[RFC] Run fast RTL DCE pass at -O0

[Eric Botcazou <ebotcazou at adacore dot com>]

Hi,

flow.c used to run (integrated) DCE at -O0, this was ditched with the switch 
to full DF.  As a result, we measured significant code size increases at -O0, 
for example 10 MB on a big Ada application (85 MB).

This patch adds a fast RTL DCE pass at -O0; in our 4.3-based compiler, this 
entirely fixed the code size regression with no compilation time increases.

Tested (GCC, GDB) on i586-suse-linux.  Thoughts?


2009-04-18  Eric Botcazou  <ebotcazou@adacore.com>

	* tree-pass.h (pass_fast_rtl_dce_no_opt): Declare.
	* dce.c (insn_is_nop): New function.
	(delete_unmarked_insns): Do not delete NOPs at -O0.
	(gate_fast_dce_no_opt): New gate function.
	(pass_fast_rtl_dce_no_opt): New pass.
	* passes.c (init_optimization_passes): Register it.


-- 
Eric Botcazou

Index: tree-pass.h
===================================================================
--- tree-pass.h	(revision 146259)
+++ tree-pass.h	(working copy)
@@ -431,6 +431,7 @@ extern struct rtl_opt_pass pass_jump2;
 extern struct rtl_opt_pass pass_lower_subreg;
 extern struct rtl_opt_pass pass_cse;
 extern struct rtl_opt_pass pass_fast_rtl_dce;
+extern struct rtl_opt_pass pass_fast_rtl_dce_no_opt;
 extern struct rtl_opt_pass pass_ud_rtl_dce;
 extern struct rtl_opt_pass pass_rtl_dce;
 extern struct rtl_opt_pass pass_rtl_dse1;
Index: passes.c
===================================================================
--- passes.c	(revision 146259)
+++ passes.c	(working copy)
@@ -766,6 +766,7 @@ init_optimization_passes (void)
       NEXT_PASS (pass_split_all_insns);
       NEXT_PASS (pass_lower_subreg2);
       NEXT_PASS (pass_df_initialize_no_opt);
+      NEXT_PASS (pass_fast_rtl_dce_no_opt);
       NEXT_PASS (pass_stack_ptr_mod);
       NEXT_PASS (pass_mode_switching);
       NEXT_PASS (pass_see);
Index: dce.c
===================================================================
--- dce.c	(revision 146259)
+++ dce.c	(working copy)
@@ -501,6 +501,36 @@ delete_corresponding_reg_eq_notes (rtx i
 }
 
 
+/* Return true if INSN contains only a formal nop pattern.
+
+   NOTE: This function is coded for speed and thus not meant to be
+   bullet-proof.  Its failure to recognize a nop pattern shouldn't
+   be allowed to generate wrong code.  */
+
+static bool
+insn_is_nop (const_rtx insn)
+{
+  rtx pat;
+
+  if (!INSN_P (insn))
+    return false;
+
+  pat = PATTERN (insn);
+
+  /* This is the pattern used by almost all back-ends.  */
+  if (pat == const0_rtx)
+    return true;
+
+  /* This is the pattern used by the SPU back-end.  */
+  if (GET_CODE (pat) == UNSPEC_VOLATILE
+      && XVECLEN (pat, 0) == 1
+      && XVECEXP (pat, 0, 0) == const0_rtx)
+    return true;
+
+  return false;
+}
+
+
 /* Delete every instruction that hasn't been marked.  */
 
 static void
@@ -522,6 +552,10 @@ delete_unmarked_insns (void)
 	  else if (marked_insn_p (insn))
 	    continue;
 
+	  /* Do not delete NOPs created to carry source line info at -O0.  */
+	  if (!optimize && insn_is_nop (insn))
+	    continue;
+
 	  /* Beware that reaching a dbg counter limit here can result
 	     in miscompiled file.  This occurs when a group of insns
 	     must be deleted together, typically because the kept insn
@@ -1115,8 +1149,7 @@ run_fast_dce (void)
 static bool
 gate_fast_dce (void)
 {
-  return optimize > 0 && flag_dce
-    && dbg_cnt (dce_fast);
+  return optimize > 0 && flag_dce && dbg_cnt (dce_fast);
 }
 
 struct rtl_opt_pass pass_fast_rtl_dce =
@@ -1140,6 +1173,35 @@ struct rtl_opt_pass pass_fast_rtl_dce =
  }
 };
 
+
+static bool
+gate_fast_dce_no_opt (void)
+{
+  return optimize == 0 && flag_dce && dbg_cnt (dce_fast);
+}
+
+struct rtl_opt_pass pass_fast_rtl_dce_no_opt =
+{
+ {
+  RTL_PASS,
+  "dce",                                /* name */
+  gate_fast_dce_no_opt,                 /* gate */
+  rest_of_handle_fast_dce,              /* execute */
+  NULL,                                 /* sub */
+  NULL,                                 /* next */
+  0,                                    /* static_pass_number */
+  TV_DCE,                               /* tv_id */
+  0,                                    /* properties_required */
+  0,                                    /* properties_provided */
+  0,                                    /* properties_destroyed */
+  0,                                    /* todo_flags_start */
+  TODO_dump_func |
+  TODO_df_finish | TODO_verify_rtl_sharing |
+  TODO_ggc_collect                      /* todo_flags_finish */
+ }
+};
+
+
 struct rtl_opt_pass pass_fast_rtl_byte_dce =
 {
  {