Introduce flag_cunroll_grow_size for cunroll
authorguojiufu <guojiufu@linux.ibm.com>
Thu, 28 May 2020 05:42:23 +0000 (13:42 +0800)
committerguojiufu <guojiufu@linux.ibm.com>
Sun, 7 Jun 2020 09:35:01 +0000 (17:35 +0800)
Currently GIMPLE complete unroller(cunroll) is checking
flag_unroll_loops and flag_peel_loops to see if allow size growth.
Beside affects curnoll, flag_unroll_loops also controls RTL unroler.
To have more freedom to control cunroll and RTL unroller, this patch
introduces flag_cunroll_grow_size.  With this patch, we can control
cunroll and RTL unroller indepently.

gcc/ChangeLog
2020-06-07  Jiufu Guo  <guojiufu@linux.ibm.com>

* common.opt (flag_cunroll_grow_size): New flag.
* toplev.c (process_options): Set flag_cunroll_grow_size.
* tree-ssa-loop-ivcanon.c (pass_complete_unroll::execute):
Use flag_cunroll_grow_size.

gcc/common.opt
gcc/toplev.c
gcc/tree-ssa-loop-ivcanon.c

index 1b770bc..df8af36 100644 (file)
@@ -2863,6 +2863,10 @@ funroll-all-loops
 Common Report Var(flag_unroll_all_loops) Optimization
 Perform loop unrolling for all loops.
 
+funroll-completely-grow-size
+Undocumented Var(flag_cunroll_grow_size) Init(2) Optimization
+; Internal undocumented flag, allow size growth during complete unrolling
+
 ; Nonzero means that loop optimizer may assume that the induction variables
 ; that control loops do not overflow and that the loops with nontrivial
 ; exit condition are not infinite
index 1a75485..95eea63 100644 (file)
@@ -1482,6 +1482,11 @@ process_options (void)
   if (flag_unroll_all_loops)
     flag_unroll_loops = 1;
 
+  /* Allow cunroll to grow size accordingly.  */
+  if (flag_cunroll_grow_size == AUTODETECT_VALUE)
+    flag_cunroll_grow_size
+      = flag_unroll_loops || flag_peel_loops || optimize >= 3;
+
   /* web and rename-registers help when run after loop unrolling.  */
   if (flag_web == AUTODETECT_VALUE)
     flag_web = flag_unroll_loops;
index 8ab6ab3..298ab21 100644 (file)
@@ -1603,9 +1603,8 @@ pass_complete_unroll::execute (function *fun)
      re-peeling the same loop multiple times.  */
   if (flag_peel_loops)
     peeled_loops = BITMAP_ALLOC (NULL);
-  unsigned int val = tree_unroll_loops_completely (flag_unroll_loops
-                                                  || flag_peel_loops
-                                                  || optimize >= 3, true);
+  unsigned int val = tree_unroll_loops_completely (flag_cunroll_grow_size, 
+                                                  true);
   if (peeled_loops)
     {
       BITMAP_FREE (peeled_loops);