removed firm proj num translation function, existing projs are renumbered instead
[libfirm] / ir / be / ia32 / ia32_optimize.c
index 5f7ac53..212676a 100644 (file)
@@ -35,6 +35,8 @@
 #include "ia32_dbg_stat.h"
 #include "ia32_util.h"
 
+#define AGGRESSIVE_AM
+
 typedef enum {
        IA32_AM_CAND_NONE  = 0,
        IA32_AM_CAND_LEFT  = 1,
@@ -239,11 +241,29 @@ static void ia32_transform_const(ir_node *irn, void *env) {
        tenv.irn  = irn;
        DEBUG_ONLY(tenv.mod = cg->mod;)
 
+#if 0
        /* 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
+       {
+               ir_node *afterstart = NULL;
+               ir_node *startblock = get_irg_start_block(tenv.irg);
+               const ir_edge_t *edge;
+
+               foreach_block_succ(startblock, edge) {
+                       ir_node *block = get_edge_src_irn(edge);
+                       if (block != startblock) {
+                               afterstart = block;
+                               break;
+                       }
+               }
+               assert(afterstart != NULL);
+               tenv.block = afterstart;
+       }
+#endif
 
        switch (get_irn_opcode(irn)) {
                case iro_Const:
@@ -447,78 +467,6 @@ static void ia32_optimize_CondJmp(ir_node *irn, ia32_code_gen_t *cg) {
        }
 }
 
-#if 0
-/**
- * Creates a Push from Store(IncSP(gp_reg_size))
- */
-static void ia32_create_Push(ir_node *irn, ia32_code_gen_t *cg) {
-       ir_node  *sp  = get_irn_n(irn, 0);
-       ir_graph *irg = cg->irg;
-       ir_node *val, *next, *push, *bl, *proj_M, *proj_res, *old_proj_M, *mem;
-       const ir_edge_t *edge;
-       heights_t *h;
-
-       /* do not create push if store has already an offset assigned or base is not a IncSP */
-       if (get_ia32_am_offs(irn) || ! be_is_IncSP(sp))
-               return;
-
-       /* do not create push if index is not NOREG */
-       if (arch_get_irn_register(cg->arch_env, get_irn_n(irn, 1)) !=
-               &ia32_gp_regs[REG_GP_NOREG])
-               return;
-
-       /* do not create push for floating point */
-       val = get_irn_n(irn, 2);
-       if (mode_is_float(get_irn_mode(val)))
-               return;
-
-       /* do not create push if IncSp doesn't expand stack or expand size is different from register size */
-       if (be_get_IncSP_offset(sp) != get_mode_size_bytes(ia32_reg_classes[CLASS_ia32_gp].mode))
-               return;
-
-       /* do not create push, if there is a path (inside the block) from the push value to IncSP */
-       h = heights_new(cg->irg);
-       if (get_nodes_block(val) == get_nodes_block(sp) &&
-               heights_reachable_in_block(h, val, sp))
-       {
-               heights_free(h);
-               return;
-       }
-       heights_free(h);
-
-       /* ok, translate into Push */
-       edge       = get_irn_out_edge_first(irn);
-       old_proj_M = get_edge_src_irn(edge);
-       bl         = get_nodes_block(irn);
-
-       next = sched_next(irn);
-       sched_remove(irn);
-       sched_remove(sp);
-
-       /*
-               build memory input:
-               if the IncSP points to NoMem -> just use the memory input from store
-               if IncSP points to somewhere else -> sync memory of IncSP and Store
-       */
-       mem  = get_irn_n(irn, 3);
-       push = new_rd_ia32_Push(NULL, irg, bl, be_get_IncSP_pred(sp), val, mem);
-       proj_res = new_r_Proj(irg, bl, push, get_irn_mode(sp), pn_ia32_Push_stack);
-       proj_M   = new_r_Proj(irg, bl, push, mode_M, pn_ia32_Push_M);
-
-       /* copy a possible constant from the store */
-       set_ia32_id_cnst(push, get_ia32_id_cnst(irn));
-       set_ia32_immop_type(push, get_ia32_immop_type(irn));
-
-       /* the push must have SP out register */
-       arch_set_irn_register(cg->arch_env, push, arch_get_irn_register(cg->arch_env, sp));
-
-       exchange(old_proj_M, proj_M);
-       exchange(sp, proj_res);
-       sched_add_before(next, push);
-       sched_add_after(push, proj_res);
-}
-#endif
-
 // only optimize up to 48 stores behind IncSPs
 #define MAXPUSH_OPTIMIZE       48
 
