OSDN Git Service

Daily bump.
[pf3gnuchains/gcc-fork.git] / gcc / builtins.c
index 7b24a0c..04980cc 100644 (file)
@@ -1,7 +1,7 @@
 /* Expand builtin functions.
    Copyright (C) 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
-   2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011
-   Free Software Foundation, Inc.
+   2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011,
+   2012 Free Software Foundation, Inc.
 
 This file is part of GCC.
 
@@ -75,11 +75,7 @@ const char * built_in_names[(int) END_BUILTINS] =
 
 /* Setup an array of _DECL trees, make sure each element is
    initialized to NULL_TREE.  */
-tree built_in_decls[(int) END_BUILTINS];
-/* Declarations used when constructing the builtin implicitly in the compiler.
-   It may be NULL_TREE when this is invalid (for instance runtime is not
-   required to implement the function call in all cases).  */
-tree implicit_built_in_decls[(int) END_BUILTINS];
+builtin_info_type builtin_info;
 
 static const char *c_getstr (tree);
 static rtx c_readstr (const char *, enum machine_mode);
@@ -227,16 +223,19 @@ static tree do_mpfr_bessel_n (tree, tree, tree,
                              const REAL_VALUE_TYPE *, bool);
 static tree do_mpfr_remquo (tree, tree, tree);
 static tree do_mpfr_lgamma_r (tree, tree, tree);
+static void expand_builtin_sync_synchronize (void);
 
 /* Return true if NAME starts with __builtin_ or __sync_.  */
 
-bool
+static bool
 is_builtin_name (const char *name)
 {
   if (strncmp (name, "__builtin_", 10) == 0)
     return true;
   if (strncmp (name, "__sync_", 7) == 0)
     return true;
+  if (strncmp (name, "__atomic_", 9) == 0)
+    return true;
   return false;
 }
 
@@ -264,8 +263,15 @@ called_as_built_in (tree node)
   return is_builtin_name (name);
 }
 
-/* Return the alignment in bits of EXP, an object.
-   Don't return more than MAX_ALIGN no matter what.  */
+/* Compute values M and N such that M divides (address of EXP - N) and
+   such that N < M.  Store N in *BITPOSP and return M.
+
+   Note that the address (and thus the alignment) computed here is based
+   on the address to which a symbol resolves, whereas DECL_ALIGN is based
+   on the address at which an object is actually located.  These two
+   addresses are not always the same.  For example, on ARM targets,
+   the address &foo of a Thumb function foo() has the lowest bit set,
+   whereas foo() itself starts on an even address.  */
 
 unsigned int
 get_object_alignment_1 (tree exp, unsigned HOST_WIDE_INT *bitposp)
@@ -287,7 +293,21 @@ get_object_alignment_1 (tree exp, unsigned HOST_WIDE_INT *bitposp)
     exp = DECL_INITIAL (exp);
   if (DECL_P (exp)
       && TREE_CODE (exp) != LABEL_DECL)
-    align = DECL_ALIGN (exp);
+    {
+      if (TREE_CODE (exp) == FUNCTION_DECL)
+       {
+         /* Function addresses can encode extra information besides their
+            alignment.  However, if TARGET_PTRMEMFUNC_VBIT_LOCATION
+            allows the low bit to be used as a virtual bit, we know
+            that the address itself must be 2-byte aligned.  */
+         if (TARGET_PTRMEMFUNC_VBIT_LOCATION == ptrmemfunc_vbit_in_pfn)
+           align = 2 * BITS_PER_UNIT;
+         else
+           align = BITS_PER_UNIT;
+       }
+      else
+       align = DECL_ALIGN (exp);
+    }
   else if (CONSTANT_CLASS_P (exp))
     {
       align = TYPE_ALIGN (TREE_TYPE (exp));
@@ -320,7 +340,7 @@ get_object_alignment_1 (tree exp, unsigned HOST_WIDE_INT *bitposp)
          align = MAX (pi->align * BITS_PER_UNIT, align);
        }
       else if (TREE_CODE (addr) == ADDR_EXPR)
-       align = MAX (align, get_object_alignment (TREE_OPERAND (addr, 0), ~0U));
+       align = MAX (align, get_object_alignment (TREE_OPERAND (addr, 0)));
       bitpos += mem_ref_offset (exp).low * BITS_PER_UNIT;
     }
   else if (TREE_CODE (exp) == TARGET_MEM_REF)
@@ -344,7 +364,7 @@ get_object_alignment_1 (tree exp, unsigned HOST_WIDE_INT *bitposp)
          align = MAX (pi->align * BITS_PER_UNIT, align);
        }
       else if (TREE_CODE (addr) == ADDR_EXPR)
-       align = MAX (align, get_object_alignment (TREE_OPERAND (addr, 0), ~0U));
+       align = MAX (align, get_object_alignment (TREE_OPERAND (addr, 0)));
       if (TMR_OFFSET (exp))
        bitpos += TREE_INT_CST_LOW (TMR_OFFSET (exp)) * BITS_PER_UNIT;
       if (TMR_INDEX (exp) && TMR_STEP (exp))
@@ -413,11 +433,10 @@ get_object_alignment_1 (tree exp, unsigned HOST_WIDE_INT *bitposp)
   return align;
 }
 
-/* Return the alignment in bits of EXP, an object.
-   Don't return more than MAX_ALIGN no matter what.  */
+/* Return the alignment in bits of EXP, an object.  */
 
 unsigned int
-get_object_alignment (tree exp, unsigned int max_align)
+get_object_alignment (tree exp)
 {
   unsigned HOST_WIDE_INT bitpos = 0;
   unsigned int align;
@@ -430,51 +449,87 @@ get_object_alignment (tree exp, unsigned int max_align)
   if (bitpos != 0)
     align = (bitpos & -bitpos);
 
-  return MIN (align, max_align);
+  return align;
 }
 
-/* Returns true iff we can trust that alignment information has been
-   calculated properly.  */
+/* Return the alignment of object EXP, also considering its type when we do
+   not know of explicit misalignment.  Only handle MEM_REF and TARGET_MEM_REF.
 
-bool
-can_trust_pointer_alignment (void)
+   ??? Note that, in the general case, the type of an expression is not kept
+   consistent with misalignment information by the front-end, for example when
+   taking the address of a member of a packed structure.  However, in most of
+   the cases, expressions have the alignment of their type so we optimistically
+   fall back to this alignment when we cannot compute a misalignment.  */
+
+unsigned int
+get_object_or_type_alignment (tree exp)
 {
-  /* We rely on TER to compute accurate alignment information.  */
-  return (optimize && flag_tree_ter);
+  unsigned HOST_WIDE_INT misalign;
+  unsigned int align = get_object_alignment_1 (exp, &misalign);
+
+  gcc_assert (TREE_CODE (exp) == MEM_REF || TREE_CODE (exp) == TARGET_MEM_REF);
+
+  if (misalign != 0)
+    align = (misalign & -misalign);
+  else
+    align = MAX (TYPE_ALIGN (TREE_TYPE (exp)), align);
+
+  return align;
 }
 
-/* Return the alignment in bits of EXP, a pointer valued expression.
-   But don't return more than MAX_ALIGN no matter what.
-   The alignment returned is, by default, the alignment of the thing that
-   EXP points to.  If it is not a POINTER_TYPE, 0 is returned.
+/* For a pointer valued expression EXP compute values M and N such that
+   M divides (EXP - N) and such that N < M.  Store N in *BITPOSP and return M.
 
-   Otherwise, look at the expression to see if we can do better, i.e., if the
-   expression is actually pointing at an object whose alignment is tighter.  */
+   If EXP is not a pointer, 0 is returned.  */
 
 unsigned int
-get_pointer_alignment (tree exp, unsigned int max_align)
+get_pointer_alignment_1 (tree exp, unsigned HOST_WIDE_INT *bitposp)
 {
   STRIP_NOPS (exp);
 
   if (TREE_CODE (exp) == ADDR_EXPR)
-    return get_object_alignment (TREE_OPERAND (exp, 0), max_align);
+    return get_object_alignment_1 (TREE_OPERAND (exp, 0), bitposp);
   else if (TREE_CODE (exp) == SSA_NAME
           && POINTER_TYPE_P (TREE_TYPE (exp)))
     {
       struct ptr_info_def *pi = SSA_NAME_PTR_INFO (exp);
-      unsigned align;
       if (!pi)
-       return BITS_PER_UNIT;
-      if (pi->misalign != 0)
-       align = (pi->misalign & -pi->misalign);
-      else
-       align = pi->align;
-      return MIN (max_align, align * BITS_PER_UNIT);
+       {
+         *bitposp = 0;
+         return BITS_PER_UNIT;
+       }
+      *bitposp = pi->misalign * BITS_PER_UNIT;
+      return pi->align * BITS_PER_UNIT;
     }
 
+  *bitposp = 0;
   return POINTER_TYPE_P (TREE_TYPE (exp)) ? BITS_PER_UNIT : 0;
 }
 
+/* Return the alignment in bits of EXP, a pointer valued expression.
+   The alignment returned is, by default, the alignment of the thing that
+   EXP points to.  If it is not a POINTER_TYPE, 0 is returned.
+
+   Otherwise, look at the expression to see if we can do better, i.e., if the
+   expression is actually pointing at an object whose alignment is tighter.  */
+
+unsigned int
+get_pointer_alignment (tree exp)
+{
+  unsigned HOST_WIDE_INT bitpos = 0;
+  unsigned int align;
+  
+  align = get_pointer_alignment_1 (exp, &bitpos);
+
+  /* align and bitpos now specify known low bits of the pointer.
+     ptr & (align - 1) == bitpos.  */
+
+  if (bitpos != 0)
+    align = (bitpos & -bitpos);
+
+  return align;
+}
+
 /* Compute the length of a C string.  TREE_STRING_LENGTH is not the right
    way, because it could contain a zero byte in the middle.
    TREE_STRING_LENGTH is the size of the character array, not the string.
@@ -1217,9 +1272,8 @@ get_memory_rtx (tree exp, tree len)
 
          gcc_assert (TREE_CODE (inner) == COMPONENT_REF);
 
-         if (MEM_OFFSET (mem)
-             && CONST_INT_P (MEM_OFFSET (mem)))
-           offset = INTVAL (MEM_OFFSET (mem));
+         if (MEM_OFFSET_KNOWN_P (mem))
+           offset = MEM_OFFSET (mem);
 
          if (offset >= 0 && len && host_integerp (len, 0))
            length = tree_low_cst (len, 0);
@@ -1274,11 +1328,14 @@ get_memory_rtx (tree exp, tree len)
          if (mem_expr != MEM_EXPR (mem))
            {
              set_mem_expr (mem, mem_expr);
-             set_mem_offset (mem, offset >= 0 ? GEN_INT (offset) : NULL_RTX);
+             if (offset >= 0)
+               set_mem_offset (mem, offset);
+             else
+               clear_mem_offset (mem);
            }
        }
       set_mem_alias_set (mem, 0);
-      set_mem_size (mem, NULL_RTX);
+      clear_mem_size (mem);
     }
 
   return mem;
@@ -1669,6 +1726,7 @@ expand_builtin_apply (rtx function, rtx arguments, rtx argsize)
   else
 #endif
     emit_stack_restore (SAVE_BLOCK, old_stack_level);
+  fixup_args_size_notes (call_insn, get_last_insn(), 0);
 
   OK_DEFER_POP;
 
@@ -1779,17 +1837,15 @@ expand_builtin_classify_type (tree exp)
   fcode = BUILT_IN_MATHFN##_R; fcodef = BUILT_IN_MATHFN##F_R ; \
   fcodel = BUILT_IN_MATHFN##L_R ; break;
 
-/* Return mathematic function equivalent to FN but operating directly
-   on TYPE, if available.  If IMPLICIT is true find the function in
-   implicit_built_in_decls[], otherwise use built_in_decls[].  If we
-   can't do the conversion, return zero.  */
+/* Return mathematic function equivalent to FN but operating directly on TYPE,
+   if available.  If IMPLICIT is true use the implicit builtin declaration,
+   otherwise use the explicit declaration.  If we can't do the conversion,
+   return zero.  */
 
 static tree
-mathfn_built_in_1 (tree type, enum built_in_function fn, bool implicit)
+mathfn_built_in_1 (tree type, enum built_in_function fn, bool implicit_p)
 {
-  tree const *const fn_arr
-    = implicit ? implicit_built_in_decls : built_in_decls;
-  enum built_in_function fcode, fcodef, fcodel;
+  enum built_in_function fcode, fcodef, fcodel, fcode2;
 
   switch (fn)
     {
@@ -1826,7 +1882,11 @@ mathfn_built_in_1 (tree type, enum built_in_function fn, bool implicit)
       CASE_MATHFN (BUILT_IN_HUGE_VAL)
       CASE_MATHFN (BUILT_IN_HYPOT)
       CASE_MATHFN (BUILT_IN_ILOGB)
+      CASE_MATHFN (BUILT_IN_ICEIL)
+      CASE_MATHFN (BUILT_IN_IFLOOR)
       CASE_MATHFN (BUILT_IN_INF)
+      CASE_MATHFN (BUILT_IN_IRINT)
+      CASE_MATHFN (BUILT_IN_IROUND)
       CASE_MATHFN (BUILT_IN_ISINF)
       CASE_MATHFN (BUILT_IN_J0)
       CASE_MATHFN (BUILT_IN_J1)
@@ -1882,13 +1942,18 @@ mathfn_built_in_1 (tree type, enum built_in_function fn, bool implicit)
       }
 
   if (TYPE_MAIN_VARIANT (type) == double_type_node)
-    return fn_arr[fcode];
+    fcode2 = fcode;
   else if (TYPE_MAIN_VARIANT (type) == float_type_node)
-    return fn_arr[fcodef];
+    fcode2 = fcodef;
   else if (TYPE_MAIN_VARIANT (type) == long_double_type_node)
-    return fn_arr[fcodel];
+    fcode2 = fcodel;
   else
     return NULL_TREE;
+
+  if (implicit_p && !builtin_decl_implicit_p (fcode2))
+    return NULL_TREE;
+
+  return builtin_decl_explicit (fcode2);
 }
 
 /* Like mathfn_built_in_1(), but always use the implicit array.  */
@@ -2538,11 +2603,11 @@ expand_builtin_cexpi (tree exp, rtx target)
       rtx op1a, op2a;
 
       if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CEXPIF)
-       fn = built_in_decls[BUILT_IN_SINCOSF];
+       fn = builtin_decl_explicit (BUILT_IN_SINCOSF);
       else if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CEXPI)
-       fn = built_in_decls[BUILT_IN_SINCOS];
+       fn = builtin_decl_explicit (BUILT_IN_SINCOS);
       else if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CEXPIL)
-       fn = built_in_decls[BUILT_IN_SINCOSL];
+       fn = builtin_decl_explicit (BUILT_IN_SINCOSL);
       else
        gcc_unreachable ();
 
@@ -2564,11 +2629,11 @@ expand_builtin_cexpi (tree exp, rtx target)
       tree ctype = build_complex_type (type);
 
       if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CEXPIF)
-       fn = built_in_decls[BUILT_IN_CEXPF];
+       fn = builtin_decl_explicit (BUILT_IN_CEXPF);
       else if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CEXPI)
-       fn = built_in_decls[BUILT_IN_CEXP];
+       fn = builtin_decl_explicit (BUILT_IN_CEXP);
       else if (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CEXPIL)
-       fn = built_in_decls[BUILT_IN_CEXPL];
+       fn = builtin_decl_explicit (BUILT_IN_CEXPL);
       else
        gcc_unreachable ();
 
@@ -2651,12 +2716,14 @@ expand_builtin_int_roundingfn (tree exp, rtx target)
 
   switch (DECL_FUNCTION_CODE (fndecl))
     {
+    CASE_FLT_FN (BUILT_IN_ICEIL):
     CASE_FLT_FN (BUILT_IN_LCEIL):
     CASE_FLT_FN (BUILT_IN_LLCEIL):
       builtin_optab = lceil_optab;
       fallback_fn = BUILT_IN_CEIL;
       break;
 
+    CASE_FLT_FN (BUILT_IN_IFLOOR):
     CASE_FLT_FN (BUILT_IN_LFLOOR):
     CASE_FLT_FN (BUILT_IN_LLFLOOR):
       builtin_optab = lfloor_optab;
@@ -2709,26 +2776,32 @@ expand_builtin_int_roundingfn (tree exp, rtx target)
 
       switch (DECL_FUNCTION_CODE (fndecl))
        {
+       case BUILT_IN_ICEIL:
        case BUILT_IN_LCEIL:
        case BUILT_IN_LLCEIL:
          name = "ceil";
          break;
+       case BUILT_IN_ICEILF:
        case BUILT_IN_LCEILF:
        case BUILT_IN_LLCEILF:
          name = "ceilf";
          break;
+       case BUILT_IN_ICEILL:
        case BUILT_IN_LCEILL:
        case BUILT_IN_LLCEILL:
          name = "ceill";
          break;
+       case BUILT_IN_IFLOOR:
        case BUILT_IN_LFLOOR:
        case BUILT_IN_LLFLOOR:
          name = "floor";
          break;
+       case BUILT_IN_IFLOORF:
        case BUILT_IN_LFLOORF:
        case BUILT_IN_LLFLOORF:
          name = "floorf";
          break;
+       case BUILT_IN_IFLOORL:
        case BUILT_IN_LFLOORL:
        case BUILT_IN_LLFLOORL:
          name = "floorl";
@@ -2768,10 +2841,7 @@ expand_builtin_int_roundingfn_2 (tree exp, rtx target)
   tree fndecl = get_callee_fndecl (exp);
   tree arg;
   enum machine_mode mode;
-
-  /* There's no easy way to detect the case we need to set EDOM.  */
-  if (flag_errno_math)
-    return NULL_RTX;
+  enum built_in_function fallback_fn = BUILT_IN_NONE;
 
   if (!validate_arglist (exp, REAL_TYPE, VOID_TYPE))
      gcc_unreachable ();
@@ -2780,43 +2850,79 @@ expand_builtin_int_roundingfn_2 (tree exp, rtx target)
 
   switch (DECL_FUNCTION_CODE (fndecl))
     {
+    CASE_FLT_FN (BUILT_IN_IRINT):
+      fallback_fn = BUILT_IN_LRINT;
+      /* FALLTHRU */
     CASE_FLT_FN (BUILT_IN_LRINT):
     CASE_FLT_FN (BUILT_IN_LLRINT):
-      builtin_optab = lrint_optab; break;
+      builtin_optab = lrint_optab;
+      break;
+
+    CASE_FLT_FN (BUILT_IN_IROUND):
+      fallback_fn = BUILT_IN_LROUND;
+      /* FALLTHRU */
     CASE_FLT_FN (BUILT_IN_LROUND):
     CASE_FLT_FN (BUILT_IN_LLROUND):
-      builtin_optab = lround_optab; break;
+      builtin_optab = lround_optab;
+      break;
+
     default:
       gcc_unreachable ();
     }
 
