn_n(node, be_pos_Spill_val);
        ir_mode *mode  = get_spill_mode(spillval);
        ir_node *noreg = ia32_new_NoReg_gp(cg);
-       ir_node *nomem = new_rd_NoMem(irg);
+       ir_node *nomem = new_NoMem();
        ir_node *ptr   = get_irg_frame(irg);
        ir_node *val   = get_irn_n(node, be_pos_Spill_val);
        ir_node *store;
@@ -1199,23 +1147,24 @@ static void transform_to_Store(ia32_code_gen_t *cg, ir_node *node) {
 
        if (mode_is_float(mode)) {
                if (ia32_cg_config.use_sse2)
-                       store = new_rd_ia32_xStore(dbg, irg, block, ptr, noreg, nomem, val);
+                       store = new_bd_ia32_xStore(dbg, block, ptr, noreg, nomem, val);
                else
-                       store = new_rd_ia32_vfst(dbg, irg, block, ptr, noreg, nomem, val, mode);
+                       store = new_bd_ia32_vfst(dbg, block, ptr, noreg, nomem, val, mode);
        } else if (get_mode_size_bits(mode) == 128) {
                /* Spill 128 bit SSE registers */
-               store = new_rd_ia32_xxStore(dbg, irg, block, ptr, noreg, nomem, val);
+               store = new_bd_ia32_xxStore(dbg, block, ptr, noreg, nomem, val);
        } else if (get_mode_size_bits(mode) == 8) {
-               store = new_rd_ia32_Store8Bit(dbg, irg, block, ptr, noreg, nomem, val);
+               store = new_bd_ia32_Store8Bit(dbg, block, ptr, noreg, nomem, val);
        } else {
-               store = new_rd_ia32_Store(dbg, irg, block, ptr, noreg, nomem, val);
+               store = new_bd_ia32_Store(dbg, block, ptr, noreg, nomem, val);
        }
 
        set_ia32_op_type(store, ia32_AddrModeD);
        set_ia32_ls_mode(store, mode);
        set_ia32_frame_ent(store, ent);
        set_ia32_use_frame(store);
-       SET_IA32_ORIG_NODE(store, ia32_get_old_node_name(cg, node));
+       set_ia32_is_spill(store);
+       SET_IA32_ORIG_NODE(store, node);
        DBG_OPT_SPILL2ST(node, store);
 
        if (sched_point) {
@@ -1227,43 +1176,46 @@ static void transform_to_Store(ia32_code_gen_t *cg, ir_node *node) {
 }
 
 static ir_node *create_push(ia32_code_gen_t *cg, ir_node *node, ir_node *schedpoint, ir_node *sp, ir_node *mem, ir_entity *ent) {
-       ir_graph *irg = get_irn_irg(node);
        dbg_info *dbg = get_irn_dbg_info(node);
        ir_node *block = get_nodes_block(node);
        ir_node *noreg = ia32_new_NoReg_gp(cg);
+       ir_graph *irg = get_irn_irg(node);
        ir_node *frame = get_irg_frame(irg);
 
-       ir_node *push = new_rd_ia32_Push(dbg, irg, block, frame, noreg, mem, noreg, sp);
+       ir_node *push = new_bd_ia32_Push(dbg, block, frame, noreg, mem, noreg, sp);
 
        set_ia32_frame_ent(push, ent);
        set_ia32_use_frame(push);
        set_ia32_op_type(push, ia32_AddrModeS);
        set_ia32_ls_mode(push, mode_Is);
+       set_ia32_is_spill(push);
 
        sched_add_before(schedpoint, push);
        return push;
 }
 
 static ir_node *create_pop(ia32_code_gen_t *cg, ir_node *node, ir_node *schedpoint, ir_node *sp, ir_entity *ent) {
-       ir_graph *irg = get_irn_irg(node);
        dbg_info *dbg = get_irn_dbg_info(node);
        ir_node *block = get_nodes_block(node);
        ir_node *noreg = ia32_new_NoReg_gp(cg);
+       ir_graph *irg = get_irn_irg(node);
        ir_node *frame = get_irg_frame(irg);
 
-       ir_node *pop = new_rd_ia32_PopMem(dbg, irg, block, frame, noreg, new_NoMem(), sp);
+       ir_node *pop = new_bd_ia32_PopMem(dbg, block, frame, noreg, new_NoMem(), sp);
 
        set_ia32_frame_ent(pop, ent);
        set_ia32_use_frame(pop);
        set_ia32_op_type(pop, ia32_AddrModeD);
        set_ia32_ls_mode(pop, mode_Is);
+       set_ia32_is_reload(pop);
 
        sched_add_before(schedpoint, pop);
 
        return pop;
 }
 
-static ir_node* create_spproj(ia32_code_gen_t *cg, ir_node *node, ir_node *pred, int pos) {
+static ir_node* create_spproj(ir_node *node, ir_node *pred, int pos)
+{
        ir_graph *irg = get_irn_irg(node);
        dbg_info *dbg = get_irn_dbg_info(node);
        ir_node *block = get_nodes_block(node);
@@ -1272,7 +1224,7 @@ static ir_node* create_spproj(ia32_code_gen_t *cg, ir_node *node, ir_node *pred,
        ir_node *sp;
 
