public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] Improve SCCVN with devirtualization to pure/const calls
@ 2016-09-22 13:03 Richard Biener
  0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2016-09-22 13:03 UTC (permalink / raw)
  To: gcc-patches; +Cc: dje.gcc


Not because it might be terribly important but the recent change
triggered a latent(?) issue on AIX which broke bootstrap.  Let's
try if another change mitigates that.

The change basically allows devirt-to-const/pure to be taken
into account during the value-numbering rather than only in
followup passes (after we committed the devirtualization).

Bootstrap / regtest running on x86_64-unknown-linux-gnu.

Richard.

2016-09-22  Richard Biener  <rguenther@suse.de>

	* tree-ssa-sccvn.c (visit_reference_op_call): Value number
	virtual definition to virtual use if the call devirtualizes
	to a const or pure function.
	(visit_use): Also visit calls we can devirtualize to a
	const or pure function.

Index: gcc/tree-ssa-sccvn.c
===================================================================
*** gcc/tree-ssa-sccvn.c	(revision 240353)
--- gcc/tree-ssa-sccvn.c	(working copy)
*************** visit_reference_op_call (tree lhs, gcall
*** 3486,3492 ****
        vn_reference_t vr2;
        vn_reference_s **slot;
        if (vdef)
! 	changed |= set_ssa_val_to (vdef, vdef);
        if (lhs)
  	changed |= set_ssa_val_to (lhs, lhs);
        vr2 = current_info->references_pool->allocate ();
--- 3593,3615 ----
        vn_reference_t vr2;
        vn_reference_s **slot;
        if (vdef)
! 	{
! 	  /* If we value numbered an indirect functions function to
! 	     one not clobbering memory value number its VDEF to its
! 	     VUSE.  */
! 	  tree fn = gimple_call_fn (stmt);
! 	  tree vdef_val = vdef;
! 	  if (fn && TREE_CODE (fn) == SSA_NAME)
! 	    {
! 	      fn = SSA_VAL (fn);
! 	      if (TREE_CODE (fn) == ADDR_EXPR
! 		  && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL
! 		  && (flags_from_decl_or_type (TREE_OPERAND (fn, 0))
! 		      & (ECF_CONST | ECF_PURE)))
! 		vdef_val = vuse_ssa_val (gimple_vuse (stmt));
! 	    }
! 	  changed |= set_ssa_val_to (vdef, vdef_val);
! 	}
        if (lhs)
  	changed |= set_ssa_val_to (lhs, lhs);
        vr2 = current_info->references_pool->allocate ();
*************** visit_use (tree use)
*** 3913,3923 ****
  	    }
  	}
  
        if (!gimple_call_internal_p (call_stmt)
  	  && (/* Calls to the same function with the same vuse
  		 and the same operands do not necessarily return the same
  		 value, unless they're pure or const.  */
! 	      gimple_call_flags (call_stmt) & (ECF_PURE | ECF_CONST)
  	      /* If calls have a vdef, subsequent calls won't have
  		 the same incoming vuse.  So, if 2 calls with vdef have the
  		 same vuse, we know they're not subsequent.
--- 4035,4056 ----
  	    }
  	}
  
+       /* Pick up flags from a devirtualization target.  */
+       tree fn = gimple_call_fn (stmt);
+       int extra_fnflags = 0;
+       if (fn && TREE_CODE (fn) == SSA_NAME)
+ 	{
+ 	  fn = SSA_VAL (fn);
+ 	  if (TREE_CODE (fn) == ADDR_EXPR
+ 	      && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL)
+ 	    extra_fnflags = flags_from_decl_or_type (TREE_OPERAND (fn, 0));
+ 	}
        if (!gimple_call_internal_p (call_stmt)
  	  && (/* Calls to the same function with the same vuse
  		 and the same operands do not necessarily return the same
  		 value, unless they're pure or const.  */
! 	      ((gimple_call_flags (call_stmt) | extra_fnflags)
! 	       & (ECF_PURE | ECF_CONST))
  	      /* If calls have a vdef, subsequent calls won't have
  		 the same incoming vuse.  So, if 2 calls with vdef have the
  		 same vuse, we know they're not subsequent.
Index: gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-56.c
===================================================================
*** gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-56.c	(revision 0)
--- gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-56.c	(working copy)
***************
*** 0 ****
--- 1,26 ----
+ /* { dg-do compile } */
+ /* { dg-options "-O -fdump-tree-fre1" } */
+ 
+ int x = 1;
+ int __attribute__((noinline,noclone)) fn ()
+ {
+   return x;
+ }
+ int (*f)();
+ int main ()
+ {
+   int res;
+   f = fn;
+   x = 0;
+   res = f ();
+   res += x;
+   if (res != 0)
+     __builtin_abort ();
+   return 0;
+ }
+ 
+ /* We should be able to optimize the load from x in main and thus the
+    addition.  */
+ 
+ /* { dg-final { scan-tree-dump-times "= x;" 1 "fre1" } } */
+ /* { dg-final { scan-tree-dump-times " \\\+ " 0 "fre1" } } */

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2016-09-22 12:55 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-09-22 13:03 [PATCH] Improve SCCVN with devirtualization to pure/const calls Richard Biener

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).