+  /* There's no easy way to detect the case we need to set EDOM.  */
+  if (flag_errno_math && fallback_fn == BUILT_IN_NONE)
+    return NULL_RTX;
+
   /* Make a suitable register to place result in.  */
   mode = TYPE_MODE (TREE_TYPE (exp));
 
-  target = gen_reg_rtx (mode);
+  /* There's no easy way to detect the case we need to set EDOM.  */
+  if (!flag_errno_math)
+    {
+      target = gen_reg_rtx (mode);
 
-  /* Wrap the computation of the argument in a SAVE_EXPR, as we may
-     need to expand the argument again.  This way, we will not perform
-     side-effects more the once.  */
-  CALL_EXPR_ARG (exp, 0) = arg = builtin_save_expr (arg);
+      /* Wrap the computation of the argument in a SAVE_EXPR, as we may
+        need to expand the argument again.  This way, we will not perform
+        side-effects more the once.  */
+      CALL_EXPR_ARG (exp, 0) = arg = builtin_save_expr (arg);
 
-  op0 = expand_expr (arg, NULL, VOIDmode, EXPAND_NORMAL);
+      op0 = expand_expr (arg, NULL, VOIDmode, EXPAND_NORMAL);
 
-  start_sequence ();
+      start_sequence ();
 
-  if (expand_sfix_optab (target, op0, builtin_optab))
-    {
-      /* Output the entire sequence.  */
-      insns = get_insns ();
+      if (expand_sfix_optab (target, op0, builtin_optab))
+       {
+         /* Output the entire sequence.  */
+         insns = get_insns ();
+         end_sequence ();
+         emit_insn (insns);
+         return target;
+       }
+
+      /* If we were unable to expand via the builtin, stop the sequence
+        (without outputting the insns) and call to the library function
+        with the stabilized argument list.  */
       end_sequence ();
-      emit_insn (insns);
-      return target;
     }
 
-  /* If we were unable to expand via the builtin, stop the sequence
-     (without outputting the insns) and call to the library function
-     with the stabilized argument list.  */
-  end_sequence ();
+  if (fallback_fn != BUILT_IN_NONE)
+    {
+      /* Fall back to rounding to long int.  Use implicit_p 0 - for non-C99
+        targets, (int) round (x) should never be transformed into
+        BUILT_IN_IROUND and if __builtin_iround is called directly, emit
+        a call to lround in the hope that the target provides at least some
+        C99 functions.  This should result in the best user experience for
+        not full C99 targets.  */
+      tree fallback_fndecl = mathfn_built_in_1 (TREE_TYPE (arg),
+                                               fallback_fn, 0);
+
+      exp = build_call_nofold_loc (EXPR_LOCATION (exp),
+                                  fallback_fndecl, 1, arg);
+
+      target = expand_call (exp, NULL_RTX, target == const0_rtx);
+      return convert_to_mode (mode, target, 0);
+    }
 
   target = expand_call (exp, target, target == const0_rtx);
 
@@ -2903,7 +3009,7 @@ expand_builtin_strlen (tree exp, rtx target,
          return expand_expr (len, target, target_mode, EXPAND_NORMAL);
        }
 
-      align = get_pointer_alignment (src, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
+      align = get_pointer_alignment (src) / BITS_PER_UNIT;
 
       /* If SRC is not a pointer type, don't do this operation inline.  */
       if (align == 0)
@@ -2939,9 +3045,16 @@ expand_builtin_strlen (tree exp, rtx target,
 
       /* Now that we are assured of success, expand the source.  */
       start_sequence ();
-      pat = expand_expr (src, src_reg, ptr_mode, EXPAND_NORMAL);
+      pat = expand_expr (src, src_reg, Pmode, EXPAND_NORMAL);
       if (pat != src_reg)
-       emit_move_insn (src_reg, pat);
+       {
+#ifdef POINTERS_EXTEND_UNSIGNED
+         if (GET_MODE (pat) != Pmode)
+           pat = convert_to_mode (Pmode, pat,
+                                  POINTERS_EXTEND_UNSIGNED);
+#endif
+         emit_move_insn (src_reg, pat);
+       }
       pat = get_insns ();
       end_sequence ();
 
@@ -2996,9 +3109,8 @@ expand_builtin_memcpy (tree exp, rtx target)
       tree src = CALL_EXPR_ARG (exp, 1);
       tree len = CALL_EXPR_ARG (exp, 2);
       const char *src_str;
-      unsigned int src_align = get_pointer_alignment (src, BIGGEST_ALIGNMENT);
-      unsigned int dest_align
-       = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+      unsigned int src_align = get_pointer_alignment (src);
+      unsigned int dest_align = get_pointer_alignment (dest);
       rtx dest_mem, src_mem, dest_addr, len_rtx;
       HOST_WIDE_INT expected_size = -1;
       unsigned int expected_align = 0;
@@ -3095,9 +3207,9 @@ expand_builtin_mempcpy_args (tree dest, tree src, tree len,
                             rtx target, enum machine_mode mode, int endp)
 {
     /* If return value is ignored, transform mempcpy into memcpy.  */
-  if (target == const0_rtx && implicit_built_in_decls[BUILT_IN_MEMCPY])
+  if (target == const0_rtx && builtin_decl_implicit_p (BUILT_IN_MEMCPY))
     {
-      tree fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+      tree fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
       tree result = build_call_nofold_loc (UNKNOWN_LOCATION, fn, 3,
                                           dest, src, len);
       return expand_expr (result, target, mode, EXPAND_NORMAL);
@@ -3105,9 +3217,8 @@ expand_builtin_mempcpy_args (tree dest, tree src, tree len,
   else
     {
       const char *src_str;
-      unsigned int src_align = get_pointer_alignment (src, BIGGEST_ALIGNMENT);
-      unsigned int dest_align
-       = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+      unsigned int src_align = get_pointer_alignment (src);
+      unsigned int dest_align = get_pointer_alignment (dest);
       rtx dest_mem, src_mem, len_rtx;
 
       /* If either SRC or DEST is not a pointer type, don't do this
@@ -3259,9 +3370,9 @@ expand_builtin_stpcpy (tree exp, rtx target, enum machine_mode mode)
   src = CALL_EXPR_ARG (exp, 1);
 
   /* If return value is ignored, transform stpcpy into strcpy.  */
-  if (target == const0_rtx && implicit_built_in_decls[BUILT_IN_STRCPY])
+  if (target == const0_rtx && builtin_decl_implicit (BUILT_IN_STRCPY))
     {
-      tree fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+      tree fn = builtin_decl_implicit (BUILT_IN_STRCPY);
       tree result = build_call_nofold_loc (loc, fn, 2, dst, src);
       return expand_expr (result, target, mode, EXPAND_NORMAL);
     }
@@ -3360,8 +3471,7 @@ expand_builtin_strncpy (tree exp, rtx target)
         use store_by_pieces, if it fails, punt.  */
       if (tree_int_cst_lt (slen, len))
        {
-         unsigned int dest_align
-           = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+         unsigned int dest_align = get_pointer_alignment (dest);
          const char *p = c_getstr (src);
          rtx dest_mem;
 
@@ -3465,7 +3575,7 @@ expand_builtin_memset_args (tree dest, tree val, tree len,
   HOST_WIDE_INT expected_size = -1;
   unsigned int expected_align = 0;
 
-  dest_align = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+  dest_align = get_pointer_alignment (dest);
 
   /* If DEST is not a pointer type, don't do this operation in-line.  */
   if (dest_align == 0)
@@ -3599,14 +3709,15 @@ expand_builtin_bzero (tree exp)
      calling bzero instead of memset.  */
 
   return expand_builtin_memset_args (dest, integer_zero_node,
-                                    fold_convert_loc (loc, sizetype, size),
+                                    fold_convert_loc (loc,
+                                                      size_type_node, size),
                                     const0_rtx, VOIDmode, exp);
 }
 
 /* Expand expression EXP, which is a call to the memcmp built-in function.
-   Return NULL_RTX if we failed and the
-   caller should emit a normal call, otherwise try to get the result in
-   TARGET, if convenient (and in mode MODE, if that's convenient).  */
+   Return NULL_RTX if we failed and the caller should emit a normal call,
+   otherwise try to get the result in TARGET, if convenient (and in mode
+   MODE, if that's convenient).  */
 
 static rtx
 expand_builtin_memcmp (tree exp, ATTRIBUTE_UNUSED rtx target,
@@ -3618,7 +3729,10 @@ expand_builtin_memcmp (tree exp, ATTRIBUTE_UNUSED rtx target,
                         POINTER_TYPE, POINTER_TYPE, INTEGER_TYPE, VOID_TYPE))
     return NULL_RTX;
 
-#if defined HAVE_cmpmemsi || defined HAVE_cmpstrnsi
+  /* Note: The cmpstrnsi pattern, if it exists, is not suitable for
+     implementing memcmp because it will stop if it encounters two
+     zero bytes.  */
+#if defined HAVE_cmpmemsi
   {
     rtx arg1_rtx, arg2_rtx, arg3_rtx;
     rtx result;
@@ -3627,22 +3741,13 @@ expand_builtin_memcmp (tree exp, ATTRIBUTE_UNUSED rtx target,
     tree arg2 = CALL_EXPR_ARG (exp, 1);
     tree len = CALL_EXPR_ARG (exp, 2);
 
-    unsigned int arg1_align
-      = get_pointer_alignment (arg1, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
-    unsigned int arg2_align
-      = get_pointer_alignment (arg2, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
+    unsigned int arg1_align = get_pointer_alignment (arg1) / BITS_PER_UNIT;
+    unsigned int arg2_align = get_pointer_alignment (arg2) / BITS_PER_UNIT;
     enum machine_mode insn_mode;
 
-#ifdef HAVE_cmpmemsi
     if (HAVE_cmpmemsi)
       insn_mode = insn_data[(int) CODE_FOR_cmpmemsi].operand[0].mode;
     else
-#endif
-#ifdef HAVE_cmpstrnsi
-    if (HAVE_cmpstrnsi)
-      insn_mode = insn_data[(int) CODE_FOR_cmpstrnsi].operand[0].mode;
-    else
-#endif
       return NULL_RTX;
 
     /* If we don't have POINTER_TYPE, call the function.  */
@@ -3663,22 +3768,14 @@ expand_builtin_memcmp (tree exp, ATTRIBUTE_UNUSED rtx target,
     /* Set MEM_SIZE as appropriate.  */
     if (CONST_INT_P (arg3_rtx))
       {
-       set_mem_size (arg1_rtx, arg3_rtx);
-       set_mem_size (arg2_rtx, arg3_rtx);
+       set_mem_size (arg1_rtx, INTVAL (arg3_rtx));
+       set_mem_size (arg2_rtx, INTVAL (arg3_rtx));
       }
 
-#ifdef HAVE_cmpmemsi
     if (HAVE_cmpmemsi)
       insn = gen_cmpmemsi (result, arg1_rtx, arg2_rtx, arg3_rtx,
                           GEN_INT (MIN (arg1_align, arg2_align)));
     else
-#endif
-#ifdef HAVE_cmpstrnsi
-    if (HAVE_cmpstrnsi)
-      insn = gen_cmpstrnsi (result, arg1_rtx, arg2_rtx, arg3_rtx,
-                           GEN_INT (MIN (arg1_align, arg2_align)));
-    else
-#endif
       gcc_unreachable ();
 
     if (insn)
@@ -3704,7 +3801,7 @@ expand_builtin_memcmp (tree exp, ATTRIBUTE_UNUSED rtx target,
     else
       return convert_to_mode (mode, result, 0);
   }
-#endif
+#endif /* HAVE_cmpmemsi.  */
 
   return NULL_RTX;
 }
@@ -3729,10 +3826,8 @@ expand_builtin_strcmp (tree exp, ATTRIBUTE_UNUSED rtx target)
       tree arg1 = CALL_EXPR_ARG (exp, 0);
       tree arg2 = CALL_EXPR_ARG (exp, 1);
 
-      unsigned int arg1_align
-       = get_pointer_alignment (arg1, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
-      unsigned int arg2_align
-       = get_pointer_alignment (arg2, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
+      unsigned int arg1_align = get_pointer_alignment (arg1) / BITS_PER_UNIT;
+      unsigned int arg2_align = get_pointer_alignment (arg2) / BITS_PER_UNIT;
 
       /* If we don't have POINTER_TYPE, call the function.  */
       if (arg1_align == 0 || arg2_align == 0)
@@ -3880,10 +3975,8 @@ expand_builtin_strncmp (tree exp, ATTRIBUTE_UNUSED rtx target,
     tree arg2 = CALL_EXPR_ARG (exp, 1);
     tree arg3 = CALL_EXPR_ARG (exp, 2);
 
-    unsigned int arg1_align
-      = get_pointer_alignment (arg1, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
-    unsigned int arg2_align
-      = get_pointer_alignment (arg2, BIGGEST_ALIGNMENT) / BITS_PER_UNIT;
+    unsigned int arg1_align = get_pointer_alignment (arg1) / BITS_PER_UNIT;
+    unsigned int arg2_align = get_pointer_alignment (arg2) / BITS_PER_UNIT;
     enum machine_mode insn_mode
       = insn_data[(int) CODE_FOR_cmpstrnsi].operand[0].mode;
 
@@ -4208,16 +4301,13 @@ std_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p,
       && !integer_zerop (TYPE_SIZE (type)))
     {
       t = build2 (MODIFY_EXPR, TREE_TYPE (valist), valist_tmp,
-                 fold_build2 (POINTER_PLUS_EXPR,
-                              TREE_TYPE (valist),
-                              valist_tmp, size_int (boundary - 1)));
+                 fold_build_pointer_plus_hwi (valist_tmp, boundary - 1));
       gimplify_and_add (t, pre_p);
 
-      t = fold_convert (sizetype, valist_tmp);
       t = build2 (MODIFY_EXPR, TREE_TYPE (valist), valist_tmp,
-                 fold_convert (TREE_TYPE (valist),
-                               fold_build2 (BIT_AND_EXPR, sizetype, t,
-                                            size_int (-boundary))));
+                 fold_build2 (BIT_AND_EXPR, TREE_TYPE (valist),
+                              valist_tmp,
+                              build_int_cst (TREE_TYPE (valist), -boundary)));
       gimplify_and_add (t, pre_p);
     }
   else
@@ -4249,12 +4339,11 @@ std_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p,
                       rounded_size, size_int (align));
       t = fold_build3 (COND_EXPR, sizetype, t, size_zero_node,
                       size_binop (MINUS_EXPR, rounded_size, type_size));
-      addr = fold_build2 (POINTER_PLUS_EXPR,
-                         TREE_TYPE (addr), addr, t);
+      addr = fold_build_pointer_plus (addr, t);
     }
 
   /* Compute new value for AP.  */
-  t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (valist), valist_tmp, rounded_size);
+  t = fold_build_pointer_plus (valist_tmp, rounded_size);
   t = build2 (MODIFY_EXPR, TREE_TYPE (valist), valist, t);
   gimplify_and_add (t, pre_p);
 
@@ -4342,7 +4431,7 @@ gimplify_va_arg_expr (tree *expr_p, gimple_seq *pre_p, gimple_seq *post_p)
         expression to exit or longjmp.  */
       gimplify_and_add (valist, pre_p);
       t = build_call_expr_loc (loc,
-                              implicit_built_in_decls[BUILT_IN_TRAP], 0);
+                              builtin_decl_implicit (BUILT_IN_TRAP), 0);
       gimplify_and_add (t, pre_p);
 
       /* This is dead code, but go ahead and finish so that the
@@ -4505,20 +4594,33 @@ expand_builtin_alloca (tree exp, bool cannot_accumulate)
 {
   rtx op0;
   rtx result;
+  bool valid_arglist;
+  unsigned int align;
+  bool alloca_with_align = (DECL_FUNCTION_CODE (get_callee_fndecl (exp))
+                           == BUILT_IN_ALLOCA_WITH_ALIGN);
 
-  /* Emit normal call if marked not-inlineable.  */
-  if (CALL_CANNOT_INLINE_P (exp))
+  /* Emit normal call if we use mudflap.  */
+  if (flag_mudflap)
     return NULL_RTX;
 
-  if (!validate_arglist (exp, INTEGER_TYPE, VOID_TYPE))
+  valid_arglist
+    = (alloca_with_align
+       ? validate_arglist (exp, INTEGER_TYPE, INTEGER_TYPE, VOID_TYPE)
+       : validate_arglist (exp, INTEGER_TYPE, VOID_TYPE));
+
+  if (!valid_arglist)
     return NULL_RTX;
 
   /* Compute the argument.  */
   op0 = expand_normal (CALL_EXPR_ARG (exp, 0));
 
+  /* Compute the alignment.  */
+  align = (alloca_with_align
+          ? TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 1))
+          : BIGGEST_ALIGNMENT);
+
   /* Allocate the desired space.  */
-  result = allocate_dynamic_stack_space (op0, 0, BIGGEST_ALIGNMENT,
-                                        cannot_accumulate);
+  result = allocate_dynamic_stack_space (op0, 0, align, cannot_accumulate);
   result = convert_memory_address (ptr_mode, result);
 
   return result;
@@ -4571,7 +4673,7 @@ expand_builtin_unop (enum machine_mode target_mode, tree exp, rtx target,
   /* Compute op, into TARGET if possible.
      Set TARGET to wherever the result comes back.  */
   target = expand_unop (TYPE_MODE (TREE_TYPE (CALL_EXPR_ARG (exp, 0))),
-                       op_optab, op0, target, 1);
+                       op_optab, op0, target, op_optab != clrsb_optab);
   gcc_assert (target);
 
   return convert_to_mode (target_mode, target, 0);
@@ -4597,6 +4699,23 @@ expand_builtin_expect (tree exp, rtx target)
   return target;
 }
 
