- allow implicit cast from two-complements signed/unsigned to be killed
authorMichael Beck <beck@ipd.info.uni-karlsruhe.de>
Fri, 12 Jan 2007 18:07:54 +0000 (18:07 +0000)
committerMichael Beck <beck@ipd.info.uni-karlsruhe.de>
Fri, 12 Jan 2007 18:07:54 +0000 (18:07 +0000)
- updated for new memory disambiguator

[r8520]

ir/opt/ldstopt.c

index cee7142..de17fd0 100644 (file)
@@ -431,6 +431,23 @@ static void reduce_adr_usage(ir_node *ptr) {
        }
 }  /* reduce_adr_usage */
 
+/**
+ * Check, if an already existing value of mode old_mode can be converted
+ * into the needed one new_mode without loss.
+ */
+static int can_use_stored_value(ir_mode *old_mode, ir_mode *new_mode) {
+       if (old_mode == new_mode)
+               return 1;
+
+       /* if both modes are two-complement ones, we can always convert the
+          Stored value into the needed one. */
+       if (get_mode_size_bits(old_mode) >= get_mode_size_bits(new_mode) &&
+                 get_mode_arithmetic(old_mode) == irma_twos_complement &&
+                 get_mode_arithmetic(new_mode) == irma_twos_complement)
+               return 1;
+       return 0;
+}  /* can_use_stored_value */
+
 /**
  * Follow the memory chain as long as there are only Loads
  * and alias free Stores and try to replace current Load or Store
@@ -459,7 +476,7 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                 * is possible a = *(ir_type1 *)p; b = *(ir_type2 *)p ...
                 */
                if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
-                       get_irn_mode(get_Store_value(pred)) == load_mode) {
+                   can_use_stored_value(get_irn_mode(get_Store_value(pred)), load_mode)) {
                        /*
                         * a Load immediately after a Store -- a read after write.
                         * We may remove the Load, if both Load & Store does not have an exception handler
@@ -477,6 +494,12 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                                ir_node *value = get_Store_value(pred);
 
                                DBG_OPT_RAW(load, value);
+
+                               /* add an convert if needed */
+                               if (get_irn_mode(get_Store_value(pred)) != load_mode) {
+                                       value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
+                               }
+
                                if (info->projs[pn_Load_M])
                                        exchange(info->projs[pn_Load_M], mem);
 
@@ -494,7 +517,7 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                                return res | DF_CHANGED;
                        }
                } else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
-                          get_Load_mode(pred) == load_mode) {
+                          can_use_stored_value(get_Load_mode(pred), load_mode)) {
                        /*
                         * a Load after a Load -- a read after read.
                         * We may remove the second Load, if it does not have an exception handler
@@ -508,6 +531,13 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                                DBG_OPT_RAR(load, pred);
 
                                if (pred_info->projs[pn_Load_res]) {
+                                       ir_node *value = pred_info->projs[pn_Load_res];
+
+                                       /* add an convert if needed */
+                                       if (get_Load_mode(pred) != load_mode) {
+                                               value = new_r_Conv(current_ir_graph, get_nodes_block(load), value, load_mode);
+                                       }
+
                                        /* we need a data proj from the previous load for this optimization */
                                        if (info->projs[pn_Load_res])
                                                exchange(info->projs[pn_Load_res], pred_info->projs[pn_Load_res]);
@@ -545,7 +575,7 @@ static unsigned follow_Mem_chain(ir_node *load, ir_node *curr) {
                                current_ir_graph,
                                get_Store_ptr(pred),
                                get_irn_mode(get_Store_value(pred)),
-                               ptr, load_mode, opt_non_opt);
+                               ptr, load_mode);
                        /* if the might be an alias, we cannot pass this Store */
                        if (rel != no_alias)
                                break;
@@ -779,6 +809,15 @@ static unsigned optimize_load(ir_node *load)
        return res;
 }  /* optimize_load */
 
+/**
+ * Check whether a value of mode new_mode would completely overwrite a value
+ * of mode old_mode in memory.
+ */
+static int is_completely_overwritten(ir_mode *old_mode, ir_mode *new_mode)
+{
+       return get_mode_size_bits(new_mode) >= get_mode_size_bits(old_mode);
+}  /* is_completely_overwritten */
+
 /**
  * follow the memory chain as long as there are only Loads and alias free Stores.
  *
@@ -802,9 +841,12 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
                 * if the pointers are identical, they refer to the same object.
                 * This is only true in strong typed languages, not is C were the following
                 * is possible *(ir_type1 *)p = a; *(ir_type2 *)p = b ...
+                * However, if the mode that is written have a bigger  or equal size the the old
+                * one, the old value is completely overwritten and can be killed ...
                 */
                if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
-                   get_nodes_block(pred) == block && get_irn_mode(get_Store_value(pred)) == mode) {
+                   get_nodes_block(pred) == block &&
+                   is_completely_overwritten(get_irn_mode(get_Store_value(pred)), mode)) {
                        /*
                         * a Store after a Store in the same block -- a write after write.
                         * We may remove the first Store, if it does not have an exception handler.
@@ -839,7 +881,7 @@ static unsigned follow_Mem_chain_for_Store(ir_node *store, ir_node *curr) {
                                current_ir_graph,
                                get_Store_ptr(pred),
                                get_irn_mode(get_Store_value(pred)),
-                               ptr, mode, opt_non_opt);
+                               ptr, mode);
                        /* if the might be an alias, we cannot pass this Store */
                        if (rel != no_alias)
                                break;
@@ -1098,19 +1140,19 @@ static void do_load_store_optimize(ir_node *n, void *env) {
 
        switch (get_irn_opcode(n)) {
 
-  case iro_Load:
-         wenv->changes |= optimize_load(n);
-         break;
+       case iro_Load:
+               wenv->changes |= optimize_load(n);
+               break;
 
-  case iro_Store:
-         wenv->changes |= optimize_store(n);
-         break;
+       case iro_Store:
+               wenv->changes |= optimize_store(n);
+               break;
 
-  case iro_Phi:
-         wenv->changes |= optimize_phi(n, wenv);
+       case iro_Phi:
+               wenv->changes |= optimize_phi(n, wenv);
 
-  default:
-         ;
+       default:
+               ;
        }
 }  /* do_load_store_optimize */