        sp = new_rd_Proj(dbg, irg, block, pred, spmode, pos);
-       arch_set_irn_register(cg->arch_env, sp, spreg);
+       arch_set_irn_register(sp, spreg);
 
        return sp;
 }
@@ -1282,19 +1234,18 @@ static ir_node* create_spproj(ia32_code_gen_t *cg, ir_node *node, ir_node *pred,
  * push/pop into/from memory cascades. This is possible without using
  * any registers.
  */
-static void transform_MemPerm(ia32_code_gen_t *cg, ir_node *node) {
-       ir_graph *irg = get_irn_irg(node);
-       ir_node *block = get_nodes_block(node);
-       ir_node *in[1];
-       ir_node *keep;
-       int i, arity;
-       ir_node *sp = be_abi_get_ignore_irn(cg->birg->abi, &ia32_gp_regs[REG_ESP]);
+static void transform_MemPerm(ia32_code_gen_t *cg, ir_node *node)
+{
+       ir_graph        *irg   = get_irn_irg(node);
+       ir_node         *block = get_nodes_block(node);
+       ir_node         *sp    = be_abi_get_ignore_irn(cg->birg->abi, &ia32_gp_regs[REG_ESP]);
+       int              arity = be_get_MemPerm_entity_arity(node);
+       ir_node        **pops  = ALLOCAN(ir_node*, arity);
+       ir_node         *in[1];
+       ir_node         *keep;
+       int              i;
        const ir_edge_t *edge;
        const ir_edge_t *next;
-       ir_node **pops;
-
-       arity = be_get_MemPerm_entity_arity(node);
-       pops = alloca(arity * sizeof(pops[0]));
 
        /* create Pushs */
        for(i = 0; i < arity; ++i) {
@@ -1312,12 +1263,12 @@ static void transform_MemPerm(ia32_code_gen_t *cg, ir_node *node) {
                assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
 
                push = create_push(cg, node, node, sp, mem, inent);
-               sp = create_spproj(cg, node, push, pn_ia32_Push_stack);
+               sp = create_spproj(node, push, pn_ia32_Push_stack);
                if(entsize == 8) {
                        /* add another push after the first one */
                        push = create_push(cg, node, node, sp, mem, inent);
                        add_ia32_am_offs_int(push, 4);
-                       sp = create_spproj(cg, node, push, pn_ia32_Push_stack);
+                       sp = create_spproj(node, push, pn_ia32_Push_stack);
                }
 
                set_irn_n(node, i, new_Bad());
@@ -1338,13 +1289,13 @@ static void transform_MemPerm(ia32_code_gen_t *cg, ir_node *node) {
                assert( (entsize == 4 || entsize == 8) && "spillslot on x86 should be 32 or 64 bit");
 
                pop = create_pop(cg, node, node, sp, outent);
-               sp = create_spproj(cg, node, pop, pn_ia32_Pop_stack);
+               sp = create_spproj(node, pop, pn_ia32_Pop_stack);
                if(entsize == 8) {
                        add_ia32_am_offs_int(pop, 4);
 
                        /* add another pop after the first one */
                        pop = create_pop(cg, node, node, sp, outent);
-                       sp = create_spproj(cg, node, pop, pn_ia32_Pop_stack);
+                       sp = create_spproj(node, pop, pn_ia32_Pop_stack);
                }
 
                pops[i] = pop;
@@ -1399,47 +1350,73 @@ static void ia32_after_ra_walker(ir_node *block, void *env) {
  */
 static void ia32_collect_frame_entity_nodes(ir_node *node, void *data)
 {
-       be_fec_env_t *env = data;
+       be_fec_env_t  *env = data;
+       const ir_mode *mode;
+       int            align;
 
