i386, fab: Optimize __atomic_{add,sub,and,or,xor}_fetch (x, y, z) {==,!=,<,<=,>,...
authorJakub Jelinek <jakub@redhat.com>
Mon, 3 Jan 2022 13:02:23 +0000 (14:02 +0100)
committerJakub Jelinek <jakub@redhat.com>
Mon, 3 Jan 2022 13:17:26 +0000 (14:17 +0100)
On Wed, Jan 27, 2021 at 12:27:13PM +0100, Ulrich Drepper via Gcc-patches wrote:
> On 1/27/21 11:37 AM, Jakub Jelinek wrote:
> > Would equality comparison against 0 handle the most common cases.
> >
> > The user can write it as
> > __atomic_sub_fetch (x, y, z) == 0
> > or
> > __atomic_fetch_sub (x, y, z) - y == 0
> > thouch, so the expansion code would need to be able to cope with both.
>
> Please also keep !=0, <0, <=0, >0, and >=0 in mind.  They all can be
> useful and can be handled with the flags.

<= 0 and > 0 don't really work well with lock {add,sub,inc,dec}, x86 doesn't
have comparisons that would look solely at both SF and ZF and not at other
flags (and emitting two separate conditional jumps or two setcc insns and
oring them together looks awful).

But the rest can work.

Here is a patch that adds internal functions and optabs for these,
recognizes them at the same spot as e.g. .ATOMIC_BIT_TEST_AND* internal
functions (fold all builtins pass) and expands them appropriately (or for
the <= 0 and > 0 cases of +/- FAILs and let's middle-end fall back).

So far I have handled just the op_fetch builtins, IMHO instead of handling
also __atomic_fetch_sub (x, y, z) - y == 0 etc. we should canonicalize
__atomic_fetch_sub (x, y, z) - y to __atomic_sub_fetch (x, y, z) (and vice
versa).

2022-01-03  Jakub Jelinek  <jakub@redhat.com>

PR target/98737
* internal-fn.def (ATOMIC_ADD_FETCH_CMP_0, ATOMIC_SUB_FETCH_CMP_0,
ATOMIC_AND_FETCH_CMP_0, ATOMIC_OR_FETCH_CMP_0, ATOMIC_XOR_FETCH_CMP_0):
New internal fns.
* internal-fn.h (ATOMIC_OP_FETCH_CMP_0_EQ, ATOMIC_OP_FETCH_CMP_0_NE,
ATOMIC_OP_FETCH_CMP_0_LT, ATOMIC_OP_FETCH_CMP_0_LE,
ATOMIC_OP_FETCH_CMP_0_GT, ATOMIC_OP_FETCH_CMP_0_GE): New enumerators.
* internal-fn.c (expand_ATOMIC_ADD_FETCH_CMP_0,
expand_ATOMIC_SUB_FETCH_CMP_0, expand_ATOMIC_AND_FETCH_CMP_0,
expand_ATOMIC_OR_FETCH_CMP_0, expand_ATOMIC_XOR_FETCH_CMP_0): New
functions.
* optabs.def (atomic_add_fetch_cmp_0_optab,
atomic_sub_fetch_cmp_0_optab, atomic_and_fetch_cmp_0_optab,
atomic_or_fetch_cmp_0_optab, atomic_xor_fetch_cmp_0_optab): New
direct optabs.
* builtins.h (expand_ifn_atomic_op_fetch_cmp_0): Declare.
* builtins.c (expand_ifn_atomic_op_fetch_cmp_0): New function.
* tree-ssa-ccp.c: Include internal-fn.h.
(optimize_atomic_bit_test_and): Add . before internal fn call
in function comment.  Change return type from void to bool and
return true only if successfully replaced.
(optimize_atomic_op_fetch_cmp_0): New function.
(pass_fold_builtins::execute): Use optimize_atomic_op_fetch_cmp_0
for BUILT_IN_ATOMIC_{ADD,SUB,AND,OR,XOR}_FETCH_{1,2,4,8,16} and
BUILT_IN_SYNC_{ADD,SUB,AND,OR,XOR}_AND_FETCH_{1,2,4,8,16},
for *XOR* ones only if optimize_atomic_bit_test_and failed.
* config/i386/sync.md (atomic_<plusminus_mnemonic>_fetch_cmp_0<mode>,
atomic_<logic>_fetch_cmp_0<mode>): New define_expand patterns.
(atomic_add_fetch_cmp_0<mode>_1, atomic_sub_fetch_cmp_0<mode>_1,
atomic_<logic>_fetch_cmp_0<mode>_1): New define_insn patterns.
* doc/md.texi (atomic_add_fetch_cmp_0<mode>,
atomic_sub_fetch_cmp_0<mode>, atomic_and_fetch_cmp_0<mode>,
atomic_or_fetch_cmp_0<mode>, atomic_xor_fetch_cmp_0<mode>): Document
new named patterns.

* gcc.target/i386/pr98737-1.c: New test.
* gcc.target/i386/pr98737-2.c: New test.
* gcc.target/i386/pr98737-3.c: New test.
* gcc.target/i386/pr98737-4.c: New test.
* gcc.target/i386/pr98737-5.c: New test.
* gcc.target/i386/pr98737-6.c: New test.
* gcc.target/i386/pr98737-7.c: New test.

16 files changed:
gcc/builtins.c
gcc/builtins.h
gcc/config/i386/sync.md
gcc/doc/md.texi
gcc/internal-fn.c
gcc/internal-fn.def
gcc/internal-fn.h
gcc/optabs.def
gcc/testsuite/gcc.target/i386/pr98737-1.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/pr98737-2.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/pr98737-3.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/pr98737-4.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/pr98737-5.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/pr98737-6.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/pr98737-7.c [new file with mode: 0644]
gcc/tree-ssa-ccp.c

index 77c5755..39b08af 100644 (file)
@@ -6276,6 +6276,93 @@ expand_ifn_atomic_bit_test_and (gcall *call)
     emit_move_insn (target, result);
 }
 
