diff mbox series

[2/2] rs6000: Turn on -frtl-unroll-loops instead -funroll-loops at -O2

Message ID 20200525111040.8876-2-guojiufu@linux.ibm.com
State New
Headers show
Series [1/2] Seperate -funroll-loops for GIMPLE unroller and RTL unroller | expand

Commit Message

Jiufu Guo May 25, 2020, 11:10 a.m. UTC 
Previously, turning -funroll-loops on at -O2, which also turn on
GIMPLE cunroll fully.  While cunroll unrolls some complex loops.

This patch turn on -frtl-unroll-loops at -O2 only, and continue to
use previous tuned rs6000 heurisitics for small loops.  While this
patch does not turn on GIMPLE cunroll any more.  We may tune
cunroll in near future at -O2.

In this patch, it become simpler to check/set -fweb, -frename-register
and -munroll-only-small-loops.  Together with -frtl-unroll-loops, -fweb
is useful, then turn -fweb on;  and -frename-registers is no need to
be checked, because it is affected by -frtl-unroll-loops.


Bootstrap and regtest pass on powerpc64le, is this ok for trunk?
And backport to GCC10 together with the patch "Seperate -funroll-loops
 for GIMPLE unroller and RTL unroller"

Jiufu


gcc/ChangeLog
2020-05-25  Jiufu Guo   <guojiufu@cn.ibm.com>

	PR target/95018
	* common/config/rs6000/rs6000-common.c
	(rs6000_option_optimization_table)
	[OPT_LEVELS_2_PLUS_SPEED_ONLY]: Replace -funroll-loops
	with -frtl-unroll-loops.  Remove -munroll-only-small-loops
	and add -fweb.
	[OPT_LEVELS_ALL]: Remove turn off -frename-registers.
	* config/rs6000/rs6000.c (rs6000_option_override_internal):
	-funroll-loops overrides -munroll-only-small-loops and
	-frtl-unroll-loops.
---
 gcc/common/config/rs6000/rs6000-common.c | 11 +++--------
 gcc/config/rs6000/rs6000.c               | 21 ++++++++++-----------
 2 files changed, 13 insertions(+), 19 deletions(-)
diff mbox series

Patch

diff --git a/gcc/common/config/rs6000/rs6000-common.c b/gcc/common/config/rs6000/rs6000-common.c
index ee37b9dc90b..c7388edb867 100644
--- a/gcc/common/config/rs6000/rs6000-common.c
+++ b/gcc/common/config/rs6000/rs6000-common.c
@@ -34,14 +34,9 @@  static const struct default_options rs6000_option_optimization_table[] =
     { OPT_LEVELS_ALL, OPT_fsplit_wide_types_early, NULL, 1 },
     /* Enable -fsched-pressure for first pass instruction scheduling.  */
     { OPT_LEVELS_1_PLUS, OPT_fsched_pressure, NULL, 1 },
-    /* Enable -munroll-only-small-loops with -funroll-loops to unroll small
-       loops at -O2 and above by default.  */
-    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_funroll_loops, NULL, 1 },
-    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_munroll_only_small_loops, NULL, 1 },
-
-    /* -frename-registers leads to non-optimal codegen and performance
-       on rs6000, turn it off by default.  */
-    { OPT_LEVELS_ALL, OPT_frename_registers, NULL, 0 },
+    /* Enable -frtl-unroll-loops and -fweb at -O2 and above by default.  */
+    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_frtl_unroll_loops, NULL, 1 },
+    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_fweb, NULL, 1 },
 
     /* Double growth factor to counter reduced min jump length.  */
     { OPT_LEVELS_ALL, OPT__param_max_grow_copy_bb_insns_, NULL, 16 },
diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index 8435bc15d72..96620651a59 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -4557,17 +4557,16 @@  rs6000_option_override_internal (bool global_init_p)
 			   param_sched_pressure_algorithm,
 			   SCHED_PRESSURE_MODEL);
 
-      /* Explicit -funroll-loops turns -munroll-only-small-loops off, and
-	 turns -frename-registers on.  */
-      if ((global_options_set.x_flag_unroll_loops && flag_unroll_loops)
-	   || (global_options_set.x_flag_unroll_all_loops
-	       && flag_unroll_all_loops))
-	{
-	  if (!global_options_set.x_unroll_only_small_loops)
-	    unroll_only_small_loops = 0;
-	  if (!global_options_set.x_flag_rename_registers)
-	    flag_rename_registers = 1;
-	}
+      /* if -f[no-]unroll-loops is specified explicitly, turn [off/]on
+	 -frtl-unroll-loops.  */
+      if (global_options_set.x_flag_unroll_loops
+	  && !global_options_set.x_flag_rtl_unroll_loops)
+	flag_rtl_unroll_loops = flag_unroll_loops;
+	
+      /* If flag_unroll_loops is effect, not _only_ small loops, but
+	 large loops are unrolled if possible.  */
+      if (!global_options_set.x_unroll_only_small_loops)
+	unroll_only_small_loops = flag_unroll_loops ? 0 : 1;
 
       /* If using typedef char *va_list, signal that
 	 __builtin_va_start (&ap, 0) can be optimized to