@@ -614,6 +562,7 @@ static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
                ir_node *push;
                ir_node *val, *mem;
                ir_node *store = stores[i];
+               ir_node *noreg = ia32_new_NoReg_gp(cg);
 
                if(store == NULL || is_Bad(store))
                        break;
@@ -623,7 +572,7 @@ static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
                spreg = arch_get_irn_register(cg->arch_env, curr_sp);
 
                // create a push
-               push = new_rd_ia32_Push(NULL, irg, block, curr_sp, val, mem);
+               push = new_rd_ia32_Push(NULL, irg, block, noreg, noreg, val, curr_sp, mem);
                if(get_ia32_immop_type(store) != ia32_ImmNone) {
                        copy_ia32_Immop_attr(push, store);
                }
@@ -642,7 +591,11 @@ static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
                        set_irn_n(succ, 0, push);
                }
 
-               // we can remove the store from schedule now
+               // we can remove the store now
+               set_irn_n(store, 0, new_Bad());
+               set_irn_n(store, 1, new_Bad());
+               set_irn_n(store, 2, new_Bad());
+               set_irn_n(store, 3, new_Bad());
                sched_remove(store);
 
                offset -= 4;
@@ -654,9 +607,6 @@ static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
        if(offset == 0) {
                const ir_edge_t *edge, *next;
 
-               sched_remove(irn);
-               set_irn_n(irn, 0, new_Bad());
-
                foreach_out_edge_safe(irn, edge, next) {
                        ir_node *arg = get_edge_src_irn(edge);
                        int pos = get_edge_src_pos(edge);
@@ -664,8 +614,8 @@ static void ia32_create_Pushs(ir_node *irn, ia32_code_gen_t *cg) {
                        set_irn_n(arg, pos, curr_sp);
                }
 
-               sched_remove(irn);
                set_irn_n(irn, 0, new_Bad());
+               sched_remove(irn);
        } else {
                set_irn_n(irn, 0, curr_sp);
        }
@@ -687,6 +637,8 @@ static void ia32_optimize_IncSP(ir_node *irn, ia32_code_gen_t *cg) {
 
                /* Omit the optimized IncSP */
                be_set_IncSP_pred(irn, be_get_IncSP_pred(prev));
+
+               set_irn_n(prev, 0, new_Bad());
                sched_remove(prev);
        }
 }
@@ -708,6 +660,7 @@ static void ia32_peephole_optimize_node(ir_node *irn, void *env) {
        if (be_is_IncSP(irn)) {
                // optimize_IncSP doesn't respect dependency edges yet...
                //ia32_optimize_IncSP(irn, cg);
+               (void) ia32_optimize_IncSP;
                ia32_create_Pushs(irn, cg);
        }
 }
@@ -716,7 +669,6 @@ void ia32_peephole_optimization(ir_graph *irg, ia32_code_gen_t *cg) {
        irg_walk_graph(irg, ia32_peephole_optimize_node, NULL, cg);
 }
 
-
 /******************************************************************
  *              _     _                   __  __           _
  *     /\      | |   | |                 |  \/  |         | |
@@ -804,6 +756,7 @@ static int is_addr_candidate(const ir_node *block, const ir_node *irn) {
 
        in = left;
 
+#ifndef AGGRESSIVE_AM
        if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
                n         = ia32_get_irn_n_edges(in);
                is_cand   = (n == 1) ? 0 : is_cand;  /* load with only one user: don't create LEA */
@@ -815,6 +768,9 @@ static int is_addr_candidate(const ir_node *block, const ir_node *irn) {
                n         = ia32_get_irn_n_edges(in);
                is_cand   = (n == 1) ? 0 : is_cand;  /* load with only one user: don't create LEA */
        }
+#else
+       (void) n;
+#endif
 
        is_cand = get_ia32_frame_ent(irn) ? 1 : is_cand;
 