+/* Expand IFN_ATOMIC_*_FETCH_CMP_0 internal function.  */
+
+void
+expand_ifn_atomic_op_fetch_cmp_0 (gcall *call)
+{
+  tree cmp = gimple_call_arg (call, 0);
+  tree ptr = gimple_call_arg (call, 1);
+  tree arg = gimple_call_arg (call, 2);
+  tree lhs = gimple_call_lhs (call);
+  enum memmodel model = MEMMODEL_SYNC_SEQ_CST;
+  machine_mode mode = TYPE_MODE (TREE_TYPE (cmp));
+  optab optab;
+  rtx_code code;
+  class expand_operand ops[5];
+
+  gcc_assert (flag_inline_atomics);
+
+  if (gimple_call_num_args (call) == 4)
+    model = get_memmodel (gimple_call_arg (call, 3));
+
+  rtx mem = get_builtin_sync_mem (ptr, mode);
+  rtx op = expand_expr_force_mode (arg, mode);
+
+  switch (gimple_call_internal_fn (call))
+    {
+    case IFN_ATOMIC_ADD_FETCH_CMP_0:
+      code = PLUS;
+      optab = atomic_add_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_SUB_FETCH_CMP_0:
+      code = MINUS;
+      optab = atomic_sub_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_AND_FETCH_CMP_0:
+      code = AND;
+      optab = atomic_and_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_OR_FETCH_CMP_0:
+      code = IOR;
+      optab = atomic_or_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_XOR_FETCH_CMP_0:
+      code = XOR;
+      optab = atomic_xor_fetch_cmp_0_optab;
+      break;
+    default:
+      gcc_unreachable ();
+    }
+
+  enum rtx_code comp = UNKNOWN;
+  switch (tree_to_uhwi (cmp))
+    {
+    case ATOMIC_OP_FETCH_CMP_0_EQ: comp = EQ; break;
+    case ATOMIC_OP_FETCH_CMP_0_NE: comp = NE; break;
+    case ATOMIC_OP_FETCH_CMP_0_GT: comp = GT; break;
+    case ATOMIC_OP_FETCH_CMP_0_GE: comp = GE; break;
+    case ATOMIC_OP_FETCH_CMP_0_LT: comp = LT; break;
+    case ATOMIC_OP_FETCH_CMP_0_LE: comp = LE; break;
+    default: gcc_unreachable ();
+    }
+
+  rtx target;
+  if (lhs == NULL_TREE)
+    target = gen_reg_rtx (TYPE_MODE (boolean_type_node));
+  else
+    target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
+  enum insn_code icode = direct_optab_handler (optab, mode);
+  gcc_assert (icode != CODE_FOR_nothing);
+  create_output_operand (&ops[0], target, TYPE_MODE (boolean_type_node));
+  create_fixed_operand (&ops[1], mem);
+  create_convert_operand_to (&ops[2], op, mode, true);
+  create_integer_operand (&ops[3], model);
+  create_integer_operand (&ops[4], comp);
+  if (maybe_expand_insn (icode, 5, ops))
+    return;
+
+  rtx result = expand_atomic_fetch_op (gen_reg_rtx (mode), mem, op,
+                                      code, model, true);
+  if (lhs)
+    {
+      result = emit_store_flag_force (target, comp, result, const0_rtx, mode,
+                                     0, 1);
+      if (result != target)
+       emit_move_insn (target, result);
+    }
+}
+
 /* Expand an atomic clear operation.
        void _atomic_clear (BOOL *obj, enum memmodel)
    EXP is the call expression.  */
index 2d82741..ea10b4b 100644 (file)
@@ -123,6 +123,7 @@ extern void std_expand_builtin_va_start (tree, rtx);
 extern void expand_builtin_trap (void);
 extern void expand_ifn_atomic_bit_test_and (gcall *);
 extern void expand_ifn_atomic_compare_exchange (gcall *);
+extern void expand_ifn_atomic_op_fetch_cmp_0 (gcall *);
 extern rtx expand_builtin (tree, rtx, rtx, machine_mode, int);
 extern enum built_in_function builtin_mathfn_code (const_tree);
 extern tree fold_builtin_expect (location_t, tree, tree, tree, tree);
index f87fa5e..36417c5 100644 (file)
        (const_int 0))]
   ""
   "lock{%;} %K2btr{<imodesuffix>}\t{%1, %0|%0, %1}")
+
+(define_expand "atomic_<plusminus_mnemonic>_fetch_cmp_0<mode>"
+  [(match_operand:QI 0 "register_operand")
+   (plusminus:SWI (match_operand:SWI 1 "memory_operand")
+                 (match_operand:SWI 2 "nonmemory_operand"))
+   (match_operand:SI 3 "const_int_operand") ;; model
+   (match_operand:SI 4 "const_int_operand")]
+  ""
+{
+  if (INTVAL (operands[4]) == GT || INTVAL (operands[4]) == LE)
+    FAIL;
+  emit_insn (gen_atomic_<plusminus_mnemonic>_fetch_cmp_0<mode>_1 (operands[1],
+                                                                 operands[2],
+                                                                 operands[3]));
+  ix86_expand_setcc (operands[0], (enum rtx_code) INTVAL (operands[4]),
+                    gen_rtx_REG (CCGOCmode, FLAGS_REG), const0_rtx);
+  DONE;
+})
+
+(define_insn "atomic_add_fetch_cmp_0<mode>_1"
+  [(set (reg:CCGOC FLAGS_REG)
+       (compare:CCGOC
+         (plus:SWI
+           (unspec_volatile:SWI
+             [(match_operand:SWI 0 "memory_operand" "+m")
+              (match_operand:SI 2 "const_int_operand")]                ;; model
+             UNSPECV_XCHG)
+           (match_operand:SWI 1 "nonmemory_operand" "<r><i>"))
+         (const_int 0)))
+   (set (match_dup 0)
+       (plus:SWI (match_dup 0) (match_dup 1)))]
+  ""
+{
+  if (incdec_operand (operands[1], <MODE>mode))
+    {
+      if (operands[1] == const1_rtx)
+       return "lock{%;} %K2inc{<imodesuffix>}\t%0";
+      else
+       return "lock{%;} %K2dec{<imodesuffix>}\t%0";
+    }
+
+  if (x86_maybe_negate_const_int (&operands[1], <MODE>mode))
+    return "lock{%;} %K2sub{<imodesuffix>}\t{%1, %0|%0, %1}";
+
+  return "lock{%;} %K2add{<imodesuffix>}\t{%1, %0|%0, %1}";
+})
+
+(define_insn "atomic_sub_fetch_cmp_0<mode>_1"
+  [(set (reg:CCGOC FLAGS_REG)
+       (compare:CCGOC
+         (minus:SWI
+           (unspec_volatile:SWI
+             [(match_operand:SWI 0 "memory_operand" "+m")
+              (match_operand:SI 2 "const_int_operand")]                ;; model
+             UNSPECV_XCHG)
+           (match_operand:SWI 1 "nonmemory_operand" "<r><i>"))
+         (const_int 0)))
+   (set (match_dup 0)
+       (minus:SWI (match_dup 0) (match_dup 1)))]
+  ""
+{
+  if (incdec_operand (operands[1], <MODE>mode))
+    {
+      if (operands[1] != const1_rtx)
+       return "lock{%;} %K2inc{<imodesuffix>}\t%0";
+      else
+       return "lock{%;} %K2dec{<imodesuffix>}\t%0";
+    }
+
+  if (x86_maybe_negate_const_int (&operands[1], <MODE>mode))
+    return "lock{%;} %K2add{<imodesuffix>}\t{%1, %0|%0, %1}";
+
+  return "lock{%;} %K2sub{<imodesuffix>}\t{%1, %0|%0, %1}";
+})
+
+(define_expand "atomic_<logic>_fetch_cmp_0<mode>"
+  [(match_operand:QI 0 "register_operand")
+   (any_logic:SWI (match_operand:SWI 1 "memory_operand")
+                 (match_operand:SWI 2 "nonmemory_operand"))
+   (match_operand:SI 3 "const_int_operand") ;; model
+   (match_operand:SI 4 "const_int_operand")]
+  ""
+{
+  emit_insn (gen_atomic_<logic>_fetch_cmp_0<mode>_1 (operands[1], operands[2],
+                                                    operands[3]));
+  ix86_expand_setcc (operands[0], (enum rtx_code) INTVAL (operands[4]),
+                    gen_rtx_REG (CCNOmode, FLAGS_REG), const0_rtx);
+  DONE;
+})
+
+(define_insn "atomic_<logic>_fetch_cmp_0<mode>_1"
+  [(set (reg:CCNO FLAGS_REG)
+       (compare:CCNO
+         (any_logic:SWI
+           (unspec_volatile:SWI
+             [(match_operand:SWI 0 "memory_operand" "+m")
+              (match_operand:SI 2 "const_int_operand")]                ;; model
+             UNSPECV_XCHG)
+           (match_operand:SWI 1 "nonmemory_operand" "<r><i>"))
+         (const_int 0)))
+   (set (match_dup 0)
+       (any_logic:SWI (match_dup 0) (match_dup 1)))]
+  ""
+  "lock{%;} %K2<logic>{<imodesuffix>}\t{%1, %0|%0, %1}")
index fc7dc28..19e89ae 100644 (file)
@@ -7918,6 +7918,30 @@ If these patterns are not defined, attempts will be made to use
 counterparts.  If none of these are available a compare-and-swap
 loop will be used.
 
