synchronized with edgfe, using new inline heuristic

[cparser] / driver / firm_opt.c

diff --git a/driver/firm_opt.c b/driver/firm_opt.c
index 803b2f8..95cae02 100644 (file)
--- a/driver/firm_opt.c
+++ b/driver/firm_opt.c
@@ -20,6 +20,10 @@
 #include "firm_cmdline.h"
 #include "firm_timing.h"
 
+#ifdef _WIN32
+#define snprintf _snprintf
+#endif
+
 #if defined(_DEBUG) || defined(FIRM_DEBUG)
 #define DBG(x)  dbg_printf x
 #else
@@ -327,13 +331,16 @@ static void do_firm_optimizations(const char *input_filename, int firm_const_exi
     CHECK_ALL(firm_opt.check_all);
   }
 
-  if (firm_opt.tail_rec) {
-    timer_push(TV_TAIL_REC);
-      opt_tail_recursion();
+  /* first step: kill dead code */
+  for (i = 0; i < get_irp_n_irgs(); i++) {
+    irg = current_ir_graph = get_irp_irg(i);
+    /* Confirm construction currently can only handle blocks with only one control
+       flow predecessor. Calling optimize_cf here removes Bad predecessors and help
+       the optimization of switch constructs. */
+    timer_push(TV_CF_OPT);
+      optimize_graph_df(irg);
+      optimize_cf(irg);
     timer_pop();
-
-    DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "tail_rec");
-    CHECK_ALL(firm_opt.check_all);
   }
 
   if (firm_opt.func_calls) {
@@ -344,18 +351,12 @@ static void do_firm_optimizations(const char *input_filename, int firm_const_exi
     CHECK_ALL(firm_opt.check_all);
   }
 
-  if (firm_opt.do_inline) {
-    timer_push(TV_INLINE);
-      inline_leave_functions(500, 80, 30, FALSE);
-    timer_pop();
-    DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "inl");
-    CHECK_ALL(firm_opt.check_all);
-  }
+  /* do lowering on the const code irg */
+  lower_const_code();
 
   for (i = 0; i < get_irp_n_irgs(); i++) {
     irg = current_ir_graph = get_irp_irg(i);
 
-
 #ifdef FIRM_EXT_GRS
   /* If SIMD optimization is on, make sure we have only 1 return */
   if (firm_ext_grs.create_pattern || firm_ext_grs.simd_opt)
@@ -450,14 +451,6 @@ static void do_firm_optimizations(const char *input_filename, int firm_const_exi
       irg_verify(irg, VRFY_ENFORCE_SSA);
     }
 
-    if (firm_opt.loop_unrolling) {
-      timer_push(TV_LOOP_UNROLL);
-        optimize_loop_unrolling(irg);
-      timer_pop();
-      DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "loop");
-      CHECK_ONE(firm_opt.check_all, irg);
-       }
-
     if (firm_opt.load_store) {
       timer_push(TV_LOAD_STORE);
         optimize_load_store(irg);
@@ -521,7 +514,7 @@ static void do_firm_optimizations(const char *input_filename, int firm_const_exi
     }
 
     timer_push(TV_OSR);
-      opt_osr(current_ir_graph, osr_flag_default /*| osr_flag_ignore_x86_shift*/);
+      opt_osr(current_ir_graph, osr_flag_default | osr_flag_keep_reg_pressure | osr_flag_ignore_x86_shift);
     timer_pop();
     DUMP_ONE_C(firm_dump.ir_graph && firm_dump.all_phases, irg, "stred");
     CHECK_ONE(firm_opt.check_all, irg);
@@ -540,11 +533,24 @@ static void do_firm_optimizations(const char *input_filename, int firm_const_exi
     CHECK_ONE(firm_opt.check_all, irg);
   }
 
+  if (firm_opt.do_inline) {
+    inline_functions(50);
+    DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "inl");
+    CHECK_ALL(firm_opt.check_all);
+  }
   if (firm_opt.cloning) {
     proc_cloning((float)firm_opt.clone_threshold);
     DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "clone");
     CHECK_ALL(firm_opt.check_all);
   }
+  if (firm_opt.tail_rec) {
+    timer_push(TV_TAIL_REC);
+      opt_tail_recursion();
+    timer_pop();
+
+    DUMP_ALL_C(firm_dump.ir_graph && firm_dump.all_phases, "tail_rec");
+    CHECK_ALL(firm_opt.check_all);
+  }
 
   if (firm_dump.ir_graph) {
     /* recompute backedges for nicer dumps */
@@ -571,8 +577,10 @@ static void do_firm_optimizations(const char *input_filename, int firm_const_exi
   DUMP_ALL(firm_dump.ir_graph, "-opt");
 