+/* Expand a call to __builtin_assume_aligned.  We just return our first
+   argument as the builtin_assume_aligned semantic should've been already
+   executed by CCP.  */
+
+static rtx
+expand_builtin_assume_aligned (tree exp, rtx target)
+{
+  if (call_expr_nargs (exp) < 2)
+    return const0_rtx;
+  target = expand_expr (CALL_EXPR_ARG (exp, 0), target, VOIDmode,
+                       EXPAND_NORMAL);
+  gcc_assert (!TREE_SIDE_EFFECTS (CALL_EXPR_ARG (exp, 1))
+             && (call_expr_nargs (exp) < 3
+                 || !TREE_SIDE_EFFECTS (CALL_EXPR_ARG (exp, 2))));
+  return target;
+}
+
 void
 expand_builtin_trap (void)
 {
@@ -4764,7 +4883,7 @@ round_trampoline_addr (rtx tramp)
 }
 
 static rtx
-expand_builtin_init_trampoline (tree exp)
+expand_builtin_init_trampoline (tree exp, bool onstack)
 {
   tree t_tramp, t_func, t_chain;
   rtx m_tramp, r_tramp, r_chain, tmp;
@@ -4781,19 +4900,22 @@ expand_builtin_init_trampoline (tree exp)
   m_tramp = gen_rtx_MEM (BLKmode, r_tramp);
   MEM_NOTRAP_P (m_tramp) = 1;
 
-  /* The TRAMP argument should be the address of a field within the
-     local function's FRAME decl.  Let's see if we can fill in the
-     to fill in the MEM_ATTRs for this memory.  */
+  /* If ONSTACK, the TRAMP argument should be the address of a field
+     within the local function's FRAME decl.  Either way, let's see if
+     we can fill in the MEM_ATTRs for this memory.  */
   if (TREE_CODE (t_tramp) == ADDR_EXPR)
     set_mem_attributes_minus_bitpos (m_tramp, TREE_OPERAND (t_tramp, 0),
                                     true, 0);
 
+  /* Creator of a heap trampoline is responsible for making sure the
+     address is aligned to at least STACK_BOUNDARY.  Normally malloc
+     will ensure this anyhow.  */
   tmp = round_trampoline_addr (r_tramp);
   if (tmp != r_tramp)
     {
       m_tramp = change_address (m_tramp, BLKmode, tmp);
       set_mem_align (m_tramp, TRAMPOLINE_ALIGNMENT);
-      set_mem_size (m_tramp, GEN_INT (TRAMPOLINE_SIZE));
+      set_mem_size (m_tramp, TRAMPOLINE_SIZE);
     }
 
   /* The FUNC argument should be the address of the nested function.
@@ -4807,10 +4929,13 @@ expand_builtin_init_trampoline (tree exp)
   /* Generate insns to initialize the trampoline.  */
   targetm.calls.trampoline_init (m_tramp, t_func, r_chain);
 
-  trampolines_created = 1;
+  if (onstack)
+    {
+      trampolines_created = 1;
 
-  warning_at (DECL_SOURCE_LOCATION (t_func), OPT_Wtrampolines,
-              "trampoline generated for nested function %qD", t_func);
+      warning_at (DECL_SOURCE_LOCATION (t_func), OPT_Wtrampolines,
+                 "trampoline generated for nested function %qD", t_func);
+    }
 
   return const0_rtx;
 }
@@ -5043,28 +5168,48 @@ get_builtin_sync_mem (tree loc, enum machine_mode mode)
 
   /* The alignment needs to be at least according to that of the mode.  */
   set_mem_align (mem, MAX (GET_MODE_ALIGNMENT (mode),
-                          get_pointer_alignment (loc, BIGGEST_ALIGNMENT)));
+                          get_pointer_alignment (loc)));
   set_mem_alias_set (mem, ALIAS_SET_MEMORY_BARRIER);
   MEM_VOLATILE_P (mem) = 1;
 
   return mem;
 }
 
+/* Make sure an argument is in the right mode.
+   EXP is the tree argument. 
+   MODE is the mode it should be in.  */
+
+static rtx
+expand_expr_force_mode (tree exp, enum machine_mode mode)
+{
+  rtx val;
+  enum machine_mode old_mode;
+
+  val = expand_expr (exp, NULL_RTX, mode, EXPAND_NORMAL);
+  /* If VAL is promoted to a wider mode, convert it back to MODE.  Take care
+     of CONST_INTs, where we know the old_mode only from the call argument.  */
+
+  old_mode = GET_MODE (val);
+  if (old_mode == VOIDmode)
+    old_mode = TYPE_MODE (TREE_TYPE (exp));
+  val = convert_modes (mode, old_mode, val, 1);
+  return val;
+}
+
+
 /* Expand the __sync_xxx_and_fetch and __sync_fetch_and_xxx intrinsics.
    EXP is the CALL_EXPR.  CODE is the rtx code
    that corresponds to the arithmetic or logical operation from the name;
    an exception here is that NOT actually means NAND.  TARGET is an optional
    place for us to store the results; AFTER is true if this is the
-   fetch_and_xxx form.  IGNORE is true if we don't actually care about
-   the result of the operation at all.  */
+   fetch_and_xxx form.  */
 
 static rtx
 expand_builtin_sync_operation (enum machine_mode mode, tree exp,
                               enum rtx_code code, bool after,
-                              rtx target, bool ignore)
+                              rtx target)
 {
   rtx val, mem;
-  enum machine_mode old_mode;
   location_t loc = EXPR_LOCATION (exp);
 
   if (code == NOT && warn_sync_nand)
@@ -5076,181 +5221,567 @@ expand_builtin_sync_operation (enum machine_mode mode, tree exp,
 
       switch (fcode)
        {
-       case BUILT_IN_FETCH_AND_NAND_1:
-       case BUILT_IN_FETCH_AND_NAND_2:
-       case BUILT_IN_FETCH_AND_NAND_4:
-       case BUILT_IN_FETCH_AND_NAND_8:
-       case BUILT_IN_FETCH_AND_NAND_16:
-
+       case BUILT_IN_SYNC_FETCH_AND_NAND_1:
+       case BUILT_IN_SYNC_FETCH_AND_NAND_2:
+       case BUILT_IN_SYNC_FETCH_AND_NAND_4:
+       case BUILT_IN_SYNC_FETCH_AND_NAND_8:
+       case BUILT_IN_SYNC_FETCH_AND_NAND_16:
          if (warned_f_a_n)
            break;
 
-         fndecl = implicit_built_in_decls[BUILT_IN_FETCH_AND_NAND_N];
+         fndecl = builtin_decl_implicit (BUILT_IN_SYNC_FETCH_AND_NAND_N);
          inform (loc, "%qD changed semantics in GCC 4.4", fndecl);
          warned_f_a_n = true;
          break;
 
-       case BUILT_IN_NAND_AND_FETCH_1:
-       case BUILT_IN_NAND_AND_FETCH_2:
-       case BUILT_IN_NAND_AND_FETCH_4:
-       case BUILT_IN_NAND_AND_FETCH_8:
-       case BUILT_IN_NAND_AND_FETCH_16:
+       case BUILT_IN_SYNC_NAND_AND_FETCH_1:
+       case BUILT_IN_SYNC_NAND_AND_FETCH_2:
+       case BUILT_IN_SYNC_NAND_AND_FETCH_4:
+       case BUILT_IN_SYNC_NAND_AND_FETCH_8:
+       case BUILT_IN_SYNC_NAND_AND_FETCH_16:
+         if (warned_n_a_f)
+           break;
+
+        fndecl = builtin_decl_implicit (BUILT_IN_SYNC_NAND_AND_FETCH_N);
+         inform (loc, "%qD changed semantics in GCC 4.4", fndecl);
+         warned_n_a_f = true;
+         break;
+
+       default:
+         gcc_unreachable ();
+       }
+    }
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 1), mode);
+
+  return expand_atomic_fetch_op (target, mem, val, code, MEMMODEL_SEQ_CST,
+                                after);
+}
+
+/* Expand the __sync_val_compare_and_swap and __sync_bool_compare_and_swap
+   intrinsics. EXP is the CALL_EXPR.  IS_BOOL is
+   true if this is the boolean form.  TARGET is a place for us to store the
+   results; this is NOT optional if IS_BOOL is true.  */
+
+static rtx
+expand_builtin_compare_and_swap (enum machine_mode mode, tree exp,
+                                bool is_bool, rtx target)
+{
+  rtx old_val, new_val, mem;
+  rtx *pbool, *poval;
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  old_val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 1), mode);
+  new_val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 2), mode);
+
+  pbool = poval = NULL;
+  if (target != const0_rtx)
+    {
+      if (is_bool)
+       pbool = &target;
+      else
+       poval = &target;
+    }
+  if (!expand_atomic_compare_and_swap (pbool, poval, mem, old_val, new_val,
+                                      false, MEMMODEL_SEQ_CST,
+                                      MEMMODEL_SEQ_CST))
+    return NULL_RTX;
+
+  return target;
+}
+
+/* Expand the __sync_lock_test_and_set intrinsic.  Note that the most
+   general form is actually an atomic exchange, and some targets only
+   support a reduced form with the second argument being a constant 1.
+   EXP is the CALL_EXPR; TARGET is an optional place for us to store
+   the results.  */
+
+static rtx
+expand_builtin_sync_lock_test_and_set (enum machine_mode mode, tree exp,
+                                      rtx target)
+{
+  rtx val, mem;
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 1), mode);
+
+  return expand_sync_lock_test_and_set (target, mem, val);
+}
+
+/* Expand the __sync_lock_release intrinsic.  EXP is the CALL_EXPR.  */
+
+static void
+expand_builtin_sync_lock_release (enum machine_mode mode, tree exp)
+{
+  rtx mem;
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+
+  expand_atomic_store (mem, const0_rtx, MEMMODEL_RELEASE, true);
+}
+
+/* Given an integer representing an ``enum memmodel'', verify its
+   correctness and return the memory model enum.  */
+
+static enum memmodel
+get_memmodel (tree exp)
+{
+  rtx op;
+
+  /* If the parameter is not a constant, it's a run time value so we'll just
+     convert it to MEMMODEL_SEQ_CST to avoid annoying runtime checking.  */
+  if (TREE_CODE (exp) != INTEGER_CST)
+    return MEMMODEL_SEQ_CST;
+
+  op = expand_normal (exp);
+  if (INTVAL (op) < 0 || INTVAL (op) >= MEMMODEL_LAST)
+    {
+      warning (OPT_Winvalid_memory_model,
+              "invalid memory model argument to builtin");
+      return MEMMODEL_SEQ_CST;
+    }
+  return (enum memmodel) INTVAL (op);
+}
+
+/* Expand the __atomic_exchange intrinsic:
+       TYPE __atomic_exchange (TYPE *object, TYPE desired, enum memmodel)
+   EXP is the CALL_EXPR.
+   TARGET is an optional place for us to store the results.  */
+
+static rtx
+expand_builtin_atomic_exchange (enum machine_mode mode, tree exp, rtx target)
+{
+  rtx val, mem;
+  enum memmodel model;
+
+  model = get_memmodel (CALL_EXPR_ARG (exp, 2));
+  if (model == MEMMODEL_CONSUME)
+    {
+      error ("invalid memory model for %<__atomic_exchange%>");
+      return NULL_RTX;
+    }
+
+  if (!flag_inline_atomics)
+    return NULL_RTX;
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 1), mode);
+
+  return expand_atomic_exchange (target, mem, val, model);
+}
+
+/* Expand the __atomic_compare_exchange intrinsic:
+       bool __atomic_compare_exchange (TYPE *object, TYPE *expect, 
+                                       TYPE desired, BOOL weak, 
+                                       enum memmodel success,
+                                       enum memmodel failure)
+   EXP is the CALL_EXPR.
+   TARGET is an optional place for us to store the results.  */
+
+static rtx
+expand_builtin_atomic_compare_exchange (enum machine_mode mode, tree exp, 
+                                       rtx target)
+{
+  rtx expect, desired, mem, oldval;
+  enum memmodel success, failure;
+  tree weak;
+  bool is_weak;
+
+  success = get_memmodel (CALL_EXPR_ARG (exp, 4));
+  failure = get_memmodel (CALL_EXPR_ARG (exp, 5));
+
+  if (failure == MEMMODEL_RELEASE || failure == MEMMODEL_ACQ_REL)
+    {
+      error ("invalid failure memory model for %<__atomic_compare_exchange%>");
+      return NULL_RTX;
+    }
+
+  if (failure > success)
+    {
+      error ("failure memory model cannot be stronger than success "
+            "memory model for %<__atomic_compare_exchange%>");
+      return NULL_RTX;
+    }
+  
+  if (!flag_inline_atomics)
+    return NULL_RTX;
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+
+  expect = expand_normal (CALL_EXPR_ARG (exp, 1));
+  expect = convert_memory_address (Pmode, expect);
+  desired = expand_expr_force_mode (CALL_EXPR_ARG (exp, 2), mode);
+
+  weak = CALL_EXPR_ARG (exp, 3);
+  is_weak = false;
+  if (host_integerp (weak, 0) && tree_low_cst (weak, 0) != 0)
+    is_weak = true;
+
+  oldval = copy_to_reg (gen_rtx_MEM (mode, expect));
+
+  if (!expand_atomic_compare_and_swap ((target == const0_rtx ? NULL : &target),
+                                      &oldval, mem, oldval, desired,
+                                      is_weak, success, failure))
+    return NULL_RTX;
+
+  emit_move_insn (gen_rtx_MEM (mode, expect), oldval);
+  return target;
+}
+
+/* Expand the __atomic_load intrinsic:
+       TYPE __atomic_load (TYPE *object, enum memmodel)
+   EXP is the CALL_EXPR.
+   TARGET is an optional place for us to store the results.  */
+
+static rtx
+expand_builtin_atomic_load (enum machine_mode mode, tree exp, rtx target)
+{
+  rtx mem;
+  enum memmodel model;
+
+  model = get_memmodel (CALL_EXPR_ARG (exp, 1));
+  if (model == MEMMODEL_RELEASE
+      || model == MEMMODEL_ACQ_REL)
+    {
+      error ("invalid memory model for %<__atomic_load%>");
+      return NULL_RTX;
+    }
+
+  if (!flag_inline_atomics)
+    return NULL_RTX;
+
+  /* Expand the operand.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+
+  return expand_atomic_load (target, mem, model);
+}
+
+
+/* Expand the __atomic_store intrinsic:
+       void __atomic_store (TYPE *object, TYPE desired, enum memmodel)
+   EXP is the CALL_EXPR.
+   TARGET is an optional place for us to store the results.  */
+
+static rtx
+expand_builtin_atomic_store (enum machine_mode mode, tree exp)
+{
+  rtx mem, val;
+  enum memmodel model;
+
+  model = get_memmodel (CALL_EXPR_ARG (exp, 2));
+  if (model != MEMMODEL_RELAXED
+      && model != MEMMODEL_SEQ_CST
+      && model != MEMMODEL_RELEASE)
+    {
+      error ("invalid memory model for %<__atomic_store%>");
+      return NULL_RTX;
+    }
+
+  if (!flag_inline_atomics)
+    return NULL_RTX;
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 1), mode);
+
+  return expand_atomic_store (mem, val, model, false);
+}
+
+/* Expand the __atomic_fetch_XXX intrinsic:
+       TYPE __atomic_fetch_XXX (TYPE *object, TYPE val, enum memmodel)
+   EXP is the CALL_EXPR.
+   TARGET is an optional place for us to store the results.
+   CODE is the operation, PLUS, MINUS, ADD, XOR, or IOR.
+   FETCH_AFTER is true if returning the result of the operation.
+   FETCH_AFTER is false if returning the value before the operation.
+   IGNORE is true if the result is not used.
+   EXT_CALL is the correct builtin for an external call if this cannot be
+   resolved to an instruction sequence.  */
+
+static rtx
+expand_builtin_atomic_fetch_op (enum machine_mode mode, tree exp, rtx target,
+                               enum rtx_code code, bool fetch_after,
+                               bool ignore, enum built_in_function ext_call)
+{
+  rtx val, mem, ret;
+  enum memmodel model;
+  tree fndecl;
+  tree addr;
+
+  model = get_memmodel (CALL_EXPR_ARG (exp, 2));
+
+  /* Expand the operands.  */
+  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  val = expand_expr_force_mode (CALL_EXPR_ARG (exp, 1), mode);
+
+  /* Only try generating instructions if inlining is turned on.  */
+  if (flag_inline_atomics)
+    {
+      ret = expand_atomic_fetch_op (target, mem, val, code, model, fetch_after);
+      if (ret)
+       return ret;
+    }
+
+  /* Return if a different routine isn't needed for the library call.  */
+  if (ext_call == BUILT_IN_NONE)
+    return NULL_RTX;
+
+  /* Change the call to the specified function.  */
+  fndecl = get_callee_fndecl (exp);
+  addr = CALL_EXPR_FN (exp);
+  STRIP_NOPS (addr);
+
+  gcc_assert (TREE_OPERAND (addr, 0) == fndecl);
+  TREE_OPERAND (addr, 0) = builtin_decl_explicit(ext_call);
+
+  /* Expand the call here so we can emit trailing code.  */
+  ret = expand_call (exp, target, ignore);
+
+  /* Replace the original function just in case it matters.  */
+  TREE_OPERAND (addr, 0) = fndecl;
+
+  /* Then issue the arithmetic correction to return the right result.  */
+  if (!ignore)
+    {
+      if (code == NOT)
+       {
+         ret = expand_simple_binop (mode, AND, ret, val, NULL_RTX, true,
+                                    OPTAB_LIB_WIDEN);
+         ret = expand_simple_unop (mode, NOT, ret, target, true);
+       }
+      else
+       ret = expand_simple_binop (mode, code, ret, val, target, true,
+                                  OPTAB_LIB_WIDEN);
+    }
+  return ret;
+}
+
 
-         if (warned_n_a_f)
-           break;
+#ifndef HAVE_atomic_clear
+# define HAVE_atomic_clear 0
+# define gen_atomic_clear(x,y) (gcc_unreachable (), NULL_RTX)
+#endif
 
