common.opt: Add description of the new -fgcse-after-reload flag.

2004-03-03  Mostafa Hagog  <mustafa@il.ibm.com>

        * common.opt: Add description of the new -fgcse-after-reload flag.

        * flags.h (flag_gcse_after_reload): Declaration of global variable.

        * gcse.c (reg_used_on_edge ,reg_set_between_after_reload_p,
        reg_used_between_after_reload_p, rtx get_avail_load_store_reg,
        is_jump_table_basic_block, bb_has_well_behaved_predecessors,
        get_bb_avail_insn, hash_scan_set_after_reload,
        compute_hash_table_after_reload, eliminate_partially_redundant_loads,
        gcse_after_reload, get_bb_avail_insn): New functions to implement
        gcse-after-reload.
        (gcse_after_reload_main): New function, the main entry point to
        gcse-after-reload.

        * rtl.h (gcse_after_reload_main): Declaration of the new function.

        * opts.c (common_handle_option): Handle the -fgcse-after-reload flag.

        * toplev.c (flag_gcse_after_reload): Initialization.

        * passes.c (rest_of_handl_gcse2): Call gcse_after_reload_main.

        * params.def (PARAM_GCSE_AFTER_RELOAD_PARTIAL_FRACTION,
        PARAM_GCSE_AFTER_RELOAD_CRITICAL_FRACTION): New parameters for tuning
        the gcse after reload optimization.

        * params.h (GCSE_AFTER_RELOAD_PARTIAL_FRACTION,
        GCSE_AFTER_RELOAD_CRITICAL_FRACTION): Two macros to access the tuning
        parameters.

        * doc/invoke.texi: Documentation for the new flag gcse-after-reload.

From-SVN: r78842

gcc/ChangeLog

+2004-03-03  Mostafa Hagog  <mustafa@il.ibm.com>
+
+	* common.opt: Add description of the new -fgcse-after-reload flag.
+
+	* flags.h (flag_gcse_after_reload): Declaration of global variable.
+
+	* gcse.c (reg_used_on_edge ,reg_set_between_after_reload_p,
+	reg_used_between_after_reload_p, rtx get_avail_load_store_reg,
+	is_jump_table_basic_block, bb_has_well_behaved_predecessors,
+	get_bb_avail_insn, hash_scan_set_after_reload,
+	compute_hash_table_after_reload, eliminate_partially_redundant_loads,
+	gcse_after_reload, get_bb_avail_insn): New functions to implement
+	gcse-after-reload.
+	(gcse_after_reload_main): New function, the main entry point to
+	gcse-after-reload.
+
+	* rtl.h (gcse_after_reload_main): Declaration of the new function.
+
+	* opts.c (common_handle_option): Handle the -fgcse-after-reload flag.
+
+	* toplev.c (flag_gcse_after_reload): Initialization.
+
+	* passes.c (rest_of_handl_gcse2): Call gcse_after_reload_main.
+
+	* params.def (PARAM_GCSE_AFTER_RELOAD_PARTIAL_FRACTION,
+	PARAM_GCSE_AFTER_RELOAD_CRITICAL_FRACTION): New parameters for tuning
+	the gcse after reload optimization.
+
+	* params.h (GCSE_AFTER_RELOAD_PARTIAL_FRACTION,
+	GCSE_AFTER_RELOAD_CRITICAL_FRACTION): Two macros to access the tuning
+	parameters.
+
+	* doc/invoke.texi: Documentation for the new flag gcse-after-reload.
+
 2004-03-03  Nicolas Pitre <nico@cam.org>
 
 	* config/arm/ieee754-df.S (muldf3, divdf3): Fix denormalization of

gcc/common.opt

@@ -371,7 +371,13 @@ Perform store motion after global common subexpression elimination
 
 fgcse-las
 Common
-Perform redundant load after store elimination in global common subexpression elimination
+Perform redundant load after store elimination in global common subexpression
+elimination
+
+fgcse-after-reload
+Common
+Perform global common subexpression elimination after register allocation
+has finished.
 