+@cindex @code{atomic_add_fetch_cmp_0@var{mode}} instruction pattern
+@cindex @code{atomic_sub_fetch_cmp_0@var{mode}} instruction pattern
+@cindex @code{atomic_and_fetch_cmp_0@var{mode}} instruction pattern
+@cindex @code{atomic_or_fetch_cmp_0@var{mode}} instruction pattern
+@cindex @code{atomic_xor_fetch_cmp_0@var{mode}} instruction pattern
+@item @samp{atomic_add_fetch_cmp_0@var{mode}}
+@itemx @samp{atomic_sub_fetch_cmp_0@var{mode}}
+@itemx @samp{atomic_and_fetch_cmp_0@var{mode}}
+@itemx @samp{atomic_or_fetch_cmp_0@var{mode}}
+@itemx @samp{atomic_xor_fetch_cmp_0@var{mode}}
+These patterns emit code for an atomic operation on memory with memory
+model semantics if the fetch result is used only in a comparison against
+zero.
+Operand 0 is an output operand which contains a boolean result of comparison
+of the value after the operation against zero.  Operand 1 is the memory on
+which the atomic operation is performed.  Operand 2 is the second operand
+to the binary operator.  Operand 3 is the memory model to be used by the
+operation.  Operand 4 is an integer holding the comparison code, one of
+@code{EQ}, @code{NE}, @code{LT}, @code{GT}, @code{LE} or @code{GE}.
+
+If these patterns are not defined, attempts will be made to use separate
+atomic operation and fetch pattern followed by comparison of the result
+against zero.
+
 @cindex @code{mem_thread_fence} instruction pattern
 @item @samp{mem_thread_fence}
 This pattern emits code required to implement a thread fence with
index afd4fd0..b24102a 100644 (file)
@@ -3238,6 +3238,46 @@ expand_ATOMIC_COMPARE_EXCHANGE (internal_fn, gcall *call)
   expand_ifn_atomic_compare_exchange (call);
 }
 
+/* Expand atomic add fetch and cmp with 0.  */
+
+static void
+expand_ATOMIC_ADD_FETCH_CMP_0 (internal_fn, gcall *call)
+{
+  expand_ifn_atomic_op_fetch_cmp_0 (call);
+}
+
+/* Expand atomic sub fetch and cmp with 0.  */
+
+static void
+expand_ATOMIC_SUB_FETCH_CMP_0 (internal_fn, gcall *call)
+{
+  expand_ifn_atomic_op_fetch_cmp_0 (call);
+}
+
+/* Expand atomic and fetch and cmp with 0.  */
+
+static void
+expand_ATOMIC_AND_FETCH_CMP_0 (internal_fn, gcall *call)
+{
+  expand_ifn_atomic_op_fetch_cmp_0 (call);
+}
+
+/* Expand atomic or fetch and cmp with 0.  */
+
+static void
+expand_ATOMIC_OR_FETCH_CMP_0 (internal_fn, gcall *call)
+{
+  expand_ifn_atomic_op_fetch_cmp_0 (call);
+}
+
+/* Expand atomic xor fetch and cmp with 0.  */
+
+static void
+expand_ATOMIC_XOR_FETCH_CMP_0 (internal_fn, gcall *call)
+{
+  expand_ifn_atomic_op_fetch_cmp_0 (call);
+}
+
 /* Expand LAUNDER to assignment, lhs = arg0.  */
 
 static void
index 68d9632..8891071 100644 (file)
@@ -403,6 +403,11 @@ DEF_INTERNAL_FN (ATOMIC_BIT_TEST_AND_SET, ECF_LEAF, NULL)
 DEF_INTERNAL_FN (ATOMIC_BIT_TEST_AND_COMPLEMENT, ECF_LEAF, NULL)
 DEF_INTERNAL_FN (ATOMIC_BIT_TEST_AND_RESET, ECF_LEAF, NULL)
 DEF_INTERNAL_FN (ATOMIC_COMPARE_EXCHANGE, ECF_LEAF, NULL)
+DEF_INTERNAL_FN (ATOMIC_ADD_FETCH_CMP_0, ECF_LEAF, NULL)
+DEF_INTERNAL_FN (ATOMIC_SUB_FETCH_CMP_0, ECF_LEAF, NULL)
+DEF_INTERNAL_FN (ATOMIC_AND_FETCH_CMP_0, ECF_LEAF, NULL)
+DEF_INTERNAL_FN (ATOMIC_OR_FETCH_CMP_0, ECF_LEAF, NULL)
+DEF_INTERNAL_FN (ATOMIC_XOR_FETCH_CMP_0, ECF_LEAF, NULL)
 
 /* To implement [[fallthrough]].  */
 DEF_INTERNAL_FN (FALLTHROUGH, ECF_LEAF | ECF_NOTHROW, NULL)
index 1d28c22..743888b 100644 (file)
@@ -240,4 +240,13 @@ extern void expand_SHUFFLEVECTOR (internal_fn, gcall *);
 
 extern bool vectorized_internal_fn_supported_p (internal_fn, tree);
 
+enum {
+  ATOMIC_OP_FETCH_CMP_0_EQ = 0,
+  ATOMIC_OP_FETCH_CMP_0_NE = 1,
+  ATOMIC_OP_FETCH_CMP_0_LT = 2,
+  ATOMIC_OP_FETCH_CMP_0_LE = 3,
+  ATOMIC_OP_FETCH_CMP_0_GT = 4,
+  ATOMIC_OP_FETCH_CMP_0_GE = 5
+};
+
 #endif
index f1c4a10..5fcf538 100644 (file)
@@ -451,6 +451,11 @@ OPTAB_D (atomic_sub_fetch_optab, "atomic_sub_fetch$I$a")
 OPTAB_D (atomic_sub_optab, "atomic_sub$I$a")
 OPTAB_D (atomic_xor_fetch_optab, "atomic_xor_fetch$I$a")
 OPTAB_D (atomic_xor_optab, "atomic_xor$I$a")
+OPTAB_D (atomic_add_fetch_cmp_0_optab, "atomic_add_fetch_cmp_0$I$a")
+OPTAB_D (atomic_sub_fetch_cmp_0_optab, "atomic_sub_fetch_cmp_0$I$a")
+OPTAB_D (atomic_and_fetch_cmp_0_optab, "atomic_and_fetch_cmp_0$I$a")
+OPTAB_D (atomic_or_fetch_cmp_0_optab, "atomic_or_fetch_cmp_0$I$a")
+OPTAB_D (atomic_xor_fetch_cmp_0_optab, "atomic_xor_fetch_cmp_0$I$a")
 
