simplify-rtx: Fix the case where subreg of a vec_concat could fail [PR123822]

So combine (rightfully) produces:
(subreg:DD (vec_concat:V2DI (const_int -1 [0xffffffffffffffff])
        (reg:DI 145 [ _22 ])) 0)

And then calls simplify_subreg on it.
simplify_subreg then sees it can remove the vec_concat as this is
the lower part and try to do `(subreg:DD (const_int -1 [0xffffffffffffffff]))`
but that is an invalid constant for DF64 so simplify_subreg rejects
that and returns NULL but there is code in simplify_subreg to create
a SUBREG then if the subreg validates. The problem is we are using DImode
for the inner mode and that validates but when gen_rtx_SUBREG is called
the inner mode of DImode is not there; only VOIDmode.
So the fix is to check for VOIDmode before the call to validate_subreg.

Bootstrapped and tested on x86_64-linux-gnu. And lightly tested for aarch64-linux-gnu.

	PR rtl-optimization/123822

gcc/ChangeLog:

	* simplify-rtx.cc (simplify_context::simplify_subreg): Check the
	mode of the part to be non-VOIDmode before calling gen_rtx_SUBREG.

gcc/testsuite/ChangeLog:

	* gcc.dg/torture/pr123822-1.c: New test.

Signed-off-by: Andrew Pinski <andrew.pinski@oss.qualcomm.com>

gcc/simplify-rtx.cc

@@ -8491,7 +8491,8 @@ simplify_context::simplify_subreg (machine_mode outermode, rtx op,
       res = simplify_subreg (outermode, part, part_mode, final_offset);
       if (res)
 	return res;
-      if (validate_subreg (outermode, part_mode, part, final_offset))
+      if (GET_MODE (part) != VOIDmode
+	  && validate_subreg (outermode, part_mode, part, final_offset))
 	return gen_rtx_SUBREG (outermode, part, final_offset);
       return NULL_RTX;
     }

gcc/testsuite/gcc.dg/torture/pr123822-1.c

@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-ffast-math" } */
+/* { dg-require-effective-target dfp } */
+
+/* PR rtl-optimization/123822 */
+
+typedef __attribute__((__vector_size__(16))) _Decimal64 D;
+
+int i;
+D d;
+
+void
+foo()
+{
+  d /= (D)(d >= *(D *)__builtin_memset(&d, i, 3));
+}