-         fndecl = implicit_built_in_decls[BUILT_IN_NAND_AND_FETCH_N];
-         inform (loc, "%qD changed semantics in GCC 4.4", fndecl);
-         warned_n_a_f = true;
-         break;
+/* Expand an atomic clear operation.
+       void _atomic_clear (BOOL *obj, enum memmodel)
+   EXP is the call expression.  */
 
-       default:
-         gcc_unreachable ();
-       }
-    }
+static rtx
+expand_builtin_atomic_clear (tree exp) 
+{
+  enum machine_mode mode;
+  rtx mem, ret;
+  enum memmodel model;
 
-  /* Expand the operands.  */
+  mode = mode_for_size (BOOL_TYPE_SIZE, MODE_INT, 0);
   mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  model = get_memmodel (CALL_EXPR_ARG (exp, 1));
 
-  val = expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX, mode, EXPAND_NORMAL);
-  /* If VAL is promoted to a wider mode, convert it back to MODE.  Take care
-     of CONST_INTs, where we know the old_mode only from the call argument.  */
-  old_mode = GET_MODE (val);
-  if (old_mode == VOIDmode)
-    old_mode = TYPE_MODE (TREE_TYPE (CALL_EXPR_ARG (exp, 1)));
-  val = convert_modes (mode, old_mode, val, 1);
+  if (model == MEMMODEL_ACQUIRE || model == MEMMODEL_ACQ_REL)
+    {
+      error ("invalid memory model for %<__atomic_store%>");
+      return const0_rtx;
+    }
 
-  if (ignore)
-    return expand_sync_operation (mem, val, code);
-  else
-    return expand_sync_fetch_operation (mem, val, code, after, target);
+  if (HAVE_atomic_clear)
+    {
+      emit_insn (gen_atomic_clear (mem, model));
+      return const0_rtx;
+    }
+
+  /* Try issuing an __atomic_store, and allow fallback to __sync_lock_release.
+     Failing that, a store is issued by __atomic_store.  The only way this can
+     fail is if the bool type is larger than a word size.  Unlikely, but
+     handle it anyway for completeness.  Assume a single threaded model since
+     there is no atomic support in this case, and no barriers are required.  */
+  ret = expand_atomic_store (mem, const0_rtx, model, true);
+  if (!ret)
+    emit_move_insn (mem, const0_rtx);
+  return const0_rtx;
 }
 
-/* Expand the __sync_val_compare_and_swap and __sync_bool_compare_and_swap
-   intrinsics. EXP is the CALL_EXPR.  IS_BOOL is
-   true if this is the boolean form.  TARGET is a place for us to store the
-   results; this is NOT optional if IS_BOOL is true.  */
+/* Expand an atomic test_and_set operation.
+       bool _atomic_test_and_set (BOOL *obj, enum memmodel)
+   EXP is the call expression.  */
 
 static rtx
-expand_builtin_compare_and_swap (enum machine_mode mode, tree exp,
-                                bool is_bool, rtx target)
+expand_builtin_atomic_test_and_set (tree exp, rtx target)
 {
-  rtx old_val, new_val, mem;
-  enum machine_mode old_mode;
+  rtx mem;
+  enum memmodel model;
+  enum machine_mode mode;
 
-  /* Expand the operands.  */
+  mode = mode_for_size (BOOL_TYPE_SIZE, MODE_INT, 0);
   mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+  model = get_memmodel (CALL_EXPR_ARG (exp, 1));
 
+  return expand_atomic_test_and_set (target, mem, model);
+}
 
-  old_val = expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX,
-                        mode, EXPAND_NORMAL);
-  /* If VAL is promoted to a wider mode, convert it back to MODE.  Take care
-     of CONST_INTs, where we know the old_mode only from the call argument.  */
-  old_mode = GET_MODE (old_val);
-  if (old_mode == VOIDmode)
-    old_mode = TYPE_MODE (TREE_TYPE (CALL_EXPR_ARG (exp, 1)));
-  old_val = convert_modes (mode, old_mode, old_val, 1);
 
-  new_val = expand_expr (CALL_EXPR_ARG (exp, 2), NULL_RTX,
-                        mode, EXPAND_NORMAL);
-  /* If VAL is promoted to a wider mode, convert it back to MODE.  Take care
-     of CONST_INTs, where we know the old_mode only from the call argument.  */
-  old_mode = GET_MODE (new_val);
-  if (old_mode == VOIDmode)
-    old_mode = TYPE_MODE (TREE_TYPE (CALL_EXPR_ARG (exp, 2)));
-  new_val = convert_modes (mode, old_mode, new_val, 1);
+/* Return true if (optional) argument ARG1 of size ARG0 is always lock free on
+   this architecture.  If ARG1 is NULL, use typical alignment for size ARG0.  */
+
+static tree
+fold_builtin_atomic_always_lock_free (tree arg0, tree arg1)
+{
+  int size;
+  enum machine_mode mode;
+  unsigned int mode_align, type_align;
+
+  if (TREE_CODE (arg0) != INTEGER_CST)
+    return NULL_TREE;
+
+  size = INTVAL (expand_normal (arg0)) * BITS_PER_UNIT;
+  mode = mode_for_size (size, MODE_INT, 0);
+  mode_align = GET_MODE_ALIGNMENT (mode);
+
+  if (TREE_CODE (arg1) == INTEGER_CST && INTVAL (expand_normal (arg1)) == 0)
+    type_align = mode_align;
+  else
+    {
+      tree ttype = TREE_TYPE (arg1);
+
+      /* This function is usually invoked and folded immediately by the front
+        end before anything else has a chance to look at it.  The pointer
+        parameter at this point is usually cast to a void *, so check for that
+        and look past the cast.  */
+      if (TREE_CODE (arg1) == NOP_EXPR && POINTER_TYPE_P (ttype)
+         && VOID_TYPE_P (TREE_TYPE (ttype)))
+       arg1 = TREE_OPERAND (arg1, 0);
+
+      ttype = TREE_TYPE (arg1);
+      gcc_assert (POINTER_TYPE_P (ttype));
+
+      /* Get the underlying type of the object.  */
+      ttype = TREE_TYPE (ttype);
+      type_align = TYPE_ALIGN (ttype);
+    }
+
+  /* If the object has smaller alignment, the the lock free routines cannot
+     be used.  */
+  if (type_align < mode_align)
+    return boolean_false_node;
 
-  if (is_bool)
-    return expand_bool_compare_and_swap (mem, old_val, new_val, target);
+  /* Check if a compare_and_swap pattern exists for the mode which represents
+     the required size.  The pattern is not allowed to fail, so the existence
+     of the pattern indicates support is present.  */
+  if (can_compare_and_swap_p (mode, true))
+    return boolean_true_node;
   else
-    return expand_val_compare_and_swap (mem, old_val, new_val, target);
+    return boolean_false_node;
 }
 
-/* Expand the __sync_lock_test_and_set intrinsic.  Note that the most
-   general form is actually an atomic exchange, and some targets only
-   support a reduced form with the second argument being a constant 1.
-   EXP is the CALL_EXPR; TARGET is an optional place for us to store
-   the results.  */
+/* Return true if the parameters to call EXP represent an object which will
+   always generate lock free instructions.  The first argument represents the
+   size of the object, and the second parameter is a pointer to the object 
+   itself.  If NULL is passed for the object, then the result is based on 
+   typical alignment for an object of the specified size.  Otherwise return 
+   false.  */
 
 static rtx
-expand_builtin_lock_test_and_set (enum machine_mode mode, tree exp,
-                                 rtx target)
+expand_builtin_atomic_always_lock_free (tree exp)
 {
-  rtx val, mem;
-  enum machine_mode old_mode;
+  tree size;
+  tree arg0 = CALL_EXPR_ARG (exp, 0);
+  tree arg1 = CALL_EXPR_ARG (exp, 1);
 
-  /* Expand the operands.  */
-  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
-  val = expand_expr (CALL_EXPR_ARG (exp, 1), NULL_RTX, mode, EXPAND_NORMAL);
-  /* If VAL is promoted to a wider mode, convert it back to MODE.  Take care
-     of CONST_INTs, where we know the old_mode only from the call argument.  */
-  old_mode = GET_MODE (val);
-  if (old_mode == VOIDmode)
-    old_mode = TYPE_MODE (TREE_TYPE (CALL_EXPR_ARG (exp, 1)));
-  val = convert_modes (mode, old_mode, val, 1);
+  if (TREE_CODE (arg0) != INTEGER_CST)
+    {
+      error ("non-constant argument 1 to __atomic_always_lock_free");
+      return const0_rtx;
+    }
 
-  return expand_sync_lock_test_and_set (mem, val, target);
+  size = fold_builtin_atomic_always_lock_free (arg0, arg1);
+  if (size == boolean_true_node)
+    return const1_rtx;
+  return const0_rtx;
 }
 
-/* Expand the __sync_synchronize intrinsic.  */
+/* Return a one or zero if it can be determined that object ARG1 of size ARG 
+   is lock free on this architecture.  */
 
-static void
-expand_builtin_synchronize (void)
+static tree
+fold_builtin_atomic_is_lock_free (tree arg0, tree arg1)
 {
-  gimple x;
-  VEC (tree, gc) *v_clobbers;
+  if (!flag_inline_atomics)
+    return NULL_TREE;
+  
+  /* If it isn't always lock free, don't generate a result.  */
+  if (fold_builtin_atomic_always_lock_free (arg0, arg1) == boolean_true_node)
+    return boolean_true_node;
 
-#ifdef HAVE_memory_barrier
-  if (HAVE_memory_barrier)
-    {
-      emit_insn (gen_memory_barrier ());
-      return;
-    }
-#endif
+  return NULL_TREE;
+}
+
+/* Return true if the parameters to call EXP represent an object which will
+   always generate lock free instructions.  The first argument represents the
+   size of the object, and the second parameter is a pointer to the object 
+   itself.  If NULL is passed for the object, then the result is based on 
+   typical alignment for an object of the specified size.  Otherwise return 
+   NULL*/
 
-  if (synchronize_libfunc != NULL_RTX)
+static rtx
+expand_builtin_atomic_is_lock_free (tree exp)
+{
+  tree size;
+  tree arg0 = CALL_EXPR_ARG (exp, 0);
+  tree arg1 = CALL_EXPR_ARG (exp, 1);
+
+  if (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)))
     {
-      emit_library_call (synchronize_libfunc, LCT_NORMAL, VOIDmode, 0);
-      return;
+      error ("non-integer argument 1 to __atomic_is_lock_free");
+      return NULL_RTX;
     }
 
-  /* If no explicit memory barrier instruction is available, create an
-     empty asm stmt with a memory clobber.  */
-  v_clobbers = VEC_alloc (tree, gc, 1);
-  VEC_quick_push (tree, v_clobbers,
-                 tree_cons (NULL, build_string (6, "memory"), NULL));
-  x = gimple_build_asm_vec ("", NULL, NULL, v_clobbers, NULL);
-  gimple_asm_set_volatile (x, true);
-  expand_asm_stmt (x);
+  if (!flag_inline_atomics)
+    return NULL_RTX; 
+
+  /* If the value is known at compile time, return the RTX for it.  */
+  size = fold_builtin_atomic_is_lock_free (arg0, arg1);
+  if (size == boolean_true_node)
+    return const1_rtx;
+
+  return NULL_RTX;
 }
 
-/* Expand the __sync_lock_release intrinsic.  EXP is the CALL_EXPR.  */
+/* Expand the __atomic_thread_fence intrinsic:
+       void __atomic_thread_fence (enum memmodel)
+   EXP is the CALL_EXPR.  */
 
 static void
-expand_builtin_lock_release (enum machine_mode mode, tree exp)
+expand_builtin_atomic_thread_fence (tree exp)
 {
-  struct expand_operand ops[2];
-  enum insn_code icode;
-  rtx mem;
+  enum memmodel model = get_memmodel (CALL_EXPR_ARG (exp, 0));
+  expand_mem_thread_fence (model);
+}
 
-  /* Expand the operands.  */
-  mem = get_builtin_sync_mem (CALL_EXPR_ARG (exp, 0), mode);
+/* Expand the __atomic_signal_fence intrinsic:
+       void __atomic_signal_fence (enum memmodel)
+   EXP is the CALL_EXPR.  */
 
-  /* If there is an explicit operation in the md file, use it.  */
-  icode = direct_optab_handler (sync_lock_release_optab, mode);
-  if (icode != CODE_FOR_nothing)
-    {
-      create_fixed_operand (&ops[0], mem);
-      create_input_operand (&ops[1], const0_rtx, mode);
-      if (maybe_expand_insn (icode, 2, ops))
-       return;
-    }
+static void
+expand_builtin_atomic_signal_fence (tree exp)
+{
+  enum memmodel model = get_memmodel (CALL_EXPR_ARG (exp, 0));
+  expand_mem_signal_fence (model);
+}
+
+/* Expand the __sync_synchronize intrinsic.  */
 
-  /* Otherwise we can implement this operation by emitting a barrier
-     followed by a store of zero.  */
-  expand_builtin_synchronize ();
-  emit_move_insn (mem, const0_rtx);
+static void
+expand_builtin_sync_synchronize (void)
+{
+  expand_mem_thread_fence (MEMMODEL_SEQ_CST);
 }
+
 \f
 /* Expand an expression EXP that calls a built-in function,
    with result going to TARGET if that's convenient
@@ -5274,8 +5805,8 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
      set of builtins.  */
   if (!optimize
       && !called_as_built_in (fndecl)
-      && DECL_ASSEMBLER_NAME_SET_P (fndecl)
       && fcode != BUILT_IN_ALLOCA
+      && fcode != BUILT_IN_ALLOCA_WITH_ALIGN
       && fcode != BUILT_IN_FREE)
     return expand_call (exp, target, ignore);
 
@@ -5378,17 +5909,21 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
        return target;
       break;
 
+    CASE_FLT_FN (BUILT_IN_ICEIL):
     CASE_FLT_FN (BUILT_IN_LCEIL):
     CASE_FLT_FN (BUILT_IN_LLCEIL):
     CASE_FLT_FN (BUILT_IN_LFLOOR):
+    CASE_FLT_FN (BUILT_IN_IFLOOR):
     CASE_FLT_FN (BUILT_IN_LLFLOOR):
       target = expand_builtin_int_roundingfn (exp, target);
       if (target)
        return target;
       break;
 
+    CASE_FLT_FN (BUILT_IN_IRINT):
     CASE_FLT_FN (BUILT_IN_LRINT):
     CASE_FLT_FN (BUILT_IN_LLRINT):
+    CASE_FLT_FN (BUILT_IN_IROUND):
     CASE_FLT_FN (BUILT_IN_LROUND):
     CASE_FLT_FN (BUILT_IN_LLROUND):
       target = expand_builtin_int_roundingfn_2 (exp, target);
@@ -5527,6 +6062,7 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
        return XEXP (DECL_RTL (DECL_RESULT (current_function_decl)), 0);
 
     case BUILT_IN_ALLOCA:
+    case BUILT_IN_ALLOCA_WITH_ALIGN:
       /* If the allocation stems from the declaration of a variable-sized
         object, it cannot accumulate.  */
       target = expand_builtin_alloca (exp, CALL_ALLOCA_FOR_VAR_P (exp));
@@ -5573,6 +6109,14 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
        return target;
       break;
 
+    CASE_INT_FN (BUILT_IN_CLRSB):
+    case BUILT_IN_CLRSBIMAX:
+      target = expand_builtin_unop (target_mode, exp, target,
+                                   subtarget, clrsb_optab);
+      if (target)
+       return target;
+      break;
+
     CASE_INT_FN (BUILT_IN_POPCOUNT):
     case BUILT_IN_POPCOUNTIMAX:
       target = expand_builtin_unop (target_mode, exp, target,
@@ -5808,12 +6352,16 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
       return expand_builtin_va_copy (exp);
     case BUILT_IN_EXPECT:
       return expand_builtin_expect (exp, target);
+    case BUILT_IN_ASSUME_ALIGNED:
+      return expand_builtin_assume_aligned (exp, target);
     case BUILT_IN_PREFETCH:
       expand_builtin_prefetch (exp);
       return const0_rtx;
 
     case BUILT_IN_INIT_TRAMPOLINE:
-      return expand_builtin_init_trampoline (exp);
+      return expand_builtin_init_trampoline (exp, true);
+    case BUILT_IN_INIT_HEAP_TRAMPOLINE:
+      return expand_builtin_init_trampoline (exp, false);
     case BUILT_IN_ADJUST_TRAMPOLINE:
       return expand_builtin_adjust_trampoline (exp);
 
@@ -5829,199 +6377,441 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
        return target;
       break;
 