 OPTAB_D (get_thread_pointer_optab, "get_thread_pointer$I$a")
 OPTAB_D (set_thread_pointer_optab, "set_thread_pointer$I$a")
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-1.c b/gcc/testsuite/gcc.target/i386/pr98737-1.c
new file mode 100644 (file)
index 0000000..33c84da
--- /dev/null
@@ -0,0 +1,207 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*subq\t" { target lp64 } } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*subl\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*subw\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*subb\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_sub_fetch (&a, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_sub_fetch (&b, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_sub_fetch (&c, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_sub_fetch (&d, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_sub_fetch (&a, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_sub_fetch (&b, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_sub_fetch (&c, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_sub_fetch (&d, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f9 (long x)
+{
+  return __atomic_sub_fetch (&a, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f10 (int x)
+{
+  return __atomic_sub_fetch (&b, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f11 (short x)
+{
+  return __atomic_sub_fetch (&c, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f12 (char x)
+{
+  return __atomic_sub_fetch (&d, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f13 (long x)
+{
+  return __atomic_sub_fetch (&a, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f14 (int x)
+{
+  return __atomic_sub_fetch (&b, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f15 (short x)
+{
+  return __atomic_sub_fetch (&c, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f16 (char x)
+{
+  return __atomic_sub_fetch (&d, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f17 (long x)
+{
+  return __sync_sub_and_fetch (&a, x) == 0;
+}
+
+int
+f18 (int x)
+{
+  return __sync_sub_and_fetch (&b, x) == 0;
+}
+
+int
+f19 (short x)
+{
+  return __sync_sub_and_fetch (&c, x) == 0;
+}
+
+int
+f20 (char x)
+{
+  return __sync_sub_and_fetch (&d, x) == 0;
+}
+
+int
+f21 (long x)
+{
+  return __sync_sub_and_fetch (&a, x) != 0;
+}
+
+int
+f22 (int x)
+{
+  return __sync_sub_and_fetch (&b, x) != 0;
+}
+
+int
+f23 (short x)
+{
+  return __sync_sub_and_fetch (&c, x) != 0;
+}
+
+int
+f24 (char x)
+{
+  return __sync_sub_and_fetch (&d, x) != 0;
+}
+
+int
+f25 (long x)
+{
+  return __sync_sub_and_fetch (&a, x) < 0;
+}
+
+int
+f26 (int x)
+{
+  return __sync_sub_and_fetch (&b, x) < 0;
+}
+
+int
+f27 (short x)
+{
+  return __sync_sub_and_fetch (&c, x) < 0;
+}
+
+int
+f28 (char x)
+{
+  return __sync_sub_and_fetch (&d, x) < 0;
+}
+
+int
+f29 (long x)
+{
+  return __sync_sub_and_fetch (&a, x) >= 0;
+}
+
+int
+f30 (int x)
+{
+  return __sync_sub_and_fetch (&b, x) >= 0;
+}
+
+int
+f31 (short x)
+{
+  return __sync_sub_and_fetch (&c, x) >= 0;
+}
+
+int
+f32 (char x)
+{
+  return __sync_sub_and_fetch (&d, x) >= 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-2.c b/gcc/testsuite/gcc.target/i386/pr98737-2.c
new file mode 100644 (file)
index 0000000..53b674e
--- /dev/null
@@ -0,0 +1,111 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*subq\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*subl\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*subw\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*subb\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_sub_fetch (&a, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_sub_fetch (&b, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_sub_fetch (&c, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_sub_fetch (&d, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_sub_fetch (&a, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_sub_fetch (&b, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_sub_fetch (&c, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_sub_fetch (&d, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f9 (long x)
+{
+  return __sync_sub_and_fetch (&a, x) <= 0;
+}
+
+int
+f10 (int x)
+{
+  return __sync_sub_and_fetch (&b, x) <= 0;
+}
+
+int
+f11 (short x)
+{
+  return __sync_sub_and_fetch (&c, x) <= 0;
+}
+
+int
+f12 (char x)
+{
+  return __sync_sub_and_fetch (&d, x) <= 0;
+}
+
+int
+f13 (long x)
+{
+  return __sync_sub_and_fetch (&a, x) > 0;
+}
+
+int
+f14 (int x)
+{
+  return __sync_sub_and_fetch (&b, x) > 0;
+}
+
+int
+f15 (short x)
+{
+  return __sync_sub_and_fetch (&c, x) > 0;
+}
+
+int
+f16 (char x)
+{
+  return __sync_sub_and_fetch (&d, x) > 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-3.c b/gcc/testsuite/gcc.target/i386/pr98737-3.c
new file mode 100644 (file)
index 0000000..0e7108a
--- /dev/null
@@ -0,0 +1,207 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*addq\t" { target lp64 } } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*addl\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*addw\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*addb\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_add_fetch (&a, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_add_fetch (&b, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_add_fetch (&c, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_add_fetch (&d, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_add_fetch (&a, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_add_fetch (&b, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_add_fetch (&c, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_add_fetch (&d, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f9 (long x)
+{
+  return __atomic_add_fetch (&a, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f10 (int x)
+{
+  return __atomic_add_fetch (&b, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f11 (short x)
+{
+  return __atomic_add_fetch (&c, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f12 (char x)
+{
+  return __atomic_add_fetch (&d, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f13 (long x)
+{
+  return __atomic_add_fetch (&a, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f14 (int x)
+{
+  return __atomic_add_fetch (&b, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f15 (short x)
+{
+  return __atomic_add_fetch (&c, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f16 (char x)
+{
+  return __atomic_add_fetch (&d, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f17 (long x)
+{
+  return __sync_add_and_fetch (&a, x) == 0;
+}
+
+int
+f18 (int x)
+{
+  return __sync_add_and_fetch (&b, x) == 0;
+}
+
+int
+f19 (short x)
+{
+  return __sync_add_and_fetch (&c, x) == 0;
+}
+
+int
+f20 (char x)
+{
+  return __sync_add_and_fetch (&d, x) == 0;
+}
+
+int
+f21 (long x)
+{
+  return __sync_add_and_fetch (&a, x) != 0;
+}
+
+int
+f22 (int x)
+{
+  return __sync_add_and_fetch (&b, x) != 0;
+}
+
+int
+f23 (short x)
+{
+  return __sync_add_and_fetch (&c, x) != 0;
+}
+
+int
+f24 (char x)
+{
+  return __sync_add_and_fetch (&d, x) != 0;
+}
+
+int
+f25 (long x)
+{
+  return __sync_add_and_fetch (&a, x) < 0;
+}
+
+int
+f26 (int x)
+{
+  return __sync_add_and_fetch (&b, x) < 0;
+}
+
+int
+f27 (short x)
+{
+  return __sync_add_and_fetch (&c, x) < 0;
+}
+
+int
+f28 (char x)
+{
+  return __sync_add_and_fetch (&d, x) < 0;
+}
+
+int
+f29 (long x)
+{
+  return __sync_add_and_fetch (&a, x) >= 0;
+}
+
+int
+f30 (int x)
+{
+  return __sync_add_and_fetch (&b, x) >= 0;
+}
+
+int
+f31 (short x)
+{
+  return __sync_add_and_fetch (&c, x) >= 0;
+}
+
+int
+f32 (char x)
+{
+  return __sync_add_and_fetch (&d, x) >= 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-4.c b/gcc/testsuite/gcc.target/i386/pr98737-4.c
new file mode 100644 (file)
index 0000000..8228d52
--- /dev/null
@@ -0,0 +1,111 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\rx]\*addq\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\rx]\*addl\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\rx]\*addw\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\rx]\*addb\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_add_fetch (&a, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_add_fetch (&b, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_add_fetch (&c, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_add_fetch (&d, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_add_fetch (&a, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_add_fetch (&b, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_add_fetch (&c, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_add_fetch (&d, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f9 (long x)
+{
+  return __sync_add_and_fetch (&a, x) <= 0;
+}
+
+int
+f10 (int x)
+{
+  return __sync_add_and_fetch (&b, x) <= 0;
+}
+
+int
+f11 (short x)
+{
+  return __sync_add_and_fetch (&c, x) <= 0;
+}
+
+int
+f12 (char x)
+{
+  return __sync_add_and_fetch (&d, x) <= 0;
+}
+
+int
+f13 (long x)
+{
+  return __sync_add_and_fetch (&a, x) > 0;
+}
+
+int
+f14 (int x)
+{
+  return __sync_add_and_fetch (&b, x) > 0;
+}
+
+int
+f15 (short x)
+{
+  return __sync_add_and_fetch (&c, x) > 0;
+}
+
+int
+f16 (char x)
+{
+  return __sync_add_and_fetch (&d, x) > 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-5.c b/gcc/testsuite/gcc.target/i386/pr98737-5.c
new file mode 100644 (file)
index 0000000..6d3e063
--- /dev/null
@@ -0,0 +1,303 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*andq\t" { target lp64 } } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*andl\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*andw\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*andb\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_and_fetch (&a, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_and_fetch (&b, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_and_fetch (&c, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_and_fetch (&d, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_and_fetch (&a, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_and_fetch (&b, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_and_fetch (&c, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_and_fetch (&d, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f9 (long x)
+{
+  return __atomic_and_fetch (&a, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f10 (int x)
+{
+  return __atomic_and_fetch (&b, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f11 (short x)
+{
+  return __atomic_and_fetch (&c, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f12 (char x)
+{
+  return __atomic_and_fetch (&d, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f13 (long x)
+{
+  return __atomic_and_fetch (&a, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f14 (int x)
+{
+  return __atomic_and_fetch (&b, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f15 (short x)
+{
+  return __atomic_and_fetch (&c, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f16 (char x)
+{
+  return __atomic_and_fetch (&d, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f17 (long x)
+{
+  return __sync_and_and_fetch (&a, x) == 0;
+}
+
+int
+f18 (int x)
+{
+  return __sync_and_and_fetch (&b, x) == 0;
+}
+
+int
+f19 (short x)
+{
+  return __sync_and_and_fetch (&c, x) == 0;
+}
+
+int
+f20 (char x)
+{
+  return __sync_and_and_fetch (&d, x) == 0;
+}
+
+int
+f21 (long x)
+{
+  return __sync_and_and_fetch (&a, x) != 0;
+}
+
+int
+f22 (int x)
+{