 fguess-branch-probability
 Common

gcc/doc/invoke.texi

@@ -270,8 +270,8 @@ in the following sections.
 -fdelayed-branch  -fdelete-null-pointer-checks @gol
 -fexpensive-optimizations  -ffast-math  -ffloat-store @gol
 -fforce-addr  -fforce-mem  -ffunction-sections @gol
--fgcse  -fgcse-lm  -fgcse-sm  -fgcse-las  -floop-optimize @gol
--fcrossjumping  -fif-conversion  -fif-conversion2 @gol
+-fgcse  -fgcse-lm  -fgcse-sm  -fgcse-las  -fgcse-after-reload @gol
+-floop-optimize -fcrossjumping  -fif-conversion  -fif-conversion2 @gol
 -finline-functions  -finline-limit=@var{n}  -fkeep-inline-functions @gol
 -fkeep-static-consts  -fmerge-constants  -fmerge-all-constants @gol
 -fmove-all-movables  -fnew-ra  -fno-branch-count-reg @gol
@@ -3646,7 +3646,8 @@ invoking @option{-O2} on programs that use computed gotos.
 @opindex O3
 Optimize yet more.  @option{-O3} turns on all optimizations specified by
 @option{-O2} and also turns on the @option{-finline-functions},
-@option{-fweb} and @option{-frename-registers} options.
+@option{-fweb}, @option{-frename-registers}
+and @option{-fgcse-after-reload} options.
 
 @item -O0
 @opindex O0
@@ -3957,6 +3958,12 @@ same memory location (both partial and full redundancies).
 
 Enabled by default when gcse is enabled.
 
+@item -fgcse-after-reload
+@opindex fgcse-after-reload
+When @option{-fgcse-after-reload} is enabled, a redundant load elimination
+pass is performed after reload. The purpose of this pass is to cleanup
+redundant spilling.
+
 @item -floop-optimize
 @opindex floop-optimize
 Perform loop optimizations: move constant expressions out of loops, simplify

gcc/flags.h

@@ -672,6 +672,11 @@ extern int flag_gcse_sm;
 
 extern int flag_gcse_las;
 
+/* Nonzero if we want to perform global redundancy elimination after
+   register allocation.  */
+
+extern int flag_gcse_after_reload;
+
 /* Nonzero if value histograms should be used to optimize code.  */
 extern int flag_value_profile_transformations;
 

gcc/opts.c

@@ -574,6 +574,7 @@ decode_options (unsigned int argc, const char **argv)
       flag_rename_registers = 1;
       flag_unswitch_loops = 1;
       flag_web = 1;
+      flag_gcse_after_reload = 1;
     }
 
   if (optimize < 2 || optimize_size)
@@ -1035,6 +1036,10 @@ common_handle_option (size_t scode, const char *arg,
       flag_gcse_sm = value;
       break;
 
+    case OPT_fgcse_after_reload:
+      flag_gcse_after_reload = value;
+      break;
+
     case OPT_fgcse_las:
       flag_gcse_las = value;
       break;

gcc/params.def

@@ -131,7 +131,25 @@ DEFPARAM(PARAM_MAX_GCSE_PASSES,
 	"max-gcse-passes",
 	"The maximum number of passes to make when doing GCSE",
 	1)
-
+/* This is the threshold ratio when to perform partial redundancy
+   elimination after reload. We perform partial redundancy elimination
+   when the following holds:
+   (Redundant load execution count)
+   ------------------------------- >= GCSE_AFTER_RELOAD_PARTIAL_FRACTION
+   (Added loads execution count)					  */
+DEFPARAM(PARAM_GCSE_AFTER_RELOAD_PARTIAL_FRACTION,
+	"gcse-after-reload-partial-fraction",
+	"The threshold ratio for performing partial redundancy elimination \
+         after reload.",
+        3)
+/* This is the threshold ratio of the critical edges execution count compared to
+   the redundant loads execution count that permits performing the load
+   redundancy elimination in gcse after reload.  */
+DEFPARAM(PARAM_GCSE_AFTER_RELOAD_CRITICAL_FRACTION,
+	"gcse-after-reload-critical-fraction",
+	"The threshold ratio of critical edges execution count that permit \
+         performing redundancy elimination after reload.",
+        10)
 /* This parameter limits the number of insns in a loop that will be unrolled,
    and by how much the loop is unrolled.
    