-    case BUILT_IN_FETCH_AND_ADD_1:
-    case BUILT_IN_FETCH_AND_ADD_2:
-    case BUILT_IN_FETCH_AND_ADD_4:
-    case BUILT_IN_FETCH_AND_ADD_8:
-    case BUILT_IN_FETCH_AND_ADD_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_FETCH_AND_ADD_1);
-      target = expand_builtin_sync_operation (mode, exp, PLUS,
-                                             false, target, ignore);
+    case BUILT_IN_SYNC_FETCH_AND_ADD_1:
+    case BUILT_IN_SYNC_FETCH_AND_ADD_2:
+    case BUILT_IN_SYNC_FETCH_AND_ADD_4:
+    case BUILT_IN_SYNC_FETCH_AND_ADD_8:
+    case BUILT_IN_SYNC_FETCH_AND_ADD_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_FETCH_AND_ADD_1);
+      target = expand_builtin_sync_operation (mode, exp, PLUS, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_FETCH_AND_SUB_1:
-    case BUILT_IN_FETCH_AND_SUB_2:
-    case BUILT_IN_FETCH_AND_SUB_4:
-    case BUILT_IN_FETCH_AND_SUB_8:
-    case BUILT_IN_FETCH_AND_SUB_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_FETCH_AND_SUB_1);
-      target = expand_builtin_sync_operation (mode, exp, MINUS,
-                                             false, target, ignore);
+    case BUILT_IN_SYNC_FETCH_AND_SUB_1:
+    case BUILT_IN_SYNC_FETCH_AND_SUB_2:
+    case BUILT_IN_SYNC_FETCH_AND_SUB_4:
+    case BUILT_IN_SYNC_FETCH_AND_SUB_8:
+    case BUILT_IN_SYNC_FETCH_AND_SUB_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_FETCH_AND_SUB_1);
+      target = expand_builtin_sync_operation (mode, exp, MINUS, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_FETCH_AND_OR_1:
-    case BUILT_IN_FETCH_AND_OR_2:
-    case BUILT_IN_FETCH_AND_OR_4:
-    case BUILT_IN_FETCH_AND_OR_8:
-    case BUILT_IN_FETCH_AND_OR_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_FETCH_AND_OR_1);
-      target = expand_builtin_sync_operation (mode, exp, IOR,
-                                             false, target, ignore);
+    case BUILT_IN_SYNC_FETCH_AND_OR_1:
+    case BUILT_IN_SYNC_FETCH_AND_OR_2:
+    case BUILT_IN_SYNC_FETCH_AND_OR_4:
+    case BUILT_IN_SYNC_FETCH_AND_OR_8:
+    case BUILT_IN_SYNC_FETCH_AND_OR_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_FETCH_AND_OR_1);
+      target = expand_builtin_sync_operation (mode, exp, IOR, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_FETCH_AND_AND_1:
-    case BUILT_IN_FETCH_AND_AND_2:
-    case BUILT_IN_FETCH_AND_AND_4:
-    case BUILT_IN_FETCH_AND_AND_8:
-    case BUILT_IN_FETCH_AND_AND_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_FETCH_AND_AND_1);
-      target = expand_builtin_sync_operation (mode, exp, AND,
-                                             false, target, ignore);
+    case BUILT_IN_SYNC_FETCH_AND_AND_1:
+    case BUILT_IN_SYNC_FETCH_AND_AND_2:
+    case BUILT_IN_SYNC_FETCH_AND_AND_4:
+    case BUILT_IN_SYNC_FETCH_AND_AND_8:
+    case BUILT_IN_SYNC_FETCH_AND_AND_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_FETCH_AND_AND_1);
+      target = expand_builtin_sync_operation (mode, exp, AND, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_FETCH_AND_XOR_1:
-    case BUILT_IN_FETCH_AND_XOR_2:
-    case BUILT_IN_FETCH_AND_XOR_4:
-    case BUILT_IN_FETCH_AND_XOR_8:
-    case BUILT_IN_FETCH_AND_XOR_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_FETCH_AND_XOR_1);
-      target = expand_builtin_sync_operation (mode, exp, XOR,
-                                             false, target, ignore);
+    case BUILT_IN_SYNC_FETCH_AND_XOR_1:
+    case BUILT_IN_SYNC_FETCH_AND_XOR_2:
+    case BUILT_IN_SYNC_FETCH_AND_XOR_4:
+    case BUILT_IN_SYNC_FETCH_AND_XOR_8:
+    case BUILT_IN_SYNC_FETCH_AND_XOR_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_FETCH_AND_XOR_1);
+      target = expand_builtin_sync_operation (mode, exp, XOR, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_FETCH_AND_NAND_1:
-    case BUILT_IN_FETCH_AND_NAND_2:
-    case BUILT_IN_FETCH_AND_NAND_4:
-    case BUILT_IN_FETCH_AND_NAND_8:
-    case BUILT_IN_FETCH_AND_NAND_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_FETCH_AND_NAND_1);
-      target = expand_builtin_sync_operation (mode, exp, NOT,
-                                             false, target, ignore);
+    case BUILT_IN_SYNC_FETCH_AND_NAND_1:
+    case BUILT_IN_SYNC_FETCH_AND_NAND_2:
+    case BUILT_IN_SYNC_FETCH_AND_NAND_4:
+    case BUILT_IN_SYNC_FETCH_AND_NAND_8:
+    case BUILT_IN_SYNC_FETCH_AND_NAND_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_FETCH_AND_NAND_1);
+      target = expand_builtin_sync_operation (mode, exp, NOT, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_ADD_AND_FETCH_1:
-    case BUILT_IN_ADD_AND_FETCH_2:
-    case BUILT_IN_ADD_AND_FETCH_4:
-    case BUILT_IN_ADD_AND_FETCH_8:
-    case BUILT_IN_ADD_AND_FETCH_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_ADD_AND_FETCH_1);
-      target = expand_builtin_sync_operation (mode, exp, PLUS,
-                                             true, target, ignore);
+    case BUILT_IN_SYNC_ADD_AND_FETCH_1:
+    case BUILT_IN_SYNC_ADD_AND_FETCH_2:
+    case BUILT_IN_SYNC_ADD_AND_FETCH_4:
+    case BUILT_IN_SYNC_ADD_AND_FETCH_8:
+    case BUILT_IN_SYNC_ADD_AND_FETCH_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_ADD_AND_FETCH_1);
+      target = expand_builtin_sync_operation (mode, exp, PLUS, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_SUB_AND_FETCH_1:
-    case BUILT_IN_SUB_AND_FETCH_2:
-    case BUILT_IN_SUB_AND_FETCH_4:
-    case BUILT_IN_SUB_AND_FETCH_8:
-    case BUILT_IN_SUB_AND_FETCH_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_SUB_AND_FETCH_1);
-      target = expand_builtin_sync_operation (mode, exp, MINUS,
-                                             true, target, ignore);
+    case BUILT_IN_SYNC_SUB_AND_FETCH_1:
+    case BUILT_IN_SYNC_SUB_AND_FETCH_2:
+    case BUILT_IN_SYNC_SUB_AND_FETCH_4:
+    case BUILT_IN_SYNC_SUB_AND_FETCH_8:
+    case BUILT_IN_SYNC_SUB_AND_FETCH_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_SUB_AND_FETCH_1);
+      target = expand_builtin_sync_operation (mode, exp, MINUS, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_OR_AND_FETCH_1:
-    case BUILT_IN_OR_AND_FETCH_2:
-    case BUILT_IN_OR_AND_FETCH_4:
-    case BUILT_IN_OR_AND_FETCH_8:
-    case BUILT_IN_OR_AND_FETCH_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_OR_AND_FETCH_1);
-      target = expand_builtin_sync_operation (mode, exp, IOR,
-                                             true, target, ignore);
+    case BUILT_IN_SYNC_OR_AND_FETCH_1:
+    case BUILT_IN_SYNC_OR_AND_FETCH_2:
+    case BUILT_IN_SYNC_OR_AND_FETCH_4:
+    case BUILT_IN_SYNC_OR_AND_FETCH_8:
+    case BUILT_IN_SYNC_OR_AND_FETCH_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_OR_AND_FETCH_1);
+      target = expand_builtin_sync_operation (mode, exp, IOR, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_AND_AND_FETCH_1:
-    case BUILT_IN_AND_AND_FETCH_2:
-    case BUILT_IN_AND_AND_FETCH_4:
-    case BUILT_IN_AND_AND_FETCH_8:
-    case BUILT_IN_AND_AND_FETCH_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_AND_AND_FETCH_1);
-      target = expand_builtin_sync_operation (mode, exp, AND,
-                                             true, target, ignore);
+    case BUILT_IN_SYNC_AND_AND_FETCH_1:
+    case BUILT_IN_SYNC_AND_AND_FETCH_2:
+    case BUILT_IN_SYNC_AND_AND_FETCH_4:
+    case BUILT_IN_SYNC_AND_AND_FETCH_8:
+    case BUILT_IN_SYNC_AND_AND_FETCH_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_AND_AND_FETCH_1);
+      target = expand_builtin_sync_operation (mode, exp, AND, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_XOR_AND_FETCH_1:
-    case BUILT_IN_XOR_AND_FETCH_2:
-    case BUILT_IN_XOR_AND_FETCH_4:
-    case BUILT_IN_XOR_AND_FETCH_8:
-    case BUILT_IN_XOR_AND_FETCH_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_XOR_AND_FETCH_1);
-      target = expand_builtin_sync_operation (mode, exp, XOR,
-                                             true, target, ignore);
+    case BUILT_IN_SYNC_XOR_AND_FETCH_1:
+    case BUILT_IN_SYNC_XOR_AND_FETCH_2:
+    case BUILT_IN_SYNC_XOR_AND_FETCH_4:
+    case BUILT_IN_SYNC_XOR_AND_FETCH_8:
+    case BUILT_IN_SYNC_XOR_AND_FETCH_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_XOR_AND_FETCH_1);
+      target = expand_builtin_sync_operation (mode, exp, XOR, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_NAND_AND_FETCH_1:
-    case BUILT_IN_NAND_AND_FETCH_2:
-    case BUILT_IN_NAND_AND_FETCH_4:
-    case BUILT_IN_NAND_AND_FETCH_8:
-    case BUILT_IN_NAND_AND_FETCH_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_NAND_AND_FETCH_1);
-      target = expand_builtin_sync_operation (mode, exp, NOT,
-                                             true, target, ignore);
+    case BUILT_IN_SYNC_NAND_AND_FETCH_1:
+    case BUILT_IN_SYNC_NAND_AND_FETCH_2:
+    case BUILT_IN_SYNC_NAND_AND_FETCH_4:
+    case BUILT_IN_SYNC_NAND_AND_FETCH_8:
+    case BUILT_IN_SYNC_NAND_AND_FETCH_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_NAND_AND_FETCH_1);
+      target = expand_builtin_sync_operation (mode, exp, NOT, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_BOOL_COMPARE_AND_SWAP_1:
-    case BUILT_IN_BOOL_COMPARE_AND_SWAP_2:
-    case BUILT_IN_BOOL_COMPARE_AND_SWAP_4:
-    case BUILT_IN_BOOL_COMPARE_AND_SWAP_8:
-    case BUILT_IN_BOOL_COMPARE_AND_SWAP_16:
+    case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1:
+    case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_2:
+    case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_4:
+    case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_8:
+    case BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_16:
       if (mode == VOIDmode)
        mode = TYPE_MODE (boolean_type_node);
       if (!target || !register_operand (target, mode))
        target = gen_reg_rtx (mode);
 
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_BOOL_COMPARE_AND_SWAP_1);
+      mode = get_builtin_sync_mode 
+                               (fcode - BUILT_IN_SYNC_BOOL_COMPARE_AND_SWAP_1);
       target = expand_builtin_compare_and_swap (mode, exp, true, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_VAL_COMPARE_AND_SWAP_1:
-    case BUILT_IN_VAL_COMPARE_AND_SWAP_2:
-    case BUILT_IN_VAL_COMPARE_AND_SWAP_4:
-    case BUILT_IN_VAL_COMPARE_AND_SWAP_8:
-    case BUILT_IN_VAL_COMPARE_AND_SWAP_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_VAL_COMPARE_AND_SWAP_1);
+    case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_1:
+    case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_2:
+    case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_4:
+    case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_8:
+    case BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_16:
+      mode = get_builtin_sync_mode 
+                               (fcode - BUILT_IN_SYNC_VAL_COMPARE_AND_SWAP_1);
       target = expand_builtin_compare_and_swap (mode, exp, false, target);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_LOCK_TEST_AND_SET_1:
-    case BUILT_IN_LOCK_TEST_AND_SET_2:
-    case BUILT_IN_LOCK_TEST_AND_SET_4:
-    case BUILT_IN_LOCK_TEST_AND_SET_8:
-    case BUILT_IN_LOCK_TEST_AND_SET_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_LOCK_TEST_AND_SET_1);
-      target = expand_builtin_lock_test_and_set (mode, exp, target);
+    case BUILT_IN_SYNC_LOCK_TEST_AND_SET_1:
+    case BUILT_IN_SYNC_LOCK_TEST_AND_SET_2:
+    case BUILT_IN_SYNC_LOCK_TEST_AND_SET_4:
+    case BUILT_IN_SYNC_LOCK_TEST_AND_SET_8:
+    case BUILT_IN_SYNC_LOCK_TEST_AND_SET_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_LOCK_TEST_AND_SET_1);
+      target = expand_builtin_sync_lock_test_and_set (mode, exp, target);
+      if (target)
+       return target;
+      break;
+
+    case BUILT_IN_SYNC_LOCK_RELEASE_1:
+    case BUILT_IN_SYNC_LOCK_RELEASE_2:
+    case BUILT_IN_SYNC_LOCK_RELEASE_4:
+    case BUILT_IN_SYNC_LOCK_RELEASE_8:
+    case BUILT_IN_SYNC_LOCK_RELEASE_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_SYNC_LOCK_RELEASE_1);
+      expand_builtin_sync_lock_release (mode, exp);
+      return const0_rtx;
+
+    case BUILT_IN_SYNC_SYNCHRONIZE:
+      expand_builtin_sync_synchronize ();
+      return const0_rtx;
+
+    case BUILT_IN_ATOMIC_EXCHANGE_1:
+    case BUILT_IN_ATOMIC_EXCHANGE_2:
+    case BUILT_IN_ATOMIC_EXCHANGE_4:
+    case BUILT_IN_ATOMIC_EXCHANGE_8:
+    case BUILT_IN_ATOMIC_EXCHANGE_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_EXCHANGE_1);
+      target = expand_builtin_atomic_exchange (mode, exp, target);
+      if (target)
+       return target;
+      break;
+
+    case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1:
+    case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2:
+    case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4:
+    case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8:
+    case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16:
+      {
+       unsigned int nargs, z;
+       VEC(tree,gc) *vec;
+
+       mode = 
+           get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1);
+       target = expand_builtin_atomic_compare_exchange (mode, exp, target);
+       if (target)
+         return target;
+
+       /* If this is turned into an external library call, the weak parameter
+          must be dropped to match the expected parameter list.  */
+       nargs = call_expr_nargs (exp);
+       vec = VEC_alloc (tree, gc, nargs - 1);
+       for (z = 0; z < 3; z++)
+         VEC_quick_push (tree, vec, CALL_EXPR_ARG (exp, z));
+       /* Skip the boolean weak parameter.  */
+       for (z = 4; z < 6; z++)
+         VEC_quick_push (tree, vec, CALL_EXPR_ARG (exp, z));
+       exp = build_call_vec (TREE_TYPE (exp), CALL_EXPR_FN (exp), vec);
+       break;
+      }
+
+    case BUILT_IN_ATOMIC_LOAD_1:
+    case BUILT_IN_ATOMIC_LOAD_2:
+    case BUILT_IN_ATOMIC_LOAD_4:
+    case BUILT_IN_ATOMIC_LOAD_8:
+    case BUILT_IN_ATOMIC_LOAD_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_LOAD_1);
+      target = expand_builtin_atomic_load (mode, exp, target);
+      if (target)
+       return target;
+      break;
+
+    case BUILT_IN_ATOMIC_STORE_1:
+    case BUILT_IN_ATOMIC_STORE_2:
+    case BUILT_IN_ATOMIC_STORE_4:
+    case BUILT_IN_ATOMIC_STORE_8:
+    case BUILT_IN_ATOMIC_STORE_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_STORE_1);
+      target = expand_builtin_atomic_store (mode, exp);
+      if (target)
+       return const0_rtx;
+      break;
+
+    case BUILT_IN_ATOMIC_ADD_FETCH_1:
+    case BUILT_IN_ATOMIC_ADD_FETCH_2:
+    case BUILT_IN_ATOMIC_ADD_FETCH_4:
+    case BUILT_IN_ATOMIC_ADD_FETCH_8:
+    case BUILT_IN_ATOMIC_ADD_FETCH_16:
+      {
+       enum built_in_function lib;
+       mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_ADD_FETCH_1);
+       lib = (enum built_in_function)((int)BUILT_IN_ATOMIC_FETCH_ADD_1 + 
+                                      (fcode - BUILT_IN_ATOMIC_ADD_FETCH_1));
+       target = expand_builtin_atomic_fetch_op (mode, exp, target, PLUS, true,
+                                                ignore, lib);
+       if (target)
+         return target;
+       break;
+      }
+    case BUILT_IN_ATOMIC_SUB_FETCH_1:
+    case BUILT_IN_ATOMIC_SUB_FETCH_2:
+    case BUILT_IN_ATOMIC_SUB_FETCH_4:
+    case BUILT_IN_ATOMIC_SUB_FETCH_8:
+    case BUILT_IN_ATOMIC_SUB_FETCH_16:
+      {
+       enum built_in_function lib;
+       mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_SUB_FETCH_1);
+       lib = (enum built_in_function)((int)BUILT_IN_ATOMIC_FETCH_SUB_1 + 
+                                      (fcode - BUILT_IN_ATOMIC_SUB_FETCH_1));
+       target = expand_builtin_atomic_fetch_op (mode, exp, target, MINUS, true,
+                                                ignore, lib);
+       if (target)
+         return target;
+       break;
+      }
+    case BUILT_IN_ATOMIC_AND_FETCH_1:
+    case BUILT_IN_ATOMIC_AND_FETCH_2:
+    case BUILT_IN_ATOMIC_AND_FETCH_4:
+    case BUILT_IN_ATOMIC_AND_FETCH_8:
+    case BUILT_IN_ATOMIC_AND_FETCH_16:
+      {
+       enum built_in_function lib;
+       mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_AND_FETCH_1);
+       lib = (enum built_in_function)((int)BUILT_IN_ATOMIC_FETCH_AND_1 + 
+                                      (fcode - BUILT_IN_ATOMIC_AND_FETCH_1));
+       target = expand_builtin_atomic_fetch_op (mode, exp, target, AND, true,
+                                                ignore, lib);
+       if (target)
+         return target;
+       break;
+      }
+    case BUILT_IN_ATOMIC_NAND_FETCH_1:
+    case BUILT_IN_ATOMIC_NAND_FETCH_2:
+    case BUILT_IN_ATOMIC_NAND_FETCH_4:
+    case BUILT_IN_ATOMIC_NAND_FETCH_8:
+    case BUILT_IN_ATOMIC_NAND_FETCH_16:
+      {
+       enum built_in_function lib;
+       mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_NAND_FETCH_1);
+       lib = (enum built_in_function)((int)BUILT_IN_ATOMIC_FETCH_NAND_1 + 
+                                      (fcode - BUILT_IN_ATOMIC_NAND_FETCH_1));
+       target = expand_builtin_atomic_fetch_op (mode, exp, target, NOT, true,
+                                                ignore, lib);
+       if (target)
+         return target;
+       break;
+      }
+    case BUILT_IN_ATOMIC_XOR_FETCH_1:
+    case BUILT_IN_ATOMIC_XOR_FETCH_2:
+    case BUILT_IN_ATOMIC_XOR_FETCH_4:
+    case BUILT_IN_ATOMIC_XOR_FETCH_8:
+    case BUILT_IN_ATOMIC_XOR_FETCH_16:
+      {
+       enum built_in_function lib;
+       mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_XOR_FETCH_1);
+       lib = (enum built_in_function)((int)BUILT_IN_ATOMIC_FETCH_XOR_1 + 
+                                      (fcode - BUILT_IN_ATOMIC_XOR_FETCH_1));
+       target = expand_builtin_atomic_fetch_op (mode, exp, target, XOR, true,
+                                                ignore, lib);
+       if (target)
+         return target;
+       break;
+      }
+    case BUILT_IN_ATOMIC_OR_FETCH_1:
+    case BUILT_IN_ATOMIC_OR_FETCH_2:
+    case BUILT_IN_ATOMIC_OR_FETCH_4:
+    case BUILT_IN_ATOMIC_OR_FETCH_8:
+    case BUILT_IN_ATOMIC_OR_FETCH_16:
+      {
+       enum built_in_function lib;
+       mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_OR_FETCH_1);
+       lib = (enum built_in_function)((int)BUILT_IN_ATOMIC_FETCH_OR_1 + 
+                                      (fcode - BUILT_IN_ATOMIC_OR_FETCH_1));
+       target = expand_builtin_atomic_fetch_op (mode, exp, target, IOR, true,
+                                                ignore, lib);
+       if (target)
+         return target;
+       break;
+      }
+    case BUILT_IN_ATOMIC_FETCH_ADD_1:
+    case BUILT_IN_ATOMIC_FETCH_ADD_2:
+    case BUILT_IN_ATOMIC_FETCH_ADD_4:
+    case BUILT_IN_ATOMIC_FETCH_ADD_8:
+    case BUILT_IN_ATOMIC_FETCH_ADD_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_FETCH_ADD_1);
+      target = expand_builtin_atomic_fetch_op (mode, exp, target, PLUS, false,
+                                              ignore, BUILT_IN_NONE);
+      if (target)
+       return target;
+      break;
+    case BUILT_IN_ATOMIC_FETCH_SUB_1:
+    case BUILT_IN_ATOMIC_FETCH_SUB_2:
+    case BUILT_IN_ATOMIC_FETCH_SUB_4:
+    case BUILT_IN_ATOMIC_FETCH_SUB_8:
+    case BUILT_IN_ATOMIC_FETCH_SUB_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_FETCH_SUB_1);
+      target = expand_builtin_atomic_fetch_op (mode, exp, target, MINUS, false,
+                                              ignore, BUILT_IN_NONE);
+      if (target)
+       return target;
+      break;
+
+    case BUILT_IN_ATOMIC_FETCH_AND_1:
+    case BUILT_IN_ATOMIC_FETCH_AND_2:
+    case BUILT_IN_ATOMIC_FETCH_AND_4:
+    case BUILT_IN_ATOMIC_FETCH_AND_8:
+    case BUILT_IN_ATOMIC_FETCH_AND_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_FETCH_AND_1);
+      target = expand_builtin_atomic_fetch_op (mode, exp, target, AND, false,
+                                              ignore, BUILT_IN_NONE);
+      if (target)
+       return target;
+      break;
+  
+    case BUILT_IN_ATOMIC_FETCH_NAND_1:
+    case BUILT_IN_ATOMIC_FETCH_NAND_2:
+    case BUILT_IN_ATOMIC_FETCH_NAND_4:
+    case BUILT_IN_ATOMIC_FETCH_NAND_8:
+    case BUILT_IN_ATOMIC_FETCH_NAND_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_FETCH_NAND_1);
+      target = expand_builtin_atomic_fetch_op (mode, exp, target, NOT, false,
+                                              ignore, BUILT_IN_NONE);
+      if (target)
+       return target;
+      break;
+    case BUILT_IN_ATOMIC_FETCH_XOR_1:
+    case BUILT_IN_ATOMIC_FETCH_XOR_2:
+    case BUILT_IN_ATOMIC_FETCH_XOR_4:
+    case BUILT_IN_ATOMIC_FETCH_XOR_8:
+    case BUILT_IN_ATOMIC_FETCH_XOR_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_FETCH_XOR_1);
+      target = expand_builtin_atomic_fetch_op (mode, exp, target, XOR, false,
+                                              ignore, BUILT_IN_NONE);
+      if (target)
+       return target;
+      break;
+    case BUILT_IN_ATOMIC_FETCH_OR_1:
+    case BUILT_IN_ATOMIC_FETCH_OR_2:
+    case BUILT_IN_ATOMIC_FETCH_OR_4:
+    case BUILT_IN_ATOMIC_FETCH_OR_8:
+    case BUILT_IN_ATOMIC_FETCH_OR_16:
+      mode = get_builtin_sync_mode (fcode - BUILT_IN_ATOMIC_FETCH_OR_1);
+      target = expand_builtin_atomic_fetch_op (mode, exp, target, IOR, false,
+                                              ignore, BUILT_IN_NONE);
       if (target)
        return target;
       break;
 