+  return __sync_and_and_fetch (&b, x) != 0;
+}
+
+int
+f23 (short x)
+{
+  return __sync_and_and_fetch (&c, x) != 0;
+}
+
+int
+f24 (char x)
+{
+  return __sync_and_and_fetch (&d, x) != 0;
+}
+
+int
+f25 (long x)
+{
+  return __sync_and_and_fetch (&a, x) < 0;
+}
+
+int
+f26 (int x)
+{
+  return __sync_and_and_fetch (&b, x) < 0;
+}
+
+int
+f27 (short x)
+{
+  return __sync_and_and_fetch (&c, x) < 0;
+}
+
+int
+f28 (char x)
+{
+  return __sync_and_and_fetch (&d, x) < 0;
+}
+
+int
+f29 (long x)
+{
+  return __sync_and_and_fetch (&a, x) >= 0;
+}
+
+int
+f30 (int x)
+{
+  return __sync_and_and_fetch (&b, x) >= 0;
+}
+
+int
+f31 (short x)
+{
+  return __sync_and_and_fetch (&c, x) >= 0;
+}
+
+int
+f32 (char x)
+{
+  return __sync_and_and_fetch (&d, x) >= 0;
+}
+
+int
+f33 (long x)
+{
+  return __atomic_and_fetch (&a, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f34 (int x)
+{
+  return __atomic_and_fetch (&b, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f35 (short x)
+{
+  return __atomic_and_fetch (&c, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f36 (char x)
+{
+  return __atomic_and_fetch (&d, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f37 (long x)
+{
+  return __atomic_and_fetch (&a, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f38 (int x)
+{
+  return __atomic_and_fetch (&b, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f39 (short x)
+{
+  return __atomic_and_fetch (&c, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f40 (char x)
+{
+  return __atomic_and_fetch (&d, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f41 (long x)
+{
+  return __sync_and_and_fetch (&a, x) <= 0;
+}
+
+int
+f42 (int x)
+{
+  return __sync_and_and_fetch (&b, x) <= 0;
+}
+
+int
+f43 (short x)
+{
+  return __sync_and_and_fetch (&c, x) <= 0;
+}
+
+int
+f44 (char x)
+{
+  return __sync_and_and_fetch (&d, x) <= 0;
+}
+
+int
+f45 (long x)
+{
+  return __sync_and_and_fetch (&a, x) > 0;
+}
+
+int
+f46 (int x)
+{
+  return __sync_and_and_fetch (&b, x) > 0;
+}
+
+int
+f47 (short x)
+{
+  return __sync_and_and_fetch (&c, x) > 0;
+}
+
+int
+f48 (char x)
+{
+  return __sync_and_and_fetch (&d, x) > 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-6.c b/gcc/testsuite/gcc.target/i386/pr98737-6.c
new file mode 100644 (file)
index 0000000..6cc1c3a
--- /dev/null
@@ -0,0 +1,303 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*orq\t" { target lp64 } } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*orl\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*orw\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*orb\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_or_fetch (&a, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_or_fetch (&b, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_or_fetch (&c, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_or_fetch (&d, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_or_fetch (&a, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_or_fetch (&b, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_or_fetch (&c, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_or_fetch (&d, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f9 (long x)
+{
+  return __atomic_or_fetch (&a, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f10 (int x)
+{
+  return __atomic_or_fetch (&b, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f11 (short x)
+{
+  return __atomic_or_fetch (&c, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f12 (char x)
+{
+  return __atomic_or_fetch (&d, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f13 (long x)
+{
+  return __atomic_or_fetch (&a, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f14 (int x)
+{
+  return __atomic_or_fetch (&b, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f15 (short x)
+{
+  return __atomic_or_fetch (&c, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f16 (char x)
+{
+  return __atomic_or_fetch (&d, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f17 (long x)
+{
+  return __sync_or_and_fetch (&a, x) == 0;
+}
+
+int
+f18 (int x)
+{
+  return __sync_or_and_fetch (&b, x) == 0;
+}
+
+int
+f19 (short x)
+{
+  return __sync_or_and_fetch (&c, x) == 0;
+}
+
+int
+f20 (char x)
+{
+  return __sync_or_and_fetch (&d, x) == 0;
+}
+
+int
+f21 (long x)
+{
+  return __sync_or_and_fetch (&a, x) != 0;
+}
+
+int
+f22 (int x)
+{
+  return __sync_or_and_fetch (&b, x) != 0;
+}
+
+int
+f23 (short x)
+{
+  return __sync_or_and_fetch (&c, x) != 0;
+}
+
+int
+f24 (char x)
+{
+  return __sync_or_and_fetch (&d, x) != 0;
+}
+
+int
+f25 (long x)
+{
+  return __sync_or_and_fetch (&a, x) < 0;
+}
+
+int
+f26 (int x)
+{
+  return __sync_or_and_fetch (&b, x) < 0;
+}
+
+int
+f27 (short x)
+{
+  return __sync_or_and_fetch (&c, x) < 0;
+}
+
+int
+f28 (char x)
+{
+  return __sync_or_and_fetch (&d, x) < 0;
+}
+
+int
+f29 (long x)
+{
+  return __sync_or_and_fetch (&a, x) >= 0;
+}
+
+int
+f30 (int x)
+{
+  return __sync_or_and_fetch (&b, x) >= 0;
+}
+
+int
+f31 (short x)
+{
+  return __sync_or_and_fetch (&c, x) >= 0;
+}
+
+int
+f32 (char x)
+{
+  return __sync_or_and_fetch (&d, x) >= 0;
+}
+
+int
+f33 (long x)
+{
+  return __atomic_or_fetch (&a, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f34 (int x)
+{
+  return __atomic_or_fetch (&b, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f35 (short x)
+{
+  return __atomic_or_fetch (&c, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f36 (char x)
+{
+  return __atomic_or_fetch (&d, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f37 (long x)
+{
+  return __atomic_or_fetch (&a, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f38 (int x)
+{
+  return __atomic_or_fetch (&b, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f39 (short x)
+{
+  return __atomic_or_fetch (&c, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f40 (char x)
+{
+  return __atomic_or_fetch (&d, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f41 (long x)
+{
+  return __sync_or_and_fetch (&a, x) <= 0;
+}
+
+int
+f42 (int x)
+{
+  return __sync_or_and_fetch (&b, x) <= 0;
+}
+
+int
+f43 (short x)
+{
+  return __sync_or_and_fetch (&c, x) <= 0;
+}
+
+int
+f44 (char x)
+{
+  return __sync_or_and_fetch (&d, x) <= 0;
+}
+
+int
+f45 (long x)
+{
+  return __sync_or_and_fetch (&a, x) > 0;
+}
+
+int
+f46 (int x)
+{
+  return __sync_or_and_fetch (&b, x) > 0;
+}
+
+int
+f47 (short x)
+{
+  return __sync_or_and_fetch (&c, x) > 0;
+}
+
+int
+f48 (char x)
+{
+  return __sync_or_and_fetch (&d, x) > 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/pr98737-7.c b/gcc/testsuite/gcc.target/i386/pr98737-7.c
new file mode 100644 (file)
index 0000000..2da23c4
--- /dev/null
@@ -0,0 +1,303 @@
+/* PR target/98737 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -masm=att" } */
+/* { dg-additional-options "-march=i686" { target ia32 } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*xorq\t" { target lp64 } } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*xorl\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*xorw\t" } } */
+/* { dg-final { scan-assembler "lock\[^\n\r]\*xorb\t" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*xadd" } } */
+/* { dg-final { scan-assembler-not "lock\[^\n\r]\*cmpxchg" } } */
+
+long a;
+int b;
+short c;
+char d;
+
+int
+f1 (long x)
+{
+  return __atomic_xor_fetch (&a, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f2 (int x)
+{
+  return __atomic_xor_fetch (&b, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f3 (short x)
+{
+  return __atomic_xor_fetch (&c, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f4 (char x)
+{
+  return __atomic_xor_fetch (&d, x, __ATOMIC_RELEASE) == 0;
+}
+
+int
+f5 (long x)
+{
+  return __atomic_xor_fetch (&a, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f6 (int x)
+{
+  return __atomic_xor_fetch (&b, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f7 (short x)
+{
+  return __atomic_xor_fetch (&c, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f8 (char x)
+{
+  return __atomic_xor_fetch (&d, x, __ATOMIC_RELEASE) != 0;
+}
+
+int
+f9 (long x)
+{
+  return __atomic_xor_fetch (&a, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f10 (int x)
+{
+  return __atomic_xor_fetch (&b, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f11 (short x)
+{
+  return __atomic_xor_fetch (&c, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f12 (char x)
+{
+  return __atomic_xor_fetch (&d, x, __ATOMIC_RELEASE) < 0;
+}
+
+int
+f13 (long x)
+{
+  return __atomic_xor_fetch (&a, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f14 (int x)
+{
+  return __atomic_xor_fetch (&b, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f15 (short x)
+{
+  return __atomic_xor_fetch (&c, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f16 (char x)
+{
+  return __atomic_xor_fetch (&d, x, __ATOMIC_RELEASE) >= 0;
+}
+
+int
+f17 (long x)
+{
+  return __sync_xor_and_fetch (&a, x) == 0;
+}
+
+int
+f18 (int x)
+{
+  return __sync_xor_and_fetch (&b, x) == 0;
+}
+
+int
+f19 (short x)
+{
+  return __sync_xor_and_fetch (&c, x) == 0;
+}
+
+int
+f20 (char x)
+{
+  return __sync_xor_and_fetch (&d, x) == 0;
+}
+
+int
+f21 (long x)
+{
+  return __sync_xor_and_fetch (&a, x) != 0;
+}
+
+int
+f22 (int x)
+{
+  return __sync_xor_and_fetch (&b, x) != 0;
+}
+
+int
+f23 (short x)
+{
+  return __sync_xor_and_fetch (&c, x) != 0;
+}
+
+int
+f24 (char x)
+{
+  return __sync_xor_and_fetch (&d, x) != 0;
+}
+
+int
+f25 (long x)
+{
+  return __sync_xor_and_fetch (&a, x) < 0;
+}
+
+int
+f26 (int x)
+{
+  return __sync_xor_and_fetch (&b, x) < 0;
+}
+
+int
+f27 (short x)
+{
+  return __sync_xor_and_fetch (&c, x) < 0;
+}
+
+int
+f28 (char x)
+{
+  return __sync_xor_and_fetch (&d, x) < 0;
+}
+
+int
+f29 (long x)
+{
+  return __sync_xor_and_fetch (&a, x) >= 0;
+}
+
+int
+f30 (int x)
+{
+  return __sync_xor_and_fetch (&b, x) >= 0;
+}
+
+int
+f31 (short x)
+{
+  return __sync_xor_and_fetch (&c, x) >= 0;
+}
+
+int
+f32 (char x)
+{
+  return __sync_xor_and_fetch (&d, x) >= 0;
+}
+
+int
+f33 (long x)
+{
+  return __atomic_xor_fetch (&a, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f34 (int x)
+{
+  return __atomic_xor_fetch (&b, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f35 (short x)
+{
+  return __atomic_xor_fetch (&c, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f36 (char x)
+{
+  return __atomic_xor_fetch (&d, x, __ATOMIC_RELEASE) <= 0;
+}
+
+int
+f37 (long x)
+{
+  return __atomic_xor_fetch (&a, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f38 (int x)
+{
+  return __atomic_xor_fetch (&b, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f39 (short x)
+{
+  return __atomic_xor_fetch (&c, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f40 (char x)
+{
+  return __atomic_xor_fetch (&d, x, __ATOMIC_RELEASE) > 0;
+}
+
+int
+f41 (long x)
+{
+  return __sync_xor_and_fetch (&a, x) <= 0;
+}
+
+int
+f42 (int x)
+{
+  return __sync_xor_and_fetch (&b, x) <= 0;
+}
+
+int
+f43 (short x)
+{
+  return __sync_xor_and_fetch (&c, x) <= 0;
+}
+
+int
+f44 (char x)
+{
+  return __sync_xor_and_fetch (&d, x) <= 0;
+}
+
+int
+f45 (long x)
+{
+  return __sync_xor_and_fetch (&a, x) > 0;
+}
+
+int
+f46 (int x)
+{
+  return __sync_xor_and_fetch (&b, x) > 0;
+}
+
+int
+f47 (short x)
+{
+  return __sync_xor_and_fetch (&c, x) > 0;
+}
+
+int
+f48 (char x)
+{
+  return __sync_xor_and_fetch (&d, x) > 0;
+}
index bd3f2bc..93fa536 100644 (file)
@@ -151,6 +151,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "symbol-summary.h"
 #include "ipa-utils.h"
 #include "ipa-prop.h"
+#include "internal-fn.h"
 