gcc/params.h

@@ -104,6 +104,10 @@ typedef enum compiler_param
   ((size_t) PARAM_VALUE (PARAM_MAX_GCSE_MEMORY))
 #define MAX_GCSE_PASSES \
   PARAM_VALUE (PARAM_MAX_GCSE_PASSES)
+#define GCSE_AFTER_RELOAD_PARTIAL_FRACTION \
+  PARAM_VALUE (PARAM_GCSE_AFTER_RELOAD_PARTIAL_FRACTION)
+#define GCSE_AFTER_RELOAD_CRITICAL_FRACTION \
+  PARAM_VALUE (PARAM_GCSE_AFTER_RELOAD_CRITICAL_FRACTION)
 #define MAX_UNROLLED_INSNS \
   PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS)
 #endif /* ! GCC_PARAMS_H */

gcc/passes.c

@@ -158,6 +158,7 @@ enum dump_file_index
   DFI_lreg,
   DFI_greg,
   DFI_postreload,
+  DFI_gcse2,
   DFI_flow2,
   DFI_peephole2,
   DFI_ce3,
@@ -178,7 +179,7 @@ enum dump_file_index
    Remaining -d letters:
 
 	"   e        m   q         "
-	"         JK   O Q     WXY "
+	"          K   O Q     WXY "
 */
 
 static struct dump_file_info dump_file_tbl[DFI_MAX] =
@@ -210,6 +211,7 @@ static struct dump_file_info dump_file_tbl[DFI_MAX] =
   { "lreg",	'l', 1, 0, 0 },
   { "greg",	'g', 1, 0, 0 },
   { "postreload", 'o', 1, 0, 0 },
+  { "gcse2",    'J', 0, 0, 0 },
   { "flow2",	'w', 1, 0, 0 },
   { "peephole2", 'z', 1, 0, 0 },
   { "ce3",	'E', 1, 0, 0 },
@@ -788,6 +790,23 @@ rest_of_handle_sched2 (tree decl, rtx insns)
 }
 #endif
 
+static void
+rest_of_handle_gcse2 (tree decl, rtx insns)
+{
+  open_dump_file (DFI_gcse2, decl);
+
+  gcse_after_reload_main (insns, dump_file);
+  rebuild_jump_labels (insns);
+  delete_trivially_dead_insns (insns, max_reg_num ());
+  close_dump_file (DFI_gcse2, print_rtl_with_bb, insns);
+
+  ggc_collect ();
+
+#ifdef ENABLE_CHECKING
+  verify_flow_info ();
+#endif
+}
+
 /* Register allocation pre-pass, to reduce number of moves necessary
    for two-address machines.  */
 static void
@@ -1842,6 +1861,9 @@ rest_of_compilation (tree decl)
 
   close_dump_file (DFI_postreload, print_rtl_with_bb, insns);
 
+  if (optimize > 0 && flag_gcse_after_reload)
+    rest_of_handle_gcse2 (decl, insns);
+
   /* Re-create the death notes which were deleted during reload.  */
   timevar_push (TV_FLOW2);
   open_dump_file (DFI_flow2, decl);

gcc/rtl.h

@@ -2289,6 +2289,7 @@ extern rtx fis_get_condition (rtx);
 #ifdef BUFSIZ
 extern int gcse_main (rtx, FILE *);
 extern int bypass_jumps (FILE *);
+extern void gcse_after_reload_main (rtx, FILE *);
 #endif
 
 /* In global.c */

gcc/toplev.c

@@ -526,6 +526,9 @@ int flag_gcse_sm = 1;
 
 int flag_gcse_las = 1;
 
+/* Nonzero means perform global cse after register allocation.  */
+int flag_gcse_after_reload = 0;
+
 /* Perform target register optimization before prologue / epilogue
    threading.  */
 
@@ -915,6 +918,7 @@ static const lang_independent_options f_options[] =
   {"gcse-lm", &flag_gcse_lm, 1 },
   {"gcse-sm", &flag_gcse_sm, 1 },
   {"gcse-las", &flag_gcse_las, 1 },
+  {"gcse-after-reload", &flag_gcse_after_reload, 1},
   {"branch-target-load-optimize", &flag_branch_target_load_optimize, 1 },
   {"branch-target-load-optimize2", &flag_branch_target_load_optimize2, 1 },
   {"btr-bb-exclusive", &flag_btr_bb_exclusive, 1 },