[Bug rtl-optimization/53861] New: Assignment of an array element from pointer is not taken as ARRAY_TYPE when expand_assignment

["zhenqiang.chen at linaro dot org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53861

             Bug #: 53861
           Summary: Assignment of an array element from pointer is not
                    taken as ARRAY_TYPE when expand_assignment
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zhenqiang.chen@linaro.org


In expr.c, there are comments and codes in function expand_assignment.

 /* Assignment of a structure component needs special treatment
     if the structure component's rtx is not simply a MEM.
     Assignment of an array element at a constant index, and assignment of
     an array element in an unaligned packed structure field, has the same
     problem.  Same for (partially) storing into a non-memory object.  */
  if (handled_component_p (to)
      || (TREE_CODE (to) == MEM_REF
          && mem_ref_refers_to_non_mem_p (to))
      || TREE_CODE (TREE_TYPE (to)) == ARRAY_TYPE)

But if an array element is accessed from a pointer, the condition check will
fail. Here is an example:

void test1( unsigned char* t, unsigned char* t1);
void test (int i, int j)
{
  unsigned char *p;
  unsigned char a1[16];
  unsigned char a[8];
  p = &a;
  p[1] = 1;
  p[3] = 3;
  a1[2] = 6;
  test1(a, a1);
}

Compile it with -O2/Os. The t.c.149t.optimized is like:

;; Function test (test, funcdef_no=0, decl_uid=4059, cgraph_uid=0)
test (int i, int j)
{
  unsigned char a[8];
  unsigned char a1[16];
<bb 2>:
  MEM[(unsigned char *)&a + 1B] = 1;
  MEM[(unsigned char *)&a + 3B] = 3;
  a1[2] = 6;
  test1 (&a, &a1);
  a1 ={v} {CLOBBER};
  a ={v} {CLOBBER};
  return;
}

The MEM_REF is not taken as an array element access. Then their address will be
simplified based on the sp during expanding. But in some targets (like ARM
THUMB1), sp can not be used in some sore instructions, we have to reset the
base address from sp before each reference.

If we take the MEM_REF as ARRAY_TYPE (by tracing its operand's TREE_TYPE, we
can find it is from ARRAY_TYPE), we can keep the index mode during expanding.

After that, if the targets support sp used in store instructions (like X86,
MIPS, ARM THUMB2 etc), fwprop1 can optimized it. Otherwise, just keep the index
mode, then we need only set the base address once.