        if (be_is_Reload(node) && be_get_frame_entity(node) == NULL) {
-               const ir_mode *mode = get_spill_mode_mode(get_irn_mode(node));
-               int align = get_mode_size_bytes(mode);
-               be_node_needs_frame_entity(env, node, mode, align);
-       } else if(is_ia32_irn(node) && get_ia32_frame_ent(node) == NULL
-                 && is_ia32_use_frame(node)) {
-               if (is_ia32_need_stackent(node) || is_ia32_Load(node)) {
-                       const ir_mode     *mode  = get_ia32_ls_mode(node);
-                       const ia32_attr_t *attr  = get_ia32_attr_const(node);
-                       int                align = get_mode_size_bytes(mode);
-
-                       if(attr->data.need_64bit_stackent) {
-                               mode = mode_Ls;
+               mode  = get_spill_mode_mode(get_irn_mode(node));
+               align = get_mode_size_bytes(mode);
+       } else if (is_ia32_irn(node)         &&
+                       get_ia32_frame_ent(node) == NULL &&
+                       is_ia32_use_frame(node)) {
+               if (is_ia32_need_stackent(node))
+                       goto need_stackent;
+
+               switch (get_ia32_irn_opcode(node)) {
+need_stackent:
+                       case iro_ia32_Load: {
+                               const ia32_attr_t *attr = get_ia32_attr_const(node);
+
+                               if (attr->data.need_32bit_stackent) {
+                                       mode = mode_Is;
+                               } else if (attr->data.need_64bit_stackent) {
+                                       mode = mode_Ls;
+                               } else {
+                                       mode = get_ia32_ls_mode(node);
+                                       if (is_ia32_is_reload(node))
+                                               mode = get_spill_mode_mode(mode);
+                               }
+                               align = get_mode_size_bytes(mode);
+                               break;
                        }
-                       if(attr->data.need_32bit_stackent) {
-                               mode = mode_Is;
+
+                       case iro_ia32_vfild:
+                       case iro_ia32_vfld:
+                       case iro_ia32_xLoad: {
+                               mode  = get_ia32_ls_mode(node);
+                               align = 4;
+                               break;
                        }
-                       be_node_needs_frame_entity(env, node, mode, align);
-               } else if (is_ia32_vfild(node) || is_ia32_xLoad(node)
-                          || is_ia32_vfld(node)) {
-                       const ir_mode *mode  = get_ia32_ls_mode(node);
-                       int            align = 4;
-                       be_node_needs_frame_entity(env, node, mode, align);
-               } else if(is_ia32_FldCW(node)) {
-                       /* although 2 byte would be enough 4 byte performs best */
-                       const ir_mode *mode  = mode_Iu;
-                       int            align = 4;
-                       be_node_needs_frame_entity(env, node, mode, align);
-               } else {
+
+                       case iro_ia32_FldCW: {
+                               /* although 2 byte would be enough 4 byte performs best */
+                               mode  = mode_Iu;
+                               align = 4;
+                               break;
+                       }
+
+                       default:
 #ifndef NDEBUG
-                       assert(is_ia32_St(node) ||
-                                  is_ia32_xStoreSimple(node) ||
-                                  is_ia32_vfst(node) ||
-                                  is_ia32_vfist(node) ||
-                                  is_ia32_vfisttp(node) ||
-                              is_ia32_FnstCW(node));
+                               panic("unexpected frame user while collection frame entity nodes");
+
+                       case iro_ia32_FnstCW:
+                       case iro_ia32_Store8Bit:
+                       case iro_ia32_Store:
+                       case iro_ia32_fst:
+                       case iro_ia32_fstp:
+                       case iro_ia32_vfist:
+                       case iro_ia32_vfisttp:
+                       case iro_ia32_vfst:
+                       case iro_ia32_xStore:
+                       case iro_ia32_xStoreSimple:
 #endif
+                               return;
                }
+       } else {
+               return;
        }
+       be_node_needs_frame_entity(env, node, mode, align);
 }
 
 /**
@@ -1472,7 +1449,7 @@ static void ia32_finish(void *self) {
 
        /* we might have to rewrite x87 virtual registers */
        if (cg->do_x87_sim) {
-               x87_simulate_graph(cg->arch_env, cg->birg);
+               x87_simulate_graph(cg->birg);
        }
 
        /* do peephole optimisations */
@@ -1517,11 +1494,10 @@ static ir_node *ia32_get_pic_base(void *self) {
                return get_eip;
 
        block       = get_irg_start_block(cg->irg);
-       get_eip     = new_rd_ia32_GetEIP(NULL, cg->irg, block);
+       get_eip     = new_bd_ia32_GetEIP(NULL, block);
        cg->get_eip = get_eip;
 
-       add_irn_dep(get_eip, get_irg_frame(cg->irg));
-
+       be_dep_on_frame(get_eip);
        return get_eip;
 }
 
@@ -1533,7 +1509,6 @@ static const arch_code_generator_if_t ia32_code_gen_if = {
        ia32_before_abi,     /* before abi introduce hook */
        ia32_prepare_graph,
        NULL,                /* spill */
-       ia32_before_sched,   /* before scheduling hook */
        ia32_before_ra,      /* before register allocation hook */
        ia32_after_ra,       /* after register allocation hook */
        ia32_finish,         /* called before codegen */
@@ -1545,13 +1520,12 @@ static const arch_code_generator_if_t ia32_code_gen_if = {
  */
 static void *ia32_cg_init(be_irg_t *birg) {
        ia32_isa_t      *isa = (ia32_isa_t *)birg->main_env->arch_env;
-       ia32_code_gen_t *cg  = xcalloc(1, sizeof(*cg));
+       ia32_code_gen_t *cg  = XMALLOCZ(ia32_code_gen_t);
 
        cg->impl      = &ia32_code_gen_if;
        cg->irg       = birg->irg;
        cg->reg_set   = new_set(ia32_cmp_irn_reg_assoc, 1024);
        cg->isa       = isa;
-       cg->arch_env  = birg->main_env->arch_env;
        cg->birg      = birg;
        cg->blk_sched = NULL;
        cg->dump      = (birg->main_env->options->dump_flags & DUMP_BE) ? 1 : 0;
@@ -1629,7 +1603,7 @@ static ia32_isa_t ia32_isa_template = {
                &ia32_gp_regs[REG_ESP],  /* stack pointer register */
                &ia32_gp_regs[REG_EBP],  /* base pointer register */
                -1,                      /* stack direction */
-               16,                      /* stack alignment */
+               2,                       /* power of two stack alignment, 2^2 == 4 */
                NULL,                    /* main environment */
                7,                       /* costs for a spill instruction */
                5,                       /* costs for a reload instruction */
@@ -1646,12 +1620,60 @@ static ia32_isa_t ia32_isa_template = {
 #endif
 };
 
