On Wed, Nov 13, 2013 at 08:12:41PM +0100, Uros Bizjak wrote:
> Patch is OK everywhere, with or without the suggested change.
Ok, here is what I've committed, so far to trunk/4.8, after another
round of bootstrap/regtest:
2013-11-14 Jakub Jelinek <ja...@redhat.com>
Uros Bizjak <ubiz...@gmail.com>
PR target/59101
* config/i386/i386.md (*anddi_2): Only allow CCZmode if
operands[2] satisfies_constraint_Z that might have bit 31 set.
* gcc.c-torture/execute/pr59101.c: New test.
--- gcc/config/i386/i386.md.jj 2013-11-13 19:17:58.281733148 +0100
+++ gcc/config/i386/i386.md 2013-11-13 20:41:08.941030127 +0100
@@ -7978,7 +7978,18 @@ (define_insn "*anddi_2"
(const_int 0)))
(set (match_operand:DI 0 "nonimmediate_operand" "=r,r,rm")
(and:DI (match_dup 1) (match_dup 2)))]
- "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
+ "TARGET_64BIT
+ && ix86_match_ccmode
+ (insn,
+ /* If we are going to emit andl instead of andq, and the operands[2]
+ constant might have the SImode sign bit set, make sure the sign
+ flag isn't tested, because the instruction will set the sign flag
+ based on bit 31 rather than bit 63. If it isn't CONST_INT,
+ conservatively assume it might have bit 31 set. */
+ (satisfies_constraint_Z (operands[2])
+ && (!CONST_INT_P (operands[2])
+ || val_signbit_known_set_p (SImode, INTVAL (operands[2]))))
+ ? CCZmode : CCNOmode)
&& ix86_binary_operator_ok (AND, DImode, operands)"
"@
and{l}\t{%k2, %k0|%k0, %k2}
--- gcc/testsuite/gcc.c-torture/execute/pr59101.c.jj 2013-11-13
20:37:47.795031458 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr59101.c 2013-11-13
20:37:47.795031458 +0100
@@ -0,0 +1,15 @@
+/* PR target/59101 */
+
+__attribute__((noinline, noclone)) int
+foo (int a)
+{
+ return (~a & 4102790424LL) > 0 | 6;
+}
+
+int
+main ()
+{
+ if (foo (0) != 7)
+ __builtin_abort ();
+ return 0;
+}
Jakub