-    case BUILT_IN_LOCK_RELEASE_1:
-    case BUILT_IN_LOCK_RELEASE_2:
-    case BUILT_IN_LOCK_RELEASE_4:
-    case BUILT_IN_LOCK_RELEASE_8:
-    case BUILT_IN_LOCK_RELEASE_16:
-      mode = get_builtin_sync_mode (fcode - BUILT_IN_LOCK_RELEASE_1);
-      expand_builtin_lock_release (mode, exp);
+    case BUILT_IN_ATOMIC_TEST_AND_SET:
+      return expand_builtin_atomic_test_and_set (exp, target);
+
+    case BUILT_IN_ATOMIC_CLEAR:
+      return expand_builtin_atomic_clear (exp);
+    case BUILT_IN_ATOMIC_ALWAYS_LOCK_FREE:
+      return expand_builtin_atomic_always_lock_free (exp);
+
+    case BUILT_IN_ATOMIC_IS_LOCK_FREE:
+      target = expand_builtin_atomic_is_lock_free (exp);
+      if (target)
+        return target;
+      break;
+
+    case BUILT_IN_ATOMIC_THREAD_FENCE:
+      expand_builtin_atomic_thread_fence (exp);
       return const0_rtx;
 
-    case BUILT_IN_SYNCHRONIZE:
-      expand_builtin_synchronize ();
+    case BUILT_IN_ATOMIC_SIGNAL_FENCE:
+      expand_builtin_atomic_signal_fence (exp);
       return const0_rtx;
 
     case BUILT_IN_OBJECT_SIZE:
@@ -6039,6 +6829,7 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
     case BUILT_IN_STRCPY_CHK:
     case BUILT_IN_STPCPY_CHK:
     case BUILT_IN_STRNCPY_CHK:
+    case BUILT_IN_STPNCPY_CHK:
     case BUILT_IN_STRCAT_CHK:
     case BUILT_IN_STRNCAT_CHK:
     case BUILT_IN_SNPRINTF_CHK:
@@ -6052,7 +6843,8 @@ expand_builtin (tree exp, rtx target, rtx subtarget, enum machine_mode mode,
       break;
 
     case BUILT_IN_FREE:
-      maybe_emit_free_warning (exp);
+      if (warn_free_nonheap_object)
+       maybe_emit_free_warning (exp);
       break;
 
     default:   /* just do library call, if unknown builtin */
@@ -6187,7 +6979,7 @@ build_builtin_expect_predicate (location_t loc, tree pred, tree expected)
 {
   tree fn, arg_types, pred_type, expected_type, call_expr, ret_type;
 
-  fn = built_in_decls[BUILT_IN_EXPECT];
+  fn = builtin_decl_explicit (BUILT_IN_EXPECT);
   arg_types = TYPE_ARG_TYPES (TREE_TYPE (fn));
   ret_type = TREE_TYPE (TREE_TYPE (fn));
   pred_type = TREE_VALUE (arg_types);
@@ -6207,13 +6999,22 @@ build_builtin_expect_predicate (location_t loc, tree pred, tree expected)
 static tree
 fold_builtin_expect (location_t loc, tree arg0, tree arg1)
 {
-  tree inner, fndecl;
+  tree inner, fndecl, inner_arg0;
   enum tree_code code;
 
+  /* Distribute the expected value over short-circuiting operators.
+     See through the cast from truthvalue_type_node to long.  */
+  inner_arg0 = arg0;
+  while (TREE_CODE (inner_arg0) == NOP_EXPR
+        && INTEGRAL_TYPE_P (TREE_TYPE (inner_arg0))
+        && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (inner_arg0, 0))))
+    inner_arg0 = TREE_OPERAND (inner_arg0, 0);
+
   /* If this is a builtin_expect within a builtin_expect keep the
      inner one.  See through a comparison against a constant.  It
      might have been added to create a thruthvalue.  */
-  inner = arg0;
+  inner = inner_arg0;
+
   if (COMPARISON_CLASS_P (inner)
       && TREE_CODE (TREE_OPERAND (inner, 1)) == INTEGER_CST)
     inner = TREE_OPERAND (inner, 0);
@@ -6224,14 +7025,7 @@ fold_builtin_expect (location_t loc, tree arg0, tree arg1)
       && DECL_FUNCTION_CODE (fndecl) == BUILT_IN_EXPECT)
     return arg0;
 
-  /* Distribute the expected value over short-circuiting operators.
-     See through the cast from truthvalue_type_node to long.  */
-  inner = arg0;
-  while (TREE_CODE (inner) == NOP_EXPR
-        && INTEGRAL_TYPE_P (TREE_TYPE (inner))
-        && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (inner, 0))))
-    inner = TREE_OPERAND (inner, 0);
-
+  inner = inner_arg0;
   code = TREE_CODE (inner);
   if (code == TRUTH_ANDIF_EXPR || code == TRUTH_ORIF_EXPR)
     {
@@ -6246,13 +7040,13 @@ fold_builtin_expect (location_t loc, tree arg0, tree arg1)
     }
 
   /* If the argument isn't invariant then there's nothing else we can do.  */
-  if (!TREE_CONSTANT (arg0))
+  if (!TREE_CONSTANT (inner_arg0))
     return NULL_TREE;
 
   /* If we expect that a comparison against the argument will fold to
      a constant return the constant.  In practice, this means a true
      constant or the address of a non-weak symbol.  */
-  inner = arg0;
+  inner = inner_arg0;
   STRIP_NOPS (inner);
   if (TREE_CODE (inner) == ADDR_EXPR)
     {
@@ -6484,6 +7278,42 @@ fold_fixed_mathfn (location_t loc, tree fndecl, tree arg)
                                fold_convert_loc (loc, newtype, arg0));
     }
 
+  /* Canonicalize iround (x) to lround (x) on ILP32 targets where
+     sizeof (int) == sizeof (long).  */
+  if (TYPE_PRECISION (integer_type_node)
+      == TYPE_PRECISION (long_integer_type_node))
+    {
+      tree newfn = NULL_TREE;
+      switch (fcode)
+       {
+       CASE_FLT_FN (BUILT_IN_ICEIL):
+         newfn = mathfn_built_in (TREE_TYPE (arg), BUILT_IN_LCEIL);
+         break;
+
+       CASE_FLT_FN (BUILT_IN_IFLOOR):
+         newfn = mathfn_built_in (TREE_TYPE (arg), BUILT_IN_LFLOOR);
+         break;
+
+       CASE_FLT_FN (BUILT_IN_IROUND):
+         newfn = mathfn_built_in (TREE_TYPE (arg), BUILT_IN_LROUND);
+         break;
+
+       CASE_FLT_FN (BUILT_IN_IRINT):
+         newfn = mathfn_built_in (TREE_TYPE (arg), BUILT_IN_LRINT);
+         break;
+
+       default:
+         break;
+       }
+
+      if (newfn)
+       {
+         tree newcall = build_call_expr_loc (loc, newfn, 1, arg);
+         return fold_convert_loc (loc,
+                                  TREE_TYPE (TREE_TYPE (fndecl)), newcall);
+       }
+    }
+
   /* Canonicalize llround (x) to lround (x) on LP64 targets where
      sizeof (long long) == sizeof (long).  */
   if (TYPE_PRECISION (long_long_integer_type_node)
@@ -7183,16 +8013,19 @@ fold_builtin_int_roundingfn (location_t loc, tree fndecl, tree arg)
 
          switch (DECL_FUNCTION_CODE (fndecl))
            {
+           CASE_FLT_FN (BUILT_IN_IFLOOR):
            CASE_FLT_FN (BUILT_IN_LFLOOR):
            CASE_FLT_FN (BUILT_IN_LLFLOOR):
              real_floor (&r, TYPE_MODE (ftype), &x);
              break;
 
+           CASE_FLT_FN (BUILT_IN_ICEIL):
            CASE_FLT_FN (BUILT_IN_LCEIL):
            CASE_FLT_FN (BUILT_IN_LLCEIL):
              real_ceil (&r, TYPE_MODE (ftype), &x);
              break;
 
+           CASE_FLT_FN (BUILT_IN_IROUND):
            CASE_FLT_FN (BUILT_IN_LROUND):
            CASE_FLT_FN (BUILT_IN_LLROUND):
              real_round (&r, TYPE_MODE (ftype), &x);
@@ -7250,7 +8083,8 @@ fold_builtin_bitop (tree fndecl, tree arg)
        {
          hi = TREE_INT_CST_HIGH (arg);
          if (width < 2 * HOST_BITS_PER_WIDE_INT)
-           hi &= ~((HOST_WIDE_INT) (-1) >> (width - HOST_BITS_PER_WIDE_INT));
+           hi &= ~((unsigned HOST_WIDE_INT) (-1)
+                   << (width - HOST_BITS_PER_WIDE_INT));
        }
       else
        {
@@ -7288,6 +8122,26 @@ fold_builtin_bitop (tree fndecl, tree arg)
            result = width;
          break;
 
+       CASE_INT_FN (BUILT_IN_CLRSB):
+         if (width > HOST_BITS_PER_WIDE_INT
+             && (hi & ((unsigned HOST_WIDE_INT) 1
+                       << (width - HOST_BITS_PER_WIDE_INT - 1))) != 0)
+           {
+             hi = ~hi & ~((unsigned HOST_WIDE_INT) (-1)
+                          << (width - HOST_BITS_PER_WIDE_INT - 1));
+             lo = ~lo;
+           }
+         else if (width <= HOST_BITS_PER_WIDE_INT
+                  && (lo & ((unsigned HOST_WIDE_INT) 1 << (width - 1))) != 0)
+           lo = ~lo & ~((unsigned HOST_WIDE_INT) (-1) << (width - 1));
+         if (hi != 0)
+           result = width - floor_log2 (hi) - 2 - HOST_BITS_PER_WIDE_INT;
+         else if (lo != 0)
+           result = width - floor_log2 (lo) - 2;
+         else
+           result = width - 1;
+         break;
+
        CASE_INT_FN (BUILT_IN_POPCOUNT):
          result = 0;
          while (lo)
@@ -7812,8 +8666,7 @@ fold_builtin_memset (location_t loc, tree dest, tree c, tree len,
 
   length = tree_low_cst (len, 1);
   if (GET_MODE_SIZE (TYPE_MODE (etype)) != length
-      || get_pointer_alignment (dest, BIGGEST_ALIGNMENT) / BITS_PER_UNIT
-        < length)
+      || get_pointer_alignment (dest) / BITS_PER_UNIT < length)
     return NULL_TREE;
 
   if (length > HOST_BITS_PER_WIDE_INT / BITS_PER_UNIT)
@@ -7864,7 +8717,7 @@ fold_builtin_bzero (location_t loc, tree dest, tree size, bool ignore)
      calling bzero instead of memset.  */
 
   return fold_builtin_memset (loc, dest, integer_zero_node,
-                             fold_convert_loc (loc, sizetype, size),
+                             fold_convert_loc (loc, size_type_node, size),
                              void_type_node, ignore);
 }
 
@@ -7903,8 +8756,8 @@ fold_builtin_memory_op (location_t loc, tree dest, tree src,
 
       if (endp == 3)
        {
-         src_align = get_pointer_alignment (src, BIGGEST_ALIGNMENT);
-         dest_align = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+         src_align = get_pointer_alignment (src);
+         dest_align = get_pointer_alignment (dest);
 
          /* Both DEST and SRC must be pointer types.
             ??? This is what old code did.  Is the testing for pointer types
@@ -7918,7 +8771,7 @@ fold_builtin_memory_op (location_t loc, tree dest, tree src,
                  && (MIN (src_align, dest_align) / BITS_PER_UNIT
                      >= (unsigned HOST_WIDE_INT) tree_low_cst (len, 1))))
            {
-             tree fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+             tree fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
              if (!fn)
                return NULL_TREE;
               return build_call_expr_loc (loc, fn, 3, dest, src, len);
@@ -7977,7 +8830,7 @@ fold_builtin_memory_op (location_t loc, tree dest, tree src,
              else
                return NULL_TREE;
 
-             fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+             fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
              if (!fn)
                return NULL_TREE;
              return build_call_expr_loc (loc, fn, 3, dest, src, len);
@@ -7996,7 +8849,7 @@ fold_builtin_memory_op (location_t loc, tree dest, tree src,
              if (!refs_may_alias_p_1 (&destr, &srcr, false))
                {
                  tree fn;
-                 fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+                 fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
                  if (!fn)
                    return NULL_TREE;
                  return build_call_expr_loc (loc, fn, 3, dest, src, len);
@@ -8052,8 +8905,8 @@ fold_builtin_memory_op (location_t loc, tree dest, tree src,
          || TREE_ADDRESSABLE (desttype))
        return NULL_TREE;
 
-      src_align = get_pointer_alignment (src, BIGGEST_ALIGNMENT);
-      dest_align = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+      src_align = get_pointer_alignment (src);
+      dest_align = get_pointer_alignment (dest);
       if (dest_align < TYPE_ALIGN (desttype)
          || src_align < TYPE_ALIGN (srctype))
        return NULL_TREE;
@@ -8144,8 +8997,7 @@ fold_builtin_memory_op (location_t loc, tree dest, tree src,
     len = fold_build2_loc (loc, MINUS_EXPR, TREE_TYPE (len), len,
                       ssize_int (1));
 