+static void init_asm_constraints(void)
+{
+       be_init_default_asm_constraint_flags();
+
+       asm_constraint_flags['a'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['b'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['c'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['d'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['D'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['S'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['Q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['q'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['A'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['l'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['R'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['r'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['p'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['f'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['t'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['u'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['Y'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_SUPPORTS_REGISTER;
+       asm_constraint_flags['n'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
+       asm_constraint_flags['g'] = ASM_CONSTRAINT_FLAG_SUPPORTS_IMMEDIATE;
+
+       /* no support for autodecrement/autoincrement */
+       asm_constraint_flags['<'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       asm_constraint_flags['>'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no float consts */
+       asm_constraint_flags['E'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       asm_constraint_flags['F'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* makes no sense on x86 */
+       asm_constraint_flags['s'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no support for sse consts yet */
+       asm_constraint_flags['C'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no support for x87 consts yet */
+       asm_constraint_flags['G'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* no support for mmx registers yet */
+       asm_constraint_flags['y'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       /* not available in 32bit mode */
+       asm_constraint_flags['Z'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+       asm_constraint_flags['e'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+
+       /* no code yet to determine register class needed... */
+       asm_constraint_flags['X'] = ASM_CONSTRAINT_FLAG_NO_SUPPORT;
+}
+
 /**
  * Initializes the backend ISA.
  */
 static arch_env_t *ia32_init(FILE *file_handle) {
        static int inited = 0;
        ia32_isa_t *isa;
+       int        i, n;
 
        if (inited)
                return NULL;
@@ -1659,7 +1681,7 @@ static arch_env_t *ia32_init(FILE *file_handle) {
 
        set_tarval_output_modes();
 
-       isa = xmalloc(sizeof(*isa));
+       isa = XMALLOC(ia32_isa_t);
        memcpy(isa, &ia32_isa_template, sizeof(*isa));
 
        if(mode_fpcw == NULL) {
@@ -1682,13 +1704,20 @@ static arch_env_t *ia32_init(FILE *file_handle) {
        ia32_build_8bit_reg_map_high(isa->regs_8bit_high);
 
 #ifndef NDEBUG
-       isa->name_obst = xmalloc(sizeof(*isa->name_obst));
+       isa->name_obst = XMALLOC(struct obstack);
        obstack_init(isa->name_obst);
 #endif /* NDEBUG */
 
        /* enter the ISA object into the intrinsic environment */
        intrinsic_env.isa = isa;
-       ia32_handle_intrinsics();
+
+       /* emit asm includes */
+       n = get_irp_n_asms();
+       for (i = 0; i < n; ++i) {
+               be_emit_cstring("#APP\n");
+               be_emit_ident(get_irp_asm(i));
+               be_emit_cstring("\n#NO_APP\n");
+       }
 
        /* needed for the debug support */
        be_gas_emit_switch_section(GAS_SECTION_TEXT);
@@ -1786,7 +1815,9 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
        ir_mode  *mode;
        unsigned  cc;
        int       n, i, regnum;
+       int                 pop_amount = 0;
        be_abi_call_flags_t call_flags = be_abi_call_get_flags(abi);
+
        (void) self;
 
        /* set abi flags for calls */
@@ -1794,7 +1825,7 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
        call_flags.bits.store_args_sequential = 0;
        /* call_flags.bits.try_omit_fp                 not changed: can handle both settings */
        call_flags.bits.fp_free               = 0;  /* the frame pointer is fixed in IA32 */
-       call_flags.bits.call_has_imm          = 1;  /* IA32 calls can have immediate address */
+       call_flags.bits.call_has_imm          = 0;  /* No call immediates, we handle this by ourselves */
 
        /* set parameter passing style */
        be_abi_call_set_flags(abi, call_flags, &ia32_abi_callbacks);
@@ -1804,8 +1835,8 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
                cc = cc_cdecl_set;
        } else {
                cc = get_method_calling_convention(method_type);
-               if (get_method_additional_properties(method_type) & mtp_property_private
-                               && (ia32_cg_config.optimize_cc)) {
+               if (get_method_additional_properties(method_type) & mtp_property_private &&
+                   ia32_cg_config.optimize_cc) {
                        /* set the calling conventions to register parameter */
                        cc = (cc & ~cc_bits) | cc_reg_param;
                }
@@ -1814,7 +1845,7 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
        /* we have to pop the shadow parameter ourself for compound calls */
        if( (get_method_calling_convention(method_type) & cc_compound_ret)
                        && !(cc & cc_reg_param)) {
-               be_abi_call_set_pop(abi, get_mode_size_bytes(mode_P_data));
+               pop_amount += get_mode_size_bytes(mode_P_data);
        }
 
        n = get_method_n_params(method_type);
@@ -1834,11 +1865,23 @@ static void ia32_get_call_abi(const void *self, ir_type *method_type,
                        /* Micro optimisation: if the mode is shorter than 4 bytes, load 4 bytes.
                         * movl has a shorter opcode than mov[sz][bw]l */
                        ir_mode *load_mode = mode;
-                       if (mode != NULL && get_mode_size_bytes(mode) < 4) load_mode = mode_Iu;
+
+                       if (mode != NULL) {
+                               unsigned size = get_mode_size_bytes(mode);
+
+                               if (cc & cc_callee_clear_stk) {
+                                       pop_amount += (size + 3U) & ~3U;
+                               }
+
+                               if (size < 4) load_mode = mode_Iu;
+                       }
+
                        be_abi_call_param_stack(abi, i, load_mode, 4, 0, 0);
                }
        }
 