 /* Possible lattice values.  */
 typedef enum
@@ -3333,7 +3334,7 @@ extern bool gimple_nop_convert (tree, tree*, tree (*) (tree));
      _4 = __atomic_fetch_or_* (ptr_6, mask_2, _3);
      _5 = _4 & mask_2;
    to
-     _4 = ATOMIC_BIT_TEST_AND_SET (ptr_6, cnt_1, 0, _3);
+     _4 = .ATOMIC_BIT_TEST_AND_SET (ptr_6, cnt_1, 0, _3);
      _5 = _4;
    If _5 is only used in _5 != 0 or _5 == 0 comparisons, 1
    is passed instead of 0, and the builtin just returns a zero
@@ -3345,7 +3346,7 @@ extern bool gimple_nop_convert (tree, tree*, tree (*) (tree));
    the second argument to the builtin needs to be one's complement
    of the mask instead of mask.  */
 
-static void
+static bool
 optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
                              enum internal_fn fn, bool has_model_arg,
                              bool after)
@@ -3365,7 +3366,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
       || !single_imm_use (lhs, &use_p, &use_stmt)
       || !is_gimple_assign (use_stmt)
       || !gimple_vdef (call))
-    return;
+    return false;
 
   switch (fn)
     {
@@ -3379,7 +3380,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
       optab = atomic_bit_test_and_reset_optab;
       break;
     default:
-      return;
+      return false;
     }
 
   tree bit = nullptr;
@@ -3389,20 +3390,20 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
   if (rhs_code != BIT_AND_EXPR)
     {
       if (rhs_code != NOP_EXPR && rhs_code != BIT_NOT_EXPR)
-       return;
+       return false;
 
       tree use_lhs = gimple_assign_lhs (use_stmt);
       if (TREE_CODE (use_lhs) == SSA_NAME
          && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use_lhs))
-       return;
+       return false;
 
       tree use_rhs = gimple_assign_rhs1 (use_stmt);
       if (lhs != use_rhs)
-       return;
+       return false;
 
       if (optab_handler (optab, TYPE_MODE (TREE_TYPE (lhs)))
          == CODE_FOR_nothing)
-       return;
+       return false;
 
       gimple *g;
       gimple_stmt_iterator gsi;
@@ -3413,7 +3414,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
        {
          g = convert_atomic_bit_not (fn, use_stmt, lhs, mask);
          if (!g)
-           return;
+           return false;
          use_stmt = g;
          ibit = 0;
        }
@@ -3426,7 +3427,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
              if (!operand_equal_p (build_int_cst (TREE_TYPE (lhs),
                                                   ~HOST_WIDE_INT_1),
                                    mask, 0))
-               return;
+               return false;
 
              /* Convert
                 _1 = __atomic_fetch_and_* (ptr_6, ~1, _3);
@@ -3442,7 +3443,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
            {
              and_mask = build_int_cst (TREE_TYPE (lhs), 1);
              if (!operand_equal_p (and_mask, mask, 0))
-               return;
+               return false;
 