   /* verify optimized graphs */
-  for (i = get_irp_n_irgs() - 1; i >= 0; --i)
-    irg_verify(get_irp_irg(i), VRFY_ENFORCE_SSA);
+  for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
+    ir_graph *irg = get_irp_irg(i);
+    irg_verify(irg, VRFY_ENFORCE_SSA);
+  }
 
   if (firm_dump.statistic & STAT_AFTER_OPT)
     stat_dump_snapshot(input_filename, "opt");
@@ -715,24 +723,14 @@ static int compute_type_size(ir_type *ty)
 }  /* compute_type_size */
 
 /**
- * layout all types of the Firm graph
+ * layout all non-frame types of the Firm graph
  */
 static void compute_type_sizes(void)
 {
   int i;
   ir_type *tp;
-  ir_graph *irg;
-
-  /* all frame types */
-  for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
-    irg = get_irp_irg(i);
-    /* do not optimize away variables in debug mode */
-    if (firm_opt.debug_mode == DBG_MODE_NONE)
-      opt_frame_irg(irg);
-    compute_type_size(get_irg_frame_type(irg));
-  }
 
-  /* all other types */
+  /* all non-frame other types */
   for (i = get_irp_n_types() - 1; i >= 0; --i) {
     tp = get_irp_type(i);
     compute_type_size(tp);
@@ -748,6 +746,24 @@ static void compute_type_sizes(void)
   }
 }  /* compute_type_sizes */
 
+/**
+ * layout all frame-types of the Firm graph
+ */
+static void compute_frame_type_sizes(void)
+{
+  int i;
+  ir_graph *irg;
+
+  /* all frame types */
+  for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
+    irg = get_irp_irg(i);
+    /* do not optimize away variables in debug mode */
+    if (firm_opt.debug_mode == DBG_MODE_NONE)
+      opt_frame_irg(irg);
+    compute_type_size(get_irg_frame_type(irg));
+  }
+}  /* compute_frame_type_sizes */
+
 /**
  * do Firm lowering
  *
@@ -783,7 +799,6 @@ static void do_firm_lowering(const char *input_filename)
     }
     timer_push(TV_DW_LOWER);
       lower_dw_ops(&init);
-      DUMP_ALL(firm_dump.ir_graph, "-dw");
     timer_pop();
   }
 
@@ -835,7 +850,6 @@ static void do_firm_lowering(const char *input_filename)
         CHECK_ONE(firm_opt.check_all, current_ir_graph);
       }
 
-//      set_opt_global_cse(0);
       timer_push(TV_LOAD_STORE);
         optimize_load_store(current_ir_graph);
       timer_pop();
@@ -947,6 +961,9 @@ void gen_firm_init(void)
   set_opt_precise_exc_context(firm_opt.precise_exc);
   set_opt_fragile_ops(firm_opt.fragile_ops);
 
+  /* Sel node cannot produce NULL pointers */
+  set_opt_sel_based_null_check_elim(1);
+
   /* dynamic dispatch works currently only if whole world scenarios */
   set_opt_dyn_meth_dispatch(0);
 
@@ -1045,6 +1062,12 @@ void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int firm
   /* computes the sizes of all types that are still not computed */
   compute_type_sizes();
 
+  /* lower copyb nodes */
+  for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
+    ir_graph *irg = get_irp_irg(i);
+    lower_CopyB(irg, 128, 4);
+  }
+
   if (firm_dump.statistic & STAT_BEFORE_OPT) {
     stat_dump_snapshot(input_filename, "noopt");
   }
@@ -1062,6 +1085,9 @@ void gen_firm_finish(FILE *out, const char *input_filename, int c_mode, int firm
   if (firm_opt.lower)
     do_firm_lowering(input_filename);
 
+  /* computes the sizes of all frame types */
+  compute_frame_type_sizes();
+
   /* set the phase to low */
   for (i = get_irp_n_irgs() - 1; i >= 0; --i)
     set_irg_phase_low(get_irp_irg(i));