+       be_abi_call_set_pop(abi, pop_amount);
+
        /* set return registers */
        n = get_method_n_ress(method_type);
 
@@ -1981,16 +2024,13 @@ static const be_execution_unit_t ***ia32_get_allowed_execution_units(
 
        if (is_ia32_irn(irn)) {
                ret = get_ia32_exec_units(irn);
-       }
-       else if (is_be_node(irn)) {
-               if (be_is_Call(irn) || be_is_Return(irn)) {
+       } else if (is_be_node(irn)) {
+               if (be_is_Return(irn)) {
                        ret = _units_callret;
-               }
-               else if (be_is_Barrier(irn)) {
+               } else if (be_is_Barrier(irn)) {
                        ret = _units_dummy;
-               }
-               else {
-                        ret = _units_other;
+               } else {
+                       ret = _units_other;
                }
        }
        else {
@@ -2018,34 +2058,131 @@ static ir_graph **ia32_get_irg_list(const void *self, ir_graph ***irg_list)
        return NULL;
 }
 
+static void ia32_mark_remat(const void *self, ir_node *node) {
+       (void) self;
+       if (is_ia32_irn(node)) {
+               set_ia32_is_remat(node);
+       }
+}
+
 /**
- * Allows or disallows the creation of Psi nodes for the given Phi nodes.
+ * Check for Abs or -Abs.
+ */
+static int psi_is_Abs_or_Nabs(ir_node *cmp, ir_node *sel, ir_node *t, ir_node *f) {
+       ir_node *l, *r;
+       pn_Cmp  pnc;
+
+       if (cmp == NULL)
+               return 0;
+
+       /* must be <, <=, >=, > */
+       pnc = get_Proj_proj(sel);
+       if (pnc != pn_Cmp_Ge && pnc != pn_Cmp_Gt &&
+               pnc != pn_Cmp_Le && pnc != pn_Cmp_Lt)
+               return 0;
+
+       l = get_Cmp_left(cmp);
+       r = get_Cmp_right(cmp);
+
+       /* must be x cmp 0 */
+       if ((l != t && l != f) || !is_Const(r) || !is_Const_null(r))
+               return 0;
+
+       if ((!is_Minus(t) || get_Minus_op(t) != f) &&
+               (!is_Minus(f) || get_Minus_op(f) != t))
+               return 0;
+       return 1;
+}
+
+/**
+ * Check for Abs only
+ */
+static int psi_is_Abs(ir_node *cmp, ir_node *sel, ir_node *t, ir_node *f) {
+       ir_node *l, *r;
+       pn_Cmp  pnc;
+
+       if (cmp == NULL)
+               return 0;
+
+       /* must be <, <=, >=, > */
+       pnc = get_Proj_proj(sel);
+       if (pnc != pn_Cmp_Ge && pnc != pn_Cmp_Gt &&
+               pnc != pn_Cmp_Le && pnc != pn_Cmp_Lt)
+               return 0;
+
+       l = get_Cmp_left(cmp);
+       r = get_Cmp_right(cmp);
+
+       /* must be x cmp 0 */
+       if ((l != t && l != f) || !is_Const(r) || !is_Const_null(r))
+               return 0;
+
+       if ((!is_Minus(t) || get_Minus_op(t) != f) &&
+               (!is_Minus(f) || get_Minus_op(f) != t))
+               return 0;
+
+       if (pnc & pn_Cmp_Gt) {
+               /* x >= 0 ? -x : x is NABS */
+               if (is_Minus(t))
+                       return 0;
+       } else {
+               /* x < 0 ? x : -x is NABS */
+               if (is_Minus(f))
+                       return 0;
+       }
+       return 1;
+}
+
+
+/**
+ * Allows or disallows the creation of Mux nodes for the given Phi nodes.
+ *
+ * @param sel        A selector of a Cond.
+ * @param phi_list   List of Phi nodes about to be converted (linked via get_Phi_next() field)
+ * @param i          First data predecessor involved in if conversion
+ * @param j          Second data predecessor involved in if conversion
+ *
  * @return 1 if allowed, 0 otherwise
  */
-static int ia32_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j)
+static int ia32_is_mux_allowed(ir_node *sel, ir_node *phi_list, int i, int j)
 {
        ir_node *phi;
-       ir_node *cmp = NULL;
+       ir_node *cmp;
+       pn_Cmp  pn;
+       ir_node *cl, *cr;
 