              /* Convert
                 _1 = __atomic_fetch_or_* (ptr_6, 1, _3);
@@ -3468,20 +3469,20 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
          gimple *use_nop_stmt;
          if (!single_imm_use (use_lhs, &use_p, &use_nop_stmt)
              || !is_gimple_assign (use_nop_stmt))
-           return;
+           return false;
          tree use_nop_lhs = gimple_assign_lhs (use_nop_stmt);
          rhs_code = gimple_assign_rhs_code (use_nop_stmt);
          if (rhs_code != BIT_AND_EXPR)
            {
              if (TREE_CODE (use_nop_lhs) == SSA_NAME
                  && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use_nop_lhs))
-               return;
+               return false;
              if (rhs_code == BIT_NOT_EXPR)
                {
                  g = convert_atomic_bit_not (fn, use_nop_stmt, lhs,
                                              mask);
                  if (!g)
-                   return;
+                   return false;
                  /* Convert
                     _1 = __atomic_fetch_or_4 (ptr_6, 1, _3);
                     _2 = (int) _1;
@@ -3509,15 +3510,15 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
              else
                {
                  if (TREE_CODE (TREE_TYPE (use_nop_lhs)) != BOOLEAN_TYPE)
-                   return;
+                   return false;
                  if (rhs_code != GE_EXPR && rhs_code != LT_EXPR)
-                   return;
+                   return false;
                  tree cmp_rhs1 = gimple_assign_rhs1 (use_nop_stmt);
                  if (use_lhs != cmp_rhs1)
-                   return;
+                   return false;
                  tree cmp_rhs2 = gimple_assign_rhs2 (use_nop_stmt);
                  if (!integer_zerop (cmp_rhs2))
-                   return;
+                   return false;
 
                  tree and_mask;
 
@@ -3533,7 +3534,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
                      and_mask = build_int_cst (TREE_TYPE (use_rhs),
                                                highest - 1);
                      if (!operand_equal_p (and_mask, mask, 0))
-                       return;
+                       return false;
 
                      /* Convert
                         _1 = __atomic_fetch_and_4 (ptr_6, 0x7fffffff, _3);
@@ -3553,7 +3554,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
                      and_mask = build_int_cst (TREE_TYPE (use_rhs),
                                                highest);
                      if (!operand_equal_p (and_mask, mask, 0))
-                       return;
+                       return false;
 
                      /* Convert
                         _1 = __atomic_fetch_or_4 (ptr_6, 0x80000000, _3);
@@ -3592,7 +3593,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
                  || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (match_op[2])
                  || !single_imm_use (match_op[2], &use_p, &g)
                  || !is_gimple_assign (g))
-               return;
+               return false;
              mask = match_op[0];
              if (TREE_CODE (match_op[1]) == INTEGER_CST)
                {
@@ -3650,7 +3651,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
            }
        }
       else
-       return;
+       return false;
 
       if (!bit)
        {
@@ -3661,11 +3662,11 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
     }
   else if (optab_handler (optab, TYPE_MODE (TREE_TYPE (lhs)))
           == CODE_FOR_nothing)
-    return;
+    return false;
 
   tree use_lhs = gimple_assign_lhs (use_stmt);
   if (!use_lhs)
-    return;
+    return false;
 
   if (!bit)
     {
@@ -3676,7 +3677,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
          mask = fold_convert (TREE_TYPE (lhs), mask);
          int ibit = tree_log2 (mask);
          if (ibit < 0)
-           return;
+           return false;
          bit = build_int_cst (TREE_TYPE (lhs), ibit);
        }
       else if (TREE_CODE (mask) == SSA_NAME)
@@ -3687,30 +3688,30 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
            {
              mask = match_op;
              if (TREE_CODE (mask) != SSA_NAME)
-               return;
+               return false;
              g = SSA_NAME_DEF_STMT (mask);
            }
          if (!is_gimple_assign (g))
-           return;
+           return false;
 
          if (fn == IFN_ATOMIC_BIT_TEST_AND_RESET)
            {
              if (gimple_assign_rhs_code (g) != BIT_NOT_EXPR)
-               return;
+               return false;
              mask = gimple_assign_rhs1 (g);
              if (TREE_CODE (mask) != SSA_NAME)
-               return;
+               return false;
              g = SSA_NAME_DEF_STMT (mask);
            }
 
          if (!is_gimple_assign (g)
              || gimple_assign_rhs_code (g) != LSHIFT_EXPR
              || !integer_onep (gimple_assign_rhs1 (g)))
-           return;
+           return false;
          bit = gimple_assign_rhs2 (g);
        }
       else
-       return;
+       return false;
 
       tree cmp_mask;
       if (gimple_assign_rhs1 (use_stmt) == lhs)
@@ -3723,7 +3724,7 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
        cmp_mask = match_op;
 
       if (!operand_equal_p (cmp_mask, mask, 0))
-       return;
+       return false;
     }
 
