[PATCH] Fix PR78847

[PATCH] Fix PR78847

[Richard Biener]

The following patch fixes a missed optimization caused by niter analysis
returning a complex pointer subtraction expression instead of a constant
for a loop with a pointer IV (it looks like pointer IVs are quite common
in libstdc++).

Bootstrapped and tested on x86_64-unknown-linux-gnu.

Not a regression (well, didn't try to find a release where we optimized
the testcase, but it needs c++14 library support anyway).  I'm also
not quite sure associate_trees handles pointer types correctly
(there is some pointer type handling in the associate case but
POINTER_PLUS_EXPR folding for example does not dispatch to associate).

I guess trying to add some more exhaustive testcases for GENERIC input
is required.

Thanks,
Richard.

2017-02-09  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/78847
	* fold-const.c (split_tree): Handle POINTER_PLUS_EXPR.

	* g++.dg/tree-ssa/pr78847.C: New testcase.

Index: gcc/fold-const.c
===================================================================
*** gcc/fold-const.c	(revision 245276)
--- gcc/fold-const.c	(working copy)
*************** split_tree (location_t loc, tree in, tre
*** 785,791 ****
  		  the value is not affected.  For reals, the value might be
  		  affected, so we can't.  */
  	       && ((code == PLUS_EXPR && TREE_CODE (in) == MINUS_EXPR)
! 		   || (code == MINUS_EXPR && TREE_CODE (in) == PLUS_EXPR))))
      {
        tree op0 = TREE_OPERAND (in, 0);
        tree op1 = TREE_OPERAND (in, 1);
--- 785,793 ----
  		  the value is not affected.  For reals, the value might be
  		  affected, so we can't.  */
  	       && ((code == PLUS_EXPR && TREE_CODE (in) == MINUS_EXPR)
! 		   || (code == MINUS_EXPR
! 		       && (TREE_CODE (in) == PLUS_EXPR
! 			   || TREE_CODE (in) == POINTER_PLUS_EXPR)))))
      {
        tree op0 = TREE_OPERAND (in, 0);
        tree op1 = TREE_OPERAND (in, 1);
Index: gcc/testsuite/g++.dg/tree-ssa/pr78847.C
===================================================================
*** gcc/testsuite/g++.dg/tree-ssa/pr78847.C	(nonexistent)
--- gcc/testsuite/g++.dg/tree-ssa/pr78847.C	(working copy)
***************
*** 0 ****
--- 1,26 ----
+ /* { dg-do compile } */
+ /* { dg-require-effective-target c++14 } */
+ /* { dg-options "-O3 -fdump-tree-ldist" } */
+ 
+ #include <stddef.h>
+ #include <cstring>
+ #include <experimental/string_view>
+ 
+ using string_view = std::experimental::string_view;
+ 
+ class Foo {
+     constexpr static size_t Length = 9;
+     char ascii_[Length];
+ public:
+     Foo();
+     string_view view() const {
+ 	return string_view(ascii_, Length);
+     }
+ };
+ 
+ void testWithLoopValue(const Foo foo, size_t ptr, char *buf_) {
+     for (auto c : foo.view())
+       buf_[ptr++] = c;
+ }
+ 
+ /* { dg-final { scan-tree-dump "memcpy\[^\n\r\]*, 9\\);" "ldist" } } */

Re: [PATCH] Fix PR78847

[Richard Biener]

On Thu, 9 Feb 2017, Richard Biener wrote:

> 
> The following patch fixes a missed optimization caused by niter analysis
> returning a complex pointer subtraction expression instead of a constant
> for a loop with a pointer IV (it looks like pointer IVs are quite common
> in libstdc++).
> 
> Bootstrapped and tested on x86_64-unknown-linux-gnu.
> 
> Not a regression (well, didn't try to find a release where we optimized
> the testcase, but it needs c++14 library support anyway).  I'm also
> not quite sure associate_trees handles pointer types correctly
> (there is some pointer type handling in the associate case but
> POINTER_PLUS_EXPR folding for example does not dispatch to associate).
> 
> I guess trying to add some more exhaustive testcases for GENERIC input
> is required.

I've generalized a bit and convinced myself it is safe.

Bootstrapped and tested on x86_64-unknow-linux-gnu, applied to trunk.

Richard.

2017-04-21  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/78847
	* fold-const.c (split_tree): Handle POINTER_PLUS_EXPR.

	* g++.dg/tree-ssa/pr78847.C: New testcase.

Index: gcc/testsuite/g++.dg/tree-ssa/pr78847.C
===================================================================
*** gcc/testsuite/g++.dg/tree-ssa/pr78847.C	(nonexistent)
--- gcc/testsuite/g++.dg/tree-ssa/pr78847.C	(working copy)
***************
*** 0 ****
--- 1,26 ----
+ /* { dg-do compile } */
+ /* { dg-require-effective-target c++14 } */
+ /* { dg-options "-O3 -fdump-tree-ldist" } */
+ 
+ #include <stddef.h>
+ #include <cstring>
+ #include <experimental/string_view>
+ 
+ using string_view = std::experimental::string_view;
+ 
+ class Foo {
+     constexpr static size_t Length = 9;
+     char ascii_[Length];
+ public:
+     Foo();
+     string_view view() const {
+ 	return string_view(ascii_, Length);
+     }
+ };
+ 
+ void testWithLoopValue(const Foo foo, size_t ptr, char *buf_) {
+     for (auto c : foo.view())
+       buf_[ptr++] = c;
+ }
+ 
+ /* { dg-final { scan-tree-dump "memcpy\[^\n\r\]*, 9\\);" "ldist" } } */
Index: gcc/fold-const.c
===================================================================
--- gcc/fold-const.c	(revision 247049)
+++ gcc/fold-const.c	(working copy)
@@ -798,8 +798,11 @@ split_tree (location_t loc, tree in, tre
 		  though the C standard doesn't say so) for integers because
 		  the value is not affected.  For reals, the value might be
 		  affected, so we can't.  */
-	       && ((code == PLUS_EXPR && TREE_CODE (in) == MINUS_EXPR)
-		   || (code == MINUS_EXPR && TREE_CODE (in) == PLUS_EXPR))))
+	       && ((code == PLUS_EXPR && TREE_CODE (in) == POINTER_PLUS_EXPR)
+		   || (code == PLUS_EXPR && TREE_CODE (in) == MINUS_EXPR)
+		   || (code == MINUS_EXPR
+		       && (TREE_CODE (in) == PLUS_EXPR
+			   || TREE_CODE (in) == POINTER_PLUS_EXPR)))))
     {
       tree op0 = TREE_OPERAND (in, 0);
       tree op1 = TREE_OPERAND (in, 1);