-       /* we can't handle psis with 64bit compares yet */
+       /* we can't handle Muxs with 64bit compares yet */
        if (is_Proj(sel)) {
                cmp = get_Proj_pred(sel);
                if (is_Cmp(cmp)) {
                        ir_node *left     = get_Cmp_left(cmp);
                        ir_mode *cmp_mode = get_irn_mode(left);
-                       if (!mode_is_float(cmp_mode) && get_mode_size_bits(cmp_mode) > 32)
-                               return 0;
+                       if (!mode_is_float(cmp_mode) && get_mode_size_bits(cmp_mode) > 32) {
+                               /* 64bit Abs IS supported */
+                               for (phi = phi_list; phi; phi = get_Phi_next(phi)) {
+                                       ir_node *t = get_Phi_pred(phi, i);
+                                       ir_node *f = get_Phi_pred(phi, j);
+
+                                       if (! psi_is_Abs(cmp, sel, t, f))
+                                               return 0;
+                               }
+                               return 1;
+                       }
                } else {
-                       cmp = NULL;
+                       /* we do not support nodes without Cmp yet */
+                       return 0;
                }
+       } else {
+               /* we do not support nodes without Cmp yet */
+               return 0;
        }
 
-       if (ia32_cg_config.use_cmov) {
-               if (ia32_cg_config.use_sse2 && cmp != NULL) {
-                       pn_Cmp pn   = get_Proj_proj(sel);
-                       ir_node *cl = get_Cmp_left(cmp);
-                       ir_node *cr = get_Cmp_right(cmp);
+       pn = get_Proj_proj(sel);
+       cl = get_Cmp_left(cmp);
+       cr = get_Cmp_right(cmp);
 
+       if (ia32_cg_config.use_cmov) {
+               if (ia32_cg_config.use_sse2) {
                        /* check the Phi nodes: no 64bit and no floating point cmov */
                        for (phi = phi_list; phi; phi = get_Phi_next(phi)) {
                                ir_mode *mode = get_irn_mode(phi);
@@ -2054,97 +2191,105 @@ static int ia32_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j)
                                        /* check for Min, Max */
                                        ir_node *t = get_Phi_pred(phi, i);
                                        ir_node *f = get_Phi_pred(phi, j);
-                                       int res    = 0;
 
                                        /* SSE2 supports Min & Max */
                                        if (pn == pn_Cmp_Lt || pn == pn_Cmp_Le || pn == pn_Cmp_Ge || pn == pn_Cmp_Gt) {
                                                if (cl == t && cr == f) {
-                                                       /* Psi(a <=/>= b, a, b) => MIN, MAX */
-                                                       res = 1;
+                                                       /* Mux(a <=/>= b, a, b) => MIN, MAX */
+                                                       continue;
                                                } else if (cl == f && cr == t) {
-                                                       /* Psi(a <=/>= b, b, a) => MAX, MIN */
-                                                       res = 1;
+                                                       /* Mux(a <=/>= b, b, a) => MAX, MIN */
+                                                       continue;
                                                }
                                        }
-                                       if (! res)
-                                               return 0;
-
-                               } else if (get_mode_size_bits(mode) > 32)
                                        return 0;
+                               } else if (get_mode_size_bits(mode) > 32) {
+                                       /* no 64bit cmov */
+                                       return 0;
+                               }
                        }
                } else {
                        /* check the Phi nodes: no 64bit and no floating point cmov */
                        for (phi = phi_list; phi; phi = get_Phi_next(phi)) {
                                ir_mode *mode = get_irn_mode(phi);
 
-                               if (mode_is_float(mode) || get_mode_size_bits(mode) > 32)
+                               if (mode_is_float(mode)) {
+                                       ir_node *t = get_Phi_pred(phi, i);
+                                       ir_node *f = get_Phi_pred(phi, j);
+
+                                       /* always support Mux(!float, C1, C2) */
+                                       if (is_Const(t) && is_Const(f) && !mode_is_float(get_irn_mode(cl)))
+                                               continue;
+                                       /* only abs or nabs supported */
+                                       if (! psi_is_Abs_or_Nabs(cmp, sel, t, f))
+                                               return 0;
+                               } else if (get_mode_size_bits(mode) > 32)
                                        return 0;
                        }
                }
 
                return 1;
-       } else {
-               ir_node *cl, *cr;
-               pn_Cmp  pn;
-
-               /* No cmov, only some special cases */
-               if (cmp == NULL)
-                       return 0;
+       } else { /* No Cmov, only some special cases */
 
                /* Now some supported cases here */
-               pn = get_Proj_proj(sel);
-               cl = get_Cmp_left(cmp);
-               cr = get_Cmp_right(cmp);
-
                for (phi = phi_list; phi; phi = get_Phi_next(phi)) {
                        ir_mode *mode = get_irn_mode(phi);
-                       int res = 0;
                        ir_node *t, *f;
 
                        t = get_Phi_pred(phi, i);
                        f = get_Phi_pred(phi, j);
 
-                       /* no floating point and no 64bit yet */
-                       if (mode_is_float(mode) || get_mode_size_bits(mode) > 32)
+                       if (mode_is_float(mode)) {
+                               /* always support Mux(!float, C1, C2) */
+                               if (is_Const(t) && is_Const(f) && !mode_is_float(get_irn_mode(cl)))
+                                       continue;
+                               /* only abs or nabs supported */
+                               if (! psi_is_Abs_or_Nabs(cmp, sel, t, f))
+                                       return 0;
+                       } else if (get_mode_size_bits(mode) > 32) {
+                               /* no 64bit yet */
                                return 0;
+                       }
 