   bool use_bool = true;
@@ -3748,6 +3749,8 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
          case COND_EXPR:
            op1 = gimple_assign_rhs1 (g);
            code = TREE_CODE (op1);
+           if (TREE_CODE_CLASS (code) != tcc_comparison)
+             break;
            op0 = TREE_OPERAND (op1, 0);
            op1 = TREE_OPERAND (op1, 1);
            break;
@@ -3864,6 +3867,196 @@ optimize_atomic_bit_test_and (gimple_stmt_iterator *gsip,
   release_defs (use_stmt);
   gsi_remove (gsip, true);
   release_ssa_name (lhs);
+  return true;
+}
+
+/* Optimize
+     _4 = __atomic_add_fetch_* (ptr_6, arg_2, _3);
+     _5 = _4 == 0;
+   to
+     _4 = .ATOMIC_ADD_FETCH_CMP_0 (EQ_EXPR, ptr_6, arg_2, _3);
+     _5 = _4;
+   Similarly for __sync_add_and_fetch_* (without the ", _3" part
+   in there).  */
+
+static bool
+optimize_atomic_op_fetch_cmp_0 (gimple_stmt_iterator *gsip,
+                               enum internal_fn fn, bool has_model_arg)
+{
+  gimple *call = gsi_stmt (*gsip);
+  tree lhs = gimple_call_lhs (call);
+  use_operand_p use_p;
+  gimple *use_stmt;
+
+  if (!flag_inline_atomics
+      || optimize_debug
+      || !gimple_call_builtin_p (call, BUILT_IN_NORMAL)
+      || !lhs
+      || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs)
+      || !single_imm_use (lhs, &use_p, &use_stmt)
+      || !gimple_vdef (call))
+    return false;
+
+  optab optab;
+  switch (fn)
+    {
+    case IFN_ATOMIC_ADD_FETCH_CMP_0:
+      optab = atomic_add_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_SUB_FETCH_CMP_0:
+      optab = atomic_sub_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_AND_FETCH_CMP_0:
+      optab = atomic_and_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_OR_FETCH_CMP_0:
+      optab = atomic_or_fetch_cmp_0_optab;
+      break;
+    case IFN_ATOMIC_XOR_FETCH_CMP_0:
+      optab = atomic_xor_fetch_cmp_0_optab;
+      break;
+    default:
+      return false;
+    }
+
+  if (optab_handler (optab, TYPE_MODE (TREE_TYPE (lhs)))
+      == CODE_FOR_nothing)
+    return false;
+
+  tree use_lhs = lhs;
+  if (gimple_assign_cast_p (use_stmt))
+    {
+      use_lhs = gimple_assign_lhs (use_stmt);
+      if (!tree_nop_conversion_p (TREE_TYPE (use_lhs), TREE_TYPE (lhs))
+         || (!INTEGRAL_TYPE_P (TREE_TYPE (use_lhs))
+             && !POINTER_TYPE_P (TREE_TYPE (use_lhs)))
+         || SSA_NAME_OCCURS_IN_ABNORMAL_PHI (use_lhs)
+         || !single_imm_use (use_lhs, &use_p, &use_stmt))
+       return false;
+    }
+  enum tree_code code = ERROR_MARK;
+  tree op0 = NULL_TREE, op1 = NULL_TREE;
+  if (is_gimple_assign (use_stmt))
+    switch (gimple_assign_rhs_code (use_stmt))
+      {
+      case COND_EXPR:
+       op1 = gimple_assign_rhs1 (use_stmt);
+       code = TREE_CODE (op1);
+       if (TREE_CODE_CLASS (code) == tcc_comparison)
+         {
+           op0 = TREE_OPERAND (op1, 0);
+           op1 = TREE_OPERAND (op1, 1);
+         }
+       break;
+      default:
+       code = gimple_assign_rhs_code (use_stmt);
+       if (TREE_CODE_CLASS (code) == tcc_comparison)
+         {
+           op0 = gimple_assign_rhs1 (use_stmt);
+           op1 = gimple_assign_rhs2 (use_stmt);
+         }
+       break;
+      }
+  else if (gimple_code (use_stmt) == GIMPLE_COND)
+    {
+      code = gimple_cond_code (use_stmt);
+      op0 = gimple_cond_lhs (use_stmt);
+      op1 = gimple_cond_rhs (use_stmt);
+    }
+
+  switch (code)
+    {
+    case LT_EXPR:
+    case LE_EXPR:
+    case GT_EXPR:
+    case GE_EXPR:
+      if (!INTEGRAL_TYPE_P (TREE_TYPE (use_lhs))
+         || TREE_CODE (TREE_TYPE (use_lhs)) == BOOLEAN_TYPE
+         || TYPE_UNSIGNED (TREE_TYPE (use_lhs)))
+       return false;
+      /* FALLTHRU */
+    case EQ_EXPR:
+    case NE_EXPR:
+      if (op0 == use_lhs && integer_zerop (op1))
+       break;
+      return false;
+    default:
+      return false;
+    }
+
+  int encoded;
+  switch (code)
+    {
+    /* Use special encoding of the operation.  We want to also
+       encode the mode in the first argument and for neither EQ_EXPR
+       etc. nor EQ etc. we can rely it will fit into QImode.  */
+    case EQ_EXPR: encoded = ATOMIC_OP_FETCH_CMP_0_EQ; break;
+    case NE_EXPR: encoded = ATOMIC_OP_FETCH_CMP_0_NE; break;
+    case LT_EXPR: encoded = ATOMIC_OP_FETCH_CMP_0_LT; break;
+    case LE_EXPR: encoded = ATOMIC_OP_FETCH_CMP_0_LE; break;
+    case GT_EXPR: encoded = ATOMIC_OP_FETCH_CMP_0_GT; break;
+    case GE_EXPR: encoded = ATOMIC_OP_FETCH_CMP_0_GE; break;
+    default: gcc_unreachable ();
+    }
+
+  tree new_lhs = make_ssa_name (boolean_type_node);
+  gimple *g;
+  tree flag = build_int_cst (TREE_TYPE (lhs), encoded);
+  if (has_model_arg)
+    g = gimple_build_call_internal (fn, 4, flag,
+                                   gimple_call_arg (call, 0),
+                                   gimple_call_arg (call, 1),
+                                   gimple_call_arg (call, 2));
+  else
+    g = gimple_build_call_internal (fn, 3, flag,
+                                   gimple_call_arg (call, 0),
+                                   gimple_call_arg (call, 1));
+  gimple_call_set_lhs (g, new_lhs);
+  gimple_set_location (g, gimple_location (call));
+  gimple_move_vops (g, call);
+  bool throws = stmt_can_throw_internal (cfun, call);
+  gimple_call_set_nothrow (as_a <gcall *> (g),
+                          gimple_call_nothrow_p (as_a <gcall *> (call)));
+  gimple_stmt_iterator gsi = *gsip;
+  gsi_insert_after (&gsi, g, GSI_SAME_STMT);
+  if (throws)
+    maybe_clean_or_replace_eh_stmt (call, g);
+  if (is_gimple_assign (use_stmt))
+    switch (gimple_assign_rhs_code (use_stmt))
+      {
+      case COND_EXPR:
+       gimple_assign_set_rhs1 (use_stmt, new_lhs);
+       break;
+      default:
+       gsi = gsi_for_stmt (use_stmt);
+       if (tree ulhs = gimple_assign_lhs (use_stmt))
+         if (useless_type_conversion_p (TREE_TYPE (ulhs),
+                                        boolean_type_node))
+           {
+             gimple_assign_set_rhs_with_ops (&gsi, SSA_NAME, new_lhs);
+             break;
+           }
+       gimple_assign_set_rhs_with_ops (&gsi, NOP_EXPR, new_lhs);
+       break;
+      }
+  else if (gimple_code (use_stmt) == GIMPLE_COND)
+    {
+      gcond *use_cond = as_a <gcond *> (use_stmt);
+      gimple_cond_set_code (use_cond, NE_EXPR);
+      gimple_cond_set_lhs (use_cond, new_lhs);
+      gimple_cond_set_rhs (use_cond, boolean_false_node);
+    }
+
+  update_stmt (use_stmt);
+  if (use_lhs != lhs)
+    {
+      gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (use_lhs));
+      gsi_remove (&gsi, true);
+      release_ssa_name (use_lhs);
+    }
+  gsi_remove (gsip, true);
+  release_ssa_name (lhs);
+  return true;
 }
 
 /* Optimize
@@ -4092,6 +4285,44 @@ pass_fold_builtins::execute (function *fun)
                    cfg_changed = true;
                  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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_ADD_FETCH_CMP_0,
+                                                 true);
+                 break;
+               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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_ADD_FETCH_CMP_0,
+                                                 false);
+                 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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_SUB_FETCH_CMP_0,
+                                                 true);
+                 break;
+               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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_SUB_FETCH_CMP_0,
+                                                 false);
+                 break;
+
                case BUILT_IN_ATOMIC_FETCH_OR_1:
                case BUILT_IN_ATOMIC_FETCH_OR_2:
                case BUILT_IN_ATOMIC_FETCH_OR_4:
@@ -4133,16 +4364,24 @@ pass_fold_builtins::execute (function *fun)
                case BUILT_IN_ATOMIC_XOR_FETCH_4:
                case BUILT_IN_ATOMIC_XOR_FETCH_8:
                case BUILT_IN_ATOMIC_XOR_FETCH_16:
-                 optimize_atomic_bit_test_and
-                       (&i, IFN_ATOMIC_BIT_TEST_AND_COMPLEMENT, true, true);
+                 if (optimize_atomic_bit_test_and
+                       (&i, IFN_ATOMIC_BIT_TEST_AND_COMPLEMENT, true, true))
+                   break;
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_XOR_FETCH_CMP_0,
+                                                 true);
                  break;
                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:
-                 optimize_atomic_bit_test_and
-                       (&i, IFN_ATOMIC_BIT_TEST_AND_COMPLEMENT, false, true);
+                 if (optimize_atomic_bit_test_and
+                       (&i, IFN_ATOMIC_BIT_TEST_AND_COMPLEMENT, false, true))
+                   break;
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_XOR_FETCH_CMP_0,
+                                                 false);
                  break;
 
                case BUILT_IN_ATOMIC_FETCH_AND_1:
@@ -4164,6 +4403,44 @@ pass_fold_builtins::execute (function *fun)
                                                false, false);
                  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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_AND_FETCH_CMP_0,
+                                                 true);
+                 break;
+               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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_AND_FETCH_CMP_0,
+                                                 false);
+                 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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_OR_FETCH_CMP_0,
+                                                 true);
+                 break;
+               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:
+                 optimize_atomic_op_fetch_cmp_0 (&i,
+                                                 IFN_ATOMIC_OR_FETCH_CMP_0,
+                                                 false);
+                 break;
+
                case BUILT_IN_MEMCPY:
                  if (gimple_call_builtin_p (stmt, BUILT_IN_NORMAL)
                      && TREE_CODE (gimple_call_arg (stmt, 0)) == ADDR_EXPR