-  len = fold_convert_loc (loc, sizetype, len);
-  dest = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (dest), dest, len);
+  dest = fold_build_pointer_plus_loc (loc, dest, len);
   dest = fold_convert_loc (loc, type, dest);
   if (expr)
     dest = omit_one_operand_loc (loc, type, dest, expr);
@@ -8172,7 +9024,7 @@ fold_builtin_strcpy (location_t loc, tree fndecl, tree dest, tree src, tree len)
   if (optimize_function_for_size_p (cfun))
     return NULL_TREE;
 
-  fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+  fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
   if (!fn)
     return NULL_TREE;
 
@@ -8183,7 +9035,8 @@ fold_builtin_strcpy (location_t loc, tree fndecl, tree dest, tree src, tree len)
        return NULL_TREE;
     }
 
-  len = size_binop_loc (loc, PLUS_EXPR, len, ssize_int (1));
+  len = fold_convert_loc (loc, size_type_node, len);
+  len = size_binop_loc (loc, PLUS_EXPR, len, build_int_cst (size_type_node, 1));
   return fold_convert_loc (loc, TREE_TYPE (TREE_TYPE (fndecl)),
                           build_call_expr_loc (loc, fn, 3, dest, src, len));
 }
@@ -8210,19 +9063,20 @@ fold_builtin_stpcpy (location_t loc, tree fndecl, tree dest, tree src)
       && !integer_zerop (len))
     return NULL_TREE;
 
-  fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+  fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
   if (!fn)
     return NULL_TREE;
 
-  lenp1 = size_binop_loc (loc, PLUS_EXPR, len, ssize_int (1));
+  lenp1 = size_binop_loc (loc, PLUS_EXPR,
+                         fold_convert_loc (loc, size_type_node, len),
+                         build_int_cst (size_type_node, 1));
   /* We use dest twice in building our expression.  Save it from
      multiple expansions.  */
   dest = builtin_save_expr (dest);
   call = build_call_expr_loc (loc, fn, 3, dest, src, lenp1);
 
   type = TREE_TYPE (TREE_TYPE (fndecl));
-  len = fold_convert_loc (loc, sizetype, len);
-  dest = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (dest), dest, len);
+  dest = fold_build_pointer_plus_loc (loc, dest, len);
   dest = fold_convert_loc (loc, type, dest);
   dest = omit_one_operand_loc (loc, type, dest, call);
   return dest;
@@ -8268,9 +9122,11 @@ fold_builtin_strncpy (location_t loc, tree fndecl, tree dest,
     return NULL_TREE;
 
   /* OK transform into builtin memcpy.  */
-  fn = implicit_built_in_decls[BUILT_IN_MEMCPY];
+  fn = builtin_decl_implicit (BUILT_IN_MEMCPY);
   if (!fn)
     return NULL_TREE;
+
+  len = fold_convert_loc (loc, size_type_node, len);
   return fold_convert_loc (loc, TREE_TYPE (TREE_TYPE (fndecl)),
                           build_call_expr_loc (loc, fn, 3, dest, src, len));
 }
@@ -8304,13 +9160,12 @@ fold_builtin_memchr (location_t loc, tree arg1, tree arg2, tree len, tree type)
          if (target_char_cast (arg2, &c))
            return NULL_TREE;
 
-         r = (char *) memchr (p1, c, tree_low_cst (len, 1));
+         r = (const char *) memchr (p1, c, tree_low_cst (len, 1));
 
          if (r == NULL)
            return build_int_cst (TREE_TYPE (arg1), 0);
 
-         tem = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (arg1), arg1,
-                            size_int (r - p1));
+         tem = fold_build_pointer_plus_hwi_loc (loc, arg1, r - p1);
          return fold_convert_loc (loc, type, tem);
        }
       return NULL_TREE;
@@ -8569,8 +9424,9 @@ fold_builtin_signbit (location_t loc, tree arg, tree type)
 
   /* If ARG's format doesn't have signed zeros, return "arg < 0.0".  */
   if (!HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg))))
-    return fold_build2_loc (loc, LT_EXPR, type, arg,
-                       build_real (TREE_TYPE (arg), dconst0));
+    return fold_convert (type,
+                        fold_build2_loc (loc, LT_EXPR, boolean_type_node, arg,
+                       build_real (TREE_TYPE (arg), dconst0)));
 
   return NULL_TREE;
 }
@@ -9089,7 +9945,7 @@ fold_builtin_interclass_mathfn (location_t loc, tree fndecl, tree arg)
     CASE_FLT_FN (BUILT_IN_ISINF):
       {
        /* isinf(x) -> isgreater(fabs(x),DBL_MAX).  */
-       tree const isgr_fn = built_in_decls[BUILT_IN_ISGREATER];
+       tree const isgr_fn = builtin_decl_explicit (BUILT_IN_ISGREATER);
        tree const type = TREE_TYPE (arg);
        REAL_VALUE_TYPE r;
        char buf[128];
@@ -9105,7 +9961,7 @@ fold_builtin_interclass_mathfn (location_t loc, tree fndecl, tree arg)
     case BUILT_IN_ISFINITE:
       {
        /* isfinite(x) -> islessequal(fabs(x),DBL_MAX).  */
-       tree const isle_fn = built_in_decls[BUILT_IN_ISLESSEQUAL];
+       tree const isle_fn = builtin_decl_explicit (BUILT_IN_ISLESSEQUAL);
        tree const type = TREE_TYPE (arg);
        REAL_VALUE_TYPE r;
        char buf[128];
@@ -9128,8 +9984,8 @@ fold_builtin_interclass_mathfn (location_t loc, tree fndecl, tree arg)
       {
        /* isnormal(x) -> isgreaterequal(fabs(x),DBL_MIN) &
           islessequal(fabs(x),DBL_MAX).  */
-       tree const isle_fn = built_in_decls[BUILT_IN_ISLESSEQUAL];
-       tree const isge_fn = built_in_decls[BUILT_IN_ISGREATEREQUAL];
+       tree const isle_fn = builtin_decl_explicit (BUILT_IN_ISLESSEQUAL);
+       tree const isge_fn = builtin_decl_explicit (BUILT_IN_ISGREATEREQUAL);
        tree const type = TREE_TYPE (arg);
        REAL_VALUE_TYPE rmax, rmin;
        char buf[128];
@@ -9190,7 +10046,7 @@ fold_builtin_classify (location_t loc, tree fndecl, tree arg, int builtin_index)
           1.  So e.g. "if (isinf_sign(x))" would be folded to just
           "if (isinf(x) ? 1 : 0)" which becomes "if (isinf(x))". */
        tree signbit_fn = mathfn_built_in_1 (TREE_TYPE (arg), BUILT_IN_SIGNBIT, 0);
-       tree isinf_fn = built_in_decls[BUILT_IN_ISINF];
+       tree isinf_fn = builtin_decl_explicit (BUILT_IN_ISINF);
        tree tmp = NULL_TREE;
 
        arg = builtin_save_expr (arg);
@@ -9703,14 +10559,18 @@ fold_builtin_1 (location_t loc, tree fndecl, tree arg0, bool ignore)
     CASE_FLT_FN (BUILT_IN_RINT):
       return fold_trunc_transparent_mathfn (loc, fndecl, arg0);
 
+    CASE_FLT_FN (BUILT_IN_ICEIL):
     CASE_FLT_FN (BUILT_IN_LCEIL):
     CASE_FLT_FN (BUILT_IN_LLCEIL):
     CASE_FLT_FN (BUILT_IN_LFLOOR):
+    CASE_FLT_FN (BUILT_IN_IFLOOR):
     CASE_FLT_FN (BUILT_IN_LLFLOOR):
+    CASE_FLT_FN (BUILT_IN_IROUND):
     CASE_FLT_FN (BUILT_IN_LROUND):
     CASE_FLT_FN (BUILT_IN_LLROUND):
       return fold_builtin_int_roundingfn (loc, fndecl, arg0);
 
+    CASE_FLT_FN (BUILT_IN_IRINT):
     CASE_FLT_FN (BUILT_IN_LRINT):
     CASE_FLT_FN (BUILT_IN_LLRINT):
       return fold_fixed_mathfn (loc, fndecl, arg0);
@@ -9722,6 +10582,7 @@ fold_builtin_1 (location_t loc, tree fndecl, tree arg0, bool ignore)
     CASE_INT_FN (BUILT_IN_FFS):
     CASE_INT_FN (BUILT_IN_CLZ):
     CASE_INT_FN (BUILT_IN_CTZ):
+    CASE_INT_FN (BUILT_IN_CLRSB):
     CASE_INT_FN (BUILT_IN_POPCOUNT):
     CASE_INT_FN (BUILT_IN_PARITY):
       return fold_builtin_bitop (fndecl, arg0);
@@ -9908,7 +10769,7 @@ fold_builtin_2 (location_t loc, tree fndecl, tree arg0, tree arg1, bool ignore)
     case BUILT_IN_STPCPY:
       if (ignore)
        {
-         tree fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+         tree fn = builtin_decl_implicit (BUILT_IN_STRCPY);
          if (!fn)
            break;
 
@@ -9993,6 +10854,12 @@ fold_builtin_2 (location_t loc, tree fndecl, tree arg0, tree arg1, bool ignore)
       return fold_builtin_fprintf (loc, fndecl, arg0, arg1, NULL_TREE,
                                   ignore, fcode);
 
+    case BUILT_IN_ATOMIC_ALWAYS_LOCK_FREE:
+      return fold_builtin_atomic_always_lock_free (arg0, arg1);
+
+    case BUILT_IN_ATOMIC_IS_LOCK_FREE:
+      return fold_builtin_atomic_is_lock_free (arg0, arg1);
+
     default:
       break;
     }
@@ -10127,7 +10994,9 @@ fold_builtin_4 (location_t loc, tree fndecl,
                                      DECL_FUNCTION_CODE (fndecl));
 
     case BUILT_IN_STRNCPY_CHK:
-      return fold_builtin_strncpy_chk (loc, arg0, arg1, arg2, arg3, NULL_TREE);
+    case BUILT_IN_STPNCPY_CHK:
+      return fold_builtin_stxncpy_chk (loc, arg0, arg1, arg2, arg3, NULL_TREE,
+                                       ignore, fcode);
 
     case BUILT_IN_STRNCAT_CHK:
       return fold_builtin_strncat_chk (loc, fndecl, arg0, arg1, arg2, arg3);
@@ -10246,7 +11115,7 @@ fold_builtin_varargs (location_t loc, tree fndecl, tree exp,
    been inlined, otherwise e.g. -D_FORTIFY_SOURCE checking
    might not be performed.  */
 
-static bool
+bool
 avoid_folding_inline_builtin (tree fndecl)
 {
   return (DECL_DECLARED_INLINE_P (fndecl)
@@ -10696,8 +11565,7 @@ fold_builtin_strstr (location_t loc, tree s1, tree s2, tree type)
            return build_int_cst (TREE_TYPE (s1), 0);
 
          /* Return an offset into the constant string argument.  */
-         tem = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (s1),
-                            s1, size_int (r - p1));
+         tem = fold_build_pointer_plus_hwi_loc (loc, s1, r - p1);
          return fold_convert_loc (loc, type, tem);
        }
 
@@ -10709,7 +11577,7 @@ fold_builtin_strstr (location_t loc, tree s1, tree s2, tree type)
       if (p2[1] != '\0')
        return NULL_TREE;
 
-      fn = implicit_built_in_decls[BUILT_IN_STRCHR];
+      fn = builtin_decl_implicit (BUILT_IN_STRCHR);
       if (!fn)
        return NULL_TREE;
 
@@ -10767,8 +11635,7 @@ fold_builtin_strchr (location_t loc, tree s1, tree s2, tree type)
            return build_int_cst (TREE_TYPE (s1), 0);
 
          /* Return an offset into the constant string argument.  */
-         tem = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (s1),
-                            s1, size_int (r - p1));
+         tem = fold_build_pointer_plus_hwi_loc (loc, s1, r - p1);
          return fold_convert_loc (loc, type, tem);
        }
       return NULL_TREE;
@@ -10823,15 +11690,14 @@ fold_builtin_strrchr (location_t loc, tree s1, tree s2, tree type)
            return build_int_cst (TREE_TYPE (s1), 0);
 
          /* Return an offset into the constant string argument.  */
-         tem = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (s1),
-                            s1, size_int (r - p1));
+         tem = fold_build_pointer_plus_hwi_loc (loc, s1, r - p1);
          return fold_convert_loc (loc, type, tem);
        }
 
       if (! integer_zerop (s2))
        return NULL_TREE;
 
-      fn = implicit_built_in_decls[BUILT_IN_STRCHR];
+      fn = builtin_decl_implicit (BUILT_IN_STRCHR);
       if (!fn)
        return NULL_TREE;
 
@@ -10883,8 +11749,7 @@ fold_builtin_strpbrk (location_t loc, tree s1, tree s2, tree type)
            return build_int_cst (TREE_TYPE (s1), 0);
 
          /* Return an offset into the constant string argument.  */
-         tem = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (s1),
-                            s1, size_int (r - p1));
+         tem = fold_build_pointer_plus_hwi_loc (loc, s1, r - p1);
          return fold_convert_loc (loc, type, tem);
        }
 
@@ -10896,7 +11761,7 @@ fold_builtin_strpbrk (location_t loc, tree s1, tree s2, tree type)
       if (p2[1] != '\0')
        return NULL_TREE;  /* Really call strpbrk.  */
 
-      fn = implicit_built_in_decls[BUILT_IN_STRCHR];
+      fn = builtin_decl_implicit (BUILT_IN_STRCHR);
       if (!fn)
        return NULL_TREE;
 
@@ -10943,8 +11808,8 @@ fold_builtin_strcat (location_t loc ATTRIBUTE_UNUSED, tree dst, tree src)
        {
          /* See if we can store by pieces into (dst + strlen(dst)).  */
          tree newdst, call;
-         tree strlen_fn = implicit_built_in_decls[BUILT_IN_STRLEN];
-         tree strcpy_fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+         tree strlen_fn = builtin_decl_implicit (BUILT_IN_STRLEN);
+         tree strcpy_fn = builtin_decl_implicit (BUILT_IN_STRCPY);
 
          if (!strlen_fn || !strcpy_fn)
            return NULL_TREE;
@@ -10968,8 +11833,7 @@ fold_builtin_strcat (location_t loc ATTRIBUTE_UNUSED, tree dst, tree src)
          newdst = build_call_expr_loc (loc, strlen_fn, 1, dst);
          /* Create (dst p+ strlen (dst)).  */
 
-         newdst = fold_build2_loc (loc, POINTER_PLUS_EXPR,
-                               TREE_TYPE (dst), dst, newdst);
+         newdst = fold_build_pointer_plus_loc (loc, dst, newdst);
          newdst = builtin_save_expr (newdst);
 
          call = build_call_expr_loc (loc, strcpy_fn, 2, newdst, src);
@@ -11018,7 +11882,7 @@ fold_builtin_strncat (location_t loc, tree dst, tree src, tree len)
       if (TREE_CODE (len) == INTEGER_CST && p
          && compare_tree_int (len, strlen (p)) >= 0)
        {
-         tree fn = implicit_built_in_decls[BUILT_IN_STRCAT];
+         tree fn = builtin_decl_implicit (BUILT_IN_STRCAT);
 
          /* If the replacement _DECL isn't initialized, don't do the
             transformation.  */
@@ -11063,7 +11927,7 @@ fold_builtin_strspn (location_t loc, tree s1, tree s2)
       if (p1 && p2)
        {
          const size_t r = strspn (p1, p2);
-         return size_int (r);
+         return build_int_cst (size_type_node, r);
        }
 
       /* If either argument is "", return NULL_TREE.  */
@@ -11108,7 +11972,7 @@ fold_builtin_strcspn (location_t loc, tree s1, tree s2)
       if (p1 && p2)
        {
          const size_t r = strcspn (p1, p2);
-         return size_int (r);
+         return build_int_cst (size_type_node, r);
        }
 
       /* If the first argument is "", return NULL_TREE.  */
@@ -11123,7 +11987,7 @@ fold_builtin_strcspn (location_t loc, tree s1, tree s2)
       /* If the second argument is "", return __builtin_strlen(s1).  */
       if (p2 && *p2 == '\0')
        {
-         tree fn = implicit_built_in_decls[BUILT_IN_STRLEN];
+         tree fn = builtin_decl_implicit (BUILT_IN_STRLEN);
 
          /* If the replacement _DECL isn't initialized, don't do the
             transformation.  */
@@ -11149,10 +12013,12 @@ fold_builtin_fputs (location_t loc, tree arg0, tree arg1,
 {
   /* If we're using an unlocked function, assume the other unlocked
      functions exist explicitly.  */
-  tree const fn_fputc = unlocked ? built_in_decls[BUILT_IN_FPUTC_UNLOCKED]
-    : implicit_built_in_decls[BUILT_IN_FPUTC];
-  tree const fn_fwrite = unlocked ? built_in_decls[BUILT_IN_FWRITE_UNLOCKED]
-    : implicit_built_in_decls[BUILT_IN_FWRITE];
+  tree const fn_fputc = (unlocked
+                        ? builtin_decl_explicit (BUILT_IN_FPUTC_UNLOCKED)
+                        : builtin_decl_implicit (BUILT_IN_FPUTC));
+  tree const fn_fwrite = (unlocked
+                         ? builtin_decl_explicit (BUILT_IN_FWRITE_UNLOCKED)
+                         : builtin_decl_implicit (BUILT_IN_FWRITE));
 
   /* If the return value is used, don't do the transformation.  */
   if (!ignore)
@@ -11346,7 +12212,7 @@ fold_builtin_sprintf (location_t loc, tree dest, tree fmt,
   /* If the format doesn't contain % args or %%, use strcpy.  */
   if (strchr (fmt_str, target_percent) == NULL)
     {
-      tree fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+      tree fn = builtin_decl_implicit (BUILT_IN_STRCPY);
 
       if (!fn)
        return NULL_TREE;
@@ -11366,7 +12232,7 @@ fold_builtin_sprintf (location_t loc, tree dest, tree fmt,
   else if (fmt_str && strcmp (fmt_str, target_percent_s) == 0)
     {
       tree fn;
-      fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+      fn = builtin_decl_implicit (BUILT_IN_STRCPY);
 
       if (!fn)
        return NULL_TREE;
@@ -11388,7 +12254,7 @@ fold_builtin_sprintf (location_t loc, tree dest, tree fmt,
   if (call && retval)
     {
       retval = fold_convert_loc
-       (loc, TREE_TYPE (TREE_TYPE (implicit_built_in_decls[BUILT_IN_SPRINTF])),
+       (loc, TREE_TYPE (TREE_TYPE (builtin_decl_implicit (BUILT_IN_SPRINTF))),
         retval);
       return build2 (COMPOUND_EXPR, TREE_TYPE (retval), call, retval);
     }
@@ -11441,7 +12307,7 @@ fold_builtin_snprintf (location_t loc, tree dest, tree destsize, tree fmt,
   /* If the format doesn't contain % args or %%, use strcpy.  */
   if (strchr (fmt_str, target_percent) == NULL)
     {
-      tree fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+      tree fn = builtin_decl_implicit (BUILT_IN_STRCPY);
       size_t len = strlen (fmt_str);
 