@@ -837,7 +793,7 @@ static int is_addr_candidate(const ir_node *block, const ir_node *irn) {
  */
 static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const ir_node *block, ir_node *irn) {
        ir_node *in, *load, *other, *left, *right;
-       int      n, is_cand = 0, cand;
+       int      is_cand = 0, cand;
 
        if (is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn) || is_ia32_vfild(irn) || is_ia32_vfist(irn) ||
                is_ia32_GetST0(irn) || is_ia32_SetST0(irn) || is_ia32_xStoreSimple(irn))
@@ -849,8 +805,11 @@ static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const i
        in = left;
 
        if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
+#ifndef AGGRESSIVE_AM
+               int n;
                n         = ia32_get_irn_n_edges(in);
                is_cand   = (n == 1) ? 1 : is_cand;  /* load with more than one user: no AM */
+#endif
 
                load  = get_Proj_pred(in);
                other = right;
@@ -873,8 +832,11 @@ static ia32_am_cand_t is_am_candidate(ia32_code_gen_t *cg, heights_t *h, const i
        is_cand = 0;
 
        if (pred_is_specific_nodeblock(block, in, is_ia32_Ld)) {
+#ifndef AGGRESSIVE_AM
+               int n;
                n         = ia32_get_irn_n_edges(in);
                is_cand   = (n == 1) ? 1 : is_cand;  /* load with more than one user: no AM */
+#endif
 
                load  = get_Proj_pred(in);
                other = left;
@@ -1064,15 +1026,24 @@ static INLINE void try_add_to_sched(ir_node *irn, ir_node *res) {
  * @param irn  The irn to be removed from schedule
  */
 static INLINE void try_remove_from_sched(ir_node *irn) {
+       int i, arity;
+
        if (sched_is_scheduled(irn)) {
                if (get_irn_mode(irn) == mode_T) {
                        const ir_edge_t *edge;
                        foreach_out_edge(irn, edge) {
                                ir_node *proj = get_edge_src_irn(edge);
-                               if (sched_is_scheduled(proj))
+                               if (sched_is_scheduled(proj)) {
+                                       set_irn_n(proj, 0, new_Bad());
                                        sched_remove(proj);
+                               }
                        }
                }
+
+               arity = get_irn_arity(irn);
+               for(i = 0; i < arity; ++i) {
+                       set_irn_n(irn, i, new_Bad());
+               }
                sched_remove(irn);
        }
 }
@@ -1520,7 +1491,6 @@ static void optimize_lea(ir_node *irn, void *env) {
        }
 }
 
-
 /**
  * Checks for address mode patterns and performs the
  * necessary transformations.
@@ -1530,8 +1500,7 @@ static void optimize_am(ir_node *irn, void *env) {
        ia32_am_opt_env_t *am_opt_env = env;
        ia32_code_gen_t   *cg         = am_opt_env->cg;
        heights_t         *h          = am_opt_env->h;
-       ir_node           *block, *noreg_gp, *noreg_fp;
-       ir_node           *left, *right;
+       ir_node           *block, *left, *right;
        ir_node           *store, *load, *mem_proj;
        ir_node           *succ, *addr_b, *addr_i;
        int               check_am_src          = 0;
@@ -1541,9 +1510,7 @@ static void optimize_am(ir_node *irn, void *env) {
        if (! is_ia32_irn(irn) || is_ia32_Ld(irn) || is_ia32_St(irn) || is_ia32_Store8Bit(irn))
                return;
 
-       block    = get_nodes_block(irn);
-       noreg_gp = ia32_new_NoReg_gp(cg);
-       noreg_fp = ia32_new_NoReg_fp(cg);
+       block = get_nodes_block(irn);
 
        DBG((mod, LEVEL_1, "checking for AM\n"));
 
@@ -1670,12 +1637,12 @@ static void optimize_am(ir_node *irn, void *env) {
                                        if (get_irn_arity(irn) == 5) {
                                                /* binary AMop */
                                                set_irn_n(irn, 4, get_irn_n(load, 2));
-                                               set_irn_n(irn, 2, noreg_gp);
+                                               set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2));
                                        }
                                        else {
                                                /* unary AMop */
                                                set_irn_n(irn, 3, get_irn_n(load, 2));
-                                               set_irn_n(irn, 2, noreg_gp);
+                                               set_irn_n(irn, 2, ia32_get_admissible_noreg(cg, irn, 2));
                                        }
 
                                        /* connect the memory Proj of the Store to the op */
@@ -1768,7 +1735,7 @@ static void optimize_am(ir_node *irn, void *env) {
                        set_ia32_pncode(irn, get_inversed_pnc(get_ia32_pncode(irn)));
 
                        /* disconnect from Load */
-                       set_irn_n(irn, 3, noreg_gp);
+                       set_irn_n(irn, 3, ia32_get_admissible_noreg(cg, irn, 3));
 
                        DBG_OPT_AM_S(right, irn);