[PATCH] Fix nested vector shift ICE (PR middle-end/26379)

Jakub Jelinek jakub@redhat.com
Mon Feb 20 11:38:00 GMT 2006


Hi!

The following testcase ICEs on x86_64-linux (both -m32 and -m64).
The problem is that the nested shift optimization (in this case
((m >> 2) << 8)) in combine.c doesn't handle vector shifts
and ICEs when trying to optimize V4HImode shift 0x3fffffffffffffff << 8
(the mask itself is wrong, it would need to be vector { 0x3fff, 0x3fff,
0x3fff, 0x3fff }).  The following patch simply disables the optimization
for vector shifts, which is a safe thing to do for 4.0/4.1.  For
trunk I'm not sure if we want this as well or rewrite the optimization
to handle vector shift as well.
This is a 4.0/4.1/4.2 regression.
Ok to commit to 4.0/4.1 (to 4.1 probably after 4.1 is released)?  Is it
ok also for trunk followed by another patch to handle the optimization
later on?

2006-02-20  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/26379
	* combine.c (simplify_shift_const_1): Disable nested shifts
	optimization for vector shifts.

	* gcc.target/i386/mmx-7.c: New test.

--- gcc/combine.c.jj	2006-02-16 08:22:59.000000000 +0100
+++ gcc/combine.c	2006-02-20 12:23:27.000000000 +0100
@@ -8737,7 +8737,8 @@ simplify_shift_const_1 (enum rtx_code co
 	      && INTVAL (XEXP (varop, 1)) >= 0
 	      && INTVAL (XEXP (varop, 1)) < GET_MODE_BITSIZE (GET_MODE (varop))
 	      && GET_MODE_BITSIZE (result_mode) <= HOST_BITS_PER_WIDE_INT
-	      && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
+	      && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
+	      && !VECTOR_MODE_P (result_mode))
 	    {
 	      enum rtx_code first_code = GET_CODE (varop);
 	      unsigned int first_count = INTVAL (XEXP (varop, 1));
--- gcc/testsuite/gcc.target/i386/mmx-7.c.jj	2006-02-20 12:26:28.000000000 +0100
+++ gcc/testsuite/gcc.target/i386/mmx-7.c	2006-02-20 12:25:32.000000000 +0100
@@ -0,0 +1,18 @@
+/* PR middle-end/26379 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mmmx" } */
+
+#include <mmintrin.h>
+
+void
+foo (__m64 *p)
+{
+  __m64 m;
+
+  m = p[0];
+  m = _mm_srli_pi16(m, 2);
+  m = _mm_slli_pi16(m, 8);
+
+  p[0] = m;
+  _mm_empty();
+}

	Jakub



More information about the Gcc-patches mailing list