expand_call_mem_ref checks for TARGET_MEM_REFs that have compatible
type, but it didn't then go on to install the specific type we need,
which might have different alignment due to:

  if (TYPE_ALIGN (type) != align)
    type = build_aligned_type (type, align);

This was causing masked stores to be incorrectly marked as
aligned on AVX512.

Tested on aarch64-linux-gnu, x86_64-linux-gnu and powerpc64-linux-gnu.
Also tested by Alexander on AVX512 hw.  OK to install?

Richard


2018-02-20  Richard Sandiford  <richard.sandif...@linaro.org>

gcc/
        PR tree-optimization/84419
        * internal-fn.c (expand_call_mem_ref): Create a TARGET_MEM_REF
        with the required type if its current type is compatible but
        different.

gcc/testsuite/
        PR tree-optimization/84419
        * gcc.dg/vect/pr84419.c: New test.

Index: gcc/internal-fn.c
===================================================================
--- gcc/internal-fn.c   2018-01-13 18:01:51.235735290 +0000
+++ gcc/internal-fn.c   2018-02-20 09:40:47.547217244 +0000
@@ -2444,11 +2444,14 @@ expand_call_mem_ref (tree type, gcall *s
          && types_compatible_p (TREE_TYPE (mem), type))
        {
          tree offset = TMR_OFFSET (mem);
-         if (alias_ptr_type != TREE_TYPE (offset) || !integer_zerop (offset))
+         if (type != TREE_TYPE (mem)
+             || alias_ptr_type != TREE_TYPE (offset)
+             || !integer_zerop (offset))
            {
              mem = copy_node (mem);
              TMR_OFFSET (mem) = wide_int_to_tree (alias_ptr_type,
                                                   wi::to_poly_wide (offset));
+             TREE_TYPE (mem) = type;
            }
          return mem;
        }
Index: gcc/testsuite/gcc.dg/vect/pr84419.c
===================================================================
--- /dev/null   2018-02-19 19:34:42.906488063 +0000
+++ gcc/testsuite/gcc.dg/vect/pr84419.c 2018-02-20 09:40:47.548217201 +0000
@@ -0,0 +1,21 @@
+#include <string.h>
+
+#define SIZE 400
+
+int  foo[SIZE];
+char bar[SIZE];
+
+void __attribute__ ((noinline)) foo_func(void)
+{
+  int i;
+  for (i = 1; i < SIZE; i++)
+    if (bar[i])
+      foo[i] = 1;
+}
+
+int main()
+{
+  memset(bar, 1, sizeof(bar));
+  foo_func();
+  return 0;
+}

Reply via email to