- Improved addressmode optimisation for conv nodes
[libfirm] / ir / be / ia32 / ia32_optimize.c
index 212676a..0d434a7 100644 (file)
@@ -108,12 +108,12 @@ static ir_node *gen_SymConst(ia32_transform_env_t *env) {
        if (mode_is_float(mode)) {
                FP_USED(env->cg);
                if (USE_SSE2(env->cg))
-                       cnst = new_rd_ia32_xConst(dbg, irg, block, get_irg_no_mem(irg), mode);
+                       cnst = new_rd_ia32_xConst(dbg, irg, block, mode);
                else
-                       cnst = new_rd_ia32_vfConst(dbg, irg, block, get_irg_no_mem(irg), mode);
+                       cnst = new_rd_ia32_vfConst(dbg, irg, block, mode);
        }
        else
-               cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), mode);
+               cnst = new_rd_ia32_Const(dbg, irg, block, mode);
 
        set_ia32_Const_attr(cnst, env->irn);
 
@@ -213,13 +213,16 @@ static ir_node *gen_Const(ia32_transform_env_t *env) {
                env->irn  = cnst;
                env->mode = mode_P;
                cnst      = gen_SymConst(env);
+               add_irn_dep(cnst, be_abi_get_start_barrier(env->cg->birg->abi));
                set_Load_ptr(get_Proj_pred(load), cnst);
                cnst      = load;
        }
        else {
-               cnst = new_rd_ia32_Const(dbg, irg, block, get_irg_no_mem(irg), get_irn_mode(node));
+               cnst = new_rd_ia32_Const(dbg, irg, block, get_irn_mode(node));
+               add_irn_dep(cnst, be_abi_get_start_barrier(env->cg->birg->abi));
                set_ia32_Const_attr(cnst, node);
        }
+
        return cnst;
 }
 
@@ -241,13 +244,15 @@ static void ia32_transform_const(ir_node *irn, void *env) {
        tenv.irn  = irn;
        DEBUG_ONLY(tenv.mod = cg->mod;)
 
-#if 0
+#if 1
        /* place const either in the smallest dominator of all its users or the original block */
        if (cg->opt & IA32_OPT_PLACECNST)
                tenv.block = node_users_smallest_common_dominator(irn, 1);
        else
                tenv.block = get_nodes_block(irn);
 #else
+       /* Actually, there is no real sense in placing     */
+       /* the Consts in the successor of the start block. */
        {
                ir_node *afterstart = NULL;
                ir_node *startblock = get_irg_start_block(tenv.irg);
@@ -288,7 +293,7 @@ static void ia32_transform_const(ir_node *irn, void *env) {
 static void ia32_place_consts_walker(ir_node *irn, void *env) {
        ia32_code_gen_t *cg = env;
 
-       if(!is_Const(irn) && !is_SymConst(irn))
+       if (! is_Const(irn) && ! is_SymConst(irn))
                return;
 
        ia32_transform_const(irn, cg);
@@ -809,14 +814,19 @@ static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const i
                int n;
                n         = ia32_get_irn_n_edges(in);
                is_cand   = (n == 1) ? 1 : is_cand;  /* load with more than one user: no AM */
+#else
+               is_cand   = 1;
 #endif
 
                load  = get_Proj_pred(in);
                other = right;
 
-               /* 8bit Loads are not supported, they cannot be used with every register */
-               if (get_mode_size_bits(get_ia32_ls_mode(load)) < 16)
+               /* 8bit Loads are not supported (for binary ops),
+                * they cannot be used with every register */
+               if (get_irn_arity(irn) != 4 && get_mode_size_bits(get_ia32_ls_mode(load)) < 16) {
+                       assert(get_irn_arity(irn) == 5);
                        is_cand = 0;
+               }
 
