re PR target/48986 (Missed optimization in atomic decrement on x86/x64)
authorJakub Jelinek <jakub@redhat.com>
Tue, 17 May 2011 07:37:59 +0000 (09:37 +0200)
committerJakub Jelinek <jakub@gcc.gnu.org>
Tue, 17 May 2011 07:37:59 +0000 (09:37 +0200)
PR target/48986
* config/i386/sync.md (sync_old_add<mode>): Relax operand 2
predicate to allow CONST_INT.
(*sync_old_add_cmp<mode>): New insn and peephole2 for it.

From-SVN: r173816

gcc/ChangeLog
gcc/config/i386/sync.md

index 7f140b84fab8dc4dc5b8ea11db02344c5bfc38f1..50b035615f8bc931348fc8490efc806f9ebd5b21 100644 (file)
@@ -1,3 +1,10 @@
+2011-05-17  Jakub Jelinek  <jakub@redhat.com>
+
+       PR target/48986
+       * config/i386/sync.md (sync_old_add<mode>): Relax operand 2
+       predicate to allow CONST_INT.
+       (*sync_old_add_cmp<mode>): New insn and peephole2 for it.
+
 2011-05-16  Joseph Myers  <joseph@codesourcery.com>
 
        * opts-common.c (opt_enum_arg_to_value): New.
index 22f36487579cef8053e18076d07c427a3c2e95e3..20378d090bf63e4b1f8c6329b05305e0e3ef909a 100644 (file)
@@ -1,5 +1,5 @@
 ;; GCC machine description for i386 synchronization instructions.
-;; Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010
+;; Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011
 ;; Free Software Foundation, Inc.
 ;;
 ;; This file is part of GCC.
   "!TARGET_64BIT && TARGET_CMPXCHG8B && flag_pic"
   "xchg{l}\t%%ebx, %3\;lock{%;} cmpxchg8b\t%1\;xchg{l}\t%%ebx, %3")
 
+;; For operand 2 nonmemory_operand predicate is used instead of
+;; register_operand to allow combiner to better optimize atomic
+;; additions of constants.
 (define_insn "sync_old_add<mode>"
   [(set (match_operand:SWI 0 "register_operand" "=<r>")
        (unspec_volatile:SWI
          [(match_operand:SWI 1 "memory_operand" "+m")] UNSPECV_XCHG))
    (set (match_dup 1)
        (plus:SWI (match_dup 1)
-                 (match_operand:SWI 2 "register_operand" "0")))
+                 (match_operand:SWI 2 "nonmemory_operand" "0")))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_XADD"
   "lock{%;} xadd{<imodesuffix>}\t{%0, %1|%1, %0}")
 
+;; This peephole2 and following insn optimize
+;; __sync_fetch_and_add (x, -N) == N into just lock {add,sub,inc,dec}
+;; followed by testing of flags instead of lock xadd and comparisons.
+(define_peephole2
+  [(set (match_operand:SWI 0 "register_operand" "")
+       (match_operand:SWI 2 "const_int_operand" ""))
+   (parallel [(set (match_dup 0)
+                  (unspec_volatile:SWI
+                    [(match_operand:SWI 1 "memory_operand" "")] UNSPECV_XCHG))
+             (set (match_dup 1)
+                  (plus:SWI (match_dup 1)
+                            (match_dup 0)))
+             (clobber (reg:CC FLAGS_REG))])
+   (set (reg:CCZ FLAGS_REG)
+       (compare:CCZ (match_dup 0)
+                    (match_operand:SWI 3 "const_int_operand" "")))]
+  "peep2_reg_dead_p (3, operands[0])
+   && (unsigned HOST_WIDE_INT) INTVAL (operands[2])
+      == -(unsigned HOST_WIDE_INT) INTVAL (operands[3])
+   && !reg_overlap_mentioned_p (operands[0], operands[1])"
+  [(parallel [(set (reg:CCZ FLAGS_REG)
+                  (compare:CCZ (unspec_volatile:SWI [(match_dup 1)]
+                                                    UNSPECV_XCHG)
+                               (match_dup 3)))
+             (set (match_dup 1)
+                  (plus:SWI (match_dup 1)
+                            (match_dup 2)))])])
+
+(define_insn "*sync_old_add_cmp<mode>"
+  [(set (reg:CCZ FLAGS_REG)
+       (compare:CCZ (unspec_volatile:SWI
+                      [(match_operand:SWI 0 "memory_operand" "+m")]
+                      UNSPECV_XCHG)
+                    (match_operand:SWI 2 "const_int_operand" "i")))
+   (set (match_dup 0)
+       (plus:SWI (match_dup 0)
+                 (match_operand:SWI 1 "const_int_operand" "i")))]
+  "(unsigned HOST_WIDE_INT) INTVAL (operands[1])
+   == -(unsigned HOST_WIDE_INT) INTVAL (operands[2])"
+{
+  if (TARGET_USE_INCDEC)
+    {
+      if (operands[1] == const1_rtx)
+       return "lock{%;} inc{<imodesuffix>}\t%0";
+      if (operands[1] == constm1_rtx)
+       return "lock{%;} dec{<imodesuffix>}\t%0";
+    }
+
+  if (x86_maybe_negate_const_int (&operands[1], <MODE>mode))
+    return "lock{%;} sub{<imodesuffix>}\t{%1, %0|%0, %1}";
+
+  return "lock{%;} add{<imodesuffix>}\t{%1, %0|%0, %1}";
+})
+
 ;; Recall that xchg implicitly sets LOCK#, so adding it again wastes space.
 (define_insn "sync_lock_test_and_set<mode>"
   [(set (match_operand:SWI 0 "register_operand" "=<r>")