re PR tree-optimization/50444 (-ftree-sra ignores alignment)

2012-01-27  Richard Guenther  <rguenther@suse.de>

	PR tree-optimization/50444
	* tree-sra.c (build_ref_for_offset): Properly adjust the
	MEM_REF type for unaligned accesses.

	* gcc.dg/torture/pr50444.c: New testcase.

From-SVN: r183630

gcc/ChangeLog

@@ -1,3 +1,9 @@
+2012-01-27  Richard Guenther  <rguenther@suse.de>
+
+	PR tree-optimization/50444
+	* tree-sra.c (build_ref_for_offset): Properly adjust the
+	MEM_REF type for unaligned accesses.
+
 2012-01-27  Richard Guenther  <rguenther@suse.de>
 
 	PR tree-optimization/50444

gcc/testsuite/ChangeLog

@@ -1,3 +1,8 @@
+2012-01-27  Richard Guenther  <rguenther@suse.de>
+
+	PR tree-optimization/50444
+	* gcc.dg/torture/pr50444.c: New testcase.
+
 2012-01-27  Tobias Burnus  <burnus@net-b.de>
 
 	PR fortran/51970

gcc/testsuite/gcc.dg/torture/pr50444.c

@@ -0,0 +1,76 @@
+/* { dg-require-effective-target sse2_runtime } */
+/* { dg-do run } */
+/* { dg-options "-msse2" } */
+
+typedef long long __m128i __attribute__ ((__vector_size__ (16),
+__may_alias__));
+typedef int __v4si __attribute__ ((__vector_size__ (16)));
+typedef long long __v2di __attribute__ ((__vector_size__ (16)));
+typedef unsigned int uint32_t;
+
+typedef struct {
+    uint32_t v[4];
+} a4x32;
+
+a4x32* incr(a4x32* x)
+{
+  x->v[0] += 1;
+  return x;
+}
+
+typedef struct {
+    __m128i m;
+} a1xm128i;
+
+static inline  a1xm128i ssefunc( a1xm128i in,  a1xm128i k)
+{
+  a1xm128i ret;
+  ret.m = (__m128i)__builtin_ia32_pxor128 ((__v2di)in.m, (__v2di)k.m);
+  return ret;
+}
+
+static  a4x32  caster( a4x32 c4x32,  a1xm128i k)
+{
+  a1xm128i c1x128;
+  if( sizeof(c4x32) != sizeof(c1x128) ) __builtin_abort();
+  __builtin_memcpy(&c1x128, &c4x32, sizeof(c1x128));
+  c1x128 = ssefunc(c1x128, k);
+  __builtin_memcpy(&c4x32, &c1x128, sizeof(c4x32));
+  return c4x32;
+}
+
+typedef struct  {
+    a1xm128i key;
+    a4x32 c;
+    __SIZE_TYPE__ elem;
+    a4x32 v;
+} Engine;
+
+void ctor(Engine *e)
+{
+  e->elem = 0;
+  e->key.m = (__m128i)(__v4si){ 0, 0, 0, 0 };
+  e->c.v[0] = 0;
+  e->c.v[1] = 0;
+  e->c.v[2] = 0;
+  e->c.v[3] = 0;
+}
+
+uint32_t method( Engine *e)
+{
+  if( e->elem == 0 )
+    {
+      e->v = caster(*incr(&e->c), e->key);
+      e->elem = 4;
+    }
+  return e->v.v[--e->elem];
+}
+
+int main()
+{
+  Engine e4; ctor(&e4);
+  Engine e5; ctor(&e5);
+  if(method(&e4)!=method(&e5))
+    __builtin_abort ();
+  return 0;
+}

gcc/tree-sra.c

@@ -1461,6 +1461,8 @@ build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
   tree prev_base = base;
   tree off;
   HOST_WIDE_INT base_offset;
+  unsigned HOST_WIDE_INT misalign;
+  unsigned int align;
 
   gcc_checking_assert (offset % BITS_PER_UNIT == 0);
 
@@ -1506,6 +1508,23 @@ build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
       base = build_fold_addr_expr (unshare_expr (base));
     }
 
+  /* If prev_base were always an originally performed access
+     we can extract more optimistic alignment information
+     by looking at the access mode.  That would constrain the
+     alignment of base + base_offset which we would need to
+     adjust according to offset.
+     ???  But it is not at all clear that prev_base is an access
+     that was in the IL that way, so be conservative for now.  */
+  align = get_pointer_alignment_1 (base, &misalign);
+  misalign += (double_int_sext (tree_to_double_int (off),
+				TYPE_PRECISION (TREE_TYPE (off))).low
+	       * BITS_PER_UNIT);
+  misalign = misalign & (align - 1);
+  if (misalign != 0)
+    align = (misalign & -misalign);
+  if (align < TYPE_ALIGN (exp_type))
+    exp_type = build_aligned_type (exp_type, align);
+
   return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
 }