                /* If there is a data dependency of other irn from load: cannot use AM */
                if (is_cand && get_nodes_block(other) == block) {
@@ -1543,6 +1553,7 @@ static void optimize_am(ir_node *irn, void *env) {
                if (get_irn_arity(irn) == 4) {
                        /* it's an "unary" operation */
                        right = left;
+                       cand = IA32_AM_CAND_BOTH;
                }
                else {
                        right = get_irn_n(irn, 3);
@@ -1699,56 +1710,61 @@ static void optimize_am(ir_node *irn, void *env) {
                /* and right operand is a Load which only used by this irn */
                if (check_am_src                &&
                        (cand & IA32_AM_CAND_RIGHT) &&
-                       (get_irn_arity(irn) == 5)   &&
                        (ia32_get_irn_n_edges(right) == 1))
                {
-                       right = get_Proj_pred(right);
+                       ir_node *load = get_Proj_pred(right);
 
-                       addr_b = get_irn_n(right, 0);
-                       addr_i = get_irn_n(right, 1);
+                       addr_b = get_irn_n(load, 0);
+                       addr_i = get_irn_n(load, 1);
 
                        /* set new base, index and attributes */
                        set_irn_n(irn, 0, addr_b);
                        set_irn_n(irn, 1, addr_i);
-                       add_ia32_am_offs(irn, get_ia32_am_offs(right));
-                       set_ia32_am_scale(irn, get_ia32_am_scale(right));
-                       set_ia32_am_flavour(irn, get_ia32_am_flavour(right));
+                       add_ia32_am_offs(irn, get_ia32_am_offs(load));
+                       set_ia32_am_scale(irn, get_ia32_am_scale(load));
+                       set_ia32_am_flavour(irn, get_ia32_am_flavour(load));
                        set_ia32_op_type(irn, ia32_AddrModeS);
-                       set_ia32_frame_ent(irn, get_ia32_frame_ent(right));
-                       set_ia32_ls_mode(irn, get_ia32_ls_mode(right));
+                       set_ia32_frame_ent(irn, get_ia32_frame_ent(load));
+                       set_ia32_ls_mode(irn, get_ia32_ls_mode(load));
 
-                       set_ia32_am_sc(irn, get_ia32_am_sc(right));
-                       if (is_ia32_am_sc_sign(right))
+                       set_ia32_am_sc(irn, get_ia32_am_sc(load));
+                       if (is_ia32_am_sc_sign(load))
                                set_ia32_am_sc_sign(irn);
 
                        /* clear remat flag */
                        set_ia32_flags(irn, get_ia32_flags(irn) & ~arch_irn_flags_rematerializable);
 
-                       if (is_ia32_use_frame(right))
+                       if (is_ia32_use_frame(load))
                                set_ia32_use_frame(irn);
 
-                       /* connect to Load memory */
-                       set_irn_n(irn, 4, get_irn_n(right, 2));
+                       /* connect to Load memory and disconnect Load */
+                       if (get_irn_arity(irn) == 5) {
+                               /* binary AMop */
+                               set_irn_n(irn, 4, get_irn_n(load, 2));
+                               set_irn_n(irn, 3, ia32_get_admissible_noreg(cg, irn, 3));
+                       } else {
+                               assert(get_irn_arity(irn) == 4);
+                               /* unary AMop */
+                               set_irn_n(irn, 3, get_irn_n(load, 2));
+                               set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 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 */
-                       set_irn_n(irn, 3, ia32_get_admissible_noreg(cg, irn, 3));
-
-                       DBG_OPT_AM_S(right, irn);
+                       DBG_OPT_AM_S(load, irn);
 
                        /* If Load has a memory Proj, connect it to the op */
-                       mem_proj = ia32_get_proj_for_mode(right, mode_M);
+                       mem_proj = ia32_get_proj_for_mode(load, mode_M);
                        if (mem_proj) {
                                set_Proj_pred(mem_proj, irn);
                                set_Proj_proj(mem_proj, 1);
                        }
 
-                       try_remove_from_sched(right);
+                       try_remove_from_sched(load);
 
-                       DB((mod, LEVEL_1, "merged with %+F into source AM\n", right));
+                       DB((mod, LEVEL_1, "merged with %+F into source AM\n", load));
                }
                else {
                        /* was exchanged but optimize failed: exchange back */