                        if (is_Const(t) && is_Const(f)) {
                                if ((is_Const_null(t) && is_Const_one(f)) || (is_Const_one(t) && is_Const_null(f))) {
-                                       /* always support Psi(x, C1, C2) */
-                                       res = 1;
+                                       /* always support Mux(x, C1, C2) */
+                                       continue;
                                }
                        } else if (pn == pn_Cmp_Lt || pn == pn_Cmp_Le || pn == pn_Cmp_Ge || pn == pn_Cmp_Gt) {
-                               if (0) {
 #if 0
-                               } else if (cl == t && cr == f) {
-                                       /* Psi(a <=/>= b, a, b) => Min, Max */
-                                       res = 1;
-                               } else if (cl == f && cr == t) {
-                                       /* Psi(a <=/>= b, b, a) => Max, Min */
-                                       res = 1;
+                               if (cl == t && cr == f) {
+                                       /* Mux(a <=/>= b, a, b) => Min, Max */
+                                       continue;
+                               }
+                               if (cl == f && cr == t) {
+                                       /* Mux(a <=/>= b, b, a) => Max, Min */
+                                       continue;
+                               }
 #endif
-                               } else if ((pn & pn_Cmp_Gt) && !mode_is_signed(mode) &&
-                                          is_Const(f) && is_Const_null(f) && is_Sub(t) &&
-                                          get_Sub_left(t) == cl && get_Sub_right(t) == cr) {
-                                       /* Psi(a >=u b, a - b, 0) unsigned Doz */
-                                       res = 1;
-                               } else if ((pn & pn_Cmp_Lt) && !mode_is_signed(mode) &&
-                                          is_Const(t) && is_Const_null(t) && is_Sub(f) &&
-                                          get_Sub_left(f) == cl && get_Sub_right(f) == cr) {
-                                       /* Psi(a <=u b, 0, a - b) unsigned Doz */
-                                       res = 1;
-                               } else if (is_Const(cr) && is_Const_null(cr)) {
+                               if ((pn & pn_Cmp_Gt) && !mode_is_signed(mode) &&
+                                   is_Const(f) && is_Const_null(f) && is_Sub(t) &&
+                                   get_Sub_left(t) == cl && get_Sub_right(t) == cr) {
+                                       /* Mux(a >=u b, a - b, 0) unsigned Doz */
+                                       continue;
+                               }
+                               if ((pn & pn_Cmp_Lt) && !mode_is_signed(mode) &&
+                                   is_Const(t) && is_Const_null(t) && is_Sub(f) &&
+                                   get_Sub_left(f) == cl && get_Sub_right(f) == cr) {
+                                       /* Mux(a <=u b, 0, a - b) unsigned Doz */
+                                       continue;
+                               }
+                               if (is_Const(cr) && is_Const_null(cr)) {
                                        if (cl == t && is_Minus(f) && get_Minus_op(f) == cl) {
-                                               /* Psi(a <=/>= 0 ? a : -a) Nabs/Abs */
-                                               res = 1;
+                                               /* Mux(a <=/>= 0 ? a : -a) Nabs/Abs */
+                                               continue;
                                        } else if (cl == f && is_Minus(t) && get_Minus_op(t) == cl) {
-                                               /* Psi(a <=/>= 0 ? -a : a) Abs/Nabs */
-                                               res = 1;
+                                               /* Mux(a <=/>= 0 ? -a : a) Abs/Nabs */
+                                               continue;
                                        }
                                }
                        }
-                       if (! res)
-                               return 0;
+                       return 0;
                }
                /* all checks passed */
                return 1;
@@ -2152,13 +2297,57 @@ static int ia32_is_psi_allowed(ir_node *sel, ir_node *phi_list, int i, int j)
        return 0;
 }
 
