===================================================================
@@ -5685,8 +5685,11 @@ simplify_subreg (enum machine_mode outer
gcc_assert (GET_MODE (op) == innermode
|| GET_MODE (op) == VOIDmode);
- gcc_assert ((byte % GET_MODE_SIZE (outermode)) == 0);
- gcc_assert (byte < GET_MODE_SIZE (innermode));
+ if ((byte % GET_MODE_SIZE (outermode)) != 0)
+ return NULL_RTX;
+
+ if (byte >= GET_MODE_SIZE (innermode))
+ return NULL_RTX;
if (outermode == innermode && !byte)
return op;
===================================================================
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-mavx2 -O2" } */
+
+#include <x86intrin.h>
+
+__m128d
+f (__m256d x)
+{
+ return *((__m128d*) ((double *) &x + 1));
+}
This patch fixes PR 55359, which is an ICE caused by my removal of a validate_subreg call in: http://gcc.gnu.org/ml/gcc-patches/2012-10/msg01353.html The question then is whether the caller of simplify_(gen_)subreg is responsible for checking whether a subreg is valid, or whether the simplification routines should return null for cases they can't handle. simplify_gen_subreg already contains a validate_subreg call, so the latter seemed better. At first I wanted to make simplify_subreg return null if !validate_subreg, but that hit a case in the x86 backend where we tried to create an SFmode constant by taking the lowpart of a DImode CONST_INT. This patch instead takes the more conservative approach of returning null for offsets that can't be handled. Tested on x86_64-linux-gnu. OK to install? Richard gcc/ PR middle-end/55359 * simplify-rtx.c (simplify_subreg): Return null for invalid offsets. gcc/testsuite/ * gcc.target/i386/pr55359.c: New test.