This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[PATCH] Fix A < 0 ? C : 0 optimization (PR tree-optimization/78720)


Hi!

This patch fixes the recently added A < 0 ? C : 0 for power of 2 C
optimization.  The if (!TYPE_UNSIGNED (TREE_TYPE (@0))) part is just
for safety, I'd hope that unsigned < 0 is folded into 0 earlier, but just
in case this is handled first.

The issue which breaks the testcase is that the patch actually assumed
that C has type narrower or as wide as A, which generally doesn't have to be
the case.  If the type of C is narrower or as wide as A's type, then the
shift count is necessarily non-negative, but if A is narrower than C, we
might need a left shift instead of right shift (and in that case have to
convert A to C's type first, then left shift so that the MSB of A is moved
to the right position and finally and it.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2016-12-09  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/78720
	* match.pd (A < 0 ? C : 0): Only optimize for signed A.  If shift
	is negative, first convert to @1's type and then lshift it by -shift.

	* gcc.c-torture/execute/pr78720.c: New test.

--- gcc/match.pd.jj	2016-12-07 17:19:11.000000000 +0100
+++ gcc/match.pd	2016-12-09 10:13:58.531705042 +0100
@@ -2768,17 +2768,23 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
     (ncmp (convert:stype @0) { build_zero_cst (stype); })))))
 
 /* If we have A < 0 ? C : 0 where C is a power of 2, convert
-   this into a right shift followed by ANDing with C.  */
+   this into a right or left shift followed by ANDing with C.  */
 (simplify
  (cond
   (lt @0 integer_zerop)
   integer_pow2p@1 integer_zerop)
- (with {
+ (if (!TYPE_UNSIGNED (TREE_TYPE (@0)))
+  (with {
     int shift = element_precision (@0) - wi::exact_log2 (@1) - 1;
-  }
-  (bit_and
-   (convert (rshift @0 { build_int_cst (integer_type_node, shift); }))
-   @1)))
+    tree ctype = TREE_TYPE (@1);
+   }
+   (if (shift >= 0)
+    (bit_and
+     (convert (rshift @0 { build_int_cst (integer_type_node, shift); }))
+     @1)
+    (bit_and
+     (lshift (convert:ctype @0) { build_int_cst (integer_type_node, -shift); })
+     @1)))))
 
 /* When the addresses are not directly of decls compare base and offset.
    This implements some remaining parts of fold_comparison address
--- gcc/testsuite/gcc.c-torture/execute/pr78720.c.jj	2016-12-09 10:09:09.655375273 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr78720.c	2016-12-09 10:07:22.000000000 +0100
@@ -0,0 +1,29 @@
+/* PR tree-optimization/78720 */
+
+__attribute__((noinline, noclone)) long int
+foo (signed char x)
+{
+  return x < 0 ? 0x80000L : 0L;
+}
+
+__attribute__((noinline, noclone)) long int
+bar (signed char x)
+{
+  return x < 0 ? 0x80L : 0L;
+}
+
+__attribute__((noinline, noclone)) long int
+baz (signed char x)
+{
+  return x < 0 ? 0x20L : 0L;
+}
+
+int
+main ()
+{
+  if (foo (-1) != 0x80000L || bar (-1) != 0x80L || baz (-1) != 0x20L
+      || foo (0) != 0L || bar (0) != 0L || baz (0) != 0L
+      || foo (31) != 0L || bar (31) != 0L || baz (31) != 0L)
+    __builtin_abort ();
+  return 0;
+}

	Jakub


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]