Hi!

When the base of a handled component (BIT_FIELD_REF in the testcase)
is SSA_NAME which happens to be expanded as some MEM (on the testcase
it is SSA_NAME set to VIEW_CONVERT_EXPR of an SSA_NAME that has MEM as
DECL_RTL), expand_expr_real_1 can try to update the MEM attributes from
exp, but that is wrong, it might change the alias set of the MEM, etc.
If the base is SSA_NAME, we should keep the attributes unmodified.
The patch actually also tests for !MEM_P (orig_op0), so that if
the SSA_NAME expanded to non-MEM (e.g. constant or REG), but the
normal_inner_ref expansion forces it for whatever reason in memory,
we still set the attributes of such a MEM, it is a temporary in that
case, rather than the original read.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2016-02-23  Jakub Jelinek  <ja...@redhat.com>
            Richard Biener  <rgue...@suse.de>

        PR middle-end/69909
        * expr.c (expand_expr_real_1) <normal_inner_ref>: Avoid
        set_mem_attributes if tem is SSA_NAME which got expanded
        as a MEM.

        * gcc.dg/torture/pr69909.c: New test.

--- gcc/expr.c.jj       2016-02-23 13:54:02.000000000 +0100
+++ gcc/expr.c  2016-02-23 14:30:23.810657866 +0100
@@ -10521,7 +10521,11 @@ expand_expr_real_1 (tree exp, rtx target
        if (op0 == orig_op0)
          op0 = copy_rtx (op0);
 
-       set_mem_attributes (op0, exp, 0);
+       /* Don't set memory attributes if the base expression is
+          SSA_NAME that got expanded as a MEM.  In that case, we should
+          just honor its original memory attributes.  */
+       if (TREE_CODE (tem) != SSA_NAME || !MEM_P (orig_op0))
+         set_mem_attributes (op0, exp, 0);
 
        if (REG_P (XEXP (op0, 0)))
          mark_reg_pointer (XEXP (op0, 0), MEM_ALIGN (op0));
--- gcc/testsuite/gcc.dg/torture/pr69909.c.jj   2016-02-23 14:25:27.819719259 
+0100
+++ gcc/testsuite/gcc.dg/torture/pr69909.c      2016-02-23 14:25:27.818719272 
+0100
@@ -0,0 +1,35 @@
+/* PR middle-end/69909 */
+/* { dg-do run { target int128 } } */
+/* { dg-additional-options "-w" } */
+
+typedef unsigned V __attribute__ ((vector_size (32)));
+typedef __int128 T;
+typedef __int128 U __attribute__ ((vector_size (32)));
+
+__attribute__((noinline, noclone)) T
+foo (T a, V b, V c, V d, V e, U f)
+{
+  d[6] ^= 0x10;
+  f -= (U) d;
+  f[1] |= f[1] << (a & 127);
+  c ^= d;
+  return b[7] + c[2] + c[2] + d[6] + e[2] + f[1];
+}
+
+int
+main ()
+{
+  if (__CHAR_BIT__ != 8 || sizeof (unsigned) != 4 || sizeof (T) != 16)
+    return 0;
+
+  T x = foo (1, (V) { 9, 2, 5, 8, 1, 2, 9, 3 },
+               (V) { 1, 2, 3, 4, 5, 6, 7, 8 },
+               (V) { 4, 1, 2, 9, 8, 3, 5, 2 },
+               (V) { 3, 6, 1, 3, 2, 9, 4, 8 }, (U) { 3, 5 });
+  if (((unsigned long long) (x >> 64) != 0xffffffffffffffffULL
+       || (unsigned long long) x != 0xfffffffe0000001aULL)
+      && ((unsigned long long) (x >> 64) != 0xfffffffffffffffdULL
+         || (unsigned long long) x != 0xffffffff00000022ULL))
+    __builtin_abort ();
+  return 0;
+}

        Jakub

Reply via email to