[PATCH 2/2] combine: Fix for PR81423

[Segher Boessenkool <segher@kernel.crashing.org>]

We here have an AND of a SUBREG of an LSHIFTRT.  If that SUBREG is
paradoxical, the extraction we form is the length of the size of the
inner mode, which includes some bits that should not be in the result.
Just give up in that case.

Tested on powerpc64-linux {-m32,-m64} and on x86_64-linux.  Committing
to trunk.


Segher


2018-07-18  Segher Boessenkool  <segher@kernel.crashing.org>

	PR rtl-optimization/81423
	* combine.c (make_compound_operation_int): Don't try to optimize
	the AND of a SUBREG of an LSHIFTRT if that SUBREG is paradoxical.

---
 gcc/combine.c | 15 +++------------
 1 file changed, 3 insertions(+), 12 deletions(-)

diff --git a/gcc/combine.c b/gcc/combine.c
index c5200db..c486f12 100644
--- a/gcc/combine.c
+++ b/gcc/combine.c
@@ -7990,18 +7990,9 @@ make_compound_operation_int (machine_mode mode, rtx *x_ptr,
 				     XEXP (inner_x0, 1),
 				     i, 1, 0, in_code == COMPARE);
 
-	  if (new_rtx)
-	    {
-	      /* If we narrowed the mode when dropping the subreg, then
-		 we must zero-extend to keep the semantics of the AND.  */
-	      if (GET_MODE_SIZE (inner_mode) >= GET_MODE_SIZE (mode))
-		;
-	      else if (SCALAR_INT_MODE_P (inner_mode))
-		new_rtx = simplify_gen_unary (ZERO_EXTEND, mode,
-					      new_rtx, inner_mode);
-	      else
-		new_rtx = NULL;
-	    }
+	  /* If we narrowed the mode when dropping the subreg, then we lose.  */
+	  if (GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (mode))
+	    new_rtx = NULL;
 
 	  /* If that didn't give anything, see if the AND simplifies on
 	     its own.  */
-- 
1.9.3