+static asm_constraint_flags_t ia32_parse_asm_constraint(const void *self, const char **c)
+{
+       (void) self;
+       (void) c;
+
+       /* we already added all our simple flags to the flags modifier list in
+        * init, so this flag we don't know. */
+       return ASM_CONSTRAINT_FLAG_INVALID;
+}
+
+static int ia32_is_valid_clobber(const void *self, const char *clobber)
+{
+       (void) self;
+
+       return ia32_get_clobber_register(clobber) != NULL;
+}
+
+/**
+ * Create the trampoline code.
+ */
+static ir_node *ia32_create_trampoline_fkt(ir_node *block, ir_node *mem, ir_node *trampoline, ir_node *env, ir_node *callee)
+{
+       ir_graph *irg    = get_Block_irg(block);
+       ir_node  *st, *p = trampoline;
+       ir_mode *mode    = get_irn_mode(p);
+
+       /* mov  ecx,<env> */
+       st  = new_r_Store(irg, block, mem, p, new_Const_long(mode_Bu, 0xb9), 0);
+       mem = new_r_Proj(irg, block, st, mode_M, pn_Store_M);
+       p   = new_r_Add(irg, block, p, new_Const_long(mode_Iu, 1), mode);
+       st  = new_r_Store(irg, block, mem, p, env, 0);
+       mem = new_r_Proj(irg, block, st, mode_M, pn_Store_M);
+       p   = new_r_Add(irg, block, p, new_Const_long(mode_Iu, 4), mode);
+       /* jmp  <callee> */
+       st  = new_r_Store(irg, block, mem, p, new_Const_long(mode_Bu, 0xe9), 0);
+       mem = new_r_Proj(irg, block, st, mode_M, pn_Store_M);
+       p   = new_r_Add(irg, block, p, new_Const_long(mode_Iu, 1), mode);
+       st  = new_r_Store(irg, block, mem, p, callee, 0);
+       mem = new_r_Proj(irg, block, st, mode_M, pn_Store_M);
+       p   = new_r_Add(irg, block, p, new_Const_long(mode_Iu, 4), mode);
+
+       return mem;
+}
+
 /**
  * Returns the libFirm configuration parameter for this backend.
  */
 static const backend_params *ia32_get_libfirm_params(void) {
        static const ir_settings_if_conv_t ifconv = {
-               4,                    /* maxdepth, doesn't matter for Psi-conversion */
-               ia32_is_psi_allowed   /* allows or disallows Psi creation for given selector */
+               4,                    /* maxdepth, doesn't matter for Mux-conversion */
+               ia32_is_mux_allowed   /* allows or disallows Mux creation for given selector */
        };
        static const ir_settings_arch_dep_t ad = {
                1,                   /* also use subs */
@@ -2168,22 +2357,30 @@ static const backend_params *ia32_get_libfirm_params(void) {
 
                1,  /* allow Mulhs */
                1,  /* allow Mulus */
-               32  /* Mulh allowed up to 32 bit */
+               32, /* Mulh allowed up to 32 bit */
        };
        static backend_params p = {
                1,     /* need dword lowering */
                1,     /* support inline assembly */
-               NULL,  /* no additional opcodes */
                NULL,  /* will be set later */
                ia32_create_intrinsic_fkt,
                &intrinsic_env,  /* context for ia32_create_intrinsic_fkt */
-               NULL,  /* will be set below */
+               NULL,  /* ifconv info will be set below */
+               NULL,  /* float arithmetic mode, will be set below */
+               12,    /* size of trampoline code */
+               4,     /* alignment of trampoline code */
+               ia32_create_trampoline_fkt,
        };
 
        ia32_setup_cg_config();
 
-       p.dep_param    = &ad;
-       p.if_conv_info = &ifconv;
+       /* doesn't really belong here, but this is the earliest place the backend
+        * is called... */
+       init_asm_constraints();
+
+       p.dep_param             = &ad;
+       p.if_conv_info          = &ifconv;
+       p.mode_float_arithmetic = mode_E;
        return &p;
 }
 
@@ -2199,9 +2396,25 @@ static lc_opt_enum_int_var_t gas_var = {
        (int*) &be_gas_flavour, gas_items
 };
 
+#ifdef FIRM_GRGEN_BE
+static const lc_opt_enum_int_items_t transformer_items[] = {
+       { "default", TRANSFORMER_DEFAULT },
+       { "pbqp",    TRANSFORMER_PBQP    },
+       { "random",  TRANSFORMER_RAND    },
+       { NULL,      0                   }
+};
+
+static lc_opt_enum_int_var_t transformer_var = {
+       (int*)&be_transformer, transformer_items
+};
+#endif
+
 static const lc_opt_table_entry_t ia32_options[] = {
        LC_OPT_ENT_ENUM_INT("gasmode", "set the GAS compatibility mode", &gas_var),
-       LC_OPT_ENT_INT("stackalign", "set stack alignment for calls",
+#ifdef FIRM_GRGEN_BE
+       LC_OPT_ENT_ENUM_INT("transformer", "the transformer used for code selection", &transformer_var),
+#endif
+       LC_OPT_ENT_INT("stackalign", "set power of two stack alignment for calls",
                       &ia32_isa_template.arch_env.stack_alignment),
        LC_OPT_LAST
 };
@@ -2209,6 +2422,7 @@ static const lc_opt_table_entry_t ia32_options[] = {
 const arch_isa_if_t ia32_isa_if = {
        ia32_init,
        ia32_done,
+       ia32_handle_intrinsics,
        ia32_get_n_reg_class,
        ia32_get_reg_class,
        ia32_get_reg_class_for_mode,
@@ -2221,14 +2435,11 @@ const arch_isa_if_t ia32_isa_if = {
        ia32_get_allowed_execution_units,
        ia32_get_machine,
        ia32_get_irg_list,
+       ia32_mark_remat,
+       ia32_parse_asm_constraint,
+       ia32_is_valid_clobber
 };
 
-void ia32_init_emitter(void);
-void ia32_init_finish(void);
-void ia32_init_optimize(void);
-void ia32_init_transform(void);
-void ia32_init_x87(void);
-
 void be_init_arch_ia32(void)
 {
        lc_opt_entry_t *be_grp   = lc_opt_get_grp(firm_opt_get_root(), "be");