       /* Don't optimize snprintf (buf, 4, "abc", ptr++).  */
@@ -11473,7 +12339,7 @@ fold_builtin_snprintf (location_t loc, tree dest, tree destsize, tree fmt,
   /* If the format is "%s", use strcpy if the result isn't used.  */
   else if (fmt_str && strcmp (fmt_str, target_percent_s) == 0)
     {
-      tree fn = implicit_built_in_decls[BUILT_IN_STRCPY];
+      tree fn = builtin_decl_implicit (BUILT_IN_STRCPY);
       unsigned HOST_WIDE_INT origlen;
 
       /* Don't crash on snprintf (str1, cst, "%s").  */
@@ -11508,7 +12374,7 @@ fold_builtin_snprintf (location_t loc, tree dest, tree destsize, tree fmt,
 
   if (call && retval)
     {
-      tree fn = built_in_decls[BUILT_IN_SNPRINTF];
+      tree fn = builtin_decl_explicit (BUILT_IN_SNPRINTF);
       retval = fold_convert_loc (loc, TREE_TYPE (TREE_TYPE (fn)), retval);
       return build2 (COMPOUND_EXPR, TREE_TYPE (retval), call, retval);
     }
@@ -11596,16 +12462,16 @@ expand_builtin_memory_chk (tree exp, rtx target, enum machine_mode mode,
       switch (fcode)
        {
        case BUILT_IN_MEMCPY_CHK:
-         fn = built_in_decls[BUILT_IN_MEMCPY];
+         fn = builtin_decl_explicit (BUILT_IN_MEMCPY);
          break;
        case BUILT_IN_MEMPCPY_CHK:
-         fn = built_in_decls[BUILT_IN_MEMPCPY];
+         fn = builtin_decl_explicit (BUILT_IN_MEMPCPY);
          break;
        case BUILT_IN_MEMMOVE_CHK:
-         fn = built_in_decls[BUILT_IN_MEMMOVE];
+         fn = builtin_decl_explicit (BUILT_IN_MEMMOVE);
          break;
        case BUILT_IN_MEMSET_CHK:
-         fn = built_in_decls[BUILT_IN_MEMSET];
+         fn = builtin_decl_explicit (BUILT_IN_MEMSET);
          break;
        default:
          break;
@@ -11623,8 +12489,7 @@ expand_builtin_memory_chk (tree exp, rtx target, enum machine_mode mode,
     return NULL_RTX;
   else
     {
-      unsigned int dest_align
-       = get_pointer_alignment (dest, BIGGEST_ALIGNMENT);
+      unsigned int dest_align = get_pointer_alignment (dest);
 
       /* If DEST is not a pointer type, call the normal function.  */
       if (dest_align == 0)
@@ -11642,15 +12507,14 @@ expand_builtin_memory_chk (tree exp, rtx target, enum machine_mode mode,
              return expand_expr (dest, target, mode, EXPAND_NORMAL);
            }
 
-         expr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (dest), dest, len);
+         expr = fold_build_pointer_plus (dest, len);
          return expand_expr (expr, target, mode, EXPAND_NORMAL);
        }
 
       /* __memmove_chk special case.  */
       if (fcode == BUILT_IN_MEMMOVE_CHK)
        {
-         unsigned int src_align
-           = get_pointer_alignment (src, BIGGEST_ALIGNMENT);
+         unsigned int src_align = get_pointer_alignment (src);
 
          if (src_align == 0)
            return NULL_RTX;
@@ -11659,7 +12523,7 @@ expand_builtin_memory_chk (tree exp, rtx target, enum machine_mode mode,
             normal __memcpy_chk.  */
          if (readonly_data_expr (src))
            {
-             tree fn = built_in_decls[BUILT_IN_MEMCPY_CHK];
+             tree fn = builtin_decl_explicit (BUILT_IN_MEMCPY_CHK);
              if (!fn)
                return NULL_RTX;
              fn = build_call_nofold_loc (EXPR_LOCATION (exp), fn, 4,
@@ -11695,6 +12559,7 @@ maybe_emit_chk_warning (tree exp, enum built_in_function fcode)
       break;
     case BUILT_IN_STRNCAT_CHK:
     case BUILT_IN_STRNCPY_CHK:
+    case BUILT_IN_STPNCPY_CHK:
       len = CALL_EXPR_ARG (exp, 2);
       size = CALL_EXPR_ARG (exp, 3);
       break;
@@ -11814,11 +12679,11 @@ maybe_emit_free_warning (tree exp)
     return;
 
   if (SSA_VAR_P (arg))
-    warning_at (tree_nonartificial_location (exp),
-               0, "%Kattempt to free a non-heap object %qD", exp, arg);
+    warning_at (tree_nonartificial_location (exp), OPT_Wfree_nonheap_object,
+               "%Kattempt to free a non-heap object %qD", exp, arg);
   else
-    warning_at (tree_nonartificial_location (exp),
-               0, "%Kattempt to free a non-heap object", exp);
+    warning_at (tree_nonartificial_location (exp), OPT_Wfree_nonheap_object,
+               "%Kattempt to free a non-heap object", exp);
 }
 
 /* Fold a call to __builtin_object_size with arguments PTR and OST,
@@ -11902,8 +12767,7 @@ fold_builtin_memory_chk (location_t loc, tree fndecl,
                                 dest, len);
       else
        {
-         tree temp = fold_build2_loc (loc, POINTER_PLUS_EXPR, TREE_TYPE (dest),
-                                  dest, len);
+         tree temp = fold_build_pointer_plus_loc (loc, dest, len);
          return fold_convert_loc (loc, TREE_TYPE (TREE_TYPE (fndecl)), temp);
        }
     }
@@ -11924,7 +12788,7 @@ fold_builtin_memory_chk (location_t loc, tree fndecl,
                {
                  /* (void) __mempcpy_chk () can be optimized into
                     (void) __memcpy_chk ().  */
-                 fn = built_in_decls[BUILT_IN_MEMCPY_CHK];
+                 fn = builtin_decl_explicit (BUILT_IN_MEMCPY_CHK);
                  if (!fn)
                    return NULL_TREE;
 
@@ -11946,16 +12810,16 @@ fold_builtin_memory_chk (location_t loc, tree fndecl,
   switch (fcode)
     {
     case BUILT_IN_MEMCPY_CHK:
-      fn = built_in_decls[BUILT_IN_MEMCPY];
+      fn = builtin_decl_explicit (BUILT_IN_MEMCPY);
       break;
     case BUILT_IN_MEMPCPY_CHK:
-      fn = built_in_decls[BUILT_IN_MEMPCPY];
+      fn = builtin_decl_explicit (BUILT_IN_MEMPCPY);
       break;
     case BUILT_IN_MEMMOVE_CHK:
-      fn = built_in_decls[BUILT_IN_MEMMOVE];
+      fn = builtin_decl_explicit (BUILT_IN_MEMMOVE);
       break;
     case BUILT_IN_MEMSET_CHK:
-      fn = built_in_decls[BUILT_IN_MEMSET];
+      fn = builtin_decl_explicit (BUILT_IN_MEMSET);
       break;
     default:
       break;
@@ -12010,7 +12874,7 @@ fold_builtin_stxcpy_chk (location_t loc, tree fndecl, tree dest,
 
                  /* If return value of __stpcpy_chk is ignored,
                     optimize into __strcpy_chk.  */
-                 fn = built_in_decls[BUILT_IN_STRCPY_CHK];
+                 fn = builtin_decl_explicit (BUILT_IN_STRCPY_CHK);
                  if (!fn)
                    return NULL_TREE;
 
@@ -12022,11 +12886,13 @@ fold_builtin_stxcpy_chk (location_t loc, tree fndecl, tree dest,
 
              /* If c_strlen returned something, but not a constant,
                 transform __strcpy_chk into __memcpy_chk.  */
-             fn = built_in_decls[BUILT_IN_MEMCPY_CHK];
+             fn = builtin_decl_explicit (BUILT_IN_MEMCPY_CHK);
              if (!fn)
                return NULL_TREE;
 
-             len = size_binop_loc (loc, PLUS_EXPR, len, ssize_int (1));
+             len = fold_convert_loc (loc, size_type_node, len);
+             len = size_binop_loc (loc, PLUS_EXPR, len,
+                                   build_int_cst (size_type_node, 1));
              return fold_convert_loc (loc, TREE_TYPE (TREE_TYPE (fndecl)),
                                       build_call_expr_loc (loc, fn, 4,
                                                        dest, src, len, size));
@@ -12040,21 +12906,23 @@ fold_builtin_stxcpy_chk (location_t loc, tree fndecl, tree dest,
     }
 
   /* If __builtin_st{r,p}cpy_chk is used, assume st{r,p}cpy is available.  */
-  fn = built_in_decls[fcode == BUILT_IN_STPCPY_CHK
-                     ? BUILT_IN_STPCPY : BUILT_IN_STRCPY];
+  fn = builtin_decl_explicit (fcode == BUILT_IN_STPCPY_CHK
+                             ? BUILT_IN_STPCPY : BUILT_IN_STRCPY);
   if (!fn)
     return NULL_TREE;
 
   return build_call_expr_loc (loc, fn, 2, dest, src);
 }
 
-/* Fold a call to the __strncpy_chk builtin.  DEST, SRC, LEN, and SIZE
+/* Fold a call to the __st{r,p}ncpy_chk builtin.  DEST, SRC, LEN, and SIZE
    are the arguments to the call.  If MAXLEN is not NULL, it is maximum
-   length passed as third argument.  */
+   length passed as third argument. IGNORE is true if return value can be
+   ignored. FCODE is the BUILT_IN_* code of the builtin. */
 
 tree
-fold_builtin_strncpy_chk (location_t loc, tree dest, tree src,
-                         tree len, tree size, tree maxlen)
+fold_builtin_stxncpy_chk (location_t loc, tree dest, tree src,
+                         tree len, tree size, tree maxlen, bool ignore,
+                         enum built_in_function fcode)
 {
   tree fn;
 
@@ -12064,6 +12932,15 @@ fold_builtin_strncpy_chk (location_t loc, tree dest, tree src,
       || !validate_arg (size, INTEGER_TYPE))
     return NULL_TREE;
 
+  if (fcode == BUILT_IN_STPNCPY_CHK && ignore)
+    {
+       /* If return value of __stpncpy_chk is ignored,
+          optimize into __strncpy_chk.  */
+       fn = builtin_decl_explicit (BUILT_IN_STRNCPY_CHK);
+       if (fn)
+         return build_call_expr_loc (loc, fn, 4, dest, src, len, size);
+    }
+
   if (! host_integerp (size, 1))
     return NULL_TREE;
 
@@ -12084,8 +12961,9 @@ fold_builtin_strncpy_chk (location_t loc, tree dest, tree src,
        return NULL_TREE;
     }
 
-  /* If __builtin_strncpy_chk is used, assume strncpy is available.  */
-  fn = built_in_decls[BUILT_IN_STRNCPY];
+  /* If __builtin_st{r,p}ncpy_chk is used, assume st{r,p}ncpy is available.  */
+  fn = builtin_decl_explicit (fcode == BUILT_IN_STPNCPY_CHK
+                             ? BUILT_IN_STPNCPY : BUILT_IN_STRNCPY);
   if (!fn)
     return NULL_TREE;
 
@@ -12116,7 +12994,7 @@ fold_builtin_strcat_chk (location_t loc, tree fndecl, tree dest,
     return NULL_TREE;
 
   /* If __builtin_strcat_chk is used, assume strcat is available.  */
-  fn = built_in_decls[BUILT_IN_STRCAT];
+  fn = builtin_decl_explicit (BUILT_IN_STRCAT);
   if (!fn)
     return NULL_TREE;
 
@@ -12158,7 +13036,7 @@ fold_builtin_strncat_chk (location_t loc, tree fndecl,
          && ! tree_int_cst_lt (len, src_len))
        {
          /* If LEN >= strlen (SRC), optimize into __strcat_chk.  */
-         fn = built_in_decls[BUILT_IN_STRCAT_CHK];
+         fn = builtin_decl_explicit (BUILT_IN_STRCAT_CHK);
          if (!fn)
            return NULL_TREE;
 
@@ -12168,7 +13046,7 @@ fold_builtin_strncat_chk (location_t loc, tree fndecl,
     }
 
   /* If __builtin_strncat_chk is used, assume strncat is available.  */
-  fn = built_in_decls[BUILT_IN_STRNCAT];
+  fn = builtin_decl_explicit (BUILT_IN_STRNCAT);
   if (!fn)
     return NULL_TREE;
 
@@ -12259,8 +13137,8 @@ fold_builtin_sprintf_chk_1 (location_t loc, int nargs, tree *args,
     }
 
   /* If __builtin_{,v}sprintf_chk is used, assume {,v}sprintf is available.  */
-  fn = built_in_decls[fcode == BUILT_IN_VSPRINTF_CHK
-                     ? BUILT_IN_VSPRINTF : BUILT_IN_SPRINTF];
+  fn = builtin_decl_explicit (fcode == BUILT_IN_VSPRINTF_CHK
+                             ? BUILT_IN_VSPRINTF : BUILT_IN_SPRINTF);
   if (!fn)
     return NULL_TREE;
 
@@ -12348,8 +13226,8 @@ fold_builtin_snprintf_chk_1 (location_t loc, int nargs, tree *args,
 
   /* If __builtin_{,v}snprintf_chk is used, assume {,v}snprintf is
      available.  */
-  fn = built_in_decls[fcode == BUILT_IN_VSNPRINTF_CHK
-                     ? BUILT_IN_VSNPRINTF : BUILT_IN_SNPRINTF];
+  fn = builtin_decl_explicit (fcode == BUILT_IN_VSNPRINTF_CHK
+                             ? BUILT_IN_VSNPRINTF : BUILT_IN_SNPRINTF);
   if (!fn)
     return NULL_TREE;
 
@@ -12403,13 +13281,13 @@ fold_builtin_printf (location_t loc, tree fndecl, tree fmt,
     {
       /* If we're using an unlocked function, assume the other
         unlocked functions exist explicitly.  */
-      fn_putchar = built_in_decls[BUILT_IN_PUTCHAR_UNLOCKED];
-      fn_puts = built_in_decls[BUILT_IN_PUTS_UNLOCKED];
+      fn_putchar = builtin_decl_explicit (BUILT_IN_PUTCHAR_UNLOCKED);
+      fn_puts = builtin_decl_explicit (BUILT_IN_PUTS_UNLOCKED);
     }
   else
     {
-      fn_putchar = implicit_built_in_decls[BUILT_IN_PUTCHAR];
-      fn_puts = implicit_built_in_decls[BUILT_IN_PUTS];
+      fn_putchar = builtin_decl_implicit (BUILT_IN_PUTCHAR);
+      fn_puts = builtin_decl_implicit (BUILT_IN_PUTS);
     }
 
   if (!init_target_chars ())
@@ -12554,13 +13432,13 @@ fold_builtin_fprintf (location_t loc, tree fndecl, tree fp,
     {
       /* If we're using an unlocked function, assume the other
         unlocked functions exist explicitly.  */
-      fn_fputc = built_in_decls[BUILT_IN_FPUTC_UNLOCKED];
-      fn_fputs = built_in_decls[BUILT_IN_FPUTS_UNLOCKED];
+      fn_fputc = builtin_decl_explicit (BUILT_IN_FPUTC_UNLOCKED);
+      fn_fputs = builtin_decl_explicit (BUILT_IN_FPUTS_UNLOCKED);
     }
   else
     {
-      fn_fputc = implicit_built_in_decls[BUILT_IN_FPUTC];
-      fn_fputs = implicit_built_in_decls[BUILT_IN_FPUTS];
+      fn_fputc = builtin_decl_implicit (BUILT_IN_FPUTC);
+      fn_fputs = builtin_decl_implicit (BUILT_IN_FPUTS);
     }
 
   if (!init_target_chars ())
@@ -13362,7 +14240,7 @@ fold_call_stmt (gimple stmt, bool ignore)
   return NULL_TREE;
 }
 
-/* Look up the function in built_in_decls that corresponds to DECL
+/* Look up the function in builtin_decl that corresponds to DECL
    and set ASMSPEC as its user assembler name.  DECL must be a
    function decl that declares a builtin.  */
 
@@ -13374,7 +14252,7 @@ set_builtin_user_assembler_name (tree decl, const char *asmspec)
              && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL
              && asmspec != 0);
 
-  builtin = built_in_decls [DECL_FUNCTION_CODE (decl)];
+  builtin = builtin_decl_explicit (DECL_FUNCTION_CODE (decl));
   set_user_assembler_name (builtin, asmspec);
   switch (DECL_FUNCTION_CODE (decl))
     {
@@ -13422,6 +14300,7 @@ is_simple_builtin (tree decl)
       case BUILT_IN_OBJECT_SIZE:
       case BUILT_IN_UNREACHABLE:
        /* Simple register moves or loads from stack.  */
+      case BUILT_IN_ASSUME_ALIGNED:
       case BUILT_IN_RETURN_ADDRESS:
       case BUILT_IN_EXTRACT_RETURN_ADDR:
       case BUILT_IN_FROB_RETURN_ADDR:
@@ -13459,6 +14338,7 @@ is_inexpensive_builtin (tree decl)
       {
       case BUILT_IN_ABS:
       case BUILT_IN_ALLOCA:
+      case BUILT_IN_ALLOCA_WITH_ALIGN:
       case BUILT_IN_BSWAP32:
       case BUILT_IN_BSWAP64:
       case BUILT_IN_CLZ: